xref: /haiku/src/libs/compat/freebsd_iflib/iflib.c (revision 4c8e85b316c35a9161f5a1c50ad70bc91c83a76f)
1 /*-
2  * Copyright (c) 2014-2018, Matthew Macy <mmacy@mattmacy.io>
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions are met:
7  *
8  *  1. Redistributions of source code must retain the above copyright notice,
9  *     this list of conditions and the following disclaimer.
10  *
11  *  2. Neither the name of Matthew Macy nor the names of its
12  *     contributors may be used to endorse or promote products derived from
13  *     this software without specific prior written permission.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
16  * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18  * ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
19  * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
20  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
21  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
22  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
23  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
24  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
25  * POSSIBILITY OF SUCH DAMAGE.
26  */
27 
28 #include <sys/cdefs.h>
29 #include <stdlib.h>
30 __FBSDID("$FreeBSD$");
31 
32 #ifndef __HAIKU__
33 #include "opt_inet.h"
34 #include "opt_inet6.h"
35 #include "opt_acpi.h"
36 #include "opt_sched.h"
37 #endif
38 
39 #include <sys/param.h>
40 #include <sys/types.h>
41 #include <sys/bus.h>
42 #include <sys/eventhandler.h>
43 #ifndef __HAIKU__
44 #include <sys/jail.h>
45 #endif
46 #include <sys/kernel.h>
47 #include <sys/lock.h>
48 #include <sys/mutex.h>
49 #include <sys/sx.h>
50 #include <sys/module.h>
51 #include <sys/kobj.h>
52 #include <sys/rman.h>
53 #include <sys/sbuf.h>
54 #include <sys/smp.h>
55 #include <sys/socket.h>
56 #include <sys/sockio.h>
57 #include <sys/sysctl.h>
58 #include <sys/syslog.h>
59 #include <sys/taskqueue.h>
60 #include <sys/limits.h>
61 
62 #include <net/if.h>
63 #include <net/if_var.h>
64 #include <net/if_types.h>
65 #include <net/if_media.h>
66 #include <net/bpf.h>
67 #include <net/ethernet.h>
68 #include <net/if_vlan_var.h>
69 #include <net/mp_ring.h>
70 #include <net/vnet.h>
71 #include <net/debugnet.h>
72 
73 #include <netinet/in.h>
74 #ifndef __HAIKU__
75 #include <netinet/in_pcb.h>
76 #include <netinet/tcp_lro.h>
77 #include <netinet/in_systm.h>
78 #endif
79 #include <netinet/if_ether.h>
80 #include <netinet/ip.h>
81 #include <netinet/ip6.h>
82 #include <netinet/tcp.h>
83 #include <netinet/ip_var.h>
84 #ifndef __HAIKU__
85 #include <netinet6/ip6_var.h>
86 #endif
87 
88 #include <machine/bus.h>
89 #ifndef __HAIKU__
90 #include <machine/in_cksum.h>
91 #endif
92 
93 #include <vm/vm.h>
94 #include <vm/pmap.h>
95 
96 #include <dev/led/led.h>
97 #include <dev/pci/pcireg.h>
98 #include <dev/pci/pcivar.h>
99 #ifndef __HAIKU__
100 #include <dev/pci/pci_private.h>
101 #endif
102 
103 #include <net/iflib.h>
104 #include <net/iflib_private.h>
105 
106 #include <ifdi_if.h>
107 #include <device_if.h>
108 
109 #ifdef PCI_IOV
110 #include <dev/pci/pci_iov.h>
111 #endif
112 
113 #include <sys/bitstring.h>
114 
115 /*
116  * enable accounting of every mbuf as it comes in to and goes out of
117  * iflib's software descriptor references
118  */
119 #define MEMORY_LOGGING 0
120 /*
121  * Enable mbuf vectors for compressing long mbuf chains
122  */
123 
124 /*
125  * NB:
126  * - Prefetching in tx cleaning should perhaps be a tunable. The distance ahead
127  *   we prefetch needs to be determined by the time spent in m_free vis a vis
128  *   the cost of a prefetch. This will of course vary based on the workload:
129  *      - NFLX's m_free path is dominated by vm-based M_EXT manipulation which
130  *        is quite expensive, thus suggesting very little prefetch.
131  *      - small packet forwarding which is just returning a single mbuf to
132  *        UMA will typically be very fast vis a vis the cost of a memory
133  *        access.
134  */
135 
136 /*
137  * File organization:
138  *  - private structures
139  *  - iflib private utility functions
140  *  - ifnet functions
141  *  - vlan registry and other exported functions
142  *  - iflib public core functions
143  *
144  *
145  */
146 MALLOC_DEFINE(M_IFLIB, "iflib", "ifnet library");
147 
148 #define	IFLIB_RXEOF_MORE (1U << 0)
149 #define	IFLIB_RXEOF_EMPTY (2U << 0)
150 
151 struct iflib_txq;
152 typedef struct iflib_txq *iflib_txq_t;
153 struct iflib_rxq;
154 typedef struct iflib_rxq *iflib_rxq_t;
155 struct iflib_fl;
156 typedef struct iflib_fl *iflib_fl_t;
157 
158 struct iflib_ctx;
159 
160 static void iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid);
161 static void iflib_timer(void *arg);
162 static void iflib_tqg_detach(if_ctx_t ctx);
163 
164 typedef struct iflib_filter_info {
165 	driver_filter_t *ifi_filter;
166 	void *ifi_filter_arg;
167 	struct grouptask *ifi_task;
168 	void *ifi_ctx;
169 } *iflib_filter_info_t;
170 
171 struct iflib_ctx {
172 	KOBJ_FIELDS;
173 	/*
174 	 * Pointer to hardware driver's softc
175 	 */
176 	void *ifc_softc;
177 	device_t ifc_dev;
178 	if_t ifc_ifp;
179 
180 #ifndef __HAIKU__
181 	cpuset_t ifc_cpus;
182 #endif
183 	if_shared_ctx_t ifc_sctx;
184 	struct if_softc_ctx ifc_softc_ctx;
185 
186 	struct sx ifc_ctx_sx;
187 	struct mtx ifc_state_mtx;
188 
189 	iflib_txq_t ifc_txqs;
190 	iflib_rxq_t ifc_rxqs;
191 	uint32_t ifc_if_flags;
192 	uint32_t ifc_flags;
193 	uint32_t ifc_max_fl_buf_size;
194 	uint32_t ifc_rx_mbuf_sz;
195 
196 	int ifc_link_state;
197 	int ifc_watchdog_events;
198 	struct cdev *ifc_led_dev;
199 	struct resource *ifc_msix_mem;
200 
201 	struct if_irq ifc_legacy_irq;
202 	struct grouptask ifc_admin_task;
203 	struct grouptask ifc_vflr_task;
204 	struct iflib_filter_info ifc_filter_info;
205 	struct ifmedia	ifc_media;
206 	struct ifmedia	*ifc_mediap;
207 
208 	struct sysctl_oid *ifc_sysctl_node;
209 	uint16_t ifc_sysctl_ntxqs;
210 	uint16_t ifc_sysctl_nrxqs;
211 	uint16_t ifc_sysctl_qs_eq_override;
212 	uint16_t ifc_sysctl_rx_budget;
213 	uint16_t ifc_sysctl_tx_abdicate;
214 	uint16_t ifc_sysctl_core_offset;
215 #define	CORE_OFFSET_UNSPECIFIED	0xffff
216 	uint8_t  ifc_sysctl_separate_txrx;
217 	uint8_t  ifc_sysctl_use_logical_cores;
218 	bool	 ifc_cpus_are_physical_cores;
219 
220 	qidx_t ifc_sysctl_ntxds[8];
221 	qidx_t ifc_sysctl_nrxds[8];
222 	struct if_txrx ifc_txrx;
223 #define isc_txd_encap  ifc_txrx.ift_txd_encap
224 #define isc_txd_flush  ifc_txrx.ift_txd_flush
225 #define isc_txd_credits_update  ifc_txrx.ift_txd_credits_update
226 #define isc_rxd_available ifc_txrx.ift_rxd_available
227 #define isc_rxd_pkt_get ifc_txrx.ift_rxd_pkt_get
228 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
229 #define isc_rxd_flush ifc_txrx.ift_rxd_flush
230 #define isc_legacy_intr ifc_txrx.ift_legacy_intr
231 	eventhandler_tag ifc_vlan_attach_event;
232 	eventhandler_tag ifc_vlan_detach_event;
233 	struct ether_addr ifc_mac;
234 };
235 
236 void *
237 iflib_get_softc(if_ctx_t ctx)
238 {
239 
240 	return (ctx->ifc_softc);
241 }
242 
243 device_t
244 iflib_get_dev(if_ctx_t ctx)
245 {
246 
247 	return (ctx->ifc_dev);
248 }
249 
250 if_t
251 iflib_get_ifp(if_ctx_t ctx)
252 {
253 
254 	return (ctx->ifc_ifp);
255 }
256 
257 struct ifmedia *
258 iflib_get_media(if_ctx_t ctx)
259 {
260 
261 	return (ctx->ifc_mediap);
262 }
263 
264 uint32_t
265 iflib_get_flags(if_ctx_t ctx)
266 {
267 	return (ctx->ifc_flags);
268 }
269 
270 void
271 iflib_set_mac(if_ctx_t ctx, uint8_t mac[ETHER_ADDR_LEN])
272 {
273 
274 	bcopy(mac, ctx->ifc_mac.octet, ETHER_ADDR_LEN);
275 }
276 
277 if_softc_ctx_t
278 iflib_get_softc_ctx(if_ctx_t ctx)
279 {
280 
281 	return (&ctx->ifc_softc_ctx);
282 }
283 
284 if_shared_ctx_t
285 iflib_get_sctx(if_ctx_t ctx)
286 {
287 
288 	return (ctx->ifc_sctx);
289 }
290 
291 #define IP_ALIGNED(m) ((((uintptr_t)(m)->m_data) & 0x3) == 0x2)
292 #define CACHE_PTR_INCREMENT (CACHE_LINE_SIZE/sizeof(void*))
293 #define CACHE_PTR_NEXT(ptr) ((void *)(((uintptr_t)(ptr)+CACHE_LINE_SIZE-1) & (CACHE_LINE_SIZE-1)))
294 
295 #define LINK_ACTIVE(ctx) ((ctx)->ifc_link_state == LINK_STATE_UP)
296 #define CTX_IS_VF(ctx) ((ctx)->ifc_sctx->isc_flags & IFLIB_IS_VF)
297 
298 typedef struct iflib_sw_rx_desc_array {
299 	bus_dmamap_t	*ifsd_map;         /* bus_dma maps for packet */
300 	struct mbuf	**ifsd_m;           /* pkthdr mbufs */
301 	caddr_t		*ifsd_cl;          /* direct cluster pointer for rx */
302 	bus_addr_t	*ifsd_ba;          /* bus addr of cluster for rx */
303 } iflib_rxsd_array_t;
304 
305 typedef struct iflib_sw_tx_desc_array {
306 	bus_dmamap_t    *ifsd_map;         /* bus_dma maps for packet */
307 	bus_dmamap_t	*ifsd_tso_map;     /* bus_dma maps for TSO packet */
308 	struct mbuf    **ifsd_m;           /* pkthdr mbufs */
309 } if_txsd_vec_t;
310 
311 /* magic number that should be high enough for any hardware */
312 #define IFLIB_MAX_TX_SEGS		128
313 #define IFLIB_RX_COPY_THRESH		128
314 #define IFLIB_MAX_RX_REFRESH		32
315 /* The minimum descriptors per second before we start coalescing */
316 #define IFLIB_MIN_DESC_SEC		16384
317 #define IFLIB_DEFAULT_TX_UPDATE_FREQ	16
318 #define IFLIB_QUEUE_IDLE		0
319 #define IFLIB_QUEUE_HUNG		1
320 #define IFLIB_QUEUE_WORKING		2
321 /* maximum number of txqs that can share an rx interrupt */
322 #define IFLIB_MAX_TX_SHARED_INTR	4
323 
324 /* this should really scale with ring size - this is a fairly arbitrary value */
325 #define TX_BATCH_SIZE			32
326 
327 #define IFLIB_RESTART_BUDGET		8
328 
329 #define CSUM_OFFLOAD		(CSUM_IP_TSO|CSUM_IP6_TSO|CSUM_IP| \
330 				 CSUM_IP_UDP|CSUM_IP_TCP|CSUM_IP_SCTP| \
331 				 CSUM_IP6_UDP|CSUM_IP6_TCP|CSUM_IP6_SCTP)
332 
333 struct iflib_txq {
334 	qidx_t		ift_in_use;
335 	qidx_t		ift_cidx;
336 	qidx_t		ift_cidx_processed;
337 	qidx_t		ift_pidx;
338 	uint8_t		ift_gen;
339 	uint8_t		ift_br_offset;
340 	uint16_t	ift_npending;
341 	uint16_t	ift_db_pending;
342 	uint16_t	ift_rs_pending;
343 	/* implicit pad */
344 	uint8_t		ift_txd_size[8];
345 	uint64_t	ift_processed;
346 	uint64_t	ift_cleaned;
347 	uint64_t	ift_cleaned_prev;
348 #if MEMORY_LOGGING
349 	uint64_t	ift_enqueued;
350 	uint64_t	ift_dequeued;
351 #endif
352 	uint64_t	ift_no_tx_dma_setup;
353 	uint64_t	ift_no_desc_avail;
354 	uint64_t	ift_mbuf_defrag_failed;
355 	uint64_t	ift_mbuf_defrag;
356 	uint64_t	ift_map_failed;
357 	uint64_t	ift_txd_encap_efbig;
358 	uint64_t	ift_pullups;
359 	uint64_t	ift_last_timer_tick;
360 
361 	struct mtx	ift_mtx;
362 	struct mtx	ift_db_mtx;
363 
364 	/* constant values */
365 	if_ctx_t	ift_ctx;
366 	struct ifmp_ring        *ift_br;
367 	struct grouptask	ift_task;
368 	qidx_t		ift_size;
369 	uint16_t	ift_id;
370 	struct callout	ift_timer;
371 #ifdef DEV_NETMAP
372 	struct callout	ift_netmap_timer;
373 #endif /* DEV_NETMAP */
374 
375 	if_txsd_vec_t	ift_sds;
376 	uint8_t		ift_qstatus;
377 	uint8_t		ift_closed;
378 	uint8_t		ift_update_freq;
379 	struct iflib_filter_info ift_filter_info;
380 	bus_dma_tag_t	ift_buf_tag;
381 	bus_dma_tag_t	ift_tso_buf_tag;
382 	iflib_dma_info_t	ift_ifdi;
383 #define	MTX_NAME_LEN	32
384 	char                    ift_mtx_name[MTX_NAME_LEN];
385 	bus_dma_segment_t	ift_segs[IFLIB_MAX_TX_SEGS]  __aligned(CACHE_LINE_SIZE);
386 #ifdef IFLIB_DIAGNOSTICS
387 	uint64_t ift_cpu_exec_count[256];
388 #endif
389 } __aligned(CACHE_LINE_SIZE);
390 
391 struct iflib_fl {
392 	qidx_t		ifl_cidx;
393 	qidx_t		ifl_pidx;
394 	qidx_t		ifl_credits;
395 	uint8_t		ifl_gen;
396 	uint8_t		ifl_rxd_size;
397 #if MEMORY_LOGGING
398 	uint64_t	ifl_m_enqueued;
399 	uint64_t	ifl_m_dequeued;
400 	uint64_t	ifl_cl_enqueued;
401 	uint64_t	ifl_cl_dequeued;
402 #endif
403 	/* implicit pad */
404 	bitstr_t 	*ifl_rx_bitmap;
405 	qidx_t		ifl_fragidx;
406 	/* constant */
407 	qidx_t		ifl_size;
408 	uint16_t	ifl_buf_size;
409 	uint16_t	ifl_cltype;
410 #ifndef __HAIKU__
411 	uma_zone_t	ifl_zone;
412 #endif
413 	iflib_rxsd_array_t	ifl_sds;
414 	iflib_rxq_t	ifl_rxq;
415 	uint8_t		ifl_id;
416 	bus_dma_tag_t	ifl_buf_tag;
417 	iflib_dma_info_t	ifl_ifdi;
418 	uint64_t	ifl_bus_addrs[IFLIB_MAX_RX_REFRESH] __aligned(CACHE_LINE_SIZE);
419 	qidx_t		ifl_rxd_idxs[IFLIB_MAX_RX_REFRESH];
420 }  __aligned(CACHE_LINE_SIZE);
421 
422 static inline qidx_t
423 get_inuse(int size, qidx_t cidx, qidx_t pidx, uint8_t gen)
424 {
425 	qidx_t used;
426 
427 	if (pidx > cidx)
428 		used = pidx - cidx;
429 	else if (pidx < cidx)
430 		used = size - cidx + pidx;
431 	else if (gen == 0 && pidx == cidx)
432 		used = 0;
433 	else if (gen == 1 && pidx == cidx)
434 		used = size;
435 	else
436 		panic("bad state");
437 
438 	return (used);
439 }
440 
441 #define TXQ_AVAIL(txq) (txq->ift_size - get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx, txq->ift_gen))
442 
443 #define IDXDIFF(head, tail, wrap) \
444 	((head) >= (tail) ? (head) - (tail) : (wrap) - (tail) + (head))
445 
446 struct iflib_rxq {
447 	if_ctx_t	ifr_ctx;
448 	iflib_fl_t	ifr_fl;
449 	uint64_t	ifr_rx_irq;
450 #ifndef __HAIKU__
451 	struct pfil_head	*pfil;
452 #else
453 #define PFIL_PASS 0
454 #endif
455 	/*
456 	 * If there is a separate completion queue (IFLIB_HAS_RXCQ), this is
457 	 * the completion queue consumer index.  Otherwise it's unused.
458 	 */
459 	qidx_t		ifr_cq_cidx;
460 	uint16_t	ifr_id;
461 	uint8_t		ifr_nfl;
462 	uint8_t		ifr_ntxqirq;
463 	uint8_t		ifr_txqid[IFLIB_MAX_TX_SHARED_INTR];
464 	uint8_t		ifr_fl_offset;
465 #ifndef __HAIKU__
466 	struct lro_ctrl			ifr_lc;
467 #endif
468 	struct grouptask        ifr_task;
469 	struct callout		ifr_watchdog;
470 	struct iflib_filter_info ifr_filter_info;
471 	iflib_dma_info_t		ifr_ifdi;
472 
473 	/* dynamically allocate if any drivers need a value substantially larger than this */
474 	struct if_rxd_frag	ifr_frags[IFLIB_MAX_RX_SEGS] __aligned(CACHE_LINE_SIZE);
475 #ifdef IFLIB_DIAGNOSTICS
476 	uint64_t ifr_cpu_exec_count[256];
477 #endif
478 }  __aligned(CACHE_LINE_SIZE);
479 
480 typedef struct if_rxsd {
481 	caddr_t *ifsd_cl;
482 	iflib_fl_t ifsd_fl;
483 } *if_rxsd_t;
484 
485 /* multiple of word size */
486 #ifdef __LP64__
487 #define PKT_INFO_SIZE	6
488 #define RXD_INFO_SIZE	5
489 #define PKT_TYPE uint64_t
490 #else
491 #define PKT_INFO_SIZE	11
492 #define RXD_INFO_SIZE	8
493 #define PKT_TYPE uint32_t
494 #endif
495 #define PKT_LOOP_BOUND  ((PKT_INFO_SIZE/3)*3)
496 #define RXD_LOOP_BOUND  ((RXD_INFO_SIZE/4)*4)
497 
498 typedef struct if_pkt_info_pad {
499 	PKT_TYPE pkt_val[PKT_INFO_SIZE];
500 } *if_pkt_info_pad_t;
501 typedef struct if_rxd_info_pad {
502 	PKT_TYPE rxd_val[RXD_INFO_SIZE];
503 } *if_rxd_info_pad_t;
504 
505 CTASSERT(sizeof(struct if_pkt_info_pad) == sizeof(struct if_pkt_info));
506 CTASSERT(sizeof(struct if_rxd_info_pad) == sizeof(struct if_rxd_info));
507 
508 static inline void
509 pkt_info_zero(if_pkt_info_t pi)
510 {
511 	if_pkt_info_pad_t pi_pad;
512 
513 	pi_pad = (if_pkt_info_pad_t)pi;
514 	pi_pad->pkt_val[0] = 0; pi_pad->pkt_val[1] = 0; pi_pad->pkt_val[2] = 0;
515 	pi_pad->pkt_val[3] = 0; pi_pad->pkt_val[4] = 0; pi_pad->pkt_val[5] = 0;
516 #ifndef __LP64__
517 	pi_pad->pkt_val[6] = 0; pi_pad->pkt_val[7] = 0; pi_pad->pkt_val[8] = 0;
518 	pi_pad->pkt_val[9] = 0; pi_pad->pkt_val[10] = 0;
519 #endif
520 }
521 
522 #ifndef __HAIKU__
523 static device_method_t iflib_pseudo_methods[] = {
524 	DEVMETHOD(device_attach, noop_attach),
525 	DEVMETHOD(device_detach, iflib_pseudo_detach),
526 	DEVMETHOD_END
527 };
528 
529 driver_t iflib_pseudodriver = {
530 	"iflib_pseudo", iflib_pseudo_methods, sizeof(struct iflib_ctx),
531 };
532 #endif
533 
534 static inline void
535 rxd_info_zero(if_rxd_info_t ri)
536 {
537 	if_rxd_info_pad_t ri_pad;
538 	int i;
539 
540 	ri_pad = (if_rxd_info_pad_t)ri;
541 	for (i = 0; i < RXD_LOOP_BOUND; i += 4) {
542 		ri_pad->rxd_val[i] = 0;
543 		ri_pad->rxd_val[i+1] = 0;
544 		ri_pad->rxd_val[i+2] = 0;
545 		ri_pad->rxd_val[i+3] = 0;
546 	}
547 #ifdef __LP64__
548 	ri_pad->rxd_val[RXD_INFO_SIZE-1] = 0;
549 #endif
550 }
551 
552 /*
553  * Only allow a single packet to take up most 1/nth of the tx ring
554  */
555 #define MAX_SINGLE_PACKET_FRACTION 12
556 #define IF_BAD_DMA (bus_addr_t)-1
557 
558 #define CTX_ACTIVE(ctx) ((if_getdrvflags((ctx)->ifc_ifp) & IFF_DRV_RUNNING))
559 
560 #define CTX_LOCK_INIT(_sc)  sx_init(&(_sc)->ifc_ctx_sx, "iflib ctx lock")
561 #define CTX_LOCK(ctx) sx_xlock(&(ctx)->ifc_ctx_sx)
562 #define CTX_UNLOCK(ctx) sx_xunlock(&(ctx)->ifc_ctx_sx)
563 #define CTX_LOCK_DESTROY(ctx) sx_destroy(&(ctx)->ifc_ctx_sx)
564 
565 #define STATE_LOCK_INIT(_sc, _name)  mtx_init(&(_sc)->ifc_state_mtx, _name, "iflib state lock", MTX_DEF)
566 #define STATE_LOCK(ctx) mtx_lock(&(ctx)->ifc_state_mtx)
567 #define STATE_UNLOCK(ctx) mtx_unlock(&(ctx)->ifc_state_mtx)
568 #define STATE_LOCK_DESTROY(ctx) mtx_destroy(&(ctx)->ifc_state_mtx)
569 
570 #define CALLOUT_LOCK(txq)	mtx_lock(&txq->ift_mtx)
571 #define CALLOUT_UNLOCK(txq) 	mtx_unlock(&txq->ift_mtx)
572 
573 void
574 iflib_set_detach(if_ctx_t ctx)
575 {
576 	STATE_LOCK(ctx);
577 	ctx->ifc_flags |= IFC_IN_DETACH;
578 	STATE_UNLOCK(ctx);
579 }
580 
581 /* Our boot-time initialization hook */
582 static int	iflib_module_event_handler(module_t, int, void *);
583 
584 #ifndef __HAIKU__
585 static moduledata_t iflib_moduledata = {
586 	"iflib",
587 	iflib_module_event_handler,
588 	NULL
589 };
590 #endif
591 
592 DECLARE_MODULE(iflib, iflib_moduledata, SI_SUB_INIT_IF, SI_ORDER_ANY);
593 MODULE_VERSION(iflib, 1);
594 
595 MODULE_DEPEND(iflib, pci, 1, 1, 1);
596 MODULE_DEPEND(iflib, ether, 1, 1, 1);
597 
598 TASKQGROUP_DEFINE(if_io_tqg, mp_ncpus, 1);
599 TASKQGROUP_DEFINE(if_config_tqg, 1, 1);
600 
601 #ifndef IFLIB_DEBUG_COUNTERS
602 #ifdef INVARIANTS
603 #define IFLIB_DEBUG_COUNTERS 1
604 #else
605 #define IFLIB_DEBUG_COUNTERS 0
606 #endif /* !INVARIANTS */
607 #endif
608 
609 static SYSCTL_NODE(_net, OID_AUTO, iflib, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
610     "iflib driver parameters");
611 
612 /*
613  * XXX need to ensure that this can't accidentally cause the head to be moved backwards
614  */
615 static int iflib_min_tx_latency = 0;
616 SYSCTL_INT(_net_iflib, OID_AUTO, min_tx_latency, CTLFLAG_RW,
617 		   &iflib_min_tx_latency, 0, "minimize transmit latency at the possible expense of throughput");
618 static int iflib_no_tx_batch = 0;
619 SYSCTL_INT(_net_iflib, OID_AUTO, no_tx_batch, CTLFLAG_RW,
620 		   &iflib_no_tx_batch, 0, "minimize transmit latency at the possible expense of throughput");
621 static int iflib_timer_default = 1000;
622 SYSCTL_INT(_net_iflib, OID_AUTO, timer_default, CTLFLAG_RW,
623 		   &iflib_timer_default, 0, "number of ticks between iflib_timer calls");
624 
625 
626 #if IFLIB_DEBUG_COUNTERS
627 
628 static int iflib_tx_seen;
629 static int iflib_tx_sent;
630 static int iflib_tx_encap;
631 static int iflib_rx_allocs;
632 static int iflib_fl_refills;
633 static int iflib_fl_refills_large;
634 static int iflib_tx_frees;
635 
636 SYSCTL_INT(_net_iflib, OID_AUTO, tx_seen, CTLFLAG_RD,
637 		   &iflib_tx_seen, 0, "# TX mbufs seen");
638 SYSCTL_INT(_net_iflib, OID_AUTO, tx_sent, CTLFLAG_RD,
639 		   &iflib_tx_sent, 0, "# TX mbufs sent");
640 SYSCTL_INT(_net_iflib, OID_AUTO, tx_encap, CTLFLAG_RD,
641 		   &iflib_tx_encap, 0, "# TX mbufs encapped");
642 SYSCTL_INT(_net_iflib, OID_AUTO, tx_frees, CTLFLAG_RD,
643 		   &iflib_tx_frees, 0, "# TX frees");
644 SYSCTL_INT(_net_iflib, OID_AUTO, rx_allocs, CTLFLAG_RD,
645 		   &iflib_rx_allocs, 0, "# RX allocations");
646 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills, CTLFLAG_RD,
647 		   &iflib_fl_refills, 0, "# refills");
648 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills_large, CTLFLAG_RD,
649 		   &iflib_fl_refills_large, 0, "# large refills");
650 
651 static int iflib_txq_drain_flushing;
652 static int iflib_txq_drain_oactive;
653 static int iflib_txq_drain_notready;
654 
655 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_flushing, CTLFLAG_RD,
656 		   &iflib_txq_drain_flushing, 0, "# drain flushes");
657 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_oactive, CTLFLAG_RD,
658 		   &iflib_txq_drain_oactive, 0, "# drain oactives");
659 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_notready, CTLFLAG_RD,
660 		   &iflib_txq_drain_notready, 0, "# drain notready");
661 
662 static int iflib_encap_load_mbuf_fail;
663 static int iflib_encap_pad_mbuf_fail;
664 static int iflib_encap_txq_avail_fail;
665 static int iflib_encap_txd_encap_fail;
666 
667 SYSCTL_INT(_net_iflib, OID_AUTO, encap_load_mbuf_fail, CTLFLAG_RD,
668 		   &iflib_encap_load_mbuf_fail, 0, "# busdma load failures");
669 SYSCTL_INT(_net_iflib, OID_AUTO, encap_pad_mbuf_fail, CTLFLAG_RD,
670 		   &iflib_encap_pad_mbuf_fail, 0, "# runt frame pad failures");
671 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txq_avail_fail, CTLFLAG_RD,
672 		   &iflib_encap_txq_avail_fail, 0, "# txq avail failures");
673 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txd_encap_fail, CTLFLAG_RD,
674 		   &iflib_encap_txd_encap_fail, 0, "# driver encap failures");
675 
676 static int iflib_task_fn_rxs;
677 static int iflib_rx_intr_enables;
678 static int iflib_fast_intrs;
679 static int iflib_rx_unavail;
680 static int iflib_rx_ctx_inactive;
681 static int iflib_rx_if_input;
682 static int iflib_rxd_flush;
683 
684 static int iflib_verbose_debug;
685 
686 SYSCTL_INT(_net_iflib, OID_AUTO, task_fn_rx, CTLFLAG_RD,
687 		   &iflib_task_fn_rxs, 0, "# task_fn_rx calls");
688 SYSCTL_INT(_net_iflib, OID_AUTO, rx_intr_enables, CTLFLAG_RD,
689 		   &iflib_rx_intr_enables, 0, "# RX intr enables");
690 SYSCTL_INT(_net_iflib, OID_AUTO, fast_intrs, CTLFLAG_RD,
691 		   &iflib_fast_intrs, 0, "# fast_intr calls");
692 SYSCTL_INT(_net_iflib, OID_AUTO, rx_unavail, CTLFLAG_RD,
693 		   &iflib_rx_unavail, 0, "# times rxeof called with no available data");
694 SYSCTL_INT(_net_iflib, OID_AUTO, rx_ctx_inactive, CTLFLAG_RD,
695 		   &iflib_rx_ctx_inactive, 0, "# times rxeof called with inactive context");
696 SYSCTL_INT(_net_iflib, OID_AUTO, rx_if_input, CTLFLAG_RD,
697 		   &iflib_rx_if_input, 0, "# times rxeof called if_input");
698 SYSCTL_INT(_net_iflib, OID_AUTO, rxd_flush, CTLFLAG_RD,
699 	         &iflib_rxd_flush, 0, "# times rxd_flush called");
700 SYSCTL_INT(_net_iflib, OID_AUTO, verbose_debug, CTLFLAG_RW,
701 		   &iflib_verbose_debug, 0, "enable verbose debugging");
702 
703 #define DBG_COUNTER_INC(name) atomic_add_int(&(iflib_ ## name), 1)
704 static void
705 iflib_debug_reset(void)
706 {
707 	iflib_tx_seen = iflib_tx_sent = iflib_tx_encap = iflib_rx_allocs =
708 		iflib_fl_refills = iflib_fl_refills_large = iflib_tx_frees =
709 		iflib_txq_drain_flushing = iflib_txq_drain_oactive =
710 		iflib_txq_drain_notready =
711 		iflib_encap_load_mbuf_fail = iflib_encap_pad_mbuf_fail =
712 		iflib_encap_txq_avail_fail = iflib_encap_txd_encap_fail =
713 		iflib_task_fn_rxs = iflib_rx_intr_enables = iflib_fast_intrs =
714 		iflib_rx_unavail =
715 		iflib_rx_ctx_inactive = iflib_rx_if_input =
716 		iflib_rxd_flush = 0;
717 }
718 
719 #else
720 #define DBG_COUNTER_INC(name)
721 static void iflib_debug_reset(void) {}
722 #endif
723 
724 #define IFLIB_DEBUG 0
725 
726 static void iflib_tx_structures_free(if_ctx_t ctx);
727 static void iflib_rx_structures_free(if_ctx_t ctx);
728 static int iflib_queues_alloc(if_ctx_t ctx);
729 static int iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq);
730 static int iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget);
731 static int iflib_qset_structures_setup(if_ctx_t ctx);
732 static int iflib_msix_init(if_ctx_t ctx);
733 static int iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filterarg, int *rid, const char *str);
734 static void iflib_txq_check_drain(iflib_txq_t txq, int budget);
735 static uint32_t iflib_txq_can_drain(struct ifmp_ring *);
736 #ifdef ALTQ
737 static void iflib_altq_if_start(if_t ifp);
738 static int iflib_altq_if_transmit(if_t ifp, struct mbuf *m);
739 #endif
740 static int iflib_register(if_ctx_t);
741 static void iflib_deregister(if_ctx_t);
742 static void iflib_unregister_vlan_handlers(if_ctx_t ctx);
743 static uint16_t iflib_get_mbuf_size_for(unsigned int size);
744 static void iflib_init_locked(if_ctx_t ctx);
745 static void iflib_add_device_sysctl_pre(if_ctx_t ctx);
746 static void iflib_add_device_sysctl_post(if_ctx_t ctx);
747 static void iflib_ifmp_purge(iflib_txq_t txq);
748 static void _iflib_pre_assert(if_softc_ctx_t scctx);
749 static void iflib_if_init_locked(if_ctx_t ctx);
750 static void iflib_free_intr_mem(if_ctx_t ctx);
751 #ifndef __NO_STRICT_ALIGNMENT
752 static struct mbuf * iflib_fixup_rx(struct mbuf *m);
753 #endif
754 
755 #ifndef __HAIKU__
756 static SLIST_HEAD(cpu_offset_list, cpu_offset) cpu_offsets =
757     SLIST_HEAD_INITIALIZER(cpu_offsets);
758 struct cpu_offset {
759 	SLIST_ENTRY(cpu_offset) entries;
760 	cpuset_t	set;
761 	unsigned int	refcount;
762 	uint16_t	next_cpuid;
763 };
764 static struct mtx cpu_offset_mtx;
765 MTX_SYSINIT(iflib_cpu_offset, &cpu_offset_mtx, "iflib_cpu_offset lock",
766     MTX_DEF);
767 #endif
768 
769 DEBUGNET_DEFINE(iflib);
770 
771 static int
772 iflib_num_rx_descs(if_ctx_t ctx)
773 {
774 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
775 	if_shared_ctx_t sctx = ctx->ifc_sctx;
776 	uint16_t first_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
777 
778 	return scctx->isc_nrxd[first_rxq];
779 }
780 
781 static int
782 iflib_num_tx_descs(if_ctx_t ctx)
783 {
784 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
785 	if_shared_ctx_t sctx = ctx->ifc_sctx;
786 	uint16_t first_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
787 
788 	return scctx->isc_ntxd[first_txq];
789 }
790 
791 #ifdef DEV_NETMAP
792 #include <sys/selinfo.h>
793 #include <net/netmap.h>
794 #include <dev/netmap/netmap_kern.h>
795 
796 MODULE_DEPEND(iflib, netmap, 1, 1, 1);
797 
798 static int netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, bool init);
799 static void iflib_netmap_timer(void *arg);
800 
801 /*
802  * device-specific sysctl variables:
803  *
804  * iflib_crcstrip: 0: keep CRC in rx frames (default), 1: strip it.
805  *	During regular operations the CRC is stripped, but on some
806  *	hardware reception of frames not multiple of 64 is slower,
807  *	so using crcstrip=0 helps in benchmarks.
808  *
809  * iflib_rx_miss, iflib_rx_miss_bufs:
810  *	count packets that might be missed due to lost interrupts.
811  */
812 SYSCTL_DECL(_dev_netmap);
813 /*
814  * The xl driver by default strips CRCs and we do not override it.
815  */
816 
817 int iflib_crcstrip = 1;
818 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_crcstrip,
819     CTLFLAG_RW, &iflib_crcstrip, 1, "strip CRC on RX frames");
820 
821 int iflib_rx_miss, iflib_rx_miss_bufs;
822 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss,
823     CTLFLAG_RW, &iflib_rx_miss, 0, "potentially missed RX intr");
824 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss_bufs,
825     CTLFLAG_RW, &iflib_rx_miss_bufs, 0, "potentially missed RX intr bufs");
826 
827 /*
828  * Register/unregister. We are already under netmap lock.
