xref: /haiku/src/libs/compat/freebsd_iflib/iflib.c (revision fc7456e9b1ec38c941134ed6d01c438cf289381e)
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 
31 #ifndef __HAIKU__
32 #include "opt_inet.h"
33 #include "opt_inet6.h"
34 #include "opt_acpi.h"
35 #include "opt_sched.h"
36 #else
37 #include <sys/epoch.h>
38 #endif
39 
40 #include <sys/param.h>
41 #include <sys/types.h>
42 #include <sys/bus.h>
43 #include <sys/eventhandler.h>
44 #include <sys/kernel.h>
45 #include <sys/lock.h>
46 #include <sys/mutex.h>
47 #include <sys/sx.h>
48 #include <sys/module.h>
49 #include <sys/kobj.h>
50 #include <sys/rman.h>
51 #include <sys/sbuf.h>
52 #include <sys/smp.h>
53 #include <sys/socket.h>
54 #include <sys/sockio.h>
55 #include <sys/sysctl.h>
56 #include <sys/syslog.h>
57 #include <sys/taskqueue.h>
58 #include <sys/limits.h>
59 
60 #include <net/if.h>
61 #include <net/if_var.h>
62 #include <net/if_types.h>
63 #include <net/if_media.h>
64 #include <net/bpf.h>
65 #include <net/ethernet.h>
66 #include <net/if_vlan_var.h>
67 #include <net/mp_ring.h>
68 #include <net/debugnet.h>
69 #include <net/vnet.h>
70 
71 #include <netinet/in.h>
72 #ifndef __HAIKU__
73 #include <netinet/in_pcb.h>
74 #include <netinet/tcp_lro.h>
75 #include <netinet/in_systm.h>
76 #endif
77 #include <netinet/if_ether.h>
78 #include <netinet/ip.h>
79 #include <netinet/ip6.h>
80 #include <netinet/tcp.h>
81 #include <netinet/ip_var.h>
82 #ifndef __HAIKU__
83 #include <netinet6/ip6_var.h>
84 #endif
85 
86 #include <machine/bus.h>
87 #ifndef __HAIKU__
88 #include <machine/in_cksum.h>
89 #endif
90 
91 #include <vm/vm.h>
92 #include <vm/pmap.h>
93 
94 #include <dev/led/led.h>
95 #include <dev/pci/pcireg.h>
96 #include <dev/pci/pcivar.h>
97 #ifndef __HAIKU__
98 #include <dev/pci/pci_private.h>
99 #endif
100 
101 #include <net/iflib.h>
102 
103 #include <ifdi_if.h>
104 #include <device_if.h>
105 
106 #ifdef PCI_IOV
107 #include <dev/pci/pci_iov.h>
108 #endif
109 
110 #include <sys/bitstring.h>
111 /*
112  * enable accounting of every mbuf as it comes in to and goes out of
113  * iflib's software descriptor references
114  */
115 #define MEMORY_LOGGING 0
116 /*
117  * Enable mbuf vectors for compressing long mbuf chains
118  */
119 
120 /*
121  * NB:
122  * - Prefetching in tx cleaning should perhaps be a tunable. The distance ahead
123  *   we prefetch needs to be determined by the time spent in m_free vis a vis
124  *   the cost of a prefetch. This will of course vary based on the workload:
125  *      - NFLX's m_free path is dominated by vm-based M_EXT manipulation which
126  *        is quite expensive, thus suggesting very little prefetch.
127  *      - small packet forwarding which is just returning a single mbuf to
128  *        UMA will typically be very fast vis a vis the cost of a memory
129  *        access.
130  */
131 
132 /*
133  * File organization:
134  *  - private structures
135  *  - iflib private utility functions
136  *  - ifnet functions
137  *  - vlan registry and other exported functions
138  *  - iflib public core functions
139  *
140  *
141  */
142 static MALLOC_DEFINE(M_IFLIB, "iflib", "ifnet library");
143 
144 #define	IFLIB_RXEOF_MORE (1U << 0)
145 #define	IFLIB_RXEOF_EMPTY (2U << 0)
146 
147 struct iflib_txq;
148 typedef struct iflib_txq *iflib_txq_t;
149 struct iflib_rxq;
150 typedef struct iflib_rxq *iflib_rxq_t;
151 struct iflib_fl;
152 typedef struct iflib_fl *iflib_fl_t;
153 
154 struct iflib_ctx;
155 
156 static void iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid);
157 static void iflib_timer(void *arg);
158 static void iflib_tqg_detach(if_ctx_t ctx);
159 
160 typedef struct iflib_filter_info {
161 	driver_filter_t *ifi_filter;
162 	void *ifi_filter_arg;
163 	struct grouptask *ifi_task;
164 	void *ifi_ctx;
165 } *iflib_filter_info_t;
166 
167 struct iflib_ctx {
168 	KOBJ_FIELDS;
169 	/*
170 	 * Pointer to hardware driver's softc
171 	 */
172 	void *ifc_softc;
173 	device_t ifc_dev;
174 	if_t ifc_ifp;
175 
176 #ifndef __HAIKU__
177 	cpuset_t ifc_cpus;
178 #endif
179 	if_shared_ctx_t ifc_sctx;
180 	struct if_softc_ctx ifc_softc_ctx;
181 
182 	struct sx ifc_ctx_sx;
183 	struct mtx ifc_state_mtx;
184 
185 	iflib_txq_t ifc_txqs;
186 	iflib_rxq_t ifc_rxqs;
187 	uint32_t ifc_if_flags;
188 	uint32_t ifc_flags;
189 	uint32_t ifc_max_fl_buf_size;
190 	uint32_t ifc_rx_mbuf_sz;
191 
192 	int ifc_link_state;
193 	int ifc_watchdog_events;
194 	struct cdev *ifc_led_dev;
195 	struct resource *ifc_msix_mem;
196 
197 	struct if_irq ifc_legacy_irq;
198 	struct grouptask ifc_admin_task;
199 	struct grouptask ifc_vflr_task;
200 	struct iflib_filter_info ifc_filter_info;
201 	struct ifmedia	ifc_media;
202 	struct ifmedia	*ifc_mediap;
203 
204 	struct sysctl_oid *ifc_sysctl_node;
205 	uint16_t ifc_sysctl_ntxqs;
206 	uint16_t ifc_sysctl_nrxqs;
207 	uint16_t ifc_sysctl_qs_eq_override;
208 	uint16_t ifc_sysctl_rx_budget;
209 	uint16_t ifc_sysctl_tx_abdicate;
210 	uint16_t ifc_sysctl_core_offset;
211 #define	CORE_OFFSET_UNSPECIFIED	0xffff
212 	uint8_t  ifc_sysctl_separate_txrx;
213 	uint8_t  ifc_sysctl_use_logical_cores;
214 	uint16_t ifc_sysctl_extra_msix_vectors;
215 	bool	 ifc_cpus_are_physical_cores;
216 
217 	qidx_t ifc_sysctl_ntxds[8];
218 	qidx_t ifc_sysctl_nrxds[8];
219 	struct if_txrx ifc_txrx;
220 #define isc_txd_encap  ifc_txrx.ift_txd_encap
221 #define isc_txd_flush  ifc_txrx.ift_txd_flush
222 #define isc_txd_credits_update  ifc_txrx.ift_txd_credits_update
223 #define isc_rxd_available ifc_txrx.ift_rxd_available
224 #define isc_rxd_pkt_get ifc_txrx.ift_rxd_pkt_get
225 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
226 #define isc_rxd_flush ifc_txrx.ift_rxd_flush
227 #define isc_legacy_intr ifc_txrx.ift_legacy_intr
228 #define isc_txq_select ifc_txrx.ift_txq_select
229 #define isc_txq_select_v2 ifc_txrx.ift_txq_select_v2
230 
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 void
265 iflib_set_mac(if_ctx_t ctx, uint8_t mac[ETHER_ADDR_LEN])
266 {
267 
268 	bcopy(mac, ctx->ifc_mac.octet, ETHER_ADDR_LEN);
269 }
270 
271 if_softc_ctx_t
272 iflib_get_softc_ctx(if_ctx_t ctx)
273 {
274 
275 	return (&ctx->ifc_softc_ctx);
276 }
277 
278 if_shared_ctx_t
279 iflib_get_sctx(if_ctx_t ctx)
280 {
281 
282 	return (ctx->ifc_sctx);
283 }
284 
285 uint16_t
286 iflib_get_extra_msix_vectors_sysctl(if_ctx_t ctx)
287 {
288 
289 	return (ctx->ifc_sysctl_extra_msix_vectors);
290 }
291 
292 #define IP_ALIGNED(m) ((((uintptr_t)(m)->m_data) & 0x3) == 0x2)
293 #define CACHE_PTR_INCREMENT (CACHE_LINE_SIZE/sizeof(void*))
294 #define CACHE_PTR_NEXT(ptr) ((void *)(((uintptr_t)(ptr)+CACHE_LINE_SIZE-1) & (CACHE_LINE_SIZE-1)))
295 
296 #define LINK_ACTIVE(ctx) ((ctx)->ifc_link_state == LINK_STATE_UP)
297 #define CTX_IS_VF(ctx) ((ctx)->ifc_sctx->isc_flags & IFLIB_IS_VF)
298 
299 typedef struct iflib_sw_rx_desc_array {
300 	bus_dmamap_t	*ifsd_map;         /* bus_dma maps for packet */
301 	struct mbuf	**ifsd_m;           /* pkthdr mbufs */
302 	caddr_t		*ifsd_cl;          /* direct cluster pointer for rx */
303 	bus_addr_t	*ifsd_ba;          /* bus addr of cluster for rx */
304 } iflib_rxsd_array_t;
305 
306 typedef struct iflib_sw_tx_desc_array {
307 	bus_dmamap_t    *ifsd_map;         /* bus_dma maps for packet */
308 	bus_dmamap_t	*ifsd_tso_map;     /* bus_dma maps for TSO packet */
309 	struct mbuf    **ifsd_m;           /* pkthdr mbufs */
310 } if_txsd_vec_t;
311 
312 /* magic number that should be high enough for any hardware */
313 #define IFLIB_MAX_TX_SEGS		128
314 #define IFLIB_RX_COPY_THRESH		128
315 #define IFLIB_MAX_RX_REFRESH		32
316 /* The minimum descriptors per second before we start coalescing */
317 #define IFLIB_MIN_DESC_SEC		16384
318 #define IFLIB_DEFAULT_TX_UPDATE_FREQ	16
319 #define IFLIB_QUEUE_IDLE		0
320 #define IFLIB_QUEUE_HUNG		1
321 #define IFLIB_QUEUE_WORKING		2
322 /* maximum number of txqs that can share an rx interrupt */
323 #define IFLIB_MAX_TX_SHARED_INTR	4
324 
325 /* this should really scale with ring size - this is a fairly arbitrary value */
326 #define TX_BATCH_SIZE			32
327 
328 #define IFLIB_RESTART_BUDGET		8
329 
330 #define	IFC_LEGACY		0x001
331 #define	IFC_QFLUSH		0x002
332 #define	IFC_MULTISEG		0x004
333 #define	IFC_SPARE1		0x008
334 #define	IFC_SC_ALLOCATED	0x010
335 #define	IFC_INIT_DONE		0x020
336 #define	IFC_PREFETCH		0x040
337 #define	IFC_DO_RESET		0x080
338 #define	IFC_DO_WATCHDOG		0x100
339 #define	IFC_SPARE0		0x200
340 #define	IFC_SPARE2		0x400
341 #define	IFC_IN_DETACH		0x800
342 
343 #define	IFC_NETMAP_TX_IRQ	0x80000000
344 
345 #define CSUM_OFFLOAD		(CSUM_IP_TSO|CSUM_IP6_TSO|CSUM_IP| \
346 				 CSUM_IP_UDP|CSUM_IP_TCP|CSUM_IP_SCTP| \
347 				 CSUM_IP6_UDP|CSUM_IP6_TCP|CSUM_IP6_SCTP)
348 
349 struct iflib_txq {
350 	qidx_t		ift_in_use;
351 	qidx_t		ift_cidx;
352 	qidx_t		ift_cidx_processed;
353 	qidx_t		ift_pidx;
354 	uint8_t		ift_gen;
355 	uint8_t		ift_br_offset;
356 	uint16_t	ift_npending;
357 	uint16_t	ift_db_pending;
358 	uint16_t	ift_rs_pending;
359 	/* implicit pad */
360 	uint8_t		ift_txd_size[8];
361 	uint64_t	ift_processed;
362 	uint64_t	ift_cleaned;
363 	uint64_t	ift_cleaned_prev;
364 #if MEMORY_LOGGING
365 	uint64_t	ift_enqueued;
366 	uint64_t	ift_dequeued;
367 #endif
368 	uint64_t	ift_no_tx_dma_setup;
369 	uint64_t	ift_no_desc_avail;
370 	uint64_t	ift_mbuf_defrag_failed;
371 	uint64_t	ift_mbuf_defrag;
372 	uint64_t	ift_map_failed;
373 	uint64_t	ift_txd_encap_efbig;
374 	uint64_t	ift_pullups;
375 	uint64_t	ift_last_timer_tick;
376 
377 	struct mtx	ift_mtx;
378 	struct mtx	ift_db_mtx;
379 
380 	/* constant values */
381 	if_ctx_t	ift_ctx;
382 	struct ifmp_ring        *ift_br;
383 	struct grouptask	ift_task;
384 	qidx_t		ift_size;
385 	uint16_t	ift_id;
386 	struct callout	ift_timer;
387 #ifdef DEV_NETMAP
388 	struct callout	ift_netmap_timer;
389 #endif /* DEV_NETMAP */
390 
391 	if_txsd_vec_t	ift_sds;
392 	uint8_t		ift_qstatus;
393 	uint8_t		ift_closed;
394 	uint8_t		ift_update_freq;
395 	struct iflib_filter_info ift_filter_info;
396 	bus_dma_tag_t	ift_buf_tag;
397 	bus_dma_tag_t	ift_tso_buf_tag;
398 	iflib_dma_info_t	ift_ifdi;
399 #define	MTX_NAME_LEN	32
400 	char                    ift_mtx_name[MTX_NAME_LEN];
401 	bus_dma_segment_t	ift_segs[IFLIB_MAX_TX_SEGS]  __aligned(CACHE_LINE_SIZE);
402 #ifdef IFLIB_DIAGNOSTICS
403 	uint64_t ift_cpu_exec_count[256];
404 #endif
405 } __aligned(CACHE_LINE_SIZE);
406 
407 struct iflib_fl {
408 	qidx_t		ifl_cidx;
409 	qidx_t		ifl_pidx;
410 	qidx_t		ifl_credits;
411 	uint8_t		ifl_gen;
412 	uint8_t		ifl_rxd_size;
413 #if MEMORY_LOGGING
414 	uint64_t	ifl_m_enqueued;
415 	uint64_t	ifl_m_dequeued;
416 	uint64_t	ifl_cl_enqueued;
417 	uint64_t	ifl_cl_dequeued;
418 #endif
419 	/* implicit pad */
420 	bitstr_t 	*ifl_rx_bitmap;
421 	qidx_t		ifl_fragidx;
422 	/* constant */
423 	qidx_t		ifl_size;
424 	uint16_t	ifl_buf_size;
425 	uint16_t	ifl_cltype;
426 #ifndef __HAIKU__
427 	uma_zone_t	ifl_zone;
428 #endif
429 	iflib_rxsd_array_t	ifl_sds;
430 	iflib_rxq_t	ifl_rxq;
431 	uint8_t		ifl_id;
432 	bus_dma_tag_t	ifl_buf_tag;
433 	iflib_dma_info_t	ifl_ifdi;
434 	uint64_t	ifl_bus_addrs[IFLIB_MAX_RX_REFRESH] __aligned(CACHE_LINE_SIZE);
435 	qidx_t		ifl_rxd_idxs[IFLIB_MAX_RX_REFRESH];
436 }  __aligned(CACHE_LINE_SIZE);
437 
438 static inline qidx_t
439 get_inuse(int size, qidx_t cidx, qidx_t pidx, uint8_t gen)
440 {
441 	qidx_t used;
442 
443 	if (pidx > cidx)
444 		used = pidx - cidx;
445 	else if (pidx < cidx)
446 		used = size - cidx + pidx;
447 	else if (gen == 0 && pidx == cidx)
448 		used = 0;
449 	else if (gen == 1 && pidx == cidx)
450 		used = size;
451 	else
452 		panic("bad state");
453 
454 	return (used);
455 }
456 
457 #define TXQ_AVAIL(txq) (txq->ift_size - get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx, txq->ift_gen))
458 
459 #define IDXDIFF(head, tail, wrap) \
460 	((head) >= (tail) ? (head) - (tail) : (wrap) - (tail) + (head))
461 
462 struct iflib_rxq {
463 	if_ctx_t	ifr_ctx;
464 	iflib_fl_t	ifr_fl;
465 	uint64_t	ifr_rx_irq;
466 #ifndef __HAIKU__
467 	struct pfil_head	*pfil;
468 #else
469 #define PFIL_PASS 0
470 #endif
471 	/*
472 	 * If there is a separate completion queue (IFLIB_HAS_RXCQ), this is
473 	 * the completion queue consumer index.  Otherwise it's unused.
474 	 */
475 	qidx_t		ifr_cq_cidx;
476 	uint16_t	ifr_id;
477 	uint8_t		ifr_nfl;
478 	uint8_t		ifr_ntxqirq;
479 	uint8_t		ifr_txqid[IFLIB_MAX_TX_SHARED_INTR];
480 	uint8_t		ifr_fl_offset;
481 #ifndef __HAIKU__
482 	struct lro_ctrl			ifr_lc;
483 #endif
484 	struct grouptask        ifr_task;
485 	struct callout		ifr_watchdog;
486 	struct iflib_filter_info ifr_filter_info;
487 	iflib_dma_info_t		ifr_ifdi;
488 
489 	/* dynamically allocate if any drivers need a value substantially larger than this */
490 	struct if_rxd_frag	ifr_frags[IFLIB_MAX_RX_SEGS] __aligned(CACHE_LINE_SIZE);
491 #ifdef IFLIB_DIAGNOSTICS
492 	uint64_t ifr_cpu_exec_count[256];
493 #endif
494 }  __aligned(CACHE_LINE_SIZE);
495 
496 typedef struct if_rxsd {
497 	caddr_t *ifsd_cl;
498 	iflib_fl_t ifsd_fl;
499 } *if_rxsd_t;
500 
501 /* multiple of word size */
502 #ifdef __LP64__
503 #define PKT_INFO_SIZE	6
504 #define RXD_INFO_SIZE	5
505 #define PKT_TYPE uint64_t
506 #else
507 #define PKT_INFO_SIZE	11
508 #define RXD_INFO_SIZE	8
509 #define PKT_TYPE uint32_t
510 #endif
511 #define PKT_LOOP_BOUND  ((PKT_INFO_SIZE/3)*3)
512 #define RXD_LOOP_BOUND  ((RXD_INFO_SIZE/4)*4)
513 
514 typedef struct if_pkt_info_pad {
515 	PKT_TYPE pkt_val[PKT_INFO_SIZE];
516 } *if_pkt_info_pad_t;
517 typedef struct if_rxd_info_pad {
518 	PKT_TYPE rxd_val[RXD_INFO_SIZE];
519 } *if_rxd_info_pad_t;
520 
521 CTASSERT(sizeof(struct if_pkt_info_pad) == sizeof(struct if_pkt_info));
522 CTASSERT(sizeof(struct if_rxd_info_pad) == sizeof(struct if_rxd_info));
523 
524 static inline void
525 pkt_info_zero(if_pkt_info_t pi)
526 {
527 	if_pkt_info_pad_t pi_pad;
528 
529 	pi_pad = (if_pkt_info_pad_t)pi;
530 	pi_pad->pkt_val[0] = 0; pi_pad->pkt_val[1] = 0; pi_pad->pkt_val[2] = 0;
531 	pi_pad->pkt_val[3] = 0; pi_pad->pkt_val[4] = 0; pi_pad->pkt_val[5] = 0;
532 #ifndef __LP64__
533 	pi_pad->pkt_val[6] = 0; pi_pad->pkt_val[7] = 0; pi_pad->pkt_val[8] = 0;
534 	pi_pad->pkt_val[9] = 0; pi_pad->pkt_val[10] = 0;
535 #endif
536 }
537 
538 static inline void
539 rxd_info_zero(if_rxd_info_t ri)
540 {
541 	if_rxd_info_pad_t ri_pad;
542 	int i;
543 
544 	ri_pad = (if_rxd_info_pad_t)ri;
545 	for (i = 0; i < RXD_LOOP_BOUND; i += 4) {
546 		ri_pad->rxd_val[i] = 0;
547 		ri_pad->rxd_val[i+1] = 0;
548 		ri_pad->rxd_val[i+2] = 0;
549 		ri_pad->rxd_val[i+3] = 0;
550 	}
551 #ifdef __LP64__
552 	ri_pad->rxd_val[RXD_INFO_SIZE-1] = 0;
553 #endif
554 }
555 
556 /*
557  * Only allow a single packet to take up most 1/nth of the tx ring
558  */
559 #define MAX_SINGLE_PACKET_FRACTION 12
560 #define IF_BAD_DMA (bus_addr_t)-1
561 
562 #define CTX_ACTIVE(ctx) ((if_getdrvflags((ctx)->ifc_ifp) & IFF_DRV_RUNNING))
563 
564 #define CTX_LOCK_INIT(_sc)  sx_init(&(_sc)->ifc_ctx_sx, "iflib ctx lock")
565 #define CTX_LOCK(ctx) sx_xlock(&(ctx)->ifc_ctx_sx)
566 #define CTX_UNLOCK(ctx) sx_xunlock(&(ctx)->ifc_ctx_sx)
567 #define CTX_LOCK_DESTROY(ctx) sx_destroy(&(ctx)->ifc_ctx_sx)
568 
569 #define STATE_LOCK_INIT(_sc, _name)  mtx_init(&(_sc)->ifc_state_mtx, _name, "iflib state lock", MTX_DEF)
570 #define STATE_LOCK(ctx) mtx_lock(&(ctx)->ifc_state_mtx)
571 #define STATE_UNLOCK(ctx) mtx_unlock(&(ctx)->ifc_state_mtx)
572 #define STATE_LOCK_DESTROY(ctx) mtx_destroy(&(ctx)->ifc_state_mtx)
573 
574 #define CALLOUT_LOCK(txq)	mtx_lock(&txq->ift_mtx)
575 #define CALLOUT_UNLOCK(txq) 	mtx_unlock(&txq->ift_mtx)
576 
577 /* Our boot-time initialization hook */
578 static int	iflib_module_event_handler(module_t, int, void *);
579 
580 #ifndef __HAIKU__
581 static moduledata_t iflib_moduledata = {
582 	"iflib",
583 	iflib_module_event_handler,
584 	NULL
585 };
586 #endif
587 
588 DECLARE_MODULE(iflib, iflib_moduledata, SI_SUB_INIT_IF, SI_ORDER_ANY);
589 MODULE_VERSION(iflib, 1);
590 
591 MODULE_DEPEND(iflib, pci, 1, 1, 1);
592 MODULE_DEPEND(iflib, ether, 1, 1, 1);
593 
594 TASKQGROUP_DEFINE(if_io_tqg, mp_ncpus, 1);
595 TASKQGROUP_DEFINE(if_config_tqg, 1, 1);
596 
597 #ifndef IFLIB_DEBUG_COUNTERS
598 #ifdef INVARIANTS
599 #define IFLIB_DEBUG_COUNTERS 1
600 #else
601 #define IFLIB_DEBUG_COUNTERS 0
602 #endif /* !INVARIANTS */
603 #endif
604 
605 static SYSCTL_NODE(_net, OID_AUTO, iflib, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
606     "iflib driver parameters");
607 
608 /*
609  * XXX need to ensure that this can't accidentally cause the head to be moved backwards
610  */
611 static int iflib_min_tx_latency = 0;
612 SYSCTL_INT(_net_iflib, OID_AUTO, min_tx_latency, CTLFLAG_RW,
613 		   &iflib_min_tx_latency, 0, "minimize transmit latency at the possible expense of throughput");
614 static int iflib_no_tx_batch = 0;
615 SYSCTL_INT(_net_iflib, OID_AUTO, no_tx_batch, CTLFLAG_RW,
616 		   &iflib_no_tx_batch, 0, "minimize transmit latency at the possible expense of throughput");
617 static int iflib_timer_default = 1000;
618 SYSCTL_INT(_net_iflib, OID_AUTO, timer_default, CTLFLAG_RW,
619 		   &iflib_timer_default, 0, "number of ticks between iflib_timer calls");
620 
621 
622 #if IFLIB_DEBUG_COUNTERS
623 
624 static int iflib_tx_seen;
625 static int iflib_tx_sent;
626 static int iflib_tx_encap;
627 static int iflib_rx_allocs;
628 static int iflib_fl_refills;
629 static int iflib_fl_refills_large;
630 static int iflib_tx_frees;
631 
632 SYSCTL_INT(_net_iflib, OID_AUTO, tx_seen, CTLFLAG_RD,
633 		   &iflib_tx_seen, 0, "# TX mbufs seen");
634 SYSCTL_INT(_net_iflib, OID_AUTO, tx_sent, CTLFLAG_RD,
635 		   &iflib_tx_sent, 0, "# TX mbufs sent");
636 SYSCTL_INT(_net_iflib, OID_AUTO, tx_encap, CTLFLAG_RD,
637 		   &iflib_tx_encap, 0, "# TX mbufs encapped");
638 SYSCTL_INT(_net_iflib, OID_AUTO, tx_frees, CTLFLAG_RD,
639 		   &iflib_tx_frees, 0, "# TX frees");
640 SYSCTL_INT(_net_iflib, OID_AUTO, rx_allocs, CTLFLAG_RD,
641 		   &iflib_rx_allocs, 0, "# RX allocations");
642 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills, CTLFLAG_RD,
643 		   &iflib_fl_refills, 0, "# refills");
644 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills_large, CTLFLAG_RD,
645 		   &iflib_fl_refills_large, 0, "# large refills");
646 
647 static int iflib_txq_drain_flushing;
648 static int iflib_txq_drain_oactive;
649 static int iflib_txq_drain_notready;
650 
651 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_flushing, CTLFLAG_RD,
652 		   &iflib_txq_drain_flushing, 0, "# drain flushes");
653 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_oactive, CTLFLAG_RD,
654 		   &iflib_txq_drain_oactive, 0, "# drain oactives");
655 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_notready, CTLFLAG_RD,
656 		   &iflib_txq_drain_notready, 0, "# drain notready");
657 
658 static int iflib_encap_load_mbuf_fail;
659 static int iflib_encap_pad_mbuf_fail;
660 static int iflib_encap_txq_avail_fail;
661 static int iflib_encap_txd_encap_fail;
662 
663 SYSCTL_INT(_net_iflib, OID_AUTO, encap_load_mbuf_fail, CTLFLAG_RD,
664 		   &iflib_encap_load_mbuf_fail, 0, "# busdma load failures");
665 SYSCTL_INT(_net_iflib, OID_AUTO, encap_pad_mbuf_fail, CTLFLAG_RD,
666 		   &iflib_encap_pad_mbuf_fail, 0, "# runt frame pad failures");
667 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txq_avail_fail, CTLFLAG_RD,
668 		   &iflib_encap_txq_avail_fail, 0, "# txq avail failures");
669 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txd_encap_fail, CTLFLAG_RD,
670 		   &iflib_encap_txd_encap_fail, 0, "# driver encap failures");
671 
672 static int iflib_task_fn_rxs;
673 static int iflib_rx_intr_enables;
674 static int iflib_fast_intrs;
675 static int iflib_rx_unavail;
676 static int iflib_rx_ctx_inactive;
677 static int iflib_rx_if_input;
678 static int iflib_rxd_flush;
679 
680 static int iflib_verbose_debug;
681 
682 SYSCTL_INT(_net_iflib, OID_AUTO, task_fn_rx, CTLFLAG_RD,
683 		   &iflib_task_fn_rxs, 0, "# task_fn_rx calls");
684 SYSCTL_INT(_net_iflib, OID_AUTO, rx_intr_enables, CTLFLAG_RD,
685 		   &iflib_rx_intr_enables, 0, "# RX intr enables");
686 SYSCTL_INT(_net_iflib, OID_AUTO, fast_intrs, CTLFLAG_RD,
687 		   &iflib_fast_intrs, 0, "# fast_intr calls");
688 SYSCTL_INT(_net_iflib, OID_AUTO, rx_unavail, CTLFLAG_RD,
689 		   &iflib_rx_unavail, 0, "# times rxeof called with no available data");
690 SYSCTL_INT(_net_iflib, OID_AUTO, rx_ctx_inactive, CTLFLAG_RD,
691 		   &iflib_rx_ctx_inactive, 0, "# times rxeof called with inactive context");
692 SYSCTL_INT(_net_iflib, OID_AUTO, rx_if_input, CTLFLAG_RD,
693 		   &iflib_rx_if_input, 0, "# times rxeof called if_input");
694 SYSCTL_INT(_net_iflib, OID_AUTO, rxd_flush, CTLFLAG_RD,
695 	         &iflib_rxd_flush, 0, "# times rxd_flush called");
696 SYSCTL_INT(_net_iflib, OID_AUTO, verbose_debug, CTLFLAG_RW,
697 		   &iflib_verbose_debug, 0, "enable verbose debugging");
698 
699 #define DBG_COUNTER_INC(name) atomic_add_int(&(iflib_ ## name), 1)
700 static void
701 iflib_debug_reset(void)
702 {
703 	iflib_tx_seen = iflib_tx_sent = iflib_tx_encap = iflib_rx_allocs =
704 		iflib_fl_refills = iflib_fl_refills_large = iflib_tx_frees =
705 		iflib_txq_drain_flushing = iflib_txq_drain_oactive =
706 		iflib_txq_drain_notready =
707 		iflib_encap_load_mbuf_fail = iflib_encap_pad_mbuf_fail =
708 		iflib_encap_txq_avail_fail = iflib_encap_txd_encap_fail =
709 		iflib_task_fn_rxs = iflib_rx_intr_enables = iflib_fast_intrs =
710 		iflib_rx_unavail =
711 		iflib_rx_ctx_inactive = iflib_rx_if_input =
712 		iflib_rxd_flush = 0;
713 }
714 
715 #else
716 #define DBG_COUNTER_INC(name)
717 static void iflib_debug_reset(void) {}
718 #endif
719 
720 #define IFLIB_DEBUG 0
721 
722 static void iflib_tx_structures_free(if_ctx_t ctx);
723 static void iflib_rx_structures_free(if_ctx_t ctx);
724 static int iflib_queues_alloc(if_ctx_t ctx);
725 static int iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq);
726 static int iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget);
727 static int iflib_qset_structures_setup(if_ctx_t ctx);
728 static int iflib_msix_init(if_ctx_t ctx);
729 static int iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filterarg, int *rid, const char *str);
730 static void iflib_txq_check_drain(iflib_txq_t txq, int budget);
731 static uint32_t iflib_txq_can_drain(struct ifmp_ring *);
732 #ifdef ALTQ
733 static void iflib_altq_if_start(if_t ifp);
734 static int iflib_altq_if_transmit(if_t ifp, struct mbuf *m);
735 #endif
736 static int iflib_register(if_ctx_t);
737 static void iflib_deregister(if_ctx_t);
738 static void iflib_unregister_vlan_handlers(if_ctx_t ctx);
739 static uint16_t iflib_get_mbuf_size_for(unsigned int size);
740 static void iflib_init_locked(if_ctx_t ctx);
741 static void iflib_add_device_sysctl_pre(if_ctx_t ctx);
742 static void iflib_add_device_sysctl_post(if_ctx_t ctx);
743 static void iflib_ifmp_purge(iflib_txq_t txq);
744 static void _iflib_pre_assert(if_softc_ctx_t scctx);
745 static void iflib_stop(if_ctx_t ctx);
746 static void iflib_if_init_locked(if_ctx_t ctx);
747 static void iflib_free_intr_mem(if_ctx_t ctx);
748 #ifndef __NO_STRICT_ALIGNMENT
749 static struct mbuf * iflib_fixup_rx(struct mbuf *m);
750 #endif
751 
752 #ifndef __HAIKU__
753 static SLIST_HEAD(cpu_offset_list, cpu_offset) cpu_offsets =
754     SLIST_HEAD_INITIALIZER(cpu_offsets);
755 struct cpu_offset {
756 	SLIST_ENTRY(cpu_offset) entries;
757 	cpuset_t	set;
758 	unsigned int	refcount;
759 	uint16_t	next_cpuid;
760 };
761 static struct mtx cpu_offset_mtx;
762 MTX_SYSINIT(iflib_cpu_offset, &cpu_offset_mtx, "iflib_cpu_offset lock",
763     MTX_DEF);
764 #endif
765 
766 DEBUGNET_DEFINE(iflib);
767 
768 static int
769 iflib_num_rx_descs(if_ctx_t ctx)
770 {
771 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
772 	if_shared_ctx_t sctx = ctx->ifc_sctx;
773 	uint16_t first_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
774 
775 	return scctx->isc_nrxd[first_rxq];
776 }
777 
778 static int
779 iflib_num_tx_descs(if_ctx_t ctx)
780 {
781 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
782 	if_shared_ctx_t sctx = ctx->ifc_sctx;
783 	uint16_t first_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
784 
785 	return scctx->isc_ntxd[first_txq];
786 }
787 
788 #ifdef DEV_NETMAP
789 #include <sys/selinfo.h>
790 #include <net/netmap.h>
791 #include <dev/netmap/netmap_kern.h>
792 
793 MODULE_DEPEND(iflib, netmap, 1, 1, 1);
794 
795 static int netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, bool init);
796 static void iflib_netmap_timer(void *arg);
797 
798 /*
799  * device-specific sysctl variables:
800  *
801  * iflib_crcstrip: 0: keep CRC in rx frames (default), 1: strip it.
