xref: /freebsd/sys/net/iflib.c (revision 6132212808e8dccedc9e5d85fea4390c2f38059a)
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 
350 	if_txsd_vec_t	ift_sds;
351 	uint8_t		ift_qstatus;
352 	uint8_t		ift_closed;
353 	uint8_t		ift_update_freq;
354 	struct iflib_filter_info ift_filter_info;
355 	bus_dma_tag_t	ift_buf_tag;
356 	bus_dma_tag_t	ift_tso_buf_tag;
357 	iflib_dma_info_t	ift_ifdi;
358 #define	MTX_NAME_LEN	32
359 	char                    ift_mtx_name[MTX_NAME_LEN];
360 	bus_dma_segment_t	ift_segs[IFLIB_MAX_TX_SEGS]  __aligned(CACHE_LINE_SIZE);
361 #ifdef IFLIB_DIAGNOSTICS
362 	uint64_t ift_cpu_exec_count[256];
363 #endif
364 } __aligned(CACHE_LINE_SIZE);
365 
366 struct iflib_fl {
367 	qidx_t		ifl_cidx;
368 	qidx_t		ifl_pidx;
369 	qidx_t		ifl_credits;
370 	uint8_t		ifl_gen;
371 	uint8_t		ifl_rxd_size;
372 #if MEMORY_LOGGING
373 	uint64_t	ifl_m_enqueued;
374 	uint64_t	ifl_m_dequeued;
375 	uint64_t	ifl_cl_enqueued;
376 	uint64_t	ifl_cl_dequeued;
377 #endif
378 	/* implicit pad */
379 	bitstr_t 	*ifl_rx_bitmap;
380 	qidx_t		ifl_fragidx;
381 	/* constant */
382 	qidx_t		ifl_size;
383 	uint16_t	ifl_buf_size;
384 	uint16_t	ifl_cltype;
385 	uma_zone_t	ifl_zone;
386 	iflib_rxsd_array_t	ifl_sds;
387 	iflib_rxq_t	ifl_rxq;
388 	uint8_t		ifl_id;
389 	bus_dma_tag_t	ifl_buf_tag;
390 	iflib_dma_info_t	ifl_ifdi;
391 	uint64_t	ifl_bus_addrs[IFLIB_MAX_RX_REFRESH] __aligned(CACHE_LINE_SIZE);
392 	qidx_t		ifl_rxd_idxs[IFLIB_MAX_RX_REFRESH];
393 }  __aligned(CACHE_LINE_SIZE);
394 
395 static inline qidx_t
396 get_inuse(int size, qidx_t cidx, qidx_t pidx, uint8_t gen)
397 {
398 	qidx_t used;
399 
400 	if (pidx > cidx)
401 		used = pidx - cidx;
402 	else if (pidx < cidx)
403 		used = size - cidx + pidx;
404 	else if (gen == 0 && pidx == cidx)
405 		used = 0;
406 	else if (gen == 1 && pidx == cidx)
407 		used = size;
408 	else
409 		panic("bad state");
410 
411 	return (used);
412 }
413 
414 #define TXQ_AVAIL(txq) (txq->ift_size - get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx, txq->ift_gen))
415 
416 #define IDXDIFF(head, tail, wrap) \
417 	((head) >= (tail) ? (head) - (tail) : (wrap) - (tail) + (head))
418 
419 struct iflib_rxq {
420 	if_ctx_t	ifr_ctx;
421 	iflib_fl_t	ifr_fl;
422 	uint64_t	ifr_rx_irq;
423 	struct pfil_head	*pfil;
424 	/*
425 	 * If there is a separate completion queue (IFLIB_HAS_RXCQ), this is
426 	 * the completion queue consumer index.  Otherwise it's unused.
427 	 */
428 	qidx_t		ifr_cq_cidx;
429 	uint16_t	ifr_id;
430 	uint8_t		ifr_nfl;
431 	uint8_t		ifr_ntxqirq;
432 	uint8_t		ifr_txqid[IFLIB_MAX_TX_SHARED_INTR];
433 	uint8_t		ifr_fl_offset;
434 	struct lro_ctrl			ifr_lc;
435 	struct grouptask        ifr_task;
436 	struct callout		ifr_watchdog;
437 	struct iflib_filter_info ifr_filter_info;
438 	iflib_dma_info_t		ifr_ifdi;
439 
440 	/* dynamically allocate if any drivers need a value substantially larger than this */
441 	struct if_rxd_frag	ifr_frags[IFLIB_MAX_RX_SEGS] __aligned(CACHE_LINE_SIZE);
442 #ifdef IFLIB_DIAGNOSTICS
443 	uint64_t ifr_cpu_exec_count[256];
444 #endif
445 }  __aligned(CACHE_LINE_SIZE);
446 
447 typedef struct if_rxsd {
448 	caddr_t *ifsd_cl;
449 	iflib_fl_t ifsd_fl;
450 } *if_rxsd_t;
451 
452 /* multiple of word size */
453 #ifdef __LP64__
454 #define PKT_INFO_SIZE	6
455 #define RXD_INFO_SIZE	5
456 #define PKT_TYPE uint64_t
457 #else
458 #define PKT_INFO_SIZE	11
459 #define RXD_INFO_SIZE	8
460 #define PKT_TYPE uint32_t
461 #endif
462 #define PKT_LOOP_BOUND  ((PKT_INFO_SIZE/3)*3)
463 #define RXD_LOOP_BOUND  ((RXD_INFO_SIZE/4)*4)
464 
465 typedef struct if_pkt_info_pad {
466 	PKT_TYPE pkt_val[PKT_INFO_SIZE];
467 } *if_pkt_info_pad_t;
468 typedef struct if_rxd_info_pad {
469 	PKT_TYPE rxd_val[RXD_INFO_SIZE];
470 } *if_rxd_info_pad_t;
471 
472 CTASSERT(sizeof(struct if_pkt_info_pad) == sizeof(struct if_pkt_info));
473 CTASSERT(sizeof(struct if_rxd_info_pad) == sizeof(struct if_rxd_info));
474 
475 static inline void
476 pkt_info_zero(if_pkt_info_t pi)
477 {
478 	if_pkt_info_pad_t pi_pad;
479 
480 	pi_pad = (if_pkt_info_pad_t)pi;
481 	pi_pad->pkt_val[0] = 0; pi_pad->pkt_val[1] = 0; pi_pad->pkt_val[2] = 0;
482 	pi_pad->pkt_val[3] = 0; pi_pad->pkt_val[4] = 0; pi_pad->pkt_val[5] = 0;
483 #ifndef __LP64__
484 	pi_pad->pkt_val[6] = 0; pi_pad->pkt_val[7] = 0; pi_pad->pkt_val[8] = 0;
485 	pi_pad->pkt_val[9] = 0; pi_pad->pkt_val[10] = 0;
486 #endif
487 }
488 
489 static device_method_t iflib_pseudo_methods[] = {
490 	DEVMETHOD(device_attach, noop_attach),
491 	DEVMETHOD(device_detach, iflib_pseudo_detach),
492 	DEVMETHOD_END
493 };
494 
495 driver_t iflib_pseudodriver = {
496 	"iflib_pseudo", iflib_pseudo_methods, sizeof(struct iflib_ctx),
497 };
498 
499 static inline void
500 rxd_info_zero(if_rxd_info_t ri)
501 {
502 	if_rxd_info_pad_t ri_pad;
503 	int i;
504 
505 	ri_pad = (if_rxd_info_pad_t)ri;
506 	for (i = 0; i < RXD_LOOP_BOUND; i += 4) {
507 		ri_pad->rxd_val[i] = 0;
508 		ri_pad->rxd_val[i+1] = 0;
509 		ri_pad->rxd_val[i+2] = 0;
510 		ri_pad->rxd_val[i+3] = 0;
511 	}
512 #ifdef __LP64__
513 	ri_pad->rxd_val[RXD_INFO_SIZE-1] = 0;
514 #endif
515 }
516 
517 /*
518  * Only allow a single packet to take up most 1/nth of the tx ring
519  */
520 #define MAX_SINGLE_PACKET_FRACTION 12
521 #define IF_BAD_DMA (bus_addr_t)-1
522 
523 #define CTX_ACTIVE(ctx) ((if_getdrvflags((ctx)->ifc_ifp) & IFF_DRV_RUNNING))
524 
525 #define CTX_LOCK_INIT(_sc)  sx_init(&(_sc)->ifc_ctx_sx, "iflib ctx lock")
526 #define CTX_LOCK(ctx) sx_xlock(&(ctx)->ifc_ctx_sx)
527 #define CTX_UNLOCK(ctx) sx_xunlock(&(ctx)->ifc_ctx_sx)
528 #define CTX_LOCK_DESTROY(ctx) sx_destroy(&(ctx)->ifc_ctx_sx)
529 
530 #define STATE_LOCK_INIT(_sc, _name)  mtx_init(&(_sc)->ifc_state_mtx, _name, "iflib state lock", MTX_DEF)
531 #define STATE_LOCK(ctx) mtx_lock(&(ctx)->ifc_state_mtx)
532 #define STATE_UNLOCK(ctx) mtx_unlock(&(ctx)->ifc_state_mtx)
533 #define STATE_LOCK_DESTROY(ctx) mtx_destroy(&(ctx)->ifc_state_mtx)
534 
535 #define CALLOUT_LOCK(txq)	mtx_lock(&txq->ift_mtx)
536 #define CALLOUT_UNLOCK(txq) 	mtx_unlock(&txq->ift_mtx)
537 
538 void
539 iflib_set_detach(if_ctx_t ctx)
540 {
541 	STATE_LOCK(ctx);
542 	ctx->ifc_flags |= IFC_IN_DETACH;
543 	STATE_UNLOCK(ctx);
544 }
545 
546 /* Our boot-time initialization hook */
547 static int	iflib_module_event_handler(module_t, int, void *);
548 
549 static moduledata_t iflib_moduledata = {
550 	"iflib",
551 	iflib_module_event_handler,
552 	NULL
553 };
554 
555 DECLARE_MODULE(iflib, iflib_moduledata, SI_SUB_INIT_IF, SI_ORDER_ANY);
556 MODULE_VERSION(iflib, 1);
557 
558 MODULE_DEPEND(iflib, pci, 1, 1, 1);
559 MODULE_DEPEND(iflib, ether, 1, 1, 1);
560 
561 TASKQGROUP_DEFINE(if_io_tqg, mp_ncpus, 1);
562 TASKQGROUP_DEFINE(if_config_tqg, 1, 1);
563 
564 #ifndef IFLIB_DEBUG_COUNTERS
565 #ifdef INVARIANTS
566 #define IFLIB_DEBUG_COUNTERS 1
567 #else
568 #define IFLIB_DEBUG_COUNTERS 0
569 #endif /* !INVARIANTS */
570 #endif
571 
572 static SYSCTL_NODE(_net, OID_AUTO, iflib, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
573     "iflib driver parameters");
574 
575 /*
576  * XXX need to ensure that this can't accidentally cause the head to be moved backwards
577  */
578 static int iflib_min_tx_latency = 0;
579 SYSCTL_INT(_net_iflib, OID_AUTO, min_tx_latency, CTLFLAG_RW,
580 		   &iflib_min_tx_latency, 0, "minimize transmit latency at the possible expense of throughput");
581 static int iflib_no_tx_batch = 0;
582 SYSCTL_INT(_net_iflib, OID_AUTO, no_tx_batch, CTLFLAG_RW,
583 		   &iflib_no_tx_batch, 0, "minimize transmit latency at the possible expense of throughput");
584 
585 #if IFLIB_DEBUG_COUNTERS
586 
587 static int iflib_tx_seen;
588 static int iflib_tx_sent;
589 static int iflib_tx_encap;
590 static int iflib_rx_allocs;
591 static int iflib_fl_refills;
592 static int iflib_fl_refills_large;
593 static int iflib_tx_frees;
594 
595 SYSCTL_INT(_net_iflib, OID_AUTO, tx_seen, CTLFLAG_RD,
596 		   &iflib_tx_seen, 0, "# TX mbufs seen");
597 SYSCTL_INT(_net_iflib, OID_AUTO, tx_sent, CTLFLAG_RD,
598 		   &iflib_tx_sent, 0, "# TX mbufs sent");
599 SYSCTL_INT(_net_iflib, OID_AUTO, tx_encap, CTLFLAG_RD,
600 		   &iflib_tx_encap, 0, "# TX mbufs encapped");
601 SYSCTL_INT(_net_iflib, OID_AUTO, tx_frees, CTLFLAG_RD,
602 		   &iflib_tx_frees, 0, "# TX frees");
603 SYSCTL_INT(_net_iflib, OID_AUTO, rx_allocs, CTLFLAG_RD,
604 		   &iflib_rx_allocs, 0, "# RX allocations");
605 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills, CTLFLAG_RD,
606 		   &iflib_fl_refills, 0, "# refills");
607 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills_large, CTLFLAG_RD,
608 		   &iflib_fl_refills_large, 0, "# large refills");
609 
610 static int iflib_txq_drain_flushing;
611 static int iflib_txq_drain_oactive;
612 static int iflib_txq_drain_notready;
613 
614 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_flushing, CTLFLAG_RD,
615 		   &iflib_txq_drain_flushing, 0, "# drain flushes");
616 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_oactive, CTLFLAG_RD,
617 		   &iflib_txq_drain_oactive, 0, "# drain oactives");
618 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_notready, CTLFLAG_RD,
619 		   &iflib_txq_drain_notready, 0, "# drain notready");
620 
621 static int iflib_encap_load_mbuf_fail;
622 static int iflib_encap_pad_mbuf_fail;
623 static int iflib_encap_txq_avail_fail;
624 static int iflib_encap_txd_encap_fail;
625 
626 SYSCTL_INT(_net_iflib, OID_AUTO, encap_load_mbuf_fail, CTLFLAG_RD,
627 		   &iflib_encap_load_mbuf_fail, 0, "# busdma load failures");
628 SYSCTL_INT(_net_iflib, OID_AUTO, encap_pad_mbuf_fail, CTLFLAG_RD,
629 		   &iflib_encap_pad_mbuf_fail, 0, "# runt frame pad failures");
630 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txq_avail_fail, CTLFLAG_RD,
631 		   &iflib_encap_txq_avail_fail, 0, "# txq avail failures");
632 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txd_encap_fail, CTLFLAG_RD,
633 		   &iflib_encap_txd_encap_fail, 0, "# driver encap failures");
634 
635 static int iflib_task_fn_rxs;
636 static int iflib_rx_intr_enables;
637 static int iflib_fast_intrs;
638 static int iflib_rx_unavail;
639 static int iflib_rx_ctx_inactive;
640 static int iflib_rx_if_input;
641 static int iflib_rxd_flush;
642 
643 static int iflib_verbose_debug;
644 
645 SYSCTL_INT(_net_iflib, OID_AUTO, task_fn_rx, CTLFLAG_RD,
646 		   &iflib_task_fn_rxs, 0, "# task_fn_rx calls");
647 SYSCTL_INT(_net_iflib, OID_AUTO, rx_intr_enables, CTLFLAG_RD,
648 		   &iflib_rx_intr_enables, 0, "# RX intr enables");
649 SYSCTL_INT(_net_iflib, OID_AUTO, fast_intrs, CTLFLAG_RD,
650 		   &iflib_fast_intrs, 0, "# fast_intr calls");
651 SYSCTL_INT(_net_iflib, OID_AUTO, rx_unavail, CTLFLAG_RD,
652 		   &iflib_rx_unavail, 0, "# times rxeof called with no available data");
653 SYSCTL_INT(_net_iflib, OID_AUTO, rx_ctx_inactive, CTLFLAG_RD,
654 		   &iflib_rx_ctx_inactive, 0, "# times rxeof called with inactive context");
655 SYSCTL_INT(_net_iflib, OID_AUTO, rx_if_input, CTLFLAG_RD,
656 		   &iflib_rx_if_input, 0, "# times rxeof called if_input");
657 SYSCTL_INT(_net_iflib, OID_AUTO, rxd_flush, CTLFLAG_RD,
658 	         &iflib_rxd_flush, 0, "# times rxd_flush called");
659 SYSCTL_INT(_net_iflib, OID_AUTO, verbose_debug, CTLFLAG_RW,
660 		   &iflib_verbose_debug, 0, "enable verbose debugging");
661 
662 #define DBG_COUNTER_INC(name) atomic_add_int(&(iflib_ ## name), 1)
663 static void
664 iflib_debug_reset(void)
665 {
666 	iflib_tx_seen = iflib_tx_sent = iflib_tx_encap = iflib_rx_allocs =
667 		iflib_fl_refills = iflib_fl_refills_large = iflib_tx_frees =
668 		iflib_txq_drain_flushing = iflib_txq_drain_oactive =
669 		iflib_txq_drain_notready =
670 		iflib_encap_load_mbuf_fail = iflib_encap_pad_mbuf_fail =
671 		iflib_encap_txq_avail_fail = iflib_encap_txd_encap_fail =
672 		iflib_task_fn_rxs = iflib_rx_intr_enables = iflib_fast_intrs =
673 		iflib_rx_unavail =
674 		iflib_rx_ctx_inactive = iflib_rx_if_input =
675 		iflib_rxd_flush = 0;
676 }
677 
678 #else
679 #define DBG_COUNTER_INC(name)
680 static void iflib_debug_reset(void) {}
681 #endif
682 
683 #define IFLIB_DEBUG 0
684 
685 static void iflib_tx_structures_free(if_ctx_t ctx);
686 static void iflib_rx_structures_free(if_ctx_t ctx);
687 static int iflib_queues_alloc(if_ctx_t ctx);
688 static int iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq);
689 static int iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget);
690 static int iflib_qset_structures_setup(if_ctx_t ctx);
691 static int iflib_msix_init(if_ctx_t ctx);
692 static int iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filterarg, int *rid, const char *str);
693 static void iflib_txq_check_drain(iflib_txq_t txq, int budget);
694 static uint32_t iflib_txq_can_drain(struct ifmp_ring *);
695 #ifdef ALTQ
696 static void iflib_altq_if_start(if_t ifp);
697 static int iflib_altq_if_transmit(if_t ifp, struct mbuf *m);
698 #endif
699 static int iflib_register(if_ctx_t);
700 static void iflib_deregister(if_ctx_t);
701 static void iflib_unregister_vlan_handlers(if_ctx_t ctx);
702 static uint16_t iflib_get_mbuf_size_for(unsigned int size);
703 static void iflib_init_locked(if_ctx_t ctx);
704 static void iflib_add_device_sysctl_pre(if_ctx_t ctx);
705 static void iflib_add_device_sysctl_post(if_ctx_t ctx);
706 static void iflib_ifmp_purge(iflib_txq_t txq);
707 static void _iflib_pre_assert(if_softc_ctx_t scctx);
708 static void iflib_if_init_locked(if_ctx_t ctx);
709 static void iflib_free_intr_mem(if_ctx_t ctx);
710 #ifndef __NO_STRICT_ALIGNMENT
711 static struct mbuf * iflib_fixup_rx(struct mbuf *m);
712 #endif
713 
714 static SLIST_HEAD(cpu_offset_list, cpu_offset) cpu_offsets =
715     SLIST_HEAD_INITIALIZER(cpu_offsets);
716 struct cpu_offset {
717 	SLIST_ENTRY(cpu_offset) entries;
718 	cpuset_t	set;
719 	unsigned int	refcount;
720 	uint16_t	offset;
721 };
722 static struct mtx cpu_offset_mtx;
723 MTX_SYSINIT(iflib_cpu_offset, &cpu_offset_mtx, "iflib_cpu_offset lock",
724     MTX_DEF);
725 
726 DEBUGNET_DEFINE(iflib);
727 
728 static int
729 iflib_num_rx_descs(if_ctx_t ctx)
730 {
731 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
732 	if_shared_ctx_t sctx = ctx->ifc_sctx;
733 	uint16_t first_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
734 
735 	return scctx->isc_nrxd[first_rxq];
736 }
737 
738 static int
739 iflib_num_tx_descs(if_ctx_t ctx)
740 {
741 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
742 	if_shared_ctx_t sctx = ctx->ifc_sctx;
743 	uint16_t first_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
744 
745 	return scctx->isc_ntxd[first_txq];
746 }
747 
748 #ifdef DEV_NETMAP
749 #include <sys/selinfo.h>
750 #include <net/netmap.h>
751 #include <dev/netmap/netmap_kern.h>
752 
753 MODULE_DEPEND(iflib, netmap, 1, 1, 1);
754 
755 static int netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, bool init);
756 
757 /*
758  * device-specific sysctl variables:
759  *
760  * iflib_crcstrip: 0: keep CRC in rx frames (default), 1: strip it.
761  *	During regular operations the CRC is stripped, but on some
762  *	hardware reception of frames not multiple of 64 is slower,
763  *	so using crcstrip=0 helps in benchmarks.
764  *
765  * iflib_rx_miss, iflib_rx_miss_bufs:
766  *	count packets that might be missed due to lost interrupts.
767  */
768 SYSCTL_DECL(_dev_netmap);
769 /*
770  * The xl driver by default strips CRCs and we do not override it.
771  */
772 
773 int iflib_crcstrip = 1;
774 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_crcstrip,
775     CTLFLAG_RW, &iflib_crcstrip, 1, "strip CRC on RX frames");
776 
777 int iflib_rx_miss, iflib_rx_miss_bufs;
778 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss,
779     CTLFLAG_RW, &iflib_rx_miss, 0, "potentially missed RX intr");
780 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss_bufs,
781     CTLFLAG_RW, &iflib_rx_miss_bufs, 0, "potentially missed RX intr bufs");
782 
783 /*
784  * Register/unregister. We are already under netmap lock.
785  * Only called on the first register or the last unregister.
786  */
787 static int
788 iflib_netmap_register(struct netmap_adapter *na, int onoff)
789 {
790 	if_t ifp = na->ifp;
791 	if_ctx_t ctx = ifp->if_softc;
792 	int status;
793 
794 	CTX_LOCK(ctx);
795 	IFDI_INTR_DISABLE(ctx);
796 
797 	/* Tell the stack that the interface is no longer active */
798 	ifp->if_drv_flags &= ~(IFF_DRV_RUNNING | IFF_DRV_OACTIVE);
799 
800 	if (!CTX_IS_VF(ctx))
801 		IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip);
802 
803 	iflib_stop(ctx);
804 
805 	/*
806 	 * Enable (or disable) netmap flags, and intercept (or restore)
807 	 * ifp->if_transmit. This is done once the device has been stopped
808 	 * to prevent race conditions.
809 	 */
810 	if (onoff) {
811 		nm_set_native_flags(na);
812 	} else {
813 		nm_clear_native_flags(na);
814 	}
815 
816 	iflib_init_locked(ctx);
817 	IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip); // XXX why twice ?
818 	status = ifp->if_drv_flags & IFF_DRV_RUNNING ? 0 : 1;
819 	if (status)
820 		nm_clear_native_flags(na);
821 	CTX_UNLOCK(ctx);
822 	return (status);
823 }
824 
825 static int
826 netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, bool init)
827 {
828 	struct netmap_adapter *na = kring->na;
829 	u_int const lim = kring->nkr_num_slots - 1;
830 	u_int nm_i = kring->nr_hwcur;
831 	struct netmap_ring *ring = kring->ring;
832 	bus_dmamap_t *map;
833 	struct if_rxd_update iru;
834 	if_ctx_t ctx = rxq->ifr_ctx;
835 	iflib_fl_t fl = &rxq->ifr_fl[0];
836 	u_int nic_i_first, nic_i;
837 	int i, n;
838 #if IFLIB_DEBUG_COUNTERS
839 	int rf_count = 0;
840 #endif
841 
842 	/*
843 	 * This function is used both at initialization and in rxsync.
844 	 * At initialization we need to prepare (with isc_rxd_refill())
845 	 * all the (N) netmap buffers in the ring, in such a way to keep
846 	 * fl->ifl_pidx and kring->nr_hwcur in sync (except for
847 	 * kring->nkr_hwofs); at rxsync time, both indexes point to the
848 	 * next buffer to be refilled.
849 	 * In any case we publish (with isc_rxd_flush()) up to
850 	 * (fl->ifl_pidx - 1) % N (included), to avoid the NIC tail/prod
851 	 * pointer to overrun the head/cons pointer, although this is
852 	 * not necessary for some NICs (e.g. vmx).
853 	 */
854 	if (__predict_false(init))
855 		n = kring->nkr_num_slots;
856 	else {
857 		n = kring->rhead - nm_i;
858 		if (n == 0)
859 			return (0); /* Nothing to do. */
860 		if (n < 0)
861 			n += kring->nkr_num_slots;
862 	}
863 
864 	/* Start to refill from nr_hwcur, publishing n buffers. */
865 	iru_init(&iru, rxq, 0 /* flid */);
866 	map = fl->ifl_sds.ifsd_map;
867 	nic_i = fl->ifl_pidx;
868 	MPASS(nic_i == netmap_idx_k2n(kring, nm_i));
869 	DBG_COUNTER_INC(fl_refills);
870 	while (n > 0) {
871 #if IFLIB_DEBUG_COUNTERS
872 		if (++rf_count == 9)
873 			DBG_COUNTER_INC(fl_refills_large);
874 #endif
875 		nic_i_first = nic_i;
876 		for (i = 0; n > 0 && i < IFLIB_MAX_RX_REFRESH; n--, i++) {
877 			struct netmap_slot *slot = &ring->slot[nm_i];
878 			void *addr = PNMB(na, slot, &fl->ifl_bus_addrs[i]);
879 
880 			MPASS(i < IFLIB_MAX_RX_REFRESH);
881 
882 			if (addr == NETMAP_BUF_BASE(na)) /* bad buf */
883 			        return netmap_ring_reinit(kring);
884 
885 			fl->ifl_rxd_idxs[i] = nic_i;
886 
887 			if (__predict_false(init)) {
888 				netmap_load_map(na, fl->ifl_buf_tag,
889 				    map[nic_i], addr);
890 			} else if (slot->flags & NS_BUF_CHANGED) {
891 				/* buffer has changed, reload map */
892 				netmap_reload_map(na, fl->ifl_buf_tag,
893 				    map[nic_i], addr);
894 			}
895 			bus_dmamap_sync(fl->ifl_buf_tag, map[nic_i],
896 			    BUS_DMASYNC_PREREAD);
897 			slot->flags &= ~NS_BUF_CHANGED;
898 
899 			nm_i = nm_next(nm_i, lim);
900 			nic_i = nm_next(nic_i, lim);
901 		}
902 
903 		iru.iru_pidx = nic_i_first;
904 		iru.iru_count = i;
905 		ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
906 	}
907 	fl->ifl_pidx = nic_i;
908 	MPASS(!init || nm_i == 0);
909 	MPASS(nm_i == kring->rhead);
910 	kring->nr_hwcur = nm_i;
911 
912 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
913 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
914 	ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id,
915 	    nm_prev(nic_i, lim));
916 	DBG_COUNTER_INC(rxd_flush);
917 
918 	return (0);
919 }
920 
921 /*
922  * Reconcile kernel and user view of the transmit ring.
923  *
924  * All information is in the kring.
925  * Userspace wants to send packets up to the one before kring->rhead,
926  * kernel knows kring->nr_hwcur is the first unsent packet.
927  *
928  * Here we push packets out (as many as possible), and possibly
929  * reclaim buffers from previously completed transmission.
930  *
931  * The caller (netmap) guarantees that there is only one instance
932  * running at any time. Any interference with other driver
933  * methods should be handled by the individual drivers.
934  */
935 static int
936 iflib_netmap_txsync(struct netmap_kring *kring, int flags)
937 {
938 	struct netmap_adapter *na = kring->na;
939 	if_t ifp = na->ifp;
940 	struct netmap_ring *ring = kring->ring;
941 	u_int nm_i;	/* index into the netmap kring */
942 	u_int nic_i;	/* index into the NIC ring */
943 	u_int n;
944 	u_int const lim = kring->nkr_num_slots - 1;
945 	u_int const head = kring->rhead;
946 	struct if_pkt_info pi;
947 
948 	/*
949 	 * interrupts on every tx packet are expensive so request
950 	 * them every half ring, or where NS_REPORT is set
951 	 */
952 	u_int report_frequency = kring->nkr_num_slots >> 1;
953 	/* device-specific */
954 	if_ctx_t ctx = ifp->if_softc;
955 	iflib_txq_t txq = &ctx->ifc_txqs[kring->ring_id];
956 
957 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
958 	    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
959 
960 	/*
961 	 * First part: process new packets to send.
962 	 * nm_i is the current index in the netmap kring,
963 	 * nic_i is the corresponding index in the NIC ring.
964 	 *
965 	 * If we have packets to send (nm_i != head)
966 	 * iterate over the netmap ring, fetch length and update
967 	 * the corresponding slot in the NIC ring. Some drivers also
968 	 * need to update the buffer's physical address in the NIC slot
969 	 * even NS_BUF_CHANGED is not set (PNMB computes the addresses).
970 	 *
971 	 * The netmap_reload_map() calls is especially expensive,
972 	 * even when (as in this case) the tag is 0, so do only
973 	 * when the buffer has actually changed.
974 	 *
975 	 * If possible do not set the report/intr bit on all slots,
976 	 * but only a few times per ring or when NS_REPORT is set.
977 	 *
978 	 * Finally, on 10G and faster drivers, it might be useful
979 	 * to prefetch the next slot and txr entry.
980 	 */
981 
982 	nm_i = kring->nr_hwcur;
983 	if (nm_i != head) {	/* we have new packets to send */
984 		pkt_info_zero(&pi);
985 		pi.ipi_segs = txq->ift_segs;
986 		pi.ipi_qsidx = kring->ring_id;
987 		nic_i = netmap_idx_k2n(kring, nm_i);
988 
989 		__builtin_prefetch(&ring->slot[nm_i]);
990 		__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i]);
991 		__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i]);
992 
993 		for (n = 0; nm_i != head; n++) {
994 			struct netmap_slot *slot = &ring->slot[nm_i];
995 			u_int len = slot->len;
996 			uint64_t paddr;
997 			void *addr = PNMB(na, slot, &paddr);
998 			int flags = (slot->flags & NS_REPORT ||
999 				nic_i == 0 || nic_i == report_frequency) ?
1000 				IPI_TX_INTR : 0;
1001 
1002 			/* device-specific */
1003 			pi.ipi_len = len;
1004 			pi.ipi_segs[0].ds_addr = paddr;
1005 			pi.ipi_segs[0].ds_len = len;
1006 			pi.ipi_nsegs = 1;
1007 			pi.ipi_ndescs = 0;
1008 			pi.ipi_pidx = nic_i;
1009 			pi.ipi_flags = flags;
1010 
1011 			/* Fill the slot in the NIC ring. */
1012 			ctx->isc_txd_encap(ctx->ifc_softc, &pi);
1013 			DBG_COUNTER_INC(tx_encap);
1014 
1015 			/* prefetch for next round */
1016 			__builtin_prefetch(&ring->slot[nm_i + 1]);
1017 			__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i + 1]);
1018 			__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i + 1]);
1019 
1020 			NM_CHECK_ADDR_LEN(na, addr, len);
1021 
1022 			if (slot->flags & NS_BUF_CHANGED) {
1023 				/* buffer has changed, reload map */
1024 				netmap_reload_map(na, txq->ift_buf_tag,
1025 				    txq->ift_sds.ifsd_map[nic_i], addr);
1026 			}
1027 			/* make sure changes to the buffer are synced */
1028 			bus_dmamap_sync(txq->ift_buf_tag,
1029 			    txq->ift_sds.ifsd_map[nic_i],
1030 			    BUS_DMASYNC_PREWRITE);
1031 
1032 			slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED);
1033 			nm_i = nm_next(nm_i, lim);
1034 			nic_i = nm_next(nic_i, lim);
1035 		}
1036 		kring->nr_hwcur = nm_i;
1037 
1038 		/* synchronize the NIC ring */
1039 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1040 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1041 
1042 		/* (re)start the tx unit up to slot nic_i (excluded) */
1043 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, nic_i);
1044 	}
1045 
1046 	/*
1047 	 * Second part: reclaim buffers for completed transmissions.
1048 	 *
1049 	 * If there are unclaimed buffers, attempt to reclaim them.