829  * Only called on the first register or the last unregister.
830  */
831 static int
832 iflib_netmap_register(struct netmap_adapter *na, int onoff)
833 {
834 	if_t ifp = na->ifp;
835 	if_ctx_t ctx = ifp->if_softc;
836 	int status;
837 
838 	CTX_LOCK(ctx);
839 	if (!CTX_IS_VF(ctx))
840 		IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip);
841 
842 	iflib_stop(ctx);
843 
844 	/*
845 	 * Enable (or disable) netmap flags, and intercept (or restore)
846 	 * ifp->if_transmit. This is done once the device has been stopped
847 	 * to prevent race conditions. Also, this must be done after
848 	 * calling netmap_disable_all_rings() and before calling
849 	 * netmap_enable_all_rings(), so that these two functions see the
850 	 * updated state of the NAF_NETMAP_ON bit.
851 	 */
852 	if (onoff) {
853 		nm_set_native_flags(na);
854 	} else {
855 		nm_clear_native_flags(na);
856 	}
857 
858 	iflib_init_locked(ctx);
859 	IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip); // XXX why twice ?
860 	status = ifp->if_drv_flags & IFF_DRV_RUNNING ? 0 : 1;
861 	if (status)
862 		nm_clear_native_flags(na);
863 	CTX_UNLOCK(ctx);
864 	return (status);
865 }
866 
867 static int
868 iflib_netmap_config(struct netmap_adapter *na, struct nm_config_info *info)
869 {
870 	if_t ifp = na->ifp;
871 	if_ctx_t ctx = ifp->if_softc;
872 	iflib_rxq_t rxq = &ctx->ifc_rxqs[0];
873 	iflib_fl_t fl = &rxq->ifr_fl[0];
874 
875 	info->num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
876 	info->num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
877 	info->num_tx_descs = iflib_num_tx_descs(ctx);
878 	info->num_rx_descs = iflib_num_rx_descs(ctx);
879 	info->rx_buf_maxsize = fl->ifl_buf_size;
880 	nm_prinf("txr %u rxr %u txd %u rxd %u rbufsz %u",
881 		info->num_tx_rings, info->num_rx_rings, info->num_tx_descs,
882 		info->num_rx_descs, info->rx_buf_maxsize);
883 
884 	return 0;
885 }
886 
887 static int
888 netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, bool init)
889 {
890 	struct netmap_adapter *na = kring->na;
891 	u_int const lim = kring->nkr_num_slots - 1;
892 	struct netmap_ring *ring = kring->ring;
893 	bus_dmamap_t *map;
894 	struct if_rxd_update iru;
895 	if_ctx_t ctx = rxq->ifr_ctx;
896 	iflib_fl_t fl = &rxq->ifr_fl[0];
897 	u_int nic_i_first, nic_i;
898 	u_int nm_i;
899 	int i, n;
900 #if IFLIB_DEBUG_COUNTERS
901 	int rf_count = 0;
902 #endif
903 
904 	/*
905 	 * This function is used both at initialization and in rxsync.
906 	 * At initialization we need to prepare (with isc_rxd_refill())
907 	 * all the netmap buffers currently owned by the kernel, in
908 	 * such a way to keep fl->ifl_pidx and kring->nr_hwcur in sync
909 	 * (except for kring->nkr_hwofs). These may be less than
910 	 * kring->nkr_num_slots if netmap_reset() was called while
911 	 * an application using the kring that still owned some
912 	 * buffers.
913 	 * At rxsync time, both indexes point to the next buffer to be
914 	 * refilled.
915 	 * In any case we publish (with isc_rxd_flush()) up to
916 	 * (fl->ifl_pidx - 1) % N (included), to avoid the NIC tail/prod
917 	 * pointer to overrun the head/cons pointer, although this is
918 	 * not necessary for some NICs (e.g. vmx).
919 	 */
920 	if (__predict_false(init)) {
921 		n = kring->nkr_num_slots - nm_kr_rxspace(kring);
922 	} else {
923 		n = kring->rhead - kring->nr_hwcur;
924 		if (n == 0)
925 			return (0); /* Nothing to do. */
926 		if (n < 0)
927 			n += kring->nkr_num_slots;
928 	}
929 
930 	iru_init(&iru, rxq, 0 /* flid */);
931 	map = fl->ifl_sds.ifsd_map;
932 	nic_i = fl->ifl_pidx;
933 	nm_i = netmap_idx_n2k(kring, nic_i);
934 	if (__predict_false(init)) {
935 		/*
936 		 * On init/reset, nic_i must be 0, and we must
937 		 * start to refill from hwtail (see netmap_reset()).
938 		 */
939 		MPASS(nic_i == 0);
940 		MPASS(nm_i == kring->nr_hwtail);
941 	} else
942 		MPASS(nm_i == kring->nr_hwcur);
943 	DBG_COUNTER_INC(fl_refills);
944 	while (n > 0) {
945 #if IFLIB_DEBUG_COUNTERS
946 		if (++rf_count == 9)
947 			DBG_COUNTER_INC(fl_refills_large);
948 #endif
949 		nic_i_first = nic_i;
950 		for (i = 0; n > 0 && i < IFLIB_MAX_RX_REFRESH; n--, i++) {
951 			struct netmap_slot *slot = &ring->slot[nm_i];
952 			void *addr = PNMB(na, slot, &fl->ifl_bus_addrs[i]);
953 
954 			MPASS(i < IFLIB_MAX_RX_REFRESH);
955 
956 			if (addr == NETMAP_BUF_BASE(na)) /* bad buf */
957 			        return netmap_ring_reinit(kring);
958 
959 			fl->ifl_rxd_idxs[i] = nic_i;
960 
961 			if (__predict_false(init)) {
962 				netmap_load_map(na, fl->ifl_buf_tag,
963 				    map[nic_i], addr);
964 			} else if (slot->flags & NS_BUF_CHANGED) {
965 				/* buffer has changed, reload map */
966 				netmap_reload_map(na, fl->ifl_buf_tag,
967 				    map[nic_i], addr);
968 			}
969 			bus_dmamap_sync(fl->ifl_buf_tag, map[nic_i],
970 			    BUS_DMASYNC_PREREAD);
971 			slot->flags &= ~NS_BUF_CHANGED;
972 
973 			nm_i = nm_next(nm_i, lim);
974 			nic_i = nm_next(nic_i, lim);
975 		}
976 
977 		iru.iru_pidx = nic_i_first;
978 		iru.iru_count = i;
979 		ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
980 	}
981 	fl->ifl_pidx = nic_i;
982 	/*
983 	 * At the end of the loop we must have refilled everything
984 	 * we could possibly refill.
985 	 */
986 	MPASS(nm_i == kring->rhead);
987 	kring->nr_hwcur = nm_i;
988 
989 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
990 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
991 	ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id,
992 	    nm_prev(nic_i, lim));
993 	DBG_COUNTER_INC(rxd_flush);
994 
995 	return (0);
996 }
997 
998 #define NETMAP_TX_TIMER_US	90
999 
1000 /*
1001  * Reconcile kernel and user view of the transmit ring.
1002  *
1003  * All information is in the kring.
1004  * Userspace wants to send packets up to the one before kring->rhead,
1005  * kernel knows kring->nr_hwcur is the first unsent packet.
1006  *
1007  * Here we push packets out (as many as possible), and possibly
1008  * reclaim buffers from previously completed transmission.
1009  *
1010  * The caller (netmap) guarantees that there is only one instance
1011  * running at any time. Any interference with other driver
1012  * methods should be handled by the individual drivers.
1013  */
1014 static int
1015 iflib_netmap_txsync(struct netmap_kring *kring, int flags)
1016 {
1017 	struct netmap_adapter *na = kring->na;
1018 	if_t ifp = na->ifp;
1019 	struct netmap_ring *ring = kring->ring;
1020 	u_int nm_i;	/* index into the netmap kring */
1021 	u_int nic_i;	/* index into the NIC ring */
1022 	u_int n;
1023 	u_int const lim = kring->nkr_num_slots - 1;
1024 	u_int const head = kring->rhead;
1025 	struct if_pkt_info pi;
1026 	int tx_pkts = 0, tx_bytes = 0;
1027 
1028 	/*
1029 	 * interrupts on every tx packet are expensive so request
1030 	 * them every half ring, or where NS_REPORT is set
1031 	 */
1032 	u_int report_frequency = kring->nkr_num_slots >> 1;
1033 	/* device-specific */
1034 	if_ctx_t ctx = ifp->if_softc;
1035 	iflib_txq_t txq = &ctx->ifc_txqs[kring->ring_id];
1036 
1037 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1038 	    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1039 
1040 	/*
1041 	 * First part: process new packets to send.
1042 	 * nm_i is the current index in the netmap kring,
1043 	 * nic_i is the corresponding index in the NIC ring.
1044 	 *
1045 	 * If we have packets to send (nm_i != head)
1046 	 * iterate over the netmap ring, fetch length and update
1047 	 * the corresponding slot in the NIC ring. Some drivers also
1048 	 * need to update the buffer's physical address in the NIC slot
1049 	 * even NS_BUF_CHANGED is not set (PNMB computes the addresses).
1050 	 *
1051 	 * The netmap_reload_map() calls is especially expensive,
1052 	 * even when (as in this case) the tag is 0, so do only
1053 	 * when the buffer has actually changed.
1054 	 *
1055 	 * If possible do not set the report/intr bit on all slots,
1056 	 * but only a few times per ring or when NS_REPORT is set.
1057 	 *
1058 	 * Finally, on 10G and faster drivers, it might be useful
1059 	 * to prefetch the next slot and txr entry.
1060 	 */
1061 
1062 	nm_i = kring->nr_hwcur;
1063 	if (nm_i != head) {	/* we have new packets to send */
1064 		uint32_t pkt_len = 0, seg_idx = 0;
1065 		int nic_i_start = -1, flags = 0;
1066 		pkt_info_zero(&pi);
1067 		pi.ipi_segs = txq->ift_segs;
1068 		pi.ipi_qsidx = kring->ring_id;
1069 		nic_i = netmap_idx_k2n(kring, nm_i);
1070 
1071 		__builtin_prefetch(&ring->slot[nm_i]);
1072 		__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i]);
1073 		__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i]);
1074 
1075 		for (n = 0; nm_i != head; n++) {
1076 			struct netmap_slot *slot = &ring->slot[nm_i];
1077 			u_int len = slot->len;
1078 			uint64_t paddr;
1079 			void *addr = PNMB(na, slot, &paddr);
1080 
1081 			flags |= (slot->flags & NS_REPORT ||
1082 				nic_i == 0 || nic_i == report_frequency) ?
1083 				IPI_TX_INTR : 0;
1084 
1085 			/*
1086 			 * If this is the first packet fragment, save the
1087 			 * index of the first NIC slot for later.
1088 			 */
1089 			if (nic_i_start < 0)
1090 				nic_i_start = nic_i;
1091 
1092 			pi.ipi_segs[seg_idx].ds_addr = paddr;
1093 			pi.ipi_segs[seg_idx].ds_len = len;
1094 			if (len) {
1095 				pkt_len += len;
1096 				seg_idx++;
1097 			}
1098 
1099 			if (!(slot->flags & NS_MOREFRAG)) {
1100 				pi.ipi_len = pkt_len;
1101 				pi.ipi_nsegs = seg_idx;
1102 				pi.ipi_pidx = nic_i_start;
1103 				pi.ipi_ndescs = 0;
1104 				pi.ipi_flags = flags;
1105 
1106 				/* Prepare the NIC TX ring. */
1107 				ctx->isc_txd_encap(ctx->ifc_softc, &pi);
1108 				DBG_COUNTER_INC(tx_encap);
1109 
1110 				/* Update transmit counters */
1111 				tx_bytes += pi.ipi_len;
1112 				tx_pkts++;
1113 
1114 				/* Reinit per-packet info for the next one. */
1115 				flags = seg_idx = pkt_len = 0;
1116 				nic_i_start = -1;
1117 			}
1118 
1119 			/* prefetch for next round */
1120 			__builtin_prefetch(&ring->slot[nm_i + 1]);
1121 			__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i + 1]);
1122 			__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i + 1]);
1123 
1124 			NM_CHECK_ADDR_LEN(na, addr, len);
1125 
1126 			if (slot->flags & NS_BUF_CHANGED) {
1127 				/* buffer has changed, reload map */
1128 				netmap_reload_map(na, txq->ift_buf_tag,
1129 				    txq->ift_sds.ifsd_map[nic_i], addr);
1130 			}
1131 			/* make sure changes to the buffer are synced */
1132 			bus_dmamap_sync(txq->ift_buf_tag,
1133 			    txq->ift_sds.ifsd_map[nic_i],
1134 			    BUS_DMASYNC_PREWRITE);
1135 
1136 			slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED | NS_MOREFRAG);
1137 			nm_i = nm_next(nm_i, lim);
1138 			nic_i = nm_next(nic_i, lim);
1139 		}
1140 		kring->nr_hwcur = nm_i;
1141 
1142 		/* synchronize the NIC ring */
1143 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1144 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1145 
1146 		/* (re)start the tx unit up to slot nic_i (excluded) */
1147 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, nic_i);
1148 	}
1149 
1150 	/*
1151 	 * Second part: reclaim buffers for completed transmissions.
1152 	 *
1153 	 * If there are unclaimed buffers, attempt to reclaim them.
1154 	 * If we don't manage to reclaim them all, and TX IRQs are not in use,
1155 	 * trigger a per-tx-queue timer to try again later.
1156 	 */
1157 	if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1158 		if (iflib_tx_credits_update(ctx, txq)) {
1159 			/* some tx completed, increment avail */
1160 			nic_i = txq->ift_cidx_processed;
1161 			kring->nr_hwtail = nm_prev(netmap_idx_n2k(kring, nic_i), lim);
1162 		}
1163 	}
1164 
1165 	if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ))
1166 		if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1167 			callout_reset_sbt_on(&txq->ift_netmap_timer,
1168 			    NETMAP_TX_TIMER_US * SBT_1US, SBT_1US,
1169 			    iflib_netmap_timer, txq,
1170 			    txq->ift_netmap_timer.c_cpu, 0);
1171 		}
1172 
1173 	if_inc_counter(ifp, IFCOUNTER_OBYTES, tx_bytes);
1174 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, tx_pkts);
1175 
1176 	return (0);
1177 }
1178 
1179 /*
1180  * Reconcile kernel and user view of the receive ring.
1181  * Same as for the txsync, this routine must be efficient.
1182  * The caller guarantees a single invocations, but races against
1183  * the rest of the driver should be handled here.
1184  *
1185  * On call, kring->rhead is the first packet that userspace wants
1186  * to keep, and kring->rcur is the wakeup point.
1187  * The kernel has previously reported packets up to kring->rtail.
1188  *
1189  * If (flags & NAF_FORCE_READ) also check for incoming packets irrespective
1190  * of whether or not we received an interrupt.
1191  */
1192 static int
1193 iflib_netmap_rxsync(struct netmap_kring *kring, int flags)
1194 {
1195 	struct netmap_adapter *na = kring->na;
1196 	struct netmap_ring *ring = kring->ring;
1197 	if_t ifp = na->ifp;
1198 	uint32_t nm_i;	/* index into the netmap ring */
1199 	uint32_t nic_i;	/* index into the NIC ring */
1200 	u_int n;
1201 	u_int const lim = kring->nkr_num_slots - 1;
1202 	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
1203 	int i = 0, rx_bytes = 0, rx_pkts = 0;
1204 
1205 	if_ctx_t ctx = ifp->if_softc;
1206 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1207 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1208 	iflib_rxq_t rxq = &ctx->ifc_rxqs[kring->ring_id];
1209 	iflib_fl_t fl = &rxq->ifr_fl[0];
1210 	struct if_rxd_info ri;
1211 	qidx_t *cidxp;
1212 
1213 	/*
1214 	 * netmap only uses free list 0, to avoid out of order consumption
1215 	 * of receive buffers
1216 	 */
1217 
1218 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
1219 	    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1220 
1221 	/*
1222 	 * First part: import newly received packets.
1223 	 *
1224 	 * nm_i is the index of the next free slot in the netmap ring,
1225 	 * nic_i is the index of the next received packet in the NIC ring
1226 	 * (or in the free list 0 if IFLIB_HAS_RXCQ is set), and they may
1227 	 * differ in case if_init() has been called while
1228 	 * in netmap mode. For the receive ring we have
1229 	 *
1230 	 *	nic_i = fl->ifl_cidx;
1231 	 *	nm_i = kring->nr_hwtail (previous)
1232 	 * and
1233 	 *	nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1234 	 *
1235 	 * fl->ifl_cidx is set to 0 on a ring reinit
1236 	 */
1237 	if (netmap_no_pendintr || force_update) {
1238 		uint32_t hwtail_lim = nm_prev(kring->nr_hwcur, lim);
1239 		bool have_rxcq = sctx->isc_flags & IFLIB_HAS_RXCQ;
1240 		int crclen = iflib_crcstrip ? 0 : 4;
1241 		int error, avail;
1242 
1243 		/*
1244 		 * For the free list consumer index, we use the same
1245 		 * logic as in iflib_rxeof().
1246 		 */
1247 		if (have_rxcq)
1248 			cidxp = &rxq->ifr_cq_cidx;
1249 		else
1250 			cidxp = &fl->ifl_cidx;
1251 		avail = ctx->isc_rxd_available(ctx->ifc_softc,
1252 		    rxq->ifr_id, *cidxp, USHRT_MAX);
1253 
1254 		nic_i = fl->ifl_cidx;
1255 		nm_i = netmap_idx_n2k(kring, nic_i);
1256 		MPASS(nm_i == kring->nr_hwtail);
1257 		for (n = 0; avail > 0 && nm_i != hwtail_lim; n++, avail--) {
1258 			rxd_info_zero(&ri);
1259 			ri.iri_frags = rxq->ifr_frags;
1260 			ri.iri_qsidx = kring->ring_id;
1261 			ri.iri_ifp = ctx->ifc_ifp;
1262 			ri.iri_cidx = *cidxp;
1263 
1264 			error = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
1265 			for (i = 0; i < ri.iri_nfrags; i++) {
1266 				if (error) {
1267 					ring->slot[nm_i].len = 0;
1268 					ring->slot[nm_i].flags = 0;
1269 				} else {
1270 					ring->slot[nm_i].len = ri.iri_frags[i].irf_len;
1271 					if (i == (ri.iri_nfrags - 1)) {
1272 						ring->slot[nm_i].len -= crclen;
1273 						ring->slot[nm_i].flags = 0;
1274 
1275 						/* Update receive counters */
1276 						rx_bytes += ri.iri_len;
1277 						rx_pkts++;
1278 					} else
1279 						ring->slot[nm_i].flags = NS_MOREFRAG;
1280 				}
1281 
1282 				bus_dmamap_sync(fl->ifl_buf_tag,
1283 				    fl->ifl_sds.ifsd_map[nic_i], BUS_DMASYNC_POSTREAD);
1284 				nm_i = nm_next(nm_i, lim);
1285 				fl->ifl_cidx = nic_i = nm_next(nic_i, lim);
1286 			}
1287 
1288 			if (have_rxcq) {
1289 				*cidxp = ri.iri_cidx;
1290 				while (*cidxp >= scctx->isc_nrxd[0])
1291 					*cidxp -= scctx->isc_nrxd[0];
1292 			}
1293 
1294 		}
1295 		if (n) { /* update the state variables */
1296 			if (netmap_no_pendintr && !force_update) {
1297 				/* diagnostics */
1298 				iflib_rx_miss ++;
1299 				iflib_rx_miss_bufs += n;
1300 			}
1301 			kring->nr_hwtail = nm_i;
1302 		}
1303 		kring->nr_kflags &= ~NKR_PENDINTR;
1304 	}
1305 	/*
1306 	 * Second part: skip past packets that userspace has released.
1307 	 * (kring->nr_hwcur to head excluded),
1308 	 * and make the buffers available for reception.
1309 	 * As usual nm_i is the index in the netmap ring,
1310 	 * nic_i is the index in the NIC ring, and
1311 	 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1312 	 */
1313 	netmap_fl_refill(rxq, kring, false);
1314 
1315 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
1316 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
1317 
1318 	return (0);
1319 }
1320 
1321 static void
1322 iflib_netmap_intr(struct netmap_adapter *na, int onoff)
1323 {
1324 	if_ctx_t ctx = na->ifp->if_softc;
1325 
1326 	CTX_LOCK(ctx);
1327 	if (onoff) {
1328 		IFDI_INTR_ENABLE(ctx);
1329 	} else {
1330 		IFDI_INTR_DISABLE(ctx);
1331 	}
1332 	CTX_UNLOCK(ctx);
1333 }
1334 
1335 static int
1336 iflib_netmap_attach(if_ctx_t ctx)
1337 {
1338 	struct netmap_adapter na;
1339 
1340 	bzero(&na, sizeof(na));
1341 
1342 	na.ifp = ctx->ifc_ifp;
1343 	na.na_flags = NAF_BDG_MAYSLEEP | NAF_MOREFRAG;
1344 	MPASS(ctx->ifc_softc_ctx.isc_ntxqsets);
1345 	MPASS(ctx->ifc_softc_ctx.isc_nrxqsets);
1346 
1347 	na.num_tx_desc = iflib_num_tx_descs(ctx);
1348 	na.num_rx_desc = iflib_num_rx_descs(ctx);
1349 	na.nm_txsync = iflib_netmap_txsync;
1350 	na.nm_rxsync = iflib_netmap_rxsync;
1351 	na.nm_register = iflib_netmap_register;
1352 	na.nm_intr = iflib_netmap_intr;
1353 	na.nm_config = iflib_netmap_config;
1354 	na.num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
1355 	na.num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
1356 	return (netmap_attach(&na));
1357 }
1358 
1359 static int
1360 iflib_netmap_txq_init(if_ctx_t ctx, iflib_txq_t txq)
1361 {
1362 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1363 	struct netmap_slot *slot;
1364 
1365 	slot = netmap_reset(na, NR_TX, txq->ift_id, 0);
1366 	if (slot == NULL)
1367 		return (0);
1368 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxd[0]; i++) {
1369 		/*
1370 		 * In netmap mode, set the map for the packet buffer.
1371 		 * NOTE: Some drivers (not this one) also need to set
1372 		 * the physical buffer address in the NIC ring.
1373 		 * netmap_idx_n2k() maps a nic index, i, into the corresponding
1374 		 * netmap slot index, si
1375 		 */
1376 		int si = netmap_idx_n2k(na->tx_rings[txq->ift_id], i);
1377 		netmap_load_map(na, txq->ift_buf_tag, txq->ift_sds.ifsd_map[i],
1378 		    NMB(na, slot + si));
1379 	}
1380 	return (1);
1381 }
1382 
1383 static int
1384 iflib_netmap_rxq_init(if_ctx_t ctx, iflib_rxq_t rxq)
1385 {
1386 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1387 	struct netmap_kring *kring;
1388 	struct netmap_slot *slot;
1389 
1390 	slot = netmap_reset(na, NR_RX, rxq->ifr_id, 0);
1391 	if (slot == NULL)
1392 		return (0);
1393 	kring = na->rx_rings[rxq->ifr_id];
1394 	netmap_fl_refill(rxq, kring, true);
1395 	return (1);
1396 }
1397 
1398 static void
1399 iflib_netmap_timer(void *arg)
1400 {
1401 	iflib_txq_t txq = arg;
1402 	if_ctx_t ctx = txq->ift_ctx;
1403 
1404 	/*
1405 	 * Wake up the netmap application, to give it a chance to
1406 	 * call txsync and reclaim more completed TX buffers.
1407 	 */
1408 	netmap_tx_irq(ctx->ifc_ifp, txq->ift_id);
1409 }
1410 
1411 #define iflib_netmap_detach(ifp) netmap_detach(ifp)
1412 
1413 #else
1414 #define iflib_netmap_txq_init(ctx, txq) (0)
1415 #define iflib_netmap_rxq_init(ctx, rxq) (0)
1416 #define iflib_netmap_detach(ifp)
1417 #define netmap_enable_all_rings(ifp)
1418 #define netmap_disable_all_rings(ifp)
1419 
1420 #define iflib_netmap_attach(ctx) (0)
1421 #define netmap_rx_irq(ifp, qid, budget) (0)
1422 #endif
1423 
1424 #if defined(__i386__) || defined(__amd64__)
1425 static __inline void
1426 prefetch(void *x)
1427 {
1428 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1429 }
1430 static __inline void
1431 prefetch2cachelines(void *x)
1432 {
1433 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1434 #if (CACHE_LINE_SIZE < 128)
1435 	__asm volatile("prefetcht0 %0" :: "m" (*(((unsigned long *)x)+CACHE_LINE_SIZE/(sizeof(unsigned long)))));
1436 #endif
1437 }
1438 #else
1439 #define prefetch(x)
1440 #define prefetch2cachelines(x)
1441 #endif
1442 
1443 static void
1444 iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid)
1445 {
1446 	iflib_fl_t fl;
1447 
1448 	fl = &rxq->ifr_fl[flid];
1449 	iru->iru_paddrs = fl->ifl_bus_addrs;
1450 	iru->iru_idxs = fl->ifl_rxd_idxs;
1451 	iru->iru_qsidx = rxq->ifr_id;
1452 	iru->iru_buf_size = fl->ifl_buf_size;
1453 	iru->iru_flidx = fl->ifl_id;
1454 }
1455 
1456 static void
1457 _iflib_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int err)
1458 {
1459 	if (err)
1460 		return;
1461 	*(bus_addr_t *) arg = segs[0].ds_addr;
1462 }
1463 
1464 int
1465 iflib_dma_alloc_align(if_ctx_t ctx, int size, int align, iflib_dma_info_t dma, int mapflags)
1466 {
1467 	int err;
1468 	device_t dev = ctx->ifc_dev;
1469 
1470 	err = bus_dma_tag_create(bus_get_dma_tag(dev),	/* parent */
1471 				align, 0,		/* alignment, bounds */
1472 				BUS_SPACE_MAXADDR,	/* lowaddr */
1473 				BUS_SPACE_MAXADDR,	/* highaddr */
1474 				NULL, NULL,		/* filter, filterarg */
1475 				size,			/* maxsize */
1476 				1,			/* nsegments */
1477 				size,			/* maxsegsize */
1478 				BUS_DMA_ALLOCNOW,	/* flags */
1479 				NULL,			/* lockfunc */
1480 				NULL,			/* lockarg */
1481 				&dma->idi_tag);
1482 	if (err) {
1483 		device_printf(dev,
1484 		    "%s: bus_dma_tag_create failed: %d\n",
1485 		    __func__, err);
1486 		goto fail_0;
1487 	}
1488 
1489 	err = bus_dmamem_alloc(dma->idi_tag, (void**) &dma->idi_vaddr,
1490 	    BUS_DMA_NOWAIT | BUS_DMA_COHERENT | BUS_DMA_ZERO, &dma->idi_map);
1491 	if (err) {
1492 		device_printf(dev,
1493 		    "%s: bus_dmamem_alloc(%ju) failed: %d\n",
1494 		    __func__, (uintmax_t)size, err);
1495 		goto fail_1;
1496 	}
1497 
1498 	dma->idi_paddr = IF_BAD_DMA;
1499 	err = bus_dmamap_load(dma->idi_tag, dma->idi_map, dma->idi_vaddr,
1500 	    size, _iflib_dmamap_cb, &dma->idi_paddr, mapflags | BUS_DMA_NOWAIT);
1501 	if (err || dma->idi_paddr == IF_BAD_DMA) {
1502 		device_printf(dev,
1503 		    "%s: bus_dmamap_load failed: %d\n",
1504 		    __func__, err);
1505 		goto fail_2;
1506 	}
1507 
1508 	dma->idi_size = size;
1509 	return (0);
1510 
1511 fail_2:
1512 	bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1513 fail_1:
1514 	bus_dma_tag_destroy(dma->idi_tag);
1515 fail_0:
1516 	dma->idi_tag = NULL;
1517 
1518 	return (err);
1519 }
1520 
1521 int
1522 iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags)
1523 {
1524 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1525 
1526 	KASSERT(sctx->isc_q_align != 0, ("alignment value not initialized"));
1527 
1528 	return (iflib_dma_alloc_align(ctx, size, sctx->isc_q_align, dma, mapflags));
1529 }
1530 
1531 int
1532 iflib_dma_alloc_multi(if_ctx_t ctx, int *sizes, iflib_dma_info_t *dmalist, int mapflags, int count)
1533 {
1534 	int i, err;
1535 	iflib_dma_info_t *dmaiter;
1536 
1537 	dmaiter = dmalist;
1538 	for (i = 0; i < count; i++, dmaiter++) {
1539 		if ((err = iflib_dma_alloc(ctx, sizes[i], *dmaiter, mapflags)) != 0)
1540 			break;
1541 	}
1542 	if (err)
1543 		iflib_dma_free_multi(dmalist, i);
1544 	return (err);
1545 }
1546 
1547 void
1548 iflib_dma_free(iflib_dma_info_t dma)
1549 {
1550 	if (dma->idi_tag == NULL)
1551 		return;
1552 	if (dma->idi_paddr != IF_BAD_DMA) {
1553 		bus_dmamap_sync(dma->idi_tag, dma->idi_map,
1554 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1555 		bus_dmamap_unload(dma->idi_tag, dma->idi_map);
1556 		dma->idi_paddr = IF_BAD_DMA;
1557 	}
1558 	if (dma->idi_vaddr != NULL) {
1559 		bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1560 		dma->idi_vaddr = NULL;
1561 	}
1562 	bus_dma_tag_destroy(dma->idi_tag);
1563 	dma->idi_tag = NULL;
1564 }
1565 
1566 void
1567 iflib_dma_free_multi(iflib_dma_info_t *dmalist, int count)
1568 {
1569 	int i;
1570 	iflib_dma_info_t *dmaiter = dmalist;
1571 
1572 	for (i = 0; i < count; i++, dmaiter++)
1573 		iflib_dma_free(*dmaiter);
1574 }
1575 
1576 static int
1577 iflib_fast_intr(void *arg)
1578 {
1579 	iflib_filter_info_t info = arg;
1580 	struct grouptask *gtask = info->ifi_task;
1581 	int result;
1582 
1583 	DBG_COUNTER_INC(fast_intrs);
1584 	if (info->ifi_filter != NULL) {
1585 		result = info->ifi_filter(info->ifi_filter_arg);
1586 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1587 			return (result);
1588 	}
1589 
1590 	GROUPTASK_ENQUEUE(gtask);
1591 	return (FILTER_HANDLED);
1592 }
1593 
1594 static int
1595 iflib_fast_intr_rxtx(void *arg)
1596 {
1597 	iflib_filter_info_t info = arg;
1598 	struct grouptask *gtask = info->ifi_task;
1599 	if_ctx_t ctx;
1600 	iflib_rxq_t rxq = (iflib_rxq_t)info->ifi_ctx;
1601 	iflib_txq_t txq;
1602 	void *sc;
1603 	int i, cidx, result;
1604 	qidx_t txqid;
1605 	bool intr_enable, intr_legacy;
1606 
1607 	DBG_COUNTER_INC(fast_intrs);
1608 	if (info->ifi_filter != NULL) {
1609 		result = info->ifi_filter(info->ifi_filter_arg);
1610 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1611 			return (result);
1612 	}
1613 
1614 	ctx = rxq->ifr_ctx;
1615 	sc = ctx->ifc_softc;
1616 	intr_enable = false;
1617 	intr_legacy = !!(ctx->ifc_flags & IFC_LEGACY);
1618 	MPASS(rxq->ifr_ntxqirq);
1619 	for (i = 0; i < rxq->ifr_ntxqirq; i++) {
1620 		txqid = rxq->ifr_txqid[i];
1621 		txq = &ctx->ifc_txqs[txqid];
1622 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1623 		    BUS_DMASYNC_POSTREAD);
1624 		if (!ctx->isc_txd_credits_update(sc, txqid, false)) {
1625 			if (intr_legacy)
1626 				intr_enable = true;
1627 			else
1628 				IFDI_TX_QUEUE_INTR_ENABLE(ctx, txqid);
1629 			continue;
1630 		}
1631 		GROUPTASK_ENQUEUE(&txq->ift_task);
1632 	}
1633 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_RXCQ)
1634 		cidx = rxq->ifr_cq_cidx;
1635 	else
1636 		cidx = rxq->ifr_fl[0].ifl_cidx;
1637 	if (iflib_rxd_avail(ctx, rxq, cidx, 1))
1638 		GROUPTASK_ENQUEUE(gtask);
1639 	else {
1640 		if (intr_legacy)
1641 			intr_enable = true;
1642 		else
1643 			IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
1644 		DBG_COUNTER_INC(rx_intr_enables);
1645 	}
1646 	if (intr_enable)
1647 		IFDI_INTR_ENABLE(ctx);
1648 	return (FILTER_HANDLED);
1649 }
1650 
1651 static int
1652 iflib_fast_intr_ctx(void *arg)
1653 {
1654 	iflib_filter_info_t info = arg;
1655 	struct grouptask *gtask = info->ifi_task;
1656 	int result;
1657 
1658 	DBG_COUNTER_INC(fast_intrs);
1659 	if (info->ifi_filter != NULL) {
1660 		result = info->ifi_filter(info->ifi_filter_arg);
1661 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1662 			return (result);
1663 	}
1664 
1665 	GROUPTASK_ENQUEUE(gtask);
1666 	return (FILTER_HANDLED);
1667 }
1668 
1669 static int
1670 _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
1671 		 driver_filter_t filter, driver_intr_t handler, void *arg,
1672 		 const char *name)
1673 {
1674 	struct resource *res;
1675 	void *tag = NULL;
1676 	device_t dev = ctx->ifc_dev;
1677 	int flags, i, rc;
1678 
1679 	flags = RF_ACTIVE;
1680 	if (ctx->ifc_flags & IFC_LEGACY)
1681 		flags |= RF_SHAREABLE;
1682 	MPASS(rid < 512);
1683 	i = rid;
1684 	res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &i, flags);
1685 	if (res == NULL) {
1686 		device_printf(dev,
1687 		    "failed to allocate IRQ for rid %d, name %s.\n", rid, name);
1688 		return (ENOMEM);
1689 	}
1690 	irq->ii_res = res;
1691 	KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL"));
1692 	rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET,
1693 						filter, handler, arg, &tag);
1694 	if (rc != 0) {
1695 		device_printf(dev,
1696 		    "failed to setup interrupt for rid %d, name %s: %d\n",
1697 					  rid, name ? name : "unknown", rc);
1698 		return (rc);
1699 	} else if (name)
1700 		bus_describe_intr(dev, res, tag, "%s", name);
1701 
1702 	irq->ii_tag = tag;
1703 	return (0);
1704 }
1705 
1706 /*********************************************************************
1707  *
1708  *  Allocate DMA resources for TX buffers as well as memory for the TX
1709  *  mbuf map.  TX DMA maps (non-TSO/TSO) and TX mbuf map are kept in a
1710  *  iflib_sw_tx_desc_array structure, storing all the information that
1711  *  is needed to transmit a packet on the wire.  This is called only
1712  *  once at attach, setup is done every reset.
1713  *
1714  **********************************************************************/
1715 static int
1716 iflib_txsd_alloc(iflib_txq_t txq)
1717 {
1718 	if_ctx_t ctx = txq->ift_ctx;
1719 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1720 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1721 	device_t dev = ctx->ifc_dev;
1722 	bus_size_t tsomaxsize;
1723 	int err, nsegments, ntsosegments;
1724 	bool tso;
1725 
1726 	nsegments = scctx->isc_tx_nsegments;
1727 	ntsosegments = scctx->isc_tx_tso_segments_max;
1728 	tsomaxsize = scctx->isc_tx_tso_size_max;
1729 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_VLAN_MTU)
1730 		tsomaxsize += sizeof(struct ether_vlan_header);
1731 	MPASS(scctx->isc_ntxd[0] > 0);
1732 	MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0);
1733 	MPASS(nsegments > 0);
1734 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) {
1735 		MPASS(ntsosegments > 0);
1736 		MPASS(sctx->isc_tso_maxsize >= tsomaxsize);
1737 	}
1738 
1739 	/*
1740 	 * Set up DMA tags for TX buffers.