802  *	During regular operations the CRC is stripped, but on some
803  *	hardware reception of frames not multiple of 64 is slower,
804  *	so using crcstrip=0 helps in benchmarks.
805  *
806  * iflib_rx_miss, iflib_rx_miss_bufs:
807  *	count packets that might be missed due to lost interrupts.
808  */
809 SYSCTL_DECL(_dev_netmap);
810 /*
811  * The xl driver by default strips CRCs and we do not override it.
812  */
813 
814 int iflib_crcstrip = 1;
815 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_crcstrip,
816     CTLFLAG_RW, &iflib_crcstrip, 1, "strip CRC on RX frames");
817 
818 int iflib_rx_miss, iflib_rx_miss_bufs;
819 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss,
820     CTLFLAG_RW, &iflib_rx_miss, 0, "potentially missed RX intr");
821 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss_bufs,
822     CTLFLAG_RW, &iflib_rx_miss_bufs, 0, "potentially missed RX intr bufs");
823 
824 /*
825  * Register/unregister. We are already under netmap lock.
826  * Only called on the first register or the last unregister.
827  */
828 static int
829 iflib_netmap_register(struct netmap_adapter *na, int onoff)
830 {
831 	if_t ifp = na->ifp;
832 	if_ctx_t ctx = if_getsoftc(ifp);
833 	int status;
834 
835 	CTX_LOCK(ctx);
836 	if (!CTX_IS_VF(ctx))
837 		IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip);
838 
839 	iflib_stop(ctx);
840 
841 	/*
842 	 * Enable (or disable) netmap flags, and intercept (or restore)
843 	 * ifp->if_transmit. This is done once the device has been stopped
844 	 * to prevent race conditions. Also, this must be done after
845 	 * calling netmap_disable_all_rings() and before calling
846 	 * netmap_enable_all_rings(), so that these two functions see the
847 	 * updated state of the NAF_NETMAP_ON bit.
848 	 */
849 	if (onoff) {
850 		nm_set_native_flags(na);
851 	} else {
852 		nm_clear_native_flags(na);
853 	}
854 
855 	iflib_init_locked(ctx);
856 	IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip); // XXX why twice ?
857 	status = if_getdrvflags(ifp) & IFF_DRV_RUNNING ? 0 : 1;
858 	if (status)
859 		nm_clear_native_flags(na);
860 	CTX_UNLOCK(ctx);
861 	return (status);
862 }
863 
864 static int
865 iflib_netmap_config(struct netmap_adapter *na, struct nm_config_info *info)
866 {
867 	if_t ifp = na->ifp;
868 	if_ctx_t ctx = if_getsoftc(ifp);
869 	iflib_rxq_t rxq = &ctx->ifc_rxqs[0];
870 	iflib_fl_t fl = &rxq->ifr_fl[0];
871 
872 	info->num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
873 	info->num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
874 	info->num_tx_descs = iflib_num_tx_descs(ctx);
875 	info->num_rx_descs = iflib_num_rx_descs(ctx);
876 	info->rx_buf_maxsize = fl->ifl_buf_size;
877 	nm_prinf("txr %u rxr %u txd %u rxd %u rbufsz %u",
878 		info->num_tx_rings, info->num_rx_rings, info->num_tx_descs,
879 		info->num_rx_descs, info->rx_buf_maxsize);
880 
881 	return 0;
882 }
883 
884 static int
885 netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, bool init)
886 {
887 	struct netmap_adapter *na = kring->na;
888 	u_int const lim = kring->nkr_num_slots - 1;
889 	struct netmap_ring *ring = kring->ring;
890 	bus_dmamap_t *map;
891 	struct if_rxd_update iru;
892 	if_ctx_t ctx = rxq->ifr_ctx;
893 	iflib_fl_t fl = &rxq->ifr_fl[0];
894 	u_int nic_i_first, nic_i;
895 	u_int nm_i;
896 	int i, n;
897 #if IFLIB_DEBUG_COUNTERS
898 	int rf_count = 0;
899 #endif
900 
901 	/*
902 	 * This function is used both at initialization and in rxsync.
903 	 * At initialization we need to prepare (with isc_rxd_refill())
904 	 * all the netmap buffers currently owned by the kernel, in
905 	 * such a way to keep fl->ifl_pidx and kring->nr_hwcur in sync
906 	 * (except for kring->nkr_hwofs). These may be less than
907 	 * kring->nkr_num_slots if netmap_reset() was called while
908 	 * an application using the kring that still owned some
909 	 * buffers.
910 	 * At rxsync time, both indexes point to the next buffer to be
911 	 * refilled.
912 	 * In any case we publish (with isc_rxd_flush()) up to
913 	 * (fl->ifl_pidx - 1) % N (included), to avoid the NIC tail/prod
914 	 * pointer to overrun the head/cons pointer, although this is
915 	 * not necessary for some NICs (e.g. vmx).
916 	 */
917 	if (__predict_false(init)) {
918 		n = kring->nkr_num_slots - nm_kr_rxspace(kring);
919 	} else {
920 		n = kring->rhead - kring->nr_hwcur;
921 		if (n == 0)
922 			return (0); /* Nothing to do. */
923 		if (n < 0)
924 			n += kring->nkr_num_slots;
925 	}
926 
927 	iru_init(&iru, rxq, 0 /* flid */);
928 	map = fl->ifl_sds.ifsd_map;
929 	nic_i = fl->ifl_pidx;
930 	nm_i = netmap_idx_n2k(kring, nic_i);
931 	if (__predict_false(init)) {
932 		/*
933 		 * On init/reset, nic_i must be 0, and we must
934 		 * start to refill from hwtail (see netmap_reset()).
935 		 */
936 		MPASS(nic_i == 0);
937 		MPASS(nm_i == kring->nr_hwtail);
938 	} else
939 		MPASS(nm_i == kring->nr_hwcur);
940 	DBG_COUNTER_INC(fl_refills);
941 	while (n > 0) {
942 #if IFLIB_DEBUG_COUNTERS
943 		if (++rf_count == 9)
944 			DBG_COUNTER_INC(fl_refills_large);
945 #endif
946 		nic_i_first = nic_i;
947 		for (i = 0; n > 0 && i < IFLIB_MAX_RX_REFRESH; n--, i++) {
948 			struct netmap_slot *slot = &ring->slot[nm_i];
949 			uint64_t paddr;
950 			void *addr = PNMB(na, slot, &paddr);
951 
952 			MPASS(i < IFLIB_MAX_RX_REFRESH);
953 
954 			if (addr == NETMAP_BUF_BASE(na)) /* bad buf */
955 			        return netmap_ring_reinit(kring);
956 
957 			fl->ifl_bus_addrs[i] = paddr +
958 			    nm_get_offset(kring, slot);
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 const lim = kring->nkr_num_slots - 1;
1023 	u_int const head = kring->rhead;
1024 	struct if_pkt_info pi;
1025 	int tx_pkts = 0, tx_bytes = 0;
1026 
1027 	/*
1028 	 * interrupts on every tx packet are expensive so request
1029 	 * them every half ring, or where NS_REPORT is set
1030 	 */
1031 	u_int report_frequency = kring->nkr_num_slots >> 1;
1032 	/* device-specific */
1033 	if_ctx_t ctx = if_getsoftc(ifp);
1034 	iflib_txq_t txq = &ctx->ifc_txqs[kring->ring_id];
1035 
1036 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1037 	    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1038 
1039 	/*
1040 	 * First part: process new packets to send.
1041 	 * nm_i is the current index in the netmap kring,
1042 	 * nic_i is the corresponding index in the NIC ring.
1043 	 *
1044 	 * If we have packets to send (nm_i != head)
1045 	 * iterate over the netmap ring, fetch length and update
1046 	 * the corresponding slot in the NIC ring. Some drivers also
1047 	 * need to update the buffer's physical address in the NIC slot
1048 	 * even NS_BUF_CHANGED is not set (PNMB computes the addresses).
1049 	 *
1050 	 * The netmap_reload_map() calls is especially expensive,
1051 	 * even when (as in this case) the tag is 0, so do only
1052 	 * when the buffer has actually changed.
1053 	 *
1054 	 * If possible do not set the report/intr bit on all slots,
1055 	 * but only a few times per ring or when NS_REPORT is set.
1056 	 *
1057 	 * Finally, on 10G and faster drivers, it might be useful
1058 	 * to prefetch the next slot and txr entry.
1059 	 */
1060 
1061 	nm_i = kring->nr_hwcur;
1062 	if (nm_i != head) {	/* we have new packets to send */
1063 		uint32_t pkt_len = 0, seg_idx = 0;
1064 		int nic_i_start = -1, flags = 0;
1065 		pkt_info_zero(&pi);
1066 		pi.ipi_segs = txq->ift_segs;
1067 		pi.ipi_qsidx = kring->ring_id;
1068 		nic_i = netmap_idx_k2n(kring, nm_i);
1069 
1070 		__builtin_prefetch(&ring->slot[nm_i]);
1071 		__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i]);
1072 		__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i]);
1073 
1074 		while (nm_i != head) {
1075 			struct netmap_slot *slot = &ring->slot[nm_i];
1076 			uint64_t offset = nm_get_offset(kring, slot);
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 + offset;
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_OFF(na, len, offset);
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 = if_getsoftc(ifp);
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 = if_getsoftc(na->ifp);
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 | NAF_OFFSETS;
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 
1431 static __inline void
1432 prefetch2cachelines(void *x)
1433 {
1434 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1435 #if (CACHE_LINE_SIZE < 128)
1436 	__asm volatile("prefetcht0 %0" :: "m" (*(((unsigned long *)x)+CACHE_LINE_SIZE/(sizeof(unsigned long)))));
1437 #endif
1438 }
1439 #else
1440 static __inline void
1441 prefetch(void *x)
1442 {
1443 }
1444 
1445 static __inline void
1446 prefetch2cachelines(void *x)
1447 {
1448 }
1449 #endif
1450 
1451 static void
1452 iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid)
1453 {
1454 	iflib_fl_t fl;
1455 
1456 	fl = &rxq->ifr_fl[flid];
1457 	iru->iru_paddrs = fl->ifl_bus_addrs;
1458 	iru->iru_idxs = fl->ifl_rxd_idxs;
1459 	iru->iru_qsidx = rxq->ifr_id;
1460 	iru->iru_buf_size = fl->ifl_buf_size;
1461 	iru->iru_flidx = fl->ifl_id;
1462 }
1463 
1464 static void
1465 _iflib_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int err)
1466 {
1467 	if (err)
1468 		return;
1469 	*(bus_addr_t *) arg = segs[0].ds_addr;
1470 }
1471 
1472 #define	DMA_WIDTH_TO_BUS_LOWADDR(width)				\
1473 	(((width) == 0) || (width) == flsll(BUS_SPACE_MAXADDR) ?	\
1474 	    BUS_SPACE_MAXADDR : (1ULL << (width)) - 1ULL)
1475 
1476 int
1477 iflib_dma_alloc_align(if_ctx_t ctx, int size, int align, iflib_dma_info_t dma, int mapflags)
1478 {
1479 	int err;
1480 	device_t dev = ctx->ifc_dev;
1481 	bus_addr_t lowaddr;
1482 
1483 	lowaddr = DMA_WIDTH_TO_BUS_LOWADDR(ctx->ifc_softc_ctx.isc_dma_width);
1484 
1485 	err = bus_dma_tag_create(bus_get_dma_tag(dev),	/* parent */
1486 				align, 0,		/* alignment, bounds */
1487 				lowaddr,		/* lowaddr */
1488 				BUS_SPACE_MAXADDR,	/* highaddr */
1489 				NULL, NULL,		/* filter, filterarg */
1490 				size,			/* maxsize */
1491 				1,			/* nsegments */
1492 				size,			/* maxsegsize */
1493 				BUS_DMA_ALLOCNOW,	/* flags */
1494 				NULL,			/* lockfunc */
1495 				NULL,			/* lockarg */
1496 				&dma->idi_tag);
1497 	if (err) {
1498 		device_printf(dev,
1499 		    "%s: bus_dma_tag_create failed: %d (size=%d, align=%d)\n",
1500 		    __func__, err, size, align);
1501 		goto fail_0;
1502 	}
1503 
1504 	err = bus_dmamem_alloc(dma->idi_tag, (void**) &dma->idi_vaddr,
1505 	    BUS_DMA_NOWAIT | BUS_DMA_COHERENT | BUS_DMA_ZERO, &dma->idi_map);
1506 	if (err) {
1507 		device_printf(dev,
1508 		    "%s: bus_dmamem_alloc(%ju) failed: %d\n",
1509 		    __func__, (uintmax_t)size, err);
1510 		goto fail_1;
1511 	}
1512 
1513 	dma->idi_paddr = IF_BAD_DMA;
1514 	err = bus_dmamap_load(dma->idi_tag, dma->idi_map, dma->idi_vaddr,
1515 	    size, _iflib_dmamap_cb, &dma->idi_paddr, mapflags | BUS_DMA_NOWAIT);
1516 	if (err || dma->idi_paddr == IF_BAD_DMA) {
1517 		device_printf(dev,
1518 		    "%s: bus_dmamap_load failed: %d\n",
1519 		    __func__, err);
1520 		goto fail_2;
1521 	}
1522 
1523 	dma->idi_size = size;
1524 	return (0);
1525 
1526 fail_2:
1527 	bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1528 fail_1:
1529 	bus_dma_tag_destroy(dma->idi_tag);
1530 fail_0:
1531 	dma->idi_tag = NULL;
1532 
1533 	return (err);
1534 }
1535 
1536 int
1537 iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags)
1538 {
1539 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1540 
1541 	KASSERT(sctx->isc_q_align != 0, ("alignment value not initialized"));
1542 
1543 	return (iflib_dma_alloc_align(ctx, size, sctx->isc_q_align, dma, mapflags));
1544 }
1545 
1546 int
1547 iflib_dma_alloc_multi(if_ctx_t ctx, int *sizes, iflib_dma_info_t *dmalist, int mapflags, int count)
1548 {
1549 	int i, err;
1550 	iflib_dma_info_t *dmaiter;
1551 
1552 	dmaiter = dmalist;
1553 	for (i = 0; i < count; i++, dmaiter++) {
1554 		if ((err = iflib_dma_alloc(ctx, sizes[i], *dmaiter, mapflags)) != 0)
1555 			break;
1556 	}
1557 	if (err)
1558 		iflib_dma_free_multi(dmalist, i);
1559 	return (err);
1560 }
1561 
1562 void
1563 iflib_dma_free(iflib_dma_info_t dma)
1564 {
1565 	if (dma->idi_tag == NULL)
1566 		return;
1567 	if (dma->idi_paddr != IF_BAD_DMA) {
1568 		bus_dmamap_sync(dma->idi_tag, dma->idi_map,
1569 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1570 		bus_dmamap_unload(dma->idi_tag, dma->idi_map);
1571 		dma->idi_paddr = IF_BAD_DMA;
1572 	}
1573 	if (dma->idi_vaddr != NULL) {
1574 		bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1575 		dma->idi_vaddr = NULL;
1576 	}
1577 	bus_dma_tag_destroy(dma->idi_tag);
1578 	dma->idi_tag = NULL;
1579 }
1580 
1581 void
1582 iflib_dma_free_multi(iflib_dma_info_t *dmalist, int count)
1583 {
1584 	int i;
1585 	iflib_dma_info_t *dmaiter = dmalist;
1586 
1587 	for (i = 0; i < count; i++, dmaiter++)
1588 		iflib_dma_free(*dmaiter);
1589 }
1590 
1591 static int
1592 iflib_fast_intr(void *arg)
1593 {
1594 	iflib_filter_info_t info = arg;
1595 	struct grouptask *gtask = info->ifi_task;
1596 	int result;
1597 
1598 	DBG_COUNTER_INC(fast_intrs);
1599 	if (info->ifi_filter != NULL) {
1600 		result = info->ifi_filter(info->ifi_filter_arg);
1601 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1602 			return (result);
1603 	}
1604 
1605 	GROUPTASK_ENQUEUE(gtask);
1606 	return (FILTER_HANDLED);
1607 }
1608 
1609 static int
1610 iflib_fast_intr_rxtx(void *arg)
1611 {
1612 	iflib_filter_info_t info = arg;
1613 	struct grouptask *gtask = info->ifi_task;
1614 	if_ctx_t ctx;
1615 	iflib_rxq_t rxq = (iflib_rxq_t)info->ifi_ctx;
1616 	iflib_txq_t txq;
1617 	void *sc;
1618 	int i, cidx, result;
1619 	qidx_t txqid;
1620 	bool intr_enable, intr_legacy;
1621 
1622 	DBG_COUNTER_INC(fast_intrs);
1623 	if (info->ifi_filter != NULL) {
1624 		result = info->ifi_filter(info->ifi_filter_arg);
1625 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1626 			return (result);
1627 	}
1628 
1629 	ctx = rxq->ifr_ctx;
1630 	sc = ctx->ifc_softc;
1631 	intr_enable = false;
1632 	intr_legacy = !!(ctx->ifc_flags & IFC_LEGACY);
1633 	MPASS(rxq->ifr_ntxqirq);
1634 	for (i = 0; i < rxq->ifr_ntxqirq; i++) {
1635 		txqid = rxq->ifr_txqid[i];
1636 		txq = &ctx->ifc_txqs[txqid];
1637 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1638 		    BUS_DMASYNC_POSTREAD);
1639 		if (!ctx->isc_txd_credits_update(sc, txqid, false)) {
1640 			if (intr_legacy)
1641 				intr_enable = true;
1642 			else
1643 				IFDI_TX_QUEUE_INTR_ENABLE(ctx, txqid);
1644 			continue;
1645 		}
1646 		GROUPTASK_ENQUEUE(&txq->ift_task);
1647 	}
1648 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_RXCQ)
1649 		cidx = rxq->ifr_cq_cidx;
1650 	else
1651 		cidx = rxq->ifr_fl[0].ifl_cidx;
1652 	if (iflib_rxd_avail(ctx, rxq, cidx, 1))
1653 		GROUPTASK_ENQUEUE(gtask);
1654 	else {
1655 		if (intr_legacy)
1656 			intr_enable = true;
1657 		else
1658 			IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
1659 		DBG_COUNTER_INC(rx_intr_enables);
1660 	}
1661 	if (intr_enable)
1662 		IFDI_INTR_ENABLE(ctx);
1663 	return (FILTER_HANDLED);
1664 }
1665 
1666 static int
1667 iflib_fast_intr_ctx(void *arg)
1668 {
1669 	iflib_filter_info_t info = arg;
1670 	struct grouptask *gtask = info->ifi_task;
1671 	int result;
1672 
1673 	DBG_COUNTER_INC(fast_intrs);
1674 	if (info->ifi_filter != NULL) {
1675 		result = info->ifi_filter(info->ifi_filter_arg);
1676 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1677 			return (result);
1678 	}
1679 
1680 	if (gtask->gt_taskqueue != NULL)
1681 		GROUPTASK_ENQUEUE(gtask);
1682 	return (FILTER_HANDLED);
1683 }
1684 
1685 static int
1686 _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
1687 		 driver_filter_t filter, driver_intr_t handler, void *arg,
1688 		 const char *name)
1689 {
1690 	struct resource *res;
1691 	void *tag = NULL;
1692 	device_t dev = ctx->ifc_dev;
1693 	int flags, i, rc;
1694 
1695 	flags = RF_ACTIVE;
1696 	if (ctx->ifc_flags & IFC_LEGACY)
1697 		flags |= RF_SHAREABLE;
1698 	MPASS(rid < 512);
1699 	i = rid;
1700 	res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &i, flags);
1701 	if (res == NULL) {
1702 		device_printf(dev,
1703 		    "failed to allocate IRQ for rid %d, name %s.\n", rid, name);
1704 		return (ENOMEM);
1705 	}
1706 	irq->ii_res = res;
1707 	KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL"));
1708 	rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET,
1709 						filter, handler, arg, &tag);
1710 	if (rc != 0) {
1711 		device_printf(dev,
1712 		    "failed to setup interrupt for rid %d, name %s: %d\n",
1713 					  rid, name ? name : "unknown", rc);
1714 		return (rc);
1715 	} else if (name)
1716 		bus_describe_intr(dev, res, tag, "%s", name);
1717 
1718 	irq->ii_tag = tag;
1719 	return (0);
1720 }
1721 
1722 /*********************************************************************
1723  *
1724  *  Allocate DMA resources for TX buffers as well as memory for the TX
1725  *  mbuf map.  TX DMA maps (non-TSO/TSO) and TX mbuf map are kept in a
1726  *  iflib_sw_tx_desc_array structure, storing all the information that
1727  *  is needed to transmit a packet on the wire.  This is called only
1728  *  once at attach, setup is done every reset.
1729  *
1730  **********************************************************************/
1731 static int
1732 iflib_txsd_alloc(iflib_txq_t txq)
1733 {
1734 	if_ctx_t ctx = txq->ift_ctx;
1735 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1736 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1737 	device_t dev = ctx->ifc_dev;
1738 	bus_size_t tsomaxsize;
1739 	bus_addr_t lowaddr;
1740 	int err, nsegments, ntsosegments;
1741 	bool tso;
1742 
1743 	nsegments = scctx->isc_tx_nsegments;
1744 	ntsosegments = scctx->isc_tx_tso_segments_max;
1745 	tsomaxsize = scctx->isc_tx_tso_size_max;
1746 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_VLAN_MTU)
1747 		tsomaxsize += sizeof(struct ether_vlan_header);
1748 	MPASS(scctx->isc_ntxd[0] > 0);
1749 	MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0);
1750 	MPASS(nsegments > 0);
1751 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) {
1752 		MPASS(ntsosegments > 0);
1753 		MPASS(sctx->isc_tso_maxsize >= tsomaxsize);
1754 	}
1755 
1756 	lowaddr = DMA_WIDTH_TO_BUS_LOWADDR(scctx->isc_dma_width);
1757 
1758 	/*
1759 	 * Set up DMA tags for TX buffers.
1760 	 */
1761 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1762 			       1, 0,			/* alignment, bounds */
1763 			       lowaddr,			/* lowaddr */
1764 			       BUS_SPACE_MAXADDR,	/* highaddr */
1765 			       NULL, NULL,		/* filter, filterarg */
1766 			       sctx->isc_tx_maxsize,		/* maxsize */
1767 			       nsegments,	/* nsegments */
1768 			       sctx->isc_tx_maxsegsize,	/* maxsegsize */
1769 			       0,			/* flags */
1770 			       NULL,			/* lockfunc */
1771 			       NULL,			/* lockfuncarg */
1772 			       &txq->ift_buf_tag))) {
1773 		device_printf(dev,"Unable to allocate TX DMA tag: %d\n", err);
1774 		device_printf(dev,"maxsize: %ju nsegments: %d maxsegsize: %ju\n",
1775 		    (uintmax_t)sctx->isc_tx_maxsize, nsegments, (uintmax_t)sctx->isc_tx_maxsegsize);
1776 		goto fail;
1777 	}
1778 	tso = (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) != 0;
1779 	if (tso && (err = bus_dma_tag_create(bus_get_dma_tag(dev),
1780 			       1, 0,			/* alignment, bounds */
1781 			       lowaddr,			/* lowaddr */
1782 			       BUS_SPACE_MAXADDR,	/* highaddr */
1783 			       NULL, NULL,		/* filter, filterarg */
1784 			       tsomaxsize,		/* maxsize */
1785 			       ntsosegments,	/* nsegments */
1786 			       sctx->isc_tso_maxsegsize,/* maxsegsize */
1787 			       0,			/* flags */
1788 			       NULL,			/* lockfunc */
1789 			       NULL,			/* lockfuncarg */
1790 			       &txq->ift_tso_buf_tag))) {
1791 		device_printf(dev, "Unable to allocate TSO TX DMA tag: %d\n",
1792 		    err);
1793 		goto fail;
1794 	}
1795 
1796 	/* Allocate memory for the TX mbuf map. */
1797 	if (!(txq->ift_sds.ifsd_m =
1798 	    (struct mbuf **) malloc(sizeof(struct mbuf *) *
1799 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1800 		device_printf(dev, "Unable to allocate TX mbuf map memory\n");
1801 		err = ENOMEM;
1802 		goto fail;
1803 	}
1804 
1805 	/*
1806 	 * Create the DMA maps for TX buffers.
1807 	 */
1808 	if ((txq->ift_sds.ifsd_map = (bus_dmamap_t *)malloc(
1809 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1810 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1811 		device_printf(dev,
1812 		    "Unable to allocate TX buffer DMA map memory\n");
1813 		err = ENOMEM;
1814 		goto fail;
1815 	}
1816 	if (tso && (txq->ift_sds.ifsd_tso_map = (bus_dmamap_t *)malloc(
1817 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1818 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1819 		device_printf(dev,
1820 		    "Unable to allocate TSO TX buffer map memory\n");
1821 		err = ENOMEM;
1822 		goto fail;
1823 	}
1824 	for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) {
1825 		err = bus_dmamap_create(txq->ift_buf_tag, 0,
1826 		    &txq->ift_sds.ifsd_map[i]);
1827 		if (err != 0) {
1828 			device_printf(dev, "Unable to create TX DMA map\n");
1829 			goto fail;
1830 		}
1831 		if (!tso)
1832 			continue;
1833 		err = bus_dmamap_create(txq->ift_tso_buf_tag, 0,
1834 		    &txq->ift_sds.ifsd_tso_map[i]);
1835 		if (err != 0) {
1836 			device_printf(dev, "Unable to create TSO TX DMA map\n");
1837 			goto fail;
1838 		}
1839 	}
1840 	return (0);
1841 fail:
1842 	/* We free all, it handles case where we are in the middle */
1843 	iflib_tx_structures_free(ctx);
1844 	return (err);
1845 }
1846 
1847 static void
1848 iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i)
1849 {
1850 	bus_dmamap_t map;
1851 
1852 	if (txq->ift_sds.ifsd_map != NULL) {
1853 		map = txq->ift_sds.ifsd_map[i];
1854 		bus_dmamap_sync(txq->ift_buf_tag, map, BUS_DMASYNC_POSTWRITE);
1855 		bus_dmamap_unload(txq->ift_buf_tag, map);
1856 		bus_dmamap_destroy(txq->ift_buf_tag, map);
1857 		txq->ift_sds.ifsd_map[i] = NULL;
1858 	}
1859 
1860 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1861 		map = txq->ift_sds.ifsd_tso_map[i];
1862 		bus_dmamap_sync(txq->ift_tso_buf_tag, map,
1863 		    BUS_DMASYNC_POSTWRITE);
1864 		bus_dmamap_unload(txq->ift_tso_buf_tag, map);
1865 		bus_dmamap_destroy(txq->ift_tso_buf_tag, map);
1866 		txq->ift_sds.ifsd_tso_map[i] = NULL;
1867 	}
1868 }
1869 
1870 static void
1871 iflib_txq_destroy(iflib_txq_t txq)
1872 {
1873 	if_ctx_t ctx = txq->ift_ctx;
1874 
1875 	for (int i = 0; i < txq->ift_size; i++)
1876 		iflib_txsd_destroy(ctx, txq, i);
1877 
1878 	if (txq->ift_br != NULL) {
1879 		ifmp_ring_free(txq->ift_br);
1880 		txq->ift_br = NULL;
1881 	}
1882 
1883 	mtx_destroy(&txq->ift_mtx);
1884 
1885 	if (txq->ift_sds.ifsd_map != NULL) {
1886 		free(txq->ift_sds.ifsd_map, M_IFLIB);
1887 		txq->ift_sds.ifsd_map = NULL;
1888 	}
1889 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1890 		free(txq->ift_sds.ifsd_tso_map, M_IFLIB);
1891 		txq->ift_sds.ifsd_tso_map = NULL;
1892 	}
1893 	if (txq->ift_sds.ifsd_m != NULL) {
1894 		free(txq->ift_sds.ifsd_m, M_IFLIB);
1895 		txq->ift_sds.ifsd_m = NULL;
1896 	}
1897 	if (txq->ift_buf_tag != NULL) {
1898 		bus_dma_tag_destroy(txq->ift_buf_tag);
1899 		txq->ift_buf_tag = NULL;
1900 	}
1901 	if (txq->ift_tso_buf_tag != NULL) {
1902 		bus_dma_tag_destroy(txq->ift_tso_buf_tag);
1903 		txq->ift_tso_buf_tag = NULL;
1904 	}
1905 	if (txq->ift_ifdi != NULL) {
1906 		free(txq->ift_ifdi, M_IFLIB);
1907 	}
1908 }
1909 
1910 static void
1911 iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i)
1912 {
1913 	struct mbuf **mp;
1914 
1915 	mp = &txq->ift_sds.ifsd_m[i];
1916 	if (*mp == NULL)
1917 		return;
1918 
1919 	if (txq->ift_sds.ifsd_map != NULL) {
1920 		bus_dmamap_sync(txq->ift_buf_tag,
1921 		    txq->ift_sds.ifsd_map[i], BUS_DMASYNC_POSTWRITE);
1922 		bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[i]);
1923 	}
1924 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1925 		bus_dmamap_sync(txq->ift_tso_buf_tag,
1926 		    txq->ift_sds.ifsd_tso_map[i], BUS_DMASYNC_POSTWRITE);
1927 		bus_dmamap_unload(txq->ift_tso_buf_tag,
1928 		    txq->ift_sds.ifsd_tso_map[i]);
1929 	}
1930 	m_freem(*mp);
1931 	DBG_COUNTER_INC(tx_frees);
1932 	*mp = NULL;
1933 }
1934 
1935 static int
1936 iflib_txq_setup(iflib_txq_t txq)
1937 {
1938 	if_ctx_t ctx = txq->ift_ctx;
1939 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1940 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1941 	iflib_dma_info_t di;
1942 	int i;
1943 
1944 	/* Set number of descriptors available */
1945 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
1946 	/* XXX make configurable */
1947 	txq->ift_update_freq = IFLIB_DEFAULT_TX_UPDATE_FREQ;
1948 
1949 	/* Reset indices */
1950 	txq->ift_cidx_processed = 0;
1951 	txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0;
1952 	txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset];
1953 
1954 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1955 		bzero((void *)di->idi_vaddr, di->idi_size);
1956 
1957 	IFDI_TXQ_SETUP(ctx, txq->ift_id);
1958 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1959 		bus_dmamap_sync(di->idi_tag, di->idi_map,
1960 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1961 	return (0);
1962 }
1963 
1964 /*********************************************************************
1965  *
1966  *  Allocate DMA resources for RX buffers as well as memory for the RX
1967  *  mbuf map, direct RX cluster pointer map and RX cluster bus address
1968  *  map.  RX DMA map, RX mbuf map, direct RX cluster pointer map and
1969  *  RX cluster map are kept in a iflib_sw_rx_desc_array structure.
1970  *  Since we use use one entry in iflib_sw_rx_desc_array per received
1971  *  packet, the maximum number of entries we'll need is equal to the
1972  *  number of hardware receive descriptors that we've allocated.
1973  *
1974  **********************************************************************/
1975 static int
1976 iflib_rxsd_alloc(iflib_rxq_t rxq)
1977 {
1978 	if_ctx_t ctx = rxq->ifr_ctx;
1979 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1980 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1981 	device_t dev = ctx->ifc_dev;
1982 	iflib_fl_t fl;
1983 	bus_addr_t lowaddr;
1984 	int			err;
1985 
1986 	MPASS(scctx->isc_nrxd[0] > 0);
1987 	MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0);
1988 
1989 	lowaddr = DMA_WIDTH_TO_BUS_LOWADDR(scctx->isc_dma_width);
1990 
1991 	fl = rxq->ifr_fl;
1992 	for (int i = 0; i < rxq->ifr_nfl; i++, fl++) {
1993 		fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */
1994 		/* Set up DMA tag for RX buffers. */
1995 		err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1996 					 1, 0,			/* alignment, bounds */
1997 					 lowaddr,		/* lowaddr */
1998 					 BUS_SPACE_MAXADDR,	/* highaddr */
1999 					 NULL, NULL,		/* filter, filterarg */
2000 					 sctx->isc_rx_maxsize,	/* maxsize */
2001 					 sctx->isc_rx_nsegments,	/* nsegments */
2002 					 sctx->isc_rx_maxsegsize,	/* maxsegsize */
2003 					 0,			/* flags */
2004 					 NULL,			/* lockfunc */
2005 					 NULL,			/* lockarg */
2006 					 &fl->ifl_buf_tag);
2007 		if (err) {
2008 			device_printf(dev,
2009 			    "Unable to allocate RX DMA tag: %d\n", err);
2010 			goto fail;
2011 		}
2012 
2013 		/* Allocate memory for the RX mbuf map. */
2014 		if (!(fl->ifl_sds.ifsd_m =
2015 		      (struct mbuf **) malloc(sizeof(struct mbuf *) *
2016 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
2017 			device_printf(dev,
2018 			    "Unable to allocate RX mbuf map memory\n");
2019 			err = ENOMEM;
2020 			goto fail;
2021 		}
2022 
2023 		/* Allocate memory for the direct RX cluster pointer map. */
2024 		if (!(fl->ifl_sds.ifsd_cl =
2025 		      (caddr_t *) malloc(sizeof(caddr_t) *
2026 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
2027 			device_printf(dev,
2028 			    "Unable to allocate RX cluster map memory\n");
2029 			err = ENOMEM;
2030 			goto fail;
2031 		}
2032 
2033 		/* Allocate memory for the RX cluster bus address map. */
2034 		if (!(fl->ifl_sds.ifsd_ba =
2035 		      (bus_addr_t *) malloc(sizeof(bus_addr_t) *
2036 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
2037 			device_printf(dev,
2038 			    "Unable to allocate RX bus address map memory\n");
2039 			err = ENOMEM;
2040 			goto fail;
2041 		}
2042 
2043 		/*
2044 		 * Create the DMA maps for RX buffers.
