xref: /freebsd/sys/net/iflib.c (revision 0a5eb308d3006e2f386620f7266b6593b300b987)
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 /*
121  * File organization:
122  *  - private structures
123  *  - iflib private utility functions
124  *  - ifnet functions
125  *  - vlan registry and other exported functions
126  *  - iflib public core functions
127  *
128  *
129  */
130 MALLOC_DEFINE(M_IFLIB, "iflib", "ifnet library");
131 
132 #define	IFLIB_RXEOF_MORE (1U << 0)
133 #define	IFLIB_RXEOF_EMPTY (2U << 0)
134 
135 struct iflib_txq;
136 typedef struct iflib_txq *iflib_txq_t;
137 struct iflib_rxq;
138 typedef struct iflib_rxq *iflib_rxq_t;
139 struct iflib_fl;
140 typedef struct iflib_fl *iflib_fl_t;
141 
142 struct iflib_ctx;
143 
144 static void iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid);
145 static void iflib_timer(void *arg);
146 
147 typedef struct iflib_filter_info {
148 	driver_filter_t *ifi_filter;
149 	void *ifi_filter_arg;
150 	struct grouptask *ifi_task;
151 	void *ifi_ctx;
152 } *iflib_filter_info_t;
153 
154 struct iflib_ctx {
155 	KOBJ_FIELDS;
156 	/*
157 	 * Pointer to hardware driver's softc
158 	 */
159 	void *ifc_softc;
160 	device_t ifc_dev;
161 	if_t ifc_ifp;
162 
163 	cpuset_t ifc_cpus;
164 	if_shared_ctx_t ifc_sctx;
165 	struct if_softc_ctx ifc_softc_ctx;
166 
167 	struct sx ifc_ctx_sx;
168 	struct mtx ifc_state_mtx;
169 
170 	iflib_txq_t ifc_txqs;
171 	iflib_rxq_t ifc_rxqs;
172 	uint32_t ifc_if_flags;
173 	uint32_t ifc_flags;
174 	uint32_t ifc_max_fl_buf_size;
175 	uint32_t ifc_rx_mbuf_sz;
176 
177 	int ifc_link_state;
178 	int ifc_watchdog_events;
179 	struct cdev *ifc_led_dev;
180 	struct resource *ifc_msix_mem;
181 
182 	struct if_irq ifc_legacy_irq;
183 	struct grouptask ifc_admin_task;
184 	struct grouptask ifc_vflr_task;
185 	struct iflib_filter_info ifc_filter_info;
186 	struct ifmedia	ifc_media;
187 	struct ifmedia	*ifc_mediap;
188 
189 	struct sysctl_oid *ifc_sysctl_node;
190 	uint16_t ifc_sysctl_ntxqs;
191 	uint16_t ifc_sysctl_nrxqs;
192 	uint16_t ifc_sysctl_qs_eq_override;
193 	uint16_t ifc_sysctl_rx_budget;
194 	uint16_t ifc_sysctl_tx_abdicate;
195 	uint16_t ifc_sysctl_core_offset;
196 #define	CORE_OFFSET_UNSPECIFIED	0xffff
197 	uint8_t  ifc_sysctl_separate_txrx;
198 
199 	qidx_t ifc_sysctl_ntxds[8];
200 	qidx_t ifc_sysctl_nrxds[8];
201 	struct if_txrx ifc_txrx;
202 #define isc_txd_encap  ifc_txrx.ift_txd_encap
203 #define isc_txd_flush  ifc_txrx.ift_txd_flush
204 #define isc_txd_credits_update  ifc_txrx.ift_txd_credits_update
205 #define isc_rxd_available ifc_txrx.ift_rxd_available
206 #define isc_rxd_pkt_get ifc_txrx.ift_rxd_pkt_get
207 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
208 #define isc_rxd_flush ifc_txrx.ift_rxd_flush
