xref: /freebsd/sys/net/iflib.c (revision bce7ee9d412b6410e6d799c4a417617cbb148e09)
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(&txq->ift_netmap_timer,
1070 			    NETMAP_TX_TIMER_US * SBT_1US, SBT_1US,
1071 			    iflib_netmap_timer, txq, txq->ift_netmap_timer.c_cpu);
1072 		}
1073 	return (0);
1074 }
1075 
1076 /*
1077  * Reconcile kernel and user view of the receive ring.
1078  * Same as for the txsync, this routine must be efficient.
1079  * The caller guarantees a single invocations, but races against
1080  * the rest of the driver should be handled here.
1081  *
1082  * On call, kring->rhead is the first packet that userspace wants
1083  * to keep, and kring->rcur is the wakeup point.
1084  * The kernel has previously reported packets up to kring->rtail.
1085  *
1086  * If (flags & NAF_FORCE_READ) also check for incoming packets irrespective
1087  * of whether or not we received an interrupt.
1088  */
1089 static int
1090 iflib_netmap_rxsync(struct netmap_kring *kring, int flags)
1091 {
1092 	struct netmap_adapter *na = kring->na;
1093 	struct netmap_ring *ring = kring->ring;
1094 	if_t ifp = na->ifp;
1095 	uint32_t nm_i;	/* index into the netmap ring */
1096 	uint32_t nic_i;	/* index into the NIC ring */
1097 	u_int n;
1098 	u_int const lim = kring->nkr_num_slots - 1;
1099 	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
1100 
1101 	if_ctx_t ctx = ifp->if_softc;
1102 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1103 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1104 	iflib_rxq_t rxq = &ctx->ifc_rxqs[kring->ring_id];
1105 	iflib_fl_t fl = &rxq->ifr_fl[0];
1106 	struct if_rxd_info ri;
1107 	qidx_t *cidxp;
1108 
1109 	/*
1110 	 * netmap only uses free list 0, to avoid out of order consumption
1111 	 * of receive buffers
1112 	 */
1113 
1114 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
1115 	    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1116 
1117 	/*
1118 	 * First part: import newly received packets.
1119 	 *
1120 	 * nm_i is the index of the next free slot in the netmap ring,
1121 	 * nic_i is the index of the next received packet in the NIC ring
1122 	 * (or in the free list 0 if IFLIB_HAS_RXCQ is set), and they may
1123 	 * differ in case if_init() has been called while
1124 	 * in netmap mode. For the receive ring we have
1125 	 *
1126 	 *	nic_i = fl->ifl_cidx;
1127 	 *	nm_i = kring->nr_hwtail (previous)
1128 	 * and
1129 	 *	nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1130 	 *
1131 	 * fl->ifl_cidx is set to 0 on a ring reinit
1132 	 */
1133 	if (netmap_no_pendintr || force_update) {
1134 		uint32_t hwtail_lim = nm_prev(kring->nr_hwcur, lim);
1135 		bool have_rxcq = sctx->isc_flags & IFLIB_HAS_RXCQ;
1136 		int crclen = iflib_crcstrip ? 0 : 4;
1137 		int error, avail;
1138 
1139 		/*
1140 		 * For the free list consumer index, we use the same
1141 		 * logic as in iflib_rxeof().
1142 		 */
1143 		if (have_rxcq)
1144 			cidxp = &rxq->ifr_cq_cidx;
1145 		else
1146 			cidxp = &fl->ifl_cidx;
1147 		avail = ctx->isc_rxd_available(ctx->ifc_softc,
1148 		    rxq->ifr_id, *cidxp, USHRT_MAX);
1149 
1150 		nic_i = fl->ifl_cidx;
1151 		nm_i = netmap_idx_n2k(kring, nic_i);
1152 		MPASS(nm_i == kring->nr_hwtail);
1153 		for (n = 0; avail > 0 && nm_i != hwtail_lim; n++, avail--) {
1154 			rxd_info_zero(&ri);
1155 			ri.iri_frags = rxq->ifr_frags;
1156 			ri.iri_qsidx = kring->ring_id;
1157 			ri.iri_ifp = ctx->ifc_ifp;
1158 			ri.iri_cidx = *cidxp;
1159 
1160 			error = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
1161 			ring->slot[nm_i].len = error ? 0 : ri.iri_len - crclen;
1162 			ring->slot[nm_i].flags = 0;
1163 			if (have_rxcq) {
1164 				*cidxp = ri.iri_cidx;
1165 				while (*cidxp >= scctx->isc_nrxd[0])
1166 					*cidxp -= scctx->isc_nrxd[0];
1167 			}
1168 			bus_dmamap_sync(fl->ifl_buf_tag,
1169 			    fl->ifl_sds.ifsd_map[nic_i], BUS_DMASYNC_POSTREAD);
1170 			nm_i = nm_next(nm_i, lim);
1171 			fl->ifl_cidx = nic_i = nm_next(nic_i, lim);
1172 		}
1173 		if (n) { /* update the state variables */
1174 			if (netmap_no_pendintr && !force_update) {
1175 				/* diagnostics */
1176 				iflib_rx_miss ++;
1177 				iflib_rx_miss_bufs += n;
1178 			}
1179 			kring->nr_hwtail = nm_i;
1180 		}
1181 		kring->nr_kflags &= ~NKR_PENDINTR;
1182 	}
1183 	/*
1184 	 * Second part: skip past packets that userspace has released.
1185 	 * (kring->nr_hwcur to head excluded),
1186 	 * and make the buffers available for reception.
1187 	 * As usual nm_i is the index in the netmap ring,
1188 	 * nic_i is the index in the NIC ring, and
1189 	 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1190 	 */
1191 	netmap_fl_refill(rxq, kring, false);
1192 
1193 	return (0);
1194 }
1195 
1196 static void
1197 iflib_netmap_intr(struct netmap_adapter *na, int onoff)
1198 {
1199 	if_ctx_t ctx = na->ifp->if_softc;
1200 
1201 	CTX_LOCK(ctx);
1202 	if (onoff) {
1203 		IFDI_INTR_ENABLE(ctx);
1204 	} else {
1205 		IFDI_INTR_DISABLE(ctx);
1206 	}
1207 	CTX_UNLOCK(ctx);
1208 }
1209 
1210 static int
1211 iflib_netmap_attach(if_ctx_t ctx)
1212 {
1213 	struct netmap_adapter na;
1214 
1215 	bzero(&na, sizeof(na));
1216 
1217 	na.ifp = ctx->ifc_ifp;
1218 	na.na_flags = NAF_BDG_MAYSLEEP;
1219 	MPASS(ctx->ifc_softc_ctx.isc_ntxqsets);
1220 	MPASS(ctx->ifc_softc_ctx.isc_nrxqsets);
1221 
1222 	na.num_tx_desc = iflib_num_tx_descs(ctx);
1223 	na.num_rx_desc = iflib_num_rx_descs(ctx);
1224 	na.nm_txsync = iflib_netmap_txsync;
1225 	na.nm_rxsync = iflib_netmap_rxsync;
1226 	na.nm_register = iflib_netmap_register;
1227 	na.nm_intr = iflib_netmap_intr;
1228 	na.num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
1229 	na.num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
1230 	return (netmap_attach(&na));
1231 }
1232 
1233 static int
1234 iflib_netmap_txq_init(if_ctx_t ctx, iflib_txq_t txq)
1235 {
1236 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1237 	struct netmap_slot *slot;
1238 
1239 	slot = netmap_reset(na, NR_TX, txq->ift_id, 0);
1240 	if (slot == NULL)
1241 		return (0);
1242 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxd[0]; i++) {
1243 		/*
1244 		 * In netmap mode, set the map for the packet buffer.
1245 		 * NOTE: Some drivers (not this one) also need to set
1246 		 * the physical buffer address in the NIC ring.
1247 		 * netmap_idx_n2k() maps a nic index, i, into the corresponding
1248 		 * netmap slot index, si
1249 		 */
1250 		int si = netmap_idx_n2k(na->tx_rings[txq->ift_id], i);
1251 		netmap_load_map(na, txq->ift_buf_tag, txq->ift_sds.ifsd_map[i],
1252 		    NMB(na, slot + si));
1253 	}
1254 	return (1);
1255 }
1256 
1257 static int
1258 iflib_netmap_rxq_init(if_ctx_t ctx, iflib_rxq_t rxq)
1259 {
1260 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1261 	struct netmap_kring *kring;
1262 	struct netmap_slot *slot;
1263 
1264 	slot = netmap_reset(na, NR_RX, rxq->ifr_id, 0);
1265 	if (slot == NULL)
1266 		return (0);
1267 	kring = na->rx_rings[rxq->ifr_id];
1268 	netmap_fl_refill(rxq, kring, true);
1269 	return (1);
1270 }
1271 
1272 static void
1273 iflib_netmap_timer(void *arg)
1274 {
1275 	iflib_txq_t txq = arg;
1276 	if_ctx_t ctx = txq->ift_ctx;
1277 
1278 	/*
1279 	 * Wake up the netmap application, to give it a chance to
1280 	 * call txsync and reclaim more completed TX buffers.
1281 	 */
1282 	netmap_tx_irq(ctx->ifc_ifp, txq->ift_id);
1283 }
1284 
1285 #define iflib_netmap_detach(ifp) netmap_detach(ifp)
1286 
1287 #else
1288 #define iflib_netmap_txq_init(ctx, txq) (0)
1289 #define iflib_netmap_rxq_init(ctx, rxq) (0)
1290 #define iflib_netmap_detach(ifp)
1291 
1292 #define iflib_netmap_attach(ctx) (0)
1293 #define netmap_rx_irq(ifp, qid, budget) (0)
1294 #endif
1295 
1296 #if defined(__i386__) || defined(__amd64__)
1297 static __inline void
1298 prefetch(void *x)
1299 {
1300 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1301 }
1302 static __inline void
1303 prefetch2cachelines(void *x)
1304 {
1305 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1306 #if (CACHE_LINE_SIZE < 128)
1307 	__asm volatile("prefetcht0 %0" :: "m" (*(((unsigned long *)x)+CACHE_LINE_SIZE/(sizeof(unsigned long)))));
1308 #endif
1309 }
1310 #else
1311 #define prefetch(x)
1312 #define prefetch2cachelines(x)
1313 #endif
1314 
1315 static void
1316 iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid)
1317 {
1318 	iflib_fl_t fl;
1319 
1320 	fl = &rxq->ifr_fl[flid];
1321 	iru->iru_paddrs = fl->ifl_bus_addrs;
1322 	iru->iru_idxs = fl->ifl_rxd_idxs;
1323 	iru->iru_qsidx = rxq->ifr_id;
1324 	iru->iru_buf_size = fl->ifl_buf_size;
1325 	iru->iru_flidx = fl->ifl_id;
1326 }
1327 
1328 static void
1329 _iflib_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int err)
1330 {
1331 	if (err)
1332 		return;
1333 	*(bus_addr_t *) arg = segs[0].ds_addr;
1334 }
1335 
1336 int
1337 iflib_dma_alloc_align(if_ctx_t ctx, int size, int align, iflib_dma_info_t dma, int mapflags)
1338 {
1339 	int err;
1340 	device_t dev = ctx->ifc_dev;
1341 
1342 	err = bus_dma_tag_create(bus_get_dma_tag(dev),	/* parent */
1343 				align, 0,		/* alignment, bounds */
1344 				BUS_SPACE_MAXADDR,	/* lowaddr */
1345 				BUS_SPACE_MAXADDR,	/* highaddr */
1346 				NULL, NULL,		/* filter, filterarg */
1347 				size,			/* maxsize */
1348 				1,			/* nsegments */
1349 				size,			/* maxsegsize */
1350 				BUS_DMA_ALLOCNOW,	/* flags */
1351 				NULL,			/* lockfunc */
1352 				NULL,			/* lockarg */
1353 				&dma->idi_tag);
1354 	if (err) {
1355 		device_printf(dev,
1356 		    "%s: bus_dma_tag_create failed: %d\n",
1357 		    __func__, err);
1358 		goto fail_0;
1359 	}
1360 
1361 	err = bus_dmamem_alloc(dma->idi_tag, (void**) &dma->idi_vaddr,
1362 	    BUS_DMA_NOWAIT | BUS_DMA_COHERENT | BUS_DMA_ZERO, &dma->idi_map);
1363 	if (err) {
1364 		device_printf(dev,
1365 		    "%s: bus_dmamem_alloc(%ju) failed: %d\n",
1366 		    __func__, (uintmax_t)size, err);
1367 		goto fail_1;
1368 	}
1369 
1370 	dma->idi_paddr = IF_BAD_DMA;
1371 	err = bus_dmamap_load(dma->idi_tag, dma->idi_map, dma->idi_vaddr,
1372 	    size, _iflib_dmamap_cb, &dma->idi_paddr, mapflags | BUS_DMA_NOWAIT);
1373 	if (err || dma->idi_paddr == IF_BAD_DMA) {
1374 		device_printf(dev,
1375 		    "%s: bus_dmamap_load failed: %d\n",
1376 		    __func__, err);
1377 		goto fail_2;
1378 	}
1379 
1380 	dma->idi_size = size;
1381 	return (0);
1382 
1383 fail_2:
1384 	bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1385 fail_1:
1386 	bus_dma_tag_destroy(dma->idi_tag);
1387 fail_0:
1388 	dma->idi_tag = NULL;
1389 
1390 	return (err);
1391 }
1392 
1393 int
1394 iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags)
1395 {
1396 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1397 
1398 	KASSERT(sctx->isc_q_align != 0, ("alignment value not initialized"));
1399 
1400 	return (iflib_dma_alloc_align(ctx, size, sctx->isc_q_align, dma, mapflags));
1401 }
1402 
1403 int
1404 iflib_dma_alloc_multi(if_ctx_t ctx, int *sizes, iflib_dma_info_t *dmalist, int mapflags, int count)
1405 {
1406 	int i, err;
1407 	iflib_dma_info_t *dmaiter;
1408 
1409 	dmaiter = dmalist;
1410 	for (i = 0; i < count; i++, dmaiter++) {
1411 		if ((err = iflib_dma_alloc(ctx, sizes[i], *dmaiter, mapflags)) != 0)
1412 			break;
1413 	}
1414 	if (err)
1415 		iflib_dma_free_multi(dmalist, i);
1416 	return (err);
1417 }
1418 
1419 void
1420 iflib_dma_free(iflib_dma_info_t dma)
1421 {
1422 	if (dma->idi_tag == NULL)
1423 		return;
1424 	if (dma->idi_paddr != IF_BAD_DMA) {
1425 		bus_dmamap_sync(dma->idi_tag, dma->idi_map,
1426 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1427 		bus_dmamap_unload(dma->idi_tag, dma->idi_map);
1428 		dma->idi_paddr = IF_BAD_DMA;
1429 	}
1430 	if (dma->idi_vaddr != NULL) {
1431 		bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1432 		dma->idi_vaddr = NULL;
1433 	}
1434 	bus_dma_tag_destroy(dma->idi_tag);
1435 	dma->idi_tag = NULL;
1436 }
1437 
1438 void
1439 iflib_dma_free_multi(iflib_dma_info_t *dmalist, int count)
1440 {
1441 	int i;
1442 	iflib_dma_info_t *dmaiter = dmalist;
1443 
1444 	for (i = 0; i < count; i++, dmaiter++)
1445 		iflib_dma_free(*dmaiter);
1446 }
1447 
1448 static int
1449 iflib_fast_intr(void *arg)
1450 {
1451 	iflib_filter_info_t info = arg;
1452 	struct grouptask *gtask = info->ifi_task;
1453 	int result;
1454 
1455 	DBG_COUNTER_INC(fast_intrs);
1456 	if (info->ifi_filter != NULL) {
1457 		result = info->ifi_filter(info->ifi_filter_arg);
1458 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1459 			return (result);
1460 	}
1461 
1462 	GROUPTASK_ENQUEUE(gtask);
1463 	return (FILTER_HANDLED);
1464 }
1465 
1466 static int
1467 iflib_fast_intr_rxtx(void *arg)
1468 {
1469 	iflib_filter_info_t info = arg;
1470 	struct grouptask *gtask = info->ifi_task;
1471 	if_ctx_t ctx;
1472 	iflib_rxq_t rxq = (iflib_rxq_t)info->ifi_ctx;
1473 	iflib_txq_t txq;
1474 	void *sc;
1475 	int i, cidx, result;
1476 	qidx_t txqid;
1477 	bool intr_enable, intr_legacy;
1478 
1479 	DBG_COUNTER_INC(fast_intrs);
1480 	if (info->ifi_filter != NULL) {
1481 		result = info->ifi_filter(info->ifi_filter_arg);
1482 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1483 			return (result);
1484 	}
1485 
1486 	ctx = rxq->ifr_ctx;
1487 	sc = ctx->ifc_softc;
1488 	intr_enable = false;
1489 	intr_legacy = !!(ctx->ifc_flags & IFC_LEGACY);
1490 	MPASS(rxq->ifr_ntxqirq);
1491 	for (i = 0; i < rxq->ifr_ntxqirq; i++) {
1492 		txqid = rxq->ifr_txqid[i];
1493 		txq = &ctx->ifc_txqs[txqid];
1494 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1495 		    BUS_DMASYNC_POSTREAD);
1496 		if (!ctx->isc_txd_credits_update(sc, txqid, false)) {
1497 			if (intr_legacy)
1498 				intr_enable = true;
1499 			else
1500 				IFDI_TX_QUEUE_INTR_ENABLE(ctx, txqid);
1501 			continue;
1502 		}
1503 		GROUPTASK_ENQUEUE(&txq->ift_task);
1504 	}
1505 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_RXCQ)
1506 		cidx = rxq->ifr_cq_cidx;
1507 	else
1508 		cidx = rxq->ifr_fl[0].ifl_cidx;
1509 	if (iflib_rxd_avail(ctx, rxq, cidx, 1))
1510 		GROUPTASK_ENQUEUE(gtask);
1511 	else {
1512 		if (intr_legacy)
1513 			intr_enable = true;
1514 		else
1515 			IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
1516 		DBG_COUNTER_INC(rx_intr_enables);
1517 	}
1518 	if (intr_enable)
1519 		IFDI_INTR_ENABLE(ctx);
1520 	return (FILTER_HANDLED);
1521 }
1522 
1523 static int
1524 iflib_fast_intr_ctx(void *arg)
1525 {
1526 	iflib_filter_info_t info = arg;
1527 	struct grouptask *gtask = info->ifi_task;
1528 	int result;
1529 
1530 	DBG_COUNTER_INC(fast_intrs);
1531 	if (info->ifi_filter != NULL) {
1532 		result = info->ifi_filter(info->ifi_filter_arg);
1533 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1534 			return (result);
1535 	}
1536 
1537 	GROUPTASK_ENQUEUE(gtask);
1538 	return (FILTER_HANDLED);
1539 }
1540 
1541 static int
1542 _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
1543 		 driver_filter_t filter, driver_intr_t handler, void *arg,
1544 		 const char *name)
1545 {
1546 	struct resource *res;
1547 	void *tag = NULL;
1548 	device_t dev = ctx->ifc_dev;
1549 	int flags, i, rc;
1550 
1551 	flags = RF_ACTIVE;
1552 	if (ctx->ifc_flags & IFC_LEGACY)
1553 		flags |= RF_SHAREABLE;
1554 	MPASS(rid < 512);
1555 	i = rid;
1556 	res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &i, flags);
1557 	if (res == NULL) {
1558 		device_printf(dev,
1559 		    "failed to allocate IRQ for rid %d, name %s.\n", rid, name);
1560 		return (ENOMEM);
1561 	}
1562 	irq->ii_res = res;
1563 	KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL"));
1564 	rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET,
1565 						filter, handler, arg, &tag);
1566 	if (rc != 0) {
1567 		device_printf(dev,
1568 		    "failed to setup interrupt for rid %d, name %s: %d\n",
1569 					  rid, name ? name : "unknown", rc);
1570 		return (rc);
1571 	} else if (name)
1572 		bus_describe_intr(dev, res, tag, "%s", name);
1573 
1574 	irq->ii_tag = tag;
1575 	return (0);
1576 }
1577 
1578 /*********************************************************************
1579  *
1580  *  Allocate DMA resources for TX buffers as well as memory for the TX
1581  *  mbuf map.  TX DMA maps (non-TSO/TSO) and TX mbuf map are kept in a
1582  *  iflib_sw_tx_desc_array structure, storing all the information that
1583  *  is needed to transmit a packet on the wire.  This is called only
1584  *  once at attach, setup is done every reset.
1585  *
1586  **********************************************************************/
1587 static int
1588 iflib_txsd_alloc(iflib_txq_t txq)
1589 {
1590 	if_ctx_t ctx = txq->ift_ctx;
1591 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1592 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1593 	device_t dev = ctx->ifc_dev;
1594 	bus_size_t tsomaxsize;
1595 	int err, nsegments, ntsosegments;
1596 	bool tso;
1597 
1598 	nsegments = scctx->isc_tx_nsegments;
1599 	ntsosegments = scctx->isc_tx_tso_segments_max;
1600 	tsomaxsize = scctx->isc_tx_tso_size_max;
1601 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_VLAN_MTU)
1602 		tsomaxsize += sizeof(struct ether_vlan_header);
1603 	MPASS(scctx->isc_ntxd[0] > 0);
1604 	MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0);
1605 	MPASS(nsegments > 0);
1606 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) {
1607 		MPASS(ntsosegments > 0);
1608 		MPASS(sctx->isc_tso_maxsize >= tsomaxsize);
1609 	}
1610 
1611 	/*
1612 	 * Set up DMA tags for TX buffers.
1613 	 */
1614 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1615 			       1, 0,			/* alignment, bounds */
1616 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1617 			       BUS_SPACE_MAXADDR,	/* highaddr */
1618 			       NULL, NULL,		/* filter, filterarg */
1619 			       sctx->isc_tx_maxsize,		/* maxsize */
1620 			       nsegments,	/* nsegments */
1621 			       sctx->isc_tx_maxsegsize,	/* maxsegsize */
1622 			       0,			/* flags */
1623 			       NULL,			/* lockfunc */
1624 			       NULL,			/* lockfuncarg */
1625 			       &txq->ift_buf_tag))) {
1626 		device_printf(dev,"Unable to allocate TX DMA tag: %d\n", err);
1627 		device_printf(dev,"maxsize: %ju nsegments: %d maxsegsize: %ju\n",
1628 		    (uintmax_t)sctx->isc_tx_maxsize, nsegments, (uintmax_t)sctx->isc_tx_maxsegsize);
1629 		goto fail;
1630 	}
1631 	tso = (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) != 0;
1632 	if (tso && (err = bus_dma_tag_create(bus_get_dma_tag(dev),
1633 			       1, 0,			/* alignment, bounds */
1634 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1635 			       BUS_SPACE_MAXADDR,	/* highaddr */
1636 			       NULL, NULL,		/* filter, filterarg */
1637 			       tsomaxsize,		/* maxsize */
1638 			       ntsosegments,	/* nsegments */
1639 			       sctx->isc_tso_maxsegsize,/* maxsegsize */
1640 			       0,			/* flags */
1641 			       NULL,			/* lockfunc */
1642 			       NULL,			/* lockfuncarg */
1643 			       &txq->ift_tso_buf_tag))) {
1644 		device_printf(dev, "Unable to allocate TSO TX DMA tag: %d\n",
1645 		    err);
1646 		goto fail;
1647 	}
1648 
1649 	/* Allocate memory for the TX mbuf map. */
1650 	if (!(txq->ift_sds.ifsd_m =
1651 	    (struct mbuf **) malloc(sizeof(struct mbuf *) *
1652 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1653 		device_printf(dev, "Unable to allocate TX mbuf map memory\n");
1654 		err = ENOMEM;
1655 		goto fail;
1656 	}
1657 
1658 	/*
1659 	 * Create the DMA maps for TX buffers.
1660 	 */
1661 	if ((txq->ift_sds.ifsd_map = (bus_dmamap_t *)malloc(
1662 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1663 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1664 		device_printf(dev,
1665 		    "Unable to allocate TX buffer DMA map memory\n");
1666 		err = ENOMEM;
1667 		goto fail;
1668 	}
1669 	if (tso && (txq->ift_sds.ifsd_tso_map = (bus_dmamap_t *)malloc(
1670 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1671 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1672 		device_printf(dev,
1673 		    "Unable to allocate TSO TX buffer map memory\n");
1674 		err = ENOMEM;
1675 		goto fail;
1676 	}
1677 	for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) {
1678 		err = bus_dmamap_create(txq->ift_buf_tag, 0,
1679 		    &txq->ift_sds.ifsd_map[i]);
1680 		if (err != 0) {
1681 			device_printf(dev, "Unable to create TX DMA map\n");
1682 			goto fail;
1683 		}
1684 		if (!tso)
1685 			continue;
1686 		err = bus_dmamap_create(txq->ift_tso_buf_tag, 0,
1687 		    &txq->ift_sds.ifsd_tso_map[i]);
1688 		if (err != 0) {
1689 			device_printf(dev, "Unable to create TSO TX DMA map\n");
1690 			goto fail;
1691 		}
1692 	}
1693 	return (0);
1694 fail:
1695 	/* We free all, it handles case where we are in the middle */
1696 	iflib_tx_structures_free(ctx);
1697 	return (err);
1698 }
1699 
1700 static void
1701 iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i)
1702 {
1703 	bus_dmamap_t map;
1704 
1705 	if (txq->ift_sds.ifsd_map != NULL) {
1706 		map = txq->ift_sds.ifsd_map[i];
1707 		bus_dmamap_sync(txq->ift_buf_tag, map, BUS_DMASYNC_POSTWRITE);
1708 		bus_dmamap_unload(txq->ift_buf_tag, map);
1709 		bus_dmamap_destroy(txq->ift_buf_tag, map);
1710 		txq->ift_sds.ifsd_map[i] = NULL;
1711 	}
1712 
1713 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1714 		map = txq->ift_sds.ifsd_tso_map[i];
1715 		bus_dmamap_sync(txq->ift_tso_buf_tag, map,
1716 		    BUS_DMASYNC_POSTWRITE);
1717 		bus_dmamap_unload(txq->ift_tso_buf_tag, map);
1718 		bus_dmamap_destroy(txq->ift_tso_buf_tag, map);
1719 		txq->ift_sds.ifsd_tso_map[i] = NULL;
1720 	}
1721 }
1722 
1723 static void
1724 iflib_txq_destroy(iflib_txq_t txq)
1725 {
1726 	if_ctx_t ctx = txq->ift_ctx;
1727 
1728 	for (int i = 0; i < txq->ift_size; i++)
1729 		iflib_txsd_destroy(ctx, txq, i);
1730 
1731 	if (txq->ift_br != NULL) {
1732 		ifmp_ring_free(txq->ift_br);
1733 		txq->ift_br = NULL;
1734 	}
1735 
1736 	mtx_destroy(&txq->ift_mtx);
1737 
1738 	if (txq->ift_sds.ifsd_map != NULL) {
1739 		free(txq->ift_sds.ifsd_map, M_IFLIB);
1740 		txq->ift_sds.ifsd_map = NULL;
1741 	}
1742 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1743 		free(txq->ift_sds.ifsd_tso_map, M_IFLIB);
1744 		txq->ift_sds.ifsd_tso_map = NULL;
1745 	}
1746 	if (txq->ift_sds.ifsd_m != NULL) {
1747 		free(txq->ift_sds.ifsd_m, M_IFLIB);
1748 		txq->ift_sds.ifsd_m = NULL;
1749 	}
1750 	if (txq->ift_buf_tag != NULL) {
1751 		bus_dma_tag_destroy(txq->ift_buf_tag);
1752 		txq->ift_buf_tag = NULL;
1753 	}
1754 	if (txq->ift_tso_buf_tag != NULL) {
1755 		bus_dma_tag_destroy(txq->ift_tso_buf_tag);
1756 		txq->ift_tso_buf_tag = NULL;
1757 	}
1758 	if (txq->ift_ifdi != NULL) {
1759 		free(txq->ift_ifdi, M_IFLIB);
1760 	}
1761 }
1762 
1763 static void
1764 iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i)
1765 {
1766 	struct mbuf **mp;
1767 
1768 	mp = &txq->ift_sds.ifsd_m[i];
1769 	if (*mp == NULL)
1770 		return;
1771 
1772 	if (txq->ift_sds.ifsd_map != NULL) {
1773 		bus_dmamap_sync(txq->ift_buf_tag,
1774 		    txq->ift_sds.ifsd_map[i], BUS_DMASYNC_POSTWRITE);
1775 		bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[i]);
1776 	}
1777 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1778 		bus_dmamap_sync(txq->ift_tso_buf_tag,
1779 		    txq->ift_sds.ifsd_tso_map[i], BUS_DMASYNC_POSTWRITE);
1780 		bus_dmamap_unload(txq->ift_tso_buf_tag,
1781 		    txq->ift_sds.ifsd_tso_map[i]);
1782 	}
1783 	m_free(*mp);
1784 	DBG_COUNTER_INC(tx_frees);
1785 	*mp = NULL;
1786 }
1787 
1788 static int
1789 iflib_txq_setup(iflib_txq_t txq)
1790 {
1791 	if_ctx_t ctx = txq->ift_ctx;
1792 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1793 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1794 	iflib_dma_info_t di;
1795 	int i;
1796 
1797 	/* Set number of descriptors available */
1798 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
1799 	/* XXX make configurable */
1800 	txq->ift_update_freq = IFLIB_DEFAULT_TX_UPDATE_FREQ;
1801 
1802 	/* Reset indices */
1803 	txq->ift_cidx_processed = 0;
1804 	txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0;
1805 	txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset];
1806 
1807 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1808 		bzero((void *)di->idi_vaddr, di->idi_size);
1809 
1810 	IFDI_TXQ_SETUP(ctx, txq->ift_id);
1811 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1812 		bus_dmamap_sync(di->idi_tag, di->idi_map,
1813 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1814 	return (0);
1815 }
1816 
1817 /*********************************************************************
1818  *
1819  *  Allocate DMA resources for RX buffers as well as memory for the RX
1820  *  mbuf map, direct RX cluster pointer map and RX cluster bus address
1821  *  map.  RX DMA map, RX mbuf map, direct RX cluster pointer map and
1822  *  RX cluster map are kept in a iflib_sw_rx_desc_array structure.