2045 		 */
2046 		if (!(fl->ifl_sds.ifsd_map =
2047 		      (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
2048 			device_printf(dev,
2049 			    "Unable to allocate RX buffer DMA map memory\n");
2050 			err = ENOMEM;
2051 			goto fail;
2052 		}
2053 		for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++) {
2054 			err = bus_dmamap_create(fl->ifl_buf_tag, 0,
2055 			    &fl->ifl_sds.ifsd_map[i]);
2056 			if (err != 0) {
2057 				device_printf(dev, "Unable to create RX buffer DMA map\n");
2058 				goto fail;
2059 			}
2060 		}
2061 	}
2062 	return (0);
2063 
2064 fail:
2065 	iflib_rx_structures_free(ctx);
2066 	return (err);
2067 }
2068 
2069 /*
2070  * Internal service routines
2071  */
2072 
2073 struct rxq_refill_cb_arg {
2074 	int               error;
2075 	bus_dma_segment_t seg;
2076 	int               nseg;
2077 };
2078 
2079 static void
2080 _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
2081 {
2082 	struct rxq_refill_cb_arg *cb_arg = arg;
2083 
2084 	cb_arg->error = error;
2085 	cb_arg->seg = segs[0];
2086 	cb_arg->nseg = nseg;
2087 }
2088 
2089 /**
2090  * iflib_fl_refill - refill an rxq free-buffer list
2091  * @ctx: the iflib context
2092  * @fl: the free list to refill
2093  * @count: the number of new buffers to allocate
2094  *
2095  * (Re)populate an rxq free-buffer list with up to @count new packet buffers.
2096  * The caller must assure that @count does not exceed the queue's capacity
2097  * minus one (since we always leave a descriptor unavailable).
2098  */
2099 static uint8_t
2100 iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
2101 {
2102 	struct if_rxd_update iru;
2103 	struct rxq_refill_cb_arg cb_arg;
2104 	struct mbuf *m;
2105 	caddr_t cl, *sd_cl;
2106 	struct mbuf **sd_m;
2107 	bus_dmamap_t *sd_map;
2108 	bus_addr_t bus_addr, *sd_ba;
2109 	int err, frag_idx, i, idx, n, pidx;
2110 	qidx_t credits;
2111 
2112 	MPASS(count <= fl->ifl_size - fl->ifl_credits - 1);
2113 
2114 	sd_m = fl->ifl_sds.ifsd_m;
2115 	sd_map = fl->ifl_sds.ifsd_map;
2116 	sd_cl = fl->ifl_sds.ifsd_cl;
2117 	sd_ba = fl->ifl_sds.ifsd_ba;
2118 	pidx = fl->ifl_pidx;
2119 	idx = pidx;
2120 	frag_idx = fl->ifl_fragidx;
2121 	credits = fl->ifl_credits;
2122 
2123 	i = 0;
2124 	n = count;
2125 	MPASS(n > 0);
2126 	MPASS(credits + n <= fl->ifl_size);
2127 
2128 	if (pidx < fl->ifl_cidx)
2129 		MPASS(pidx + n <= fl->ifl_cidx);
2130 	if (pidx == fl->ifl_cidx && (credits < fl->ifl_size))
2131 		MPASS(fl->ifl_gen == 0);
2132 	if (pidx > fl->ifl_cidx)
2133 		MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx);
2134 
2135 	DBG_COUNTER_INC(fl_refills);
2136 	if (n > 8)
2137 		DBG_COUNTER_INC(fl_refills_large);
2138 	iru_init(&iru, fl->ifl_rxq, fl->ifl_id);
2139 	while (n-- > 0) {
2140 		/*
2141 		 * We allocate an uninitialized mbuf + cluster, mbuf is
2142 		 * initialized after rx.
2143 		 *
2144 		 * If the cluster is still set then we know a minimum sized
2145 		 * packet was received
2146 		 */
2147 		bit_ffc_at(fl->ifl_rx_bitmap, frag_idx, fl->ifl_size,
2148 		    &frag_idx);
2149 		if (frag_idx < 0)
2150 			bit_ffc(fl->ifl_rx_bitmap, fl->ifl_size, &frag_idx);
2151 		MPASS(frag_idx >= 0);
2152 		if ((cl = sd_cl[frag_idx]) == NULL) {
2153 #ifndef __HAIKU__
2154 			cl = uma_zalloc(fl->ifl_zone, M_NOWAIT);
2155 			if (__predict_false(cl == NULL))
2156 #else
2157 			if ((cl = m_cljget(NULL, M_NOWAIT, fl->ifl_buf_size)) == NULL)
2158 #endif
2159 				break;
2160 
2161 			cb_arg.error = 0;
2162 			MPASS(sd_map != NULL);
2163 			err = bus_dmamap_load(fl->ifl_buf_tag, sd_map[frag_idx],
2164 			    cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg,
2165 			    BUS_DMA_NOWAIT);
2166 			if (__predict_false(err != 0 || cb_arg.error)) {
2167 #ifndef __HAIKU__
2168 				uma_zfree(fl->ifl_zone, cl);
2169 #else
2170 				m_free(cl);
2171 #endif
2172 				break;
2173 			}
2174 
2175 			sd_ba[frag_idx] = bus_addr = cb_arg.seg.ds_addr;
2176 			sd_cl[frag_idx] = cl;
2177 #if MEMORY_LOGGING
2178 			fl->ifl_cl_enqueued++;
2179 #endif
2180 		} else {
2181 			bus_addr = sd_ba[frag_idx];
2182 		}
2183 		bus_dmamap_sync(fl->ifl_buf_tag, sd_map[frag_idx],
2184 		    BUS_DMASYNC_PREREAD);
2185 
2186 		if (sd_m[frag_idx] == NULL) {
2187 			m = m_gethdr(M_NOWAIT, MT_NOINIT);
2188 			if (__predict_false(m == NULL))
2189 				break;
2190 			sd_m[frag_idx] = m;
2191 		}
2192 		bit_set(fl->ifl_rx_bitmap, frag_idx);
2193 #if MEMORY_LOGGING
2194 		fl->ifl_m_enqueued++;
2195 #endif
2196 
2197 		DBG_COUNTER_INC(rx_allocs);
2198 		fl->ifl_rxd_idxs[i] = frag_idx;
2199 		fl->ifl_bus_addrs[i] = bus_addr;
2200 		credits++;
2201 		i++;
2202 		MPASS(credits <= fl->ifl_size);
2203 		if (++idx == fl->ifl_size) {
2204 #ifdef INVARIANTS
2205 			fl->ifl_gen = 1;
2206 #endif
2207 			idx = 0;
2208 		}
2209 		if (n == 0 || i == IFLIB_MAX_RX_REFRESH) {
2210 			iru.iru_pidx = pidx;
2211 			iru.iru_count = i;
2212 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2213 			fl->ifl_pidx = idx;
2214 			fl->ifl_credits = credits;
2215 			pidx = idx;
2216 			i = 0;
2217 		}
2218 	}
2219 
2220 	if (n < count - 1) {
2221 		if (i != 0) {
2222 			iru.iru_pidx = pidx;
2223 			iru.iru_count = i;
2224 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2225 			fl->ifl_pidx = idx;
2226 			fl->ifl_credits = credits;
2227 		}
2228 		DBG_COUNTER_INC(rxd_flush);
2229 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2230 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2231 		ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id,
2232 		    fl->ifl_id, fl->ifl_pidx);
2233 		if (__predict_true(bit_test(fl->ifl_rx_bitmap, frag_idx))) {
2234 			fl->ifl_fragidx = frag_idx + 1;
2235 			if (fl->ifl_fragidx == fl->ifl_size)
2236 				fl->ifl_fragidx = 0;
2237 		} else {
2238 			fl->ifl_fragidx = frag_idx;
2239 		}
2240 	}
2241 
2242 	return (n == -1 ? 0 : IFLIB_RXEOF_EMPTY);
2243 }
2244 
2245 static inline uint8_t
2246 iflib_fl_refill_all(if_ctx_t ctx, iflib_fl_t fl)
2247 {
2248 	/*
2249 	 * We leave an unused descriptor to avoid pidx to catch up with cidx.
2250 	 * This is important as it confuses most NICs. For instance,
2251 	 * Intel NICs have (per receive ring) RDH and RDT registers, where
2252 	 * RDH points to the next receive descriptor to be used by the NIC,
2253 	 * and RDT for the next receive descriptor to be published by the
2254 	 * driver to the NIC (RDT - 1 is thus the last valid one).
2255 	 * The condition RDH == RDT means no descriptors are available to
2256 	 * the NIC, and thus it would be ambiguous if it also meant that
2257 	 * all the descriptors are available to the NIC.
2258 	 */
2259 	int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1;
2260 #ifdef INVARIANTS
2261 	int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1;
2262 #endif
2263 
2264 	MPASS(fl->ifl_credits <= fl->ifl_size);
2265 	MPASS(reclaimable == delta);
2266 
2267 	if (reclaimable > 0)
2268 		return (iflib_fl_refill(ctx, fl, reclaimable));
2269 	return (0);
2270 }
2271 
2272 uint8_t
2273 iflib_in_detach(if_ctx_t ctx)
2274 {
2275 	bool in_detach;
2276 
2277 	STATE_LOCK(ctx);
2278 	in_detach = !!(ctx->ifc_flags & IFC_IN_DETACH);
2279 	STATE_UNLOCK(ctx);
2280 	return (in_detach);
2281 }
2282 
2283 static void
2284 iflib_fl_bufs_free(iflib_fl_t fl)
2285 {
2286 	iflib_dma_info_t idi = fl->ifl_ifdi;
2287 	bus_dmamap_t sd_map;
2288 	uint32_t i;
2289 
2290 	for (i = 0; i < fl->ifl_size; i++) {
2291 		struct mbuf **sd_m = &fl->ifl_sds.ifsd_m[i];
2292 		caddr_t *sd_cl = &fl->ifl_sds.ifsd_cl[i];
2293 
2294 		if (*sd_cl != NULL) {
2295 			sd_map = fl->ifl_sds.ifsd_map[i];
2296 			bus_dmamap_sync(fl->ifl_buf_tag, sd_map,
2297 			    BUS_DMASYNC_POSTREAD);
2298 			bus_dmamap_unload(fl->ifl_buf_tag, sd_map);
2299 #ifndef __HAIKU__
2300 			uma_zfree(fl->ifl_zone, *sd_cl);
2301 #else
2302 			struct mbuf* mb = m_get(0, MT_DATA);
2303 			m_cljset(mb, *sd_cl, fl->ifl_cltype);
2304 			m_free(mb);
2305 #endif
2306 			*sd_cl = NULL;
2307 			if (*sd_m != NULL) {
2308 				m_init(*sd_m, M_NOWAIT, MT_DATA, 0);
2309 #ifndef __HAIKU__
2310 				m_free_raw(*sd_m);
2311 #else
2312 				m_free(*sd_m);
2313 #endif
2314 				*sd_m = NULL;
2315 			}
2316 		} else {
2317 			MPASS(*sd_m == NULL);
2318 		}
2319 #if MEMORY_LOGGING
2320 		fl->ifl_m_dequeued++;
2321 		fl->ifl_cl_dequeued++;
2322 #endif
2323 	}
2324 #ifdef INVARIANTS
2325 	for (i = 0; i < fl->ifl_size; i++) {
2326 		MPASS(fl->ifl_sds.ifsd_cl[i] == NULL);
2327 		MPASS(fl->ifl_sds.ifsd_m[i] == NULL);
2328 	}
2329 #endif
2330 	/*
2331 	 * Reset free list values
2332 	 */
2333 	fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = fl->ifl_fragidx = 0;
2334 	bzero(idi->idi_vaddr, idi->idi_size);
2335 }
2336 
2337 /*********************************************************************
2338  *
2339  *  Initialize a free list and its buffers.
2340  *
2341  **********************************************************************/
2342 static int
2343 iflib_fl_setup(iflib_fl_t fl)
2344 {
2345 	iflib_rxq_t rxq = fl->ifl_rxq;
2346 	if_ctx_t ctx = rxq->ifr_ctx;
2347 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2348 	int qidx;
2349 
2350 	bit_nclear(fl->ifl_rx_bitmap, 0, fl->ifl_size - 1);
2351 	/*
2352 	** Free current RX buffer structs and their mbufs
2353 	*/
2354 	iflib_fl_bufs_free(fl);
2355 	/* Now replenish the mbufs */
2356 	MPASS(fl->ifl_credits == 0);
2357 	qidx = rxq->ifr_fl_offset + fl->ifl_id;
2358 	if (scctx->isc_rxd_buf_size[qidx] != 0)
2359 		fl->ifl_buf_size = scctx->isc_rxd_buf_size[qidx];
2360 	else
2361 		fl->ifl_buf_size = ctx->ifc_rx_mbuf_sz;
2362 	/*
2363 	 * ifl_buf_size may be a driver-supplied value, so pull it up
2364 	 * to the selected mbuf size.
2365 	 */
2366 	fl->ifl_buf_size = iflib_get_mbuf_size_for(fl->ifl_buf_size);
2367 	if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size)
2368 		ctx->ifc_max_fl_buf_size = fl->ifl_buf_size;
2369 	fl->ifl_cltype = m_gettype(fl->ifl_buf_size);
2370 #ifndef __HAIKU__
2371 	fl->ifl_zone = m_getzone(fl->ifl_buf_size);
2372 #endif
2373 
2374 	/*
2375 	 * Avoid pre-allocating zillions of clusters to an idle card
2376 	 * potentially speeding up attach. In any case make sure
2377 	 * to leave a descriptor unavailable. See the comment in
2378 	 * iflib_fl_refill_all().
2379 	 */
2380 	MPASS(fl->ifl_size > 0);
2381 	(void)iflib_fl_refill(ctx, fl, min(128, fl->ifl_size - 1));
2382 	if (min(128, fl->ifl_size - 1) != fl->ifl_credits)
2383 		return (ENOBUFS);
2384 	/*
2385 	 * handle failure
2386 	 */
2387 	MPASS(rxq != NULL);
2388 	MPASS(fl->ifl_ifdi != NULL);
2389 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2390 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2391 	return (0);
2392 }
2393 
2394 /*********************************************************************
2395  *
2396  *  Free receive ring data structures
2397  *
2398  **********************************************************************/
2399 static void
2400 iflib_rx_sds_free(iflib_rxq_t rxq)
2401 {
2402 	iflib_fl_t fl;
2403 	int i, j;
2404 
2405 	if (rxq->ifr_fl != NULL) {
2406 		for (i = 0; i < rxq->ifr_nfl; i++) {
2407 			fl = &rxq->ifr_fl[i];
2408 			if (fl->ifl_buf_tag != NULL) {
2409 				if (fl->ifl_sds.ifsd_map != NULL) {
2410 					for (j = 0; j < fl->ifl_size; j++) {
2411 						bus_dmamap_sync(
2412 						    fl->ifl_buf_tag,
2413 						    fl->ifl_sds.ifsd_map[j],
2414 						    BUS_DMASYNC_POSTREAD);
2415 						bus_dmamap_unload(
2416 						    fl->ifl_buf_tag,
2417 						    fl->ifl_sds.ifsd_map[j]);
2418 						bus_dmamap_destroy(
2419 						    fl->ifl_buf_tag,
2420 						    fl->ifl_sds.ifsd_map[j]);
2421 					}
2422 				}
2423 				bus_dma_tag_destroy(fl->ifl_buf_tag);
2424 				fl->ifl_buf_tag = NULL;
2425 			}
2426 			free(fl->ifl_sds.ifsd_m, M_IFLIB);
2427 			free(fl->ifl_sds.ifsd_cl, M_IFLIB);
2428 			free(fl->ifl_sds.ifsd_ba, M_IFLIB);
2429 			free(fl->ifl_sds.ifsd_map, M_IFLIB);
2430 			free(fl->ifl_rx_bitmap, M_IFLIB);
2431 			fl->ifl_sds.ifsd_m = NULL;
2432 			fl->ifl_sds.ifsd_cl = NULL;
2433 			fl->ifl_sds.ifsd_ba = NULL;
2434 			fl->ifl_sds.ifsd_map = NULL;
2435 			fl->ifl_rx_bitmap = NULL;
2436 		}
2437 		free(rxq->ifr_fl, M_IFLIB);
2438 		rxq->ifr_fl = NULL;
2439 		free(rxq->ifr_ifdi, M_IFLIB);
2440 		rxq->ifr_ifdi = NULL;
2441 		rxq->ifr_cq_cidx = 0;
2442 	}
2443 }
2444 
2445 /*
2446  * Timer routine
2447  */
2448 static void
2449 iflib_timer(void *arg)
2450 {
2451 	iflib_txq_t txq = arg;
2452 	if_ctx_t ctx = txq->ift_ctx;
2453 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2454 	uint64_t this_tick = ticks;
2455 
2456 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
2457 		return;
2458 
2459 	/*
2460 	** Check on the state of the TX queue(s), this
2461 	** can be done without the lock because its RO
2462 	** and the HUNG state will be static if set.
2463 	*/
2464 	if (this_tick - txq->ift_last_timer_tick >= iflib_timer_default) {
2465 		txq->ift_last_timer_tick = this_tick;
2466 		IFDI_TIMER(ctx, txq->ift_id);
2467 		if ((txq->ift_qstatus == IFLIB_QUEUE_HUNG) &&
2468 		    ((txq->ift_cleaned_prev == txq->ift_cleaned) ||
2469 		     (sctx->isc_pause_frames == 0)))
2470 			goto hung;
2471 
2472 		if (txq->ift_qstatus != IFLIB_QUEUE_IDLE &&
2473 		    ifmp_ring_is_stalled(txq->ift_br)) {
2474 			KASSERT(ctx->ifc_link_state == LINK_STATE_UP,
2475 			    ("queue can't be marked as hung if interface is down"));
2476 			txq->ift_qstatus = IFLIB_QUEUE_HUNG;
2477 		}
2478 		txq->ift_cleaned_prev = txq->ift_cleaned;
2479 	}
2480 	/* handle any laggards */
2481 	if (txq->ift_db_pending)
2482 		GROUPTASK_ENQUEUE(&txq->ift_task);
2483 
2484 	sctx->isc_pause_frames = 0;
2485 	if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)
2486 		callout_reset_on(&txq->ift_timer, iflib_timer_default, iflib_timer,
2487 		    txq, txq->ift_timer.c_cpu);
2488 	return;
2489 
2490  hung:
2491 	device_printf(ctx->ifc_dev,
2492 	    "Watchdog timeout (TX: %d desc avail: %d pidx: %d) -- resetting\n",
2493 	    txq->ift_id, TXQ_AVAIL(txq), txq->ift_pidx);
2494 	STATE_LOCK(ctx);
2495 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2496 	ctx->ifc_flags |= (IFC_DO_WATCHDOG|IFC_DO_RESET);
2497 	iflib_admin_intr_deferred(ctx);
2498 	STATE_UNLOCK(ctx);
2499 }
2500 
2501 static uint16_t
2502 iflib_get_mbuf_size_for(unsigned int size)
2503 {
2504 
2505 	if (size <= MCLBYTES)
2506 		return (MCLBYTES);
2507 	else
2508 		return (MJUMPAGESIZE);
2509 }
2510 
2511 static void
2512 iflib_calc_rx_mbuf_sz(if_ctx_t ctx)
2513 {
2514 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2515 
2516 	/*
2517 	 * XXX don't set the max_frame_size to larger
2518 	 * than the hardware can handle
2519 	 */
2520 	ctx->ifc_rx_mbuf_sz =
2521 	    iflib_get_mbuf_size_for(sctx->isc_max_frame_size);
2522 }
2523 
2524 uint32_t
2525 iflib_get_rx_mbuf_sz(if_ctx_t ctx)
2526 {
2527 
2528 	return (ctx->ifc_rx_mbuf_sz);
2529 }
2530 
2531 static void
2532 iflib_init_locked(if_ctx_t ctx)
2533 {
2534 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2535 	if_t ifp = ctx->ifc_ifp;
2536 	iflib_fl_t fl;
2537 	iflib_txq_t txq;
2538 	iflib_rxq_t rxq;
2539 	int i, j, tx_ip_csum_flags, tx_ip6_csum_flags;
2540 
2541 	if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2542 	IFDI_INTR_DISABLE(ctx);
2543 
2544 	/*
2545 	 * See iflib_stop(). Useful in case iflib_init_locked() is
2546 	 * called without first calling iflib_stop().
2547 	 */
2548 	netmap_disable_all_rings(ifp);
2549 
2550 	tx_ip_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP);
2551 	tx_ip6_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | CSUM_IP6_SCTP);
2552 	/* Set hardware offload abilities */
2553 	if_clearhwassist(ifp);
2554 	if (if_getcapenable(ifp) & IFCAP_TXCSUM)
2555 		if_sethwassistbits(ifp, tx_ip_csum_flags, 0);
2556 	if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6)
2557 		if_sethwassistbits(ifp,  tx_ip6_csum_flags, 0);
2558 	if (if_getcapenable(ifp) & IFCAP_TSO4)
2559 		if_sethwassistbits(ifp, CSUM_IP_TSO, 0);
2560 	if (if_getcapenable(ifp) & IFCAP_TSO6)
2561 		if_sethwassistbits(ifp, CSUM_IP6_TSO, 0);
2562 
2563 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
2564 		CALLOUT_LOCK(txq);
2565 		callout_stop(&txq->ift_timer);
2566 #ifdef DEV_NETMAP
2567 		callout_stop(&txq->ift_netmap_timer);
2568 #endif /* DEV_NETMAP */
2569 		CALLOUT_UNLOCK(txq);
2570 		(void)iflib_netmap_txq_init(ctx, txq);
2571 	}
2572 
2573 	/*
2574 	 * Calculate a suitable Rx mbuf size prior to calling IFDI_INIT, so
2575 	 * that drivers can use the value when setting up the hardware receive
2576 	 * buffers.
2577 	 */
2578 	iflib_calc_rx_mbuf_sz(ctx);
2579 
2580 #ifdef INVARIANTS
2581 	i = if_getdrvflags(ifp);
2582 #endif
2583 	IFDI_INIT(ctx);
2584 	MPASS(if_getdrvflags(ifp) == i);
2585 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
2586 		if (iflib_netmap_rxq_init(ctx, rxq) > 0) {
2587 			/* This rxq is in netmap mode. Skip normal init. */
2588 			continue;
2589 		}
2590 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
2591 			if (iflib_fl_setup(fl)) {
2592 				device_printf(ctx->ifc_dev,
2593 				    "setting up free list %d failed - "
2594 				    "check cluster settings\n", j);
2595 				goto done;
2596 			}
2597 		}
2598 	}
2599 done:
2600 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
2601 	IFDI_INTR_ENABLE(ctx);
2602 	txq = ctx->ifc_txqs;
2603 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++)
2604 		callout_reset_on(&txq->ift_timer, iflib_timer_default, iflib_timer, txq,
2605 			txq->ift_timer.c_cpu);
2606 
2607         /* Re-enable txsync/rxsync. */
2608 	netmap_enable_all_rings(ifp);
2609 }
2610 
2611 static int
2612 iflib_media_change(if_t ifp)
2613 {
2614 	if_ctx_t ctx = if_getsoftc(ifp);
2615 	int err;
2616 
2617 	CTX_LOCK(ctx);
2618 	if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0)
2619 		iflib_if_init_locked(ctx);
2620 	CTX_UNLOCK(ctx);
2621 	return (err);
2622 }
2623 
2624 static void
2625 iflib_media_status(if_t ifp, struct ifmediareq *ifmr)
2626 {
2627 	if_ctx_t ctx = if_getsoftc(ifp);
2628 
2629 	CTX_LOCK(ctx);
2630 	IFDI_UPDATE_ADMIN_STATUS(ctx);
2631 	IFDI_MEDIA_STATUS(ctx, ifmr);
2632 	CTX_UNLOCK(ctx);
2633 }
2634 
2635 void
2636 iflib_stop(if_ctx_t ctx)
2637 {
2638 	iflib_txq_t txq = ctx->ifc_txqs;
2639 	iflib_rxq_t rxq = ctx->ifc_rxqs;
2640 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2641 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2642 	iflib_dma_info_t di;
2643 	iflib_fl_t fl;
2644 	int i, j;
2645 
2646 	/* Tell the stack that the interface is no longer active */
2647 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2648 
2649 	IFDI_INTR_DISABLE(ctx);
2650 	DELAY(1000);
2651 	IFDI_STOP(ctx);
2652 	DELAY(1000);
2653 
2654 	/*
2655 	 * Stop any pending txsync/rxsync and prevent new ones
2656 	 * form starting. Processes blocked in poll() will get
2657 	 * POLLERR.
2658 	 */
2659 	netmap_disable_all_rings(ctx->ifc_ifp);
2660 
2661 	iflib_debug_reset();
2662 	/* Wait for current tx queue users to exit to disarm watchdog timer. */
2663 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) {
2664 		/* make sure all transmitters have completed before proceeding XXX */
2665 
2666 		CALLOUT_LOCK(txq);
2667 		callout_stop(&txq->ift_timer);
2668 #ifdef DEV_NETMAP
2669 		callout_stop(&txq->ift_netmap_timer);
2670 #endif /* DEV_NETMAP */
2671 		CALLOUT_UNLOCK(txq);
2672 
2673 		/* clean any enqueued buffers */
2674 		iflib_ifmp_purge(txq);
2675 		/* Free any existing tx buffers. */
2676 		for (j = 0; j < txq->ift_size; j++) {
2677 			iflib_txsd_free(ctx, txq, j);
2678 		}
2679 		txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0;
2680 		txq->ift_in_use = txq->ift_gen = txq->ift_no_desc_avail = 0;
2681 		if (sctx->isc_flags & IFLIB_PRESERVE_TX_INDICES)
2682 			txq->ift_cidx = txq->ift_pidx;
2683 		else
2684 			txq->ift_cidx = txq->ift_pidx = 0;
2685 
2686 		txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0;
2687 		txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0;
2688 		txq->ift_pullups = 0;
2689 		ifmp_ring_reset_stats(txq->ift_br);
2690 		for (j = 0, di = txq->ift_ifdi; j < sctx->isc_ntxqs; j++, di++)
2691 			bzero((void *)di->idi_vaddr, di->idi_size);
2692 	}
2693 	for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) {
2694 		if (rxq->ifr_task.gt_taskqueue != NULL)
2695 			gtaskqueue_drain(rxq->ifr_task.gt_taskqueue,
2696 				 &rxq->ifr_task.gt_task);
2697 
2698 		rxq->ifr_cq_cidx = 0;
2699 		for (j = 0, di = rxq->ifr_ifdi; j < sctx->isc_nrxqs; j++, di++)
2700 			bzero((void *)di->idi_vaddr, di->idi_size);
2701 		/* also resets the free lists pidx/cidx */
2702 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
2703 			iflib_fl_bufs_free(fl);
2704 	}
2705 }
2706 
2707 static inline caddr_t
2708 calc_next_rxd(iflib_fl_t fl, int cidx)
2709 {
2710 	qidx_t size;
2711 	int nrxd;
2712 	caddr_t start, end, cur, next;
2713 
2714 	nrxd = fl->ifl_size;
2715 	size = fl->ifl_rxd_size;
2716 	start = fl->ifl_ifdi->idi_vaddr;
2717 
2718 	if (__predict_false(size == 0))
2719 		return (start);
2720 	cur = start + size*cidx;
2721 	end = start + size*nrxd;
2722 	next = CACHE_PTR_NEXT(cur);
2723 	return (next < end ? next : start);
2724 }
2725 
2726 static inline void
2727 prefetch_pkts(iflib_fl_t fl, int cidx)
2728 {
2729 	int nextptr;
2730 	int nrxd = fl->ifl_size;
2731 	caddr_t next_rxd;
2732 
2733 	nextptr = (cidx + CACHE_PTR_INCREMENT) & (nrxd-1);
2734 	prefetch(&fl->ifl_sds.ifsd_m[nextptr]);
2735 	prefetch(&fl->ifl_sds.ifsd_cl[nextptr]);
2736 	next_rxd = calc_next_rxd(fl, cidx);
2737 	prefetch(next_rxd);
2738 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 1) & (nrxd-1)]);
2739 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 2) & (nrxd-1)]);
2740 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 3) & (nrxd-1)]);
2741 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 4) & (nrxd-1)]);
2742 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 1) & (nrxd-1)]);
2743 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 2) & (nrxd-1)]);
2744 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 3) & (nrxd-1)]);
2745 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 4) & (nrxd-1)]);
2746 }
2747 
2748 static struct mbuf *
2749 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, bool unload, if_rxsd_t sd,
2750     int *pf_rv, if_rxd_info_t ri)
2751 {
2752 	bus_dmamap_t map;
2753 	iflib_fl_t fl;
2754 	caddr_t payload;
2755 	struct mbuf *m;
2756 	int flid, cidx, len, next;
2757 
2758 	map = NULL;
2759 	flid = irf->irf_flid;
2760 	cidx = irf->irf_idx;
2761 	fl = &rxq->ifr_fl[flid];
2762 	sd->ifsd_fl = fl;
2763 	sd->ifsd_cl = &fl->ifl_sds.ifsd_cl[cidx];
2764 	fl->ifl_credits--;
2765 #if MEMORY_LOGGING
2766 	fl->ifl_m_dequeued++;
2767 #endif
2768 	if (rxq->ifr_ctx->ifc_flags & IFC_PREFETCH)
2769 		prefetch_pkts(fl, cidx);
2770 	next = (cidx + CACHE_PTR_INCREMENT) & (fl->ifl_size-1);
2771 	prefetch(&fl->ifl_sds.ifsd_map[next]);
2772 	map = fl->ifl_sds.ifsd_map[cidx];
2773 
2774 	bus_dmamap_sync(fl->ifl_buf_tag, map, BUS_DMASYNC_POSTREAD);
2775 
2776 #ifndef __HAIKU__
2777 	if (rxq->pfil != NULL && PFIL_HOOKED_IN(rxq->pfil) && pf_rv != NULL &&
2778 	    irf->irf_len != 0) {
2779 		payload  = *sd->ifsd_cl;
2780 		payload +=  ri->iri_pad;
2781 		len = ri->iri_len - ri->iri_pad;
2782 		*pf_rv = pfil_mem_in(rxq->pfil, payload, len, ri->iri_ifp, &m);
2783 		switch (*pf_rv) {
2784 		case PFIL_DROPPED:
2785 		case PFIL_CONSUMED:
2786 			/*
2787 			 * The filter ate it.  Everything is recycled.
2788 			 */
2789 			m = NULL;
2790 			unload = 0;
2791 			break;
2792 		case PFIL_REALLOCED:
2793 			/*
2794 			 * The filter copied it.  Everything is recycled.
2795 			 * 'm' points at new mbuf.