209 #define isc_legacy_intr ifc_txrx.ift_legacy_intr
210 	eventhandler_tag ifc_vlan_attach_event;
211 	eventhandler_tag ifc_vlan_detach_event;
212 	struct ether_addr ifc_mac;
213 };
214 
215 void *
216 iflib_get_softc(if_ctx_t ctx)
217 {
218 
219 	return (ctx->ifc_softc);
220 }
221 
222 device_t
223 iflib_get_dev(if_ctx_t ctx)
224 {
225 
226 	return (ctx->ifc_dev);
227 }
228 
229 if_t
230 iflib_get_ifp(if_ctx_t ctx)
231 {
232 
233 	return (ctx->ifc_ifp);
234 }
235 
236 struct ifmedia *
237 iflib_get_media(if_ctx_t ctx)
238 {
239 
240 	return (ctx->ifc_mediap);
241 }
242 
243 uint32_t
244 iflib_get_flags(if_ctx_t ctx)
245 {
246 	return (ctx->ifc_flags);
247 }
248 
249 void
250 iflib_set_mac(if_ctx_t ctx, uint8_t mac[ETHER_ADDR_LEN])
251 {
252 
253 	bcopy(mac, ctx->ifc_mac.octet, ETHER_ADDR_LEN);
254 }
255 
256 if_softc_ctx_t
257 iflib_get_softc_ctx(if_ctx_t ctx)
258 {
259 
260 	return (&ctx->ifc_softc_ctx);
261 }
262 
263 if_shared_ctx_t
264 iflib_get_sctx(if_ctx_t ctx)
265 {
266 
267 	return (ctx->ifc_sctx);
268 }
269 
270 #define IP_ALIGNED(m) ((((uintptr_t)(m)->m_data) & 0x3) == 0x2)
271 #define CACHE_PTR_INCREMENT (CACHE_LINE_SIZE/sizeof(void*))
272 #define CACHE_PTR_NEXT(ptr) ((void *)(((uintptr_t)(ptr)+CACHE_LINE_SIZE-1) & (CACHE_LINE_SIZE-1)))
273 
274 #define LINK_ACTIVE(ctx) ((ctx)->ifc_link_state == LINK_STATE_UP)
275 #define CTX_IS_VF(ctx) ((ctx)->ifc_sctx->isc_flags & IFLIB_IS_VF)
276 
277 typedef struct iflib_sw_rx_desc_array {
278 	bus_dmamap_t	*ifsd_map;         /* bus_dma maps for packet */
279 	struct mbuf	**ifsd_m;           /* pkthdr mbufs */
280 	caddr_t		*ifsd_cl;          /* direct cluster pointer for rx */
281 	bus_addr_t	*ifsd_ba;          /* bus addr of cluster for rx */
282 } iflib_rxsd_array_t;
283 
284 typedef struct iflib_sw_tx_desc_array {
285 	bus_dmamap_t    *ifsd_map;         /* bus_dma maps for packet */
286 	bus_dmamap_t	*ifsd_tso_map;     /* bus_dma maps for TSO packet */
287 	struct mbuf    **ifsd_m;           /* pkthdr mbufs */
288 } if_txsd_vec_t;
289 
290 /* magic number that should be high enough for any hardware */
291 #define IFLIB_MAX_TX_SEGS		128
292 #define IFLIB_RX_COPY_THRESH		128
293 #define IFLIB_MAX_RX_REFRESH		32
294 /* The minimum descriptors per second before we start coalescing */
295 #define IFLIB_MIN_DESC_SEC		16384
296 #define IFLIB_DEFAULT_TX_UPDATE_FREQ	16
297 #define IFLIB_QUEUE_IDLE		0
298 #define IFLIB_QUEUE_HUNG		1
299 #define IFLIB_QUEUE_WORKING		2
300 /* maximum number of txqs that can share an rx interrupt */
301 #define IFLIB_MAX_TX_SHARED_INTR	4
302 
303 /* this should really scale with ring size - this is a fairly arbitrary value */
304 #define TX_BATCH_SIZE			32
305 
306 #define IFLIB_RESTART_BUDGET		8
307 
308 #define CSUM_OFFLOAD		(CSUM_IP_TSO|CSUM_IP6_TSO|CSUM_IP| \