1741 	 */
1742 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1743 			       1, 0,			/* alignment, bounds */
1744 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1745 			       BUS_SPACE_MAXADDR,	/* highaddr */
1746 			       NULL, NULL,		/* filter, filterarg */
1747 			       sctx->isc_tx_maxsize,		/* maxsize */
1748 			       nsegments,	/* nsegments */
1749 			       sctx->isc_tx_maxsegsize,	/* maxsegsize */
1750 			       0,			/* flags */
1751 			       NULL,			/* lockfunc */
1752 			       NULL,			/* lockfuncarg */
1753 			       &txq->ift_buf_tag))) {
1754 		device_printf(dev,"Unable to allocate TX DMA tag: %d\n", err);
1755 		device_printf(dev,"maxsize: %ju nsegments: %d maxsegsize: %ju\n",
1756 		    (uintmax_t)sctx->isc_tx_maxsize, nsegments, (uintmax_t)sctx->isc_tx_maxsegsize);
1757 		goto fail;
1758 	}
1759 	tso = (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) != 0;
1760 	if (tso && (err = bus_dma_tag_create(bus_get_dma_tag(dev),
1761 			       1, 0,			/* alignment, bounds */
1762 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1763 			       BUS_SPACE_MAXADDR,	/* highaddr */
1764 			       NULL, NULL,		/* filter, filterarg */
1765 			       tsomaxsize,		/* maxsize */
1766 			       ntsosegments,	/* nsegments */
1767 			       sctx->isc_tso_maxsegsize,/* maxsegsize */
1768 			       0,			/* flags */
1769 			       NULL,			/* lockfunc */
1770 			       NULL,			/* lockfuncarg */
1771 			       &txq->ift_tso_buf_tag))) {
1772 		device_printf(dev, "Unable to allocate TSO TX DMA tag: %d\n",
1773 		    err);
1774 		goto fail;
1775 	}
1776 
1777 	/* Allocate memory for the TX mbuf map. */
1778 	if (!(txq->ift_sds.ifsd_m =
1779 	    (struct mbuf **) malloc(sizeof(struct mbuf *) *
1780 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1781 		device_printf(dev, "Unable to allocate TX mbuf map memory\n");
1782 		err = ENOMEM;
1783 		goto fail;
1784 	}
1785 
1786 	/*
1787 	 * Create the DMA maps for TX buffers.
1788 	 */
1789 	if ((txq->ift_sds.ifsd_map = (bus_dmamap_t *)malloc(
1790 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1791 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1792 		device_printf(dev,
1793 		    "Unable to allocate TX buffer DMA map memory\n");
1794 		err = ENOMEM;
1795 		goto fail;
1796 	}
1797 	if (tso && (txq->ift_sds.ifsd_tso_map = (bus_dmamap_t *)malloc(
1798 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1799 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1800 		device_printf(dev,
1801 		    "Unable to allocate TSO TX buffer map memory\n");
1802 		err = ENOMEM;
1803 		goto fail;
1804 	}
1805 	for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) {
1806 		err = bus_dmamap_create(txq->ift_buf_tag, 0,
1807 		    &txq->ift_sds.ifsd_map[i]);
1808 		if (err != 0) {
1809 			device_printf(dev, "Unable to create TX DMA map\n");
1810 			goto fail;
1811 		}
1812 		if (!tso)
1813 			continue;
1814 		err = bus_dmamap_create(txq->ift_tso_buf_tag, 0,
1815 		    &txq->ift_sds.ifsd_tso_map[i]);
1816 		if (err != 0) {
1817 			device_printf(dev, "Unable to create TSO TX DMA map\n");
1818 			goto fail;
1819 		}
1820 	}
1821 	return (0);
1822 fail:
1823 	/* We free all, it handles case where we are in the middle */
1824 	iflib_tx_structures_free(ctx);
1825 	return (err);
1826 }
1827 
1828 static void
1829 iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i)
1830 {
1831 	bus_dmamap_t map;
1832 
1833 	if (txq->ift_sds.ifsd_map != NULL) {
1834 		map = txq->ift_sds.ifsd_map[i];
1835 		bus_dmamap_sync(txq->ift_buf_tag, map, BUS_DMASYNC_POSTWRITE);
1836 		bus_dmamap_unload(txq->ift_buf_tag, map);
1837 		bus_dmamap_destroy(txq->ift_buf_tag, map);
1838 		txq->ift_sds.ifsd_map[i] = NULL;
1839 	}
1840 
1841 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1842 		map = txq->ift_sds.ifsd_tso_map[i];
1843 		bus_dmamap_sync(txq->ift_tso_buf_tag, map,
1844 		    BUS_DMASYNC_POSTWRITE);
1845 		bus_dmamap_unload(txq->ift_tso_buf_tag, map);
1846 		bus_dmamap_destroy(txq->ift_tso_buf_tag, map);
1847 		txq->ift_sds.ifsd_tso_map[i] = NULL;
1848 	}
1849 }
1850 
1851 static void
1852 iflib_txq_destroy(iflib_txq_t txq)
1853 {
1854 	if_ctx_t ctx = txq->ift_ctx;
1855 
1856 	for (int i = 0; i < txq->ift_size; i++)
1857 		iflib_txsd_destroy(ctx, txq, i);
1858 
1859 	if (txq->ift_br != NULL) {
1860 		ifmp_ring_free(txq->ift_br);
1861 		txq->ift_br = NULL;
1862 	}
1863 
1864 	mtx_destroy(&txq->ift_mtx);
1865 
1866 	if (txq->ift_sds.ifsd_map != NULL) {
1867 		free(txq->ift_sds.ifsd_map, M_IFLIB);
1868 		txq->ift_sds.ifsd_map = NULL;
1869 	}
1870 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1871 		free(txq->ift_sds.ifsd_tso_map, M_IFLIB);
1872 		txq->ift_sds.ifsd_tso_map = NULL;
1873 	}
1874 	if (txq->ift_sds.ifsd_m != NULL) {
1875 		free(txq->ift_sds.ifsd_m, M_IFLIB);
1876 		txq->ift_sds.ifsd_m = NULL;
1877 	}
1878 	if (txq->ift_buf_tag != NULL) {
1879 		bus_dma_tag_destroy(txq->ift_buf_tag);
1880 		txq->ift_buf_tag = NULL;
1881 	}
1882 	if (txq->ift_tso_buf_tag != NULL) {
1883 		bus_dma_tag_destroy(txq->ift_tso_buf_tag);
1884 		txq->ift_tso_buf_tag = NULL;
1885 	}
1886 	if (txq->ift_ifdi != NULL) {
1887 		free(txq->ift_ifdi, M_IFLIB);
1888 	}
1889 }
1890 
1891 static void
1892 iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i)
1893 {
1894 	struct mbuf **mp;
1895 
1896 	mp = &txq->ift_sds.ifsd_m[i];
1897 	if (*mp == NULL)
1898 		return;
1899 
1900 	if (txq->ift_sds.ifsd_map != NULL) {
1901 		bus_dmamap_sync(txq->ift_buf_tag,
1902 		    txq->ift_sds.ifsd_map[i], BUS_DMASYNC_POSTWRITE);
1903 		bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[i]);
1904 	}
1905 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1906 		bus_dmamap_sync(txq->ift_tso_buf_tag,
1907 		    txq->ift_sds.ifsd_tso_map[i], BUS_DMASYNC_POSTWRITE);
1908 		bus_dmamap_unload(txq->ift_tso_buf_tag,
1909 		    txq->ift_sds.ifsd_tso_map[i]);
1910 	}
1911 	m_freem(*mp);
1912 	DBG_COUNTER_INC(tx_frees);
1913 	*mp = NULL;
1914 }
1915 
1916 static int
1917 iflib_txq_setup(iflib_txq_t txq)
1918 {
1919 	if_ctx_t ctx = txq->ift_ctx;
1920 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1921 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1922 	iflib_dma_info_t di;
1923 	int i;
1924 
1925 	/* Set number of descriptors available */
1926 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
1927 	/* XXX make configurable */
1928 	txq->ift_update_freq = IFLIB_DEFAULT_TX_UPDATE_FREQ;
1929 
1930 	/* Reset indices */
1931 	txq->ift_cidx_processed = 0;
1932 	txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0;
1933 	txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset];
1934 
1935 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1936 		bzero((void *)di->idi_vaddr, di->idi_size);
1937 
1938 	IFDI_TXQ_SETUP(ctx, txq->ift_id);
1939 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1940 		bus_dmamap_sync(di->idi_tag, di->idi_map,
1941 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1942 	return (0);
1943 }
1944 
1945 /*********************************************************************
1946  *
1947  *  Allocate DMA resources for RX buffers as well as memory for the RX
1948  *  mbuf map, direct RX cluster pointer map and RX cluster bus address
1949  *  map.  RX DMA map, RX mbuf map, direct RX cluster pointer map and
1950  *  RX cluster map are kept in a iflib_sw_rx_desc_array structure.
1951  *  Since we use use one entry in iflib_sw_rx_desc_array per received
1952  *  packet, the maximum number of entries we'll need is equal to the
1953  *  number of hardware receive descriptors that we've allocated.
1954  *
1955  **********************************************************************/
1956 static int
1957 iflib_rxsd_alloc(iflib_rxq_t rxq)
1958 {
1959 	if_ctx_t ctx = rxq->ifr_ctx;
1960 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1961 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1962 	device_t dev = ctx->ifc_dev;
1963 	iflib_fl_t fl;
1964 	int			err;
1965 
1966 	MPASS(scctx->isc_nrxd[0] > 0);
1967 	MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0);
1968 
1969 	fl = rxq->ifr_fl;
1970 	for (int i = 0; i <  rxq->ifr_nfl; i++, fl++) {
1971 		fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */
1972 		/* Set up DMA tag for RX buffers. */
1973 		err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1974 					 1, 0,			/* alignment, bounds */
1975 					 BUS_SPACE_MAXADDR,	/* lowaddr */
1976 					 BUS_SPACE_MAXADDR,	/* highaddr */
1977 					 NULL, NULL,		/* filter, filterarg */
1978 					 sctx->isc_rx_maxsize,	/* maxsize */
1979 					 sctx->isc_rx_nsegments,	/* nsegments */
1980 					 sctx->isc_rx_maxsegsize,	/* maxsegsize */
1981 					 0,			/* flags */
1982 					 NULL,			/* lockfunc */
1983 					 NULL,			/* lockarg */
1984 					 &fl->ifl_buf_tag);
1985 		if (err) {
1986 			device_printf(dev,
1987 			    "Unable to allocate RX DMA tag: %d\n", err);
1988 			goto fail;
1989 		}
1990 
1991 		/* Allocate memory for the RX mbuf map. */
1992 		if (!(fl->ifl_sds.ifsd_m =
1993 		      (struct mbuf **) malloc(sizeof(struct mbuf *) *
1994 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1995 			device_printf(dev,
1996 			    "Unable to allocate RX mbuf map memory\n");
1997 			err = ENOMEM;
1998 			goto fail;
1999 		}
2000 
2001 		/* Allocate memory for the direct RX cluster pointer map. */
2002 		if (!(fl->ifl_sds.ifsd_cl =
2003 		      (caddr_t *) malloc(sizeof(caddr_t) *
2004 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
2005 			device_printf(dev,
2006 			    "Unable to allocate RX cluster map memory\n");
2007 			err = ENOMEM;
2008 			goto fail;
2009 		}
2010 
2011 		/* Allocate memory for the RX cluster bus address map. */
2012 		if (!(fl->ifl_sds.ifsd_ba =
2013 		      (bus_addr_t *) malloc(sizeof(bus_addr_t) *
2014 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
2015 			device_printf(dev,
2016 			    "Unable to allocate RX bus address map memory\n");
2017 			err = ENOMEM;
2018 			goto fail;
2019 		}
2020 
2021 		/*
2022 		 * Create the DMA maps for RX buffers.
2023 		 */
2024 		if (!(fl->ifl_sds.ifsd_map =
2025 		      (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
2026 			device_printf(dev,
2027 			    "Unable to allocate RX buffer DMA map memory\n");
2028 			err = ENOMEM;
2029 			goto fail;
2030 		}
2031 		for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++) {
2032 			err = bus_dmamap_create(fl->ifl_buf_tag, 0,
2033 			    &fl->ifl_sds.ifsd_map[i]);
2034 			if (err != 0) {
2035 				device_printf(dev, "Unable to create RX buffer DMA map\n");
2036 				goto fail;
2037 			}
2038 		}
2039 	}
2040 	return (0);
2041 
2042 fail:
2043 	iflib_rx_structures_free(ctx);
2044 	return (err);
2045 }
2046 
2047 /*
2048  * Internal service routines
2049  */
2050 
2051 struct rxq_refill_cb_arg {
2052 	int               error;
2053 	bus_dma_segment_t seg;
2054 	int               nseg;
2055 };
2056 
2057 static void
2058 _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
2059 {
2060 	struct rxq_refill_cb_arg *cb_arg = arg;
2061 
2062 	cb_arg->error = error;
2063 	cb_arg->seg = segs[0];
2064 	cb_arg->nseg = nseg;
2065 }
2066 
2067 /**
2068  * iflib_fl_refill - refill an rxq free-buffer list
2069  * @ctx: the iflib context
2070  * @fl: the free list to refill
2071  * @count: the number of new buffers to allocate
2072  *
2073  * (Re)populate an rxq free-buffer list with up to @count new packet buffers.
2074  * The caller must assure that @count does not exceed the queue's capacity
2075  * minus one (since we always leave a descriptor unavailable).
2076  */
2077 static uint8_t
2078 iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
2079 {
2080 	struct if_rxd_update iru;
2081 	struct rxq_refill_cb_arg cb_arg;
2082 	struct mbuf *m;
2083 	caddr_t cl, *sd_cl;
2084 	struct mbuf **sd_m;
2085 	bus_dmamap_t *sd_map;
2086 	bus_addr_t bus_addr, *sd_ba;
2087 	int err, frag_idx, i, idx, n, pidx;
2088 	qidx_t credits;
2089 
2090 	MPASS(count <= fl->ifl_size - fl->ifl_credits - 1);
2091 
2092 	sd_m = fl->ifl_sds.ifsd_m;
2093 	sd_map = fl->ifl_sds.ifsd_map;
2094 	sd_cl = fl->ifl_sds.ifsd_cl;
2095 	sd_ba = fl->ifl_sds.ifsd_ba;
2096 	pidx = fl->ifl_pidx;
2097 	idx = pidx;
2098 	frag_idx = fl->ifl_fragidx;
2099 	credits = fl->ifl_credits;
2100 
2101 	i = 0;
2102 	n = count;
2103 	MPASS(n > 0);
2104 	MPASS(credits + n <= fl->ifl_size);
2105 
2106 	if (pidx < fl->ifl_cidx)
2107 		MPASS(pidx + n <= fl->ifl_cidx);
2108 	if (pidx == fl->ifl_cidx && (credits < fl->ifl_size))
2109 		MPASS(fl->ifl_gen == 0);
2110 	if (pidx > fl->ifl_cidx)
2111 		MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx);
2112 
2113 	DBG_COUNTER_INC(fl_refills);
2114 	if (n > 8)
2115 		DBG_COUNTER_INC(fl_refills_large);
2116 	iru_init(&iru, fl->ifl_rxq, fl->ifl_id);
2117 	while (n-- > 0) {
2118 		/*
2119 		 * We allocate an uninitialized mbuf + cluster, mbuf is
2120 		 * initialized after rx.
2121 		 *
2122 		 * If the cluster is still set then we know a minimum sized
2123 		 * packet was received
2124 		 */
2125 		bit_ffc_at(fl->ifl_rx_bitmap, frag_idx, fl->ifl_size,
2126 		    &frag_idx);
2127 		if (frag_idx < 0)
2128 			bit_ffc(fl->ifl_rx_bitmap, fl->ifl_size, &frag_idx);
2129 		MPASS(frag_idx >= 0);
2130 		if ((cl = sd_cl[frag_idx]) == NULL) {
2131 #ifndef __HAIKU__
2132 			cl = uma_zalloc(fl->ifl_zone, M_NOWAIT);
2133 			if (__predict_false(cl == NULL))
2134 #else
2135 			if ((cl = m_cljget(NULL, M_NOWAIT, fl->ifl_buf_size)) == NULL)
2136 #endif
2137 				break;
2138 
2139 			cb_arg.error = 0;
2140 			MPASS(sd_map != NULL);
2141 			err = bus_dmamap_load(fl->ifl_buf_tag, sd_map[frag_idx],
2142 			    cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg,
2143 			    BUS_DMA_NOWAIT);
2144 			if (__predict_false(err != 0 || cb_arg.error)) {
2145 #ifndef __HAIKU__
2146 				uma_zfree(fl->ifl_zone, cl);
2147 #else
2148 				m_free(cl);
2149 #endif
2150 				break;
2151 			}
2152 
2153 			sd_ba[frag_idx] = bus_addr = cb_arg.seg.ds_addr;
2154 			sd_cl[frag_idx] = cl;
2155 #if MEMORY_LOGGING
2156 			fl->ifl_cl_enqueued++;
2157 #endif
2158 		} else {
2159 			bus_addr = sd_ba[frag_idx];
2160 		}
2161 		bus_dmamap_sync(fl->ifl_buf_tag, sd_map[frag_idx],
2162 		    BUS_DMASYNC_PREREAD);
2163 
2164 		if (sd_m[frag_idx] == NULL) {
2165 			m = m_gethdr(M_NOWAIT, MT_NOINIT);
2166 			if (__predict_false(m == NULL))
2167 				break;
2168 			sd_m[frag_idx] = m;
2169 		}
2170 		bit_set(fl->ifl_rx_bitmap, frag_idx);
2171 #if MEMORY_LOGGING
2172 		fl->ifl_m_enqueued++;
2173 #endif
2174 
2175 		DBG_COUNTER_INC(rx_allocs);
2176 		fl->ifl_rxd_idxs[i] = frag_idx;
2177 		fl->ifl_bus_addrs[i] = bus_addr;
2178 		credits++;
2179 		i++;
2180 		MPASS(credits <= fl->ifl_size);
2181 		if (++idx == fl->ifl_size) {
2182 #ifdef INVARIANTS
2183 			fl->ifl_gen = 1;
2184 #endif
2185 			idx = 0;
2186 		}
2187 		if (n == 0 || i == IFLIB_MAX_RX_REFRESH) {
2188 			iru.iru_pidx = pidx;
2189 			iru.iru_count = i;
2190 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2191 			fl->ifl_pidx = idx;
2192 			fl->ifl_credits = credits;
2193 			pidx = idx;
2194 			i = 0;
2195 		}
2196 	}
2197 
2198 	if (n < count - 1) {
2199 		if (i != 0) {
2200 			iru.iru_pidx = pidx;
2201 			iru.iru_count = i;
2202 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2203 			fl->ifl_pidx = idx;
2204 			fl->ifl_credits = credits;
2205 		}
2206 		DBG_COUNTER_INC(rxd_flush);
2207 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2208 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2209 		ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id,
2210 		    fl->ifl_id, fl->ifl_pidx);
2211 		if (__predict_true(bit_test(fl->ifl_rx_bitmap, frag_idx))) {
2212 			fl->ifl_fragidx = frag_idx + 1;
2213 			if (fl->ifl_fragidx == fl->ifl_size)
2214 				fl->ifl_fragidx = 0;
2215 		} else {
2216 			fl->ifl_fragidx = frag_idx;
2217 		}
2218 	}
2219 
2220 	return (n == -1 ? 0 : IFLIB_RXEOF_EMPTY);
2221 }
2222 
2223 static inline uint8_t
2224 iflib_fl_refill_all(if_ctx_t ctx, iflib_fl_t fl)
2225 {
2226 	/*
2227 	 * We leave an unused descriptor to avoid pidx to catch up with cidx.
2228 	 * This is important as it confuses most NICs. For instance,
2229 	 * Intel NICs have (per receive ring) RDH and RDT registers, where
2230 	 * RDH points to the next receive descriptor to be used by the NIC,
2231 	 * and RDT for the next receive descriptor to be published by the
2232 	 * driver to the NIC (RDT - 1 is thus the last valid one).
2233 	 * The condition RDH == RDT means no descriptors are available to
2234 	 * the NIC, and thus it would be ambiguous if it also meant that
2235 	 * all the descriptors are available to the NIC.
2236 	 */
2237 	int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1;
2238 #ifdef INVARIANTS
2239 	int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1;
2240 #endif
2241 
2242 	MPASS(fl->ifl_credits <= fl->ifl_size);
2243 	MPASS(reclaimable == delta);
2244 
2245 	if (reclaimable > 0)
2246 		return (iflib_fl_refill(ctx, fl, reclaimable));
2247 	return (0);
2248 }
2249 
2250 uint8_t
2251 iflib_in_detach(if_ctx_t ctx)
2252 {
2253 	bool in_detach;
2254 
2255 	STATE_LOCK(ctx);
2256 	in_detach = !!(ctx->ifc_flags & IFC_IN_DETACH);
2257 	STATE_UNLOCK(ctx);
2258 	return (in_detach);
2259 }
2260 
2261 static void
2262 iflib_fl_bufs_free(iflib_fl_t fl)
2263 {
2264 	iflib_dma_info_t idi = fl->ifl_ifdi;
2265 	bus_dmamap_t sd_map;
2266 	uint32_t i;
2267 
2268 	for (i = 0; i < fl->ifl_size; i++) {
2269 		struct mbuf **sd_m = &fl->ifl_sds.ifsd_m[i];
2270 		caddr_t *sd_cl = &fl->ifl_sds.ifsd_cl[i];
2271 
2272 		if (*sd_cl != NULL) {
2273 			sd_map = fl->ifl_sds.ifsd_map[i];
2274 			bus_dmamap_sync(fl->ifl_buf_tag, sd_map,
2275 			    BUS_DMASYNC_POSTREAD);
2276 			bus_dmamap_unload(fl->ifl_buf_tag, sd_map);
2277 #ifndef __HAIKU__
2278 			uma_zfree(fl->ifl_zone, *sd_cl);
2279 #else
2280 			struct mbuf* mb = m_get(0, MT_DATA);
2281 			m_cljset(mb, *sd_cl, fl->ifl_cltype);
2282 			m_free(mb);
2283 #endif
2284 			*sd_cl = NULL;
2285 			if (*sd_m != NULL) {
2286 				m_init(*sd_m, M_NOWAIT, MT_DATA, 0);
2287 #ifndef __HAIKU__
2288 				m_free_raw(*sd_m);
2289 #else
2290 				m_free(*sd_m);
2291 #endif
2292 				*sd_m = NULL;
2293 			}
2294 		} else {
2295 			MPASS(*sd_m == NULL);
2296 		}
2297 #if MEMORY_LOGGING
2298 		fl->ifl_m_dequeued++;
2299 		fl->ifl_cl_dequeued++;
2300 #endif
2301 	}
2302 #ifdef INVARIANTS
2303 	for (i = 0; i < fl->ifl_size; i++) {
2304 		MPASS(fl->ifl_sds.ifsd_cl[i] == NULL);
2305 		MPASS(fl->ifl_sds.ifsd_m[i] == NULL);
2306 	}
2307 #endif
2308 	/*
2309 	 * Reset free list values
2310 	 */
2311 	fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = fl->ifl_fragidx = 0;
2312 	bzero(idi->idi_vaddr, idi->idi_size);
2313 }
2314 
2315 /*********************************************************************
2316  *
2317  *  Initialize a free list and its buffers.
2318  *
2319  **********************************************************************/
2320 static int
2321 iflib_fl_setup(iflib_fl_t fl)
2322 {
2323 	iflib_rxq_t rxq = fl->ifl_rxq;
2324 	if_ctx_t ctx = rxq->ifr_ctx;
2325 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2326 	int qidx;
2327 
2328 	bit_nclear(fl->ifl_rx_bitmap, 0, fl->ifl_size - 1);
2329 	/*
2330 	** Free current RX buffer structs and their mbufs
2331 	*/
2332 	iflib_fl_bufs_free(fl);
2333 	/* Now replenish the mbufs */
2334 	MPASS(fl->ifl_credits == 0);
2335 	qidx = rxq->ifr_fl_offset + fl->ifl_id;
2336 	if (scctx->isc_rxd_buf_size[qidx] != 0)
2337 		fl->ifl_buf_size = scctx->isc_rxd_buf_size[qidx];
2338 	else
2339 		fl->ifl_buf_size = ctx->ifc_rx_mbuf_sz;
2340 	/*
2341 	 * ifl_buf_size may be a driver-supplied value, so pull it up
2342 	 * to the selected mbuf size.
2343 	 */
2344 	fl->ifl_buf_size = iflib_get_mbuf_size_for(fl->ifl_buf_size);
2345 	if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size)
2346 		ctx->ifc_max_fl_buf_size = fl->ifl_buf_size;
2347 	fl->ifl_cltype = m_gettype(fl->ifl_buf_size);
2348 #ifndef __HAIKU__
2349 	fl->ifl_zone = m_getzone(fl->ifl_buf_size);
2350 #endif
2351 
2352 	/*
2353 	 * Avoid pre-allocating zillions of clusters to an idle card
2354 	 * potentially speeding up attach. In any case make sure
2355 	 * to leave a descriptor unavailable. See the comment in
2356 	 * iflib_fl_refill_all().
2357 	 */
2358 	MPASS(fl->ifl_size > 0);
2359 	(void)iflib_fl_refill(ctx, fl, min(128, fl->ifl_size - 1));
2360 	if (min(128, fl->ifl_size - 1) != fl->ifl_credits)
2361 		return (ENOBUFS);
2362 	/*
2363 	 * handle failure
2364 	 */
2365 	MPASS(rxq != NULL);
2366 	MPASS(fl->ifl_ifdi != NULL);
2367 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2368 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2369 	return (0);
2370 }
2371 
2372 /*********************************************************************
2373  *
2374  *  Free receive ring data structures
2375  *
2376  **********************************************************************/
2377 static void
2378 iflib_rx_sds_free(iflib_rxq_t rxq)
2379 {
2380 	iflib_fl_t fl;
2381 	int i, j;
2382 
2383 	if (rxq->ifr_fl != NULL) {
2384 		for (i = 0; i < rxq->ifr_nfl; i++) {
2385 			fl = &rxq->ifr_fl[i];
2386 			if (fl->ifl_buf_tag != NULL) {
2387 				if (fl->ifl_sds.ifsd_map != NULL) {
2388 					for (j = 0; j < fl->ifl_size; j++) {
2389 						bus_dmamap_sync(
2390 						    fl->ifl_buf_tag,
2391 						    fl->ifl_sds.ifsd_map[j],
2392 						    BUS_DMASYNC_POSTREAD);
2393 						bus_dmamap_unload(
2394 						    fl->ifl_buf_tag,
2395 						    fl->ifl_sds.ifsd_map[j]);
2396 						bus_dmamap_destroy(
2397 						    fl->ifl_buf_tag,
2398 						    fl->ifl_sds.ifsd_map[j]);
2399 					}
2400 				}
2401 				bus_dma_tag_destroy(fl->ifl_buf_tag);
2402 				fl->ifl_buf_tag = NULL;
2403 			}
2404 			free(fl->ifl_sds.ifsd_m, M_IFLIB);
2405 			free(fl->ifl_sds.ifsd_cl, M_IFLIB);
2406 			free(fl->ifl_sds.ifsd_ba, M_IFLIB);
2407 			free(fl->ifl_sds.ifsd_map, M_IFLIB);
2408 			free(fl->ifl_rx_bitmap, M_IFLIB);
2409 			fl->ifl_sds.ifsd_m = NULL;
2410 			fl->ifl_sds.ifsd_cl = NULL;
2411 			fl->ifl_sds.ifsd_ba = NULL;
2412 			fl->ifl_sds.ifsd_map = NULL;
2413 			fl->ifl_rx_bitmap = NULL;
2414 		}
2415 		free(rxq->ifr_fl, M_IFLIB);
2416 		rxq->ifr_fl = NULL;
2417 		free(rxq->ifr_ifdi, M_IFLIB);
2418 		rxq->ifr_ifdi = NULL;
2419 		rxq->ifr_cq_cidx = 0;
2420 	}
2421 }
2422 
2423 /*
2424  * Timer routine
2425  */
2426 static void
2427 iflib_timer(void *arg)
2428 {
2429 	iflib_txq_t txq = arg;
2430 	if_ctx_t ctx = txq->ift_ctx;
2431 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2432 	uint64_t this_tick = ticks;
2433 
2434 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
2435 		return;
2436 
2437 	/*
2438 	** Check on the state of the TX queue(s), this
2439 	** can be done without the lock because its RO
2440 	** and the HUNG state will be static if set.
2441 	*/
2442 	if (this_tick - txq->ift_last_timer_tick >= iflib_timer_default) {
2443 		txq->ift_last_timer_tick = this_tick;
2444 		IFDI_TIMER(ctx, txq->ift_id);
2445 		if ((txq->ift_qstatus == IFLIB_QUEUE_HUNG) &&
2446 		    ((txq->ift_cleaned_prev == txq->ift_cleaned) ||
2447 		     (sctx->isc_pause_frames == 0)))
2448 			goto hung;
2449 
2450 		if (txq->ift_qstatus != IFLIB_QUEUE_IDLE &&
2451 		    ifmp_ring_is_stalled(txq->ift_br)) {
2452 			KASSERT(ctx->ifc_link_state == LINK_STATE_UP,
2453 			    ("queue can't be marked as hung if interface is down"));
2454 			txq->ift_qstatus = IFLIB_QUEUE_HUNG;
2455 		}
2456 		txq->ift_cleaned_prev = txq->ift_cleaned;
2457 	}
2458 	/* handle any laggards */
2459 	if (txq->ift_db_pending)
2460 		GROUPTASK_ENQUEUE(&txq->ift_task);
2461 
2462 	sctx->isc_pause_frames = 0;
2463 	if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)
2464 		callout_reset_on(&txq->ift_timer, iflib_timer_default, iflib_timer,
2465 		    txq, txq->ift_timer.c_cpu);
2466 	return;
2467 
2468  hung:
2469 	device_printf(ctx->ifc_dev,
2470 	    "Watchdog timeout (TX: %d desc avail: %d pidx: %d) -- resetting\n",
2471 	    txq->ift_id, TXQ_AVAIL(txq), txq->ift_pidx);
2472 	STATE_LOCK(ctx);
2473 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2474 	ctx->ifc_flags |= (IFC_DO_WATCHDOG|IFC_DO_RESET);
2475 	iflib_admin_intr_deferred(ctx);
2476 	STATE_UNLOCK(ctx);
2477 }
2478 
2479 static uint16_t
2480 iflib_get_mbuf_size_for(unsigned int size)
2481 {
2482 
2483 	if (size <= MCLBYTES)
2484 		return (MCLBYTES);
2485 	else
2486 		return (MJUMPAGESIZE);
2487 }
2488 
2489 static void
2490 iflib_calc_rx_mbuf_sz(if_ctx_t ctx)
2491 {
2492 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2493 
2494 	/*
2495 	 * XXX don't set the max_frame_size to larger
2496 	 * than the hardware can handle
2497 	 */
2498 	ctx->ifc_rx_mbuf_sz =
2499 	    iflib_get_mbuf_size_for(sctx->isc_max_frame_size);
2500 }
2501 
2502 uint32_t
2503 iflib_get_rx_mbuf_sz(if_ctx_t ctx)
2504 {
2505 
2506 	return (ctx->ifc_rx_mbuf_sz);
2507 }
2508 
2509 static void
2510 iflib_init_locked(if_ctx_t ctx)
2511 {
2512 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2513 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2514 	if_t ifp = ctx->ifc_ifp;
2515 	iflib_fl_t fl;
2516 	iflib_txq_t txq;
2517 	iflib_rxq_t rxq;
2518 	int i, j, tx_ip_csum_flags, tx_ip6_csum_flags;
2519 
2520 	if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2521 	IFDI_INTR_DISABLE(ctx);
2522 
2523 	/*
2524 	 * See iflib_stop(). Useful in case iflib_init_locked() is
2525 	 * called without first calling iflib_stop().
2526 	 */
2527 	netmap_disable_all_rings(ifp);
2528 
2529 	tx_ip_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP);
2530 	tx_ip6_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | CSUM_IP6_SCTP);
2531 	/* Set hardware offload abilities */
2532 	if_clearhwassist(ifp);
2533 	if (if_getcapenable(ifp) & IFCAP_TXCSUM)
2534 		if_sethwassistbits(ifp, tx_ip_csum_flags, 0);
2535 	if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6)
2536 		if_sethwassistbits(ifp,  tx_ip6_csum_flags, 0);
2537 	if (if_getcapenable(ifp) & IFCAP_TSO4)
2538 		if_sethwassistbits(ifp, CSUM_IP_TSO, 0);
2539 	if (if_getcapenable(ifp) & IFCAP_TSO6)
2540 		if_sethwassistbits(ifp, CSUM_IP6_TSO, 0);
2541 
2542 	for (i = 0, txq = ctx->ifc_txqs; i < sctx->isc_ntxqsets; i++, txq++) {
2543 		CALLOUT_LOCK(txq);
2544 		callout_stop(&txq->ift_timer);
2545 #ifdef DEV_NETMAP
2546 		callout_stop(&txq->ift_netmap_timer);
2547 #endif /* DEV_NETMAP */
2548 		CALLOUT_UNLOCK(txq);
2549 		iflib_netmap_txq_init(ctx, txq);
2550 	}
2551 
2552 	/*
2553 	 * Calculate a suitable Rx mbuf size prior to calling IFDI_INIT, so
2554 	 * that drivers can use the value when setting up the hardware receive
2555 	 * buffers.
2556 	 */
2557 	iflib_calc_rx_mbuf_sz(ctx);
2558 
2559 #ifdef INVARIANTS
2560 	i = if_getdrvflags(ifp);
2561 #endif
2562 	IFDI_INIT(ctx);
2563 	MPASS(if_getdrvflags(ifp) == i);
2564 	for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
2565 		if (iflib_netmap_rxq_init(ctx, rxq) > 0) {
2566 			/* This rxq is in netmap mode. Skip normal init. */
2567 			continue;
2568 		}
2569 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
2570 			if (iflib_fl_setup(fl)) {
2571 				device_printf(ctx->ifc_dev,
2572 				    "setting up free list %d failed - "
2573 				    "check cluster settings\n", j);
2574 				goto done;
2575 			}
2576 		}
2577 	}
2578 done:
2579 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
2580 	IFDI_INTR_ENABLE(ctx);
2581 	txq = ctx->ifc_txqs;
2582 	for (i = 0; i < sctx->isc_ntxqsets; i++, txq++)
2583 		callout_reset_on(&txq->ift_timer, iflib_timer_default, iflib_timer, txq,
2584 			txq->ift_timer.c_cpu);
2585 
2586         /* Re-enable txsync/rxsync. */
2587 	netmap_enable_all_rings(ifp);
2588 }
2589 
2590 static int
2591 iflib_media_change(if_t ifp)
2592 {
2593 	if_ctx_t ctx = if_getsoftc(ifp);
2594 	int err;
2595 
2596 	CTX_LOCK(ctx);
2597 	if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0)
2598 		iflib_if_init_locked(ctx);
2599 	CTX_UNLOCK(ctx);
2600 	return (err);
2601 }
2602 
2603 static void
2604 iflib_media_status(if_t ifp, struct ifmediareq *ifmr)
2605 {
2606 	if_ctx_t ctx = if_getsoftc(ifp);
2607 
2608 	CTX_LOCK(ctx);
2609 	IFDI_UPDATE_ADMIN_STATUS(ctx);
2610 	IFDI_MEDIA_STATUS(ctx, ifmr);
2611 	CTX_UNLOCK(ctx);
2612 }
2613 
2614 void
2615 iflib_stop(if_ctx_t ctx)
2616 {
2617 	iflib_txq_t txq = ctx->ifc_txqs;
2618 	iflib_rxq_t rxq = ctx->ifc_rxqs;
2619 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2620 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2621 	iflib_dma_info_t di;
2622 	iflib_fl_t fl;
2623 	int i, j;
2624 
2625 	/* Tell the stack that the interface is no longer active */
2626 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2627 
2628 	IFDI_INTR_DISABLE(ctx);
2629 	DELAY(1000);
2630 	IFDI_STOP(ctx);
2631 	DELAY(1000);
2632 
2633 	/*
2634 	 * Stop any pending txsync/rxsync and prevent new ones
2635 	 * form starting. Processes blocked in poll() will get
2636 	 * POLLERR.