1050 	 * If none are reclaimed, and TX IRQs are not in use, do an initial
1051 	 * minimal delay, then trigger the tx handler which will spin in the
1052 	 * group task queue.
1053 	 */
1054 	if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1055 		if (iflib_tx_credits_update(ctx, txq)) {
1056 			/* some tx completed, increment avail */
1057 			nic_i = txq->ift_cidx_processed;
1058 			kring->nr_hwtail = nm_prev(netmap_idx_n2k(kring, nic_i), lim);
1059 		}
1060 	}
1061 	if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ))
1062 		if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1063 			callout_reset_on(&txq->ift_timer, hz < 2000 ? 1 : hz / 1000,
1064 			    iflib_timer, txq, txq->ift_timer.c_cpu);
1065 	}
1066 	return (0);
1067 }
1068 
1069 /*
1070  * Reconcile kernel and user view of the receive ring.
1071  * Same as for the txsync, this routine must be efficient.
1072  * The caller guarantees a single invocations, but races against
1073  * the rest of the driver should be handled here.
1074  *
1075  * On call, kring->rhead is the first packet that userspace wants
1076  * to keep, and kring->rcur is the wakeup point.
1077  * The kernel has previously reported packets up to kring->rtail.
1078  *
1079  * If (flags & NAF_FORCE_READ) also check for incoming packets irrespective
1080  * of whether or not we received an interrupt.
1081  */
1082 static int
1083 iflib_netmap_rxsync(struct netmap_kring *kring, int flags)
1084 {
1085 	struct netmap_adapter *na = kring->na;
1086 	struct netmap_ring *ring = kring->ring;
1087 	if_t ifp = na->ifp;
1088 	uint32_t nm_i;	/* index into the netmap ring */
1089 	uint32_t nic_i;	/* index into the NIC ring */
1090 	u_int n;
1091 	u_int const lim = kring->nkr_num_slots - 1;
1092 	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
1093 
1094 	if_ctx_t ctx = ifp->if_softc;
1095 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1096 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1097 	iflib_rxq_t rxq = &ctx->ifc_rxqs[kring->ring_id];
1098 	iflib_fl_t fl = &rxq->ifr_fl[0];
1099 	struct if_rxd_info ri;
1100 	qidx_t *cidxp;
1101 
1102 	/*
1103 	 * netmap only uses free list 0, to avoid out of order consumption
1104 	 * of receive buffers
1105 	 */
1106 
1107 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
1108 	    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1109 
1110 	/*
1111 	 * First part: import newly received packets.
1112 	 *
1113 	 * nm_i is the index of the next free slot in the netmap ring,
1114 	 * nic_i is the index of the next received packet in the NIC ring
1115 	 * (or in the free list 0 if IFLIB_HAS_RXCQ is set), and they may
1116 	 * differ in case if_init() has been called while
1117 	 * in netmap mode. For the receive ring we have
1118 	 *
1119 	 *	nic_i = fl->ifl_cidx;
1120 	 *	nm_i = kring->nr_hwtail (previous)
1121 	 * and
1122 	 *	nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1123 	 *
1124 	 * fl->ifl_cidx is set to 0 on a ring reinit
1125 	 */
1126 	if (netmap_no_pendintr || force_update) {
1127 		uint32_t hwtail_lim = nm_prev(kring->nr_hwcur, lim);
1128 		bool have_rxcq = sctx->isc_flags & IFLIB_HAS_RXCQ;
1129 		int crclen = iflib_crcstrip ? 0 : 4;
1130 		int error, avail;
1131 
1132 		/*
1133 		 * For the free list consumer index, we use the same
1134 		 * logic as in iflib_rxeof().
1135 		 */
1136 		if (have_rxcq)
1137 			cidxp = &rxq->ifr_cq_cidx;
1138 		else
1139 			cidxp = &fl->ifl_cidx;
1140 		avail = ctx->isc_rxd_available(ctx->ifc_softc,
1141 		    rxq->ifr_id, *cidxp, USHRT_MAX);
1142 
1143 		nic_i = fl->ifl_cidx;
1144 		nm_i = netmap_idx_n2k(kring, nic_i);
1145 		MPASS(nm_i == kring->nr_hwtail);
1146 		for (n = 0; avail > 0 && nm_i != hwtail_lim; n++, avail--) {
1147 			rxd_info_zero(&ri);
1148 			ri.iri_frags = rxq->ifr_frags;
1149 			ri.iri_qsidx = kring->ring_id;
1150 			ri.iri_ifp = ctx->ifc_ifp;
1151 			ri.iri_cidx = *cidxp;
1152 
1153 			error = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
1154 			ring->slot[nm_i].len = error ? 0 : ri.iri_len - crclen;
1155 			ring->slot[nm_i].flags = 0;
1156 			if (have_rxcq) {
1157 				*cidxp = ri.iri_cidx;
1158 				while (*cidxp >= scctx->isc_nrxd[0])
1159 					*cidxp -= scctx->isc_nrxd[0];
1160 			}
1161 			bus_dmamap_sync(fl->ifl_buf_tag,
1162 			    fl->ifl_sds.ifsd_map[nic_i], BUS_DMASYNC_POSTREAD);
1163 			nm_i = nm_next(nm_i, lim);
1164 			fl->ifl_cidx = nic_i = nm_next(nic_i, lim);
1165 		}
1166 		if (n) { /* update the state variables */
1167 			if (netmap_no_pendintr && !force_update) {
1168 				/* diagnostics */
1169 				iflib_rx_miss ++;
1170 				iflib_rx_miss_bufs += n;
1171 			}
1172 			kring->nr_hwtail = nm_i;
1173 		}
1174 		kring->nr_kflags &= ~NKR_PENDINTR;
1175 	}
1176 	/*
1177 	 * Second part: skip past packets that userspace has released.
1178 	 * (kring->nr_hwcur to head excluded),
1179 	 * and make the buffers available for reception.
1180 	 * As usual nm_i is the index in the netmap ring,
1181 	 * nic_i is the index in the NIC ring, and
1182 	 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1183 	 */
1184 	netmap_fl_refill(rxq, kring, false);
1185 
1186 	return (0);
1187 }
1188 
1189 static void
1190 iflib_netmap_intr(struct netmap_adapter *na, int onoff)
1191 {
1192 	if_ctx_t ctx = na->ifp->if_softc;
1193 
1194 	CTX_LOCK(ctx);
1195 	if (onoff) {
1196 		IFDI_INTR_ENABLE(ctx);
1197 	} else {
1198 		IFDI_INTR_DISABLE(ctx);
1199 	}
1200 	CTX_UNLOCK(ctx);
1201 }
1202 
1203 static int
1204 iflib_netmap_attach(if_ctx_t ctx)
1205 {
1206 	struct netmap_adapter na;
1207 
1208 	bzero(&na, sizeof(na));
1209 
1210 	na.ifp = ctx->ifc_ifp;
1211 	na.na_flags = NAF_BDG_MAYSLEEP;
1212 	MPASS(ctx->ifc_softc_ctx.isc_ntxqsets);
1213 	MPASS(ctx->ifc_softc_ctx.isc_nrxqsets);
1214 
1215 	na.num_tx_desc = iflib_num_tx_descs(ctx);
1216 	na.num_rx_desc = iflib_num_rx_descs(ctx);
1217 	na.nm_txsync = iflib_netmap_txsync;
1218 	na.nm_rxsync = iflib_netmap_rxsync;
1219 	na.nm_register = iflib_netmap_register;
1220 	na.nm_intr = iflib_netmap_intr;
1221 	na.num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
1222 	na.num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
1223 	return (netmap_attach(&na));
1224 }
1225 
1226 static int
1227 iflib_netmap_txq_init(if_ctx_t ctx, iflib_txq_t txq)
1228 {
1229 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1230 	struct netmap_slot *slot;
1231 
1232 	slot = netmap_reset(na, NR_TX, txq->ift_id, 0);
1233 	if (slot == NULL)
1234 		return (0);
1235 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxd[0]; i++) {
1236 		/*
1237 		 * In netmap mode, set the map for the packet buffer.
1238 		 * NOTE: Some drivers (not this one) also need to set
1239 		 * the physical buffer address in the NIC ring.
1240 		 * netmap_idx_n2k() maps a nic index, i, into the corresponding
1241 		 * netmap slot index, si
1242 		 */
1243 		int si = netmap_idx_n2k(na->tx_rings[txq->ift_id], i);
1244 		netmap_load_map(na, txq->ift_buf_tag, txq->ift_sds.ifsd_map[i],
1245 		    NMB(na, slot + si));
1246 	}
1247 	return (1);
1248 }
1249 
1250 static int
1251 iflib_netmap_rxq_init(if_ctx_t ctx, iflib_rxq_t rxq)
1252 {
1253 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1254 	struct netmap_kring *kring;
1255 	struct netmap_slot *slot;
1256 
1257 	slot = netmap_reset(na, NR_RX, rxq->ifr_id, 0);
1258 	if (slot == NULL)
1259 		return (0);
1260 	kring = na->rx_rings[rxq->ifr_id];
1261 	netmap_fl_refill(rxq, kring, true);
1262 	return (1);
1263 }
1264 
1265 static void
1266 iflib_netmap_timer_adjust(if_ctx_t ctx, iflib_txq_t txq, uint32_t *reset_on)
1267 {
1268 	struct netmap_kring *kring;
1269 	uint16_t txqid;
1270 
1271 	txqid = txq->ift_id;
1272 	kring = netmap_kring_on(NA(ctx->ifc_ifp), txqid, NR_TX);
1273 	if (kring == NULL)
1274 		return;
1275 
1276 	if (kring->nr_hwcur != nm_next(kring->nr_hwtail, kring->nkr_num_slots - 1)) {
1277 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1278 		    BUS_DMASYNC_POSTREAD);
1279 		if (ctx->isc_txd_credits_update(ctx->ifc_softc, txqid, false))
1280 			netmap_tx_irq(ctx->ifc_ifp, txqid);
1281 		if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ)) {
1282 			if (hz < 2000)
1283 				*reset_on = 1;
1284 			else
1285 				*reset_on = hz / 1000;
1286 		}
1287 	}
1288 }
1289 
1290 #define iflib_netmap_detach(ifp) netmap_detach(ifp)
1291 
1292 #else
1293 #define iflib_netmap_txq_init(ctx, txq) (0)
1294 #define iflib_netmap_rxq_init(ctx, rxq) (0)
1295 #define iflib_netmap_detach(ifp)
1296 
1297 #define iflib_netmap_attach(ctx) (0)
1298 #define netmap_rx_irq(ifp, qid, budget) (0)
1299 #define netmap_tx_irq(ifp, qid) do {} while (0)
1300 #define iflib_netmap_timer_adjust(ctx, txq, reset_on)
1301 #endif
1302 
1303 #if defined(__i386__) || defined(__amd64__)
1304 static __inline void
1305 prefetch(void *x)
1306 {
1307 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1308 }
1309 static __inline void
1310 prefetch2cachelines(void *x)
1311 {
1312 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1313 #if (CACHE_LINE_SIZE < 128)
1314 	__asm volatile("prefetcht0 %0" :: "m" (*(((unsigned long *)x)+CACHE_LINE_SIZE/(sizeof(unsigned long)))));
1315 #endif
1316 }
1317 #else
1318 #define prefetch(x)
1319 #define prefetch2cachelines(x)
1320 #endif
1321 
1322 static void
1323 iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid)
1324 {
1325 	iflib_fl_t fl;
1326 
1327 	fl = &rxq->ifr_fl[flid];
1328 	iru->iru_paddrs = fl->ifl_bus_addrs;
1329 	iru->iru_idxs = fl->ifl_rxd_idxs;
1330 	iru->iru_qsidx = rxq->ifr_id;
1331 	iru->iru_buf_size = fl->ifl_buf_size;
1332 	iru->iru_flidx = fl->ifl_id;
1333 }
1334 
1335 static void
1336 _iflib_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int err)
1337 {
1338 	if (err)
1339 		return;
1340 	*(bus_addr_t *) arg = segs[0].ds_addr;
1341 }
1342 
1343 int
1344 iflib_dma_alloc_align(if_ctx_t ctx, int size, int align, iflib_dma_info_t dma, int mapflags)
1345 {
1346 	int err;
1347 	device_t dev = ctx->ifc_dev;
1348 
1349 	err = bus_dma_tag_create(bus_get_dma_tag(dev),	/* parent */
1350 				align, 0,		/* alignment, bounds */
1351 				BUS_SPACE_MAXADDR,	/* lowaddr */
1352 				BUS_SPACE_MAXADDR,	/* highaddr */
1353 				NULL, NULL,		/* filter, filterarg */
1354 				size,			/* maxsize */
1355 				1,			/* nsegments */
1356 				size,			/* maxsegsize */
1357 				BUS_DMA_ALLOCNOW,	/* flags */
1358 				NULL,			/* lockfunc */
1359 				NULL,			/* lockarg */
1360 				&dma->idi_tag);
1361 	if (err) {
1362 		device_printf(dev,
1363 		    "%s: bus_dma_tag_create failed: %d\n",
1364 		    __func__, err);
1365 		goto fail_0;
1366 	}
1367 
1368 	err = bus_dmamem_alloc(dma->idi_tag, (void**) &dma->idi_vaddr,
1369 	    BUS_DMA_NOWAIT | BUS_DMA_COHERENT | BUS_DMA_ZERO, &dma->idi_map);
1370 	if (err) {
1371 		device_printf(dev,
1372 		    "%s: bus_dmamem_alloc(%ju) failed: %d\n",
1373 		    __func__, (uintmax_t)size, err);
1374 		goto fail_1;
1375 	}
1376 
1377 	dma->idi_paddr = IF_BAD_DMA;
1378 	err = bus_dmamap_load(dma->idi_tag, dma->idi_map, dma->idi_vaddr,
1379 	    size, _iflib_dmamap_cb, &dma->idi_paddr, mapflags | BUS_DMA_NOWAIT);
1380 	if (err || dma->idi_paddr == IF_BAD_DMA) {
1381 		device_printf(dev,
1382 		    "%s: bus_dmamap_load failed: %d\n",
1383 		    __func__, err);
1384 		goto fail_2;
1385 	}
1386 
1387 	dma->idi_size = size;
1388 	return (0);
1389 
1390 fail_2:
1391 	bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1392 fail_1:
1393 	bus_dma_tag_destroy(dma->idi_tag);
1394 fail_0:
1395 	dma->idi_tag = NULL;
1396 
1397 	return (err);
1398 }
1399 
1400 int
1401 iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags)
1402 {
1403 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1404 
1405 	KASSERT(sctx->isc_q_align != 0, ("alignment value not initialized"));
1406 
1407 	return (iflib_dma_alloc_align(ctx, size, sctx->isc_q_align, dma, mapflags));
1408 }
1409 
1410 int
1411 iflib_dma_alloc_multi(if_ctx_t ctx, int *sizes, iflib_dma_info_t *dmalist, int mapflags, int count)
1412 {
1413 	int i, err;
1414 	iflib_dma_info_t *dmaiter;
1415 
1416 	dmaiter = dmalist;
1417 	for (i = 0; i < count; i++, dmaiter++) {
1418 		if ((err = iflib_dma_alloc(ctx, sizes[i], *dmaiter, mapflags)) != 0)
1419 			break;
1420 	}
1421 	if (err)
1422 		iflib_dma_free_multi(dmalist, i);
1423 	return (err);
1424 }
1425 
1426 void
1427 iflib_dma_free(iflib_dma_info_t dma)
1428 {
1429 	if (dma->idi_tag == NULL)
1430 		return;
1431 	if (dma->idi_paddr != IF_BAD_DMA) {
1432 		bus_dmamap_sync(dma->idi_tag, dma->idi_map,
1433 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1434 		bus_dmamap_unload(dma->idi_tag, dma->idi_map);
1435 		dma->idi_paddr = IF_BAD_DMA;
1436 	}
1437 	if (dma->idi_vaddr != NULL) {
1438 		bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1439 		dma->idi_vaddr = NULL;
1440 	}
1441 	bus_dma_tag_destroy(dma->idi_tag);
1442 	dma->idi_tag = NULL;
1443 }
1444 
1445 void
1446 iflib_dma_free_multi(iflib_dma_info_t *dmalist, int count)
1447 {
1448 	int i;
1449 	iflib_dma_info_t *dmaiter = dmalist;
1450 
1451 	for (i = 0; i < count; i++, dmaiter++)
1452 		iflib_dma_free(*dmaiter);
1453 }
1454 
1455 static int
1456 iflib_fast_intr(void *arg)
1457 {
1458 	iflib_filter_info_t info = arg;
1459 	struct grouptask *gtask = info->ifi_task;
1460 	int result;
1461 
1462 	DBG_COUNTER_INC(fast_intrs);
1463 	if (info->ifi_filter != NULL) {
1464 		result = info->ifi_filter(info->ifi_filter_arg);
1465 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1466 			return (result);
1467 	}
1468 
1469 	GROUPTASK_ENQUEUE(gtask);
1470 	return (FILTER_HANDLED);
1471 }
1472 
1473 static int
1474 iflib_fast_intr_rxtx(void *arg)
1475 {
1476 	iflib_filter_info_t info = arg;
1477 	struct grouptask *gtask = info->ifi_task;
1478 	if_ctx_t ctx;
1479 	iflib_rxq_t rxq = (iflib_rxq_t)info->ifi_ctx;
1480 	iflib_txq_t txq;
1481 	void *sc;
1482 	int i, cidx, result;
1483 	qidx_t txqid;
1484 	bool intr_enable, intr_legacy;
1485 
1486 	DBG_COUNTER_INC(fast_intrs);
1487 	if (info->ifi_filter != NULL) {
1488 		result = info->ifi_filter(info->ifi_filter_arg);
1489 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1490 			return (result);
1491 	}
1492 
1493 	ctx = rxq->ifr_ctx;
1494 	sc = ctx->ifc_softc;
1495 	intr_enable = false;
1496 	intr_legacy = !!(ctx->ifc_flags & IFC_LEGACY);
1497 	MPASS(rxq->ifr_ntxqirq);
1498 	for (i = 0; i < rxq->ifr_ntxqirq; i++) {
1499 		txqid = rxq->ifr_txqid[i];
1500 		txq = &ctx->ifc_txqs[txqid];
1501 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1502 		    BUS_DMASYNC_POSTREAD);
1503 		if (!ctx->isc_txd_credits_update(sc, txqid, false)) {
1504 			if (intr_legacy)
1505 				intr_enable = true;
1506 			else
1507 				IFDI_TX_QUEUE_INTR_ENABLE(ctx, txqid);
1508 			continue;
1509 		}
1510 		GROUPTASK_ENQUEUE(&txq->ift_task);
1511 	}
1512 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_RXCQ)
1513 		cidx = rxq->ifr_cq_cidx;
1514 	else
1515 		cidx = rxq->ifr_fl[0].ifl_cidx;
1516 	if (iflib_rxd_avail(ctx, rxq, cidx, 1))
1517 		GROUPTASK_ENQUEUE(gtask);
1518 	else {
1519 		if (intr_legacy)
1520 			intr_enable = true;
1521 		else
1522 			IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
1523 		DBG_COUNTER_INC(rx_intr_enables);
1524 	}
1525 	if (intr_enable)
1526 		IFDI_INTR_ENABLE(ctx);
1527 	return (FILTER_HANDLED);
1528 }
1529 
1530 static int
1531 iflib_fast_intr_ctx(void *arg)
1532 {
1533 	iflib_filter_info_t info = arg;
1534 	struct grouptask *gtask = info->ifi_task;
1535 	int result;
1536 
1537 	DBG_COUNTER_INC(fast_intrs);
1538 	if (info->ifi_filter != NULL) {
1539 		result = info->ifi_filter(info->ifi_filter_arg);
1540 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1541 			return (result);
1542 	}
1543 
1544 	GROUPTASK_ENQUEUE(gtask);
1545 	return (FILTER_HANDLED);
1546 }
1547 
1548 static int
1549 _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
1550 		 driver_filter_t filter, driver_intr_t handler, void *arg,
1551 		 const char *name)
1552 {
1553 	struct resource *res;
1554 	void *tag = NULL;
1555 	device_t dev = ctx->ifc_dev;
1556 	int flags, i, rc;
1557 
1558 	flags = RF_ACTIVE;
1559 	if (ctx->ifc_flags & IFC_LEGACY)
1560 		flags |= RF_SHAREABLE;
1561 	MPASS(rid < 512);
1562 	i = rid;
1563 	res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &i, flags);
1564 	if (res == NULL) {
1565 		device_printf(dev,
1566 		    "failed to allocate IRQ for rid %d, name %s.\n", rid, name);
1567 		return (ENOMEM);
1568 	}
1569 	irq->ii_res = res;
1570 	KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL"));
1571 	rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET,
1572 						filter, handler, arg, &tag);
1573 	if (rc != 0) {
1574 		device_printf(dev,
1575 		    "failed to setup interrupt for rid %d, name %s: %d\n",
1576 					  rid, name ? name : "unknown", rc);
1577 		return (rc);
1578 	} else if (name)
1579 		bus_describe_intr(dev, res, tag, "%s", name);
1580 
1581 	irq->ii_tag = tag;
1582 	return (0);
1583 }
1584 
1585 /*********************************************************************
1586  *
1587  *  Allocate DMA resources for TX buffers as well as memory for the TX
1588  *  mbuf map.  TX DMA maps (non-TSO/TSO) and TX mbuf map are kept in a
1589  *  iflib_sw_tx_desc_array structure, storing all the information that
1590  *  is needed to transmit a packet on the wire.  This is called only
1591  *  once at attach, setup is done every reset.
1592  *
1593  **********************************************************************/
1594 static int
1595 iflib_txsd_alloc(iflib_txq_t txq)
1596 {
1597 	if_ctx_t ctx = txq->ift_ctx;
1598 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1599 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1600 	device_t dev = ctx->ifc_dev;
1601 	bus_size_t tsomaxsize;
1602 	int err, nsegments, ntsosegments;
1603 	bool tso;
1604 
1605 	nsegments = scctx->isc_tx_nsegments;
1606 	ntsosegments = scctx->isc_tx_tso_segments_max;
1607 	tsomaxsize = scctx->isc_tx_tso_size_max;
1608 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_VLAN_MTU)
1609 		tsomaxsize += sizeof(struct ether_vlan_header);
1610 	MPASS(scctx->isc_ntxd[0] > 0);
1611 	MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0);
1612 	MPASS(nsegments > 0);
1613 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) {
1614 		MPASS(ntsosegments > 0);
1615 		MPASS(sctx->isc_tso_maxsize >= tsomaxsize);
1616 	}
1617 
1618 	/*
1619 	 * Set up DMA tags for TX buffers.
1620 	 */
1621 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1622 			       1, 0,			/* alignment, bounds */
1623 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1624 			       BUS_SPACE_MAXADDR,	/* highaddr */
1625 			       NULL, NULL,		/* filter, filterarg */
1626 			       sctx->isc_tx_maxsize,		/* maxsize */
1627 			       nsegments,	/* nsegments */
1628 			       sctx->isc_tx_maxsegsize,	/* maxsegsize */
1629 			       0,			/* flags */
1630 			       NULL,			/* lockfunc */
1631 			       NULL,			/* lockfuncarg */
1632 			       &txq->ift_buf_tag))) {
1633 		device_printf(dev,"Unable to allocate TX DMA tag: %d\n", err);
1634 		device_printf(dev,"maxsize: %ju nsegments: %d maxsegsize: %ju\n",
1635 		    (uintmax_t)sctx->isc_tx_maxsize, nsegments, (uintmax_t)sctx->isc_tx_maxsegsize);
1636 		goto fail;
1637 	}
1638 	tso = (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) != 0;
1639 	if (tso && (err = bus_dma_tag_create(bus_get_dma_tag(dev),
1640 			       1, 0,			/* alignment, bounds */
1641 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1642 			       BUS_SPACE_MAXADDR,	/* highaddr */
1643 			       NULL, NULL,		/* filter, filterarg */
1644 			       tsomaxsize,		/* maxsize */
1645 			       ntsosegments,	/* nsegments */
1646 			       sctx->isc_tso_maxsegsize,/* maxsegsize */
1647 			       0,			/* flags */
1648 			       NULL,			/* lockfunc */
1649 			       NULL,			/* lockfuncarg */
1650 			       &txq->ift_tso_buf_tag))) {
1651 		device_printf(dev, "Unable to allocate TSO TX DMA tag: %d\n",
1652 		    err);
1653 		goto fail;
1654 	}
1655 
1656 	/* Allocate memory for the TX mbuf map. */
1657 	if (!(txq->ift_sds.ifsd_m =
1658 	    (struct mbuf **) malloc(sizeof(struct mbuf *) *
1659 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1660 		device_printf(dev, "Unable to allocate TX mbuf map memory\n");
1661 		err = ENOMEM;
1662 		goto fail;
1663 	}
1664 
1665 	/*
1666 	 * Create the DMA maps for TX buffers.
1667 	 */
1668 	if ((txq->ift_sds.ifsd_map = (bus_dmamap_t *)malloc(
1669 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1670 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1671 		device_printf(dev,
1672 		    "Unable to allocate TX buffer DMA map memory\n");
1673 		err = ENOMEM;
1674 		goto fail;
1675 	}
1676 	if (tso && (txq->ift_sds.ifsd_tso_map = (bus_dmamap_t *)malloc(
1677 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1678 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1679 		device_printf(dev,
1680 		    "Unable to allocate TSO TX buffer map memory\n");
1681 		err = ENOMEM;
1682 		goto fail;
1683 	}
1684 	for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) {
1685 		err = bus_dmamap_create(txq->ift_buf_tag, 0,
1686 		    &txq->ift_sds.ifsd_map[i]);
1687 		if (err != 0) {
1688 			device_printf(dev, "Unable to create TX DMA map\n");
1689 			goto fail;
1690 		}
1691 		if (!tso)
1692 			continue;
1693 		err = bus_dmamap_create(txq->ift_tso_buf_tag, 0,
1694 		    &txq->ift_sds.ifsd_tso_map[i]);
1695 		if (err != 0) {
1696 			device_printf(dev, "Unable to create TSO TX DMA map\n");
1697 			goto fail;
1698 		}
1699 	}
1700 	return (0);
1701 fail:
1702 	/* We free all, it handles case where we are in the middle */
1703 	iflib_tx_structures_free(ctx);
1704 	return (err);
1705 }
1706 
1707 static void
1708 iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i)
1709 {
1710 	bus_dmamap_t map;
1711 
1712 	if (txq->ift_sds.ifsd_map != NULL) {
1713 		map = txq->ift_sds.ifsd_map[i];
1714 		bus_dmamap_sync(txq->ift_buf_tag, map, BUS_DMASYNC_POSTWRITE);
1715 		bus_dmamap_unload(txq->ift_buf_tag, map);
1716 		bus_dmamap_destroy(txq->ift_buf_tag, map);
1717 		txq->ift_sds.ifsd_map[i] = NULL;
1718 	}
1719 
1720 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1721 		map = txq->ift_sds.ifsd_tso_map[i];
1722 		bus_dmamap_sync(txq->ift_tso_buf_tag, map,
1723 		    BUS_DMASYNC_POSTWRITE);
1724 		bus_dmamap_unload(txq->ift_tso_buf_tag, map);
1725 		bus_dmamap_destroy(txq->ift_tso_buf_tag, map);
1726 		txq->ift_sds.ifsd_tso_map[i] = NULL;
1727 	}
1728 }
1729 
1730 static void
1731 iflib_txq_destroy(iflib_txq_t txq)
1732 {
1733 	if_ctx_t ctx = txq->ift_ctx;
1734 
1735 	for (int i = 0; i < txq->ift_size; i++)
1736 		iflib_txsd_destroy(ctx, txq, i);
1737 
1738 	if (txq->ift_br != NULL) {
1739 		ifmp_ring_free(txq->ift_br);
1740 		txq->ift_br = NULL;
1741 	}
1742 
1743 	mtx_destroy(&txq->ift_mtx);
1744 
1745 	if (txq->ift_sds.ifsd_map != NULL) {
1746 		free(txq->ift_sds.ifsd_map, M_IFLIB);
1747 		txq->ift_sds.ifsd_map = NULL;
1748 	}
1749 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1750 		free(txq->ift_sds.ifsd_tso_map, M_IFLIB);
1751 		txq->ift_sds.ifsd_tso_map = NULL;
1752 	}
1753 	if (txq->ift_sds.ifsd_m != NULL) {
1754 		free(txq->ift_sds.ifsd_m, M_IFLIB);
1755 		txq->ift_sds.ifsd_m = NULL;
1756 	}
1757 	if (txq->ift_buf_tag != NULL) {
1758 		bus_dma_tag_destroy(txq->ift_buf_tag);
1759 		txq->ift_buf_tag = NULL;
1760 	}
1761 	if (txq->ift_tso_buf_tag != NULL) {
1762 		bus_dma_tag_destroy(txq->ift_tso_buf_tag);
1763 		txq->ift_tso_buf_tag = NULL;
1764 	}
1765 	if (txq->ift_ifdi != NULL) {
1766 		free(txq->ift_ifdi, M_IFLIB);
1767 	}
1768 }
1769 
1770 static void
1771 iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i)
1772 {
1773 	struct mbuf **mp;
1774 
1775 	mp = &txq->ift_sds.ifsd_m[i];
1776 	if (*mp == NULL)
1777 		return;
1778 
1779 	if (txq->ift_sds.ifsd_map != NULL) {
1780 		bus_dmamap_sync(txq->ift_buf_tag,
1781 		    txq->ift_sds.ifsd_map[i], BUS_DMASYNC_POSTWRITE);
1782 		bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[i]);
1783 	}
1784 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1785 		bus_dmamap_sync(txq->ift_tso_buf_tag,
1786 		    txq->ift_sds.ifsd_tso_map[i], BUS_DMASYNC_POSTWRITE);
1787 		bus_dmamap_unload(txq->ift_tso_buf_tag,
1788 		    txq->ift_sds.ifsd_tso_map[i]);
1789 	}
1790 	m_free(*mp);
1791 	DBG_COUNTER_INC(tx_frees);
1792 	*mp = NULL;
1793 }
1794 
1795 static int
1796 iflib_txq_setup(iflib_txq_t txq)
1797 {
1798 	if_ctx_t ctx = txq->ift_ctx;
1799 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1800 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1801 	iflib_dma_info_t di;
1802 	int i;
1803 
1804 	/* Set number of descriptors available */
1805 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
1806 	/* XXX make configurable */
1807 	txq->ift_update_freq = IFLIB_DEFAULT_TX_UPDATE_FREQ;
1808 
1809 	/* Reset indices */
1810 	txq->ift_cidx_processed = 0;
1811 	txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0;
1812 	txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset];
1813 
1814 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1815 		bzero((void *)di->idi_vaddr, di->idi_size);
1816 
1817 	IFDI_TXQ_SETUP(ctx, txq->ift_id);
1818 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1819 		bus_dmamap_sync(di->idi_tag, di->idi_map,
1820 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1821 	return (0);
1822 }
1823 
1824 /*********************************************************************
1825  *
1826  *  Allocate DMA resources for RX buffers as well as memory for the RX
1827  *  mbuf map, direct RX cluster pointer map and RX cluster bus address
1828  *  map.  RX DMA map, RX mbuf map, direct RX cluster pointer map and
1829  *  RX cluster map are kept in a iflib_sw_rx_desc_array structure.
1830  *  Since we use use one entry in iflib_sw_rx_desc_array per received
1831  *  packet, the maximum number of entries we'll need is equal to the
1832  *  number of hardware receive descriptors that we've allocated.