1823  *  Since we use use one entry in iflib_sw_rx_desc_array per received
1824  *  packet, the maximum number of entries we'll need is equal to the
1825  *  number of hardware receive descriptors that we've allocated.
1826  *
1827  **********************************************************************/
1828 static int
1829 iflib_rxsd_alloc(iflib_rxq_t rxq)
1830 {
1831 	if_ctx_t ctx = rxq->ifr_ctx;
1832 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1833 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1834 	device_t dev = ctx->ifc_dev;
1835 	iflib_fl_t fl;
1836 	int			err;
1837 
1838 	MPASS(scctx->isc_nrxd[0] > 0);
1839 	MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0);
1840 
1841 	fl = rxq->ifr_fl;
1842 	for (int i = 0; i <  rxq->ifr_nfl; i++, fl++) {
1843 		fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */
1844 		/* Set up DMA tag for RX buffers. */
1845 		err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1846 					 1, 0,			/* alignment, bounds */
1847 					 BUS_SPACE_MAXADDR,	/* lowaddr */
1848 					 BUS_SPACE_MAXADDR,	/* highaddr */
1849 					 NULL, NULL,		/* filter, filterarg */
1850 					 sctx->isc_rx_maxsize,	/* maxsize */
1851 					 sctx->isc_rx_nsegments,	/* nsegments */
1852 					 sctx->isc_rx_maxsegsize,	/* maxsegsize */
1853 					 0,			/* flags */
1854 					 NULL,			/* lockfunc */
1855 					 NULL,			/* lockarg */
1856 					 &fl->ifl_buf_tag);
1857 		if (err) {
1858 			device_printf(dev,
1859 			    "Unable to allocate RX DMA tag: %d\n", err);
1860 			goto fail;
1861 		}
1862 
1863 		/* Allocate memory for the RX mbuf map. */
1864 		if (!(fl->ifl_sds.ifsd_m =
1865 		      (struct mbuf **) malloc(sizeof(struct mbuf *) *
1866 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1867 			device_printf(dev,
1868 			    "Unable to allocate RX mbuf map memory\n");
1869 			err = ENOMEM;
1870 			goto fail;
1871 		}
1872 
1873 		/* Allocate memory for the direct RX cluster pointer map. */
1874 		if (!(fl->ifl_sds.ifsd_cl =
1875 		      (caddr_t *) malloc(sizeof(caddr_t) *
1876 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1877 			device_printf(dev,
1878 			    "Unable to allocate RX cluster map memory\n");
1879 			err = ENOMEM;
1880 			goto fail;
1881 		}
1882 
1883 		/* Allocate memory for the RX cluster bus address map. */
1884 		if (!(fl->ifl_sds.ifsd_ba =
1885 		      (bus_addr_t *) malloc(sizeof(bus_addr_t) *
1886 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1887 			device_printf(dev,
1888 			    "Unable to allocate RX bus address map memory\n");
1889 			err = ENOMEM;
1890 			goto fail;
1891 		}
1892 
1893 		/*
1894 		 * Create the DMA maps for RX buffers.
1895 		 */
1896 		if (!(fl->ifl_sds.ifsd_map =
1897 		      (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1898 			device_printf(dev,
1899 			    "Unable to allocate RX buffer DMA map memory\n");
1900 			err = ENOMEM;
1901 			goto fail;
1902 		}
1903 		for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++) {
1904 			err = bus_dmamap_create(fl->ifl_buf_tag, 0,
1905 			    &fl->ifl_sds.ifsd_map[i]);
1906 			if (err != 0) {
1907 				device_printf(dev, "Unable to create RX buffer DMA map\n");
1908 				goto fail;
1909 			}
1910 		}
1911 	}
1912 	return (0);
1913 
1914 fail:
1915 	iflib_rx_structures_free(ctx);
1916 	return (err);
1917 }
1918 
1919 /*
1920  * Internal service routines
1921  */
1922 
1923 struct rxq_refill_cb_arg {
1924 	int               error;
1925 	bus_dma_segment_t seg;
1926 	int               nseg;
1927 };
1928 
1929 static void
1930 _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
1931 {
1932 	struct rxq_refill_cb_arg *cb_arg = arg;
1933 
1934 	cb_arg->error = error;
1935 	cb_arg->seg = segs[0];
1936 	cb_arg->nseg = nseg;
1937 }
1938 
1939 /**
1940  * iflib_fl_refill - refill an rxq free-buffer list
1941  * @ctx: the iflib context
1942  * @fl: the free list to refill
1943  * @count: the number of new buffers to allocate
1944  *
1945  * (Re)populate an rxq free-buffer list with up to @count new packet buffers.
1946  * The caller must assure that @count does not exceed the queue's capacity
1947  * minus one (since we always leave a descriptor unavailable).
1948  */
1949 static uint8_t
1950 iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
1951 {
1952 	struct if_rxd_update iru;
1953 	struct rxq_refill_cb_arg cb_arg;
1954 	struct mbuf *m;
1955 	caddr_t cl, *sd_cl;
1956 	struct mbuf **sd_m;
1957 	bus_dmamap_t *sd_map;
1958 	bus_addr_t bus_addr, *sd_ba;
1959 	int err, frag_idx, i, idx, n, pidx;
1960 	qidx_t credits;
1961 
1962 	MPASS(count <= fl->ifl_size - fl->ifl_credits - 1);
1963 
1964 	sd_m = fl->ifl_sds.ifsd_m;
1965 	sd_map = fl->ifl_sds.ifsd_map;
1966 	sd_cl = fl->ifl_sds.ifsd_cl;
1967 	sd_ba = fl->ifl_sds.ifsd_ba;
1968 	pidx = fl->ifl_pidx;
1969 	idx = pidx;
1970 	frag_idx = fl->ifl_fragidx;
1971 	credits = fl->ifl_credits;
1972 
1973 	i = 0;
1974 	n = count;
1975 	MPASS(n > 0);
1976 	MPASS(credits + n <= fl->ifl_size);
1977 
1978 	if (pidx < fl->ifl_cidx)
1979 		MPASS(pidx + n <= fl->ifl_cidx);
1980 	if (pidx == fl->ifl_cidx && (credits < fl->ifl_size))
1981 		MPASS(fl->ifl_gen == 0);
1982 	if (pidx > fl->ifl_cidx)
1983 		MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx);
1984 
1985 	DBG_COUNTER_INC(fl_refills);
1986 	if (n > 8)
1987 		DBG_COUNTER_INC(fl_refills_large);
1988 	iru_init(&iru, fl->ifl_rxq, fl->ifl_id);
1989 	while (n-- > 0) {
1990 		/*
1991 		 * We allocate an uninitialized mbuf + cluster, mbuf is
1992 		 * initialized after rx.
1993 		 *
1994 		 * If the cluster is still set then we know a minimum sized
1995 		 * packet was received
1996 		 */
1997 		bit_ffc_at(fl->ifl_rx_bitmap, frag_idx, fl->ifl_size,
1998 		    &frag_idx);
1999 		if (frag_idx < 0)
2000 			bit_ffc(fl->ifl_rx_bitmap, fl->ifl_size, &frag_idx);
2001 		MPASS(frag_idx >= 0);
2002 		if ((cl = sd_cl[frag_idx]) == NULL) {
2003 			cl = uma_zalloc(fl->ifl_zone, M_NOWAIT);
2004 			if (__predict_false(cl == NULL))
2005 				break;
2006 
2007 			cb_arg.error = 0;
2008 			MPASS(sd_map != NULL);
2009 			err = bus_dmamap_load(fl->ifl_buf_tag, sd_map[frag_idx],
2010 			    cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg,
2011 			    BUS_DMA_NOWAIT);
2012 			if (__predict_false(err != 0 || cb_arg.error)) {
2013 				uma_zfree(fl->ifl_zone, cl);
2014 				break;
2015 			}
2016 
2017 			sd_ba[frag_idx] = bus_addr = cb_arg.seg.ds_addr;
2018 			sd_cl[frag_idx] = cl;
2019 #if MEMORY_LOGGING
2020 			fl->ifl_cl_enqueued++;
2021 #endif
2022 		} else {
2023 			bus_addr = sd_ba[frag_idx];
2024 		}
2025 		bus_dmamap_sync(fl->ifl_buf_tag, sd_map[frag_idx],
2026 		    BUS_DMASYNC_PREREAD);
2027 
2028 		if (sd_m[frag_idx] == NULL) {
2029 			m = m_gethdr(M_NOWAIT, MT_NOINIT);
2030 			if (__predict_false(m == NULL))
2031 				break;
2032 			sd_m[frag_idx] = m;
2033 		}
2034 		bit_set(fl->ifl_rx_bitmap, frag_idx);
2035 #if MEMORY_LOGGING
2036 		fl->ifl_m_enqueued++;
2037 #endif
2038 
2039 		DBG_COUNTER_INC(rx_allocs);
2040 		fl->ifl_rxd_idxs[i] = frag_idx;
2041 		fl->ifl_bus_addrs[i] = bus_addr;
2042 		credits++;
2043 		i++;
2044 		MPASS(credits <= fl->ifl_size);
2045 		if (++idx == fl->ifl_size) {
2046 #ifdef INVARIANTS
2047 			fl->ifl_gen = 1;
2048 #endif
2049 			idx = 0;
2050 		}
2051 		if (n == 0 || i == IFLIB_MAX_RX_REFRESH) {
2052 			iru.iru_pidx = pidx;
2053 			iru.iru_count = i;
2054 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2055 			fl->ifl_pidx = idx;
2056 			fl->ifl_credits = credits;
2057 			pidx = idx;
2058 			i = 0;
2059 		}
2060 	}
2061 
2062 	if (n < count - 1) {
2063 		if (i != 0) {
2064 			iru.iru_pidx = pidx;
2065 			iru.iru_count = i;
2066 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2067 			fl->ifl_pidx = idx;
2068 			fl->ifl_credits = credits;
2069 		}
2070 		DBG_COUNTER_INC(rxd_flush);
2071 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2072 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2073 		ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id,
2074 		    fl->ifl_id, fl->ifl_pidx);
2075 		if (__predict_true(bit_test(fl->ifl_rx_bitmap, frag_idx))) {
2076 			fl->ifl_fragidx = frag_idx + 1;
2077 			if (fl->ifl_fragidx == fl->ifl_size)
2078 				fl->ifl_fragidx = 0;
2079 		} else {
2080 			fl->ifl_fragidx = frag_idx;
2081 		}
2082 	}
2083 
2084 	return (n == -1 ? 0 : IFLIB_RXEOF_EMPTY);
2085 }
2086 
2087 static inline uint8_t
2088 iflib_fl_refill_all(if_ctx_t ctx, iflib_fl_t fl)
2089 {
2090 	/*
2091 	 * We leave an unused descriptor to avoid pidx to catch up with cidx.
2092 	 * This is important as it confuses most NICs. For instance,
2093 	 * Intel NICs have (per receive ring) RDH and RDT registers, where
2094 	 * RDH points to the next receive descriptor to be used by the NIC,
2095 	 * and RDT for the next receive descriptor to be published by the
2096 	 * driver to the NIC (RDT - 1 is thus the last valid one).
2097 	 * The condition RDH == RDT means no descriptors are available to
2098 	 * the NIC, and thus it would be ambiguous if it also meant that
2099 	 * all the descriptors are available to the NIC.
2100 	 */
2101 	int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1;
2102 #ifdef INVARIANTS
2103 	int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1;
2104 #endif
2105 
2106 	MPASS(fl->ifl_credits <= fl->ifl_size);
2107 	MPASS(reclaimable == delta);
2108 
2109 	if (reclaimable > 0)
2110 		return (iflib_fl_refill(ctx, fl, reclaimable));
2111 	return (0);
2112 }
2113 
2114 uint8_t
2115 iflib_in_detach(if_ctx_t ctx)
2116 {
2117 	bool in_detach;
2118 
2119 	STATE_LOCK(ctx);
2120 	in_detach = !!(ctx->ifc_flags & IFC_IN_DETACH);
2121 	STATE_UNLOCK(ctx);
2122 	return (in_detach);
2123 }
2124 
2125 static void
2126 iflib_fl_bufs_free(iflib_fl_t fl)
2127 {
2128 	iflib_dma_info_t idi = fl->ifl_ifdi;
2129 	bus_dmamap_t sd_map;
2130 	uint32_t i;
2131 
2132 	for (i = 0; i < fl->ifl_size; i++) {
2133 		struct mbuf **sd_m = &fl->ifl_sds.ifsd_m[i];
2134 		caddr_t *sd_cl = &fl->ifl_sds.ifsd_cl[i];
2135 
2136 		if (*sd_cl != NULL) {
2137 			sd_map = fl->ifl_sds.ifsd_map[i];
2138 			bus_dmamap_sync(fl->ifl_buf_tag, sd_map,
2139 			    BUS_DMASYNC_POSTREAD);
2140 			bus_dmamap_unload(fl->ifl_buf_tag, sd_map);
2141 			uma_zfree(fl->ifl_zone, *sd_cl);
2142 			*sd_cl = NULL;
2143 			if (*sd_m != NULL) {
2144 				m_init(*sd_m, M_NOWAIT, MT_DATA, 0);
2145 				uma_zfree(zone_mbuf, *sd_m);
2146 				*sd_m = NULL;
2147 			}
2148 		} else {
2149 			MPASS(*sd_m == NULL);
2150 		}
2151 #if MEMORY_LOGGING
2152 		fl->ifl_m_dequeued++;
2153 		fl->ifl_cl_dequeued++;
2154 #endif
2155 	}
2156 #ifdef INVARIANTS
2157 	for (i = 0; i < fl->ifl_size; i++) {
2158 		MPASS(fl->ifl_sds.ifsd_cl[i] == NULL);
2159 		MPASS(fl->ifl_sds.ifsd_m[i] == NULL);
2160 	}
2161 #endif
2162 	/*
2163 	 * Reset free list values
2164 	 */
2165 	fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = fl->ifl_fragidx = 0;
2166 	bzero(idi->idi_vaddr, idi->idi_size);
2167 }
2168 
2169 /*********************************************************************
2170  *
2171  *  Initialize a free list and its buffers.
2172  *
2173  **********************************************************************/
2174 static int
2175 iflib_fl_setup(iflib_fl_t fl)
2176 {
2177 	iflib_rxq_t rxq = fl->ifl_rxq;
2178 	if_ctx_t ctx = rxq->ifr_ctx;
2179 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2180 	int qidx;
2181 
2182 	bit_nclear(fl->ifl_rx_bitmap, 0, fl->ifl_size - 1);
2183 	/*
2184 	** Free current RX buffer structs and their mbufs
2185 	*/
2186 	iflib_fl_bufs_free(fl);
2187 	/* Now replenish the mbufs */
2188 	MPASS(fl->ifl_credits == 0);
2189 	qidx = rxq->ifr_fl_offset + fl->ifl_id;
2190 	if (scctx->isc_rxd_buf_size[qidx] != 0)
2191 		fl->ifl_buf_size = scctx->isc_rxd_buf_size[qidx];
2192 	else
2193 		fl->ifl_buf_size = ctx->ifc_rx_mbuf_sz;
2194 	/*
2195 	 * ifl_buf_size may be a driver-supplied value, so pull it up
2196 	 * to the selected mbuf size.
2197 	 */
2198 	fl->ifl_buf_size = iflib_get_mbuf_size_for(fl->ifl_buf_size);
2199 	if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size)
2200 		ctx->ifc_max_fl_buf_size = fl->ifl_buf_size;
2201 	fl->ifl_cltype = m_gettype(fl->ifl_buf_size);
2202 	fl->ifl_zone = m_getzone(fl->ifl_buf_size);
2203 
2204 	/*
2205 	 * Avoid pre-allocating zillions of clusters to an idle card
2206 	 * potentially speeding up attach. In any case make sure
2207 	 * to leave a descriptor unavailable. See the comment in
2208 	 * iflib_fl_refill_all().
2209 	 */
2210 	MPASS(fl->ifl_size > 0);
2211 	(void)iflib_fl_refill(ctx, fl, min(128, fl->ifl_size - 1));
2212 	if (min(128, fl->ifl_size - 1) != fl->ifl_credits)
2213 		return (ENOBUFS);
2214 	/*
2215 	 * handle failure
2216 	 */
2217 	MPASS(rxq != NULL);
2218 	MPASS(fl->ifl_ifdi != NULL);
2219 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2220 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2221 	return (0);
2222 }
2223 
2224 /*********************************************************************
2225  *
2226  *  Free receive ring data structures
2227  *
2228  **********************************************************************/
2229 static void
2230 iflib_rx_sds_free(iflib_rxq_t rxq)
2231 {
2232 	iflib_fl_t fl;
2233 	int i, j;
2234 
2235 	if (rxq->ifr_fl != NULL) {
2236 		for (i = 0; i < rxq->ifr_nfl; i++) {
2237 			fl = &rxq->ifr_fl[i];
2238 			if (fl->ifl_buf_tag != NULL) {
2239 				if (fl->ifl_sds.ifsd_map != NULL) {
2240 					for (j = 0; j < fl->ifl_size; j++) {
2241 						bus_dmamap_sync(
2242 						    fl->ifl_buf_tag,
2243 						    fl->ifl_sds.ifsd_map[j],
2244 						    BUS_DMASYNC_POSTREAD);
2245 						bus_dmamap_unload(
2246 						    fl->ifl_buf_tag,
2247 						    fl->ifl_sds.ifsd_map[j]);
2248 						bus_dmamap_destroy(
2249 						    fl->ifl_buf_tag,
2250 						    fl->ifl_sds.ifsd_map[j]);
2251 					}
2252 				}
2253 				bus_dma_tag_destroy(fl->ifl_buf_tag);
2254 				fl->ifl_buf_tag = NULL;
2255 			}
2256 			free(fl->ifl_sds.ifsd_m, M_IFLIB);
2257 			free(fl->ifl_sds.ifsd_cl, M_IFLIB);
2258 			free(fl->ifl_sds.ifsd_ba, M_IFLIB);
2259 			free(fl->ifl_sds.ifsd_map, M_IFLIB);
2260 			fl->ifl_sds.ifsd_m = NULL;
2261 			fl->ifl_sds.ifsd_cl = NULL;
2262 			fl->ifl_sds.ifsd_ba = NULL;
2263 			fl->ifl_sds.ifsd_map = NULL;
2264 		}
2265 		free(rxq->ifr_fl, M_IFLIB);
2266 		rxq->ifr_fl = NULL;
2267 		free(rxq->ifr_ifdi, M_IFLIB);
2268 		rxq->ifr_ifdi = NULL;
2269 		rxq->ifr_cq_cidx = 0;
2270 	}
2271 }
2272 
2273 /*
2274  * Timer routine
2275  */
2276 static void
2277 iflib_timer(void *arg)
2278 {
2279 	iflib_txq_t txq = arg;
2280 	if_ctx_t ctx = txq->ift_ctx;
2281 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2282 	uint64_t this_tick = ticks;
2283 
2284 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
2285 		return;
2286 
2287 	/*
2288 	** Check on the state of the TX queue(s), this
2289 	** can be done without the lock because its RO
2290 	** and the HUNG state will be static if set.
2291 	*/
2292 	if (this_tick - txq->ift_last_timer_tick >= hz / 2) {
2293 		txq->ift_last_timer_tick = this_tick;
2294 		IFDI_TIMER(ctx, txq->ift_id);
2295 		if ((txq->ift_qstatus == IFLIB_QUEUE_HUNG) &&
2296 		    ((txq->ift_cleaned_prev == txq->ift_cleaned) ||
2297 		     (sctx->isc_pause_frames == 0)))
2298 			goto hung;
2299 
2300 		if (txq->ift_qstatus != IFLIB_QUEUE_IDLE &&
2301 		    ifmp_ring_is_stalled(txq->ift_br)) {
2302 			KASSERT(ctx->ifc_link_state == LINK_STATE_UP, ("queue can't be marked as hung if interface is down"));
2303 			txq->ift_qstatus = IFLIB_QUEUE_HUNG;
2304 		}
2305 		txq->ift_cleaned_prev = txq->ift_cleaned;
2306 	}
2307 	/* handle any laggards */
2308 	if (txq->ift_db_pending)
2309 		GROUPTASK_ENQUEUE(&txq->ift_task);
2310 
2311 	sctx->isc_pause_frames = 0;
2312 	if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)
2313 		callout_reset_on(&txq->ift_timer, hz / 2, iflib_timer, txq, txq->ift_timer.c_cpu);
2314 	return;
2315 
2316  hung:
2317 	device_printf(ctx->ifc_dev,
2318 	    "Watchdog timeout (TX: %d desc avail: %d pidx: %d) -- resetting\n",
2319 	    txq->ift_id, TXQ_AVAIL(txq), txq->ift_pidx);
2320 	STATE_LOCK(ctx);
2321 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2322 	ctx->ifc_flags |= (IFC_DO_WATCHDOG|IFC_DO_RESET);
2323 	iflib_admin_intr_deferred(ctx);
2324 	STATE_UNLOCK(ctx);
2325 }
2326 
2327 static uint16_t
2328 iflib_get_mbuf_size_for(unsigned int size)
2329 {
2330 
2331 	if (size <= MCLBYTES)
2332 		return (MCLBYTES);
2333 	else
2334 		return (MJUMPAGESIZE);
2335 }
2336 
2337 static void
2338 iflib_calc_rx_mbuf_sz(if_ctx_t ctx)
2339 {
2340 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2341 
2342 	/*
2343 	 * XXX don't set the max_frame_size to larger
2344 	 * than the hardware can handle
2345 	 */
2346 	ctx->ifc_rx_mbuf_sz =
2347 	    iflib_get_mbuf_size_for(sctx->isc_max_frame_size);
2348 }
2349 
2350 uint32_t
2351 iflib_get_rx_mbuf_sz(if_ctx_t ctx)
2352 {
2353 
2354 	return (ctx->ifc_rx_mbuf_sz);
2355 }
2356 
2357 static void
2358 iflib_init_locked(if_ctx_t ctx)
2359 {
2360 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2361 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2362 	if_t ifp = ctx->ifc_ifp;
2363 	iflib_fl_t fl;
2364 	iflib_txq_t txq;
2365 	iflib_rxq_t rxq;
2366 	int i, j, tx_ip_csum_flags, tx_ip6_csum_flags;
2367 
2368 	if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2369 	IFDI_INTR_DISABLE(ctx);
2370 
2371 	tx_ip_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP);
2372 	tx_ip6_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | CSUM_IP6_SCTP);
2373 	/* Set hardware offload abilities */
2374 	if_clearhwassist(ifp);
2375 	if (if_getcapenable(ifp) & IFCAP_TXCSUM)
2376 		if_sethwassistbits(ifp, tx_ip_csum_flags, 0);
2377 	if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6)
2378 		if_sethwassistbits(ifp,  tx_ip6_csum_flags, 0);
2379 	if (if_getcapenable(ifp) & IFCAP_TSO4)
2380 		if_sethwassistbits(ifp, CSUM_IP_TSO, 0);
2381 	if (if_getcapenable(ifp) & IFCAP_TSO6)
2382 		if_sethwassistbits(ifp, CSUM_IP6_TSO, 0);
2383 
2384 	for (i = 0, txq = ctx->ifc_txqs; i < sctx->isc_ntxqsets; i++, txq++) {
2385 		CALLOUT_LOCK(txq);
2386 		callout_stop(&txq->ift_timer);
2387 #ifdef DEV_NETMAP
2388 		callout_stop(&txq->ift_netmap_timer);
2389 #endif /* DEV_NETMAP */
2390 		CALLOUT_UNLOCK(txq);
2391 		iflib_netmap_txq_init(ctx, txq);
2392 	}
2393 
2394 	/*
2395 	 * Calculate a suitable Rx mbuf size prior to calling IFDI_INIT, so
2396 	 * that drivers can use the value when setting up the hardware receive
2397 	 * buffers.