2637 	 */
2638 	netmap_disable_all_rings(ctx->ifc_ifp);
2639 
2640 	iflib_debug_reset();
2641 	/* Wait for current tx queue users to exit to disarm watchdog timer. */
2642 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) {
2643 		/* make sure all transmitters have completed before proceeding XXX */
2644 
2645 		CALLOUT_LOCK(txq);
2646 		callout_stop(&txq->ift_timer);
2647 #ifdef DEV_NETMAP
2648 		callout_stop(&txq->ift_netmap_timer);
2649 #endif /* DEV_NETMAP */
2650 		CALLOUT_UNLOCK(txq);
2651 
2652 		/* clean any enqueued buffers */
2653 		iflib_ifmp_purge(txq);
2654 		/* Free any existing tx buffers. */
2655 		for (j = 0; j < txq->ift_size; j++) {
2656 			iflib_txsd_free(ctx, txq, j);
2657 		}
2658 		txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0;
2659 		txq->ift_in_use = txq->ift_gen = txq->ift_cidx = txq->ift_pidx = txq->ift_no_desc_avail = 0;
2660 		txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0;
2661 		txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0;
2662 		txq->ift_pullups = 0;
2663 		ifmp_ring_reset_stats(txq->ift_br);
2664 		for (j = 0, di = txq->ift_ifdi; j < sctx->isc_ntxqs; j++, di++)
2665 			bzero((void *)di->idi_vaddr, di->idi_size);
2666 	}
2667 	for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) {
2668 		/* make sure all transmitters have completed before proceeding XXX */
2669 
2670 		rxq->ifr_cq_cidx = 0;
2671 		for (j = 0, di = rxq->ifr_ifdi; j < sctx->isc_nrxqs; j++, di++)
2672 			bzero((void *)di->idi_vaddr, di->idi_size);
2673 		/* also resets the free lists pidx/cidx */
2674 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
2675 			iflib_fl_bufs_free(fl);
2676 	}
2677 }
2678 
2679 static inline caddr_t
2680 calc_next_rxd(iflib_fl_t fl, int cidx)
2681 {
2682 	qidx_t size;
2683 	int nrxd;
2684 	caddr_t start, end, cur, next;
2685 
2686 	nrxd = fl->ifl_size;
2687 	size = fl->ifl_rxd_size;
2688 	start = fl->ifl_ifdi->idi_vaddr;
2689 
2690 	if (__predict_false(size == 0))
2691 		return (start);
2692 	cur = start + size*cidx;
2693 	end = start + size*nrxd;
2694 	next = CACHE_PTR_NEXT(cur);
2695 	return (next < end ? next : start);
2696 }
2697 
2698 static inline void
2699 prefetch_pkts(iflib_fl_t fl, int cidx)
2700 {
2701 	int nextptr;
2702 	int nrxd = fl->ifl_size;
2703 	caddr_t next_rxd;
2704 
2705 	nextptr = (cidx + CACHE_PTR_INCREMENT) & (nrxd-1);
2706 	prefetch(&fl->ifl_sds.ifsd_m[nextptr]);
2707 	prefetch(&fl->ifl_sds.ifsd_cl[nextptr]);
2708 	next_rxd = calc_next_rxd(fl, cidx);
2709 	prefetch(next_rxd);
2710 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 1) & (nrxd-1)]);
2711 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 2) & (nrxd-1)]);
2712 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 3) & (nrxd-1)]);
2713 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 4) & (nrxd-1)]);
2714 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 1) & (nrxd-1)]);
2715 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 2) & (nrxd-1)]);
2716 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 3) & (nrxd-1)]);
2717 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 4) & (nrxd-1)]);
2718 }
2719 
2720 static struct mbuf *
2721 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, bool unload, if_rxsd_t sd,
2722     int *pf_rv, if_rxd_info_t ri)
2723 {
2724 	bus_dmamap_t map;
2725 	iflib_fl_t fl;
2726 	caddr_t payload;
2727 	struct mbuf *m;
2728 	int flid, cidx, len, next;
2729 
2730 	map = NULL;
2731 	flid = irf->irf_flid;
2732 	cidx = irf->irf_idx;
2733 	fl = &rxq->ifr_fl[flid];
2734 	sd->ifsd_fl = fl;
2735 	m = fl->ifl_sds.ifsd_m[cidx];
2736 	sd->ifsd_cl = &fl->ifl_sds.ifsd_cl[cidx];
2737 	fl->ifl_credits--;
2738 #if MEMORY_LOGGING
2739 	fl->ifl_m_dequeued++;
2740 #endif
2741 	if (rxq->ifr_ctx->ifc_flags & IFC_PREFETCH)
2742 		prefetch_pkts(fl, cidx);
2743 	next = (cidx + CACHE_PTR_INCREMENT) & (fl->ifl_size-1);
2744 	prefetch(&fl->ifl_sds.ifsd_map[next]);
2745 	map = fl->ifl_sds.ifsd_map[cidx];
2746 
2747 	bus_dmamap_sync(fl->ifl_buf_tag, map, BUS_DMASYNC_POSTREAD);
2748 
2749 #ifndef __HAIKU__
2750 	if (rxq->pfil != NULL && PFIL_HOOKED_IN(rxq->pfil) && pf_rv != NULL &&
2751 	    irf->irf_len != 0) {
2752 		payload  = *sd->ifsd_cl;
2753 		payload +=  ri->iri_pad;
2754 		len = ri->iri_len - ri->iri_pad;
2755 		*pf_rv = pfil_run_hooks(rxq->pfil, payload, ri->iri_ifp,
2756 		    len | PFIL_MEMPTR | PFIL_IN, NULL);
2757 		switch (*pf_rv) {
2758 		case PFIL_DROPPED:
2759 		case PFIL_CONSUMED:
2760 			/*
2761 			 * The filter ate it.  Everything is recycled.
2762 			 */
2763 			m = NULL;
2764 			unload = 0;
2765 			break;
2766 		case PFIL_REALLOCED:
2767 			/*
2768 			 * The filter copied it.  Everything is recycled.
2769 			 */
2770 			m = pfil_mem2mbuf(payload);
2771 			unload = 0;
2772 			break;
2773 		case PFIL_PASS:
2774 			/*
2775 			 * Filter said it was OK, so receive like
2776 			 * normal
2777 			 */
2778 			fl->ifl_sds.ifsd_m[cidx] = NULL;
2779 			break;
2780 		default:
2781 			MPASS(0);
2782 		}
2783 	} else
2784 #endif
2785 	{
2786 		fl->ifl_sds.ifsd_m[cidx] = NULL;
2787 		if (pf_rv != NULL)
2788 			*pf_rv = PFIL_PASS;
2789 	}
2790 
2791 	if (unload && irf->irf_len != 0)
2792 		bus_dmamap_unload(fl->ifl_buf_tag, map);
2793 	fl->ifl_cidx = (fl->ifl_cidx + 1) & (fl->ifl_size-1);
2794 	if (__predict_false(fl->ifl_cidx == 0))
2795 		fl->ifl_gen = 0;
2796 	bit_clear(fl->ifl_rx_bitmap, cidx);
2797 	return (m);
2798 }
2799 
2800 static struct mbuf *
2801 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri, if_rxsd_t sd, int *pf_rv)
2802 {
2803 	struct mbuf *m, *mh, *mt;
2804 	caddr_t cl;
2805 	int  *pf_rv_ptr, flags, i, padlen;
2806 	bool consumed;
2807 
2808 	i = 0;
2809 	mh = NULL;
2810 	consumed = false;
2811 	*pf_rv = PFIL_PASS;
2812 	pf_rv_ptr = pf_rv;
2813 	do {
2814 		m = rxd_frag_to_sd(rxq, &ri->iri_frags[i], !consumed, sd,
2815 		    pf_rv_ptr, ri);
2816 
2817 		MPASS(*sd->ifsd_cl != NULL);
2818 
2819 		/*
2820 		 * Exclude zero-length frags & frags from
2821 		 * packets the filter has consumed or dropped
2822 		 */
2823 		if (ri->iri_frags[i].irf_len == 0 || consumed ||
2824 #ifndef __HAIKU__
2825 		    *pf_rv == PFIL_CONSUMED || *pf_rv == PFIL_DROPPED
2826 #else
2827 			0
2828 #endif
2829 		     ) {
2830 			if (mh == NULL) {
2831 				/* everything saved here */
2832 				consumed = true;
2833 				pf_rv_ptr = NULL;
2834 				continue;
2835 			}
2836 			/* XXX we can save the cluster here, but not the mbuf */
2837 			m_init(m, M_NOWAIT, MT_DATA, 0);
2838 			m_free(m);
2839 			continue;
2840 		}
2841 		if (mh == NULL) {
2842 			flags = M_PKTHDR|M_EXT;
2843 			mh = mt = m;
2844 			padlen = ri->iri_pad;
2845 		} else {
2846 			flags = M_EXT;
2847 			mt->m_next = m;
2848 			mt = m;
2849 			/* assuming padding is only on the first fragment */
2850 			padlen = 0;
2851 		}
2852 		cl = *sd->ifsd_cl;
2853 		*sd->ifsd_cl = NULL;
2854 
2855 		/* Can these two be made one ? */
2856 		m_init(m, M_NOWAIT, MT_DATA, flags);
2857 		m_cljset(m, cl, sd->ifsd_fl->ifl_cltype);
2858 		/*
2859 		 * These must follow m_init and m_cljset
2860 		 */
2861 		m->m_data += padlen;
2862 		ri->iri_len -= padlen;
2863 		m->m_len = ri->iri_frags[i].irf_len;
2864 	} while (++i < ri->iri_nfrags);
2865 
2866 	return (mh);
2867 }
2868 
2869 /*
2870  * Process one software descriptor
2871  */
2872 static struct mbuf *
2873 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri)
2874 {
2875 	struct if_rxsd sd;
2876 	struct mbuf *m;
2877 	int pf_rv;
2878 
2879 	/* should I merge this back in now that the two paths are basically duplicated? */
2880 	if (ri->iri_nfrags == 1 &&
2881 	    ri->iri_frags[0].irf_len != 0 &&
2882 	    ri->iri_frags[0].irf_len <= MIN(IFLIB_RX_COPY_THRESH, MHLEN)) {
2883 		m = rxd_frag_to_sd(rxq, &ri->iri_frags[0], false, &sd,
2884 		    &pf_rv, ri);
2885 		if (pf_rv != PFIL_PASS
2886 #ifndef __HAIKU__
2887 		        && pf_rv != PFIL_REALLOCED
2888 #endif
2889 		        )
2890 			return (m);
2891 		if (pf_rv == PFIL_PASS) {
2892 			m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR);
2893 #ifndef __NO_STRICT_ALIGNMENT
2894 			if (!IP_ALIGNED(m))
2895 				m->m_data += 2;
2896 #endif
2897 			memcpy(m->m_data, *sd.ifsd_cl, ri->iri_len);
2898 			m->m_len = ri->iri_frags[0].irf_len;
2899 		}
2900 	} else {
2901 		m = assemble_segments(rxq, ri, &sd, &pf_rv);
2902 		if (m == NULL)
2903 			return (NULL);
2904 		if (pf_rv != PFIL_PASS
2905 #ifndef __HAIKU__
2906 		        && pf_rv != PFIL_REALLOCED
2907 #endif
2908 				)
2909 			return (m);
2910 	}
2911 	m->m_pkthdr.len = ri->iri_len;
2912 	m->m_pkthdr.rcvif = ri->iri_ifp;
2913 	m->m_flags |= ri->iri_flags;
2914 	m->m_pkthdr.ether_vtag = ri->iri_vtag;
2915 	m->m_pkthdr.flowid = ri->iri_flowid;
2916 	M_HASHTYPE_SET(m, ri->iri_rsstype);
2917 	m->m_pkthdr.csum_flags = ri->iri_csum_flags;
2918 	m->m_pkthdr.csum_data = ri->iri_csum_data;
2919 	return (m);
2920 }
2921 
2922 #if defined(INET6) || defined(INET)
2923 static void
2924 iflib_get_ip_forwarding(struct lro_ctrl *lc, bool *v4, bool *v6)
2925 {
2926 	CURVNET_SET(lc->ifp->if_vnet);
2927 #if defined(INET6)
2928 	*v6 = V_ip6_forwarding;
2929 #endif
2930 #if defined(INET)
2931 	*v4 = V_ipforwarding;
2932 #endif
2933 	CURVNET_RESTORE();
2934 }
2935 
2936 /*
2937  * Returns true if it's possible this packet could be LROed.
2938  * if it returns false, it is guaranteed that tcp_lro_rx()
2939  * would not return zero.
2940  */
2941 static bool
2942 iflib_check_lro_possible(struct mbuf *m, bool v4_forwarding, bool v6_forwarding)
2943 {
2944 #ifndef __HAIKU__
2945 	struct ether_header *eh;
2946 
2947 	eh = mtod(m, struct ether_header *);
2948 	switch (eh->ether_type) {
2949 #if defined(INET6)
2950 		case htons(ETHERTYPE_IPV6):
2951 			return (!v6_forwarding);
2952 #endif
2953 #if defined (INET)
2954 		case htons(ETHERTYPE_IP):
2955 			return (!v4_forwarding);
2956 #endif
2957 	}
2958 #endif
2959 
2960 	return false;
2961 }
2962 #else
2963 static void
2964 iflib_get_ip_forwarding(struct lro_ctrl *lc __unused, bool *v4 __unused, bool *v6 __unused)
2965 {
2966 }
2967 #endif
2968 
2969 static void
2970 _task_fn_rx_watchdog(void *context)
2971 {
2972 	iflib_rxq_t rxq = context;
2973 
2974 	GROUPTASK_ENQUEUE(&rxq->ifr_task);
2975 }
2976 
2977 static uint8_t
2978 iflib_rxeof(iflib_rxq_t rxq, qidx_t budget)
2979 {
2980 	if_t ifp;
2981 	if_ctx_t ctx = rxq->ifr_ctx;
2982 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2983 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2984 	int avail, i;
2985 	qidx_t *cidxp;
2986 	struct if_rxd_info ri;
2987 	int err, budget_left, rx_bytes, rx_pkts;
2988 	iflib_fl_t fl;
2989 	int lro_enabled;
2990 	bool v4_forwarding, v6_forwarding, lro_possible;
2991 	uint8_t retval = 0;
2992 
2993 	/*
2994 	 * XXX early demux data packets so that if_input processing only handles
2995 	 * acks in interrupt context
2996 	 */
2997 	struct mbuf *m, *mh, *mt, *mf;
2998 
2999 #ifndef __HAIKU__
3000 	NET_EPOCH_ASSERT();
3001 #endif
3002 
3003 	lro_possible = v4_forwarding = v6_forwarding = false;
3004 	ifp = ctx->ifc_ifp;
3005 	mh = mt = NULL;
3006 	MPASS(budget > 0);
3007 	rx_pkts	= rx_bytes = 0;
3008 	if (sctx->isc_flags & IFLIB_HAS_RXCQ)
3009 		cidxp = &rxq->ifr_cq_cidx;
3010 	else
3011 		cidxp = &rxq->ifr_fl[0].ifl_cidx;
3012 	if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) {
3013 		for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
3014 			retval |= iflib_fl_refill_all(ctx, fl);
3015 		DBG_COUNTER_INC(rx_unavail);
3016 		return (retval);
3017 	}
3018 
3019 	/* pfil needs the vnet to be set */
3020 	CURVNET_SET_QUIET(ifp->if_vnet);
3021 	for (budget_left = budget; budget_left > 0 && avail > 0;) {
3022 		if (__predict_false(!CTX_ACTIVE(ctx))) {
3023 			DBG_COUNTER_INC(rx_ctx_inactive);
3024 			break;
3025 		}
3026 		/*
3027 		 * Reset client set fields to their default values
3028 		 */
3029 		rxd_info_zero(&ri);
3030 		ri.iri_qsidx = rxq->ifr_id;
3031 		ri.iri_cidx = *cidxp;
3032 		ri.iri_ifp = ifp;
3033 		ri.iri_frags = rxq->ifr_frags;
3034 		err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
3035 
3036 		if (err)
3037 			goto err;
3038 		rx_pkts += 1;
3039 		rx_bytes += ri.iri_len;
3040 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
3041 			*cidxp = ri.iri_cidx;
3042 			/* Update our consumer index */
3043 			/* XXX NB: shurd - check if this is still safe */
3044 			while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0])
3045 				rxq->ifr_cq_cidx -= scctx->isc_nrxd[0];
3046 			/* was this only a completion queue message? */
3047 			if (__predict_false(ri.iri_nfrags == 0))
3048 				continue;
3049 		}
3050 		MPASS(ri.iri_nfrags != 0);
3051 		MPASS(ri.iri_len != 0);
3052 
3053 		/* will advance the cidx on the corresponding free lists */
3054 		m = iflib_rxd_pkt_get(rxq, &ri);
3055 		avail--;
3056 		budget_left--;
3057 		if (avail == 0 && budget_left)
3058 			avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left);
3059 
3060 		if (__predict_false(m == NULL))
3061 			continue;
3062 
3063 		/* imm_pkt: -- cxgb */
3064 		if (mh == NULL)
3065 			mh = mt = m;
3066 		else {
3067 			mt->m_nextpkt = m;
3068 			mt = m;
3069 		}
3070 	}
3071 	CURVNET_RESTORE();
3072 	/* make sure that we can refill faster than drain */
3073 	for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
3074 		retval |= iflib_fl_refill_all(ctx, fl);
3075 
3076 	lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO);
3077 #ifndef __HAIKU__
3078 	if (lro_enabled)
3079 		iflib_get_ip_forwarding(&rxq->ifr_lc, &v4_forwarding, &v6_forwarding);
3080 #endif
3081 	mt = mf = NULL;
3082 	while (mh != NULL) {
3083 		m = mh;
3084 		mh = mh->m_nextpkt;
3085 		m->m_nextpkt = NULL;
3086 #ifndef __NO_STRICT_ALIGNMENT
3087 		if (!IP_ALIGNED(m) && (m = iflib_fixup_rx(m)) == NULL)
3088 			continue;
3089 #endif
3090 #ifndef __HAIKU__
3091 #if defined(INET6) || defined(INET)
3092 		if (lro_enabled) {
3093 			if (!lro_possible) {
3094 				lro_possible = iflib_check_lro_possible(m, v4_forwarding, v6_forwarding);
3095 				if (lro_possible && mf != NULL) {
3096 					ifp->if_input(ifp, mf);
3097 					DBG_COUNTER_INC(rx_if_input);
3098 					mt = mf = NULL;
3099 				}
3100 			}
3101 			if ((m->m_pkthdr.csum_flags & (CSUM_L4_CALC|CSUM_L4_VALID)) ==
3102 			    (CSUM_L4_CALC|CSUM_L4_VALID)) {
3103 				if (lro_possible && tcp_lro_rx(&rxq->ifr_lc, m, 0) == 0)
3104 					continue;
3105 			}
3106 		}
3107 #endif
3108 		if (lro_possible) {
3109 			ifp->if_input(ifp, m);
3110 			DBG_COUNTER_INC(rx_if_input);
3111 			continue;
3112 		}
3113 #else /* __HAIKU __*/
3114 		if (mf != NULL) {
3115 			ifp->if_input(ifp, mf);
3116 			DBG_COUNTER_INC(rx_if_input);
3117 			mt = mf = NULL;
3118 		}
3119 		ifp->if_input(ifp, m);
3120 		DBG_COUNTER_INC(rx_if_input);
3121 		continue;
3122 #endif
3123 
3124 		if (mf == NULL)
3125 			mf = m;
3126 		if (mt != NULL)
3127 			mt->m_nextpkt = m;
3128 		mt = m;
3129 	}
3130 	if (mf != NULL) {
3131 		ifp->if_input(ifp, mf);
3132 		DBG_COUNTER_INC(rx_if_input);
3133 	}
3134 
3135 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
3136 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
3137 
3138 	/*
3139 	 * Flush any outstanding LRO work
3140 	 */
3141 #if defined(INET6) || defined(INET)
3142 #ifndef __HAIKU__
3143 	tcp_lro_flush_all(&rxq->ifr_lc);
3144 #endif
3145 #endif
3146 	if (avail != 0 || iflib_rxd_avail(ctx, rxq, *cidxp, 1) != 0)
3147 		retval |= IFLIB_RXEOF_MORE;
3148 	return (retval);
3149 err:
3150 	STATE_LOCK(ctx);
3151 	ctx->ifc_flags |= IFC_DO_RESET;
3152 	iflib_admin_intr_deferred(ctx);
3153 	STATE_UNLOCK(ctx);
3154 	return (0);
3155 }
3156 
3157 #define TXD_NOTIFY_COUNT(txq) (((txq)->ift_size / (txq)->ift_update_freq)-1)
3158 static inline qidx_t
3159 txq_max_db_deferred(iflib_txq_t txq, qidx_t in_use)
3160 {
3161 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
3162 	qidx_t minthresh = txq->ift_size / 8;
3163 	if (in_use > 4*minthresh)
3164 		return (notify_count);
3165 	if (in_use > 2*minthresh)
3166 		return (notify_count >> 1);
3167 	if (in_use > minthresh)
3168 		return (notify_count >> 3);
3169 	return (0);
3170 }
3171 
3172 static inline qidx_t
3173 txq_max_rs_deferred(iflib_txq_t txq)
3174 {
3175 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
3176 	qidx_t minthresh = txq->ift_size / 8;
3177 	if (txq->ift_in_use > 4*minthresh)
3178 		return (notify_count);
3179 	if (txq->ift_in_use > 2*minthresh)
3180 		return (notify_count >> 1);
3181 	if (txq->ift_in_use > minthresh)
3182 		return (notify_count >> 2);
3183 	return (2);
3184 }
3185 
3186 #define M_CSUM_FLAGS(m) ((m)->m_pkthdr.csum_flags)
3187 #define M_HAS_VLANTAG(m) (m->m_flags & M_VLANTAG)
3188 
3189 #define TXQ_MAX_DB_DEFERRED(txq, in_use) txq_max_db_deferred((txq), (in_use))
3190 #define TXQ_MAX_RS_DEFERRED(txq) txq_max_rs_deferred(txq)
3191 #define TXQ_MAX_DB_CONSUMED(size) (size >> 4)
3192 
3193 /* forward compatibility for cxgb */
3194 #define FIRST_QSET(ctx) 0
3195 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets)
3196 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets)
3197 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx))
3198 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments))
3199 
3200 /* XXX we should be setting this to something other than zero */
3201 #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh)
3202 #define	MAX_TX_DESC(ctx) MAX((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max, \
3203     (ctx)->ifc_softc_ctx.isc_tx_nsegments)
3204 
3205 static inline bool
3206 iflib_txd_db_check(iflib_txq_t txq, int ring)
3207 {
3208 	if_ctx_t ctx = txq->ift_ctx;
3209 	qidx_t dbval, max;
3210 
3211 	max = TXQ_MAX_DB_DEFERRED(txq, txq->ift_in_use);
3212 
3213 	/* force || threshold exceeded || at the edge of the ring */
3214 	if (ring || (txq->ift_db_pending >= max) || (TXQ_AVAIL(txq) <= MAX_TX_DESC(ctx) + 2)) {
3215 
3216 		/*
3217 		 * 'npending' is used if the card's doorbell is in terms of the number of descriptors
3218 		 * pending flush (BRCM). 'pidx' is used in cases where the card's doorbeel uses the
3219 		 * producer index explicitly (INTC).
3220 		 */
3221 		dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx;
3222 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3223 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
3224 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval);
3225 
3226 		/*
3227 		 * Absent bugs there are zero packets pending so reset pending counts to zero.
3228 		 */
3229 		txq->ift_db_pending = txq->ift_npending = 0;
3230 		return (true);
3231 	}
3232 	return (false);
3233 }
3234 
3235 #ifdef PKT_DEBUG
3236 static void
3237 print_pkt(if_pkt_info_t pi)
3238 {
3239 	printf("pi len:  %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n",
3240 	       pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx);
3241 	printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n",
3242 	       pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag);
3243 	printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n",
3244 	       pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto);
3245 }
3246 #endif
3247 
3248 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO)
3249 #define IS_TX_OFFLOAD4(pi) ((pi)->ipi_csum_flags & (CSUM_IP_TCP | CSUM_IP_TSO))
3250 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO)
3251 #define IS_TX_OFFLOAD6(pi) ((pi)->ipi_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_TSO))
3252 
3253 static int
3254 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp)
3255 {
3256 	if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx;
3257 	struct ether_vlan_header *eh;
3258 	struct mbuf *m;
3259 
3260 	m = *mp;
3261 	if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) &&
3262 	    M_WRITABLE(m) == 0) {
3263 		if ((m = m_dup(m, M_NOWAIT)) == NULL) {
3264 			return (ENOMEM);
3265 		} else {
3266 			m_freem(*mp);
3267 			DBG_COUNTER_INC(tx_frees);
3268 			*mp = m;
3269 		}
3270 	}
3271 
3272 	/*
3273 	 * Determine where frame payload starts.
3274 	 * Jump over vlan headers if already present,
3275 	 * helpful for QinQ too.
3276 	 */
3277 	if (__predict_false(m->m_len < sizeof(*eh))) {
3278 		txq->ift_pullups++;
3279 		if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL))
3280 			return (ENOMEM);
3281 	}
3282 	eh = mtod(m, struct ether_vlan_header *);
3283 	if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
3284 		pi->ipi_etype = ntohs(eh->evl_proto);
3285 		pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
3286 	} else {
3287 		pi->ipi_etype = ntohs(eh->evl_encap_proto);
3288 		pi->ipi_ehdrlen = ETHER_HDR_LEN;
3289 	}
3290 
3291 	switch (pi->ipi_etype) {
3292 #ifdef INET
3293 	case ETHERTYPE_IP:
3294 	{
3295 		struct mbuf *n;
3296 		struct ip *ip = NULL;
3297 		struct tcphdr *th = NULL;
3298 		int minthlen;
3299 
3300 		minthlen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip) + sizeof(*th));
3301 		if (__predict_false(m->m_len < minthlen)) {
3302 			/*
3303 			 * if this code bloat is causing too much of a hit
3304 			 * move it to a separate function and mark it noinline
3305 			 */
3306 			if (m->m_len == pi->ipi_ehdrlen) {
3307 				n = m->m_next;
3308 				MPASS(n);
3309 				if (n->m_len >= sizeof(*ip))  {
3310 					ip = (struct ip *)n->m_data;
3311 					if (n->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3312 						th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3313 				} else {
3314 					txq->ift_pullups++;
3315 					if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
3316 						return (ENOMEM);
3317 					ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3318 				}
3319 			} else {
3320 				txq->ift_pullups++;
3321 				if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
3322 					return (ENOMEM);
3323 				ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3324 				if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3325 					th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3326 			}
3327 		} else {
3328 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3329 			if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3330 				th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3331 		}
3332 		pi->ipi_ip_hlen = ip->ip_hl << 2;
3333 		pi->ipi_ipproto = ip->ip_p;
3334 		pi->ipi_flags |= IPI_TX_IPV4;
3335 
3336 		/* TCP checksum offload may require TCP header length */
3337 		if (IS_TX_OFFLOAD4(pi)) {
3338 			if (__predict_true(pi->ipi_ipproto == IPPROTO_TCP)) {
3339 				if (__predict_false(th == NULL)) {
3340 					txq->ift_pullups++;
3341 					if (__predict_false((m = m_pullup(m, (ip->ip_hl << 2) + sizeof(*th))) == NULL))
3342 						return (ENOMEM);
3343 					th = (struct tcphdr *)((caddr_t)ip + pi->ipi_ip_hlen);
3344 				}
3345 				pi->ipi_tcp_hflags = th->th_flags;
3346 				pi->ipi_tcp_hlen = th->th_off << 2;
3347 				pi->ipi_tcp_seq = th->th_seq;
3348 			}
3349 			if (IS_TSO4(pi)) {
3350 				if (__predict_false(ip->ip_p != IPPROTO_TCP))
3351 					return (ENXIO);
3352 				/*
3353 				 * TSO always requires hardware checksum offload.
3354 				 */
3355 				pi->ipi_csum_flags |= (CSUM_IP_TCP | CSUM_IP);
3356 				th->th_sum = in_pseudo(ip->ip_src.s_addr,
3357 						       ip->ip_dst.s_addr, htons(IPPROTO_TCP));
3358 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3359 				if (sctx->isc_flags & IFLIB_TSO_INIT_IP) {
3360 					ip->ip_sum = 0;
3361 					ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz);
3362 				}
3363 			}
3364 		}
3365 		if ((sctx->isc_flags & IFLIB_NEED_ZERO_CSUM) && (pi->ipi_csum_flags & CSUM_IP))
3366                        ip->ip_sum = 0;
3367 
3368 		break;
3369 	}
3370 #endif
3371 #ifdef INET6
3372 	case ETHERTYPE_IPV6:
3373 	{
3374 		struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
3375 		struct tcphdr *th;
3376 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
3377 
3378 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
3379 			txq->ift_pullups++;
3380 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
3381 				return (ENOMEM);
3382 		}
3383 		th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
3384 
3385 		/* XXX-BZ this will go badly in case of ext hdrs. */
3386 		pi->ipi_ipproto = ip6->ip6_nxt;
3387 		pi->ipi_flags |= IPI_TX_IPV6;
3388 
3389 		/* TCP checksum offload may require TCP header length */
3390 		if (IS_TX_OFFLOAD6(pi)) {
3391 			if (pi->ipi_ipproto == IPPROTO_TCP) {
3392 				if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) {
3393 					txq->ift_pullups++;
3394 					if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL))
3395 						return (ENOMEM);
3396 				}
3397 				pi->ipi_tcp_hflags = th->th_flags;
3398 				pi->ipi_tcp_hlen = th->th_off << 2;
3399 				pi->ipi_tcp_seq = th->th_seq;
3400 			}
3401 			if (IS_TSO6(pi)) {
3402 				if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP))
3403 					return (ENXIO);
3404 				/*
3405 				 * TSO always requires hardware checksum offload.
3406 				 */
3407 				pi->ipi_csum_flags |= CSUM_IP6_TCP;
3408 				th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0);
3409 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3410 			}
3411 		}
3412 		break;
3413 	}
3414 #endif
3415 	default:
3416 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
3417 		pi->ipi_ip_hlen = 0;
3418 		break;
3419 	}
3420 	*mp = m;
3421 
3422 	return (0);
3423 }
3424 
3425 /*
3426  * If dodgy hardware rejects the scatter gather chain we've handed it
3427  * we'll need to remove the mbuf chain from ifsg_m[] before we can add the
3428  * m_defrag'd mbufs
3429  */
3430 static __noinline struct mbuf *
3431 iflib_remove_mbuf(iflib_txq_t txq)
3432 {
3433 	int ntxd, pidx;
3434 	struct mbuf *m, **ifsd_m;
3435 
3436 	ifsd_m = txq->ift_sds.ifsd_m;
3437 	ntxd = txq->ift_size;
3438 	pidx = txq->ift_pidx & (ntxd - 1);
3439 	ifsd_m = txq->ift_sds.ifsd_m;
3440 	m = ifsd_m[pidx];
3441 	ifsd_m[pidx] = NULL;
3442 	bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[pidx]);
3443 	if (txq->ift_sds.ifsd_tso_map != NULL)
3444 		bus_dmamap_unload(txq->ift_tso_buf_tag,
3445 		    txq->ift_sds.ifsd_tso_map[pidx]);
3446 #if MEMORY_LOGGING
3447 	txq->ift_dequeued++;
3448 #endif
3449 	return (m);
3450 }
3451 
3452 static inline caddr_t
3453 calc_next_txd(iflib_txq_t txq, int cidx, uint8_t qid)
3454 {
3455 	qidx_t size;
3456 	int ntxd;
3457 	caddr_t start, end, cur, next;
3458 
3459 	ntxd = txq->ift_size;
3460 	size = txq->ift_txd_size[qid];
3461 	start = txq->ift_ifdi[qid].idi_vaddr;
3462 
3463 	if (__predict_false(size == 0))
3464 		return (start);
3465 	cur = start + size*cidx;
3466 	end = start + size*ntxd;
3467 	next = CACHE_PTR_NEXT(cur);
3468 	return (next < end ? next : start);
3469 }
3470 
3471 /*
3472  * Pad an mbuf to ensure a minimum ethernet frame size.
