xref: /freebsd/sys/net/iflib.c (revision e2eeea75eb8b6dd50c1298067a0655880d186734)
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 __FBSDID("$FreeBSD$");
30 
31 #include "opt_inet.h"
32 #include "opt_inet6.h"
33 #include "opt_acpi.h"
34 #include "opt_sched.h"
35 
36 #include <sys/param.h>
37 #include <sys/types.h>
38 #include <sys/bus.h>
39 #include <sys/eventhandler.h>
40 #include <sys/kernel.h>
41 #include <sys/lock.h>
42 #include <sys/mutex.h>
43 #include <sys/module.h>
44 #include <sys/kobj.h>
45 #include <sys/rman.h>
46 #include <sys/sbuf.h>
47 #include <sys/smp.h>
48 #include <sys/socket.h>
49 #include <sys/sockio.h>
50 #include <sys/sysctl.h>
51 #include <sys/syslog.h>
52 #include <sys/taskqueue.h>
53 #include <sys/limits.h>
54 
55 #include <net/if.h>
56 #include <net/if_var.h>
57 #include <net/if_types.h>
58 #include <net/if_media.h>
59 #include <net/bpf.h>
60 #include <net/ethernet.h>
61 #include <net/mp_ring.h>
62 #include <net/debugnet.h>
63 #include <net/pfil.h>
64 #include <net/vnet.h>
65 
66 #include <netinet/in.h>
67 #include <netinet/in_pcb.h>
68 #include <netinet/tcp_lro.h>
69 #include <netinet/in_systm.h>
70 #include <netinet/if_ether.h>
71 #include <netinet/ip.h>
72 #include <netinet/ip6.h>
73 #include <netinet/tcp.h>
74 #include <netinet/ip_var.h>
75 #include <netinet6/ip6_var.h>
76 
77 #include <machine/bus.h>
78 #include <machine/in_cksum.h>
79 
80 #include <vm/vm.h>
81 #include <vm/pmap.h>
82 
83 #include <dev/led/led.h>
84 #include <dev/pci/pcireg.h>
85 #include <dev/pci/pcivar.h>
86 #include <dev/pci/pci_private.h>
87 
88 #include <net/iflib.h>
89 #include <net/iflib_private.h>
90 
91 #include "ifdi_if.h"
92 
93 #ifdef PCI_IOV
94 #include <dev/pci/pci_iov.h>
95 #endif
96 
97 #include <sys/bitstring.h>
98 /*
99  * enable accounting of every mbuf as it comes in to and goes out of
100  * iflib's software descriptor references
101  */
102 #define MEMORY_LOGGING 0
103 /*
104  * Enable mbuf vectors for compressing long mbuf chains
105  */
106 
107 /*
108  * NB:
109  * - Prefetching in tx cleaning should perhaps be a tunable. The distance ahead
110  *   we prefetch needs to be determined by the time spent in m_free vis a vis
111  *   the cost of a prefetch. This will of course vary based on the workload:
112  *      - NFLX's m_free path is dominated by vm-based M_EXT manipulation which
113  *        is quite expensive, thus suggesting very little prefetch.
114  *      - small packet forwarding which is just returning a single mbuf to
115  *        UMA will typically be very fast vis a vis the cost of a memory
116  *        access.
117  */
118 
119 /*
120  * File organization:
121  *  - private structures
122  *  - iflib private utility functions
123  *  - ifnet functions
124  *  - vlan registry and other exported functions
125  *  - iflib public core functions
126  *
127  *
128  */
129 MALLOC_DEFINE(M_IFLIB, "iflib", "ifnet library");
130 
131 #define	IFLIB_RXEOF_MORE (1U << 0)
132 #define	IFLIB_RXEOF_EMPTY (2U << 0)
133 
134 struct iflib_txq;
135 typedef struct iflib_txq *iflib_txq_t;
136 struct iflib_rxq;
137 typedef struct iflib_rxq *iflib_rxq_t;
138 struct iflib_fl;
139 typedef struct iflib_fl *iflib_fl_t;
140 
141 struct iflib_ctx;
142 
143 static void iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid);
144 static void iflib_timer(void *arg);
145 
146 typedef struct iflib_filter_info {
147 	driver_filter_t *ifi_filter;
148 	void *ifi_filter_arg;
149 	struct grouptask *ifi_task;
150 	void *ifi_ctx;
151 } *iflib_filter_info_t;
152 
153 struct iflib_ctx {
154 	KOBJ_FIELDS;
155 	/*
156 	 * Pointer to hardware driver's softc
157 	 */
158 	void *ifc_softc;
159 	device_t ifc_dev;
160 	if_t ifc_ifp;
161 
162 	cpuset_t ifc_cpus;
163 	if_shared_ctx_t ifc_sctx;
164 	struct if_softc_ctx ifc_softc_ctx;
165 
166 	struct sx ifc_ctx_sx;
167 	struct mtx ifc_state_mtx;
168 
169 	iflib_txq_t ifc_txqs;
170 	iflib_rxq_t ifc_rxqs;
171 	uint32_t ifc_if_flags;
172 	uint32_t ifc_flags;
173 	uint32_t ifc_max_fl_buf_size;
174 	uint32_t ifc_rx_mbuf_sz;
175 
176 	int ifc_link_state;
177 	int ifc_watchdog_events;
178 	struct cdev *ifc_led_dev;
179 	struct resource *ifc_msix_mem;
180 
181 	struct if_irq ifc_legacy_irq;
182 	struct grouptask ifc_admin_task;
183 	struct grouptask ifc_vflr_task;
184 	struct iflib_filter_info ifc_filter_info;
185 	struct ifmedia	ifc_media;
186 	struct ifmedia	*ifc_mediap;
187 
188 	struct sysctl_oid *ifc_sysctl_node;
189 	uint16_t ifc_sysctl_ntxqs;
190 	uint16_t ifc_sysctl_nrxqs;
191 	uint16_t ifc_sysctl_qs_eq_override;
192 	uint16_t ifc_sysctl_rx_budget;
193 	uint16_t ifc_sysctl_tx_abdicate;
194 	uint16_t ifc_sysctl_core_offset;
195 #define	CORE_OFFSET_UNSPECIFIED	0xffff
196 	uint8_t  ifc_sysctl_separate_txrx;
197 
198 	qidx_t ifc_sysctl_ntxds[8];
199 	qidx_t ifc_sysctl_nrxds[8];
200 	struct if_txrx ifc_txrx;
201 #define isc_txd_encap  ifc_txrx.ift_txd_encap
202 #define isc_txd_flush  ifc_txrx.ift_txd_flush
203 #define isc_txd_credits_update  ifc_txrx.ift_txd_credits_update
204 #define isc_rxd_available ifc_txrx.ift_rxd_available
205 #define isc_rxd_pkt_get ifc_txrx.ift_rxd_pkt_get
206 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
207 #define isc_rxd_flush ifc_txrx.ift_rxd_flush
208 #define isc_legacy_intr ifc_txrx.ift_legacy_intr
209 	eventhandler_tag ifc_vlan_attach_event;
210 	eventhandler_tag ifc_vlan_detach_event;
211 	struct ether_addr ifc_mac;
212 };
213 
214 void *
215 iflib_get_softc(if_ctx_t ctx)
216 {
217 
218 	return (ctx->ifc_softc);
219 }
220 
221 device_t
222 iflib_get_dev(if_ctx_t ctx)
223 {
224 
225 	return (ctx->ifc_dev);
226 }
227 
228 if_t
229 iflib_get_ifp(if_ctx_t ctx)
230 {
231 
232 	return (ctx->ifc_ifp);
233 }
234 
235 struct ifmedia *
236 iflib_get_media(if_ctx_t ctx)
237 {
238 
239 	return (ctx->ifc_mediap);
240 }
241 
242 uint32_t
243 iflib_get_flags(if_ctx_t ctx)
244 {
245 	return (ctx->ifc_flags);
246 }
247 
248 void
249 iflib_set_mac(if_ctx_t ctx, uint8_t mac[ETHER_ADDR_LEN])
250 {
251 
252 	bcopy(mac, ctx->ifc_mac.octet, ETHER_ADDR_LEN);
253 }
254 
255 if_softc_ctx_t
256 iflib_get_softc_ctx(if_ctx_t ctx)
257 {
258 
259 	return (&ctx->ifc_softc_ctx);
260 }
261 
262 if_shared_ctx_t
263 iflib_get_sctx(if_ctx_t ctx)
264 {
265 
266 	return (ctx->ifc_sctx);
267 }
268 
269 #define IP_ALIGNED(m) ((((uintptr_t)(m)->m_data) & 0x3) == 0x2)
270 #define CACHE_PTR_INCREMENT (CACHE_LINE_SIZE/sizeof(void*))
271 #define CACHE_PTR_NEXT(ptr) ((void *)(((uintptr_t)(ptr)+CACHE_LINE_SIZE-1) & (CACHE_LINE_SIZE-1)))
272 
273 #define LINK_ACTIVE(ctx) ((ctx)->ifc_link_state == LINK_STATE_UP)
274 #define CTX_IS_VF(ctx) ((ctx)->ifc_sctx->isc_flags & IFLIB_IS_VF)
275 
276 typedef struct iflib_sw_rx_desc_array {
277 	bus_dmamap_t	*ifsd_map;         /* bus_dma maps for packet */
278 	struct mbuf	**ifsd_m;           /* pkthdr mbufs */
279 	caddr_t		*ifsd_cl;          /* direct cluster pointer for rx */
280 	bus_addr_t	*ifsd_ba;          /* bus addr of cluster for rx */
281 } iflib_rxsd_array_t;
282 
283 typedef struct iflib_sw_tx_desc_array {
284 	bus_dmamap_t    *ifsd_map;         /* bus_dma maps for packet */
285 	bus_dmamap_t	*ifsd_tso_map;     /* bus_dma maps for TSO packet */
286 	struct mbuf    **ifsd_m;           /* pkthdr mbufs */
287 } if_txsd_vec_t;
288 
289 /* magic number that should be high enough for any hardware */
290 #define IFLIB_MAX_TX_SEGS		128
291 #define IFLIB_RX_COPY_THRESH		128
292 #define IFLIB_MAX_RX_REFRESH		32
293 /* The minimum descriptors per second before we start coalescing */
294 #define IFLIB_MIN_DESC_SEC		16384
295 #define IFLIB_DEFAULT_TX_UPDATE_FREQ	16
296 #define IFLIB_QUEUE_IDLE		0
297 #define IFLIB_QUEUE_HUNG		1
298 #define IFLIB_QUEUE_WORKING		2
299 /* maximum number of txqs that can share an rx interrupt */
300 #define IFLIB_MAX_TX_SHARED_INTR	4
301 
302 /* this should really scale with ring size - this is a fairly arbitrary value */
303 #define TX_BATCH_SIZE			32
304 
305 #define IFLIB_RESTART_BUDGET		8
306 
307 #define CSUM_OFFLOAD		(CSUM_IP_TSO|CSUM_IP6_TSO|CSUM_IP| \
308 				 CSUM_IP_UDP|CSUM_IP_TCP|CSUM_IP_SCTP| \
309 				 CSUM_IP6_UDP|CSUM_IP6_TCP|CSUM_IP6_SCTP)
310 
311 struct iflib_txq {
312 	qidx_t		ift_in_use;
313 	qidx_t		ift_cidx;
314 	qidx_t		ift_cidx_processed;
315 	qidx_t		ift_pidx;
316 	uint8_t		ift_gen;
317 	uint8_t		ift_br_offset;
318 	uint16_t	ift_npending;
319 	uint16_t	ift_db_pending;
320 	uint16_t	ift_rs_pending;
321 	/* implicit pad */
322 	uint8_t		ift_txd_size[8];
323 	uint64_t	ift_processed;
324 	uint64_t	ift_cleaned;
325 	uint64_t	ift_cleaned_prev;
326 #if MEMORY_LOGGING
327 	uint64_t	ift_enqueued;
328 	uint64_t	ift_dequeued;
329 #endif
330 	uint64_t	ift_no_tx_dma_setup;
331 	uint64_t	ift_no_desc_avail;
332 	uint64_t	ift_mbuf_defrag_failed;
333 	uint64_t	ift_mbuf_defrag;
334 	uint64_t	ift_map_failed;
335 	uint64_t	ift_txd_encap_efbig;
336 	uint64_t	ift_pullups;
337 	uint64_t	ift_last_timer_tick;
338 
339 	struct mtx	ift_mtx;
340 	struct mtx	ift_db_mtx;
341 
342 	/* constant values */
343 	if_ctx_t	ift_ctx;
344 	struct ifmp_ring        *ift_br;
345 	struct grouptask	ift_task;
346 	qidx_t		ift_size;
347 	uint16_t	ift_id;
348 	struct callout	ift_timer;
349 #ifdef DEV_NETMAP
350 	struct callout	ift_netmap_timer;
351 #endif /* DEV_NETMAP */
352 
353 	if_txsd_vec_t	ift_sds;
354 	uint8_t		ift_qstatus;
355 	uint8_t		ift_closed;
356 	uint8_t		ift_update_freq;
357 	struct iflib_filter_info ift_filter_info;
358 	bus_dma_tag_t	ift_buf_tag;
359 	bus_dma_tag_t	ift_tso_buf_tag;
360 	iflib_dma_info_t	ift_ifdi;
361 #define	MTX_NAME_LEN	32
362 	char                    ift_mtx_name[MTX_NAME_LEN];
363 	bus_dma_segment_t	ift_segs[IFLIB_MAX_TX_SEGS]  __aligned(CACHE_LINE_SIZE);
364 #ifdef IFLIB_DIAGNOSTICS
365 	uint64_t ift_cpu_exec_count[256];
366 #endif
367 } __aligned(CACHE_LINE_SIZE);
368 
369 struct iflib_fl {
370 	qidx_t		ifl_cidx;
371 	qidx_t		ifl_pidx;
372 	qidx_t		ifl_credits;
373 	uint8_t		ifl_gen;
374 	uint8_t		ifl_rxd_size;
375 #if MEMORY_LOGGING
376 	uint64_t	ifl_m_enqueued;
377 	uint64_t	ifl_m_dequeued;
378 	uint64_t	ifl_cl_enqueued;
379 	uint64_t	ifl_cl_dequeued;
380 #endif
381 	/* implicit pad */
382 	bitstr_t 	*ifl_rx_bitmap;
383 	qidx_t		ifl_fragidx;
384 	/* constant */
385 	qidx_t		ifl_size;
386 	uint16_t	ifl_buf_size;
387 	uint16_t	ifl_cltype;
388 	uma_zone_t	ifl_zone;
389 	iflib_rxsd_array_t	ifl_sds;
390 	iflib_rxq_t	ifl_rxq;
391 	uint8_t		ifl_id;
392 	bus_dma_tag_t	ifl_buf_tag;
393 	iflib_dma_info_t	ifl_ifdi;
394 	uint64_t	ifl_bus_addrs[IFLIB_MAX_RX_REFRESH] __aligned(CACHE_LINE_SIZE);
395 	qidx_t		ifl_rxd_idxs[IFLIB_MAX_RX_REFRESH];
396 }  __aligned(CACHE_LINE_SIZE);
397 
398 static inline qidx_t
399 get_inuse(int size, qidx_t cidx, qidx_t pidx, uint8_t gen)
400 {
401 	qidx_t used;
402 
403 	if (pidx > cidx)
404 		used = pidx - cidx;
405 	else if (pidx < cidx)
406 		used = size - cidx + pidx;
407 	else if (gen == 0 && pidx == cidx)
408 		used = 0;
409 	else if (gen == 1 && pidx == cidx)
410 		used = size;
411 	else
412 		panic("bad state");
413 
414 	return (used);
415 }
416 
417 #define TXQ_AVAIL(txq) (txq->ift_size - get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx, txq->ift_gen))
418 
419 #define IDXDIFF(head, tail, wrap) \
420 	((head) >= (tail) ? (head) - (tail) : (wrap) - (tail) + (head))
421 
422 struct iflib_rxq {
423 	if_ctx_t	ifr_ctx;
424 	iflib_fl_t	ifr_fl;
425 	uint64_t	ifr_rx_irq;
426 	struct pfil_head	*pfil;
427 	/*
428 	 * If there is a separate completion queue (IFLIB_HAS_RXCQ), this is
429 	 * the completion queue consumer index.  Otherwise it's unused.
430 	 */
431 	qidx_t		ifr_cq_cidx;
432 	uint16_t	ifr_id;
433 	uint8_t		ifr_nfl;
434 	uint8_t		ifr_ntxqirq;
435 	uint8_t		ifr_txqid[IFLIB_MAX_TX_SHARED_INTR];
436 	uint8_t		ifr_fl_offset;
437 	struct lro_ctrl			ifr_lc;
438 	struct grouptask        ifr_task;
439 	struct callout		ifr_watchdog;
440 	struct iflib_filter_info ifr_filter_info;
441 	iflib_dma_info_t		ifr_ifdi;
442 
443 	/* dynamically allocate if any drivers need a value substantially larger than this */
444 	struct if_rxd_frag	ifr_frags[IFLIB_MAX_RX_SEGS] __aligned(CACHE_LINE_SIZE);
445 #ifdef IFLIB_DIAGNOSTICS
446 	uint64_t ifr_cpu_exec_count[256];
447 #endif
448 }  __aligned(CACHE_LINE_SIZE);
449 
450 typedef struct if_rxsd {
451 	caddr_t *ifsd_cl;
452 	iflib_fl_t ifsd_fl;
453 } *if_rxsd_t;
454 
455 /* multiple of word size */
456 #ifdef __LP64__
457 #define PKT_INFO_SIZE	6
458 #define RXD_INFO_SIZE	5
459 #define PKT_TYPE uint64_t
460 #else
461 #define PKT_INFO_SIZE	11
462 #define RXD_INFO_SIZE	8
463 #define PKT_TYPE uint32_t
464 #endif
465 #define PKT_LOOP_BOUND  ((PKT_INFO_SIZE/3)*3)
466 #define RXD_LOOP_BOUND  ((RXD_INFO_SIZE/4)*4)
467 
468 typedef struct if_pkt_info_pad {
469 	PKT_TYPE pkt_val[PKT_INFO_SIZE];
470 } *if_pkt_info_pad_t;
471 typedef struct if_rxd_info_pad {
472 	PKT_TYPE rxd_val[RXD_INFO_SIZE];
473 } *if_rxd_info_pad_t;
474 
475 CTASSERT(sizeof(struct if_pkt_info_pad) == sizeof(struct if_pkt_info));
476 CTASSERT(sizeof(struct if_rxd_info_pad) == sizeof(struct if_rxd_info));
477 
478 static inline void
479 pkt_info_zero(if_pkt_info_t pi)
480 {
481 	if_pkt_info_pad_t pi_pad;
482 
483 	pi_pad = (if_pkt_info_pad_t)pi;
484 	pi_pad->pkt_val[0] = 0; pi_pad->pkt_val[1] = 0; pi_pad->pkt_val[2] = 0;
485 	pi_pad->pkt_val[3] = 0; pi_pad->pkt_val[4] = 0; pi_pad->pkt_val[5] = 0;
486 #ifndef __LP64__
487 	pi_pad->pkt_val[6] = 0; pi_pad->pkt_val[7] = 0; pi_pad->pkt_val[8] = 0;
488 	pi_pad->pkt_val[9] = 0; pi_pad->pkt_val[10] = 0;
489 #endif
490 }
491 
492 static device_method_t iflib_pseudo_methods[] = {
493 	DEVMETHOD(device_attach, noop_attach),
494 	DEVMETHOD(device_detach, iflib_pseudo_detach),
495 	DEVMETHOD_END
496 };
497 
498 driver_t iflib_pseudodriver = {
499 	"iflib_pseudo", iflib_pseudo_methods, sizeof(struct iflib_ctx),
500 };
501 
502 static inline void
503 rxd_info_zero(if_rxd_info_t ri)
504 {
505 	if_rxd_info_pad_t ri_pad;
506 	int i;
507 
508 	ri_pad = (if_rxd_info_pad_t)ri;
509 	for (i = 0; i < RXD_LOOP_BOUND; i += 4) {
510 		ri_pad->rxd_val[i] = 0;
511 		ri_pad->rxd_val[i+1] = 0;
512 		ri_pad->rxd_val[i+2] = 0;
513 		ri_pad->rxd_val[i+3] = 0;
514 	}
515 #ifdef __LP64__
516 	ri_pad->rxd_val[RXD_INFO_SIZE-1] = 0;
517 #endif
518 }
519 
520 /*
521  * Only allow a single packet to take up most 1/nth of the tx ring
522  */
523 #define MAX_SINGLE_PACKET_FRACTION 12
524 #define IF_BAD_DMA (bus_addr_t)-1
525 
526 #define CTX_ACTIVE(ctx) ((if_getdrvflags((ctx)->ifc_ifp) & IFF_DRV_RUNNING))
527 
528 #define CTX_LOCK_INIT(_sc)  sx_init(&(_sc)->ifc_ctx_sx, "iflib ctx lock")
529 #define CTX_LOCK(ctx) sx_xlock(&(ctx)->ifc_ctx_sx)
530 #define CTX_UNLOCK(ctx) sx_xunlock(&(ctx)->ifc_ctx_sx)
531 #define CTX_LOCK_DESTROY(ctx) sx_destroy(&(ctx)->ifc_ctx_sx)
532 
533 #define STATE_LOCK_INIT(_sc, _name)  mtx_init(&(_sc)->ifc_state_mtx, _name, "iflib state lock", MTX_DEF)
534 #define STATE_LOCK(ctx) mtx_lock(&(ctx)->ifc_state_mtx)
535 #define STATE_UNLOCK(ctx) mtx_unlock(&(ctx)->ifc_state_mtx)
536 #define STATE_LOCK_DESTROY(ctx) mtx_destroy(&(ctx)->ifc_state_mtx)
537 
538 #define CALLOUT_LOCK(txq)	mtx_lock(&txq->ift_mtx)
539 #define CALLOUT_UNLOCK(txq) 	mtx_unlock(&txq->ift_mtx)
540 
541 void
542 iflib_set_detach(if_ctx_t ctx)
543 {
544 	STATE_LOCK(ctx);
545 	ctx->ifc_flags |= IFC_IN_DETACH;
546 	STATE_UNLOCK(ctx);
547 }
548 
549 /* Our boot-time initialization hook */
550 static int	iflib_module_event_handler(module_t, int, void *);
551 
552 static moduledata_t iflib_moduledata = {
553 	"iflib",
554 	iflib_module_event_handler,
555 	NULL
556 };
557 
558 DECLARE_MODULE(iflib, iflib_moduledata, SI_SUB_INIT_IF, SI_ORDER_ANY);
559 MODULE_VERSION(iflib, 1);
560 
561 MODULE_DEPEND(iflib, pci, 1, 1, 1);
562 MODULE_DEPEND(iflib, ether, 1, 1, 1);
563 
564 TASKQGROUP_DEFINE(if_io_tqg, mp_ncpus, 1);
565 TASKQGROUP_DEFINE(if_config_tqg, 1, 1);
566 
567 #ifndef IFLIB_DEBUG_COUNTERS
568 #ifdef INVARIANTS
569 #define IFLIB_DEBUG_COUNTERS 1
570 #else
571 #define IFLIB_DEBUG_COUNTERS 0
572 #endif /* !INVARIANTS */
573 #endif
574 
575 static SYSCTL_NODE(_net, OID_AUTO, iflib, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
576     "iflib driver parameters");
577 
578 /*
579  * XXX need to ensure that this can't accidentally cause the head to be moved backwards
580  */
581 static int iflib_min_tx_latency = 0;
582 SYSCTL_INT(_net_iflib, OID_AUTO, min_tx_latency, CTLFLAG_RW,
583 		   &iflib_min_tx_latency, 0, "minimize transmit latency at the possible expense of throughput");
584 static int iflib_no_tx_batch = 0;
585 SYSCTL_INT(_net_iflib, OID_AUTO, no_tx_batch, CTLFLAG_RW,
586 		   &iflib_no_tx_batch, 0, "minimize transmit latency at the possible expense of throughput");
587 
588 #if IFLIB_DEBUG_COUNTERS
589 
590 static int iflib_tx_seen;
591 static int iflib_tx_sent;
592 static int iflib_tx_encap;
593 static int iflib_rx_allocs;
594 static int iflib_fl_refills;
595 static int iflib_fl_refills_large;
596 static int iflib_tx_frees;
597 
598 SYSCTL_INT(_net_iflib, OID_AUTO, tx_seen, CTLFLAG_RD,
599 		   &iflib_tx_seen, 0, "# TX mbufs seen");
600 SYSCTL_INT(_net_iflib, OID_AUTO, tx_sent, CTLFLAG_RD,
601 		   &iflib_tx_sent, 0, "# TX mbufs sent");
602 SYSCTL_INT(_net_iflib, OID_AUTO, tx_encap, CTLFLAG_RD,
603 		   &iflib_tx_encap, 0, "# TX mbufs encapped");
604 SYSCTL_INT(_net_iflib, OID_AUTO, tx_frees, CTLFLAG_RD,
605 		   &iflib_tx_frees, 0, "# TX frees");
606 SYSCTL_INT(_net_iflib, OID_AUTO, rx_allocs, CTLFLAG_RD,
607 		   &iflib_rx_allocs, 0, "# RX allocations");
608 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills, CTLFLAG_RD,
609 		   &iflib_fl_refills, 0, "# refills");
610 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills_large, CTLFLAG_RD,
611 		   &iflib_fl_refills_large, 0, "# large refills");
612 
613 static int iflib_txq_drain_flushing;
614 static int iflib_txq_drain_oactive;
615 static int iflib_txq_drain_notready;
616 
617 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_flushing, CTLFLAG_RD,
618 		   &iflib_txq_drain_flushing, 0, "# drain flushes");
619 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_oactive, CTLFLAG_RD,
620 		   &iflib_txq_drain_oactive, 0, "# drain oactives");
621 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_notready, CTLFLAG_RD,
622 		   &iflib_txq_drain_notready, 0, "# drain notready");
623 
624 static int iflib_encap_load_mbuf_fail;
625 static int iflib_encap_pad_mbuf_fail;
626 static int iflib_encap_txq_avail_fail;
627 static int iflib_encap_txd_encap_fail;
628 
629 SYSCTL_INT(_net_iflib, OID_AUTO, encap_load_mbuf_fail, CTLFLAG_RD,
630 		   &iflib_encap_load_mbuf_fail, 0, "# busdma load failures");
631 SYSCTL_INT(_net_iflib, OID_AUTO, encap_pad_mbuf_fail, CTLFLAG_RD,
632 		   &iflib_encap_pad_mbuf_fail, 0, "# runt frame pad failures");
633 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txq_avail_fail, CTLFLAG_RD,
634 		   &iflib_encap_txq_avail_fail, 0, "# txq avail failures");
635 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txd_encap_fail, CTLFLAG_RD,
636 		   &iflib_encap_txd_encap_fail, 0, "# driver encap failures");
637 
638 static int iflib_task_fn_rxs;
639 static int iflib_rx_intr_enables;
640 static int iflib_fast_intrs;
641 static int iflib_rx_unavail;
642 static int iflib_rx_ctx_inactive;
643 static int iflib_rx_if_input;
644 static int iflib_rxd_flush;
645 
646 static int iflib_verbose_debug;
647 
648 SYSCTL_INT(_net_iflib, OID_AUTO, task_fn_rx, CTLFLAG_RD,
649 		   &iflib_task_fn_rxs, 0, "# task_fn_rx calls");
650 SYSCTL_INT(_net_iflib, OID_AUTO, rx_intr_enables, CTLFLAG_RD,
651 		   &iflib_rx_intr_enables, 0, "# RX intr enables");
652 SYSCTL_INT(_net_iflib, OID_AUTO, fast_intrs, CTLFLAG_RD,
653 		   &iflib_fast_intrs, 0, "# fast_intr calls");
654 SYSCTL_INT(_net_iflib, OID_AUTO, rx_unavail, CTLFLAG_RD,
655 		   &iflib_rx_unavail, 0, "# times rxeof called with no available data");
656 SYSCTL_INT(_net_iflib, OID_AUTO, rx_ctx_inactive, CTLFLAG_RD,
657 		   &iflib_rx_ctx_inactive, 0, "# times rxeof called with inactive context");
658 SYSCTL_INT(_net_iflib, OID_AUTO, rx_if_input, CTLFLAG_RD,
659 		   &iflib_rx_if_input, 0, "# times rxeof called if_input");
660 SYSCTL_INT(_net_iflib, OID_AUTO, rxd_flush, CTLFLAG_RD,
661 	         &iflib_rxd_flush, 0, "# times rxd_flush called");
662 SYSCTL_INT(_net_iflib, OID_AUTO, verbose_debug, CTLFLAG_RW,
663 		   &iflib_verbose_debug, 0, "enable verbose debugging");
664 
665 #define DBG_COUNTER_INC(name) atomic_add_int(&(iflib_ ## name), 1)
666 static void
667 iflib_debug_reset(void)
668 {
669 	iflib_tx_seen = iflib_tx_sent = iflib_tx_encap = iflib_rx_allocs =
670 		iflib_fl_refills = iflib_fl_refills_large = iflib_tx_frees =
671 		iflib_txq_drain_flushing = iflib_txq_drain_oactive =
672 		iflib_txq_drain_notready =
673 		iflib_encap_load_mbuf_fail = iflib_encap_pad_mbuf_fail =
674 		iflib_encap_txq_avail_fail = iflib_encap_txd_encap_fail =
675 		iflib_task_fn_rxs = iflib_rx_intr_enables = iflib_fast_intrs =
676 		iflib_rx_unavail =
677 		iflib_rx_ctx_inactive = iflib_rx_if_input =
678 		iflib_rxd_flush = 0;
679 }
680 
681 #else
682 #define DBG_COUNTER_INC(name)
683 static void iflib_debug_reset(void) {}
684 #endif
685 
686 #define IFLIB_DEBUG 0
687 
688 static void iflib_tx_structures_free(if_ctx_t ctx);
689 static void iflib_rx_structures_free(if_ctx_t ctx);
690 static int iflib_queues_alloc(if_ctx_t ctx);
691 static int iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq);
692 static int iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget);
693 static int iflib_qset_structures_setup(if_ctx_t ctx);
694 static int iflib_msix_init(if_ctx_t ctx);
695 static int iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filterarg, int *rid, const char *str);
696 static void iflib_txq_check_drain(iflib_txq_t txq, int budget);
697 static uint32_t iflib_txq_can_drain(struct ifmp_ring *);
698 #ifdef ALTQ
699 static void iflib_altq_if_start(if_t ifp);
700 static int iflib_altq_if_transmit(if_t ifp, struct mbuf *m);
701 #endif
702 static int iflib_register(if_ctx_t);
703 static void iflib_deregister(if_ctx_t);
704 static void iflib_unregister_vlan_handlers(if_ctx_t ctx);
705 static uint16_t iflib_get_mbuf_size_for(unsigned int size);
706 static void iflib_init_locked(if_ctx_t ctx);
707 static void iflib_add_device_sysctl_pre(if_ctx_t ctx);
708 static void iflib_add_device_sysctl_post(if_ctx_t ctx);
709 static void iflib_ifmp_purge(iflib_txq_t txq);
710 static void _iflib_pre_assert(if_softc_ctx_t scctx);
711 static void iflib_if_init_locked(if_ctx_t ctx);
712 static void iflib_free_intr_mem(if_ctx_t ctx);
713 #ifndef __NO_STRICT_ALIGNMENT
714 static struct mbuf * iflib_fixup_rx(struct mbuf *m);
715 #endif
716 
717 static SLIST_HEAD(cpu_offset_list, cpu_offset) cpu_offsets =
718     SLIST_HEAD_INITIALIZER(cpu_offsets);
719 struct cpu_offset {
720 	SLIST_ENTRY(cpu_offset) entries;
721 	cpuset_t	set;
722 	unsigned int	refcount;
723 	uint16_t	offset;
724 };
725 static struct mtx cpu_offset_mtx;
726 MTX_SYSINIT(iflib_cpu_offset, &cpu_offset_mtx, "iflib_cpu_offset lock",
727     MTX_DEF);
728 
729 DEBUGNET_DEFINE(iflib);
730 
731 static int
732 iflib_num_rx_descs(if_ctx_t ctx)
733 {
734 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
735 	if_shared_ctx_t sctx = ctx->ifc_sctx;
736 	uint16_t first_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
737 
738 	return scctx->isc_nrxd[first_rxq];
739 }
740 
741 static int
742 iflib_num_tx_descs(if_ctx_t ctx)
743 {
744 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
745 	if_shared_ctx_t sctx = ctx->ifc_sctx;
746 	uint16_t first_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
747 
748 	return scctx->isc_ntxd[first_txq];
749 }
750 
751 #ifdef DEV_NETMAP
752 #include <sys/selinfo.h>
753 #include <net/netmap.h>
754 #include <dev/netmap/netmap_kern.h>
755 
756 MODULE_DEPEND(iflib, netmap, 1, 1, 1);
757 
758 static int netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, bool init);
759 static void iflib_netmap_timer(void *arg);
760 
761 /*
762  * device-specific sysctl variables:
763  *
764  * iflib_crcstrip: 0: keep CRC in rx frames (default), 1: strip it.
765  *	During regular operations the CRC is stripped, but on some
766  *	hardware reception of frames not multiple of 64 is slower,
767  *	so using crcstrip=0 helps in benchmarks.
768  *
769  * iflib_rx_miss, iflib_rx_miss_bufs:
770  *	count packets that might be missed due to lost interrupts.
771  */
772 SYSCTL_DECL(_dev_netmap);
773 /*
774  * The xl driver by default strips CRCs and we do not override it.