2398 	 */
2399 	iflib_calc_rx_mbuf_sz(ctx);
2400 
2401 #ifdef INVARIANTS
2402 	i = if_getdrvflags(ifp);
2403 #endif
2404 	IFDI_INIT(ctx);
2405 	MPASS(if_getdrvflags(ifp) == i);
2406 	for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
2407 		if (iflib_netmap_rxq_init(ctx, rxq) > 0) {
2408 			/* This rxq is in netmap mode. Skip normal init. */
2409 			continue;
2410 		}
2411 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
2412 			if (iflib_fl_setup(fl)) {
2413 				device_printf(ctx->ifc_dev,
2414 				    "setting up free list %d failed - "
2415 				    "check cluster settings\n", j);
2416 				goto done;
2417 			}
2418 		}
2419 	}
2420 done:
2421 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
2422 	IFDI_INTR_ENABLE(ctx);
2423 	txq = ctx->ifc_txqs;
2424 	for (i = 0; i < sctx->isc_ntxqsets; i++, txq++)
2425 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq,
2426 			txq->ift_timer.c_cpu);
2427 }
2428 
2429 static int
2430 iflib_media_change(if_t ifp)
2431 {
2432 	if_ctx_t ctx = if_getsoftc(ifp);
2433 	int err;
2434 
2435 	CTX_LOCK(ctx);
2436 	if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0)
2437 		iflib_init_locked(ctx);
2438 	CTX_UNLOCK(ctx);
2439 	return (err);
2440 }
2441 
2442 static void
2443 iflib_media_status(if_t ifp, struct ifmediareq *ifmr)
2444 {
2445 	if_ctx_t ctx = if_getsoftc(ifp);
2446 
2447 	CTX_LOCK(ctx);
2448 	IFDI_UPDATE_ADMIN_STATUS(ctx);
2449 	IFDI_MEDIA_STATUS(ctx, ifmr);
2450 	CTX_UNLOCK(ctx);
2451 }
2452 
2453 void
2454 iflib_stop(if_ctx_t ctx)
2455 {
2456 	iflib_txq_t txq = ctx->ifc_txqs;
2457 	iflib_rxq_t rxq = ctx->ifc_rxqs;
2458 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2459 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2460 	iflib_dma_info_t di;
2461 	iflib_fl_t fl;
2462 	int i, j;
2463 
2464 	/* Tell the stack that the interface is no longer active */
2465 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2466 
2467 	IFDI_INTR_DISABLE(ctx);
2468 	DELAY(1000);
2469 	IFDI_STOP(ctx);
2470 	DELAY(1000);
2471 
2472 	iflib_debug_reset();
2473 	/* Wait for current tx queue users to exit to disarm watchdog timer. */
2474 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) {
2475 		/* make sure all transmitters have completed before proceeding XXX */
2476 
2477 		CALLOUT_LOCK(txq);
2478 		callout_stop(&txq->ift_timer);
2479 #ifdef DEV_NETMAP
2480 		callout_stop(&txq->ift_netmap_timer);
2481 #endif /* DEV_NETMAP */
2482 		CALLOUT_UNLOCK(txq);
2483 
2484 		/* clean any enqueued buffers */
2485 		iflib_ifmp_purge(txq);
2486 		/* Free any existing tx buffers. */
2487 		for (j = 0; j < txq->ift_size; j++) {
2488 			iflib_txsd_free(ctx, txq, j);
2489 		}
2490 		txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0;
2491 		txq->ift_in_use = txq->ift_gen = txq->ift_cidx = txq->ift_pidx = txq->ift_no_desc_avail = 0;
2492 		txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0;
2493 		txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0;
2494 		txq->ift_pullups = 0;
2495 		ifmp_ring_reset_stats(txq->ift_br);
2496 		for (j = 0, di = txq->ift_ifdi; j < sctx->isc_ntxqs; j++, di++)
2497 			bzero((void *)di->idi_vaddr, di->idi_size);
2498 	}
2499 	for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) {
2500 		/* make sure all transmitters have completed before proceeding XXX */
2501 
2502 		rxq->ifr_cq_cidx = 0;
2503 		for (j = 0, di = rxq->ifr_ifdi; j < sctx->isc_nrxqs; j++, di++)
2504 			bzero((void *)di->idi_vaddr, di->idi_size);
2505 		/* also resets the free lists pidx/cidx */
2506 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
2507 			iflib_fl_bufs_free(fl);
2508 	}
2509 }
2510 
2511 static inline caddr_t
2512 calc_next_rxd(iflib_fl_t fl, int cidx)
2513 {
2514 	qidx_t size;
2515 	int nrxd;
2516 	caddr_t start, end, cur, next;
2517 
2518 	nrxd = fl->ifl_size;
2519 	size = fl->ifl_rxd_size;
2520 	start = fl->ifl_ifdi->idi_vaddr;
2521 
2522 	if (__predict_false(size == 0))
2523 		return (start);
2524 	cur = start + size*cidx;
2525 	end = start + size*nrxd;
2526 	next = CACHE_PTR_NEXT(cur);
2527 	return (next < end ? next : start);
2528 }
2529 
2530 static inline void
2531 prefetch_pkts(iflib_fl_t fl, int cidx)
2532 {
2533 	int nextptr;
2534 	int nrxd = fl->ifl_size;
2535 	caddr_t next_rxd;
2536 
2537 	nextptr = (cidx + CACHE_PTR_INCREMENT) & (nrxd-1);
2538 	prefetch(&fl->ifl_sds.ifsd_m[nextptr]);
2539 	prefetch(&fl->ifl_sds.ifsd_cl[nextptr]);
2540 	next_rxd = calc_next_rxd(fl, cidx);
2541 	prefetch(next_rxd);
2542 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 1) & (nrxd-1)]);
2543 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 2) & (nrxd-1)]);
2544 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 3) & (nrxd-1)]);
2545 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 4) & (nrxd-1)]);
2546 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 1) & (nrxd-1)]);
2547 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 2) & (nrxd-1)]);
2548 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 3) & (nrxd-1)]);
2549 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 4) & (nrxd-1)]);
2550 }
2551 
2552 static struct mbuf *
2553 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, bool unload, if_rxsd_t sd,
2554     int *pf_rv, if_rxd_info_t ri)
2555 {
2556 	bus_dmamap_t map;
2557 	iflib_fl_t fl;
2558 	caddr_t payload;
2559 	struct mbuf *m;
2560 	int flid, cidx, len, next;
2561 
2562 	map = NULL;
2563 	flid = irf->irf_flid;
2564 	cidx = irf->irf_idx;
2565 	fl = &rxq->ifr_fl[flid];
2566 	sd->ifsd_fl = fl;
2567 	m = fl->ifl_sds.ifsd_m[cidx];
2568 	sd->ifsd_cl = &fl->ifl_sds.ifsd_cl[cidx];
2569 	fl->ifl_credits--;
2570 #if MEMORY_LOGGING
2571 	fl->ifl_m_dequeued++;
2572 #endif
2573 	if (rxq->ifr_ctx->ifc_flags & IFC_PREFETCH)
2574 		prefetch_pkts(fl, cidx);
2575 	next = (cidx + CACHE_PTR_INCREMENT) & (fl->ifl_size-1);
2576 	prefetch(&fl->ifl_sds.ifsd_map[next]);
2577 	map = fl->ifl_sds.ifsd_map[cidx];
2578 
2579 	bus_dmamap_sync(fl->ifl_buf_tag, map, BUS_DMASYNC_POSTREAD);
2580 
2581 	if (rxq->pfil != NULL && PFIL_HOOKED_IN(rxq->pfil) && pf_rv != NULL &&
2582 	    irf->irf_len != 0) {
2583 		payload  = *sd->ifsd_cl;
2584 		payload +=  ri->iri_pad;
2585 		len = ri->iri_len - ri->iri_pad;
2586 		*pf_rv = pfil_run_hooks(rxq->pfil, payload, ri->iri_ifp,
2587 		    len | PFIL_MEMPTR | PFIL_IN, NULL);
2588 		switch (*pf_rv) {
2589 		case PFIL_DROPPED:
2590 		case PFIL_CONSUMED:
2591 			/*
2592 			 * The filter ate it.  Everything is recycled.
2593 			 */
2594 			m = NULL;
2595 			unload = 0;
2596 			break;
2597 		case PFIL_REALLOCED:
2598 			/*
2599 			 * The filter copied it.  Everything is recycled.
2600 			 */
2601 			m = pfil_mem2mbuf(payload);
2602 			unload = 0;
2603 			break;
2604 		case PFIL_PASS:
2605 			/*
2606 			 * Filter said it was OK, so receive like
2607 			 * normal
2608 			 */
2609 			fl->ifl_sds.ifsd_m[cidx] = NULL;
2610 			break;
2611 		default:
2612 			MPASS(0);
2613 		}
2614 	} else {
2615 		fl->ifl_sds.ifsd_m[cidx] = NULL;
2616 		*pf_rv = PFIL_PASS;
2617 	}
2618 
2619 	if (unload && irf->irf_len != 0)
2620 		bus_dmamap_unload(fl->ifl_buf_tag, map);
2621 	fl->ifl_cidx = (fl->ifl_cidx + 1) & (fl->ifl_size-1);
2622 	if (__predict_false(fl->ifl_cidx == 0))
2623 		fl->ifl_gen = 0;
2624 	bit_clear(fl->ifl_rx_bitmap, cidx);
2625 	return (m);
2626 }
2627 
2628 static struct mbuf *
2629 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri, if_rxsd_t sd, int *pf_rv)
2630 {
2631 	struct mbuf *m, *mh, *mt;
2632 	caddr_t cl;
2633 	int  *pf_rv_ptr, flags, i, padlen;
2634 	bool consumed;
2635 
2636 	i = 0;
2637 	mh = NULL;
2638 	consumed = false;
2639 	*pf_rv = PFIL_PASS;
2640 	pf_rv_ptr = pf_rv;
2641 	do {
2642 		m = rxd_frag_to_sd(rxq, &ri->iri_frags[i], !consumed, sd,
2643 		    pf_rv_ptr, ri);
2644 
2645 		MPASS(*sd->ifsd_cl != NULL);
2646 
2647 		/*
2648 		 * Exclude zero-length frags & frags from
2649 		 * packets the filter has consumed or dropped
2650 		 */
2651 		if (ri->iri_frags[i].irf_len == 0 || consumed ||
2652 		    *pf_rv == PFIL_CONSUMED || *pf_rv == PFIL_DROPPED) {
2653 			if (mh == NULL) {
2654 				/* everything saved here */
2655 				consumed = true;
2656 				pf_rv_ptr = NULL;
2657 				continue;
2658 			}
2659 			/* XXX we can save the cluster here, but not the mbuf */
2660 			m_init(m, M_NOWAIT, MT_DATA, 0);
2661 			m_free(m);
2662 			continue;
2663 		}
2664 		if (mh == NULL) {
2665 			flags = M_PKTHDR|M_EXT;
2666 			mh = mt = m;
2667 			padlen = ri->iri_pad;
2668 		} else {
2669 			flags = M_EXT;
2670 			mt->m_next = m;
2671 			mt = m;
2672 			/* assuming padding is only on the first fragment */
2673 			padlen = 0;
2674 		}
2675 		cl = *sd->ifsd_cl;
2676 		*sd->ifsd_cl = NULL;
2677 
2678 		/* Can these two be made one ? */
2679 		m_init(m, M_NOWAIT, MT_DATA, flags);
2680 		m_cljset(m, cl, sd->ifsd_fl->ifl_cltype);
2681 		/*
2682 		 * These must follow m_init and m_cljset
2683 		 */
2684 		m->m_data += padlen;
2685 		ri->iri_len -= padlen;
2686 		m->m_len = ri->iri_frags[i].irf_len;
2687 	} while (++i < ri->iri_nfrags);
2688 
2689 	return (mh);
2690 }
2691 
2692 /*
2693  * Process one software descriptor
2694  */
2695 static struct mbuf *
2696 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri)
2697 {
2698 	struct if_rxsd sd;
2699 	struct mbuf *m;
2700 	int pf_rv;
2701 
2702 	/* should I merge this back in now that the two paths are basically duplicated? */
2703 	if (ri->iri_nfrags == 1 &&
2704 	    ri->iri_frags[0].irf_len != 0 &&
2705 	    ri->iri_frags[0].irf_len <= MIN(IFLIB_RX_COPY_THRESH, MHLEN)) {
2706 		m = rxd_frag_to_sd(rxq, &ri->iri_frags[0], false, &sd,
2707 		    &pf_rv, ri);
2708 		if (pf_rv != PFIL_PASS && pf_rv != PFIL_REALLOCED)
2709 			return (m);
2710 		if (pf_rv == PFIL_PASS) {
2711 			m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR);
2712 #ifndef __NO_STRICT_ALIGNMENT
2713 			if (!IP_ALIGNED(m))
2714 				m->m_data += 2;
2715 #endif
2716 			memcpy(m->m_data, *sd.ifsd_cl, ri->iri_len);
2717 			m->m_len = ri->iri_frags[0].irf_len;
2718 		}
2719 	} else {
2720 		m = assemble_segments(rxq, ri, &sd, &pf_rv);
2721 		if (m == NULL)
2722 			return (NULL);
2723 		if (pf_rv != PFIL_PASS && pf_rv != PFIL_REALLOCED)
2724 			return (m);
2725 	}
2726 	m->m_pkthdr.len = ri->iri_len;
2727 	m->m_pkthdr.rcvif = ri->iri_ifp;
2728 	m->m_flags |= ri->iri_flags;
2729 	m->m_pkthdr.ether_vtag = ri->iri_vtag;
2730 	m->m_pkthdr.flowid = ri->iri_flowid;
2731 	M_HASHTYPE_SET(m, ri->iri_rsstype);
2732 	m->m_pkthdr.csum_flags = ri->iri_csum_flags;
2733 	m->m_pkthdr.csum_data = ri->iri_csum_data;
2734 	return (m);
2735 }
2736 
2737 #if defined(INET6) || defined(INET)
2738 static void
2739 iflib_get_ip_forwarding(struct lro_ctrl *lc, bool *v4, bool *v6)
2740 {
2741 	CURVNET_SET(lc->ifp->if_vnet);
2742 #if defined(INET6)
2743 	*v6 = V_ip6_forwarding;
2744 #endif
2745 #if defined(INET)
2746 	*v4 = V_ipforwarding;
2747 #endif
2748 	CURVNET_RESTORE();
2749 }
2750 
2751 /*
2752  * Returns true if it's possible this packet could be LROed.
2753  * if it returns false, it is guaranteed that tcp_lro_rx()
2754  * would not return zero.
2755  */
2756 static bool
2757 iflib_check_lro_possible(struct mbuf *m, bool v4_forwarding, bool v6_forwarding)
2758 {
2759 	struct ether_header *eh;
2760 
2761 	eh = mtod(m, struct ether_header *);
2762 	switch (eh->ether_type) {
2763 #if defined(INET6)
2764 		case htons(ETHERTYPE_IPV6):
2765 			return (!v6_forwarding);
2766 #endif
2767 #if defined (INET)
2768 		case htons(ETHERTYPE_IP):
2769 			return (!v4_forwarding);
2770 #endif
2771 	}
2772 
2773 	return false;
2774 }
2775 #else
2776 static void
2777 iflib_get_ip_forwarding(struct lro_ctrl *lc __unused, bool *v4 __unused, bool *v6 __unused)
2778 {
2779 }
2780 #endif
2781 
2782 static void
2783 _task_fn_rx_watchdog(void *context)
2784 {
2785 	iflib_rxq_t rxq = context;
2786 
2787 	GROUPTASK_ENQUEUE(&rxq->ifr_task);
2788 }
2789 
2790 static uint8_t
2791 iflib_rxeof(iflib_rxq_t rxq, qidx_t budget)
2792 {
2793 	if_t ifp;
2794 	if_ctx_t ctx = rxq->ifr_ctx;
2795 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2796 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2797 	int avail, i;
2798 	qidx_t *cidxp;
2799 	struct if_rxd_info ri;
2800 	int err, budget_left, rx_bytes, rx_pkts;
2801 	iflib_fl_t fl;
2802 	int lro_enabled;
2803 	bool v4_forwarding, v6_forwarding, lro_possible;
2804 	uint8_t retval = 0;
2805 
2806 	/*
2807 	 * XXX early demux data packets so that if_input processing only handles
2808 	 * acks in interrupt context
2809 	 */
2810 	struct mbuf *m, *mh, *mt, *mf;
2811 
2812 	NET_EPOCH_ASSERT();
2813 
2814 	lro_possible = v4_forwarding = v6_forwarding = false;
2815 	ifp = ctx->ifc_ifp;
2816 	mh = mt = NULL;
2817 	MPASS(budget > 0);
2818 	rx_pkts	= rx_bytes = 0;
2819 	if (sctx->isc_flags & IFLIB_HAS_RXCQ)
2820 		cidxp = &rxq->ifr_cq_cidx;
2821 	else
2822 		cidxp = &rxq->ifr_fl[0].ifl_cidx;
2823 	if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) {
2824 		for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2825 			retval |= iflib_fl_refill_all(ctx, fl);
2826 		DBG_COUNTER_INC(rx_unavail);
2827 		return (retval);
2828 	}
2829 
2830 	/* pfil needs the vnet to be set */
2831 	CURVNET_SET_QUIET(ifp->if_vnet);
2832 	for (budget_left = budget; budget_left > 0 && avail > 0;) {
2833 		if (__predict_false(!CTX_ACTIVE(ctx))) {
2834 			DBG_COUNTER_INC(rx_ctx_inactive);
2835 			break;
2836 		}
2837 		/*
2838 		 * Reset client set fields to their default values
2839 		 */
2840 		rxd_info_zero(&ri);
2841 		ri.iri_qsidx = rxq->ifr_id;
2842 		ri.iri_cidx = *cidxp;
2843 		ri.iri_ifp = ifp;
2844 		ri.iri_frags = rxq->ifr_frags;
2845 		err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
2846 
2847 		if (err)
2848 			goto err;
2849 		rx_pkts += 1;
2850 		rx_bytes += ri.iri_len;
2851 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
2852 			*cidxp = ri.iri_cidx;
2853 			/* Update our consumer index */
2854 			/* XXX NB: shurd - check if this is still safe */
2855 			while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0])
2856 				rxq->ifr_cq_cidx -= scctx->isc_nrxd[0];
2857 			/* was this only a completion queue message? */
2858 			if (__predict_false(ri.iri_nfrags == 0))
2859 				continue;
2860 		}
2861 		MPASS(ri.iri_nfrags != 0);
2862 		MPASS(ri.iri_len != 0);
2863 
2864 		/* will advance the cidx on the corresponding free lists */
2865 		m = iflib_rxd_pkt_get(rxq, &ri);
2866 		avail--;
2867 		budget_left--;
2868 		if (avail == 0 && budget_left)
2869 			avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left);
2870 
2871 		if (__predict_false(m == NULL))
2872 			continue;
2873 
2874 		/* imm_pkt: -- cxgb */
2875 		if (mh == NULL)
2876 			mh = mt = m;
2877 		else {
2878 			mt->m_nextpkt = m;
2879 			mt = m;
2880 		}
2881 	}
2882 	CURVNET_RESTORE();
2883 	/* make sure that we can refill faster than drain */
2884 	for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2885 		retval |= iflib_fl_refill_all(ctx, fl);
2886 
2887 	lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO);
2888 	if (lro_enabled)
2889 		iflib_get_ip_forwarding(&rxq->ifr_lc, &v4_forwarding, &v6_forwarding);
2890 	mt = mf = NULL;
2891 	while (mh != NULL) {
2892 		m = mh;
2893 		mh = mh->m_nextpkt;
2894 		m->m_nextpkt = NULL;
2895 #ifndef __NO_STRICT_ALIGNMENT
2896 		if (!IP_ALIGNED(m) && (m = iflib_fixup_rx(m)) == NULL)
2897 			continue;
2898 #endif
2899 		rx_bytes += m->m_pkthdr.len;
2900 		rx_pkts++;
2901 #if defined(INET6) || defined(INET)
2902 		if (lro_enabled) {
2903 			if (!lro_possible) {
2904 				lro_possible = iflib_check_lro_possible(m, v4_forwarding, v6_forwarding);
2905 				if (lro_possible && mf != NULL) {
2906 					ifp->if_input(ifp, mf);
2907 					DBG_COUNTER_INC(rx_if_input);
2908 					mt = mf = NULL;
2909 				}
2910 			}
2911 			if ((m->m_pkthdr.csum_flags & (CSUM_L4_CALC|CSUM_L4_VALID)) ==
2912 			    (CSUM_L4_CALC|CSUM_L4_VALID)) {
2913 				if (lro_possible && tcp_lro_rx(&rxq->ifr_lc, m, 0) == 0)
2914 					continue;
2915 			}
2916 		}
2917 #endif
2918 		if (lro_possible) {
2919 			ifp->if_input(ifp, m);
2920 			DBG_COUNTER_INC(rx_if_input);
2921 			continue;
2922 		}
2923 
2924 		if (mf == NULL)
2925 			mf = m;
2926 		if (mt != NULL)
2927 			mt->m_nextpkt = m;
2928 		mt = m;
2929 	}
2930 	if (mf != NULL) {
2931 		ifp->if_input(ifp, mf);
2932 		DBG_COUNTER_INC(rx_if_input);
2933 	}
2934 
2935 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
2936 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
2937 
2938 	/*
2939 	 * Flush any outstanding LRO work
2940 	 */
2941 #if defined(INET6) || defined(INET)
2942 	tcp_lro_flush_all(&rxq->ifr_lc);
2943 #endif
2944 	if (avail != 0 || iflib_rxd_avail(ctx, rxq, *cidxp, 1) != 0)
2945 		retval |= IFLIB_RXEOF_MORE;
2946 	return (retval);
2947 err:
2948 	STATE_LOCK(ctx);
2949 	ctx->ifc_flags |= IFC_DO_RESET;
2950 	iflib_admin_intr_deferred(ctx);
2951 	STATE_UNLOCK(ctx);
2952 	return (0);
2953 }
2954 
2955 #define TXD_NOTIFY_COUNT(txq) (((txq)->ift_size / (txq)->ift_update_freq)-1)
2956 static inline qidx_t
2957 txq_max_db_deferred(iflib_txq_t txq, qidx_t in_use)
2958 {
2959 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2960 	qidx_t minthresh = txq->ift_size / 8;
2961 	if (in_use > 4*minthresh)
2962 		return (notify_count);
2963 	if (in_use > 2*minthresh)
2964 		return (notify_count >> 1);
2965 	if (in_use > minthresh)
2966 		return (notify_count >> 3);
2967 	return (0);
2968 }
2969 
2970 static inline qidx_t
2971 txq_max_rs_deferred(iflib_txq_t txq)
2972 {
2973 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2974 	qidx_t minthresh = txq->ift_size / 8;
2975 	if (txq->ift_in_use > 4*minthresh)
2976 		return (notify_count);
2977 	if (txq->ift_in_use > 2*minthresh)
2978 		return (notify_count >> 1);
2979 	if (txq->ift_in_use > minthresh)
2980 		return (notify_count >> 2);
2981 	return (2);
2982 }
2983 
2984 #define M_CSUM_FLAGS(m) ((m)->m_pkthdr.csum_flags)
2985 #define M_HAS_VLANTAG(m) (m->m_flags & M_VLANTAG)
2986 
2987 #define TXQ_MAX_DB_DEFERRED(txq, in_use) txq_max_db_deferred((txq), (in_use))
2988 #define TXQ_MAX_RS_DEFERRED(txq) txq_max_rs_deferred(txq)
2989 #define TXQ_MAX_DB_CONSUMED(size) (size >> 4)
2990 
2991 /* forward compatibility for cxgb */
2992 #define FIRST_QSET(ctx) 0
2993 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets)
2994 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets)
2995 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx))
2996 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments))
2997 
2998 /* XXX we should be setting this to something other than zero */
2999 #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh)
3000 #define	MAX_TX_DESC(ctx) max((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max, \
3001     (ctx)->ifc_softc_ctx.isc_tx_nsegments)
3002 
3003 static inline bool
3004 iflib_txd_db_check(if_ctx_t ctx, iflib_txq_t txq, int ring, qidx_t in_use)
3005 {
3006 	qidx_t dbval, max;
3007 	bool rang;
3008 
3009 	rang = false;
3010 	max = TXQ_MAX_DB_DEFERRED(txq, in_use);
3011 	if (ring || txq->ift_db_pending >= max) {
3012 		dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx;
3013 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3014 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
3015 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval);
3016 		txq->ift_db_pending = txq->ift_npending = 0;
3017 		rang = true;
3018 	}
3019 	return (rang);
3020 }
3021 
3022 #ifdef PKT_DEBUG
3023 static void
3024 print_pkt(if_pkt_info_t pi)
3025 {
3026 	printf("pi len:  %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n",
3027 	       pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx);
3028 	printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n",
3029 	       pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag);
3030 	printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n",
3031 	       pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto);
3032 }
3033 #endif
3034 
3035 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO)
3036 #define IS_TX_OFFLOAD4(pi) ((pi)->ipi_csum_flags & (CSUM_IP_TCP | CSUM_IP_TSO))
3037 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO)
3038 #define IS_TX_OFFLOAD6(pi) ((pi)->ipi_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_TSO))
3039 
3040 static int
3041 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp)
3042 {
3043 	if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx;
3044 	struct ether_vlan_header *eh;
3045 	struct mbuf *m;
3046 
3047 	m = *mp;
3048 	if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) &&
3049 	    M_WRITABLE(m) == 0) {
3050 		if ((m = m_dup(m, M_NOWAIT)) == NULL) {
3051 			return (ENOMEM);
3052 		} else {
3053 			m_freem(*mp);
3054 			DBG_COUNTER_INC(tx_frees);
3055 			*mp = m;
3056 		}
3057 	}
3058 
3059 	/*
3060 	 * Determine where frame payload starts.
3061 	 * Jump over vlan headers if already present,
3062 	 * helpful for QinQ too.
3063 	 */
3064 	if (__predict_false(m->m_len < sizeof(*eh))) {
3065 		txq->ift_pullups++;
3066 		if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL))
3067 			return (ENOMEM);
3068 	}
3069 	eh = mtod(m, struct ether_vlan_header *);
3070 	if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
3071 		pi->ipi_etype = ntohs(eh->evl_proto);
3072 		pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
3073 	} else {
3074 		pi->ipi_etype = ntohs(eh->evl_encap_proto);
3075 		pi->ipi_ehdrlen = ETHER_HDR_LEN;
3076 	}
3077 
3078 	switch (pi->ipi_etype) {
3079 #ifdef INET
3080 	case ETHERTYPE_IP:
3081 	{
3082 		struct mbuf *n;
3083 		struct ip *ip = NULL;
3084 		struct tcphdr *th = NULL;
3085 		int minthlen;
3086 
3087 		minthlen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip) + sizeof(*th));
3088 		if (__predict_false(m->m_len < minthlen)) {
3089 			/*
3090 			 * if this code bloat is causing too much of a hit
3091 			 * move it to a separate function and mark it noinline
3092 			 */
3093 			if (m->m_len == pi->ipi_ehdrlen) {
3094 				n = m->m_next;
3095 				MPASS(n);
3096 				if (n->m_len >= sizeof(*ip))  {
3097 					ip = (struct ip *)n->m_data;
3098 					if (n->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3099 						th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3100 				} else {
3101 					txq->ift_pullups++;
3102 					if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
3103 						return (ENOMEM);
3104 					ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3105 				}
3106 			} else {
3107 				txq->ift_pullups++;
3108 				if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
3109 					return (ENOMEM);
3110 				ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3111 				if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3112 					th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3113 			}
3114 		} else {
3115 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3116 			if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3117 				th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3118 		}
3119 		pi->ipi_ip_hlen = ip->ip_hl << 2;
3120 		pi->ipi_ipproto = ip->ip_p;
3121 		pi->ipi_flags |= IPI_TX_IPV4;
3122 
3123 		/* TCP checksum offload may require TCP header length */
3124 		if (IS_TX_OFFLOAD4(pi)) {
3125 			if (__predict_true(pi->ipi_ipproto == IPPROTO_TCP)) {
3126 				if (__predict_false(th == NULL)) {
3127 					txq->ift_pullups++;
3128 					if (__predict_false((m = m_pullup(m, (ip->ip_hl << 2) + sizeof(*th))) == NULL))
3129 						return (ENOMEM);
3130 					th = (struct tcphdr *)((caddr_t)ip + pi->ipi_ip_hlen);
3131 				}
3132 				pi->ipi_tcp_hflags = th->th_flags;
3133 				pi->ipi_tcp_hlen = th->th_off << 2;
3134 				pi->ipi_tcp_seq = th->th_seq;
3135 			}
3136 			if (IS_TSO4(pi)) {
3137 				if (__predict_false(ip->ip_p != IPPROTO_TCP))
3138 					return (ENXIO);
3139 				/*
3140 				 * TSO always requires hardware checksum offload.
3141 				 */
3142 				pi->ipi_csum_flags |= (CSUM_IP_TCP | CSUM_IP);
3143 				th->th_sum = in_pseudo(ip->ip_src.s_addr,
3144 						       ip->ip_dst.s_addr, htons(IPPROTO_TCP));
3145 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3146 				if (sctx->isc_flags & IFLIB_TSO_INIT_IP) {
3147 					ip->ip_sum = 0;
3148 					ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz);
3149 				}
3150 			}
3151 		}
3152 		if ((sctx->isc_flags & IFLIB_NEED_ZERO_CSUM) && (pi->ipi_csum_flags & CSUM_IP))
3153                        ip->ip_sum = 0;
3154 
3155 		break;
3156 	}
3157 #endif
3158 #ifdef INET6
3159 	case ETHERTYPE_IPV6:
3160 	{
3161 		struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
3162 		struct tcphdr *th;
3163 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
3164 
3165 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
3166 			txq->ift_pullups++;
3167 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
3168 				return (ENOMEM);
3169 		}
3170 		th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
3171 
3172 		/* XXX-BZ this will go badly in case of ext hdrs. */
3173 		pi->ipi_ipproto = ip6->ip6_nxt;
3174 		pi->ipi_flags |= IPI_TX_IPV6;
3175 
3176 		/* TCP checksum offload may require TCP header length */
3177 		if (IS_TX_OFFLOAD6(pi)) {
3178 			if (pi->ipi_ipproto == IPPROTO_TCP) {
3179 				if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) {
3180 					txq->ift_pullups++;
3181 					if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL))
3182 						return (ENOMEM);
3183 				}
3184 				pi->ipi_tcp_hflags = th->th_flags;
3185 				pi->ipi_tcp_hlen = th->th_off << 2;
3186 				pi->ipi_tcp_seq = th->th_seq;
3187 			}
3188 			if (IS_TSO6(pi)) {
3189 				if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP))
3190 					return (ENXIO);
3191 				/*
3192 				 * TSO always requires hardware checksum offload.