1833  *
1834  **********************************************************************/
1835 static int
1836 iflib_rxsd_alloc(iflib_rxq_t rxq)
1837 {
1838 	if_ctx_t ctx = rxq->ifr_ctx;
1839 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1840 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1841 	device_t dev = ctx->ifc_dev;
1842 	iflib_fl_t fl;
1843 	int			err;
1844 
1845 	MPASS(scctx->isc_nrxd[0] > 0);
1846 	MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0);
1847 
1848 	fl = rxq->ifr_fl;
1849 	for (int i = 0; i <  rxq->ifr_nfl; i++, fl++) {
1850 		fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */
1851 		/* Set up DMA tag for RX buffers. */
1852 		err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1853 					 1, 0,			/* alignment, bounds */
1854 					 BUS_SPACE_MAXADDR,	/* lowaddr */
1855 					 BUS_SPACE_MAXADDR,	/* highaddr */
1856 					 NULL, NULL,		/* filter, filterarg */
1857 					 sctx->isc_rx_maxsize,	/* maxsize */
1858 					 sctx->isc_rx_nsegments,	/* nsegments */
1859 					 sctx->isc_rx_maxsegsize,	/* maxsegsize */
1860 					 0,			/* flags */
1861 					 NULL,			/* lockfunc */
1862 					 NULL,			/* lockarg */
1863 					 &fl->ifl_buf_tag);
1864 		if (err) {
1865 			device_printf(dev,
1866 			    "Unable to allocate RX DMA tag: %d\n", err);
1867 			goto fail;
1868 		}
1869 
1870 		/* Allocate memory for the RX mbuf map. */
1871 		if (!(fl->ifl_sds.ifsd_m =
1872 		      (struct mbuf **) malloc(sizeof(struct mbuf *) *
1873 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1874 			device_printf(dev,
1875 			    "Unable to allocate RX mbuf map memory\n");
1876 			err = ENOMEM;
1877 			goto fail;
1878 		}
1879 
1880 		/* Allocate memory for the direct RX cluster pointer map. */
1881 		if (!(fl->ifl_sds.ifsd_cl =
1882 		      (caddr_t *) malloc(sizeof(caddr_t) *
1883 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1884 			device_printf(dev,
1885 			    "Unable to allocate RX cluster map memory\n");
1886 			err = ENOMEM;
1887 			goto fail;
1888 		}
1889 
1890 		/* Allocate memory for the RX cluster bus address map. */
1891 		if (!(fl->ifl_sds.ifsd_ba =
1892 		      (bus_addr_t *) malloc(sizeof(bus_addr_t) *
1893 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1894 			device_printf(dev,
1895 			    "Unable to allocate RX bus address map memory\n");
1896 			err = ENOMEM;
1897 			goto fail;
1898 		}
1899 
1900 		/*
1901 		 * Create the DMA maps for RX buffers.
1902 		 */
1903 		if (!(fl->ifl_sds.ifsd_map =
1904 		      (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1905 			device_printf(dev,
1906 			    "Unable to allocate RX buffer DMA map memory\n");
1907 			err = ENOMEM;
1908 			goto fail;
1909 		}
1910 		for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++) {
1911 			err = bus_dmamap_create(fl->ifl_buf_tag, 0,
1912 			    &fl->ifl_sds.ifsd_map[i]);
1913 			if (err != 0) {
1914 				device_printf(dev, "Unable to create RX buffer DMA map\n");
1915 				goto fail;
1916 			}
1917 		}
1918 	}
1919 	return (0);
1920 
1921 fail:
1922 	iflib_rx_structures_free(ctx);
1923 	return (err);
1924 }
1925 
1926 /*
1927  * Internal service routines
1928  */
1929 
1930 struct rxq_refill_cb_arg {
1931 	int               error;
1932 	bus_dma_segment_t seg;
1933 	int               nseg;
1934 };
1935 
1936 static void
1937 _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
1938 {
1939 	struct rxq_refill_cb_arg *cb_arg = arg;
1940 
1941 	cb_arg->error = error;
1942 	cb_arg->seg = segs[0];
1943 	cb_arg->nseg = nseg;
1944 }
1945 
1946 /**
1947  * iflib_fl_refill - refill an rxq free-buffer list
1948  * @ctx: the iflib context
1949  * @fl: the free list to refill
1950  * @count: the number of new buffers to allocate
1951  *
1952  * (Re)populate an rxq free-buffer list with up to @count new packet buffers.
1953  * The caller must assure that @count does not exceed the queue's capacity
1954  * minus one (since we always leave a descriptor unavailable).
1955  */
1956 static uint8_t
1957 iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
1958 {
1959 	struct if_rxd_update iru;
1960 	struct rxq_refill_cb_arg cb_arg;
1961 	struct mbuf *m;
1962 	caddr_t cl, *sd_cl;
1963 	struct mbuf **sd_m;
1964 	bus_dmamap_t *sd_map;
1965 	bus_addr_t bus_addr, *sd_ba;
1966 	int err, frag_idx, i, idx, n, pidx;
1967 	qidx_t credits;
1968 
1969 	MPASS(count <= fl->ifl_size - fl->ifl_credits - 1);
1970 
1971 	sd_m = fl->ifl_sds.ifsd_m;
1972 	sd_map = fl->ifl_sds.ifsd_map;
1973 	sd_cl = fl->ifl_sds.ifsd_cl;
1974 	sd_ba = fl->ifl_sds.ifsd_ba;
1975 	pidx = fl->ifl_pidx;
1976 	idx = pidx;
1977 	frag_idx = fl->ifl_fragidx;
1978 	credits = fl->ifl_credits;
1979 
1980 	i = 0;
1981 	n = count;
1982 	MPASS(n > 0);
1983 	MPASS(credits + n <= fl->ifl_size);
1984 
1985 	if (pidx < fl->ifl_cidx)
1986 		MPASS(pidx + n <= fl->ifl_cidx);
1987 	if (pidx == fl->ifl_cidx && (credits < fl->ifl_size))
1988 		MPASS(fl->ifl_gen == 0);
1989 	if (pidx > fl->ifl_cidx)
1990 		MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx);
1991 
1992 	DBG_COUNTER_INC(fl_refills);
1993 	if (n > 8)
1994 		DBG_COUNTER_INC(fl_refills_large);
1995 	iru_init(&iru, fl->ifl_rxq, fl->ifl_id);
1996 	while (n-- > 0) {
1997 		/*
1998 		 * We allocate an uninitialized mbuf + cluster, mbuf is
1999 		 * initialized after rx.
2000 		 *
2001 		 * If the cluster is still set then we know a minimum sized
2002 		 * packet was received
2003 		 */
2004 		bit_ffc_at(fl->ifl_rx_bitmap, frag_idx, fl->ifl_size,
2005 		    &frag_idx);
2006 		if (frag_idx < 0)
2007 			bit_ffc(fl->ifl_rx_bitmap, fl->ifl_size, &frag_idx);
2008 		MPASS(frag_idx >= 0);
2009 		if ((cl = sd_cl[frag_idx]) == NULL) {
2010 			cl = uma_zalloc(fl->ifl_zone, M_NOWAIT);
2011 			if (__predict_false(cl == NULL))
2012 				break;
2013 
2014 			cb_arg.error = 0;
2015 			MPASS(sd_map != NULL);
2016 			err = bus_dmamap_load(fl->ifl_buf_tag, sd_map[frag_idx],
2017 			    cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg,
2018 			    BUS_DMA_NOWAIT);
2019 			if (__predict_false(err != 0 || cb_arg.error)) {
2020 				uma_zfree(fl->ifl_zone, cl);
2021 				break;
2022 			}
2023 
2024 			sd_ba[frag_idx] = bus_addr = cb_arg.seg.ds_addr;
2025 			sd_cl[frag_idx] = cl;
2026 #if MEMORY_LOGGING
2027 			fl->ifl_cl_enqueued++;
2028 #endif
2029 		} else {
2030 			bus_addr = sd_ba[frag_idx];
2031 		}
2032 		bus_dmamap_sync(fl->ifl_buf_tag, sd_map[frag_idx],
2033 		    BUS_DMASYNC_PREREAD);
2034 
2035 		if (sd_m[frag_idx] == NULL) {
2036 			m = m_gethdr(M_NOWAIT, MT_NOINIT);
2037 			if (__predict_false(m == NULL))
2038 				break;
2039 			sd_m[frag_idx] = m;
2040 		}
2041 		bit_set(fl->ifl_rx_bitmap, frag_idx);
2042 #if MEMORY_LOGGING
2043 		fl->ifl_m_enqueued++;
2044 #endif
2045 
2046 		DBG_COUNTER_INC(rx_allocs);
2047 		fl->ifl_rxd_idxs[i] = frag_idx;
2048 		fl->ifl_bus_addrs[i] = bus_addr;
2049 		credits++;
2050 		i++;
2051 		MPASS(credits <= fl->ifl_size);
2052 		if (++idx == fl->ifl_size) {
2053 #ifdef INVARIANTS
2054 			fl->ifl_gen = 1;
2055 #endif
2056 			idx = 0;
2057 		}
2058 		if (n == 0 || i == IFLIB_MAX_RX_REFRESH) {
2059 			iru.iru_pidx = pidx;
2060 			iru.iru_count = i;
2061 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2062 			fl->ifl_pidx = idx;
2063 			fl->ifl_credits = credits;
2064 			pidx = idx;
2065 			i = 0;
2066 		}
2067 	}
2068 
2069 	if (n < count - 1) {
2070 		if (i != 0) {
2071 			iru.iru_pidx = pidx;
2072 			iru.iru_count = i;
2073 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2074 			fl->ifl_pidx = idx;
2075 			fl->ifl_credits = credits;
2076 		}
2077 		DBG_COUNTER_INC(rxd_flush);
2078 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2079 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2080 		ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id,
2081 		    fl->ifl_id, fl->ifl_pidx);
2082 		if (__predict_true(bit_test(fl->ifl_rx_bitmap, frag_idx))) {
2083 			fl->ifl_fragidx = frag_idx + 1;
2084 			if (fl->ifl_fragidx == fl->ifl_size)
2085 				fl->ifl_fragidx = 0;
2086 		} else {
2087 			fl->ifl_fragidx = frag_idx;
2088 		}
2089 	}
2090 
2091 	return (n == -1 ? 0 : IFLIB_RXEOF_EMPTY);
2092 }
2093 
2094 static inline uint8_t
2095 iflib_fl_refill_all(if_ctx_t ctx, iflib_fl_t fl)
2096 {
2097 	/*
2098 	 * We leave an unused descriptor to avoid pidx to catch up with cidx.
2099 	 * This is important as it confuses most NICs. For instance,
2100 	 * Intel NICs have (per receive ring) RDH and RDT registers, where
2101 	 * RDH points to the next receive descriptor to be used by the NIC,
2102 	 * and RDT for the next receive descriptor to be published by the
2103 	 * driver to the NIC (RDT - 1 is thus the last valid one).
2104 	 * The condition RDH == RDT means no descriptors are available to
2105 	 * the NIC, and thus it would be ambiguous if it also meant that
2106 	 * all the descriptors are available to the NIC.
2107 	 */
2108 	int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1;
2109 #ifdef INVARIANTS
2110 	int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1;
2111 #endif
2112 
2113 	MPASS(fl->ifl_credits <= fl->ifl_size);
2114 	MPASS(reclaimable == delta);
2115 
2116 	if (reclaimable > 0)
2117 		return (iflib_fl_refill(ctx, fl, reclaimable));
2118 	return (0);
2119 }
2120 
2121 uint8_t
2122 iflib_in_detach(if_ctx_t ctx)
2123 {
2124 	bool in_detach;
2125 
2126 	STATE_LOCK(ctx);
2127 	in_detach = !!(ctx->ifc_flags & IFC_IN_DETACH);
2128 	STATE_UNLOCK(ctx);
2129 	return (in_detach);
2130 }
2131 
2132 static void
2133 iflib_fl_bufs_free(iflib_fl_t fl)
2134 {
2135 	iflib_dma_info_t idi = fl->ifl_ifdi;
2136 	bus_dmamap_t sd_map;
2137 	uint32_t i;
2138 
2139 	for (i = 0; i < fl->ifl_size; i++) {
2140 		struct mbuf **sd_m = &fl->ifl_sds.ifsd_m[i];
2141 		caddr_t *sd_cl = &fl->ifl_sds.ifsd_cl[i];
2142 
2143 		if (*sd_cl != NULL) {
2144 			sd_map = fl->ifl_sds.ifsd_map[i];
2145 			bus_dmamap_sync(fl->ifl_buf_tag, sd_map,
2146 			    BUS_DMASYNC_POSTREAD);
2147 			bus_dmamap_unload(fl->ifl_buf_tag, sd_map);
2148 			uma_zfree(fl->ifl_zone, *sd_cl);
2149 			*sd_cl = NULL;
2150 			if (*sd_m != NULL) {
2151 				m_init(*sd_m, M_NOWAIT, MT_DATA, 0);
2152 				uma_zfree(zone_mbuf, *sd_m);
2153 				*sd_m = NULL;
2154 			}
2155 		} else {
2156 			MPASS(*sd_m == NULL);
2157 		}
2158 #if MEMORY_LOGGING
2159 		fl->ifl_m_dequeued++;
2160 		fl->ifl_cl_dequeued++;
2161 #endif
2162 	}
2163 #ifdef INVARIANTS
2164 	for (i = 0; i < fl->ifl_size; i++) {
2165 		MPASS(fl->ifl_sds.ifsd_cl[i] == NULL);
2166 		MPASS(fl->ifl_sds.ifsd_m[i] == NULL);
2167 	}
2168 #endif
2169 	/*
2170 	 * Reset free list values
2171 	 */
2172 	fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = fl->ifl_fragidx = 0;
2173 	bzero(idi->idi_vaddr, idi->idi_size);
2174 }
2175 
2176 /*********************************************************************
2177  *
2178  *  Initialize a free list and its buffers.
2179  *
2180  **********************************************************************/
2181 static int
2182 iflib_fl_setup(iflib_fl_t fl)
2183 {
2184 	iflib_rxq_t rxq = fl->ifl_rxq;
2185 	if_ctx_t ctx = rxq->ifr_ctx;
2186 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2187 	int qidx;
2188 
2189 	bit_nclear(fl->ifl_rx_bitmap, 0, fl->ifl_size - 1);
2190 	/*
2191 	** Free current RX buffer structs and their mbufs
2192 	*/
2193 	iflib_fl_bufs_free(fl);
2194 	/* Now replenish the mbufs */
2195 	MPASS(fl->ifl_credits == 0);
2196 	qidx = rxq->ifr_fl_offset + fl->ifl_id;
2197 	if (scctx->isc_rxd_buf_size[qidx] != 0)
2198 		fl->ifl_buf_size = scctx->isc_rxd_buf_size[qidx];
2199 	else
2200 		fl->ifl_buf_size = ctx->ifc_rx_mbuf_sz;
2201 	/*
2202 	 * ifl_buf_size may be a driver-supplied value, so pull it up
2203 	 * to the selected mbuf size.
2204 	 */
2205 	fl->ifl_buf_size = iflib_get_mbuf_size_for(fl->ifl_buf_size);
2206 	if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size)
2207 		ctx->ifc_max_fl_buf_size = fl->ifl_buf_size;
2208 	fl->ifl_cltype = m_gettype(fl->ifl_buf_size);
2209 	fl->ifl_zone = m_getzone(fl->ifl_buf_size);
2210 
2211 	/*
2212 	 * Avoid pre-allocating zillions of clusters to an idle card
2213 	 * potentially speeding up attach. In any case make sure
2214 	 * to leave a descriptor unavailable. See the comment in
2215 	 * iflib_fl_refill_all().
2216 	 */
2217 	MPASS(fl->ifl_size > 0);
2218 	(void)iflib_fl_refill(ctx, fl, min(128, fl->ifl_size - 1));
2219 	if (min(128, fl->ifl_size - 1) != fl->ifl_credits)
2220 		return (ENOBUFS);
2221 	/*
2222 	 * handle failure
2223 	 */
2224 	MPASS(rxq != NULL);
2225 	MPASS(fl->ifl_ifdi != NULL);
2226 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2227 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2228 	return (0);
2229 }
2230 
2231 /*********************************************************************
2232  *
2233  *  Free receive ring data structures
2234  *
2235  **********************************************************************/
2236 static void
2237 iflib_rx_sds_free(iflib_rxq_t rxq)
2238 {
2239 	iflib_fl_t fl;
2240 	int i, j;
2241 
2242 	if (rxq->ifr_fl != NULL) {
2243 		for (i = 0; i < rxq->ifr_nfl; i++) {
2244 			fl = &rxq->ifr_fl[i];
2245 			if (fl->ifl_buf_tag != NULL) {
2246 				if (fl->ifl_sds.ifsd_map != NULL) {
2247 					for (j = 0; j < fl->ifl_size; j++) {
2248 						bus_dmamap_sync(
2249 						    fl->ifl_buf_tag,
2250 						    fl->ifl_sds.ifsd_map[j],
2251 						    BUS_DMASYNC_POSTREAD);
2252 						bus_dmamap_unload(
2253 						    fl->ifl_buf_tag,
2254 						    fl->ifl_sds.ifsd_map[j]);
2255 						bus_dmamap_destroy(
2256 						    fl->ifl_buf_tag,
2257 						    fl->ifl_sds.ifsd_map[j]);
2258 					}
2259 				}
2260 				bus_dma_tag_destroy(fl->ifl_buf_tag);
2261 				fl->ifl_buf_tag = NULL;
2262 			}
2263 			free(fl->ifl_sds.ifsd_m, M_IFLIB);
2264 			free(fl->ifl_sds.ifsd_cl, M_IFLIB);
2265 			free(fl->ifl_sds.ifsd_ba, M_IFLIB);
2266 			free(fl->ifl_sds.ifsd_map, M_IFLIB);
2267 			fl->ifl_sds.ifsd_m = NULL;
2268 			fl->ifl_sds.ifsd_cl = NULL;
2269 			fl->ifl_sds.ifsd_ba = NULL;
2270 			fl->ifl_sds.ifsd_map = NULL;
2271 		}
2272 		free(rxq->ifr_fl, M_IFLIB);
2273 		rxq->ifr_fl = NULL;
2274 		free(rxq->ifr_ifdi, M_IFLIB);
2275 		rxq->ifr_ifdi = NULL;
2276 		rxq->ifr_cq_cidx = 0;
2277 	}
2278 }
2279 
2280 /*
2281  * Timer routine
2282  */
2283 static void
2284 iflib_timer(void *arg)
2285 {
2286 	iflib_txq_t txq = arg;
2287 	if_ctx_t ctx = txq->ift_ctx;
2288 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2289 	uint64_t this_tick = ticks;
2290 	uint32_t reset_on = hz / 2;
2291 
2292 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
2293 		return;
2294 
2295 	/*
2296 	** Check on the state of the TX queue(s), this
2297 	** can be done without the lock because its RO
2298 	** and the HUNG state will be static if set.
2299 	*/
2300 	if (this_tick - txq->ift_last_timer_tick >= hz / 2) {
2301 		txq->ift_last_timer_tick = this_tick;
2302 		IFDI_TIMER(ctx, txq->ift_id);
2303 		if ((txq->ift_qstatus == IFLIB_QUEUE_HUNG) &&
2304 		    ((txq->ift_cleaned_prev == txq->ift_cleaned) ||
2305 		     (sctx->isc_pause_frames == 0)))
2306 			goto hung;
2307 
2308 		if (txq->ift_qstatus != IFLIB_QUEUE_IDLE &&
2309 		    ifmp_ring_is_stalled(txq->ift_br)) {
2310 			KASSERT(ctx->ifc_link_state == LINK_STATE_UP, ("queue can't be marked as hung if interface is down"));
2311 			txq->ift_qstatus = IFLIB_QUEUE_HUNG;
2312 		}
2313 		txq->ift_cleaned_prev = txq->ift_cleaned;
2314 	}
2315 #ifdef DEV_NETMAP
2316 	if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP)
2317 		iflib_netmap_timer_adjust(ctx, txq, &reset_on);
2318 #endif
2319 	/* handle any laggards */
2320 	if (txq->ift_db_pending)
2321 		GROUPTASK_ENQUEUE(&txq->ift_task);
2322 
2323 	sctx->isc_pause_frames = 0;
2324 	if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)
2325 		callout_reset_on(&txq->ift_timer, reset_on, iflib_timer, txq, txq->ift_timer.c_cpu);
2326 	return;
2327 
2328  hung:
2329 	device_printf(ctx->ifc_dev,
2330 	    "Watchdog timeout (TX: %d desc avail: %d pidx: %d) -- resetting\n",
2331 	    txq->ift_id, TXQ_AVAIL(txq), txq->ift_pidx);
2332 	STATE_LOCK(ctx);
2333 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2334 	ctx->ifc_flags |= (IFC_DO_WATCHDOG|IFC_DO_RESET);
2335 	iflib_admin_intr_deferred(ctx);
2336 	STATE_UNLOCK(ctx);
2337 }
2338 
2339 static uint16_t
2340 iflib_get_mbuf_size_for(unsigned int size)
2341 {
2342 
2343 	if (size <= MCLBYTES)
2344 		return (MCLBYTES);
2345 	else
2346 		return (MJUMPAGESIZE);
2347 }
2348 
2349 static void
2350 iflib_calc_rx_mbuf_sz(if_ctx_t ctx)
2351 {
2352 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2353 
2354 	/*
2355 	 * XXX don't set the max_frame_size to larger
2356 	 * than the hardware can handle
2357 	 */
2358 	ctx->ifc_rx_mbuf_sz =
2359 	    iflib_get_mbuf_size_for(sctx->isc_max_frame_size);
2360 }
2361 
2362 uint32_t
2363 iflib_get_rx_mbuf_sz(if_ctx_t ctx)
2364 {
2365 
2366 	return (ctx->ifc_rx_mbuf_sz);
2367 }
2368 
2369 static void
2370 iflib_init_locked(if_ctx_t ctx)
2371 {
2372 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2373 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2374 	if_t ifp = ctx->ifc_ifp;
2375 	iflib_fl_t fl;
2376 	iflib_txq_t txq;
2377 	iflib_rxq_t rxq;
2378 	int i, j, tx_ip_csum_flags, tx_ip6_csum_flags;
2379 
2380 	if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2381 	IFDI_INTR_DISABLE(ctx);
2382 
2383 	tx_ip_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP);
2384 	tx_ip6_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | CSUM_IP6_SCTP);
2385 	/* Set hardware offload abilities */
2386 	if_clearhwassist(ifp);
2387 	if (if_getcapenable(ifp) & IFCAP_TXCSUM)
2388 		if_sethwassistbits(ifp, tx_ip_csum_flags, 0);
2389 	if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6)
2390 		if_sethwassistbits(ifp,  tx_ip6_csum_flags, 0);
2391 	if (if_getcapenable(ifp) & IFCAP_TSO4)
2392 		if_sethwassistbits(ifp, CSUM_IP_TSO, 0);
2393 	if (if_getcapenable(ifp) & IFCAP_TSO6)
2394 		if_sethwassistbits(ifp, CSUM_IP6_TSO, 0);
2395 
2396 	for (i = 0, txq = ctx->ifc_txqs; i < sctx->isc_ntxqsets; i++, txq++) {
2397 		CALLOUT_LOCK(txq);
2398 		callout_stop(&txq->ift_timer);
2399 		CALLOUT_UNLOCK(txq);
2400 		iflib_netmap_txq_init(ctx, txq);
2401 	}
2402 
2403 	/*
2404 	 * Calculate a suitable Rx mbuf size prior to calling IFDI_INIT, so
2405 	 * that drivers can use the value when setting up the hardware receive
2406 	 * buffers.
2407 	 */
2408 	iflib_calc_rx_mbuf_sz(ctx);
2409 
2410 #ifdef INVARIANTS
2411 	i = if_getdrvflags(ifp);
2412 #endif
2413 	IFDI_INIT(ctx);
2414 	MPASS(if_getdrvflags(ifp) == i);
2415 	for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
2416 		if (iflib_netmap_rxq_init(ctx, rxq) > 0) {
2417 			/* This rxq is in netmap mode. Skip normal init. */
2418 			continue;
2419 		}
2420 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
2421 			if (iflib_fl_setup(fl)) {
2422 				device_printf(ctx->ifc_dev,
2423 				    "setting up free list %d failed - "
2424 				    "check cluster settings\n", j);
2425 				goto done;
2426 			}
2427 		}
2428 	}
2429 done:
2430 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
2431 	IFDI_INTR_ENABLE(ctx);
2432 	txq = ctx->ifc_txqs;
2433 	for (i = 0; i < sctx->isc_ntxqsets; i++, txq++)
2434 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq,
2435 			txq->ift_timer.c_cpu);
2436 }
2437 
2438 static int
2439 iflib_media_change(if_t ifp)
2440 {
2441 	if_ctx_t ctx = if_getsoftc(ifp);
2442 	int err;
2443 
2444 	CTX_LOCK(ctx);
2445 	if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0)
2446 		iflib_init_locked(ctx);
2447 	CTX_UNLOCK(ctx);
2448 	return (err);
2449 }
2450 
2451 static void
2452 iflib_media_status(if_t ifp, struct ifmediareq *ifmr)
2453 {
2454 	if_ctx_t ctx = if_getsoftc(ifp);
2455 
2456 	CTX_LOCK(ctx);
2457 	IFDI_UPDATE_ADMIN_STATUS(ctx);
2458 	IFDI_MEDIA_STATUS(ctx, ifmr);
2459 	CTX_UNLOCK(ctx);
2460 }
2461 
2462 void
2463 iflib_stop(if_ctx_t ctx)
2464 {
2465 	iflib_txq_t txq = ctx->ifc_txqs;
2466 	iflib_rxq_t rxq = ctx->ifc_rxqs;
2467 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2468 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2469 	iflib_dma_info_t di;
2470 	iflib_fl_t fl;
2471 	int i, j;
2472 
2473 	/* Tell the stack that the interface is no longer active */
2474 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2475 
2476 	IFDI_INTR_DISABLE(ctx);
2477 	DELAY(1000);
2478 	IFDI_STOP(ctx);
2479 	DELAY(1000);
2480 
2481 	iflib_debug_reset();
2482 	/* Wait for current tx queue users to exit to disarm watchdog timer. */
2483 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) {
2484 		/* make sure all transmitters have completed before proceeding XXX */
2485 
2486 		CALLOUT_LOCK(txq);
2487 		callout_stop(&txq->ift_timer);
2488 		CALLOUT_UNLOCK(txq);
2489 
2490 		/* clean any enqueued buffers */
2491 		iflib_ifmp_purge(txq);
2492 		/* Free any existing tx buffers. */
2493 		for (j = 0; j < txq->ift_size; j++) {
2494 			iflib_txsd_free(ctx, txq, j);
2495 		}
2496 		txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0;
2497 		txq->ift_in_use = txq->ift_gen = txq->ift_cidx = txq->ift_pidx = txq->ift_no_desc_avail = 0;
2498 		txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0;
2499 		txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0;
2500 		txq->ift_pullups = 0;
2501 		ifmp_ring_reset_stats(txq->ift_br);
2502 		for (j = 0, di = txq->ift_ifdi; j < sctx->isc_ntxqs; j++, di++)
2503 			bzero((void *)di->idi_vaddr, di->idi_size);
2504 	}
2505 	for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) {
2506 		/* make sure all transmitters have completed before proceeding XXX */
2507 
2508 		rxq->ifr_cq_cidx = 0;
2509 		for (j = 0, di = rxq->ifr_ifdi; j < sctx->isc_nrxqs; j++, di++)
2510 			bzero((void *)di->idi_vaddr, di->idi_size);
2511 		/* also resets the free lists pidx/cidx */
2512 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
2513 			iflib_fl_bufs_free(fl);
2514 	}
2515 }
2516 
2517 static inline caddr_t
2518 calc_next_rxd(iflib_fl_t fl, int cidx)
2519 {
2520 	qidx_t size;
2521 	int nrxd;
2522 	caddr_t start, end, cur, next;
2523 
2524 	nrxd = fl->ifl_size;
2525 	size = fl->ifl_rxd_size;
2526 	start = fl->ifl_ifdi->idi_vaddr;
2527 
2528 	if (__predict_false(size == 0))
2529 		return (start);
2530 	cur = start + size*cidx;
2531 	end = start + size*nrxd;
2532 	next = CACHE_PTR_NEXT(cur);
2533 	return (next < end ? next : start);
2534 }
2535 
2536 static inline void
2537 prefetch_pkts(iflib_fl_t fl, int cidx)
2538 {
2539 	int nextptr;
2540 	int nrxd = fl->ifl_size;
2541 	caddr_t next_rxd;
2542 
2543 	nextptr = (cidx + CACHE_PTR_INCREMENT) & (nrxd-1);
2544 	prefetch(&fl->ifl_sds.ifsd_m[nextptr]);
2545 	prefetch(&fl->ifl_sds.ifsd_cl[nextptr]);
2546 	next_rxd = calc_next_rxd(fl, cidx);
2547 	prefetch(next_rxd);
2548 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 1) & (nrxd-1)]);
2549 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 2) & (nrxd-1)]);
2550 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 3) & (nrxd-1)]);
2551 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 4) & (nrxd-1)]);
2552 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 1) & (nrxd-1)]);
2553 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 2) & (nrxd-1)]);
2554 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 3) & (nrxd-1)]);
2555 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 4) & (nrxd-1)]);
2556 }
2557 
2558 static struct mbuf *
2559 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, bool unload, if_rxsd_t sd,
2560     int *pf_rv, if_rxd_info_t ri)
2561 {
2562 	bus_dmamap_t map;
2563 	iflib_fl_t fl;
2564 	caddr_t payload;
2565 	struct mbuf *m;
2566 	int flid, cidx, len, next;
2567 
2568 	map = NULL;
2569 	flid = irf->irf_flid;
2570 	cidx = irf->irf_idx;
2571 	fl = &rxq->ifr_fl[flid];
2572 	sd->ifsd_fl = fl;
2573 	m = fl->ifl_sds.ifsd_m[cidx];
2574 	sd->ifsd_cl = &fl->ifl_sds.ifsd_cl[cidx];
2575 	fl->ifl_credits--;
2576 #if MEMORY_LOGGING
2577 	fl->ifl_m_dequeued++;
2578 #endif
2579 	if (rxq->ifr_ctx->ifc_flags & IFC_PREFETCH)
2580 		prefetch_pkts(fl, cidx);
2581 	next = (cidx + CACHE_PTR_INCREMENT) & (fl->ifl_size-1);
2582 	prefetch(&fl->ifl_sds.ifsd_map[next]);
2583 	map = fl->ifl_sds.ifsd_map[cidx];
2584 
2585 	bus_dmamap_sync(fl->ifl_buf_tag, map, BUS_DMASYNC_POSTREAD);
2586 
2587 	if (rxq->pfil != NULL && PFIL_HOOKED_IN(rxq->pfil) && pf_rv != NULL &&
2588 	    irf->irf_len != 0) {
2589 		payload  = *sd->ifsd_cl;
2590 		payload +=  ri->iri_pad;
2591 		len = ri->iri_len - ri->iri_pad;
2592 		*pf_rv = pfil_run_hooks(rxq->pfil, payload, ri->iri_ifp,
2593 		    len | PFIL_MEMPTR | PFIL_IN, NULL);
2594 		switch (*pf_rv) {
2595 		case PFIL_DROPPED:
2596 		case PFIL_CONSUMED:
2597 			/*
2598 			 * The filter ate it.  Everything is recycled.
2599 			 */
2600 			m = NULL;
2601 			unload = 0;
2602 			break;
2603 		case PFIL_REALLOCED:
2604 			/*
2605 			 * The filter copied it.  Everything is recycled.