775  */
776 
777 int iflib_crcstrip = 1;
778 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_crcstrip,
779     CTLFLAG_RW, &iflib_crcstrip, 1, "strip CRC on RX frames");
780 
781 int iflib_rx_miss, iflib_rx_miss_bufs;
782 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss,
783     CTLFLAG_RW, &iflib_rx_miss, 0, "potentially missed RX intr");
784 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss_bufs,
785     CTLFLAG_RW, &iflib_rx_miss_bufs, 0, "potentially missed RX intr bufs");
786 
787 /*
788  * Register/unregister. We are already under netmap lock.
789  * Only called on the first register or the last unregister.
790  */
791 static int
792 iflib_netmap_register(struct netmap_adapter *na, int onoff)
793 {
794 	if_t ifp = na->ifp;
795 	if_ctx_t ctx = ifp->if_softc;
796 	int status;
797 
798 	CTX_LOCK(ctx);
799 	IFDI_INTR_DISABLE(ctx);
800 
801 	/* Tell the stack that the interface is no longer active */
802 	ifp->if_drv_flags &= ~(IFF_DRV_RUNNING | IFF_DRV_OACTIVE);
803 
804 	if (!CTX_IS_VF(ctx))
805 		IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip);
806 
807 	iflib_stop(ctx);
808 
809 	/*
810 	 * Enable (or disable) netmap flags, and intercept (or restore)
811 	 * ifp->if_transmit. This is done once the device has been stopped
812 	 * to prevent race conditions.
813 	 */
814 	if (onoff) {
815 		nm_set_native_flags(na);
816 	} else {
817 		nm_clear_native_flags(na);
818 	}
819 
820 	iflib_init_locked(ctx);
821 	IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip); // XXX why twice ?
822 	status = ifp->if_drv_flags & IFF_DRV_RUNNING ? 0 : 1;
823 	if (status)
824 		nm_clear_native_flags(na);
825 	CTX_UNLOCK(ctx);
826 	return (status);
827 }
828 
829 static int
830 netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, bool init)
831 {
832 	struct netmap_adapter *na = kring->na;
833 	u_int const lim = kring->nkr_num_slots - 1;
834 	u_int nm_i = kring->nr_hwcur;
835 	struct netmap_ring *ring = kring->ring;
836 	bus_dmamap_t *map;
837 	struct if_rxd_update iru;
838 	if_ctx_t ctx = rxq->ifr_ctx;
839 	iflib_fl_t fl = &rxq->ifr_fl[0];
840 	u_int nic_i_first, nic_i;
841 	int i, n;
842 #if IFLIB_DEBUG_COUNTERS
843 	int rf_count = 0;
844 #endif
845 
846 	/*
847 	 * This function is used both at initialization and in rxsync.
848 	 * At initialization we need to prepare (with isc_rxd_refill())
849 	 * all the (N) netmap buffers in the ring, in such a way to keep
850 	 * fl->ifl_pidx and kring->nr_hwcur in sync (except for
851 	 * kring->nkr_hwofs); at rxsync time, both indexes point to the
852 	 * next buffer to be refilled.
853 	 * In any case we publish (with isc_rxd_flush()) up to
854 	 * (fl->ifl_pidx - 1) % N (included), to avoid the NIC tail/prod
855 	 * pointer to overrun the head/cons pointer, although this is
856 	 * not necessary for some NICs (e.g. vmx).
857 	 */
858 	if (__predict_false(init))
859 		n = kring->nkr_num_slots;
860 	else {
861 		n = kring->rhead - nm_i;
862 		if (n == 0)
863 			return (0); /* Nothing to do. */
864 		if (n < 0)
865 			n += kring->nkr_num_slots;
866 	}
867 
868 	/* Start to refill from nr_hwcur, publishing n buffers. */
869 	iru_init(&iru, rxq, 0 /* flid */);
870 	map = fl->ifl_sds.ifsd_map;
871 	nic_i = fl->ifl_pidx;
872 	MPASS(nic_i == netmap_idx_k2n(kring, nm_i));
873 	DBG_COUNTER_INC(fl_refills);
874 	while (n > 0) {
875 #if IFLIB_DEBUG_COUNTERS
876 		if (++rf_count == 9)
877 			DBG_COUNTER_INC(fl_refills_large);
878 #endif
879 		nic_i_first = nic_i;
880 		for (i = 0; n > 0 && i < IFLIB_MAX_RX_REFRESH; n--, i++) {
881 			struct netmap_slot *slot = &ring->slot[nm_i];
882 			void *addr = PNMB(na, slot, &fl->ifl_bus_addrs[i]);
883 
884 			MPASS(i < IFLIB_MAX_RX_REFRESH);
885 
886 			if (addr == NETMAP_BUF_BASE(na)) /* bad buf */
887 			        return netmap_ring_reinit(kring);
888 
889 			fl->ifl_rxd_idxs[i] = nic_i;
890 
891 			if (__predict_false(init)) {
892 				netmap_load_map(na, fl->ifl_buf_tag,
893 				    map[nic_i], addr);
894 			} else if (slot->flags & NS_BUF_CHANGED) {
895 				/* buffer has changed, reload map */
896 				netmap_reload_map(na, fl->ifl_buf_tag,
897 				    map[nic_i], addr);
898 			}
899 			bus_dmamap_sync(fl->ifl_buf_tag, map[nic_i],
900 			    BUS_DMASYNC_PREREAD);
901 			slot->flags &= ~NS_BUF_CHANGED;
902 
903 			nm_i = nm_next(nm_i, lim);
904 			nic_i = nm_next(nic_i, lim);
905 		}
906 
907 		iru.iru_pidx = nic_i_first;
908 		iru.iru_count = i;
909 		ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
910 	}
911 	fl->ifl_pidx = nic_i;
912 	MPASS(!init || nm_i == 0);
913 	MPASS(nm_i == kring->rhead);
914 	kring->nr_hwcur = nm_i;
915 
916 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
917 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
918 	ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id,
919 	    nm_prev(nic_i, lim));
920 	DBG_COUNTER_INC(rxd_flush);
921 
922 	return (0);
923 }
924 
925 #define NETMAP_TX_TIMER_US	90
926 
927 /*
928  * Reconcile kernel and user view of the transmit ring.
929  *
930  * All information is in the kring.
931  * Userspace wants to send packets up to the one before kring->rhead,
932  * kernel knows kring->nr_hwcur is the first unsent packet.
933  *
934  * Here we push packets out (as many as possible), and possibly
935  * reclaim buffers from previously completed transmission.
936  *
937  * The caller (netmap) guarantees that there is only one instance
938  * running at any time. Any interference with other driver
939  * methods should be handled by the individual drivers.
940  */
941 static int
942 iflib_netmap_txsync(struct netmap_kring *kring, int flags)
943 {
944 	struct netmap_adapter *na = kring->na;
945 	if_t ifp = na->ifp;
946 	struct netmap_ring *ring = kring->ring;
947 	u_int nm_i;	/* index into the netmap kring */
948 	u_int nic_i;	/* index into the NIC ring */
949 	u_int n;
950 	u_int const lim = kring->nkr_num_slots - 1;
951 	u_int const head = kring->rhead;
952 	struct if_pkt_info pi;
953 
954 	/*
955 	 * interrupts on every tx packet are expensive so request
956 	 * them every half ring, or where NS_REPORT is set
957 	 */
958 	u_int report_frequency = kring->nkr_num_slots >> 1;
959 	/* device-specific */
960 	if_ctx_t ctx = ifp->if_softc;
961 	iflib_txq_t txq = &ctx->ifc_txqs[kring->ring_id];
962 
963 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
964 	    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
965 
966 	/*
967 	 * First part: process new packets to send.
968 	 * nm_i is the current index in the netmap kring,
969 	 * nic_i is the corresponding index in the NIC ring.
970 	 *
971 	 * If we have packets to send (nm_i != head)
972 	 * iterate over the netmap ring, fetch length and update
973 	 * the corresponding slot in the NIC ring. Some drivers also
974 	 * need to update the buffer's physical address in the NIC slot
975 	 * even NS_BUF_CHANGED is not set (PNMB computes the addresses).
976 	 *
977 	 * The netmap_reload_map() calls is especially expensive,
978 	 * even when (as in this case) the tag is 0, so do only
979 	 * when the buffer has actually changed.
980 	 *
981 	 * If possible do not set the report/intr bit on all slots,
982 	 * but only a few times per ring or when NS_REPORT is set.
983 	 *
984 	 * Finally, on 10G and faster drivers, it might be useful
985 	 * to prefetch the next slot and txr entry.
986 	 */
987 
988 	nm_i = kring->nr_hwcur;
989 	if (nm_i != head) {	/* we have new packets to send */
990 		pkt_info_zero(&pi);
991 		pi.ipi_segs = txq->ift_segs;
992 		pi.ipi_qsidx = kring->ring_id;
993 		nic_i = netmap_idx_k2n(kring, nm_i);
994 
995 		__builtin_prefetch(&ring->slot[nm_i]);
996 		__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i]);
997 		__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i]);
998 
999 		for (n = 0; nm_i != head; n++) {
1000 			struct netmap_slot *slot = &ring->slot[nm_i];
1001 			u_int len = slot->len;
1002 			uint64_t paddr;
1003 			void *addr = PNMB(na, slot, &paddr);
1004 			int flags = (slot->flags & NS_REPORT ||
1005 				nic_i == 0 || nic_i == report_frequency) ?
1006 				IPI_TX_INTR : 0;
1007 
1008 			/* device-specific */
1009 			pi.ipi_len = len;
1010 			pi.ipi_segs[0].ds_addr = paddr;
1011 			pi.ipi_segs[0].ds_len = len;
1012 			pi.ipi_nsegs = 1;
1013 			pi.ipi_ndescs = 0;
1014 			pi.ipi_pidx = nic_i;
1015 			pi.ipi_flags = flags;
1016 
1017 			/* Fill the slot in the NIC ring. */
1018 			ctx->isc_txd_encap(ctx->ifc_softc, &pi);
1019 			DBG_COUNTER_INC(tx_encap);
1020 
1021 			/* prefetch for next round */
1022 			__builtin_prefetch(&ring->slot[nm_i + 1]);
1023 			__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i + 1]);
1024 			__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i + 1]);
1025 
1026 			NM_CHECK_ADDR_LEN(na, addr, len);
1027 
1028 			if (slot->flags & NS_BUF_CHANGED) {
1029 				/* buffer has changed, reload map */
1030 				netmap_reload_map(na, txq->ift_buf_tag,
1031 				    txq->ift_sds.ifsd_map[nic_i], addr);
1032 			}
1033 			/* make sure changes to the buffer are synced */
1034 			bus_dmamap_sync(txq->ift_buf_tag,
1035 			    txq->ift_sds.ifsd_map[nic_i],
1036 			    BUS_DMASYNC_PREWRITE);
1037 
1038 			slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED);
1039 			nm_i = nm_next(nm_i, lim);
1040 			nic_i = nm_next(nic_i, lim);
1041 		}
1042 		kring->nr_hwcur = nm_i;
1043 
1044 		/* synchronize the NIC ring */
1045 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1046 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1047 
1048 		/* (re)start the tx unit up to slot nic_i (excluded) */
1049 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, nic_i);
1050 	}
1051 
1052 	/*
1053 	 * Second part: reclaim buffers for completed transmissions.
1054 	 *
1055 	 * If there are unclaimed buffers, attempt to reclaim them.
1056 	 * If we don't manage to reclaim them all, and TX IRQs are not in use,
1057 	 * trigger a per-tx-queue timer to try again later.
1058 	 */
1059 	if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1060 		if (iflib_tx_credits_update(ctx, txq)) {
1061 			/* some tx completed, increment avail */
1062 			nic_i = txq->ift_cidx_processed;
1063 			kring->nr_hwtail = nm_prev(netmap_idx_n2k(kring, nic_i), lim);
1064 		}
1065 	}
1066 
1067 	if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ))
1068 		if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1069 			callout_reset_sbt_on(&txq->ift_netmap_timer,
1070 			    NETMAP_TX_TIMER_US * SBT_1US, SBT_1US,
1071 			    iflib_netmap_timer, txq,
1072 			    txq->ift_netmap_timer.c_cpu, 0);
1073 		}
1074 	return (0);
1075 }
1076 
1077 /*
1078  * Reconcile kernel and user view of the receive ring.
1079  * Same as for the txsync, this routine must be efficient.
1080  * The caller guarantees a single invocations, but races against
1081  * the rest of the driver should be handled here.
1082  *
1083  * On call, kring->rhead is the first packet that userspace wants
1084  * to keep, and kring->rcur is the wakeup point.
1085  * The kernel has previously reported packets up to kring->rtail.
1086  *
1087  * If (flags & NAF_FORCE_READ) also check for incoming packets irrespective
1088  * of whether or not we received an interrupt.
1089  */
1090 static int
1091 iflib_netmap_rxsync(struct netmap_kring *kring, int flags)
1092 {
1093 	struct netmap_adapter *na = kring->na;
1094 	struct netmap_ring *ring = kring->ring;
1095 	if_t ifp = na->ifp;
1096 	uint32_t nm_i;	/* index into the netmap ring */
1097 	uint32_t nic_i;	/* index into the NIC ring */
1098 	u_int n;
1099 	u_int const lim = kring->nkr_num_slots - 1;
1100 	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
1101 
1102 	if_ctx_t ctx = ifp->if_softc;
1103 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1104 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1105 	iflib_rxq_t rxq = &ctx->ifc_rxqs[kring->ring_id];
1106 	iflib_fl_t fl = &rxq->ifr_fl[0];
1107 	struct if_rxd_info ri;
1108 	qidx_t *cidxp;
1109 
1110 	/*
1111 	 * netmap only uses free list 0, to avoid out of order consumption
1112 	 * of receive buffers
1113 	 */
1114 
1115 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
1116 	    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1117 
1118 	/*
1119 	 * First part: import newly received packets.
1120 	 *
1121 	 * nm_i is the index of the next free slot in the netmap ring,
1122 	 * nic_i is the index of the next received packet in the NIC ring
1123 	 * (or in the free list 0 if IFLIB_HAS_RXCQ is set), and they may
1124 	 * differ in case if_init() has been called while
1125 	 * in netmap mode. For the receive ring we have
1126 	 *
1127 	 *	nic_i = fl->ifl_cidx;
1128 	 *	nm_i = kring->nr_hwtail (previous)
1129 	 * and
1130 	 *	nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1131 	 *
1132 	 * fl->ifl_cidx is set to 0 on a ring reinit
1133 	 */
1134 	if (netmap_no_pendintr || force_update) {
1135 		uint32_t hwtail_lim = nm_prev(kring->nr_hwcur, lim);
1136 		bool have_rxcq = sctx->isc_flags & IFLIB_HAS_RXCQ;
1137 		int crclen = iflib_crcstrip ? 0 : 4;
1138 		int error, avail;
1139 
1140 		/*
1141 		 * For the free list consumer index, we use the same
1142 		 * logic as in iflib_rxeof().
1143 		 */
1144 		if (have_rxcq)
1145 			cidxp = &rxq->ifr_cq_cidx;
1146 		else
1147 			cidxp = &fl->ifl_cidx;
1148 		avail = ctx->isc_rxd_available(ctx->ifc_softc,
1149 		    rxq->ifr_id, *cidxp, USHRT_MAX);
1150 
1151 		nic_i = fl->ifl_cidx;
1152 		nm_i = netmap_idx_n2k(kring, nic_i);
1153 		MPASS(nm_i == kring->nr_hwtail);
1154 		for (n = 0; avail > 0 && nm_i != hwtail_lim; n++, avail--) {
1155 			rxd_info_zero(&ri);
1156 			ri.iri_frags = rxq->ifr_frags;
1157 			ri.iri_qsidx = kring->ring_id;
1158 			ri.iri_ifp = ctx->ifc_ifp;
1159 			ri.iri_cidx = *cidxp;
1160 
1161 			error = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
1162 			ring->slot[nm_i].len = error ? 0 : ri.iri_len - crclen;
1163 			ring->slot[nm_i].flags = 0;
1164 			if (have_rxcq) {
1165 				*cidxp = ri.iri_cidx;
1166 				while (*cidxp >= scctx->isc_nrxd[0])
1167 					*cidxp -= scctx->isc_nrxd[0];
1168 			}
1169 			bus_dmamap_sync(fl->ifl_buf_tag,
1170 			    fl->ifl_sds.ifsd_map[nic_i], BUS_DMASYNC_POSTREAD);
1171 			nm_i = nm_next(nm_i, lim);
1172 			fl->ifl_cidx = nic_i = nm_next(nic_i, lim);
1173 		}
1174 		if (n) { /* update the state variables */
1175 			if (netmap_no_pendintr && !force_update) {
1176 				/* diagnostics */
1177 				iflib_rx_miss ++;
1178 				iflib_rx_miss_bufs += n;
1179 			}
1180 			kring->nr_hwtail = nm_i;
1181 		}
1182 		kring->nr_kflags &= ~NKR_PENDINTR;
1183 	}
1184 	/*
1185 	 * Second part: skip past packets that userspace has released.
1186 	 * (kring->nr_hwcur to head excluded),
1187 	 * and make the buffers available for reception.
1188 	 * As usual nm_i is the index in the netmap ring,
1189 	 * nic_i is the index in the NIC ring, and
1190 	 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1191 	 */
1192 	netmap_fl_refill(rxq, kring, false);
1193 
1194 	return (0);
1195 }
1196 
1197 static void
1198 iflib_netmap_intr(struct netmap_adapter *na, int onoff)
1199 {
1200 	if_ctx_t ctx = na->ifp->if_softc;
1201 
1202 	CTX_LOCK(ctx);
1203 	if (onoff) {
1204 		IFDI_INTR_ENABLE(ctx);
1205 	} else {
1206 		IFDI_INTR_DISABLE(ctx);
1207 	}
1208 	CTX_UNLOCK(ctx);
1209 }
1210 
1211 static int
1212 iflib_netmap_attach(if_ctx_t ctx)
1213 {
1214 	struct netmap_adapter na;
1215 
1216 	bzero(&na, sizeof(na));
1217 
1218 	na.ifp = ctx->ifc_ifp;
1219 	na.na_flags = NAF_BDG_MAYSLEEP;
1220 	MPASS(ctx->ifc_softc_ctx.isc_ntxqsets);
1221 	MPASS(ctx->ifc_softc_ctx.isc_nrxqsets);
1222 
1223 	na.num_tx_desc = iflib_num_tx_descs(ctx);
1224 	na.num_rx_desc = iflib_num_rx_descs(ctx);
1225 	na.nm_txsync = iflib_netmap_txsync;
1226 	na.nm_rxsync = iflib_netmap_rxsync;
1227 	na.nm_register = iflib_netmap_register;
1228 	na.nm_intr = iflib_netmap_intr;
1229 	na.num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
1230 	na.num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
1231 	return (netmap_attach(&na));
1232 }
1233 
1234 static int
1235 iflib_netmap_txq_init(if_ctx_t ctx, iflib_txq_t txq)
1236 {
1237 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1238 	struct netmap_slot *slot;
1239 
1240 	slot = netmap_reset(na, NR_TX, txq->ift_id, 0);
1241 	if (slot == NULL)
1242 		return (0);
1243 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxd[0]; i++) {
1244 		/*
1245 		 * In netmap mode, set the map for the packet buffer.
1246 		 * NOTE: Some drivers (not this one) also need to set
1247 		 * the physical buffer address in the NIC ring.
1248 		 * netmap_idx_n2k() maps a nic index, i, into the corresponding
1249 		 * netmap slot index, si
1250 		 */
1251 		int si = netmap_idx_n2k(na->tx_rings[txq->ift_id], i);
1252 		netmap_load_map(na, txq->ift_buf_tag, txq->ift_sds.ifsd_map[i],
1253 		    NMB(na, slot + si));
1254 	}
1255 	return (1);
1256 }
1257 
1258 static int
1259 iflib_netmap_rxq_init(if_ctx_t ctx, iflib_rxq_t rxq)
1260 {
1261 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1262 	struct netmap_kring *kring;
1263 	struct netmap_slot *slot;
1264 
1265 	slot = netmap_reset(na, NR_RX, rxq->ifr_id, 0);
1266 	if (slot == NULL)
1267 		return (0);
1268 	kring = na->rx_rings[rxq->ifr_id];
1269 	netmap_fl_refill(rxq, kring, true);
1270 	return (1);
1271 }
1272 
1273 static void
1274 iflib_netmap_timer(void *arg)
1275 {
1276 	iflib_txq_t txq = arg;
1277 	if_ctx_t ctx = txq->ift_ctx;
1278 
1279 	/*
1280 	 * Wake up the netmap application, to give it a chance to
1281 	 * call txsync and reclaim more completed TX buffers.
1282 	 */
1283 	netmap_tx_irq(ctx->ifc_ifp, txq->ift_id);
1284 }
1285 
1286 #define iflib_netmap_detach(ifp) netmap_detach(ifp)
1287 
1288 #else
1289 #define iflib_netmap_txq_init(ctx, txq) (0)
1290 #define iflib_netmap_rxq_init(ctx, rxq) (0)
1291 #define iflib_netmap_detach(ifp)
1292 
1293 #define iflib_netmap_attach(ctx) (0)
1294 #define netmap_rx_irq(ifp, qid, budget) (0)
1295 #endif
1296 
1297 #if defined(__i386__) || defined(__amd64__)
1298 static __inline void
1299 prefetch(void *x)
1300 {
1301 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1302 }
1303 static __inline void
1304 prefetch2cachelines(void *x)
1305 {
1306 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1307 #if (CACHE_LINE_SIZE < 128)
1308 	__asm volatile("prefetcht0 %0" :: "m" (*(((unsigned long *)x)+CACHE_LINE_SIZE/(sizeof(unsigned long)))));
1309 #endif
1310 }
1311 #else
1312 #define prefetch(x)
1313 #define prefetch2cachelines(x)
1314 #endif
1315 
1316 static void
1317 iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid)
1318 {
1319 	iflib_fl_t fl;
1320 
1321 	fl = &rxq->ifr_fl[flid];
1322 	iru->iru_paddrs = fl->ifl_bus_addrs;
1323 	iru->iru_idxs = fl->ifl_rxd_idxs;
1324 	iru->iru_qsidx = rxq->ifr_id;
1325 	iru->iru_buf_size = fl->ifl_buf_size;
1326 	iru->iru_flidx = fl->ifl_id;
1327 }
1328 
1329 static void
1330 _iflib_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int err)
1331 {
1332 	if (err)
1333 		return;
1334 	*(bus_addr_t *) arg = segs[0].ds_addr;
1335 }
1336 
1337 int
1338 iflib_dma_alloc_align(if_ctx_t ctx, int size, int align, iflib_dma_info_t dma, int mapflags)
1339 {
1340 	int err;
1341 	device_t dev = ctx->ifc_dev;
1342 
1343 	err = bus_dma_tag_create(bus_get_dma_tag(dev),	/* parent */
1344 				align, 0,		/* alignment, bounds */
1345 				BUS_SPACE_MAXADDR,	/* lowaddr */
1346 				BUS_SPACE_MAXADDR,	/* highaddr */
1347 				NULL, NULL,		/* filter, filterarg */
1348 				size,			/* maxsize */
1349 				1,			/* nsegments */
1350 				size,			/* maxsegsize */
1351 				BUS_DMA_ALLOCNOW,	/* flags */
1352 				NULL,			/* lockfunc */
1353 				NULL,			/* lockarg */
1354 				&dma->idi_tag);
1355 	if (err) {
1356 		device_printf(dev,
1357 		    "%s: bus_dma_tag_create failed: %d\n",
1358 		    __func__, err);
1359 		goto fail_0;
1360 	}
1361 
1362 	err = bus_dmamem_alloc(dma->idi_tag, (void**) &dma->idi_vaddr,
1363 	    BUS_DMA_NOWAIT | BUS_DMA_COHERENT | BUS_DMA_ZERO, &dma->idi_map);
1364 	if (err) {
1365 		device_printf(dev,
1366 		    "%s: bus_dmamem_alloc(%ju) failed: %d\n",
1367 		    __func__, (uintmax_t)size, err);
1368 		goto fail_1;
1369 	}
1370 
1371 	dma->idi_paddr = IF_BAD_DMA;
1372 	err = bus_dmamap_load(dma->idi_tag, dma->idi_map, dma->idi_vaddr,
1373 	    size, _iflib_dmamap_cb, &dma->idi_paddr, mapflags | BUS_DMA_NOWAIT);
1374 	if (err || dma->idi_paddr == IF_BAD_DMA) {
1375 		device_printf(dev,
1376 		    "%s: bus_dmamap_load failed: %d\n",
1377 		    __func__, err);
1378 		goto fail_2;
1379 	}
1380 
1381 	dma->idi_size = size;
1382 	return (0);
1383 
1384 fail_2:
1385 	bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1386 fail_1:
1387 	bus_dma_tag_destroy(dma->idi_tag);
1388 fail_0:
1389 	dma->idi_tag = NULL;
1390 
1391 	return (err);
1392 }
1393 
1394 int
1395 iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags)
1396 {
1397 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1398 
1399 	KASSERT(sctx->isc_q_align != 0, ("alignment value not initialized"));
1400 
1401 	return (iflib_dma_alloc_align(ctx, size, sctx->isc_q_align, dma, mapflags));
1402 }
1403 
1404 int
1405 iflib_dma_alloc_multi(if_ctx_t ctx, int *sizes, iflib_dma_info_t *dmalist, int mapflags, int count)
1406 {
1407 	int i, err;
1408 	iflib_dma_info_t *dmaiter;
1409 
1410 	dmaiter = dmalist;
1411 	for (i = 0; i < count; i++, dmaiter++) {
1412 		if ((err = iflib_dma_alloc(ctx, sizes[i], *dmaiter, mapflags)) != 0)
1413 			break;
1414 	}
1415 	if (err)
1416 		iflib_dma_free_multi(dmalist, i);
1417 	return (err);
1418 }
1419 
1420 void
1421 iflib_dma_free(iflib_dma_info_t dma)
1422 {
1423 	if (dma->idi_tag == NULL)
1424 		return;
1425 	if (dma->idi_paddr != IF_BAD_DMA) {
1426 		bus_dmamap_sync(dma->idi_tag, dma->idi_map,
1427 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1428 		bus_dmamap_unload(dma->idi_tag, dma->idi_map);
1429 		dma->idi_paddr = IF_BAD_DMA;
1430 	}
1431 	if (dma->idi_vaddr != NULL) {
1432 		bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1433 		dma->idi_vaddr = NULL;
1434 	}
1435 	bus_dma_tag_destroy(dma->idi_tag);
1436 	dma->idi_tag = NULL;
1437 }
1438 
1439 void
1440 iflib_dma_free_multi(iflib_dma_info_t *dmalist, int count)
1441 {
1442 	int i;
1443 	iflib_dma_info_t *dmaiter = dmalist;
1444 
1445 	for (i = 0; i < count; i++, dmaiter++)
1446 		iflib_dma_free(*dmaiter);
1447 }
1448 
1449 static int
1450 iflib_fast_intr(void *arg)
1451 {
1452 	iflib_filter_info_t info = arg;
1453 	struct grouptask *gtask = info->ifi_task;
1454 	int result;
1455 
1456 	DBG_COUNTER_INC(fast_intrs);
1457 	if (info->ifi_filter != NULL) {
1458 		result = info->ifi_filter(info->ifi_filter_arg);
1459 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1460 			return (result);
1461 	}
1462 
1463 	GROUPTASK_ENQUEUE(gtask);
1464 	return (FILTER_HANDLED);
1465 }
1466 
1467 static int
1468 iflib_fast_intr_rxtx(void *arg)
1469 {
1470 	iflib_filter_info_t info = arg;
1471 	struct grouptask *gtask = info->ifi_task;
1472 	if_ctx_t ctx;
1473 	iflib_rxq_t rxq = (iflib_rxq_t)info->ifi_ctx;
1474 	iflib_txq_t txq;
1475 	void *sc;
1476 	int i, cidx, result;
1477 	qidx_t txqid;
1478 	bool intr_enable, intr_legacy;
1479 
1480 	DBG_COUNTER_INC(fast_intrs);
1481 	if (info->ifi_filter != NULL) {
1482 		result = info->ifi_filter(info->ifi_filter_arg);
1483 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1484 			return (result);
1485 	}
1486 
1487 	ctx = rxq->ifr_ctx;
1488 	sc = ctx->ifc_softc;
1489 	intr_enable = false;
1490 	intr_legacy = !!(ctx->ifc_flags & IFC_LEGACY);
1491 	MPASS(rxq->ifr_ntxqirq);
1492 	for (i = 0; i < rxq->ifr_ntxqirq; i++) {
1493 		txqid = rxq->ifr_txqid[i];
1494 		txq = &ctx->ifc_txqs[txqid];
1495 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1496 		    BUS_DMASYNC_POSTREAD);
1497 		if (!ctx->isc_txd_credits_update(sc, txqid, false)) {
1498 			if (intr_legacy)
1499 				intr_enable = true;
1500 			else
1501 				IFDI_TX_QUEUE_INTR_ENABLE(ctx, txqid);
1502 			continue;
1503 		}
1504 		GROUPTASK_ENQUEUE(&txq->ift_task);
1505 	}
1506 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_RXCQ)
1507 		cidx = rxq->ifr_cq_cidx;
1508 	else
1509 		cidx = rxq->ifr_fl[0].ifl_cidx;
1510 	if (iflib_rxd_avail(ctx, rxq, cidx, 1))
1511 		GROUPTASK_ENQUEUE(gtask);
1512 	else {
1513 		if (intr_legacy)
1514 			intr_enable = true;
1515 		else
1516 			IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
1517 		DBG_COUNTER_INC(rx_intr_enables);
1518 	}
1519 	if (intr_enable)
1520 		IFDI_INTR_ENABLE(ctx);
1521 	return (FILTER_HANDLED);
1522 }
1523 
1524 static int
1525 iflib_fast_intr_ctx(void *arg)
1526 {
1527 	iflib_filter_info_t info = arg;
1528 	struct grouptask *gtask = info->ifi_task;
1529 	int result;
1530 
1531 	DBG_COUNTER_INC(fast_intrs);
1532 	if (info->ifi_filter != NULL) {
1533 		result = info->ifi_filter(info->ifi_filter_arg);
1534 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1535 			return (result);
1536 	}
1537 
1538 	GROUPTASK_ENQUEUE(gtask);
1539 	return (FILTER_HANDLED);
1540 }
1541 
1542 static int
1543 _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
1544 		 driver_filter_t filter, driver_intr_t handler, void *arg,
1545 		 const char *name)
1546 {
1547 	struct resource *res;
1548 	void *tag = NULL;
1549 	device_t dev = ctx->ifc_dev;
1550 	int flags, i, rc;
1551 
1552 	flags = RF_ACTIVE;
1553 	if (ctx->ifc_flags & IFC_LEGACY)
1554 		flags |= RF_SHAREABLE;
1555 	MPASS(rid < 512);
1556 	i = rid;
1557 	res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &i, flags);
1558 	if (res == NULL) {
1559 		device_printf(dev,
1560 		    "failed to allocate IRQ for rid %d, name %s.\n", rid, name);
1561 		return (ENOMEM);
1562 	}
1563 	irq->ii_res = res;
1564 	KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL"));
1565 	rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET,
1566 						filter, handler, arg, &tag);
1567 	if (rc != 0) {
1568 		device_printf(dev,
1569 		    "failed to setup interrupt for rid %d, name %s: %d\n",
1570 					  rid, name ? name : "unknown", rc);
1571 		return (rc);
1572 	} else if (name)
1573 		bus_describe_intr(dev, res, tag, "%s", name);
1574 
1575 	irq->ii_tag = tag;
1576 	return (0);
1577 }
1578 
1579 /*********************************************************************
1580  *
1581  *  Allocate DMA resources for TX buffers as well as memory for the TX
1582  *  mbuf map.  TX DMA maps (non-TSO/TSO) and TX mbuf map are kept in a
1583  *  iflib_sw_tx_desc_array structure, storing all the information that
1584  *  is needed to transmit a packet on the wire.  This is called only
1585  *  once at attach, setup is done every reset.
1586  *
1587  **********************************************************************/
1588 static int
1589 iflib_txsd_alloc(iflib_txq_t txq)
1590 {
1591 	if_ctx_t ctx = txq->ift_ctx;
1592 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1593 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1594 	device_t dev = ctx->ifc_dev;
1595 	bus_size_t tsomaxsize;
1596 	int err, nsegments, ntsosegments;
1597 	bool tso;
1598 
1599 	nsegments = scctx->isc_tx_nsegments;
1600 	ntsosegments = scctx->isc_tx_tso_segments_max;
1601 	tsomaxsize = scctx->isc_tx_tso_size_max;
1602 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_VLAN_MTU)
1603 		tsomaxsize += sizeof(struct ether_vlan_header);
1604 	MPASS(scctx->isc_ntxd[0] > 0);
1605 	MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0);
1606 	MPASS(nsegments > 0);
1607 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) {
1608 		MPASS(ntsosegments > 0);
1609 		MPASS(sctx->isc_tso_maxsize >= tsomaxsize);
1610 	}
1611 
1612 	/*
1613 	 * Set up DMA tags for TX buffers.