309 				 CSUM_IP_UDP|CSUM_IP_TCP|CSUM_IP_SCTP| \
310 				 CSUM_IP6_UDP|CSUM_IP6_TCP|CSUM_IP6_SCTP)
311 
312 struct iflib_txq {
313 	qidx_t		ift_in_use;
314 	qidx_t		ift_cidx;
315 	qidx_t		ift_cidx_processed;
316 	qidx_t		ift_pidx;
317 	uint8_t		ift_gen;
318 	uint8_t		ift_br_offset;
319 	uint16_t	ift_npending;
320 	uint16_t	ift_db_pending;
321 	uint16_t	ift_rs_pending;
322 	/* implicit pad */
323 	uint8_t		ift_txd_size[8];
324 	uint64_t	ift_processed;
325 	uint64_t	ift_cleaned;
326 	uint64_t	ift_cleaned_prev;
327 #if MEMORY_LOGGING
328 	uint64_t	ift_enqueued;
329 	uint64_t	ift_dequeued;
330 #endif
331 	uint64_t	ift_no_tx_dma_setup;
332 	uint64_t	ift_no_desc_avail;
333 	uint64_t	ift_mbuf_defrag_failed;
334 	uint64_t	ift_mbuf_defrag;
335 	uint64_t	ift_map_failed;
336 	uint64_t	ift_txd_encap_efbig;
337 	uint64_t	ift_pullups;
338 	uint64_t	ift_last_timer_tick;
339 
340 	struct mtx	ift_mtx;
341 	struct mtx	ift_db_mtx;
342 
343 	/* constant values */
344 	if_ctx_t	ift_ctx;
345 	struct ifmp_ring        *ift_br;
346 	struct grouptask	ift_task;
347 	qidx_t		ift_size;
348 	uint16_t	ift_id;
349 	struct callout	ift_timer;
350 
351 	if_txsd_vec_t	ift_sds;
352 	uint8_t		ift_qstatus;
353 	uint8_t		ift_closed;
354 	uint8_t		ift_update_freq;
355 	struct iflib_filter_info ift_filter_info;
356 	bus_dma_tag_t	ift_buf_tag;
357 	bus_dma_tag_t	ift_tso_buf_tag;
358 	iflib_dma_info_t	ift_ifdi;
359 #define	MTX_NAME_LEN	32
360 	char                    ift_mtx_name[MTX_NAME_LEN];
361 	bus_dma_segment_t	ift_segs[IFLIB_MAX_TX_SEGS]  __aligned(CACHE_LINE_SIZE);
362 #ifdef IFLIB_DIAGNOSTICS
363 	uint64_t ift_cpu_exec_count[256];
364 #endif
365 } __aligned(CACHE_LINE_SIZE);
366 
367 struct iflib_fl {
368 	qidx_t		ifl_cidx;
369 	qidx_t		ifl_pidx;
370 	qidx_t		ifl_credits;
371 	uint8_t		ifl_gen;
372 	uint8_t		ifl_rxd_size;
373 #if MEMORY_LOGGING
374 	uint64_t	ifl_m_enqueued;
375 	uint64_t	ifl_m_dequeued;
376 	uint64_t	ifl_cl_enqueued;
377 	uint64_t	ifl_cl_dequeued;
378 #endif
379 	/* implicit pad */
380 	bitstr_t 	*ifl_rx_bitmap;
381 	qidx_t		ifl_fragidx;
382 	/* constant */
383 	qidx_t		ifl_size;
384 	uint16_t	ifl_buf_size;
385 	uint16_t	ifl_cltype;
386 	uma_zone_t	ifl_zone;
387 	iflib_rxsd_array_t	ifl_sds;
388 	iflib_rxq_t	ifl_rxq;
389 	uint8_t		ifl_id;
390 	bus_dma_tag_t	ifl_buf_tag;
391 	iflib_dma_info_t	ifl_ifdi;
392 	uint64_t	ifl_bus_addrs[IFLIB_MAX_RX_REFRESH] __aligned(CACHE_LINE_SIZE);
393 	qidx_t		ifl_rxd_idxs[IFLIB_MAX_RX_REFRESH];
394 }  __aligned(CACHE_LINE_SIZE);
395 
396 static inline qidx_t
397 get_inuse(int size, qidx_t cidx, qidx_t pidx, uint8_t gen)
398 {
399 	qidx_t used;
400 
401 	if (pidx > cidx)
402 		used = pidx - cidx;
403 	else if (pidx < cidx)
404 		used = size - cidx + pidx;
405 	else if (gen == 0 && pidx == cidx)