3193 				 */
3194 				pi->ipi_csum_flags |= CSUM_IP6_TCP;
3195 				th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0);
3196 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3197 			}
3198 		}
3199 		break;
3200 	}
3201 #endif
3202 	default:
3203 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
3204 		pi->ipi_ip_hlen = 0;
3205 		break;
3206 	}
3207 	*mp = m;
3208 
3209 	return (0);
3210 }
3211 
3212 /*
3213  * If dodgy hardware rejects the scatter gather chain we've handed it
3214  * we'll need to remove the mbuf chain from ifsg_m[] before we can add the
3215  * m_defrag'd mbufs
3216  */
3217 static __noinline struct mbuf *
3218 iflib_remove_mbuf(iflib_txq_t txq)
3219 {
3220 	int ntxd, pidx;
3221 	struct mbuf *m, **ifsd_m;
3222 
3223 	ifsd_m = txq->ift_sds.ifsd_m;
3224 	ntxd = txq->ift_size;
3225 	pidx = txq->ift_pidx & (ntxd - 1);
3226 	ifsd_m = txq->ift_sds.ifsd_m;
3227 	m = ifsd_m[pidx];
3228 	ifsd_m[pidx] = NULL;
3229 	bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[pidx]);
3230 	if (txq->ift_sds.ifsd_tso_map != NULL)
3231 		bus_dmamap_unload(txq->ift_tso_buf_tag,
3232 		    txq->ift_sds.ifsd_tso_map[pidx]);
3233 #if MEMORY_LOGGING
3234 	txq->ift_dequeued++;
3235 #endif
3236 	return (m);
3237 }
3238 
3239 static inline caddr_t
3240 calc_next_txd(iflib_txq_t txq, int cidx, uint8_t qid)
3241 {
3242 	qidx_t size;
3243 	int ntxd;
3244 	caddr_t start, end, cur, next;
3245 
3246 	ntxd = txq->ift_size;
3247 	size = txq->ift_txd_size[qid];
3248 	start = txq->ift_ifdi[qid].idi_vaddr;
3249 
3250 	if (__predict_false(size == 0))
3251 		return (start);
3252 	cur = start + size*cidx;
3253 	end = start + size*ntxd;
3254 	next = CACHE_PTR_NEXT(cur);
3255 	return (next < end ? next : start);
3256 }
3257 
3258 /*
3259  * Pad an mbuf to ensure a minimum ethernet frame size.
3260  * min_frame_size is the frame size (less CRC) to pad the mbuf to
3261  */
3262 static __noinline int
3263 iflib_ether_pad(device_t dev, struct mbuf **m_head, uint16_t min_frame_size)
3264 {
3265 	/*
3266 	 * 18 is enough bytes to pad an ARP packet to 46 bytes, and
3267 	 * and ARP message is the smallest common payload I can think of
3268 	 */
3269 	static char pad[18];	/* just zeros */
3270 	int n;
3271 	struct mbuf *new_head;
3272 
3273 	if (!M_WRITABLE(*m_head)) {
3274 		new_head = m_dup(*m_head, M_NOWAIT);
3275 		if (new_head == NULL) {
3276 			m_freem(*m_head);
3277 			device_printf(dev, "cannot pad short frame, m_dup() failed");
3278 			DBG_COUNTER_INC(encap_pad_mbuf_fail);
3279 			DBG_COUNTER_INC(tx_frees);
3280 			return ENOMEM;
3281 		}
3282 		m_freem(*m_head);
3283 		*m_head = new_head;
3284 	}
3285 
3286 	for (n = min_frame_size - (*m_head)->m_pkthdr.len;
3287 	     n > 0; n -= sizeof(pad))
3288 		if (!m_append(*m_head, min(n, sizeof(pad)), pad))
3289 			break;
3290 
3291 	if (n > 0) {
3292 		m_freem(*m_head);
3293 		device_printf(dev, "cannot pad short frame\n");
3294 		DBG_COUNTER_INC(encap_pad_mbuf_fail);
3295 		DBG_COUNTER_INC(tx_frees);
3296 		return (ENOBUFS);
3297 	}
3298 
3299 	return 0;
3300 }
3301 
3302 static int
3303 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp)
3304 {
3305 	if_ctx_t		ctx;
3306 	if_shared_ctx_t		sctx;
3307 	if_softc_ctx_t		scctx;
3308 	bus_dma_tag_t		buf_tag;
3309 	bus_dma_segment_t	*segs;
3310 	struct mbuf		*m_head, **ifsd_m;
3311 	void			*next_txd;
3312 	bus_dmamap_t		map;
3313 	struct if_pkt_info	pi;
3314 	int remap = 0;
3315 	int err, nsegs, ndesc, max_segs, pidx, cidx, next, ntxd;
3316 
3317 	ctx = txq->ift_ctx;
3318 	sctx = ctx->ifc_sctx;
3319 	scctx = &ctx->ifc_softc_ctx;
3320 	segs = txq->ift_segs;
3321 	ntxd = txq->ift_size;
3322 	m_head = *m_headp;
3323 	map = NULL;
3324 
3325 	/*
3326 	 * If we're doing TSO the next descriptor to clean may be quite far ahead
3327 	 */
3328 	cidx = txq->ift_cidx;
3329 	pidx = txq->ift_pidx;
3330 	if (ctx->ifc_flags & IFC_PREFETCH) {
3331 		next = (cidx + CACHE_PTR_INCREMENT) & (ntxd-1);
3332 		if (!(ctx->ifc_flags & IFLIB_HAS_TXCQ)) {
3333 			next_txd = calc_next_txd(txq, cidx, 0);
3334 			prefetch(next_txd);
3335 		}
3336 
3337 		/* prefetch the next cache line of mbuf pointers and flags */
3338 		prefetch(&txq->ift_sds.ifsd_m[next]);
3339 		prefetch(&txq->ift_sds.ifsd_map[next]);
3340 		next = (cidx + CACHE_LINE_SIZE) & (ntxd-1);
3341 	}
3342 	map = txq->ift_sds.ifsd_map[pidx];
3343 	ifsd_m = txq->ift_sds.ifsd_m;
3344 
3345 	if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
3346 		buf_tag = txq->ift_tso_buf_tag;
3347 		max_segs = scctx->isc_tx_tso_segments_max;
3348 		map = txq->ift_sds.ifsd_tso_map[pidx];
3349 		MPASS(buf_tag != NULL);
3350 		MPASS(max_segs > 0);
3351 	} else {
3352 		buf_tag = txq->ift_buf_tag;
3353 		max_segs = scctx->isc_tx_nsegments;
3354 		map = txq->ift_sds.ifsd_map[pidx];
3355 	}
3356 	if ((sctx->isc_flags & IFLIB_NEED_ETHER_PAD) &&
3357 	    __predict_false(m_head->m_pkthdr.len < scctx->isc_min_frame_size)) {
3358 		err = iflib_ether_pad(ctx->ifc_dev, m_headp, scctx->isc_min_frame_size);
3359 		if (err) {
3360 			DBG_COUNTER_INC(encap_txd_encap_fail);
3361 			return err;
3362 		}
3363 	}
3364 	m_head = *m_headp;
3365 
3366 	pkt_info_zero(&pi);
3367 	pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG|M_BCAST|M_MCAST));
3368 	pi.ipi_pidx = pidx;
3369 	pi.ipi_qsidx = txq->ift_id;
3370 	pi.ipi_len = m_head->m_pkthdr.len;
3371 	pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags;
3372 	pi.ipi_vtag = M_HAS_VLANTAG(m_head) ? m_head->m_pkthdr.ether_vtag : 0;
3373 
3374 	/* deliberate bitwise OR to make one condition */
3375 	if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) {
3376 		if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0)) {
3377 			DBG_COUNTER_INC(encap_txd_encap_fail);
3378 			return (err);
3379 		}
3380 		m_head = *m_headp;
3381 	}
3382 
3383 retry:
3384 	err = bus_dmamap_load_mbuf_sg(buf_tag, map, m_head, segs, &nsegs,
3385 	    BUS_DMA_NOWAIT);
3386 defrag:
3387 	if (__predict_false(err)) {
3388 		switch (err) {
3389 		case EFBIG:
3390 			/* try collapse once and defrag once */
3391 			if (remap == 0) {
3392 				m_head = m_collapse(*m_headp, M_NOWAIT, max_segs);
3393 				/* try defrag if collapsing fails */
3394 				if (m_head == NULL)
3395 					remap++;
3396 			}
3397 			if (remap == 1) {
3398 				txq->ift_mbuf_defrag++;
3399 				m_head = m_defrag(*m_headp, M_NOWAIT);
3400 			}
3401 			/*
3402 			 * remap should never be >1 unless bus_dmamap_load_mbuf_sg
3403 			 * failed to map an mbuf that was run through m_defrag
3404 			 */
3405 			MPASS(remap <= 1);
3406 			if (__predict_false(m_head == NULL || remap > 1))
3407 				goto defrag_failed;
3408 			remap++;
3409 			*m_headp = m_head;
3410 			goto retry;
3411 			break;
3412 		case ENOMEM:
3413 			txq->ift_no_tx_dma_setup++;
3414 			break;
3415 		default:
3416 			txq->ift_no_tx_dma_setup++;
3417 			m_freem(*m_headp);
3418 			DBG_COUNTER_INC(tx_frees);
3419 			*m_headp = NULL;
3420 			break;
3421 		}
3422 		txq->ift_map_failed++;
3423 		DBG_COUNTER_INC(encap_load_mbuf_fail);
3424 		DBG_COUNTER_INC(encap_txd_encap_fail);
3425 		return (err);
3426 	}
3427 	ifsd_m[pidx] = m_head;
3428 	/*
3429 	 * XXX assumes a 1 to 1 relationship between segments and
3430 	 *        descriptors - this does not hold true on all drivers, e.g.
3431 	 *        cxgb
3432 	 */
3433 	if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) {
3434 		txq->ift_no_desc_avail++;
3435 		bus_dmamap_unload(buf_tag, map);
3436 		DBG_COUNTER_INC(encap_txq_avail_fail);
3437 		DBG_COUNTER_INC(encap_txd_encap_fail);
3438 		if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0)
3439 			GROUPTASK_ENQUEUE(&txq->ift_task);
3440 		return (ENOBUFS);
3441 	}
3442 	/*
3443 	 * On Intel cards we can greatly reduce the number of TX interrupts
3444 	 * we see by only setting report status on every Nth descriptor.
3445 	 * However, this also means that the driver will need to keep track
3446 	 * of the descriptors that RS was set on to check them for the DD bit.
3447 	 */
3448 	txq->ift_rs_pending += nsegs + 1;
3449 	if (txq->ift_rs_pending > TXQ_MAX_RS_DEFERRED(txq) ||
3450 	     iflib_no_tx_batch || (TXQ_AVAIL(txq) - nsegs) <= MAX_TX_DESC(ctx) + 2) {
3451 		pi.ipi_flags |= IPI_TX_INTR;
3452 		txq->ift_rs_pending = 0;
3453 	}
3454 
3455 	pi.ipi_segs = segs;
3456 	pi.ipi_nsegs = nsegs;
3457 
3458 	MPASS(pidx >= 0 && pidx < txq->ift_size);
3459 #ifdef PKT_DEBUG
3460 	print_pkt(&pi);
3461 #endif
3462 	if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) {
3463 		bus_dmamap_sync(buf_tag, map, BUS_DMASYNC_PREWRITE);
3464 		DBG_COUNTER_INC(tx_encap);
3465 		MPASS(pi.ipi_new_pidx < txq->ift_size);
3466 
3467 		ndesc = pi.ipi_new_pidx - pi.ipi_pidx;
3468 		if (pi.ipi_new_pidx < pi.ipi_pidx) {
3469 			ndesc += txq->ift_size;
3470 			txq->ift_gen = 1;
3471 		}
3472 		/*
3473 		 * drivers can need as many as
3474 		 * two sentinels
3475 		 */
3476 		MPASS(ndesc <= pi.ipi_nsegs + 2);
3477 		MPASS(pi.ipi_new_pidx != pidx);
3478 		MPASS(ndesc > 0);
3479 		txq->ift_in_use += ndesc;
3480 
3481 		/*
3482 		 * We update the last software descriptor again here because there may
3483 		 * be a sentinel and/or there may be more mbufs than segments
3484 		 */
3485 		txq->ift_pidx = pi.ipi_new_pidx;
3486 		txq->ift_npending += pi.ipi_ndescs;
3487 	} else {
3488 		*m_headp = m_head = iflib_remove_mbuf(txq);
3489 		if (err == EFBIG) {
3490 			txq->ift_txd_encap_efbig++;
3491 			if (remap < 2) {
3492 				remap = 1;
3493 				goto defrag;
3494 			}
3495 		}
3496 		goto defrag_failed;
3497 	}
3498 	/*
3499 	 * err can't possibly be non-zero here, so we don't neet to test it
3500 	 * to see if we need to DBG_COUNTER_INC(encap_txd_encap_fail).
3501 	 */
3502 	return (err);
3503 
3504 defrag_failed:
3505 	txq->ift_mbuf_defrag_failed++;
3506 	txq->ift_map_failed++;
3507 	m_freem(*m_headp);
3508 	DBG_COUNTER_INC(tx_frees);
3509 	*m_headp = NULL;
3510 	DBG_COUNTER_INC(encap_txd_encap_fail);
3511 	return (ENOMEM);
3512 }
3513 
3514 static void
3515 iflib_tx_desc_free(iflib_txq_t txq, int n)
3516 {
3517 	uint32_t qsize, cidx, mask, gen;
3518 	struct mbuf *m, **ifsd_m;
3519 	bool do_prefetch;
3520 
3521 	cidx = txq->ift_cidx;
3522 	gen = txq->ift_gen;
3523 	qsize = txq->ift_size;
3524 	mask = qsize-1;
3525 	ifsd_m = txq->ift_sds.ifsd_m;
3526 	do_prefetch = (txq->ift_ctx->ifc_flags & IFC_PREFETCH);
3527 
3528 	while (n-- > 0) {
3529 		if (do_prefetch) {
3530 			prefetch(ifsd_m[(cidx + 3) & mask]);
3531 			prefetch(ifsd_m[(cidx + 4) & mask]);
3532 		}
3533 		if ((m = ifsd_m[cidx]) != NULL) {
3534 			prefetch(&ifsd_m[(cidx + CACHE_PTR_INCREMENT) & mask]);
3535 			if (m->m_pkthdr.csum_flags & CSUM_TSO) {
3536 				bus_dmamap_sync(txq->ift_tso_buf_tag,
3537 				    txq->ift_sds.ifsd_tso_map[cidx],
3538 				    BUS_DMASYNC_POSTWRITE);
3539 				bus_dmamap_unload(txq->ift_tso_buf_tag,
3540 				    txq->ift_sds.ifsd_tso_map[cidx]);
3541 			} else {
3542 				bus_dmamap_sync(txq->ift_buf_tag,
3543 				    txq->ift_sds.ifsd_map[cidx],
3544 				    BUS_DMASYNC_POSTWRITE);
3545 				bus_dmamap_unload(txq->ift_buf_tag,
3546 				    txq->ift_sds.ifsd_map[cidx]);
3547 			}
3548 			/* XXX we don't support any drivers that batch packets yet */
3549 			MPASS(m->m_nextpkt == NULL);
3550 			m_freem(m);
3551 			ifsd_m[cidx] = NULL;
3552 #if MEMORY_LOGGING
3553 			txq->ift_dequeued++;
3554 #endif
3555 			DBG_COUNTER_INC(tx_frees);
3556 		}
3557 		if (__predict_false(++cidx == qsize)) {
3558 			cidx = 0;
3559 			gen = 0;
3560 		}
3561 	}
3562 	txq->ift_cidx = cidx;
3563 	txq->ift_gen = gen;
3564 }
3565 
3566 static __inline int
3567 iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh)
3568 {
3569 	int reclaim;
3570 	if_ctx_t ctx = txq->ift_ctx;
3571 
3572 	KASSERT(thresh >= 0, ("invalid threshold to reclaim"));
3573 	MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size);
3574 
3575 	/*
3576 	 * Need a rate-limiting check so that this isn't called every time
3577 	 */
3578 	iflib_tx_credits_update(ctx, txq);
3579 	reclaim = DESC_RECLAIMABLE(txq);
3580 
3581 	if (reclaim <= thresh /* + MAX_TX_DESC(txq->ift_ctx) */) {
3582 #ifdef INVARIANTS
3583 		if (iflib_verbose_debug) {
3584 			printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __FUNCTION__,
3585 			       txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments,
3586 			       reclaim, thresh);
3587 		}
3588 #endif
3589 		return (0);
3590 	}
3591 	iflib_tx_desc_free(txq, reclaim);
3592 	txq->ift_cleaned += reclaim;
3593 	txq->ift_in_use -= reclaim;
3594 
3595 	return (reclaim);
3596 }
3597 
3598 static struct mbuf **
3599 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset, int remaining)
3600 {
3601 	int next, size;
3602 	struct mbuf **items;
3603 
3604 	size = r->size;
3605 	next = (cidx + CACHE_PTR_INCREMENT) & (size-1);
3606 	items = __DEVOLATILE(struct mbuf **, &r->items[0]);
3607 
3608 	prefetch(items[(cidx + offset) & (size-1)]);
3609 	if (remaining > 1) {
3610 		prefetch2cachelines(&items[next]);
3611 		prefetch2cachelines(items[(cidx + offset + 1) & (size-1)]);
3612 		prefetch2cachelines(items[(cidx + offset + 2) & (size-1)]);
3613 		prefetch2cachelines(items[(cidx + offset + 3) & (size-1)]);
3614 	}
3615 	return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (size-1)]));
3616 }
3617 
3618 static void
3619 iflib_txq_check_drain(iflib_txq_t txq, int budget)
3620 {
3621 
3622 	ifmp_ring_check_drainage(txq->ift_br, budget);
3623 }
3624 
3625 static uint32_t
3626 iflib_txq_can_drain(struct ifmp_ring *r)
3627 {
3628 	iflib_txq_t txq = r->cookie;
3629 	if_ctx_t ctx = txq->ift_ctx;
3630 
3631 	if (TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2)
3632 		return (1);
3633 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3634 	    BUS_DMASYNC_POSTREAD);
3635 	return (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id,
3636 	    false));
3637 }
3638 
3639 static uint32_t
3640 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3641 {
3642 	iflib_txq_t txq = r->cookie;
3643 	if_ctx_t ctx = txq->ift_ctx;
3644 	if_t ifp = ctx->ifc_ifp;
3645 	struct mbuf *m, **mp;
3646 	int avail, bytes_sent, consumed, count, err, i, in_use_prev;
3647 	int mcast_sent, pkt_sent, reclaimed, txq_avail;
3648 	bool do_prefetch, rang, ring;
3649 
3650 	if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) ||
3651 			    !LINK_ACTIVE(ctx))) {
3652 		DBG_COUNTER_INC(txq_drain_notready);
3653 		return (0);
3654 	}
3655 	reclaimed = iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
3656 	rang = iflib_txd_db_check(ctx, txq, reclaimed, txq->ift_in_use);
3657 	avail = IDXDIFF(pidx, cidx, r->size);
3658 	if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) {
3659 		DBG_COUNTER_INC(txq_drain_flushing);
3660 		for (i = 0; i < avail; i++) {
3661 			if (__predict_true(r->items[(cidx + i) & (r->size-1)] != (void *)txq))
3662 				m_free(r->items[(cidx + i) & (r->size-1)]);
3663 			r->items[(cidx + i) & (r->size-1)] = NULL;
3664 		}
3665 		return (avail);
3666 	}
3667 
3668 	if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3669 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3670 		CALLOUT_LOCK(txq);
3671 		callout_stop(&txq->ift_timer);
3672 		CALLOUT_UNLOCK(txq);
3673 		DBG_COUNTER_INC(txq_drain_oactive);
3674 		return (0);
3675 	}
3676 	if (reclaimed)
3677 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3678 	consumed = mcast_sent = bytes_sent = pkt_sent = 0;
3679 	count = MIN(avail, TX_BATCH_SIZE);
3680 #ifdef INVARIANTS
3681 	if (iflib_verbose_debug)
3682 		printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __FUNCTION__,
3683 		       avail, ctx->ifc_flags, TXQ_AVAIL(txq));
3684 #endif
3685 	do_prefetch = (ctx->ifc_flags & IFC_PREFETCH);
3686 	txq_avail = TXQ_AVAIL(txq);
3687 	err = 0;
3688 	for (i = 0; i < count && txq_avail > MAX_TX_DESC(ctx) + 2; i++) {
3689 		int rem = do_prefetch ? count - i : 0;
3690 
3691 		mp = _ring_peek_one(r, cidx, i, rem);
3692 		MPASS(mp != NULL && *mp != NULL);
3693 		if (__predict_false(*mp == (struct mbuf *)txq)) {
3694 			consumed++;
3695 			continue;
3696 		}
3697 		in_use_prev = txq->ift_in_use;
3698 		err = iflib_encap(txq, mp);
3699 		if (__predict_false(err)) {
3700 			/* no room - bail out */
3701 			if (err == ENOBUFS)
3702 				break;
3703 			consumed++;
3704 			/* we can't send this packet - skip it */
3705 			continue;
3706 		}
3707 		consumed++;
3708 		pkt_sent++;
3709 		m = *mp;
3710 		DBG_COUNTER_INC(tx_sent);
3711 		bytes_sent += m->m_pkthdr.len;
3712 		mcast_sent += !!(m->m_flags & M_MCAST);
3713 		txq_avail = TXQ_AVAIL(txq);
3714 
3715 		txq->ift_db_pending += (txq->ift_in_use - in_use_prev);
3716 		ETHER_BPF_MTAP(ifp, m);
3717 		if (__predict_false(!(ifp->if_drv_flags & IFF_DRV_RUNNING)))
3718 			break;
3719 		rang = iflib_txd_db_check(ctx, txq, false, in_use_prev);
3720 	}
3721 
3722 	/* deliberate use of bitwise or to avoid gratuitous short-circuit */
3723 	ring = rang ? false  : (iflib_min_tx_latency | err) || (TXQ_AVAIL(txq) < MAX_TX_DESC(ctx));
3724 	iflib_txd_db_check(ctx, txq, ring, txq->ift_in_use);
3725 	if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
3726 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
3727 	if (mcast_sent)
3728 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
3729 #ifdef INVARIANTS
3730 	if (iflib_verbose_debug)
3731 		printf("consumed=%d\n", consumed);
3732 #endif
3733 	return (consumed);
3734 }
3735 
3736 static uint32_t
3737 iflib_txq_drain_always(struct ifmp_ring *r)
3738 {
3739 	return (1);
3740 }
3741 
3742 static uint32_t
3743 iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3744 {
3745 	int i, avail;
3746 	struct mbuf **mp;
3747 	iflib_txq_t txq;
3748 
3749 	txq = r->cookie;
3750 
3751 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3752 	CALLOUT_LOCK(txq);
3753 	callout_stop(&txq->ift_timer);
3754 	CALLOUT_UNLOCK(txq);
3755 
3756 	avail = IDXDIFF(pidx, cidx, r->size);
3757 	for (i = 0; i < avail; i++) {
3758 		mp = _ring_peek_one(r, cidx, i, avail - i);
3759 		if (__predict_false(*mp == (struct mbuf *)txq))
3760 			continue;
3761 		m_freem(*mp);
3762 		DBG_COUNTER_INC(tx_frees);
3763 	}
3764 	MPASS(ifmp_ring_is_stalled(r) == 0);
3765 	return (avail);
3766 }
3767 
3768 static void
3769 iflib_ifmp_purge(iflib_txq_t txq)
3770 {
3771 	struct ifmp_ring *r;
3772 
3773 	r = txq->ift_br;
3774 	r->drain = iflib_txq_drain_free;
3775 	r->can_drain = iflib_txq_drain_always;
3776 
3777 	ifmp_ring_check_drainage(r, r->size);
3778 
3779 	r->drain = iflib_txq_drain;
3780 	r->can_drain = iflib_txq_can_drain;
3781 }
3782 
3783 static void
3784 _task_fn_tx(void *context)
3785 {
3786 	iflib_txq_t txq = context;
3787 	if_ctx_t ctx = txq->ift_ctx;
3788 	if_t ifp = ctx->ifc_ifp;
3789 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
3790 
3791 #ifdef IFLIB_DIAGNOSTICS
3792 	txq->ift_cpu_exec_count[curcpu]++;
3793 #endif
3794 	if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
3795 		return;
3796 #ifdef DEV_NETMAP
3797 	if ((if_getcapenable(ifp) & IFCAP_NETMAP) &&
3798 	    netmap_tx_irq(ifp, txq->ift_id))
3799 		goto skip_ifmp;
3800 #endif
3801 #ifdef ALTQ
3802 	if (ALTQ_IS_ENABLED(&ifp->if_snd))
3803 		iflib_altq_if_start(ifp);
3804 #endif
3805 	if (txq->ift_db_pending)
3806 		ifmp_ring_enqueue(txq->ift_br, (void **)&txq, 1, TX_BATCH_SIZE, abdicate);
3807 	else if (!abdicate)
3808 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3809 	/*
3810 	 * When abdicating, we always need to check drainage, not just when we don't enqueue
3811 	 */
3812 	if (abdicate)
3813 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3814 #ifdef DEV_NETMAP
3815 skip_ifmp:
3816 #endif
3817 	if (ctx->ifc_flags & IFC_LEGACY)
3818 		IFDI_INTR_ENABLE(ctx);
3819 	else
3820 		IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3821 }
3822 
3823 static void
3824 _task_fn_rx(void *context)
3825 {
3826 	iflib_rxq_t rxq = context;
3827 	if_ctx_t ctx = rxq->ifr_ctx;
3828 	uint8_t more;
3829 	uint16_t budget;
3830 #ifdef DEV_NETMAP
3831 	u_int work = 0;
3832 	int nmirq;
3833 #endif
3834 
3835 #ifdef IFLIB_DIAGNOSTICS
3836 	rxq->ifr_cpu_exec_count[curcpu]++;
3837 #endif
3838 	DBG_COUNTER_INC(task_fn_rxs);
3839 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3840 		return;
3841 #ifdef DEV_NETMAP
3842 	nmirq = netmap_rx_irq(ctx->ifc_ifp, rxq->ifr_id, &work);
3843 	if (nmirq != NM_IRQ_PASS) {
3844 		more = (nmirq == NM_IRQ_RESCHED) ? IFLIB_RXEOF_MORE : 0;
3845 		goto skip_rxeof;
3846 	}
3847 #endif
3848 	budget = ctx->ifc_sysctl_rx_budget;
3849 	if (budget == 0)
3850 		budget = 16;	/* XXX */
3851 	more = iflib_rxeof(rxq, budget);
3852 #ifdef DEV_NETMAP
3853 skip_rxeof:
3854 #endif
3855 	if ((more & IFLIB_RXEOF_MORE) == 0) {
3856 		if (ctx->ifc_flags & IFC_LEGACY)
3857 			IFDI_INTR_ENABLE(ctx);
3858 		else
3859 			IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
3860 		DBG_COUNTER_INC(rx_intr_enables);
3861 	}
3862 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3863 		return;
3864 
3865 	if (more & IFLIB_RXEOF_MORE)
3866 		GROUPTASK_ENQUEUE(&rxq->ifr_task);
3867 	else if (more & IFLIB_RXEOF_EMPTY)
3868 		callout_reset_curcpu(&rxq->ifr_watchdog, 1, &_task_fn_rx_watchdog, rxq);
3869 }
3870 
3871 static void
3872 _task_fn_admin(void *context)
3873 {
3874 	if_ctx_t ctx = context;
3875 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
3876 	iflib_txq_t txq;
3877 	int i;
3878 	bool oactive, running, do_reset, do_watchdog, in_detach;
3879 
3880 	STATE_LOCK(ctx);
3881 	running = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING);
3882 	oactive = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE);
3883 	do_reset = (ctx->ifc_flags & IFC_DO_RESET);
3884 	do_watchdog = (ctx->ifc_flags & IFC_DO_WATCHDOG);
3885 	in_detach = (ctx->ifc_flags & IFC_IN_DETACH);
3886 	ctx->ifc_flags &= ~(IFC_DO_RESET|IFC_DO_WATCHDOG);
3887 	STATE_UNLOCK(ctx);
3888 
3889 	if ((!running && !oactive) && !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
3890 		return;
3891 	if (in_detach)
3892 		return;
3893 
3894 	CTX_LOCK(ctx);
3895 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3896 		CALLOUT_LOCK(txq);
3897 		callout_stop(&txq->ift_timer);
3898 		CALLOUT_UNLOCK(txq);
3899 	}
3900 	if (do_watchdog) {
3901 		ctx->ifc_watchdog_events++;
3902 		IFDI_WATCHDOG_RESET(ctx);
3903 	}
3904 	IFDI_UPDATE_ADMIN_STATUS(ctx);
3905 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3906 		callout_reset_on(&txq->ift_timer, hz / 2, iflib_timer, txq,
3907 		    txq->ift_timer.c_cpu);
3908 	}
3909 	IFDI_LINK_INTR_ENABLE(ctx);
3910 	if (do_reset)
3911 		iflib_if_init_locked(ctx);
3912 	CTX_UNLOCK(ctx);
3913 
3914 	if (LINK_ACTIVE(ctx) == 0)
3915 		return;
3916 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3917 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
3918 }
3919 
3920 static void
3921 _task_fn_iov(void *context)
3922 {
3923 	if_ctx_t ctx = context;
3924 
3925 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING) &&
3926 	    !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
3927 		return;
3928 
3929 	CTX_LOCK(ctx);
3930 	IFDI_VFLR_HANDLE(ctx);
3931 	CTX_UNLOCK(ctx);
3932 }
3933 
3934 static int
3935 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)
3936 {
3937 	int err;
3938 	if_int_delay_info_t info;
3939 	if_ctx_t ctx;
3940 
3941 	info = (if_int_delay_info_t)arg1;
3942 	ctx = info->iidi_ctx;
3943 	info->iidi_req = req;
3944 	info->iidi_oidp = oidp;
3945 	CTX_LOCK(ctx);
3946 	err = IFDI_SYSCTL_INT_DELAY(ctx, info);
3947 	CTX_UNLOCK(ctx);
3948 	return (err);
3949 }
3950 
3951 /*********************************************************************
3952  *
3953  *  IFNET FUNCTIONS
3954  *
3955  **********************************************************************/
3956 
3957 static void
3958 iflib_if_init_locked(if_ctx_t ctx)
3959 {
3960 	iflib_stop(ctx);
3961 	iflib_init_locked(ctx);
3962 }
3963 
3964 static void
3965 iflib_if_init(void *arg)
3966 {
3967 	if_ctx_t ctx = arg;
3968 
3969 	CTX_LOCK(ctx);
3970 	iflib_if_init_locked(ctx);
3971 	CTX_UNLOCK(ctx);
3972 }
3973 
3974 static int
3975 iflib_if_transmit(if_t ifp, struct mbuf *m)
3976 {
3977 	if_ctx_t	ctx = if_getsoftc(ifp);
3978 
3979 	iflib_txq_t txq;
3980 	int err, qidx;
3981 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
3982 
3983 	if (__predict_false((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) {
3984 		DBG_COUNTER_INC(tx_frees);
3985 		m_freem(m);
3986 		return (ENETDOWN);
3987 	}
3988 
3989 	MPASS(m->m_nextpkt == NULL);
3990 	/* ALTQ-enabled interfaces always use queue 0. */
3991 	qidx = 0;
3992 	if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m) && !ALTQ_IS_ENABLED(&ifp->if_snd))
3993 		qidx = QIDX(ctx, m);
3994 	/*
3995 	 * XXX calculate buf_ring based on flowid (divvy up bits?)