1614 	 */
1615 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1616 			       1, 0,			/* alignment, bounds */
1617 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1618 			       BUS_SPACE_MAXADDR,	/* highaddr */
1619 			       NULL, NULL,		/* filter, filterarg */
1620 			       sctx->isc_tx_maxsize,		/* maxsize */
1621 			       nsegments,	/* nsegments */
1622 			       sctx->isc_tx_maxsegsize,	/* maxsegsize */
1623 			       0,			/* flags */
1624 			       NULL,			/* lockfunc */
1625 			       NULL,			/* lockfuncarg */
1626 			       &txq->ift_buf_tag))) {
1627 		device_printf(dev,"Unable to allocate TX DMA tag: %d\n", err);
1628 		device_printf(dev,"maxsize: %ju nsegments: %d maxsegsize: %ju\n",
1629 		    (uintmax_t)sctx->isc_tx_maxsize, nsegments, (uintmax_t)sctx->isc_tx_maxsegsize);
1630 		goto fail;
1631 	}
1632 	tso = (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) != 0;
1633 	if (tso && (err = bus_dma_tag_create(bus_get_dma_tag(dev),
1634 			       1, 0,			/* alignment, bounds */
1635 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1636 			       BUS_SPACE_MAXADDR,	/* highaddr */
1637 			       NULL, NULL,		/* filter, filterarg */
1638 			       tsomaxsize,		/* maxsize */
1639 			       ntsosegments,	/* nsegments */
1640 			       sctx->isc_tso_maxsegsize,/* maxsegsize */
1641 			       0,			/* flags */
1642 			       NULL,			/* lockfunc */
1643 			       NULL,			/* lockfuncarg */
1644 			       &txq->ift_tso_buf_tag))) {
1645 		device_printf(dev, "Unable to allocate TSO TX DMA tag: %d\n",
1646 		    err);
1647 		goto fail;
1648 	}
1649 
1650 	/* Allocate memory for the TX mbuf map. */
1651 	if (!(txq->ift_sds.ifsd_m =
1652 	    (struct mbuf **) malloc(sizeof(struct mbuf *) *
1653 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1654 		device_printf(dev, "Unable to allocate TX mbuf map memory\n");
1655 		err = ENOMEM;
1656 		goto fail;
1657 	}
1658 
1659 	/*
1660 	 * Create the DMA maps for TX buffers.
1661 	 */
1662 	if ((txq->ift_sds.ifsd_map = (bus_dmamap_t *)malloc(
1663 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1664 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1665 		device_printf(dev,
1666 		    "Unable to allocate TX buffer DMA map memory\n");
1667 		err = ENOMEM;
1668 		goto fail;
1669 	}
1670 	if (tso && (txq->ift_sds.ifsd_tso_map = (bus_dmamap_t *)malloc(
1671 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1672 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1673 		device_printf(dev,
1674 		    "Unable to allocate TSO TX buffer map memory\n");
1675 		err = ENOMEM;
1676 		goto fail;
1677 	}
1678 	for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) {
1679 		err = bus_dmamap_create(txq->ift_buf_tag, 0,
1680 		    &txq->ift_sds.ifsd_map[i]);
1681 		if (err != 0) {
1682 			device_printf(dev, "Unable to create TX DMA map\n");
1683 			goto fail;
1684 		}
1685 		if (!tso)
1686 			continue;
1687 		err = bus_dmamap_create(txq->ift_tso_buf_tag, 0,
1688 		    &txq->ift_sds.ifsd_tso_map[i]);
1689 		if (err != 0) {
1690 			device_printf(dev, "Unable to create TSO TX DMA map\n");
1691 			goto fail;
1692 		}
1693 	}
1694 	return (0);
1695 fail:
1696 	/* We free all, it handles case where we are in the middle */
1697 	iflib_tx_structures_free(ctx);
1698 	return (err);
1699 }
1700 
1701 static void
1702 iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i)
1703 {
1704 	bus_dmamap_t map;
1705 
1706 	if (txq->ift_sds.ifsd_map != NULL) {
1707 		map = txq->ift_sds.ifsd_map[i];
1708 		bus_dmamap_sync(txq->ift_buf_tag, map, BUS_DMASYNC_POSTWRITE);
1709 		bus_dmamap_unload(txq->ift_buf_tag, map);
1710 		bus_dmamap_destroy(txq->ift_buf_tag, map);
1711 		txq->ift_sds.ifsd_map[i] = NULL;
1712 	}
1713 
1714 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1715 		map = txq->ift_sds.ifsd_tso_map[i];
1716 		bus_dmamap_sync(txq->ift_tso_buf_tag, map,
1717 		    BUS_DMASYNC_POSTWRITE);
1718 		bus_dmamap_unload(txq->ift_tso_buf_tag, map);
1719 		bus_dmamap_destroy(txq->ift_tso_buf_tag, map);
1720 		txq->ift_sds.ifsd_tso_map[i] = NULL;
1721 	}
1722 }
1723 
1724 static void
1725 iflib_txq_destroy(iflib_txq_t txq)
1726 {
1727 	if_ctx_t ctx = txq->ift_ctx;
1728 
1729 	for (int i = 0; i < txq->ift_size; i++)
1730 		iflib_txsd_destroy(ctx, txq, i);
1731 
1732 	if (txq->ift_br != NULL) {
1733 		ifmp_ring_free(txq->ift_br);
1734 		txq->ift_br = NULL;
1735 	}
1736 
1737 	mtx_destroy(&txq->ift_mtx);
1738 
1739 	if (txq->ift_sds.ifsd_map != NULL) {
1740 		free(txq->ift_sds.ifsd_map, M_IFLIB);
1741 		txq->ift_sds.ifsd_map = NULL;
1742 	}
1743 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1744 		free(txq->ift_sds.ifsd_tso_map, M_IFLIB);
1745 		txq->ift_sds.ifsd_tso_map = NULL;
1746 	}
1747 	if (txq->ift_sds.ifsd_m != NULL) {
1748 		free(txq->ift_sds.ifsd_m, M_IFLIB);
1749 		txq->ift_sds.ifsd_m = NULL;
1750 	}
1751 	if (txq->ift_buf_tag != NULL) {
1752 		bus_dma_tag_destroy(txq->ift_buf_tag);
1753 		txq->ift_buf_tag = NULL;
1754 	}
1755 	if (txq->ift_tso_buf_tag != NULL) {
1756 		bus_dma_tag_destroy(txq->ift_tso_buf_tag);
1757 		txq->ift_tso_buf_tag = NULL;
1758 	}
1759 	if (txq->ift_ifdi != NULL) {
1760 		free(txq->ift_ifdi, M_IFLIB);
1761 	}
1762 }
1763 
1764 static void
1765 iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i)
1766 {
1767 	struct mbuf **mp;
1768 
1769 	mp = &txq->ift_sds.ifsd_m[i];
1770 	if (*mp == NULL)
1771 		return;
1772 
1773 	if (txq->ift_sds.ifsd_map != NULL) {
1774 		bus_dmamap_sync(txq->ift_buf_tag,
1775 		    txq->ift_sds.ifsd_map[i], BUS_DMASYNC_POSTWRITE);
1776 		bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[i]);
1777 	}
1778 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1779 		bus_dmamap_sync(txq->ift_tso_buf_tag,
1780 		    txq->ift_sds.ifsd_tso_map[i], BUS_DMASYNC_POSTWRITE);
1781 		bus_dmamap_unload(txq->ift_tso_buf_tag,
1782 		    txq->ift_sds.ifsd_tso_map[i]);
1783 	}
1784 	m_freem(*mp);
1785 	DBG_COUNTER_INC(tx_frees);
1786 	*mp = NULL;
1787 }
1788 
1789 static int
1790 iflib_txq_setup(iflib_txq_t txq)
1791 {
1792 	if_ctx_t ctx = txq->ift_ctx;
1793 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1794 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1795 	iflib_dma_info_t di;
1796 	int i;
1797 
1798 	/* Set number of descriptors available */
1799 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
1800 	/* XXX make configurable */
1801 	txq->ift_update_freq = IFLIB_DEFAULT_TX_UPDATE_FREQ;
1802 
1803 	/* Reset indices */
1804 	txq->ift_cidx_processed = 0;
1805 	txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0;
1806 	txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset];
1807 
1808 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1809 		bzero((void *)di->idi_vaddr, di->idi_size);
1810 
1811 	IFDI_TXQ_SETUP(ctx, txq->ift_id);
1812 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1813 		bus_dmamap_sync(di->idi_tag, di->idi_map,
1814 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1815 	return (0);
1816 }
1817 
1818 /*********************************************************************
1819  *
1820  *  Allocate DMA resources for RX buffers as well as memory for the RX
1821  *  mbuf map, direct RX cluster pointer map and RX cluster bus address
1822  *  map.  RX DMA map, RX mbuf map, direct RX cluster pointer map and
1823  *  RX cluster map are kept in a iflib_sw_rx_desc_array structure.
1824  *  Since we use use one entry in iflib_sw_rx_desc_array per received
1825  *  packet, the maximum number of entries we'll need is equal to the
1826  *  number of hardware receive descriptors that we've allocated.
1827  *
1828  **********************************************************************/
1829 static int
1830 iflib_rxsd_alloc(iflib_rxq_t rxq)
1831 {
1832 	if_ctx_t ctx = rxq->ifr_ctx;
1833 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1834 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1835 	device_t dev = ctx->ifc_dev;
1836 	iflib_fl_t fl;
1837 	int			err;
1838 
1839 	MPASS(scctx->isc_nrxd[0] > 0);
1840 	MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0);
1841 
1842 	fl = rxq->ifr_fl;
1843 	for (int i = 0; i <  rxq->ifr_nfl; i++, fl++) {
1844 		fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */
1845 		/* Set up DMA tag for RX buffers. */
1846 		err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1847 					 1, 0,			/* alignment, bounds */
1848 					 BUS_SPACE_MAXADDR,	/* lowaddr */
1849 					 BUS_SPACE_MAXADDR,	/* highaddr */
1850 					 NULL, NULL,		/* filter, filterarg */
1851 					 sctx->isc_rx_maxsize,	/* maxsize */
1852 					 sctx->isc_rx_nsegments,	/* nsegments */
1853 					 sctx->isc_rx_maxsegsize,	/* maxsegsize */
1854 					 0,			/* flags */
1855 					 NULL,			/* lockfunc */
1856 					 NULL,			/* lockarg */
1857 					 &fl->ifl_buf_tag);
1858 		if (err) {
1859 			device_printf(dev,
1860 			    "Unable to allocate RX DMA tag: %d\n", err);
1861 			goto fail;
1862 		}
1863 
1864 		/* Allocate memory for the RX mbuf map. */
1865 		if (!(fl->ifl_sds.ifsd_m =
1866 		      (struct mbuf **) malloc(sizeof(struct mbuf *) *
1867 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1868 			device_printf(dev,
1869 			    "Unable to allocate RX mbuf map memory\n");
1870 			err = ENOMEM;
1871 			goto fail;
1872 		}
1873 
1874 		/* Allocate memory for the direct RX cluster pointer map. */
1875 		if (!(fl->ifl_sds.ifsd_cl =
1876 		      (caddr_t *) malloc(sizeof(caddr_t) *
1877 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1878 			device_printf(dev,
1879 			    "Unable to allocate RX cluster map memory\n");
1880 			err = ENOMEM;
1881 			goto fail;
1882 		}
1883 
1884 		/* Allocate memory for the RX cluster bus address map. */
1885 		if (!(fl->ifl_sds.ifsd_ba =
1886 		      (bus_addr_t *) malloc(sizeof(bus_addr_t) *
1887 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1888 			device_printf(dev,
1889 			    "Unable to allocate RX bus address map memory\n");
1890 			err = ENOMEM;
1891 			goto fail;
1892 		}
1893 
1894 		/*
1895 		 * Create the DMA maps for RX buffers.
1896 		 */
1897 		if (!(fl->ifl_sds.ifsd_map =
1898 		      (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1899 			device_printf(dev,
1900 			    "Unable to allocate RX buffer DMA map memory\n");
1901 			err = ENOMEM;
1902 			goto fail;
1903 		}
1904 		for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++) {
1905 			err = bus_dmamap_create(fl->ifl_buf_tag, 0,
1906 			    &fl->ifl_sds.ifsd_map[i]);
1907 			if (err != 0) {
1908 				device_printf(dev, "Unable to create RX buffer DMA map\n");
1909 				goto fail;
1910 			}
1911 		}
1912 	}
1913 	return (0);
1914 
1915 fail:
1916 	iflib_rx_structures_free(ctx);
1917 	return (err);
1918 }
1919 
1920 /*
1921  * Internal service routines
1922  */
1923 
1924 struct rxq_refill_cb_arg {
1925 	int               error;
1926 	bus_dma_segment_t seg;
1927 	int               nseg;
1928 };
1929 
1930 static void
1931 _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
1932 {
1933 	struct rxq_refill_cb_arg *cb_arg = arg;
1934 
1935 	cb_arg->error = error;
1936 	cb_arg->seg = segs[0];
1937 	cb_arg->nseg = nseg;
1938 }
1939 
1940 /**
1941  * iflib_fl_refill - refill an rxq free-buffer list
1942  * @ctx: the iflib context
1943  * @fl: the free list to refill
1944  * @count: the number of new buffers to allocate
1945  *
1946  * (Re)populate an rxq free-buffer list with up to @count new packet buffers.
1947  * The caller must assure that @count does not exceed the queue's capacity
1948  * minus one (since we always leave a descriptor unavailable).
1949  */
1950 static uint8_t
1951 iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
1952 {
1953 	struct if_rxd_update iru;
1954 	struct rxq_refill_cb_arg cb_arg;
1955 	struct mbuf *m;
1956 	caddr_t cl, *sd_cl;
1957 	struct mbuf **sd_m;
1958 	bus_dmamap_t *sd_map;
1959 	bus_addr_t bus_addr, *sd_ba;
1960 	int err, frag_idx, i, idx, n, pidx;
1961 	qidx_t credits;
1962 
1963 	MPASS(count <= fl->ifl_size - fl->ifl_credits - 1);
1964 
1965 	sd_m = fl->ifl_sds.ifsd_m;
1966 	sd_map = fl->ifl_sds.ifsd_map;
1967 	sd_cl = fl->ifl_sds.ifsd_cl;
1968 	sd_ba = fl->ifl_sds.ifsd_ba;
1969 	pidx = fl->ifl_pidx;
1970 	idx = pidx;
1971 	frag_idx = fl->ifl_fragidx;
1972 	credits = fl->ifl_credits;
1973 
1974 	i = 0;
1975 	n = count;
1976 	MPASS(n > 0);
1977 	MPASS(credits + n <= fl->ifl_size);
1978 
1979 	if (pidx < fl->ifl_cidx)
1980 		MPASS(pidx + n <= fl->ifl_cidx);
1981 	if (pidx == fl->ifl_cidx && (credits < fl->ifl_size))
1982 		MPASS(fl->ifl_gen == 0);
1983 	if (pidx > fl->ifl_cidx)
1984 		MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx);
1985 
1986 	DBG_COUNTER_INC(fl_refills);
1987 	if (n > 8)
1988 		DBG_COUNTER_INC(fl_refills_large);
1989 	iru_init(&iru, fl->ifl_rxq, fl->ifl_id);
1990 	while (n-- > 0) {
1991 		/*
1992 		 * We allocate an uninitialized mbuf + cluster, mbuf is
1993 		 * initialized after rx.
1994 		 *
1995 		 * If the cluster is still set then we know a minimum sized
1996 		 * packet was received
1997 		 */
1998 		bit_ffc_at(fl->ifl_rx_bitmap, frag_idx, fl->ifl_size,
1999 		    &frag_idx);
2000 		if (frag_idx < 0)
2001 			bit_ffc(fl->ifl_rx_bitmap, fl->ifl_size, &frag_idx);
2002 		MPASS(frag_idx >= 0);
2003 		if ((cl = sd_cl[frag_idx]) == NULL) {
2004 			cl = uma_zalloc(fl->ifl_zone, M_NOWAIT);
2005 			if (__predict_false(cl == NULL))
2006 				break;
2007 
2008 			cb_arg.error = 0;
2009 			MPASS(sd_map != NULL);
2010 			err = bus_dmamap_load(fl->ifl_buf_tag, sd_map[frag_idx],
2011 			    cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg,
2012 			    BUS_DMA_NOWAIT);
2013 			if (__predict_false(err != 0 || cb_arg.error)) {
2014 				uma_zfree(fl->ifl_zone, cl);
2015 				break;
2016 			}
2017 
2018 			sd_ba[frag_idx] = bus_addr = cb_arg.seg.ds_addr;
2019 			sd_cl[frag_idx] = cl;
2020 #if MEMORY_LOGGING
2021 			fl->ifl_cl_enqueued++;
2022 #endif
2023 		} else {
2024 			bus_addr = sd_ba[frag_idx];
2025 		}
2026 		bus_dmamap_sync(fl->ifl_buf_tag, sd_map[frag_idx],
2027 		    BUS_DMASYNC_PREREAD);
2028 
2029 		if (sd_m[frag_idx] == NULL) {
2030 			m = m_gethdr(M_NOWAIT, MT_NOINIT);
2031 			if (__predict_false(m == NULL))
2032 				break;
2033 			sd_m[frag_idx] = m;
2034 		}
2035 		bit_set(fl->ifl_rx_bitmap, frag_idx);
2036 #if MEMORY_LOGGING
2037 		fl->ifl_m_enqueued++;
2038 #endif
2039 
2040 		DBG_COUNTER_INC(rx_allocs);
2041 		fl->ifl_rxd_idxs[i] = frag_idx;
2042 		fl->ifl_bus_addrs[i] = bus_addr;
2043 		credits++;
2044 		i++;
2045 		MPASS(credits <= fl->ifl_size);
2046 		if (++idx == fl->ifl_size) {
2047 #ifdef INVARIANTS
2048 			fl->ifl_gen = 1;
2049 #endif
2050 			idx = 0;
2051 		}
2052 		if (n == 0 || i == IFLIB_MAX_RX_REFRESH) {
2053 			iru.iru_pidx = pidx;
2054 			iru.iru_count = i;
2055 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2056 			fl->ifl_pidx = idx;
2057 			fl->ifl_credits = credits;
2058 			pidx = idx;
2059 			i = 0;
2060 		}
2061 	}
2062 
2063 	if (n < count - 1) {
2064 		if (i != 0) {
2065 			iru.iru_pidx = pidx;
2066 			iru.iru_count = i;
2067 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2068 			fl->ifl_pidx = idx;
2069 			fl->ifl_credits = credits;
2070 		}
2071 		DBG_COUNTER_INC(rxd_flush);
2072 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2073 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2074 		ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id,
2075 		    fl->ifl_id, fl->ifl_pidx);
2076 		if (__predict_true(bit_test(fl->ifl_rx_bitmap, frag_idx))) {
2077 			fl->ifl_fragidx = frag_idx + 1;
2078 			if (fl->ifl_fragidx == fl->ifl_size)
2079 				fl->ifl_fragidx = 0;
2080 		} else {
2081 			fl->ifl_fragidx = frag_idx;
2082 		}
2083 	}
2084 
2085 	return (n == -1 ? 0 : IFLIB_RXEOF_EMPTY);
2086 }
2087 
2088 static inline uint8_t
2089 iflib_fl_refill_all(if_ctx_t ctx, iflib_fl_t fl)
2090 {
2091 	/*
2092 	 * We leave an unused descriptor to avoid pidx to catch up with cidx.
2093 	 * This is important as it confuses most NICs. For instance,
2094 	 * Intel NICs have (per receive ring) RDH and RDT registers, where
2095 	 * RDH points to the next receive descriptor to be used by the NIC,
2096 	 * and RDT for the next receive descriptor to be published by the
2097 	 * driver to the NIC (RDT - 1 is thus the last valid one).
2098 	 * The condition RDH == RDT means no descriptors are available to
2099 	 * the NIC, and thus it would be ambiguous if it also meant that
2100 	 * all the descriptors are available to the NIC.
2101 	 */
2102 	int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1;
2103 #ifdef INVARIANTS
2104 	int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1;
2105 #endif
2106 
2107 	MPASS(fl->ifl_credits <= fl->ifl_size);
2108 	MPASS(reclaimable == delta);
2109 
2110 	if (reclaimable > 0)
2111 		return (iflib_fl_refill(ctx, fl, reclaimable));
2112 	return (0);
2113 }
2114 
2115 uint8_t
2116 iflib_in_detach(if_ctx_t ctx)
2117 {
2118 	bool in_detach;
2119 
2120 	STATE_LOCK(ctx);
2121 	in_detach = !!(ctx->ifc_flags & IFC_IN_DETACH);
2122 	STATE_UNLOCK(ctx);
2123 	return (in_detach);
2124 }
2125 
2126 static void
2127 iflib_fl_bufs_free(iflib_fl_t fl)
2128 {
2129 	iflib_dma_info_t idi = fl->ifl_ifdi;
2130 	bus_dmamap_t sd_map;
2131 	uint32_t i;
2132 
2133 	for (i = 0; i < fl->ifl_size; i++) {
2134 		struct mbuf **sd_m = &fl->ifl_sds.ifsd_m[i];
2135 		caddr_t *sd_cl = &fl->ifl_sds.ifsd_cl[i];
2136 
2137 		if (*sd_cl != NULL) {
2138 			sd_map = fl->ifl_sds.ifsd_map[i];
2139 			bus_dmamap_sync(fl->ifl_buf_tag, sd_map,
2140 			    BUS_DMASYNC_POSTREAD);
2141 			bus_dmamap_unload(fl->ifl_buf_tag, sd_map);
2142 			uma_zfree(fl->ifl_zone, *sd_cl);
2143 			*sd_cl = NULL;
2144 			if (*sd_m != NULL) {
2145 				m_init(*sd_m, M_NOWAIT, MT_DATA, 0);
2146 				uma_zfree(zone_mbuf, *sd_m);
2147 				*sd_m = NULL;
2148 			}
2149 		} else {
2150 			MPASS(*sd_m == NULL);
2151 		}
2152 #if MEMORY_LOGGING
2153 		fl->ifl_m_dequeued++;
2154 		fl->ifl_cl_dequeued++;
2155 #endif
2156 	}
2157 #ifdef INVARIANTS
2158 	for (i = 0; i < fl->ifl_size; i++) {
2159 		MPASS(fl->ifl_sds.ifsd_cl[i] == NULL);
2160 		MPASS(fl->ifl_sds.ifsd_m[i] == NULL);
2161 	}
2162 #endif
2163 	/*
2164 	 * Reset free list values
2165 	 */
2166 	fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = fl->ifl_fragidx = 0;
2167 	bzero(idi->idi_vaddr, idi->idi_size);
2168 }
2169 
2170 /*********************************************************************
2171  *
2172  *  Initialize a free list and its buffers.
2173  *
2174  **********************************************************************/
2175 static int
2176 iflib_fl_setup(iflib_fl_t fl)
2177 {
2178 	iflib_rxq_t rxq = fl->ifl_rxq;
2179 	if_ctx_t ctx = rxq->ifr_ctx;
2180 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2181 	int qidx;
2182 
2183 	bit_nclear(fl->ifl_rx_bitmap, 0, fl->ifl_size - 1);
2184 	/*
2185 	** Free current RX buffer structs and their mbufs
2186 	*/
2187 	iflib_fl_bufs_free(fl);
2188 	/* Now replenish the mbufs */
2189 	MPASS(fl->ifl_credits == 0);
2190 	qidx = rxq->ifr_fl_offset + fl->ifl_id;
2191 	if (scctx->isc_rxd_buf_size[qidx] != 0)
2192 		fl->ifl_buf_size = scctx->isc_rxd_buf_size[qidx];
2193 	else
2194 		fl->ifl_buf_size = ctx->ifc_rx_mbuf_sz;
2195 	/*
2196 	 * ifl_buf_size may be a driver-supplied value, so pull it up
2197 	 * to the selected mbuf size.
2198 	 */
2199 	fl->ifl_buf_size = iflib_get_mbuf_size_for(fl->ifl_buf_size);
2200 	if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size)
2201 		ctx->ifc_max_fl_buf_size = fl->ifl_buf_size;
2202 	fl->ifl_cltype = m_gettype(fl->ifl_buf_size);
2203 	fl->ifl_zone = m_getzone(fl->ifl_buf_size);
2204 
2205 	/*
2206 	 * Avoid pre-allocating zillions of clusters to an idle card
2207 	 * potentially speeding up attach. In any case make sure
2208 	 * to leave a descriptor unavailable. See the comment in
2209 	 * iflib_fl_refill_all().
2210 	 */
2211 	MPASS(fl->ifl_size > 0);
2212 	(void)iflib_fl_refill(ctx, fl, min(128, fl->ifl_size - 1));
2213 	if (min(128, fl->ifl_size - 1) != fl->ifl_credits)
2214 		return (ENOBUFS);
2215 	/*
2216 	 * handle failure
2217 	 */
2218 	MPASS(rxq != NULL);
2219 	MPASS(fl->ifl_ifdi != NULL);
2220 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2221 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2222 	return (0);
2223 }
2224 
2225 /*********************************************************************
2226  *
2227  *  Free receive ring data structures
2228  *
2229  **********************************************************************/
2230 static void
2231 iflib_rx_sds_free(iflib_rxq_t rxq)
2232 {
2233 	iflib_fl_t fl;
2234 	int i, j;
2235 
2236 	if (rxq->ifr_fl != NULL) {
2237 		for (i = 0; i < rxq->ifr_nfl; i++) {
2238 			fl = &rxq->ifr_fl[i];
2239 			if (fl->ifl_buf_tag != NULL) {
2240 				if (fl->ifl_sds.ifsd_map != NULL) {
2241 					for (j = 0; j < fl->ifl_size; j++) {
2242 						bus_dmamap_sync(
2243 						    fl->ifl_buf_tag,
2244 						    fl->ifl_sds.ifsd_map[j],
2245 						    BUS_DMASYNC_POSTREAD);
2246 						bus_dmamap_unload(
2247 						    fl->ifl_buf_tag,
2248 						    fl->ifl_sds.ifsd_map[j]);
2249 						bus_dmamap_destroy(
2250 						    fl->ifl_buf_tag,
2251 						    fl->ifl_sds.ifsd_map[j]);
2252 					}
2253 				}
2254 				bus_dma_tag_destroy(fl->ifl_buf_tag);
2255 				fl->ifl_buf_tag = NULL;
2256 			}
2257 			free(fl->ifl_sds.ifsd_m, M_IFLIB);
2258 			free(fl->ifl_sds.ifsd_cl, M_IFLIB);
2259 			free(fl->ifl_sds.ifsd_ba, M_IFLIB);
2260 			free(fl->ifl_sds.ifsd_map, M_IFLIB);
2261 			fl->ifl_sds.ifsd_m = NULL;
2262 			fl->ifl_sds.ifsd_cl = NULL;
2263 			fl->ifl_sds.ifsd_ba = NULL;
2264 			fl->ifl_sds.ifsd_map = NULL;
2265 		}
2266 		free(rxq->ifr_fl, M_IFLIB);
2267 		rxq->ifr_fl = NULL;
2268 		free(rxq->ifr_ifdi, M_IFLIB);
2269 		rxq->ifr_ifdi = NULL;
2270 		rxq->ifr_cq_cidx = 0;
2271 	}
2272 }
2273 
2274 /*
2275  * Timer routine
2276  */
2277 static void
2278 iflib_timer(void *arg)
2279 {
2280 	iflib_txq_t txq = arg;
2281 	if_ctx_t ctx = txq->ift_ctx;
2282 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2283 	uint64_t this_tick = ticks;
2284 
2285 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
2286 		return;
2287 
2288 	/*
2289 	** Check on the state of the TX queue(s), this
2290 	** can be done without the lock because its RO
2291 	** and the HUNG state will be static if set.
2292 	*/
2293 	if (this_tick - txq->ift_last_timer_tick >= hz / 2) {
2294 		txq->ift_last_timer_tick = this_tick;
2295 		IFDI_TIMER(ctx, txq->ift_id);
2296 		if ((txq->ift_qstatus == IFLIB_QUEUE_HUNG) &&
2297 		    ((txq->ift_cleaned_prev == txq->ift_cleaned) ||
2298 		     (sctx->isc_pause_frames == 0)))
2299 			goto hung;
2300 
2301 		if (txq->ift_qstatus != IFLIB_QUEUE_IDLE &&
2302 		    ifmp_ring_is_stalled(txq->ift_br)) {
2303 			KASSERT(ctx->ifc_link_state == LINK_STATE_UP, ("queue can't be marked as hung if interface is down"));
2304 			txq->ift_qstatus = IFLIB_QUEUE_HUNG;
2305 		}
2306 		txq->ift_cleaned_prev = txq->ift_cleaned;
2307 	}
2308 	/* handle any laggards */
2309 	if (txq->ift_db_pending)
2310 		GROUPTASK_ENQUEUE(&txq->ift_task);
2311 
2312 	sctx->isc_pause_frames = 0;
2313 	if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)
2314 		callout_reset_on(&txq->ift_timer, hz / 2, iflib_timer, txq, txq->ift_timer.c_cpu);
2315 	return;
2316 
2317  hung:
2318 	device_printf(ctx->ifc_dev,
2319 	    "Watchdog timeout (TX: %d desc avail: %d pidx: %d) -- resetting\n",
2320 	    txq->ift_id, TXQ_AVAIL(txq), txq->ift_pidx);
2321 	STATE_LOCK(ctx);
2322 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2323 	ctx->ifc_flags |= (IFC_DO_WATCHDOG|IFC_DO_RESET);
2324 	iflib_admin_intr_deferred(ctx);
2325 	STATE_UNLOCK(ctx);
2326 }
2327 
2328 static uint16_t
2329 iflib_get_mbuf_size_for(unsigned int size)
2330 {
2331 
2332 	if (size <= MCLBYTES)
2333 		return (MCLBYTES);
2334 	else
2335 		return (MJUMPAGESIZE);
2336 }
2337 
2338 static void
2339 iflib_calc_rx_mbuf_sz(if_ctx_t ctx)
2340 {
2341 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2342 
2343 	/*
2344 	 * XXX don't set the max_frame_size to larger
2345 	 * than the hardware can handle
2346 	 */
2347 	ctx->ifc_rx_mbuf_sz =
2348 	    iflib_get_mbuf_size_for(sctx->isc_max_frame_size);
2349 }
2350 
2351 uint32_t
2352 iflib_get_rx_mbuf_sz(if_ctx_t ctx)
2353 {
2354 
2355 	return (ctx->ifc_rx_mbuf_sz);
2356 }
2357 
2358 static void
2359 iflib_init_locked(if_ctx_t ctx)
2360 {
2361 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2362 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2363 	if_t ifp = ctx->ifc_ifp;
2364 	iflib_fl_t fl;
2365 	iflib_txq_t txq;
2366 	iflib_rxq_t rxq;
2367 	int i, j, tx_ip_csum_flags, tx_ip6_csum_flags;
2368 
2369 	if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2370 	IFDI_INTR_DISABLE(ctx);
2371 
2372 	tx_ip_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP);
2373 	tx_ip6_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | CSUM_IP6_SCTP);
2374 	/* Set hardware offload abilities */
2375 	if_clearhwassist(ifp);
2376 	if (if_getcapenable(ifp) & IFCAP_TXCSUM)
2377 		if_sethwassistbits(ifp, tx_ip_csum_flags, 0);
2378 	if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6)
2379 		if_sethwassistbits(ifp,  tx_ip6_csum_flags, 0);
2380 	if (if_getcapenable(ifp) & IFCAP_TSO4)
2381 		if_sethwassistbits(ifp, CSUM_IP_TSO, 0);
2382 	if (if_getcapenable(ifp) & IFCAP_TSO6)
2383 		if_sethwassistbits(ifp, CSUM_IP6_TSO, 0);
2384 
2385 	for (i = 0, txq = ctx->ifc_txqs; i < sctx->isc_ntxqsets; i++, txq++) {
2386 		CALLOUT_LOCK(txq);
2387 		callout_stop(&txq->ift_timer);
2388 #ifdef DEV_NETMAP
2389 		callout_stop(&txq->ift_netmap_timer);
2390 #endif /* DEV_NETMAP */
2391 		CALLOUT_UNLOCK(txq);
2392 		iflib_netmap_txq_init(ctx, txq);
2393 	}
2394 
2395 	/*
2396 	 * Calculate a suitable Rx mbuf size prior to calling IFDI_INIT, so
2397 	 * that drivers can use the value when setting up the hardware receive
2398 	 * buffers.