2796 			 */
2797 			unload = 0;
2798 			break;
2799 		case PFIL_PASS:
2800 			/*
2801 			 * Filter said it was OK, so receive like
2802 			 * normal
2803 			 */
2804 			m = fl->ifl_sds.ifsd_m[cidx];
2805 			fl->ifl_sds.ifsd_m[cidx] = NULL;
2806 			break;
2807 		default:
2808 			MPASS(0);
2809 		}
2810 	} else
2811 #endif
2812 	{
2813 		m = fl->ifl_sds.ifsd_m[cidx];
2814 		fl->ifl_sds.ifsd_m[cidx] = NULL;
2815 		if (pf_rv != NULL)
2816 			*pf_rv = PFIL_PASS;
2817 	}
2818 
2819 	if (unload && irf->irf_len != 0)
2820 		bus_dmamap_unload(fl->ifl_buf_tag, map);
2821 	fl->ifl_cidx = (fl->ifl_cidx + 1) & (fl->ifl_size-1);
2822 	if (__predict_false(fl->ifl_cidx == 0))
2823 		fl->ifl_gen = 0;
2824 	bit_clear(fl->ifl_rx_bitmap, cidx);
2825 	return (m);
2826 }
2827 
2828 static struct mbuf *
2829 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri, if_rxsd_t sd, int *pf_rv)
2830 {
2831 	struct mbuf *m, *mh, *mt;
2832 	caddr_t cl;
2833 	int  *pf_rv_ptr, flags, i, padlen;
2834 	bool consumed;
2835 
2836 	i = 0;
2837 	mh = NULL;
2838 	consumed = false;
2839 	*pf_rv = PFIL_PASS;
2840 	pf_rv_ptr = pf_rv;
2841 	do {
2842 		m = rxd_frag_to_sd(rxq, &ri->iri_frags[i], !consumed, sd,
2843 		    pf_rv_ptr, ri);
2844 
2845 		MPASS(*sd->ifsd_cl != NULL);
2846 
2847 		/*
2848 		 * Exclude zero-length frags & frags from
2849 		 * packets the filter has consumed or dropped
2850 		 */
2851 		if (ri->iri_frags[i].irf_len == 0 || consumed ||
2852 #ifndef __HAIKU__
2853 		    *pf_rv == PFIL_CONSUMED || *pf_rv == PFIL_DROPPED
2854 #else
2855 			0
2856 #endif
2857 		     ) {
2858 			if (mh == NULL) {
2859 				/* everything saved here */
2860 				consumed = true;
2861 				pf_rv_ptr = NULL;
2862 				continue;
2863 			}
2864 			/* XXX we can save the cluster here, but not the mbuf */
2865 			m_init(m, M_NOWAIT, MT_DATA, 0);
2866 			m_free(m);
2867 			continue;
2868 		}
2869 		if (mh == NULL) {
2870 			flags = M_PKTHDR|M_EXT;
2871 			mh = mt = m;
2872 			padlen = ri->iri_pad;
2873 		} else {
2874 			flags = M_EXT;
2875 			mt->m_next = m;
2876 			mt = m;
2877 			/* assuming padding is only on the first fragment */
2878 			padlen = 0;
2879 		}
2880 		cl = *sd->ifsd_cl;
2881 		*sd->ifsd_cl = NULL;
2882 
2883 		/* Can these two be made one ? */
2884 		m_init(m, M_NOWAIT, MT_DATA, flags);
2885 		m_cljset(m, cl, sd->ifsd_fl->ifl_cltype);
2886 		/*
2887 		 * These must follow m_init and m_cljset
2888 		 */
2889 		m->m_data += padlen;
2890 		ri->iri_len -= padlen;
2891 		m->m_len = ri->iri_frags[i].irf_len;
2892 	} while (++i < ri->iri_nfrags);
2893 
2894 	return (mh);
2895 }
2896 
2897 /*
2898  * Process one software descriptor
2899  */
2900 static struct mbuf *
2901 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri)
2902 {
2903 	struct if_rxsd sd;
2904 	struct mbuf *m;
2905 	int pf_rv;
2906 
2907 	/* should I merge this back in now that the two paths are basically duplicated? */
2908 	if (ri->iri_nfrags == 1 &&
2909 	    ri->iri_frags[0].irf_len != 0 &&
2910 	    ri->iri_frags[0].irf_len <= MIN(IFLIB_RX_COPY_THRESH, MHLEN)) {
2911 		m = rxd_frag_to_sd(rxq, &ri->iri_frags[0], false, &sd,
2912 		    &pf_rv, ri);
2913 		if (pf_rv != PFIL_PASS
2914 #ifndef __HAIKU__
2915 		        && pf_rv != PFIL_REALLOCED
2916 #endif
2917 		        )
2918 			return (m);
2919 		if (pf_rv == PFIL_PASS) {
2920 			m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR);
2921 #ifndef __NO_STRICT_ALIGNMENT
2922 			if (!IP_ALIGNED(m) && ri->iri_pad == 0)
2923 				m->m_data += 2;
2924 #endif
2925 			memcpy(m->m_data, *sd.ifsd_cl, ri->iri_len);
2926 			m->m_len = ri->iri_frags[0].irf_len;
2927 			m->m_data += ri->iri_pad;
2928 			ri->iri_len -= ri->iri_pad;
2929 		}
2930 	} else {
2931 		m = assemble_segments(rxq, ri, &sd, &pf_rv);
2932 		if (m == NULL)
2933 			return (NULL);
2934 		if (pf_rv != PFIL_PASS
2935 #ifndef __HAIKU__
2936 		        && pf_rv != PFIL_REALLOCED
2937 #endif
2938 				)
2939 			return (m);
2940 	}
2941 	m->m_pkthdr.len = ri->iri_len;
2942 	m->m_pkthdr.rcvif = ri->iri_ifp;
2943 	m->m_flags |= ri->iri_flags;
2944 	m->m_pkthdr.ether_vtag = ri->iri_vtag;
2945 	m->m_pkthdr.flowid = ri->iri_flowid;
2946 	M_HASHTYPE_SET(m, ri->iri_rsstype);
2947 	m->m_pkthdr.csum_flags = ri->iri_csum_flags;
2948 	m->m_pkthdr.csum_data = ri->iri_csum_data;
2949 	return (m);
2950 }
2951 
2952 #if defined(INET6) || defined(INET)
2953 static void
2954 iflib_get_ip_forwarding(struct lro_ctrl *lc, bool *v4, bool *v6)
2955 {
2956 	CURVNET_SET(if_getvnet(lc->ifp));
2957 #if defined(INET6)
2958 	*v6 = V_ip6_forwarding;
2959 #endif
2960 #if defined(INET)
2961 	*v4 = V_ipforwarding;
2962 #endif
2963 	CURVNET_RESTORE();
2964 }
2965 
2966 /*
2967  * Returns true if it's possible this packet could be LROed.
2968  * if it returns false, it is guaranteed that tcp_lro_rx()
2969  * would not return zero.
2970  */
2971 static bool
2972 iflib_check_lro_possible(struct mbuf *m, bool v4_forwarding, bool v6_forwarding)
2973 {
2974 #ifndef __HAIKU__
2975 	struct ether_header *eh;
2976 
2977 	eh = mtod(m, struct ether_header *);
2978 	switch (eh->ether_type) {
2979 #if defined(INET6)
2980 		case htons(ETHERTYPE_IPV6):
2981 			return (!v6_forwarding);
2982 #endif
2983 #if defined (INET)
2984 		case htons(ETHERTYPE_IP):
2985 			return (!v4_forwarding);
2986 #endif
2987 	}
2988 #endif
2989 
2990 	return false;
2991 }
2992 #else
2993 static void
2994 iflib_get_ip_forwarding(struct lro_ctrl *lc __unused, bool *v4 __unused, bool *v6 __unused)
2995 {
2996 }
2997 #endif
2998 
2999 static void
3000 _task_fn_rx_watchdog(void *context)
3001 {
3002 	iflib_rxq_t rxq = context;
3003 
3004 	GROUPTASK_ENQUEUE(&rxq->ifr_task);
3005 }
3006 
3007 static uint8_t
3008 iflib_rxeof(iflib_rxq_t rxq, qidx_t budget)
3009 {
3010 	if_t ifp;
3011 	if_ctx_t ctx = rxq->ifr_ctx;
3012 	if_shared_ctx_t sctx = ctx->ifc_sctx;
3013 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
3014 	int avail, i;
3015 	qidx_t *cidxp;
3016 	struct if_rxd_info ri;
3017 	int err, budget_left, rx_bytes, rx_pkts;
3018 	iflib_fl_t fl;
3019 	int lro_enabled;
3020 	bool v4_forwarding, v6_forwarding, lro_possible;
3021 	uint8_t retval = 0;
3022 
3023 	/*
3024 	 * XXX early demux data packets so that if_input processing only handles
3025 	 * acks in interrupt context
3026 	 */
3027 	struct mbuf *m, *mh, *mt, *mf;
3028 
3029 	NET_EPOCH_ASSERT();
3030 
3031 	lro_possible = v4_forwarding = v6_forwarding = false;
3032 	ifp = ctx->ifc_ifp;
3033 	mh = mt = NULL;
3034 	MPASS(budget > 0);
3035 	rx_pkts	= rx_bytes = 0;
3036 	if (sctx->isc_flags & IFLIB_HAS_RXCQ)
3037 		cidxp = &rxq->ifr_cq_cidx;
3038 	else
3039 		cidxp = &rxq->ifr_fl[0].ifl_cidx;
3040 	if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) {
3041 		for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
3042 			retval |= iflib_fl_refill_all(ctx, fl);
3043 		DBG_COUNTER_INC(rx_unavail);
3044 		return (retval);
3045 	}
3046 
3047 	/* pfil needs the vnet to be set */
3048 	CURVNET_SET_QUIET(if_getvnet(ifp));
3049 	for (budget_left = budget; budget_left > 0 && avail > 0;) {
3050 		if (__predict_false(!CTX_ACTIVE(ctx))) {
3051 			DBG_COUNTER_INC(rx_ctx_inactive);
3052 			break;
3053 		}
3054 		/*
3055 		 * Reset client set fields to their default values
3056 		 */
3057 		rxd_info_zero(&ri);
3058 		ri.iri_qsidx = rxq->ifr_id;
3059 		ri.iri_cidx = *cidxp;
3060 		ri.iri_ifp = ifp;
3061 		ri.iri_frags = rxq->ifr_frags;
3062 		err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
3063 
3064 		if (err)
3065 			goto err;
3066 		rx_pkts += 1;
3067 		rx_bytes += ri.iri_len;
3068 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
3069 			*cidxp = ri.iri_cidx;
3070 			/* Update our consumer index */
3071 			/* XXX NB: shurd - check if this is still safe */
3072 			while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0])
3073 				rxq->ifr_cq_cidx -= scctx->isc_nrxd[0];
3074 			/* was this only a completion queue message? */
3075 			if (__predict_false(ri.iri_nfrags == 0))
3076 				continue;
3077 		}
3078 		MPASS(ri.iri_nfrags != 0);
3079 		MPASS(ri.iri_len != 0);
3080 
3081 		/* will advance the cidx on the corresponding free lists */
3082 		m = iflib_rxd_pkt_get(rxq, &ri);
3083 		avail--;
3084 		budget_left--;
3085 		if (avail == 0 && budget_left)
3086 			avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left);
3087 
3088 		if (__predict_false(m == NULL))
3089 			continue;
3090 
3091 		/* imm_pkt: -- cxgb */
3092 		if (mh == NULL)
3093 			mh = mt = m;
3094 		else {
3095 			mt->m_nextpkt = m;
3096 			mt = m;
3097 		}
3098 	}
3099 	CURVNET_RESTORE();
3100 	/* make sure that we can refill faster than drain */
3101 	for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
3102 		retval |= iflib_fl_refill_all(ctx, fl);
3103 
3104 	lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO);
3105 #ifndef __HAIKU__
3106 	if (lro_enabled)
3107 		iflib_get_ip_forwarding(&rxq->ifr_lc, &v4_forwarding, &v6_forwarding);
3108 #endif
3109 	mt = mf = NULL;
3110 	while (mh != NULL) {
3111 		m = mh;
3112 		mh = mh->m_nextpkt;
3113 		m->m_nextpkt = NULL;
3114 #ifndef __NO_STRICT_ALIGNMENT
3115 		if (!IP_ALIGNED(m) && (m = iflib_fixup_rx(m)) == NULL)
3116 			continue;
3117 #endif
3118 #ifndef __HAIKU__
3119 #if defined(INET6) || defined(INET)
3120 		if (lro_enabled) {
3121 			if (!lro_possible) {
3122 				lro_possible = iflib_check_lro_possible(m, v4_forwarding, v6_forwarding);
3123 				if (lro_possible && mf != NULL) {
3124 					ifp->if_input(ifp, mf);
3125 					DBG_COUNTER_INC(rx_if_input);
3126 					mt = mf = NULL;
3127 				}
3128 			}
3129 			if ((m->m_pkthdr.csum_flags & (CSUM_L4_CALC|CSUM_L4_VALID)) ==
3130 			    (CSUM_L4_CALC|CSUM_L4_VALID)) {
3131 				if (lro_possible && tcp_lro_rx(&rxq->ifr_lc, m, 0) == 0)
3132 					continue;
3133 			}
3134 		}
3135 #endif
3136 		if (lro_possible) {
3137 			if_input(ifp, m);
3138 			DBG_COUNTER_INC(rx_if_input);
3139 			continue;
3140 		}
3141 #endif
3142 
3143 		if (mf == NULL)
3144 			mf = m;
3145 		if (mt != NULL)
3146 			mt->m_nextpkt = m;
3147 		mt = m;
3148 	}
3149 	if (mf != NULL) {
3150 		if_input(ifp, mf);
3151 		DBG_COUNTER_INC(rx_if_input);
3152 	}
3153 
3154 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
3155 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
3156 
3157 	/*
3158 	 * Flush any outstanding LRO work
3159 	 */
3160 #if defined(INET6) || defined(INET)
3161 #ifndef __HAIKU__
3162 	tcp_lro_flush_all(&rxq->ifr_lc);
3163 #endif
3164 #endif
3165 	if (avail != 0 || iflib_rxd_avail(ctx, rxq, *cidxp, 1) != 0)
3166 		retval |= IFLIB_RXEOF_MORE;
3167 	return (retval);
3168 err:
3169 	STATE_LOCK(ctx);
3170 	ctx->ifc_flags |= IFC_DO_RESET;
3171 	iflib_admin_intr_deferred(ctx);
3172 	STATE_UNLOCK(ctx);
3173 	return (0);
3174 }
3175 
3176 #define TXD_NOTIFY_COUNT(txq) (((txq)->ift_size / (txq)->ift_update_freq)-1)
3177 static inline qidx_t
3178 txq_max_db_deferred(iflib_txq_t txq, qidx_t in_use)
3179 {
3180 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
3181 	qidx_t minthresh = txq->ift_size / 8;
3182 	if (in_use > 4*minthresh)
3183 		return (notify_count);
3184 	if (in_use > 2*minthresh)
3185 		return (notify_count >> 1);
3186 	if (in_use > minthresh)
3187 		return (notify_count >> 3);
3188 	return (0);
3189 }
3190 
3191 static inline qidx_t
3192 txq_max_rs_deferred(iflib_txq_t txq)
3193 {
3194 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
3195 	qidx_t minthresh = txq->ift_size / 8;
3196 	if (txq->ift_in_use > 4*minthresh)
3197 		return (notify_count);
3198 	if (txq->ift_in_use > 2*minthresh)
3199 		return (notify_count >> 1);
3200 	if (txq->ift_in_use > minthresh)
3201 		return (notify_count >> 2);
3202 	return (2);
3203 }
3204 
3205 #define M_CSUM_FLAGS(m) ((m)->m_pkthdr.csum_flags)
3206 #define M_HAS_VLANTAG(m) (m->m_flags & M_VLANTAG)
3207 
3208 #define TXQ_MAX_DB_DEFERRED(txq, in_use) txq_max_db_deferred((txq), (in_use))
3209 #define TXQ_MAX_RS_DEFERRED(txq) txq_max_rs_deferred(txq)
3210 #define TXQ_MAX_DB_CONSUMED(size) (size >> 4)
3211 
3212 /* forward compatibility for cxgb */
3213 #define FIRST_QSET(ctx) 0
3214 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets)
3215 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets)
3216 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx))
3217 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments))
3218 
3219 /* XXX we should be setting this to something other than zero */
3220 #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh)
3221 #define	MAX_TX_DESC(ctx) MAX((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max, \
3222     (ctx)->ifc_softc_ctx.isc_tx_nsegments)
3223 
3224 static inline bool
3225 iflib_txd_db_check(iflib_txq_t txq, int ring)
3226 {
3227 	if_ctx_t ctx = txq->ift_ctx;
3228 	qidx_t dbval, max;
3229 
3230 	max = TXQ_MAX_DB_DEFERRED(txq, txq->ift_in_use);
3231 
3232 	/* force || threshold exceeded || at the edge of the ring */
3233 	if (ring || (txq->ift_db_pending >= max) || (TXQ_AVAIL(txq) <= MAX_TX_DESC(ctx) + 2)) {
3234 
3235 		/*
3236 		 * 'npending' is used if the card's doorbell is in terms of the number of descriptors
3237 		 * pending flush (BRCM). 'pidx' is used in cases where the card's doorbeel uses the
3238 		 * producer index explicitly (INTC).
3239 		 */
3240 		dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx;
3241 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3242 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
3243 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval);
3244 
3245 		/*
3246 		 * Absent bugs there are zero packets pending so reset pending counts to zero.
3247 		 */
3248 		txq->ift_db_pending = txq->ift_npending = 0;
3249 		return (true);
3250 	}
3251 	return (false);
3252 }
3253 
3254 #ifdef PKT_DEBUG
3255 static void
3256 print_pkt(if_pkt_info_t pi)
3257 {
3258 	printf("pi len:  %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n",
3259 	       pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx);
3260 	printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n",
3261 	       pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag);
3262 	printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n",
3263 	       pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto);
3264 }
3265 #endif
3266 
3267 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO)
3268 #define IS_TX_OFFLOAD4(pi) ((pi)->ipi_csum_flags & (CSUM_IP_TCP | CSUM_IP_TSO))
3269 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO)
3270 #define IS_TX_OFFLOAD6(pi) ((pi)->ipi_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_TSO))
3271 
3272 /**
3273  * Parses out ethernet header information in the given mbuf.
3274  * Returns in pi: ipi_etype (EtherType) and ipi_ehdrlen (Ethernet header length)
3275  *
3276  * This will account for the VLAN header if present.
3277  *
3278  * XXX: This doesn't handle QinQ, which could prevent TX offloads for those
3279  * types of packets.
3280  */
3281 static int
3282 iflib_parse_ether_header(if_pkt_info_t pi, struct mbuf **mp, uint64_t *pullups)
3283 {
3284 	struct ether_vlan_header *eh;
3285 	struct mbuf *m;
3286 
3287 	m = *mp;
3288 	if (__predict_false(m->m_len < sizeof(*eh))) {
3289 		(*pullups)++;
3290 		if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL))
3291 			return (ENOMEM);
3292 	}
3293 	eh = mtod(m, struct ether_vlan_header *);
3294 	if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
3295 		pi->ipi_etype = ntohs(eh->evl_proto);
3296 		pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
3297 	} else {
3298 		pi->ipi_etype = ntohs(eh->evl_encap_proto);
3299 		pi->ipi_ehdrlen = ETHER_HDR_LEN;
3300 	}
3301 	*mp = m;
3302 
3303 	return (0);
3304 }
3305 
3306 /**
3307  * Parse up to the L3 header and extract IPv4/IPv6 header information into pi.
3308  * Currently this information includes: IP ToS value, IP header version/presence
3309  *
3310  * This is missing some checks and doesn't edit the packet content as it goes,
3311  * unlike iflib_parse_header(), in order to keep the amount of code here minimal.
3312  */
3313 static int
3314 iflib_parse_header_partial(if_pkt_info_t pi, struct mbuf **mp, uint64_t *pullups)
3315 {
3316 	struct mbuf *m;
3317 	int err;
3318 
3319 	*pullups = 0;
3320 	m = *mp;
3321 	if (!M_WRITABLE(m)) {
3322 		if ((m = m_dup(m, M_NOWAIT)) == NULL) {
3323 			return (ENOMEM);
3324 		} else {
3325 			m_freem(*mp);
3326 			DBG_COUNTER_INC(tx_frees);
3327 			*mp = m;
3328 		}
3329 	}
3330 
3331 	/* Fills out pi->ipi_etype */
3332 	err = iflib_parse_ether_header(pi, mp, pullups);
3333 	if (err)
3334 		return (err);
3335 	m = *mp;
3336 
3337 	switch (pi->ipi_etype) {
3338 #ifdef INET
3339 	case ETHERTYPE_IP:
3340 	{
3341 		struct mbuf *n;
3342 		struct ip *ip = NULL;
3343 		int miniplen;
3344 
3345 		miniplen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip));
3346 		if (__predict_false(m->m_len < miniplen)) {
3347 			/*
3348 			 * Check for common case where the first mbuf only contains
3349 			 * the Ethernet header
3350 			 */
3351 			if (m->m_len == pi->ipi_ehdrlen) {
3352 				n = m->m_next;
3353 				MPASS(n);
3354 				/* If next mbuf contains at least the minimal IP header, then stop */
3355 				if (n->m_len >= sizeof(*ip)) {
3356 					ip = (struct ip *)n->m_data;
3357 				} else {
3358 					(*pullups)++;
3359 					if (__predict_false((m = m_pullup(m, miniplen)) == NULL))
3360 						return (ENOMEM);
3361 					ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3362 				}
3363 			} else {
3364 				(*pullups)++;
3365 				if (__predict_false((m = m_pullup(m, miniplen)) == NULL))
3366 					return (ENOMEM);
3367 				ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3368 			}
3369 		} else {
3370 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3371 		}
3372 
3373 		/* Have the IPv4 header w/ no options here */
3374 		pi->ipi_ip_hlen = ip->ip_hl << 2;
3375 		pi->ipi_ipproto = ip->ip_p;
3376 		pi->ipi_ip_tos = ip->ip_tos;
3377 		pi->ipi_flags |= IPI_TX_IPV4;
3378 
3379 		break;
3380 	}
3381 #endif
3382 #ifdef INET6
3383 	case ETHERTYPE_IPV6:
3384 	{
3385 		struct ip6_hdr *ip6;
3386 
3387 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
3388 			(*pullups)++;
3389 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
3390 				return (ENOMEM);
3391 		}
3392 		ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
3393 
3394 		/* Have the IPv6 fixed header here */
3395 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
3396 		pi->ipi_ipproto = ip6->ip6_nxt;
3397 		pi->ipi_ip_tos = IPV6_TRAFFIC_CLASS(ip6);
3398 		pi->ipi_flags |= IPI_TX_IPV6;
3399 
3400 		break;
3401 	}
3402 #endif
3403 	default:
3404 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
3405 		pi->ipi_ip_hlen = 0;
3406 		break;
3407 	}
3408 	*mp = m;
3409 
3410 	return (0);
3411 
3412 }
3413 
3414 static int
3415 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp)
3416 {
3417 	if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx;
3418 	struct mbuf *m;
3419 	int err;
3420 
3421 	m = *mp;
3422 	if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) &&
3423 	    M_WRITABLE(m) == 0) {
3424 		if ((m = m_dup(m, M_NOWAIT)) == NULL) {
3425 			return (ENOMEM);
3426 		} else {
3427 			m_freem(*mp);
3428 			DBG_COUNTER_INC(tx_frees);
3429 			*mp = m;
3430 		}
3431 	}
3432 
3433 	/* Fills out pi->ipi_etype */
3434 	err = iflib_parse_ether_header(pi, mp, &txq->ift_pullups);
3435 	if (__predict_false(err))
3436 		return (err);
3437 	m = *mp;
3438 
3439 	switch (pi->ipi_etype) {
3440 #ifdef INET
3441 	case ETHERTYPE_IP:
3442 	{
3443 		struct mbuf *n;
3444 		struct ip *ip = NULL;
3445 		struct tcphdr *th = NULL;
3446 		int minthlen;
3447 
3448 		minthlen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip) + sizeof(*th));
3449 		if (__predict_false(m->m_len < minthlen)) {
3450 			/*
3451 			 * if this code bloat is causing too much of a hit
3452 			 * move it to a separate function and mark it noinline
3453 			 */
3454 			if (m->m_len == pi->ipi_ehdrlen) {
3455 				n = m->m_next;
3456 				MPASS(n);
3457 				if (n->m_len >= sizeof(*ip))  {
3458 					ip = (struct ip *)n->m_data;
3459 					if (n->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3460 						th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3461 				} else {
3462 					txq->ift_pullups++;
3463 					if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
3464 						return (ENOMEM);
3465 					ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3466 				}
3467 			} else {
3468 				txq->ift_pullups++;
3469 				if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
3470 					return (ENOMEM);
3471 				ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3472 				if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3473 					th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3474 			}
3475 		} else {
3476 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3477 			if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3478 				th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3479 		}
3480 		pi->ipi_ip_hlen = ip->ip_hl << 2;
3481 		pi->ipi_ipproto = ip->ip_p;
3482 		pi->ipi_ip_tos = ip->ip_tos;
3483 		pi->ipi_flags |= IPI_TX_IPV4;
3484 
3485 		/* TCP checksum offload may require TCP header length */
3486 		if (IS_TX_OFFLOAD4(pi)) {
3487 			if (__predict_true(pi->ipi_ipproto == IPPROTO_TCP)) {
3488 				if (__predict_false(th == NULL)) {
3489 					txq->ift_pullups++;
3490 					if (__predict_false((m = m_pullup(m, (ip->ip_hl << 2) + sizeof(*th))) == NULL))
3491 						return (ENOMEM);
3492 					th = (struct tcphdr *)((caddr_t)ip + pi->ipi_ip_hlen);
3493 				}
3494 				pi->ipi_tcp_hflags = th->th_flags;
3495 				pi->ipi_tcp_hlen = th->th_off << 2;
3496 				pi->ipi_tcp_seq = th->th_seq;
3497 			}
3498 			if (IS_TSO4(pi)) {
3499 				if (__predict_false(ip->ip_p != IPPROTO_TCP))
3500 					return (ENXIO);
3501 				/*
3502 				 * TSO always requires hardware checksum offload.
3503 				 */
3504 				pi->ipi_csum_flags |= (CSUM_IP_TCP | CSUM_IP);
3505 				th->th_sum = in_pseudo(ip->ip_src.s_addr,
3506 						       ip->ip_dst.s_addr, htons(IPPROTO_TCP));
3507 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3508 				if (sctx->isc_flags & IFLIB_TSO_INIT_IP) {
3509 					ip->ip_sum = 0;
3510 					ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz);
3511 				}
3512 			}
3513 		}
3514 		if ((sctx->isc_flags & IFLIB_NEED_ZERO_CSUM) && (pi->ipi_csum_flags & CSUM_IP))
3515                        ip->ip_sum = 0;
3516 
3517 		break;
3518 	}
3519 #endif
3520 #ifdef INET6
3521 	case ETHERTYPE_IPV6:
3522 	{
3523 		struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
3524 		struct tcphdr *th;
3525 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
3526 
3527 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
3528 			txq->ift_pullups++;
3529 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
3530 				return (ENOMEM);
3531 		}
3532 		th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
3533 
3534 		/* XXX-BZ this will go badly in case of ext hdrs. */
3535 		pi->ipi_ipproto = ip6->ip6_nxt;
3536 		pi->ipi_ip_tos = IPV6_TRAFFIC_CLASS(ip6);
3537 		pi->ipi_flags |= IPI_TX_IPV6;
3538 
3539 		/* TCP checksum offload may require TCP header length */
3540 		if (IS_TX_OFFLOAD6(pi)) {
3541 			if (pi->ipi_ipproto == IPPROTO_TCP) {
3542 				if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) {
3543 					txq->ift_pullups++;
3544 					if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL))
3545 						return (ENOMEM);
3546 				}
3547 				pi->ipi_tcp_hflags = th->th_flags;
3548 				pi->ipi_tcp_hlen = th->th_off << 2;
3549 				pi->ipi_tcp_seq = th->th_seq;
3550 			}
3551 			if (IS_TSO6(pi)) {
3552 				if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP))
3553 					return (ENXIO);
3554 				/*
3555 				 * TSO always requires hardware checksum offload.
3556 				 */
3557 				pi->ipi_csum_flags |= CSUM_IP6_TCP;
3558 				th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0);
3559 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3560 			}
3561 		}
3562 		break;
3563 	}
3564 #endif
3565 	default:
3566 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
3567 		pi->ipi_ip_hlen = 0;
3568 		break;
3569 	}
3570 	*mp = m;
3571 
3572 	return (0);
3573 }
3574 
3575 /*
3576  * If dodgy hardware rejects the scatter gather chain we've handed it
3577  * we'll need to remove the mbuf chain from ifsg_m[] before we can add the
3578  * m_defrag'd mbufs
3579  */
3580 static __noinline struct mbuf *
3581 iflib_remove_mbuf(iflib_txq_t txq)
3582 {
3583 	int ntxd, pidx;
3584 	struct mbuf *m, **ifsd_m;
3585 
3586 	ifsd_m = txq->ift_sds.ifsd_m;
3587 	ntxd = txq->ift_size;
3588 	pidx = txq->ift_pidx & (ntxd - 1);
3589 	ifsd_m = txq->ift_sds.ifsd_m;
3590 	m = ifsd_m[pidx];
3591 	ifsd_m[pidx] = NULL;
3592 	bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[pidx]);
3593 	if (txq->ift_sds.ifsd_tso_map != NULL)
3594 		bus_dmamap_unload(txq->ift_tso_buf_tag,
3595 		    txq->ift_sds.ifsd_tso_map[pidx]);
3596 #if MEMORY_LOGGING
3597 	txq->ift_dequeued++;
3598 #endif
3599 	return (m);
3600 }
3601 
3602 static inline caddr_t
3603 calc_next_txd(iflib_txq_t txq, int cidx, uint8_t qid)
3604 {
3605 	qidx_t size;
3606 	int ntxd;
3607 	caddr_t start, end, cur, next;
3608 
3609 	ntxd = txq->ift_size;
3610 	size = txq->ift_txd_size[qid];
3611 	start = txq->ift_ifdi[qid].idi_vaddr;
3612 
3613 	if (__predict_false(size == 0))
3614 		return (start);
3615 	cur = start + size*cidx;
3616 	end = start + size*ntxd;
3617 	next = CACHE_PTR_NEXT(cur);
3618 	return (next < end ? next : start);
3619 }
3620 
3621 /*
3622  * Pad an mbuf to ensure a minimum ethernet frame size.
3623  * min_frame_size is the frame size (less CRC) to pad the mbuf to
3624  */
3625 static __noinline int
3626 iflib_ether_pad(device_t dev, struct mbuf **m_head, uint16_t min_frame_size)
3627 {
3628 	/*
3629 	 * 18 is enough bytes to pad an ARP packet to 46 bytes, and
3630 	 * and ARP message is the smallest common payload I can think of
3631 	 */
3632 	static char pad[18];	/* just zeros */
3633 	int n;
3634 	struct mbuf *new_head;
3635 
3636 	if (!M_WRITABLE(*m_head)) {
3637 		new_head = m_dup(*m_head, M_NOWAIT);
3638 		if (new_head == NULL) {
3639 			m_freem(*m_head);
3640 			device_printf(dev, "cannot pad short frame, m_dup() failed");
3641 			DBG_COUNTER_INC(encap_pad_mbuf_fail);
3642 			DBG_COUNTER_INC(tx_frees);
3643 			return ENOMEM;
3644 		}
3645 		m_freem(*m_head);
3646 		*m_head = new_head;
3647 	}
3648 
3649 	for (n = min_frame_size - (*m_head)->m_pkthdr.len;
3650 	     n > 0; n -= sizeof(pad))
3651 		if (!m_append(*m_head, min(n, sizeof(pad)), pad))
3652 			break;
3653 
3654 	if (n > 0) {
3655 		m_freem(*m_head);
3656 		device_printf(dev, "cannot pad short frame\n");
3657 		DBG_COUNTER_INC(encap_pad_mbuf_fail);
3658 		DBG_COUNTER_INC(tx_frees);
3659 		return (ENOBUFS);
3660 	}
3661 
3662 	return 0;
3663 }
3664 
3665 static int
3666 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp)
3667 {
3668 	if_ctx_t		ctx;
3669 	if_shared_ctx_t		sctx;
3670 	if_softc_ctx_t		scctx;
3671 	bus_dma_tag_t		buf_tag;
3672 	bus_dma_segment_t	*segs;
3673 	struct mbuf		*m_head, **ifsd_m;
3674 	void			*next_txd;
3675 	bus_dmamap_t		map;
3676 	struct if_pkt_info	pi;
3677 	int remap = 0;
3678 	int err, nsegs, ndesc, max_segs, pidx, cidx, next, ntxd;
3679 
3680 	ctx = txq->ift_ctx;
3681 	sctx = ctx->ifc_sctx;
3682 	scctx = &ctx->ifc_softc_ctx;
3683 	segs = txq->ift_segs;
3684 	ntxd = txq->ift_size;
3685 	m_head = *m_headp;
3686 	map = NULL;
3687 
3688 	/*
3689 	 * If we're doing TSO the next descriptor to clean may be quite far ahead
3690 	 */
3691 	cidx = txq->ift_cidx;
3692 	pidx = txq->ift_pidx;
3693 	if (ctx->ifc_flags & IFC_PREFETCH) {
3694 		next = (cidx + CACHE_PTR_INCREMENT) & (ntxd-1);
3695 		if (!(ctx->ifc_flags & IFLIB_HAS_TXCQ)) {
3696 			next_txd = calc_next_txd(txq, cidx, 0);
3697 			prefetch(next_txd);
3698 		}
3699 
3700 		/* prefetch the next cache line of mbuf pointers and flags */
3701 		prefetch(&txq->ift_sds.ifsd_m[next]);
3702 		prefetch(&txq->ift_sds.ifsd_map[next]);
3703 		next = (cidx + CACHE_LINE_SIZE) & (ntxd-1);
3704 	}
3705 	map = txq->ift_sds.ifsd_map[pidx];
3706 	ifsd_m = txq->ift_sds.ifsd_m;
3707 
3708 	if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
3709 		buf_tag = txq->ift_tso_buf_tag;
3710 		max_segs = scctx->isc_tx_tso_segments_max;
3711 		map = txq->ift_sds.ifsd_tso_map[pidx];
3712 		MPASS(buf_tag != NULL);
3713 		MPASS(max_segs > 0);
3714 	} else {
3715 		buf_tag = txq->ift_buf_tag;
3716 		max_segs = scctx->isc_tx_nsegments;
3717 		map = txq->ift_sds.ifsd_map[pidx];
3718 	}
3719 	if ((sctx->isc_flags & IFLIB_NEED_ETHER_PAD) &&
3720 	    __predict_false(m_head->m_pkthdr.len < scctx->isc_min_frame_size)) {
3721 		err = iflib_ether_pad(ctx->ifc_dev, m_headp, scctx->isc_min_frame_size);
3722 		if (err) {
3723 			DBG_COUNTER_INC(encap_txd_encap_fail);
3724 			return err;
3725 		}
3726 	}
3727 	m_head = *m_headp;
3728 
3729 	pkt_info_zero(&pi);
3730 	pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG|M_BCAST|M_MCAST));
3731 	pi.ipi_pidx = pidx;
3732 	pi.ipi_qsidx = txq->ift_id;
3733 	pi.ipi_len = m_head->m_pkthdr.len;
3734 	pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags;
3735 	pi.ipi_vtag = M_HAS_VLANTAG(m_head) ? m_head->m_pkthdr.ether_vtag : 0;
3736 
3737 	/* deliberate bitwise OR to make one condition */
3738 	if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) {
3739 		if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0)) {
3740 			DBG_COUNTER_INC(encap_txd_encap_fail);
3741 			return (err);
3742 		}
3743 		m_head = *m_headp;
3744 	}
3745 
3746 retry:
3747 	err = bus_dmamap_load_mbuf_sg(buf_tag, map, m_head, segs, &nsegs,
3748 	    BUS_DMA_NOWAIT);
3749 defrag:
3750 	if (__predict_false(err)) {
3751 		switch (err) {
3752 		case EFBIG:
3753 			/* try collapse once and defrag once */
3754 			if (remap == 0) {
3755 				m_head = m_collapse(*m_headp, M_NOWAIT, max_segs);
3756 				/* try defrag if collapsing fails */
3757 				if (m_head == NULL)
3758 					remap++;
3759 			}
3760 			if (remap == 1) {
3761 				txq->ift_mbuf_defrag++;
3762 				m_head = m_defrag(*m_headp, M_NOWAIT);
3763 			}
3764 			/*
3765 			 * remap should never be >1 unless bus_dmamap_load_mbuf_sg
3766 			 * failed to map an mbuf that was run through m_defrag
3767 			 */
3768 			MPASS(remap <= 1);
3769 			if (__predict_false(m_head == NULL || remap > 1))
3770 				goto defrag_failed;
3771 			remap++;
3772 			*m_headp = m_head;
3773 			goto retry;
3774 			break;
3775 		case ENOMEM:
3776 			txq->ift_no_tx_dma_setup++;
3777 			break;
3778 		default:
3779 			txq->ift_no_tx_dma_setup++;
3780 			m_freem(*m_headp);
3781 			DBG_COUNTER_INC(tx_frees);
3782 			*m_headp = NULL;
3783 			break;
3784 		}
3785 		txq->ift_map_failed++;
3786 		DBG_COUNTER_INC(encap_load_mbuf_fail);
3787 		DBG_COUNTER_INC(encap_txd_encap_fail);
3788 		return (err);
3789 	}
3790 	ifsd_m[pidx] = m_head;
3791 	/*
3792 	 * XXX assumes a 1 to 1 relationship between segments and
3793 	 *        descriptors - this does not hold true on all drivers, e.g.