3473  * min_frame_size is the frame size (less CRC) to pad the mbuf to
3474  */
3475 static __noinline int
3476 iflib_ether_pad(device_t dev, struct mbuf **m_head, uint16_t min_frame_size)
3477 {
3478 	/*
3479 	 * 18 is enough bytes to pad an ARP packet to 46 bytes, and
3480 	 * and ARP message is the smallest common payload I can think of
3481 	 */
3482 	static char pad[18];	/* just zeros */
3483 	int n;
3484 	struct mbuf *new_head;
3485 
3486 	if (!M_WRITABLE(*m_head)) {
3487 		new_head = m_dup(*m_head, M_NOWAIT);
3488 		if (new_head == NULL) {
3489 			m_freem(*m_head);
3490 			device_printf(dev, "cannot pad short frame, m_dup() failed");
3491 			DBG_COUNTER_INC(encap_pad_mbuf_fail);
3492 			DBG_COUNTER_INC(tx_frees);
3493 			return ENOMEM;
3494 		}
3495 		m_freem(*m_head);
3496 		*m_head = new_head;
3497 	}
3498 
3499 	for (n = min_frame_size - (*m_head)->m_pkthdr.len;
3500 	     n > 0; n -= sizeof(pad))
3501 		if (!m_append(*m_head, min(n, sizeof(pad)), pad))
3502 			break;
3503 
3504 	if (n > 0) {
3505 		m_freem(*m_head);
3506 		device_printf(dev, "cannot pad short frame\n");
3507 		DBG_COUNTER_INC(encap_pad_mbuf_fail);
3508 		DBG_COUNTER_INC(tx_frees);
3509 		return (ENOBUFS);
3510 	}
3511 
3512 	return 0;
3513 }
3514 
3515 static int
3516 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp)
3517 {
3518 	if_ctx_t		ctx;
3519 	if_shared_ctx_t		sctx;
3520 	if_softc_ctx_t		scctx;
3521 	bus_dma_tag_t		buf_tag;
3522 	bus_dma_segment_t	*segs;
3523 	struct mbuf		*m_head, **ifsd_m;
3524 	void			*next_txd;
3525 	bus_dmamap_t		map;
3526 	struct if_pkt_info	pi;
3527 	int remap = 0;
3528 	int err, nsegs, ndesc, max_segs, pidx, cidx, next, ntxd;
3529 
3530 	ctx = txq->ift_ctx;
3531 	sctx = ctx->ifc_sctx;
3532 	scctx = &ctx->ifc_softc_ctx;
3533 	segs = txq->ift_segs;
3534 	ntxd = txq->ift_size;
3535 	m_head = *m_headp;
3536 	map = NULL;
3537 
3538 	/*
3539 	 * If we're doing TSO the next descriptor to clean may be quite far ahead
3540 	 */
3541 	cidx = txq->ift_cidx;
3542 	pidx = txq->ift_pidx;
3543 	if (ctx->ifc_flags & IFC_PREFETCH) {
3544 		next = (cidx + CACHE_PTR_INCREMENT) & (ntxd-1);
3545 		if (!(ctx->ifc_flags & IFLIB_HAS_TXCQ)) {
3546 			next_txd = calc_next_txd(txq, cidx, 0);
3547 			prefetch(next_txd);
3548 		}
3549 
3550 		/* prefetch the next cache line of mbuf pointers and flags */
3551 		prefetch(&txq->ift_sds.ifsd_m[next]);
3552 		prefetch(&txq->ift_sds.ifsd_map[next]);
3553 		next = (cidx + CACHE_LINE_SIZE) & (ntxd-1);
3554 	}
3555 	map = txq->ift_sds.ifsd_map[pidx];
3556 	ifsd_m = txq->ift_sds.ifsd_m;
3557 
3558 	if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
3559 		buf_tag = txq->ift_tso_buf_tag;
3560 		max_segs = scctx->isc_tx_tso_segments_max;
3561 		map = txq->ift_sds.ifsd_tso_map[pidx];
3562 		MPASS(buf_tag != NULL);
3563 		MPASS(max_segs > 0);
3564 	} else {
3565 		buf_tag = txq->ift_buf_tag;
3566 		max_segs = scctx->isc_tx_nsegments;
3567 		map = txq->ift_sds.ifsd_map[pidx];
3568 	}
3569 	if ((sctx->isc_flags & IFLIB_NEED_ETHER_PAD) &&
3570 	    __predict_false(m_head->m_pkthdr.len < scctx->isc_min_frame_size)) {
3571 		err = iflib_ether_pad(ctx->ifc_dev, m_headp, scctx->isc_min_frame_size);
3572 		if (err) {
3573 			DBG_COUNTER_INC(encap_txd_encap_fail);
3574 			return err;
3575 		}
3576 	}
3577 	m_head = *m_headp;
3578 
3579 	pkt_info_zero(&pi);
3580 	pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG|M_BCAST|M_MCAST));
3581 	pi.ipi_pidx = pidx;
3582 	pi.ipi_qsidx = txq->ift_id;
3583 	pi.ipi_len = m_head->m_pkthdr.len;
3584 	pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags;
3585 	pi.ipi_vtag = M_HAS_VLANTAG(m_head) ? m_head->m_pkthdr.ether_vtag : 0;
3586 
3587 	/* deliberate bitwise OR to make one condition */
3588 	if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) {
3589 		if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0)) {
3590 			DBG_COUNTER_INC(encap_txd_encap_fail);
3591 			return (err);
3592 		}
3593 		m_head = *m_headp;
3594 	}
3595 
3596 retry:
3597 	err = bus_dmamap_load_mbuf_sg(buf_tag, map, m_head, segs, &nsegs,
3598 	    BUS_DMA_NOWAIT);
3599 defrag:
3600 	if (__predict_false(err)) {
3601 		switch (err) {
3602 		case EFBIG:
3603 			/* try collapse once and defrag once */
3604 			if (remap == 0) {
3605 				m_head = m_collapse(*m_headp, M_NOWAIT, max_segs);
3606 				/* try defrag if collapsing fails */
3607 				if (m_head == NULL)
3608 					remap++;
3609 			}
3610 			if (remap == 1) {
3611 				txq->ift_mbuf_defrag++;
3612 				m_head = m_defrag(*m_headp, M_NOWAIT);
3613 			}
3614 			/*
3615 			 * remap should never be >1 unless bus_dmamap_load_mbuf_sg
3616 			 * failed to map an mbuf that was run through m_defrag
3617 			 */
3618 			MPASS(remap <= 1);
3619 			if (__predict_false(m_head == NULL || remap > 1))
3620 				goto defrag_failed;
3621 			remap++;
3622 			*m_headp = m_head;
3623 			goto retry;
3624 			break;
3625 		case ENOMEM:
3626 			txq->ift_no_tx_dma_setup++;
3627 			break;
3628 		default:
3629 			txq->ift_no_tx_dma_setup++;
3630 			m_freem(*m_headp);
3631 			DBG_COUNTER_INC(tx_frees);
3632 			*m_headp = NULL;
3633 			break;
3634 		}
3635 		txq->ift_map_failed++;
3636 		DBG_COUNTER_INC(encap_load_mbuf_fail);
3637 		DBG_COUNTER_INC(encap_txd_encap_fail);
3638 		return (err);
3639 	}
3640 	ifsd_m[pidx] = m_head;
3641 	/*
3642 	 * XXX assumes a 1 to 1 relationship between segments and
3643 	 *        descriptors - this does not hold true on all drivers, e.g.
3644 	 *        cxgb
3645 	 */
3646 	if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) {
3647 		txq->ift_no_desc_avail++;
3648 		bus_dmamap_unload(buf_tag, map);
3649 		DBG_COUNTER_INC(encap_txq_avail_fail);
3650 		DBG_COUNTER_INC(encap_txd_encap_fail);
3651 		if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0)
3652 			GROUPTASK_ENQUEUE(&txq->ift_task);
3653 		return (ENOBUFS);
3654 	}
3655 	/*
3656 	 * On Intel cards we can greatly reduce the number of TX interrupts
3657 	 * we see by only setting report status on every Nth descriptor.
3658 	 * However, this also means that the driver will need to keep track
3659 	 * of the descriptors that RS was set on to check them for the DD bit.
3660 	 */
3661 	txq->ift_rs_pending += nsegs + 1;
3662 	if (txq->ift_rs_pending > TXQ_MAX_RS_DEFERRED(txq) ||
3663 	     iflib_no_tx_batch || (TXQ_AVAIL(txq) - nsegs) <= MAX_TX_DESC(ctx) + 2) {
3664 		pi.ipi_flags |= IPI_TX_INTR;
3665 		txq->ift_rs_pending = 0;
3666 	}
3667 
3668 	pi.ipi_segs = segs;
3669 	pi.ipi_nsegs = nsegs;
3670 
3671 	MPASS(pidx >= 0 && pidx < txq->ift_size);
3672 #ifdef PKT_DEBUG
3673 	print_pkt(&pi);
3674 #endif
3675 	if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) {
3676 		bus_dmamap_sync(buf_tag, map, BUS_DMASYNC_PREWRITE);
3677 		DBG_COUNTER_INC(tx_encap);
3678 		MPASS(pi.ipi_new_pidx < txq->ift_size);
3679 
3680 		ndesc = pi.ipi_new_pidx - pi.ipi_pidx;
3681 		if (pi.ipi_new_pidx < pi.ipi_pidx) {
3682 			ndesc += txq->ift_size;
3683 			txq->ift_gen = 1;
3684 		}
3685 		/*
3686 		 * drivers can need as many as
3687 		 * two sentinels
3688 		 */
3689 		MPASS(ndesc <= pi.ipi_nsegs + 2);
3690 		MPASS(pi.ipi_new_pidx != pidx);
3691 		MPASS(ndesc > 0);
3692 		txq->ift_in_use += ndesc;
3693 		txq->ift_db_pending += ndesc;
3694 
3695 		/*
3696 		 * We update the last software descriptor again here because there may
3697 		 * be a sentinel and/or there may be more mbufs than segments
3698 		 */
3699 		txq->ift_pidx = pi.ipi_new_pidx;
3700 		txq->ift_npending += pi.ipi_ndescs;
3701 	} else {
3702 		*m_headp = m_head = iflib_remove_mbuf(txq);
3703 		if (err == EFBIG) {
3704 			txq->ift_txd_encap_efbig++;
3705 			if (remap < 2) {
3706 				remap = 1;
3707 				goto defrag;
3708 			}
3709 		}
3710 		goto defrag_failed;
3711 	}
3712 	/*
3713 	 * err can't possibly be non-zero here, so we don't neet to test it
3714 	 * to see if we need to DBG_COUNTER_INC(encap_txd_encap_fail).
3715 	 */
3716 	return (err);
3717 
3718 defrag_failed:
3719 	txq->ift_mbuf_defrag_failed++;
3720 	txq->ift_map_failed++;
3721 	m_freem(*m_headp);
3722 	DBG_COUNTER_INC(tx_frees);
3723 	*m_headp = NULL;
3724 	DBG_COUNTER_INC(encap_txd_encap_fail);
3725 	return (ENOMEM);
3726 }
3727 
3728 static void
3729 iflib_tx_desc_free(iflib_txq_t txq, int n)
3730 {
3731 	uint32_t qsize, cidx, mask, gen;
3732 	struct mbuf *m, **ifsd_m;
3733 	bool do_prefetch;
3734 
3735 	cidx = txq->ift_cidx;
3736 	gen = txq->ift_gen;
3737 	qsize = txq->ift_size;
3738 	mask = qsize-1;
3739 	ifsd_m = txq->ift_sds.ifsd_m;
3740 	do_prefetch = (txq->ift_ctx->ifc_flags & IFC_PREFETCH);
3741 
3742 	while (n-- > 0) {
3743 		if (do_prefetch) {
3744 			prefetch(ifsd_m[(cidx + 3) & mask]);
3745 			prefetch(ifsd_m[(cidx + 4) & mask]);
3746 		}
3747 		if ((m = ifsd_m[cidx]) != NULL) {
3748 			prefetch(&ifsd_m[(cidx + CACHE_PTR_INCREMENT) & mask]);
3749 			if (m->m_pkthdr.csum_flags & CSUM_TSO) {
3750 				bus_dmamap_sync(txq->ift_tso_buf_tag,
3751 				    txq->ift_sds.ifsd_tso_map[cidx],
3752 				    BUS_DMASYNC_POSTWRITE);
3753 				bus_dmamap_unload(txq->ift_tso_buf_tag,
3754 				    txq->ift_sds.ifsd_tso_map[cidx]);
3755 			} else {
3756 				bus_dmamap_sync(txq->ift_buf_tag,
3757 				    txq->ift_sds.ifsd_map[cidx],
3758 				    BUS_DMASYNC_POSTWRITE);
3759 				bus_dmamap_unload(txq->ift_buf_tag,
3760 				    txq->ift_sds.ifsd_map[cidx]);
3761 			}
3762 			/* XXX we don't support any drivers that batch packets yet */
3763 			MPASS(m->m_nextpkt == NULL);
3764 			m_freem(m);
3765 			ifsd_m[cidx] = NULL;
3766 #if MEMORY_LOGGING
3767 			txq->ift_dequeued++;
3768 #endif
3769 			DBG_COUNTER_INC(tx_frees);
3770 		}
3771 		if (__predict_false(++cidx == qsize)) {
3772 			cidx = 0;
3773 			gen = 0;
3774 		}
3775 	}
3776 	txq->ift_cidx = cidx;
3777 	txq->ift_gen = gen;
3778 }
3779 
3780 static __inline int
3781 iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh)
3782 {
3783 	int reclaim;
3784 	if_ctx_t ctx = txq->ift_ctx;
3785 
3786 	KASSERT(thresh >= 0, ("invalid threshold to reclaim"));
3787 	MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size);
3788 
3789 	/*
3790 	 * Need a rate-limiting check so that this isn't called every time
3791 	 */
3792 	iflib_tx_credits_update(ctx, txq);
3793 	reclaim = DESC_RECLAIMABLE(txq);
3794 
3795 	if (reclaim <= thresh /* + MAX_TX_DESC(txq->ift_ctx) */) {
3796 #ifdef INVARIANTS
3797 		if (iflib_verbose_debug) {
3798 			printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __FUNCTION__,
3799 			       txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments,
3800 			       reclaim, thresh);
3801 		}
3802 #endif
3803 		return (0);
3804 	}
3805 	iflib_tx_desc_free(txq, reclaim);
3806 	txq->ift_cleaned += reclaim;
3807 	txq->ift_in_use -= reclaim;
3808 
3809 	return (reclaim);
3810 }
3811 
3812 static struct mbuf **
3813 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset, int remaining)
3814 {
3815 	int next, size;
3816 	struct mbuf **items;
3817 
3818 	size = r->size;
3819 	next = (cidx + CACHE_PTR_INCREMENT) & (size-1);
3820 	items = __DEVOLATILE(struct mbuf **, &r->items[0]);
3821 
3822 	prefetch(items[(cidx + offset) & (size-1)]);
3823 	if (remaining > 1) {
3824 		prefetch2cachelines(&items[next]);
3825 		prefetch2cachelines(items[(cidx + offset + 1) & (size-1)]);
3826 		prefetch2cachelines(items[(cidx + offset + 2) & (size-1)]);
3827 		prefetch2cachelines(items[(cidx + offset + 3) & (size-1)]);
3828 	}
3829 	return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (size-1)]));
3830 }
3831 
3832 static void
3833 iflib_txq_check_drain(iflib_txq_t txq, int budget)
3834 {
3835 
3836 	ifmp_ring_check_drainage(txq->ift_br, budget);
3837 }
3838 
3839 static uint32_t
3840 iflib_txq_can_drain(struct ifmp_ring *r)
3841 {
3842 	iflib_txq_t txq = r->cookie;
3843 	if_ctx_t ctx = txq->ift_ctx;
3844 
3845 	if (TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2)
3846 		return (1);
3847 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3848 	    BUS_DMASYNC_POSTREAD);
3849 	return (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id,
3850 	    false));
3851 }
3852 
3853 static uint32_t
3854 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3855 {
3856 	iflib_txq_t txq = r->cookie;
3857 	if_ctx_t ctx = txq->ift_ctx;
3858 	if_t ifp = ctx->ifc_ifp;
3859 	struct mbuf *m, **mp;
3860 	int avail, bytes_sent, skipped, count, err, i;
3861 	int mcast_sent, pkt_sent, reclaimed;
3862 	bool do_prefetch, rang, ring;
3863 
3864 	if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) ||
3865 			    !LINK_ACTIVE(ctx))) {
3866 		DBG_COUNTER_INC(txq_drain_notready);
3867 		return (0);
3868 	}
3869 	reclaimed = iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
3870 	rang = iflib_txd_db_check(txq, reclaimed && txq->ift_db_pending);
3871 	avail = IDXDIFF(pidx, cidx, r->size);
3872 
3873 	if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) {
3874 		/*
3875 		 * The driver is unloading so we need to free all pending packets.
3876 		 */
3877 		DBG_COUNTER_INC(txq_drain_flushing);
3878 		for (i = 0; i < avail; i++) {
3879 			if (__predict_true(r->items[(cidx + i) & (r->size-1)] != (void *)txq))
3880 				m_freem(r->items[(cidx + i) & (r->size-1)]);
3881 			r->items[(cidx + i) & (r->size-1)] = NULL;
3882 		}
3883 		return (avail);
3884 	}
3885 
3886 	if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3887 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3888 		CALLOUT_LOCK(txq);
3889 		callout_stop(&txq->ift_timer);
3890 		CALLOUT_UNLOCK(txq);
3891 		DBG_COUNTER_INC(txq_drain_oactive);
3892 		return (0);
3893 	}
3894 
3895 	/*
3896 	 * If we've reclaimed any packets this queue cannot be hung.
3897 	 */
3898 	if (reclaimed)
3899 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3900 	skipped = mcast_sent = bytes_sent = pkt_sent = 0;
3901 	count = MIN(avail, TX_BATCH_SIZE);
3902 #ifdef INVARIANTS
3903 	if (iflib_verbose_debug)
3904 		printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __FUNCTION__,
3905 		       avail, ctx->ifc_flags, TXQ_AVAIL(txq));
3906 #endif
3907 	do_prefetch = (ctx->ifc_flags & IFC_PREFETCH);
3908 	err = 0;
3909 	for (i = 0; i < count && TXQ_AVAIL(txq) >= MAX_TX_DESC(ctx) + 2; i++) {
3910 		int rem = do_prefetch ? count - i : 0;
3911 
3912 		mp = _ring_peek_one(r, cidx, i, rem);
3913 		MPASS(mp != NULL && *mp != NULL);
3914 
3915 		/*
3916 		 * Completion interrupts will use the address of the txq
3917 		 * as a sentinel to enqueue _something_ in order to acquire
3918 		 * the lock on the mp_ring (there's no direct lock call).
3919 		 * We obviously whave to check for these sentinel cases
3920 		 * and skip them.
3921 		 */
3922 		if (__predict_false(*mp == (struct mbuf *)txq)) {
3923 			skipped++;
3924 			continue;
3925 		}
3926 		err = iflib_encap(txq, mp);
3927 		if (__predict_false(err)) {
3928 			/* no room - bail out */
3929 			if (err == ENOBUFS)
3930 				break;
3931 			skipped++;
3932 			/* we can't send this packet - skip it */
3933 			continue;
3934 		}
3935 		pkt_sent++;
3936 		m = *mp;
3937 		DBG_COUNTER_INC(tx_sent);
3938 		bytes_sent += m->m_pkthdr.len;
3939 		mcast_sent += !!(m->m_flags & M_MCAST);
3940 
3941 		if (__predict_false(!(ifp->if_drv_flags & IFF_DRV_RUNNING)))
3942 			break;
3943 		ETHER_BPF_MTAP(ifp, m);
3944 		rang = iflib_txd_db_check(txq, false);
3945 	}
3946 
3947 	/* deliberate use of bitwise or to avoid gratuitous short-circuit */
3948 	ring = rang ? false  : (iflib_min_tx_latency | err);
3949 	iflib_txd_db_check(txq, ring);
3950 	if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
3951 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
3952 	if (mcast_sent)
3953 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
3954 #ifdef INVARIANTS
3955 	if (iflib_verbose_debug)
3956 		printf("consumed=%d\n", skipped + pkt_sent);
3957 #endif
3958 	return (skipped + pkt_sent);
3959 }
3960 
3961 static uint32_t
3962 iflib_txq_drain_always(struct ifmp_ring *r)
3963 {
3964 	return (1);
3965 }
3966 
3967 static uint32_t
3968 iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3969 {
3970 	int i, avail;
3971 	struct mbuf **mp;
3972 	iflib_txq_t txq;
3973 
3974 	txq = r->cookie;
3975 
3976 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3977 	CALLOUT_LOCK(txq);
3978 	callout_stop(&txq->ift_timer);
3979 	CALLOUT_UNLOCK(txq);
3980 
3981 	avail = IDXDIFF(pidx, cidx, r->size);
3982 	for (i = 0; i < avail; i++) {
3983 		mp = _ring_peek_one(r, cidx, i, avail - i);
3984 		if (__predict_false(*mp == (struct mbuf *)txq))
3985 			continue;
3986 		m_freem(*mp);
3987 		DBG_COUNTER_INC(tx_frees);
3988 	}
3989 	MPASS(ifmp_ring_is_stalled(r) == 0);
3990 	return (avail);
3991 }
3992 
3993 static void
3994 iflib_ifmp_purge(iflib_txq_t txq)
3995 {
3996 	struct ifmp_ring *r;
3997 
3998 	r = txq->ift_br;
3999 	r->drain = iflib_txq_drain_free;
4000 	r->can_drain = iflib_txq_drain_always;
4001 
4002 	ifmp_ring_check_drainage(r, r->size);
4003 
4004 	r->drain = iflib_txq_drain;
4005 	r->can_drain = iflib_txq_can_drain;
4006 }
4007 
4008 static void
4009 _task_fn_tx(void *context)
4010 {
4011 	iflib_txq_t txq = context;
4012 	if_ctx_t ctx = txq->ift_ctx;
4013 	if_t ifp = ctx->ifc_ifp;
4014 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
4015 
4016 #ifdef IFLIB_DIAGNOSTICS
4017 	txq->ift_cpu_exec_count[curcpu]++;
4018 #endif
4019 	if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
4020 		return;
4021 #ifdef DEV_NETMAP
4022 	if ((if_getcapenable(ifp) & IFCAP_NETMAP) &&
4023 	    netmap_tx_irq(ifp, txq->ift_id))
4024 		goto skip_ifmp;
4025 #endif
4026 #ifdef ALTQ
4027 	if (ALTQ_IS_ENABLED(&ifp->if_snd))
4028 		iflib_altq_if_start(ifp);
4029 #endif
4030 	if (txq->ift_db_pending)
4031 		ifmp_ring_enqueue(txq->ift_br, (void **)&txq, 1, TX_BATCH_SIZE, abdicate);
4032 	else if (!abdicate)
4033 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
4034 	/*
4035 	 * When abdicating, we always need to check drainage, not just when we don't enqueue
4036 	 */
4037 	if (abdicate)
4038 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
4039 #ifdef DEV_NETMAP
4040 skip_ifmp:
4041 #endif
4042 	if (ctx->ifc_flags & IFC_LEGACY)
4043 		IFDI_INTR_ENABLE(ctx);
4044 	else
4045 		IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
4046 }
4047 
4048 static void
4049 _task_fn_rx(void *context)
4050 {
4051 	iflib_rxq_t rxq = context;
4052 	if_ctx_t ctx = rxq->ifr_ctx;
4053 	uint8_t more;
4054 	uint16_t budget;
4055 #ifdef DEV_NETMAP
4056 	u_int work = 0;
4057 	int nmirq;
4058 #endif
4059 
4060 #ifdef IFLIB_DIAGNOSTICS
4061 	rxq->ifr_cpu_exec_count[curcpu]++;
4062 #endif
4063 	DBG_COUNTER_INC(task_fn_rxs);
4064 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
4065 		return;
4066 #ifdef DEV_NETMAP
4067 	nmirq = netmap_rx_irq(ctx->ifc_ifp, rxq->ifr_id, &work);
4068 	if (nmirq != NM_IRQ_PASS) {
4069 		more = (nmirq == NM_IRQ_RESCHED) ? IFLIB_RXEOF_MORE : 0;
4070 		goto skip_rxeof;
4071 	}
4072 #endif
4073 	budget = ctx->ifc_sysctl_rx_budget;
4074 	if (budget == 0)
4075 		budget = 16;	/* XXX */
4076 	more = iflib_rxeof(rxq, budget);
4077 #ifdef DEV_NETMAP
4078 skip_rxeof:
4079 #endif
4080 	if ((more & IFLIB_RXEOF_MORE) == 0) {
4081 		if (ctx->ifc_flags & IFC_LEGACY)
4082 			IFDI_INTR_ENABLE(ctx);
4083 		else
4084 			IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
4085 		DBG_COUNTER_INC(rx_intr_enables);
4086 	}
4087 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
4088 		return;
4089 
4090 	if (more & IFLIB_RXEOF_MORE)
4091 		GROUPTASK_ENQUEUE(&rxq->ifr_task);
4092 	else if (more & IFLIB_RXEOF_EMPTY)
4093 #ifndef __HAIKU__
4094 		callout_reset_curcpu(&rxq->ifr_watchdog, 1, &_task_fn_rx_watchdog, rxq);
4095 #else
4096 		callout_reset(&rxq->ifr_watchdog, 1, &_task_fn_rx_watchdog, rxq);
4097 #endif
4098 }
4099 
4100 static void
4101 _task_fn_admin(void *context)
4102 {
4103 	if_ctx_t ctx = context;
4104 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
4105 	iflib_txq_t txq;
4106 	int i;
4107 	bool oactive, running, do_reset, do_watchdog, in_detach;
4108 
4109 	STATE_LOCK(ctx);
4110 	running = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING);
4111 	oactive = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE);
4112 	do_reset = (ctx->ifc_flags & IFC_DO_RESET);
4113 	do_watchdog = (ctx->ifc_flags & IFC_DO_WATCHDOG);
4114 	in_detach = (ctx->ifc_flags & IFC_IN_DETACH);
4115 	ctx->ifc_flags &= ~(IFC_DO_RESET|IFC_DO_WATCHDOG);
4116 	STATE_UNLOCK(ctx);
4117 
4118 	if ((!running && !oactive) && !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
4119 		return;
4120 	if (in_detach)
4121 		return;
4122 
4123 	CTX_LOCK(ctx);
4124 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
4125 		CALLOUT_LOCK(txq);
4126 		callout_stop(&txq->ift_timer);
4127 		CALLOUT_UNLOCK(txq);
4128 	}
4129 	if (do_watchdog) {
4130 		ctx->ifc_watchdog_events++;
4131 		IFDI_WATCHDOG_RESET(ctx);
4132 	}
4133 	IFDI_UPDATE_ADMIN_STATUS(ctx);
4134 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
4135 		callout_reset_on(&txq->ift_timer, iflib_timer_default, iflib_timer, txq,
4136 		    txq->ift_timer.c_cpu);
4137 	}
4138 	IFDI_LINK_INTR_ENABLE(ctx);
4139 	if (do_reset)
4140 		iflib_if_init_locked(ctx);
4141 	CTX_UNLOCK(ctx);
4142 
4143 	if (LINK_ACTIVE(ctx) == 0)
4144 		return;
4145 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
4146 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
4147 }
4148 
4149 static void
4150 _task_fn_iov(void *context)
4151 {
4152 	if_ctx_t ctx = context;
4153 
4154 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING) &&
4155 	    !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
4156 		return;
4157 
4158 	CTX_LOCK(ctx);
4159 	IFDI_VFLR_HANDLE(ctx);
4160 	CTX_UNLOCK(ctx);
4161 }
4162 
4163 static int
4164 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)
4165 {
4166 	int err;
4167 	if_int_delay_info_t info;
4168 	if_ctx_t ctx;
4169 
4170 	info = (if_int_delay_info_t)arg1;
4171 	ctx = info->iidi_ctx;
4172 	info->iidi_req = req;
4173 	info->iidi_oidp = oidp;
4174 	CTX_LOCK(ctx);
4175 	err = IFDI_SYSCTL_INT_DELAY(ctx, info);
4176 	CTX_UNLOCK(ctx);
4177 	return (err);
4178 }
4179 
4180 /*********************************************************************
4181  *
4182  *  IFNET FUNCTIONS
4183  *
4184  **********************************************************************/
4185 
4186 static void
4187 iflib_if_init_locked(if_ctx_t ctx)
4188 {
4189 	iflib_stop(ctx);
4190 	iflib_init_locked(ctx);
4191 }
4192 
4193 static void
4194 iflib_if_init(void *arg)
4195 {
4196 	if_ctx_t ctx = arg;
4197 
4198 	CTX_LOCK(ctx);
4199 	iflib_if_init_locked(ctx);
4200 	CTX_UNLOCK(ctx);
4201 }
4202 
4203 static int
4204 iflib_if_transmit(if_t ifp, struct mbuf *m)
4205 {
4206 	if_ctx_t	ctx = if_getsoftc(ifp);
4207 
4208 	iflib_txq_t txq;
4209 	int err, qidx;
4210 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
4211 
4212 	if (__predict_false((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) {
4213 		DBG_COUNTER_INC(tx_frees);
4214 		m_freem(m);
4215 		return (ENETDOWN);
4216 	}
4217 
4218 	MPASS(m->m_nextpkt == NULL);
4219 	/* ALTQ-enabled interfaces always use queue 0. */
4220 	qidx = 0;
4221 	if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m) && !ALTQ_IS_ENABLED(&ifp->if_snd))
4222 		qidx = QIDX(ctx, m);
4223 	/*
4224 	 * XXX calculate buf_ring based on flowid (divvy up bits?)
4225 	 */
4226 	txq = &ctx->ifc_txqs[qidx];
4227 
4228 #ifdef DRIVER_BACKPRESSURE
4229 	if (txq->ift_closed) {
4230 		while (m != NULL) {
4231 			next = m->m_nextpkt;
4232 			m->m_nextpkt = NULL;
4233 			m_freem(m);
4234 			DBG_COUNTER_INC(tx_frees);
4235 			m = next;
4236 		}
4237 		return (ENOBUFS);
4238 	}
4239 #endif
4240 #ifdef notyet
4241 	qidx = count = 0;
4242 	mp = marr;
4243 	next = m;
4244 	do {
4245 		count++;
4246 		next = next->m_nextpkt;
4247 	} while (next != NULL);
4248 
4249 	if (count > nitems(marr))
4250 		if ((mp = malloc(count*sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) {
4251 			/* XXX check nextpkt */
4252 			m_freem(m);
4253 			/* XXX simplify for now */
4254 			DBG_COUNTER_INC(tx_frees);
4255 			return (ENOBUFS);
4256 		}
4257 	for (next = m, i = 0; next != NULL; i++) {
4258 		mp[i] = next;
4259 		next = next->m_nextpkt;
4260 		mp[i]->m_nextpkt = NULL;
4261 	}
4262 #endif
4263 	DBG_COUNTER_INC(tx_seen);
4264 	err = ifmp_ring_enqueue(txq->ift_br, (void **)&m, 1, TX_BATCH_SIZE, abdicate);
4265 
4266 	if (abdicate)
4267 		GROUPTASK_ENQUEUE(&txq->ift_task);
4268  	if (err) {
4269 		if (!abdicate)
4270 			GROUPTASK_ENQUEUE(&txq->ift_task);
4271 		/* support forthcoming later */
4272 #ifdef DRIVER_BACKPRESSURE
4273 		txq->ift_closed = TRUE;
4274 #endif
4275 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
4276 		m_freem(m);
4277 		DBG_COUNTER_INC(tx_frees);
4278 	}
4279 
4280 	return (err);
4281 }
4282 
4283 #ifdef ALTQ
4284 /*
4285  * The overall approach to integrating iflib with ALTQ is to continue to use
4286  * the iflib mp_ring machinery between the ALTQ queue(s) and the hardware
4287  * ring.  Technically, when using ALTQ, queueing to an intermediate mp_ring
4288  * is redundant/unnecessary, but doing so minimizes the amount of
4289  * ALTQ-specific code required in iflib.  It is assumed that the overhead of
4290  * redundantly queueing to an intermediate mp_ring is swamped by the
4291  * performance limitations inherent in using ALTQ.
4292  *
4293  * When ALTQ support is compiled in, all iflib drivers will use a transmit
4294  * routine, iflib_altq_if_transmit(), that checks if ALTQ is enabled for the
4295  * given interface.  If ALTQ is enabled for an interface, then all
4296  * transmitted packets for that interface will be submitted to the ALTQ
4297  * subsystem via IFQ_ENQUEUE().  We don't use the legacy if_transmit()
4298  * implementation because it uses IFQ_HANDOFF(), which will duplicatively
4299  * update stats that the iflib machinery handles, and which is sensitve to
4300  * the disused IFF_DRV_OACTIVE flag.  Additionally, iflib_altq_if_start()
4301  * will be installed as the start routine for use by ALTQ facilities that
4302  * need to trigger queue drains on a scheduled basis.
4303  *
4304  */
4305 static void
4306 iflib_altq_if_start(if_t ifp)
4307 {
4308 	struct ifaltq *ifq = &ifp->if_snd;
4309 	struct mbuf *m;
4310 
4311 	IFQ_LOCK(ifq);
4312 	IFQ_DEQUEUE_NOLOCK(ifq, m);
4313 	while (m != NULL) {
4314 		iflib_if_transmit(ifp, m);
4315 		IFQ_DEQUEUE_NOLOCK(ifq, m);
4316 	}
4317 	IFQ_UNLOCK(ifq);
4318 }
4319 
4320 static int
4321 iflib_altq_if_transmit(if_t ifp, struct mbuf *m)
4322 {
4323 	int err;
4324 
4325 	if (ALTQ_IS_ENABLED(&ifp->if_snd)) {
4326 		IFQ_ENQUEUE(&ifp->if_snd, m, err);
4327 		if (err == 0)
4328 			iflib_altq_if_start(ifp);
4329 	} else
4330 		err = iflib_if_transmit(ifp, m);
4331 
4332 	return (err);
4333 }
4334 #endif /* ALTQ */
4335 
4336 static void
4337 iflib_if_qflush(if_t ifp)
4338 {
4339 	if_ctx_t ctx = if_getsoftc(ifp);
4340 	iflib_txq_t txq = ctx->ifc_txqs;
4341 	int i;
4342 
4343 	STATE_LOCK(ctx);
4344 	ctx->ifc_flags |= IFC_QFLUSH;
4345 	STATE_UNLOCK(ctx);
4346 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
4347 		while (!(ifmp_ring_is_idle(txq->ift_br) || ifmp_ring_is_stalled(txq->ift_br)))
4348 			iflib_txq_check_drain(txq, 0);
4349 	STATE_LOCK(ctx);
4350 	ctx->ifc_flags &= ~IFC_QFLUSH;
4351 	STATE_UNLOCK(ctx);
4352 
4353 	/*
4354 	 * When ALTQ is enabled, this will also take care of purging the
4355 	 * ALTQ queue(s).
4356 	 */
4357 	if_qflush(ifp);
4358 }
4359 
4360 #define IFCAP_FLAGS (IFCAP_HWCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \
4361 		     IFCAP_TSO | IFCAP_VLAN_HWTAGGING | IFCAP_HWSTATS | \
4362 		     IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | \
4363 		     IFCAP_VLAN_HWTSO | IFCAP_VLAN_HWCSUM | IFCAP_MEXTPG)
4364 
4365 static int
4366 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data)
4367 {
4368 	if_ctx_t ctx = if_getsoftc(ifp);
4369 	struct ifreq	*ifr = (struct ifreq *)data;
4370 #if defined(INET) || defined(INET6)
4371 	struct ifaddr	*ifa = (struct ifaddr *)data;
4372 #endif
4373 	bool		avoid_reset = false;
4374 	int		err = 0, reinit = 0, bits;
4375 
4376 	switch (command) {
4377 	case SIOCSIFADDR:
4378 #ifdef INET
4379 		if (ifa->ifa_addr->sa_family == AF_INET)
4380 			avoid_reset = true;
4381 #endif
4382 #ifdef INET6
4383 		if (ifa->ifa_addr->sa_family == AF_INET6)
4384 			avoid_reset = true;
4385 #endif
4386 		/*
4387 		** Calling init results in link renegotiation,
4388 		** so we avoid doing it when possible.