2399 	 */
2400 	iflib_calc_rx_mbuf_sz(ctx);
2401 
2402 #ifdef INVARIANTS
2403 	i = if_getdrvflags(ifp);
2404 #endif
2405 	IFDI_INIT(ctx);
2406 	MPASS(if_getdrvflags(ifp) == i);
2407 	for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
2408 		if (iflib_netmap_rxq_init(ctx, rxq) > 0) {
2409 			/* This rxq is in netmap mode. Skip normal init. */
2410 			continue;
2411 		}
2412 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
2413 			if (iflib_fl_setup(fl)) {
2414 				device_printf(ctx->ifc_dev,
2415 				    "setting up free list %d failed - "
2416 				    "check cluster settings\n", j);
2417 				goto done;
2418 			}
2419 		}
2420 	}
2421 done:
2422 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
2423 	IFDI_INTR_ENABLE(ctx);
2424 	txq = ctx->ifc_txqs;
2425 	for (i = 0; i < sctx->isc_ntxqsets; i++, txq++)
2426 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq,
2427 			txq->ift_timer.c_cpu);
2428 }
2429 
2430 static int
2431 iflib_media_change(if_t ifp)
2432 {
2433 	if_ctx_t ctx = if_getsoftc(ifp);
2434 	int err;
2435 
2436 	CTX_LOCK(ctx);
2437 	if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0)
2438 		iflib_init_locked(ctx);
2439 	CTX_UNLOCK(ctx);
2440 	return (err);
2441 }
2442 
2443 static void
2444 iflib_media_status(if_t ifp, struct ifmediareq *ifmr)
2445 {
2446 	if_ctx_t ctx = if_getsoftc(ifp);
2447 
2448 	CTX_LOCK(ctx);
2449 	IFDI_UPDATE_ADMIN_STATUS(ctx);
2450 	IFDI_MEDIA_STATUS(ctx, ifmr);
2451 	CTX_UNLOCK(ctx);
2452 }
2453 
2454 void
2455 iflib_stop(if_ctx_t ctx)
2456 {
2457 	iflib_txq_t txq = ctx->ifc_txqs;
2458 	iflib_rxq_t rxq = ctx->ifc_rxqs;
2459 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2460 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2461 	iflib_dma_info_t di;
2462 	iflib_fl_t fl;
2463 	int i, j;
2464 
2465 	/* Tell the stack that the interface is no longer active */
2466 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2467 
2468 	IFDI_INTR_DISABLE(ctx);
2469 	DELAY(1000);
2470 	IFDI_STOP(ctx);
2471 	DELAY(1000);
2472 
2473 	iflib_debug_reset();
2474 	/* Wait for current tx queue users to exit to disarm watchdog timer. */
2475 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) {
2476 		/* make sure all transmitters have completed before proceeding XXX */
2477 
2478 		CALLOUT_LOCK(txq);
2479 		callout_stop(&txq->ift_timer);
2480 #ifdef DEV_NETMAP
2481 		callout_stop(&txq->ift_netmap_timer);
2482 #endif /* DEV_NETMAP */
2483 		CALLOUT_UNLOCK(txq);
2484 
2485 		/* clean any enqueued buffers */
2486 		iflib_ifmp_purge(txq);
2487 		/* Free any existing tx buffers. */
2488 		for (j = 0; j < txq->ift_size; j++) {
2489 			iflib_txsd_free(ctx, txq, j);
2490 		}
2491 		txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0;
2492 		txq->ift_in_use = txq->ift_gen = txq->ift_cidx = txq->ift_pidx = txq->ift_no_desc_avail = 0;
2493 		txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0;
2494 		txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0;
2495 		txq->ift_pullups = 0;
2496 		ifmp_ring_reset_stats(txq->ift_br);
2497 		for (j = 0, di = txq->ift_ifdi; j < sctx->isc_ntxqs; j++, di++)
2498 			bzero((void *)di->idi_vaddr, di->idi_size);
2499 	}
2500 	for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) {
2501 		/* make sure all transmitters have completed before proceeding XXX */
2502 
2503 		rxq->ifr_cq_cidx = 0;
2504 		for (j = 0, di = rxq->ifr_ifdi; j < sctx->isc_nrxqs; j++, di++)
2505 			bzero((void *)di->idi_vaddr, di->idi_size);
2506 		/* also resets the free lists pidx/cidx */
2507 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
2508 			iflib_fl_bufs_free(fl);
2509 	}
2510 }
2511 
2512 static inline caddr_t
2513 calc_next_rxd(iflib_fl_t fl, int cidx)
2514 {
2515 	qidx_t size;
2516 	int nrxd;
2517 	caddr_t start, end, cur, next;
2518 
2519 	nrxd = fl->ifl_size;
2520 	size = fl->ifl_rxd_size;
2521 	start = fl->ifl_ifdi->idi_vaddr;
2522 
2523 	if (__predict_false(size == 0))
2524 		return (start);
2525 	cur = start + size*cidx;
2526 	end = start + size*nrxd;
2527 	next = CACHE_PTR_NEXT(cur);
2528 	return (next < end ? next : start);
2529 }
2530 
2531 static inline void
2532 prefetch_pkts(iflib_fl_t fl, int cidx)
2533 {
2534 	int nextptr;
2535 	int nrxd = fl->ifl_size;
2536 	caddr_t next_rxd;
2537 
2538 	nextptr = (cidx + CACHE_PTR_INCREMENT) & (nrxd-1);
2539 	prefetch(&fl->ifl_sds.ifsd_m[nextptr]);
2540 	prefetch(&fl->ifl_sds.ifsd_cl[nextptr]);
2541 	next_rxd = calc_next_rxd(fl, cidx);
2542 	prefetch(next_rxd);
2543 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 1) & (nrxd-1)]);
2544 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 2) & (nrxd-1)]);
2545 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 3) & (nrxd-1)]);
2546 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 4) & (nrxd-1)]);
2547 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 1) & (nrxd-1)]);
2548 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 2) & (nrxd-1)]);
2549 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 3) & (nrxd-1)]);
2550 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 4) & (nrxd-1)]);
2551 }
2552 
2553 static struct mbuf *
2554 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, bool unload, if_rxsd_t sd,
2555     int *pf_rv, if_rxd_info_t ri)
2556 {
2557 	bus_dmamap_t map;
2558 	iflib_fl_t fl;
2559 	caddr_t payload;
2560 	struct mbuf *m;
2561 	int flid, cidx, len, next;
2562 
2563 	map = NULL;
2564 	flid = irf->irf_flid;
2565 	cidx = irf->irf_idx;
2566 	fl = &rxq->ifr_fl[flid];
2567 	sd->ifsd_fl = fl;
2568 	m = fl->ifl_sds.ifsd_m[cidx];
2569 	sd->ifsd_cl = &fl->ifl_sds.ifsd_cl[cidx];
2570 	fl->ifl_credits--;
2571 #if MEMORY_LOGGING
2572 	fl->ifl_m_dequeued++;
2573 #endif
2574 	if (rxq->ifr_ctx->ifc_flags & IFC_PREFETCH)
2575 		prefetch_pkts(fl, cidx);
2576 	next = (cidx + CACHE_PTR_INCREMENT) & (fl->ifl_size-1);
2577 	prefetch(&fl->ifl_sds.ifsd_map[next]);
2578 	map = fl->ifl_sds.ifsd_map[cidx];
2579 
2580 	bus_dmamap_sync(fl->ifl_buf_tag, map, BUS_DMASYNC_POSTREAD);
2581 
2582 	if (rxq->pfil != NULL && PFIL_HOOKED_IN(rxq->pfil) && pf_rv != NULL &&
2583 	    irf->irf_len != 0) {
2584 		payload  = *sd->ifsd_cl;
2585 		payload +=  ri->iri_pad;
2586 		len = ri->iri_len - ri->iri_pad;
2587 		*pf_rv = pfil_run_hooks(rxq->pfil, payload, ri->iri_ifp,
2588 		    len | PFIL_MEMPTR | PFIL_IN, NULL);
2589 		switch (*pf_rv) {
2590 		case PFIL_DROPPED:
2591 		case PFIL_CONSUMED:
2592 			/*
2593 			 * The filter ate it.  Everything is recycled.
2594 			 */
2595 			m = NULL;
2596 			unload = 0;
2597 			break;
2598 		case PFIL_REALLOCED:
2599 			/*
2600 			 * The filter copied it.  Everything is recycled.
2601 			 */
2602 			m = pfil_mem2mbuf(payload);
2603 			unload = 0;
2604 			break;
2605 		case PFIL_PASS:
2606 			/*
2607 			 * Filter said it was OK, so receive like
2608 			 * normal
2609 			 */
2610 			fl->ifl_sds.ifsd_m[cidx] = NULL;
2611 			break;
2612 		default:
2613 			MPASS(0);
2614 		}
2615 	} else {
2616 		fl->ifl_sds.ifsd_m[cidx] = NULL;
2617 		*pf_rv = PFIL_PASS;
2618 	}
2619 
2620 	if (unload && irf->irf_len != 0)
2621 		bus_dmamap_unload(fl->ifl_buf_tag, map);
2622 	fl->ifl_cidx = (fl->ifl_cidx + 1) & (fl->ifl_size-1);
2623 	if (__predict_false(fl->ifl_cidx == 0))
2624 		fl->ifl_gen = 0;
2625 	bit_clear(fl->ifl_rx_bitmap, cidx);
2626 	return (m);
2627 }
2628 
2629 static struct mbuf *
2630 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri, if_rxsd_t sd, int *pf_rv)
2631 {
2632 	struct mbuf *m, *mh, *mt;
2633 	caddr_t cl;
2634 	int  *pf_rv_ptr, flags, i, padlen;
2635 	bool consumed;
2636 
2637 	i = 0;
2638 	mh = NULL;
2639 	consumed = false;
2640 	*pf_rv = PFIL_PASS;
2641 	pf_rv_ptr = pf_rv;
2642 	do {
2643 		m = rxd_frag_to_sd(rxq, &ri->iri_frags[i], !consumed, sd,
2644 		    pf_rv_ptr, ri);
2645 
2646 		MPASS(*sd->ifsd_cl != NULL);
2647 
2648 		/*
2649 		 * Exclude zero-length frags & frags from
2650 		 * packets the filter has consumed or dropped
2651 		 */
2652 		if (ri->iri_frags[i].irf_len == 0 || consumed ||
2653 		    *pf_rv == PFIL_CONSUMED || *pf_rv == PFIL_DROPPED) {
2654 			if (mh == NULL) {
2655 				/* everything saved here */
2656 				consumed = true;
2657 				pf_rv_ptr = NULL;
2658 				continue;
2659 			}
2660 			/* XXX we can save the cluster here, but not the mbuf */
2661 			m_init(m, M_NOWAIT, MT_DATA, 0);
2662 			m_free(m);
2663 			continue;
2664 		}
2665 		if (mh == NULL) {
2666 			flags = M_PKTHDR|M_EXT;
2667 			mh = mt = m;
2668 			padlen = ri->iri_pad;
2669 		} else {
2670 			flags = M_EXT;
2671 			mt->m_next = m;
2672 			mt = m;
2673 			/* assuming padding is only on the first fragment */
2674 			padlen = 0;
2675 		}
2676 		cl = *sd->ifsd_cl;
2677 		*sd->ifsd_cl = NULL;
2678 
2679 		/* Can these two be made one ? */
2680 		m_init(m, M_NOWAIT, MT_DATA, flags);
2681 		m_cljset(m, cl, sd->ifsd_fl->ifl_cltype);
2682 		/*
2683 		 * These must follow m_init and m_cljset
2684 		 */
2685 		m->m_data += padlen;
2686 		ri->iri_len -= padlen;
2687 		m->m_len = ri->iri_frags[i].irf_len;
2688 	} while (++i < ri->iri_nfrags);
2689 
2690 	return (mh);
2691 }
2692 
2693 /*
2694  * Process one software descriptor
2695  */
2696 static struct mbuf *
2697 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri)
2698 {
2699 	struct if_rxsd sd;
2700 	struct mbuf *m;
2701 	int pf_rv;
2702 
2703 	/* should I merge this back in now that the two paths are basically duplicated? */
2704 	if (ri->iri_nfrags == 1 &&
2705 	    ri->iri_frags[0].irf_len != 0 &&
2706 	    ri->iri_frags[0].irf_len <= MIN(IFLIB_RX_COPY_THRESH, MHLEN)) {
2707 		m = rxd_frag_to_sd(rxq, &ri->iri_frags[0], false, &sd,
2708 		    &pf_rv, ri);
2709 		if (pf_rv != PFIL_PASS && pf_rv != PFIL_REALLOCED)
2710 			return (m);
2711 		if (pf_rv == PFIL_PASS) {
2712 			m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR);
2713 #ifndef __NO_STRICT_ALIGNMENT
2714 			if (!IP_ALIGNED(m))
2715 				m->m_data += 2;
2716 #endif
2717 			memcpy(m->m_data, *sd.ifsd_cl, ri->iri_len);
2718 			m->m_len = ri->iri_frags[0].irf_len;
2719 		}
2720 	} else {
2721 		m = assemble_segments(rxq, ri, &sd, &pf_rv);
2722 		if (m == NULL)
2723 			return (NULL);
2724 		if (pf_rv != PFIL_PASS && pf_rv != PFIL_REALLOCED)
2725 			return (m);
2726 	}
2727 	m->m_pkthdr.len = ri->iri_len;
2728 	m->m_pkthdr.rcvif = ri->iri_ifp;
2729 	m->m_flags |= ri->iri_flags;
2730 	m->m_pkthdr.ether_vtag = ri->iri_vtag;
2731 	m->m_pkthdr.flowid = ri->iri_flowid;
2732 	M_HASHTYPE_SET(m, ri->iri_rsstype);
2733 	m->m_pkthdr.csum_flags = ri->iri_csum_flags;
2734 	m->m_pkthdr.csum_data = ri->iri_csum_data;
2735 	return (m);
2736 }
2737 
2738 #if defined(INET6) || defined(INET)
2739 static void
2740 iflib_get_ip_forwarding(struct lro_ctrl *lc, bool *v4, bool *v6)
2741 {
2742 	CURVNET_SET(lc->ifp->if_vnet);
2743 #if defined(INET6)
2744 	*v6 = V_ip6_forwarding;
2745 #endif
2746 #if defined(INET)
2747 	*v4 = V_ipforwarding;
2748 #endif
2749 	CURVNET_RESTORE();
2750 }
2751 
2752 /*
2753  * Returns true if it's possible this packet could be LROed.
2754  * if it returns false, it is guaranteed that tcp_lro_rx()
2755  * would not return zero.
2756  */
2757 static bool
2758 iflib_check_lro_possible(struct mbuf *m, bool v4_forwarding, bool v6_forwarding)
2759 {
2760 	struct ether_header *eh;
2761 
2762 	eh = mtod(m, struct ether_header *);
2763 	switch (eh->ether_type) {
2764 #if defined(INET6)
2765 		case htons(ETHERTYPE_IPV6):
2766 			return (!v6_forwarding);
2767 #endif
2768 #if defined (INET)
2769 		case htons(ETHERTYPE_IP):
2770 			return (!v4_forwarding);
2771 #endif
2772 	}
2773 
2774 	return false;
2775 }
2776 #else
2777 static void
2778 iflib_get_ip_forwarding(struct lro_ctrl *lc __unused, bool *v4 __unused, bool *v6 __unused)
2779 {
2780 }
2781 #endif
2782 
2783 static void
2784 _task_fn_rx_watchdog(void *context)
2785 {
2786 	iflib_rxq_t rxq = context;
2787 
2788 	GROUPTASK_ENQUEUE(&rxq->ifr_task);
2789 }
2790 
2791 static uint8_t
2792 iflib_rxeof(iflib_rxq_t rxq, qidx_t budget)
2793 {
2794 	if_t ifp;
2795 	if_ctx_t ctx = rxq->ifr_ctx;
2796 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2797 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2798 	int avail, i;
2799 	qidx_t *cidxp;
2800 	struct if_rxd_info ri;
2801 	int err, budget_left, rx_bytes, rx_pkts;
2802 	iflib_fl_t fl;
2803 	int lro_enabled;
2804 	bool v4_forwarding, v6_forwarding, lro_possible;
2805 	uint8_t retval = 0;
2806 
2807 	/*
2808 	 * XXX early demux data packets so that if_input processing only handles
2809 	 * acks in interrupt context
2810 	 */
2811 	struct mbuf *m, *mh, *mt, *mf;
2812 
2813 	NET_EPOCH_ASSERT();
2814 
2815 	lro_possible = v4_forwarding = v6_forwarding = false;
2816 	ifp = ctx->ifc_ifp;
2817 	mh = mt = NULL;
2818 	MPASS(budget > 0);
2819 	rx_pkts	= rx_bytes = 0;
2820 	if (sctx->isc_flags & IFLIB_HAS_RXCQ)
2821 		cidxp = &rxq->ifr_cq_cidx;
2822 	else
2823 		cidxp = &rxq->ifr_fl[0].ifl_cidx;
2824 	if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) {
2825 		for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2826 			retval |= iflib_fl_refill_all(ctx, fl);
2827 		DBG_COUNTER_INC(rx_unavail);
2828 		return (retval);
2829 	}
2830 
2831 	/* pfil needs the vnet to be set */
2832 	CURVNET_SET_QUIET(ifp->if_vnet);
2833 	for (budget_left = budget; budget_left > 0 && avail > 0;) {
2834 		if (__predict_false(!CTX_ACTIVE(ctx))) {
2835 			DBG_COUNTER_INC(rx_ctx_inactive);
2836 			break;
2837 		}
2838 		/*
2839 		 * Reset client set fields to their default values
2840 		 */
2841 		rxd_info_zero(&ri);
2842 		ri.iri_qsidx = rxq->ifr_id;
2843 		ri.iri_cidx = *cidxp;
2844 		ri.iri_ifp = ifp;
2845 		ri.iri_frags = rxq->ifr_frags;
2846 		err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
2847 
2848 		if (err)
2849 			goto err;
2850 		rx_pkts += 1;
2851 		rx_bytes += ri.iri_len;
2852 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
2853 			*cidxp = ri.iri_cidx;
2854 			/* Update our consumer index */
2855 			/* XXX NB: shurd - check if this is still safe */
2856 			while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0])
2857 				rxq->ifr_cq_cidx -= scctx->isc_nrxd[0];
2858 			/* was this only a completion queue message? */
2859 			if (__predict_false(ri.iri_nfrags == 0))
2860 				continue;
2861 		}
2862 		MPASS(ri.iri_nfrags != 0);
2863 		MPASS(ri.iri_len != 0);
2864 
2865 		/* will advance the cidx on the corresponding free lists */
2866 		m = iflib_rxd_pkt_get(rxq, &ri);
2867 		avail--;
2868 		budget_left--;
2869 		if (avail == 0 && budget_left)
2870 			avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left);
2871 
2872 		if (__predict_false(m == NULL))
2873 			continue;
2874 
2875 		/* imm_pkt: -- cxgb */
2876 		if (mh == NULL)
2877 			mh = mt = m;
2878 		else {
2879 			mt->m_nextpkt = m;
2880 			mt = m;
2881 		}
2882 	}
2883 	CURVNET_RESTORE();
2884 	/* make sure that we can refill faster than drain */
2885 	for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2886 		retval |= iflib_fl_refill_all(ctx, fl);
2887 
2888 	lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO);
2889 	if (lro_enabled)
2890 		iflib_get_ip_forwarding(&rxq->ifr_lc, &v4_forwarding, &v6_forwarding);
2891 	mt = mf = NULL;
2892 	while (mh != NULL) {
2893 		m = mh;
2894 		mh = mh->m_nextpkt;
2895 		m->m_nextpkt = NULL;
2896 #ifndef __NO_STRICT_ALIGNMENT
2897 		if (!IP_ALIGNED(m) && (m = iflib_fixup_rx(m)) == NULL)
2898 			continue;
2899 #endif
2900 		rx_bytes += m->m_pkthdr.len;
2901 		rx_pkts++;
2902 #if defined(INET6) || defined(INET)
2903 		if (lro_enabled) {
2904 			if (!lro_possible) {
2905 				lro_possible = iflib_check_lro_possible(m, v4_forwarding, v6_forwarding);
2906 				if (lro_possible && mf != NULL) {
2907 					ifp->if_input(ifp, mf);
2908 					DBG_COUNTER_INC(rx_if_input);
2909 					mt = mf = NULL;
2910 				}
2911 			}
2912 			if ((m->m_pkthdr.csum_flags & (CSUM_L4_CALC|CSUM_L4_VALID)) ==
2913 			    (CSUM_L4_CALC|CSUM_L4_VALID)) {
2914 				if (lro_possible && tcp_lro_rx(&rxq->ifr_lc, m, 0) == 0)
2915 					continue;
2916 			}
2917 		}
2918 #endif
2919 		if (lro_possible) {
2920 			ifp->if_input(ifp, m);
2921 			DBG_COUNTER_INC(rx_if_input);
2922 			continue;
2923 		}
2924 
2925 		if (mf == NULL)
2926 			mf = m;
2927 		if (mt != NULL)
2928 			mt->m_nextpkt = m;
2929 		mt = m;
2930 	}
2931 	if (mf != NULL) {
2932 		ifp->if_input(ifp, mf);
2933 		DBG_COUNTER_INC(rx_if_input);
2934 	}
2935 
2936 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
2937 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
2938 
2939 	/*
2940 	 * Flush any outstanding LRO work
2941 	 */
2942 #if defined(INET6) || defined(INET)
2943 	tcp_lro_flush_all(&rxq->ifr_lc);
2944 #endif
2945 	if (avail != 0 || iflib_rxd_avail(ctx, rxq, *cidxp, 1) != 0)
2946 		retval |= IFLIB_RXEOF_MORE;
2947 	return (retval);
2948 err:
2949 	STATE_LOCK(ctx);
2950 	ctx->ifc_flags |= IFC_DO_RESET;
2951 	iflib_admin_intr_deferred(ctx);
2952 	STATE_UNLOCK(ctx);
2953 	return (0);
2954 }
2955 
2956 #define TXD_NOTIFY_COUNT(txq) (((txq)->ift_size / (txq)->ift_update_freq)-1)
2957 static inline qidx_t
2958 txq_max_db_deferred(iflib_txq_t txq, qidx_t in_use)
2959 {
2960 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2961 	qidx_t minthresh = txq->ift_size / 8;
2962 	if (in_use > 4*minthresh)
2963 		return (notify_count);
2964 	if (in_use > 2*minthresh)
2965 		return (notify_count >> 1);
2966 	if (in_use > minthresh)
2967 		return (notify_count >> 3);
2968 	return (0);
2969 }
2970 
2971 static inline qidx_t
2972 txq_max_rs_deferred(iflib_txq_t txq)
2973 {
2974 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2975 	qidx_t minthresh = txq->ift_size / 8;
2976 	if (txq->ift_in_use > 4*minthresh)
2977 		return (notify_count);
2978 	if (txq->ift_in_use > 2*minthresh)
2979 		return (notify_count >> 1);
2980 	if (txq->ift_in_use > minthresh)
2981 		return (notify_count >> 2);
2982 	return (2);
2983 }
2984 
2985 #define M_CSUM_FLAGS(m) ((m)->m_pkthdr.csum_flags)
2986 #define M_HAS_VLANTAG(m) (m->m_flags & M_VLANTAG)
2987 
2988 #define TXQ_MAX_DB_DEFERRED(txq, in_use) txq_max_db_deferred((txq), (in_use))
2989 #define TXQ_MAX_RS_DEFERRED(txq) txq_max_rs_deferred(txq)
2990 #define TXQ_MAX_DB_CONSUMED(size) (size >> 4)
2991 
2992 /* forward compatibility for cxgb */
2993 #define FIRST_QSET(ctx) 0
2994 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets)
2995 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets)
2996 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx))
2997 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments))
2998 
2999 /* XXX we should be setting this to something other than zero */
3000 #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh)
3001 #define	MAX_TX_DESC(ctx) max((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max, \
3002     (ctx)->ifc_softc_ctx.isc_tx_nsegments)
3003 
3004 static inline bool
3005 iflib_txd_db_check(if_ctx_t ctx, iflib_txq_t txq, int ring, qidx_t in_use)
3006 {
3007 	qidx_t dbval, max;
3008 	bool rang;
3009 
3010 	rang = false;
3011 	max = TXQ_MAX_DB_DEFERRED(txq, in_use);
3012 	if (ring || txq->ift_db_pending >= max) {
3013 		dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx;
3014 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3015 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
3016 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval);
3017 		txq->ift_db_pending = txq->ift_npending = 0;
3018 		rang = true;
3019 	}
3020 	return (rang);
3021 }
3022 
3023 #ifdef PKT_DEBUG
3024 static void
3025 print_pkt(if_pkt_info_t pi)
3026 {
3027 	printf("pi len:  %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n",
3028 	       pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx);
3029 	printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n",
3030 	       pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag);
3031 	printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n",
3032 	       pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto);
3033 }
3034 #endif
3035 
3036 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO)
3037 #define IS_TX_OFFLOAD4(pi) ((pi)->ipi_csum_flags & (CSUM_IP_TCP | CSUM_IP_TSO))
3038 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO)
3039 #define IS_TX_OFFLOAD6(pi) ((pi)->ipi_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_TSO))
3040 
3041 static int
3042 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp)
3043 {
3044 	if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx;
3045 	struct ether_vlan_header *eh;
3046 	struct mbuf *m;
3047 
3048 	m = *mp;
3049 	if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) &&
3050 	    M_WRITABLE(m) == 0) {
3051 		if ((m = m_dup(m, M_NOWAIT)) == NULL) {
3052 			return (ENOMEM);
3053 		} else {
3054 			m_freem(*mp);
3055 			DBG_COUNTER_INC(tx_frees);
3056 			*mp = m;
3057 		}
3058 	}
3059 
3060 	/*
3061 	 * Determine where frame payload starts.
3062 	 * Jump over vlan headers if already present,
3063 	 * helpful for QinQ too.
3064 	 */
3065 	if (__predict_false(m->m_len < sizeof(*eh))) {
3066 		txq->ift_pullups++;
3067 		if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL))
3068 			return (ENOMEM);
3069 	}
3070 	eh = mtod(m, struct ether_vlan_header *);
3071 	if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
3072 		pi->ipi_etype = ntohs(eh->evl_proto);
3073 		pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
3074 	} else {
3075 		pi->ipi_etype = ntohs(eh->evl_encap_proto);
3076 		pi->ipi_ehdrlen = ETHER_HDR_LEN;
3077 	}
3078 
3079 	switch (pi->ipi_etype) {
3080 #ifdef INET
3081 	case ETHERTYPE_IP:
3082 	{
3083 		struct mbuf *n;
3084 		struct ip *ip = NULL;
3085 		struct tcphdr *th = NULL;
3086 		int minthlen;
3087 
3088 		minthlen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip) + sizeof(*th));
3089 		if (__predict_false(m->m_len < minthlen)) {
3090 			/*
3091 			 * if this code bloat is causing too much of a hit
3092 			 * move it to a separate function and mark it noinline
3093 			 */
3094 			if (m->m_len == pi->ipi_ehdrlen) {
3095 				n = m->m_next;
3096 				MPASS(n);
3097 				if (n->m_len >= sizeof(*ip))  {
3098 					ip = (struct ip *)n->m_data;
3099 					if (n->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3100 						th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3101 				} else {
3102 					txq->ift_pullups++;
3103 					if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
3104 						return (ENOMEM);
3105 					ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3106 				}
3107 			} else {
3108 				txq->ift_pullups++;
3109 				if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
3110 					return (ENOMEM);
3111 				ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3112 				if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3113 					th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3114 			}
3115 		} else {
3116 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3117 			if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3118 				th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3119 		}
3120 		pi->ipi_ip_hlen = ip->ip_hl << 2;
3121 		pi->ipi_ipproto = ip->ip_p;
3122 		pi->ipi_flags |= IPI_TX_IPV4;
3123 
3124 		/* TCP checksum offload may require TCP header length */
3125 		if (IS_TX_OFFLOAD4(pi)) {
3126 			if (__predict_true(pi->ipi_ipproto == IPPROTO_TCP)) {
3127 				if (__predict_false(th == NULL)) {
3128 					txq->ift_pullups++;
3129 					if (__predict_false((m = m_pullup(m, (ip->ip_hl << 2) + sizeof(*th))) == NULL))
3130 						return (ENOMEM);
3131 					th = (struct tcphdr *)((caddr_t)ip + pi->ipi_ip_hlen);
3132 				}
3133 				pi->ipi_tcp_hflags = th->th_flags;
3134 				pi->ipi_tcp_hlen = th->th_off << 2;
3135 				pi->ipi_tcp_seq = th->th_seq;
3136 			}
3137 			if (IS_TSO4(pi)) {
3138 				if (__predict_false(ip->ip_p != IPPROTO_TCP))
3139 					return (ENXIO);
3140 				/*
3141 				 * TSO always requires hardware checksum offload.
3142 				 */
3143 				pi->ipi_csum_flags |= (CSUM_IP_TCP | CSUM_IP);
3144 				th->th_sum = in_pseudo(ip->ip_src.s_addr,
3145 						       ip->ip_dst.s_addr, htons(IPPROTO_TCP));
3146 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3147 				if (sctx->isc_flags & IFLIB_TSO_INIT_IP) {
3148 					ip->ip_sum = 0;
3149 					ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz);
3150 				}
3151 			}
3152 		}
3153 		if ((sctx->isc_flags & IFLIB_NEED_ZERO_CSUM) && (pi->ipi_csum_flags & CSUM_IP))
3154                        ip->ip_sum = 0;
3155 
3156 		break;
3157 	}
3158 #endif
3159 #ifdef INET6
3160 	case ETHERTYPE_IPV6:
3161 	{
3162 		struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
3163 		struct tcphdr *th;
3164 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
3165 
3166 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
3167 			txq->ift_pullups++;
3168 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
3169 				return (ENOMEM);
3170 		}
3171 		th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
3172 
3173 		/* XXX-BZ this will go badly in case of ext hdrs. */
3174 		pi->ipi_ipproto = ip6->ip6_nxt;
3175 		pi->ipi_flags |= IPI_TX_IPV6;
3176 
3177 		/* TCP checksum offload may require TCP header length */
3178 		if (IS_TX_OFFLOAD6(pi)) {
3179 			if (pi->ipi_ipproto == IPPROTO_TCP) {
3180 				if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) {
3181 					txq->ift_pullups++;
3182 					if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL))
3183 						return (ENOMEM);
3184 				}
3185 				pi->ipi_tcp_hflags = th->th_flags;
3186 				pi->ipi_tcp_hlen = th->th_off << 2;
3187 				pi->ipi_tcp_seq = th->th_seq;
3188 			}
3189 			if (IS_TSO6(pi)) {
3190 				if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP))
3191 					return (ENXIO);
3192 				/*
3193 				 * TSO always requires hardware checksum offload.
3194 				 */
3195 				pi->ipi_csum_flags |= CSUM_IP6_TCP;
3196 				th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0);
3197 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3198 			}
3199 		}
3200 		break;
3201 	}
3202 #endif
3203 	default:
3204 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
3205 		pi->ipi_ip_hlen = 0;
3206 		break;
3207 	}
3208 	*mp = m;
3209 
3210 	return (0);
3211 }
3212 
3213 /*
3214  * If dodgy hardware rejects the scatter gather chain we've handed it
3215  * we'll need to remove the mbuf chain from ifsg_m[] before we can add the
3216  * m_defrag'd mbufs
3217  */
3218 static __noinline struct mbuf *
3219 iflib_remove_mbuf(iflib_txq_t txq)
3220 {
3221 	int ntxd, pidx;
3222 	struct mbuf *m, **ifsd_m;
3223 
3224 	ifsd_m = txq->ift_sds.ifsd_m;
3225 	ntxd = txq->ift_size;
3226 	pidx = txq->ift_pidx & (ntxd - 1);
3227 	ifsd_m = txq->ift_sds.ifsd_m;
3228 	m = ifsd_m[pidx];
3229 	ifsd_m[pidx] = NULL;
3230 	bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[pidx]);
3231 	if (txq->ift_sds.ifsd_tso_map != NULL)
3232 		bus_dmamap_unload(txq->ift_tso_buf_tag,
3233 		    txq->ift_sds.ifsd_tso_map[pidx]);
3234 #if MEMORY_LOGGING
3235 	txq->ift_dequeued++;
3236 #endif
3237 	return (m);
3238 }
3239 
3240 static inline caddr_t
3241 calc_next_txd(iflib_txq_t txq, int cidx, uint8_t qid)
3242 {
3243 	qidx_t size;
3244 	int ntxd;
3245 	caddr_t start, end, cur, next;
3246 
3247 	ntxd = txq->ift_size;
3248 	size = txq->ift_txd_size[qid];
3249 	start = txq->ift_ifdi[qid].idi_vaddr;
3250 
3251 	if (__predict_false(size == 0))
3252 		return (start);
3253 	cur = start + size*cidx;
3254 	end = start + size*ntxd;
3255 	next = CACHE_PTR_NEXT(cur);
3256 	return (next < end ? next : start);
3257 }
3258 
3259 /*
3260  * Pad an mbuf to ensure a minimum ethernet frame size.