3794 	 *        cxgb
3795 	 */
3796 	if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) {
3797 		txq->ift_no_desc_avail++;
3798 		bus_dmamap_unload(buf_tag, map);
3799 		DBG_COUNTER_INC(encap_txq_avail_fail);
3800 		DBG_COUNTER_INC(encap_txd_encap_fail);
3801 		if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0)
3802 			GROUPTASK_ENQUEUE(&txq->ift_task);
3803 		return (ENOBUFS);
3804 	}
3805 	/*
3806 	 * On Intel cards we can greatly reduce the number of TX interrupts
3807 	 * we see by only setting report status on every Nth descriptor.
3808 	 * However, this also means that the driver will need to keep track
3809 	 * of the descriptors that RS was set on to check them for the DD bit.
3810 	 */
3811 	txq->ift_rs_pending += nsegs + 1;
3812 	if (txq->ift_rs_pending > TXQ_MAX_RS_DEFERRED(txq) ||
3813 	     iflib_no_tx_batch || (TXQ_AVAIL(txq) - nsegs) <= MAX_TX_DESC(ctx) + 2) {
3814 		pi.ipi_flags |= IPI_TX_INTR;
3815 		txq->ift_rs_pending = 0;
3816 	}
3817 
3818 	pi.ipi_segs = segs;
3819 	pi.ipi_nsegs = nsegs;
3820 
3821 	MPASS(pidx >= 0 && pidx < txq->ift_size);
3822 #ifdef PKT_DEBUG
3823 	print_pkt(&pi);
3824 #endif
3825 	if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) {
3826 		bus_dmamap_sync(buf_tag, map, BUS_DMASYNC_PREWRITE);
3827 		DBG_COUNTER_INC(tx_encap);
3828 		MPASS(pi.ipi_new_pidx < txq->ift_size);
3829 
3830 		ndesc = pi.ipi_new_pidx - pi.ipi_pidx;
3831 		if (pi.ipi_new_pidx < pi.ipi_pidx) {
3832 			ndesc += txq->ift_size;
3833 			txq->ift_gen = 1;
3834 		}
3835 		/*
3836 		 * drivers can need as many as
3837 		 * two sentinels
3838 		 */
3839 		MPASS(ndesc <= pi.ipi_nsegs + 2);
3840 		MPASS(pi.ipi_new_pidx != pidx);
3841 		MPASS(ndesc > 0);
3842 		txq->ift_in_use += ndesc;
3843 		txq->ift_db_pending += ndesc;
3844 
3845 		/*
3846 		 * We update the last software descriptor again here because there may
3847 		 * be a sentinel and/or there may be more mbufs than segments
3848 		 */
3849 		txq->ift_pidx = pi.ipi_new_pidx;
3850 		txq->ift_npending += pi.ipi_ndescs;
3851 	} else {
3852 		*m_headp = m_head = iflib_remove_mbuf(txq);
3853 		if (err == EFBIG) {
3854 			txq->ift_txd_encap_efbig++;
3855 			if (remap < 2) {
3856 				remap = 1;
3857 				goto defrag;
3858 			}
3859 		}
3860 		goto defrag_failed;
3861 	}
3862 	/*
3863 	 * err can't possibly be non-zero here, so we don't neet to test it
3864 	 * to see if we need to DBG_COUNTER_INC(encap_txd_encap_fail).
3865 	 */
3866 	return (err);
3867 
3868 defrag_failed:
3869 	txq->ift_mbuf_defrag_failed++;
3870 	txq->ift_map_failed++;
3871 	m_freem(*m_headp);
3872 	DBG_COUNTER_INC(tx_frees);
3873 	*m_headp = NULL;
3874 	DBG_COUNTER_INC(encap_txd_encap_fail);
3875 	return (ENOMEM);
3876 }
3877 
3878 static void
3879 iflib_tx_desc_free(iflib_txq_t txq, int n)
3880 {
3881 	uint32_t qsize, cidx, mask, gen;
3882 	struct mbuf *m, **ifsd_m;
3883 	bool do_prefetch;
3884 
3885 	cidx = txq->ift_cidx;
3886 	gen = txq->ift_gen;
3887 	qsize = txq->ift_size;
3888 	mask = qsize-1;
3889 	ifsd_m = txq->ift_sds.ifsd_m;
3890 	do_prefetch = (txq->ift_ctx->ifc_flags & IFC_PREFETCH);
3891 
3892 	while (n-- > 0) {
3893 		if (do_prefetch) {
3894 			prefetch(ifsd_m[(cidx + 3) & mask]);
3895 			prefetch(ifsd_m[(cidx + 4) & mask]);
3896 		}
3897 		if ((m = ifsd_m[cidx]) != NULL) {
3898 			prefetch(&ifsd_m[(cidx + CACHE_PTR_INCREMENT) & mask]);
3899 			if (m->m_pkthdr.csum_flags & CSUM_TSO) {
3900 				bus_dmamap_sync(txq->ift_tso_buf_tag,
3901 				    txq->ift_sds.ifsd_tso_map[cidx],
3902 				    BUS_DMASYNC_POSTWRITE);
3903 				bus_dmamap_unload(txq->ift_tso_buf_tag,
3904 				    txq->ift_sds.ifsd_tso_map[cidx]);
3905 			} else {
3906 				bus_dmamap_sync(txq->ift_buf_tag,
3907 				    txq->ift_sds.ifsd_map[cidx],
3908 				    BUS_DMASYNC_POSTWRITE);
3909 				bus_dmamap_unload(txq->ift_buf_tag,
3910 				    txq->ift_sds.ifsd_map[cidx]);
3911 			}
3912 			/* XXX we don't support any drivers that batch packets yet */
3913 			MPASS(m->m_nextpkt == NULL);
3914 			m_freem(m);
3915 			ifsd_m[cidx] = NULL;
3916 #if MEMORY_LOGGING
3917 			txq->ift_dequeued++;
3918 #endif
3919 			DBG_COUNTER_INC(tx_frees);
3920 		}
3921 		if (__predict_false(++cidx == qsize)) {
3922 			cidx = 0;
3923 			gen = 0;
3924 		}
3925 	}
3926 	txq->ift_cidx = cidx;
3927 	txq->ift_gen = gen;
3928 }
3929 
3930 static __inline int
3931 iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh)
3932 {
3933 	int reclaim;
3934 	if_ctx_t ctx = txq->ift_ctx;
3935 
3936 	KASSERT(thresh >= 0, ("invalid threshold to reclaim"));
3937 	MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size);
3938 
3939 	/*
3940 	 * Need a rate-limiting check so that this isn't called every time
3941 	 */
3942 	iflib_tx_credits_update(ctx, txq);
3943 	reclaim = DESC_RECLAIMABLE(txq);
3944 
3945 	if (reclaim <= thresh /* + MAX_TX_DESC(txq->ift_ctx) */) {
3946 #ifdef INVARIANTS
3947 		if (iflib_verbose_debug) {
3948 			printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __FUNCTION__,
3949 			       txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments,
3950 			       reclaim, thresh);
3951 		}
3952 #endif
3953 		return (0);
3954 	}
3955 	iflib_tx_desc_free(txq, reclaim);
3956 	txq->ift_cleaned += reclaim;
3957 	txq->ift_in_use -= reclaim;
3958 
3959 	return (reclaim);
3960 }
3961 
3962 static struct mbuf **
3963 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset, int remaining)
3964 {
3965 	int next, size;
3966 	struct mbuf **items;
3967 
3968 	size = r->size;
3969 	next = (cidx + CACHE_PTR_INCREMENT) & (size-1);
3970 	items = __DEVOLATILE(struct mbuf **, &r->items[0]);
3971 
3972 	prefetch(items[(cidx + offset) & (size-1)]);
3973 	if (remaining > 1) {
3974 		prefetch2cachelines(&items[next]);
3975 		prefetch2cachelines(items[(cidx + offset + 1) & (size-1)]);
3976 		prefetch2cachelines(items[(cidx + offset + 2) & (size-1)]);
3977 		prefetch2cachelines(items[(cidx + offset + 3) & (size-1)]);
3978 	}
3979 	return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (size-1)]));
3980 }
3981 
3982 static void
3983 iflib_txq_check_drain(iflib_txq_t txq, int budget)
3984 {
3985 
3986 	ifmp_ring_check_drainage(txq->ift_br, budget);
3987 }
3988 
3989 static uint32_t
3990 iflib_txq_can_drain(struct ifmp_ring *r)
3991 {
3992 	iflib_txq_t txq = r->cookie;
3993 	if_ctx_t ctx = txq->ift_ctx;
3994 
3995 	if (TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2)
3996 		return (1);
3997 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3998 	    BUS_DMASYNC_POSTREAD);
3999 	return (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id,
4000 	    false));
4001 }
4002 
4003 static uint32_t
4004 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
4005 {
4006 	iflib_txq_t txq = r->cookie;
4007 	if_ctx_t ctx = txq->ift_ctx;
4008 	if_t ifp = ctx->ifc_ifp;
4009 	struct mbuf *m, **mp;
4010 	int avail, bytes_sent, skipped, count, err, i;
4011 	int mcast_sent, pkt_sent, reclaimed;
4012 	bool do_prefetch, rang, ring;
4013 
4014 	if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) ||
4015 			    !LINK_ACTIVE(ctx))) {
4016 		DBG_COUNTER_INC(txq_drain_notready);
4017 		return (0);
4018 	}
4019 	reclaimed = iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
4020 	rang = iflib_txd_db_check(txq, reclaimed && txq->ift_db_pending);
4021 	avail = IDXDIFF(pidx, cidx, r->size);
4022 
4023 	if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) {
4024 		/*
4025 		 * The driver is unloading so we need to free all pending packets.
4026 		 */
4027 		DBG_COUNTER_INC(txq_drain_flushing);
4028 		for (i = 0; i < avail; i++) {
4029 			if (__predict_true(r->items[(cidx + i) & (r->size-1)] != (void *)txq))
4030 				m_freem(r->items[(cidx + i) & (r->size-1)]);
4031 			r->items[(cidx + i) & (r->size-1)] = NULL;
4032 		}
4033 		return (avail);
4034 	}
4035 
4036 	if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
4037 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
4038 		CALLOUT_LOCK(txq);
4039 		callout_stop(&txq->ift_timer);
4040 		CALLOUT_UNLOCK(txq);
4041 		DBG_COUNTER_INC(txq_drain_oactive);
4042 		return (0);
4043 	}
4044 
4045 	/*
4046 	 * If we've reclaimed any packets this queue cannot be hung.
4047 	 */
4048 	if (reclaimed)
4049 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
4050 	skipped = mcast_sent = bytes_sent = pkt_sent = 0;
4051 	count = MIN(avail, TX_BATCH_SIZE);
4052 #ifdef INVARIANTS
4053 	if (iflib_verbose_debug)
4054 		printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __FUNCTION__,
4055 		       avail, ctx->ifc_flags, TXQ_AVAIL(txq));
4056 #endif
4057 	do_prefetch = (ctx->ifc_flags & IFC_PREFETCH);
4058 	err = 0;
4059 	for (i = 0; i < count && TXQ_AVAIL(txq) >= MAX_TX_DESC(ctx) + 2; i++) {
4060 		int rem = do_prefetch ? count - i : 0;
4061 
4062 		mp = _ring_peek_one(r, cidx, i, rem);
4063 		MPASS(mp != NULL && *mp != NULL);
4064 
4065 		/*
4066 		 * Completion interrupts will use the address of the txq
4067 		 * as a sentinel to enqueue _something_ in order to acquire
4068 		 * the lock on the mp_ring (there's no direct lock call).
4069 		 * We obviously whave to check for these sentinel cases
4070 		 * and skip them.
4071 		 */
4072 		if (__predict_false(*mp == (struct mbuf *)txq)) {
4073 			skipped++;
4074 			continue;
4075 		}
4076 		err = iflib_encap(txq, mp);
4077 		if (__predict_false(err)) {
4078 			/* no room - bail out */
4079 			if (err == ENOBUFS)
4080 				break;
4081 			skipped++;
4082 			/* we can't send this packet - skip it */
4083 			continue;
4084 		}
4085 		pkt_sent++;
4086 		m = *mp;
4087 		DBG_COUNTER_INC(tx_sent);
4088 		bytes_sent += m->m_pkthdr.len;
4089 		mcast_sent += !!(m->m_flags & M_MCAST);
4090 
4091 		if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING)))
4092 			break;
4093 		ETHER_BPF_MTAP(ifp, m);
4094 		rang = iflib_txd_db_check(txq, false);
4095 	}
4096 
4097 	/* deliberate use of bitwise or to avoid gratuitous short-circuit */
4098 	ring = rang ? false  : (iflib_min_tx_latency | err);
4099 	iflib_txd_db_check(txq, ring);
4100 	if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
4101 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
4102 	if (mcast_sent)
4103 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
4104 #ifdef INVARIANTS
4105 	if (iflib_verbose_debug)
4106 		printf("consumed=%d\n", skipped + pkt_sent);
4107 #endif
4108 	return (skipped + pkt_sent);
4109 }
4110 
4111 static uint32_t
4112 iflib_txq_drain_always(struct ifmp_ring *r)
4113 {
4114 	return (1);
4115 }
4116 
4117 static uint32_t
4118 iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
4119 {
4120 	int i, avail;
4121 	struct mbuf **mp;
4122 	iflib_txq_t txq;
4123 
4124 	txq = r->cookie;
4125 
4126 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
4127 	CALLOUT_LOCK(txq);
4128 	callout_stop(&txq->ift_timer);
4129 	CALLOUT_UNLOCK(txq);
4130 
4131 	avail = IDXDIFF(pidx, cidx, r->size);
4132 	for (i = 0; i < avail; i++) {
4133 		mp = _ring_peek_one(r, cidx, i, avail - i);
4134 		if (__predict_false(*mp == (struct mbuf *)txq))
4135 			continue;
4136 		m_freem(*mp);
4137 		DBG_COUNTER_INC(tx_frees);
4138 	}
4139 	MPASS(ifmp_ring_is_stalled(r) == 0);
4140 	return (avail);
4141 }
4142 
4143 static void
4144 iflib_ifmp_purge(iflib_txq_t txq)
4145 {
4146 	struct ifmp_ring *r;
4147 
4148 	r = txq->ift_br;
4149 	r->drain = iflib_txq_drain_free;
4150 	r->can_drain = iflib_txq_drain_always;
4151 
4152 	ifmp_ring_check_drainage(r, r->size);
4153 
4154 	r->drain = iflib_txq_drain;
4155 	r->can_drain = iflib_txq_can_drain;
4156 }
4157 
4158 static void
4159 _task_fn_tx(void *context)
4160 {
4161 	iflib_txq_t txq = context;
4162 	if_ctx_t ctx = txq->ift_ctx;
4163 	if_t ifp = ctx->ifc_ifp;
4164 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
4165 
4166 #ifdef IFLIB_DIAGNOSTICS
4167 	txq->ift_cpu_exec_count[curcpu]++;
4168 #endif
4169 	if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
4170 		return;
4171 #ifdef DEV_NETMAP
4172 	if ((if_getcapenable(ifp) & IFCAP_NETMAP) &&
4173 	    netmap_tx_irq(ifp, txq->ift_id))
4174 		goto skip_ifmp;
4175 #endif
4176 #ifdef ALTQ
4177 	if (if_altq_is_enabled(ifp))
4178 		iflib_altq_if_start(ifp);
4179 #endif
4180 	if (txq->ift_db_pending)
4181 		ifmp_ring_enqueue(txq->ift_br, (void **)&txq, 1, TX_BATCH_SIZE, abdicate);
4182 	else if (!abdicate)
4183 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
4184 	/*
4185 	 * When abdicating, we always need to check drainage, not just when we don't enqueue
4186 	 */
4187 	if (abdicate)
4188 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
4189 #ifdef DEV_NETMAP
4190 skip_ifmp:
4191 #endif
4192 	if (ctx->ifc_flags & IFC_LEGACY)
4193 		IFDI_INTR_ENABLE(ctx);
4194 	else
4195 		IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
4196 }
4197 
4198 static void
4199 _task_fn_rx(void *context)
4200 {
4201 	iflib_rxq_t rxq = context;
4202 	if_ctx_t ctx = rxq->ifr_ctx;
4203 	uint8_t more;
4204 	uint16_t budget;
4205 #ifdef DEV_NETMAP
4206 	u_int work = 0;
4207 	int nmirq;
4208 #endif
4209 
4210 #ifdef IFLIB_DIAGNOSTICS
4211 	rxq->ifr_cpu_exec_count[curcpu]++;
4212 #endif
4213 	DBG_COUNTER_INC(task_fn_rxs);
4214 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
4215 		return;
4216 #ifdef DEV_NETMAP
4217 	nmirq = netmap_rx_irq(ctx->ifc_ifp, rxq->ifr_id, &work);
4218 	if (nmirq != NM_IRQ_PASS) {
4219 		more = (nmirq == NM_IRQ_RESCHED) ? IFLIB_RXEOF_MORE : 0;
4220 		goto skip_rxeof;
4221 	}
4222 #endif
4223 	budget = ctx->ifc_sysctl_rx_budget;
4224 	if (budget == 0)
4225 		budget = 16;	/* XXX */
4226 	more = iflib_rxeof(rxq, budget);
4227 #ifdef DEV_NETMAP
4228 skip_rxeof:
4229 #endif
4230 	if ((more & IFLIB_RXEOF_MORE) == 0) {
4231 		if (ctx->ifc_flags & IFC_LEGACY)
4232 			IFDI_INTR_ENABLE(ctx);
4233 		else
4234 			IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
4235 		DBG_COUNTER_INC(rx_intr_enables);
4236 	}
4237 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
4238 		return;
4239 
4240 	if (more & IFLIB_RXEOF_MORE)
4241 		GROUPTASK_ENQUEUE(&rxq->ifr_task);
4242 	else if (more & IFLIB_RXEOF_EMPTY)
4243 #ifndef __HAIKU__
4244 		callout_reset_curcpu(&rxq->ifr_watchdog, 1, &_task_fn_rx_watchdog, rxq);
4245 #else
4246 		callout_reset(&rxq->ifr_watchdog, 1, &_task_fn_rx_watchdog, rxq);
4247 #endif
4248 }
4249 
4250 static void
4251 _task_fn_admin(void *context)
4252 {
4253 	if_ctx_t ctx = context;
4254 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
4255 	iflib_txq_t txq;
4256 	int i;
4257 	bool oactive, running, do_reset, do_watchdog, in_detach;
4258 
4259 	STATE_LOCK(ctx);
4260 	running = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING);
4261 	oactive = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE);
4262 	do_reset = (ctx->ifc_flags & IFC_DO_RESET);
4263 	do_watchdog = (ctx->ifc_flags & IFC_DO_WATCHDOG);
4264 	in_detach = (ctx->ifc_flags & IFC_IN_DETACH);
4265 	ctx->ifc_flags &= ~(IFC_DO_RESET|IFC_DO_WATCHDOG);
4266 	STATE_UNLOCK(ctx);
4267 
4268 	if ((!running && !oactive) && !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
4269 		return;
4270 	if (in_detach)
4271 		return;
4272 
4273 	CTX_LOCK(ctx);
4274 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
4275 		CALLOUT_LOCK(txq);
4276 		callout_stop(&txq->ift_timer);
4277 		CALLOUT_UNLOCK(txq);
4278 	}
4279 #ifndef __HAIKU__
4280 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_ADMINCQ)
4281 		IFDI_ADMIN_COMPLETION_HANDLE(ctx);
4282 #endif
4283 	if (do_watchdog) {
4284 		ctx->ifc_watchdog_events++;
4285 		IFDI_WATCHDOG_RESET(ctx);
4286 	}
4287 	IFDI_UPDATE_ADMIN_STATUS(ctx);
4288 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
4289 		callout_reset_on(&txq->ift_timer, iflib_timer_default, iflib_timer, txq,
4290 		    txq->ift_timer.c_cpu);
4291 	}
4292 	IFDI_LINK_INTR_ENABLE(ctx);
4293 	if (do_reset)
4294 		iflib_if_init_locked(ctx);
4295 	CTX_UNLOCK(ctx);
4296 
4297 	if (LINK_ACTIVE(ctx) == 0)
4298 		return;
4299 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
4300 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
4301 }
4302 
4303 static void
4304 _task_fn_iov(void *context)
4305 {
4306 	if_ctx_t ctx = context;
4307 
4308 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING) &&
4309 	    !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
4310 		return;
4311 
4312 	CTX_LOCK(ctx);
4313 	IFDI_VFLR_HANDLE(ctx);
4314 	CTX_UNLOCK(ctx);
4315 }
4316 
4317 static int
4318 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)
4319 {
4320 	int err;
4321 	if_int_delay_info_t info;
4322 	if_ctx_t ctx;
4323 
4324 	info = (if_int_delay_info_t)arg1;
4325 	ctx = info->iidi_ctx;
4326 	info->iidi_req = req;
4327 	info->iidi_oidp = oidp;
4328 	CTX_LOCK(ctx);
4329 	err = IFDI_SYSCTL_INT_DELAY(ctx, info);
4330 	CTX_UNLOCK(ctx);
4331 	return (err);
4332 }
4333 
4334 /*********************************************************************
4335  *
4336  *  IFNET FUNCTIONS
4337  *
4338  **********************************************************************/
4339 
4340 static void
4341 iflib_if_init_locked(if_ctx_t ctx)
4342 {
4343 	iflib_stop(ctx);
4344 	iflib_init_locked(ctx);
4345 }
4346 
4347 static void
4348 iflib_if_init(void *arg)
4349 {
4350 	if_ctx_t ctx = arg;
4351 
4352 	CTX_LOCK(ctx);
4353 	iflib_if_init_locked(ctx);
4354 	CTX_UNLOCK(ctx);
4355 }
4356 
4357 static int
4358 iflib_if_transmit(if_t ifp, struct mbuf *m)
4359 {
4360 	if_ctx_t ctx = if_getsoftc(ifp);
4361 	iflib_txq_t txq;
4362 	int err, qidx;
4363 	int abdicate;
4364 
4365 	if (__predict_false((if_getdrvflags(ifp) & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) {
4366 		DBG_COUNTER_INC(tx_frees);
4367 		m_freem(m);
4368 		return (ENETDOWN);
4369 	}
4370 
4371 	MPASS(m->m_nextpkt == NULL);
4372 	/* ALTQ-enabled interfaces always use queue 0. */
4373 	qidx = 0;
4374 	/* Use driver-supplied queue selection method if it exists */
4375 	if (ctx->isc_txq_select_v2) {
4376 		struct if_pkt_info pi;
4377 		uint64_t early_pullups = 0;
4378 		pkt_info_zero(&pi);
4379 
4380 		err = iflib_parse_header_partial(&pi, &m, &early_pullups);
4381 		if (__predict_false(err != 0)) {
4382 			/* Assign pullups for bad pkts to default queue */
4383 			ctx->ifc_txqs[0].ift_pullups += early_pullups;
4384 			DBG_COUNTER_INC(encap_txd_encap_fail);
4385 			return (err);
4386 		}
4387 		/* Let driver make queueing decision */
4388 		qidx = ctx->isc_txq_select_v2(ctx->ifc_softc, m, &pi);
4389 		ctx->ifc_txqs[qidx].ift_pullups += early_pullups;
4390 	}
4391 	/* Backwards compatibility w/ simpler queue select */
4392 	else if (ctx->isc_txq_select)
4393 		qidx = ctx->isc_txq_select(ctx->ifc_softc, m);
4394 	/* If not, use iflib's standard method */
4395 	else if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m) && !if_altq_is_enabled(ifp))
4396 		qidx = QIDX(ctx, m);
4397 
4398 	/* Set TX queue */
4399 	txq = &ctx->ifc_txqs[qidx];
4400 
4401 #ifdef DRIVER_BACKPRESSURE
4402 	if (txq->ift_closed) {
4403 		while (m != NULL) {
4404 			next = m->m_nextpkt;
4405 			m->m_nextpkt = NULL;
4406 			m_freem(m);
4407 			DBG_COUNTER_INC(tx_frees);
4408 			m = next;
4409 		}
4410 		return (ENOBUFS);
4411 	}
4412 #endif
4413 #ifdef notyet
4414 	qidx = count = 0;
4415 	mp = marr;
4416 	next = m;
4417 	do {
4418 		count++;
4419 		next = next->m_nextpkt;
4420 	} while (next != NULL);
4421 
4422 	if (count > nitems(marr))
4423 		if ((mp = malloc(count*sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) {
4424 			/* XXX check nextpkt */
4425 			m_freem(m);
4426 			/* XXX simplify for now */
4427 			DBG_COUNTER_INC(tx_frees);
4428 			return (ENOBUFS);
4429 		}
4430 	for (next = m, i = 0; next != NULL; i++) {
4431 		mp[i] = next;
4432 		next = next->m_nextpkt;
4433 		mp[i]->m_nextpkt = NULL;
4434 	}
4435 #endif
4436 	DBG_COUNTER_INC(tx_seen);
4437 	abdicate = ctx->ifc_sysctl_tx_abdicate;
4438 
4439 	err = ifmp_ring_enqueue(txq->ift_br, (void **)&m, 1, TX_BATCH_SIZE, abdicate);
4440 
4441 	if (abdicate)
4442 		GROUPTASK_ENQUEUE(&txq->ift_task);
4443  	if (err) {
4444 		if (!abdicate)
4445 			GROUPTASK_ENQUEUE(&txq->ift_task);
4446 		/* support forthcoming later */
4447 #ifdef DRIVER_BACKPRESSURE
4448 		txq->ift_closed = TRUE;
4449 #endif
4450 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
4451 		m_freem(m);
4452 		DBG_COUNTER_INC(tx_frees);
4453 	}
4454 
4455 	return (err);
4456 }
4457 
4458 #ifdef ALTQ
4459 /*
4460  * The overall approach to integrating iflib with ALTQ is to continue to use
4461  * the iflib mp_ring machinery between the ALTQ queue(s) and the hardware
4462  * ring.  Technically, when using ALTQ, queueing to an intermediate mp_ring
4463  * is redundant/unnecessary, but doing so minimizes the amount of
4464  * ALTQ-specific code required in iflib.  It is assumed that the overhead of
4465  * redundantly queueing to an intermediate mp_ring is swamped by the
4466  * performance limitations inherent in using ALTQ.
4467  *
4468  * When ALTQ support is compiled in, all iflib drivers will use a transmit
4469  * routine, iflib_altq_if_transmit(), that checks if ALTQ is enabled for the
4470  * given interface.  If ALTQ is enabled for an interface, then all
4471  * transmitted packets for that interface will be submitted to the ALTQ
4472  * subsystem via IFQ_ENQUEUE().  We don't use the legacy if_transmit()
4473  * implementation because it uses IFQ_HANDOFF(), which will duplicatively
4474  * update stats that the iflib machinery handles, and which is sensitve to
4475  * the disused IFF_DRV_OACTIVE flag.  Additionally, iflib_altq_if_start()
4476  * will be installed as the start routine for use by ALTQ facilities that
4477  * need to trigger queue drains on a scheduled basis.
4478  *
4479  */
4480 static void
4481 iflib_altq_if_start(if_t ifp)
4482 {
4483 	struct ifaltq *ifq = &ifp->if_snd; /* XXX - DRVAPI */
4484 	struct mbuf *m;
4485 
4486 	IFQ_LOCK(ifq);
4487 	IFQ_DEQUEUE_NOLOCK(ifq, m);
4488 	while (m != NULL) {
4489 		iflib_if_transmit(ifp, m);
4490 		IFQ_DEQUEUE_NOLOCK(ifq, m);
4491 	}
4492 	IFQ_UNLOCK(ifq);
4493 }
4494 
4495 static int
4496 iflib_altq_if_transmit(if_t ifp, struct mbuf *m)
4497 {
4498 	int err;
4499 
4500 	if (if_altq_is_enabled(ifp)) {
4501 		IFQ_ENQUEUE(&ifp->if_snd, m, err); /* XXX - DRVAPI */
4502 		if (err == 0)
4503 			iflib_altq_if_start(ifp);
4504 	} else
4505 		err = iflib_if_transmit(ifp, m);
4506 
4507 	return (err);
4508 }
4509 #endif /* ALTQ */
4510 
4511 static void
4512 iflib_if_qflush(if_t ifp)
4513 {
4514 	if_ctx_t ctx = if_getsoftc(ifp);
4515 	iflib_txq_t txq = ctx->ifc_txqs;
4516 	int i;
4517 
4518 	STATE_LOCK(ctx);
4519 	ctx->ifc_flags |= IFC_QFLUSH;
4520 	STATE_UNLOCK(ctx);
4521 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
4522 		while (!(ifmp_ring_is_idle(txq->ift_br) || ifmp_ring_is_stalled(txq->ift_br)))
4523 			iflib_txq_check_drain(txq, 0);
4524 	STATE_LOCK(ctx);
4525 	ctx->ifc_flags &= ~IFC_QFLUSH;
4526 	STATE_UNLOCK(ctx);
4527 
4528 	/*
4529 	 * When ALTQ is enabled, this will also take care of purging the
4530 	 * ALTQ queue(s).
4531 	 */
4532 #ifdef __HAIKU__
4533 	ifp->if_qflush(ifp);
4534 #else
4535 	if_qflush(ifp);
4536 #endif
4537 }
4538 
4539 #define IFCAP_FLAGS (IFCAP_HWCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \
4540 		     IFCAP_TSO | IFCAP_VLAN_HWTAGGING | IFCAP_HWSTATS | \
4541 		     IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | \
4542 		     IFCAP_VLAN_HWTSO | IFCAP_VLAN_HWCSUM | IFCAP_MEXTPG)
4543 
4544 static int
4545 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data)
4546 {
4547 	if_ctx_t ctx = if_getsoftc(ifp);
4548 	struct ifreq	*ifr = (struct ifreq *)data;
4549 #if defined(INET) || defined(INET6)
4550 	struct ifaddr	*ifa = (struct ifaddr *)data;
4551 #endif
4552 	bool		avoid_reset = false;
4553 	int		err = 0, reinit = 0, bits;
4554 
4555 	switch (command) {
4556 	case SIOCSIFADDR:
4557 #ifdef INET
4558 		if (ifa->ifa_addr->sa_family == AF_INET)
4559 			avoid_reset = true;
4560 #endif
4561 #ifdef INET6
4562 		if (ifa->ifa_addr->sa_family == AF_INET6)
4563 			avoid_reset = true;
4564 #endif
4565 		/*
4566 		** Calling init results in link renegotiation,
4567 		** so we avoid doing it when possible.