3996 	 */
3997 	txq = &ctx->ifc_txqs[qidx];
3998 
3999 #ifdef DRIVER_BACKPRESSURE
4000 	if (txq->ift_closed) {
4001 		while (m != NULL) {
4002 			next = m->m_nextpkt;
4003 			m->m_nextpkt = NULL;
4004 			m_freem(m);
4005 			DBG_COUNTER_INC(tx_frees);
4006 			m = next;
4007 		}
4008 		return (ENOBUFS);
4009 	}
4010 #endif
4011 #ifdef notyet
4012 	qidx = count = 0;
4013 	mp = marr;
4014 	next = m;
4015 	do {
4016 		count++;
4017 		next = next->m_nextpkt;
4018 	} while (next != NULL);
4019 
4020 	if (count > nitems(marr))
4021 		if ((mp = malloc(count*sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) {
4022 			/* XXX check nextpkt */
4023 			m_freem(m);
4024 			/* XXX simplify for now */
4025 			DBG_COUNTER_INC(tx_frees);
4026 			return (ENOBUFS);
4027 		}
4028 	for (next = m, i = 0; next != NULL; i++) {
4029 		mp[i] = next;
4030 		next = next->m_nextpkt;
4031 		mp[i]->m_nextpkt = NULL;
4032 	}
4033 #endif
4034 	DBG_COUNTER_INC(tx_seen);
4035 	err = ifmp_ring_enqueue(txq->ift_br, (void **)&m, 1, TX_BATCH_SIZE, abdicate);
4036 
4037 	if (abdicate)
4038 		GROUPTASK_ENQUEUE(&txq->ift_task);
4039  	if (err) {
4040 		if (!abdicate)
4041 			GROUPTASK_ENQUEUE(&txq->ift_task);
4042 		/* support forthcoming later */
4043 #ifdef DRIVER_BACKPRESSURE
4044 		txq->ift_closed = TRUE;
4045 #endif
4046 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
4047 		m_freem(m);
4048 		DBG_COUNTER_INC(tx_frees);
4049 	}
4050 
4051 	return (err);
4052 }
4053 
4054 #ifdef ALTQ
4055 /*
4056  * The overall approach to integrating iflib with ALTQ is to continue to use
4057  * the iflib mp_ring machinery between the ALTQ queue(s) and the hardware
4058  * ring.  Technically, when using ALTQ, queueing to an intermediate mp_ring
4059  * is redundant/unnecessary, but doing so minimizes the amount of
4060  * ALTQ-specific code required in iflib.  It is assumed that the overhead of
4061  * redundantly queueing to an intermediate mp_ring is swamped by the
4062  * performance limitations inherent in using ALTQ.
4063  *
4064  * When ALTQ support is compiled in, all iflib drivers will use a transmit
4065  * routine, iflib_altq_if_transmit(), that checks if ALTQ is enabled for the
4066  * given interface.  If ALTQ is enabled for an interface, then all
4067  * transmitted packets for that interface will be submitted to the ALTQ
4068  * subsystem via IFQ_ENQUEUE().  We don't use the legacy if_transmit()
4069  * implementation because it uses IFQ_HANDOFF(), which will duplicatively
4070  * update stats that the iflib machinery handles, and which is sensitve to
4071  * the disused IFF_DRV_OACTIVE flag.  Additionally, iflib_altq_if_start()
4072  * will be installed as the start routine for use by ALTQ facilities that
4073  * need to trigger queue drains on a scheduled basis.
4074  *
4075  */
4076 static void
4077 iflib_altq_if_start(if_t ifp)
4078 {
4079 	struct ifaltq *ifq = &ifp->if_snd;
4080 	struct mbuf *m;
4081 
4082 	IFQ_LOCK(ifq);
4083 	IFQ_DEQUEUE_NOLOCK(ifq, m);
4084 	while (m != NULL) {
4085 		iflib_if_transmit(ifp, m);
4086 		IFQ_DEQUEUE_NOLOCK(ifq, m);
4087 	}
4088 	IFQ_UNLOCK(ifq);
4089 }
4090 
4091 static int
4092 iflib_altq_if_transmit(if_t ifp, struct mbuf *m)
4093 {
4094 	int err;
4095 
4096 	if (ALTQ_IS_ENABLED(&ifp->if_snd)) {
4097 		IFQ_ENQUEUE(&ifp->if_snd, m, err);
4098 		if (err == 0)
4099 			iflib_altq_if_start(ifp);
4100 	} else
4101 		err = iflib_if_transmit(ifp, m);
4102 
4103 	return (err);
4104 }
4105 #endif /* ALTQ */
4106 
4107 static void
4108 iflib_if_qflush(if_t ifp)
4109 {
4110 	if_ctx_t ctx = if_getsoftc(ifp);
4111 	iflib_txq_t txq = ctx->ifc_txqs;
4112 	int i;
4113 
4114 	STATE_LOCK(ctx);
4115 	ctx->ifc_flags |= IFC_QFLUSH;
4116 	STATE_UNLOCK(ctx);
4117 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
4118 		while (!(ifmp_ring_is_idle(txq->ift_br) || ifmp_ring_is_stalled(txq->ift_br)))
4119 			iflib_txq_check_drain(txq, 0);
4120 	STATE_LOCK(ctx);
4121 	ctx->ifc_flags &= ~IFC_QFLUSH;
4122 	STATE_UNLOCK(ctx);
4123 
4124 	/*
4125 	 * When ALTQ is enabled, this will also take care of purging the
4126 	 * ALTQ queue(s).
4127 	 */
4128 	if_qflush(ifp);
4129 }
4130 
4131 #define IFCAP_FLAGS (IFCAP_HWCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \
4132 		     IFCAP_TSO | IFCAP_VLAN_HWTAGGING | IFCAP_HWSTATS | \
4133 		     IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | \
4134 		     IFCAP_VLAN_HWTSO | IFCAP_VLAN_HWCSUM | IFCAP_NOMAP)
4135 
4136 static int
4137 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data)
4138 {
4139 	if_ctx_t ctx = if_getsoftc(ifp);
4140 	struct ifreq	*ifr = (struct ifreq *)data;
4141 #if defined(INET) || defined(INET6)
4142 	struct ifaddr	*ifa = (struct ifaddr *)data;
4143 #endif
4144 	bool		avoid_reset = false;
4145 	int		err = 0, reinit = 0, bits;
4146 
4147 	switch (command) {
4148 	case SIOCSIFADDR:
4149 #ifdef INET
4150 		if (ifa->ifa_addr->sa_family == AF_INET)
4151 			avoid_reset = true;
4152 #endif
4153 #ifdef INET6
4154 		if (ifa->ifa_addr->sa_family == AF_INET6)
4155 			avoid_reset = true;
4156 #endif
4157 		/*
4158 		** Calling init results in link renegotiation,
4159 		** so we avoid doing it when possible.
4160 		*/
4161 		if (avoid_reset) {
4162 			if_setflagbits(ifp, IFF_UP,0);
4163 			if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
4164 				reinit = 1;
4165 #ifdef INET
4166 			if (!(if_getflags(ifp) & IFF_NOARP))
4167 				arp_ifinit(ifp, ifa);
4168 #endif
4169 		} else
4170 			err = ether_ioctl(ifp, command, data);
4171 		break;
4172 	case SIOCSIFMTU:
4173 		CTX_LOCK(ctx);
4174 		if (ifr->ifr_mtu == if_getmtu(ifp)) {
4175 			CTX_UNLOCK(ctx);
4176 			break;
4177 		}
4178 		bits = if_getdrvflags(ifp);
4179 		/* stop the driver and free any clusters before proceeding */
4180 		iflib_stop(ctx);
4181 
4182 		if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) {
4183 			STATE_LOCK(ctx);
4184 			if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size)
4185 				ctx->ifc_flags |= IFC_MULTISEG;
4186 			else
4187 				ctx->ifc_flags &= ~IFC_MULTISEG;
4188 			STATE_UNLOCK(ctx);
4189 			err = if_setmtu(ifp, ifr->ifr_mtu);
4190 		}
4191 		iflib_init_locked(ctx);
4192 		STATE_LOCK(ctx);
4193 		if_setdrvflags(ifp, bits);
4194 		STATE_UNLOCK(ctx);
4195 		CTX_UNLOCK(ctx);
4196 		break;
4197 	case SIOCSIFFLAGS:
4198 		CTX_LOCK(ctx);
4199 		if (if_getflags(ifp) & IFF_UP) {
4200 			if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4201 				if ((if_getflags(ifp) ^ ctx->ifc_if_flags) &
4202 				    (IFF_PROMISC | IFF_ALLMULTI)) {
4203 					CTX_UNLOCK(ctx);
4204 					err = IFDI_PROMISC_SET(ctx, if_getflags(ifp));
4205 					CTX_LOCK(ctx);
4206 				}
4207 			} else
4208 				reinit = 1;
4209 		} else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4210 			iflib_stop(ctx);
4211 		}
4212 		ctx->ifc_if_flags = if_getflags(ifp);
4213 		CTX_UNLOCK(ctx);
4214 		break;
4215 	case SIOCADDMULTI:
4216 	case SIOCDELMULTI:
4217 		if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4218 			CTX_LOCK(ctx);
4219 			IFDI_INTR_DISABLE(ctx);
4220 			IFDI_MULTI_SET(ctx);
4221 			IFDI_INTR_ENABLE(ctx);
4222 			CTX_UNLOCK(ctx);
4223 		}
4224 		break;
4225 	case SIOCSIFMEDIA:
4226 		CTX_LOCK(ctx);
4227 		IFDI_MEDIA_SET(ctx);
4228 		CTX_UNLOCK(ctx);
4229 		/* FALLTHROUGH */
4230 	case SIOCGIFMEDIA:
4231 	case SIOCGIFXMEDIA:
4232 		err = ifmedia_ioctl(ifp, ifr, ctx->ifc_mediap, command);
4233 		break;
4234 	case SIOCGI2C:
4235 	{
4236 		struct ifi2creq i2c;
4237 
4238 		err = copyin(ifr_data_get_ptr(ifr), &i2c, sizeof(i2c));
4239 		if (err != 0)
4240 			break;
4241 		if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) {
4242 			err = EINVAL;
4243 			break;
4244 		}
4245 		if (i2c.len > sizeof(i2c.data)) {
4246 			err = EINVAL;
4247 			break;
4248 		}
4249 
4250 		if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0)
4251 			err = copyout(&i2c, ifr_data_get_ptr(ifr),
4252 			    sizeof(i2c));
4253 		break;
4254 	}
4255 	case SIOCSIFCAP:
4256 	{
4257 		int mask, setmask, oldmask;
4258 
4259 		oldmask = if_getcapenable(ifp);
4260 		mask = ifr->ifr_reqcap ^ oldmask;
4261 		mask &= ctx->ifc_softc_ctx.isc_capabilities | IFCAP_NOMAP;
4262 		setmask = 0;
4263 #ifdef TCP_OFFLOAD
4264 		setmask |= mask & (IFCAP_TOE4|IFCAP_TOE6);
4265 #endif
4266 		setmask |= (mask & IFCAP_FLAGS);
4267 		setmask |= (mask & IFCAP_WOL);
4268 
4269 		/*
4270 		 * If any RX csum has changed, change all the ones that
4271 		 * are supported by the driver.
4272 		 */
4273 		if (setmask & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6)) {
4274 			setmask |= ctx->ifc_softc_ctx.isc_capabilities &
4275 			    (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6);
4276 		}
4277 
4278 		/*
4279 		 * want to ensure that traffic has stopped before we change any of the flags
4280 		 */
4281 		if (setmask) {
4282 			CTX_LOCK(ctx);
4283 			bits = if_getdrvflags(ifp);
4284 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4285 				iflib_stop(ctx);
4286 			STATE_LOCK(ctx);
4287 			if_togglecapenable(ifp, setmask);
4288 			STATE_UNLOCK(ctx);
4289 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4290 				iflib_init_locked(ctx);
4291 			STATE_LOCK(ctx);
4292 			if_setdrvflags(ifp, bits);
4293 			STATE_UNLOCK(ctx);
4294 			CTX_UNLOCK(ctx);
4295 		}
4296 		if_vlancap(ifp);
4297 		break;
4298 	}
4299 	case SIOCGPRIVATE_0:
4300 	case SIOCSDRVSPEC:
4301 	case SIOCGDRVSPEC:
4302 		CTX_LOCK(ctx);
4303 		err = IFDI_PRIV_IOCTL(ctx, command, data);
4304 		CTX_UNLOCK(ctx);
4305 		break;
4306 	default:
4307 		err = ether_ioctl(ifp, command, data);
4308 		break;
4309 	}
4310 	if (reinit)
4311 		iflib_if_init(ctx);
4312 	return (err);
4313 }
4314 
4315 static uint64_t
4316 iflib_if_get_counter(if_t ifp, ift_counter cnt)
4317 {
4318 	if_ctx_t ctx = if_getsoftc(ifp);
4319 
4320 	return (IFDI_GET_COUNTER(ctx, cnt));
4321 }
4322 
4323 /*********************************************************************
4324  *
4325  *  OTHER FUNCTIONS EXPORTED TO THE STACK
4326  *
4327  **********************************************************************/
4328 
4329 static void
4330 iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag)
4331 {
4332 	if_ctx_t ctx = if_getsoftc(ifp);
4333 
4334 	if ((void *)ctx != arg)
4335 		return;
4336 
4337 	if ((vtag == 0) || (vtag > 4095))
4338 		return;
4339 
4340 	if (iflib_in_detach(ctx))
4341 		return;
4342 
4343 	CTX_LOCK(ctx);
4344 	/* Driver may need all untagged packets to be flushed */
4345 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4346 		iflib_stop(ctx);
4347 	IFDI_VLAN_REGISTER(ctx, vtag);
4348 	/* Re-init to load the changes, if required */
4349 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4350 		iflib_init_locked(ctx);
4351 	CTX_UNLOCK(ctx);
4352 }
4353 
4354 static void
4355 iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag)
4356 {
4357 	if_ctx_t ctx = if_getsoftc(ifp);
4358 
4359 	if ((void *)ctx != arg)
4360 		return;
4361 
4362 	if ((vtag == 0) || (vtag > 4095))
4363 		return;
4364 
4365 	CTX_LOCK(ctx);
4366 	/* Driver may need all tagged packets to be flushed */
4367 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4368 		iflib_stop(ctx);
4369 	IFDI_VLAN_UNREGISTER(ctx, vtag);
4370 	/* Re-init to load the changes, if required */
4371 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4372 		iflib_init_locked(ctx);
4373 	CTX_UNLOCK(ctx);
4374 }
4375 
4376 static void
4377 iflib_led_func(void *arg, int onoff)
4378 {
4379 	if_ctx_t ctx = arg;
4380 
4381 	CTX_LOCK(ctx);
4382 	IFDI_LED_FUNC(ctx, onoff);
4383 	CTX_UNLOCK(ctx);
4384 }
4385 
4386 /*********************************************************************
4387  *
4388  *  BUS FUNCTION DEFINITIONS
4389  *
4390  **********************************************************************/
4391 
4392 int
4393 iflib_device_probe(device_t dev)
4394 {
4395 	const pci_vendor_info_t *ent;
4396 	if_shared_ctx_t sctx;
4397 	uint16_t pci_device_id, pci_rev_id, pci_subdevice_id, pci_subvendor_id;
4398 	uint16_t pci_vendor_id;
4399 
4400 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4401 		return (ENOTSUP);
4402 
4403 	pci_vendor_id = pci_get_vendor(dev);
4404 	pci_device_id = pci_get_device(dev);
4405 	pci_subvendor_id = pci_get_subvendor(dev);
4406 	pci_subdevice_id = pci_get_subdevice(dev);
4407 	pci_rev_id = pci_get_revid(dev);
4408 	if (sctx->isc_parse_devinfo != NULL)
4409 		sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id);
4410 
4411 	ent = sctx->isc_vendor_info;
4412 	while (ent->pvi_vendor_id != 0) {
4413 		if (pci_vendor_id != ent->pvi_vendor_id) {
4414 			ent++;
4415 			continue;
4416 		}
4417 		if ((pci_device_id == ent->pvi_device_id) &&
4418 		    ((pci_subvendor_id == ent->pvi_subvendor_id) ||
4419 		     (ent->pvi_subvendor_id == 0)) &&
4420 		    ((pci_subdevice_id == ent->pvi_subdevice_id) ||
4421 		     (ent->pvi_subdevice_id == 0)) &&
4422 		    ((pci_rev_id == ent->pvi_rev_id) ||
4423 		     (ent->pvi_rev_id == 0))) {
4424 			device_set_desc_copy(dev, ent->pvi_name);
4425 			/* this needs to be changed to zero if the bus probing code
4426 			 * ever stops re-probing on best match because the sctx
4427 			 * may have its values over written by register calls
4428 			 * in subsequent probes
4429 			 */
4430 			return (BUS_PROBE_DEFAULT);
4431 		}
4432 		ent++;
4433 	}
4434 	return (ENXIO);
4435 }
4436 
4437 int
4438 iflib_device_probe_vendor(device_t dev)
4439 {
4440 	int probe;
4441 
4442 	probe = iflib_device_probe(dev);
4443 	if (probe == BUS_PROBE_DEFAULT)
4444 		return (BUS_PROBE_VENDOR);
4445 	else
4446 		return (probe);
4447 }
4448 
4449 static void
4450 iflib_reset_qvalues(if_ctx_t ctx)
4451 {
4452 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4453 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4454 	device_t dev = ctx->ifc_dev;
4455 	int i;
4456 
4457 	if (ctx->ifc_sysctl_ntxqs != 0)
4458 		scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs;
4459 	if (ctx->ifc_sysctl_nrxqs != 0)
4460 		scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs;
4461 
4462 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4463 		if (ctx->ifc_sysctl_ntxds[i] != 0)
4464 			scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i];
4465 		else
4466 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4467 	}
4468 
4469 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4470 		if (ctx->ifc_sysctl_nrxds[i] != 0)
4471 			scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i];
4472 		else
4473 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4474 	}
4475 
4476 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4477 		if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) {
4478 			device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n",
4479 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]);
4480 			scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i];
4481 		}
4482 		if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) {
4483 			device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n",
4484 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]);
4485 			scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i];
4486 		}
4487 		if (!powerof2(scctx->isc_nrxd[i])) {
4488 			device_printf(dev, "nrxd%d: %d is not a power of 2 - using default value of %d\n",
4489 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_default[i]);
4490 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4491 		}
4492 	}
4493 
4494 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4495 		if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) {
4496 			device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n",
4497 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]);
4498 			scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i];
4499 		}
4500 		if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) {
4501 			device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n",
4502 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]);
4503 			scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i];
4504 		}
4505 		if (!powerof2(scctx->isc_ntxd[i])) {
4506 			device_printf(dev, "ntxd%d: %d is not a power of 2 - using default value of %d\n",
4507 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_default[i]);
4508 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4509 		}
4510 	}
4511 }
4512 
4513 static void
4514 iflib_add_pfil(if_ctx_t ctx)
4515 {
4516 	struct pfil_head *pfil;
4517 	struct pfil_head_args pa;
4518 	iflib_rxq_t rxq;
4519 	int i;
4520 
4521 	pa.pa_version = PFIL_VERSION;
4522 	pa.pa_flags = PFIL_IN;
4523 	pa.pa_type = PFIL_TYPE_ETHERNET;
4524 	pa.pa_headname = ctx->ifc_ifp->if_xname;
4525 	pfil = pfil_head_register(&pa);
4526 
4527 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4528 		rxq->pfil = pfil;
4529 	}
4530 }
4531 
4532 static void
4533 iflib_rem_pfil(if_ctx_t ctx)
4534 {
4535 	struct pfil_head *pfil;
4536 	iflib_rxq_t rxq;
4537 	int i;
4538 
4539 	rxq = ctx->ifc_rxqs;
4540 	pfil = rxq->pfil;
4541 	for (i = 0; i < NRXQSETS(ctx); i++, rxq++) {
4542 		rxq->pfil = NULL;
4543 	}
4544 	pfil_head_unregister(pfil);
4545 }
4546 
4547 static uint16_t
4548 get_ctx_core_offset(if_ctx_t ctx)
4549 {
4550 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4551 	struct cpu_offset *op;
4552 	uint16_t qc;
4553 	uint16_t ret = ctx->ifc_sysctl_core_offset;
4554 
4555 	if (ret != CORE_OFFSET_UNSPECIFIED)
4556 		return (ret);
4557 
4558 	if (ctx->ifc_sysctl_separate_txrx)
4559 		qc = scctx->isc_ntxqsets + scctx->isc_nrxqsets;
4560 	else
4561 		qc = max(scctx->isc_ntxqsets, scctx->isc_nrxqsets);
4562 
4563 	mtx_lock(&cpu_offset_mtx);
4564 	SLIST_FOREACH(op, &cpu_offsets, entries) {
4565 		if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) {
4566 			ret = op->offset;
4567 			op->offset += qc;
4568 			MPASS(op->refcount < UINT_MAX);
4569 			op->refcount++;
4570 			break;
4571 		}
4572 	}
4573 	if (ret == CORE_OFFSET_UNSPECIFIED) {
4574 		ret = 0;
4575 		op = malloc(sizeof(struct cpu_offset), M_IFLIB,
4576 		    M_NOWAIT | M_ZERO);
4577 		if (op == NULL) {
4578 			device_printf(ctx->ifc_dev,
4579 			    "allocation for cpu offset failed.\n");
4580 		} else {
4581 			op->offset = qc;
4582 			op->refcount = 1;
4583 			CPU_COPY(&ctx->ifc_cpus, &op->set);
4584 			SLIST_INSERT_HEAD(&cpu_offsets, op, entries);
4585 		}
4586 	}
4587 	mtx_unlock(&cpu_offset_mtx);
4588 
4589 	return (ret);
4590 }
4591 
4592 static void
4593 unref_ctx_core_offset(if_ctx_t ctx)
4594 {
4595 	struct cpu_offset *op, *top;
4596 
4597 	mtx_lock(&cpu_offset_mtx);
4598 	SLIST_FOREACH_SAFE(op, &cpu_offsets, entries, top) {
4599 		if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) {
4600 			MPASS(op->refcount > 0);
4601 			op->refcount--;
4602 			if (op->refcount == 0) {
4603 				SLIST_REMOVE(&cpu_offsets, op, cpu_offset, entries);
4604 				free(op, M_IFLIB);
4605 			}
4606 			break;
4607 		}
4608 	}
4609 	mtx_unlock(&cpu_offset_mtx);
4610 }
4611 
4612 int
4613 iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp)
4614 {
4615 	if_ctx_t ctx;
4616 	if_t ifp;
4617 	if_softc_ctx_t scctx;
4618 	kobjop_desc_t kobj_desc;
4619 	kobj_method_t *kobj_method;
4620 	int err, msix, rid;
4621 	int num_txd, num_rxd;
4622 
4623 	ctx = malloc(sizeof(* ctx), M_IFLIB, M_WAITOK|M_ZERO);
4624 
4625 	if (sc == NULL) {
4626 		sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4627 		device_set_softc(dev, ctx);
4628 		ctx->ifc_flags |= IFC_SC_ALLOCATED;
4629 	}
4630 
4631 	ctx->ifc_sctx = sctx;
4632 	ctx->ifc_dev = dev;
4633 	ctx->ifc_softc = sc;
4634 
4635 	if ((err = iflib_register(ctx)) != 0) {
4636 		device_printf(dev, "iflib_register failed %d\n", err);
4637 		goto fail_ctx_free;
4638 	}
4639 	iflib_add_device_sysctl_pre(ctx);
4640 
4641 	scctx = &ctx->ifc_softc_ctx;
4642 	ifp = ctx->ifc_ifp;
4643 
4644 	iflib_reset_qvalues(ctx);
4645 	CTX_LOCK(ctx);
4646 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4647 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4648 		goto fail_unlock;
4649 	}
4650 	_iflib_pre_assert(scctx);
4651 	ctx->ifc_txrx = *scctx->isc_txrx;
4652 
4653 	if (sctx->isc_flags & IFLIB_DRIVER_MEDIA)
4654 		ctx->ifc_mediap = scctx->isc_media;
4655 
4656 #ifdef INVARIANTS
4657 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
4658 		MPASS(scctx->isc_tx_csum_flags);
4659 #endif
4660 
4661 	if_setcapabilities(ifp,
4662 	    scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_NOMAP);
4663 	if_setcapenable(ifp,
4664 	    scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_NOMAP);
4665 
4666 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4667 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4668 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4669 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4670 
4671 	num_txd = iflib_num_tx_descs(ctx);
4672 	num_rxd = iflib_num_rx_descs(ctx);
4673 
4674 	/* XXX change for per-queue sizes */
4675 	device_printf(dev, "Using %d TX descriptors and %d RX descriptors\n",
4676 	    num_txd, num_rxd);
4677 
4678 	if (scctx->isc_tx_nsegments > num_txd / MAX_SINGLE_PACKET_FRACTION)
4679 		scctx->isc_tx_nsegments = max(1, num_txd /
4680 		    MAX_SINGLE_PACKET_FRACTION);
4681 	if (scctx->isc_tx_tso_segments_max > num_txd /
4682 	    MAX_SINGLE_PACKET_FRACTION)
4683 		scctx->isc_tx_tso_segments_max = max(1,
4684 		    num_txd / MAX_SINGLE_PACKET_FRACTION);
4685 
4686 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4687 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
4688 		/*
4689 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
4690 		 * but some MACs do.