2606 			 */
2607 			m = pfil_mem2mbuf(payload);
2608 			unload = 0;
2609 			break;
2610 		case PFIL_PASS:
2611 			/*
2612 			 * Filter said it was OK, so receive like
2613 			 * normal
2614 			 */
2615 			fl->ifl_sds.ifsd_m[cidx] = NULL;
2616 			break;
2617 		default:
2618 			MPASS(0);
2619 		}
2620 	} else {
2621 		fl->ifl_sds.ifsd_m[cidx] = NULL;
2622 		*pf_rv = PFIL_PASS;
2623 	}
2624 
2625 	if (unload && irf->irf_len != 0)
2626 		bus_dmamap_unload(fl->ifl_buf_tag, map);
2627 	fl->ifl_cidx = (fl->ifl_cidx + 1) & (fl->ifl_size-1);
2628 	if (__predict_false(fl->ifl_cidx == 0))
2629 		fl->ifl_gen = 0;
2630 	bit_clear(fl->ifl_rx_bitmap, cidx);
2631 	return (m);
2632 }
2633 
2634 static struct mbuf *
2635 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri, if_rxsd_t sd, int *pf_rv)
2636 {
2637 	struct mbuf *m, *mh, *mt;
2638 	caddr_t cl;
2639 	int  *pf_rv_ptr, flags, i, padlen;
2640 	bool consumed;
2641 
2642 	i = 0;
2643 	mh = NULL;
2644 	consumed = false;
2645 	*pf_rv = PFIL_PASS;
2646 	pf_rv_ptr = pf_rv;
2647 	do {
2648 		m = rxd_frag_to_sd(rxq, &ri->iri_frags[i], !consumed, sd,
2649 		    pf_rv_ptr, ri);
2650 
2651 		MPASS(*sd->ifsd_cl != NULL);
2652 
2653 		/*
2654 		 * Exclude zero-length frags & frags from
2655 		 * packets the filter has consumed or dropped
2656 		 */
2657 		if (ri->iri_frags[i].irf_len == 0 || consumed ||
2658 		    *pf_rv == PFIL_CONSUMED || *pf_rv == PFIL_DROPPED) {
2659 			if (mh == NULL) {
2660 				/* everything saved here */
2661 				consumed = true;
2662 				pf_rv_ptr = NULL;
2663 				continue;
2664 			}
2665 			/* XXX we can save the cluster here, but not the mbuf */
2666 			m_init(m, M_NOWAIT, MT_DATA, 0);
2667 			m_free(m);
2668 			continue;
2669 		}
2670 		if (mh == NULL) {
2671 			flags = M_PKTHDR|M_EXT;
2672 			mh = mt = m;
2673 			padlen = ri->iri_pad;
2674 		} else {
2675 			flags = M_EXT;
2676 			mt->m_next = m;
2677 			mt = m;
2678 			/* assuming padding is only on the first fragment */
2679 			padlen = 0;
2680 		}
2681 		cl = *sd->ifsd_cl;
2682 		*sd->ifsd_cl = NULL;
2683 
2684 		/* Can these two be made one ? */
2685 		m_init(m, M_NOWAIT, MT_DATA, flags);
2686 		m_cljset(m, cl, sd->ifsd_fl->ifl_cltype);
2687 		/*
2688 		 * These must follow m_init and m_cljset
2689 		 */
2690 		m->m_data += padlen;
2691 		ri->iri_len -= padlen;
2692 		m->m_len = ri->iri_frags[i].irf_len;
2693 	} while (++i < ri->iri_nfrags);
2694 
2695 	return (mh);
2696 }
2697 
2698 /*
2699  * Process one software descriptor
2700  */
2701 static struct mbuf *
2702 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri)
2703 {
2704 	struct if_rxsd sd;
2705 	struct mbuf *m;
2706 	int pf_rv;
2707 
2708 	/* should I merge this back in now that the two paths are basically duplicated? */
2709 	if (ri->iri_nfrags == 1 &&
2710 	    ri->iri_frags[0].irf_len != 0 &&
2711 	    ri->iri_frags[0].irf_len <= MIN(IFLIB_RX_COPY_THRESH, MHLEN)) {
2712 		m = rxd_frag_to_sd(rxq, &ri->iri_frags[0], false, &sd,
2713 		    &pf_rv, ri);
2714 		if (pf_rv != PFIL_PASS && pf_rv != PFIL_REALLOCED)
2715 			return (m);
2716 		if (pf_rv == PFIL_PASS) {
2717 			m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR);
2718 #ifndef __NO_STRICT_ALIGNMENT
2719 			if (!IP_ALIGNED(m))
2720 				m->m_data += 2;
2721 #endif
2722 			memcpy(m->m_data, *sd.ifsd_cl, ri->iri_len);
2723 			m->m_len = ri->iri_frags[0].irf_len;
2724 		}
2725 	} else {
2726 		m = assemble_segments(rxq, ri, &sd, &pf_rv);
2727 		if (m == NULL)
2728 			return (NULL);
2729 		if (pf_rv != PFIL_PASS && pf_rv != PFIL_REALLOCED)
2730 			return (m);
2731 	}
2732 	m->m_pkthdr.len = ri->iri_len;
2733 	m->m_pkthdr.rcvif = ri->iri_ifp;
2734 	m->m_flags |= ri->iri_flags;
2735 	m->m_pkthdr.ether_vtag = ri->iri_vtag;
2736 	m->m_pkthdr.flowid = ri->iri_flowid;
2737 	M_HASHTYPE_SET(m, ri->iri_rsstype);
2738 	m->m_pkthdr.csum_flags = ri->iri_csum_flags;
2739 	m->m_pkthdr.csum_data = ri->iri_csum_data;
2740 	return (m);
2741 }
2742 
2743 #if defined(INET6) || defined(INET)
2744 static void
2745 iflib_get_ip_forwarding(struct lro_ctrl *lc, bool *v4, bool *v6)
2746 {
2747 	CURVNET_SET(lc->ifp->if_vnet);
2748 #if defined(INET6)
2749 	*v6 = V_ip6_forwarding;
2750 #endif
2751 #if defined(INET)
2752 	*v4 = V_ipforwarding;
2753 #endif
2754 	CURVNET_RESTORE();
2755 }
2756 
2757 /*
2758  * Returns true if it's possible this packet could be LROed.
2759  * if it returns false, it is guaranteed that tcp_lro_rx()
2760  * would not return zero.
2761  */
2762 static bool
2763 iflib_check_lro_possible(struct mbuf *m, bool v4_forwarding, bool v6_forwarding)
2764 {
2765 	struct ether_header *eh;
2766 
2767 	eh = mtod(m, struct ether_header *);
2768 	switch (eh->ether_type) {
2769 #if defined(INET6)
2770 		case htons(ETHERTYPE_IPV6):
2771 			return (!v6_forwarding);
2772 #endif
2773 #if defined (INET)
2774 		case htons(ETHERTYPE_IP):
2775 			return (!v4_forwarding);
2776 #endif
2777 	}
2778 
2779 	return false;
2780 }
2781 #else
2782 static void
2783 iflib_get_ip_forwarding(struct lro_ctrl *lc __unused, bool *v4 __unused, bool *v6 __unused)
2784 {
2785 }
2786 #endif
2787 
2788 static void
2789 _task_fn_rx_watchdog(void *context)
2790 {
2791 	iflib_rxq_t rxq = context;
2792 
2793 	GROUPTASK_ENQUEUE(&rxq->ifr_task);
2794 }
2795 
2796 static uint8_t
2797 iflib_rxeof(iflib_rxq_t rxq, qidx_t budget)
2798 {
2799 	if_t ifp;
2800 	if_ctx_t ctx = rxq->ifr_ctx;
2801 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2802 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2803 	int avail, i;
2804 	qidx_t *cidxp;
2805 	struct if_rxd_info ri;
2806 	int err, budget_left, rx_bytes, rx_pkts;
2807 	iflib_fl_t fl;
2808 	int lro_enabled;
2809 	bool v4_forwarding, v6_forwarding, lro_possible;
2810 	uint8_t retval = 0;
2811 
2812 	/*
2813 	 * XXX early demux data packets so that if_input processing only handles
2814 	 * acks in interrupt context
2815 	 */
2816 	struct mbuf *m, *mh, *mt, *mf;
2817 
2818 	NET_EPOCH_ASSERT();
2819 
2820 	lro_possible = v4_forwarding = v6_forwarding = false;
2821 	ifp = ctx->ifc_ifp;
2822 	mh = mt = NULL;
2823 	MPASS(budget > 0);
2824 	rx_pkts	= rx_bytes = 0;
2825 	if (sctx->isc_flags & IFLIB_HAS_RXCQ)
2826 		cidxp = &rxq->ifr_cq_cidx;
2827 	else
2828 		cidxp = &rxq->ifr_fl[0].ifl_cidx;
2829 	if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) {
2830 		for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2831 			retval |= iflib_fl_refill_all(ctx, fl);
2832 		DBG_COUNTER_INC(rx_unavail);
2833 		return (retval);
2834 	}
2835 
2836 	/* pfil needs the vnet to be set */
2837 	CURVNET_SET_QUIET(ifp->if_vnet);
2838 	for (budget_left = budget; budget_left > 0 && avail > 0;) {
2839 		if (__predict_false(!CTX_ACTIVE(ctx))) {
2840 			DBG_COUNTER_INC(rx_ctx_inactive);
2841 			break;
2842 		}
2843 		/*
2844 		 * Reset client set fields to their default values
2845 		 */
2846 		rxd_info_zero(&ri);
2847 		ri.iri_qsidx = rxq->ifr_id;
2848 		ri.iri_cidx = *cidxp;
2849 		ri.iri_ifp = ifp;
2850 		ri.iri_frags = rxq->ifr_frags;
2851 		err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
2852 
2853 		if (err)
2854 			goto err;
2855 		rx_pkts += 1;
2856 		rx_bytes += ri.iri_len;
2857 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
2858 			*cidxp = ri.iri_cidx;
2859 			/* Update our consumer index */
2860 			/* XXX NB: shurd - check if this is still safe */
2861 			while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0])
2862 				rxq->ifr_cq_cidx -= scctx->isc_nrxd[0];
2863 			/* was this only a completion queue message? */
2864 			if (__predict_false(ri.iri_nfrags == 0))
2865 				continue;
2866 		}
2867 		MPASS(ri.iri_nfrags != 0);
2868 		MPASS(ri.iri_len != 0);
2869 
2870 		/* will advance the cidx on the corresponding free lists */
2871 		m = iflib_rxd_pkt_get(rxq, &ri);
2872 		avail--;
2873 		budget_left--;
2874 		if (avail == 0 && budget_left)
2875 			avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left);
2876 
2877 		if (__predict_false(m == NULL))
2878 			continue;
2879 
2880 		/* imm_pkt: -- cxgb */
2881 		if (mh == NULL)
2882 			mh = mt = m;
2883 		else {
2884 			mt->m_nextpkt = m;
2885 			mt = m;
2886 		}
2887 	}
2888 	CURVNET_RESTORE();
2889 	/* make sure that we can refill faster than drain */
2890 	for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2891 		retval |= iflib_fl_refill_all(ctx, fl);
2892 
2893 	lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO);
2894 	if (lro_enabled)
2895 		iflib_get_ip_forwarding(&rxq->ifr_lc, &v4_forwarding, &v6_forwarding);
2896 	mt = mf = NULL;
2897 	while (mh != NULL) {
2898 		m = mh;
2899 		mh = mh->m_nextpkt;
2900 		m->m_nextpkt = NULL;
2901 #ifndef __NO_STRICT_ALIGNMENT
2902 		if (!IP_ALIGNED(m) && (m = iflib_fixup_rx(m)) == NULL)
2903 			continue;
2904 #endif
2905 		rx_bytes += m->m_pkthdr.len;
2906 		rx_pkts++;
2907 #if defined(INET6) || defined(INET)
2908 		if (lro_enabled) {
2909 			if (!lro_possible) {
2910 				lro_possible = iflib_check_lro_possible(m, v4_forwarding, v6_forwarding);
2911 				if (lro_possible && mf != NULL) {
2912 					ifp->if_input(ifp, mf);
2913 					DBG_COUNTER_INC(rx_if_input);
2914 					mt = mf = NULL;
2915 				}
2916 			}
2917 			if ((m->m_pkthdr.csum_flags & (CSUM_L4_CALC|CSUM_L4_VALID)) ==
2918 			    (CSUM_L4_CALC|CSUM_L4_VALID)) {
2919 				if (lro_possible && tcp_lro_rx(&rxq->ifr_lc, m, 0) == 0)
2920 					continue;
2921 			}
2922 		}
2923 #endif
2924 		if (lro_possible) {
2925 			ifp->if_input(ifp, m);
2926 			DBG_COUNTER_INC(rx_if_input);
2927 			continue;
2928 		}
2929 
2930 		if (mf == NULL)
2931 			mf = m;
2932 		if (mt != NULL)
2933 			mt->m_nextpkt = m;
2934 		mt = m;
2935 	}
2936 	if (mf != NULL) {
2937 		ifp->if_input(ifp, mf);
2938 		DBG_COUNTER_INC(rx_if_input);
2939 	}
2940 
2941 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
2942 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
2943 
2944 	/*
2945 	 * Flush any outstanding LRO work
2946 	 */
2947 #if defined(INET6) || defined(INET)
2948 	tcp_lro_flush_all(&rxq->ifr_lc);
2949 #endif
2950 	if (avail != 0 || iflib_rxd_avail(ctx, rxq, *cidxp, 1) != 0)
2951 		retval |= IFLIB_RXEOF_MORE;
2952 	return (retval);
2953 err:
2954 	STATE_LOCK(ctx);
2955 	ctx->ifc_flags |= IFC_DO_RESET;
2956 	iflib_admin_intr_deferred(ctx);
2957 	STATE_UNLOCK(ctx);
2958 	return (0);
2959 }
2960 
2961 #define TXD_NOTIFY_COUNT(txq) (((txq)->ift_size / (txq)->ift_update_freq)-1)
2962 static inline qidx_t
2963 txq_max_db_deferred(iflib_txq_t txq, qidx_t in_use)
2964 {
2965 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2966 	qidx_t minthresh = txq->ift_size / 8;
2967 	if (in_use > 4*minthresh)
2968 		return (notify_count);
2969 	if (in_use > 2*minthresh)
2970 		return (notify_count >> 1);
2971 	if (in_use > minthresh)
2972 		return (notify_count >> 3);
2973 	return (0);
2974 }
2975 
2976 static inline qidx_t
2977 txq_max_rs_deferred(iflib_txq_t txq)
2978 {
2979 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2980 	qidx_t minthresh = txq->ift_size / 8;
2981 	if (txq->ift_in_use > 4*minthresh)
2982 		return (notify_count);
2983 	if (txq->ift_in_use > 2*minthresh)
2984 		return (notify_count >> 1);
2985 	if (txq->ift_in_use > minthresh)
2986 		return (notify_count >> 2);
2987 	return (2);
2988 }
2989 
2990 #define M_CSUM_FLAGS(m) ((m)->m_pkthdr.csum_flags)
2991 #define M_HAS_VLANTAG(m) (m->m_flags & M_VLANTAG)
2992 
2993 #define TXQ_MAX_DB_DEFERRED(txq, in_use) txq_max_db_deferred((txq), (in_use))
2994 #define TXQ_MAX_RS_DEFERRED(txq) txq_max_rs_deferred(txq)
2995 #define TXQ_MAX_DB_CONSUMED(size) (size >> 4)
2996 
2997 /* forward compatibility for cxgb */
2998 #define FIRST_QSET(ctx) 0
2999 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets)
3000 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets)
3001 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx))
3002 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments))
3003 
3004 /* XXX we should be setting this to something other than zero */
3005 #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh)
3006 #define	MAX_TX_DESC(ctx) max((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max, \
3007     (ctx)->ifc_softc_ctx.isc_tx_nsegments)
3008 
3009 static inline bool
3010 iflib_txd_db_check(if_ctx_t ctx, iflib_txq_t txq, int ring, qidx_t in_use)
3011 {
3012 	qidx_t dbval, max;
3013 	bool rang;
3014 
3015 	rang = false;
3016 	max = TXQ_MAX_DB_DEFERRED(txq, in_use);
3017 	if (ring || txq->ift_db_pending >= max) {
3018 		dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx;
3019 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3020 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
3021 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval);
3022 		txq->ift_db_pending = txq->ift_npending = 0;
3023 		rang = true;
3024 	}
3025 	return (rang);
3026 }
3027 
3028 #ifdef PKT_DEBUG
3029 static void
3030 print_pkt(if_pkt_info_t pi)
3031 {
3032 	printf("pi len:  %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n",
3033 	       pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx);
3034 	printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n",
3035 	       pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag);
3036 	printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n",
3037 	       pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto);
3038 }
3039 #endif
3040 
3041 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO)
3042 #define IS_TX_OFFLOAD4(pi) ((pi)->ipi_csum_flags & (CSUM_IP_TCP | CSUM_IP_TSO))
3043 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO)
3044 #define IS_TX_OFFLOAD6(pi) ((pi)->ipi_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_TSO))
3045 
3046 static int
3047 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp)
3048 {
3049 	if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx;
3050 	struct ether_vlan_header *eh;
3051 	struct mbuf *m;
3052 
3053 	m = *mp;
3054 	if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) &&
3055 	    M_WRITABLE(m) == 0) {
3056 		if ((m = m_dup(m, M_NOWAIT)) == NULL) {
3057 			return (ENOMEM);
3058 		} else {
3059 			m_freem(*mp);
3060 			DBG_COUNTER_INC(tx_frees);
3061 			*mp = m;
3062 		}
3063 	}
3064 
3065 	/*
3066 	 * Determine where frame payload starts.
3067 	 * Jump over vlan headers if already present,
3068 	 * helpful for QinQ too.
3069 	 */
3070 	if (__predict_false(m->m_len < sizeof(*eh))) {
3071 		txq->ift_pullups++;
3072 		if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL))
3073 			return (ENOMEM);
3074 	}
3075 	eh = mtod(m, struct ether_vlan_header *);
3076 	if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
3077 		pi->ipi_etype = ntohs(eh->evl_proto);
3078 		pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
3079 	} else {
3080 		pi->ipi_etype = ntohs(eh->evl_encap_proto);
3081 		pi->ipi_ehdrlen = ETHER_HDR_LEN;
3082 	}
3083 
3084 	switch (pi->ipi_etype) {
3085 #ifdef INET
3086 	case ETHERTYPE_IP:
3087 	{
3088 		struct mbuf *n;
3089 		struct ip *ip = NULL;
3090 		struct tcphdr *th = NULL;
3091 		int minthlen;
3092 
3093 		minthlen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip) + sizeof(*th));
3094 		if (__predict_false(m->m_len < minthlen)) {
3095 			/*
3096 			 * if this code bloat is causing too much of a hit
3097 			 * move it to a separate function and mark it noinline
3098 			 */
3099 			if (m->m_len == pi->ipi_ehdrlen) {
3100 				n = m->m_next;
3101 				MPASS(n);
3102 				if (n->m_len >= sizeof(*ip))  {
3103 					ip = (struct ip *)n->m_data;
3104 					if (n->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3105 						th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
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 				}
3112 			} else {
3113 				txq->ift_pullups++;
3114 				if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
3115 					return (ENOMEM);
3116 				ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3117 				if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3118 					th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3119 			}
3120 		} else {
3121 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3122 			if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3123 				th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3124 		}
3125 		pi->ipi_ip_hlen = ip->ip_hl << 2;
3126 		pi->ipi_ipproto = ip->ip_p;
3127 		pi->ipi_flags |= IPI_TX_IPV4;
3128 
3129 		/* TCP checksum offload may require TCP header length */
3130 		if (IS_TX_OFFLOAD4(pi)) {
3131 			if (__predict_true(pi->ipi_ipproto == IPPROTO_TCP)) {
3132 				if (__predict_false(th == NULL)) {
3133 					txq->ift_pullups++;
3134 					if (__predict_false((m = m_pullup(m, (ip->ip_hl << 2) + sizeof(*th))) == NULL))
3135 						return (ENOMEM);
3136 					th = (struct tcphdr *)((caddr_t)ip + pi->ipi_ip_hlen);
3137 				}
3138 				pi->ipi_tcp_hflags = th->th_flags;
3139 				pi->ipi_tcp_hlen = th->th_off << 2;
3140 				pi->ipi_tcp_seq = th->th_seq;
3141 			}
3142 			if (IS_TSO4(pi)) {
3143 				if (__predict_false(ip->ip_p != IPPROTO_TCP))
3144 					return (ENXIO);
3145 				/*
3146 				 * TSO always requires hardware checksum offload.
3147 				 */
3148 				pi->ipi_csum_flags |= (CSUM_IP_TCP | CSUM_IP);
3149 				th->th_sum = in_pseudo(ip->ip_src.s_addr,
3150 						       ip->ip_dst.s_addr, htons(IPPROTO_TCP));
3151 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3152 				if (sctx->isc_flags & IFLIB_TSO_INIT_IP) {
3153 					ip->ip_sum = 0;
3154 					ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz);
3155 				}
3156 			}
3157 		}
3158 		if ((sctx->isc_flags & IFLIB_NEED_ZERO_CSUM) && (pi->ipi_csum_flags & CSUM_IP))
3159                        ip->ip_sum = 0;
3160 
3161 		break;
3162 	}
3163 #endif
3164 #ifdef INET6
3165 	case ETHERTYPE_IPV6:
3166 	{
3167 		struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
3168 		struct tcphdr *th;
3169 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
3170 
3171 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
3172 			txq->ift_pullups++;
3173 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
3174 				return (ENOMEM);
3175 		}
3176 		th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
3177 
3178 		/* XXX-BZ this will go badly in case of ext hdrs. */
3179 		pi->ipi_ipproto = ip6->ip6_nxt;
3180 		pi->ipi_flags |= IPI_TX_IPV6;
3181 
3182 		/* TCP checksum offload may require TCP header length */
3183 		if (IS_TX_OFFLOAD6(pi)) {
3184 			if (pi->ipi_ipproto == IPPROTO_TCP) {
3185 				if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) {
3186 					txq->ift_pullups++;
3187 					if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL))
3188 						return (ENOMEM);
3189 				}
3190 				pi->ipi_tcp_hflags = th->th_flags;
3191 				pi->ipi_tcp_hlen = th->th_off << 2;
3192 				pi->ipi_tcp_seq = th->th_seq;
3193 			}
3194 			if (IS_TSO6(pi)) {
3195 				if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP))
3196 					return (ENXIO);
3197 				/*
3198 				 * TSO always requires hardware checksum offload.
3199 				 */
3200 				pi->ipi_csum_flags |= CSUM_IP6_TCP;
3201 				th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0);
3202 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3203 			}
3204 		}
3205 		break;
3206 	}
3207 #endif
3208 	default:
3209 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
3210 		pi->ipi_ip_hlen = 0;
3211 		break;
3212 	}
3213 	*mp = m;
3214 
3215 	return (0);
3216 }
3217 
3218 /*
3219  * If dodgy hardware rejects the scatter gather chain we've handed it
3220  * we'll need to remove the mbuf chain from ifsg_m[] before we can add the
3221  * m_defrag'd mbufs
3222  */
3223 static __noinline struct mbuf *
3224 iflib_remove_mbuf(iflib_txq_t txq)
3225 {
3226 	int ntxd, pidx;
3227 	struct mbuf *m, **ifsd_m;
3228 
3229 	ifsd_m = txq->ift_sds.ifsd_m;
3230 	ntxd = txq->ift_size;
3231 	pidx = txq->ift_pidx & (ntxd - 1);
3232 	ifsd_m = txq->ift_sds.ifsd_m;
3233 	m = ifsd_m[pidx];
3234 	ifsd_m[pidx] = NULL;
3235 	bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[pidx]);
3236 	if (txq->ift_sds.ifsd_tso_map != NULL)
3237 		bus_dmamap_unload(txq->ift_tso_buf_tag,
3238 		    txq->ift_sds.ifsd_tso_map[pidx]);
3239 #if MEMORY_LOGGING
3240 	txq->ift_dequeued++;
3241 #endif
3242 	return (m);
3243 }
3244 
3245 static inline caddr_t
3246 calc_next_txd(iflib_txq_t txq, int cidx, uint8_t qid)
3247 {
3248 	qidx_t size;
3249 	int ntxd;
3250 	caddr_t start, end, cur, next;
3251 
3252 	ntxd = txq->ift_size;
3253 	size = txq->ift_txd_size[qid];
3254 	start = txq->ift_ifdi[qid].idi_vaddr;
3255 
3256 	if (__predict_false(size == 0))
3257 		return (start);
3258 	cur = start + size*cidx;
3259 	end = start + size*ntxd;
3260 	next = CACHE_PTR_NEXT(cur);
3261 	return (next < end ? next : start);
3262 }
3263 
3264 /*
3265  * Pad an mbuf to ensure a minimum ethernet frame size.
3266  * min_frame_size is the frame size (less CRC) to pad the mbuf to
3267  */
3268 static __noinline int
3269 iflib_ether_pad(device_t dev, struct mbuf **m_head, uint16_t min_frame_size)
3270 {
3271 	/*
3272 	 * 18 is enough bytes to pad an ARP packet to 46 bytes, and
3273 	 * and ARP message is the smallest common payload I can think of
3274 	 */
3275 	static char pad[18];	/* just zeros */
3276 	int n;
3277 	struct mbuf *new_head;
3278 
3279 	if (!M_WRITABLE(*m_head)) {
3280 		new_head = m_dup(*m_head, M_NOWAIT);
3281 		if (new_head == NULL) {
3282 			m_freem(*m_head);
3283 			device_printf(dev, "cannot pad short frame, m_dup() failed");
3284 			DBG_COUNTER_INC(encap_pad_mbuf_fail);
3285 			DBG_COUNTER_INC(tx_frees);
3286 			return ENOMEM;
3287 		}
3288 		m_freem(*m_head);
3289 		*m_head = new_head;
3290 	}
3291 
3292 	for (n = min_frame_size - (*m_head)->m_pkthdr.len;
3293 	     n > 0; n -= sizeof(pad))
3294 		if (!m_append(*m_head, min(n, sizeof(pad)), pad))
3295 			break;
3296 
3297 	if (n > 0) {
3298 		m_freem(*m_head);
3299 		device_printf(dev, "cannot pad short frame\n");
3300 		DBG_COUNTER_INC(encap_pad_mbuf_fail);
3301 		DBG_COUNTER_INC(tx_frees);
3302 		return (ENOBUFS);
3303 	}
3304 
3305 	return 0;
3306 }
3307 
3308 static int
3309 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp)
3310 {
3311 	if_ctx_t		ctx;
3312 	if_shared_ctx_t		sctx;
3313 	if_softc_ctx_t		scctx;
3314 	bus_dma_tag_t		buf_tag;
3315 	bus_dma_segment_t	*segs;
3316 	struct mbuf		*m_head, **ifsd_m;
3317 	void			*next_txd;
3318 	bus_dmamap_t		map;
3319 	struct if_pkt_info	pi;
3320 	int remap = 0;
3321 	int err, nsegs, ndesc, max_segs, pidx, cidx, next, ntxd;
3322 
3323 	ctx = txq->ift_ctx;
3324 	sctx = ctx->ifc_sctx;
3325 	scctx = &ctx->ifc_softc_ctx;
3326 	segs = txq->ift_segs;
3327 	ntxd = txq->ift_size;
3328 	m_head = *m_headp;
3329 	map = NULL;
3330 
3331 	/*
3332 	 * If we're doing TSO the next descriptor to clean may be quite far ahead
3333 	 */
3334 	cidx = txq->ift_cidx;
3335 	pidx = txq->ift_pidx;
3336 	if (ctx->ifc_flags & IFC_PREFETCH) {
3337 		next = (cidx + CACHE_PTR_INCREMENT) & (ntxd-1);
3338 		if (!(ctx->ifc_flags & IFLIB_HAS_TXCQ)) {
3339 			next_txd = calc_next_txd(txq, cidx, 0);
3340 			prefetch(next_txd);
3341 		}
3342 
3343 		/* prefetch the next cache line of mbuf pointers and flags */
3344 		prefetch(&txq->ift_sds.ifsd_m[next]);
3345 		prefetch(&txq->ift_sds.ifsd_map[next]);
3346 		next = (cidx + CACHE_LINE_SIZE) & (ntxd-1);
3347 	}
3348 	map = txq->ift_sds.ifsd_map[pidx];
3349 	ifsd_m = txq->ift_sds.ifsd_m;
3350 
3351 	if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
3352 		buf_tag = txq->ift_tso_buf_tag;
3353 		max_segs = scctx->isc_tx_tso_segments_max;
3354 		map = txq->ift_sds.ifsd_tso_map[pidx];
3355 		MPASS(buf_tag != NULL);
3356 		MPASS(max_segs > 0);
3357 	} else {
3358 		buf_tag = txq->ift_buf_tag;
3359 		max_segs = scctx->isc_tx_nsegments;
3360 		map = txq->ift_sds.ifsd_map[pidx];
3361 	}
3362 	if ((sctx->isc_flags & IFLIB_NEED_ETHER_PAD) &&
3363 	    __predict_false(m_head->m_pkthdr.len < scctx->isc_min_frame_size)) {
3364 		err = iflib_ether_pad(ctx->ifc_dev, m_headp, scctx->isc_min_frame_size);
3365 		if (err) {
3366 			DBG_COUNTER_INC(encap_txd_encap_fail);
3367 			return err;
3368 		}
3369 	}
3370 	m_head = *m_headp;
3371 
3372 	pkt_info_zero(&pi);
3373 	pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG|M_BCAST|M_MCAST));
3374 	pi.ipi_pidx = pidx;
3375 	pi.ipi_qsidx = txq->ift_id;
3376 	pi.ipi_len = m_head->m_pkthdr.len;
3377 	pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags;
3378 	pi.ipi_vtag = M_HAS_VLANTAG(m_head) ? m_head->m_pkthdr.ether_vtag : 0;
3379 
3380 	/* deliberate bitwise OR to make one condition */
3381 	if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) {
3382 		if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0)) {
3383 			DBG_COUNTER_INC(encap_txd_encap_fail);
3384 			return (err);
3385 		}
3386 		m_head = *m_headp;
3387 	}
3388 
3389 retry:
3390 	err = bus_dmamap_load_mbuf_sg(buf_tag, map, m_head, segs, &nsegs,
3391 	    BUS_DMA_NOWAIT);
3392 defrag:
3393 	if (__predict_false(err)) {
3394 		switch (err) {
3395 		case EFBIG:
3396 			/* try collapse once and defrag once */
3397 			if (remap == 0) {
3398 				m_head = m_collapse(*m_headp, M_NOWAIT, max_segs);
3399 				/* try defrag if collapsing fails */
3400 				if (m_head == NULL)
3401 					remap++;
3402 			}
3403 			if (remap == 1) {
3404 				txq->ift_mbuf_defrag++;
3405 				m_head = m_defrag(*m_headp, M_NOWAIT);
3406 			}
3407 			/*
3408 			 * remap should never be >1 unless bus_dmamap_load_mbuf_sg
3409 			 * failed to map an mbuf that was run through m_defrag
3410 			 */
3411 			MPASS(remap <= 1);
3412 			if (__predict_false(m_head == NULL || remap > 1))
3413 				goto defrag_failed;
3414 			remap++;
3415 			*m_headp = m_head;
3416 			goto retry;
3417 			break;
3418 		case ENOMEM:
3419 			txq->ift_no_tx_dma_setup++;
3420 			break;
3421 		default:
3422 			txq->ift_no_tx_dma_setup++;
3423 			m_freem(*m_headp);
3424 			DBG_COUNTER_INC(tx_frees);
3425 			*m_headp = NULL;
3426 			break;
3427 		}
3428 		txq->ift_map_failed++;
3429 		DBG_COUNTER_INC(encap_load_mbuf_fail);
3430 		DBG_COUNTER_INC(encap_txd_encap_fail);
3431 		return (err);
3432 	}
3433 	ifsd_m[pidx] = m_head;
3434 	/*
3435 	 * XXX assumes a 1 to 1 relationship between segments and
3436 	 *        descriptors - this does not hold true on all drivers, e.g.
3437 	 *        cxgb
3438 	 */
3439 	if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) {
3440 		txq->ift_no_desc_avail++;
3441 		bus_dmamap_unload(buf_tag, map);
3442 		DBG_COUNTER_INC(encap_txq_avail_fail);
3443 		DBG_COUNTER_INC(encap_txd_encap_fail);
3444 		if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0)
3445 			GROUPTASK_ENQUEUE(&txq->ift_task);
3446 		return (ENOBUFS);
3447 	}
3448 	/*
3449 	 * On Intel cards we can greatly reduce the number of TX interrupts
3450 	 * we see by only setting report status on every Nth descriptor.
3451 	 * However, this also means that the driver will need to keep track
3452 	 * of the descriptors that RS was set on to check them for the DD bit.