4389 		*/
4390 		if (avoid_reset) {
4391 			if_setflagbits(ifp, IFF_UP,0);
4392 			if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
4393 				reinit = 1;
4394 #ifdef INET
4395 			if (!(if_getflags(ifp) & IFF_NOARP))
4396 				arp_ifinit(ifp, ifa);
4397 #endif
4398 		} else
4399 			err = ether_ioctl(ifp, command, data);
4400 		break;
4401 	case SIOCSIFMTU:
4402 		CTX_LOCK(ctx);
4403 		if (ifr->ifr_mtu == if_getmtu(ifp)) {
4404 			CTX_UNLOCK(ctx);
4405 			break;
4406 		}
4407 		bits = if_getdrvflags(ifp);
4408 		/* stop the driver and free any clusters before proceeding */
4409 		iflib_stop(ctx);
4410 
4411 		if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) {
4412 			STATE_LOCK(ctx);
4413 			if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size)
4414 				ctx->ifc_flags |= IFC_MULTISEG;
4415 			else
4416 				ctx->ifc_flags &= ~IFC_MULTISEG;
4417 			STATE_UNLOCK(ctx);
4418 			err = if_setmtu(ifp, ifr->ifr_mtu);
4419 		}
4420 		iflib_init_locked(ctx);
4421 		STATE_LOCK(ctx);
4422 		if_setdrvflags(ifp, bits);
4423 		STATE_UNLOCK(ctx);
4424 		CTX_UNLOCK(ctx);
4425 		break;
4426 	case SIOCSIFFLAGS:
4427 		CTX_LOCK(ctx);
4428 		if (if_getflags(ifp) & IFF_UP) {
4429 			if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4430 				if ((if_getflags(ifp) ^ ctx->ifc_if_flags) &
4431 				    (IFF_PROMISC | IFF_ALLMULTI)) {
4432 					CTX_UNLOCK(ctx);
4433 					err = IFDI_PROMISC_SET(ctx, if_getflags(ifp));
4434 					CTX_LOCK(ctx);
4435 				}
4436 			} else
4437 				reinit = 1;
4438 		} else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4439 			iflib_stop(ctx);
4440 		}
4441 		ctx->ifc_if_flags = if_getflags(ifp);
4442 		CTX_UNLOCK(ctx);
4443 		break;
4444 	case SIOCADDMULTI:
4445 	case SIOCDELMULTI:
4446 		if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4447 			CTX_LOCK(ctx);
4448 			IFDI_INTR_DISABLE(ctx);
4449 			IFDI_MULTI_SET(ctx);
4450 			IFDI_INTR_ENABLE(ctx);
4451 			CTX_UNLOCK(ctx);
4452 		}
4453 		break;
4454 	case SIOCSIFMEDIA:
4455 		CTX_LOCK(ctx);
4456 		IFDI_MEDIA_SET(ctx);
4457 		CTX_UNLOCK(ctx);
4458 		/* FALLTHROUGH */
4459 	case SIOCGIFMEDIA:
4460 #ifndef __HAIKU__
4461 	case SIOCGIFXMEDIA:
4462 #endif
4463 		err = ifmedia_ioctl(ifp, ifr, ctx->ifc_mediap, command);
4464 		break;
4465 #ifndef __HAIKU__
4466 	case SIOCGI2C:
4467 	{
4468 		struct ifi2creq i2c;
4469 
4470 		err = copyin(ifr_data_get_ptr(ifr), &i2c, sizeof(i2c));
4471 		if (err != 0)
4472 			break;
4473 		if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) {
4474 			err = EINVAL;
4475 			break;
4476 		}
4477 		if (i2c.len > sizeof(i2c.data)) {
4478 			err = EINVAL;
4479 			break;
4480 		}
4481 
4482 		if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0)
4483 			err = copyout(&i2c, ifr_data_get_ptr(ifr),
4484 			    sizeof(i2c));
4485 		break;
4486 	}
4487 #endif
4488 	case SIOCSIFCAP:
4489 	{
4490 		int mask, setmask, oldmask;
4491 
4492 		oldmask = if_getcapenable(ifp);
4493 		mask = ifr->ifr_reqcap ^ oldmask;
4494 		mask &= ctx->ifc_softc_ctx.isc_capabilities | IFCAP_MEXTPG;
4495 		setmask = 0;
4496 #ifdef TCP_OFFLOAD
4497 		setmask |= mask & (IFCAP_TOE4|IFCAP_TOE6);
4498 #endif
4499 		setmask |= (mask & IFCAP_FLAGS);
4500 		setmask |= (mask & IFCAP_WOL);
4501 
4502 		/*
4503 		 * If any RX csum has changed, change all the ones that
4504 		 * are supported by the driver.
4505 		 */
4506 		if (setmask & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6)) {
4507 			setmask |= ctx->ifc_softc_ctx.isc_capabilities &
4508 			    (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6);
4509 		}
4510 
4511 		/*
4512 		 * want to ensure that traffic has stopped before we change any of the flags
4513 		 */
4514 		if (setmask) {
4515 			CTX_LOCK(ctx);
4516 			bits = if_getdrvflags(ifp);
4517 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4518 				iflib_stop(ctx);
4519 			STATE_LOCK(ctx);
4520 			if_togglecapenable(ifp, setmask);
4521 			STATE_UNLOCK(ctx);
4522 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4523 				iflib_init_locked(ctx);
4524 			STATE_LOCK(ctx);
4525 			if_setdrvflags(ifp, bits);
4526 			STATE_UNLOCK(ctx);
4527 			CTX_UNLOCK(ctx);
4528 		}
4529 		if_vlancap(ifp);
4530 		break;
4531 	}
4532 	case SIOCGPRIVATE_0:
4533 	case SIOCSDRVSPEC:
4534 	case SIOCGDRVSPEC:
4535 		CTX_LOCK(ctx);
4536 		err = IFDI_PRIV_IOCTL(ctx, command, data);
4537 		CTX_UNLOCK(ctx);
4538 		break;
4539 	default:
4540 		err = ether_ioctl(ifp, command, data);
4541 		break;
4542 	}
4543 	if (reinit)
4544 		iflib_if_init(ctx);
4545 	return (err);
4546 }
4547 
4548 static uint64_t
4549 iflib_if_get_counter(if_t ifp, ift_counter cnt)
4550 {
4551 	if_ctx_t ctx = if_getsoftc(ifp);
4552 
4553 	return (IFDI_GET_COUNTER(ctx, cnt));
4554 }
4555 
4556 /*********************************************************************
4557  *
4558  *  OTHER FUNCTIONS EXPORTED TO THE STACK
4559  *
4560  **********************************************************************/
4561 
4562 static void
4563 iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag)
4564 {
4565 	if_ctx_t ctx = if_getsoftc(ifp);
4566 
4567 	if ((void *)ctx != arg)
4568 		return;
4569 
4570 	if ((vtag == 0) || (vtag > 4095))
4571 		return;
4572 
4573 	if (iflib_in_detach(ctx))
4574 		return;
4575 
4576 	CTX_LOCK(ctx);
4577 	/* Driver may need all untagged packets to be flushed */
4578 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4579 		iflib_stop(ctx);
4580 	IFDI_VLAN_REGISTER(ctx, vtag);
4581 	/* Re-init to load the changes, if required */
4582 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4583 		iflib_init_locked(ctx);
4584 	CTX_UNLOCK(ctx);
4585 }
4586 
4587 static void
4588 iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag)
4589 {
4590 	if_ctx_t ctx = if_getsoftc(ifp);
4591 
4592 	if ((void *)ctx != arg)
4593 		return;
4594 
4595 	if ((vtag == 0) || (vtag > 4095))
4596 		return;
4597 
4598 	CTX_LOCK(ctx);
4599 	/* Driver may need all tagged packets to be flushed */
4600 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4601 		iflib_stop(ctx);
4602 	IFDI_VLAN_UNREGISTER(ctx, vtag);
4603 	/* Re-init to load the changes, if required */
4604 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4605 		iflib_init_locked(ctx);
4606 	CTX_UNLOCK(ctx);
4607 }
4608 
4609 static void
4610 iflib_led_func(void *arg, int onoff)
4611 {
4612 	if_ctx_t ctx = arg;
4613 
4614 	CTX_LOCK(ctx);
4615 	IFDI_LED_FUNC(ctx, onoff);
4616 	CTX_UNLOCK(ctx);
4617 }
4618 
4619 /*********************************************************************
4620  *
4621  *  BUS FUNCTION DEFINITIONS
4622  *
4623  **********************************************************************/
4624 
4625 int
4626 iflib_device_probe(device_t dev)
4627 {
4628 	const pci_vendor_info_t *ent;
4629 	if_shared_ctx_t sctx;
4630 	uint16_t pci_device_id, pci_rev_id, pci_subdevice_id, pci_subvendor_id;
4631 	uint16_t pci_vendor_id;
4632 
4633 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4634 		return (ENOTSUP);
4635 
4636 	pci_vendor_id = pci_get_vendor(dev);
4637 	pci_device_id = pci_get_device(dev);
4638 	pci_subvendor_id = pci_get_subvendor(dev);
4639 	pci_subdevice_id = pci_get_subdevice(dev);
4640 	pci_rev_id = pci_get_revid(dev);
4641 	if (sctx->isc_parse_devinfo != NULL)
4642 		sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id);
4643 
4644 	ent = sctx->isc_vendor_info;
4645 	while (ent->pvi_vendor_id != 0) {
4646 		if (pci_vendor_id != ent->pvi_vendor_id) {
4647 			ent++;
4648 			continue;
4649 		}
4650 		if ((pci_device_id == ent->pvi_device_id) &&
4651 		    ((pci_subvendor_id == ent->pvi_subvendor_id) ||
4652 		     (ent->pvi_subvendor_id == 0)) &&
4653 		    ((pci_subdevice_id == ent->pvi_subdevice_id) ||
4654 		     (ent->pvi_subdevice_id == 0)) &&
4655 		    ((pci_rev_id == ent->pvi_rev_id) ||
4656 		     (ent->pvi_rev_id == 0))) {
4657 			device_set_desc_copy(dev, ent->pvi_name);
4658 			/* this needs to be changed to zero if the bus probing code
4659 			 * ever stops re-probing on best match because the sctx
4660 			 * may have its values over written by register calls
4661 			 * in subsequent probes
4662 			 */
4663 			return (BUS_PROBE_DEFAULT);
4664 		}
4665 		ent++;
4666 	}
4667 	return (ENXIO);
4668 }
4669 
4670 int
4671 iflib_device_probe_vendor(device_t dev)
4672 {
4673 	int probe;
4674 
4675 	probe = iflib_device_probe(dev);
4676 #ifndef __HAIKU__
4677 	if (probe == BUS_PROBE_DEFAULT)
4678 		return (BUS_PROBE_VENDOR);
4679 	else
4680 #endif
4681 		return (probe);
4682 }
4683 
4684 static void
4685 iflib_reset_qvalues(if_ctx_t ctx)
4686 {
4687 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4688 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4689 	device_t dev = ctx->ifc_dev;
4690 	int i;
4691 
4692 	if (ctx->ifc_sysctl_ntxqs != 0)
4693 		scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs;
4694 	if (ctx->ifc_sysctl_nrxqs != 0)
4695 		scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs;
4696 
4697 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4698 		if (ctx->ifc_sysctl_ntxds[i] != 0)
4699 			scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i];
4700 		else
4701 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4702 	}
4703 
4704 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4705 		if (ctx->ifc_sysctl_nrxds[i] != 0)
4706 			scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i];
4707 		else
4708 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4709 	}
4710 
4711 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4712 		if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) {
4713 			device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n",
4714 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]);
4715 			scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i];
4716 		}
4717 		if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) {
4718 			device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n",
4719 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]);
4720 			scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i];
4721 		}
4722 		if (!powerof2(scctx->isc_nrxd[i])) {
4723 			device_printf(dev, "nrxd%d: %d is not a power of 2 - using default value of %d\n",
4724 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_default[i]);
4725 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4726 		}
4727 	}
4728 
4729 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4730 		if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) {
4731 			device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n",
4732 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]);
4733 			scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i];
4734 		}
4735 		if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) {
4736 			device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n",
4737 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]);
4738 			scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i];
4739 		}
4740 		if (!powerof2(scctx->isc_ntxd[i])) {
4741 			device_printf(dev, "ntxd%d: %d is not a power of 2 - using default value of %d\n",
4742 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_default[i]);
4743 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4744 		}
4745 	}
4746 }
4747 
4748 static void
4749 iflib_add_pfil(if_ctx_t ctx)
4750 {
4751 #ifndef __HAIKU__
4752 	struct pfil_head *pfil;
4753 	struct pfil_head_args pa;
4754 	iflib_rxq_t rxq;
4755 	int i;
4756 
4757 	pa.pa_version = PFIL_VERSION;
4758 	pa.pa_flags = PFIL_IN;
4759 	pa.pa_type = PFIL_TYPE_ETHERNET;
4760 	pa.pa_headname = ctx->ifc_ifp->if_xname;
4761 	pfil = pfil_head_register(&pa);
4762 
4763 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4764 		rxq->pfil = pfil;
4765 	}
4766 #endif
4767 }
4768 
4769 static void
4770 iflib_rem_pfil(if_ctx_t ctx)
4771 {
4772 #ifndef __HAIKU__
4773 	struct pfil_head *pfil;
4774 	iflib_rxq_t rxq;
4775 	int i;
4776 
4777 	rxq = ctx->ifc_rxqs;
4778 	pfil = rxq->pfil;
4779 	for (i = 0; i < NRXQSETS(ctx); i++, rxq++) {
4780 		rxq->pfil = NULL;
4781 	}
4782 	pfil_head_unregister(pfil);
4783 #endif
4784 }
4785 
4786 
4787 #ifndef __HAIKU__
4788 /*
4789  * Advance forward by n members of the cpuset ctx->ifc_cpus starting from
4790  * cpuid and wrapping as necessary.
4791  */
4792 static unsigned int
4793 cpuid_advance(if_ctx_t ctx, unsigned int cpuid, unsigned int n)
4794 {
4795 	unsigned int first_valid;
4796 	unsigned int last_valid;
4797 
4798 	/* cpuid should always be in the valid set */
4799 	MPASS(CPU_ISSET(cpuid, &ctx->ifc_cpus));
4800 
4801 	/* valid set should never be empty */
4802 	MPASS(!CPU_EMPTY(&ctx->ifc_cpus));
4803 
4804 	first_valid = CPU_FFS(&ctx->ifc_cpus) - 1;
4805 	last_valid = CPU_FLS(&ctx->ifc_cpus) - 1;
4806 	n = n % CPU_COUNT(&ctx->ifc_cpus);
4807 	while (n > 0) {
4808 		do {
4809 			cpuid++;
4810 			if (cpuid > last_valid)
4811 				cpuid = first_valid;
4812 		} while (!CPU_ISSET(cpuid, &ctx->ifc_cpus));
4813 		n--;
4814 	}
4815 
4816 	return (cpuid);
4817 }
4818 #endif
4819 
4820 #if defined(SMP) && defined(SCHED_ULE)
4821 extern struct cpu_group *cpu_top;              /* CPU topology */
4822 
4823 static int
4824 find_child_with_core(int cpu, struct cpu_group *grp)
4825 {
4826 	int i;
4827 
4828 	if (grp->cg_children == 0)
4829 		return -1;
4830 
4831 	MPASS(grp->cg_child);
4832 	for (i = 0; i < grp->cg_children; i++) {
4833 		if (CPU_ISSET(cpu, &grp->cg_child[i].cg_mask))
4834 			return i;
4835 	}
4836 
4837 	return -1;
4838 }
4839 
4840 
4841 /*
4842  * Find an L2 neighbor of the given CPU or return -1 if none found.  This
4843  * does not distinguish among multiple L2 neighbors if the given CPU has
4844  * more than one (it will always return the same result in that case).
4845  */
4846 static int
4847 find_l2_neighbor(int cpu)
4848 {
4849 	struct cpu_group *grp;
4850 	int i;
4851 
4852 	grp = cpu_top;
4853 	if (grp == NULL)
4854 		return -1;
4855 
4856 	/*
4857 	 * Find the smallest CPU group that contains the given core.
4858 	 */
4859 	i = 0;
4860 	while ((i = find_child_with_core(cpu, grp)) != -1) {
4861 		/*
4862 		 * If the smallest group containing the given CPU has less
4863 		 * than two members, we conclude the given CPU has no
4864 		 * L2 neighbor.
4865 		 */
4866 		if (grp->cg_child[i].cg_count <= 1)
4867 			return (-1);
4868 		grp = &grp->cg_child[i];
4869 	}
4870 
4871 	/* Must share L2. */
4872 	if (grp->cg_level > CG_SHARE_L2 || grp->cg_level == CG_SHARE_NONE)
4873 		return -1;
4874 
4875 	/*
4876 	 * Select the first member of the set that isn't the reference
4877 	 * CPU, which at this point is guaranteed to exist.
4878 	 */
4879 	for (i = 0; i < CPU_SETSIZE; i++) {
4880 		if (CPU_ISSET(i, &grp->cg_mask) && i != cpu)
4881 			return (i);
4882 	}
4883 
4884 	/* Should never be reached */
4885 	return (-1);
4886 }
4887 
4888 #else
4889 static int
4890 find_l2_neighbor(int cpu)
4891 {
4892 
4893 	return (-1);
4894 }
4895 #endif
4896 
4897 #ifndef __HAIKU__
4898 /*
4899  * CPU mapping behaviors
4900  * ---------------------
4901  * 'separate txrx' refers to the separate_txrx sysctl
4902  * 'use logical' refers to the use_logical_cores sysctl
4903  * 'INTR CPUS' indicates whether bus_get_cpus(INTR_CPUS) succeeded
4904  *
4905  *  separate     use     INTR
4906  *    txrx     logical   CPUS   result
4907  * ---------- --------- ------ ------------------------------------------------
4908  *     -          -       X     RX and TX queues mapped to consecutive physical
4909  *                              cores with RX/TX pairs on same core and excess
4910  *                              of either following
4911  *     -          X       X     RX and TX queues mapped to consecutive cores
4912  *                              of any type with RX/TX pairs on same core and
4913  *                              excess of either following
4914  *     X          -       X     RX and TX queues mapped to consecutive physical
4915  *                              cores; all RX then all TX
4916  *     X          X       X     RX queues mapped to consecutive physical cores
4917  *                              first, then TX queues mapped to L2 neighbor of
4918  *                              the corresponding RX queue if one exists,
4919  *                              otherwise to consecutive physical cores
4920  *     -         n/a      -     RX and TX queues mapped to consecutive cores of
4921  *                              any type with RX/TX pairs on same core and excess
4922  *                              of either following
4923  *     X         n/a      -     RX and TX queues mapped to consecutive cores of
4924  *                              any type; all RX then all TX
4925  */
4926 static unsigned int
4927 get_cpuid_for_queue(if_ctx_t ctx, unsigned int base_cpuid, unsigned int qid,
4928     bool is_tx)
4929 {
4930 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4931 	unsigned int core_index;
4932 
4933 	if (ctx->ifc_sysctl_separate_txrx) {
4934 		/*
4935 		 * When using separate CPUs for TX and RX, the assignment
4936 		 * will always be of a consecutive CPU out of the set of
4937 		 * context CPUs, except for the specific case where the
4938 		 * context CPUs are phsyical cores, the use of logical cores
4939 		 * has been enabled, the assignment is for TX, the TX qid
4940 		 * corresponds to an RX qid, and the CPU assigned to the
4941 		 * corresponding RX queue has an L2 neighbor.
4942 		 */
4943 		if (ctx->ifc_sysctl_use_logical_cores &&
4944 		    ctx->ifc_cpus_are_physical_cores &&
4945 		    is_tx && qid < scctx->isc_nrxqsets) {
4946 			int l2_neighbor;
4947 			unsigned int rx_cpuid;
4948 
4949 			rx_cpuid = cpuid_advance(ctx, base_cpuid, qid);
4950 			l2_neighbor = find_l2_neighbor(rx_cpuid);
4951 			if (l2_neighbor != -1) {
4952 				return (l2_neighbor);
4953 			}
4954 			/*
4955 			 * ... else fall through to the normal
4956 			 * consecutive-after-RX assignment scheme.
4957 			 *
4958 			 * Note that we are assuming that all RX queue CPUs
4959 			 * have an L2 neighbor, or all do not.  If a mixed
4960 			 * scenario is possible, we will have to keep track
4961 			 * separately of how many queues prior to this one
4962 			 * were not able to be assigned to an L2 neighbor.
4963 			 */
4964 		}
4965 		if (is_tx)
4966 			core_index = scctx->isc_nrxqsets + qid;
4967 		else
4968 			core_index = qid;
4969 	} else {
4970 		core_index = qid;
4971 	}
4972 
4973 	return (cpuid_advance(ctx, base_cpuid, core_index));
4974 }
4975 #else
4976 #define get_cpuid_for_queue(...) CPU_FIRST()
4977 #endif
4978 
4979 static uint16_t
4980 get_ctx_core_offset(if_ctx_t ctx)
4981 {
4982 #ifndef __HAIKU__
4983 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4984 	struct cpu_offset *op;
4985 	cpuset_t assigned_cpus;
4986 	unsigned int cores_consumed;
4987 	unsigned int base_cpuid = ctx->ifc_sysctl_core_offset;
4988 	unsigned int first_valid;
4989 	unsigned int last_valid;
4990 	unsigned int i;
4991 
4992 	first_valid = CPU_FFS(&ctx->ifc_cpus) - 1;
4993 	last_valid = CPU_FLS(&ctx->ifc_cpus) - 1;
4994 
4995 	if (base_cpuid != CORE_OFFSET_UNSPECIFIED) {
4996 		/*
4997 		 * Align the user-chosen base CPU ID to the next valid CPU
4998 		 * for this device.  If the chosen base CPU ID is smaller
4999 		 * than the first valid CPU or larger than the last valid
5000 		 * CPU, we assume the user does not know what the valid
5001 		 * range is for this device and is thinking in terms of a
5002 		 * zero-based reference frame, and so we shift the given
5003 		 * value into the valid range (and wrap accordingly) so the
5004 		 * intent is translated to the proper frame of reference.
5005 		 * If the base CPU ID is within the valid first/last, but
5006 		 * does not correspond to a valid CPU, it is advanced to the
5007 		 * next valid CPU (wrapping if necessary).
5008 		 */
5009 		if (base_cpuid < first_valid || base_cpuid > last_valid) {
5010 			/* shift from zero-based to first_valid-based */
5011 			base_cpuid += first_valid;
5012 			/* wrap to range [first_valid, last_valid] */
5013 			base_cpuid = (base_cpuid - first_valid) %
5014 			    (last_valid - first_valid + 1);
5015 		}
5016 		if (!CPU_ISSET(base_cpuid, &ctx->ifc_cpus)) {
5017 			/*
5018 			 * base_cpuid is in [first_valid, last_valid], but
5019 			 * not a member of the valid set.  In this case,
5020 			 * there will always be a member of the valid set
5021 			 * with a CPU ID that is greater than base_cpuid,
5022 			 * and we simply advance to it.
5023 			 */
5024 			while (!CPU_ISSET(base_cpuid, &ctx->ifc_cpus))
5025 				base_cpuid++;
5026 		}
5027 		return (base_cpuid);
5028 	}
5029 
5030 	/*
5031 	 * Determine how many cores will be consumed by performing the CPU
5032 	 * assignments and counting how many of the assigned CPUs correspond
5033 	 * to CPUs in the set of context CPUs.  This is done using the CPU
5034 	 * ID first_valid as the base CPU ID, as the base CPU must be within
5035 	 * the set of context CPUs.
5036 	 *
5037 	 * Note not all assigned CPUs will be in the set of context CPUs
5038 	 * when separate CPUs are being allocated to TX and RX queues,
5039 	 * assignment to logical cores has been enabled, the set of context
5040 	 * CPUs contains only physical CPUs, and TX queues are mapped to L2
5041 	 * neighbors of CPUs that RX queues have been mapped to - in this
5042 	 * case we do only want to count how many CPUs in the set of context
5043 	 * CPUs have been consumed, as that determines the next CPU in that
5044 	 * set to start allocating at for the next device for which
5045 	 * core_offset is not set.
5046 	 */
5047 	CPU_ZERO(&assigned_cpus);
5048 	for (i = 0; i < scctx->isc_ntxqsets; i++)
5049 		CPU_SET(get_cpuid_for_queue(ctx, first_valid, i, true),
5050 		    &assigned_cpus);
5051 	for (i = 0; i < scctx->isc_nrxqsets; i++)
5052 		CPU_SET(get_cpuid_for_queue(ctx, first_valid, i, false),
5053 		    &assigned_cpus);
5054 	CPU_AND(&assigned_cpus, &ctx->ifc_cpus);
5055 	cores_consumed = CPU_COUNT(&assigned_cpus);
5056 
5057 	mtx_lock(&cpu_offset_mtx);
5058 	SLIST_FOREACH(op, &cpu_offsets, entries) {
5059 		if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) {
5060 			base_cpuid = op->next_cpuid;
5061 			op->next_cpuid = cpuid_advance(ctx, op->next_cpuid,
5062 			    cores_consumed);
5063 			MPASS(op->refcount < UINT_MAX);
5064 			op->refcount++;
5065 			break;
5066 		}
5067 	}
5068 	if (base_cpuid == CORE_OFFSET_UNSPECIFIED) {
5069 		base_cpuid = first_valid;
5070 		op = malloc(sizeof(struct cpu_offset), M_IFLIB,
5071 		    M_NOWAIT | M_ZERO);
5072 		if (op == NULL) {
5073 			device_printf(ctx->ifc_dev,
5074 			    "allocation for cpu offset failed.\n");
5075 		} else {
5076 			op->next_cpuid = cpuid_advance(ctx, base_cpuid,
5077 			    cores_consumed);
5078 			op->refcount = 1;
5079 			CPU_COPY(&ctx->ifc_cpus, &op->set);
5080 			SLIST_INSERT_HEAD(&cpu_offsets, op, entries);
5081 		}
5082 	}
5083 	mtx_unlock(&cpu_offset_mtx);
5084 
5085 	return (base_cpuid);
5086 #else
5087 	return 0;
5088 #endif
5089 }
5090 
5091 static void
5092 unref_ctx_core_offset(if_ctx_t ctx)
5093 {
5094 #ifndef __HAIKU__
5095 	struct cpu_offset *op, *top;
5096 
5097 	mtx_lock(&cpu_offset_mtx);
5098 	SLIST_FOREACH_SAFE(op, &cpu_offsets, entries, top) {
5099 		if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) {
5100 			MPASS(op->refcount > 0);
5101 			op->refcount--;
5102 			if (op->refcount == 0) {
5103 				SLIST_REMOVE(&cpu_offsets, op, cpu_offset, entries);
5104 				free(op, M_IFLIB);
5105 			}
5106 			break;
5107 		}
5108 	}
5109 	mtx_unlock(&cpu_offset_mtx);
5110 #endif
5111 }
5112 
5113 int
5114 iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp)
5115 {
5116 	if_ctx_t ctx;
5117 	if_t ifp;
5118 	if_softc_ctx_t scctx;
5119 	kobjop_desc_t kobj_desc;
5120 	kobj_method_t *kobj_method;
5121 	int err, msix, rid;
5122 	int num_txd, num_rxd;
5123 
5124 	ctx = malloc(sizeof(* ctx), M_IFLIB, M_WAITOK|M_ZERO);
5125 
5126 	if (sc == NULL) {
5127 		sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
5128 		device_set_softc(dev, ctx);
5129 		ctx->ifc_flags |= IFC_SC_ALLOCATED;
5130 	}
5131 
5132 	ctx->ifc_sctx = sctx;
5133 	ctx->ifc_dev = dev;
5134 	ctx->ifc_softc = sc;
5135 
5136 	if ((err = iflib_register(ctx)) != 0) {
5137 		device_printf(dev, "iflib_register failed %d\n", err);
5138 		goto fail_ctx_free;
5139 	}
5140 	iflib_add_device_sysctl_pre(ctx);
5141 
5142 	scctx = &ctx->ifc_softc_ctx;
5143 	ifp = ctx->ifc_ifp;
5144 
5145 	iflib_reset_qvalues(ctx);
5146 	CTX_LOCK(ctx);
5147 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
5148 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
5149 		goto fail_unlock;
5150 	}
5151 	_iflib_pre_assert(scctx);
5152 	ctx->ifc_txrx = *scctx->isc_txrx;
5153 
5154 	if (sctx->isc_flags & IFLIB_DRIVER_MEDIA)
5155 		ctx->ifc_mediap = scctx->isc_media;
5156 
5157 #ifdef INVARIANTS
5158 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
5159 		MPASS(scctx->isc_tx_csum_flags);
5160 #endif
5161 
5162 	if_setcapabilities(ifp,
5163 	    scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_MEXTPG);
5164 	if_setcapenable(ifp,
5165 	    scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_MEXTPG);
5166 
5167 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
5168 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
5169 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
5170 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
5171 
5172 	num_txd = iflib_num_tx_descs(ctx);
5173 	num_rxd = iflib_num_rx_descs(ctx);
5174 
5175 	/* XXX change for per-queue sizes */
5176 	device_printf(dev, "Using %d TX descriptors and %d RX descriptors\n",
5177 	    num_txd, num_rxd);
5178 
5179 	if (scctx->isc_tx_nsegments > num_txd / MAX_SINGLE_PACKET_FRACTION)
5180 		scctx->isc_tx_nsegments = max(1, num_txd /
5181 		    MAX_SINGLE_PACKET_FRACTION);
5182 	if (scctx->isc_tx_tso_segments_max > num_txd /
5183 	    MAX_SINGLE_PACKET_FRACTION)
5184 		scctx->isc_tx_tso_segments_max = max(1,
5185 		    num_txd / MAX_SINGLE_PACKET_FRACTION);
5186 
5187 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
5188 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
5189 #ifndef __HAIKU__
5190 		/*
5191 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
5192 		 * but some MACs do.
5193 		 */
5194 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
5195 		    IP_MAXPACKET));
5196 		/*
5197 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
5198 		 * into account.  In the worst case, each of these calls will
5199 		 * add another mbuf and, thus, the requirement for another DMA
5200 		 * segment.  So for best performance, it doesn't make sense to
5201 		 * advertize a maximum of TSO segments that typically will
5202 		 * require defragmentation in iflib_encap().
5203 		 */
5204 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
5205 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
5206 #endif
5207 	}
5208 	if (scctx->isc_rss_table_size == 0)
5209 		scctx->isc_rss_table_size = 64;
5210 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
5211 
5212 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
5213 	/* XXX format name */
5214 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx,
5215 	    NULL, NULL, "admin");
5216 
5217 #ifndef __HAIKU__
5218 	/* Set up cpu set.  If it fails, use the set of all CPUs. */
5219 	if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) != 0) {
5220 		device_printf(dev, "Unable to fetch CPU list\n");
5221 		CPU_COPY(&all_cpus, &ctx->ifc_cpus);
5222 		ctx->ifc_cpus_are_physical_cores = false;
5223 	} else
5224 		ctx->ifc_cpus_are_physical_cores = true;
5225 	MPASS(CPU_COUNT(&ctx->ifc_cpus) > 0);
5226 #endif
5227 
5228 	/*
5229 	** Now set up MSI or MSI-X, should return us the number of supported
5230 	** vectors (will be 1 for a legacy interrupt and MSI).
5231 	*/
5232 	if (sctx->isc_flags & IFLIB_SKIP_MSIX) {
5233 		msix = scctx->isc_vectors;
5234 	} else if (scctx->isc_msix_bar != 0)
5235 	       /*
5236 		* The simple fact that isc_msix_bar is not 0 does not mean we
5237 		* we have a good value there that is known to work.
5238 		*/
5239 		msix = iflib_msix_init(ctx);
5240 	else {
5241 		scctx->isc_vectors = 1;
5242 		scctx->isc_ntxqsets = 1;
5243 		scctx->isc_nrxqsets = 1;
5244 		scctx->isc_intr = IFLIB_INTR_LEGACY;
5245 		msix = 0;
5246 	}
5247 	/* Get memory for the station queues */
5248 	if ((err = iflib_queues_alloc(ctx))) {
5249 		device_printf(dev, "Unable to allocate queue memory\n");
5250 		goto fail_intr_free;
5251 	}
5252 
5253 	if ((err = iflib_qset_structures_setup(ctx)))
5254 		goto fail_queues;
5255 
5256 	/*
5257 	 * Now that we know how many queues there are, get the core offset.
5258 	 */
5259 	ctx->ifc_sysctl_core_offset = get_ctx_core_offset(ctx);
5260 
5261 	if (msix > 1) {
5262 		/*
5263 		 * When using MSI-X, ensure that ifdi_{r,t}x_queue_intr_enable
5264 		 * aren't the default NULL implementation.
5265 		 */
5266 		kobj_desc = &ifdi_rx_queue_intr_enable_desc;
5267 #ifdef __HAIKU__
5268 		kobj_method = kobj_lookup_method(ctx->ops.cls, NULL,
5269 #else
5270 		kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL,
5271 #endif
5272 		    kobj_desc);
5273 		if (kobj_method == &kobj_desc->deflt) {
5274 			device_printf(dev,
5275 			    "MSI-X requires ifdi_rx_queue_intr_enable method");
5276 			err = EOPNOTSUPP;
5277 			goto fail_queues;
5278 		}
5279 		kobj_desc = &ifdi_tx_queue_intr_enable_desc;
5280 #ifdef __HAIKU__
5281 		kobj_method = kobj_lookup_method(ctx->ops.cls, NULL,
5282 #else
5283 		kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL,
5284 #endif
5285 		    kobj_desc);
5286 		if (kobj_method == &kobj_desc->deflt) {
5287 			device_printf(dev,
5288 			    "MSI-X requires ifdi_tx_queue_intr_enable method");
5289 			err = EOPNOTSUPP;
5290 			goto fail_queues;
5291 		}
5292 
5293 		/*
5294 		 * Assign the MSI-X vectors.
5295 		 * Note that the default NULL ifdi_msix_intr_assign method will
5296 		 * fail here, too.
5297 		 */
5298 		err = IFDI_MSIX_INTR_ASSIGN(ctx, msix);
5299 		if (err != 0) {
5300 			device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n",
5301 			    err);
5302 			goto fail_queues;
5303 		}
5304 	} else if (scctx->isc_intr != IFLIB_INTR_MSIX) {
5305 		rid = 0;
5306 		if (scctx->isc_intr == IFLIB_INTR_MSI) {
5307 			MPASS(msix == 1);
5308 			rid = 1;
5309 		}
5310 		if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) {
5311 			device_printf(dev, "iflib_legacy_setup failed %d\n", err);
5312 			goto fail_queues;
5313 		}
5314 	} else {
5315 		device_printf(dev,
5316 		    "Cannot use iflib with only 1 MSI-X interrupt!\n");
5317 		err = ENODEV;
5318 		goto fail_queues;
5319 	}
5320 
5321 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
5322 
5323 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
5324 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
5325 		goto fail_detach;
5326 	}
5327 
5328 	/*
5329 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
5330 	 * This must appear after the call to ether_ifattach() because
5331 	 * ether_ifattach() sets if_hdrlen to the default value.
5332 	 */
5333 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
5334 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
5335 
5336 	if ((err = iflib_netmap_attach(ctx))) {
5337 		device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err);
5338 		goto fail_detach;
5339 	}
5340 	*ctxp = ctx;
5341 
5342 	DEBUGNET_SET(ctx->ifc_ifp, iflib);
5343 
5344 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
5345 	iflib_add_device_sysctl_post(ctx);
5346 	iflib_add_pfil(ctx);
5347 	ctx->ifc_flags |= IFC_INIT_DONE;
5348 	CTX_UNLOCK(ctx);
5349 
5350 	return (0);
5351 
5352 fail_detach:
5353 	ether_ifdetach(ctx->ifc_ifp);
5354 fail_queues:
5355 	iflib_tqg_detach(ctx);
5356 	iflib_tx_structures_free(ctx);
5357 	iflib_rx_structures_free(ctx);
5358 	IFDI_DETACH(ctx);
5359 	IFDI_QUEUES_FREE(ctx);
5360 fail_intr_free:
5361 	iflib_free_intr_mem(ctx);
5362 fail_unlock:
5363 	CTX_UNLOCK(ctx);
5364 	iflib_deregister(ctx);
5365 fail_ctx_free:
5366 	device_set_softc(ctx->ifc_dev, NULL);
5367         if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5368                 free(ctx->ifc_softc, M_IFLIB);
5369         free(ctx, M_IFLIB);
5370 	return (err);
5371 }
5372 
5373 int
5374 iflib_pseudo_register(device_t dev, if_shared_ctx_t sctx, if_ctx_t *ctxp,
5375 					  struct iflib_cloneattach_ctx *clctx)
5376 {
5377 	int num_txd, num_rxd;
5378 	int err;
5379 	if_ctx_t ctx;
5380 	if_t ifp;
5381 	if_softc_ctx_t scctx;
5382 	int i;
5383 	void *sc;
5384 
5385 	ctx = malloc(sizeof(*ctx), M_IFLIB, M_WAITOK|M_ZERO);
5386 	sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
5387 	ctx->ifc_flags |= IFC_SC_ALLOCATED;
5388 	if (sctx->isc_flags & (IFLIB_PSEUDO|IFLIB_VIRTUAL))
5389 		ctx->ifc_flags |= IFC_PSEUDO;
5390 
5391 	ctx->ifc_sctx = sctx;
5392 	ctx->ifc_softc = sc;
5393 	ctx->ifc_dev = dev;
5394 
5395 	if ((err = iflib_register(ctx)) != 0) {
5396 		device_printf(dev, "%s: iflib_register failed %d\n", __func__, err);
5397 		goto fail_ctx_free;
5398 	}
5399 	iflib_add_device_sysctl_pre(ctx);
5400 
5401 	scctx = &ctx->ifc_softc_ctx;
5402 	ifp = ctx->ifc_ifp;
5403 
5404 	iflib_reset_qvalues(ctx);
5405 	CTX_LOCK(ctx);
5406 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
5407 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
5408 		goto fail_unlock;
5409 	}
5410 #ifndef __HAIKU__
5411 	if (sctx->isc_flags & IFLIB_GEN_MAC)
5412 		ether_gen_addr(ifp, &ctx->ifc_mac);
5413 #endif
5414 	if ((err = IFDI_CLONEATTACH(ctx, clctx->cc_ifc, clctx->cc_name,
5415 								clctx->cc_params)) != 0) {
5416 		device_printf(dev, "IFDI_CLONEATTACH failed %d\n", err);
5417 		goto fail_unlock;
5418 	}
5419 #ifdef INVARIANTS
5420 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
5421 		MPASS(scctx->isc_tx_csum_flags);
5422 #endif
5423 
5424 	if_setcapabilities(ifp, scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_LINKSTATE);
5425 	if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_LINKSTATE);
5426 
5427 	ifp->if_flags |= IFF_NOGROUP;
5428 	if (sctx->isc_flags & IFLIB_PSEUDO) {
5429 		ifmedia_add(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO, 0, NULL);
5430 		ifmedia_set(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO);
5431 		if (sctx->isc_flags & IFLIB_PSEUDO_ETHER) {
5432 			ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
5433 		} else {
5434 			if_attach(ctx->ifc_ifp);
5435 			bpfattach(ctx->ifc_ifp, DLT_NULL, sizeof(u_int32_t));
5436 		}
5437 
5438 		if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
5439 			device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
5440 			goto fail_detach;
5441 		}
5442 		*ctxp = ctx;
5443 
5444 		/*
5445 		 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
5446 		 * This must appear after the call to ether_ifattach() because
5447 		 * ether_ifattach() sets if_hdrlen to the default value.