4691 		 */
4692 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
4693 		    IP_MAXPACKET));
4694 		/*
4695 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
4696 		 * into account.  In the worst case, each of these calls will
4697 		 * add another mbuf and, thus, the requirement for another DMA
4698 		 * segment.  So for best performance, it doesn't make sense to
4699 		 * advertize a maximum of TSO segments that typically will
4700 		 * require defragmentation in iflib_encap().
4701 		 */
4702 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
4703 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
4704 	}
4705 	if (scctx->isc_rss_table_size == 0)
4706 		scctx->isc_rss_table_size = 64;
4707 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4708 
4709 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4710 	/* XXX format name */
4711 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx,
4712 	    NULL, NULL, "admin");
4713 
4714 	/* Set up cpu set.  If it fails, use the set of all CPUs. */
4715 	if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) != 0) {
4716 		device_printf(dev, "Unable to fetch CPU list\n");
4717 		CPU_COPY(&all_cpus, &ctx->ifc_cpus);
4718 	}
4719 	MPASS(CPU_COUNT(&ctx->ifc_cpus) > 0);
4720 
4721 	/*
4722 	** Now set up MSI or MSI-X, should return us the number of supported
4723 	** vectors (will be 1 for a legacy interrupt and MSI).
4724 	*/
4725 	if (sctx->isc_flags & IFLIB_SKIP_MSIX) {
4726 		msix = scctx->isc_vectors;
4727 	} else if (scctx->isc_msix_bar != 0)
4728 	       /*
4729 		* The simple fact that isc_msix_bar is not 0 does not mean we
4730 		* we have a good value there that is known to work.
4731 		*/
4732 		msix = iflib_msix_init(ctx);
4733 	else {
4734 		scctx->isc_vectors = 1;
4735 		scctx->isc_ntxqsets = 1;
4736 		scctx->isc_nrxqsets = 1;
4737 		scctx->isc_intr = IFLIB_INTR_LEGACY;
4738 		msix = 0;
4739 	}
4740 	/* Get memory for the station queues */
4741 	if ((err = iflib_queues_alloc(ctx))) {
4742 		device_printf(dev, "Unable to allocate queue memory\n");
4743 		goto fail_intr_free;
4744 	}
4745 
4746 	if ((err = iflib_qset_structures_setup(ctx)))
4747 		goto fail_queues;
4748 
4749 	/*
4750 	 * Now that we know how many queues there are, get the core offset.
4751 	 */
4752 	ctx->ifc_sysctl_core_offset = get_ctx_core_offset(ctx);
4753 
4754 	if (msix > 1) {
4755 		/*
4756 		 * When using MSI-X, ensure that ifdi_{r,t}x_queue_intr_enable
4757 		 * aren't the default NULL implementation.
4758 		 */
4759 		kobj_desc = &ifdi_rx_queue_intr_enable_desc;
4760 		kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL,
4761 		    kobj_desc);
4762 		if (kobj_method == &kobj_desc->deflt) {
4763 			device_printf(dev,
4764 			    "MSI-X requires ifdi_rx_queue_intr_enable method");
4765 			err = EOPNOTSUPP;
4766 			goto fail_queues;
4767 		}
4768 		kobj_desc = &ifdi_tx_queue_intr_enable_desc;
4769 		kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL,
4770 		    kobj_desc);
4771 		if (kobj_method == &kobj_desc->deflt) {
4772 			device_printf(dev,
4773 			    "MSI-X requires ifdi_tx_queue_intr_enable method");
4774 			err = EOPNOTSUPP;
4775 			goto fail_queues;
4776 		}
4777 
4778 		/*
4779 		 * Assign the MSI-X vectors.
4780 		 * Note that the default NULL ifdi_msix_intr_assign method will
4781 		 * fail here, too.
4782 		 */
4783 		err = IFDI_MSIX_INTR_ASSIGN(ctx, msix);
4784 		if (err != 0) {
4785 			device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n",
4786 			    err);
4787 			goto fail_queues;
4788 		}
4789 	} else if (scctx->isc_intr != IFLIB_INTR_MSIX) {
4790 		rid = 0;
4791 		if (scctx->isc_intr == IFLIB_INTR_MSI) {
4792 			MPASS(msix == 1);
4793 			rid = 1;
4794 		}
4795 		if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) {
4796 			device_printf(dev, "iflib_legacy_setup failed %d\n", err);
4797 			goto fail_queues;
4798 		}
4799 	} else {
4800 		device_printf(dev,
4801 		    "Cannot use iflib with only 1 MSI-X interrupt!\n");
4802 		err = ENODEV;
4803 		goto fail_intr_free;
4804 	}
4805 
4806 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
4807 
4808 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4809 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4810 		goto fail_detach;
4811 	}
4812 
4813 	/*
4814 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4815 	 * This must appear after the call to ether_ifattach() because
4816 	 * ether_ifattach() sets if_hdrlen to the default value.
4817 	 */
4818 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4819 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
4820 
4821 	if ((err = iflib_netmap_attach(ctx))) {
4822 		device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err);
4823 		goto fail_detach;
4824 	}
4825 	*ctxp = ctx;
4826 
4827 	DEBUGNET_SET(ctx->ifc_ifp, iflib);
4828 
4829 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4830 	iflib_add_device_sysctl_post(ctx);
4831 	iflib_add_pfil(ctx);
4832 	ctx->ifc_flags |= IFC_INIT_DONE;
4833 	CTX_UNLOCK(ctx);
4834 
4835 	return (0);
4836 
4837 fail_detach:
4838 	ether_ifdetach(ctx->ifc_ifp);
4839 fail_intr_free:
4840 	iflib_free_intr_mem(ctx);
4841 fail_queues:
4842 	iflib_tx_structures_free(ctx);
4843 	iflib_rx_structures_free(ctx);
4844 	taskqgroup_detach(qgroup_if_config_tqg, &ctx->ifc_admin_task);
4845 	IFDI_DETACH(ctx);
4846 fail_unlock:
4847 	CTX_UNLOCK(ctx);
4848 	iflib_deregister(ctx);
4849 fail_ctx_free:
4850 	device_set_softc(ctx->ifc_dev, NULL);
4851         if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4852                 free(ctx->ifc_softc, M_IFLIB);
4853         free(ctx, M_IFLIB);
4854 	return (err);
4855 }
4856 
4857 int
4858 iflib_pseudo_register(device_t dev, if_shared_ctx_t sctx, if_ctx_t *ctxp,
4859 					  struct iflib_cloneattach_ctx *clctx)
4860 {
4861 	int num_txd, num_rxd;
4862 	int err;
4863 	if_ctx_t ctx;
4864 	if_t ifp;
4865 	if_softc_ctx_t scctx;
4866 	int i;
4867 	void *sc;
4868 
4869 	ctx = malloc(sizeof(*ctx), M_IFLIB, M_WAITOK|M_ZERO);
4870 	sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4871 	ctx->ifc_flags |= IFC_SC_ALLOCATED;
4872 	if (sctx->isc_flags & (IFLIB_PSEUDO|IFLIB_VIRTUAL))
4873 		ctx->ifc_flags |= IFC_PSEUDO;
4874 
4875 	ctx->ifc_sctx = sctx;
4876 	ctx->ifc_softc = sc;
4877 	ctx->ifc_dev = dev;
4878 
4879 	if ((err = iflib_register(ctx)) != 0) {
4880 		device_printf(dev, "%s: iflib_register failed %d\n", __func__, err);
4881 		goto fail_ctx_free;
4882 	}
4883 	iflib_add_device_sysctl_pre(ctx);
4884 
4885 	scctx = &ctx->ifc_softc_ctx;
4886 	ifp = ctx->ifc_ifp;
4887 
4888 	iflib_reset_qvalues(ctx);
4889 	CTX_LOCK(ctx);
4890 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4891 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4892 		goto fail_unlock;
4893 	}
4894 	if (sctx->isc_flags & IFLIB_GEN_MAC)
4895 		ether_gen_addr(ifp, &ctx->ifc_mac);
4896 	if ((err = IFDI_CLONEATTACH(ctx, clctx->cc_ifc, clctx->cc_name,
4897 								clctx->cc_params)) != 0) {
4898 		device_printf(dev, "IFDI_CLONEATTACH failed %d\n", err);
4899 		goto fail_unlock;
4900 	}
4901 #ifdef INVARIANTS
4902 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
4903 		MPASS(scctx->isc_tx_csum_flags);
4904 #endif
4905 
4906 	if_setcapabilities(ifp, scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_LINKSTATE);
4907 	if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_LINKSTATE);
4908 
4909 	ifp->if_flags |= IFF_NOGROUP;
4910 	if (sctx->isc_flags & IFLIB_PSEUDO) {
4911 		ifmedia_add(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO, 0, NULL);
4912 		ifmedia_set(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO);
4913 		if (sctx->isc_flags & IFLIB_PSEUDO_ETHER) {
4914 			ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
4915 		} else {
4916 			if_attach(ctx->ifc_ifp);
4917 			bpfattach(ctx->ifc_ifp, DLT_NULL, sizeof(u_int32_t));
4918 		}
4919 
4920 		if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4921 			device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4922 			goto fail_detach;
4923 		}
4924 		*ctxp = ctx;
4925 
4926 		/*
4927 		 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4928 		 * This must appear after the call to ether_ifattach() because
4929 		 * ether_ifattach() sets if_hdrlen to the default value.
4930 		 */
4931 		if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4932 			if_setifheaderlen(ifp,
4933 			    sizeof(struct ether_vlan_header));
4934 
4935 		if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4936 		iflib_add_device_sysctl_post(ctx);
4937 		ctx->ifc_flags |= IFC_INIT_DONE;
4938 		CTX_UNLOCK(ctx);
4939 		return (0);
4940 	}
4941 	ifmedia_add(ctx->ifc_mediap, IFM_ETHER | IFM_1000_T | IFM_FDX, 0, NULL);
4942 	ifmedia_add(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO, 0, NULL);
4943 	ifmedia_set(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO);
4944 
4945 	_iflib_pre_assert(scctx);
4946 	ctx->ifc_txrx = *scctx->isc_txrx;
4947 
4948 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4949 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4950 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4951 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4952 
4953 	num_txd = iflib_num_tx_descs(ctx);
4954 	num_rxd = iflib_num_rx_descs(ctx);
4955 
4956 	/* XXX change for per-queue sizes */
4957 	device_printf(dev, "Using %d TX descriptors and %d RX descriptors\n",
4958 	    num_txd, num_rxd);
4959 
4960 	if (scctx->isc_tx_nsegments > num_txd / MAX_SINGLE_PACKET_FRACTION)
4961 		scctx->isc_tx_nsegments = max(1, num_txd /
4962 		    MAX_SINGLE_PACKET_FRACTION);
4963 	if (scctx->isc_tx_tso_segments_max > num_txd /
4964 	    MAX_SINGLE_PACKET_FRACTION)
4965 		scctx->isc_tx_tso_segments_max = max(1,
4966 		    num_txd / MAX_SINGLE_PACKET_FRACTION);
4967 
4968 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4969 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
4970 		/*
4971 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
4972 		 * but some MACs do.
4973 		 */
4974 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
4975 		    IP_MAXPACKET));
4976 		/*
4977 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
4978 		 * into account.  In the worst case, each of these calls will
4979 		 * add another mbuf and, thus, the requirement for another DMA
4980 		 * segment.  So for best performance, it doesn't make sense to
4981 		 * advertize a maximum of TSO segments that typically will
4982 		 * require defragmentation in iflib_encap().
4983 		 */
4984 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
4985 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
4986 	}
4987 	if (scctx->isc_rss_table_size == 0)
4988 		scctx->isc_rss_table_size = 64;
4989 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4990 
4991 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4992 	/* XXX format name */
4993 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx,
4994 	    NULL, NULL, "admin");
4995 
4996 	/* XXX --- can support > 1 -- but keep it simple for now */
4997 	scctx->isc_intr = IFLIB_INTR_LEGACY;
4998 
4999 	/* Get memory for the station queues */
5000 	if ((err = iflib_queues_alloc(ctx))) {
5001 		device_printf(dev, "Unable to allocate queue memory\n");
5002 		goto fail_iflib_detach;
5003 	}
5004 
5005 	if ((err = iflib_qset_structures_setup(ctx))) {
5006 		device_printf(dev, "qset structure setup failed %d\n", err);
5007 		goto fail_queues;
5008 	}
5009 
5010 	/*
5011 	 * XXX What if anything do we want to do about interrupts?
5012 	 */
5013 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
5014 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
5015 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
5016 		goto fail_detach;
5017 	}
5018 
5019 	/*
5020 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
5021 	 * This must appear after the call to ether_ifattach() because
5022 	 * ether_ifattach() sets if_hdrlen to the default value.
5023 	 */
5024 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
5025 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
5026 
5027 	/* XXX handle more than one queue */
5028 	for (i = 0; i < scctx->isc_nrxqsets; i++)
5029 		IFDI_RX_CLSET(ctx, 0, i, ctx->ifc_rxqs[i].ifr_fl[0].ifl_sds.ifsd_cl);
5030 
5031 	*ctxp = ctx;
5032 
5033 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
5034 	iflib_add_device_sysctl_post(ctx);
5035 	ctx->ifc_flags |= IFC_INIT_DONE;
5036 	CTX_UNLOCK(ctx);
5037 
5038 	return (0);
5039 fail_detach:
5040 	ether_ifdetach(ctx->ifc_ifp);
5041 fail_queues:
5042 	iflib_tx_structures_free(ctx);
5043 	iflib_rx_structures_free(ctx);
5044 fail_iflib_detach:
5045 	IFDI_DETACH(ctx);
5046 fail_unlock:
5047 	CTX_UNLOCK(ctx);
5048 	iflib_deregister(ctx);
5049 fail_ctx_free:
5050 	free(ctx->ifc_softc, M_IFLIB);
5051 	free(ctx, M_IFLIB);
5052 	return (err);
5053 }
5054 
5055 int
5056 iflib_pseudo_deregister(if_ctx_t ctx)
5057 {
5058 	if_t ifp = ctx->ifc_ifp;
5059 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5060 	iflib_txq_t txq;
5061 	iflib_rxq_t rxq;
5062 	int i, j;
5063 	struct taskqgroup *tqg;
5064 	iflib_fl_t fl;
5065 
5066 	/* Unregister VLAN event handlers early */
5067 	iflib_unregister_vlan_handlers(ctx);
5068 
5069 	if ((sctx->isc_flags & IFLIB_PSEUDO)  &&
5070 		(sctx->isc_flags & IFLIB_PSEUDO_ETHER) == 0) {
5071 		bpfdetach(ifp);
5072 		if_detach(ifp);
5073 	} else {
5074 		ether_ifdetach(ifp);
5075 	}
5076 	/* XXX drain any dependent tasks */
5077 	tqg = qgroup_if_io_tqg;
5078 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
5079 		callout_drain(&txq->ift_timer);
5080 #ifdef DEV_NETMAP
5081 		callout_drain(&txq->ift_netmap_timer);
5082 #endif /* DEV_NETMAP */
5083 		if (txq->ift_task.gt_uniq != NULL)
5084 			taskqgroup_detach(tqg, &txq->ift_task);
5085 	}
5086 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
5087 		callout_drain(&rxq->ifr_watchdog);
5088 		if (rxq->ifr_task.gt_uniq != NULL)
5089 			taskqgroup_detach(tqg, &rxq->ifr_task);
5090 
5091 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5092 			free(fl->ifl_rx_bitmap, M_IFLIB);
5093 	}
5094 	tqg = qgroup_if_config_tqg;
5095 	if (ctx->ifc_admin_task.gt_uniq != NULL)
5096 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
5097 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
5098 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
5099 
5100 	iflib_tx_structures_free(ctx);
5101 	iflib_rx_structures_free(ctx);
5102 
5103 	iflib_deregister(ctx);
5104 
5105 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5106 		free(ctx->ifc_softc, M_IFLIB);
5107 	free(ctx, M_IFLIB);
5108 	return (0);
5109 }
5110 
5111 int
5112 iflib_device_attach(device_t dev)
5113 {
5114 	if_ctx_t ctx;
5115 	if_shared_ctx_t sctx;
5116 
5117 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
5118 		return (ENOTSUP);
5119 
5120 	pci_enable_busmaster(dev);
5121 
5122 	return (iflib_device_register(dev, NULL, sctx, &ctx));
5123 }
5124 
5125 int
5126 iflib_device_deregister(if_ctx_t ctx)
5127 {
5128 	if_t ifp = ctx->ifc_ifp;
5129 	iflib_txq_t txq;
5130 	iflib_rxq_t rxq;
5131 	device_t dev = ctx->ifc_dev;
5132 	int i, j;
5133 	struct taskqgroup *tqg;
5134 	iflib_fl_t fl;
5135 
5136 	/* Make sure VLANS are not using driver */
5137 	if (if_vlantrunkinuse(ifp)) {
5138 		device_printf(dev, "Vlan in use, detach first\n");
5139 		return (EBUSY);
5140 	}
5141 #ifdef PCI_IOV
5142 	if (!CTX_IS_VF(ctx) && pci_iov_detach(dev) != 0) {
5143 		device_printf(dev, "SR-IOV in use; detach first.\n");
5144 		return (EBUSY);
5145 	}
5146 #endif
5147 
5148 	STATE_LOCK(ctx);
5149 	ctx->ifc_flags |= IFC_IN_DETACH;
5150 	STATE_UNLOCK(ctx);
5151 
5152 	/* Unregister VLAN handlers before calling iflib_stop() */
5153 	iflib_unregister_vlan_handlers(ctx);
5154 
5155 	iflib_netmap_detach(ifp);
5156 	ether_ifdetach(ifp);
5157 
5158 	CTX_LOCK(ctx);
5159 	iflib_stop(ctx);
5160 	CTX_UNLOCK(ctx);
5161 
5162 	iflib_rem_pfil(ctx);
5163 	if (ctx->ifc_led_dev != NULL)
5164 		led_destroy(ctx->ifc_led_dev);
5165 	/* XXX drain any dependent tasks */
5166 	tqg = qgroup_if_io_tqg;
5167 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
5168 		callout_drain(&txq->ift_timer);
5169 #ifdef DEV_NETMAP
5170 		callout_drain(&txq->ift_netmap_timer);
5171 #endif /* DEV_NETMAP */
5172 		if (txq->ift_task.gt_uniq != NULL)
5173 			taskqgroup_detach(tqg, &txq->ift_task);
5174 	}
5175 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
5176 		if (rxq->ifr_task.gt_uniq != NULL)
5177 			taskqgroup_detach(tqg, &rxq->ifr_task);
5178 
5179 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5180 			free(fl->ifl_rx_bitmap, M_IFLIB);
5181 	}
5182 	tqg = qgroup_if_config_tqg;
5183 	if (ctx->ifc_admin_task.gt_uniq != NULL)
5184 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
5185 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
5186 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
5187 	CTX_LOCK(ctx);
5188 	IFDI_DETACH(ctx);
5189 	CTX_UNLOCK(ctx);
5190 
5191 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
5192 	iflib_free_intr_mem(ctx);
5193 
5194 	bus_generic_detach(dev);
5195 
5196 	iflib_tx_structures_free(ctx);
5197 	iflib_rx_structures_free(ctx);
5198 
5199 	iflib_deregister(ctx);
5200 
5201 	device_set_softc(ctx->ifc_dev, NULL);
5202 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5203 		free(ctx->ifc_softc, M_IFLIB);
5204 	unref_ctx_core_offset(ctx);
5205 	free(ctx, M_IFLIB);
5206 	return (0);
5207 }
5208 
5209 static void
5210 iflib_free_intr_mem(if_ctx_t ctx)
5211 {
5212 
5213 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) {
5214 		iflib_irq_free(ctx, &ctx->ifc_legacy_irq);
5215 	}
5216 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) {
5217 		pci_release_msi(ctx->ifc_dev);
5218 	}
5219 	if (ctx->ifc_msix_mem != NULL) {
5220 		bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY,
5221 		    rman_get_rid(ctx->ifc_msix_mem), ctx->ifc_msix_mem);
5222 		ctx->ifc_msix_mem = NULL;
5223 	}
5224 }
5225 
5226 int
5227 iflib_device_detach(device_t dev)
5228 {
5229 	if_ctx_t ctx = device_get_softc(dev);
5230 
5231 	return (iflib_device_deregister(ctx));
5232 }
5233 
5234 int
5235 iflib_device_suspend(device_t dev)
5236 {
5237 	if_ctx_t ctx = device_get_softc(dev);
5238 
5239 	CTX_LOCK(ctx);
5240 	IFDI_SUSPEND(ctx);
5241 	CTX_UNLOCK(ctx);
5242 
5243 	return bus_generic_suspend(dev);
5244 }
5245 int
5246 iflib_device_shutdown(device_t dev)
5247 {
5248 	if_ctx_t ctx = device_get_softc(dev);
5249 
5250 	CTX_LOCK(ctx);
5251 	IFDI_SHUTDOWN(ctx);
5252 	CTX_UNLOCK(ctx);
5253 
5254 	return bus_generic_suspend(dev);
5255 }
5256 
5257 int
5258 iflib_device_resume(device_t dev)
5259 {
5260 	if_ctx_t ctx = device_get_softc(dev);
5261 	iflib_txq_t txq = ctx->ifc_txqs;
5262 
5263 	CTX_LOCK(ctx);
5264 	IFDI_RESUME(ctx);
5265 	iflib_if_init_locked(ctx);
5266 	CTX_UNLOCK(ctx);
5267 	for (int i = 0; i < NTXQSETS(ctx); i++, txq++)
5268 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
5269 
5270 	return (bus_generic_resume(dev));
5271 }
5272 
5273 int
5274 iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params)
5275 {
5276 	int error;
5277 	if_ctx_t ctx = device_get_softc(dev);
5278 
5279 	CTX_LOCK(ctx);
5280 	error = IFDI_IOV_INIT(ctx, num_vfs, params);
5281 	CTX_UNLOCK(ctx);
5282 
5283 	return (error);
5284 }
5285 
5286 void
5287 iflib_device_iov_uninit(device_t dev)
5288 {
5289 	if_ctx_t ctx = device_get_softc(dev);
5290 
5291 	CTX_LOCK(ctx);
5292 	IFDI_IOV_UNINIT(ctx);
5293 	CTX_UNLOCK(ctx);
5294 }
5295 
5296 int
5297 iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params)
5298 {
5299 	int error;
5300 	if_ctx_t ctx = device_get_softc(dev);
5301 
5302 	CTX_LOCK(ctx);
5303 	error = IFDI_IOV_VF_ADD(ctx, vfnum, params);
5304 	CTX_UNLOCK(ctx);
5305 
5306 	return (error);
5307 }
5308 
5309 /*********************************************************************
5310  *
5311  *  MODULE FUNCTION DEFINITIONS
5312  *
5313  **********************************************************************/
5314 
5315 /*
5316  * - Start a fast taskqueue thread for each core
5317  * - Start a taskqueue for control operations
5318  */
5319 static int
5320 iflib_module_init(void)
5321 {
5322 	return (0);
5323 }
5324 
5325 static int
5326 iflib_module_event_handler(module_t mod, int what, void *arg)
5327 {
5328 	int err;
5329 
5330 	switch (what) {
5331 	case MOD_LOAD:
5332 		if ((err = iflib_module_init()) != 0)
5333 			return (err);
5334 		break;
5335 	case MOD_UNLOAD:
5336 		return (EBUSY);
5337 	default:
5338 		return (EOPNOTSUPP);
5339 	}
5340 
5341 	return (0);
5342 }
5343 
5344 /*********************************************************************
5345  *
5346  *  PUBLIC FUNCTION DEFINITIONS
5347  *     ordered as in iflib.h
5348  *
5349  **********************************************************************/
5350 
5351 static void
5352 _iflib_assert(if_shared_ctx_t sctx)
5353 {
5354 	int i;
5355 
5356 	MPASS(sctx->isc_tx_maxsize);
5357 	MPASS(sctx->isc_tx_maxsegsize);
5358 
5359 	MPASS(sctx->isc_rx_maxsize);
5360 	MPASS(sctx->isc_rx_nsegments);
5361 	MPASS(sctx->isc_rx_maxsegsize);
5362 
5363 	MPASS(sctx->isc_nrxqs >= 1 && sctx->isc_nrxqs <= 8);
5364 	for (i = 0; i < sctx->isc_nrxqs; i++) {
5365 		MPASS(sctx->isc_nrxd_min[i]);
5366 		MPASS(powerof2(sctx->isc_nrxd_min[i]));
5367 		MPASS(sctx->isc_nrxd_max[i]);
5368 		MPASS(powerof2(sctx->isc_nrxd_max[i]));
5369 		MPASS(sctx->isc_nrxd_default[i]);
5370 		MPASS(powerof2(sctx->isc_nrxd_default[i]));
5371 	}
5372 
5373 	MPASS(sctx->isc_ntxqs >= 1 && sctx->isc_ntxqs <= 8);
5374 	for (i = 0; i < sctx->isc_ntxqs; i++) {
5375 		MPASS(sctx->isc_ntxd_min[i]);
5376 		MPASS(powerof2(sctx->isc_ntxd_min[i]));
5377 		MPASS(sctx->isc_ntxd_max[i]);
5378 		MPASS(powerof2(sctx->isc_ntxd_max[i]));
5379 		MPASS(sctx->isc_ntxd_default[i]);
5380 		MPASS(powerof2(sctx->isc_ntxd_default[i]));
5381 	}
5382 }
5383 
5384 static void
5385 _iflib_pre_assert(if_softc_ctx_t scctx)
5386 {
5387 
5388 	MPASS(scctx->isc_txrx->ift_txd_encap);
5389 	MPASS(scctx->isc_txrx->ift_txd_flush);
5390 	MPASS(scctx->isc_txrx->ift_txd_credits_update);
5391 	MPASS(scctx->isc_txrx->ift_rxd_available);
5392 	MPASS(scctx->isc_txrx->ift_rxd_pkt_get);
5393 	MPASS(scctx->isc_txrx->ift_rxd_refill);
5394 	MPASS(scctx->isc_txrx->ift_rxd_flush);
5395 }
5396 
5397 static int
5398 iflib_register(if_ctx_t ctx)
5399 {
5400 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5401 	driver_t *driver = sctx->isc_driver;
5402 	device_t dev = ctx->ifc_dev;
5403 	if_t ifp;
5404 	u_char type;
5405 	int iflags;
5406 
5407 	if ((sctx->isc_flags & IFLIB_PSEUDO) == 0)
5408 		_iflib_assert(sctx);
5409 
5410 	CTX_LOCK_INIT(ctx);
5411 	STATE_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev));
5412 	if (sctx->isc_flags & IFLIB_PSEUDO) {
5413 		if (sctx->isc_flags & IFLIB_PSEUDO_ETHER)
5414 			type = IFT_ETHER;
5415 		else
5416 			type = IFT_PPP;
5417 	} else
5418 		type = IFT_ETHER;
5419 	ifp = ctx->ifc_ifp = if_alloc(type);
5420 	if (ifp == NULL) {
5421 		device_printf(dev, "can not allocate ifnet structure\n");
5422 		return (ENOMEM);
5423 	}
5424 
5425 	/*
5426 	 * Initialize our context's device specific methods
5427 	 */
5428 	kobj_init((kobj_t) ctx, (kobj_class_t) driver);
5429 	kobj_class_compile((kobj_class_t) driver);
5430 
5431 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
5432 	if_setsoftc(ifp, ctx);
5433 	if_setdev(ifp, dev);
5434 	if_setinitfn(ifp, iflib_if_init);
5435 	if_setioctlfn(ifp, iflib_if_ioctl);
5436 #ifdef ALTQ
5437 	if_setstartfn(ifp, iflib_altq_if_start);
5438 	if_settransmitfn(ifp, iflib_altq_if_transmit);
5439 	if_setsendqready(ifp);
5440 #else
5441 	if_settransmitfn(ifp, iflib_if_transmit);
5442 #endif
5443 	if_setqflushfn(ifp, iflib_if_qflush);
5444 	iflags = IFF_MULTICAST | IFF_KNOWSEPOCH;
5445 
5446 	if ((sctx->isc_flags & IFLIB_PSEUDO) &&
5447 		(sctx->isc_flags & IFLIB_PSEUDO_ETHER) == 0)
5448 		iflags |= IFF_POINTOPOINT;
5449 	else
5450 		iflags |= IFF_BROADCAST | IFF_SIMPLEX;
5451 	if_setflags(ifp, iflags);
5452 	ctx->ifc_vlan_attach_event =
5453 		EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx,
5454 							  EVENTHANDLER_PRI_FIRST);
5455 	ctx->ifc_vlan_detach_event =
5456 		EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx,
5457 							  EVENTHANDLER_PRI_FIRST);
5458 
5459 	if ((sctx->isc_flags & IFLIB_DRIVER_MEDIA) == 0) {
5460 		ctx->ifc_mediap = &ctx->ifc_media;
5461 		ifmedia_init(ctx->ifc_mediap, IFM_IMASK,
5462 		    iflib_media_change, iflib_media_status);
5463 	}
5464 	return (0);
5465 }
5466 
5467 static void
5468 iflib_unregister_vlan_handlers(if_ctx_t ctx)
5469 {
5470 	/* Unregister VLAN events */
5471 	if (ctx->ifc_vlan_attach_event != NULL) {
5472 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
5473 		ctx->ifc_vlan_attach_event = NULL;
5474 	}
5475 	if (ctx->ifc_vlan_detach_event != NULL) {
5476 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
5477 		ctx->ifc_vlan_detach_event = NULL;
5478 	}
5479 
5480 }
5481 
5482 static void
5483 iflib_deregister(if_ctx_t ctx)
5484 {
5485 	if_t ifp = ctx->ifc_ifp;
5486 
5487 	/* Remove all media */
5488 	ifmedia_removeall(&ctx->ifc_media);
5489 
5490 	/* Ensure that VLAN event handlers are unregistered */
5491 	iflib_unregister_vlan_handlers(ctx);
5492 
5493 	/* Release kobject reference */
5494 	kobj_delete((kobj_t) ctx, NULL);
5495 
5496 	/* Free the ifnet structure */
5497 	if_free(ifp);
5498 
5499 	STATE_LOCK_DESTROY(ctx);
5500 
5501 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
5502 	CTX_LOCK_DESTROY(ctx);
5503 }
5504 
5505 static int
5506 iflib_queues_alloc(if_ctx_t ctx)
5507 {
5508 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5509 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5510 	device_t dev = ctx->ifc_dev;
5511 	int nrxqsets = scctx->isc_nrxqsets;
5512 	int ntxqsets = scctx->isc_ntxqsets;
5513 	iflib_txq_t txq;
5514 	iflib_rxq_t rxq;
5515 	iflib_fl_t fl = NULL;
5516 	int i, j, cpu, err, txconf, rxconf;
5517 	iflib_dma_info_t ifdip;
5518 	uint32_t *rxqsizes = scctx->isc_rxqsizes;
5519 	uint32_t *txqsizes = scctx->isc_txqsizes;