3453 	 */
3454 	txq->ift_rs_pending += nsegs + 1;
3455 	if (txq->ift_rs_pending > TXQ_MAX_RS_DEFERRED(txq) ||
3456 	     iflib_no_tx_batch || (TXQ_AVAIL(txq) - nsegs) <= MAX_TX_DESC(ctx) + 2) {
3457 		pi.ipi_flags |= IPI_TX_INTR;
3458 		txq->ift_rs_pending = 0;
3459 	}
3460 
3461 	pi.ipi_segs = segs;
3462 	pi.ipi_nsegs = nsegs;
3463 
3464 	MPASS(pidx >= 0 && pidx < txq->ift_size);
3465 #ifdef PKT_DEBUG
3466 	print_pkt(&pi);
3467 #endif
3468 	if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) {
3469 		bus_dmamap_sync(buf_tag, map, BUS_DMASYNC_PREWRITE);
3470 		DBG_COUNTER_INC(tx_encap);
3471 		MPASS(pi.ipi_new_pidx < txq->ift_size);
3472 
3473 		ndesc = pi.ipi_new_pidx - pi.ipi_pidx;
3474 		if (pi.ipi_new_pidx < pi.ipi_pidx) {
3475 			ndesc += txq->ift_size;
3476 			txq->ift_gen = 1;
3477 		}
3478 		/*
3479 		 * drivers can need as many as
3480 		 * two sentinels
3481 		 */
3482 		MPASS(ndesc <= pi.ipi_nsegs + 2);
3483 		MPASS(pi.ipi_new_pidx != pidx);
3484 		MPASS(ndesc > 0);
3485 		txq->ift_in_use += ndesc;
3486 
3487 		/*
3488 		 * We update the last software descriptor again here because there may
3489 		 * be a sentinel and/or there may be more mbufs than segments
3490 		 */
3491 		txq->ift_pidx = pi.ipi_new_pidx;
3492 		txq->ift_npending += pi.ipi_ndescs;
3493 	} else {
3494 		*m_headp = m_head = iflib_remove_mbuf(txq);
3495 		if (err == EFBIG) {
3496 			txq->ift_txd_encap_efbig++;
3497 			if (remap < 2) {
3498 				remap = 1;
3499 				goto defrag;
3500 			}
3501 		}
3502 		goto defrag_failed;
3503 	}
3504 	/*
3505 	 * err can't possibly be non-zero here, so we don't neet to test it
3506 	 * to see if we need to DBG_COUNTER_INC(encap_txd_encap_fail).
3507 	 */
3508 	return (err);
3509 
3510 defrag_failed:
3511 	txq->ift_mbuf_defrag_failed++;
3512 	txq->ift_map_failed++;
3513 	m_freem(*m_headp);
3514 	DBG_COUNTER_INC(tx_frees);
3515 	*m_headp = NULL;
3516 	DBG_COUNTER_INC(encap_txd_encap_fail);
3517 	return (ENOMEM);
3518 }
3519 
3520 static void
3521 iflib_tx_desc_free(iflib_txq_t txq, int n)
3522 {
3523 	uint32_t qsize, cidx, mask, gen;
3524 	struct mbuf *m, **ifsd_m;
3525 	bool do_prefetch;
3526 
3527 	cidx = txq->ift_cidx;
3528 	gen = txq->ift_gen;
3529 	qsize = txq->ift_size;
3530 	mask = qsize-1;
3531 	ifsd_m = txq->ift_sds.ifsd_m;
3532 	do_prefetch = (txq->ift_ctx->ifc_flags & IFC_PREFETCH);
3533 
3534 	while (n-- > 0) {
3535 		if (do_prefetch) {
3536 			prefetch(ifsd_m[(cidx + 3) & mask]);
3537 			prefetch(ifsd_m[(cidx + 4) & mask]);
3538 		}
3539 		if ((m = ifsd_m[cidx]) != NULL) {
3540 			prefetch(&ifsd_m[(cidx + CACHE_PTR_INCREMENT) & mask]);
3541 			if (m->m_pkthdr.csum_flags & CSUM_TSO) {
3542 				bus_dmamap_sync(txq->ift_tso_buf_tag,
3543 				    txq->ift_sds.ifsd_tso_map[cidx],
3544 				    BUS_DMASYNC_POSTWRITE);
3545 				bus_dmamap_unload(txq->ift_tso_buf_tag,
3546 				    txq->ift_sds.ifsd_tso_map[cidx]);
3547 			} else {
3548 				bus_dmamap_sync(txq->ift_buf_tag,
3549 				    txq->ift_sds.ifsd_map[cidx],
3550 				    BUS_DMASYNC_POSTWRITE);
3551 				bus_dmamap_unload(txq->ift_buf_tag,
3552 				    txq->ift_sds.ifsd_map[cidx]);
3553 			}
3554 			/* XXX we don't support any drivers that batch packets yet */
3555 			MPASS(m->m_nextpkt == NULL);
3556 			m_freem(m);
3557 			ifsd_m[cidx] = NULL;
3558 #if MEMORY_LOGGING
3559 			txq->ift_dequeued++;
3560 #endif
3561 			DBG_COUNTER_INC(tx_frees);
3562 		}
3563 		if (__predict_false(++cidx == qsize)) {
3564 			cidx = 0;
3565 			gen = 0;
3566 		}
3567 	}
3568 	txq->ift_cidx = cidx;
3569 	txq->ift_gen = gen;
3570 }
3571 
3572 static __inline int
3573 iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh)
3574 {
3575 	int reclaim;
3576 	if_ctx_t ctx = txq->ift_ctx;
3577 
3578 	KASSERT(thresh >= 0, ("invalid threshold to reclaim"));
3579 	MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size);
3580 
3581 	/*
3582 	 * Need a rate-limiting check so that this isn't called every time
3583 	 */
3584 	iflib_tx_credits_update(ctx, txq);
3585 	reclaim = DESC_RECLAIMABLE(txq);
3586 
3587 	if (reclaim <= thresh /* + MAX_TX_DESC(txq->ift_ctx) */) {
3588 #ifdef INVARIANTS
3589 		if (iflib_verbose_debug) {
3590 			printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __FUNCTION__,
3591 			       txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments,
3592 			       reclaim, thresh);
3593 		}
3594 #endif
3595 		return (0);
3596 	}
3597 	iflib_tx_desc_free(txq, reclaim);
3598 	txq->ift_cleaned += reclaim;
3599 	txq->ift_in_use -= reclaim;
3600 
3601 	return (reclaim);
3602 }
3603 
3604 static struct mbuf **
3605 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset, int remaining)
3606 {
3607 	int next, size;
3608 	struct mbuf **items;
3609 
3610 	size = r->size;
3611 	next = (cidx + CACHE_PTR_INCREMENT) & (size-1);
3612 	items = __DEVOLATILE(struct mbuf **, &r->items[0]);
3613 
3614 	prefetch(items[(cidx + offset) & (size-1)]);
3615 	if (remaining > 1) {
3616 		prefetch2cachelines(&items[next]);
3617 		prefetch2cachelines(items[(cidx + offset + 1) & (size-1)]);
3618 		prefetch2cachelines(items[(cidx + offset + 2) & (size-1)]);
3619 		prefetch2cachelines(items[(cidx + offset + 3) & (size-1)]);
3620 	}
3621 	return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (size-1)]));
3622 }
3623 
3624 static void
3625 iflib_txq_check_drain(iflib_txq_t txq, int budget)
3626 {
3627 
3628 	ifmp_ring_check_drainage(txq->ift_br, budget);
3629 }
3630 
3631 static uint32_t
3632 iflib_txq_can_drain(struct ifmp_ring *r)
3633 {
3634 	iflib_txq_t txq = r->cookie;
3635 	if_ctx_t ctx = txq->ift_ctx;
3636 
3637 	if (TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2)
3638 		return (1);
3639 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3640 	    BUS_DMASYNC_POSTREAD);
3641 	return (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id,
3642 	    false));
3643 }
3644 
3645 static uint32_t
3646 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3647 {
3648 	iflib_txq_t txq = r->cookie;
3649 	if_ctx_t ctx = txq->ift_ctx;
3650 	if_t ifp = ctx->ifc_ifp;
3651 	struct mbuf *m, **mp;
3652 	int avail, bytes_sent, consumed, count, err, i, in_use_prev;
3653 	int mcast_sent, pkt_sent, reclaimed, txq_avail;
3654 	bool do_prefetch, rang, ring;
3655 
3656 	if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) ||
3657 			    !LINK_ACTIVE(ctx))) {
3658 		DBG_COUNTER_INC(txq_drain_notready);
3659 		return (0);
3660 	}
3661 	reclaimed = iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
3662 	rang = iflib_txd_db_check(ctx, txq, reclaimed, txq->ift_in_use);
3663 	avail = IDXDIFF(pidx, cidx, r->size);
3664 	if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) {
3665 		DBG_COUNTER_INC(txq_drain_flushing);
3666 		for (i = 0; i < avail; i++) {
3667 			if (__predict_true(r->items[(cidx + i) & (r->size-1)] != (void *)txq))
3668 				m_free(r->items[(cidx + i) & (r->size-1)]);
3669 			r->items[(cidx + i) & (r->size-1)] = NULL;
3670 		}
3671 		return (avail);
3672 	}
3673 
3674 	if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3675 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3676 		CALLOUT_LOCK(txq);
3677 		callout_stop(&txq->ift_timer);
3678 		CALLOUT_UNLOCK(txq);
3679 		DBG_COUNTER_INC(txq_drain_oactive);
3680 		return (0);
3681 	}
3682 	if (reclaimed)
3683 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3684 	consumed = mcast_sent = bytes_sent = pkt_sent = 0;
3685 	count = MIN(avail, TX_BATCH_SIZE);
3686 #ifdef INVARIANTS
3687 	if (iflib_verbose_debug)
3688 		printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __FUNCTION__,
3689 		       avail, ctx->ifc_flags, TXQ_AVAIL(txq));
3690 #endif
3691 	do_prefetch = (ctx->ifc_flags & IFC_PREFETCH);
3692 	txq_avail = TXQ_AVAIL(txq);
3693 	err = 0;
3694 	for (i = 0; i < count && txq_avail > MAX_TX_DESC(ctx) + 2; i++) {
3695 		int rem = do_prefetch ? count - i : 0;
3696 
3697 		mp = _ring_peek_one(r, cidx, i, rem);
3698 		MPASS(mp != NULL && *mp != NULL);
3699 		if (__predict_false(*mp == (struct mbuf *)txq)) {
3700 			consumed++;
3701 			continue;
3702 		}
3703 		in_use_prev = txq->ift_in_use;
3704 		err = iflib_encap(txq, mp);
3705 		if (__predict_false(err)) {
3706 			/* no room - bail out */
3707 			if (err == ENOBUFS)
3708 				break;
3709 			consumed++;
3710 			/* we can't send this packet - skip it */
3711 			continue;
3712 		}
3713 		consumed++;
3714 		pkt_sent++;
3715 		m = *mp;
3716 		DBG_COUNTER_INC(tx_sent);
3717 		bytes_sent += m->m_pkthdr.len;
3718 		mcast_sent += !!(m->m_flags & M_MCAST);
3719 		txq_avail = TXQ_AVAIL(txq);
3720 
3721 		txq->ift_db_pending += (txq->ift_in_use - in_use_prev);
3722 		ETHER_BPF_MTAP(ifp, m);
3723 		if (__predict_false(!(ifp->if_drv_flags & IFF_DRV_RUNNING)))
3724 			break;
3725 		rang = iflib_txd_db_check(ctx, txq, false, in_use_prev);
3726 	}
3727 
3728 	/* deliberate use of bitwise or to avoid gratuitous short-circuit */
3729 	ring = rang ? false  : (iflib_min_tx_latency | err) || (TXQ_AVAIL(txq) < MAX_TX_DESC(ctx));
3730 	iflib_txd_db_check(ctx, txq, ring, txq->ift_in_use);
3731 	if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
3732 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
3733 	if (mcast_sent)
3734 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
3735 #ifdef INVARIANTS
3736 	if (iflib_verbose_debug)
3737 		printf("consumed=%d\n", consumed);
3738 #endif
3739 	return (consumed);
3740 }
3741 
3742 static uint32_t
3743 iflib_txq_drain_always(struct ifmp_ring *r)
3744 {
3745 	return (1);
3746 }
3747 
3748 static uint32_t
3749 iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3750 {
3751 	int i, avail;
3752 	struct mbuf **mp;
3753 	iflib_txq_t txq;
3754 
3755 	txq = r->cookie;
3756 
3757 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3758 	CALLOUT_LOCK(txq);
3759 	callout_stop(&txq->ift_timer);
3760 	CALLOUT_UNLOCK(txq);
3761 
3762 	avail = IDXDIFF(pidx, cidx, r->size);
3763 	for (i = 0; i < avail; i++) {
3764 		mp = _ring_peek_one(r, cidx, i, avail - i);
3765 		if (__predict_false(*mp == (struct mbuf *)txq))
3766 			continue;
3767 		m_freem(*mp);
3768 		DBG_COUNTER_INC(tx_frees);
3769 	}
3770 	MPASS(ifmp_ring_is_stalled(r) == 0);
3771 	return (avail);
3772 }
3773 
3774 static void
3775 iflib_ifmp_purge(iflib_txq_t txq)
3776 {
3777 	struct ifmp_ring *r;
3778 
3779 	r = txq->ift_br;
3780 	r->drain = iflib_txq_drain_free;
3781 	r->can_drain = iflib_txq_drain_always;
3782 
3783 	ifmp_ring_check_drainage(r, r->size);
3784 
3785 	r->drain = iflib_txq_drain;
3786 	r->can_drain = iflib_txq_can_drain;
3787 }
3788 
3789 static void
3790 _task_fn_tx(void *context)
3791 {
3792 	iflib_txq_t txq = context;
3793 	if_ctx_t ctx = txq->ift_ctx;
3794 	if_t ifp = ctx->ifc_ifp;
3795 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
3796 
3797 #ifdef IFLIB_DIAGNOSTICS
3798 	txq->ift_cpu_exec_count[curcpu]++;
3799 #endif
3800 	if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
3801 		return;
3802 #ifdef DEV_NETMAP
3803 	if ((if_getcapenable(ifp) & IFCAP_NETMAP) &&
3804 	    netmap_tx_irq(ifp, txq->ift_id))
3805 		goto skip_ifmp;
3806 #endif
3807 #ifdef ALTQ
3808 	if (ALTQ_IS_ENABLED(&ifp->if_snd))
3809 		iflib_altq_if_start(ifp);
3810 #endif
3811 	if (txq->ift_db_pending)
3812 		ifmp_ring_enqueue(txq->ift_br, (void **)&txq, 1, TX_BATCH_SIZE, abdicate);
3813 	else if (!abdicate)
3814 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3815 	/*
3816 	 * When abdicating, we always need to check drainage, not just when we don't enqueue
3817 	 */
3818 	if (abdicate)
3819 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3820 #ifdef DEV_NETMAP
3821 skip_ifmp:
3822 #endif
3823 	if (ctx->ifc_flags & IFC_LEGACY)
3824 		IFDI_INTR_ENABLE(ctx);
3825 	else
3826 		IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3827 }
3828 
3829 static void
3830 _task_fn_rx(void *context)
3831 {
3832 	iflib_rxq_t rxq = context;
3833 	if_ctx_t ctx = rxq->ifr_ctx;
3834 	uint8_t more;
3835 	uint16_t budget;
3836 #ifdef DEV_NETMAP
3837 	u_int work = 0;
3838 	int nmirq;
3839 #endif
3840 
3841 #ifdef IFLIB_DIAGNOSTICS
3842 	rxq->ifr_cpu_exec_count[curcpu]++;
3843 #endif
3844 	DBG_COUNTER_INC(task_fn_rxs);
3845 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3846 		return;
3847 #ifdef DEV_NETMAP
3848 	nmirq = netmap_rx_irq(ctx->ifc_ifp, rxq->ifr_id, &work);
3849 	if (nmirq != NM_IRQ_PASS) {
3850 		more = (nmirq == NM_IRQ_RESCHED) ? IFLIB_RXEOF_MORE : 0;
3851 		goto skip_rxeof;
3852 	}
3853 #endif
3854 	budget = ctx->ifc_sysctl_rx_budget;
3855 	if (budget == 0)
3856 		budget = 16;	/* XXX */
3857 	more = iflib_rxeof(rxq, budget);
3858 #ifdef DEV_NETMAP
3859 skip_rxeof:
3860 #endif
3861 	if ((more & IFLIB_RXEOF_MORE) == 0) {
3862 		if (ctx->ifc_flags & IFC_LEGACY)
3863 			IFDI_INTR_ENABLE(ctx);
3864 		else
3865 			IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
3866 		DBG_COUNTER_INC(rx_intr_enables);
3867 	}
3868 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3869 		return;
3870 
3871 	if (more & IFLIB_RXEOF_MORE)
3872 		GROUPTASK_ENQUEUE(&rxq->ifr_task);
3873 	else if (more & IFLIB_RXEOF_EMPTY)
3874 		callout_reset_curcpu(&rxq->ifr_watchdog, 1, &_task_fn_rx_watchdog, rxq);
3875 }
3876 
3877 static void
3878 _task_fn_admin(void *context)
3879 {
3880 	if_ctx_t ctx = context;
3881 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
3882 	iflib_txq_t txq;
3883 	int i;
3884 	bool oactive, running, do_reset, do_watchdog, in_detach;
3885 	uint32_t reset_on = hz / 2;
3886 
3887 	STATE_LOCK(ctx);
3888 	running = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING);
3889 	oactive = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE);
3890 	do_reset = (ctx->ifc_flags & IFC_DO_RESET);
3891 	do_watchdog = (ctx->ifc_flags & IFC_DO_WATCHDOG);
3892 	in_detach = (ctx->ifc_flags & IFC_IN_DETACH);
3893 	ctx->ifc_flags &= ~(IFC_DO_RESET|IFC_DO_WATCHDOG);
3894 	STATE_UNLOCK(ctx);
3895 
3896 	if ((!running && !oactive) && !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
3897 		return;
3898 	if (in_detach)
3899 		return;
3900 
3901 	CTX_LOCK(ctx);
3902 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3903 		CALLOUT_LOCK(txq);
3904 		callout_stop(&txq->ift_timer);
3905 		CALLOUT_UNLOCK(txq);
3906 	}
3907 	if (do_watchdog) {
3908 		ctx->ifc_watchdog_events++;
3909 		IFDI_WATCHDOG_RESET(ctx);
3910 	}
3911 	IFDI_UPDATE_ADMIN_STATUS(ctx);
3912 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3913 #ifdef DEV_NETMAP
3914 		reset_on = hz / 2;
3915 		if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP)
3916 			iflib_netmap_timer_adjust(ctx, txq, &reset_on);
3917 #endif
3918 		callout_reset_on(&txq->ift_timer, reset_on, iflib_timer, txq, txq->ift_timer.c_cpu);
3919 	}
3920 	IFDI_LINK_INTR_ENABLE(ctx);
3921 	if (do_reset)
3922 		iflib_if_init_locked(ctx);
3923 	CTX_UNLOCK(ctx);
3924 
3925 	if (LINK_ACTIVE(ctx) == 0)
3926 		return;
3927 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3928 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
3929 }
3930 
3931 static void
3932 _task_fn_iov(void *context)
3933 {
3934 	if_ctx_t ctx = context;
3935 
3936 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING) &&
3937 	    !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
3938 		return;
3939 
3940 	CTX_LOCK(ctx);
3941 	IFDI_VFLR_HANDLE(ctx);
3942 	CTX_UNLOCK(ctx);
3943 }
3944 
3945 static int
3946 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)
3947 {
3948 	int err;
3949 	if_int_delay_info_t info;
3950 	if_ctx_t ctx;
3951 
3952 	info = (if_int_delay_info_t)arg1;
3953 	ctx = info->iidi_ctx;
3954 	info->iidi_req = req;
3955 	info->iidi_oidp = oidp;
3956 	CTX_LOCK(ctx);
3957 	err = IFDI_SYSCTL_INT_DELAY(ctx, info);
3958 	CTX_UNLOCK(ctx);
3959 	return (err);
3960 }
3961 
3962 /*********************************************************************
3963  *
3964  *  IFNET FUNCTIONS
3965  *
3966  **********************************************************************/
3967 
3968 static void
3969 iflib_if_init_locked(if_ctx_t ctx)
3970 {
3971 	iflib_stop(ctx);
3972 	iflib_init_locked(ctx);
3973 }
3974 
3975 static void
3976 iflib_if_init(void *arg)
3977 {
3978 	if_ctx_t ctx = arg;
3979 
3980 	CTX_LOCK(ctx);
3981 	iflib_if_init_locked(ctx);
3982 	CTX_UNLOCK(ctx);
3983 }
3984 
3985 static int
3986 iflib_if_transmit(if_t ifp, struct mbuf *m)
3987 {
3988 	if_ctx_t	ctx = if_getsoftc(ifp);
3989 
3990 	iflib_txq_t txq;
3991 	int err, qidx;
3992 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
3993 
3994 	if (__predict_false((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) {
3995 		DBG_COUNTER_INC(tx_frees);
3996 		m_freem(m);
3997 		return (ENETDOWN);
3998 	}
3999 
4000 	MPASS(m->m_nextpkt == NULL);
4001 	/* ALTQ-enabled interfaces always use queue 0. */
4002 	qidx = 0;
4003 	if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m) && !ALTQ_IS_ENABLED(&ifp->if_snd))
4004 		qidx = QIDX(ctx, m);
4005 	/*
4006 	 * XXX calculate buf_ring based on flowid (divvy up bits?)
4007 	 */
4008 	txq = &ctx->ifc_txqs[qidx];
4009 
4010 #ifdef DRIVER_BACKPRESSURE
4011 	if (txq->ift_closed) {
4012 		while (m != NULL) {
4013 			next = m->m_nextpkt;
4014 			m->m_nextpkt = NULL;
4015 			m_freem(m);
4016 			DBG_COUNTER_INC(tx_frees);
4017 			m = next;
4018 		}
4019 		return (ENOBUFS);
4020 	}
4021 #endif
4022 #ifdef notyet
4023 	qidx = count = 0;
4024 	mp = marr;
4025 	next = m;
4026 	do {
4027 		count++;
4028 		next = next->m_nextpkt;
4029 	} while (next != NULL);
4030 
4031 	if (count > nitems(marr))
4032 		if ((mp = malloc(count*sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) {
4033 			/* XXX check nextpkt */
4034 			m_freem(m);
4035 			/* XXX simplify for now */
4036 			DBG_COUNTER_INC(tx_frees);
4037 			return (ENOBUFS);
4038 		}
4039 	for (next = m, i = 0; next != NULL; i++) {
4040 		mp[i] = next;
4041 		next = next->m_nextpkt;
4042 		mp[i]->m_nextpkt = NULL;
4043 	}
4044 #endif
4045 	DBG_COUNTER_INC(tx_seen);
4046 	err = ifmp_ring_enqueue(txq->ift_br, (void **)&m, 1, TX_BATCH_SIZE, abdicate);
4047 
4048 	if (abdicate)
4049 		GROUPTASK_ENQUEUE(&txq->ift_task);
4050  	if (err) {
4051 		if (!abdicate)
4052 			GROUPTASK_ENQUEUE(&txq->ift_task);
4053 		/* support forthcoming later */
4054 #ifdef DRIVER_BACKPRESSURE
4055 		txq->ift_closed = TRUE;
4056 #endif
4057 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
4058 		m_freem(m);
4059 		DBG_COUNTER_INC(tx_frees);
4060 	}
4061 
4062 	return (err);
4063 }
4064 
4065 #ifdef ALTQ
4066 /*
4067  * The overall approach to integrating iflib with ALTQ is to continue to use
4068  * the iflib mp_ring machinery between the ALTQ queue(s) and the hardware
4069  * ring.  Technically, when using ALTQ, queueing to an intermediate mp_ring
4070  * is redundant/unnecessary, but doing so minimizes the amount of
4071  * ALTQ-specific code required in iflib.  It is assumed that the overhead of
4072  * redundantly queueing to an intermediate mp_ring is swamped by the
4073  * performance limitations inherent in using ALTQ.
4074  *
4075  * When ALTQ support is compiled in, all iflib drivers will use a transmit
4076  * routine, iflib_altq_if_transmit(), that checks if ALTQ is enabled for the
4077  * given interface.  If ALTQ is enabled for an interface, then all
4078  * transmitted packets for that interface will be submitted to the ALTQ
4079  * subsystem via IFQ_ENQUEUE().  We don't use the legacy if_transmit()
4080  * implementation because it uses IFQ_HANDOFF(), which will duplicatively
4081  * update stats that the iflib machinery handles, and which is sensitve to
4082  * the disused IFF_DRV_OACTIVE flag.  Additionally, iflib_altq_if_start()
4083  * will be installed as the start routine for use by ALTQ facilities that
4084  * need to trigger queue drains on a scheduled basis.
4085  *
4086  */
4087 static void
4088 iflib_altq_if_start(if_t ifp)
4089 {
4090 	struct ifaltq *ifq = &ifp->if_snd;
4091 	struct mbuf *m;
4092 
4093 	IFQ_LOCK(ifq);
4094 	IFQ_DEQUEUE_NOLOCK(ifq, m);
4095 	while (m != NULL) {
4096 		iflib_if_transmit(ifp, m);
4097 		IFQ_DEQUEUE_NOLOCK(ifq, m);
4098 	}
4099 	IFQ_UNLOCK(ifq);
4100 }
4101 
4102 static int
4103 iflib_altq_if_transmit(if_t ifp, struct mbuf *m)
4104 {
4105 	int err;
4106 
4107 	if (ALTQ_IS_ENABLED(&ifp->if_snd)) {
4108 		IFQ_ENQUEUE(&ifp->if_snd, m, err);
4109 		if (err == 0)
4110 			iflib_altq_if_start(ifp);
4111 	} else
4112 		err = iflib_if_transmit(ifp, m);
4113 
4114 	return (err);
4115 }
4116 #endif /* ALTQ */
4117 
4118 static void
4119 iflib_if_qflush(if_t ifp)
4120 {
4121 	if_ctx_t ctx = if_getsoftc(ifp);
4122 	iflib_txq_t txq = ctx->ifc_txqs;
4123 	int i;
4124 
4125 	STATE_LOCK(ctx);
4126 	ctx->ifc_flags |= IFC_QFLUSH;
4127 	STATE_UNLOCK(ctx);
4128 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
4129 		while (!(ifmp_ring_is_idle(txq->ift_br) || ifmp_ring_is_stalled(txq->ift_br)))
4130 			iflib_txq_check_drain(txq, 0);
4131 	STATE_LOCK(ctx);
4132 	ctx->ifc_flags &= ~IFC_QFLUSH;
4133 	STATE_UNLOCK(ctx);
4134 
4135 	/*
4136 	 * When ALTQ is enabled, this will also take care of purging the
4137 	 * ALTQ queue(s).
4138 	 */
4139 	if_qflush(ifp);
4140 }
4141 
4142 #define IFCAP_FLAGS (IFCAP_HWCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \
4143 		     IFCAP_TSO | IFCAP_VLAN_HWTAGGING | IFCAP_HWSTATS | \
4144 		     IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | \
4145 		     IFCAP_VLAN_HWTSO | IFCAP_VLAN_HWCSUM | IFCAP_NOMAP)
4146 
4147 static int
4148 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data)
4149 {
4150 	if_ctx_t ctx = if_getsoftc(ifp);
4151 	struct ifreq	*ifr = (struct ifreq *)data;
4152 #if defined(INET) || defined(INET6)
4153 	struct ifaddr	*ifa = (struct ifaddr *)data;
4154 #endif
4155 	bool		avoid_reset = false;
4156 	int		err = 0, reinit = 0, bits;
4157 
4158 	switch (command) {
4159 	case SIOCSIFADDR:
4160 #ifdef INET
4161 		if (ifa->ifa_addr->sa_family == AF_INET)
4162 			avoid_reset = true;
4163 #endif
4164 #ifdef INET6
4165 		if (ifa->ifa_addr->sa_family == AF_INET6)
4166 			avoid_reset = true;
4167 #endif
4168 		/*
4169 		** Calling init results in link renegotiation,
4170 		** so we avoid doing it when possible.
4171 		*/
4172 		if (avoid_reset) {
4173 			if_setflagbits(ifp, IFF_UP,0);
4174 			if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
4175 				reinit = 1;
4176 #ifdef INET
4177 			if (!(if_getflags(ifp) & IFF_NOARP))
4178 				arp_ifinit(ifp, ifa);
4179 #endif
4180 		} else
4181 			err = ether_ioctl(ifp, command, data);
4182 		break;
4183 	case SIOCSIFMTU:
4184 		CTX_LOCK(ctx);
4185 		if (ifr->ifr_mtu == if_getmtu(ifp)) {
4186 			CTX_UNLOCK(ctx);
4187 			break;
4188 		}
4189 		bits = if_getdrvflags(ifp);
4190 		/* stop the driver and free any clusters before proceeding */
4191 		iflib_stop(ctx);
4192 
4193 		if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) {
4194 			STATE_LOCK(ctx);
4195 			if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size)
4196 				ctx->ifc_flags |= IFC_MULTISEG;
4197 			else
4198 				ctx->ifc_flags &= ~IFC_MULTISEG;
4199 			STATE_UNLOCK(ctx);
4200 			err = if_setmtu(ifp, ifr->ifr_mtu);
4201 		}
4202 		iflib_init_locked(ctx);
4203 		STATE_LOCK(ctx);
4204 		if_setdrvflags(ifp, bits);
4205 		STATE_UNLOCK(ctx);
4206 		CTX_UNLOCK(ctx);
4207 		break;
4208 	case SIOCSIFFLAGS:
4209 		CTX_LOCK(ctx);
4210 		if (if_getflags(ifp) & IFF_UP) {
4211 			if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4212 				if ((if_getflags(ifp) ^ ctx->ifc_if_flags) &
4213 				    (IFF_PROMISC | IFF_ALLMULTI)) {
4214 					CTX_UNLOCK(ctx);
4215 					err = IFDI_PROMISC_SET(ctx, if_getflags(ifp));
4216 					CTX_LOCK(ctx);
4217 				}
4218 			} else
4219 				reinit = 1;
4220 		} else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4221 			iflib_stop(ctx);
4222 		}
4223 		ctx->ifc_if_flags = if_getflags(ifp);
4224 		CTX_UNLOCK(ctx);
4225 		break;
4226 	case SIOCADDMULTI:
4227 	case SIOCDELMULTI:
4228 		if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4229 			CTX_LOCK(ctx);
4230 			IFDI_INTR_DISABLE(ctx);
4231 			IFDI_MULTI_SET(ctx);
4232 			IFDI_INTR_ENABLE(ctx);
4233 			CTX_UNLOCK(ctx);
4234 		}
4235 		break;
4236 	case SIOCSIFMEDIA:
4237 		CTX_LOCK(ctx);
4238 		IFDI_MEDIA_SET(ctx);
4239 		CTX_UNLOCK(ctx);
4240 		/* FALLTHROUGH */
4241 	case SIOCGIFMEDIA:
4242 	case SIOCGIFXMEDIA:
4243 		err = ifmedia_ioctl(ifp, ifr, ctx->ifc_mediap, command);
4244 		break;
4245 	case SIOCGI2C:
4246 	{
4247 		struct ifi2creq i2c;
4248 
4249 		err = copyin(ifr_data_get_ptr(ifr), &i2c, sizeof(i2c));
4250 		if (err != 0)
4251 			break;
4252 		if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) {
4253 			err = EINVAL;
4254 			break;
4255 		}
4256 		if (i2c.len > sizeof(i2c.data)) {
4257 			err = EINVAL;
4258 			break;
4259 		}
4260 
4261 		if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0)
4262 			err = copyout(&i2c, ifr_data_get_ptr(ifr),
4263 			    sizeof(i2c));
4264 		break;
4265 	}
4266 	case SIOCSIFCAP:
4267 	{
4268 		int mask, setmask, oldmask;
4269 
4270 		oldmask = if_getcapenable(ifp);
4271 		mask = ifr->ifr_reqcap ^ oldmask;
4272 		mask &= ctx->ifc_softc_ctx.isc_capabilities | IFCAP_NOMAP;
4273 		setmask = 0;
4274 #ifdef TCP_OFFLOAD
4275 		setmask |= mask & (IFCAP_TOE4|IFCAP_TOE6);
4276 #endif
4277 		setmask |= (mask & IFCAP_FLAGS);
4278 		setmask |= (mask & IFCAP_WOL);
4279 
4280 		/*
4281 		 * If any RX csum has changed, change all the ones that
4282 		 * are supported by the driver.