4568 		*/
4569 		if (avoid_reset) {
4570 			if_setflagbits(ifp, IFF_UP,0);
4571 			if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
4572 				reinit = 1;
4573 #ifdef INET
4574 			if (!(if_getflags(ifp) & IFF_NOARP))
4575 				arp_ifinit(ifp, ifa);
4576 #endif
4577 		} else
4578 			err = ether_ioctl(ifp, command, data);
4579 		break;
4580 	case SIOCSIFMTU:
4581 		CTX_LOCK(ctx);
4582 		if (ifr->ifr_mtu == if_getmtu(ifp)) {
4583 			CTX_UNLOCK(ctx);
4584 			break;
4585 		}
4586 		bits = if_getdrvflags(ifp);
4587 		/* stop the driver and free any clusters before proceeding */
4588 		iflib_stop(ctx);
4589 
4590 		if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) {
4591 			STATE_LOCK(ctx);
4592 			if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size)
4593 				ctx->ifc_flags |= IFC_MULTISEG;
4594 			else
4595 				ctx->ifc_flags &= ~IFC_MULTISEG;
4596 			STATE_UNLOCK(ctx);
4597 			err = if_setmtu(ifp, ifr->ifr_mtu);
4598 		}
4599 		iflib_init_locked(ctx);
4600 		STATE_LOCK(ctx);
4601 		if_setdrvflags(ifp, bits);
4602 		STATE_UNLOCK(ctx);
4603 		CTX_UNLOCK(ctx);
4604 		break;
4605 	case SIOCSIFFLAGS:
4606 		CTX_LOCK(ctx);
4607 		if (if_getflags(ifp) & IFF_UP) {
4608 			if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4609 				if ((if_getflags(ifp) ^ ctx->ifc_if_flags) &
4610 				    (IFF_PROMISC | IFF_ALLMULTI)) {
4611 					CTX_UNLOCK(ctx);
4612 					err = IFDI_PROMISC_SET(ctx, if_getflags(ifp));
4613 					CTX_LOCK(ctx);
4614 				}
4615 			} else
4616 				reinit = 1;
4617 		} else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4618 			iflib_stop(ctx);
4619 		}
4620 		ctx->ifc_if_flags = if_getflags(ifp);
4621 		CTX_UNLOCK(ctx);
4622 		break;
4623 	case SIOCADDMULTI:
4624 	case SIOCDELMULTI:
4625 		if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4626 			CTX_LOCK(ctx);
4627 			IFDI_INTR_DISABLE(ctx);
4628 			IFDI_MULTI_SET(ctx);
4629 			IFDI_INTR_ENABLE(ctx);
4630 			CTX_UNLOCK(ctx);
4631 		}
4632 		break;
4633 	case SIOCSIFMEDIA:
4634 		CTX_LOCK(ctx);
4635 		IFDI_MEDIA_SET(ctx);
4636 		CTX_UNLOCK(ctx);
4637 		/* FALLTHROUGH */
4638 	case SIOCGIFMEDIA:
4639 #ifndef __HAIKU__
4640 	case SIOCGIFXMEDIA:
4641 #endif
4642 		err = ifmedia_ioctl(ifp, ifr, ctx->ifc_mediap, command);
4643 		break;
4644 #ifndef __HAIKU__
4645 	case SIOCGI2C:
4646 	{
4647 		struct ifi2creq i2c;
4648 
4649 		err = copyin(ifr_data_get_ptr(ifr), &i2c, sizeof(i2c));
4650 		if (err != 0)
4651 			break;
4652 		if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) {
4653 			err = EINVAL;
4654 			break;
4655 		}
4656 		if (i2c.len > sizeof(i2c.data)) {
4657 			err = EINVAL;
4658 			break;
4659 		}
4660 
4661 		if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0)
4662 			err = copyout(&i2c, ifr_data_get_ptr(ifr),
4663 			    sizeof(i2c));
4664 		break;
4665 	}
4666 #endif
4667 	case SIOCSIFCAP:
4668 	{
4669 		int mask, setmask, oldmask;
4670 
4671 		oldmask = if_getcapenable(ifp);
4672 		mask = ifr->ifr_reqcap ^ oldmask;
4673 		mask &= ctx->ifc_softc_ctx.isc_capabilities | IFCAP_MEXTPG;
4674 		setmask = 0;
4675 #ifdef TCP_OFFLOAD
4676 		setmask |= mask & (IFCAP_TOE4|IFCAP_TOE6);
4677 #endif
4678 		setmask |= (mask & IFCAP_FLAGS);
4679 		setmask |= (mask & IFCAP_WOL);
4680 
4681 		/*
4682 		 * If any RX csum has changed, change all the ones that
4683 		 * are supported by the driver.
4684 		 */
4685 		if (setmask & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6)) {
4686 			setmask |= ctx->ifc_softc_ctx.isc_capabilities &
4687 			    (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6);
4688 		}
4689 
4690 		/*
4691 		 * want to ensure that traffic has stopped before we change any of the flags
4692 		 */
4693 		if (setmask) {
4694 			CTX_LOCK(ctx);
4695 			bits = if_getdrvflags(ifp);
4696 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4697 				iflib_stop(ctx);
4698 			STATE_LOCK(ctx);
4699 			if_togglecapenable(ifp, setmask);
4700 			ctx->ifc_softc_ctx.isc_capenable ^= setmask;
4701 			STATE_UNLOCK(ctx);
4702 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4703 				iflib_init_locked(ctx);
4704 			STATE_LOCK(ctx);
4705 			if_setdrvflags(ifp, bits);
4706 			STATE_UNLOCK(ctx);
4707 			CTX_UNLOCK(ctx);
4708 		}
4709 		if_vlancap(ifp);
4710 		break;
4711 	}
4712 	case SIOCGPRIVATE_0:
4713 	case SIOCSDRVSPEC:
4714 	case SIOCGDRVSPEC:
4715 		CTX_LOCK(ctx);
4716 		err = IFDI_PRIV_IOCTL(ctx, command, data);
4717 		CTX_UNLOCK(ctx);
4718 		break;
4719 	default:
4720 		err = ether_ioctl(ifp, command, data);
4721 		break;
4722 	}
4723 	if (reinit)
4724 		iflib_if_init(ctx);
4725 	return (err);
4726 }
4727 
4728 static uint64_t
4729 iflib_if_get_counter(if_t ifp, ift_counter cnt)
4730 {
4731 	if_ctx_t ctx = if_getsoftc(ifp);
4732 
4733 	return (IFDI_GET_COUNTER(ctx, cnt));
4734 }
4735 
4736 /*********************************************************************
4737  *
4738  *  OTHER FUNCTIONS EXPORTED TO THE STACK
4739  *
4740  **********************************************************************/
4741 
4742 static void
4743 iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag)
4744 {
4745 	if_ctx_t ctx = if_getsoftc(ifp);
4746 
4747 	if ((void *)ctx != arg)
4748 		return;
4749 
4750 	if ((vtag == 0) || (vtag > 4095))
4751 		return;
4752 
4753 	if (iflib_in_detach(ctx))
4754 		return;
4755 
4756 	CTX_LOCK(ctx);
4757 	/* Driver may need all untagged packets to be flushed */
4758 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4759 		iflib_stop(ctx);
4760 	IFDI_VLAN_REGISTER(ctx, vtag);
4761 	/* Re-init to load the changes, if required */
4762 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4763 		iflib_init_locked(ctx);
4764 	CTX_UNLOCK(ctx);
4765 }
4766 
4767 static void
4768 iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag)
4769 {
4770 	if_ctx_t ctx = if_getsoftc(ifp);
4771 
4772 	if ((void *)ctx != arg)
4773 		return;
4774 
4775 	if ((vtag == 0) || (vtag > 4095))
4776 		return;
4777 
4778 	CTX_LOCK(ctx);
4779 	/* Driver may need all tagged packets to be flushed */
4780 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4781 		iflib_stop(ctx);
4782 	IFDI_VLAN_UNREGISTER(ctx, vtag);
4783 	/* Re-init to load the changes, if required */
4784 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4785 		iflib_init_locked(ctx);
4786 	CTX_UNLOCK(ctx);
4787 }
4788 
4789 static void
4790 iflib_led_func(void *arg, int onoff)
4791 {
4792 	if_ctx_t ctx = arg;
4793 
4794 	CTX_LOCK(ctx);
4795 	IFDI_LED_FUNC(ctx, onoff);
4796 	CTX_UNLOCK(ctx);
4797 }
4798 
4799 /*********************************************************************
4800  *
4801  *  BUS FUNCTION DEFINITIONS
4802  *
4803  **********************************************************************/
4804 
4805 int
4806 iflib_device_probe(device_t dev)
4807 {
4808 	const pci_vendor_info_t *ent;
4809 	if_shared_ctx_t sctx;
4810 	uint16_t pci_device_id, pci_rev_id, pci_subdevice_id, pci_subvendor_id;
4811 	uint16_t pci_vendor_id;
4812 
4813 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4814 		return (ENOTSUP);
4815 
4816 	pci_vendor_id = pci_get_vendor(dev);
4817 	pci_device_id = pci_get_device(dev);
4818 	pci_subvendor_id = pci_get_subvendor(dev);
4819 	pci_subdevice_id = pci_get_subdevice(dev);
4820 	pci_rev_id = pci_get_revid(dev);
4821 	if (sctx->isc_parse_devinfo != NULL)
4822 		sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id);
4823 
4824 	ent = sctx->isc_vendor_info;
4825 	while (ent->pvi_vendor_id != 0) {
4826 		if (pci_vendor_id != ent->pvi_vendor_id) {
4827 			ent++;
4828 			continue;
4829 		}
4830 		if ((pci_device_id == ent->pvi_device_id) &&
4831 		    ((pci_subvendor_id == ent->pvi_subvendor_id) ||
4832 		     (ent->pvi_subvendor_id == 0)) &&
4833 		    ((pci_subdevice_id == ent->pvi_subdevice_id) ||
4834 		     (ent->pvi_subdevice_id == 0)) &&
4835 		    ((pci_rev_id == ent->pvi_rev_id) ||
4836 		     (ent->pvi_rev_id == 0))) {
4837 			device_set_desc_copy(dev, ent->pvi_name);
4838 			/* this needs to be changed to zero if the bus probing code
4839 			 * ever stops re-probing on best match because the sctx
4840 			 * may have its values over written by register calls
4841 			 * in subsequent probes
4842 			 */
4843 			return (BUS_PROBE_DEFAULT);
4844 		}
4845 		ent++;
4846 	}
4847 	return (ENXIO);
4848 }
4849 
4850 int
4851 iflib_device_probe_vendor(device_t dev)
4852 {
4853 	int probe;
4854 
4855 	probe = iflib_device_probe(dev);
4856 #ifndef __HAIKU__
4857 	if (probe == BUS_PROBE_DEFAULT)
4858 		return (BUS_PROBE_VENDOR);
4859 	else
4860 #endif
4861 		return (probe);
4862 }
4863 
4864 static void
4865 iflib_reset_qvalues(if_ctx_t ctx)
4866 {
4867 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4868 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4869 	device_t dev = ctx->ifc_dev;
4870 	int i;
4871 
4872 	if (ctx->ifc_sysctl_ntxqs != 0)
4873 		scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs;
4874 	if (ctx->ifc_sysctl_nrxqs != 0)
4875 		scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs;
4876 
4877 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4878 		if (ctx->ifc_sysctl_ntxds[i] != 0)
4879 			scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i];
4880 		else
4881 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4882 	}
4883 
4884 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4885 		if (ctx->ifc_sysctl_nrxds[i] != 0)
4886 			scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i];
4887 		else
4888 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4889 	}
4890 
4891 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4892 		if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) {
4893 			device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n",
4894 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]);
4895 			scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i];
4896 		}
4897 		if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) {
4898 			device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n",
4899 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]);
4900 			scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i];
4901 		}
4902 		if (!powerof2(scctx->isc_nrxd[i])) {
4903 			device_printf(dev, "nrxd%d: %d is not a power of 2 - using default value of %d\n",
4904 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_default[i]);
4905 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4906 		}
4907 	}
4908 
4909 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4910 		if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) {
4911 			device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n",
4912 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]);
4913 			scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i];
4914 		}
4915 		if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) {
4916 			device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n",
4917 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]);
4918 			scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i];
4919 		}
4920 		if (!powerof2(scctx->isc_ntxd[i])) {
4921 			device_printf(dev, "ntxd%d: %d is not a power of 2 - using default value of %d\n",
4922 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_default[i]);
4923 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4924 		}
4925 	}
4926 }
4927 
4928 static void
4929 iflib_add_pfil(if_ctx_t ctx)
4930 {
4931 #ifndef __HAIKU__
4932 	struct pfil_head *pfil;
4933 	struct pfil_head_args pa;
4934 	iflib_rxq_t rxq;
4935 	int i;
4936 
4937 	pa.pa_version = PFIL_VERSION;
4938 	pa.pa_flags = PFIL_IN;
4939 	pa.pa_type = PFIL_TYPE_ETHERNET;
4940 	pa.pa_headname = if_name(ctx->ifc_ifp);
4941 	pfil = pfil_head_register(&pa);
4942 
4943 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4944 		rxq->pfil = pfil;
4945 	}
4946 #endif
4947 }
4948 
4949 static void
4950 iflib_rem_pfil(if_ctx_t ctx)
4951 {
4952 #ifndef __HAIKU__
4953 	struct pfil_head *pfil;
4954 	iflib_rxq_t rxq;
4955 	int i;
4956 
4957 	rxq = ctx->ifc_rxqs;
4958 	pfil = rxq->pfil;
4959 	for (i = 0; i < NRXQSETS(ctx); i++, rxq++) {
4960 		rxq->pfil = NULL;
4961 	}
4962 	pfil_head_unregister(pfil);
4963 #endif
4964 }
4965 
4966 
4967 #ifndef __HAIKU__
4968 /*
4969  * Advance forward by n members of the cpuset ctx->ifc_cpus starting from
4970  * cpuid and wrapping as necessary.
4971  */
4972 static unsigned int
4973 cpuid_advance(if_ctx_t ctx, unsigned int cpuid, unsigned int n)
4974 {
4975 	unsigned int first_valid;
4976 	unsigned int last_valid;
4977 
4978 	/* cpuid should always be in the valid set */
4979 	MPASS(CPU_ISSET(cpuid, &ctx->ifc_cpus));
4980 
4981 	/* valid set should never be empty */
4982 	MPASS(!CPU_EMPTY(&ctx->ifc_cpus));
4983 
4984 	first_valid = CPU_FFS(&ctx->ifc_cpus) - 1;
4985 	last_valid = CPU_FLS(&ctx->ifc_cpus) - 1;
4986 	n = n % CPU_COUNT(&ctx->ifc_cpus);
4987 	while (n > 0) {
4988 		do {
4989 			cpuid++;
4990 			if (cpuid > last_valid)
4991 				cpuid = first_valid;
4992 		} while (!CPU_ISSET(cpuid, &ctx->ifc_cpus));
4993 		n--;
4994 	}
4995 
4996 	return (cpuid);
4997 }
4998 #endif
4999 
5000 #if defined(SMP) && defined(SCHED_ULE)
5001 extern struct cpu_group *cpu_top;              /* CPU topology */
5002 
5003 static int
5004 find_child_with_core(int cpu, struct cpu_group *grp)
5005 {
5006 	int i;
5007 
5008 	if (grp->cg_children == 0)
5009 		return -1;
5010 
5011 	MPASS(grp->cg_child);
5012 	for (i = 0; i < grp->cg_children; i++) {
5013 		if (CPU_ISSET(cpu, &grp->cg_child[i].cg_mask))
5014 			return i;
5015 	}
5016 
5017 	return -1;
5018 }
5019 
5020 
5021 /*
5022  * Find an L2 neighbor of the given CPU or return -1 if none found.  This
5023  * does not distinguish among multiple L2 neighbors if the given CPU has
5024  * more than one (it will always return the same result in that case).
5025  */
5026 static int
5027 find_l2_neighbor(int cpu)
5028 {
5029 	struct cpu_group *grp;
5030 	int i;
5031 
5032 	grp = cpu_top;
5033 	if (grp == NULL)
5034 		return -1;
5035 
5036 	/*
5037 	 * Find the smallest CPU group that contains the given core.
5038 	 */
5039 	i = 0;
5040 	while ((i = find_child_with_core(cpu, grp)) != -1) {
5041 		/*
5042 		 * If the smallest group containing the given CPU has less
5043 		 * than two members, we conclude the given CPU has no
5044 		 * L2 neighbor.
5045 		 */
5046 		if (grp->cg_child[i].cg_count <= 1)
5047 			return (-1);
5048 		grp = &grp->cg_child[i];
5049 	}
5050 
5051 	/* Must share L2. */
5052 	if (grp->cg_level > CG_SHARE_L2 || grp->cg_level == CG_SHARE_NONE)
5053 		return -1;
5054 
5055 	/*
5056 	 * Select the first member of the set that isn't the reference
5057 	 * CPU, which at this point is guaranteed to exist.
5058 	 */
5059 	for (i = 0; i < CPU_SETSIZE; i++) {
5060 		if (CPU_ISSET(i, &grp->cg_mask) && i != cpu)
5061 			return (i);
5062 	}
5063 
5064 	/* Should never be reached */
5065 	return (-1);
5066 }
5067 
5068 #else
5069 static int
5070 find_l2_neighbor(int cpu)
5071 {
5072 
5073 	return (-1);
5074 }
5075 #endif
5076 
5077 #ifndef __HAIKU__
5078 /*
5079  * CPU mapping behaviors
5080  * ---------------------
5081  * 'separate txrx' refers to the separate_txrx sysctl
5082  * 'use logical' refers to the use_logical_cores sysctl
5083  * 'INTR CPUS' indicates whether bus_get_cpus(INTR_CPUS) succeeded
5084  *
5085  *  separate     use     INTR
5086  *    txrx     logical   CPUS   result
5087  * ---------- --------- ------ ------------------------------------------------
5088  *     -          -       X     RX and TX queues mapped to consecutive physical
5089  *                              cores with RX/TX pairs on same core and excess
5090  *                              of either following
5091  *     -          X       X     RX and TX queues mapped to consecutive cores
5092  *                              of any type with RX/TX pairs on same core and
5093  *                              excess of either following
5094  *     X          -       X     RX and TX queues mapped to consecutive physical
5095  *                              cores; all RX then all TX
5096  *     X          X       X     RX queues mapped to consecutive physical cores
5097  *                              first, then TX queues mapped to L2 neighbor of
5098  *                              the corresponding RX queue if one exists,
5099  *                              otherwise to consecutive physical cores
5100  *     -         n/a      -     RX and TX queues mapped to consecutive cores of
5101  *                              any type with RX/TX pairs on same core and excess
5102  *                              of either following
5103  *     X         n/a      -     RX and TX queues mapped to consecutive cores of
5104  *                              any type; all RX then all TX
5105  */
5106 static unsigned int
5107 get_cpuid_for_queue(if_ctx_t ctx, unsigned int base_cpuid, unsigned int qid,
5108     bool is_tx)
5109 {
5110 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5111 	unsigned int core_index;
5112 
5113 	if (ctx->ifc_sysctl_separate_txrx) {
5114 		/*
5115 		 * When using separate CPUs for TX and RX, the assignment
5116 		 * will always be of a consecutive CPU out of the set of
5117 		 * context CPUs, except for the specific case where the
5118 		 * context CPUs are phsyical cores, the use of logical cores
5119 		 * has been enabled, the assignment is for TX, the TX qid
5120 		 * corresponds to an RX qid, and the CPU assigned to the
5121 		 * corresponding RX queue has an L2 neighbor.
5122 		 */
5123 		if (ctx->ifc_sysctl_use_logical_cores &&
5124 		    ctx->ifc_cpus_are_physical_cores &&
5125 		    is_tx && qid < scctx->isc_nrxqsets) {
5126 			int l2_neighbor;
5127 			unsigned int rx_cpuid;
5128 
5129 			rx_cpuid = cpuid_advance(ctx, base_cpuid, qid);
5130 			l2_neighbor = find_l2_neighbor(rx_cpuid);
5131 			if (l2_neighbor != -1) {
5132 				return (l2_neighbor);
5133 			}
5134 			/*
5135 			 * ... else fall through to the normal
5136 			 * consecutive-after-RX assignment scheme.
5137 			 *
5138 			 * Note that we are assuming that all RX queue CPUs
5139 			 * have an L2 neighbor, or all do not.  If a mixed
5140 			 * scenario is possible, we will have to keep track
5141 			 * separately of how many queues prior to this one
5142 			 * were not able to be assigned to an L2 neighbor.
5143 			 */
5144 		}
5145 		if (is_tx)
5146 			core_index = scctx->isc_nrxqsets + qid;
5147 		else
5148 			core_index = qid;
5149 	} else {
5150 		core_index = qid;
5151 	}
5152 
5153 	return (cpuid_advance(ctx, base_cpuid, core_index));
5154 }
5155 #else
5156 #define get_cpuid_for_queue(...) CPU_FIRST()
5157 #endif
5158 
5159 static uint16_t
5160 get_ctx_core_offset(if_ctx_t ctx)
5161 {
5162 #ifndef __HAIKU__
5163 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5164 	struct cpu_offset *op;
5165 	cpuset_t assigned_cpus;
5166 	unsigned int cores_consumed;
5167 	unsigned int base_cpuid = ctx->ifc_sysctl_core_offset;
5168 	unsigned int first_valid;
5169 	unsigned int last_valid;
5170 	unsigned int i;
5171 
5172 	first_valid = CPU_FFS(&ctx->ifc_cpus) - 1;
5173 	last_valid = CPU_FLS(&ctx->ifc_cpus) - 1;
5174 
5175 	if (base_cpuid != CORE_OFFSET_UNSPECIFIED) {
5176 		/*
5177 		 * Align the user-chosen base CPU ID to the next valid CPU
5178 		 * for this device.  If the chosen base CPU ID is smaller
5179 		 * than the first valid CPU or larger than the last valid
5180 		 * CPU, we assume the user does not know what the valid
5181 		 * range is for this device and is thinking in terms of a
5182 		 * zero-based reference frame, and so we shift the given
5183 		 * value into the valid range (and wrap accordingly) so the
5184 		 * intent is translated to the proper frame of reference.
5185 		 * If the base CPU ID is within the valid first/last, but
5186 		 * does not correspond to a valid CPU, it is advanced to the
5187 		 * next valid CPU (wrapping if necessary).
5188 		 */
5189 		if (base_cpuid < first_valid || base_cpuid > last_valid) {
5190 			/* shift from zero-based to first_valid-based */
5191 			base_cpuid += first_valid;
5192 			/* wrap to range [first_valid, last_valid] */
5193 			base_cpuid = (base_cpuid - first_valid) %
5194 			    (last_valid - first_valid + 1);
5195 		}
5196 		if (!CPU_ISSET(base_cpuid, &ctx->ifc_cpus)) {
5197 			/*
5198 			 * base_cpuid is in [first_valid, last_valid], but
5199 			 * not a member of the valid set.  In this case,
5200 			 * there will always be a member of the valid set
5201 			 * with a CPU ID that is greater than base_cpuid,
5202 			 * and we simply advance to it.
5203 			 */
5204 			while (!CPU_ISSET(base_cpuid, &ctx->ifc_cpus))
5205 				base_cpuid++;
5206 		}
5207 		return (base_cpuid);
5208 	}
5209 
5210 	/*
5211 	 * Determine how many cores will be consumed by performing the CPU
5212 	 * assignments and counting how many of the assigned CPUs correspond
5213 	 * to CPUs in the set of context CPUs.  This is done using the CPU
5214 	 * ID first_valid as the base CPU ID, as the base CPU must be within
5215 	 * the set of context CPUs.
5216 	 *
5217 	 * Note not all assigned CPUs will be in the set of context CPUs
5218 	 * when separate CPUs are being allocated to TX and RX queues,
5219 	 * assignment to logical cores has been enabled, the set of context
5220 	 * CPUs contains only physical CPUs, and TX queues are mapped to L2
5221 	 * neighbors of CPUs that RX queues have been mapped to - in this
5222 	 * case we do only want to count how many CPUs in the set of context
5223 	 * CPUs have been consumed, as that determines the next CPU in that
5224 	 * set to start allocating at for the next device for which
5225 	 * core_offset is not set.
5226 	 */
5227 	CPU_ZERO(&assigned_cpus);
5228 	for (i = 0; i < scctx->isc_ntxqsets; i++)
5229 		CPU_SET(get_cpuid_for_queue(ctx, first_valid, i, true),
5230 		    &assigned_cpus);
5231 	for (i = 0; i < scctx->isc_nrxqsets; i++)
5232 		CPU_SET(get_cpuid_for_queue(ctx, first_valid, i, false),
5233 		    &assigned_cpus);
5234 	CPU_AND(&assigned_cpus, &assigned_cpus, &ctx->ifc_cpus);
5235 	cores_consumed = CPU_COUNT(&assigned_cpus);
5236 
5237 	mtx_lock(&cpu_offset_mtx);
5238 	SLIST_FOREACH(op, &cpu_offsets, entries) {
5239 		if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) {
5240 			base_cpuid = op->next_cpuid;
5241 			op->next_cpuid = cpuid_advance(ctx, op->next_cpuid,
5242 			    cores_consumed);
5243 			MPASS(op->refcount < UINT_MAX);
5244 			op->refcount++;
5245 			break;
5246 		}
5247 	}
5248 	if (base_cpuid == CORE_OFFSET_UNSPECIFIED) {
5249 		base_cpuid = first_valid;
5250 		op = malloc(sizeof(struct cpu_offset), M_IFLIB,
5251 		    M_NOWAIT | M_ZERO);
5252 		if (op == NULL) {
5253 			device_printf(ctx->ifc_dev,
5254 			    "allocation for cpu offset failed.\n");
5255 		} else {
5256 			op->next_cpuid = cpuid_advance(ctx, base_cpuid,
5257 			    cores_consumed);
5258 			op->refcount = 1;
5259 			CPU_COPY(&ctx->ifc_cpus, &op->set);
5260 			SLIST_INSERT_HEAD(&cpu_offsets, op, entries);
5261 		}
5262 	}
5263 	mtx_unlock(&cpu_offset_mtx);
5264 
5265 	return (base_cpuid);
5266 #else
5267 	return 0;
5268 #endif
5269 }
5270 
5271 static void
5272 unref_ctx_core_offset(if_ctx_t ctx)
5273 {
5274 #ifndef __HAIKU__
5275 	struct cpu_offset *op, *top;
5276 
5277 	mtx_lock(&cpu_offset_mtx);
5278 	SLIST_FOREACH_SAFE(op, &cpu_offsets, entries, top) {
5279 		if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) {
5280 			MPASS(op->refcount > 0);
5281 			op->refcount--;
5282 			if (op->refcount == 0) {
5283 				SLIST_REMOVE(&cpu_offsets, op, cpu_offset, entries);
5284 				free(op, M_IFLIB);
5285 			}
5286 			break;
5287 		}
5288 	}
5289 	mtx_unlock(&cpu_offset_mtx);
5290 #endif
5291 }
5292 
5293 int
5294 iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp)
5295 {
5296 	if_ctx_t ctx;
5297 	if_t ifp;
5298 	if_softc_ctx_t scctx;
5299 	kobjop_desc_t kobj_desc;
5300 	kobj_method_t *kobj_method;
5301 	int err, msix, rid;
5302 	int num_txd, num_rxd;
5303 
5304 	ctx = malloc(sizeof(* ctx), M_IFLIB, M_WAITOK|M_ZERO);
5305 
5306 	if (sc == NULL) {
5307 		sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
5308 		device_set_softc(dev, ctx);
5309 		ctx->ifc_flags |= IFC_SC_ALLOCATED;
5310 	}
5311 
5312 	ctx->ifc_sctx = sctx;
5313 	ctx->ifc_dev = dev;
5314 	ctx->ifc_softc = sc;
5315 
5316 	if ((err = iflib_register(ctx)) != 0) {
5317 		device_printf(dev, "iflib_register failed %d\n", err);
5318 		goto fail_ctx_free;
5319 	}
5320 	iflib_add_device_sysctl_pre(ctx);
5321 
5322 	scctx = &ctx->ifc_softc_ctx;
5323 	ifp = ctx->ifc_ifp;
5324 
5325 	iflib_reset_qvalues(ctx);
5326 	IFNET_WLOCK();
5327 	CTX_LOCK(ctx);
5328 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
5329 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
5330 		goto fail_unlock;
5331 	}
5332 	_iflib_pre_assert(scctx);
5333 	ctx->ifc_txrx = *scctx->isc_txrx;
5334 
5335 	MPASS(scctx->isc_dma_width <= flsll(BUS_SPACE_MAXADDR));
5336 
5337 	if (sctx->isc_flags & IFLIB_DRIVER_MEDIA)
5338 		ctx->ifc_mediap = scctx->isc_media;
5339 
5340 #ifdef INVARIANTS
5341 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
5342 		MPASS(scctx->isc_tx_csum_flags);
5343 #endif
5344 
5345 	if_setcapabilities(ifp,
5346 	    scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_MEXTPG);
5347 	if_setcapenable(ifp,
5348 	    scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_MEXTPG);
5349 
5350 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
5351 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
5352 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
5353 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
5354 
5355 	num_txd = iflib_num_tx_descs(ctx);
5356 	num_rxd = iflib_num_rx_descs(ctx);
5357 
5358 	/* XXX change for per-queue sizes */
5359 	device_printf(dev, "Using %d TX descriptors and %d RX descriptors\n",
5360 	    num_txd, num_rxd);
5361 
5362 	if (scctx->isc_tx_nsegments > num_txd / MAX_SINGLE_PACKET_FRACTION)
5363 		scctx->isc_tx_nsegments = max(1, num_txd /
5364 		    MAX_SINGLE_PACKET_FRACTION);
5365 	if (scctx->isc_tx_tso_segments_max > num_txd /
5366 	    MAX_SINGLE_PACKET_FRACTION)
5367 		scctx->isc_tx_tso_segments_max = max(1,
5368 		    num_txd / MAX_SINGLE_PACKET_FRACTION);
5369 
5370 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
5371 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
5372 #ifndef __HAIKU__
5373 		/*
5374 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
5375 		 * but some MACs do.
5376 		 */
5377 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
5378 		    IP_MAXPACKET));
5379 		/*
5380 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
5381 		 * into account.  In the worst case, each of these calls will
5382 		 * add another mbuf and, thus, the requirement for another DMA
5383 		 * segment.  So for best performance, it doesn't make sense to
5384 		 * advertize a maximum of TSO segments that typically will
5385 		 * require defragmentation in iflib_encap().
5386 		 */
5387 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
5388 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
5389 #endif
5390 	}
5391 	if (scctx->isc_rss_table_size == 0)
5392 		scctx->isc_rss_table_size = 64;
5393 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
5394 
5395 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
5396 	/* XXX format name */
5397 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx,
5398 	    NULL, NULL, "admin");
5399 
5400 #ifndef __HAIKU__
5401 	/* Set up cpu set.  If it fails, use the set of all CPUs. */
5402 	if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) != 0) {
5403 		device_printf(dev, "Unable to fetch CPU list\n");
5404 		CPU_COPY(&all_cpus, &ctx->ifc_cpus);
5405 		ctx->ifc_cpus_are_physical_cores = false;
5406 	} else
5407 		ctx->ifc_cpus_are_physical_cores = true;
5408 	MPASS(CPU_COUNT(&ctx->ifc_cpus) > 0);
5409 #endif
5410 
5411 	/*
5412 	** Now set up MSI or MSI-X, should return us the number of supported
5413 	** vectors (will be 1 for a legacy interrupt and MSI).
5414 	*/
5415 	if (sctx->isc_flags & IFLIB_SKIP_MSIX) {
5416 		msix = scctx->isc_vectors;
5417 	} else if (scctx->isc_msix_bar != 0)
5418 	       /*
5419 		* The simple fact that isc_msix_bar is not 0 does not mean we
5420 		* we have a good value there that is known to work.
5421 		*/
5422 		msix = iflib_msix_init(ctx);
5423 	else {
5424 		scctx->isc_vectors = 1;
5425 		scctx->isc_ntxqsets = 1;
5426 		scctx->isc_nrxqsets = 1;
5427 		scctx->isc_intr = IFLIB_INTR_LEGACY;
5428 		msix = 0;
5429 	}
5430 	/* Get memory for the station queues */
5431 	if ((err = iflib_queues_alloc(ctx))) {
5432 		device_printf(dev, "Unable to allocate queue memory\n");
5433 		goto fail_intr_free;
5434 	}
5435 
5436 	if ((err = iflib_qset_structures_setup(ctx)))
5437 		goto fail_queues;
5438 
5439 	/*
5440 	 * Now that we know how many queues there are, get the core offset.
5441 	 */
5442 	ctx->ifc_sysctl_core_offset = get_ctx_core_offset(ctx);
5443 
5444 	if (msix > 1) {
5445 		/*
5446 		 * When using MSI-X, ensure that ifdi_{r,t}x_queue_intr_enable
5447 		 * aren't the default NULL implementation.
5448 		 */
5449 		kobj_desc = &ifdi_rx_queue_intr_enable_desc;
5450 #ifdef __HAIKU__
5451 		kobj_method = kobj_lookup_method(ctx->ops.cls, NULL,
5452 #else
5453 		kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL,
5454 #endif
5455 		    kobj_desc);
5456 		if (kobj_method == &kobj_desc->deflt) {
5457 			device_printf(dev,
5458 			    "MSI-X requires ifdi_rx_queue_intr_enable method");
5459 			err = EOPNOTSUPP;
5460 			goto fail_queues;
5461 		}
5462 		kobj_desc = &ifdi_tx_queue_intr_enable_desc;
5463 #ifdef __HAIKU__
5464 		kobj_method = kobj_lookup_method(ctx->ops.cls, NULL,
5465 #else
5466 		kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL,
5467 #endif
5468 		    kobj_desc);
5469 		if (kobj_method == &kobj_desc->deflt) {
5470 			device_printf(dev,
5471 			    "MSI-X requires ifdi_tx_queue_intr_enable method");
5472 			err = EOPNOTSUPP;
5473 			goto fail_queues;
5474 		}
5475 
5476 		/*
5477 		 * Assign the MSI-X vectors.
5478 		 * Note that the default NULL ifdi_msix_intr_assign method will
5479 		 * fail here, too.
5480 		 */
5481 		err = IFDI_MSIX_INTR_ASSIGN(ctx, msix);
5482 		if (err != 0) {
5483 			device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n",
5484 			    err);
5485 			goto fail_queues;
5486 		}
5487 	} else if (scctx->isc_intr != IFLIB_INTR_MSIX) {
5488 		rid = 0;
5489 		if (scctx->isc_intr == IFLIB_INTR_MSI) {
5490 			MPASS(msix == 1);
5491 			rid = 1;
5492 		}
5493 		if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) {
5494 			device_printf(dev, "iflib_legacy_setup failed %d\n", err);
5495 			goto fail_queues;
5496 		}
5497 	} else {
5498 		device_printf(dev,
5499 		    "Cannot use iflib with only 1 MSI-X interrupt!\n");
5500 		err = ENODEV;
5501 		goto fail_queues;
5502 	}
5503 
5504 	/*
5505 	 * It prevents a double-locking panic with iflib_media_status when
5506 	 * the driver loads.
5507 	 */
5508 	CTX_UNLOCK(ctx);
5509 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
5510 	CTX_LOCK(ctx);
5511 
5512 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
5513 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
5514 		goto fail_detach;
5515 	}
5516 
5517 	/*
5518 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
5519 	 * This must appear after the call to ether_ifattach() because
5520 	 * ether_ifattach() sets if_hdrlen to the default value.