406 		used = 0;
407 	else if (gen == 1 && pidx == cidx)
408 		used = size;
409 	else
410 		panic("bad state");
411 
412 	return (used);
413 }
414 
415 #define TXQ_AVAIL(txq) (txq->ift_size - get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx, txq->ift_gen))
416 
417 #define IDXDIFF(head, tail, wrap) \
418 	((head) >= (tail) ? (head) - (tail) : (wrap) - (tail) + (head))
419 
420 struct iflib_rxq {
421 	if_ctx_t	ifr_ctx;
422 	iflib_fl_t	ifr_fl;
423 	uint64_t	ifr_rx_irq;
424 	struct pfil_head	*pfil;
425 	/*
426 	 * If there is a separate completion queue (IFLIB_HAS_RXCQ), this is
427 	 * the completion queue consumer index.  Otherwise it's unused.
428 	 */
429 	qidx_t		ifr_cq_cidx;
430 	uint16_t	ifr_id;
431 	uint8_t		ifr_nfl;
432 	uint8_t		ifr_ntxqirq;
433 	uint8_t		ifr_txqid[IFLIB_MAX_TX_SHARED_INTR];
434 	uint8_t		ifr_fl_offset;
435 	struct lro_ctrl			ifr_lc;
436 	struct grouptask        ifr_task;
437 	struct callout		ifr_watchdog;
438 	struct iflib_filter_info ifr_filter_info;
439 	iflib_dma_info_t		ifr_ifdi;
440 
441 	/* dynamically allocate if any drivers need a value substantially larger than this */
442 	struct if_rxd_frag	ifr_frags[IFLIB_MAX_RX_SEGS] __aligned(CACHE_LINE_SIZE);
443 #ifdef IFLIB_DIAGNOSTICS
444 	uint64_t ifr_cpu_exec_count[256];
445 #endif
446 }  __aligned(CACHE_LINE_SIZE);
447 
448 typedef struct if_rxsd {
449 	caddr_t *ifsd_cl;
450 	iflib_fl_t ifsd_fl;
451 } *if_rxsd_t;
452 
453 /* multiple of word size */
454 #ifdef __LP64__
455 #define PKT_INFO_SIZE	6
456 #define RXD_INFO_SIZE	5
457 #define PKT_TYPE uint64_t
458 #else
459 #define PKT_INFO_SIZE	11
460 #define RXD_INFO_SIZE	8
461 #define PKT_TYPE uint32_t
462 #endif
463 #define PKT_LOOP_BOUND  ((PKT_INFO_SIZE/3)*3)
464 #define RXD_LOOP_BOUND  ((RXD_INFO_SIZE/4)*4)
465 
466 typedef struct if_pkt_info_pad {
467 	PKT_TYPE pkt_val[PKT_INFO_SIZE];
468 } *if_pkt_info_pad_t;
469 typedef struct if_rxd_info_pad {
470 	PKT_TYPE rxd_val[RXD_INFO_SIZE];
471 } *if_rxd_info_pad_t;
472 
473 CTASSERT(sizeof(struct if_pkt_info_pad) == sizeof(struct if_pkt_info));
474 CTASSERT(sizeof(struct if_rxd_info_pad) == sizeof(struct if_rxd_info));
475 
476 
477 static inline void
478 pkt_info_zero(if_pkt_info_t pi)
479 {
480 	if_pkt_info_pad_t pi_pad;
481 
482 	pi_pad = (if_pkt_info_pad_t)pi;
483 	pi_pad->pkt_val[0] = 0; pi_pad->pkt_val[1] = 0; pi_pad->pkt_val[2] = 0;
484 	pi_pad->pkt_val[3] = 0; pi_pad->pkt_val[4] = 0; pi_pad->pkt_val[5] = 0;
485 #ifndef __LP64__
486 	pi_pad->pkt_val[6] = 0; pi_pad->pkt_val[7] = 0; pi_pad->pkt_val[8] = 0;
487 	pi_pad->pkt_val[9] = 0; pi_pad->pkt_val[10] = 0;
488 #endif
489 }
490 
491 static device_method_t iflib_pseudo_methods[] = {
492 	DEVMETHOD(device_attach, noop_attach),
493 	DEVMETHOD(device_detach, iflib_pseudo_detach),