4283 		 */
4284 		if (setmask & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6)) {
4285 			setmask |= ctx->ifc_softc_ctx.isc_capabilities &
4286 			    (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6);
4287 		}
4288 
4289 		/*
4290 		 * want to ensure that traffic has stopped before we change any of the flags
4291 		 */
4292 		if (setmask) {
4293 			CTX_LOCK(ctx);
4294 			bits = if_getdrvflags(ifp);
4295 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4296 				iflib_stop(ctx);
4297 			STATE_LOCK(ctx);
4298 			if_togglecapenable(ifp, setmask);
4299 			STATE_UNLOCK(ctx);
4300 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4301 				iflib_init_locked(ctx);
4302 			STATE_LOCK(ctx);
4303 			if_setdrvflags(ifp, bits);
4304 			STATE_UNLOCK(ctx);
4305 			CTX_UNLOCK(ctx);
4306 		}
4307 		if_vlancap(ifp);
4308 		break;
4309 	}
4310 	case SIOCGPRIVATE_0:
4311 	case SIOCSDRVSPEC:
4312 	case SIOCGDRVSPEC:
4313 		CTX_LOCK(ctx);
4314 		err = IFDI_PRIV_IOCTL(ctx, command, data);
4315 		CTX_UNLOCK(ctx);
4316 		break;
4317 	default:
4318 		err = ether_ioctl(ifp, command, data);
4319 		break;
4320 	}
4321 	if (reinit)
4322 		iflib_if_init(ctx);
4323 	return (err);
4324 }
4325 
4326 static uint64_t
4327 iflib_if_get_counter(if_t ifp, ift_counter cnt)
4328 {
4329 	if_ctx_t ctx = if_getsoftc(ifp);
4330 
4331 	return (IFDI_GET_COUNTER(ctx, cnt));
4332 }
4333 
4334 /*********************************************************************
4335  *
4336  *  OTHER FUNCTIONS EXPORTED TO THE STACK
4337  *
4338  **********************************************************************/
4339 
4340 static void
4341 iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag)
4342 {
4343 	if_ctx_t ctx = if_getsoftc(ifp);
4344 
4345 	if ((void *)ctx != arg)
4346 		return;
4347 
4348 	if ((vtag == 0) || (vtag > 4095))
4349 		return;
4350 
4351 	if (iflib_in_detach(ctx))
4352 		return;
4353 
4354 	CTX_LOCK(ctx);
4355 	/* Driver may need all untagged packets to be flushed */
4356 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4357 		iflib_stop(ctx);
4358 	IFDI_VLAN_REGISTER(ctx, vtag);
4359 	/* Re-init to load the changes, if required */
4360 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4361 		iflib_init_locked(ctx);
4362 	CTX_UNLOCK(ctx);
4363 }
4364 
4365 static void
4366 iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag)
4367 {
4368 	if_ctx_t ctx = if_getsoftc(ifp);
4369 
4370 	if ((void *)ctx != arg)
4371 		return;
4372 
4373 	if ((vtag == 0) || (vtag > 4095))
4374 		return;
4375 
4376 	CTX_LOCK(ctx);
4377 	/* Driver may need all tagged packets to be flushed */
4378 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4379 		iflib_stop(ctx);
4380 	IFDI_VLAN_UNREGISTER(ctx, vtag);
4381 	/* Re-init to load the changes, if required */
4382 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4383 		iflib_init_locked(ctx);
4384 	CTX_UNLOCK(ctx);
4385 }
4386 
4387 static void
4388 iflib_led_func(void *arg, int onoff)
4389 {
4390 	if_ctx_t ctx = arg;
4391 
4392 	CTX_LOCK(ctx);
4393 	IFDI_LED_FUNC(ctx, onoff);
4394 	CTX_UNLOCK(ctx);
4395 }
4396 
4397 /*********************************************************************
4398  *
4399  *  BUS FUNCTION DEFINITIONS
4400  *
4401  **********************************************************************/
4402 
4403 int
4404 iflib_device_probe(device_t dev)
4405 {
4406 	const pci_vendor_info_t *ent;
4407 	if_shared_ctx_t sctx;
4408 	uint16_t pci_device_id, pci_rev_id, pci_subdevice_id, pci_subvendor_id;
4409 	uint16_t pci_vendor_id;
4410 
4411 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4412 		return (ENOTSUP);
4413 
4414 	pci_vendor_id = pci_get_vendor(dev);
4415 	pci_device_id = pci_get_device(dev);
4416 	pci_subvendor_id = pci_get_subvendor(dev);
4417 	pci_subdevice_id = pci_get_subdevice(dev);
4418 	pci_rev_id = pci_get_revid(dev);
4419 	if (sctx->isc_parse_devinfo != NULL)
4420 		sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id);
4421 
4422 	ent = sctx->isc_vendor_info;
4423 	while (ent->pvi_vendor_id != 0) {
4424 		if (pci_vendor_id != ent->pvi_vendor_id) {
4425 			ent++;
4426 			continue;
4427 		}
4428 		if ((pci_device_id == ent->pvi_device_id) &&
4429 		    ((pci_subvendor_id == ent->pvi_subvendor_id) ||
4430 		     (ent->pvi_subvendor_id == 0)) &&
4431 		    ((pci_subdevice_id == ent->pvi_subdevice_id) ||
4432 		     (ent->pvi_subdevice_id == 0)) &&
4433 		    ((pci_rev_id == ent->pvi_rev_id) ||
4434 		     (ent->pvi_rev_id == 0))) {
4435 			device_set_desc_copy(dev, ent->pvi_name);
4436 			/* this needs to be changed to zero if the bus probing code
4437 			 * ever stops re-probing on best match because the sctx
4438 			 * may have its values over written by register calls
4439 			 * in subsequent probes
4440 			 */
4441 			return (BUS_PROBE_DEFAULT);
4442 		}
4443 		ent++;
4444 	}
4445 	return (ENXIO);
4446 }
4447 
4448 int
4449 iflib_device_probe_vendor(device_t dev)
4450 {
4451 	int probe;
4452 
4453 	probe = iflib_device_probe(dev);
4454 	if (probe == BUS_PROBE_DEFAULT)
4455 		return (BUS_PROBE_VENDOR);
4456 	else
4457 		return (probe);
4458 }
4459 
4460 static void
4461 iflib_reset_qvalues(if_ctx_t ctx)
4462 {
4463 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4464 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4465 	device_t dev = ctx->ifc_dev;
4466 	int i;
4467 
4468 	if (ctx->ifc_sysctl_ntxqs != 0)
4469 		scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs;
4470 	if (ctx->ifc_sysctl_nrxqs != 0)
4471 		scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs;
4472 
4473 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4474 		if (ctx->ifc_sysctl_ntxds[i] != 0)
4475 			scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i];
4476 		else
4477 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4478 	}
4479 
4480 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4481 		if (ctx->ifc_sysctl_nrxds[i] != 0)
4482 			scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i];
4483 		else
4484 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4485 	}
4486 
4487 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4488 		if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) {
4489 			device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n",
4490 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]);
4491 			scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i];
4492 		}
4493 		if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) {
4494 			device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n",
4495 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]);
4496 			scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i];
4497 		}
4498 		if (!powerof2(scctx->isc_nrxd[i])) {
4499 			device_printf(dev, "nrxd%d: %d is not a power of 2 - using default value of %d\n",
4500 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_default[i]);
4501 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4502 		}
4503 	}
4504 
4505 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4506 		if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) {
4507 			device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n",
4508 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]);
4509 			scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i];
4510 		}
4511 		if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) {
4512 			device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n",
4513 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]);
4514 			scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i];
4515 		}
4516 		if (!powerof2(scctx->isc_ntxd[i])) {
4517 			device_printf(dev, "ntxd%d: %d is not a power of 2 - using default value of %d\n",
4518 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_default[i]);
4519 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4520 		}
4521 	}
4522 }
4523 
4524 static void
4525 iflib_add_pfil(if_ctx_t ctx)
4526 {
4527 	struct pfil_head *pfil;
4528 	struct pfil_head_args pa;
4529 	iflib_rxq_t rxq;
4530 	int i;
4531 
4532 	pa.pa_version = PFIL_VERSION;
4533 	pa.pa_flags = PFIL_IN;
4534 	pa.pa_type = PFIL_TYPE_ETHERNET;
4535 	pa.pa_headname = ctx->ifc_ifp->if_xname;
4536 	pfil = pfil_head_register(&pa);
4537 
4538 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4539 		rxq->pfil = pfil;
4540 	}
4541 }
4542 
4543 static void
4544 iflib_rem_pfil(if_ctx_t ctx)
4545 {
4546 	struct pfil_head *pfil;
4547 	iflib_rxq_t rxq;
4548 	int i;
4549 
4550 	rxq = ctx->ifc_rxqs;
4551 	pfil = rxq->pfil;
4552 	for (i = 0; i < NRXQSETS(ctx); i++, rxq++) {
4553 		rxq->pfil = NULL;
4554 	}
4555 	pfil_head_unregister(pfil);
4556 }
4557 
4558 static uint16_t
4559 get_ctx_core_offset(if_ctx_t ctx)
4560 {
4561 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4562 	struct cpu_offset *op;
4563 	uint16_t qc;
4564 	uint16_t ret = ctx->ifc_sysctl_core_offset;
4565 
4566 	if (ret != CORE_OFFSET_UNSPECIFIED)
4567 		return (ret);
4568 
4569 	if (ctx->ifc_sysctl_separate_txrx)
4570 		qc = scctx->isc_ntxqsets + scctx->isc_nrxqsets;
4571 	else
4572 		qc = max(scctx->isc_ntxqsets, scctx->isc_nrxqsets);
4573 
4574 	mtx_lock(&cpu_offset_mtx);
4575 	SLIST_FOREACH(op, &cpu_offsets, entries) {
4576 		if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) {
4577 			ret = op->offset;
4578 			op->offset += qc;
4579 			MPASS(op->refcount < UINT_MAX);
4580 			op->refcount++;
4581 			break;
4582 		}
4583 	}
4584 	if (ret == CORE_OFFSET_UNSPECIFIED) {
4585 		ret = 0;
4586 		op = malloc(sizeof(struct cpu_offset), M_IFLIB,
4587 		    M_NOWAIT | M_ZERO);
4588 		if (op == NULL) {
4589 			device_printf(ctx->ifc_dev,
4590 			    "allocation for cpu offset failed.\n");
4591 		} else {
4592 			op->offset = qc;
4593 			op->refcount = 1;
4594 			CPU_COPY(&ctx->ifc_cpus, &op->set);
4595 			SLIST_INSERT_HEAD(&cpu_offsets, op, entries);
4596 		}
4597 	}
4598 	mtx_unlock(&cpu_offset_mtx);
4599 
4600 	return (ret);
4601 }
4602 
4603 static void
4604 unref_ctx_core_offset(if_ctx_t ctx)
4605 {
4606 	struct cpu_offset *op, *top;
4607 
4608 	mtx_lock(&cpu_offset_mtx);
4609 	SLIST_FOREACH_SAFE(op, &cpu_offsets, entries, top) {
4610 		if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) {
4611 			MPASS(op->refcount > 0);
4612 			op->refcount--;
4613 			if (op->refcount == 0) {
4614 				SLIST_REMOVE(&cpu_offsets, op, cpu_offset, entries);
4615 				free(op, M_IFLIB);
4616 			}
4617 			break;
4618 		}
4619 	}
4620 	mtx_unlock(&cpu_offset_mtx);
4621 }
4622 
4623 int
4624 iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp)
4625 {
4626 	if_ctx_t ctx;
4627 	if_t ifp;
4628 	if_softc_ctx_t scctx;
4629 	kobjop_desc_t kobj_desc;
4630 	kobj_method_t *kobj_method;
4631 	int err, msix, rid;
4632 	int num_txd, num_rxd;
4633 
4634 	ctx = malloc(sizeof(* ctx), M_IFLIB, M_WAITOK|M_ZERO);
4635 
4636 	if (sc == NULL) {
4637 		sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4638 		device_set_softc(dev, ctx);
4639 		ctx->ifc_flags |= IFC_SC_ALLOCATED;
4640 	}
4641 
4642 	ctx->ifc_sctx = sctx;
4643 	ctx->ifc_dev = dev;
4644 	ctx->ifc_softc = sc;
4645 
4646 	if ((err = iflib_register(ctx)) != 0) {
4647 		device_printf(dev, "iflib_register failed %d\n", err);
4648 		goto fail_ctx_free;
4649 	}
4650 	iflib_add_device_sysctl_pre(ctx);
4651 
4652 	scctx = &ctx->ifc_softc_ctx;
4653 	ifp = ctx->ifc_ifp;
4654 
4655 	iflib_reset_qvalues(ctx);
4656 	CTX_LOCK(ctx);
4657 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4658 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4659 		goto fail_unlock;
4660 	}
4661 	_iflib_pre_assert(scctx);
4662 	ctx->ifc_txrx = *scctx->isc_txrx;
4663 
4664 	if (sctx->isc_flags & IFLIB_DRIVER_MEDIA)
4665 		ctx->ifc_mediap = scctx->isc_media;
4666 
4667 #ifdef INVARIANTS
4668 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
4669 		MPASS(scctx->isc_tx_csum_flags);
4670 #endif
4671 
4672 	if_setcapabilities(ifp,
4673 	    scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_NOMAP);
4674 	if_setcapenable(ifp,
4675 	    scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_NOMAP);
4676 
4677 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4678 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4679 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4680 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4681 
4682 	num_txd = iflib_num_tx_descs(ctx);
4683 	num_rxd = iflib_num_rx_descs(ctx);
4684 
4685 	/* XXX change for per-queue sizes */
4686 	device_printf(dev, "Using %d TX descriptors and %d RX descriptors\n",
4687 	    num_txd, num_rxd);
4688 
4689 	if (scctx->isc_tx_nsegments > num_txd / MAX_SINGLE_PACKET_FRACTION)
4690 		scctx->isc_tx_nsegments = max(1, num_txd /
4691 		    MAX_SINGLE_PACKET_FRACTION);
4692 	if (scctx->isc_tx_tso_segments_max > num_txd /
4693 	    MAX_SINGLE_PACKET_FRACTION)
4694 		scctx->isc_tx_tso_segments_max = max(1,
4695 		    num_txd / MAX_SINGLE_PACKET_FRACTION);
4696 
4697 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4698 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
4699 		/*
4700 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
4701 		 * but some MACs do.
4702 		 */
4703 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
4704 		    IP_MAXPACKET));
4705 		/*
4706 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
4707 		 * into account.  In the worst case, each of these calls will
4708 		 * add another mbuf and, thus, the requirement for another DMA
4709 		 * segment.  So for best performance, it doesn't make sense to
4710 		 * advertize a maximum of TSO segments that typically will
4711 		 * require defragmentation in iflib_encap().
4712 		 */
4713 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
4714 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
4715 	}
4716 	if (scctx->isc_rss_table_size == 0)
4717 		scctx->isc_rss_table_size = 64;
4718 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4719 
4720 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4721 	/* XXX format name */
4722 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx,
4723 	    NULL, NULL, "admin");
4724 
4725 	/* Set up cpu set.  If it fails, use the set of all CPUs. */
4726 	if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) != 0) {
4727 		device_printf(dev, "Unable to fetch CPU list\n");
4728 		CPU_COPY(&all_cpus, &ctx->ifc_cpus);
4729 	}
4730 	MPASS(CPU_COUNT(&ctx->ifc_cpus) > 0);
4731 
4732 	/*
4733 	** Now set up MSI or MSI-X, should return us the number of supported
4734 	** vectors (will be 1 for a legacy interrupt and MSI).
4735 	*/
4736 	if (sctx->isc_flags & IFLIB_SKIP_MSIX) {
4737 		msix = scctx->isc_vectors;
4738 	} else if (scctx->isc_msix_bar != 0)
4739 	       /*
4740 		* The simple fact that isc_msix_bar is not 0 does not mean we
4741 		* we have a good value there that is known to work.
4742 		*/
4743 		msix = iflib_msix_init(ctx);
4744 	else {
4745 		scctx->isc_vectors = 1;
4746 		scctx->isc_ntxqsets = 1;
4747 		scctx->isc_nrxqsets = 1;
4748 		scctx->isc_intr = IFLIB_INTR_LEGACY;
4749 		msix = 0;
4750 	}
4751 	/* Get memory for the station queues */
4752 	if ((err = iflib_queues_alloc(ctx))) {
4753 		device_printf(dev, "Unable to allocate queue memory\n");
4754 		goto fail_intr_free;
4755 	}
4756 
4757 	if ((err = iflib_qset_structures_setup(ctx)))
4758 		goto fail_queues;
4759 
4760 	/*
4761 	 * Now that we know how many queues there are, get the core offset.
4762 	 */
4763 	ctx->ifc_sysctl_core_offset = get_ctx_core_offset(ctx);
4764 
4765 	if (msix > 1) {
4766 		/*
4767 		 * When using MSI-X, ensure that ifdi_{r,t}x_queue_intr_enable
4768 		 * aren't the default NULL implementation.
4769 		 */
4770 		kobj_desc = &ifdi_rx_queue_intr_enable_desc;
4771 		kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL,
4772 		    kobj_desc);
4773 		if (kobj_method == &kobj_desc->deflt) {
4774 			device_printf(dev,
4775 			    "MSI-X requires ifdi_rx_queue_intr_enable method");
4776 			err = EOPNOTSUPP;
4777 			goto fail_queues;
4778 		}
4779 		kobj_desc = &ifdi_tx_queue_intr_enable_desc;
4780 		kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL,
4781 		    kobj_desc);
4782 		if (kobj_method == &kobj_desc->deflt) {
4783 			device_printf(dev,
4784 			    "MSI-X requires ifdi_tx_queue_intr_enable method");
4785 			err = EOPNOTSUPP;
4786 			goto fail_queues;
4787 		}
4788 
4789 		/*
4790 		 * Assign the MSI-X vectors.
4791 		 * Note that the default NULL ifdi_msix_intr_assign method will
4792 		 * fail here, too.
4793 		 */
4794 		err = IFDI_MSIX_INTR_ASSIGN(ctx, msix);
4795 		if (err != 0) {
4796 			device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n",
4797 			    err);
4798 			goto fail_queues;
4799 		}
4800 	} else if (scctx->isc_intr != IFLIB_INTR_MSIX) {
4801 		rid = 0;
4802 		if (scctx->isc_intr == IFLIB_INTR_MSI) {
4803 			MPASS(msix == 1);
4804 			rid = 1;
4805 		}
4806 		if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) {
4807 			device_printf(dev, "iflib_legacy_setup failed %d\n", err);
4808 			goto fail_queues;
4809 		}
4810 	} else {
4811 		device_printf(dev,
4812 		    "Cannot use iflib with only 1 MSI-X interrupt!\n");
4813 		err = ENODEV;
4814 		goto fail_intr_free;
4815 	}
4816 
4817 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
4818 
4819 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4820 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4821 		goto fail_detach;
4822 	}
4823 
4824 	/*
4825 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4826 	 * This must appear after the call to ether_ifattach() because
4827 	 * ether_ifattach() sets if_hdrlen to the default value.
4828 	 */
4829 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4830 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
4831 
4832 	if ((err = iflib_netmap_attach(ctx))) {
4833 		device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err);
4834 		goto fail_detach;
4835 	}
4836 	*ctxp = ctx;
4837 
4838 	DEBUGNET_SET(ctx->ifc_ifp, iflib);
4839 
4840 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4841 	iflib_add_device_sysctl_post(ctx);
4842 	iflib_add_pfil(ctx);
4843 	ctx->ifc_flags |= IFC_INIT_DONE;
4844 	CTX_UNLOCK(ctx);
4845 
4846 	return (0);
4847 
4848 fail_detach:
4849 	ether_ifdetach(ctx->ifc_ifp);
4850 fail_intr_free:
4851 	iflib_free_intr_mem(ctx);
4852 fail_queues:
4853 	iflib_tx_structures_free(ctx);
4854 	iflib_rx_structures_free(ctx);
4855 	taskqgroup_detach(qgroup_if_config_tqg, &ctx->ifc_admin_task);
4856 	IFDI_DETACH(ctx);
4857 fail_unlock:
4858 	CTX_UNLOCK(ctx);
4859 	iflib_deregister(ctx);
4860 fail_ctx_free:
4861 	device_set_softc(ctx->ifc_dev, NULL);
4862         if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4863                 free(ctx->ifc_softc, M_IFLIB);
4864         free(ctx, M_IFLIB);
4865 	return (err);
4866 }
4867 
4868 int
4869 iflib_pseudo_register(device_t dev, if_shared_ctx_t sctx, if_ctx_t *ctxp,
4870 					  struct iflib_cloneattach_ctx *clctx)
4871 {
4872 	int num_txd, num_rxd;
4873 	int err;
4874 	if_ctx_t ctx;
4875 	if_t ifp;
4876 	if_softc_ctx_t scctx;
4877 	int i;
4878 	void *sc;
4879 
4880 	ctx = malloc(sizeof(*ctx), M_IFLIB, M_WAITOK|M_ZERO);
4881 	sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4882 	ctx->ifc_flags |= IFC_SC_ALLOCATED;
4883 	if (sctx->isc_flags & (IFLIB_PSEUDO|IFLIB_VIRTUAL))
4884 		ctx->ifc_flags |= IFC_PSEUDO;
4885 
4886 	ctx->ifc_sctx = sctx;
4887 	ctx->ifc_softc = sc;
4888 	ctx->ifc_dev = dev;
4889 
4890 	if ((err = iflib_register(ctx)) != 0) {
4891 		device_printf(dev, "%s: iflib_register failed %d\n", __func__, err);
4892 		goto fail_ctx_free;
4893 	}
4894 	iflib_add_device_sysctl_pre(ctx);
4895 
4896 	scctx = &ctx->ifc_softc_ctx;
4897 	ifp = ctx->ifc_ifp;
4898 
4899 	iflib_reset_qvalues(ctx);
4900 	CTX_LOCK(ctx);
4901 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4902 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4903 		goto fail_unlock;
4904 	}
4905 	if (sctx->isc_flags & IFLIB_GEN_MAC)
4906 		ether_gen_addr(ifp, &ctx->ifc_mac);
4907 	if ((err = IFDI_CLONEATTACH(ctx, clctx->cc_ifc, clctx->cc_name,
4908 								clctx->cc_params)) != 0) {
4909 		device_printf(dev, "IFDI_CLONEATTACH failed %d\n", err);
4910 		goto fail_unlock;
4911 	}
4912 #ifdef INVARIANTS
4913 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
4914 		MPASS(scctx->isc_tx_csum_flags);
4915 #endif
4916 
4917 	if_setcapabilities(ifp, scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_LINKSTATE);
4918 	if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_LINKSTATE);
4919 
4920 	ifp->if_flags |= IFF_NOGROUP;
4921 	if (sctx->isc_flags & IFLIB_PSEUDO) {
4922 		ifmedia_add(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO, 0, NULL);
4923 		ifmedia_set(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO);
4924 		if (sctx->isc_flags & IFLIB_PSEUDO_ETHER) {
4925 			ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
4926 		} else {
4927 			if_attach(ctx->ifc_ifp);
4928 			bpfattach(ctx->ifc_ifp, DLT_NULL, sizeof(u_int32_t));
4929 		}
4930 
4931 		if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4932 			device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4933 			goto fail_detach;
4934 		}
4935 		*ctxp = ctx;
4936 
4937 		/*
4938 		 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4939 		 * This must appear after the call to ether_ifattach() because
4940 		 * ether_ifattach() sets if_hdrlen to the default value.
4941 		 */
4942 		if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4943 			if_setifheaderlen(ifp,
4944 			    sizeof(struct ether_vlan_header));
4945 
4946 		if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4947 		iflib_add_device_sysctl_post(ctx);
4948 		ctx->ifc_flags |= IFC_INIT_DONE;
4949 		CTX_UNLOCK(ctx);
4950 		return (0);
4951 	}
4952 	ifmedia_add(ctx->ifc_mediap, IFM_ETHER | IFM_1000_T | IFM_FDX, 0, NULL);
4953 	ifmedia_add(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO, 0, NULL);
4954 	ifmedia_set(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO);
4955 
4956 	_iflib_pre_assert(scctx);
4957 	ctx->ifc_txrx = *scctx->isc_txrx;
4958 
4959 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4960 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4961 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4962 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4963 
4964 	num_txd = iflib_num_tx_descs(ctx);
4965 	num_rxd = iflib_num_rx_descs(ctx);
4966 
4967 	/* XXX change for per-queue sizes */
4968 	device_printf(dev, "Using %d TX descriptors and %d RX descriptors\n",
4969 	    num_txd, num_rxd);
4970 
4971 	if (scctx->isc_tx_nsegments > num_txd / MAX_SINGLE_PACKET_FRACTION)
4972 		scctx->isc_tx_nsegments = max(1, num_txd /
4973 		    MAX_SINGLE_PACKET_FRACTION);
4974 	if (scctx->isc_tx_tso_segments_max > num_txd /
4975 	    MAX_SINGLE_PACKET_FRACTION)
4976 		scctx->isc_tx_tso_segments_max = max(1,
4977 		    num_txd / MAX_SINGLE_PACKET_FRACTION);
4978 
4979 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4980 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
4981 		/*
4982 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
4983 		 * but some MACs do.
4984 		 */
4985 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
4986 		    IP_MAXPACKET));
4987 		/*
4988 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
4989 		 * into account.  In the worst case, each of these calls will
4990 		 * add another mbuf and, thus, the requirement for another DMA
4991 		 * segment.  So for best performance, it doesn't make sense to
4992 		 * advertize a maximum of TSO segments that typically will
4993 		 * require defragmentation in iflib_encap().
4994 		 */
4995 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
4996 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
4997 	}
4998 	if (scctx->isc_rss_table_size == 0)
4999 		scctx->isc_rss_table_size = 64;
5000 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
5001 
5002 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
5003 	/* XXX format name */
5004 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx,
5005 	    NULL, NULL, "admin");
5006 
5007 	/* XXX --- can support > 1 -- but keep it simple for now */
5008 	scctx->isc_intr = IFLIB_INTR_LEGACY;
5009 
5010 	/* Get memory for the station queues */
5011 	if ((err = iflib_queues_alloc(ctx))) {
5012 		device_printf(dev, "Unable to allocate queue memory\n");
5013 		goto fail_iflib_detach;
5014 	}
5015 
5016 	if ((err = iflib_qset_structures_setup(ctx))) {
5017 		device_printf(dev, "qset structure setup failed %d\n", err);
5018 		goto fail_queues;
5019 	}
5020 
5021 	/*
5022 	 * XXX What if anything do we want to do about interrupts?
5023 	 */
5024 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
5025 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
5026 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
5027 		goto fail_detach;
5028 	}
5029 
5030 	/*
5031 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
5032 	 * This must appear after the call to ether_ifattach() because
5033 	 * ether_ifattach() sets if_hdrlen to the default value.