3261  * min_frame_size is the frame size (less CRC) to pad the mbuf to
3262  */
3263 static __noinline int
3264 iflib_ether_pad(device_t dev, struct mbuf **m_head, uint16_t min_frame_size)
3265 {
3266 	/*
3267 	 * 18 is enough bytes to pad an ARP packet to 46 bytes, and
3268 	 * and ARP message is the smallest common payload I can think of
3269 	 */
3270 	static char pad[18];	/* just zeros */
3271 	int n;
3272 	struct mbuf *new_head;
3273 
3274 	if (!M_WRITABLE(*m_head)) {
3275 		new_head = m_dup(*m_head, M_NOWAIT);
3276 		if (new_head == NULL) {
3277 			m_freem(*m_head);
3278 			device_printf(dev, "cannot pad short frame, m_dup() failed");
3279 			DBG_COUNTER_INC(encap_pad_mbuf_fail);
3280 			DBG_COUNTER_INC(tx_frees);
3281 			return ENOMEM;
3282 		}
3283 		m_freem(*m_head);
3284 		*m_head = new_head;
3285 	}
3286 
3287 	for (n = min_frame_size - (*m_head)->m_pkthdr.len;
3288 	     n > 0; n -= sizeof(pad))
3289 		if (!m_append(*m_head, min(n, sizeof(pad)), pad))
3290 			break;
3291 
3292 	if (n > 0) {
3293 		m_freem(*m_head);
3294 		device_printf(dev, "cannot pad short frame\n");
3295 		DBG_COUNTER_INC(encap_pad_mbuf_fail);
3296 		DBG_COUNTER_INC(tx_frees);
3297 		return (ENOBUFS);
3298 	}
3299 
3300 	return 0;
3301 }
3302 
3303 static int
3304 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp)
3305 {
3306 	if_ctx_t		ctx;
3307 	if_shared_ctx_t		sctx;
3308 	if_softc_ctx_t		scctx;
3309 	bus_dma_tag_t		buf_tag;
3310 	bus_dma_segment_t	*segs;
3311 	struct mbuf		*m_head, **ifsd_m;
3312 	void			*next_txd;
3313 	bus_dmamap_t		map;
3314 	struct if_pkt_info	pi;
3315 	int remap = 0;
3316 	int err, nsegs, ndesc, max_segs, pidx, cidx, next, ntxd;
3317 
3318 	ctx = txq->ift_ctx;
3319 	sctx = ctx->ifc_sctx;
3320 	scctx = &ctx->ifc_softc_ctx;
3321 	segs = txq->ift_segs;
3322 	ntxd = txq->ift_size;
3323 	m_head = *m_headp;
3324 	map = NULL;
3325 
3326 	/*
3327 	 * If we're doing TSO the next descriptor to clean may be quite far ahead
3328 	 */
3329 	cidx = txq->ift_cidx;
3330 	pidx = txq->ift_pidx;
3331 	if (ctx->ifc_flags & IFC_PREFETCH) {
3332 		next = (cidx + CACHE_PTR_INCREMENT) & (ntxd-1);
3333 		if (!(ctx->ifc_flags & IFLIB_HAS_TXCQ)) {
3334 			next_txd = calc_next_txd(txq, cidx, 0);
3335 			prefetch(next_txd);
3336 		}
3337 
3338 		/* prefetch the next cache line of mbuf pointers and flags */
3339 		prefetch(&txq->ift_sds.ifsd_m[next]);
3340 		prefetch(&txq->ift_sds.ifsd_map[next]);
3341 		next = (cidx + CACHE_LINE_SIZE) & (ntxd-1);
3342 	}
3343 	map = txq->ift_sds.ifsd_map[pidx];
3344 	ifsd_m = txq->ift_sds.ifsd_m;
3345 
3346 	if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
3347 		buf_tag = txq->ift_tso_buf_tag;
3348 		max_segs = scctx->isc_tx_tso_segments_max;
3349 		map = txq->ift_sds.ifsd_tso_map[pidx];
3350 		MPASS(buf_tag != NULL);
3351 		MPASS(max_segs > 0);
3352 	} else {
3353 		buf_tag = txq->ift_buf_tag;
3354 		max_segs = scctx->isc_tx_nsegments;
3355 		map = txq->ift_sds.ifsd_map[pidx];
3356 	}
3357 	if ((sctx->isc_flags & IFLIB_NEED_ETHER_PAD) &&
3358 	    __predict_false(m_head->m_pkthdr.len < scctx->isc_min_frame_size)) {
3359 		err = iflib_ether_pad(ctx->ifc_dev, m_headp, scctx->isc_min_frame_size);
3360 		if (err) {
3361 			DBG_COUNTER_INC(encap_txd_encap_fail);
3362 			return err;
3363 		}
3364 	}
3365 	m_head = *m_headp;
3366 
3367 	pkt_info_zero(&pi);
3368 	pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG|M_BCAST|M_MCAST));
3369 	pi.ipi_pidx = pidx;
3370 	pi.ipi_qsidx = txq->ift_id;
3371 	pi.ipi_len = m_head->m_pkthdr.len;
3372 	pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags;
3373 	pi.ipi_vtag = M_HAS_VLANTAG(m_head) ? m_head->m_pkthdr.ether_vtag : 0;
3374 
3375 	/* deliberate bitwise OR to make one condition */
3376 	if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) {
3377 		if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0)) {
3378 			DBG_COUNTER_INC(encap_txd_encap_fail);
3379 			return (err);
3380 		}
3381 		m_head = *m_headp;
3382 	}
3383 
3384 retry:
3385 	err = bus_dmamap_load_mbuf_sg(buf_tag, map, m_head, segs, &nsegs,
3386 	    BUS_DMA_NOWAIT);
3387 defrag:
3388 	if (__predict_false(err)) {
3389 		switch (err) {
3390 		case EFBIG:
3391 			/* try collapse once and defrag once */
3392 			if (remap == 0) {
3393 				m_head = m_collapse(*m_headp, M_NOWAIT, max_segs);
3394 				/* try defrag if collapsing fails */
3395 				if (m_head == NULL)
3396 					remap++;
3397 			}
3398 			if (remap == 1) {
3399 				txq->ift_mbuf_defrag++;
3400 				m_head = m_defrag(*m_headp, M_NOWAIT);
3401 			}
3402 			/*
3403 			 * remap should never be >1 unless bus_dmamap_load_mbuf_sg
3404 			 * failed to map an mbuf that was run through m_defrag
3405 			 */
3406 			MPASS(remap <= 1);
3407 			if (__predict_false(m_head == NULL || remap > 1))
3408 				goto defrag_failed;
3409 			remap++;
3410 			*m_headp = m_head;
3411 			goto retry;
3412 			break;
3413 		case ENOMEM:
3414 			txq->ift_no_tx_dma_setup++;
3415 			break;
3416 		default:
3417 			txq->ift_no_tx_dma_setup++;
3418 			m_freem(*m_headp);
3419 			DBG_COUNTER_INC(tx_frees);
3420 			*m_headp = NULL;
3421 			break;
3422 		}
3423 		txq->ift_map_failed++;
3424 		DBG_COUNTER_INC(encap_load_mbuf_fail);
3425 		DBG_COUNTER_INC(encap_txd_encap_fail);
3426 		return (err);
3427 	}
3428 	ifsd_m[pidx] = m_head;
3429 	/*
3430 	 * XXX assumes a 1 to 1 relationship between segments and
3431 	 *        descriptors - this does not hold true on all drivers, e.g.
3432 	 *        cxgb
3433 	 */
3434 	if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) {
3435 		txq->ift_no_desc_avail++;
3436 		bus_dmamap_unload(buf_tag, map);
3437 		DBG_COUNTER_INC(encap_txq_avail_fail);
3438 		DBG_COUNTER_INC(encap_txd_encap_fail);
3439 		if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0)
3440 			GROUPTASK_ENQUEUE(&txq->ift_task);
3441 		return (ENOBUFS);
3442 	}
3443 	/*
3444 	 * On Intel cards we can greatly reduce the number of TX interrupts
3445 	 * we see by only setting report status on every Nth descriptor.
3446 	 * However, this also means that the driver will need to keep track
3447 	 * of the descriptors that RS was set on to check them for the DD bit.
3448 	 */
3449 	txq->ift_rs_pending += nsegs + 1;
3450 	if (txq->ift_rs_pending > TXQ_MAX_RS_DEFERRED(txq) ||
3451 	     iflib_no_tx_batch || (TXQ_AVAIL(txq) - nsegs) <= MAX_TX_DESC(ctx) + 2) {
3452 		pi.ipi_flags |= IPI_TX_INTR;
3453 		txq->ift_rs_pending = 0;
3454 	}
3455 
3456 	pi.ipi_segs = segs;
3457 	pi.ipi_nsegs = nsegs;
3458 
3459 	MPASS(pidx >= 0 && pidx < txq->ift_size);
3460 #ifdef PKT_DEBUG
3461 	print_pkt(&pi);
3462 #endif
3463 	if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) {
3464 		bus_dmamap_sync(buf_tag, map, BUS_DMASYNC_PREWRITE);
3465 		DBG_COUNTER_INC(tx_encap);
3466 		MPASS(pi.ipi_new_pidx < txq->ift_size);
3467 
3468 		ndesc = pi.ipi_new_pidx - pi.ipi_pidx;
3469 		if (pi.ipi_new_pidx < pi.ipi_pidx) {
3470 			ndesc += txq->ift_size;
3471 			txq->ift_gen = 1;
3472 		}
3473 		/*
3474 		 * drivers can need as many as
3475 		 * two sentinels
3476 		 */
3477 		MPASS(ndesc <= pi.ipi_nsegs + 2);
3478 		MPASS(pi.ipi_new_pidx != pidx);
3479 		MPASS(ndesc > 0);
3480 		txq->ift_in_use += ndesc;
3481 
3482 		/*
3483 		 * We update the last software descriptor again here because there may
3484 		 * be a sentinel and/or there may be more mbufs than segments
3485 		 */
3486 		txq->ift_pidx = pi.ipi_new_pidx;
3487 		txq->ift_npending += pi.ipi_ndescs;
3488 	} else {
3489 		*m_headp = m_head = iflib_remove_mbuf(txq);
3490 		if (err == EFBIG) {
3491 			txq->ift_txd_encap_efbig++;
3492 			if (remap < 2) {
3493 				remap = 1;
3494 				goto defrag;
3495 			}
3496 		}
3497 		goto defrag_failed;
3498 	}
3499 	/*
3500 	 * err can't possibly be non-zero here, so we don't neet to test it
3501 	 * to see if we need to DBG_COUNTER_INC(encap_txd_encap_fail).
3502 	 */
3503 	return (err);
3504 
3505 defrag_failed:
3506 	txq->ift_mbuf_defrag_failed++;
3507 	txq->ift_map_failed++;
3508 	m_freem(*m_headp);
3509 	DBG_COUNTER_INC(tx_frees);
3510 	*m_headp = NULL;
3511 	DBG_COUNTER_INC(encap_txd_encap_fail);
3512 	return (ENOMEM);
3513 }
3514 
3515 static void
3516 iflib_tx_desc_free(iflib_txq_t txq, int n)
3517 {
3518 	uint32_t qsize, cidx, mask, gen;
3519 	struct mbuf *m, **ifsd_m;
3520 	bool do_prefetch;
3521 
3522 	cidx = txq->ift_cidx;
3523 	gen = txq->ift_gen;
3524 	qsize = txq->ift_size;
3525 	mask = qsize-1;
3526 	ifsd_m = txq->ift_sds.ifsd_m;
3527 	do_prefetch = (txq->ift_ctx->ifc_flags & IFC_PREFETCH);
3528 
3529 	while (n-- > 0) {
3530 		if (do_prefetch) {
3531 			prefetch(ifsd_m[(cidx + 3) & mask]);
3532 			prefetch(ifsd_m[(cidx + 4) & mask]);
3533 		}
3534 		if ((m = ifsd_m[cidx]) != NULL) {
3535 			prefetch(&ifsd_m[(cidx + CACHE_PTR_INCREMENT) & mask]);
3536 			if (m->m_pkthdr.csum_flags & CSUM_TSO) {
3537 				bus_dmamap_sync(txq->ift_tso_buf_tag,
3538 				    txq->ift_sds.ifsd_tso_map[cidx],
3539 				    BUS_DMASYNC_POSTWRITE);
3540 				bus_dmamap_unload(txq->ift_tso_buf_tag,
3541 				    txq->ift_sds.ifsd_tso_map[cidx]);
3542 			} else {
3543 				bus_dmamap_sync(txq->ift_buf_tag,
3544 				    txq->ift_sds.ifsd_map[cidx],
3545 				    BUS_DMASYNC_POSTWRITE);
3546 				bus_dmamap_unload(txq->ift_buf_tag,
3547 				    txq->ift_sds.ifsd_map[cidx]);
3548 			}
3549 			/* XXX we don't support any drivers that batch packets yet */
3550 			MPASS(m->m_nextpkt == NULL);
3551 			m_freem(m);
3552 			ifsd_m[cidx] = NULL;
3553 #if MEMORY_LOGGING
3554 			txq->ift_dequeued++;
3555 #endif
3556 			DBG_COUNTER_INC(tx_frees);
3557 		}
3558 		if (__predict_false(++cidx == qsize)) {
3559 			cidx = 0;
3560 			gen = 0;
3561 		}
3562 	}
3563 	txq->ift_cidx = cidx;
3564 	txq->ift_gen = gen;
3565 }
3566 
3567 static __inline int
3568 iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh)
3569 {
3570 	int reclaim;
3571 	if_ctx_t ctx = txq->ift_ctx;
3572 
3573 	KASSERT(thresh >= 0, ("invalid threshold to reclaim"));
3574 	MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size);
3575 
3576 	/*
3577 	 * Need a rate-limiting check so that this isn't called every time
3578 	 */
3579 	iflib_tx_credits_update(ctx, txq);
3580 	reclaim = DESC_RECLAIMABLE(txq);
3581 
3582 	if (reclaim <= thresh /* + MAX_TX_DESC(txq->ift_ctx) */) {
3583 #ifdef INVARIANTS
3584 		if (iflib_verbose_debug) {
3585 			printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __FUNCTION__,
3586 			       txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments,
3587 			       reclaim, thresh);
3588 		}
3589 #endif
3590 		return (0);
3591 	}
3592 	iflib_tx_desc_free(txq, reclaim);
3593 	txq->ift_cleaned += reclaim;
3594 	txq->ift_in_use -= reclaim;
3595 
3596 	return (reclaim);
3597 }
3598 
3599 static struct mbuf **
3600 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset, int remaining)
3601 {
3602 	int next, size;
3603 	struct mbuf **items;
3604 
3605 	size = r->size;
3606 	next = (cidx + CACHE_PTR_INCREMENT) & (size-1);
3607 	items = __DEVOLATILE(struct mbuf **, &r->items[0]);
3608 
3609 	prefetch(items[(cidx + offset) & (size-1)]);
3610 	if (remaining > 1) {
3611 		prefetch2cachelines(&items[next]);
3612 		prefetch2cachelines(items[(cidx + offset + 1) & (size-1)]);
3613 		prefetch2cachelines(items[(cidx + offset + 2) & (size-1)]);
3614 		prefetch2cachelines(items[(cidx + offset + 3) & (size-1)]);
3615 	}
3616 	return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (size-1)]));
3617 }
3618 
3619 static void
3620 iflib_txq_check_drain(iflib_txq_t txq, int budget)
3621 {
3622 
3623 	ifmp_ring_check_drainage(txq->ift_br, budget);
3624 }
3625 
3626 static uint32_t
3627 iflib_txq_can_drain(struct ifmp_ring *r)
3628 {
3629 	iflib_txq_t txq = r->cookie;
3630 	if_ctx_t ctx = txq->ift_ctx;
3631 
3632 	if (TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2)
3633 		return (1);
3634 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3635 	    BUS_DMASYNC_POSTREAD);
3636 	return (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id,
3637 	    false));
3638 }
3639 
3640 static uint32_t
3641 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3642 {
3643 	iflib_txq_t txq = r->cookie;
3644 	if_ctx_t ctx = txq->ift_ctx;
3645 	if_t ifp = ctx->ifc_ifp;
3646 	struct mbuf *m, **mp;
3647 	int avail, bytes_sent, consumed, count, err, i, in_use_prev;
3648 	int mcast_sent, pkt_sent, reclaimed, txq_avail;
3649 	bool do_prefetch, rang, ring;
3650 
3651 	if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) ||
3652 			    !LINK_ACTIVE(ctx))) {
3653 		DBG_COUNTER_INC(txq_drain_notready);
3654 		return (0);
3655 	}
3656 	reclaimed = iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
3657 	rang = iflib_txd_db_check(ctx, txq, reclaimed, txq->ift_in_use);
3658 	avail = IDXDIFF(pidx, cidx, r->size);
3659 	if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) {
3660 		DBG_COUNTER_INC(txq_drain_flushing);
3661 		for (i = 0; i < avail; i++) {
3662 			if (__predict_true(r->items[(cidx + i) & (r->size-1)] != (void *)txq))
3663 				m_freem(r->items[(cidx + i) & (r->size-1)]);
3664 			r->items[(cidx + i) & (r->size-1)] = NULL;
3665 		}
3666 		return (avail);
3667 	}
3668 
3669 	if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3670 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3671 		CALLOUT_LOCK(txq);
3672 		callout_stop(&txq->ift_timer);
3673 		CALLOUT_UNLOCK(txq);
3674 		DBG_COUNTER_INC(txq_drain_oactive);
3675 		return (0);
3676 	}
3677 	if (reclaimed)
3678 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3679 	consumed = mcast_sent = bytes_sent = pkt_sent = 0;
3680 	count = MIN(avail, TX_BATCH_SIZE);
3681 #ifdef INVARIANTS
3682 	if (iflib_verbose_debug)
3683 		printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __FUNCTION__,
3684 		       avail, ctx->ifc_flags, TXQ_AVAIL(txq));
3685 #endif
3686 	do_prefetch = (ctx->ifc_flags & IFC_PREFETCH);
3687 	txq_avail = TXQ_AVAIL(txq);
3688 	err = 0;
3689 	for (i = 0; i < count && txq_avail > MAX_TX_DESC(ctx) + 2; i++) {
3690 		int rem = do_prefetch ? count - i : 0;
3691 
3692 		mp = _ring_peek_one(r, cidx, i, rem);
3693 		MPASS(mp != NULL && *mp != NULL);
3694 		if (__predict_false(*mp == (struct mbuf *)txq)) {
3695 			consumed++;
3696 			continue;
3697 		}
3698 		in_use_prev = txq->ift_in_use;
3699 		err = iflib_encap(txq, mp);
3700 		if (__predict_false(err)) {
3701 			/* no room - bail out */
3702 			if (err == ENOBUFS)
3703 				break;
3704 			consumed++;
3705 			/* we can't send this packet - skip it */
3706 			continue;
3707 		}
3708 		consumed++;
3709 		pkt_sent++;
3710 		m = *mp;
3711 		DBG_COUNTER_INC(tx_sent);
3712 		bytes_sent += m->m_pkthdr.len;
3713 		mcast_sent += !!(m->m_flags & M_MCAST);
3714 		txq_avail = TXQ_AVAIL(txq);
3715 
3716 		txq->ift_db_pending += (txq->ift_in_use - in_use_prev);
3717 		ETHER_BPF_MTAP(ifp, m);
3718 		if (__predict_false(!(ifp->if_drv_flags & IFF_DRV_RUNNING)))
3719 			break;
3720 		rang = iflib_txd_db_check(ctx, txq, false, in_use_prev);
3721 	}
3722 
3723 	/* deliberate use of bitwise or to avoid gratuitous short-circuit */
3724 	ring = rang ? false  : (iflib_min_tx_latency | err) || (TXQ_AVAIL(txq) < MAX_TX_DESC(ctx));
3725 	iflib_txd_db_check(ctx, txq, ring, txq->ift_in_use);
3726 	if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
3727 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
3728 	if (mcast_sent)
3729 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
3730 #ifdef INVARIANTS
3731 	if (iflib_verbose_debug)
3732 		printf("consumed=%d\n", consumed);
3733 #endif
3734 	return (consumed);
3735 }
3736 
3737 static uint32_t
3738 iflib_txq_drain_always(struct ifmp_ring *r)
3739 {
3740 	return (1);
3741 }
3742 
3743 static uint32_t
3744 iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3745 {
3746 	int i, avail;
3747 	struct mbuf **mp;
3748 	iflib_txq_t txq;
3749 
3750 	txq = r->cookie;
3751 
3752 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3753 	CALLOUT_LOCK(txq);
3754 	callout_stop(&txq->ift_timer);
3755 	CALLOUT_UNLOCK(txq);
3756 
3757 	avail = IDXDIFF(pidx, cidx, r->size);
3758 	for (i = 0; i < avail; i++) {
3759 		mp = _ring_peek_one(r, cidx, i, avail - i);
3760 		if (__predict_false(*mp == (struct mbuf *)txq))
3761 			continue;
3762 		m_freem(*mp);
3763 		DBG_COUNTER_INC(tx_frees);
3764 	}
3765 	MPASS(ifmp_ring_is_stalled(r) == 0);
3766 	return (avail);
3767 }
3768 
3769 static void
3770 iflib_ifmp_purge(iflib_txq_t txq)
3771 {
3772 	struct ifmp_ring *r;
3773 
3774 	r = txq->ift_br;
3775 	r->drain = iflib_txq_drain_free;
3776 	r->can_drain = iflib_txq_drain_always;
3777 
3778 	ifmp_ring_check_drainage(r, r->size);
3779 
3780 	r->drain = iflib_txq_drain;
3781 	r->can_drain = iflib_txq_can_drain;
3782 }
3783 
3784 static void
3785 _task_fn_tx(void *context)
3786 {
3787 	iflib_txq_t txq = context;
3788 	if_ctx_t ctx = txq->ift_ctx;
3789 	if_t ifp = ctx->ifc_ifp;
3790 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
3791 
3792 #ifdef IFLIB_DIAGNOSTICS
3793 	txq->ift_cpu_exec_count[curcpu]++;
3794 #endif
3795 	if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
3796 		return;
3797 #ifdef DEV_NETMAP
3798 	if ((if_getcapenable(ifp) & IFCAP_NETMAP) &&
3799 	    netmap_tx_irq(ifp, txq->ift_id))
3800 		goto skip_ifmp;
3801 #endif
3802 #ifdef ALTQ
3803 	if (ALTQ_IS_ENABLED(&ifp->if_snd))
3804 		iflib_altq_if_start(ifp);
3805 #endif
3806 	if (txq->ift_db_pending)
3807 		ifmp_ring_enqueue(txq->ift_br, (void **)&txq, 1, TX_BATCH_SIZE, abdicate);
3808 	else if (!abdicate)
3809 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3810 	/*
3811 	 * When abdicating, we always need to check drainage, not just when we don't enqueue
3812 	 */
3813 	if (abdicate)
3814 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3815 #ifdef DEV_NETMAP
3816 skip_ifmp:
3817 #endif
3818 	if (ctx->ifc_flags & IFC_LEGACY)
3819 		IFDI_INTR_ENABLE(ctx);
3820 	else
3821 		IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3822 }
3823 
3824 static void
3825 _task_fn_rx(void *context)
3826 {
3827 	iflib_rxq_t rxq = context;
3828 	if_ctx_t ctx = rxq->ifr_ctx;
3829 	uint8_t more;
3830 	uint16_t budget;
3831 #ifdef DEV_NETMAP
3832 	u_int work = 0;
3833 	int nmirq;
3834 #endif
3835 
3836 #ifdef IFLIB_DIAGNOSTICS
3837 	rxq->ifr_cpu_exec_count[curcpu]++;
3838 #endif
3839 	DBG_COUNTER_INC(task_fn_rxs);
3840 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3841 		return;
3842 #ifdef DEV_NETMAP
3843 	nmirq = netmap_rx_irq(ctx->ifc_ifp, rxq->ifr_id, &work);
3844 	if (nmirq != NM_IRQ_PASS) {
3845 		more = (nmirq == NM_IRQ_RESCHED) ? IFLIB_RXEOF_MORE : 0;
3846 		goto skip_rxeof;
3847 	}
3848 #endif
3849 	budget = ctx->ifc_sysctl_rx_budget;
3850 	if (budget == 0)
3851 		budget = 16;	/* XXX */
3852 	more = iflib_rxeof(rxq, budget);
3853 #ifdef DEV_NETMAP
3854 skip_rxeof:
3855 #endif
3856 	if ((more & IFLIB_RXEOF_MORE) == 0) {
3857 		if (ctx->ifc_flags & IFC_LEGACY)
3858 			IFDI_INTR_ENABLE(ctx);
3859 		else
3860 			IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
3861 		DBG_COUNTER_INC(rx_intr_enables);
3862 	}
3863 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3864 		return;
3865 
3866 	if (more & IFLIB_RXEOF_MORE)
3867 		GROUPTASK_ENQUEUE(&rxq->ifr_task);
3868 	else if (more & IFLIB_RXEOF_EMPTY)
3869 		callout_reset_curcpu(&rxq->ifr_watchdog, 1, &_task_fn_rx_watchdog, rxq);
3870 }
3871 
3872 static void
3873 _task_fn_admin(void *context)
3874 {
3875 	if_ctx_t ctx = context;
3876 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
3877 	iflib_txq_t txq;
3878 	int i;
3879 	bool oactive, running, do_reset, do_watchdog, in_detach;
3880 
3881 	STATE_LOCK(ctx);
3882 	running = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING);
3883 	oactive = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE);
3884 	do_reset = (ctx->ifc_flags & IFC_DO_RESET);
3885 	do_watchdog = (ctx->ifc_flags & IFC_DO_WATCHDOG);
3886 	in_detach = (ctx->ifc_flags & IFC_IN_DETACH);
3887 	ctx->ifc_flags &= ~(IFC_DO_RESET|IFC_DO_WATCHDOG);
3888 	STATE_UNLOCK(ctx);
3889 
3890 	if ((!running && !oactive) && !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
3891 		return;
3892 	if (in_detach)
3893 		return;
3894 
3895 	CTX_LOCK(ctx);
3896 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3897 		CALLOUT_LOCK(txq);
3898 		callout_stop(&txq->ift_timer);
3899 		CALLOUT_UNLOCK(txq);
3900 	}
3901 	if (do_watchdog) {
3902 		ctx->ifc_watchdog_events++;
3903 		IFDI_WATCHDOG_RESET(ctx);
3904 	}
3905 	IFDI_UPDATE_ADMIN_STATUS(ctx);
3906 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3907 		callout_reset_on(&txq->ift_timer, hz / 2, iflib_timer, txq,
3908 		    txq->ift_timer.c_cpu);
3909 	}
3910 	IFDI_LINK_INTR_ENABLE(ctx);
3911 	if (do_reset)
3912 		iflib_if_init_locked(ctx);
3913 	CTX_UNLOCK(ctx);
3914 
3915 	if (LINK_ACTIVE(ctx) == 0)
3916 		return;
3917 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3918 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
3919 }
3920 
3921 static void
3922 _task_fn_iov(void *context)
3923 {
3924 	if_ctx_t ctx = context;
3925 
3926 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING) &&
3927 	    !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
3928 		return;
3929 
3930 	CTX_LOCK(ctx);
3931 	IFDI_VFLR_HANDLE(ctx);
3932 	CTX_UNLOCK(ctx);
3933 }
3934 
3935 static int
3936 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)
3937 {
3938 	int err;
3939 	if_int_delay_info_t info;
3940 	if_ctx_t ctx;
3941 
3942 	info = (if_int_delay_info_t)arg1;
3943 	ctx = info->iidi_ctx;
3944 	info->iidi_req = req;
3945 	info->iidi_oidp = oidp;
3946 	CTX_LOCK(ctx);
3947 	err = IFDI_SYSCTL_INT_DELAY(ctx, info);
3948 	CTX_UNLOCK(ctx);
3949 	return (err);
3950 }
3951 
3952 /*********************************************************************
3953  *
3954  *  IFNET FUNCTIONS
3955  *
3956  **********************************************************************/
3957 
3958 static void
3959 iflib_if_init_locked(if_ctx_t ctx)
3960 {
3961 	iflib_stop(ctx);
3962 	iflib_init_locked(ctx);
3963 }
3964 
3965 static void
3966 iflib_if_init(void *arg)
3967 {
3968 	if_ctx_t ctx = arg;
3969 
3970 	CTX_LOCK(ctx);
3971 	iflib_if_init_locked(ctx);
3972 	CTX_UNLOCK(ctx);
3973 }
3974 
3975 static int
3976 iflib_if_transmit(if_t ifp, struct mbuf *m)
3977 {
3978 	if_ctx_t	ctx = if_getsoftc(ifp);
3979 
3980 	iflib_txq_t txq;
3981 	int err, qidx;
3982 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
3983 
3984 	if (__predict_false((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) {
3985 		DBG_COUNTER_INC(tx_frees);
3986 		m_freem(m);
3987 		return (ENETDOWN);
3988 	}
3989 
3990 	MPASS(m->m_nextpkt == NULL);
3991 	/* ALTQ-enabled interfaces always use queue 0. */
3992 	qidx = 0;
3993 	if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m) && !ALTQ_IS_ENABLED(&ifp->if_snd))
3994 		qidx = QIDX(ctx, m);
3995 	/*
3996 	 * XXX calculate buf_ring based on flowid (divvy up bits?)
3997 	 */
3998 	txq = &ctx->ifc_txqs[qidx];
3999 
4000 #ifdef DRIVER_BACKPRESSURE
4001 	if (txq->ift_closed) {
4002 		while (m != NULL) {
4003 			next = m->m_nextpkt;
4004 			m->m_nextpkt = NULL;
4005 			m_freem(m);
4006 			DBG_COUNTER_INC(tx_frees);
4007 			m = next;
4008 		}
4009 		return (ENOBUFS);
4010 	}
4011 #endif
4012 #ifdef notyet
4013 	qidx = count = 0;
4014 	mp = marr;
4015 	next = m;
4016 	do {
4017 		count++;
4018 		next = next->m_nextpkt;
4019 	} while (next != NULL);
4020 
4021 	if (count > nitems(marr))
4022 		if ((mp = malloc(count*sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) {
4023 			/* XXX check nextpkt */
4024 			m_freem(m);
4025 			/* XXX simplify for now */
4026 			DBG_COUNTER_INC(tx_frees);
4027 			return (ENOBUFS);
4028 		}
4029 	for (next = m, i = 0; next != NULL; i++) {
4030 		mp[i] = next;
4031 		next = next->m_nextpkt;
4032 		mp[i]->m_nextpkt = NULL;
4033 	}
4034 #endif
4035 	DBG_COUNTER_INC(tx_seen);
4036 	err = ifmp_ring_enqueue(txq->ift_br, (void **)&m, 1, TX_BATCH_SIZE, abdicate);
4037 
4038 	if (abdicate)
4039 		GROUPTASK_ENQUEUE(&txq->ift_task);
4040  	if (err) {
4041 		if (!abdicate)
4042 			GROUPTASK_ENQUEUE(&txq->ift_task);
4043 		/* support forthcoming later */
4044 #ifdef DRIVER_BACKPRESSURE
4045 		txq->ift_closed = TRUE;
4046 #endif
4047 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
4048 		m_freem(m);
4049 		DBG_COUNTER_INC(tx_frees);
4050 	}
4051 
4052 	return (err);
4053 }
4054 
4055 #ifdef ALTQ
4056 /*
4057  * The overall approach to integrating iflib with ALTQ is to continue to use
4058  * the iflib mp_ring machinery between the ALTQ queue(s) and the hardware
4059  * ring.  Technically, when using ALTQ, queueing to an intermediate mp_ring
4060  * is redundant/unnecessary, but doing so minimizes the amount of
4061  * ALTQ-specific code required in iflib.  It is assumed that the overhead of
4062  * redundantly queueing to an intermediate mp_ring is swamped by the
4063  * performance limitations inherent in using ALTQ.
4064  *
4065  * When ALTQ support is compiled in, all iflib drivers will use a transmit
4066  * routine, iflib_altq_if_transmit(), that checks if ALTQ is enabled for the
4067  * given interface.  If ALTQ is enabled for an interface, then all
4068  * transmitted packets for that interface will be submitted to the ALTQ
4069  * subsystem via IFQ_ENQUEUE().  We don't use the legacy if_transmit()
4070  * implementation because it uses IFQ_HANDOFF(), which will duplicatively
4071  * update stats that the iflib machinery handles, and which is sensitve to
4072  * the disused IFF_DRV_OACTIVE flag.  Additionally, iflib_altq_if_start()
4073  * will be installed as the start routine for use by ALTQ facilities that
4074  * need to trigger queue drains on a scheduled basis.
4075  *
4076  */
4077 static void
4078 iflib_altq_if_start(if_t ifp)
4079 {
4080 	struct ifaltq *ifq = &ifp->if_snd;
4081 	struct mbuf *m;
4082 
4083 	IFQ_LOCK(ifq);
4084 	IFQ_DEQUEUE_NOLOCK(ifq, m);
4085 	while (m != NULL) {
4086 		iflib_if_transmit(ifp, m);
4087 		IFQ_DEQUEUE_NOLOCK(ifq, m);
4088 	}
4089 	IFQ_UNLOCK(ifq);
4090 }
4091 
4092 static int
4093 iflib_altq_if_transmit(if_t ifp, struct mbuf *m)
4094 {
4095 	int err;
4096 
4097 	if (ALTQ_IS_ENABLED(&ifp->if_snd)) {
4098 		IFQ_ENQUEUE(&ifp->if_snd, m, err);
4099 		if (err == 0)
4100 			iflib_altq_if_start(ifp);
4101 	} else
4102 		err = iflib_if_transmit(ifp, m);
4103 
4104 	return (err);
4105 }
4106 #endif /* ALTQ */
4107 
4108 static void
4109 iflib_if_qflush(if_t ifp)
4110 {
4111 	if_ctx_t ctx = if_getsoftc(ifp);
4112 	iflib_txq_t txq = ctx->ifc_txqs;
4113 	int i;
4114 
4115 	STATE_LOCK(ctx);
4116 	ctx->ifc_flags |= IFC_QFLUSH;
4117 	STATE_UNLOCK(ctx);
4118 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
4119 		while (!(ifmp_ring_is_idle(txq->ift_br) || ifmp_ring_is_stalled(txq->ift_br)))
4120 			iflib_txq_check_drain(txq, 0);
4121 	STATE_LOCK(ctx);
4122 	ctx->ifc_flags &= ~IFC_QFLUSH;
4123 	STATE_UNLOCK(ctx);
4124 
4125 	/*
4126 	 * When ALTQ is enabled, this will also take care of purging the
4127 	 * ALTQ queue(s).
4128 	 */
4129 	if_qflush(ifp);
4130 }
4131 
4132 #define IFCAP_FLAGS (IFCAP_HWCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \
4133 		     IFCAP_TSO | IFCAP_VLAN_HWTAGGING | IFCAP_HWSTATS | \
4134 		     IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | \
4135 		     IFCAP_VLAN_HWTSO | IFCAP_VLAN_HWCSUM | IFCAP_NOMAP)
4136 
4137 static int
4138 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data)
4139 {
4140 	if_ctx_t ctx = if_getsoftc(ifp);
4141 	struct ifreq	*ifr = (struct ifreq *)data;
4142 #if defined(INET) || defined(INET6)
4143 	struct ifaddr	*ifa = (struct ifaddr *)data;
4144 #endif
4145 	bool		avoid_reset = false;
4146 	int		err = 0, reinit = 0, bits;
4147 
4148 	switch (command) {
4149 	case SIOCSIFADDR:
4150 #ifdef INET
4151 		if (ifa->ifa_addr->sa_family == AF_INET)
4152 			avoid_reset = true;
4153 #endif
4154 #ifdef INET6
4155 		if (ifa->ifa_addr->sa_family == AF_INET6)
4156 			avoid_reset = true;
4157 #endif
4158 		/*
4159 		** Calling init results in link renegotiation,
4160 		** so we avoid doing it when possible.