5520 	uint8_t nrxqs = sctx->isc_nrxqs;
5521 	uint8_t ntxqs = sctx->isc_ntxqs;
5522 	int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1;
5523 	caddr_t *vaddrs;
5524 	uint64_t *paddrs;
5525 
5526 	KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1"));
5527 	KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1"));
5528 
5529 	/* Allocate the TX ring struct memory */
5530 	if (!(ctx->ifc_txqs =
5531 	    (iflib_txq_t) malloc(sizeof(struct iflib_txq) *
5532 	    ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5533 		device_printf(dev, "Unable to allocate TX ring memory\n");
5534 		err = ENOMEM;
5535 		goto fail;
5536 	}
5537 
5538 	/* Now allocate the RX */
5539 	if (!(ctx->ifc_rxqs =
5540 	    (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) *
5541 	    nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5542 		device_printf(dev, "Unable to allocate RX ring memory\n");
5543 		err = ENOMEM;
5544 		goto rx_fail;
5545 	}
5546 
5547 	txq = ctx->ifc_txqs;
5548 	rxq = ctx->ifc_rxqs;
5549 
5550 	/*
5551 	 * XXX handle allocation failure
5552 	 */
5553 	for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) {
5554 		/* Set up some basics */
5555 
5556 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs,
5557 		    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
5558 			device_printf(dev,
5559 			    "Unable to allocate TX DMA info memory\n");
5560 			err = ENOMEM;
5561 			goto err_tx_desc;
5562 		}
5563 		txq->ift_ifdi = ifdip;
5564 		for (j = 0; j < ntxqs; j++, ifdip++) {
5565 			if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, 0)) {
5566 				device_printf(dev,
5567 				    "Unable to allocate TX descriptors\n");
5568 				err = ENOMEM;
5569 				goto err_tx_desc;
5570 			}
5571 			txq->ift_txd_size[j] = scctx->isc_txd_size[j];
5572 			bzero((void *)ifdip->idi_vaddr, txqsizes[j]);
5573 		}
5574 		txq->ift_ctx = ctx;
5575 		txq->ift_id = i;
5576 		if (sctx->isc_flags & IFLIB_HAS_TXCQ) {
5577 			txq->ift_br_offset = 1;
5578 		} else {
5579 			txq->ift_br_offset = 0;
5580 		}
5581 
5582 		if (iflib_txsd_alloc(txq)) {
5583 			device_printf(dev, "Critical Failure setting up TX buffers\n");
5584 			err = ENOMEM;
5585 			goto err_tx_desc;
5586 		}
5587 
5588 		/* Initialize the TX lock */
5589 		snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:TX(%d):callout",
5590 		    device_get_nameunit(dev), txq->ift_id);
5591 		mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF);
5592 		callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0);
5593 		txq->ift_timer.c_cpu = cpu;
5594 #ifdef DEV_NETMAP
5595 		callout_init_mtx(&txq->ift_netmap_timer, &txq->ift_mtx, 0);
5596 		txq->ift_netmap_timer.c_cpu = cpu;
5597 #endif /* DEV_NETMAP */
5598 
5599 		err = ifmp_ring_alloc(&txq->ift_br, 2048, txq, iflib_txq_drain,
5600 				      iflib_txq_can_drain, M_IFLIB, M_WAITOK);
5601 		if (err) {
5602 			/* XXX free any allocated rings */
5603 			device_printf(dev, "Unable to allocate buf_ring\n");
5604 			goto err_tx_desc;
5605 		}
5606 	}
5607 
5608 	for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) {
5609 		/* Set up some basics */
5610 		callout_init(&rxq->ifr_watchdog, 1);
5611 
5612 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs,
5613 		   M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
5614 			device_printf(dev,
5615 			    "Unable to allocate RX DMA info memory\n");
5616 			err = ENOMEM;
5617 			goto err_tx_desc;
5618 		}
5619 
5620 		rxq->ifr_ifdi = ifdip;
5621 		/* XXX this needs to be changed if #rx queues != #tx queues */
5622 		rxq->ifr_ntxqirq = 1;
5623 		rxq->ifr_txqid[0] = i;
5624 		for (j = 0; j < nrxqs; j++, ifdip++) {
5625 			if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, 0)) {
5626 				device_printf(dev,
5627 				    "Unable to allocate RX descriptors\n");
5628 				err = ENOMEM;
5629 				goto err_tx_desc;
5630 			}
5631 			bzero((void *)ifdip->idi_vaddr, rxqsizes[j]);
5632 		}
5633 		rxq->ifr_ctx = ctx;
5634 		rxq->ifr_id = i;
5635 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
5636 			rxq->ifr_fl_offset = 1;
5637 		} else {
5638 			rxq->ifr_fl_offset = 0;
5639 		}
5640 		rxq->ifr_nfl = nfree_lists;
5641 		if (!(fl =
5642 			  (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) {
5643 			device_printf(dev, "Unable to allocate free list memory\n");
5644 			err = ENOMEM;
5645 			goto err_tx_desc;
5646 		}
5647 		rxq->ifr_fl = fl;
5648 		for (j = 0; j < nfree_lists; j++) {
5649 			fl[j].ifl_rxq = rxq;
5650 			fl[j].ifl_id = j;
5651 			fl[j].ifl_ifdi = &rxq->ifr_ifdi[j + rxq->ifr_fl_offset];
5652 			fl[j].ifl_rxd_size = scctx->isc_rxd_size[j];
5653 		}
5654 		/* Allocate receive buffers for the ring */
5655 		if (iflib_rxsd_alloc(rxq)) {
5656 			device_printf(dev,
5657 			    "Critical Failure setting up receive buffers\n");
5658 			err = ENOMEM;
5659 			goto err_rx_desc;
5660 		}
5661 
5662 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5663 			fl->ifl_rx_bitmap = bit_alloc(fl->ifl_size, M_IFLIB,
5664 			    M_WAITOK);
5665 	}
5666 
5667 	/* TXQs */
5668 	vaddrs = malloc(sizeof(caddr_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5669 	paddrs = malloc(sizeof(uint64_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5670 	for (i = 0; i < ntxqsets; i++) {
5671 		iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi;
5672 
5673 		for (j = 0; j < ntxqs; j++, di++) {
5674 			vaddrs[i*ntxqs + j] = di->idi_vaddr;
5675 			paddrs[i*ntxqs + j] = di->idi_paddr;
5676 		}
5677 	}
5678 	if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) {
5679 		device_printf(ctx->ifc_dev,
5680 		    "Unable to allocate device TX queue\n");
5681 		iflib_tx_structures_free(ctx);
5682 		free(vaddrs, M_IFLIB);
5683 		free(paddrs, M_IFLIB);
5684 		goto err_rx_desc;
5685 	}
5686 	free(vaddrs, M_IFLIB);
5687 	free(paddrs, M_IFLIB);
5688 
5689 	/* RXQs */
5690 	vaddrs = malloc(sizeof(caddr_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5691 	paddrs = malloc(sizeof(uint64_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5692 	for (i = 0; i < nrxqsets; i++) {
5693 		iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi;
5694 
5695 		for (j = 0; j < nrxqs; j++, di++) {
5696 			vaddrs[i*nrxqs + j] = di->idi_vaddr;
5697 			paddrs[i*nrxqs + j] = di->idi_paddr;
5698 		}
5699 	}
5700 	if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) {
5701 		device_printf(ctx->ifc_dev,
5702 		    "Unable to allocate device RX queue\n");
5703 		iflib_tx_structures_free(ctx);
5704 		free(vaddrs, M_IFLIB);
5705 		free(paddrs, M_IFLIB);
5706 		goto err_rx_desc;
5707 	}
5708 	free(vaddrs, M_IFLIB);
5709 	free(paddrs, M_IFLIB);
5710 
5711 	return (0);
5712 
5713 /* XXX handle allocation failure changes */
5714 err_rx_desc:
5715 err_tx_desc:
5716 rx_fail:
5717 	if (ctx->ifc_rxqs != NULL)
5718 		free(ctx->ifc_rxqs, M_IFLIB);
5719 	ctx->ifc_rxqs = NULL;
5720 	if (ctx->ifc_txqs != NULL)
5721 		free(ctx->ifc_txqs, M_IFLIB);
5722 	ctx->ifc_txqs = NULL;
5723 fail:
5724 	return (err);
5725 }
5726 
5727 static int
5728 iflib_tx_structures_setup(if_ctx_t ctx)
5729 {
5730 	iflib_txq_t txq = ctx->ifc_txqs;
5731 	int i;
5732 
5733 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
5734 		iflib_txq_setup(txq);
5735 
5736 	return (0);
5737 }
5738 
5739 static void
5740 iflib_tx_structures_free(if_ctx_t ctx)
5741 {
5742 	iflib_txq_t txq = ctx->ifc_txqs;
5743 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5744 	int i, j;
5745 
5746 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
5747 		for (j = 0; j < sctx->isc_ntxqs; j++)
5748 			iflib_dma_free(&txq->ift_ifdi[j]);
5749 		iflib_txq_destroy(txq);
5750 	}
5751 	free(ctx->ifc_txqs, M_IFLIB);
5752 	ctx->ifc_txqs = NULL;
5753 	IFDI_QUEUES_FREE(ctx);
5754 }
5755 
5756 /*********************************************************************
5757  *
5758  *  Initialize all receive rings.
5759  *
5760  **********************************************************************/
5761 static int
5762 iflib_rx_structures_setup(if_ctx_t ctx)
5763 {
5764 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5765 	int q;
5766 #if defined(INET6) || defined(INET)
5767 	int err, i;
5768 #endif
5769 
5770 	for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) {
5771 #if defined(INET6) || defined(INET)
5772 		if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_LRO) {
5773 			err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp,
5774 			    TCP_LRO_ENTRIES, min(1024,
5775 			    ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset]));
5776 			if (err != 0) {
5777 				device_printf(ctx->ifc_dev,
5778 				    "LRO Initialization failed!\n");
5779 				goto fail;
5780 			}
5781 		}
5782 #endif
5783 		IFDI_RXQ_SETUP(ctx, rxq->ifr_id);
5784 	}
5785 	return (0);
5786 #if defined(INET6) || defined(INET)
5787 fail:
5788 	/*
5789 	 * Free LRO resources allocated so far, we will only handle
5790 	 * the rings that completed, the failing case will have
5791 	 * cleaned up for itself.  'q' failed, so its the terminus.
5792 	 */
5793 	rxq = ctx->ifc_rxqs;
5794 	for (i = 0; i < q; ++i, rxq++) {
5795 		if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_LRO)
5796 			tcp_lro_free(&rxq->ifr_lc);
5797 	}
5798 	return (err);
5799 #endif
5800 }
5801 
5802 /*********************************************************************
5803  *
5804  *  Free all receive rings.
5805  *
5806  **********************************************************************/
5807 static void
5808 iflib_rx_structures_free(if_ctx_t ctx)
5809 {
5810 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5811 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5812 	int i, j;
5813 
5814 	for (i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) {
5815 		for (j = 0; j < sctx->isc_nrxqs; j++)
5816 			iflib_dma_free(&rxq->ifr_ifdi[j]);
5817 		iflib_rx_sds_free(rxq);
5818 #if defined(INET6) || defined(INET)
5819 		if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_LRO)
5820 			tcp_lro_free(&rxq->ifr_lc);
5821 #endif
5822 	}
5823 	free(ctx->ifc_rxqs, M_IFLIB);
5824 	ctx->ifc_rxqs = NULL;
5825 }
5826 
5827 static int
5828 iflib_qset_structures_setup(if_ctx_t ctx)
5829 {
5830 	int err;
5831 
5832 	/*
5833 	 * It is expected that the caller takes care of freeing queues if this
5834 	 * fails.
5835 	 */
5836 	if ((err = iflib_tx_structures_setup(ctx)) != 0) {
5837 		device_printf(ctx->ifc_dev, "iflib_tx_structures_setup failed: %d\n", err);
5838 		return (err);
5839 	}
5840 
5841 	if ((err = iflib_rx_structures_setup(ctx)) != 0)
5842 		device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err);
5843 
5844 	return (err);
5845 }
5846 
5847 int
5848 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
5849 		driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, const char *name)
5850 {
5851 
5852 	return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name));
5853 }
5854 
5855 #ifdef SMP
5856 static int
5857 find_nth(if_ctx_t ctx, int qid)
5858 {
5859 	cpuset_t cpus;
5860 	int i, cpuid, eqid, count;
5861 
5862 	CPU_COPY(&ctx->ifc_cpus, &cpus);
5863 	count = CPU_COUNT(&cpus);
5864 	eqid = qid % count;
5865 	/* clear up to the qid'th bit */
5866 	for (i = 0; i < eqid; i++) {
5867 		cpuid = CPU_FFS(&cpus);
5868 		MPASS(cpuid != 0);
5869 		CPU_CLR(cpuid-1, &cpus);
5870 	}
5871 	cpuid = CPU_FFS(&cpus);
5872 	MPASS(cpuid != 0);
5873 	return (cpuid-1);
5874 }
5875 
5876 #ifdef SCHED_ULE
5877 extern struct cpu_group *cpu_top;              /* CPU topology */
5878 
5879 static int
5880 find_child_with_core(int cpu, struct cpu_group *grp)
5881 {
5882 	int i;
5883 
5884 	if (grp->cg_children == 0)
5885 		return -1;
5886 
5887 	MPASS(grp->cg_child);
5888 	for (i = 0; i < grp->cg_children; i++) {
5889 		if (CPU_ISSET(cpu, &grp->cg_child[i].cg_mask))
5890 			return i;
5891 	}
5892 
5893 	return -1;
5894 }
5895 
5896 /*
5897  * Find the nth "close" core to the specified core
5898  * "close" is defined as the deepest level that shares
5899  * at least an L2 cache.  With threads, this will be
5900  * threads on the same core.  If the shared cache is L3
5901  * or higher, simply returns the same core.
5902  */
5903 static int
5904 find_close_core(int cpu, int core_offset)
5905 {
5906 	struct cpu_group *grp;
5907 	int i;
5908 	int fcpu;
5909 	cpuset_t cs;
5910 
5911 	grp = cpu_top;
5912 	if (grp == NULL)
5913 		return cpu;
5914 	i = 0;
5915 	while ((i = find_child_with_core(cpu, grp)) != -1) {
5916 		/* If the child only has one cpu, don't descend */
5917 		if (grp->cg_child[i].cg_count <= 1)
5918 			break;
5919 		grp = &grp->cg_child[i];
5920 	}
5921 
5922 	/* If they don't share at least an L2 cache, use the same CPU */
5923 	if (grp->cg_level > CG_SHARE_L2 || grp->cg_level == CG_SHARE_NONE)
5924 		return cpu;
5925 
5926 	/* Now pick one */
5927 	CPU_COPY(&grp->cg_mask, &cs);
5928 
5929 	/* Add the selected CPU offset to core offset. */
5930 	for (i = 0; (fcpu = CPU_FFS(&cs)) != 0; i++) {
5931 		if (fcpu - 1 == cpu)
5932 			break;
5933 		CPU_CLR(fcpu - 1, &cs);
5934 	}
5935 	MPASS(fcpu);
5936 
5937 	core_offset += i;
5938 
5939 	CPU_COPY(&grp->cg_mask, &cs);
5940 	for (i = core_offset % grp->cg_count; i > 0; i--) {
5941 		MPASS(CPU_FFS(&cs));
5942 		CPU_CLR(CPU_FFS(&cs) - 1, &cs);
5943 	}
5944 	MPASS(CPU_FFS(&cs));
5945 	return CPU_FFS(&cs) - 1;
5946 }
5947 #else
5948 static int
5949 find_close_core(int cpu, int core_offset __unused)
5950 {
5951 	return cpu;
5952 }
5953 #endif
5954 
5955 static int
5956 get_core_offset(if_ctx_t ctx, iflib_intr_type_t type, int qid)
5957 {
5958 	switch (type) {
5959 	case IFLIB_INTR_TX:
5960 		/* TX queues get cores which share at least an L2 cache with the corresponding RX queue */
5961 		/* XXX handle multiple RX threads per core and more than two core per L2 group */
5962 		return qid / CPU_COUNT(&ctx->ifc_cpus) + 1;
5963 	case IFLIB_INTR_RX:
5964 	case IFLIB_INTR_RXTX:
5965 		/* RX queues get the specified core */
5966 		return qid / CPU_COUNT(&ctx->ifc_cpus);
5967 	default:
5968 		return -1;
5969 	}
5970 }
5971 #else
5972 #define get_core_offset(ctx, type, qid)	CPU_FIRST()
5973 #define find_close_core(cpuid, tid)	CPU_FIRST()
5974 #define find_nth(ctx, gid)		CPU_FIRST()
5975 #endif
5976 
5977 /* Just to avoid copy/paste */
5978 static inline int
5979 iflib_irq_set_affinity(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type,
5980     int qid, struct grouptask *gtask, struct taskqgroup *tqg, void *uniq,
5981     const char *name)
5982 {
5983 	device_t dev;
5984 	int co, cpuid, err, tid;
5985 
5986 	dev = ctx->ifc_dev;
5987 	co = ctx->ifc_sysctl_core_offset;
5988 	if (ctx->ifc_sysctl_separate_txrx && type == IFLIB_INTR_TX)
5989 		co += ctx->ifc_softc_ctx.isc_nrxqsets;
5990 	cpuid = find_nth(ctx, qid + co);
5991 	tid = get_core_offset(ctx, type, qid);
5992 	if (tid < 0) {
5993 		device_printf(dev, "get_core_offset failed\n");
5994 		return (EOPNOTSUPP);
5995 	}
5996 	cpuid = find_close_core(cpuid, tid);
5997 	err = taskqgroup_attach_cpu(tqg, gtask, uniq, cpuid, dev, irq->ii_res,
5998 	    name);
5999 	if (err) {
6000 		device_printf(dev, "taskqgroup_attach_cpu failed %d\n", err);
6001 		return (err);
6002 	}
6003 #ifdef notyet
6004 	if (cpuid > ctx->ifc_cpuid_highest)
6005 		ctx->ifc_cpuid_highest = cpuid;
6006 #endif
6007 	return (0);
6008 }
6009 
6010 int
6011 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid,
6012 			iflib_intr_type_t type, driver_filter_t *filter,
6013 			void *filter_arg, int qid, const char *name)
6014 {
6015 	device_t dev;
6016 	struct grouptask *gtask;
6017 	struct taskqgroup *tqg;
6018 	iflib_filter_info_t info;
6019 	gtask_fn_t *fn;
6020 	int tqrid, err;
6021 	driver_filter_t *intr_fast;
6022 	void *q;
6023 
6024 	info = &ctx->ifc_filter_info;
6025 	tqrid = rid;
6026 
6027 	switch (type) {
6028 	/* XXX merge tx/rx for netmap? */
6029 	case IFLIB_INTR_TX:
6030 		q = &ctx->ifc_txqs[qid];
6031 		info = &ctx->ifc_txqs[qid].ift_filter_info;
6032 		gtask = &ctx->ifc_txqs[qid].ift_task;
6033 		tqg = qgroup_if_io_tqg;
6034 		fn = _task_fn_tx;
6035 		intr_fast = iflib_fast_intr;
6036 		GROUPTASK_INIT(gtask, 0, fn, q);
6037 		ctx->ifc_flags |= IFC_NETMAP_TX_IRQ;
6038 		break;
6039 	case IFLIB_INTR_RX:
6040 		q = &ctx->ifc_rxqs[qid];
6041 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
6042 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6043 		tqg = qgroup_if_io_tqg;
6044 		fn = _task_fn_rx;
6045 		intr_fast = iflib_fast_intr;
6046 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6047 		break;
6048 	case IFLIB_INTR_RXTX:
6049 		q = &ctx->ifc_rxqs[qid];
6050 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
6051 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6052 		tqg = qgroup_if_io_tqg;
6053 		fn = _task_fn_rx;
6054 		intr_fast = iflib_fast_intr_rxtx;
6055 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6056 		break;
6057 	case IFLIB_INTR_ADMIN:
6058 		q = ctx;
6059 		tqrid = -1;
6060 		info = &ctx->ifc_filter_info;
6061 		gtask = &ctx->ifc_admin_task;
6062 		tqg = qgroup_if_config_tqg;
6063 		fn = _task_fn_admin;
6064 		intr_fast = iflib_fast_intr_ctx;
6065 		break;
6066 	default:
6067 		device_printf(ctx->ifc_dev, "%s: unknown net intr type\n",
6068 		    __func__);
6069 		return (EINVAL);
6070 	}
6071 
6072 	info->ifi_filter = filter;
6073 	info->ifi_filter_arg = filter_arg;
6074 	info->ifi_task = gtask;
6075 	info->ifi_ctx = q;
6076 
6077 	dev = ctx->ifc_dev;
6078 	err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info,  name);
6079 	if (err != 0) {
6080 		device_printf(dev, "_iflib_irq_alloc failed %d\n", err);
6081 		return (err);
6082 	}
6083 	if (type == IFLIB_INTR_ADMIN)
6084 		return (0);
6085 
6086 	if (tqrid != -1) {
6087 		err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg,
6088 		    q, name);
6089 		if (err)
6090 			return (err);
6091 	} else {
6092 		taskqgroup_attach(tqg, gtask, q, dev, irq->ii_res, name);
6093 	}
6094 
6095 	return (0);
6096 }
6097 
6098 void
6099 iflib_softirq_alloc_generic(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type, void *arg, int qid, const char *name)
6100 {
6101 	struct grouptask *gtask;
6102 	struct taskqgroup *tqg;
6103 	gtask_fn_t *fn;
6104 	void *q;
6105 	int err;
6106 
6107 	switch (type) {
6108 	case IFLIB_INTR_TX:
6109 		q = &ctx->ifc_txqs[qid];
6110 		gtask = &ctx->ifc_txqs[qid].ift_task;
6111 		tqg = qgroup_if_io_tqg;
6112 		fn = _task_fn_tx;
6113 		GROUPTASK_INIT(gtask, 0, fn, q);
6114 		break;
6115 	case IFLIB_INTR_RX:
6116 		q = &ctx->ifc_rxqs[qid];
6117 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6118 		tqg = qgroup_if_io_tqg;
6119 		fn = _task_fn_rx;
6120 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6121 		break;
6122 	case IFLIB_INTR_IOV:
6123 		q = ctx;
6124 		gtask = &ctx->ifc_vflr_task;
6125 		tqg = qgroup_if_config_tqg;
6126 		fn = _task_fn_iov;
6127 		GROUPTASK_INIT(gtask, 0, fn, q);
6128 		break;
6129 	default:
6130 		panic("unknown net intr type");
6131 	}
6132 	if (irq != NULL) {
6133 		err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg,
6134 		    q, name);
6135 		if (err)
6136 			taskqgroup_attach(tqg, gtask, q, ctx->ifc_dev,
6137 			    irq->ii_res, name);
6138 	} else {
6139 		taskqgroup_attach(tqg, gtask, q, NULL, NULL, name);
6140 	}
6141 }
6142 
6143 void
6144 iflib_irq_free(if_ctx_t ctx, if_irq_t irq)
6145 {
6146 
6147 	if (irq->ii_tag)
6148 		bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag);
6149 
6150 	if (irq->ii_res)
6151 		bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ,
6152 		    rman_get_rid(irq->ii_res), irq->ii_res);
6153 }
6154 
6155 static int
6156 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, const char *name)
6157 {
6158 	iflib_txq_t txq = ctx->ifc_txqs;
6159 	iflib_rxq_t rxq = ctx->ifc_rxqs;
6160 	if_irq_t irq = &ctx->ifc_legacy_irq;
6161 	iflib_filter_info_t info;
6162 	device_t dev;
6163 	struct grouptask *gtask;
6164 	struct resource *res;
6165 	struct taskqgroup *tqg;
6166 	void *q;
6167 	int err, tqrid;
6168 	bool rx_only;
6169 
6170 	q = &ctx->ifc_rxqs[0];
6171 	info = &rxq[0].ifr_filter_info;
6172 	gtask = &rxq[0].ifr_task;
6173 	tqg = qgroup_if_io_tqg;
6174 	tqrid = *rid;
6175 	rx_only = (ctx->ifc_sctx->isc_flags & IFLIB_SINGLE_IRQ_RX_ONLY) != 0;
6176 
6177 	ctx->ifc_flags |= IFC_LEGACY;
6178 	info->ifi_filter = filter;
6179 	info->ifi_filter_arg = filter_arg;
6180 	info->ifi_task = gtask;
6181 	info->ifi_ctx = rx_only ? ctx : q;
6182 
6183 	dev = ctx->ifc_dev;
6184 	/* We allocate a single interrupt resource */
6185 	err = _iflib_irq_alloc(ctx, irq, tqrid, rx_only ? iflib_fast_intr_ctx :
6186 	    iflib_fast_intr_rxtx, NULL, info, name);
6187 	if (err != 0)
6188 		return (err);
6189 	NET_GROUPTASK_INIT(gtask, 0, _task_fn_rx, q);
6190 	res = irq->ii_res;
6191 	taskqgroup_attach(tqg, gtask, q, dev, res, name);
6192 
6193 	GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq);
6194 	taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, dev, res,
6195 	    "tx");
6196 	return (0);
6197 }
6198 
6199 void
6200 iflib_led_create(if_ctx_t ctx)
6201 {
6202 
6203 	ctx->ifc_led_dev = led_create(iflib_led_func, ctx,
6204 	    device_get_nameunit(ctx->ifc_dev));
6205 }
6206 
6207 void
6208 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid)
6209 {
6210 
6211 	GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
6212 }
6213 
6214 void
6215 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid)
6216 {
6217 
6218 	GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task);
6219 }
6220 
6221 void
6222 iflib_admin_intr_deferred(if_ctx_t ctx)
6223 {
6224 
6225 	MPASS(ctx->ifc_admin_task.gt_taskqueue != NULL);
6226 	GROUPTASK_ENQUEUE(&ctx->ifc_admin_task);
6227 }
6228 
6229 void
6230 iflib_iov_intr_deferred(if_ctx_t ctx)
6231 {
6232 
6233 	GROUPTASK_ENQUEUE(&ctx->ifc_vflr_task);
6234 }
6235 
6236 void
6237 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, const char *name)
6238 {
6239 
6240 	taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, NULL, NULL,
6241 	    name);
6242 }
6243 
6244 void
6245 iflib_config_gtask_init(void *ctx, struct grouptask *gtask, gtask_fn_t *fn,
6246 	const char *name)
6247 {
6248 
6249 	GROUPTASK_INIT(gtask, 0, fn, ctx);
6250 	taskqgroup_attach(qgroup_if_config_tqg, gtask, gtask, NULL, NULL,
6251 	    name);
6252 }
6253 
6254 void
6255 iflib_config_gtask_deinit(struct grouptask *gtask)
6256 {
6257 
6258 	taskqgroup_detach(qgroup_if_config_tqg, gtask);
6259 }
6260 
6261 void
6262 iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate)
6263 {
6264 	if_t ifp = ctx->ifc_ifp;
6265 	iflib_txq_t txq = ctx->ifc_txqs;
6266 
6267 	if_setbaudrate(ifp, baudrate);
6268 	if (baudrate >= IF_Gbps(10)) {
6269 		STATE_LOCK(ctx);
6270 		ctx->ifc_flags |= IFC_PREFETCH;
6271 		STATE_UNLOCK(ctx);
6272 	}
6273 	/* If link down, disable watchdog */
6274 	if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) {
6275 		for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++)
6276 			txq->ift_qstatus = IFLIB_QUEUE_IDLE;
6277 	}
6278 	ctx->ifc_link_state = link_state;
6279 	if_link_state_change(ifp, link_state);
6280 }
6281 
6282 static int
6283 iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq)
6284 {
6285 	int credits;
6286 #ifdef INVARIANTS
6287 	int credits_pre = txq->ift_cidx_processed;
6288 #endif
6289 
6290 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
6291 	    BUS_DMASYNC_POSTREAD);
6292 	if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, true)) == 0)
6293 		return (0);
6294 
6295 	txq->ift_processed += credits;
6296 	txq->ift_cidx_processed += credits;
6297 
6298 	MPASS(credits_pre + credits == txq->ift_cidx_processed);
6299 	if (txq->ift_cidx_processed >= txq->ift_size)
6300 		txq->ift_cidx_processed -= txq->ift_size;
6301 	return (credits);
6302 }
6303 
6304 static int
6305 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget)
6306 {
6307 	iflib_fl_t fl;
6308 	u_int i;
6309 
6310 	for (i = 0, fl = &rxq->ifr_fl[0]; i < rxq->ifr_nfl; i++, fl++)
6311 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
6312 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
6313 	return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx,
6314 	    budget));
6315 }
6316 
6317 void
6318 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name,
6319 	const char *description, if_int_delay_info_t info,
6320 	int offset, int value)
6321 {
6322 	info->iidi_ctx = ctx;
6323 	info->iidi_offset = offset;
6324 	info->iidi_value = value;
6325 	SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev),
6326 	    SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)),
6327 	    OID_AUTO, name, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE,
6328 	    info, 0, iflib_sysctl_int_delay, "I", description);
6329 }
6330 
6331 struct sx *
6332 iflib_ctx_lock_get(if_ctx_t ctx)
6333 {
6334 
6335 	return (&ctx->ifc_ctx_sx);
6336 }
6337 
6338 static int
6339 iflib_msix_init(if_ctx_t ctx)
6340 {
6341 	device_t dev = ctx->ifc_dev;
6342 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6343 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6344 	int admincnt, bar, err, iflib_num_rx_queues, iflib_num_tx_queues;
6345 	int msgs, queuemsgs, queues, rx_queues, tx_queues, vectors;
6346 
6347 	iflib_num_tx_queues = ctx->ifc_sysctl_ntxqs;
6348 	iflib_num_rx_queues = ctx->ifc_sysctl_nrxqs;
6349 
6350 	if (bootverbose)
6351 		device_printf(dev, "msix_init qsets capped at %d\n",
6352 		    imax(scctx->isc_ntxqsets, scctx->isc_nrxqsets));
6353 
6354 	/* Override by tuneable */
6355 	if (scctx->isc_disable_msix)
6356 		goto msi;
6357 
6358 	/* First try MSI-X */
6359 	if ((msgs = pci_msix_count(dev)) == 0) {
6360 		if (bootverbose)
6361 			device_printf(dev, "MSI-X not supported or disabled\n");
6362 		goto msi;
6363 	}
6364 
6365 	bar = ctx->ifc_softc_ctx.isc_msix_bar;
6366 	/*
6367 	 * bar == -1 => "trust me I know what I'm doing"
6368 	 * Some drivers are for hardware that is so shoddily
6369 	 * documented that no one knows which bars are which
6370 	 * so the developer has to map all bars. This hack
6371 	 * allows shoddy garbage to use MSI-X in this framework.