5448 		 */
5449 		if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
5450 			if_setifheaderlen(ifp,
5451 			    sizeof(struct ether_vlan_header));
5452 
5453 		if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
5454 		iflib_add_device_sysctl_post(ctx);
5455 		ctx->ifc_flags |= IFC_INIT_DONE;
5456 		CTX_UNLOCK(ctx);
5457 		return (0);
5458 	}
5459 	ifmedia_add(ctx->ifc_mediap, IFM_ETHER | IFM_1000_T | IFM_FDX, 0, NULL);
5460 	ifmedia_add(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO, 0, NULL);
5461 	ifmedia_set(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO);
5462 
5463 	_iflib_pre_assert(scctx);
5464 	ctx->ifc_txrx = *scctx->isc_txrx;
5465 
5466 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
5467 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
5468 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
5469 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
5470 
5471 	num_txd = iflib_num_tx_descs(ctx);
5472 	num_rxd = iflib_num_rx_descs(ctx);
5473 
5474 	/* XXX change for per-queue sizes */
5475 	device_printf(dev, "Using %d TX descriptors and %d RX descriptors\n",
5476 	    num_txd, num_rxd);
5477 
5478 	if (scctx->isc_tx_nsegments > num_txd / MAX_SINGLE_PACKET_FRACTION)
5479 		scctx->isc_tx_nsegments = max(1, num_txd /
5480 		    MAX_SINGLE_PACKET_FRACTION);
5481 	if (scctx->isc_tx_tso_segments_max > num_txd /
5482 	    MAX_SINGLE_PACKET_FRACTION)
5483 		scctx->isc_tx_tso_segments_max = max(1,
5484 		    num_txd / MAX_SINGLE_PACKET_FRACTION);
5485 
5486 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
5487 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
5488 #ifndef __HAIKU__
5489 		/*
5490 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
5491 		 * but some MACs do.
5492 		 */
5493 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
5494 		    IP_MAXPACKET));
5495 		/*
5496 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
5497 		 * into account.  In the worst case, each of these calls will
5498 		 * add another mbuf and, thus, the requirement for another DMA
5499 		 * segment.  So for best performance, it doesn't make sense to
5500 		 * advertize a maximum of TSO segments that typically will
5501 		 * require defragmentation in iflib_encap().
5502 		 */
5503 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
5504 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
5505 #endif
5506 	}
5507 	if (scctx->isc_rss_table_size == 0)
5508 		scctx->isc_rss_table_size = 64;
5509 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
5510 
5511 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
5512 	/* XXX format name */
5513 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx,
5514 	    NULL, NULL, "admin");
5515 
5516 	/* XXX --- can support > 1 -- but keep it simple for now */
5517 	scctx->isc_intr = IFLIB_INTR_LEGACY;
5518 
5519 	/* Get memory for the station queues */
5520 	if ((err = iflib_queues_alloc(ctx))) {
5521 		device_printf(dev, "Unable to allocate queue memory\n");
5522 		goto fail_iflib_detach;
5523 	}
5524 
5525 	if ((err = iflib_qset_structures_setup(ctx))) {
5526 		device_printf(dev, "qset structure setup failed %d\n", err);
5527 		goto fail_queues;
5528 	}
5529 
5530 	/*
5531 	 * XXX What if anything do we want to do about interrupts?
5532 	 */
5533 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
5534 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
5535 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
5536 		goto fail_detach;
5537 	}
5538 
5539 	/*
5540 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
5541 	 * This must appear after the call to ether_ifattach() because
5542 	 * ether_ifattach() sets if_hdrlen to the default value.
5543 	 */
5544 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
5545 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
5546 
5547 	/* XXX handle more than one queue */
5548 	for (i = 0; i < scctx->isc_nrxqsets; i++)
5549 		IFDI_RX_CLSET(ctx, 0, i, ctx->ifc_rxqs[i].ifr_fl[0].ifl_sds.ifsd_cl);
5550 
5551 	*ctxp = ctx;
5552 
5553 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
5554 	iflib_add_device_sysctl_post(ctx);
5555 	ctx->ifc_flags |= IFC_INIT_DONE;
5556 	CTX_UNLOCK(ctx);
5557 
5558 	return (0);
5559 fail_detach:
5560 	ether_ifdetach(ctx->ifc_ifp);
5561 fail_queues:
5562 	iflib_tqg_detach(ctx);
5563 	iflib_tx_structures_free(ctx);
5564 	iflib_rx_structures_free(ctx);
5565 fail_iflib_detach:
5566 	IFDI_DETACH(ctx);
5567 	IFDI_QUEUES_FREE(ctx);
5568 fail_unlock:
5569 	CTX_UNLOCK(ctx);
5570 	iflib_deregister(ctx);
5571 fail_ctx_free:
5572 	free(ctx->ifc_softc, M_IFLIB);
5573 	free(ctx, M_IFLIB);
5574 	return (err);
5575 }
5576 
5577 int
5578 iflib_pseudo_deregister(if_ctx_t ctx)
5579 {
5580 	if_t ifp = ctx->ifc_ifp;
5581 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5582 
5583 	/* Unregister VLAN event handlers early */
5584 	iflib_unregister_vlan_handlers(ctx);
5585 
5586 	if ((sctx->isc_flags & IFLIB_PSEUDO)  &&
5587 		(sctx->isc_flags & IFLIB_PSEUDO_ETHER) == 0) {
5588 		bpfdetach(ifp);
5589 		if_detach(ifp);
5590 	} else {
5591 		ether_ifdetach(ifp);
5592 	}
5593 
5594 	iflib_tqg_detach(ctx);
5595 	iflib_tx_structures_free(ctx);
5596 	iflib_rx_structures_free(ctx);
5597 	IFDI_DETACH(ctx);
5598 	IFDI_QUEUES_FREE(ctx);
5599 
5600 	iflib_deregister(ctx);
5601 
5602 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5603 		free(ctx->ifc_softc, M_IFLIB);
5604 	free(ctx, M_IFLIB);
5605 	return (0);
5606 }
5607 
5608 int
5609 iflib_device_attach(device_t dev)
5610 {
5611 	if_ctx_t ctx;
5612 	if_shared_ctx_t sctx;
5613 
5614 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
5615 		return (ENOTSUP);
5616 
5617 	pci_enable_busmaster(dev);
5618 
5619 	return (iflib_device_register(dev, NULL, sctx, &ctx));
5620 }
5621 
5622 int
5623 iflib_device_deregister(if_ctx_t ctx)
5624 {
5625 	if_t ifp = ctx->ifc_ifp;
5626 	device_t dev = ctx->ifc_dev;
5627 
5628 	/* Make sure VLANS are not using driver */
5629 	if (if_vlantrunkinuse(ifp)) {
5630 		device_printf(dev, "Vlan in use, detach first\n");
5631 		return (EBUSY);
5632 	}
5633 #ifdef PCI_IOV
5634 	if (!CTX_IS_VF(ctx) && pci_iov_detach(dev) != 0) {
5635 		device_printf(dev, "SR-IOV in use; detach first.\n");
5636 		return (EBUSY);
5637 	}
5638 #endif
5639 
5640 	STATE_LOCK(ctx);
5641 	ctx->ifc_flags |= IFC_IN_DETACH;
5642 	STATE_UNLOCK(ctx);
5643 
5644 	/* Unregister VLAN handlers before calling iflib_stop() */
5645 	iflib_unregister_vlan_handlers(ctx);
5646 
5647 	iflib_netmap_detach(ifp);
5648 	ether_ifdetach(ifp);
5649 
5650 	CTX_LOCK(ctx);
5651 	iflib_stop(ctx);
5652 	CTX_UNLOCK(ctx);
5653 
5654 	iflib_rem_pfil(ctx);
5655 	if (ctx->ifc_led_dev != NULL)
5656 		led_destroy(ctx->ifc_led_dev);
5657 
5658 	iflib_tqg_detach(ctx);
5659 	iflib_tx_structures_free(ctx);
5660 	iflib_rx_structures_free(ctx);
5661 
5662 	CTX_LOCK(ctx);
5663 	IFDI_DETACH(ctx);
5664 	IFDI_QUEUES_FREE(ctx);
5665 	CTX_UNLOCK(ctx);
5666 
5667 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
5668 	iflib_free_intr_mem(ctx);
5669 
5670 	bus_generic_detach(dev);
5671 
5672 	iflib_deregister(ctx);
5673 
5674 	device_set_softc(ctx->ifc_dev, NULL);
5675 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5676 		free(ctx->ifc_softc, M_IFLIB);
5677 	unref_ctx_core_offset(ctx);
5678 	free(ctx, M_IFLIB);
5679 	return (0);
5680 }
5681 
5682 static void
5683 iflib_tqg_detach(if_ctx_t ctx)
5684 {
5685 	iflib_txq_t txq;
5686 	iflib_rxq_t rxq;
5687 	int i;
5688 	struct taskqgroup *tqg;
5689 
5690 	/* XXX drain any dependent tasks */
5691 	tqg = qgroup_if_io_tqg;
5692 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
5693 		callout_drain(&txq->ift_timer);
5694 #ifdef DEV_NETMAP
5695 		callout_drain(&txq->ift_netmap_timer);
5696 #endif /* DEV_NETMAP */
5697 		if (txq->ift_task.gt_uniq != NULL)
5698 			taskqgroup_detach(tqg, &txq->ift_task);
5699 	}
5700 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
5701 		if (rxq->ifr_task.gt_uniq != NULL)
5702 			taskqgroup_detach(tqg, &rxq->ifr_task);
5703 	}
5704 	tqg = qgroup_if_config_tqg;
5705 	if (ctx->ifc_admin_task.gt_uniq != NULL)
5706 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
5707 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
5708 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
5709 }
5710 
5711 static void
5712 iflib_free_intr_mem(if_ctx_t ctx)
5713 {
5714 
5715 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) {
5716 		iflib_irq_free(ctx, &ctx->ifc_legacy_irq);
5717 	}
5718 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) {
5719 		pci_release_msi(ctx->ifc_dev);
5720 	}
5721 	if (ctx->ifc_msix_mem != NULL) {
5722 		bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY,
5723 		    rman_get_rid(ctx->ifc_msix_mem), ctx->ifc_msix_mem);
5724 		ctx->ifc_msix_mem = NULL;
5725 	}
5726 }
5727 
5728 int
5729 iflib_device_detach(device_t dev)
5730 {
5731 	if_ctx_t ctx = device_get_softc(dev);
5732 
5733 	return (iflib_device_deregister(ctx));
5734 }
5735 
5736 int
5737 iflib_device_suspend(device_t dev)
5738 {
5739 	if_ctx_t ctx = device_get_softc(dev);
5740 
5741 	CTX_LOCK(ctx);
5742 	IFDI_SUSPEND(ctx);
5743 	CTX_UNLOCK(ctx);
5744 
5745 	return bus_generic_suspend(dev);
5746 }
5747 int
5748 iflib_device_shutdown(device_t dev)
5749 {
5750 	if_ctx_t ctx = device_get_softc(dev);
5751 
5752 	CTX_LOCK(ctx);
5753 	IFDI_SHUTDOWN(ctx);
5754 	CTX_UNLOCK(ctx);
5755 
5756 	return bus_generic_suspend(dev);
5757 }
5758 
5759 int
5760 iflib_device_resume(device_t dev)
5761 {
5762 	if_ctx_t ctx = device_get_softc(dev);
5763 	iflib_txq_t txq = ctx->ifc_txqs;
5764 
5765 	CTX_LOCK(ctx);
5766 	IFDI_RESUME(ctx);
5767 	iflib_if_init_locked(ctx);
5768 	CTX_UNLOCK(ctx);
5769 	for (int i = 0; i < NTXQSETS(ctx); i++, txq++)
5770 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
5771 
5772 	return (bus_generic_resume(dev));
5773 }
5774 
5775 int
5776 iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params)
5777 {
5778 	int error;
5779 	if_ctx_t ctx = device_get_softc(dev);
5780 
5781 	CTX_LOCK(ctx);
5782 	error = IFDI_IOV_INIT(ctx, num_vfs, params);
5783 	CTX_UNLOCK(ctx);
5784 
5785 	return (error);
5786 }
5787 
5788 void
5789 iflib_device_iov_uninit(device_t dev)
5790 {
5791 	if_ctx_t ctx = device_get_softc(dev);
5792 
5793 	CTX_LOCK(ctx);
5794 	IFDI_IOV_UNINIT(ctx);
5795 	CTX_UNLOCK(ctx);
5796 }
5797 
5798 int
5799 iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params)
5800 {
5801 	int error;
5802 	if_ctx_t ctx = device_get_softc(dev);
5803 
5804 	CTX_LOCK(ctx);
5805 	error = IFDI_IOV_VF_ADD(ctx, vfnum, params);
5806 	CTX_UNLOCK(ctx);
5807 
5808 	return (error);
5809 }
5810 
5811 /*********************************************************************
5812  *
5813  *  MODULE FUNCTION DEFINITIONS
5814  *
5815  **********************************************************************/
5816 
5817 /*
5818  * - Start a fast taskqueue thread for each core
5819  * - Start a taskqueue for control operations
5820  */
5821 static int
5822 iflib_module_init(void)
5823 {
5824 	iflib_timer_default = hz / 2;
5825 	return (0);
5826 }
5827 
5828 static int
5829 iflib_module_event_handler(module_t mod, int what, void *arg)
5830 {
5831 	int err;
5832 
5833 	switch (what) {
5834 	case MOD_LOAD:
5835 		if ((err = iflib_module_init()) != 0)
5836 			return (err);
5837 		break;
5838 	case MOD_UNLOAD:
5839 		return (EBUSY);
5840 	default:
5841 		return (EOPNOTSUPP);
5842 	}
5843 
5844 	return (0);
5845 }
5846 
5847 /*********************************************************************
5848  *
5849  *  PUBLIC FUNCTION DEFINITIONS
5850  *     ordered as in iflib.h
5851  *
5852  **********************************************************************/
5853 
5854 static void
5855 _iflib_assert(if_shared_ctx_t sctx)
5856 {
5857 	int i;
5858 
5859 	MPASS(sctx->isc_tx_maxsize);
5860 	MPASS(sctx->isc_tx_maxsegsize);
5861 
5862 	MPASS(sctx->isc_rx_maxsize);
5863 	MPASS(sctx->isc_rx_nsegments);
5864 	MPASS(sctx->isc_rx_maxsegsize);
5865 
5866 	MPASS(sctx->isc_nrxqs >= 1 && sctx->isc_nrxqs <= 8);
5867 	for (i = 0; i < sctx->isc_nrxqs; i++) {
5868 		MPASS(sctx->isc_nrxd_min[i]);
5869 		MPASS(powerof2(sctx->isc_nrxd_min[i]));
5870 		MPASS(sctx->isc_nrxd_max[i]);
5871 		MPASS(powerof2(sctx->isc_nrxd_max[i]));
5872 		MPASS(sctx->isc_nrxd_default[i]);
5873 		MPASS(powerof2(sctx->isc_nrxd_default[i]));
5874 	}
5875 
5876 	MPASS(sctx->isc_ntxqs >= 1 && sctx->isc_ntxqs <= 8);
5877 	for (i = 0; i < sctx->isc_ntxqs; i++) {
5878 		MPASS(sctx->isc_ntxd_min[i]);
5879 		MPASS(powerof2(sctx->isc_ntxd_min[i]));
5880 		MPASS(sctx->isc_ntxd_max[i]);
5881 		MPASS(powerof2(sctx->isc_ntxd_max[i]));
5882 		MPASS(sctx->isc_ntxd_default[i]);
5883 		MPASS(powerof2(sctx->isc_ntxd_default[i]));
5884 	}
5885 }
5886 
5887 static void
5888 _iflib_pre_assert(if_softc_ctx_t scctx)
5889 {
5890 
5891 	MPASS(scctx->isc_txrx->ift_txd_encap);
5892 	MPASS(scctx->isc_txrx->ift_txd_flush);
5893 	MPASS(scctx->isc_txrx->ift_txd_credits_update);
5894 	MPASS(scctx->isc_txrx->ift_rxd_available);
5895 	MPASS(scctx->isc_txrx->ift_rxd_pkt_get);
5896 	MPASS(scctx->isc_txrx->ift_rxd_refill);
5897 	MPASS(scctx->isc_txrx->ift_rxd_flush);
5898 }
5899 
5900 static int
5901 iflib_register(if_ctx_t ctx)
5902 {
5903 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5904 	driver_t *driver = sctx->isc_driver;
5905 	device_t dev = ctx->ifc_dev;
5906 	if_t ifp;
5907 	u_char type;
5908 	int iflags;
5909 
5910 	if ((sctx->isc_flags & IFLIB_PSEUDO) == 0)
5911 		_iflib_assert(sctx);
5912 
5913 	CTX_LOCK_INIT(ctx);
5914 	STATE_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev));
5915 	if (sctx->isc_flags & IFLIB_PSEUDO) {
5916 		if (sctx->isc_flags & IFLIB_PSEUDO_ETHER)
5917 			type = IFT_ETHER;
5918 		else
5919 			type = IFT_PPP;
5920 	} else
5921 		type = IFT_ETHER;
5922 	ifp = ctx->ifc_ifp = if_alloc(type);
5923 	if (ifp == NULL) {
5924 		device_printf(dev, "can not allocate ifnet structure\n");
5925 		return (ENOMEM);
5926 	}
5927 
5928 	/*
5929 	 * Initialize our context's device specific methods
5930 	 */
5931 	kobj_init((kobj_t) ctx, (kobj_class_t) driver);
5932 	kobj_class_compile((kobj_class_t) driver);
5933 
5934 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
5935 	if_setsoftc(ifp, ctx);
5936 	if_setdev(ifp, dev);
5937 	if_setinitfn(ifp, iflib_if_init);
5938 	if_setioctlfn(ifp, iflib_if_ioctl);
5939 #ifdef ALTQ
5940 	if_setstartfn(ifp, iflib_altq_if_start);
5941 	if_settransmitfn(ifp, iflib_altq_if_transmit);
5942 	if_setsendqready(ifp);
5943 #else
5944 	if_settransmitfn(ifp, iflib_if_transmit);
5945 #endif
5946 	if_setqflushfn(ifp, iflib_if_qflush);
5947 #ifndef __HAIKU__
5948 	iflags = IFF_MULTICAST | IFF_KNOWSEPOCH;
5949 #else
5950 	iflags = IFF_MULTICAST;
5951 #endif
5952 
5953 	if ((sctx->isc_flags & IFLIB_PSEUDO) &&
5954 		(sctx->isc_flags & IFLIB_PSEUDO_ETHER) == 0)
5955 		iflags |= IFF_POINTOPOINT;
5956 	else
5957 		iflags |= IFF_BROADCAST | IFF_SIMPLEX;
5958 	if_setflags(ifp, iflags);
5959 	ctx->ifc_vlan_attach_event =
5960 		EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx,
5961 							  EVENTHANDLER_PRI_FIRST);
5962 	ctx->ifc_vlan_detach_event =
5963 		EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx,
5964 							  EVENTHANDLER_PRI_FIRST);
5965 
5966 	if ((sctx->isc_flags & IFLIB_DRIVER_MEDIA) == 0) {
5967 		ctx->ifc_mediap = &ctx->ifc_media;
5968 		ifmedia_init(ctx->ifc_mediap, IFM_IMASK,
5969 		    iflib_media_change, iflib_media_status);
5970 	}
5971 	return (0);
5972 }
5973 
5974 static void
5975 iflib_unregister_vlan_handlers(if_ctx_t ctx)
5976 {
5977 	/* Unregister VLAN events */
5978 	if (ctx->ifc_vlan_attach_event != NULL) {
5979 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
5980 		ctx->ifc_vlan_attach_event = NULL;
5981 	}
5982 	if (ctx->ifc_vlan_detach_event != NULL) {
5983 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
5984 		ctx->ifc_vlan_detach_event = NULL;
5985 	}
5986 
5987 }
5988 
5989 static void
5990 iflib_deregister(if_ctx_t ctx)
5991 {
5992 	if_t ifp = ctx->ifc_ifp;
5993 
5994 	/* Remove all media */
5995 	ifmedia_removeall(&ctx->ifc_media);
5996 
5997 	/* Ensure that VLAN event handlers are unregistered */
5998 	iflib_unregister_vlan_handlers(ctx);
5999 
6000 #ifndef __HAIKU__
6001 	/* Release kobject reference */
6002 	kobj_delete((kobj_t) ctx, NULL);
6003 #endif
6004 
6005 	/* Free the ifnet structure */
6006 	if_free(ifp);
6007 
6008 	STATE_LOCK_DESTROY(ctx);
6009 
6010 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
6011 	CTX_LOCK_DESTROY(ctx);
6012 }
6013 
6014 static int
6015 iflib_queues_alloc(if_ctx_t ctx)
6016 {
6017 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6018 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6019 	device_t dev = ctx->ifc_dev;
6020 	int nrxqsets = scctx->isc_nrxqsets;
6021 	int ntxqsets = scctx->isc_ntxqsets;
6022 	iflib_txq_t txq;
6023 	iflib_rxq_t rxq;
6024 	iflib_fl_t fl = NULL;
6025 	int i, j, cpu, err, txconf, rxconf;
6026 	iflib_dma_info_t ifdip;
6027 	uint32_t *rxqsizes = scctx->isc_rxqsizes;
6028 	uint32_t *txqsizes = scctx->isc_txqsizes;
6029 	uint8_t nrxqs = sctx->isc_nrxqs;
6030 	uint8_t ntxqs = sctx->isc_ntxqs;
6031 	int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1;
6032 	int fl_offset = (sctx->isc_flags & IFLIB_HAS_RXCQ ? 1 : 0);
6033 	caddr_t *vaddrs;
6034 	uint64_t *paddrs;
6035 
6036 	KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1"));
6037 	KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1"));
6038 	KASSERT(nrxqs >= fl_offset + nfree_lists,
6039            ("there must be at least a rxq for each free list"));
6040 
6041 	/* Allocate the TX ring struct memory */
6042 	if (!(ctx->ifc_txqs =
6043 	    (iflib_txq_t) malloc(sizeof(struct iflib_txq) *
6044 	    ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
6045 		device_printf(dev, "Unable to allocate TX ring memory\n");
6046 		err = ENOMEM;
6047 		goto fail;
6048 	}
6049 
6050 	/* Now allocate the RX */
6051 	if (!(ctx->ifc_rxqs =
6052 	    (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) *
6053 	    nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
6054 		device_printf(dev, "Unable to allocate RX ring memory\n");
6055 		err = ENOMEM;
6056 		goto rx_fail;
6057 	}
6058 
6059 	txq = ctx->ifc_txqs;
6060 	rxq = ctx->ifc_rxqs;
6061 
6062 	/*
6063 	 * XXX handle allocation failure
6064 	 */
6065 	for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) {
6066 		/* Set up some basics */
6067 
6068 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs,
6069 		    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
6070 			device_printf(dev,
6071 			    "Unable to allocate TX DMA info memory\n");
6072 			err = ENOMEM;
6073 			goto err_tx_desc;
6074 		}
6075 		txq->ift_ifdi = ifdip;
6076 		for (j = 0; j < ntxqs; j++, ifdip++) {
6077 			if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, 0)) {
6078 				device_printf(dev,
6079 				    "Unable to allocate TX descriptors\n");
6080 				err = ENOMEM;
6081 				goto err_tx_desc;
6082 			}
6083 			txq->ift_txd_size[j] = scctx->isc_txd_size[j];
6084 			bzero((void *)ifdip->idi_vaddr, txqsizes[j]);
6085 		}
6086 		txq->ift_ctx = ctx;
6087 		txq->ift_id = i;
6088 		if (sctx->isc_flags & IFLIB_HAS_TXCQ) {
6089 			txq->ift_br_offset = 1;
6090 		} else {
6091 			txq->ift_br_offset = 0;
6092 		}
6093 
6094 		if (iflib_txsd_alloc(txq)) {
6095 			device_printf(dev, "Critical Failure setting up TX buffers\n");
6096 			err = ENOMEM;
6097 			goto err_tx_desc;
6098 		}
6099 
6100 		/* Initialize the TX lock */
6101 		snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:TX(%d):callout",
6102 		    device_get_nameunit(dev), txq->ift_id);
6103 		mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF);
6104 		callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0);
6105 #ifndef __HAIKU__
6106 		txq->ift_timer.c_cpu = cpu;
6107 #endif
6108 #ifdef DEV_NETMAP
6109 		callout_init_mtx(&txq->ift_netmap_timer, &txq->ift_mtx, 0);
6110 		txq->ift_netmap_timer.c_cpu = cpu;
6111 #endif /* DEV_NETMAP */
6112 
6113 		err = ifmp_ring_alloc(&txq->ift_br, 2048, txq, iflib_txq_drain,
6114 				      iflib_txq_can_drain, M_IFLIB, M_WAITOK);
6115 		if (err) {
6116 			/* XXX free any allocated rings */
6117 			device_printf(dev, "Unable to allocate buf_ring\n");
6118 			goto err_tx_desc;
6119 		}
6120 	}
6121 
6122 	for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) {
6123 		/* Set up some basics */
6124 		callout_init(&rxq->ifr_watchdog, 1);
6125 
6126 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs,
6127 		   M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
6128 			device_printf(dev,
6129 			    "Unable to allocate RX DMA info memory\n");
6130 			err = ENOMEM;
6131 			goto err_tx_desc;
6132 		}
6133 
6134 		rxq->ifr_ifdi = ifdip;
6135 		/* XXX this needs to be changed if #rx queues != #tx queues */
6136 		rxq->ifr_ntxqirq = 1;
6137 		rxq->ifr_txqid[0] = i;
6138 		for (j = 0; j < nrxqs; j++, ifdip++) {
6139 			if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, 0)) {
6140 				device_printf(dev,
6141 				    "Unable to allocate RX descriptors\n");
6142 				err = ENOMEM;
6143 				goto err_tx_desc;
6144 			}
6145 			bzero((void *)ifdip->idi_vaddr, rxqsizes[j]);
6146 		}
6147 		rxq->ifr_ctx = ctx;
6148 		rxq->ifr_id = i;
6149 		rxq->ifr_fl_offset = fl_offset;
6150 		rxq->ifr_nfl = nfree_lists;
6151 		if (!(fl =
6152 			  (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) {
6153 			device_printf(dev, "Unable to allocate free list memory\n");
6154 			err = ENOMEM;
6155 			goto err_tx_desc;
6156 		}
6157 		rxq->ifr_fl = fl;
6158 		for (j = 0; j < nfree_lists; j++) {
6159 			fl[j].ifl_rxq = rxq;
6160 			fl[j].ifl_id = j;
6161 			fl[j].ifl_ifdi = &rxq->ifr_ifdi[j + rxq->ifr_fl_offset];
6162 			fl[j].ifl_rxd_size = scctx->isc_rxd_size[j];
6163 		}
6164 		/* Allocate receive buffers for the ring */
6165 		if (iflib_rxsd_alloc(rxq)) {
6166 			device_printf(dev,
6167 			    "Critical Failure setting up receive buffers\n");
6168 			err = ENOMEM;
6169 			goto err_rx_desc;
6170 		}
6171 
6172 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
6173 			fl->ifl_rx_bitmap = bit_alloc(fl->ifl_size, M_IFLIB,
6174 			    M_WAITOK);
6175 	}
6176 
6177 	/* TXQs */
6178 	vaddrs = malloc(sizeof(caddr_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
6179 	paddrs = malloc(sizeof(uint64_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
6180 	for (i = 0; i < ntxqsets; i++) {
6181 		iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi;
6182 
6183 		for (j = 0; j < ntxqs; j++, di++) {
6184 			vaddrs[i*ntxqs + j] = di->idi_vaddr;
6185 			paddrs[i*ntxqs + j] = di->idi_paddr;
6186 		}
6187 	}
6188 	if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) {
6189 		device_printf(ctx->ifc_dev,
6190 		    "Unable to allocate device TX queue\n");
6191 		iflib_tx_structures_free(ctx);
6192 		free(vaddrs, M_IFLIB);
6193 		free(paddrs, M_IFLIB);
6194 		goto err_rx_desc;
6195 	}
6196 	free(vaddrs, M_IFLIB);
6197 	free(paddrs, M_IFLIB);
6198 
6199 	/* RXQs */
6200 	vaddrs = malloc(sizeof(caddr_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
6201 	paddrs = malloc(sizeof(uint64_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
6202 	for (i = 0; i < nrxqsets; i++) {
6203 		iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi;
6204 
6205 		for (j = 0; j < nrxqs; j++, di++) {
6206 			vaddrs[i*nrxqs + j] = di->idi_vaddr;
6207 			paddrs[i*nrxqs + j] = di->idi_paddr;
6208 		}
6209 	}
6210 	if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) {
6211 		device_printf(ctx->ifc_dev,
6212 		    "Unable to allocate device RX queue\n");
6213 		iflib_tx_structures_free(ctx);
6214 		free(vaddrs, M_IFLIB);
6215 		free(paddrs, M_IFLIB);
6216 		goto err_rx_desc;
6217 	}
6218 	free(vaddrs, M_IFLIB);
6219 	free(paddrs, M_IFLIB);
6220 
6221 	return (0);
6222 
6223 /* XXX handle allocation failure changes */
6224 err_rx_desc:
6225 err_tx_desc:
6226 rx_fail:
6227 	if (ctx->ifc_rxqs != NULL)
6228 		free(ctx->ifc_rxqs, M_IFLIB);
6229 	ctx->ifc_rxqs = NULL;
6230 	if (ctx->ifc_txqs != NULL)
6231 		free(ctx->ifc_txqs, M_IFLIB);
6232 	ctx->ifc_txqs = NULL;
6233 fail:
6234 	return (err);
6235 }
6236 
6237 static int
6238 iflib_tx_structures_setup(if_ctx_t ctx)
6239 {
6240 	iflib_txq_t txq = ctx->ifc_txqs;
6241 	int i;
6242 
6243 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
6244 		iflib_txq_setup(txq);
6245 
6246 	return (0);
6247 }
6248 
6249 static void
6250 iflib_tx_structures_free(if_ctx_t ctx)
6251 {
6252 	iflib_txq_t txq = ctx->ifc_txqs;
6253 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6254 	int i, j;
6255 
6256 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
6257 		for (j = 0; j < sctx->isc_ntxqs; j++)
6258 			iflib_dma_free(&txq->ift_ifdi[j]);
6259 		iflib_txq_destroy(txq);
6260 	}
6261 	free(ctx->ifc_txqs, M_IFLIB);
6262 	ctx->ifc_txqs = NULL;
6263 }
6264 
6265 /*********************************************************************
6266  *
6267  *  Initialize all receive rings.
6268  *
6269  **********************************************************************/
6270 static int
6271 iflib_rx_structures_setup(if_ctx_t ctx)
6272 {
6273 	iflib_rxq_t rxq = ctx->ifc_rxqs;
6274 	int q;
6275 #if defined(INET6) || defined(INET)
6276 	int err, i;
6277 #endif
6278 
6279 	for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) {
6280 #if defined(INET6) || defined(INET)
6281 		if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_LRO) {
6282 			err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp,
6283 			    TCP_LRO_ENTRIES, min(1024,
6284 			    ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset]));
6285 			if (err != 0) {
6286 				device_printf(ctx->ifc_dev,
6287 				    "LRO Initialization failed!\n");
6288 				goto fail;
6289 			}
6290 		}
6291 #endif
6292 		IFDI_RXQ_SETUP(ctx, rxq->ifr_id);
6293 	}
6294 	return (0);
6295 #if defined(INET6) || defined(INET)
6296 fail:
6297 	/*
6298 	 * Free LRO resources allocated so far, we will only handle
6299 	 * the rings that completed, the failing case will have
6300 	 * cleaned up for itself.  'q' failed, so its the terminus.
6301 	 */
6302 	rxq = ctx->ifc_rxqs;
6303 	for (i = 0; i < q; ++i, rxq++) {
6304 		if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_LRO)
6305 			tcp_lro_free(&rxq->ifr_lc);
6306 	}
6307 	return (err);
6308 #endif
6309 }
6310 
6311 /*********************************************************************
6312  *
6313  *  Free all receive rings.
6314  *
6315  **********************************************************************/
6316 static void
6317 iflib_rx_structures_free(if_ctx_t ctx)
6318 {
6319 	iflib_rxq_t rxq = ctx->ifc_rxqs;
6320 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6321 	int i, j;
6322 
6323 	for (i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) {
6324 		for (j = 0; j < sctx->isc_nrxqs; j++)
6325 			iflib_dma_free(&rxq->ifr_ifdi[j]);
6326 		iflib_rx_sds_free(rxq);
6327 #if defined(INET6) || defined(INET)
6328 		if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_LRO)
6329 			tcp_lro_free(&rxq->ifr_lc);
6330 #endif
6331 	}
6332 	free(ctx->ifc_rxqs, M_IFLIB);
6333 	ctx->ifc_rxqs = NULL;
6334 }
6335 
6336 static int
6337 iflib_qset_structures_setup(if_ctx_t ctx)
6338 {
6339 	int err;
6340 
6341 	/*
6342 	 * It is expected that the caller takes care of freeing queues if this
6343 	 * fails.