5521 	 */
5522 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
5523 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
5524 
5525 	if ((err = iflib_netmap_attach(ctx))) {
5526 		device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err);
5527 		goto fail_detach;
5528 	}
5529 	*ctxp = ctx;
5530 
5531 	DEBUGNET_SET(ctx->ifc_ifp, iflib);
5532 
5533 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
5534 	iflib_add_device_sysctl_post(ctx);
5535 	iflib_add_pfil(ctx);
5536 	ctx->ifc_flags |= IFC_INIT_DONE;
5537 	CTX_UNLOCK(ctx);
5538 	IFNET_WUNLOCK();
5539 
5540 	return (0);
5541 
5542 fail_detach:
5543 	ether_ifdetach(ctx->ifc_ifp);
5544 fail_queues:
5545 	iflib_tqg_detach(ctx);
5546 	iflib_tx_structures_free(ctx);
5547 	iflib_rx_structures_free(ctx);
5548 	IFDI_DETACH(ctx);
5549 	IFDI_QUEUES_FREE(ctx);
5550 fail_intr_free:
5551 	iflib_free_intr_mem(ctx);
5552 fail_unlock:
5553 	CTX_UNLOCK(ctx);
5554 	IFNET_WUNLOCK();
5555 	iflib_deregister(ctx);
5556 fail_ctx_free:
5557 	device_set_softc(ctx->ifc_dev, NULL);
5558         if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5559                 free(ctx->ifc_softc, M_IFLIB);
5560         free(ctx, M_IFLIB);
5561 	return (err);
5562 }
5563 
5564 int
5565 iflib_device_attach(device_t dev)
5566 {
5567 	if_ctx_t ctx;
5568 	if_shared_ctx_t sctx;
5569 
5570 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
5571 		return (ENOTSUP);
5572 
5573 	pci_enable_busmaster(dev);
5574 
5575 	return (iflib_device_register(dev, NULL, sctx, &ctx));
5576 }
5577 
5578 int
5579 iflib_device_deregister(if_ctx_t ctx)
5580 {
5581 	if_t ifp = ctx->ifc_ifp;
5582 	device_t dev = ctx->ifc_dev;
5583 
5584 	/* Make sure VLANS are not using driver */
5585 	if (if_vlantrunkinuse(ifp)) {
5586 		device_printf(dev, "Vlan in use, detach first\n");
5587 		return (EBUSY);
5588 	}
5589 #ifdef PCI_IOV
5590 	if (!CTX_IS_VF(ctx) && pci_iov_detach(dev) != 0) {
5591 		device_printf(dev, "SR-IOV in use; detach first.\n");
5592 		return (EBUSY);
5593 	}
5594 #endif
5595 
5596 	STATE_LOCK(ctx);
5597 	ctx->ifc_flags |= IFC_IN_DETACH;
5598 	STATE_UNLOCK(ctx);
5599 
5600 	/* Unregister VLAN handlers before calling iflib_stop() */
5601 	iflib_unregister_vlan_handlers(ctx);
5602 
5603 	iflib_netmap_detach(ifp);
5604 	ether_ifdetach(ifp);
5605 
5606 	CTX_LOCK(ctx);
5607 	iflib_stop(ctx);
5608 	CTX_UNLOCK(ctx);
5609 
5610 	iflib_rem_pfil(ctx);
5611 	if (ctx->ifc_led_dev != NULL)
5612 		led_destroy(ctx->ifc_led_dev);
5613 
5614 	iflib_tqg_detach(ctx);
5615 	iflib_tx_structures_free(ctx);
5616 	iflib_rx_structures_free(ctx);
5617 
5618 	CTX_LOCK(ctx);
5619 	IFDI_DETACH(ctx);
5620 	IFDI_QUEUES_FREE(ctx);
5621 	CTX_UNLOCK(ctx);
5622 
5623 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
5624 	iflib_free_intr_mem(ctx);
5625 
5626 	bus_generic_detach(dev);
5627 
5628 	iflib_deregister(ctx);
5629 
5630 	device_set_softc(ctx->ifc_dev, NULL);
5631 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5632 		free(ctx->ifc_softc, M_IFLIB);
5633 	unref_ctx_core_offset(ctx);
5634 	free(ctx, M_IFLIB);
5635 	return (0);
5636 }
5637 
5638 static void
5639 iflib_tqg_detach(if_ctx_t ctx)
5640 {
5641 	iflib_txq_t txq;
5642 	iflib_rxq_t rxq;
5643 	int i;
5644 	struct taskqgroup *tqg;
5645 
5646 	/* XXX drain any dependent tasks */
5647 	tqg = qgroup_if_io_tqg;
5648 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
5649 		callout_drain(&txq->ift_timer);
5650 #ifdef DEV_NETMAP
5651 		callout_drain(&txq->ift_netmap_timer);
5652 #endif /* DEV_NETMAP */
5653 		if (txq->ift_task.gt_uniq != NULL)
5654 			taskqgroup_detach(tqg, &txq->ift_task);
5655 	}
5656 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
5657 		if (rxq->ifr_task.gt_uniq != NULL)
5658 			taskqgroup_detach(tqg, &rxq->ifr_task);
5659 	}
5660 	tqg = qgroup_if_config_tqg;
5661 	if (ctx->ifc_admin_task.gt_uniq != NULL)
5662 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
5663 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
5664 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
5665 }
5666 
5667 static void
5668 iflib_free_intr_mem(if_ctx_t ctx)
5669 {
5670 
5671 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) {
5672 		iflib_irq_free(ctx, &ctx->ifc_legacy_irq);
5673 	}
5674 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) {
5675 		pci_release_msi(ctx->ifc_dev);
5676 	}
5677 	if (ctx->ifc_msix_mem != NULL) {
5678 		bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY,
5679 		    rman_get_rid(ctx->ifc_msix_mem), ctx->ifc_msix_mem);
5680 		ctx->ifc_msix_mem = NULL;
5681 	}
5682 }
5683 
5684 int
5685 iflib_device_detach(device_t dev)
5686 {
5687 	if_ctx_t ctx = device_get_softc(dev);
5688 
5689 	return (iflib_device_deregister(ctx));
5690 }
5691 
5692 int
5693 iflib_device_suspend(device_t dev)
5694 {
5695 	if_ctx_t ctx = device_get_softc(dev);
5696 
5697 	CTX_LOCK(ctx);
5698 	IFDI_SUSPEND(ctx);
5699 	CTX_UNLOCK(ctx);
5700 
5701 	return bus_generic_suspend(dev);
5702 }
5703 int
5704 iflib_device_shutdown(device_t dev)
5705 {
5706 	if_ctx_t ctx = device_get_softc(dev);
5707 
5708 	CTX_LOCK(ctx);
5709 	IFDI_SHUTDOWN(ctx);
5710 	CTX_UNLOCK(ctx);
5711 
5712 	return bus_generic_suspend(dev);
5713 }
5714 
5715 int
5716 iflib_device_resume(device_t dev)
5717 {
5718 	if_ctx_t ctx = device_get_softc(dev);
5719 	iflib_txq_t txq = ctx->ifc_txqs;
5720 
5721 	CTX_LOCK(ctx);
5722 	IFDI_RESUME(ctx);
5723 	iflib_if_init_locked(ctx);
5724 	CTX_UNLOCK(ctx);
5725 	for (int i = 0; i < NTXQSETS(ctx); i++, txq++)
5726 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
5727 
5728 	return (bus_generic_resume(dev));
5729 }
5730 
5731 int
5732 iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params)
5733 {
5734 	int error;
5735 	if_ctx_t ctx = device_get_softc(dev);
5736 
5737 	CTX_LOCK(ctx);
5738 	error = IFDI_IOV_INIT(ctx, num_vfs, params);
5739 	CTX_UNLOCK(ctx);
5740 
5741 	return (error);
5742 }
5743 
5744 void
5745 iflib_device_iov_uninit(device_t dev)
5746 {
5747 	if_ctx_t ctx = device_get_softc(dev);
5748 
5749 	CTX_LOCK(ctx);
5750 	IFDI_IOV_UNINIT(ctx);
5751 	CTX_UNLOCK(ctx);
5752 }
5753 
5754 int
5755 iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params)
5756 {
5757 	int error;
5758 	if_ctx_t ctx = device_get_softc(dev);
5759 
5760 	CTX_LOCK(ctx);
5761 	error = IFDI_IOV_VF_ADD(ctx, vfnum, params);
5762 	CTX_UNLOCK(ctx);
5763 
5764 	return (error);
5765 }
5766 
5767 /*********************************************************************
5768  *
5769  *  MODULE FUNCTION DEFINITIONS
5770  *
5771  **********************************************************************/
5772 
5773 /*
5774  * - Start a fast taskqueue thread for each core
5775  * - Start a taskqueue for control operations
5776  */
5777 static int
5778 iflib_module_init(void)
5779 {
5780 	iflib_timer_default = hz / 2;
5781 	return (0);
5782 }
5783 
5784 static int
5785 iflib_module_event_handler(module_t mod, int what, void *arg)
5786 {
5787 	int err;
5788 
5789 	switch (what) {
5790 	case MOD_LOAD:
5791 		if ((err = iflib_module_init()) != 0)
5792 			return (err);
5793 		break;
5794 	case MOD_UNLOAD:
5795 		return (EBUSY);
5796 	default:
5797 		return (EOPNOTSUPP);
5798 	}
5799 
5800 	return (0);
5801 }
5802 
5803 /*********************************************************************
5804  *
5805  *  PUBLIC FUNCTION DEFINITIONS
5806  *     ordered as in iflib.h
5807  *
5808  **********************************************************************/
5809 
5810 static void
5811 _iflib_assert(if_shared_ctx_t sctx)
5812 {
5813 	int i;
5814 
5815 	MPASS(sctx->isc_tx_maxsize);
5816 	MPASS(sctx->isc_tx_maxsegsize);
5817 
5818 	MPASS(sctx->isc_rx_maxsize);
5819 	MPASS(sctx->isc_rx_nsegments);
5820 	MPASS(sctx->isc_rx_maxsegsize);
5821 
5822 	MPASS(sctx->isc_nrxqs >= 1 && sctx->isc_nrxqs <= 8);
5823 	for (i = 0; i < sctx->isc_nrxqs; i++) {
5824 		MPASS(sctx->isc_nrxd_min[i]);
5825 		MPASS(powerof2(sctx->isc_nrxd_min[i]));
5826 		MPASS(sctx->isc_nrxd_max[i]);
5827 		MPASS(powerof2(sctx->isc_nrxd_max[i]));
5828 		MPASS(sctx->isc_nrxd_default[i]);
5829 		MPASS(powerof2(sctx->isc_nrxd_default[i]));
5830 	}
5831 
5832 	MPASS(sctx->isc_ntxqs >= 1 && sctx->isc_ntxqs <= 8);
5833 	for (i = 0; i < sctx->isc_ntxqs; i++) {
5834 		MPASS(sctx->isc_ntxd_min[i]);
5835 		MPASS(powerof2(sctx->isc_ntxd_min[i]));
5836 		MPASS(sctx->isc_ntxd_max[i]);
5837 		MPASS(powerof2(sctx->isc_ntxd_max[i]));
5838 		MPASS(sctx->isc_ntxd_default[i]);
5839 		MPASS(powerof2(sctx->isc_ntxd_default[i]));
5840 	}
5841 }
5842 
5843 static void
5844 _iflib_pre_assert(if_softc_ctx_t scctx)
5845 {
5846 
5847 	MPASS(scctx->isc_txrx->ift_txd_encap);
5848 	MPASS(scctx->isc_txrx->ift_txd_flush);
5849 	MPASS(scctx->isc_txrx->ift_txd_credits_update);
5850 	MPASS(scctx->isc_txrx->ift_rxd_available);
5851 	MPASS(scctx->isc_txrx->ift_rxd_pkt_get);
5852 	MPASS(scctx->isc_txrx->ift_rxd_refill);
5853 	MPASS(scctx->isc_txrx->ift_rxd_flush);
5854 }
5855 
5856 static int
5857 iflib_register(if_ctx_t ctx)
5858 {
5859 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5860 	driver_t *driver = sctx->isc_driver;
5861 	device_t dev = ctx->ifc_dev;
5862 	if_t ifp;
5863 
5864 	_iflib_assert(sctx);
5865 
5866 	CTX_LOCK_INIT(ctx);
5867 	STATE_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev));
5868 	ifp = ctx->ifc_ifp = if_alloc(IFT_ETHER);
5869 	if (ifp == NULL) {
5870 		device_printf(dev, "can not allocate ifnet structure\n");
5871 		return (ENOMEM);
5872 	}
5873 
5874 	/*
5875 	 * Initialize our context's device specific methods
5876 	 */
5877 	kobj_init((kobj_t) ctx, (kobj_class_t) driver);
5878 	kobj_class_compile((kobj_class_t) driver);
5879 
5880 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
5881 	if_setsoftc(ifp, ctx);
5882 	if_setdev(ifp, dev);
5883 	if_setinitfn(ifp, iflib_if_init);
5884 	if_setioctlfn(ifp, iflib_if_ioctl);
5885 #ifdef ALTQ
5886 	if_setstartfn(ifp, iflib_altq_if_start);
5887 	if_settransmitfn(ifp, iflib_altq_if_transmit);
5888 	if_setsendqready(ifp);
5889 #else
5890 	if_settransmitfn(ifp, iflib_if_transmit);
5891 #endif
5892 	if_setqflushfn(ifp, iflib_if_qflush);
5893 	if_setflags(ifp, IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST);
5894 	ctx->ifc_vlan_attach_event =
5895 		EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx,
5896 							  EVENTHANDLER_PRI_FIRST);
5897 	ctx->ifc_vlan_detach_event =
5898 		EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx,
5899 							  EVENTHANDLER_PRI_FIRST);
5900 
5901 	if ((sctx->isc_flags & IFLIB_DRIVER_MEDIA) == 0) {
5902 		ctx->ifc_mediap = &ctx->ifc_media;
5903 		ifmedia_init(ctx->ifc_mediap, IFM_IMASK,
5904 		    iflib_media_change, iflib_media_status);
5905 	}
5906 	return (0);
5907 }
5908 
5909 static void
5910 iflib_unregister_vlan_handlers(if_ctx_t ctx)
5911 {
5912 	/* Unregister VLAN events */
5913 	if (ctx->ifc_vlan_attach_event != NULL) {
5914 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
5915 		ctx->ifc_vlan_attach_event = NULL;
5916 	}
5917 	if (ctx->ifc_vlan_detach_event != NULL) {
5918 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
5919 		ctx->ifc_vlan_detach_event = NULL;
5920 	}
5921 
5922 }
5923 
5924 static void
5925 iflib_deregister(if_ctx_t ctx)
5926 {
5927 	if_t ifp = ctx->ifc_ifp;
5928 
5929 	/* Remove all media */
5930 	ifmedia_removeall(&ctx->ifc_media);
5931 
5932 	/* Ensure that VLAN event handlers are unregistered */
5933 	iflib_unregister_vlan_handlers(ctx);
5934 
5935 #ifndef __HAIKU__
5936 	/* Release kobject reference */
5937 	kobj_delete((kobj_t) ctx, NULL);
5938 #endif
5939 
5940 	/* Free the ifnet structure */
5941 	if_free(ifp);
5942 
5943 	STATE_LOCK_DESTROY(ctx);
5944 
5945 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
5946 	CTX_LOCK_DESTROY(ctx);
5947 }
5948 
5949 static int
5950 iflib_queues_alloc(if_ctx_t ctx)
5951 {
5952 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5953 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5954 	device_t dev = ctx->ifc_dev;
5955 	int nrxqsets = scctx->isc_nrxqsets;
5956 	int ntxqsets = scctx->isc_ntxqsets;
5957 	iflib_txq_t txq;
5958 	iflib_rxq_t rxq;
5959 	iflib_fl_t fl = NULL;
5960 	int i, j, cpu, err, txconf, rxconf;
5961 	iflib_dma_info_t ifdip;
5962 	uint32_t *rxqsizes = scctx->isc_rxqsizes;
5963 	uint32_t *txqsizes = scctx->isc_txqsizes;
5964 	uint8_t nrxqs = sctx->isc_nrxqs;
5965 	uint8_t ntxqs = sctx->isc_ntxqs;
5966 	int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1;
5967 	int fl_offset = (sctx->isc_flags & IFLIB_HAS_RXCQ ? 1 : 0);
5968 	caddr_t *vaddrs;
5969 	uint64_t *paddrs;
5970 
5971 	KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1"));
5972 	KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1"));
5973 	KASSERT(nrxqs >= fl_offset + nfree_lists,
5974            ("there must be at least a rxq for each free list"));
5975 
5976 	/* Allocate the TX ring struct memory */
5977 	if (!(ctx->ifc_txqs =
5978 	    (iflib_txq_t) malloc(sizeof(struct iflib_txq) *
5979 	    ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5980 		device_printf(dev, "Unable to allocate TX ring memory\n");
5981 		err = ENOMEM;
5982 		goto fail;
5983 	}
5984 
5985 	/* Now allocate the RX */
5986 	if (!(ctx->ifc_rxqs =
5987 	    (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) *
5988 	    nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5989 		device_printf(dev, "Unable to allocate RX ring memory\n");
5990 		err = ENOMEM;
5991 		goto rx_fail;
5992 	}
5993 
5994 	txq = ctx->ifc_txqs;
5995 	rxq = ctx->ifc_rxqs;
5996 
5997 	/*
5998 	 * XXX handle allocation failure
5999 	 */
6000 	for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) {
6001 		/* Set up some basics */
6002 
6003 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs,
6004 		    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
6005 			device_printf(dev,
6006 			    "Unable to allocate TX DMA info memory\n");
6007 			err = ENOMEM;
6008 			goto err_tx_desc;
6009 		}
6010 		txq->ift_ifdi = ifdip;
6011 		for (j = 0; j < ntxqs; j++, ifdip++) {
6012 			if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, 0)) {
6013 				device_printf(dev,
6014 				    "Unable to allocate TX descriptors\n");
6015 				err = ENOMEM;
6016 				goto err_tx_desc;
6017 			}
6018 			txq->ift_txd_size[j] = scctx->isc_txd_size[j];
6019 			bzero((void *)ifdip->idi_vaddr, txqsizes[j]);
6020 		}
6021 		txq->ift_ctx = ctx;
6022 		txq->ift_id = i;
6023 		if (sctx->isc_flags & IFLIB_HAS_TXCQ) {
6024 			txq->ift_br_offset = 1;
6025 		} else {
6026 			txq->ift_br_offset = 0;
6027 		}
6028 
6029 		if (iflib_txsd_alloc(txq)) {
6030 			device_printf(dev, "Critical Failure setting up TX buffers\n");
6031 			err = ENOMEM;
6032 			goto err_tx_desc;
6033 		}
6034 
6035 		/* Initialize the TX lock */
6036 		snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:TX(%d):callout",
6037 		    device_get_nameunit(dev), txq->ift_id);
6038 		mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF);
6039 		callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0);
6040 #ifndef __HAIKU__
6041 		txq->ift_timer.c_cpu = cpu;
6042 #endif
6043 #ifdef DEV_NETMAP
6044 		callout_init_mtx(&txq->ift_netmap_timer, &txq->ift_mtx, 0);
6045 		txq->ift_netmap_timer.c_cpu = cpu;
6046 #endif /* DEV_NETMAP */
6047 
6048 		err = ifmp_ring_alloc(&txq->ift_br, 2048, txq, iflib_txq_drain,
6049 				      iflib_txq_can_drain, M_IFLIB, M_WAITOK);
6050 		if (err) {
6051 			/* XXX free any allocated rings */
6052 			device_printf(dev, "Unable to allocate buf_ring\n");
6053 			goto err_tx_desc;
6054 		}
6055 	}
6056 
6057 	for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) {
6058 		/* Set up some basics */
6059 		callout_init(&rxq->ifr_watchdog, 1);
6060 
6061 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs,
6062 		   M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
6063 			device_printf(dev,
6064 			    "Unable to allocate RX DMA info memory\n");
6065 			err = ENOMEM;
6066 			goto err_tx_desc;
6067 		}
6068 
6069 		rxq->ifr_ifdi = ifdip;
6070 		/* XXX this needs to be changed if #rx queues != #tx queues */
6071 		rxq->ifr_ntxqirq = 1;
6072 		rxq->ifr_txqid[0] = i;
6073 		for (j = 0; j < nrxqs; j++, ifdip++) {
6074 			if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, 0)) {
6075 				device_printf(dev,
6076 				    "Unable to allocate RX descriptors\n");
6077 				err = ENOMEM;
6078 				goto err_tx_desc;
6079 			}
6080 			bzero((void *)ifdip->idi_vaddr, rxqsizes[j]);
6081 		}
6082 		rxq->ifr_ctx = ctx;
6083 		rxq->ifr_id = i;
6084 		rxq->ifr_fl_offset = fl_offset;
6085 		rxq->ifr_nfl = nfree_lists;
6086 		if (!(fl =
6087 			  (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) {
6088 			device_printf(dev, "Unable to allocate free list memory\n");
6089 			err = ENOMEM;
6090 			goto err_tx_desc;
6091 		}
6092 		rxq->ifr_fl = fl;
6093 		for (j = 0; j < nfree_lists; j++) {
6094 			fl[j].ifl_rxq = rxq;
6095 			fl[j].ifl_id = j;
6096 			fl[j].ifl_ifdi = &rxq->ifr_ifdi[j + rxq->ifr_fl_offset];
6097 			fl[j].ifl_rxd_size = scctx->isc_rxd_size[j];
6098 		}
6099 		/* Allocate receive buffers for the ring */
6100 		if (iflib_rxsd_alloc(rxq)) {
6101 			device_printf(dev,
6102 			    "Critical Failure setting up receive buffers\n");
6103 			err = ENOMEM;
6104 			goto err_rx_desc;
6105 		}
6106 
6107 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
6108 			fl->ifl_rx_bitmap = bit_alloc(fl->ifl_size, M_IFLIB,
6109 			    M_WAITOK);
6110 	}
6111 
6112 	/* TXQs */
6113 	vaddrs = malloc(sizeof(caddr_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
6114 	paddrs = malloc(sizeof(uint64_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
6115 	for (i = 0; i < ntxqsets; i++) {
6116 		iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi;
6117 
6118 		for (j = 0; j < ntxqs; j++, di++) {
6119 			vaddrs[i*ntxqs + j] = di->idi_vaddr;
6120 			paddrs[i*ntxqs + j] = di->idi_paddr;
6121 		}
6122 	}
6123 	if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) {
6124 		device_printf(ctx->ifc_dev,
6125 		    "Unable to allocate device TX queue\n");
6126 		iflib_tx_structures_free(ctx);
6127 		free(vaddrs, M_IFLIB);
6128 		free(paddrs, M_IFLIB);
6129 		goto err_rx_desc;
6130 	}
6131 	free(vaddrs, M_IFLIB);
6132 	free(paddrs, M_IFLIB);
6133 
6134 	/* RXQs */
6135 	vaddrs = malloc(sizeof(caddr_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
6136 	paddrs = malloc(sizeof(uint64_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
6137 	for (i = 0; i < nrxqsets; i++) {
6138 		iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi;
6139 
6140 		for (j = 0; j < nrxqs; j++, di++) {
6141 			vaddrs[i*nrxqs + j] = di->idi_vaddr;
6142 			paddrs[i*nrxqs + j] = di->idi_paddr;
6143 		}
6144 	}
6145 	if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) {
6146 		device_printf(ctx->ifc_dev,
6147 		    "Unable to allocate device RX queue\n");
6148 		iflib_tx_structures_free(ctx);
6149 		free(vaddrs, M_IFLIB);
6150 		free(paddrs, M_IFLIB);
6151 		goto err_rx_desc;
6152 	}
6153 	free(vaddrs, M_IFLIB);
6154 	free(paddrs, M_IFLIB);
6155 
6156 	return (0);
6157 
6158 /* XXX handle allocation failure changes */
6159 err_rx_desc:
6160 err_tx_desc:
6161 rx_fail:
6162 	if (ctx->ifc_rxqs != NULL)
6163 		free(ctx->ifc_rxqs, M_IFLIB);
6164 	ctx->ifc_rxqs = NULL;
6165 	if (ctx->ifc_txqs != NULL)
6166 		free(ctx->ifc_txqs, M_IFLIB);
6167 	ctx->ifc_txqs = NULL;
6168 fail:
6169 	return (err);
6170 }
6171 
6172 static int
6173 iflib_tx_structures_setup(if_ctx_t ctx)
6174 {
6175 	iflib_txq_t txq = ctx->ifc_txqs;
6176 	int i;
6177 
6178 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
6179 		iflib_txq_setup(txq);
6180 
6181 	return (0);
6182 }
6183 
6184 static void
6185 iflib_tx_structures_free(if_ctx_t ctx)
6186 {
6187 	iflib_txq_t txq = ctx->ifc_txqs;
6188 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6189 	int i, j;
6190 
6191 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
6192 		for (j = 0; j < sctx->isc_ntxqs; j++)
6193 			iflib_dma_free(&txq->ift_ifdi[j]);
6194 		iflib_txq_destroy(txq);
6195 	}
6196 	free(ctx->ifc_txqs, M_IFLIB);
6197 	ctx->ifc_txqs = NULL;
6198 }
6199 
6200 /*********************************************************************
6201  *
6202  *  Initialize all receive rings.
6203  *
6204  **********************************************************************/
6205 static int
6206 iflib_rx_structures_setup(if_ctx_t ctx)
6207 {
6208 	iflib_rxq_t rxq = ctx->ifc_rxqs;
6209 	int q;
6210 #if defined(INET6) || defined(INET)
6211 	int err, i;
6212 #endif
6213 
6214 	for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) {
6215 #if defined(INET6) || defined(INET)
6216 		err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp,
6217 		    TCP_LRO_ENTRIES, min(1024,
6218 		    ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset]));
6219 		if (err != 0) {
6220 			device_printf(ctx->ifc_dev,
6221 			    "LRO Initialization failed!\n");
6222 			goto fail;
6223 		}
6224 #endif
6225 		IFDI_RXQ_SETUP(ctx, rxq->ifr_id);
6226 	}
6227 	return (0);
6228 #if defined(INET6) || defined(INET)
6229 fail:
6230 	/*
6231 	 * Free LRO resources allocated so far, we will only handle
6232 	 * the rings that completed, the failing case will have
6233 	 * cleaned up for itself.  'q' failed, so its the terminus.
6234 	 */
6235 	rxq = ctx->ifc_rxqs;
6236 	for (i = 0; i < q; ++i, rxq++) {
6237 		tcp_lro_free(&rxq->ifr_lc);
6238 	}
6239 	return (err);
6240 #endif
6241 }
6242 
6243 /*********************************************************************
6244  *
6245  *  Free all receive rings.
6246  *
6247  **********************************************************************/
6248 static void
6249 iflib_rx_structures_free(if_ctx_t ctx)
6250 {
6251 	iflib_rxq_t rxq = ctx->ifc_rxqs;
6252 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6253 	int i, j;
6254 
6255 	for (i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) {
6256 		for (j = 0; j < sctx->isc_nrxqs; j++)
6257 			iflib_dma_free(&rxq->ifr_ifdi[j]);
6258 		iflib_rx_sds_free(rxq);
6259 #if defined(INET6) || defined(INET)
6260 		tcp_lro_free(&rxq->ifr_lc);
6261 #endif
6262 	}
6263 	free(ctx->ifc_rxqs, M_IFLIB);
6264 	ctx->ifc_rxqs = NULL;
6265 }
6266 
6267 static int
6268 iflib_qset_structures_setup(if_ctx_t ctx)
6269 {
6270 	int err;
6271 
6272 	/*
6273 	 * It is expected that the caller takes care of freeing queues if this
6274 	 * fails.
6275 	 */
6276 	if ((err = iflib_tx_structures_setup(ctx)) != 0) {
6277 		device_printf(ctx->ifc_dev, "iflib_tx_structures_setup failed: %d\n", err);
6278 		return (err);
6279 	}
6280 
6281 	if ((err = iflib_rx_structures_setup(ctx)) != 0)
6282 		device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err);
6283 
6284 	return (err);
6285 }
6286 
6287 int
6288 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
6289 		driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, const char *name)
6290 {
6291 
6292 	return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name));
6293 }
6294 
6295 /* Just to avoid copy/paste */
6296 static inline int
6297 iflib_irq_set_affinity(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type,
6298     int qid, struct grouptask *gtask, struct taskqgroup *tqg, void *uniq,
6299     const char *name)
6300 {
6301 	device_t dev;
6302 	unsigned int base_cpuid, cpuid;
6303 	int err;
6304 
6305 	dev = ctx->ifc_dev;
6306 	base_cpuid = ctx->ifc_sysctl_core_offset;
6307 	cpuid = get_cpuid_for_queue(ctx, base_cpuid, qid, type == IFLIB_INTR_TX);
6308 	err = taskqgroup_attach_cpu(tqg, gtask, uniq, cpuid, dev,
6309 	    irq ? irq->ii_res : NULL, name);
6310 	if (err) {
6311 		device_printf(dev, "taskqgroup_attach_cpu failed %d\n", err);
6312 		return (err);
6313 	}
6314 #ifdef notyet
6315 	if (cpuid > ctx->ifc_cpuid_highest)
6316 		ctx->ifc_cpuid_highest = cpuid;
6317 #endif
6318 	return (0);
6319 }
6320 
6321 /*
6322  * Allocate a hardware interrupt for subctx using the parent (ctx)'s hardware
6323  * resources.
6324  *
6325  * Similar to iflib_irq_alloc_generic(), but for interrupt type IFLIB_INTR_RXTX
6326  * only.
6327  *
6328  * XXX: Could be removed if subctx's dev has its intr resource allocation
6329  * methods replaced with custom ones?