494 	DEVMETHOD_END
495 };
496 
497 driver_t iflib_pseudodriver = {
498 	"iflib_pseudo", iflib_pseudo_methods, sizeof(struct iflib_ctx),
499 };
500 
501 static inline void
502 rxd_info_zero(if_rxd_info_t ri)
503 {
504 	if_rxd_info_pad_t ri_pad;
505 	int i;
506 
507 	ri_pad = (if_rxd_info_pad_t)ri;
508 	for (i = 0; i < RXD_LOOP_BOUND; i += 4) {
509 		ri_pad->rxd_val[i] = 0;
510 		ri_pad->rxd_val[i+1] = 0;
511 		ri_pad->rxd_val[i+2] = 0;
512 		ri_pad->rxd_val[i+3] = 0;
513 	}
514 #ifdef __LP64__
515 	ri_pad->rxd_val[RXD_INFO_SIZE-1] = 0;
516 #endif
517 }
518 
519 /*
520  * Only allow a single packet to take up most 1/nth of the tx ring
521  */
522 #define MAX_SINGLE_PACKET_FRACTION 12
523 #define IF_BAD_DMA (bus_addr_t)-1
524 
525 #define CTX_ACTIVE(ctx) ((if_getdrvflags((ctx)->ifc_ifp) & IFF_DRV_RUNNING))
526 
527 #define CTX_LOCK_INIT(_sc)  sx_init(&(_sc)->ifc_ctx_sx, "iflib ctx lock")
528 #define CTX_LOCK(ctx) sx_xlock(&(ctx)->ifc_ctx_sx)
529 #define CTX_UNLOCK(ctx) sx_xunlock(&(ctx)->ifc_ctx_sx)
530 #define CTX_LOCK_DESTROY(ctx) sx_destroy(&(ctx)->ifc_ctx_sx)
531 
532 #define STATE_LOCK_INIT(_sc, _name)  mtx_init(&(_sc)->ifc_state_mtx, _name, "iflib state lock", MTX_DEF)
533 #define STATE_LOCK(ctx) mtx_lock(&(ctx)->ifc_state_mtx)
534 #define STATE_UNLOCK(ctx) mtx_unlock(&(ctx)->ifc_state_mtx)
535 #define STATE_LOCK_DESTROY(ctx) mtx_destroy(&(ctx)->ifc_state_mtx)
536 
537 #define CALLOUT_LOCK(txq)	mtx_lock(&txq->ift_mtx)
538 #define CALLOUT_UNLOCK(txq) 	mtx_unlock(&txq->ift_mtx)
539 
540 void
541 iflib_set_detach(if_ctx_t ctx)
542 {
543 	STATE_LOCK(ctx);
544 	ctx->ifc_flags |= IFC_IN_DETACH;
545 	STATE_UNLOCK(ctx);
546 }
547 
548 /* Our boot-time initialization hook */
549 static int	iflib_module_event_handler(module_t, int, void *);
550 
551 static moduledata_t iflib_moduledata = {
552 	"iflib",
553 	iflib_module_event_handler,
554 	NULL
555 };
556 
557 DECLARE_MODULE(iflib, iflib_moduledata, SI_SUB_INIT_IF, SI_ORDER_ANY);
558 MODULE_VERSION(iflib, 1);
559 
560 MODULE_DEPEND(iflib, pci, 1, 1, 1);
561 MODULE_DEPEND(iflib, ether, 1, 1, 1);
562 
563 TASKQGROUP_DEFINE(if_io_tqg, mp_ncpus, 1);
564 TASKQGROUP_DEFINE(if_config_tqg, 1, 1);
565 
566 #ifndef IFLIB_DEBUG_COUNTERS
567 #ifdef INVARIANTS
568 #define IFLIB_DEBUG_COUNTERS 1
569 #else
570 #define IFLIB_DEBUG_COUNTERS 0
571 #endif /* !INVARIANTS */
572 #endif
573 
574 static SYSCTL_NODE(_net, OID_AUTO, iflib, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
575     "iflib driver parameters");
576 
577 /*
578  * XXX need to ensure that this can't accidentally cause the head to be moved backwards
579  */
580 static int iflib_min_tx_latency = 0;
581 SYSCTL_INT(_net_iflib, OID_AUTO, min_tx_latency, CTLFLAG_RW,
582 		   &iflib_min_tx_latency, 0, "minimize transmit latency at the possible expense of throughput");