4161 		*/
4162 		if (avoid_reset) {
4163 			if_setflagbits(ifp, IFF_UP,0);
4164 			if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
4165 				reinit = 1;
4166 #ifdef INET
4167 			if (!(if_getflags(ifp) & IFF_NOARP))
4168 				arp_ifinit(ifp, ifa);
4169 #endif
4170 		} else
4171 			err = ether_ioctl(ifp, command, data);
4172 		break;
4173 	case SIOCSIFMTU:
4174 		CTX_LOCK(ctx);
4175 		if (ifr->ifr_mtu == if_getmtu(ifp)) {
4176 			CTX_UNLOCK(ctx);
4177 			break;
4178 		}
4179 		bits = if_getdrvflags(ifp);
4180 		/* stop the driver and free any clusters before proceeding */
4181 		iflib_stop(ctx);
4182 
4183 		if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) {
4184 			STATE_LOCK(ctx);
4185 			if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size)
4186 				ctx->ifc_flags |= IFC_MULTISEG;
4187 			else
4188 				ctx->ifc_flags &= ~IFC_MULTISEG;
4189 			STATE_UNLOCK(ctx);
4190 			err = if_setmtu(ifp, ifr->ifr_mtu);
4191 		}
4192 		iflib_init_locked(ctx);
4193 		STATE_LOCK(ctx);
4194 		if_setdrvflags(ifp, bits);
4195 		STATE_UNLOCK(ctx);
4196 		CTX_UNLOCK(ctx);
4197 		break;
4198 	case SIOCSIFFLAGS:
4199 		CTX_LOCK(ctx);
4200 		if (if_getflags(ifp) & IFF_UP) {
4201 			if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4202 				if ((if_getflags(ifp) ^ ctx->ifc_if_flags) &
4203 				    (IFF_PROMISC | IFF_ALLMULTI)) {
4204 					CTX_UNLOCK(ctx);
4205 					err = IFDI_PROMISC_SET(ctx, if_getflags(ifp));
4206 					CTX_LOCK(ctx);
4207 				}
4208 			} else
4209 				reinit = 1;
4210 		} else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4211 			iflib_stop(ctx);
4212 		}
4213 		ctx->ifc_if_flags = if_getflags(ifp);
4214 		CTX_UNLOCK(ctx);
4215 		break;
4216 	case SIOCADDMULTI:
4217 	case SIOCDELMULTI:
4218 		if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4219 			CTX_LOCK(ctx);
4220 			IFDI_INTR_DISABLE(ctx);
4221 			IFDI_MULTI_SET(ctx);
4222 			IFDI_INTR_ENABLE(ctx);
4223 			CTX_UNLOCK(ctx);
4224 		}
4225 		break;
4226 	case SIOCSIFMEDIA:
4227 		CTX_LOCK(ctx);
4228 		IFDI_MEDIA_SET(ctx);
4229 		CTX_UNLOCK(ctx);
4230 		/* FALLTHROUGH */
4231 	case SIOCGIFMEDIA:
4232 	case SIOCGIFXMEDIA:
4233 		err = ifmedia_ioctl(ifp, ifr, ctx->ifc_mediap, command);
4234 		break;
4235 	case SIOCGI2C:
4236 	{
4237 		struct ifi2creq i2c;
4238 
4239 		err = copyin(ifr_data_get_ptr(ifr), &i2c, sizeof(i2c));
4240 		if (err != 0)
4241 			break;
4242 		if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) {
4243 			err = EINVAL;
4244 			break;
4245 		}
4246 		if (i2c.len > sizeof(i2c.data)) {
4247 			err = EINVAL;
4248 			break;
4249 		}
4250 
4251 		if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0)
4252 			err = copyout(&i2c, ifr_data_get_ptr(ifr),
4253 			    sizeof(i2c));
4254 		break;
4255 	}
4256 	case SIOCSIFCAP:
4257 	{
4258 		int mask, setmask, oldmask;
4259 
4260 		oldmask = if_getcapenable(ifp);
4261 		mask = ifr->ifr_reqcap ^ oldmask;
4262 		mask &= ctx->ifc_softc_ctx.isc_capabilities | IFCAP_NOMAP;
4263 		setmask = 0;
4264 #ifdef TCP_OFFLOAD
4265 		setmask |= mask & (IFCAP_TOE4|IFCAP_TOE6);
4266 #endif
4267 		setmask |= (mask & IFCAP_FLAGS);
4268 		setmask |= (mask & IFCAP_WOL);
4269 
4270 		/*
4271 		 * If any RX csum has changed, change all the ones that
4272 		 * are supported by the driver.
4273 		 */
4274 		if (setmask & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6)) {
4275 			setmask |= ctx->ifc_softc_ctx.isc_capabilities &
4276 			    (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6);
4277 		}
4278 
4279 		/*
4280 		 * want to ensure that traffic has stopped before we change any of the flags
4281 		 */
4282 		if (setmask) {
4283 			CTX_LOCK(ctx);
4284 			bits = if_getdrvflags(ifp);
4285 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4286 				iflib_stop(ctx);
4287 			STATE_LOCK(ctx);
4288 			if_togglecapenable(ifp, setmask);
4289 			STATE_UNLOCK(ctx);
4290 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4291 				iflib_init_locked(ctx);
4292 			STATE_LOCK(ctx);
4293 			if_setdrvflags(ifp, bits);
4294 			STATE_UNLOCK(ctx);
4295 			CTX_UNLOCK(ctx);
4296 		}
4297 		if_vlancap(ifp);
4298 		break;
4299 	}
4300 	case SIOCGPRIVATE_0:
4301 	case SIOCSDRVSPEC:
4302 	case SIOCGDRVSPEC:
4303 		CTX_LOCK(ctx);
4304 		err = IFDI_PRIV_IOCTL(ctx, command, data);
4305 		CTX_UNLOCK(ctx);
4306 		break;
4307 	default:
4308 		err = ether_ioctl(ifp, command, data);
4309 		break;
4310 	}
4311 	if (reinit)
4312 		iflib_if_init(ctx);
4313 	return (err);
4314 }
4315 
4316 static uint64_t
4317 iflib_if_get_counter(if_t ifp, ift_counter cnt)
4318 {
4319 	if_ctx_t ctx = if_getsoftc(ifp);
4320 
4321 	return (IFDI_GET_COUNTER(ctx, cnt));
4322 }
4323 
4324 /*********************************************************************
4325  *
4326  *  OTHER FUNCTIONS EXPORTED TO THE STACK
4327  *
4328  **********************************************************************/
4329 
4330 static void
4331 iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag)
4332 {
4333 	if_ctx_t ctx = if_getsoftc(ifp);
4334 
4335 	if ((void *)ctx != arg)
4336 		return;
4337 
4338 	if ((vtag == 0) || (vtag > 4095))
4339 		return;
4340 
4341 	if (iflib_in_detach(ctx))
4342 		return;
4343 
4344 	CTX_LOCK(ctx);
4345 	/* Driver may need all untagged packets to be flushed */
4346 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4347 		iflib_stop(ctx);
4348 	IFDI_VLAN_REGISTER(ctx, vtag);
4349 	/* Re-init to load the changes, if required */
4350 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4351 		iflib_init_locked(ctx);
4352 	CTX_UNLOCK(ctx);
4353 }
4354 
4355 static void
4356 iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag)
4357 {
4358 	if_ctx_t ctx = if_getsoftc(ifp);
4359 
4360 	if ((void *)ctx != arg)
4361 		return;
4362 
4363 	if ((vtag == 0) || (vtag > 4095))
4364 		return;
4365 
4366 	CTX_LOCK(ctx);
4367 	/* Driver may need all tagged packets to be flushed */
4368 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4369 		iflib_stop(ctx);
4370 	IFDI_VLAN_UNREGISTER(ctx, vtag);
4371 	/* Re-init to load the changes, if required */
4372 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4373 		iflib_init_locked(ctx);
4374 	CTX_UNLOCK(ctx);
4375 }
4376 
4377 static void
4378 iflib_led_func(void *arg, int onoff)
4379 {
4380 	if_ctx_t ctx = arg;
4381 
4382 	CTX_LOCK(ctx);
4383 	IFDI_LED_FUNC(ctx, onoff);
4384 	CTX_UNLOCK(ctx);
4385 }
4386 
4387 /*********************************************************************
4388  *
4389  *  BUS FUNCTION DEFINITIONS
4390  *
4391  **********************************************************************/
4392 
4393 int
4394 iflib_device_probe(device_t dev)
4395 {
4396 	const pci_vendor_info_t *ent;
4397 	if_shared_ctx_t sctx;
4398 	uint16_t pci_device_id, pci_rev_id, pci_subdevice_id, pci_subvendor_id;
4399 	uint16_t pci_vendor_id;
4400 
4401 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4402 		return (ENOTSUP);
4403 
4404 	pci_vendor_id = pci_get_vendor(dev);
4405 	pci_device_id = pci_get_device(dev);
4406 	pci_subvendor_id = pci_get_subvendor(dev);
4407 	pci_subdevice_id = pci_get_subdevice(dev);
4408 	pci_rev_id = pci_get_revid(dev);
4409 	if (sctx->isc_parse_devinfo != NULL)
4410 		sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id);
4411 
4412 	ent = sctx->isc_vendor_info;
4413 	while (ent->pvi_vendor_id != 0) {
4414 		if (pci_vendor_id != ent->pvi_vendor_id) {
4415 			ent++;
4416 			continue;
4417 		}
4418 		if ((pci_device_id == ent->pvi_device_id) &&
4419 		    ((pci_subvendor_id == ent->pvi_subvendor_id) ||
4420 		     (ent->pvi_subvendor_id == 0)) &&
4421 		    ((pci_subdevice_id == ent->pvi_subdevice_id) ||
4422 		     (ent->pvi_subdevice_id == 0)) &&
4423 		    ((pci_rev_id == ent->pvi_rev_id) ||
4424 		     (ent->pvi_rev_id == 0))) {
4425 			device_set_desc_copy(dev, ent->pvi_name);
4426 			/* this needs to be changed to zero if the bus probing code
4427 			 * ever stops re-probing on best match because the sctx
4428 			 * may have its values over written by register calls
4429 			 * in subsequent probes
4430 			 */
4431 			return (BUS_PROBE_DEFAULT);
4432 		}
4433 		ent++;
4434 	}
4435 	return (ENXIO);
4436 }
4437 
4438 int
4439 iflib_device_probe_vendor(device_t dev)
4440 {
4441 	int probe;
4442 
4443 	probe = iflib_device_probe(dev);
4444 	if (probe == BUS_PROBE_DEFAULT)
4445 		return (BUS_PROBE_VENDOR);
4446 	else
4447 		return (probe);
4448 }
4449 
4450 static void
4451 iflib_reset_qvalues(if_ctx_t ctx)
4452 {
4453 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4454 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4455 	device_t dev = ctx->ifc_dev;
4456 	int i;
4457 
4458 	if (ctx->ifc_sysctl_ntxqs != 0)
4459 		scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs;
4460 	if (ctx->ifc_sysctl_nrxqs != 0)
4461 		scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs;
4462 
4463 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4464 		if (ctx->ifc_sysctl_ntxds[i] != 0)
4465 			scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i];
4466 		else
4467 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4468 	}
4469 
4470 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4471 		if (ctx->ifc_sysctl_nrxds[i] != 0)
4472 			scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i];
4473 		else
4474 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4475 	}
4476 
4477 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4478 		if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) {
4479 			device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n",
4480 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]);
4481 			scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i];
4482 		}
4483 		if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) {
4484 			device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n",
4485 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]);
4486 			scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i];
4487 		}
4488 		if (!powerof2(scctx->isc_nrxd[i])) {
4489 			device_printf(dev, "nrxd%d: %d is not a power of 2 - using default value of %d\n",
4490 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_default[i]);
4491 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4492 		}
4493 	}
4494 
4495 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4496 		if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) {
4497 			device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n",
4498 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]);
4499 			scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i];
4500 		}
4501 		if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) {
4502 			device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n",
4503 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]);
4504 			scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i];
4505 		}
4506 		if (!powerof2(scctx->isc_ntxd[i])) {
4507 			device_printf(dev, "ntxd%d: %d is not a power of 2 - using default value of %d\n",
4508 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_default[i]);
4509 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4510 		}
4511 	}
4512 }
4513 
4514 static void
4515 iflib_add_pfil(if_ctx_t ctx)
4516 {
4517 	struct pfil_head *pfil;
4518 	struct pfil_head_args pa;
4519 	iflib_rxq_t rxq;
4520 	int i;
4521 
4522 	pa.pa_version = PFIL_VERSION;
4523 	pa.pa_flags = PFIL_IN;
4524 	pa.pa_type = PFIL_TYPE_ETHERNET;
4525 	pa.pa_headname = ctx->ifc_ifp->if_xname;
4526 	pfil = pfil_head_register(&pa);
4527 
4528 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4529 		rxq->pfil = pfil;
4530 	}
4531 }
4532 
4533 static void
4534 iflib_rem_pfil(if_ctx_t ctx)
4535 {
4536 	struct pfil_head *pfil;
4537 	iflib_rxq_t rxq;
4538 	int i;
4539 
4540 	rxq = ctx->ifc_rxqs;
4541 	pfil = rxq->pfil;
4542 	for (i = 0; i < NRXQSETS(ctx); i++, rxq++) {
4543 		rxq->pfil = NULL;
4544 	}
4545 	pfil_head_unregister(pfil);
4546 }
4547 
4548 static uint16_t
4549 get_ctx_core_offset(if_ctx_t ctx)
4550 {
4551 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4552 	struct cpu_offset *op;
4553 	uint16_t qc;
4554 	uint16_t ret = ctx->ifc_sysctl_core_offset;
4555 
4556 	if (ret != CORE_OFFSET_UNSPECIFIED)
4557 		return (ret);
4558 
4559 	if (ctx->ifc_sysctl_separate_txrx)
4560 		qc = scctx->isc_ntxqsets + scctx->isc_nrxqsets;
4561 	else
4562 		qc = max(scctx->isc_ntxqsets, scctx->isc_nrxqsets);
4563 
4564 	mtx_lock(&cpu_offset_mtx);
4565 	SLIST_FOREACH(op, &cpu_offsets, entries) {
4566 		if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) {
4567 			ret = op->offset;
4568 			op->offset += qc;
4569 			MPASS(op->refcount < UINT_MAX);
4570 			op->refcount++;
4571 			break;
4572 		}
4573 	}
4574 	if (ret == CORE_OFFSET_UNSPECIFIED) {
4575 		ret = 0;
4576 		op = malloc(sizeof(struct cpu_offset), M_IFLIB,
4577 		    M_NOWAIT | M_ZERO);
4578 		if (op == NULL) {
4579 			device_printf(ctx->ifc_dev,
4580 			    "allocation for cpu offset failed.\n");
4581 		} else {
4582 			op->offset = qc;
4583 			op->refcount = 1;
4584 			CPU_COPY(&ctx->ifc_cpus, &op->set);
4585 			SLIST_INSERT_HEAD(&cpu_offsets, op, entries);
4586 		}
4587 	}
4588 	mtx_unlock(&cpu_offset_mtx);
4589 
4590 	return (ret);
4591 }
4592 
4593 static void
4594 unref_ctx_core_offset(if_ctx_t ctx)
4595 {
4596 	struct cpu_offset *op, *top;
4597 
4598 	mtx_lock(&cpu_offset_mtx);
4599 	SLIST_FOREACH_SAFE(op, &cpu_offsets, entries, top) {
4600 		if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) {
4601 			MPASS(op->refcount > 0);
4602 			op->refcount--;
4603 			if (op->refcount == 0) {
4604 				SLIST_REMOVE(&cpu_offsets, op, cpu_offset, entries);
4605 				free(op, M_IFLIB);
4606 			}
4607 			break;
4608 		}
4609 	}
4610 	mtx_unlock(&cpu_offset_mtx);
4611 }
4612 
4613 int
4614 iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp)
4615 {
4616 	if_ctx_t ctx;
4617 	if_t ifp;
4618 	if_softc_ctx_t scctx;
4619 	kobjop_desc_t kobj_desc;
4620 	kobj_method_t *kobj_method;
4621 	int err, msix, rid;
4622 	int num_txd, num_rxd;
4623 
4624 	ctx = malloc(sizeof(* ctx), M_IFLIB, M_WAITOK|M_ZERO);
4625 
4626 	if (sc == NULL) {
4627 		sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4628 		device_set_softc(dev, ctx);
4629 		ctx->ifc_flags |= IFC_SC_ALLOCATED;
4630 	}
4631 
4632 	ctx->ifc_sctx = sctx;
4633 	ctx->ifc_dev = dev;
4634 	ctx->ifc_softc = sc;
4635 
4636 	if ((err = iflib_register(ctx)) != 0) {
4637 		device_printf(dev, "iflib_register failed %d\n", err);
4638 		goto fail_ctx_free;
4639 	}
4640 	iflib_add_device_sysctl_pre(ctx);
4641 
4642 	scctx = &ctx->ifc_softc_ctx;
4643 	ifp = ctx->ifc_ifp;
4644 
4645 	iflib_reset_qvalues(ctx);
4646 	CTX_LOCK(ctx);
4647 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4648 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4649 		goto fail_unlock;
4650 	}
4651 	_iflib_pre_assert(scctx);
4652 	ctx->ifc_txrx = *scctx->isc_txrx;
4653 
4654 	if (sctx->isc_flags & IFLIB_DRIVER_MEDIA)
4655 		ctx->ifc_mediap = scctx->isc_media;
4656 
4657 #ifdef INVARIANTS
4658 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
4659 		MPASS(scctx->isc_tx_csum_flags);
4660 #endif
4661 
4662 	if_setcapabilities(ifp,
4663 	    scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_NOMAP);
4664 	if_setcapenable(ifp,
4665 	    scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_NOMAP);
4666 
4667 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4668 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4669 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4670 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4671 
4672 	num_txd = iflib_num_tx_descs(ctx);
4673 	num_rxd = iflib_num_rx_descs(ctx);
4674 
4675 	/* XXX change for per-queue sizes */
4676 	device_printf(dev, "Using %d TX descriptors and %d RX descriptors\n",
4677 	    num_txd, num_rxd);
4678 
4679 	if (scctx->isc_tx_nsegments > num_txd / MAX_SINGLE_PACKET_FRACTION)
4680 		scctx->isc_tx_nsegments = max(1, num_txd /
4681 		    MAX_SINGLE_PACKET_FRACTION);
4682 	if (scctx->isc_tx_tso_segments_max > num_txd /
4683 	    MAX_SINGLE_PACKET_FRACTION)
4684 		scctx->isc_tx_tso_segments_max = max(1,
4685 		    num_txd / MAX_SINGLE_PACKET_FRACTION);
4686 
4687 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4688 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
4689 		/*
4690 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
4691 		 * but some MACs do.
4692 		 */
4693 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
4694 		    IP_MAXPACKET));
4695 		/*
4696 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
4697 		 * into account.  In the worst case, each of these calls will
4698 		 * add another mbuf and, thus, the requirement for another DMA
4699 		 * segment.  So for best performance, it doesn't make sense to
4700 		 * advertize a maximum of TSO segments that typically will
4701 		 * require defragmentation in iflib_encap().
4702 		 */
4703 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
4704 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
4705 	}
4706 	if (scctx->isc_rss_table_size == 0)
4707 		scctx->isc_rss_table_size = 64;
4708 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4709 
4710 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4711 	/* XXX format name */
4712 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx,
4713 	    NULL, NULL, "admin");
4714 
4715 	/* Set up cpu set.  If it fails, use the set of all CPUs. */
4716 	if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) != 0) {
4717 		device_printf(dev, "Unable to fetch CPU list\n");
4718 		CPU_COPY(&all_cpus, &ctx->ifc_cpus);
4719 	}
4720 	MPASS(CPU_COUNT(&ctx->ifc_cpus) > 0);
4721 
4722 	/*
4723 	** Now set up MSI or MSI-X, should return us the number of supported
4724 	** vectors (will be 1 for a legacy interrupt and MSI).
4725 	*/
4726 	if (sctx->isc_flags & IFLIB_SKIP_MSIX) {
4727 		msix = scctx->isc_vectors;
4728 	} else if (scctx->isc_msix_bar != 0)
4729 	       /*
4730 		* The simple fact that isc_msix_bar is not 0 does not mean we
4731 		* we have a good value there that is known to work.
4732 		*/
4733 		msix = iflib_msix_init(ctx);
4734 	else {
4735 		scctx->isc_vectors = 1;
4736 		scctx->isc_ntxqsets = 1;
4737 		scctx->isc_nrxqsets = 1;
4738 		scctx->isc_intr = IFLIB_INTR_LEGACY;
4739 		msix = 0;
4740 	}
4741 	/* Get memory for the station queues */
4742 	if ((err = iflib_queues_alloc(ctx))) {
4743 		device_printf(dev, "Unable to allocate queue memory\n");
4744 		goto fail_intr_free;
4745 	}
4746 
4747 	if ((err = iflib_qset_structures_setup(ctx)))
4748 		goto fail_queues;
4749 
4750 	/*
4751 	 * Now that we know how many queues there are, get the core offset.
4752 	 */
4753 	ctx->ifc_sysctl_core_offset = get_ctx_core_offset(ctx);
4754 
4755 	if (msix > 1) {
4756 		/*
4757 		 * When using MSI-X, ensure that ifdi_{r,t}x_queue_intr_enable
4758 		 * aren't the default NULL implementation.
4759 		 */
4760 		kobj_desc = &ifdi_rx_queue_intr_enable_desc;
4761 		kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL,
4762 		    kobj_desc);
4763 		if (kobj_method == &kobj_desc->deflt) {
4764 			device_printf(dev,
4765 			    "MSI-X requires ifdi_rx_queue_intr_enable method");
4766 			err = EOPNOTSUPP;
4767 			goto fail_queues;
4768 		}
4769 		kobj_desc = &ifdi_tx_queue_intr_enable_desc;
4770 		kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL,
4771 		    kobj_desc);
4772 		if (kobj_method == &kobj_desc->deflt) {
4773 			device_printf(dev,
4774 			    "MSI-X requires ifdi_tx_queue_intr_enable method");
4775 			err = EOPNOTSUPP;
4776 			goto fail_queues;
4777 		}
4778 
4779 		/*
4780 		 * Assign the MSI-X vectors.
4781 		 * Note that the default NULL ifdi_msix_intr_assign method will
4782 		 * fail here, too.
4783 		 */
4784 		err = IFDI_MSIX_INTR_ASSIGN(ctx, msix);
4785 		if (err != 0) {
4786 			device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n",
4787 			    err);
4788 			goto fail_queues;
4789 		}
4790 	} else if (scctx->isc_intr != IFLIB_INTR_MSIX) {
4791 		rid = 0;
4792 		if (scctx->isc_intr == IFLIB_INTR_MSI) {
4793 			MPASS(msix == 1);
4794 			rid = 1;
4795 		}
4796 		if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) {
4797 			device_printf(dev, "iflib_legacy_setup failed %d\n", err);
4798 			goto fail_queues;
4799 		}
4800 	} else {
4801 		device_printf(dev,
4802 		    "Cannot use iflib with only 1 MSI-X interrupt!\n");
4803 		err = ENODEV;
4804 		goto fail_intr_free;
4805 	}
4806 
4807 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
4808 
4809 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4810 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4811 		goto fail_detach;
4812 	}
4813 
4814 	/*
4815 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4816 	 * This must appear after the call to ether_ifattach() because
4817 	 * ether_ifattach() sets if_hdrlen to the default value.
4818 	 */
4819 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4820 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
4821 
4822 	if ((err = iflib_netmap_attach(ctx))) {
4823 		device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err);
4824 		goto fail_detach;
4825 	}
4826 	*ctxp = ctx;
4827 
4828 	DEBUGNET_SET(ctx->ifc_ifp, iflib);
4829 
4830 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4831 	iflib_add_device_sysctl_post(ctx);
4832 	iflib_add_pfil(ctx);
4833 	ctx->ifc_flags |= IFC_INIT_DONE;
4834 	CTX_UNLOCK(ctx);
4835 
4836 	return (0);
4837 
4838 fail_detach:
4839 	ether_ifdetach(ctx->ifc_ifp);
4840 fail_intr_free:
4841 	iflib_free_intr_mem(ctx);
4842 fail_queues:
4843 	iflib_tx_structures_free(ctx);
4844 	iflib_rx_structures_free(ctx);
4845 	taskqgroup_detach(qgroup_if_config_tqg, &ctx->ifc_admin_task);
4846 	IFDI_DETACH(ctx);
4847 fail_unlock:
4848 	CTX_UNLOCK(ctx);
4849 	iflib_deregister(ctx);
4850 fail_ctx_free:
4851 	device_set_softc(ctx->ifc_dev, NULL);
4852         if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4853                 free(ctx->ifc_softc, M_IFLIB);
4854         free(ctx, M_IFLIB);
4855 	return (err);
4856 }
4857 
4858 int
4859 iflib_pseudo_register(device_t dev, if_shared_ctx_t sctx, if_ctx_t *ctxp,
4860 					  struct iflib_cloneattach_ctx *clctx)
4861 {
4862 	int num_txd, num_rxd;
4863 	int err;
4864 	if_ctx_t ctx;
4865 	if_t ifp;
4866 	if_softc_ctx_t scctx;
4867 	int i;
4868 	void *sc;
4869 
4870 	ctx = malloc(sizeof(*ctx), M_IFLIB, M_WAITOK|M_ZERO);
4871 	sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4872 	ctx->ifc_flags |= IFC_SC_ALLOCATED;
4873 	if (sctx->isc_flags & (IFLIB_PSEUDO|IFLIB_VIRTUAL))
4874 		ctx->ifc_flags |= IFC_PSEUDO;
4875 
4876 	ctx->ifc_sctx = sctx;
4877 	ctx->ifc_softc = sc;
4878 	ctx->ifc_dev = dev;
4879 
4880 	if ((err = iflib_register(ctx)) != 0) {
4881 		device_printf(dev, "%s: iflib_register failed %d\n", __func__, err);
4882 		goto fail_ctx_free;
4883 	}
4884 	iflib_add_device_sysctl_pre(ctx);
4885 
4886 	scctx = &ctx->ifc_softc_ctx;
4887 	ifp = ctx->ifc_ifp;
4888 
4889 	iflib_reset_qvalues(ctx);
4890 	CTX_LOCK(ctx);
4891 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4892 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4893 		goto fail_unlock;
4894 	}
4895 	if (sctx->isc_flags & IFLIB_GEN_MAC)
4896 		ether_gen_addr(ifp, &ctx->ifc_mac);
4897 	if ((err = IFDI_CLONEATTACH(ctx, clctx->cc_ifc, clctx->cc_name,
4898 								clctx->cc_params)) != 0) {
4899 		device_printf(dev, "IFDI_CLONEATTACH failed %d\n", err);
4900 		goto fail_unlock;
4901 	}
4902 #ifdef INVARIANTS
4903 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
4904 		MPASS(scctx->isc_tx_csum_flags);
4905 #endif
4906 
4907 	if_setcapabilities(ifp, scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_LINKSTATE);
4908 	if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_LINKSTATE);
4909 
4910 	ifp->if_flags |= IFF_NOGROUP;
4911 	if (sctx->isc_flags & IFLIB_PSEUDO) {
4912 		ifmedia_add(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO, 0, NULL);
4913 		ifmedia_set(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO);
4914 		if (sctx->isc_flags & IFLIB_PSEUDO_ETHER) {
4915 			ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
4916 		} else {
4917 			if_attach(ctx->ifc_ifp);
4918 			bpfattach(ctx->ifc_ifp, DLT_NULL, sizeof(u_int32_t));
4919 		}
4920 
4921 		if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4922 			device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4923 			goto fail_detach;
4924 		}
4925 		*ctxp = ctx;
4926 
4927 		/*
4928 		 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4929 		 * This must appear after the call to ether_ifattach() because
4930 		 * ether_ifattach() sets if_hdrlen to the default value.
4931 		 */
4932 		if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4933 			if_setifheaderlen(ifp,
4934 			    sizeof(struct ether_vlan_header));
4935 
4936 		if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4937 		iflib_add_device_sysctl_post(ctx);
4938 		ctx->ifc_flags |= IFC_INIT_DONE;
4939 		CTX_UNLOCK(ctx);
4940 		return (0);
4941 	}
4942 	ifmedia_add(ctx->ifc_mediap, IFM_ETHER | IFM_1000_T | IFM_FDX, 0, NULL);
4943 	ifmedia_add(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO, 0, NULL);
4944 	ifmedia_set(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO);
4945 
4946 	_iflib_pre_assert(scctx);
4947 	ctx->ifc_txrx = *scctx->isc_txrx;
4948 
4949 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4950 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4951 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4952 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4953 
4954 	num_txd = iflib_num_tx_descs(ctx);
4955 	num_rxd = iflib_num_rx_descs(ctx);
4956 
4957 	/* XXX change for per-queue sizes */
4958 	device_printf(dev, "Using %d TX descriptors and %d RX descriptors\n",
4959 	    num_txd, num_rxd);
4960 
4961 	if (scctx->isc_tx_nsegments > num_txd / MAX_SINGLE_PACKET_FRACTION)
4962 		scctx->isc_tx_nsegments = max(1, num_txd /
4963 		    MAX_SINGLE_PACKET_FRACTION);
4964 	if (scctx->isc_tx_tso_segments_max > num_txd /
4965 	    MAX_SINGLE_PACKET_FRACTION)
4966 		scctx->isc_tx_tso_segments_max = max(1,
4967 		    num_txd / MAX_SINGLE_PACKET_FRACTION);
4968 
4969 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4970 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
4971 		/*
4972 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
4973 		 * but some MACs do.
4974 		 */
4975 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
4976 		    IP_MAXPACKET));
4977 		/*
4978 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
4979 		 * into account.  In the worst case, each of these calls will
4980 		 * add another mbuf and, thus, the requirement for another DMA
4981 		 * segment.  So for best performance, it doesn't make sense to
4982 		 * advertize a maximum of TSO segments that typically will
4983 		 * require defragmentation in iflib_encap().
4984 		 */
4985 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
4986 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
4987 	}
4988 	if (scctx->isc_rss_table_size == 0)
4989 		scctx->isc_rss_table_size = 64;
4990 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4991 
4992 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4993 	/* XXX format name */
4994 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx,
4995 	    NULL, NULL, "admin");
4996 
4997 	/* XXX --- can support > 1 -- but keep it simple for now */
4998 	scctx->isc_intr = IFLIB_INTR_LEGACY;
4999 
5000 	/* Get memory for the station queues */
5001 	if ((err = iflib_queues_alloc(ctx))) {
5002 		device_printf(dev, "Unable to allocate queue memory\n");
5003 		goto fail_iflib_detach;
5004 	}
5005 
5006 	if ((err = iflib_qset_structures_setup(ctx))) {
5007 		device_printf(dev, "qset structure setup failed %d\n", err);
5008 		goto fail_queues;
5009 	}
5010 
5011 	/*
5012 	 * XXX What if anything do we want to do about interrupts?
5013 	 */
5014 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
5015 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
5016 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
5017 		goto fail_detach;
5018 	}
5019 
5020 	/*
5021 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
5022 	 * This must appear after the call to ether_ifattach() because
5023 	 * ether_ifattach() sets if_hdrlen to the default value.