6344 	 */
6345 	if ((err = iflib_tx_structures_setup(ctx)) != 0) {
6346 		device_printf(ctx->ifc_dev, "iflib_tx_structures_setup failed: %d\n", err);
6347 		return (err);
6348 	}
6349 
6350 	if ((err = iflib_rx_structures_setup(ctx)) != 0)
6351 		device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err);
6352 
6353 	return (err);
6354 }
6355 
6356 int
6357 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
6358 		driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, const char *name)
6359 {
6360 
6361 	return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name));
6362 }
6363 
6364 /* Just to avoid copy/paste */
6365 static inline int
6366 iflib_irq_set_affinity(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type,
6367     int qid, struct grouptask *gtask, struct taskqgroup *tqg, void *uniq,
6368     const char *name)
6369 {
6370 	device_t dev;
6371 	unsigned int base_cpuid, cpuid;
6372 	int err;
6373 
6374 	dev = ctx->ifc_dev;
6375 	base_cpuid = ctx->ifc_sysctl_core_offset;
6376 	cpuid = get_cpuid_for_queue(ctx, base_cpuid, qid, type == IFLIB_INTR_TX);
6377 	err = taskqgroup_attach_cpu(tqg, gtask, uniq, cpuid, dev,
6378 	    irq ? irq->ii_res : NULL, name);
6379 	if (err) {
6380 		device_printf(dev, "taskqgroup_attach_cpu failed %d\n", err);
6381 		return (err);
6382 	}
6383 #ifdef notyet
6384 	if (cpuid > ctx->ifc_cpuid_highest)
6385 		ctx->ifc_cpuid_highest = cpuid;
6386 #endif
6387 	return (0);
6388 }
6389 
6390 int
6391 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid,
6392 			iflib_intr_type_t type, driver_filter_t *filter,
6393 			void *filter_arg, int qid, const char *name)
6394 {
6395 	device_t dev;
6396 	struct grouptask *gtask;
6397 	struct taskqgroup *tqg;
6398 	iflib_filter_info_t info;
6399 	gtask_fn_t *fn;
6400 	int tqrid, err;
6401 	driver_filter_t *intr_fast;
6402 	void *q;
6403 
6404 	info = &ctx->ifc_filter_info;
6405 	tqrid = rid;
6406 
6407 	switch (type) {
6408 	/* XXX merge tx/rx for netmap? */
6409 	case IFLIB_INTR_TX:
6410 		q = &ctx->ifc_txqs[qid];
6411 		info = &ctx->ifc_txqs[qid].ift_filter_info;
6412 		gtask = &ctx->ifc_txqs[qid].ift_task;
6413 		tqg = qgroup_if_io_tqg;
6414 		fn = _task_fn_tx;
6415 		intr_fast = iflib_fast_intr;
6416 		GROUPTASK_INIT(gtask, 0, fn, q);
6417 		ctx->ifc_flags |= IFC_NETMAP_TX_IRQ;
6418 		break;
6419 	case IFLIB_INTR_RX:
6420 		q = &ctx->ifc_rxqs[qid];
6421 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
6422 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6423 		tqg = qgroup_if_io_tqg;
6424 		fn = _task_fn_rx;
6425 		intr_fast = iflib_fast_intr;
6426 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6427 		break;
6428 	case IFLIB_INTR_RXTX:
6429 		q = &ctx->ifc_rxqs[qid];
6430 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
6431 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6432 		tqg = qgroup_if_io_tqg;
6433 		fn = _task_fn_rx;
6434 		intr_fast = iflib_fast_intr_rxtx;
6435 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6436 		break;
6437 	case IFLIB_INTR_ADMIN:
6438 		q = ctx;
6439 		tqrid = -1;
6440 		info = &ctx->ifc_filter_info;
6441 		gtask = &ctx->ifc_admin_task;
6442 		tqg = qgroup_if_config_tqg;
6443 		fn = _task_fn_admin;
6444 		intr_fast = iflib_fast_intr_ctx;
6445 		break;
6446 	default:
6447 		device_printf(ctx->ifc_dev, "%s: unknown net intr type\n",
6448 		    __func__);
6449 		return (EINVAL);
6450 	}
6451 
6452 	info->ifi_filter = filter;
6453 	info->ifi_filter_arg = filter_arg;
6454 	info->ifi_task = gtask;
6455 	info->ifi_ctx = q;
6456 
6457 	dev = ctx->ifc_dev;
6458 	err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info,  name);
6459 	if (err != 0) {
6460 		device_printf(dev, "_iflib_irq_alloc failed %d\n", err);
6461 		return (err);
6462 	}
6463 	if (type == IFLIB_INTR_ADMIN)
6464 		return (0);
6465 
6466 	if (tqrid != -1) {
6467 		err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg, q,
6468 		    name);
6469 		if (err)
6470 			return (err);
6471 	} else {
6472 		taskqgroup_attach(tqg, gtask, q, dev, irq->ii_res, name);
6473 	}
6474 
6475 	return (0);
6476 }
6477 
6478 void
6479 iflib_softirq_alloc_generic(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type, void *arg, int qid, const char *name)
6480 {
6481 	device_t dev;
6482 	struct grouptask *gtask;
6483 	struct taskqgroup *tqg;
6484 	gtask_fn_t *fn;
6485 	void *q;
6486 	int err;
6487 
6488 	switch (type) {
6489 	case IFLIB_INTR_TX:
6490 		q = &ctx->ifc_txqs[qid];
6491 		gtask = &ctx->ifc_txqs[qid].ift_task;
6492 		tqg = qgroup_if_io_tqg;
6493 		fn = _task_fn_tx;
6494 		GROUPTASK_INIT(gtask, 0, fn, q);
6495 		break;
6496 	case IFLIB_INTR_RX:
6497 		q = &ctx->ifc_rxqs[qid];
6498 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6499 		tqg = qgroup_if_io_tqg;
6500 		fn = _task_fn_rx;
6501 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6502 		break;
6503 	case IFLIB_INTR_IOV:
6504 		q = ctx;
6505 		gtask = &ctx->ifc_vflr_task;
6506 		tqg = qgroup_if_config_tqg;
6507 		fn = _task_fn_iov;
6508 		GROUPTASK_INIT(gtask, 0, fn, q);
6509 		break;
6510 	default:
6511 		panic("unknown net intr type");
6512 	}
6513 	err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg, q, name);
6514 	if (err) {
6515 		dev = ctx->ifc_dev;
6516 		taskqgroup_attach(tqg, gtask, q, dev, irq ? irq->ii_res : NULL,
6517 		    name);
6518 	}
6519 }
6520 
6521 void
6522 iflib_irq_free(if_ctx_t ctx, if_irq_t irq)
6523 {
6524 #ifdef __HAIKU__
6525 	if (!ctx || !irq)
6526 		return;
6527 #endif
6528 
6529 	if (irq->ii_tag)
6530 		bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag);
6531 
6532 	if (irq->ii_res)
6533 		bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ,
6534 		    rman_get_rid(irq->ii_res), irq->ii_res);
6535 }
6536 
6537 static int
6538 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, const char *name)
6539 {
6540 	iflib_txq_t txq = ctx->ifc_txqs;
6541 	iflib_rxq_t rxq = ctx->ifc_rxqs;
6542 	if_irq_t irq = &ctx->ifc_legacy_irq;
6543 	iflib_filter_info_t info;
6544 	device_t dev;
6545 	struct grouptask *gtask;
6546 	struct resource *res;
6547 	struct taskqgroup *tqg;
6548 	void *q;
6549 	int err, tqrid;
6550 	bool rx_only;
6551 
6552 	q = &ctx->ifc_rxqs[0];
6553 	info = &rxq[0].ifr_filter_info;
6554 	gtask = &rxq[0].ifr_task;
6555 	tqg = qgroup_if_io_tqg;
6556 	tqrid = *rid;
6557 	rx_only = (ctx->ifc_sctx->isc_flags & IFLIB_SINGLE_IRQ_RX_ONLY) != 0;
6558 
6559 	ctx->ifc_flags |= IFC_LEGACY;
6560 	info->ifi_filter = filter;
6561 	info->ifi_filter_arg = filter_arg;
6562 	info->ifi_task = gtask;
6563 	info->ifi_ctx = rx_only ? ctx : q;
6564 
6565 	dev = ctx->ifc_dev;
6566 	/* We allocate a single interrupt resource */
6567 	err = _iflib_irq_alloc(ctx, irq, tqrid, rx_only ? iflib_fast_intr_ctx :
6568 	    iflib_fast_intr_rxtx, NULL, info, name);
6569 	if (err != 0)
6570 		return (err);
6571 	NET_GROUPTASK_INIT(gtask, 0, _task_fn_rx, q);
6572 	res = irq->ii_res;
6573 	taskqgroup_attach(tqg, gtask, q, dev, res, name);
6574 
6575 	GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq);
6576 	taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, dev, res,
6577 	    "tx");
6578 	return (0);
6579 }
6580 
6581 void
6582 iflib_led_create(if_ctx_t ctx)
6583 {
6584 
6585 	ctx->ifc_led_dev = led_create(iflib_led_func, ctx,
6586 	    device_get_nameunit(ctx->ifc_dev));
6587 }
6588 
6589 void
6590 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid)
6591 {
6592 
6593 	GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
6594 }
6595 
6596 void
6597 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid)
6598 {
6599 
6600 	GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task);
6601 }
6602 
6603 void
6604 iflib_admin_intr_deferred(if_ctx_t ctx)
6605 {
6606 
6607 	MPASS(ctx->ifc_admin_task.gt_taskqueue != NULL);
6608 	GROUPTASK_ENQUEUE(&ctx->ifc_admin_task);
6609 }
6610 
6611 void
6612 iflib_iov_intr_deferred(if_ctx_t ctx)
6613 {
6614 
6615 	GROUPTASK_ENQUEUE(&ctx->ifc_vflr_task);
6616 }
6617 
6618 void
6619 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, const char *name)
6620 {
6621 
6622 	taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, NULL, NULL,
6623 	    name);
6624 }
6625 
6626 void
6627 iflib_config_gtask_init(void *ctx, struct grouptask *gtask, gtask_fn_t *fn,
6628 	const char *name)
6629 {
6630 
6631 	GROUPTASK_INIT(gtask, 0, fn, ctx);
6632 	taskqgroup_attach(qgroup_if_config_tqg, gtask, gtask, NULL, NULL,
6633 	    name);
6634 }
6635 
6636 void
6637 iflib_config_gtask_deinit(struct grouptask *gtask)
6638 {
6639 
6640 	taskqgroup_detach(qgroup_if_config_tqg, gtask);
6641 }
6642 
6643 void
6644 iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate)
6645 {
6646 	if_t ifp = ctx->ifc_ifp;
6647 	iflib_txq_t txq = ctx->ifc_txqs;
6648 
6649 	if_setbaudrate(ifp, baudrate);
6650 	if (baudrate >= IF_Gbps(10)) {
6651 		STATE_LOCK(ctx);
6652 		ctx->ifc_flags |= IFC_PREFETCH;
6653 		STATE_UNLOCK(ctx);
6654 	}
6655 	/* If link down, disable watchdog */
6656 	if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) {
6657 		for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++)
6658 			txq->ift_qstatus = IFLIB_QUEUE_IDLE;
6659 	}
6660 	ctx->ifc_link_state = link_state;
6661 	if_link_state_change(ifp, link_state);
6662 }
6663 
6664 static int
6665 iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq)
6666 {
6667 	int credits;
6668 #ifdef INVARIANTS
6669 	int credits_pre = txq->ift_cidx_processed;
6670 #endif
6671 
6672 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
6673 	    BUS_DMASYNC_POSTREAD);
6674 	if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, true)) == 0)
6675 		return (0);
6676 
6677 	txq->ift_processed += credits;
6678 	txq->ift_cidx_processed += credits;
6679 
6680 	MPASS(credits_pre + credits == txq->ift_cidx_processed);
6681 	if (txq->ift_cidx_processed >= txq->ift_size)
6682 		txq->ift_cidx_processed -= txq->ift_size;
6683 	return (credits);
6684 }
6685 
6686 static int
6687 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget)
6688 {
6689 	iflib_fl_t fl;
6690 	u_int i;
6691 
6692 	for (i = 0, fl = &rxq->ifr_fl[0]; i < rxq->ifr_nfl; i++, fl++)
6693 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
6694 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
6695 	return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx,
6696 	    budget));
6697 }
6698 
6699 void
6700 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name,
6701 	const char *description, if_int_delay_info_t info,
6702 	int offset, int value)
6703 {
6704 	info->iidi_ctx = ctx;
6705 	info->iidi_offset = offset;
6706 	info->iidi_value = value;
6707 	SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev),
6708 	    SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)),
6709 	    OID_AUTO, name, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE,
6710 	    info, 0, iflib_sysctl_int_delay, "I", description);
6711 }
6712 
6713 struct sx *
6714 iflib_ctx_lock_get(if_ctx_t ctx)
6715 {
6716 
6717 	return (&ctx->ifc_ctx_sx);
6718 }
6719 
6720 static int
6721 iflib_msix_init(if_ctx_t ctx)
6722 {
6723 	device_t dev = ctx->ifc_dev;
6724 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6725 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6726 	int admincnt, bar, err, iflib_num_rx_queues, iflib_num_tx_queues;
6727 	int msgs, queuemsgs, queues, rx_queues, tx_queues, vectors;
6728 
6729 	iflib_num_tx_queues = ctx->ifc_sysctl_ntxqs;
6730 	iflib_num_rx_queues = ctx->ifc_sysctl_nrxqs;
6731 
6732 	if (bootverbose)
6733 		device_printf(dev, "msix_init qsets capped at %d\n",
6734 		    imax(scctx->isc_ntxqsets, scctx->isc_nrxqsets));
6735 
6736 	/* Override by tuneable */
6737 	if (scctx->isc_disable_msix)
6738 		goto msi;
6739 
6740 	/* First try MSI-X */
6741 	if ((msgs = pci_msix_count(dev)) == 0) {
6742 		if (bootverbose)
6743 			device_printf(dev, "MSI-X not supported or disabled\n");
6744 		goto msi;
6745 	}
6746 
6747 	bar = ctx->ifc_softc_ctx.isc_msix_bar;
6748 	/*
6749 	 * bar == -1 => "trust me I know what I'm doing"
6750 	 * Some drivers are for hardware that is so shoddily
6751 	 * documented that no one knows which bars are which
6752 	 * so the developer has to map all bars. This hack
6753 	 * allows shoddy garbage to use MSI-X in this framework.
6754 	 */
6755 	if (bar != -1) {
6756 		ctx->ifc_msix_mem = bus_alloc_resource_any(dev,
6757 	            SYS_RES_MEMORY, &bar, RF_ACTIVE);
6758 		if (ctx->ifc_msix_mem == NULL) {
6759 			device_printf(dev, "Unable to map MSI-X table\n");
6760 			goto msi;
6761 		}
6762 	}
6763 
6764 	admincnt = sctx->isc_admin_intrcnt;
6765 #if IFLIB_DEBUG
6766 	/* use only 1 qset in debug mode */
6767 	queuemsgs = min(msgs - admincnt, 1);
6768 #else
6769 	queuemsgs = msgs - admincnt;
6770 #endif
6771 #ifdef RSS
6772 	queues = imin(queuemsgs, rss_getnumbuckets());
6773 #else
6774 	queues = queuemsgs;
6775 #endif
6776 #ifndef __HAIKU__
6777 	queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues);
6778 	if (bootverbose)
6779 		device_printf(dev,
6780 		    "intr CPUs: %d queue msgs: %d admincnt: %d\n",
6781 		    CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt);
6782 #endif
6783 #ifdef  RSS
6784 	/* If we're doing RSS, clamp at the number of RSS buckets */
6785 	if (queues > rss_getnumbuckets())
6786 		queues = rss_getnumbuckets();
6787 #endif
6788 	if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt)
6789 		rx_queues = iflib_num_rx_queues;
6790 	else
6791 		rx_queues = queues;
6792 
6793 	if (rx_queues > scctx->isc_nrxqsets)
6794 		rx_queues = scctx->isc_nrxqsets;
6795 
6796 	/*
6797 	 * We want this to be all logical CPUs by default
6798 	 */
6799 	if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues)
6800 		tx_queues = iflib_num_tx_queues;
6801 	else
6802 		tx_queues = mp_ncpus;
6803 
6804 	if (tx_queues > scctx->isc_ntxqsets)
6805 		tx_queues = scctx->isc_ntxqsets;
6806 
6807 	if (ctx->ifc_sysctl_qs_eq_override == 0) {
6808 #ifdef INVARIANTS
6809 		if (tx_queues != rx_queues)
6810 			device_printf(dev,
6811 			    "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n",
6812 			    min(rx_queues, tx_queues), min(rx_queues, tx_queues));
6813 #endif
6814 		tx_queues = min(rx_queues, tx_queues);
6815 		rx_queues = min(rx_queues, tx_queues);
6816 	}
6817 
6818 	vectors = rx_queues + admincnt;
6819 	if (msgs < vectors) {
6820 		device_printf(dev,
6821 		    "insufficient number of MSI-X vectors "
6822 		    "(supported %d, need %d)\n", msgs, vectors);
6823 		goto msi;
6824 	}
6825 
6826 	device_printf(dev, "Using %d RX queues %d TX queues\n", rx_queues,
6827 	    tx_queues);
6828 	msgs = vectors;
6829 	if ((err = pci_alloc_msix(dev, &vectors)) == 0) {
6830 		if (vectors != msgs) {
6831 			device_printf(dev,
6832 			    "Unable to allocate sufficient MSI-X vectors "
6833 			    "(got %d, need %d)\n", vectors, msgs);
6834 			pci_release_msi(dev);
6835 			if (bar != -1) {
6836 				bus_release_resource(dev, SYS_RES_MEMORY, bar,
6837 				    ctx->ifc_msix_mem);
6838 				ctx->ifc_msix_mem = NULL;
6839 			}
6840 			goto msi;
6841 		}
6842 		device_printf(dev, "Using MSI-X interrupts with %d vectors\n",
6843 		    vectors);
6844 		scctx->isc_vectors = vectors;
6845 		scctx->isc_nrxqsets = rx_queues;
6846 		scctx->isc_ntxqsets = tx_queues;
6847 		scctx->isc_intr = IFLIB_INTR_MSIX;
6848 
6849 		return (vectors);
6850 	} else {
6851 		device_printf(dev,
6852 		    "failed to allocate %d MSI-X vectors, err: %d\n", vectors,
6853 		    err);
6854 		if (bar != -1) {
6855 			bus_release_resource(dev, SYS_RES_MEMORY, bar,
6856 			    ctx->ifc_msix_mem);
6857 			ctx->ifc_msix_mem = NULL;
6858 		}
6859 	}
6860 
6861 msi:
6862 	vectors = pci_msi_count(dev);
6863 	scctx->isc_nrxqsets = 1;
6864 	scctx->isc_ntxqsets = 1;
6865 	scctx->isc_vectors = vectors;
6866 	if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) {
6867 		device_printf(dev,"Using an MSI interrupt\n");
6868 		scctx->isc_intr = IFLIB_INTR_MSI;
6869 	} else {
6870 		scctx->isc_vectors = 1;
6871 		device_printf(dev,"Using a Legacy interrupt\n");
6872 		scctx->isc_intr = IFLIB_INTR_LEGACY;
6873 	}
6874 
6875 	return (vectors);
6876 }
6877 
6878 static const char *ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" };
6879 
6880 #ifndef __HAIKU__
6881 static int
6882 mp_ring_state_handler(SYSCTL_HANDLER_ARGS)
6883 {
6884 	int rc;
6885 	uint16_t *state = ((uint16_t *)oidp->oid_arg1);
6886 	struct sbuf *sb;
6887 	const char *ring_state = "UNKNOWN";
6888 
6889 	/* XXX needed ? */
6890 	rc = sysctl_wire_old_buffer(req, 0);
6891 	MPASS(rc == 0);
6892 	if (rc != 0)
6893 		return (rc);
6894 	sb = sbuf_new_for_sysctl(NULL, NULL, 80, req);
6895 	MPASS(sb != NULL);
6896 	if (sb == NULL)
6897 		return (ENOMEM);
6898 	if (state[3] <= 3)
6899 		ring_state = ring_states[state[3]];
6900 
6901 	sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s",
6902 		    state[0], state[1], state[2], ring_state);
6903 	rc = sbuf_finish(sb);
6904 	sbuf_delete(sb);
6905         return(rc);
6906 }
6907 #endif
6908 
6909 enum iflib_ndesc_handler {
6910 	IFLIB_NTXD_HANDLER,
6911 	IFLIB_NRXD_HANDLER,
6912 };
6913 
6914 static int
6915 mp_ndesc_handler(SYSCTL_HANDLER_ARGS)
6916 {
6917 	if_ctx_t ctx = (void *)arg1;
6918 	enum iflib_ndesc_handler type = arg2;
6919 	char buf[256] = {0};
6920 	qidx_t *ndesc;
6921 	char *p, *next;
6922 	int nqs, rc, i;
6923 
6924 	nqs = 8;
6925 	switch(type) {
6926 	case IFLIB_NTXD_HANDLER:
6927 		ndesc = ctx->ifc_sysctl_ntxds;
6928 		if (ctx->ifc_sctx)
6929 			nqs = ctx->ifc_sctx->isc_ntxqs;
6930 		break;
6931 	case IFLIB_NRXD_HANDLER:
6932 		ndesc = ctx->ifc_sysctl_nrxds;
6933 		if (ctx->ifc_sctx)
6934 			nqs = ctx->ifc_sctx->isc_nrxqs;
6935 		break;
6936 	default:
6937 		printf("%s: unhandled type\n", __func__);
6938 		return (EINVAL);
6939 	}
6940 	if (nqs == 0)
6941 		nqs = 8;
6942 
6943 	for (i=0; i<8; i++) {
6944 		if (i >= nqs)
6945 			break;
6946 		if (i)
6947 			strcat(buf, ",");
6948 		sprintf(strchr(buf, 0), "%d", ndesc[i]);
6949 	}
6950 
6951 	rc = sysctl_handle_string(oidp, buf, sizeof(buf), req);
6952 	if (rc || req->newptr == NULL)
6953 		return rc;
6954 
6955 	for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p;
6956 	    i++, p = strsep(&next, " ,")) {
6957 		ndesc[i] = strtoul(p, NULL, 10);
6958 	}
6959 
6960 	return(rc);
6961 }
6962 
6963 #define NAME_BUFLEN 32
6964 static void
6965 iflib_add_device_sysctl_pre(if_ctx_t ctx)
6966 {
6967 #ifndef __HAIKU__
6968         device_t dev = iflib_get_dev(ctx);
6969 	struct sysctl_oid_list *child, *oid_list;
6970 	struct sysctl_ctx_list *ctx_list;
6971 	struct sysctl_oid *node;
6972 
6973 	ctx_list = device_get_sysctl_ctx(dev);
6974 	child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev));
6975 	ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, "iflib",
6976 	    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "IFLIB fields");
6977 	oid_list = SYSCTL_CHILDREN(node);
6978 
6979 	SYSCTL_ADD_CONST_STRING(ctx_list, oid_list, OID_AUTO, "driver_version",
6980 		       CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version,
6981 		       "driver version");
6982 
6983 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs",
6984 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0,
6985 			"# of txqs to use, 0 => use default #");
6986 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs",
6987 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0,
6988 			"# of rxqs to use, 0 => use default #");
6989 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_qs_enable",
6990 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0,
6991                        "permit #txq != #rxq");
6992 	SYSCTL_ADD_INT(ctx_list, oid_list, OID_AUTO, "disable_msix",
6993                       CTLFLAG_RWTUN, &ctx->ifc_softc_ctx.isc_disable_msix, 0,
6994                       "disable MSI-X (default 0)");
6995 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "rx_budget",
6996 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_rx_budget, 0,
6997 		       "set the RX budget");
6998 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "tx_abdicate",
6999 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_tx_abdicate, 0,
7000 		       "cause TX to abdicate instead of running to completion");
7001 	ctx->ifc_sysctl_core_offset = CORE_OFFSET_UNSPECIFIED;
7002 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "core_offset",
7003 		       CTLFLAG_RDTUN, &ctx->ifc_sysctl_core_offset, 0,
7004 		       "offset to start using cores at");
7005 	SYSCTL_ADD_U8(ctx_list, oid_list, OID_AUTO, "separate_txrx",
7006 		       CTLFLAG_RDTUN, &ctx->ifc_sysctl_separate_txrx, 0,
7007 		       "use separate cores for TX and RX");
7008 	SYSCTL_ADD_U8(ctx_list, oid_list, OID_AUTO, "use_logical_cores",
7009 		      CTLFLAG_RDTUN, &ctx->ifc_sysctl_use_logical_cores, 0,
7010 		      "try to make use of logical cores for TX and RX");
7011 
7012 	/* XXX change for per-queue sizes */
7013 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_ntxds",
7014 	    CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NEEDGIANT, ctx,
7015 	    IFLIB_NTXD_HANDLER, mp_ndesc_handler, "A",
7016 	    "list of # of TX descriptors to use, 0 = use default #");
7017 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_nrxds",
7018 	    CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NEEDGIANT, ctx,
7019 	    IFLIB_NRXD_HANDLER, mp_ndesc_handler, "A",
7020 	    "list of # of RX descriptors to use, 0 = use default #");
7021 #endif
7022 }
7023 
7024 static void
7025 iflib_add_device_sysctl_post(if_ctx_t ctx)
7026 {
7027 #ifndef __HAIKU__
7028 	if_shared_ctx_t sctx = ctx->ifc_sctx;
7029 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
7030         device_t dev = iflib_get_dev(ctx);
7031 	struct sysctl_oid_list *child;
7032 	struct sysctl_ctx_list *ctx_list;
7033 	iflib_fl_t fl;
7034 	iflib_txq_t txq;
7035 	iflib_rxq_t rxq;
7036 	int i, j;
7037 	char namebuf[NAME_BUFLEN];
7038 	char *qfmt;
7039 	struct sysctl_oid *queue_node, *fl_node, *node;
7040 	struct sysctl_oid_list *queue_list, *fl_list;
7041 	ctx_list = device_get_sysctl_ctx(dev);
7042 
7043 	node = ctx->ifc_sysctl_node;
7044 	child = SYSCTL_CHILDREN(node);
7045 
7046 	if (scctx->isc_ntxqsets > 100)
7047 		qfmt = "txq%03d";
7048 	else if (scctx->isc_ntxqsets > 10)
7049 		qfmt = "txq%02d";
7050 	else
7051 		qfmt = "txq%d";
7052 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
7053 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
7054 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
7055 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "Queue Name");
7056 		queue_list = SYSCTL_CHILDREN(queue_node);
7057 		SYSCTL_ADD_INT(ctx_list, queue_list, OID_AUTO, "cpu",
7058 			       CTLFLAG_RD,
7059 			       &txq->ift_task.gt_cpu, 0, "cpu this queue is bound to");
7060 #if MEMORY_LOGGING
7061 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued",
7062 				CTLFLAG_RD,
7063 				&txq->ift_dequeued, "total mbufs freed");
7064 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued",
7065 				CTLFLAG_RD,
7066 				&txq->ift_enqueued, "total mbufs enqueued");
7067 #endif
7068 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag",
7069 				   CTLFLAG_RD,
7070 				   &txq->ift_mbuf_defrag, "# of times m_defrag was called");
7071 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "m_pullups",
7072 				   CTLFLAG_RD,
7073 				   &txq->ift_pullups, "# of times m_pullup was called");
7074 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag_failed",
7075 				   CTLFLAG_RD,
7076 				   &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed");
7077 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_desc_avail",
7078 				   CTLFLAG_RD,
7079 				   &txq->ift_no_desc_avail, "# of times no descriptors were available");
7080 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "tx_map_failed",
7081 				   CTLFLAG_RD,
7082 				   &txq->ift_map_failed, "# of times DMA map failed");
7083 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txd_encap_efbig",
7084 				   CTLFLAG_RD,
7085 				   &txq->ift_txd_encap_efbig, "# of times txd_encap returned EFBIG");
7086 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_tx_dma_setup",
7087 				   CTLFLAG_RD,
7088 				   &txq->ift_no_tx_dma_setup, "# of times map failed for other than EFBIG");
7089 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx",
7090 				   CTLFLAG_RD,
7091 				   &txq->ift_pidx, 1, "Producer Index");
7092 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx",
7093 				   CTLFLAG_RD,
7094 				   &txq->ift_cidx, 1, "Consumer Index");
7095 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx_processed",
7096 				   CTLFLAG_RD,
7097 				   &txq->ift_cidx_processed, 1, "Consumer Index seen by credit update");
7098 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use",
7099 				   CTLFLAG_RD,
7100 				   &txq->ift_in_use, 1, "descriptors in use");
7101 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_processed",
7102 				   CTLFLAG_RD,
7103 				   &txq->ift_processed, "descriptors procesed for clean");
7104 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned",
7105 				   CTLFLAG_RD,
7106 				   &txq->ift_cleaned, "total cleaned");
7107 		SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state",
7108 		    CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_NEEDGIANT,
7109 		    __DEVOLATILE(uint64_t *, &txq->ift_br->state), 0,
7110 		    mp_ring_state_handler, "A", "soft ring state");
7111 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_enqueues",
7112 				       CTLFLAG_RD, &txq->ift_br->enqueues,
7113 				       "# of enqueues to the mp_ring for this queue");
7114 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_drops",
7115 				       CTLFLAG_RD, &txq->ift_br->drops,
7116 				       "# of drops in the mp_ring for this queue");
7117 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_starts",
7118 				       CTLFLAG_RD, &txq->ift_br->starts,
7119 				       "# of normal consumer starts in the mp_ring for this queue");
7120 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_stalls",
7121 				       CTLFLAG_RD, &txq->ift_br->stalls,
7122 					       "# of consumer stalls in the mp_ring for this queue");
7123 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_restarts",
7124 			       CTLFLAG_RD, &txq->ift_br->restarts,
7125 				       "# of consumer restarts in the mp_ring for this queue");
7126 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_abdications",
7127 				       CTLFLAG_RD, &txq->ift_br->abdications,
7128 				       "# of consumer abdications in the mp_ring for this queue");
7129 	}
7130 
7131 	if (scctx->isc_nrxqsets > 100)
7132 		qfmt = "rxq%03d";
7133 	else if (scctx->isc_nrxqsets > 10)
7134 		qfmt = "rxq%02d";
7135 	else
7136 		qfmt = "rxq%d";
7137 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
7138 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
7139 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
7140 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "Queue Name");
7141 		queue_list = SYSCTL_CHILDREN(queue_node);
7142 		SYSCTL_ADD_INT(ctx_list, queue_list, OID_AUTO, "cpu",
7143 			       CTLFLAG_RD,
7144 			       &rxq->ifr_task.gt_cpu, 0, "cpu this queue is bound to");
7145 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
7146 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_cidx",
7147 				       CTLFLAG_RD,
7148 				       &rxq->ifr_cq_cidx, 1, "Consumer Index");
7149 		}
7150 
7151 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
7152 			snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j);
7153 			fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list, OID_AUTO, namebuf,
7154 			    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "freelist Name");
7155 			fl_list = SYSCTL_CHILDREN(fl_node);
7156 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx",
7157 				       CTLFLAG_RD,
7158 				       &fl->ifl_pidx, 1, "Producer Index");
7159 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx",
7160 				       CTLFLAG_RD,
7161 				       &fl->ifl_cidx, 1, "Consumer Index");
7162 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits",
7163 				       CTLFLAG_RD,
7164 				       &fl->ifl_credits, 1, "credits available");
7165 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "buf_size",
7166 				       CTLFLAG_RD,
7167 				       &fl->ifl_buf_size, 1, "buffer size");
7168 #if MEMORY_LOGGING
7169 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_enqueued",
7170 					CTLFLAG_RD,
7171 					&fl->ifl_m_enqueued, "mbufs allocated");
7172 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_dequeued",
7173 					CTLFLAG_RD,
7174 					&fl->ifl_m_dequeued, "mbufs freed");
7175 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_enqueued",
7176 					CTLFLAG_RD,
7177 					&fl->ifl_cl_enqueued, "clusters allocated");
7178 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_dequeued",
7179 					CTLFLAG_RD,
7180 					&fl->ifl_cl_dequeued, "clusters freed");
7181 #endif
7182 		}
7183 	}
7184 #endif
7185 }
7186 
7187 void
7188 iflib_request_reset(if_ctx_t ctx)
7189 {
7190 
7191 	STATE_LOCK(ctx);
7192 	ctx->ifc_flags |= IFC_DO_RESET;
7193 	STATE_UNLOCK(ctx);
7194 }
7195 
7196 #ifndef __NO_STRICT_ALIGNMENT
7197 static struct mbuf *
7198 iflib_fixup_rx(struct mbuf *m)
7199 {
7200 	struct mbuf *n;
7201 
7202 	if (m->m_len <= (MCLBYTES - ETHER_HDR_LEN)) {
7203 		bcopy(m->m_data, m->m_data + ETHER_HDR_LEN, m->m_len);
7204 		m->m_data += ETHER_HDR_LEN;
7205 		n = m;
7206 	} else {
7207 		MGETHDR(n, M_NOWAIT, MT_DATA);
7208 		if (n == NULL) {
7209 			m_freem(m);
7210 			return (NULL);
7211 		}
7212 		bcopy(m->m_data, n->m_data, ETHER_HDR_LEN);
7213 		m->m_data += ETHER_HDR_LEN;
7214 		m->m_len -= ETHER_HDR_LEN;
7215 		n->m_len = ETHER_HDR_LEN;
7216 		M_MOVE_PKTHDR(n, m);
7217 		n->m_next = m;
7218 	}
7219 	return (n);
7220 }
7221 #endif
7222 
7223 #ifdef DEBUGNET
7224 static void
7225 iflib_debugnet_init(if_t ifp, int *nrxr, int *ncl, int *clsize)
7226 {
7227 	if_ctx_t ctx;
7228 
7229 	ctx = if_getsoftc(ifp);
7230 	CTX_LOCK(ctx);
7231 	*nrxr = NRXQSETS(ctx);
7232 	*ncl = ctx->ifc_rxqs[0].ifr_fl->ifl_size;
7233 	*clsize = ctx->ifc_rxqs[0].ifr_fl->ifl_buf_size;
7234 	CTX_UNLOCK(ctx);
7235 }
7236 
7237 static void
7238 iflib_debugnet_event(if_t ifp, enum debugnet_ev event)
7239 {
7240 	if_ctx_t ctx;
7241 	if_softc_ctx_t scctx;
7242 	iflib_fl_t fl;
7243 	iflib_rxq_t rxq;
7244 	int i, j;
7245 
7246 	ctx = if_getsoftc(ifp);
7247 	scctx = &ctx->ifc_softc_ctx;
7248 
7249 	switch (event) {
7250 	case DEBUGNET_START:
7251 #ifndef __HAIKU__
7252 		for (i = 0; i < scctx->isc_nrxqsets; i++) {
7253 			rxq = &ctx->ifc_rxqs[i];
7254 			for (j = 0; j < rxq->ifr_nfl; j++) {
7255 				fl = rxq->ifr_fl;
7256 				fl->ifl_zone = m_getzone(fl->ifl_buf_size);
7257 			}
7258 		}
7259 		iflib_no_tx_batch = 1;
7260 		break;
7261 #endif
7262 	default:
7263 		break;
7264 	}
7265 }
7266 
7267 static int
7268 iflib_debugnet_transmit(if_t ifp, struct mbuf *m)
7269 {
7270 	if_ctx_t ctx;
7271 	iflib_txq_t txq;
7272 	int error;
7273 
7274 	ctx = if_getsoftc(ifp);
7275 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
7276 	    IFF_DRV_RUNNING)
7277 		return (EBUSY);
7278 
7279 	txq = &ctx->ifc_txqs[0];
7280 	error = iflib_encap(txq, &m);
7281 	if (error == 0)
7282 		(void)iflib_txd_db_check(txq, true);
7283 	return (error);
7284 }
7285 
7286 static int
7287 iflib_debugnet_poll(if_t ifp, int count)
7288 {
7289 	struct epoch_tracker et;
7290 	if_ctx_t ctx;
7291 	if_softc_ctx_t scctx;
7292 	iflib_txq_t txq;
7293 	int i;
7294 
7295 	ctx = if_getsoftc(ifp);
7296 	scctx = &ctx->ifc_softc_ctx;
7297 
7298 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
7299 	    IFF_DRV_RUNNING)
7300 		return (EBUSY);
7301 
7302 	txq = &ctx->ifc_txqs[0];
7303 	(void)iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
7304 
7305 	NET_EPOCH_ENTER(et);
7306 	for (i = 0; i < scctx->isc_nrxqsets; i++)
7307 		(void)iflib_rxeof(&ctx->ifc_rxqs[i], 16 /* XXX */);
7308 	NET_EPOCH_EXIT(et);
7309 	return (0);
7310 }
7311 #endif /* DEBUGNET */
7312