583 static int iflib_no_tx_batch = 0;
584 SYSCTL_INT(_net_iflib, OID_AUTO, no_tx_batch, CTLFLAG_RW,
585 		   &iflib_no_tx_batch, 0, "minimize transmit latency at the possible expense of throughput");
586 
587 
588 #if IFLIB_DEBUG_COUNTERS
589 
590 static int iflib_tx_seen;
591 static int iflib_tx_sent;
592 static int iflib_tx_encap;
593 static int iflib_rx_allocs;
594 static int iflib_fl_refills;
595 static int iflib_fl_refills_large;
596 static int iflib_tx_frees;
597 
598 SYSCTL_INT(_net_iflib, OID_AUTO, tx_seen, CTLFLAG_RD,
599 		   &iflib_tx_seen, 0, "# TX mbufs seen");
600 SYSCTL_INT(_net_iflib, OID_AUTO, tx_sent, CTLFLAG_RD,
601 		   &iflib_tx_sent, 0, "# TX mbufs sent");
602 SYSCTL_INT(_net_iflib, OID_AUTO, tx_encap, CTLFLAG_RD,
603 		   &iflib_tx_encap, 0, "# TX mbufs encapped");
604 SYSCTL_INT(_net_iflib, OID_AUTO, tx_frees, CTLFLAG_RD,
605 		   &iflib_tx_frees, 0, "# TX frees");
606 SYSCTL_INT(_net_iflib, OID_AUTO, rx_allocs, CTLFLAG_RD,
607 		   &iflib_rx_allocs, 0, "# RX allocations");
608 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills, CTLFLAG_RD,
609 		   &iflib_fl_refills, 0, "# refills");
610 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills_large, CTLFLAG_RD,
611 		   &iflib_fl_refills_large, 0, "# large refills");
612 
613 
614 static int iflib_txq_drain_flushing;
615 static int iflib_txq_drain_oactive;
616 static int iflib_txq_drain_notready;
617 
618 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_flushing, CTLFLAG_RD,
619 		   &iflib_txq_drain_flushing, 0, "# drain flushes");
620 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_oactive, CTLFLAG_RD,
621 		   &iflib_txq_drain_oactive, 0, "# drain oactives");
622 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_notready, CTLFLAG_RD,
623 		   &iflib_txq_drain_notready, 0, "# drain notready");
624 
625 
626 static int iflib_encap_load_mbuf_fail;
627 static int iflib_encap_pad_mbuf_fail;
628 static int iflib_encap_txq_avail_fail;
629 static int iflib_encap_txd_encap_fail;
630 
631 SYSCTL_INT(_net_iflib, OID_AUTO, encap_load_mbuf_fail, CTLFLAG_RD,
632 		   &iflib_encap_load_mbuf_fail, 0, "# busdma load failures");
633 SYSCTL_INT(_net_iflib, OID_AUTO, encap_pad_mbuf_fail, CTLFLAG_RD,
634 		   &iflib_encap_pad_mbuf_fail, 0, "# runt frame pad failures");
635 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txq_avail_fail, CTLFLAG_RD,
636 		   &iflib_encap_txq_avail_fail, 0, "# txq avail failures");
637 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txd_encap_fail, CTLFLAG_RD,
638 		   &iflib_encap_txd_encap_fail, 0, "# driver encap failures");
639 
640 static int iflib_task_fn_rxs;
641 static int iflib_rx_intr_enables;
642 static int iflib_fast_intrs;