5024 	 */
5025 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
5026 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
5027 
5028 	/* XXX handle more than one queue */
5029 	for (i = 0; i < scctx->isc_nrxqsets; i++)
5030 		IFDI_RX_CLSET(ctx, 0, i, ctx->ifc_rxqs[i].ifr_fl[0].ifl_sds.ifsd_cl);
5031 
5032 	*ctxp = ctx;
5033 
5034 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
5035 	iflib_add_device_sysctl_post(ctx);
5036 	ctx->ifc_flags |= IFC_INIT_DONE;
5037 	CTX_UNLOCK(ctx);
5038 
5039 	return (0);
5040 fail_detach:
5041 	ether_ifdetach(ctx->ifc_ifp);
5042 fail_queues:
5043 	iflib_tx_structures_free(ctx);
5044 	iflib_rx_structures_free(ctx);
5045 fail_iflib_detach:
5046 	IFDI_DETACH(ctx);
5047 fail_unlock:
5048 	CTX_UNLOCK(ctx);
5049 	iflib_deregister(ctx);
5050 fail_ctx_free:
5051 	free(ctx->ifc_softc, M_IFLIB);
5052 	free(ctx, M_IFLIB);
5053 	return (err);
5054 }
5055 
5056 int
5057 iflib_pseudo_deregister(if_ctx_t ctx)
5058 {
5059 	if_t ifp = ctx->ifc_ifp;
5060 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5061 	iflib_txq_t txq;
5062 	iflib_rxq_t rxq;
5063 	int i, j;
5064 	struct taskqgroup *tqg;
5065 	iflib_fl_t fl;
5066 
5067 	/* Unregister VLAN event handlers early */
5068 	iflib_unregister_vlan_handlers(ctx);
5069 
5070 	if ((sctx->isc_flags & IFLIB_PSEUDO)  &&
5071 		(sctx->isc_flags & IFLIB_PSEUDO_ETHER) == 0) {
5072 		bpfdetach(ifp);
5073 		if_detach(ifp);
5074 	} else {
5075 		ether_ifdetach(ifp);
5076 	}
5077 	/* XXX drain any dependent tasks */
5078 	tqg = qgroup_if_io_tqg;
5079 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
5080 		callout_drain(&txq->ift_timer);
5081 #ifdef DEV_NETMAP
5082 		callout_drain(&txq->ift_netmap_timer);
5083 #endif /* DEV_NETMAP */
5084 		if (txq->ift_task.gt_uniq != NULL)
5085 			taskqgroup_detach(tqg, &txq->ift_task);
5086 	}
5087 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
5088 		callout_drain(&rxq->ifr_watchdog);
5089 		if (rxq->ifr_task.gt_uniq != NULL)
5090 			taskqgroup_detach(tqg, &rxq->ifr_task);
5091 
5092 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5093 			free(fl->ifl_rx_bitmap, M_IFLIB);
5094 	}
5095 	tqg = qgroup_if_config_tqg;
5096 	if (ctx->ifc_admin_task.gt_uniq != NULL)
5097 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
5098 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
5099 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
5100 
5101 	iflib_tx_structures_free(ctx);
5102 	iflib_rx_structures_free(ctx);
5103 
5104 	iflib_deregister(ctx);
5105 
5106 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5107 		free(ctx->ifc_softc, M_IFLIB);
5108 	free(ctx, M_IFLIB);
5109 	return (0);
5110 }
5111 
5112 int
5113 iflib_device_attach(device_t dev)
5114 {
5115 	if_ctx_t ctx;
5116 	if_shared_ctx_t sctx;
5117 
5118 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
5119 		return (ENOTSUP);
5120 
5121 	pci_enable_busmaster(dev);
5122 
5123 	return (iflib_device_register(dev, NULL, sctx, &ctx));
5124 }
5125 
5126 int
5127 iflib_device_deregister(if_ctx_t ctx)
5128 {
5129 	if_t ifp = ctx->ifc_ifp;
5130 	iflib_txq_t txq;
5131 	iflib_rxq_t rxq;
5132 	device_t dev = ctx->ifc_dev;
5133 	int i, j;
5134 	struct taskqgroup *tqg;
5135 	iflib_fl_t fl;
5136 
5137 	/* Make sure VLANS are not using driver */
5138 	if (if_vlantrunkinuse(ifp)) {
5139 		device_printf(dev, "Vlan in use, detach first\n");
5140 		return (EBUSY);
5141 	}
5142 #ifdef PCI_IOV
5143 	if (!CTX_IS_VF(ctx) && pci_iov_detach(dev) != 0) {
5144 		device_printf(dev, "SR-IOV in use; detach first.\n");
5145 		return (EBUSY);
5146 	}
5147 #endif
5148 
5149 	STATE_LOCK(ctx);
5150 	ctx->ifc_flags |= IFC_IN_DETACH;
5151 	STATE_UNLOCK(ctx);
5152 
5153 	/* Unregister VLAN handlers before calling iflib_stop() */
5154 	iflib_unregister_vlan_handlers(ctx);
5155 
5156 	iflib_netmap_detach(ifp);
5157 	ether_ifdetach(ifp);
5158 
5159 	CTX_LOCK(ctx);
5160 	iflib_stop(ctx);
5161 	CTX_UNLOCK(ctx);
5162 
5163 	iflib_rem_pfil(ctx);
5164 	if (ctx->ifc_led_dev != NULL)
5165 		led_destroy(ctx->ifc_led_dev);
5166 	/* XXX drain any dependent tasks */
5167 	tqg = qgroup_if_io_tqg;
5168 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
5169 		callout_drain(&txq->ift_timer);
5170 #ifdef DEV_NETMAP
5171 		callout_drain(&txq->ift_netmap_timer);
5172 #endif /* DEV_NETMAP */
5173 		if (txq->ift_task.gt_uniq != NULL)
5174 			taskqgroup_detach(tqg, &txq->ift_task);
5175 	}
5176 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
5177 		if (rxq->ifr_task.gt_uniq != NULL)
5178 			taskqgroup_detach(tqg, &rxq->ifr_task);
5179 
5180 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5181 			free(fl->ifl_rx_bitmap, M_IFLIB);
5182 	}
5183 	tqg = qgroup_if_config_tqg;
5184 	if (ctx->ifc_admin_task.gt_uniq != NULL)
5185 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
5186 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
5187 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
5188 	CTX_LOCK(ctx);
5189 	IFDI_DETACH(ctx);
5190 	CTX_UNLOCK(ctx);
5191 
5192 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
5193 	iflib_free_intr_mem(ctx);
5194 
5195 	bus_generic_detach(dev);
5196 
5197 	iflib_tx_structures_free(ctx);
5198 	iflib_rx_structures_free(ctx);
5199 
5200 	iflib_deregister(ctx);
5201 
5202 	device_set_softc(ctx->ifc_dev, NULL);
5203 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5204 		free(ctx->ifc_softc, M_IFLIB);
5205 	unref_ctx_core_offset(ctx);
5206 	free(ctx, M_IFLIB);
5207 	return (0);
5208 }
5209 
5210 static void
5211 iflib_free_intr_mem(if_ctx_t ctx)
5212 {
5213 
5214 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) {
5215 		iflib_irq_free(ctx, &ctx->ifc_legacy_irq);
5216 	}
5217 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) {
5218 		pci_release_msi(ctx->ifc_dev);
5219 	}
5220 	if (ctx->ifc_msix_mem != NULL) {
5221 		bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY,
5222 		    rman_get_rid(ctx->ifc_msix_mem), ctx->ifc_msix_mem);
5223 		ctx->ifc_msix_mem = NULL;
5224 	}
5225 }
5226 
5227 int
5228 iflib_device_detach(device_t dev)
5229 {
5230 	if_ctx_t ctx = device_get_softc(dev);
5231 
5232 	return (iflib_device_deregister(ctx));
5233 }
5234 
5235 int
5236 iflib_device_suspend(device_t dev)
5237 {
5238 	if_ctx_t ctx = device_get_softc(dev);
5239 
5240 	CTX_LOCK(ctx);
5241 	IFDI_SUSPEND(ctx);
5242 	CTX_UNLOCK(ctx);
5243 
5244 	return bus_generic_suspend(dev);
5245 }
5246 int
5247 iflib_device_shutdown(device_t dev)
5248 {
5249 	if_ctx_t ctx = device_get_softc(dev);
5250 
5251 	CTX_LOCK(ctx);
5252 	IFDI_SHUTDOWN(ctx);
5253 	CTX_UNLOCK(ctx);
5254 
5255 	return bus_generic_suspend(dev);
5256 }
5257 
5258 int
5259 iflib_device_resume(device_t dev)
5260 {
5261 	if_ctx_t ctx = device_get_softc(dev);
5262 	iflib_txq_t txq = ctx->ifc_txqs;
5263 
5264 	CTX_LOCK(ctx);
5265 	IFDI_RESUME(ctx);
5266 	iflib_if_init_locked(ctx);
5267 	CTX_UNLOCK(ctx);
5268 	for (int i = 0; i < NTXQSETS(ctx); i++, txq++)
5269 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
5270 
5271 	return (bus_generic_resume(dev));
5272 }
5273 
5274 int
5275 iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params)
5276 {
5277 	int error;
5278 	if_ctx_t ctx = device_get_softc(dev);
5279 
5280 	CTX_LOCK(ctx);
5281 	error = IFDI_IOV_INIT(ctx, num_vfs, params);
5282 	CTX_UNLOCK(ctx);
5283 
5284 	return (error);
5285 }
5286 
5287 void
5288 iflib_device_iov_uninit(device_t dev)
5289 {
5290 	if_ctx_t ctx = device_get_softc(dev);
5291 
5292 	CTX_LOCK(ctx);
5293 	IFDI_IOV_UNINIT(ctx);
5294 	CTX_UNLOCK(ctx);
5295 }
5296 
5297 int
5298 iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params)
5299 {
5300 	int error;
5301 	if_ctx_t ctx = device_get_softc(dev);
5302 
5303 	CTX_LOCK(ctx);
5304 	error = IFDI_IOV_VF_ADD(ctx, vfnum, params);
5305 	CTX_UNLOCK(ctx);
5306 
5307 	return (error);
5308 }
5309 
5310 /*********************************************************************
5311  *
5312  *  MODULE FUNCTION DEFINITIONS
5313  *
5314  **********************************************************************/
5315 
5316 /*
5317  * - Start a fast taskqueue thread for each core
5318  * - Start a taskqueue for control operations
5319  */
5320 static int
5321 iflib_module_init(void)
5322 {
5323 	return (0);
5324 }
5325 
5326 static int
5327 iflib_module_event_handler(module_t mod, int what, void *arg)
5328 {
5329 	int err;
5330 
5331 	switch (what) {
5332 	case MOD_LOAD:
5333 		if ((err = iflib_module_init()) != 0)
5334 			return (err);
5335 		break;
5336 	case MOD_UNLOAD:
5337 		return (EBUSY);
5338 	default:
5339 		return (EOPNOTSUPP);
5340 	}
5341 
5342 	return (0);
5343 }
5344 
5345 /*********************************************************************
5346  *
5347  *  PUBLIC FUNCTION DEFINITIONS
5348  *     ordered as in iflib.h
5349  *
5350  **********************************************************************/
5351 
5352 static void
5353 _iflib_assert(if_shared_ctx_t sctx)
5354 {
5355 	int i;
5356 
5357 	MPASS(sctx->isc_tx_maxsize);
5358 	MPASS(sctx->isc_tx_maxsegsize);
5359 
5360 	MPASS(sctx->isc_rx_maxsize);
5361 	MPASS(sctx->isc_rx_nsegments);
5362 	MPASS(sctx->isc_rx_maxsegsize);
5363 
5364 	MPASS(sctx->isc_nrxqs >= 1 && sctx->isc_nrxqs <= 8);
5365 	for (i = 0; i < sctx->isc_nrxqs; i++) {
5366 		MPASS(sctx->isc_nrxd_min[i]);
5367 		MPASS(powerof2(sctx->isc_nrxd_min[i]));
5368 		MPASS(sctx->isc_nrxd_max[i]);
5369 		MPASS(powerof2(sctx->isc_nrxd_max[i]));
5370 		MPASS(sctx->isc_nrxd_default[i]);
5371 		MPASS(powerof2(sctx->isc_nrxd_default[i]));
5372 	}
5373 
5374 	MPASS(sctx->isc_ntxqs >= 1 && sctx->isc_ntxqs <= 8);
5375 	for (i = 0; i < sctx->isc_ntxqs; i++) {
5376 		MPASS(sctx->isc_ntxd_min[i]);
5377 		MPASS(powerof2(sctx->isc_ntxd_min[i]));
5378 		MPASS(sctx->isc_ntxd_max[i]);
5379 		MPASS(powerof2(sctx->isc_ntxd_max[i]));
5380 		MPASS(sctx->isc_ntxd_default[i]);
5381 		MPASS(powerof2(sctx->isc_ntxd_default[i]));
5382 	}
5383 }
5384 
5385 static void
5386 _iflib_pre_assert(if_softc_ctx_t scctx)
5387 {
5388 
5389 	MPASS(scctx->isc_txrx->ift_txd_encap);
5390 	MPASS(scctx->isc_txrx->ift_txd_flush);
5391 	MPASS(scctx->isc_txrx->ift_txd_credits_update);
5392 	MPASS(scctx->isc_txrx->ift_rxd_available);
5393 	MPASS(scctx->isc_txrx->ift_rxd_pkt_get);
5394 	MPASS(scctx->isc_txrx->ift_rxd_refill);
5395 	MPASS(scctx->isc_txrx->ift_rxd_flush);
5396 }
5397 
5398 static int
5399 iflib_register(if_ctx_t ctx)
5400 {
5401 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5402 	driver_t *driver = sctx->isc_driver;
5403 	device_t dev = ctx->ifc_dev;
5404 	if_t ifp;
5405 	u_char type;
5406 	int iflags;
5407 
5408 	if ((sctx->isc_flags & IFLIB_PSEUDO) == 0)
5409 		_iflib_assert(sctx);
5410 
5411 	CTX_LOCK_INIT(ctx);
5412 	STATE_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev));
5413 	if (sctx->isc_flags & IFLIB_PSEUDO) {
5414 		if (sctx->isc_flags & IFLIB_PSEUDO_ETHER)
5415 			type = IFT_ETHER;
5416 		else
5417 			type = IFT_PPP;
5418 	} else
5419 		type = IFT_ETHER;
5420 	ifp = ctx->ifc_ifp = if_alloc(type);
5421 	if (ifp == NULL) {
5422 		device_printf(dev, "can not allocate ifnet structure\n");
5423 		return (ENOMEM);
5424 	}
5425 
5426 	/*
5427 	 * Initialize our context's device specific methods
5428 	 */
5429 	kobj_init((kobj_t) ctx, (kobj_class_t) driver);
5430 	kobj_class_compile((kobj_class_t) driver);
5431 
5432 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
5433 	if_setsoftc(ifp, ctx);
5434 	if_setdev(ifp, dev);
5435 	if_setinitfn(ifp, iflib_if_init);
5436 	if_setioctlfn(ifp, iflib_if_ioctl);
5437 #ifdef ALTQ
5438 	if_setstartfn(ifp, iflib_altq_if_start);
5439 	if_settransmitfn(ifp, iflib_altq_if_transmit);
5440 	if_setsendqready(ifp);
5441 #else
5442 	if_settransmitfn(ifp, iflib_if_transmit);
5443 #endif
5444 	if_setqflushfn(ifp, iflib_if_qflush);
5445 	iflags = IFF_MULTICAST | IFF_KNOWSEPOCH;
5446 
5447 	if ((sctx->isc_flags & IFLIB_PSEUDO) &&
5448 		(sctx->isc_flags & IFLIB_PSEUDO_ETHER) == 0)
5449 		iflags |= IFF_POINTOPOINT;
5450 	else
5451 		iflags |= IFF_BROADCAST | IFF_SIMPLEX;
5452 	if_setflags(ifp, iflags);
5453 	ctx->ifc_vlan_attach_event =
5454 		EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx,
5455 							  EVENTHANDLER_PRI_FIRST);
5456 	ctx->ifc_vlan_detach_event =
5457 		EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx,
5458 							  EVENTHANDLER_PRI_FIRST);
5459 
5460 	if ((sctx->isc_flags & IFLIB_DRIVER_MEDIA) == 0) {
5461 		ctx->ifc_mediap = &ctx->ifc_media;
5462 		ifmedia_init(ctx->ifc_mediap, IFM_IMASK,
5463 		    iflib_media_change, iflib_media_status);
5464 	}
5465 	return (0);
5466 }
5467 
5468 static void
5469 iflib_unregister_vlan_handlers(if_ctx_t ctx)
5470 {
5471 	/* Unregister VLAN events */
5472 	if (ctx->ifc_vlan_attach_event != NULL) {
5473 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
5474 		ctx->ifc_vlan_attach_event = NULL;
5475 	}
5476 	if (ctx->ifc_vlan_detach_event != NULL) {
5477 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
5478 		ctx->ifc_vlan_detach_event = NULL;
5479 	}
5480 
5481 }
5482 
5483 static void
5484 iflib_deregister(if_ctx_t ctx)
5485 {
5486 	if_t ifp = ctx->ifc_ifp;
5487 
5488 	/* Remove all media */
5489 	ifmedia_removeall(&ctx->ifc_media);
5490 
5491 	/* Ensure that VLAN event handlers are unregistered */
5492 	iflib_unregister_vlan_handlers(ctx);
5493 
5494 	/* Release kobject reference */
5495 	kobj_delete((kobj_t) ctx, NULL);
5496 
5497 	/* Free the ifnet structure */
5498 	if_free(ifp);
5499 
5500 	STATE_LOCK_DESTROY(ctx);
5501 
5502 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
5503 	CTX_LOCK_DESTROY(ctx);
5504 }
5505 
5506 static int
5507 iflib_queues_alloc(if_ctx_t ctx)
5508 {
5509 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5510 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5511 	device_t dev = ctx->ifc_dev;
5512 	int nrxqsets = scctx->isc_nrxqsets;
5513 	int ntxqsets = scctx->isc_ntxqsets;
5514 	iflib_txq_t txq;
5515 	iflib_rxq_t rxq;
5516 	iflib_fl_t fl = NULL;
5517 	int i, j, cpu, err, txconf, rxconf;
5518 	iflib_dma_info_t ifdip;
5519 	uint32_t *rxqsizes = scctx->isc_rxqsizes;
5520 	uint32_t *txqsizes = scctx->isc_txqsizes;
5521 	uint8_t nrxqs = sctx->isc_nrxqs;
5522 	uint8_t ntxqs = sctx->isc_ntxqs;
5523 	int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1;
5524 	caddr_t *vaddrs;
5525 	uint64_t *paddrs;
5526 
5527 	KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1"));
5528 	KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1"));
5529 
5530 	/* Allocate the TX ring struct memory */
5531 	if (!(ctx->ifc_txqs =
5532 	    (iflib_txq_t) malloc(sizeof(struct iflib_txq) *
5533 	    ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5534 		device_printf(dev, "Unable to allocate TX ring memory\n");
5535 		err = ENOMEM;
5536 		goto fail;
5537 	}
5538 
5539 	/* Now allocate the RX */
5540 	if (!(ctx->ifc_rxqs =
5541 	    (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) *
5542 	    nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5543 		device_printf(dev, "Unable to allocate RX ring memory\n");
5544 		err = ENOMEM;
5545 		goto rx_fail;
5546 	}
5547 
5548 	txq = ctx->ifc_txqs;
5549 	rxq = ctx->ifc_rxqs;
5550 
5551 	/*
5552 	 * XXX handle allocation failure
5553 	 */
5554 	for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) {
5555 		/* Set up some basics */
5556 
5557 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs,
5558 		    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
5559 			device_printf(dev,
5560 			    "Unable to allocate TX DMA info memory\n");
5561 			err = ENOMEM;
5562 			goto err_tx_desc;
5563 		}
5564 		txq->ift_ifdi = ifdip;
5565 		for (j = 0; j < ntxqs; j++, ifdip++) {
5566 			if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, 0)) {
5567 				device_printf(dev,
5568 				    "Unable to allocate TX descriptors\n");
5569 				err = ENOMEM;
5570 				goto err_tx_desc;
5571 			}
5572 			txq->ift_txd_size[j] = scctx->isc_txd_size[j];
5573 			bzero((void *)ifdip->idi_vaddr, txqsizes[j]);
5574 		}
5575 		txq->ift_ctx = ctx;
5576 		txq->ift_id = i;
5577 		if (sctx->isc_flags & IFLIB_HAS_TXCQ) {
5578 			txq->ift_br_offset = 1;
5579 		} else {
5580 			txq->ift_br_offset = 0;
5581 		}
5582 
5583 		if (iflib_txsd_alloc(txq)) {
5584 			device_printf(dev, "Critical Failure setting up TX buffers\n");
5585 			err = ENOMEM;
5586 			goto err_tx_desc;
5587 		}
5588 
5589 		/* Initialize the TX lock */
5590 		snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:TX(%d):callout",
5591 		    device_get_nameunit(dev), txq->ift_id);
5592 		mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF);
5593 		callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0);
5594 		txq->ift_timer.c_cpu = cpu;
5595 #ifdef DEV_NETMAP
5596 		callout_init_mtx(&txq->ift_netmap_timer, &txq->ift_mtx, 0);
5597 		txq->ift_netmap_timer.c_cpu = cpu;
5598 #endif /* DEV_NETMAP */
5599 
5600 		err = ifmp_ring_alloc(&txq->ift_br, 2048, txq, iflib_txq_drain,
5601 				      iflib_txq_can_drain, M_IFLIB, M_WAITOK);
5602 		if (err) {
5603 			/* XXX free any allocated rings */
5604 			device_printf(dev, "Unable to allocate buf_ring\n");
5605 			goto err_tx_desc;
5606 		}
5607 	}
5608 
5609 	for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) {
5610 		/* Set up some basics */
5611 		callout_init(&rxq->ifr_watchdog, 1);
5612 
5613 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs,
5614 		   M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
5615 			device_printf(dev,
5616 			    "Unable to allocate RX DMA info memory\n");
5617 			err = ENOMEM;
5618 			goto err_tx_desc;
5619 		}
5620 
5621 		rxq->ifr_ifdi = ifdip;
5622 		/* XXX this needs to be changed if #rx queues != #tx queues */
5623 		rxq->ifr_ntxqirq = 1;
5624 		rxq->ifr_txqid[0] = i;
5625 		for (j = 0; j < nrxqs; j++, ifdip++) {
5626 			if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, 0)) {
5627 				device_printf(dev,
5628 				    "Unable to allocate RX descriptors\n");
5629 				err = ENOMEM;
5630 				goto err_tx_desc;
5631 			}
5632 			bzero((void *)ifdip->idi_vaddr, rxqsizes[j]);
5633 		}
5634 		rxq->ifr_ctx = ctx;
5635 		rxq->ifr_id = i;
5636 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
5637 			rxq->ifr_fl_offset = 1;
5638 		} else {
5639 			rxq->ifr_fl_offset = 0;
5640 		}
5641 		rxq->ifr_nfl = nfree_lists;
5642 		if (!(fl =
5643 			  (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) {
5644 			device_printf(dev, "Unable to allocate free list memory\n");
5645 			err = ENOMEM;
5646 			goto err_tx_desc;
5647 		}
5648 		rxq->ifr_fl = fl;
5649 		for (j = 0; j < nfree_lists; j++) {
5650 			fl[j].ifl_rxq = rxq;
5651 			fl[j].ifl_id = j;
5652 			fl[j].ifl_ifdi = &rxq->ifr_ifdi[j + rxq->ifr_fl_offset];
5653 			fl[j].ifl_rxd_size = scctx->isc_rxd_size[j];
5654 		}
5655 		/* Allocate receive buffers for the ring */
5656 		if (iflib_rxsd_alloc(rxq)) {
5657 			device_printf(dev,
5658 			    "Critical Failure setting up receive buffers\n");
5659 			err = ENOMEM;
5660 			goto err_rx_desc;
5661 		}
5662 
5663 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5664 			fl->ifl_rx_bitmap = bit_alloc(fl->ifl_size, M_IFLIB,
5665 			    M_WAITOK);
5666 	}
5667 
5668 	/* TXQs */
5669 	vaddrs = malloc(sizeof(caddr_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5670 	paddrs = malloc(sizeof(uint64_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5671 	for (i = 0; i < ntxqsets; i++) {
5672 		iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi;
5673 
5674 		for (j = 0; j < ntxqs; j++, di++) {
5675 			vaddrs[i*ntxqs + j] = di->idi_vaddr;
5676 			paddrs[i*ntxqs + j] = di->idi_paddr;
5677 		}
5678 	}
5679 	if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) {
5680 		device_printf(ctx->ifc_dev,
5681 		    "Unable to allocate device TX queue\n");
5682 		iflib_tx_structures_free(ctx);
5683 		free(vaddrs, M_IFLIB);
5684 		free(paddrs, M_IFLIB);
5685 		goto err_rx_desc;
5686 	}
5687 	free(vaddrs, M_IFLIB);
5688 	free(paddrs, M_IFLIB);
5689 
5690 	/* RXQs */
5691 	vaddrs = malloc(sizeof(caddr_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5692 	paddrs = malloc(sizeof(uint64_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5693 	for (i = 0; i < nrxqsets; i++) {
5694 		iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi;
5695 
5696 		for (j = 0; j < nrxqs; j++, di++) {
5697 			vaddrs[i*nrxqs + j] = di->idi_vaddr;
5698 			paddrs[i*nrxqs + j] = di->idi_paddr;
5699 		}
5700 	}
5701 	if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) {
5702 		device_printf(ctx->ifc_dev,
5703 		    "Unable to allocate device RX queue\n");
5704 		iflib_tx_structures_free(ctx);
5705 		free(vaddrs, M_IFLIB);
5706 		free(paddrs, M_IFLIB);
5707 		goto err_rx_desc;
5708 	}
5709 	free(vaddrs, M_IFLIB);
5710 	free(paddrs, M_IFLIB);
5711 
5712 	return (0);
5713 
5714 /* XXX handle allocation failure changes */
5715 err_rx_desc:
5716 err_tx_desc:
5717 rx_fail:
5718 	if (ctx->ifc_rxqs != NULL)
5719 		free(ctx->ifc_rxqs, M_IFLIB);
5720 	ctx->ifc_rxqs = NULL;
5721 	if (ctx->ifc_txqs != NULL)
5722 		free(ctx->ifc_txqs, M_IFLIB);
5723 	ctx->ifc_txqs = NULL;
5724 fail:
5725 	return (err);
5726 }
5727 
5728 static int
5729 iflib_tx_structures_setup(if_ctx_t ctx)
5730 {
5731 	iflib_txq_t txq = ctx->ifc_txqs;
5732 	int i;
5733 
5734 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
5735 		iflib_txq_setup(txq);
5736 
5737 	return (0);
5738 }
5739 
5740 static void
5741 iflib_tx_structures_free(if_ctx_t ctx)
5742 {
5743 	iflib_txq_t txq = ctx->ifc_txqs;
5744 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5745 	int i, j;
5746 
5747 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
5748 		for (j = 0; j < sctx->isc_ntxqs; j++)
5749 			iflib_dma_free(&txq->ift_ifdi[j]);
5750 		iflib_txq_destroy(txq);
5751 	}
5752 	free(ctx->ifc_txqs, M_IFLIB);
5753 	ctx->ifc_txqs = NULL;
5754 	IFDI_QUEUES_FREE(ctx);
5755 }
5756 
5757 /*********************************************************************
5758  *
5759  *  Initialize all receive rings.
5760  *
5761  **********************************************************************/
5762 static int
5763 iflib_rx_structures_setup(if_ctx_t ctx)
5764 {
5765 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5766 	int q;
5767 #if defined(INET6) || defined(INET)
5768 	int err, i;
5769 #endif
5770 
5771 	for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) {
5772 #if defined(INET6) || defined(INET)
5773 		if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_LRO) {
5774 			err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp,
5775 			    TCP_LRO_ENTRIES, min(1024,
5776 			    ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset]));
5777 			if (err != 0) {
5778 				device_printf(ctx->ifc_dev,
5779 				    "LRO Initialization failed!\n");
5780 				goto fail;
5781 			}
5782 		}
5783 #endif
5784 		IFDI_RXQ_SETUP(ctx, rxq->ifr_id);
5785 	}
5786 	return (0);
5787 #if defined(INET6) || defined(INET)
5788 fail:
5789 	/*
5790 	 * Free LRO resources allocated so far, we will only handle
5791 	 * the rings that completed, the failing case will have
5792 	 * cleaned up for itself.  'q' failed, so its the terminus.
5793 	 */
5794 	rxq = ctx->ifc_rxqs;
5795 	for (i = 0; i < q; ++i, rxq++) {
5796 		if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_LRO)
5797 			tcp_lro_free(&rxq->ifr_lc);
5798 	}
5799 	return (err);
5800 #endif
5801 }
5802 
5803 /*********************************************************************
5804  *
5805  *  Free all receive rings.
5806  *
5807  **********************************************************************/
5808 static void
5809 iflib_rx_structures_free(if_ctx_t ctx)
5810 {
5811 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5812 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5813 	int i, j;
5814 
5815 	for (i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) {
5816 		for (j = 0; j < sctx->isc_nrxqs; j++)
5817 			iflib_dma_free(&rxq->ifr_ifdi[j]);
5818 		iflib_rx_sds_free(rxq);
5819 #if defined(INET6) || defined(INET)
5820 		if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_LRO)
5821 			tcp_lro_free(&rxq->ifr_lc);
5822 #endif
5823 	}
5824 	free(ctx->ifc_rxqs, M_IFLIB);
5825 	ctx->ifc_rxqs = NULL;
5826 }
5827 
5828 static int
5829 iflib_qset_structures_setup(if_ctx_t ctx)
5830 {
5831 	int err;
5832 
5833 	/*
5834 	 * It is expected that the caller takes care of freeing queues if this
5835 	 * fails.
5836 	 */
5837 	if ((err = iflib_tx_structures_setup(ctx)) != 0) {
5838 		device_printf(ctx->ifc_dev, "iflib_tx_structures_setup failed: %d\n", err);
5839 		return (err);
5840 	}
5841 
5842 	if ((err = iflib_rx_structures_setup(ctx)) != 0)
5843 		device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err);
5844 
5845 	return (err);
5846 }
5847 
5848 int
5849 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
5850 		driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, const char *name)
5851 {
5852 
5853 	return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name));
5854 }
5855 
5856 #ifdef SMP
5857 static int
5858 find_nth(if_ctx_t ctx, int qid)
5859 {
5860 	cpuset_t cpus;
5861 	int i, cpuid, eqid, count;
5862 
5863 	CPU_COPY(&ctx->ifc_cpus, &cpus);
5864 	count = CPU_COUNT(&cpus);
5865 	eqid = qid % count;
5866 	/* clear up to the qid'th bit */
5867 	for (i = 0; i < eqid; i++) {
5868 		cpuid = CPU_FFS(&cpus);
5869 		MPASS(cpuid != 0);
5870 		CPU_CLR(cpuid-1, &cpus);
5871 	}
5872 	cpuid = CPU_FFS(&cpus);
5873 	MPASS(cpuid != 0);
5874 	return (cpuid-1);
5875 }
5876 
5877 #ifdef SCHED_ULE
5878 extern struct cpu_group *cpu_top;              /* CPU topology */
5879 
5880 static int
5881 find_child_with_core(int cpu, struct cpu_group *grp)
5882 {
5883 	int i;
5884 
5885 	if (grp->cg_children == 0)
5886 		return -1;
5887 
5888 	MPASS(grp->cg_child);
5889 	for (i = 0; i < grp->cg_children; i++) {
5890 		if (CPU_ISSET(cpu, &grp->cg_child[i].cg_mask))
5891 			return i;
5892 	}
5893 
5894 	return -1;
5895 }
5896 
5897 /*
5898  * Find the nth "close" core to the specified core
5899  * "close" is defined as the deepest level that shares
5900  * at least an L2 cache.  With threads, this will be
5901  * threads on the same core.  If the shared cache is L3
5902  * or higher, simply returns the same core.