5034 	 */
5035 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
5036 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
5037 
5038 	/* XXX handle more than one queue */
5039 	for (i = 0; i < scctx->isc_nrxqsets; i++)
5040 		IFDI_RX_CLSET(ctx, 0, i, ctx->ifc_rxqs[i].ifr_fl[0].ifl_sds.ifsd_cl);
5041 
5042 	*ctxp = ctx;
5043 
5044 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
5045 	iflib_add_device_sysctl_post(ctx);
5046 	ctx->ifc_flags |= IFC_INIT_DONE;
5047 	CTX_UNLOCK(ctx);
5048 
5049 	return (0);
5050 fail_detach:
5051 	ether_ifdetach(ctx->ifc_ifp);
5052 fail_queues:
5053 	iflib_tx_structures_free(ctx);
5054 	iflib_rx_structures_free(ctx);
5055 fail_iflib_detach:
5056 	IFDI_DETACH(ctx);
5057 fail_unlock:
5058 	CTX_UNLOCK(ctx);
5059 	iflib_deregister(ctx);
5060 fail_ctx_free:
5061 	free(ctx->ifc_softc, M_IFLIB);
5062 	free(ctx, M_IFLIB);
5063 	return (err);
5064 }
5065 
5066 int
5067 iflib_pseudo_deregister(if_ctx_t ctx)
5068 {
5069 	if_t ifp = ctx->ifc_ifp;
5070 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5071 	iflib_txq_t txq;
5072 	iflib_rxq_t rxq;
5073 	int i, j;
5074 	struct taskqgroup *tqg;
5075 	iflib_fl_t fl;
5076 
5077 	/* Unregister VLAN event handlers early */
5078 	iflib_unregister_vlan_handlers(ctx);
5079 
5080 	if ((sctx->isc_flags & IFLIB_PSEUDO)  &&
5081 		(sctx->isc_flags & IFLIB_PSEUDO_ETHER) == 0) {
5082 		bpfdetach(ifp);
5083 		if_detach(ifp);
5084 	} else {
5085 		ether_ifdetach(ifp);
5086 	}
5087 	/* XXX drain any dependent tasks */
5088 	tqg = qgroup_if_io_tqg;
5089 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
5090 		callout_drain(&txq->ift_timer);
5091 		if (txq->ift_task.gt_uniq != NULL)
5092 			taskqgroup_detach(tqg, &txq->ift_task);
5093 	}
5094 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
5095 		callout_drain(&rxq->ifr_watchdog);
5096 		if (rxq->ifr_task.gt_uniq != NULL)
5097 			taskqgroup_detach(tqg, &rxq->ifr_task);
5098 
5099 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5100 			free(fl->ifl_rx_bitmap, M_IFLIB);
5101 	}
5102 	tqg = qgroup_if_config_tqg;
5103 	if (ctx->ifc_admin_task.gt_uniq != NULL)
5104 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
5105 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
5106 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
5107 
5108 	iflib_tx_structures_free(ctx);
5109 	iflib_rx_structures_free(ctx);
5110 
5111 	iflib_deregister(ctx);
5112 
5113 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5114 		free(ctx->ifc_softc, M_IFLIB);
5115 	free(ctx, M_IFLIB);
5116 	return (0);
5117 }
5118 
5119 int
5120 iflib_device_attach(device_t dev)
5121 {
5122 	if_ctx_t ctx;
5123 	if_shared_ctx_t sctx;
5124 
5125 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
5126 		return (ENOTSUP);
5127 
5128 	pci_enable_busmaster(dev);
5129 
5130 	return (iflib_device_register(dev, NULL, sctx, &ctx));
5131 }
5132 
5133 int
5134 iflib_device_deregister(if_ctx_t ctx)
5135 {
5136 	if_t ifp = ctx->ifc_ifp;
5137 	iflib_txq_t txq;
5138 	iflib_rxq_t rxq;
5139 	device_t dev = ctx->ifc_dev;
5140 	int i, j;
5141 	struct taskqgroup *tqg;
5142 	iflib_fl_t fl;
5143 
5144 	/* Make sure VLANS are not using driver */
5145 	if (if_vlantrunkinuse(ifp)) {
5146 		device_printf(dev, "Vlan in use, detach first\n");
5147 		return (EBUSY);
5148 	}
5149 #ifdef PCI_IOV
5150 	if (!CTX_IS_VF(ctx) && pci_iov_detach(dev) != 0) {
5151 		device_printf(dev, "SR-IOV in use; detach first.\n");
5152 		return (EBUSY);
5153 	}
5154 #endif
5155 
5156 	STATE_LOCK(ctx);
5157 	ctx->ifc_flags |= IFC_IN_DETACH;
5158 	STATE_UNLOCK(ctx);
5159 
5160 	/* Unregister VLAN handlers before calling iflib_stop() */
5161 	iflib_unregister_vlan_handlers(ctx);
5162 
5163 	iflib_netmap_detach(ifp);
5164 	ether_ifdetach(ifp);
5165 
5166 	CTX_LOCK(ctx);
5167 	iflib_stop(ctx);
5168 	CTX_UNLOCK(ctx);
5169 
5170 	iflib_rem_pfil(ctx);
5171 	if (ctx->ifc_led_dev != NULL)
5172 		led_destroy(ctx->ifc_led_dev);
5173 	/* XXX drain any dependent tasks */
5174 	tqg = qgroup_if_io_tqg;
5175 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
5176 		callout_drain(&txq->ift_timer);
5177 		if (txq->ift_task.gt_uniq != NULL)
5178 			taskqgroup_detach(tqg, &txq->ift_task);
5179 	}
5180 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
5181 		if (rxq->ifr_task.gt_uniq != NULL)
5182 			taskqgroup_detach(tqg, &rxq->ifr_task);
5183 
5184 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5185 			free(fl->ifl_rx_bitmap, M_IFLIB);
5186 	}
5187 	tqg = qgroup_if_config_tqg;
5188 	if (ctx->ifc_admin_task.gt_uniq != NULL)
5189 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
5190 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
5191 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
5192 	CTX_LOCK(ctx);
5193 	IFDI_DETACH(ctx);
5194 	CTX_UNLOCK(ctx);
5195 
5196 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
5197 	iflib_free_intr_mem(ctx);
5198 
5199 	bus_generic_detach(dev);
5200 
5201 	iflib_tx_structures_free(ctx);
5202 	iflib_rx_structures_free(ctx);
5203 
5204 	iflib_deregister(ctx);
5205 
5206 	device_set_softc(ctx->ifc_dev, NULL);
5207 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5208 		free(ctx->ifc_softc, M_IFLIB);
5209 	unref_ctx_core_offset(ctx);
5210 	free(ctx, M_IFLIB);
5211 	return (0);
5212 }
5213 
5214 static void
5215 iflib_free_intr_mem(if_ctx_t ctx)
5216 {
5217 
5218 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) {
5219 		iflib_irq_free(ctx, &ctx->ifc_legacy_irq);
5220 	}
5221 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) {
5222 		pci_release_msi(ctx->ifc_dev);
5223 	}
5224 	if (ctx->ifc_msix_mem != NULL) {
5225 		bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY,
5226 		    rman_get_rid(ctx->ifc_msix_mem), ctx->ifc_msix_mem);
5227 		ctx->ifc_msix_mem = NULL;
5228 	}
5229 }
5230 
5231 int
5232 iflib_device_detach(device_t dev)
5233 {
5234 	if_ctx_t ctx = device_get_softc(dev);
5235 
5236 	return (iflib_device_deregister(ctx));
5237 }
5238 
5239 int
5240 iflib_device_suspend(device_t dev)
5241 {
5242 	if_ctx_t ctx = device_get_softc(dev);
5243 
5244 	CTX_LOCK(ctx);
5245 	IFDI_SUSPEND(ctx);
5246 	CTX_UNLOCK(ctx);
5247 
5248 	return bus_generic_suspend(dev);
5249 }
5250 int
5251 iflib_device_shutdown(device_t dev)
5252 {
5253 	if_ctx_t ctx = device_get_softc(dev);
5254 
5255 	CTX_LOCK(ctx);
5256 	IFDI_SHUTDOWN(ctx);
5257 	CTX_UNLOCK(ctx);
5258 
5259 	return bus_generic_suspend(dev);
5260 }
5261 
5262 int
5263 iflib_device_resume(device_t dev)
5264 {
5265 	if_ctx_t ctx = device_get_softc(dev);
5266 	iflib_txq_t txq = ctx->ifc_txqs;
5267 
5268 	CTX_LOCK(ctx);
5269 	IFDI_RESUME(ctx);
5270 	iflib_if_init_locked(ctx);
5271 	CTX_UNLOCK(ctx);
5272 	for (int i = 0; i < NTXQSETS(ctx); i++, txq++)
5273 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
5274 
5275 	return (bus_generic_resume(dev));
5276 }
5277 
5278 int
5279 iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params)
5280 {
5281 	int error;
5282 	if_ctx_t ctx = device_get_softc(dev);
5283 
5284 	CTX_LOCK(ctx);
5285 	error = IFDI_IOV_INIT(ctx, num_vfs, params);
5286 	CTX_UNLOCK(ctx);
5287 
5288 	return (error);
5289 }
5290 
5291 void
5292 iflib_device_iov_uninit(device_t dev)
5293 {
5294 	if_ctx_t ctx = device_get_softc(dev);
5295 
5296 	CTX_LOCK(ctx);
5297 	IFDI_IOV_UNINIT(ctx);
5298 	CTX_UNLOCK(ctx);
5299 }
5300 
5301 int
5302 iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params)
5303 {
5304 	int error;
5305 	if_ctx_t ctx = device_get_softc(dev);
5306 
5307 	CTX_LOCK(ctx);
5308 	error = IFDI_IOV_VF_ADD(ctx, vfnum, params);
5309 	CTX_UNLOCK(ctx);
5310 
5311 	return (error);
5312 }
5313 
5314 /*********************************************************************
5315  *
5316  *  MODULE FUNCTION DEFINITIONS
5317  *
5318  **********************************************************************/
5319 
5320 /*
5321  * - Start a fast taskqueue thread for each core
5322  * - Start a taskqueue for control operations
5323  */
5324 static int
5325 iflib_module_init(void)
5326 {
5327 	return (0);
5328 }
5329 
5330 static int
5331 iflib_module_event_handler(module_t mod, int what, void *arg)
5332 {
5333 	int err;
5334 
5335 	switch (what) {
5336 	case MOD_LOAD:
5337 		if ((err = iflib_module_init()) != 0)
5338 			return (err);
5339 		break;
5340 	case MOD_UNLOAD:
5341 		return (EBUSY);
5342 	default:
5343 		return (EOPNOTSUPP);
5344 	}
5345 
5346 	return (0);
5347 }
5348 
5349 /*********************************************************************
5350  *
5351  *  PUBLIC FUNCTION DEFINITIONS
5352  *     ordered as in iflib.h
5353  *
5354  **********************************************************************/
5355 
5356 static void
5357 _iflib_assert(if_shared_ctx_t sctx)
5358 {
5359 	int i;
5360 
5361 	MPASS(sctx->isc_tx_maxsize);
5362 	MPASS(sctx->isc_tx_maxsegsize);
5363 
5364 	MPASS(sctx->isc_rx_maxsize);
5365 	MPASS(sctx->isc_rx_nsegments);
5366 	MPASS(sctx->isc_rx_maxsegsize);
5367 
5368 	MPASS(sctx->isc_nrxqs >= 1 && sctx->isc_nrxqs <= 8);
5369 	for (i = 0; i < sctx->isc_nrxqs; i++) {
5370 		MPASS(sctx->isc_nrxd_min[i]);
5371 		MPASS(powerof2(sctx->isc_nrxd_min[i]));
5372 		MPASS(sctx->isc_nrxd_max[i]);
5373 		MPASS(powerof2(sctx->isc_nrxd_max[i]));
5374 		MPASS(sctx->isc_nrxd_default[i]);
5375 		MPASS(powerof2(sctx->isc_nrxd_default[i]));
5376 	}
5377 
5378 	MPASS(sctx->isc_ntxqs >= 1 && sctx->isc_ntxqs <= 8);
5379 	for (i = 0; i < sctx->isc_ntxqs; i++) {
5380 		MPASS(sctx->isc_ntxd_min[i]);
5381 		MPASS(powerof2(sctx->isc_ntxd_min[i]));
5382 		MPASS(sctx->isc_ntxd_max[i]);
5383 		MPASS(powerof2(sctx->isc_ntxd_max[i]));
5384 		MPASS(sctx->isc_ntxd_default[i]);
5385 		MPASS(powerof2(sctx->isc_ntxd_default[i]));
5386 	}
5387 }
5388 
5389 static void
5390 _iflib_pre_assert(if_softc_ctx_t scctx)
5391 {
5392 
5393 	MPASS(scctx->isc_txrx->ift_txd_encap);
5394 	MPASS(scctx->isc_txrx->ift_txd_flush);
5395 	MPASS(scctx->isc_txrx->ift_txd_credits_update);
5396 	MPASS(scctx->isc_txrx->ift_rxd_available);
5397 	MPASS(scctx->isc_txrx->ift_rxd_pkt_get);
5398 	MPASS(scctx->isc_txrx->ift_rxd_refill);
5399 	MPASS(scctx->isc_txrx->ift_rxd_flush);
5400 }
5401 
5402 static int
5403 iflib_register(if_ctx_t ctx)
5404 {
5405 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5406 	driver_t *driver = sctx->isc_driver;
5407 	device_t dev = ctx->ifc_dev;
5408 	if_t ifp;
5409 	u_char type;
5410 	int iflags;
5411 
5412 	if ((sctx->isc_flags & IFLIB_PSEUDO) == 0)
5413 		_iflib_assert(sctx);
5414 
5415 	CTX_LOCK_INIT(ctx);
5416 	STATE_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev));
5417 	if (sctx->isc_flags & IFLIB_PSEUDO) {
5418 		if (sctx->isc_flags & IFLIB_PSEUDO_ETHER)
5419 			type = IFT_ETHER;
5420 		else
5421 			type = IFT_PPP;
5422 	} else
5423 		type = IFT_ETHER;
5424 	ifp = ctx->ifc_ifp = if_alloc(type);
5425 	if (ifp == NULL) {
5426 		device_printf(dev, "can not allocate ifnet structure\n");
5427 		return (ENOMEM);
5428 	}
5429 
5430 	/*
5431 	 * Initialize our context's device specific methods
5432 	 */
5433 	kobj_init((kobj_t) ctx, (kobj_class_t) driver);
5434 	kobj_class_compile((kobj_class_t) driver);
5435 
5436 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
5437 	if_setsoftc(ifp, ctx);
5438 	if_setdev(ifp, dev);
5439 	if_setinitfn(ifp, iflib_if_init);
5440 	if_setioctlfn(ifp, iflib_if_ioctl);
5441 #ifdef ALTQ
5442 	if_setstartfn(ifp, iflib_altq_if_start);
5443 	if_settransmitfn(ifp, iflib_altq_if_transmit);
5444 	if_setsendqready(ifp);
5445 #else
5446 	if_settransmitfn(ifp, iflib_if_transmit);
5447 #endif
5448 	if_setqflushfn(ifp, iflib_if_qflush);
5449 	iflags = IFF_MULTICAST | IFF_KNOWSEPOCH;
5450 
5451 	if ((sctx->isc_flags & IFLIB_PSEUDO) &&
5452 		(sctx->isc_flags & IFLIB_PSEUDO_ETHER) == 0)
5453 		iflags |= IFF_POINTOPOINT;
5454 	else
5455 		iflags |= IFF_BROADCAST | IFF_SIMPLEX;
5456 	if_setflags(ifp, iflags);
5457 	ctx->ifc_vlan_attach_event =
5458 		EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx,
5459 							  EVENTHANDLER_PRI_FIRST);
5460 	ctx->ifc_vlan_detach_event =
5461 		EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx,
5462 							  EVENTHANDLER_PRI_FIRST);
5463 
5464 	if ((sctx->isc_flags & IFLIB_DRIVER_MEDIA) == 0) {
5465 		ctx->ifc_mediap = &ctx->ifc_media;
5466 		ifmedia_init(ctx->ifc_mediap, IFM_IMASK,
5467 		    iflib_media_change, iflib_media_status);
5468 	}
5469 	return (0);
5470 }
5471 
5472 static void
5473 iflib_unregister_vlan_handlers(if_ctx_t ctx)
5474 {
5475 	/* Unregister VLAN events */
5476 	if (ctx->ifc_vlan_attach_event != NULL) {
5477 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
5478 		ctx->ifc_vlan_attach_event = NULL;
5479 	}
5480 	if (ctx->ifc_vlan_detach_event != NULL) {
5481 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
5482 		ctx->ifc_vlan_detach_event = NULL;
5483 	}
5484 
5485 }
5486 
5487 static void
5488 iflib_deregister(if_ctx_t ctx)
5489 {
5490 	if_t ifp = ctx->ifc_ifp;
5491 
5492 	/* Remove all media */
5493 	ifmedia_removeall(&ctx->ifc_media);
5494 
5495 	/* Ensure that VLAN event handlers are unregistered */
5496 	iflib_unregister_vlan_handlers(ctx);
5497 
5498 	/* Release kobject reference */
5499 	kobj_delete((kobj_t) ctx, NULL);
5500 
5501 	/* Free the ifnet structure */
5502 	if_free(ifp);
5503 
5504 	STATE_LOCK_DESTROY(ctx);
5505 
5506 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
5507 	CTX_LOCK_DESTROY(ctx);
5508 }
5509 
5510 static int
5511 iflib_queues_alloc(if_ctx_t ctx)
5512 {
5513 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5514 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5515 	device_t dev = ctx->ifc_dev;
5516 	int nrxqsets = scctx->isc_nrxqsets;
5517 	int ntxqsets = scctx->isc_ntxqsets;
5518 	iflib_txq_t txq;
5519 	iflib_rxq_t rxq;
5520 	iflib_fl_t fl = NULL;
5521 	int i, j, cpu, err, txconf, rxconf;
5522 	iflib_dma_info_t ifdip;
5523 	uint32_t *rxqsizes = scctx->isc_rxqsizes;
5524 	uint32_t *txqsizes = scctx->isc_txqsizes;
5525 	uint8_t nrxqs = sctx->isc_nrxqs;
5526 	uint8_t ntxqs = sctx->isc_ntxqs;
5527 	int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1;
5528 	caddr_t *vaddrs;
5529 	uint64_t *paddrs;
5530 
5531 	KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1"));
5532 	KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1"));
5533 
5534 	/* Allocate the TX ring struct memory */
5535 	if (!(ctx->ifc_txqs =
5536 	    (iflib_txq_t) malloc(sizeof(struct iflib_txq) *
5537 	    ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5538 		device_printf(dev, "Unable to allocate TX ring memory\n");
5539 		err = ENOMEM;
5540 		goto fail;
5541 	}
5542 
5543 	/* Now allocate the RX */
5544 	if (!(ctx->ifc_rxqs =
5545 	    (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) *
5546 	    nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5547 		device_printf(dev, "Unable to allocate RX ring memory\n");
5548 		err = ENOMEM;
5549 		goto rx_fail;
5550 	}
5551 
5552 	txq = ctx->ifc_txqs;
5553 	rxq = ctx->ifc_rxqs;
5554 
5555 	/*
5556 	 * XXX handle allocation failure
5557 	 */
5558 	for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) {
5559 		/* Set up some basics */
5560 
5561 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs,
5562 		    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
5563 			device_printf(dev,
5564 			    "Unable to allocate TX DMA info memory\n");
5565 			err = ENOMEM;
5566 			goto err_tx_desc;
5567 		}
5568 		txq->ift_ifdi = ifdip;
5569 		for (j = 0; j < ntxqs; j++, ifdip++) {
5570 			if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, 0)) {
5571 				device_printf(dev,
5572 				    "Unable to allocate TX descriptors\n");
5573 				err = ENOMEM;
5574 				goto err_tx_desc;
5575 			}
5576 			txq->ift_txd_size[j] = scctx->isc_txd_size[j];
5577 			bzero((void *)ifdip->idi_vaddr, txqsizes[j]);
5578 		}
5579 		txq->ift_ctx = ctx;
5580 		txq->ift_id = i;
5581 		if (sctx->isc_flags & IFLIB_HAS_TXCQ) {
5582 			txq->ift_br_offset = 1;
5583 		} else {
5584 			txq->ift_br_offset = 0;
5585 		}
5586 		/* XXX fix this */
5587 		txq->ift_timer.c_cpu = cpu;
5588 
5589 		if (iflib_txsd_alloc(txq)) {
5590 			device_printf(dev, "Critical Failure setting up TX buffers\n");
5591 			err = ENOMEM;
5592 			goto err_tx_desc;
5593 		}
5594 
5595 		/* Initialize the TX lock */
5596 		snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:TX(%d):callout",
5597 		    device_get_nameunit(dev), txq->ift_id);
5598 		mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF);
5599 		callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0);
5600 
5601 		err = ifmp_ring_alloc(&txq->ift_br, 2048, txq, iflib_txq_drain,
5602 				      iflib_txq_can_drain, M_IFLIB, M_WAITOK);
5603 		if (err) {
5604 			/* XXX free any allocated rings */
5605 			device_printf(dev, "Unable to allocate buf_ring\n");
5606 			goto err_tx_desc;
5607 		}
5608 	}
5609 
5610 	for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) {
5611 		/* Set up some basics */
5612 		callout_init(&rxq->ifr_watchdog, 1);
5613 
5614 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs,
5615 		   M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
5616 			device_printf(dev,
5617 			    "Unable to allocate RX DMA info memory\n");
5618 			err = ENOMEM;
5619 			goto err_tx_desc;
5620 		}
5621 
5622 		rxq->ifr_ifdi = ifdip;
5623 		/* XXX this needs to be changed if #rx queues != #tx queues */
5624 		rxq->ifr_ntxqirq = 1;
5625 		rxq->ifr_txqid[0] = i;
5626 		for (j = 0; j < nrxqs; j++, ifdip++) {
5627 			if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, 0)) {
5628 				device_printf(dev,
5629 				    "Unable to allocate RX descriptors\n");
5630 				err = ENOMEM;
5631 				goto err_tx_desc;
5632 			}
5633 			bzero((void *)ifdip->idi_vaddr, rxqsizes[j]);
5634 		}
5635 		rxq->ifr_ctx = ctx;
5636 		rxq->ifr_id = i;
5637 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
5638 			rxq->ifr_fl_offset = 1;
5639 		} else {
5640 			rxq->ifr_fl_offset = 0;
5641 		}
5642 		rxq->ifr_nfl = nfree_lists;
5643 		if (!(fl =
5644 			  (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) {
5645 			device_printf(dev, "Unable to allocate free list memory\n");
5646 			err = ENOMEM;
5647 			goto err_tx_desc;
5648 		}
5649 		rxq->ifr_fl = fl;
5650 		for (j = 0; j < nfree_lists; j++) {
5651 			fl[j].ifl_rxq = rxq;
5652 			fl[j].ifl_id = j;
5653 			fl[j].ifl_ifdi = &rxq->ifr_ifdi[j + rxq->ifr_fl_offset];
5654 			fl[j].ifl_rxd_size = scctx->isc_rxd_size[j];
5655 		}
5656 		/* Allocate receive buffers for the ring */
5657 		if (iflib_rxsd_alloc(rxq)) {
5658 			device_printf(dev,
5659 			    "Critical Failure setting up receive buffers\n");
5660 			err = ENOMEM;
5661 			goto err_rx_desc;
5662 		}
5663 
5664 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5665 			fl->ifl_rx_bitmap = bit_alloc(fl->ifl_size, M_IFLIB,
5666 			    M_WAITOK);
5667 	}
5668 
5669 	/* TXQs */
5670 	vaddrs = malloc(sizeof(caddr_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5671 	paddrs = malloc(sizeof(uint64_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5672 	for (i = 0; i < ntxqsets; i++) {
5673 		iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi;
5674 
5675 		for (j = 0; j < ntxqs; j++, di++) {
5676 			vaddrs[i*ntxqs + j] = di->idi_vaddr;
5677 			paddrs[i*ntxqs + j] = di->idi_paddr;
5678 		}
5679 	}
5680 	if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) {
5681 		device_printf(ctx->ifc_dev,
5682 		    "Unable to allocate device TX queue\n");
5683 		iflib_tx_structures_free(ctx);
5684 		free(vaddrs, M_IFLIB);
5685 		free(paddrs, M_IFLIB);
5686 		goto err_rx_desc;
5687 	}
5688 	free(vaddrs, M_IFLIB);
5689 	free(paddrs, M_IFLIB);
5690 
5691 	/* RXQs */
5692 	vaddrs = malloc(sizeof(caddr_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5693 	paddrs = malloc(sizeof(uint64_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5694 	for (i = 0; i < nrxqsets; i++) {
5695 		iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi;
5696 
5697 		for (j = 0; j < nrxqs; j++, di++) {
5698 			vaddrs[i*nrxqs + j] = di->idi_vaddr;
5699 			paddrs[i*nrxqs + j] = di->idi_paddr;
5700 		}
5701 	}
5702 	if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) {
5703 		device_printf(ctx->ifc_dev,
5704 		    "Unable to allocate device RX queue\n");
5705 		iflib_tx_structures_free(ctx);
5706 		free(vaddrs, M_IFLIB);
5707 		free(paddrs, M_IFLIB);
5708 		goto err_rx_desc;
5709 	}
5710 	free(vaddrs, M_IFLIB);
5711 	free(paddrs, M_IFLIB);
5712 
5713 	return (0);
5714 
5715 /* XXX handle allocation failure changes */
5716 err_rx_desc:
5717 err_tx_desc:
5718 rx_fail:
5719 	if (ctx->ifc_rxqs != NULL)
5720 		free(ctx->ifc_rxqs, M_IFLIB);
5721 	ctx->ifc_rxqs = NULL;
5722 	if (ctx->ifc_txqs != NULL)
5723 		free(ctx->ifc_txqs, M_IFLIB);
5724 	ctx->ifc_txqs = NULL;
5725 fail:
5726 	return (err);
5727 }
5728 
5729 static int
5730 iflib_tx_structures_setup(if_ctx_t ctx)
5731 {
5732 	iflib_txq_t txq = ctx->ifc_txqs;
5733 	int i;
5734 
5735 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
5736 		iflib_txq_setup(txq);
5737 
5738 	return (0);
5739 }
5740 
5741 static void
5742 iflib_tx_structures_free(if_ctx_t ctx)
5743 {
5744 	iflib_txq_t txq = ctx->ifc_txqs;
5745 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5746 	int i, j;
5747 
5748 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
5749 		for (j = 0; j < sctx->isc_ntxqs; j++)
5750 			iflib_dma_free(&txq->ift_ifdi[j]);
5751 		iflib_txq_destroy(txq);
5752 	}
5753 	free(ctx->ifc_txqs, M_IFLIB);
5754 	ctx->ifc_txqs = NULL;
5755 	IFDI_QUEUES_FREE(ctx);
5756 }
5757 
5758 /*********************************************************************
5759  *
5760  *  Initialize all receive rings.
5761  *
5762  **********************************************************************/
5763 static int
5764 iflib_rx_structures_setup(if_ctx_t ctx)
5765 {
5766 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5767 	int q;
5768 #if defined(INET6) || defined(INET)
5769 	int err, i;
5770 #endif
5771 
5772 	for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) {
5773 #if defined(INET6) || defined(INET)
5774 		if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_LRO) {
5775 			err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp,
5776 			    TCP_LRO_ENTRIES, min(1024,
5777 			    ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset]));
5778 			if (err != 0) {
5779 				device_printf(ctx->ifc_dev,
5780 				    "LRO Initialization failed!\n");
5781 				goto fail;
5782 			}
5783 		}
5784 #endif
5785 		IFDI_RXQ_SETUP(ctx, rxq->ifr_id);
5786 	}
5787 	return (0);
5788 #if defined(INET6) || defined(INET)
5789 fail:
5790 	/*
5791 	 * Free LRO resources allocated so far, we will only handle
5792 	 * the rings that completed, the failing case will have
5793 	 * cleaned up for itself.  'q' failed, so its the terminus.
5794 	 */
5795 	rxq = ctx->ifc_rxqs;
5796 	for (i = 0; i < q; ++i, rxq++) {
5797 		if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_LRO)
5798 			tcp_lro_free(&rxq->ifr_lc);
5799 	}
5800 	return (err);
5801 #endif
5802 }
5803 
5804 /*********************************************************************
5805  *
5806  *  Free all receive rings.
5807  *
5808  **********************************************************************/
5809 static void
5810 iflib_rx_structures_free(if_ctx_t ctx)
5811 {
5812 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5813 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5814 	int i, j;
5815 
5816 	for (i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) {
5817 		for (j = 0; j < sctx->isc_nrxqs; j++)
5818 			iflib_dma_free(&rxq->ifr_ifdi[j]);
5819 		iflib_rx_sds_free(rxq);
5820 #if defined(INET6) || defined(INET)
5821 		if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_LRO)
5822 			tcp_lro_free(&rxq->ifr_lc);
5823 #endif
5824 	}
5825 	free(ctx->ifc_rxqs, M_IFLIB);
5826 	ctx->ifc_rxqs = NULL;
5827 }
5828 
5829 static int
5830 iflib_qset_structures_setup(if_ctx_t ctx)
5831 {
5832 	int err;
5833 
5834 	/*
5835 	 * It is expected that the caller takes care of freeing queues if this
5836 	 * fails.
5837 	 */
5838 	if ((err = iflib_tx_structures_setup(ctx)) != 0) {
5839 		device_printf(ctx->ifc_dev, "iflib_tx_structures_setup failed: %d\n", err);
5840 		return (err);
5841 	}
5842 
5843 	if ((err = iflib_rx_structures_setup(ctx)) != 0)
5844 		device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err);
5845 
5846 	return (err);
5847 }
5848 
5849 int
5850 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
5851 		driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, const char *name)
5852 {
5853 
5854 	return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name));
5855 }
5856 
5857 #ifdef SMP
5858 static int
5859 find_nth(if_ctx_t ctx, int qid)
5860 {
5861 	cpuset_t cpus;
5862 	int i, cpuid, eqid, count;
5863 
5864 	CPU_COPY(&ctx->ifc_cpus, &cpus);
5865 	count = CPU_COUNT(&cpus);
5866 	eqid = qid % count;
5867 	/* clear up to the qid'th bit */
5868 	for (i = 0; i < eqid; i++) {
5869 		cpuid = CPU_FFS(&cpus);
5870 		MPASS(cpuid != 0);
5871 		CPU_CLR(cpuid-1, &cpus);
5872 	}
5873 	cpuid = CPU_FFS(&cpus);
5874 	MPASS(cpuid != 0);
5875 	return (cpuid-1);
5876 }
5877 
5878 #ifdef SCHED_ULE
5879 extern struct cpu_group *cpu_top;              /* CPU topology */
5880 
5881 static int
5882 find_child_with_core(int cpu, struct cpu_group *grp)
5883 {
5884 	int i;
5885 
5886 	if (grp->cg_children == 0)
5887 		return -1;
5888 
5889 	MPASS(grp->cg_child);
5890 	for (i = 0; i < grp->cg_children; i++) {
5891 		if (CPU_ISSET(cpu, &grp->cg_child[i].cg_mask))
5892 			return i;
5893 	}
5894 
5895 	return -1;
5896 }
5897 
5898 /*
5899  * Find the nth "close" core to the specified core
5900  * "close" is defined as the deepest level that shares
5901  * at least an L2 cache.  With threads, this will be
5902  * threads on the same core.  If the shared cache is L3
5903  * or higher, simply returns the same core.