6330  */
6331 int
6332 iflib_irq_alloc_generic_subctx(if_ctx_t ctx, if_ctx_t subctx, if_irq_t irq,
6333 			       int rid, iflib_intr_type_t type,
6334 			       driver_filter_t *filter, void *filter_arg,
6335 			       int qid, const char *name)
6336 {
6337 	device_t dev, subdev;
6338 	struct grouptask *gtask;
6339 	struct taskqgroup *tqg;
6340 	iflib_filter_info_t info;
6341 	gtask_fn_t *fn;
6342 	int tqrid, err;
6343 	driver_filter_t *intr_fast;
6344 	void *q;
6345 
6346 	MPASS(ctx != NULL);
6347 	MPASS(subctx != NULL);
6348 
6349 	tqrid = rid;
6350 	dev = ctx->ifc_dev;
6351 	subdev = subctx->ifc_dev;
6352 
6353 	switch (type) {
6354 	case IFLIB_INTR_RXTX:
6355 		q = &subctx->ifc_rxqs[qid];
6356 		info = &subctx->ifc_rxqs[qid].ifr_filter_info;
6357 		gtask = &subctx->ifc_rxqs[qid].ifr_task;
6358 		tqg = qgroup_if_io_tqg;
6359 		fn = _task_fn_rx;
6360 		intr_fast = iflib_fast_intr_rxtx;
6361 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6362 		break;
6363 	default:
6364 		device_printf(dev, "%s: unknown net intr type for subctx %s (%d)\n",
6365 		    __func__, device_get_nameunit(subdev), type);
6366 		return (EINVAL);
6367 	}
6368 
6369 	info->ifi_filter = filter;
6370 	info->ifi_filter_arg = filter_arg;
6371 	info->ifi_task = gtask;
6372 	info->ifi_ctx = q;
6373 
6374 	NET_GROUPTASK_INIT(gtask, 0, fn, q);
6375 
6376 	/* Allocate interrupts from hardware using parent context */
6377 	err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info, name);
6378 	if (err != 0) {
6379 		device_printf(dev, "_iflib_irq_alloc failed for subctx %s: %d\n",
6380 		    device_get_nameunit(subdev), err);
6381 		return (err);
6382 	}
6383 
6384 	if (tqrid != -1) {
6385 		err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg, q,
6386 		    name);
6387 		if (err)
6388 			return (err);
6389 	} else {
6390 		taskqgroup_attach(tqg, gtask, q, dev, irq->ii_res, name);
6391 	}
6392 
6393 	return (0);
6394 }
6395 
6396 int
6397 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid,
6398 			iflib_intr_type_t type, driver_filter_t *filter,
6399 			void *filter_arg, int qid, const char *name)
6400 {
6401 	device_t dev;
6402 	struct grouptask *gtask;
6403 	struct taskqgroup *tqg;
6404 	iflib_filter_info_t info;
6405 	gtask_fn_t *fn;
6406 	int tqrid, err;
6407 	driver_filter_t *intr_fast;
6408 	void *q;
6409 
6410 	info = &ctx->ifc_filter_info;
6411 	tqrid = rid;
6412 
6413 	switch (type) {
6414 	/* XXX merge tx/rx for netmap? */
6415 	case IFLIB_INTR_TX:
6416 		q = &ctx->ifc_txqs[qid];
6417 		info = &ctx->ifc_txqs[qid].ift_filter_info;
6418 		gtask = &ctx->ifc_txqs[qid].ift_task;
6419 		tqg = qgroup_if_io_tqg;
6420 		fn = _task_fn_tx;
6421 		intr_fast = iflib_fast_intr;
6422 		GROUPTASK_INIT(gtask, 0, fn, q);
6423 		ctx->ifc_flags |= IFC_NETMAP_TX_IRQ;
6424 		break;
6425 	case IFLIB_INTR_RX:
6426 		q = &ctx->ifc_rxqs[qid];
6427 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
6428 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6429 		tqg = qgroup_if_io_tqg;
6430 		fn = _task_fn_rx;
6431 		intr_fast = iflib_fast_intr;
6432 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6433 		break;
6434 	case IFLIB_INTR_RXTX:
6435 		q = &ctx->ifc_rxqs[qid];
6436 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
6437 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6438 		tqg = qgroup_if_io_tqg;
6439 		fn = _task_fn_rx;
6440 		intr_fast = iflib_fast_intr_rxtx;
6441 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6442 		break;
6443 	case IFLIB_INTR_ADMIN:
6444 		q = ctx;
6445 		tqrid = -1;
6446 		info = &ctx->ifc_filter_info;
6447 		gtask = &ctx->ifc_admin_task;
6448 		tqg = qgroup_if_config_tqg;
6449 		fn = _task_fn_admin;
6450 		intr_fast = iflib_fast_intr_ctx;
6451 		break;
6452 	default:
6453 		device_printf(ctx->ifc_dev, "%s: unknown net intr type\n",
6454 		    __func__);
6455 		return (EINVAL);
6456 	}
6457 
6458 	info->ifi_filter = filter;
6459 	info->ifi_filter_arg = filter_arg;
6460 	info->ifi_task = gtask;
6461 	info->ifi_ctx = q;
6462 
6463 	dev = ctx->ifc_dev;
6464 	err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info,  name);
6465 	if (err != 0) {
6466 		device_printf(dev, "_iflib_irq_alloc failed %d\n", err);
6467 		return (err);
6468 	}
6469 	if (type == IFLIB_INTR_ADMIN)
6470 		return (0);
6471 
6472 	if (tqrid != -1) {
6473 		err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg, q,
6474 		    name);
6475 		if (err)
6476 			return (err);
6477 	} else {
6478 		taskqgroup_attach(tqg, gtask, q, dev, irq->ii_res, name);
6479 	}
6480 
6481 	return (0);
6482 }
6483 
6484 void
6485 iflib_softirq_alloc_generic(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type,
6486 			    void *arg, int qid, const char *name)
6487 {
6488 	device_t dev;
6489 	struct grouptask *gtask;
6490 	struct taskqgroup *tqg;
6491 	gtask_fn_t *fn;
6492 	void *q;
6493 	int err;
6494 
6495 	switch (type) {
6496 	case IFLIB_INTR_TX:
6497 		q = &ctx->ifc_txqs[qid];
6498 		gtask = &ctx->ifc_txqs[qid].ift_task;
6499 		tqg = qgroup_if_io_tqg;
6500 		fn = _task_fn_tx;
6501 		GROUPTASK_INIT(gtask, 0, fn, q);
6502 		break;
6503 	case IFLIB_INTR_RX:
6504 		q = &ctx->ifc_rxqs[qid];
6505 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6506 		tqg = qgroup_if_io_tqg;
6507 		fn = _task_fn_rx;
6508 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6509 		break;
6510 	case IFLIB_INTR_IOV:
6511 		q = ctx;
6512 		gtask = &ctx->ifc_vflr_task;
6513 		tqg = qgroup_if_config_tqg;
6514 		fn = _task_fn_iov;
6515 		GROUPTASK_INIT(gtask, 0, fn, q);
6516 		break;
6517 	default:
6518 		panic("unknown net intr type");
6519 	}
6520 	err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg, q, name);
6521 	if (err) {
6522 		dev = ctx->ifc_dev;
6523 		taskqgroup_attach(tqg, gtask, q, dev, irq ? irq->ii_res : NULL,
6524 		    name);
6525 	}
6526 }
6527 
6528 void
6529 iflib_irq_free(if_ctx_t ctx, if_irq_t irq)
6530 {
6531 #ifdef __HAIKU__
6532 	if (!ctx || !irq)
6533 		return;
6534 #endif
6535 
6536 	if (irq->ii_tag)
6537 		bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag);
6538 
6539 	if (irq->ii_res)
6540 		bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ,
6541 		    rman_get_rid(irq->ii_res), irq->ii_res);
6542 }
6543 
6544 static int
6545 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, const char *name)
6546 {
6547 	iflib_txq_t txq = ctx->ifc_txqs;
6548 	iflib_rxq_t rxq = ctx->ifc_rxqs;
6549 	if_irq_t irq = &ctx->ifc_legacy_irq;
6550 	iflib_filter_info_t info;
6551 	device_t dev;
6552 	struct grouptask *gtask;
6553 	struct resource *res;
6554 	struct taskqgroup *tqg;
6555 	void *q;
6556 	int err, tqrid;
6557 	bool rx_only;
6558 
6559 	q = &ctx->ifc_rxqs[0];
6560 	info = &rxq[0].ifr_filter_info;
6561 	gtask = &rxq[0].ifr_task;
6562 	tqg = qgroup_if_io_tqg;
6563 	tqrid = *rid;
6564 	rx_only = (ctx->ifc_sctx->isc_flags & IFLIB_SINGLE_IRQ_RX_ONLY) != 0;
6565 
6566 	ctx->ifc_flags |= IFC_LEGACY;
6567 	info->ifi_filter = filter;
6568 	info->ifi_filter_arg = filter_arg;
6569 	info->ifi_task = gtask;
6570 	info->ifi_ctx = rx_only ? ctx : q;
6571 
6572 	dev = ctx->ifc_dev;
6573 	/* We allocate a single interrupt resource */
6574 	err = _iflib_irq_alloc(ctx, irq, tqrid, rx_only ? iflib_fast_intr_ctx :
6575 	    iflib_fast_intr_rxtx, NULL, info, name);
6576 	if (err != 0)
6577 		return (err);
6578 	NET_GROUPTASK_INIT(gtask, 0, _task_fn_rx, q);
6579 	res = irq->ii_res;
6580 	taskqgroup_attach(tqg, gtask, q, dev, res, name);
6581 
6582 	GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq);
6583 	taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, dev, res,
6584 	    "tx");
6585 	return (0);
6586 }
6587 
6588 void
6589 iflib_led_create(if_ctx_t ctx)
6590 {
6591 
6592 	ctx->ifc_led_dev = led_create(iflib_led_func, ctx,
6593 	    device_get_nameunit(ctx->ifc_dev));
6594 }
6595 
6596 void
6597 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid)
6598 {
6599 
6600 	GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
6601 }
6602 
6603 void
6604 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid)
6605 {
6606 
6607 	GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task);
6608 }
6609 
6610 void
6611 iflib_admin_intr_deferred(if_ctx_t ctx)
6612 {
6613 
6614 	MPASS(ctx->ifc_admin_task.gt_taskqueue != NULL);
6615 	GROUPTASK_ENQUEUE(&ctx->ifc_admin_task);
6616 }
6617 
6618 void
6619 iflib_iov_intr_deferred(if_ctx_t ctx)
6620 {
6621 
6622 	GROUPTASK_ENQUEUE(&ctx->ifc_vflr_task);
6623 }
6624 
6625 void
6626 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, const char *name)
6627 {
6628 
6629 	taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, NULL, NULL,
6630 	    name);
6631 }
6632 
6633 void
6634 iflib_config_gtask_init(void *ctx, struct grouptask *gtask, gtask_fn_t *fn,
6635 	const char *name)
6636 {
6637 
6638 	GROUPTASK_INIT(gtask, 0, fn, ctx);
6639 	taskqgroup_attach(qgroup_if_config_tqg, gtask, gtask, NULL, NULL,
6640 	    name);
6641 }
6642 
6643 void
6644 iflib_config_gtask_deinit(struct grouptask *gtask)
6645 {
6646 
6647 	taskqgroup_detach(qgroup_if_config_tqg, gtask);
6648 }
6649 
6650 void
6651 iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate)
6652 {
6653 	if_t ifp = ctx->ifc_ifp;
6654 	iflib_txq_t txq = ctx->ifc_txqs;
6655 
6656 	if_setbaudrate(ifp, baudrate);
6657 	if (baudrate >= IF_Gbps(10)) {
6658 		STATE_LOCK(ctx);
6659 		ctx->ifc_flags |= IFC_PREFETCH;
6660 		STATE_UNLOCK(ctx);
6661 	}
6662 	/* If link down, disable watchdog */
6663 	if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) {
6664 		for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++)
6665 			txq->ift_qstatus = IFLIB_QUEUE_IDLE;
6666 	}
6667 	ctx->ifc_link_state = link_state;
6668 	if_link_state_change(ifp, link_state);
6669 }
6670 
6671 static int
6672 iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq)
6673 {
6674 	int credits;
6675 #ifdef INVARIANTS
6676 	int credits_pre = txq->ift_cidx_processed;
6677 #endif
6678 
6679 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
6680 	    BUS_DMASYNC_POSTREAD);
6681 	if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, true)) == 0)
6682 		return (0);
6683 
6684 	txq->ift_processed += credits;
6685 	txq->ift_cidx_processed += credits;
6686 
6687 	MPASS(credits_pre + credits == txq->ift_cidx_processed);
6688 	if (txq->ift_cidx_processed >= txq->ift_size)
6689 		txq->ift_cidx_processed -= txq->ift_size;
6690 	return (credits);
6691 }
6692 
6693 static int
6694 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget)
6695 {
6696 	iflib_fl_t fl;
6697 	u_int i;
6698 
6699 	for (i = 0, fl = &rxq->ifr_fl[0]; i < rxq->ifr_nfl; i++, fl++)
6700 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
6701 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
6702 	return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx,
6703 	    budget));
6704 }
6705 
6706 void
6707 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name,
6708 	const char *description, if_int_delay_info_t info,
6709 	int offset, int value)
6710 {
6711 	info->iidi_ctx = ctx;
6712 	info->iidi_offset = offset;
6713 	info->iidi_value = value;
6714 	SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev),
6715 	    SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)),
6716 	    OID_AUTO, name, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE,
6717 	    info, 0, iflib_sysctl_int_delay, "I", description);
6718 }
6719 
6720 struct sx *
6721 iflib_ctx_lock_get(if_ctx_t ctx)
6722 {
6723 
6724 	return (&ctx->ifc_ctx_sx);
6725 }
6726 
6727 static int
6728 iflib_msix_init(if_ctx_t ctx)
6729 {
6730 	device_t dev = ctx->ifc_dev;
6731 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6732 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6733 	int admincnt, bar, err, iflib_num_rx_queues, iflib_num_tx_queues;
6734 	int msgs, queuemsgs, queues, rx_queues, tx_queues, vectors;
6735 
6736 	iflib_num_tx_queues = ctx->ifc_sysctl_ntxqs;
6737 	iflib_num_rx_queues = ctx->ifc_sysctl_nrxqs;
6738 
6739 	if (bootverbose)
6740 		device_printf(dev, "msix_init qsets capped at %d\n",
6741 		    imax(scctx->isc_ntxqsets, scctx->isc_nrxqsets));
6742 
6743 	/* Override by tuneable */
6744 	if (scctx->isc_disable_msix)
6745 		goto msi;
6746 
6747 	/* First try MSI-X */
6748 	if ((msgs = pci_msix_count(dev)) == 0) {
6749 		if (bootverbose)
6750 			device_printf(dev, "MSI-X not supported or disabled\n");
6751 		goto msi;
6752 	}
6753 
6754 	bar = ctx->ifc_softc_ctx.isc_msix_bar;
6755 	/*
6756 	 * bar == -1 => "trust me I know what I'm doing"
6757 	 * Some drivers are for hardware that is so shoddily
6758 	 * documented that no one knows which bars are which
6759 	 * so the developer has to map all bars. This hack
6760 	 * allows shoddy garbage to use MSI-X in this framework.
6761 	 */
6762 	if (bar != -1) {
6763 		ctx->ifc_msix_mem = bus_alloc_resource_any(dev,
6764 	            SYS_RES_MEMORY, &bar, RF_ACTIVE);
6765 		if (ctx->ifc_msix_mem == NULL) {
6766 			device_printf(dev, "Unable to map MSI-X table\n");
6767 			goto msi;
6768 		}
6769 	}
6770 
6771 	admincnt = sctx->isc_admin_intrcnt;
6772 #if IFLIB_DEBUG
6773 	/* use only 1 qset in debug mode */
6774 	queuemsgs = min(msgs - admincnt, 1);
6775 #else
6776 	queuemsgs = msgs - admincnt;
6777 #endif
6778 #ifdef RSS
6779 	queues = imin(queuemsgs, rss_getnumbuckets());
6780 #else
6781 	queues = queuemsgs;
6782 #endif
6783 #ifndef __HAIKU__
6784 	queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues);
6785 	if (bootverbose)
6786 		device_printf(dev,
6787 		    "intr CPUs: %d queue msgs: %d admincnt: %d\n",
6788 		    CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt);
6789 #endif
6790 #ifdef  RSS
6791 	/* If we're doing RSS, clamp at the number of RSS buckets */
6792 	if (queues > rss_getnumbuckets())
6793 		queues = rss_getnumbuckets();
6794 #endif
6795 	if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt)
6796 		rx_queues = iflib_num_rx_queues;
6797 	else
6798 		rx_queues = queues;
6799 
6800 	if (rx_queues > scctx->isc_nrxqsets)
6801 		rx_queues = scctx->isc_nrxqsets;
6802 
6803 	/*
6804 	 * We want this to be all logical CPUs by default
6805 	 */
6806 	if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues)
6807 		tx_queues = iflib_num_tx_queues;
6808 	else
6809 		tx_queues = mp_ncpus;
6810 
6811 	if (tx_queues > scctx->isc_ntxqsets)
6812 		tx_queues = scctx->isc_ntxqsets;
6813 
6814 	if (ctx->ifc_sysctl_qs_eq_override == 0) {
6815 #ifdef INVARIANTS
6816 		if (tx_queues != rx_queues)
6817 			device_printf(dev,
6818 			    "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n",
6819 			    min(rx_queues, tx_queues), min(rx_queues, tx_queues));
6820 #endif
6821 		tx_queues = min(rx_queues, tx_queues);
6822 		rx_queues = min(rx_queues, tx_queues);
6823 	}
6824 
6825 	vectors = rx_queues + admincnt;
6826 	if (msgs < vectors) {
6827 		device_printf(dev,
6828 		    "insufficient number of MSI-X vectors "
6829 		    "(supported %d, need %d)\n", msgs, vectors);
6830 		goto msi;
6831 	}
6832 
6833 	device_printf(dev, "Using %d RX queues %d TX queues\n", rx_queues,
6834 	    tx_queues);
6835 	msgs = vectors;
6836 	if ((err = pci_alloc_msix(dev, &vectors)) == 0) {
6837 		if (vectors != msgs) {
6838 			device_printf(dev,
6839 			    "Unable to allocate sufficient MSI-X vectors "
6840 			    "(got %d, need %d)\n", vectors, msgs);
6841 			pci_release_msi(dev);
6842 			if (bar != -1) {
6843 				bus_release_resource(dev, SYS_RES_MEMORY, bar,
6844 				    ctx->ifc_msix_mem);
6845 				ctx->ifc_msix_mem = NULL;
6846 			}
6847 			goto msi;
6848 		}
6849 		device_printf(dev, "Using MSI-X interrupts with %d vectors\n",
6850 		    vectors);
6851 		scctx->isc_vectors = vectors;
6852 		scctx->isc_nrxqsets = rx_queues;
6853 		scctx->isc_ntxqsets = tx_queues;
6854 		scctx->isc_intr = IFLIB_INTR_MSIX;
6855 
6856 		return (vectors);
6857 	} else {
6858 		device_printf(dev,
6859 		    "failed to allocate %d MSI-X vectors, err: %d\n", vectors,
6860 		    err);
6861 		if (bar != -1) {
6862 			bus_release_resource(dev, SYS_RES_MEMORY, bar,
6863 			    ctx->ifc_msix_mem);
6864 			ctx->ifc_msix_mem = NULL;
6865 		}
6866 	}
6867 
6868 msi:
6869 	vectors = pci_msi_count(dev);
6870 	scctx->isc_nrxqsets = 1;
6871 	scctx->isc_ntxqsets = 1;
6872 	scctx->isc_vectors = vectors;
6873 	if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) {
6874 		device_printf(dev,"Using an MSI interrupt\n");
6875 		scctx->isc_intr = IFLIB_INTR_MSI;
6876 	} else {
6877 		scctx->isc_vectors = 1;
6878 		device_printf(dev,"Using a Legacy interrupt\n");
6879 		scctx->isc_intr = IFLIB_INTR_LEGACY;
6880 	}
6881 
6882 	return (vectors);
6883 }
6884 
6885 static const char *ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" };
6886 
6887 #ifndef __HAIKU__
6888 static int
6889 mp_ring_state_handler(SYSCTL_HANDLER_ARGS)
6890 {
6891 	int rc;
6892 	uint16_t *state = ((uint16_t *)oidp->oid_arg1);
6893 	struct sbuf *sb;
6894 	const char *ring_state = "UNKNOWN";
6895 
6896 	/* XXX needed ? */
6897 	rc = sysctl_wire_old_buffer(req, 0);
6898 	MPASS(rc == 0);
6899 	if (rc != 0)
6900 		return (rc);
6901 	sb = sbuf_new_for_sysctl(NULL, NULL, 80, req);
6902 	MPASS(sb != NULL);
6903 	if (sb == NULL)
6904 		return (ENOMEM);
6905 	if (state[3] <= 3)
6906 		ring_state = ring_states[state[3]];
6907 
6908 	sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s",
6909 		    state[0], state[1], state[2], ring_state);
6910 	rc = sbuf_finish(sb);
6911 	sbuf_delete(sb);
6912         return(rc);
6913 }
6914 #endif
6915 
6916 enum iflib_ndesc_handler {
6917 	IFLIB_NTXD_HANDLER,
6918 	IFLIB_NRXD_HANDLER,
6919 };
6920 
6921 static int
6922 mp_ndesc_handler(SYSCTL_HANDLER_ARGS)
6923 {
6924 	if_ctx_t ctx = (void *)arg1;
6925 	enum iflib_ndesc_handler type = arg2;
6926 	char buf[256] = {0};
6927 	qidx_t *ndesc;
6928 	char *p, *next;
6929 	int nqs, rc, i;
6930 
6931 	nqs = 8;
6932 	switch(type) {
6933 	case IFLIB_NTXD_HANDLER:
6934 		ndesc = ctx->ifc_sysctl_ntxds;
6935 		if (ctx->ifc_sctx)
6936 			nqs = ctx->ifc_sctx->isc_ntxqs;
6937 		break;
6938 	case IFLIB_NRXD_HANDLER:
6939 		ndesc = ctx->ifc_sysctl_nrxds;
6940 		if (ctx->ifc_sctx)
6941 			nqs = ctx->ifc_sctx->isc_nrxqs;
6942 		break;
6943 	default:
6944 		printf("%s: unhandled type\n", __func__);
6945 		return (EINVAL);
6946 	}
6947 	if (nqs == 0)
6948 		nqs = 8;
6949 
6950 	for (i=0; i<8; i++) {
6951 		if (i >= nqs)
6952 			break;
6953 		if (i)
6954 			strcat(buf, ",");
6955 		sprintf(strchr(buf, 0), "%d", ndesc[i]);
6956 	}
6957 
6958 	rc = sysctl_handle_string(oidp, buf, sizeof(buf), req);
6959 	if (rc || req->newptr == NULL)
6960 		return rc;
6961 
6962 	for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p;
6963 	    i++, p = strsep(&next, " ,")) {
6964 		ndesc[i] = strtoul(p, NULL, 10);
6965 	}
6966 
6967 	return(rc);
6968 }
6969 
6970 #define NAME_BUFLEN 32
6971 static void
6972 iflib_add_device_sysctl_pre(if_ctx_t ctx)
6973 {
6974 #ifndef __HAIKU__
6975         device_t dev = iflib_get_dev(ctx);
6976 	struct sysctl_oid_list *child, *oid_list;
6977 	struct sysctl_ctx_list *ctx_list;
6978 	struct sysctl_oid *node;
6979 
6980 	ctx_list = device_get_sysctl_ctx(dev);
6981 	child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev));
6982 	ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, "iflib",
6983 	    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "IFLIB fields");
6984 	oid_list = SYSCTL_CHILDREN(node);
6985 
6986 	SYSCTL_ADD_CONST_STRING(ctx_list, oid_list, OID_AUTO, "driver_version",
6987 		       CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version,
6988 		       "driver version");
6989 
6990 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs",
6991 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0,
6992 		       "# of txqs to use, 0 => use default #");
6993 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs",
6994 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0,
6995 		       "# of rxqs to use, 0 => use default #");
6996 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_qs_enable",
6997 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0,
6998                        "permit #txq != #rxq");
6999 	SYSCTL_ADD_INT(ctx_list, oid_list, OID_AUTO, "disable_msix",
7000 		       CTLFLAG_RWTUN, &ctx->ifc_softc_ctx.isc_disable_msix, 0,
7001 		       "disable MSI-X (default 0)");
7002 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "rx_budget",
7003 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_rx_budget, 0,
7004 		       "set the RX budget");
7005 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "tx_abdicate",
7006 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_tx_abdicate, 0,
7007 		       "cause TX to abdicate instead of running to completion");
7008 	ctx->ifc_sysctl_core_offset = CORE_OFFSET_UNSPECIFIED;
7009 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "core_offset",
7010 		       CTLFLAG_RDTUN, &ctx->ifc_sysctl_core_offset, 0,
7011 		       "offset to start using cores at");
7012 	SYSCTL_ADD_U8(ctx_list, oid_list, OID_AUTO, "separate_txrx",
7013 		      CTLFLAG_RDTUN, &ctx->ifc_sysctl_separate_txrx, 0,
7014 		      "use separate cores for TX and RX");
7015 	SYSCTL_ADD_U8(ctx_list, oid_list, OID_AUTO, "use_logical_cores",
7016 		      CTLFLAG_RDTUN, &ctx->ifc_sysctl_use_logical_cores, 0,
7017 		      "try to make use of logical cores for TX and RX");
7018 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "use_extra_msix_vectors",
7019 	    CTLFLAG_RDTUN, &ctx->ifc_sysctl_extra_msix_vectors, 0,
7020 	    "attempt to reserve the given number of extra MSI-X vectors during driver load for the creation of additional interfaces later");
7021 	SYSCTL_ADD_INT(ctx_list, oid_list, OID_AUTO, "allocated_msix_vectors",
7022        	    CTLFLAG_RDTUN, &ctx->ifc_softc_ctx.isc_vectors, 0,
7023 	    "total # of MSI-X vectors allocated by driver");
7024 
7025 	/* XXX change for per-queue sizes */
7026 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_ntxds",
7027 	    CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NEEDGIANT, ctx,
7028 	    IFLIB_NTXD_HANDLER, mp_ndesc_handler, "A",
7029 	    "list of # of TX descriptors to use, 0 = use default #");
7030 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_nrxds",
7031 	    CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NEEDGIANT, ctx,
7032 	    IFLIB_NRXD_HANDLER, mp_ndesc_handler, "A",
7033 	    "list of # of RX descriptors to use, 0 = use default #");
7034 #endif
7035 }
7036 
7037 static void
7038 iflib_add_device_sysctl_post(if_ctx_t ctx)
7039 {
7040 #ifndef __HAIKU__
7041 	if_shared_ctx_t sctx = ctx->ifc_sctx;
7042 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
7043         device_t dev = iflib_get_dev(ctx);
7044 	struct sysctl_oid_list *child;
7045 	struct sysctl_ctx_list *ctx_list;
7046 	iflib_fl_t fl;
7047 	iflib_txq_t txq;
7048 	iflib_rxq_t rxq;
7049 	int i, j;
7050 	char namebuf[NAME_BUFLEN];
7051 	char *qfmt;
7052 	struct sysctl_oid *queue_node, *fl_node, *node;
7053 	struct sysctl_oid_list *queue_list, *fl_list;
7054 	ctx_list = device_get_sysctl_ctx(dev);
7055 
7056 	node = ctx->ifc_sysctl_node;
7057 	child = SYSCTL_CHILDREN(node);
7058 
7059 	if (scctx->isc_ntxqsets > 100)
7060 		qfmt = "txq%03d";
7061 	else if (scctx->isc_ntxqsets > 10)
7062 		qfmt = "txq%02d";
7063 	else
7064 		qfmt = "txq%d";
7065 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
7066 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
7067 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
7068 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "Queue Name");
7069 		queue_list = SYSCTL_CHILDREN(queue_node);
7070 		SYSCTL_ADD_INT(ctx_list, queue_list, OID_AUTO, "cpu",
7071 			       CTLFLAG_RD,
7072 			       &txq->ift_task.gt_cpu, 0, "cpu this queue is bound to");
7073 #if MEMORY_LOGGING
7074 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued",
7075 				CTLFLAG_RD,
7076 				&txq->ift_dequeued, "total mbufs freed");
7077 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued",
7078 				CTLFLAG_RD,
7079 				&txq->ift_enqueued, "total mbufs enqueued");
7080 #endif
7081 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag",
7082 				   CTLFLAG_RD,
7083 				   &txq->ift_mbuf_defrag, "# of times m_defrag was called");
7084 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "m_pullups",
7085 				   CTLFLAG_RD,
7086 				   &txq->ift_pullups, "# of times m_pullup was called");
7087 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag_failed",
7088 				   CTLFLAG_RD,
7089 				   &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed");
7090 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_desc_avail",
7091 				   CTLFLAG_RD,
7092 				   &txq->ift_no_desc_avail, "# of times no descriptors were available");
7093 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "tx_map_failed",
7094 				   CTLFLAG_RD,
7095 				   &txq->ift_map_failed, "# of times DMA map failed");
7096 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txd_encap_efbig",
7097 				   CTLFLAG_RD,
7098 				   &txq->ift_txd_encap_efbig, "# of times txd_encap returned EFBIG");
7099 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_tx_dma_setup",
7100 				   CTLFLAG_RD,
7101 				   &txq->ift_no_tx_dma_setup, "# of times map failed for other than EFBIG");
7102 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx",
7103 				   CTLFLAG_RD,
7104 				   &txq->ift_pidx, 1, "Producer Index");
7105 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx",
7106 				   CTLFLAG_RD,
7107 				   &txq->ift_cidx, 1, "Consumer Index");
7108 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx_processed",
7109 				   CTLFLAG_RD,
7110 				   &txq->ift_cidx_processed, 1, "Consumer Index seen by credit update");
7111 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use",
7112 				   CTLFLAG_RD,
7113 				   &txq->ift_in_use, 1, "descriptors in use");
7114 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_processed",
7115 				   CTLFLAG_RD,
7116 				   &txq->ift_processed, "descriptors procesed for clean");
7117 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned",
7118 				   CTLFLAG_RD,
7119 				   &txq->ift_cleaned, "total cleaned");
7120 		SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state",
7121 		    CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_NEEDGIANT,
7122 		    __DEVOLATILE(uint64_t *, &txq->ift_br->state), 0,
7123 		    mp_ring_state_handler, "A", "soft ring state");
7124 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_enqueues",
7125 				       CTLFLAG_RD, &txq->ift_br->enqueues,
7126 				       "# of enqueues to the mp_ring for this queue");
7127 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_drops",
7128 				       CTLFLAG_RD, &txq->ift_br->drops,
7129 				       "# of drops in the mp_ring for this queue");
7130 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_starts",
7131 				       CTLFLAG_RD, &txq->ift_br->starts,
7132 				       "# of normal consumer starts in the mp_ring for this queue");
7133 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_stalls",
7134 				       CTLFLAG_RD, &txq->ift_br->stalls,
7135 					       "# of consumer stalls in the mp_ring for this queue");
7136 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_restarts",
7137 			       CTLFLAG_RD, &txq->ift_br->restarts,
7138 				       "# of consumer restarts in the mp_ring for this queue");
7139 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_abdications",
7140 				       CTLFLAG_RD, &txq->ift_br->abdications,
7141 				       "# of consumer abdications in the mp_ring for this queue");
7142 	}
7143 
7144 	if (scctx->isc_nrxqsets > 100)
7145 		qfmt = "rxq%03d";
7146 	else if (scctx->isc_nrxqsets > 10)
7147 		qfmt = "rxq%02d";
7148 	else
7149 		qfmt = "rxq%d";
7150 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
7151 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
7152 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
7153 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "Queue Name");
7154 		queue_list = SYSCTL_CHILDREN(queue_node);
7155 		SYSCTL_ADD_INT(ctx_list, queue_list, OID_AUTO, "cpu",
7156 			       CTLFLAG_RD,
7157 			       &rxq->ifr_task.gt_cpu, 0, "cpu this queue is bound to");
7158 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
7159 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_cidx",
7160 				       CTLFLAG_RD,
7161 				       &rxq->ifr_cq_cidx, 1, "Consumer Index");
7162 		}
7163 
7164 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
7165 			snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j);
7166 			fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list, OID_AUTO, namebuf,
7167 			    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "freelist Name");
7168 			fl_list = SYSCTL_CHILDREN(fl_node);
7169 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx",
7170 				       CTLFLAG_RD,
7171 				       &fl->ifl_pidx, 1, "Producer Index");
7172 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx",
7173 				       CTLFLAG_RD,
7174 				       &fl->ifl_cidx, 1, "Consumer Index");
7175 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits",
7176 				       CTLFLAG_RD,
7177 				       &fl->ifl_credits, 1, "credits available");
7178 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "buf_size",
7179 				       CTLFLAG_RD,
7180 				       &fl->ifl_buf_size, 1, "buffer size");
7181 #if MEMORY_LOGGING
7182 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_enqueued",
7183 					CTLFLAG_RD,
7184 					&fl->ifl_m_enqueued, "mbufs allocated");
7185 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_dequeued",
7186 					CTLFLAG_RD,
7187 					&fl->ifl_m_dequeued, "mbufs freed");
7188 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_enqueued",
7189 					CTLFLAG_RD,
7190 					&fl->ifl_cl_enqueued, "clusters allocated");
7191 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_dequeued",
7192 					CTLFLAG_RD,
7193 					&fl->ifl_cl_dequeued, "clusters freed");
7194 #endif
7195 		}
7196 	}
7197 #endif
7198 }
7199 
7200 void
7201 iflib_request_reset(if_ctx_t ctx)
7202 {
7203 
7204 	STATE_LOCK(ctx);
7205 	ctx->ifc_flags |= IFC_DO_RESET;
7206 	STATE_UNLOCK(ctx);
7207 }
7208 
7209 #ifndef __NO_STRICT_ALIGNMENT
7210 static struct mbuf *
7211 iflib_fixup_rx(struct mbuf *m)
7212 {
7213 	struct mbuf *n;
7214 
7215 	if (m->m_len <= (MCLBYTES - ETHER_HDR_LEN)) {
7216 		bcopy(m->m_data, m->m_data + ETHER_HDR_LEN, m->m_len);
7217 		m->m_data += ETHER_HDR_LEN;
7218 		n = m;
7219 	} else {
7220 		MGETHDR(n, M_NOWAIT, MT_DATA);
7221 		if (n == NULL) {
7222 			m_freem(m);
7223 			return (NULL);
7224 		}
7225 		bcopy(m->m_data, n->m_data, ETHER_HDR_LEN);
7226 		m->m_data += ETHER_HDR_LEN;
7227 		m->m_len -= ETHER_HDR_LEN;
7228 		n->m_len = ETHER_HDR_LEN;
7229 		M_MOVE_PKTHDR(n, m);
7230 		n->m_next = m;
7231 	}
7232 	return (n);
7233 }
7234 #endif
7235 
7236 #ifdef DEBUGNET
7237 static void
7238 iflib_debugnet_init(if_t ifp, int *nrxr, int *ncl, int *clsize)
7239 {
7240 	if_ctx_t ctx;
7241 
7242 	ctx = if_getsoftc(ifp);
7243 	CTX_LOCK(ctx);
7244 	*nrxr = NRXQSETS(ctx);
7245 	*ncl = ctx->ifc_rxqs[0].ifr_fl->ifl_size;
7246 	*clsize = ctx->ifc_rxqs[0].ifr_fl->ifl_buf_size;
7247 	CTX_UNLOCK(ctx);
7248 }
7249 
7250 static void
7251 iflib_debugnet_event(if_t ifp, enum debugnet_ev event)
7252 {
7253 	if_ctx_t ctx;
7254 	if_softc_ctx_t scctx;
7255 	iflib_fl_t fl;
7256 	iflib_rxq_t rxq;
7257 	int i, j;
7258 
7259 	ctx = if_getsoftc(ifp);
7260 	scctx = &ctx->ifc_softc_ctx;
7261 
7262 	switch (event) {
7263 	case DEBUGNET_START:
7264 #ifndef __HAIKU__
7265 		for (i = 0; i < scctx->isc_nrxqsets; i++) {
7266 			rxq = &ctx->ifc_rxqs[i];
7267 			for (j = 0; j < rxq->ifr_nfl; j++) {
7268 				fl = rxq->ifr_fl;
7269 				fl->ifl_zone = m_getzone(fl->ifl_buf_size);
7270 			}
7271 		}
7272 		iflib_no_tx_batch = 1;
7273 		break;
7274 #endif
7275 	default:
7276 		break;
7277 	}
7278 }
7279 
7280 static int
7281 iflib_debugnet_transmit(if_t ifp, struct mbuf *m)
7282 {
7283 	if_ctx_t ctx;
7284 	iflib_txq_t txq;
7285 	int error;
7286 
7287 	ctx = if_getsoftc(ifp);
7288 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
7289 	    IFF_DRV_RUNNING)
7290 		return (EBUSY);
7291 
7292 	txq = &ctx->ifc_txqs[0];
7293 	error = iflib_encap(txq, &m);
7294 	if (error == 0)
7295 		(void)iflib_txd_db_check(txq, true);
7296 	return (error);
7297 }
7298 
7299 static int
7300 iflib_debugnet_poll(if_t ifp, int count)
7301 {
7302 	struct epoch_tracker et;
7303 	if_ctx_t ctx;
7304 	if_softc_ctx_t scctx;
7305 	iflib_txq_t txq;
7306 	int i;
7307 
7308 	ctx = if_getsoftc(ifp);
7309 	scctx = &ctx->ifc_softc_ctx;
7310 
7311 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
7312 	    IFF_DRV_RUNNING)
7313 		return (EBUSY);
7314 
7315 	txq = &ctx->ifc_txqs[0];
7316 	(void)iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
7317 
7318 	NET_EPOCH_ENTER(et);
7319 	for (i = 0; i < scctx->isc_nrxqsets; i++)
7320 		(void)iflib_rxeof(&ctx->ifc_rxqs[i], 16 /* XXX */);
7321 	NET_EPOCH_EXIT(et);
7322 	return (0);
7323 }
7324 #endif /* DEBUGNET */
7325