6372 	 */
6373 	if (bar != -1) {
6374 		ctx->ifc_msix_mem = bus_alloc_resource_any(dev,
6375 	            SYS_RES_MEMORY, &bar, RF_ACTIVE);
6376 		if (ctx->ifc_msix_mem == NULL) {
6377 			device_printf(dev, "Unable to map MSI-X table\n");
6378 			goto msi;
6379 		}
6380 	}
6381 
6382 	admincnt = sctx->isc_admin_intrcnt;
6383 #if IFLIB_DEBUG
6384 	/* use only 1 qset in debug mode */
6385 	queuemsgs = min(msgs - admincnt, 1);
6386 #else
6387 	queuemsgs = msgs - admincnt;
6388 #endif
6389 #ifdef RSS
6390 	queues = imin(queuemsgs, rss_getnumbuckets());
6391 #else
6392 	queues = queuemsgs;
6393 #endif
6394 	queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues);
6395 	if (bootverbose)
6396 		device_printf(dev,
6397 		    "intr CPUs: %d queue msgs: %d admincnt: %d\n",
6398 		    CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt);
6399 #ifdef  RSS
6400 	/* If we're doing RSS, clamp at the number of RSS buckets */
6401 	if (queues > rss_getnumbuckets())
6402 		queues = rss_getnumbuckets();
6403 #endif
6404 	if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt)
6405 		rx_queues = iflib_num_rx_queues;
6406 	else
6407 		rx_queues = queues;
6408 
6409 	if (rx_queues > scctx->isc_nrxqsets)
6410 		rx_queues = scctx->isc_nrxqsets;
6411 
6412 	/*
6413 	 * We want this to be all logical CPUs by default
6414 	 */
6415 	if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues)
6416 		tx_queues = iflib_num_tx_queues;
6417 	else
6418 		tx_queues = mp_ncpus;
6419 
6420 	if (tx_queues > scctx->isc_ntxqsets)
6421 		tx_queues = scctx->isc_ntxqsets;
6422 
6423 	if (ctx->ifc_sysctl_qs_eq_override == 0) {
6424 #ifdef INVARIANTS
6425 		if (tx_queues != rx_queues)
6426 			device_printf(dev,
6427 			    "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n",
6428 			    min(rx_queues, tx_queues), min(rx_queues, tx_queues));
6429 #endif
6430 		tx_queues = min(rx_queues, tx_queues);
6431 		rx_queues = min(rx_queues, tx_queues);
6432 	}
6433 
6434 	vectors = rx_queues + admincnt;
6435 	if (msgs < vectors) {
6436 		device_printf(dev,
6437 		    "insufficient number of MSI-X vectors "
6438 		    "(supported %d, need %d)\n", msgs, vectors);
6439 		goto msi;
6440 	}
6441 
6442 	device_printf(dev, "Using %d RX queues %d TX queues\n", rx_queues,
6443 	    tx_queues);
6444 	msgs = vectors;
6445 	if ((err = pci_alloc_msix(dev, &vectors)) == 0) {
6446 		if (vectors != msgs) {
6447 			device_printf(dev,
6448 			    "Unable to allocate sufficient MSI-X vectors "
6449 			    "(got %d, need %d)\n", vectors, msgs);
6450 			pci_release_msi(dev);
6451 			if (bar != -1) {
6452 				bus_release_resource(dev, SYS_RES_MEMORY, bar,
6453 				    ctx->ifc_msix_mem);
6454 				ctx->ifc_msix_mem = NULL;
6455 			}
6456 			goto msi;
6457 		}
6458 		device_printf(dev, "Using MSI-X interrupts with %d vectors\n",
6459 		    vectors);
6460 		scctx->isc_vectors = vectors;
6461 		scctx->isc_nrxqsets = rx_queues;
6462 		scctx->isc_ntxqsets = tx_queues;
6463 		scctx->isc_intr = IFLIB_INTR_MSIX;
6464 
6465 		return (vectors);
6466 	} else {
6467 		device_printf(dev,
6468 		    "failed to allocate %d MSI-X vectors, err: %d\n", vectors,
6469 		    err);
6470 		if (bar != -1) {
6471 			bus_release_resource(dev, SYS_RES_MEMORY, bar,
6472 			    ctx->ifc_msix_mem);
6473 			ctx->ifc_msix_mem = NULL;
6474 		}
6475 	}
6476 
6477 msi:
6478 	vectors = pci_msi_count(dev);
6479 	scctx->isc_nrxqsets = 1;
6480 	scctx->isc_ntxqsets = 1;
6481 	scctx->isc_vectors = vectors;
6482 	if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) {
6483 		device_printf(dev,"Using an MSI interrupt\n");
6484 		scctx->isc_intr = IFLIB_INTR_MSI;
6485 	} else {
6486 		scctx->isc_vectors = 1;
6487 		device_printf(dev,"Using a Legacy interrupt\n");
6488 		scctx->isc_intr = IFLIB_INTR_LEGACY;
6489 	}
6490 
6491 	return (vectors);
6492 }
6493 
6494 static const char *ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" };
6495 
6496 static int
6497 mp_ring_state_handler(SYSCTL_HANDLER_ARGS)
6498 {
6499 	int rc;
6500 	uint16_t *state = ((uint16_t *)oidp->oid_arg1);
6501 	struct sbuf *sb;
6502 	const char *ring_state = "UNKNOWN";
6503 
6504 	/* XXX needed ? */
6505 	rc = sysctl_wire_old_buffer(req, 0);
6506 	MPASS(rc == 0);
6507 	if (rc != 0)
6508 		return (rc);
6509 	sb = sbuf_new_for_sysctl(NULL, NULL, 80, req);
6510 	MPASS(sb != NULL);
6511 	if (sb == NULL)
6512 		return (ENOMEM);
6513 	if (state[3] <= 3)
6514 		ring_state = ring_states[state[3]];
6515 
6516 	sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s",
6517 		    state[0], state[1], state[2], ring_state);
6518 	rc = sbuf_finish(sb);
6519 	sbuf_delete(sb);
6520         return(rc);
6521 }
6522 
6523 enum iflib_ndesc_handler {
6524 	IFLIB_NTXD_HANDLER,
6525 	IFLIB_NRXD_HANDLER,
6526 };
6527 
6528 static int
6529 mp_ndesc_handler(SYSCTL_HANDLER_ARGS)
6530 {
6531 	if_ctx_t ctx = (void *)arg1;
6532 	enum iflib_ndesc_handler type = arg2;
6533 	char buf[256] = {0};
6534 	qidx_t *ndesc;
6535 	char *p, *next;
6536 	int nqs, rc, i;
6537 
6538 	nqs = 8;
6539 	switch(type) {
6540 	case IFLIB_NTXD_HANDLER:
6541 		ndesc = ctx->ifc_sysctl_ntxds;
6542 		if (ctx->ifc_sctx)
6543 			nqs = ctx->ifc_sctx->isc_ntxqs;
6544 		break;
6545 	case IFLIB_NRXD_HANDLER:
6546 		ndesc = ctx->ifc_sysctl_nrxds;
6547 		if (ctx->ifc_sctx)
6548 			nqs = ctx->ifc_sctx->isc_nrxqs;
6549 		break;
6550 	default:
6551 		printf("%s: unhandled type\n", __func__);
6552 		return (EINVAL);
6553 	}
6554 	if (nqs == 0)
6555 		nqs = 8;
6556 
6557 	for (i=0; i<8; i++) {
6558 		if (i >= nqs)
6559 			break;
6560 		if (i)
6561 			strcat(buf, ",");
6562 		sprintf(strchr(buf, 0), "%d", ndesc[i]);
6563 	}
6564 
6565 	rc = sysctl_handle_string(oidp, buf, sizeof(buf), req);
6566 	if (rc || req->newptr == NULL)
6567 		return rc;
6568 
6569 	for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p;
6570 	    i++, p = strsep(&next, " ,")) {
6571 		ndesc[i] = strtoul(p, NULL, 10);
6572 	}
6573 
6574 	return(rc);
6575 }
6576 
6577 #define NAME_BUFLEN 32
6578 static void
6579 iflib_add_device_sysctl_pre(if_ctx_t ctx)
6580 {
6581         device_t dev = iflib_get_dev(ctx);
6582 	struct sysctl_oid_list *child, *oid_list;
6583 	struct sysctl_ctx_list *ctx_list;
6584 	struct sysctl_oid *node;
6585 
6586 	ctx_list = device_get_sysctl_ctx(dev);
6587 	child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev));
6588 	ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, "iflib",
6589 	    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "IFLIB fields");
6590 	oid_list = SYSCTL_CHILDREN(node);
6591 
6592 	SYSCTL_ADD_CONST_STRING(ctx_list, oid_list, OID_AUTO, "driver_version",
6593 		       CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version,
6594 		       "driver version");
6595 
6596 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs",
6597 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0,
6598 			"# of txqs to use, 0 => use default #");
6599 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs",
6600 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0,
6601 			"# of rxqs to use, 0 => use default #");
6602 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_qs_enable",
6603 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0,
6604                        "permit #txq != #rxq");
6605 	SYSCTL_ADD_INT(ctx_list, oid_list, OID_AUTO, "disable_msix",
6606                       CTLFLAG_RWTUN, &ctx->ifc_softc_ctx.isc_disable_msix, 0,
6607                       "disable MSI-X (default 0)");
6608 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "rx_budget",
6609 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_rx_budget, 0,
6610 		       "set the RX budget");
6611 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "tx_abdicate",
6612 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_tx_abdicate, 0,
6613 		       "cause TX to abdicate instead of running to completion");
6614 	ctx->ifc_sysctl_core_offset = CORE_OFFSET_UNSPECIFIED;
6615 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "core_offset",
6616 		       CTLFLAG_RDTUN, &ctx->ifc_sysctl_core_offset, 0,
6617 		       "offset to start using cores at");
6618 	SYSCTL_ADD_U8(ctx_list, oid_list, OID_AUTO, "separate_txrx",
6619 		       CTLFLAG_RDTUN, &ctx->ifc_sysctl_separate_txrx, 0,
6620 		       "use separate cores for TX and RX");
6621 
6622 	/* XXX change for per-queue sizes */
6623 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_ntxds",
6624 	    CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NEEDGIANT, ctx,
6625 	    IFLIB_NTXD_HANDLER, mp_ndesc_handler, "A",
6626 	    "list of # of TX descriptors to use, 0 = use default #");
6627 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_nrxds",
6628 	    CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NEEDGIANT, ctx,
6629 	    IFLIB_NRXD_HANDLER, mp_ndesc_handler, "A",
6630 	    "list of # of RX descriptors to use, 0 = use default #");
6631 }
6632 
6633 static void
6634 iflib_add_device_sysctl_post(if_ctx_t ctx)
6635 {
6636 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6637 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6638         device_t dev = iflib_get_dev(ctx);
6639 	struct sysctl_oid_list *child;
6640 	struct sysctl_ctx_list *ctx_list;
6641 	iflib_fl_t fl;
6642 	iflib_txq_t txq;
6643 	iflib_rxq_t rxq;
6644 	int i, j;
6645 	char namebuf[NAME_BUFLEN];
6646 	char *qfmt;
6647 	struct sysctl_oid *queue_node, *fl_node, *node;
6648 	struct sysctl_oid_list *queue_list, *fl_list;
6649 	ctx_list = device_get_sysctl_ctx(dev);
6650 
6651 	node = ctx->ifc_sysctl_node;
6652 	child = SYSCTL_CHILDREN(node);
6653 
6654 	if (scctx->isc_ntxqsets > 100)
6655 		qfmt = "txq%03d";
6656 	else if (scctx->isc_ntxqsets > 10)
6657 		qfmt = "txq%02d";
6658 	else
6659 		qfmt = "txq%d";
6660 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
6661 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6662 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6663 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "Queue Name");
6664 		queue_list = SYSCTL_CHILDREN(queue_node);
6665 #if MEMORY_LOGGING
6666 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued",
6667 				CTLFLAG_RD,
6668 				&txq->ift_dequeued, "total mbufs freed");
6669 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued",
6670 				CTLFLAG_RD,
6671 				&txq->ift_enqueued, "total mbufs enqueued");
6672 #endif
6673 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag",
6674 				   CTLFLAG_RD,
6675 				   &txq->ift_mbuf_defrag, "# of times m_defrag was called");
6676 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "m_pullups",
6677 				   CTLFLAG_RD,
6678 				   &txq->ift_pullups, "# of times m_pullup was called");
6679 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag_failed",
6680 				   CTLFLAG_RD,
6681 				   &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed");
6682 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_desc_avail",
6683 				   CTLFLAG_RD,
6684 				   &txq->ift_no_desc_avail, "# of times no descriptors were available");
6685 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "tx_map_failed",
6686 				   CTLFLAG_RD,
6687 				   &txq->ift_map_failed, "# of times DMA map failed");
6688 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txd_encap_efbig",
6689 				   CTLFLAG_RD,
6690 				   &txq->ift_txd_encap_efbig, "# of times txd_encap returned EFBIG");
6691 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_tx_dma_setup",
6692 				   CTLFLAG_RD,
6693 				   &txq->ift_no_tx_dma_setup, "# of times map failed for other than EFBIG");
6694 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx",
6695 				   CTLFLAG_RD,
6696 				   &txq->ift_pidx, 1, "Producer Index");
6697 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx",
6698 				   CTLFLAG_RD,
6699 				   &txq->ift_cidx, 1, "Consumer Index");
6700 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx_processed",
6701 				   CTLFLAG_RD,
6702 				   &txq->ift_cidx_processed, 1, "Consumer Index seen by credit update");
6703 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use",
6704 				   CTLFLAG_RD,
6705 				   &txq->ift_in_use, 1, "descriptors in use");
6706 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_processed",
6707 				   CTLFLAG_RD,
6708 				   &txq->ift_processed, "descriptors procesed for clean");
6709 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned",
6710 				   CTLFLAG_RD,
6711 				   &txq->ift_cleaned, "total cleaned");
6712 		SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state",
6713 		    CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_NEEDGIANT,
6714 		    __DEVOLATILE(uint64_t *, &txq->ift_br->state), 0,
6715 		    mp_ring_state_handler, "A", "soft ring state");
6716 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_enqueues",
6717 				       CTLFLAG_RD, &txq->ift_br->enqueues,
6718 				       "# of enqueues to the mp_ring for this queue");
6719 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_drops",
6720 				       CTLFLAG_RD, &txq->ift_br->drops,
6721 				       "# of drops in the mp_ring for this queue");
6722 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_starts",
6723 				       CTLFLAG_RD, &txq->ift_br->starts,
6724 				       "# of normal consumer starts in the mp_ring for this queue");
6725 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_stalls",
6726 				       CTLFLAG_RD, &txq->ift_br->stalls,
6727 					       "# of consumer stalls in the mp_ring for this queue");
6728 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_restarts",
6729 			       CTLFLAG_RD, &txq->ift_br->restarts,
6730 				       "# of consumer restarts in the mp_ring for this queue");
6731 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_abdications",
6732 				       CTLFLAG_RD, &txq->ift_br->abdications,
6733 				       "# of consumer abdications in the mp_ring for this queue");
6734 	}
6735 
6736 	if (scctx->isc_nrxqsets > 100)
6737 		qfmt = "rxq%03d";
6738 	else if (scctx->isc_nrxqsets > 10)
6739 		qfmt = "rxq%02d";
6740 	else
6741 		qfmt = "rxq%d";
6742 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
6743 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6744 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6745 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "Queue Name");
6746 		queue_list = SYSCTL_CHILDREN(queue_node);
6747 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
6748 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_cidx",
6749 				       CTLFLAG_RD,
6750 				       &rxq->ifr_cq_cidx, 1, "Consumer Index");
6751 		}
6752 
6753 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
6754 			snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j);
6755 			fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list, OID_AUTO, namebuf,
6756 			    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "freelist Name");
6757 			fl_list = SYSCTL_CHILDREN(fl_node);
6758 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx",
6759 				       CTLFLAG_RD,
6760 				       &fl->ifl_pidx, 1, "Producer Index");
6761 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx",
6762 				       CTLFLAG_RD,
6763 				       &fl->ifl_cidx, 1, "Consumer Index");
6764 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits",
6765 				       CTLFLAG_RD,
6766 				       &fl->ifl_credits, 1, "credits available");
6767 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "buf_size",
6768 				       CTLFLAG_RD,
6769 				       &fl->ifl_buf_size, 1, "buffer size");
6770 #if MEMORY_LOGGING
6771 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_enqueued",
6772 					CTLFLAG_RD,
6773 					&fl->ifl_m_enqueued, "mbufs allocated");
6774 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_dequeued",
6775 					CTLFLAG_RD,
6776 					&fl->ifl_m_dequeued, "mbufs freed");
6777 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_enqueued",
6778 					CTLFLAG_RD,
6779 					&fl->ifl_cl_enqueued, "clusters allocated");
6780 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_dequeued",
6781 					CTLFLAG_RD,
6782 					&fl->ifl_cl_dequeued, "clusters freed");
6783 #endif
6784 		}
6785 	}
6786 
6787 }
6788 
6789 void
6790 iflib_request_reset(if_ctx_t ctx)
6791 {
6792 
6793 	STATE_LOCK(ctx);
6794 	ctx->ifc_flags |= IFC_DO_RESET;
6795 	STATE_UNLOCK(ctx);
6796 }
6797 
6798 #ifndef __NO_STRICT_ALIGNMENT
6799 static struct mbuf *
6800 iflib_fixup_rx(struct mbuf *m)
6801 {
6802 	struct mbuf *n;
6803 
6804 	if (m->m_len <= (MCLBYTES - ETHER_HDR_LEN)) {
6805 		bcopy(m->m_data, m->m_data + ETHER_HDR_LEN, m->m_len);
6806 		m->m_data += ETHER_HDR_LEN;
6807 		n = m;
6808 	} else {
6809 		MGETHDR(n, M_NOWAIT, MT_DATA);
6810 		if (n == NULL) {
6811 			m_freem(m);
6812 			return (NULL);
6813 		}
6814 		bcopy(m->m_data, n->m_data, ETHER_HDR_LEN);
6815 		m->m_data += ETHER_HDR_LEN;
6816 		m->m_len -= ETHER_HDR_LEN;
6817 		n->m_len = ETHER_HDR_LEN;
6818 		M_MOVE_PKTHDR(n, m);
6819 		n->m_next = m;
6820 	}
6821 	return (n);
6822 }
6823 #endif
6824 
6825 #ifdef DEBUGNET
6826 static void
6827 iflib_debugnet_init(if_t ifp, int *nrxr, int *ncl, int *clsize)
6828 {
6829 	if_ctx_t ctx;
6830 
6831 	ctx = if_getsoftc(ifp);
6832 	CTX_LOCK(ctx);
6833 	*nrxr = NRXQSETS(ctx);
6834 	*ncl = ctx->ifc_rxqs[0].ifr_fl->ifl_size;
6835 	*clsize = ctx->ifc_rxqs[0].ifr_fl->ifl_buf_size;
6836 	CTX_UNLOCK(ctx);
6837 }
6838 
6839 static void
6840 iflib_debugnet_event(if_t ifp, enum debugnet_ev event)
6841 {
6842 	if_ctx_t ctx;
6843 	if_softc_ctx_t scctx;
6844 	iflib_fl_t fl;
6845 	iflib_rxq_t rxq;
6846 	int i, j;
6847 
6848 	ctx = if_getsoftc(ifp);
6849 	scctx = &ctx->ifc_softc_ctx;
6850 
6851 	switch (event) {
6852 	case DEBUGNET_START:
6853 		for (i = 0; i < scctx->isc_nrxqsets; i++) {
6854 			rxq = &ctx->ifc_rxqs[i];
6855 			for (j = 0; j < rxq->ifr_nfl; j++) {
6856 				fl = rxq->ifr_fl;
6857 				fl->ifl_zone = m_getzone(fl->ifl_buf_size);
6858 			}
6859 		}
6860 		iflib_no_tx_batch = 1;
6861 		break;
6862 	default:
6863 		break;
6864 	}
6865 }
6866 
6867 static int
6868 iflib_debugnet_transmit(if_t ifp, struct mbuf *m)
6869 {
6870 	if_ctx_t ctx;
6871 	iflib_txq_t txq;
6872 	int error;
6873 
6874 	ctx = if_getsoftc(ifp);
6875 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
6876 	    IFF_DRV_RUNNING)
6877 		return (EBUSY);
6878 
6879 	txq = &ctx->ifc_txqs[0];
6880 	error = iflib_encap(txq, &m);
6881 	if (error == 0)
6882 		(void)iflib_txd_db_check(ctx, txq, true, txq->ift_in_use);
6883 	return (error);
6884 }
6885 
6886 static int
6887 iflib_debugnet_poll(if_t ifp, int count)
6888 {
6889 	struct epoch_tracker et;
6890 	if_ctx_t ctx;
6891 	if_softc_ctx_t scctx;
6892 	iflib_txq_t txq;
6893 	int i;
6894 
6895 	ctx = if_getsoftc(ifp);
6896 	scctx = &ctx->ifc_softc_ctx;
6897 
6898 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
6899 	    IFF_DRV_RUNNING)
6900 		return (EBUSY);
6901 
6902 	txq = &ctx->ifc_txqs[0];
6903 	(void)iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
6904 
6905 	NET_EPOCH_ENTER(et);
6906 	for (i = 0; i < scctx->isc_nrxqsets; i++)
6907 		(void)iflib_rxeof(&ctx->ifc_rxqs[i], 16 /* XXX */);
6908 	NET_EPOCH_EXIT(et);
6909 	return (0);
6910 }
6911 #endif /* DEBUGNET */
6912