5903  */
5904 static int
5905 find_close_core(int cpu, int core_offset)
5906 {
5907 	struct cpu_group *grp;
5908 	int i;
5909 	int fcpu;
5910 	cpuset_t cs;
5911 
5912 	grp = cpu_top;
5913 	if (grp == NULL)
5914 		return cpu;
5915 	i = 0;
5916 	while ((i = find_child_with_core(cpu, grp)) != -1) {
5917 		/* If the child only has one cpu, don't descend */
5918 		if (grp->cg_child[i].cg_count <= 1)
5919 			break;
5920 		grp = &grp->cg_child[i];
5921 	}
5922 
5923 	/* If they don't share at least an L2 cache, use the same CPU */
5924 	if (grp->cg_level > CG_SHARE_L2 || grp->cg_level == CG_SHARE_NONE)
5925 		return cpu;
5926 
5927 	/* Now pick one */
5928 	CPU_COPY(&grp->cg_mask, &cs);
5929 
5930 	/* Add the selected CPU offset to core offset. */
5931 	for (i = 0; (fcpu = CPU_FFS(&cs)) != 0; i++) {
5932 		if (fcpu - 1 == cpu)
5933 			break;
5934 		CPU_CLR(fcpu - 1, &cs);
5935 	}
5936 	MPASS(fcpu);
5937 
5938 	core_offset += i;
5939 
5940 	CPU_COPY(&grp->cg_mask, &cs);
5941 	for (i = core_offset % grp->cg_count; i > 0; i--) {
5942 		MPASS(CPU_FFS(&cs));
5943 		CPU_CLR(CPU_FFS(&cs) - 1, &cs);
5944 	}
5945 	MPASS(CPU_FFS(&cs));
5946 	return CPU_FFS(&cs) - 1;
5947 }
5948 #else
5949 static int
5950 find_close_core(int cpu, int core_offset __unused)
5951 {
5952 	return cpu;
5953 }
5954 #endif
5955 
5956 static int
5957 get_core_offset(if_ctx_t ctx, iflib_intr_type_t type, int qid)
5958 {
5959 	switch (type) {
5960 	case IFLIB_INTR_TX:
5961 		/* TX queues get cores which share at least an L2 cache with the corresponding RX queue */
5962 		/* XXX handle multiple RX threads per core and more than two core per L2 group */
5963 		return qid / CPU_COUNT(&ctx->ifc_cpus) + 1;
5964 	case IFLIB_INTR_RX:
5965 	case IFLIB_INTR_RXTX:
5966 		/* RX queues get the specified core */
5967 		return qid / CPU_COUNT(&ctx->ifc_cpus);
5968 	default:
5969 		return -1;
5970 	}
5971 }
5972 #else
5973 #define get_core_offset(ctx, type, qid)	CPU_FIRST()
5974 #define find_close_core(cpuid, tid)	CPU_FIRST()
5975 #define find_nth(ctx, gid)		CPU_FIRST()
5976 #endif
5977 
5978 /* Just to avoid copy/paste */
5979 static inline int
5980 iflib_irq_set_affinity(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type,
5981     int qid, struct grouptask *gtask, struct taskqgroup *tqg, void *uniq,
5982     const char *name)
5983 {
5984 	device_t dev;
5985 	int co, cpuid, err, tid;
5986 
5987 	dev = ctx->ifc_dev;
5988 	co = ctx->ifc_sysctl_core_offset;
5989 	if (ctx->ifc_sysctl_separate_txrx && type == IFLIB_INTR_TX)
5990 		co += ctx->ifc_softc_ctx.isc_nrxqsets;
5991 	cpuid = find_nth(ctx, qid + co);
5992 	tid = get_core_offset(ctx, type, qid);
5993 	if (tid < 0) {
5994 		device_printf(dev, "get_core_offset failed\n");
5995 		return (EOPNOTSUPP);
5996 	}
5997 	cpuid = find_close_core(cpuid, tid);
5998 	err = taskqgroup_attach_cpu(tqg, gtask, uniq, cpuid, dev, irq->ii_res,
5999 	    name);
6000 	if (err) {
6001 		device_printf(dev, "taskqgroup_attach_cpu failed %d\n", err);
6002 		return (err);
6003 	}
6004 #ifdef notyet
6005 	if (cpuid > ctx->ifc_cpuid_highest)
6006 		ctx->ifc_cpuid_highest = cpuid;
6007 #endif
6008 	return (0);
6009 }
6010 
6011 int
6012 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid,
6013 			iflib_intr_type_t type, driver_filter_t *filter,
6014 			void *filter_arg, int qid, const char *name)
6015 {
6016 	device_t dev;
6017 	struct grouptask *gtask;
6018 	struct taskqgroup *tqg;
6019 	iflib_filter_info_t info;
6020 	gtask_fn_t *fn;
6021 	int tqrid, err;
6022 	driver_filter_t *intr_fast;
6023 	void *q;
6024 
6025 	info = &ctx->ifc_filter_info;
6026 	tqrid = rid;
6027 
6028 	switch (type) {
6029 	/* XXX merge tx/rx for netmap? */
6030 	case IFLIB_INTR_TX:
6031 		q = &ctx->ifc_txqs[qid];
6032 		info = &ctx->ifc_txqs[qid].ift_filter_info;
6033 		gtask = &ctx->ifc_txqs[qid].ift_task;
6034 		tqg = qgroup_if_io_tqg;
6035 		fn = _task_fn_tx;
6036 		intr_fast = iflib_fast_intr;
6037 		GROUPTASK_INIT(gtask, 0, fn, q);
6038 		ctx->ifc_flags |= IFC_NETMAP_TX_IRQ;
6039 		break;
6040 	case IFLIB_INTR_RX:
6041 		q = &ctx->ifc_rxqs[qid];
6042 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
6043 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6044 		tqg = qgroup_if_io_tqg;
6045 		fn = _task_fn_rx;
6046 		intr_fast = iflib_fast_intr;
6047 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6048 		break;
6049 	case IFLIB_INTR_RXTX:
6050 		q = &ctx->ifc_rxqs[qid];
6051 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
6052 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6053 		tqg = qgroup_if_io_tqg;
6054 		fn = _task_fn_rx;
6055 		intr_fast = iflib_fast_intr_rxtx;
6056 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6057 		break;
6058 	case IFLIB_INTR_ADMIN:
6059 		q = ctx;
6060 		tqrid = -1;
6061 		info = &ctx->ifc_filter_info;
6062 		gtask = &ctx->ifc_admin_task;
6063 		tqg = qgroup_if_config_tqg;
6064 		fn = _task_fn_admin;
6065 		intr_fast = iflib_fast_intr_ctx;
6066 		break;
6067 	default:
6068 		device_printf(ctx->ifc_dev, "%s: unknown net intr type\n",
6069 		    __func__);
6070 		return (EINVAL);
6071 	}
6072 
6073 	info->ifi_filter = filter;
6074 	info->ifi_filter_arg = filter_arg;
6075 	info->ifi_task = gtask;
6076 	info->ifi_ctx = q;
6077 
6078 	dev = ctx->ifc_dev;
6079 	err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info,  name);
6080 	if (err != 0) {
6081 		device_printf(dev, "_iflib_irq_alloc failed %d\n", err);
6082 		return (err);
6083 	}
6084 	if (type == IFLIB_INTR_ADMIN)
6085 		return (0);
6086 
6087 	if (tqrid != -1) {
6088 		err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg,
6089 		    q, name);
6090 		if (err)
6091 			return (err);
6092 	} else {
6093 		taskqgroup_attach(tqg, gtask, q, dev, irq->ii_res, name);
6094 	}
6095 
6096 	return (0);
6097 }
6098 
6099 void
6100 iflib_softirq_alloc_generic(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type, void *arg, int qid, const char *name)
6101 {
6102 	struct grouptask *gtask;
6103 	struct taskqgroup *tqg;
6104 	gtask_fn_t *fn;
6105 	void *q;
6106 	int err;
6107 
6108 	switch (type) {
6109 	case IFLIB_INTR_TX:
6110 		q = &ctx->ifc_txqs[qid];
6111 		gtask = &ctx->ifc_txqs[qid].ift_task;
6112 		tqg = qgroup_if_io_tqg;
6113 		fn = _task_fn_tx;
6114 		GROUPTASK_INIT(gtask, 0, fn, q);
6115 		break;
6116 	case IFLIB_INTR_RX:
6117 		q = &ctx->ifc_rxqs[qid];
6118 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6119 		tqg = qgroup_if_io_tqg;
6120 		fn = _task_fn_rx;
6121 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6122 		break;
6123 	case IFLIB_INTR_IOV:
6124 		q = ctx;
6125 		gtask = &ctx->ifc_vflr_task;
6126 		tqg = qgroup_if_config_tqg;
6127 		fn = _task_fn_iov;
6128 		GROUPTASK_INIT(gtask, 0, fn, q);
6129 		break;
6130 	default:
6131 		panic("unknown net intr type");
6132 	}
6133 	if (irq != NULL) {
6134 		err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg,
6135 		    q, name);
6136 		if (err)
6137 			taskqgroup_attach(tqg, gtask, q, ctx->ifc_dev,
6138 			    irq->ii_res, name);
6139 	} else {
6140 		taskqgroup_attach(tqg, gtask, q, NULL, NULL, name);
6141 	}
6142 }
6143 
6144 void
6145 iflib_irq_free(if_ctx_t ctx, if_irq_t irq)
6146 {
6147 
6148 	if (irq->ii_tag)
6149 		bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag);
6150 
6151 	if (irq->ii_res)
6152 		bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ,
6153 		    rman_get_rid(irq->ii_res), irq->ii_res);
6154 }
6155 
6156 static int
6157 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, const char *name)
6158 {
6159 	iflib_txq_t txq = ctx->ifc_txqs;
6160 	iflib_rxq_t rxq = ctx->ifc_rxqs;
6161 	if_irq_t irq = &ctx->ifc_legacy_irq;
6162 	iflib_filter_info_t info;
6163 	device_t dev;
6164 	struct grouptask *gtask;
6165 	struct resource *res;
6166 	struct taskqgroup *tqg;
6167 	void *q;
6168 	int err, tqrid;
6169 	bool rx_only;
6170 
6171 	q = &ctx->ifc_rxqs[0];
6172 	info = &rxq[0].ifr_filter_info;
6173 	gtask = &rxq[0].ifr_task;
6174 	tqg = qgroup_if_io_tqg;
6175 	tqrid = *rid;
6176 	rx_only = (ctx->ifc_sctx->isc_flags & IFLIB_SINGLE_IRQ_RX_ONLY) != 0;
6177 
6178 	ctx->ifc_flags |= IFC_LEGACY;
6179 	info->ifi_filter = filter;
6180 	info->ifi_filter_arg = filter_arg;
6181 	info->ifi_task = gtask;
6182 	info->ifi_ctx = rx_only ? ctx : q;
6183 
6184 	dev = ctx->ifc_dev;
6185 	/* We allocate a single interrupt resource */
6186 	err = _iflib_irq_alloc(ctx, irq, tqrid, rx_only ? iflib_fast_intr_ctx :
6187 	    iflib_fast_intr_rxtx, NULL, info, name);
6188 	if (err != 0)
6189 		return (err);
6190 	NET_GROUPTASK_INIT(gtask, 0, _task_fn_rx, q);
6191 	res = irq->ii_res;
6192 	taskqgroup_attach(tqg, gtask, q, dev, res, name);
6193 
6194 	GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq);
6195 	taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, dev, res,
6196 	    "tx");
6197 	return (0);
6198 }
6199 
6200 void
6201 iflib_led_create(if_ctx_t ctx)
6202 {
6203 
6204 	ctx->ifc_led_dev = led_create(iflib_led_func, ctx,
6205 	    device_get_nameunit(ctx->ifc_dev));
6206 }
6207 
6208 void
6209 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid)
6210 {
6211 
6212 	GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
6213 }
6214 
6215 void
6216 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid)
6217 {
6218 
6219 	GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task);
6220 }
6221 
6222 void
6223 iflib_admin_intr_deferred(if_ctx_t ctx)
6224 {
6225 
6226 	MPASS(ctx->ifc_admin_task.gt_taskqueue != NULL);
6227 	GROUPTASK_ENQUEUE(&ctx->ifc_admin_task);
6228 }
6229 
6230 void
6231 iflib_iov_intr_deferred(if_ctx_t ctx)
6232 {
6233 
6234 	GROUPTASK_ENQUEUE(&ctx->ifc_vflr_task);
6235 }
6236 
6237 void
6238 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, const char *name)
6239 {
6240 
6241 	taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, NULL, NULL,
6242 	    name);
6243 }
6244 
6245 void
6246 iflib_config_gtask_init(void *ctx, struct grouptask *gtask, gtask_fn_t *fn,
6247 	const char *name)
6248 {
6249 
6250 	GROUPTASK_INIT(gtask, 0, fn, ctx);
6251 	taskqgroup_attach(qgroup_if_config_tqg, gtask, gtask, NULL, NULL,
6252 	    name);
6253 }
6254 
6255 void
6256 iflib_config_gtask_deinit(struct grouptask *gtask)
6257 {
6258 
6259 	taskqgroup_detach(qgroup_if_config_tqg, gtask);
6260 }
6261 
6262 void
6263 iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate)
6264 {
6265 	if_t ifp = ctx->ifc_ifp;
6266 	iflib_txq_t txq = ctx->ifc_txqs;
6267 
6268 	if_setbaudrate(ifp, baudrate);
6269 	if (baudrate >= IF_Gbps(10)) {
6270 		STATE_LOCK(ctx);
6271 		ctx->ifc_flags |= IFC_PREFETCH;
6272 		STATE_UNLOCK(ctx);
6273 	}
6274 	/* If link down, disable watchdog */
6275 	if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) {
6276 		for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++)
6277 			txq->ift_qstatus = IFLIB_QUEUE_IDLE;
6278 	}
6279 	ctx->ifc_link_state = link_state;
6280 	if_link_state_change(ifp, link_state);
6281 }
6282 
6283 static int
6284 iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq)
6285 {
6286 	int credits;
6287 #ifdef INVARIANTS
6288 	int credits_pre = txq->ift_cidx_processed;
6289 #endif
6290 
6291 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
6292 	    BUS_DMASYNC_POSTREAD);
6293 	if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, true)) == 0)
6294 		return (0);
6295 
6296 	txq->ift_processed += credits;
6297 	txq->ift_cidx_processed += credits;
6298 
6299 	MPASS(credits_pre + credits == txq->ift_cidx_processed);
6300 	if (txq->ift_cidx_processed >= txq->ift_size)
6301 		txq->ift_cidx_processed -= txq->ift_size;
6302 	return (credits);
6303 }
6304 
6305 static int
6306 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget)
6307 {
6308 	iflib_fl_t fl;
6309 	u_int i;
6310 
6311 	for (i = 0, fl = &rxq->ifr_fl[0]; i < rxq->ifr_nfl; i++, fl++)
6312 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
6313 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
6314 	return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx,
6315 	    budget));
6316 }
6317 
6318 void
6319 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name,
6320 	const char *description, if_int_delay_info_t info,
6321 	int offset, int value)
6322 {
6323 	info->iidi_ctx = ctx;
6324 	info->iidi_offset = offset;
6325 	info->iidi_value = value;
6326 	SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev),
6327 	    SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)),
6328 	    OID_AUTO, name, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE,
6329 	    info, 0, iflib_sysctl_int_delay, "I", description);
6330 }
6331 
6332 struct sx *
6333 iflib_ctx_lock_get(if_ctx_t ctx)
6334 {
6335 
6336 	return (&ctx->ifc_ctx_sx);
6337 }
6338 
6339 static int
6340 iflib_msix_init(if_ctx_t ctx)
6341 {
6342 	device_t dev = ctx->ifc_dev;
6343 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6344 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6345 	int admincnt, bar, err, iflib_num_rx_queues, iflib_num_tx_queues;
6346 	int msgs, queuemsgs, queues, rx_queues, tx_queues, vectors;
6347 
6348 	iflib_num_tx_queues = ctx->ifc_sysctl_ntxqs;
6349 	iflib_num_rx_queues = ctx->ifc_sysctl_nrxqs;
6350 
6351 	if (bootverbose)
6352 		device_printf(dev, "msix_init qsets capped at %d\n",
6353 		    imax(scctx->isc_ntxqsets, scctx->isc_nrxqsets));
6354 
6355 	/* Override by tuneable */
6356 	if (scctx->isc_disable_msix)
6357 		goto msi;
6358 
6359 	/* First try MSI-X */
6360 	if ((msgs = pci_msix_count(dev)) == 0) {
6361 		if (bootverbose)
6362 			device_printf(dev, "MSI-X not supported or disabled\n");
6363 		goto msi;
6364 	}
6365 
6366 	bar = ctx->ifc_softc_ctx.isc_msix_bar;
6367 	/*
6368 	 * bar == -1 => "trust me I know what I'm doing"
6369 	 * Some drivers are for hardware that is so shoddily
6370 	 * documented that no one knows which bars are which
6371 	 * so the developer has to map all bars. This hack
6372 	 * allows shoddy garbage to use MSI-X in this framework.
6373 	 */
6374 	if (bar != -1) {
6375 		ctx->ifc_msix_mem = bus_alloc_resource_any(dev,
6376 	            SYS_RES_MEMORY, &bar, RF_ACTIVE);
6377 		if (ctx->ifc_msix_mem == NULL) {
6378 			device_printf(dev, "Unable to map MSI-X table\n");
6379 			goto msi;
6380 		}
6381 	}
6382 
6383 	admincnt = sctx->isc_admin_intrcnt;
6384 #if IFLIB_DEBUG
6385 	/* use only 1 qset in debug mode */
6386 	queuemsgs = min(msgs - admincnt, 1);
6387 #else
6388 	queuemsgs = msgs - admincnt;
6389 #endif
6390 #ifdef RSS
6391 	queues = imin(queuemsgs, rss_getnumbuckets());
6392 #else
6393 	queues = queuemsgs;
6394 #endif
6395 	queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues);
6396 	if (bootverbose)
6397 		device_printf(dev,
6398 		    "intr CPUs: %d queue msgs: %d admincnt: %d\n",
6399 		    CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt);
6400 #ifdef  RSS
6401 	/* If we're doing RSS, clamp at the number of RSS buckets */
6402 	if (queues > rss_getnumbuckets())
6403 		queues = rss_getnumbuckets();
6404 #endif
6405 	if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt)
6406 		rx_queues = iflib_num_rx_queues;
6407 	else
6408 		rx_queues = queues;
6409 
6410 	if (rx_queues > scctx->isc_nrxqsets)
6411 		rx_queues = scctx->isc_nrxqsets;
6412 
6413 	/*
6414 	 * We want this to be all logical CPUs by default
6415 	 */
6416 	if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues)
6417 		tx_queues = iflib_num_tx_queues;
6418 	else
6419 		tx_queues = mp_ncpus;
6420 
6421 	if (tx_queues > scctx->isc_ntxqsets)
6422 		tx_queues = scctx->isc_ntxqsets;
6423 
6424 	if (ctx->ifc_sysctl_qs_eq_override == 0) {
6425 #ifdef INVARIANTS
6426 		if (tx_queues != rx_queues)
6427 			device_printf(dev,
6428 			    "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n",
6429 			    min(rx_queues, tx_queues), min(rx_queues, tx_queues));
6430 #endif
6431 		tx_queues = min(rx_queues, tx_queues);
6432 		rx_queues = min(rx_queues, tx_queues);
6433 	}
6434 
6435 	vectors = rx_queues + admincnt;
6436 	if (msgs < vectors) {
6437 		device_printf(dev,
6438 		    "insufficient number of MSI-X vectors "
6439 		    "(supported %d, need %d)\n", msgs, vectors);
6440 		goto msi;
6441 	}
6442 
6443 	device_printf(dev, "Using %d RX queues %d TX queues\n", rx_queues,
6444 	    tx_queues);
6445 	msgs = vectors;
6446 	if ((err = pci_alloc_msix(dev, &vectors)) == 0) {
6447 		if (vectors != msgs) {
6448 			device_printf(dev,
6449 			    "Unable to allocate sufficient MSI-X vectors "
6450 			    "(got %d, need %d)\n", vectors, msgs);
6451 			pci_release_msi(dev);
6452 			if (bar != -1) {
6453 				bus_release_resource(dev, SYS_RES_MEMORY, bar,
6454 				    ctx->ifc_msix_mem);
6455 				ctx->ifc_msix_mem = NULL;
6456 			}
6457 			goto msi;
6458 		}
6459 		device_printf(dev, "Using MSI-X interrupts with %d vectors\n",
6460 		    vectors);
6461 		scctx->isc_vectors = vectors;
6462 		scctx->isc_nrxqsets = rx_queues;
6463 		scctx->isc_ntxqsets = tx_queues;
6464 		scctx->isc_intr = IFLIB_INTR_MSIX;
6465 
6466 		return (vectors);
6467 	} else {
6468 		device_printf(dev,
6469 		    "failed to allocate %d MSI-X vectors, err: %d\n", vectors,
6470 		    err);
6471 		if (bar != -1) {
6472 			bus_release_resource(dev, SYS_RES_MEMORY, bar,
6473 			    ctx->ifc_msix_mem);
6474 			ctx->ifc_msix_mem = NULL;
6475 		}
6476 	}
6477 
6478 msi:
6479 	vectors = pci_msi_count(dev);
6480 	scctx->isc_nrxqsets = 1;
6481 	scctx->isc_ntxqsets = 1;
6482 	scctx->isc_vectors = vectors;
6483 	if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) {
6484 		device_printf(dev,"Using an MSI interrupt\n");
6485 		scctx->isc_intr = IFLIB_INTR_MSI;
6486 	} else {
6487 		scctx->isc_vectors = 1;
6488 		device_printf(dev,"Using a Legacy interrupt\n");
6489 		scctx->isc_intr = IFLIB_INTR_LEGACY;
6490 	}
6491 
6492 	return (vectors);
6493 }
6494 
6495 static const char *ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" };
6496 
6497 static int
6498 mp_ring_state_handler(SYSCTL_HANDLER_ARGS)
6499 {
6500 	int rc;
6501 	uint16_t *state = ((uint16_t *)oidp->oid_arg1);
6502 	struct sbuf *sb;
6503 	const char *ring_state = "UNKNOWN";
6504 
6505 	/* XXX needed ? */
6506 	rc = sysctl_wire_old_buffer(req, 0);
6507 	MPASS(rc == 0);
6508 	if (rc != 0)
6509 		return (rc);
6510 	sb = sbuf_new_for_sysctl(NULL, NULL, 80, req);
6511 	MPASS(sb != NULL);
6512 	if (sb == NULL)
6513 		return (ENOMEM);
6514 	if (state[3] <= 3)
6515 		ring_state = ring_states[state[3]];
6516 
6517 	sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s",
6518 		    state[0], state[1], state[2], ring_state);
6519 	rc = sbuf_finish(sb);
6520 	sbuf_delete(sb);
6521         return(rc);
6522 }
6523 
6524 enum iflib_ndesc_handler {
6525 	IFLIB_NTXD_HANDLER,
6526 	IFLIB_NRXD_HANDLER,
6527 };
6528 
6529 static int
6530 mp_ndesc_handler(SYSCTL_HANDLER_ARGS)
6531 {
6532 	if_ctx_t ctx = (void *)arg1;
6533 	enum iflib_ndesc_handler type = arg2;
6534 	char buf[256] = {0};
6535 	qidx_t *ndesc;
6536 	char *p, *next;
6537 	int nqs, rc, i;
6538 
6539 	nqs = 8;
6540 	switch(type) {
6541 	case IFLIB_NTXD_HANDLER:
6542 		ndesc = ctx->ifc_sysctl_ntxds;
6543 		if (ctx->ifc_sctx)
6544 			nqs = ctx->ifc_sctx->isc_ntxqs;
6545 		break;
6546 	case IFLIB_NRXD_HANDLER:
6547 		ndesc = ctx->ifc_sysctl_nrxds;
6548 		if (ctx->ifc_sctx)
6549 			nqs = ctx->ifc_sctx->isc_nrxqs;
6550 		break;
6551 	default:
6552 		printf("%s: unhandled type\n", __func__);
6553 		return (EINVAL);
6554 	}
6555 	if (nqs == 0)
6556 		nqs = 8;
6557 
6558 	for (i=0; i<8; i++) {
6559 		if (i >= nqs)
6560 			break;
6561 		if (i)
6562 			strcat(buf, ",");
6563 		sprintf(strchr(buf, 0), "%d", ndesc[i]);
6564 	}
6565 
6566 	rc = sysctl_handle_string(oidp, buf, sizeof(buf), req);
6567 	if (rc || req->newptr == NULL)
6568 		return rc;
6569 
6570 	for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p;
6571 	    i++, p = strsep(&next, " ,")) {
6572 		ndesc[i] = strtoul(p, NULL, 10);
6573 	}
6574 
6575 	return(rc);
6576 }
6577 
6578 #define NAME_BUFLEN 32
6579 static void
6580 iflib_add_device_sysctl_pre(if_ctx_t ctx)
6581 {
6582         device_t dev = iflib_get_dev(ctx);
6583 	struct sysctl_oid_list *child, *oid_list;
6584 	struct sysctl_ctx_list *ctx_list;
6585 	struct sysctl_oid *node;
6586 
6587 	ctx_list = device_get_sysctl_ctx(dev);
6588 	child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev));
6589 	ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, "iflib",
6590 	    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "IFLIB fields");
6591 	oid_list = SYSCTL_CHILDREN(node);
6592 
6593 	SYSCTL_ADD_CONST_STRING(ctx_list, oid_list, OID_AUTO, "driver_version",
6594 		       CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version,
6595 		       "driver version");
6596 
6597 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs",
6598 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0,
6599 			"# of txqs to use, 0 => use default #");
6600 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs",
6601 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0,
6602 			"# of rxqs to use, 0 => use default #");
6603 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_qs_enable",
6604 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0,
6605                        "permit #txq != #rxq");
6606 	SYSCTL_ADD_INT(ctx_list, oid_list, OID_AUTO, "disable_msix",
6607                       CTLFLAG_RWTUN, &ctx->ifc_softc_ctx.isc_disable_msix, 0,
6608                       "disable MSI-X (default 0)");
6609 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "rx_budget",
6610 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_rx_budget, 0,
6611 		       "set the RX budget");
6612 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "tx_abdicate",
6613 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_tx_abdicate, 0,
6614 		       "cause TX to abdicate instead of running to completion");
6615 	ctx->ifc_sysctl_core_offset = CORE_OFFSET_UNSPECIFIED;
6616 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "core_offset",
6617 		       CTLFLAG_RDTUN, &ctx->ifc_sysctl_core_offset, 0,
6618 		       "offset to start using cores at");
6619 	SYSCTL_ADD_U8(ctx_list, oid_list, OID_AUTO, "separate_txrx",
6620 		       CTLFLAG_RDTUN, &ctx->ifc_sysctl_separate_txrx, 0,
6621 		       "use separate cores for TX and RX");
6622 
6623 	/* XXX change for per-queue sizes */
6624 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_ntxds",
6625 	    CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NEEDGIANT, ctx,
6626 	    IFLIB_NTXD_HANDLER, mp_ndesc_handler, "A",
6627 	    "list of # of TX descriptors to use, 0 = use default #");
6628 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_nrxds",
6629 	    CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NEEDGIANT, ctx,
6630 	    IFLIB_NRXD_HANDLER, mp_ndesc_handler, "A",
6631 	    "list of # of RX descriptors to use, 0 = use default #");
6632 }
6633 
6634 static void
6635 iflib_add_device_sysctl_post(if_ctx_t ctx)
6636 {
6637 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6638 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6639         device_t dev = iflib_get_dev(ctx);
6640 	struct sysctl_oid_list *child;
6641 	struct sysctl_ctx_list *ctx_list;
6642 	iflib_fl_t fl;
6643 	iflib_txq_t txq;
6644 	iflib_rxq_t rxq;
6645 	int i, j;
6646 	char namebuf[NAME_BUFLEN];
6647 	char *qfmt;
6648 	struct sysctl_oid *queue_node, *fl_node, *node;
6649 	struct sysctl_oid_list *queue_list, *fl_list;
6650 	ctx_list = device_get_sysctl_ctx(dev);
6651 
6652 	node = ctx->ifc_sysctl_node;
6653 	child = SYSCTL_CHILDREN(node);
6654 
6655 	if (scctx->isc_ntxqsets > 100)
6656 		qfmt = "txq%03d";
6657 	else if (scctx->isc_ntxqsets > 10)
6658 		qfmt = "txq%02d";
6659 	else
6660 		qfmt = "txq%d";
6661 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
6662 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6663 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6664 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "Queue Name");
6665 		queue_list = SYSCTL_CHILDREN(queue_node);
6666 #if MEMORY_LOGGING
6667 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued",
6668 				CTLFLAG_RD,
6669 				&txq->ift_dequeued, "total mbufs freed");
6670 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued",
6671 				CTLFLAG_RD,
6672 				&txq->ift_enqueued, "total mbufs enqueued");
6673 #endif
6674 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag",
6675 				   CTLFLAG_RD,
6676 				   &txq->ift_mbuf_defrag, "# of times m_defrag was called");
6677 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "m_pullups",
6678 				   CTLFLAG_RD,
6679 				   &txq->ift_pullups, "# of times m_pullup was called");
6680 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag_failed",
6681 				   CTLFLAG_RD,
6682 				   &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed");
6683 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_desc_avail",
6684 				   CTLFLAG_RD,
6685 				   &txq->ift_no_desc_avail, "# of times no descriptors were available");
6686 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "tx_map_failed",
6687 				   CTLFLAG_RD,
6688 				   &txq->ift_map_failed, "# of times DMA map failed");
6689 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txd_encap_efbig",
6690 				   CTLFLAG_RD,
6691 				   &txq->ift_txd_encap_efbig, "# of times txd_encap returned EFBIG");
6692 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_tx_dma_setup",
6693 				   CTLFLAG_RD,
6694 				   &txq->ift_no_tx_dma_setup, "# of times map failed for other than EFBIG");
6695 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx",
6696 				   CTLFLAG_RD,
6697 				   &txq->ift_pidx, 1, "Producer Index");
6698 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx",
6699 				   CTLFLAG_RD,
6700 				   &txq->ift_cidx, 1, "Consumer Index");
6701 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx_processed",
6702 				   CTLFLAG_RD,
6703 				   &txq->ift_cidx_processed, 1, "Consumer Index seen by credit update");
6704 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use",
6705 				   CTLFLAG_RD,
6706 				   &txq->ift_in_use, 1, "descriptors in use");
6707 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_processed",
6708 				   CTLFLAG_RD,
6709 				   &txq->ift_processed, "descriptors procesed for clean");
6710 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned",
6711 				   CTLFLAG_RD,
6712 				   &txq->ift_cleaned, "total cleaned");
6713 		SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state",
6714 		    CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_NEEDGIANT,
6715 		    __DEVOLATILE(uint64_t *, &txq->ift_br->state), 0,
6716 		    mp_ring_state_handler, "A", "soft ring state");
6717 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_enqueues",
6718 				       CTLFLAG_RD, &txq->ift_br->enqueues,
6719 				       "# of enqueues to the mp_ring for this queue");
6720 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_drops",
6721 				       CTLFLAG_RD, &txq->ift_br->drops,
6722 				       "# of drops in the mp_ring for this queue");
6723 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_starts",
6724 				       CTLFLAG_RD, &txq->ift_br->starts,
6725 				       "# of normal consumer starts in the mp_ring for this queue");
6726 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_stalls",
6727 				       CTLFLAG_RD, &txq->ift_br->stalls,
6728 					       "# of consumer stalls in the mp_ring for this queue");
6729 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_restarts",
6730 			       CTLFLAG_RD, &txq->ift_br->restarts,
6731 				       "# of consumer restarts in the mp_ring for this queue");
6732 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_abdications",
6733 				       CTLFLAG_RD, &txq->ift_br->abdications,
6734 				       "# of consumer abdications in the mp_ring for this queue");
6735 	}
6736 
6737 	if (scctx->isc_nrxqsets > 100)
6738 		qfmt = "rxq%03d";
6739 	else if (scctx->isc_nrxqsets > 10)
6740 		qfmt = "rxq%02d";
6741 	else
6742 		qfmt = "rxq%d";
6743 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
6744 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6745 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6746 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "Queue Name");
6747 		queue_list = SYSCTL_CHILDREN(queue_node);
6748 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
6749 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_cidx",
6750 				       CTLFLAG_RD,
6751 				       &rxq->ifr_cq_cidx, 1, "Consumer Index");
6752 		}
6753 
6754 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
6755 			snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j);
6756 			fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list, OID_AUTO, namebuf,
6757 			    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "freelist Name");
6758 			fl_list = SYSCTL_CHILDREN(fl_node);
6759 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx",
6760 				       CTLFLAG_RD,
6761 				       &fl->ifl_pidx, 1, "Producer Index");
6762 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx",
6763 				       CTLFLAG_RD,
6764 				       &fl->ifl_cidx, 1, "Consumer Index");
6765 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits",
6766 				       CTLFLAG_RD,
6767 				       &fl->ifl_credits, 1, "credits available");
6768 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "buf_size",
6769 				       CTLFLAG_RD,
6770 				       &fl->ifl_buf_size, 1, "buffer size");
6771 #if MEMORY_LOGGING
6772 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_enqueued",
6773 					CTLFLAG_RD,
6774 					&fl->ifl_m_enqueued, "mbufs allocated");
6775 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_dequeued",
6776 					CTLFLAG_RD,
6777 					&fl->ifl_m_dequeued, "mbufs freed");
6778 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_enqueued",
6779 					CTLFLAG_RD,
6780 					&fl->ifl_cl_enqueued, "clusters allocated");
6781 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_dequeued",
6782 					CTLFLAG_RD,
6783 					&fl->ifl_cl_dequeued, "clusters freed");
6784 #endif
6785 		}
6786 	}
6787 
6788 }
6789 
6790 void
6791 iflib_request_reset(if_ctx_t ctx)
6792 {
6793 
6794 	STATE_LOCK(ctx);
6795 	ctx->ifc_flags |= IFC_DO_RESET;
6796 	STATE_UNLOCK(ctx);
6797 }
6798 
6799 #ifndef __NO_STRICT_ALIGNMENT
6800 static struct mbuf *
6801 iflib_fixup_rx(struct mbuf *m)
6802 {
6803 	struct mbuf *n;
6804 
6805 	if (m->m_len <= (MCLBYTES - ETHER_HDR_LEN)) {
6806 		bcopy(m->m_data, m->m_data + ETHER_HDR_LEN, m->m_len);
6807 		m->m_data += ETHER_HDR_LEN;
6808 		n = m;
6809 	} else {
6810 		MGETHDR(n, M_NOWAIT, MT_DATA);
6811 		if (n == NULL) {
6812 			m_freem(m);
6813 			return (NULL);
6814 		}
6815 		bcopy(m->m_data, n->m_data, ETHER_HDR_LEN);
6816 		m->m_data += ETHER_HDR_LEN;
6817 		m->m_len -= ETHER_HDR_LEN;
6818 		n->m_len = ETHER_HDR_LEN;
6819 		M_MOVE_PKTHDR(n, m);
6820 		n->m_next = m;
6821 	}
6822 	return (n);
6823 }
6824 #endif
6825 
6826 #ifdef DEBUGNET
6827 static void
6828 iflib_debugnet_init(if_t ifp, int *nrxr, int *ncl, int *clsize)
6829 {
6830 	if_ctx_t ctx;
6831 
6832 	ctx = if_getsoftc(ifp);
6833 	CTX_LOCK(ctx);
6834 	*nrxr = NRXQSETS(ctx);
6835 	*ncl = ctx->ifc_rxqs[0].ifr_fl->ifl_size;
6836 	*clsize = ctx->ifc_rxqs[0].ifr_fl->ifl_buf_size;
6837 	CTX_UNLOCK(ctx);
6838 }
6839 
6840 static void
6841 iflib_debugnet_event(if_t ifp, enum debugnet_ev event)
6842 {
6843 	if_ctx_t ctx;
6844 	if_softc_ctx_t scctx;
6845 	iflib_fl_t fl;
6846 	iflib_rxq_t rxq;
6847 	int i, j;
6848 
6849 	ctx = if_getsoftc(ifp);
6850 	scctx = &ctx->ifc_softc_ctx;
6851 
6852 	switch (event) {
6853 	case DEBUGNET_START:
6854 		for (i = 0; i < scctx->isc_nrxqsets; i++) {
6855 			rxq = &ctx->ifc_rxqs[i];
6856 			for (j = 0; j < rxq->ifr_nfl; j++) {
6857 				fl = rxq->ifr_fl;
6858 				fl->ifl_zone = m_getzone(fl->ifl_buf_size);
6859 			}
6860 		}
6861 		iflib_no_tx_batch = 1;
6862 		break;
6863 	default:
6864 		break;
6865 	}
6866 }
6867 
6868 static int
6869 iflib_debugnet_transmit(if_t ifp, struct mbuf *m)
6870 {
6871 	if_ctx_t ctx;
6872 	iflib_txq_t txq;
6873 	int error;
6874 
6875 	ctx = if_getsoftc(ifp);
6876 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
6877 	    IFF_DRV_RUNNING)
6878 		return (EBUSY);
6879 
6880 	txq = &ctx->ifc_txqs[0];
6881 	error = iflib_encap(txq, &m);
6882 	if (error == 0)
6883 		(void)iflib_txd_db_check(ctx, txq, true, txq->ift_in_use);
6884 	return (error);
6885 }
6886 
6887 static int
6888 iflib_debugnet_poll(if_t ifp, int count)
6889 {
6890 	struct epoch_tracker et;
6891 	if_ctx_t ctx;
6892 	if_softc_ctx_t scctx;
6893 	iflib_txq_t txq;
6894 	int i;
6895 
6896 	ctx = if_getsoftc(ifp);
6897 	scctx = &ctx->ifc_softc_ctx;
6898 
6899 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
6900 	    IFF_DRV_RUNNING)
6901 		return (EBUSY);
6902 
6903 	txq = &ctx->ifc_txqs[0];
6904 	(void)iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
6905 
6906 	NET_EPOCH_ENTER(et);
6907 	for (i = 0; i < scctx->isc_nrxqsets; i++)
6908 		(void)iflib_rxeof(&ctx->ifc_rxqs[i], 16 /* XXX */);
6909 	NET_EPOCH_EXIT(et);
6910 	return (0);
6911 }
6912 #endif /* DEBUGNET */
6913