5904  */
5905 static int
5906 find_close_core(int cpu, int core_offset)
5907 {
5908 	struct cpu_group *grp;
5909 	int i;
5910 	int fcpu;
5911 	cpuset_t cs;
5912 
5913 	grp = cpu_top;
5914 	if (grp == NULL)
5915 		return cpu;
5916 	i = 0;
5917 	while ((i = find_child_with_core(cpu, grp)) != -1) {
5918 		/* If the child only has one cpu, don't descend */
5919 		if (grp->cg_child[i].cg_count <= 1)
5920 			break;
5921 		grp = &grp->cg_child[i];
5922 	}
5923 
5924 	/* If they don't share at least an L2 cache, use the same CPU */
5925 	if (grp->cg_level > CG_SHARE_L2 || grp->cg_level == CG_SHARE_NONE)
5926 		return cpu;
5927 
5928 	/* Now pick one */
5929 	CPU_COPY(&grp->cg_mask, &cs);
5930 
5931 	/* Add the selected CPU offset to core offset. */
5932 	for (i = 0; (fcpu = CPU_FFS(&cs)) != 0; i++) {
5933 		if (fcpu - 1 == cpu)
5934 			break;
5935 		CPU_CLR(fcpu - 1, &cs);
5936 	}
5937 	MPASS(fcpu);
5938 
5939 	core_offset += i;
5940 
5941 	CPU_COPY(&grp->cg_mask, &cs);
5942 	for (i = core_offset % grp->cg_count; i > 0; i--) {
5943 		MPASS(CPU_FFS(&cs));
5944 		CPU_CLR(CPU_FFS(&cs) - 1, &cs);
5945 	}
5946 	MPASS(CPU_FFS(&cs));
5947 	return CPU_FFS(&cs) - 1;
5948 }
5949 #else
5950 static int
5951 find_close_core(int cpu, int core_offset __unused)
5952 {
5953 	return cpu;
5954 }
5955 #endif
5956 
5957 static int
5958 get_core_offset(if_ctx_t ctx, iflib_intr_type_t type, int qid)
5959 {
5960 	switch (type) {
5961 	case IFLIB_INTR_TX:
5962 		/* TX queues get cores which share at least an L2 cache with the corresponding RX queue */
5963 		/* XXX handle multiple RX threads per core and more than two core per L2 group */
5964 		return qid / CPU_COUNT(&ctx->ifc_cpus) + 1;
5965 	case IFLIB_INTR_RX:
5966 	case IFLIB_INTR_RXTX:
5967 		/* RX queues get the specified core */
5968 		return qid / CPU_COUNT(&ctx->ifc_cpus);
5969 	default:
5970 		return -1;
5971 	}
5972 }
5973 #else
5974 #define get_core_offset(ctx, type, qid)	CPU_FIRST()
5975 #define find_close_core(cpuid, tid)	CPU_FIRST()
5976 #define find_nth(ctx, gid)		CPU_FIRST()
5977 #endif
5978 
5979 /* Just to avoid copy/paste */
5980 static inline int
5981 iflib_irq_set_affinity(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type,
5982     int qid, struct grouptask *gtask, struct taskqgroup *tqg, void *uniq,
5983     const char *name)
5984 {
5985 	device_t dev;
5986 	int co, cpuid, err, tid;
5987 
5988 	dev = ctx->ifc_dev;
5989 	co = ctx->ifc_sysctl_core_offset;
5990 	if (ctx->ifc_sysctl_separate_txrx && type == IFLIB_INTR_TX)
5991 		co += ctx->ifc_softc_ctx.isc_nrxqsets;
5992 	cpuid = find_nth(ctx, qid + co);
5993 	tid = get_core_offset(ctx, type, qid);
5994 	if (tid < 0) {
5995 		device_printf(dev, "get_core_offset failed\n");
5996 		return (EOPNOTSUPP);
5997 	}
5998 	cpuid = find_close_core(cpuid, tid);
5999 	err = taskqgroup_attach_cpu(tqg, gtask, uniq, cpuid, dev, irq->ii_res,
6000 	    name);
6001 	if (err) {
6002 		device_printf(dev, "taskqgroup_attach_cpu failed %d\n", err);
6003 		return (err);
6004 	}
6005 #ifdef notyet
6006 	if (cpuid > ctx->ifc_cpuid_highest)
6007 		ctx->ifc_cpuid_highest = cpuid;
6008 #endif
6009 	return (0);
6010 }
6011 
6012 int
6013 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid,
6014 			iflib_intr_type_t type, driver_filter_t *filter,
6015 			void *filter_arg, int qid, const char *name)
6016 {
6017 	device_t dev;
6018 	struct grouptask *gtask;
6019 	struct taskqgroup *tqg;
6020 	iflib_filter_info_t info;
6021 	gtask_fn_t *fn;
6022 	int tqrid, err;
6023 	driver_filter_t *intr_fast;
6024 	void *q;
6025 
6026 	info = &ctx->ifc_filter_info;
6027 	tqrid = rid;
6028 
6029 	switch (type) {
6030 	/* XXX merge tx/rx for netmap? */
6031 	case IFLIB_INTR_TX:
6032 		q = &ctx->ifc_txqs[qid];
6033 		info = &ctx->ifc_txqs[qid].ift_filter_info;
6034 		gtask = &ctx->ifc_txqs[qid].ift_task;
6035 		tqg = qgroup_if_io_tqg;
6036 		fn = _task_fn_tx;
6037 		intr_fast = iflib_fast_intr;
6038 		GROUPTASK_INIT(gtask, 0, fn, q);
6039 		ctx->ifc_flags |= IFC_NETMAP_TX_IRQ;
6040 		break;
6041 	case IFLIB_INTR_RX:
6042 		q = &ctx->ifc_rxqs[qid];
6043 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
6044 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6045 		tqg = qgroup_if_io_tqg;
6046 		fn = _task_fn_rx;
6047 		intr_fast = iflib_fast_intr;
6048 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6049 		break;
6050 	case IFLIB_INTR_RXTX:
6051 		q = &ctx->ifc_rxqs[qid];
6052 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
6053 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6054 		tqg = qgroup_if_io_tqg;
6055 		fn = _task_fn_rx;
6056 		intr_fast = iflib_fast_intr_rxtx;
6057 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6058 		break;
6059 	case IFLIB_INTR_ADMIN:
6060 		q = ctx;
6061 		tqrid = -1;
6062 		info = &ctx->ifc_filter_info;
6063 		gtask = &ctx->ifc_admin_task;
6064 		tqg = qgroup_if_config_tqg;
6065 		fn = _task_fn_admin;
6066 		intr_fast = iflib_fast_intr_ctx;
6067 		break;
6068 	default:
6069 		device_printf(ctx->ifc_dev, "%s: unknown net intr type\n",
6070 		    __func__);
6071 		return (EINVAL);
6072 	}
6073 
6074 	info->ifi_filter = filter;
6075 	info->ifi_filter_arg = filter_arg;
6076 	info->ifi_task = gtask;
6077 	info->ifi_ctx = q;
6078 
6079 	dev = ctx->ifc_dev;
6080 	err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info,  name);
6081 	if (err != 0) {
6082 		device_printf(dev, "_iflib_irq_alloc failed %d\n", err);
6083 		return (err);
6084 	}
6085 	if (type == IFLIB_INTR_ADMIN)
6086 		return (0);
6087 
6088 	if (tqrid != -1) {
6089 		err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg,
6090 		    q, name);
6091 		if (err)
6092 			return (err);
6093 	} else {
6094 		taskqgroup_attach(tqg, gtask, q, dev, irq->ii_res, name);
6095 	}
6096 
6097 	return (0);
6098 }
6099 
6100 void
6101 iflib_softirq_alloc_generic(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type, void *arg, int qid, const char *name)
6102 {
6103 	struct grouptask *gtask;
6104 	struct taskqgroup *tqg;
6105 	gtask_fn_t *fn;
6106 	void *q;
6107 	int err;
6108 
6109 	switch (type) {
6110 	case IFLIB_INTR_TX:
6111 		q = &ctx->ifc_txqs[qid];
6112 		gtask = &ctx->ifc_txqs[qid].ift_task;
6113 		tqg = qgroup_if_io_tqg;
6114 		fn = _task_fn_tx;
6115 		GROUPTASK_INIT(gtask, 0, fn, q);
6116 		break;
6117 	case IFLIB_INTR_RX:
6118 		q = &ctx->ifc_rxqs[qid];
6119 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6120 		tqg = qgroup_if_io_tqg;
6121 		fn = _task_fn_rx;
6122 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6123 		break;
6124 	case IFLIB_INTR_IOV:
6125 		q = ctx;
6126 		gtask = &ctx->ifc_vflr_task;
6127 		tqg = qgroup_if_config_tqg;
6128 		fn = _task_fn_iov;
6129 		GROUPTASK_INIT(gtask, 0, fn, q);
6130 		break;
6131 	default:
6132 		panic("unknown net intr type");
6133 	}
6134 	if (irq != NULL) {
6135 		err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg,
6136 		    q, name);
6137 		if (err)
6138 			taskqgroup_attach(tqg, gtask, q, ctx->ifc_dev,
6139 			    irq->ii_res, name);
6140 	} else {
6141 		taskqgroup_attach(tqg, gtask, q, NULL, NULL, name);
6142 	}
6143 }
6144 
6145 void
6146 iflib_irq_free(if_ctx_t ctx, if_irq_t irq)
6147 {
6148 
6149 	if (irq->ii_tag)
6150 		bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag);
6151 
6152 	if (irq->ii_res)
6153 		bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ,
6154 		    rman_get_rid(irq->ii_res), irq->ii_res);
6155 }
6156 
6157 static int
6158 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, const char *name)
6159 {
6160 	iflib_txq_t txq = ctx->ifc_txqs;
6161 	iflib_rxq_t rxq = ctx->ifc_rxqs;
6162 	if_irq_t irq = &ctx->ifc_legacy_irq;
6163 	iflib_filter_info_t info;
6164 	device_t dev;
6165 	struct grouptask *gtask;
6166 	struct resource *res;
6167 	struct taskqgroup *tqg;
6168 	void *q;
6169 	int err, tqrid;
6170 	bool rx_only;
6171 
6172 	q = &ctx->ifc_rxqs[0];
6173 	info = &rxq[0].ifr_filter_info;
6174 	gtask = &rxq[0].ifr_task;
6175 	tqg = qgroup_if_io_tqg;
6176 	tqrid = *rid;
6177 	rx_only = (ctx->ifc_sctx->isc_flags & IFLIB_SINGLE_IRQ_RX_ONLY) != 0;
6178 
6179 	ctx->ifc_flags |= IFC_LEGACY;
6180 	info->ifi_filter = filter;
6181 	info->ifi_filter_arg = filter_arg;
6182 	info->ifi_task = gtask;
6183 	info->ifi_ctx = rx_only ? ctx : q;
6184 
6185 	dev = ctx->ifc_dev;
6186 	/* We allocate a single interrupt resource */
6187 	err = _iflib_irq_alloc(ctx, irq, tqrid, rx_only ? iflib_fast_intr_ctx :
6188 	    iflib_fast_intr_rxtx, NULL, info, name);
6189 	if (err != 0)
6190 		return (err);
6191 	NET_GROUPTASK_INIT(gtask, 0, _task_fn_rx, q);
6192 	res = irq->ii_res;
6193 	taskqgroup_attach(tqg, gtask, q, dev, res, name);
6194 
6195 	GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq);
6196 	taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, dev, res,
6197 	    "tx");
6198 	return (0);
6199 }
6200 
6201 void
6202 iflib_led_create(if_ctx_t ctx)
6203 {
6204 
6205 	ctx->ifc_led_dev = led_create(iflib_led_func, ctx,
6206 	    device_get_nameunit(ctx->ifc_dev));
6207 }
6208 
6209 void
6210 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid)
6211 {
6212 
6213 	GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
6214 }
6215 
6216 void
6217 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid)
6218 {
6219 
6220 	GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task);
6221 }
6222 
6223 void
6224 iflib_admin_intr_deferred(if_ctx_t ctx)
6225 {
6226 
6227 	MPASS(ctx->ifc_admin_task.gt_taskqueue != NULL);
6228 	GROUPTASK_ENQUEUE(&ctx->ifc_admin_task);
6229 }
6230 
6231 void
6232 iflib_iov_intr_deferred(if_ctx_t ctx)
6233 {
6234 
6235 	GROUPTASK_ENQUEUE(&ctx->ifc_vflr_task);
6236 }
6237 
6238 void
6239 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, const char *name)
6240 {
6241 
6242 	taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, NULL, NULL,
6243 	    name);
6244 }
6245 
6246 void
6247 iflib_config_gtask_init(void *ctx, struct grouptask *gtask, gtask_fn_t *fn,
6248 	const char *name)
6249 {
6250 
6251 	GROUPTASK_INIT(gtask, 0, fn, ctx);
6252 	taskqgroup_attach(qgroup_if_config_tqg, gtask, gtask, NULL, NULL,
6253 	    name);
6254 }
6255 
6256 void
6257 iflib_config_gtask_deinit(struct grouptask *gtask)
6258 {
6259 
6260 	taskqgroup_detach(qgroup_if_config_tqg, gtask);
6261 }
6262 
6263 void
6264 iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate)
6265 {
6266 	if_t ifp = ctx->ifc_ifp;
6267 	iflib_txq_t txq = ctx->ifc_txqs;
6268 
6269 	if_setbaudrate(ifp, baudrate);
6270 	if (baudrate >= IF_Gbps(10)) {
6271 		STATE_LOCK(ctx);
6272 		ctx->ifc_flags |= IFC_PREFETCH;
6273 		STATE_UNLOCK(ctx);
6274 	}
6275 	/* If link down, disable watchdog */
6276 	if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) {
6277 		for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++)
6278 			txq->ift_qstatus = IFLIB_QUEUE_IDLE;
6279 	}
6280 	ctx->ifc_link_state = link_state;
6281 	if_link_state_change(ifp, link_state);
6282 }
6283 
6284 static int
6285 iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq)
6286 {
6287 	int credits;
6288 #ifdef INVARIANTS
6289 	int credits_pre = txq->ift_cidx_processed;
6290 #endif
6291 
6292 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
6293 	    BUS_DMASYNC_POSTREAD);
6294 	if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, true)) == 0)
6295 		return (0);
6296 
6297 	txq->ift_processed += credits;
6298 	txq->ift_cidx_processed += credits;
6299 
6300 	MPASS(credits_pre + credits == txq->ift_cidx_processed);
6301 	if (txq->ift_cidx_processed >= txq->ift_size)
6302 		txq->ift_cidx_processed -= txq->ift_size;
6303 	return (credits);
6304 }
6305 
6306 static int
6307 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget)
6308 {
6309 	iflib_fl_t fl;
6310 	u_int i;
6311 
6312 	for (i = 0, fl = &rxq->ifr_fl[0]; i < rxq->ifr_nfl; i++, fl++)
6313 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
6314 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
6315 	return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx,
6316 	    budget));
6317 }
6318 
6319 void
6320 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name,
6321 	const char *description, if_int_delay_info_t info,
6322 	int offset, int value)
6323 {
6324 	info->iidi_ctx = ctx;
6325 	info->iidi_offset = offset;
6326 	info->iidi_value = value;
6327 	SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev),
6328 	    SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)),
6329 	    OID_AUTO, name, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE,
6330 	    info, 0, iflib_sysctl_int_delay, "I", description);
6331 }
6332 
6333 struct sx *
6334 iflib_ctx_lock_get(if_ctx_t ctx)
6335 {
6336 
6337 	return (&ctx->ifc_ctx_sx);
6338 }
6339 
6340 static int
6341 iflib_msix_init(if_ctx_t ctx)
6342 {
6343 	device_t dev = ctx->ifc_dev;
6344 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6345 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6346 	int admincnt, bar, err, iflib_num_rx_queues, iflib_num_tx_queues;
6347 	int msgs, queuemsgs, queues, rx_queues, tx_queues, vectors;
6348 
6349 	iflib_num_tx_queues = ctx->ifc_sysctl_ntxqs;
6350 	iflib_num_rx_queues = ctx->ifc_sysctl_nrxqs;
6351 
6352 	if (bootverbose)
6353 		device_printf(dev, "msix_init qsets capped at %d\n",
6354 		    imax(scctx->isc_ntxqsets, scctx->isc_nrxqsets));
6355 
6356 	/* Override by tuneable */
6357 	if (scctx->isc_disable_msix)
6358 		goto msi;
6359 
6360 	/* First try MSI-X */
6361 	if ((msgs = pci_msix_count(dev)) == 0) {
6362 		if (bootverbose)
6363 			device_printf(dev, "MSI-X not supported or disabled\n");
6364 		goto msi;
6365 	}
6366 
6367 	bar = ctx->ifc_softc_ctx.isc_msix_bar;
6368 	/*
6369 	 * bar == -1 => "trust me I know what I'm doing"
6370 	 * Some drivers are for hardware that is so shoddily
6371 	 * documented that no one knows which bars are which
6372 	 * so the developer has to map all bars. This hack
6373 	 * allows shoddy garbage to use MSI-X in this framework.
6374 	 */
6375 	if (bar != -1) {
6376 		ctx->ifc_msix_mem = bus_alloc_resource_any(dev,
6377 	            SYS_RES_MEMORY, &bar, RF_ACTIVE);
6378 		if (ctx->ifc_msix_mem == NULL) {
6379 			device_printf(dev, "Unable to map MSI-X table\n");
6380 			goto msi;
6381 		}
6382 	}
6383 
6384 	admincnt = sctx->isc_admin_intrcnt;
6385 #if IFLIB_DEBUG
6386 	/* use only 1 qset in debug mode */
6387 	queuemsgs = min(msgs - admincnt, 1);
6388 #else
6389 	queuemsgs = msgs - admincnt;
6390 #endif
6391 #ifdef RSS
6392 	queues = imin(queuemsgs, rss_getnumbuckets());
6393 #else
6394 	queues = queuemsgs;
6395 #endif
6396 	queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues);
6397 	if (bootverbose)
6398 		device_printf(dev,
6399 		    "intr CPUs: %d queue msgs: %d admincnt: %d\n",
6400 		    CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt);
6401 #ifdef  RSS
6402 	/* If we're doing RSS, clamp at the number of RSS buckets */
6403 	if (queues > rss_getnumbuckets())
6404 		queues = rss_getnumbuckets();
6405 #endif
6406 	if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt)
6407 		rx_queues = iflib_num_rx_queues;
6408 	else
6409 		rx_queues = queues;
6410 
6411 	if (rx_queues > scctx->isc_nrxqsets)
6412 		rx_queues = scctx->isc_nrxqsets;
6413 
6414 	/*
6415 	 * We want this to be all logical CPUs by default
6416 	 */
6417 	if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues)
6418 		tx_queues = iflib_num_tx_queues;
6419 	else
6420 		tx_queues = mp_ncpus;
6421 
6422 	if (tx_queues > scctx->isc_ntxqsets)
6423 		tx_queues = scctx->isc_ntxqsets;
6424 
6425 	if (ctx->ifc_sysctl_qs_eq_override == 0) {
6426 #ifdef INVARIANTS
6427 		if (tx_queues != rx_queues)
6428 			device_printf(dev,
6429 			    "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n",
6430 			    min(rx_queues, tx_queues), min(rx_queues, tx_queues));
6431 #endif
6432 		tx_queues = min(rx_queues, tx_queues);
6433 		rx_queues = min(rx_queues, tx_queues);
6434 	}
6435 
6436 	vectors = rx_queues + admincnt;
6437 	if (msgs < vectors) {
6438 		device_printf(dev,
6439 		    "insufficient number of MSI-X vectors "
6440 		    "(supported %d, need %d)\n", msgs, vectors);
6441 		goto msi;
6442 	}
6443 
6444 	device_printf(dev, "Using %d RX queues %d TX queues\n", rx_queues,
6445 	    tx_queues);
6446 	msgs = vectors;
6447 	if ((err = pci_alloc_msix(dev, &vectors)) == 0) {
6448 		if (vectors != msgs) {
6449 			device_printf(dev,
6450 			    "Unable to allocate sufficient MSI-X vectors "
6451 			    "(got %d, need %d)\n", vectors, msgs);
6452 			pci_release_msi(dev);
6453 			if (bar != -1) {
6454 				bus_release_resource(dev, SYS_RES_MEMORY, bar,
6455 				    ctx->ifc_msix_mem);
6456 				ctx->ifc_msix_mem = NULL;
6457 			}
6458 			goto msi;
6459 		}
6460 		device_printf(dev, "Using MSI-X interrupts with %d vectors\n",
6461 		    vectors);
6462 		scctx->isc_vectors = vectors;
6463 		scctx->isc_nrxqsets = rx_queues;
6464 		scctx->isc_ntxqsets = tx_queues;
6465 		scctx->isc_intr = IFLIB_INTR_MSIX;
6466 
6467 		return (vectors);
6468 	} else {
6469 		device_printf(dev,
6470 		    "failed to allocate %d MSI-X vectors, err: %d\n", vectors,
6471 		    err);
6472 		if (bar != -1) {
6473 			bus_release_resource(dev, SYS_RES_MEMORY, bar,
6474 			    ctx->ifc_msix_mem);
6475 			ctx->ifc_msix_mem = NULL;
6476 		}
6477 	}
6478 
6479 msi:
6480 	vectors = pci_msi_count(dev);
6481 	scctx->isc_nrxqsets = 1;
6482 	scctx->isc_ntxqsets = 1;
6483 	scctx->isc_vectors = vectors;
6484 	if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) {
6485 		device_printf(dev,"Using an MSI interrupt\n");
6486 		scctx->isc_intr = IFLIB_INTR_MSI;
6487 	} else {
6488 		scctx->isc_vectors = 1;
6489 		device_printf(dev,"Using a Legacy interrupt\n");
6490 		scctx->isc_intr = IFLIB_INTR_LEGACY;
6491 	}
6492 
6493 	return (vectors);
6494 }
6495 
6496 static const char *ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" };
6497 
6498 static int
6499 mp_ring_state_handler(SYSCTL_HANDLER_ARGS)
6500 {
6501 	int rc;
6502 	uint16_t *state = ((uint16_t *)oidp->oid_arg1);
6503 	struct sbuf *sb;
6504 	const char *ring_state = "UNKNOWN";
6505 
6506 	/* XXX needed ? */
6507 	rc = sysctl_wire_old_buffer(req, 0);
6508 	MPASS(rc == 0);
6509 	if (rc != 0)
6510 		return (rc);
6511 	sb = sbuf_new_for_sysctl(NULL, NULL, 80, req);
6512 	MPASS(sb != NULL);
6513 	if (sb == NULL)
6514 		return (ENOMEM);
6515 	if (state[3] <= 3)
6516 		ring_state = ring_states[state[3]];
6517 
6518 	sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s",
6519 		    state[0], state[1], state[2], ring_state);
6520 	rc = sbuf_finish(sb);
6521 	sbuf_delete(sb);
6522         return(rc);
6523 }
6524 
6525 enum iflib_ndesc_handler {
6526 	IFLIB_NTXD_HANDLER,
6527 	IFLIB_NRXD_HANDLER,
6528 };
6529 
6530 static int
6531 mp_ndesc_handler(SYSCTL_HANDLER_ARGS)
6532 {
6533 	if_ctx_t ctx = (void *)arg1;
6534 	enum iflib_ndesc_handler type = arg2;
6535 	char buf[256] = {0};
6536 	qidx_t *ndesc;
6537 	char *p, *next;
6538 	int nqs, rc, i;
6539 
6540 	nqs = 8;
6541 	switch(type) {
6542 	case IFLIB_NTXD_HANDLER:
6543 		ndesc = ctx->ifc_sysctl_ntxds;
6544 		if (ctx->ifc_sctx)
6545 			nqs = ctx->ifc_sctx->isc_ntxqs;
6546 		break;
6547 	case IFLIB_NRXD_HANDLER:
6548 		ndesc = ctx->ifc_sysctl_nrxds;
6549 		if (ctx->ifc_sctx)
6550 			nqs = ctx->ifc_sctx->isc_nrxqs;
6551 		break;
6552 	default:
6553 		printf("%s: unhandled type\n", __func__);
6554 		return (EINVAL);
6555 	}
6556 	if (nqs == 0)
6557 		nqs = 8;
6558 
6559 	for (i=0; i<8; i++) {
6560 		if (i >= nqs)
6561 			break;
6562 		if (i)
6563 			strcat(buf, ",");
6564 		sprintf(strchr(buf, 0), "%d", ndesc[i]);
6565 	}
6566 
6567 	rc = sysctl_handle_string(oidp, buf, sizeof(buf), req);
6568 	if (rc || req->newptr == NULL)
6569 		return rc;
6570 
6571 	for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p;
6572 	    i++, p = strsep(&next, " ,")) {
6573 		ndesc[i] = strtoul(p, NULL, 10);
6574 	}
6575 
6576 	return(rc);
6577 }
6578 
6579 #define NAME_BUFLEN 32
6580 static void
6581 iflib_add_device_sysctl_pre(if_ctx_t ctx)
6582 {
6583         device_t dev = iflib_get_dev(ctx);
6584 	struct sysctl_oid_list *child, *oid_list;
6585 	struct sysctl_ctx_list *ctx_list;
6586 	struct sysctl_oid *node;
6587 
6588 	ctx_list = device_get_sysctl_ctx(dev);
6589 	child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev));
6590 	ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, "iflib",
6591 	    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "IFLIB fields");
6592 	oid_list = SYSCTL_CHILDREN(node);
6593 
6594 	SYSCTL_ADD_CONST_STRING(ctx_list, oid_list, OID_AUTO, "driver_version",
6595 		       CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version,
6596 		       "driver version");
6597 
6598 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs",
6599 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0,
6600 			"# of txqs to use, 0 => use default #");
6601 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs",
6602 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0,
6603 			"# of rxqs to use, 0 => use default #");
6604 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_qs_enable",
6605 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0,
6606                        "permit #txq != #rxq");
6607 	SYSCTL_ADD_INT(ctx_list, oid_list, OID_AUTO, "disable_msix",
6608                       CTLFLAG_RWTUN, &ctx->ifc_softc_ctx.isc_disable_msix, 0,
6609                       "disable MSI-X (default 0)");
6610 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "rx_budget",
6611 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_rx_budget, 0,
6612 		       "set the RX budget");
6613 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "tx_abdicate",
6614 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_tx_abdicate, 0,
6615 		       "cause TX to abdicate instead of running to completion");
6616 	ctx->ifc_sysctl_core_offset = CORE_OFFSET_UNSPECIFIED;
6617 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "core_offset",
6618 		       CTLFLAG_RDTUN, &ctx->ifc_sysctl_core_offset, 0,
6619 		       "offset to start using cores at");
6620 	SYSCTL_ADD_U8(ctx_list, oid_list, OID_AUTO, "separate_txrx",
6621 		       CTLFLAG_RDTUN, &ctx->ifc_sysctl_separate_txrx, 0,
6622 		       "use separate cores for TX and RX");
6623 
6624 	/* XXX change for per-queue sizes */
6625 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_ntxds",
6626 	    CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NEEDGIANT, ctx,
6627 	    IFLIB_NTXD_HANDLER, mp_ndesc_handler, "A",
6628 	    "list of # of TX descriptors to use, 0 = use default #");
6629 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_nrxds",
6630 	    CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NEEDGIANT, ctx,
6631 	    IFLIB_NRXD_HANDLER, mp_ndesc_handler, "A",
6632 	    "list of # of RX descriptors to use, 0 = use default #");
6633 }
6634 
6635 static void
6636 iflib_add_device_sysctl_post(if_ctx_t ctx)
6637 {
6638 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6639 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6640         device_t dev = iflib_get_dev(ctx);
6641 	struct sysctl_oid_list *child;
6642 	struct sysctl_ctx_list *ctx_list;
6643 	iflib_fl_t fl;
6644 	iflib_txq_t txq;
6645 	iflib_rxq_t rxq;
6646 	int i, j;
6647 	char namebuf[NAME_BUFLEN];
6648 	char *qfmt;
6649 	struct sysctl_oid *queue_node, *fl_node, *node;
6650 	struct sysctl_oid_list *queue_list, *fl_list;
6651 	ctx_list = device_get_sysctl_ctx(dev);
6652 
6653 	node = ctx->ifc_sysctl_node;
6654 	child = SYSCTL_CHILDREN(node);
6655 
6656 	if (scctx->isc_ntxqsets > 100)
6657 		qfmt = "txq%03d";
6658 	else if (scctx->isc_ntxqsets > 10)
6659 		qfmt = "txq%02d";
6660 	else
6661 		qfmt = "txq%d";
6662 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
6663 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6664 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6665 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "Queue Name");
6666 		queue_list = SYSCTL_CHILDREN(queue_node);
6667 #if MEMORY_LOGGING
6668 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued",
6669 				CTLFLAG_RD,
6670 				&txq->ift_dequeued, "total mbufs freed");
6671 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued",
6672 				CTLFLAG_RD,
6673 				&txq->ift_enqueued, "total mbufs enqueued");
6674 #endif
6675 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag",
6676 				   CTLFLAG_RD,
6677 				   &txq->ift_mbuf_defrag, "# of times m_defrag was called");
6678 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "m_pullups",
6679 				   CTLFLAG_RD,
6680 				   &txq->ift_pullups, "# of times m_pullup was called");
6681 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag_failed",
6682 				   CTLFLAG_RD,
6683 				   &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed");
6684 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_desc_avail",
6685 				   CTLFLAG_RD,
6686 				   &txq->ift_no_desc_avail, "# of times no descriptors were available");
6687 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "tx_map_failed",
6688 				   CTLFLAG_RD,
6689 				   &txq->ift_map_failed, "# of times DMA map failed");
6690 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txd_encap_efbig",
6691 				   CTLFLAG_RD,
6692 				   &txq->ift_txd_encap_efbig, "# of times txd_encap returned EFBIG");
6693 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_tx_dma_setup",
6694 				   CTLFLAG_RD,
6695 				   &txq->ift_no_tx_dma_setup, "# of times map failed for other than EFBIG");
6696 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx",
6697 				   CTLFLAG_RD,
6698 				   &txq->ift_pidx, 1, "Producer Index");
6699 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx",
6700 				   CTLFLAG_RD,
6701 				   &txq->ift_cidx, 1, "Consumer Index");
6702 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx_processed",
6703 				   CTLFLAG_RD,
6704 				   &txq->ift_cidx_processed, 1, "Consumer Index seen by credit update");
6705 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use",
6706 				   CTLFLAG_RD,
6707 				   &txq->ift_in_use, 1, "descriptors in use");
6708 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_processed",
6709 				   CTLFLAG_RD,
6710 				   &txq->ift_processed, "descriptors procesed for clean");
6711 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned",
6712 				   CTLFLAG_RD,
6713 				   &txq->ift_cleaned, "total cleaned");
6714 		SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state",
6715 		    CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_NEEDGIANT,
6716 		    __DEVOLATILE(uint64_t *, &txq->ift_br->state), 0,
6717 		    mp_ring_state_handler, "A", "soft ring state");
6718 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_enqueues",
6719 				       CTLFLAG_RD, &txq->ift_br->enqueues,
6720 				       "# of enqueues to the mp_ring for this queue");
6721 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_drops",
6722 				       CTLFLAG_RD, &txq->ift_br->drops,
6723 				       "# of drops in the mp_ring for this queue");
6724 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_starts",
6725 				       CTLFLAG_RD, &txq->ift_br->starts,
6726 				       "# of normal consumer starts in the mp_ring for this queue");
6727 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_stalls",
6728 				       CTLFLAG_RD, &txq->ift_br->stalls,
6729 					       "# of consumer stalls in the mp_ring for this queue");
6730 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_restarts",
6731 			       CTLFLAG_RD, &txq->ift_br->restarts,
6732 				       "# of consumer restarts in the mp_ring for this queue");
6733 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_abdications",
6734 				       CTLFLAG_RD, &txq->ift_br->abdications,
6735 				       "# of consumer abdications in the mp_ring for this queue");
6736 	}
6737 
6738 	if (scctx->isc_nrxqsets > 100)
6739 		qfmt = "rxq%03d";
6740 	else if (scctx->isc_nrxqsets > 10)
6741 		qfmt = "rxq%02d";
6742 	else
6743 		qfmt = "rxq%d";
6744 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
6745 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6746 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6747 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "Queue Name");
6748 		queue_list = SYSCTL_CHILDREN(queue_node);
6749 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
6750 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_cidx",
6751 				       CTLFLAG_RD,
6752 				       &rxq->ifr_cq_cidx, 1, "Consumer Index");
6753 		}
6754 
6755 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
6756 			snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j);
6757 			fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list, OID_AUTO, namebuf,
6758 			    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "freelist Name");
6759 			fl_list = SYSCTL_CHILDREN(fl_node);
6760 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx",
6761 				       CTLFLAG_RD,
6762 				       &fl->ifl_pidx, 1, "Producer Index");
6763 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx",
6764 				       CTLFLAG_RD,
6765 				       &fl->ifl_cidx, 1, "Consumer Index");
6766 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits",
6767 				       CTLFLAG_RD,
6768 				       &fl->ifl_credits, 1, "credits available");
6769 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "buf_size",
6770 				       CTLFLAG_RD,
6771 				       &fl->ifl_buf_size, 1, "buffer size");
6772 #if MEMORY_LOGGING
6773 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_enqueued",
6774 					CTLFLAG_RD,
6775 					&fl->ifl_m_enqueued, "mbufs allocated");
6776 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_dequeued",
6777 					CTLFLAG_RD,
6778 					&fl->ifl_m_dequeued, "mbufs freed");
6779 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_enqueued",
6780 					CTLFLAG_RD,
6781 					&fl->ifl_cl_enqueued, "clusters allocated");
6782 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_dequeued",
6783 					CTLFLAG_RD,
6784 					&fl->ifl_cl_dequeued, "clusters freed");
6785 #endif
6786 		}
6787 	}
6788 
6789 }
6790 
6791 void
6792 iflib_request_reset(if_ctx_t ctx)
6793 {
6794 
6795 	STATE_LOCK(ctx);
6796 	ctx->ifc_flags |= IFC_DO_RESET;
6797 	STATE_UNLOCK(ctx);
6798 }
6799 
6800 #ifndef __NO_STRICT_ALIGNMENT
6801 static struct mbuf *
6802 iflib_fixup_rx(struct mbuf *m)
6803 {
6804 	struct mbuf *n;
6805 
6806 	if (m->m_len <= (MCLBYTES - ETHER_HDR_LEN)) {
6807 		bcopy(m->m_data, m->m_data + ETHER_HDR_LEN, m->m_len);
6808 		m->m_data += ETHER_HDR_LEN;
6809 		n = m;
6810 	} else {
6811 		MGETHDR(n, M_NOWAIT, MT_DATA);
6812 		if (n == NULL) {
6813 			m_freem(m);
6814 			return (NULL);
6815 		}
6816 		bcopy(m->m_data, n->m_data, ETHER_HDR_LEN);
6817 		m->m_data += ETHER_HDR_LEN;
6818 		m->m_len -= ETHER_HDR_LEN;
6819 		n->m_len = ETHER_HDR_LEN;
6820 		M_MOVE_PKTHDR(n, m);
6821 		n->m_next = m;
6822 	}
6823 	return (n);
6824 }
6825 #endif
6826 
6827 #ifdef DEBUGNET
6828 static void
6829 iflib_debugnet_init(if_t ifp, int *nrxr, int *ncl, int *clsize)
6830 {
6831 	if_ctx_t ctx;
6832 
6833 	ctx = if_getsoftc(ifp);
6834 	CTX_LOCK(ctx);
6835 	*nrxr = NRXQSETS(ctx);
6836 	*ncl = ctx->ifc_rxqs[0].ifr_fl->ifl_size;
6837 	*clsize = ctx->ifc_rxqs[0].ifr_fl->ifl_buf_size;
6838 	CTX_UNLOCK(ctx);
6839 }
6840 
6841 static void
6842 iflib_debugnet_event(if_t ifp, enum debugnet_ev event)
6843 {
6844 	if_ctx_t ctx;
6845 	if_softc_ctx_t scctx;
6846 	iflib_fl_t fl;
6847 	iflib_rxq_t rxq;
6848 	int i, j;
6849 
6850 	ctx = if_getsoftc(ifp);
6851 	scctx = &ctx->ifc_softc_ctx;
6852 
6853 	switch (event) {
6854 	case DEBUGNET_START:
6855 		for (i = 0; i < scctx->isc_nrxqsets; i++) {
6856 			rxq = &ctx->ifc_rxqs[i];
6857 			for (j = 0; j < rxq->ifr_nfl; j++) {
6858 				fl = rxq->ifr_fl;
6859 				fl->ifl_zone = m_getzone(fl->ifl_buf_size);
6860 			}
6861 		}
6862 		iflib_no_tx_batch = 1;
6863 		break;
6864 	default:
6865 		break;
6866 	}
6867 }
6868 
6869 static int
6870 iflib_debugnet_transmit(if_t ifp, struct mbuf *m)
6871 {
6872 	if_ctx_t ctx;
6873 	iflib_txq_t txq;
6874 	int error;
6875 
6876 	ctx = if_getsoftc(ifp);
6877 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
6878 	    IFF_DRV_RUNNING)
6879 		return (EBUSY);
6880 
6881 	txq = &ctx->ifc_txqs[0];
6882 	error = iflib_encap(txq, &m);
6883 	if (error == 0)
6884 		(void)iflib_txd_db_check(ctx, txq, true, txq->ift_in_use);
6885 	return (error);
6886 }
6887 
6888 static int
6889 iflib_debugnet_poll(if_t ifp, int count)
6890 {
6891 	struct epoch_tracker et;
6892 	if_ctx_t ctx;
6893 	if_softc_ctx_t scctx;
6894 	iflib_txq_t txq;
6895 	int i;
6896 
6897 	ctx = if_getsoftc(ifp);
6898 	scctx = &ctx->ifc_softc_ctx;
6899 
6900 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
6901 	    IFF_DRV_RUNNING)
6902 		return (EBUSY);
6903 
6904 	txq = &ctx->ifc_txqs[0];
6905 	(void)iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
6906 
6907 	NET_EPOCH_ENTER(et);
6908 	for (i = 0; i < scctx->isc_nrxqsets; i++)
6909 		(void)iflib_rxeof(&ctx->ifc_rxqs[i], 16 /* XXX */);
6910 	NET_EPOCH_EXIT(et);
6911 	return (0);
6912 }
6913 #endif /* DEBUGNET */
6914