643 static int iflib_rx_unavail;
644 static int iflib_rx_ctx_inactive;
645 static int iflib_rx_if_input;
646 static int iflib_rxd_flush;
647 
648 static int iflib_verbose_debug;
649 
650 SYSCTL_INT(_net_iflib, OID_AUTO, task_fn_rx, CTLFLAG_RD,
651 		   &iflib_task_fn_rxs, 0, "# task_fn_rx calls");
652 SYSCTL_INT(_net_iflib, OID_AUTO, rx_intr_enables, CTLFLAG_RD,
653 		   &iflib_rx_intr_enables, 0, "# RX intr enables");
654 SYSCTL_INT(_net_iflib, OID_AUTO, fast_intrs, CTLFLAG_RD,
655 		   &iflib_fast_intrs, 0, "# fast_intr calls");
656 SYSCTL_INT(_net_iflib, OID_AUTO, rx_unavail, CTLFLAG_RD,
657 		   &iflib_rx_unavail, 0, "# times rxeof called with no available data");
658 SYSCTL_INT(_net_iflib, OID_AUTO, rx_ctx_inactive, CTLFLAG_RD,
659 		   &iflib_rx_ctx_inactive, 0, "# times rxeof called with inactive context");
660 SYSCTL_INT(_net_iflib, OID_AUTO, rx_if_input, CTLFLAG_RD,
661 		   &iflib_rx_if_input, 0, "# times rxeof called if_input");
662 SYSCTL_INT(_net_iflib, OID_AUTO, rxd_flush, CTLFLAG_RD,
663 	         &iflib_rxd_flush, 0, "# times rxd_flush called");
664 SYSCTL_INT(_net_iflib, OID_AUTO, verbose_debug, CTLFLAG_RW,
665 		   &iflib_verbose_debug, 0, "enable verbose debugging");
666 
667 #define DBG_COUNTER_INC(name) atomic_add_int(&(iflib_ ## name), 1)
668 static void
669 iflib_debug_reset(void)
670 {
671 	iflib_tx_seen = iflib_tx_sent = iflib_tx_encap = iflib_rx_allocs =
672 		iflib_fl_refills = iflib_fl_refills_large = iflib_tx_frees =
673 		iflib_txq_drain_flushing = iflib_txq_drain_oactive =
674 		iflib_txq_drain_notready =
675 		iflib_encap_load_mbuf_fail = iflib_encap_pad_mbuf_fail =
676 		iflib_encap_txq_avail_fail = iflib_encap_txd_encap_fail =
677 		iflib_task_fn_rxs = iflib_rx_intr_enables = iflib_fast_intrs =
678 		iflib_rx_unavail =
679 		iflib_rx_ctx_inactive = iflib_rx_if_input =
680 		iflib_rxd_flush = 0;
681 }
682 
683 #else
684 #define DBG_COUNTER_INC(name)
685 static void iflib_debug_reset(void) {}
686 #endif
687 
688 #define IFLIB_DEBUG 0
689 
690 static void iflib_tx_structures_free(if_ctx_t ctx);
691 static void iflib_rx_structures_free(if_ctx_t ctx);
692 static int iflib_queues_alloc(if_ctx_t ctx);
693 static int iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq);
694 static int iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget);
695 static int iflib_qset_structures_setup(if_ctx_t ctx);
696 static int iflib_msix_init(if_ctx_t ctx);
697 static int iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filterarg, int *rid, const char *str);
698 static void iflib_txq_check_drain(iflib_txq_t txq, int budget);
699 static uint32_t iflib_txq_can_drain(struct ifmp_ring *);
700 #ifdef ALTQ
701 static void iflib_altq_if_start(if_t ifp);
702 static int iflib_altq_if_transmit(if_t ifp, struct mbuf *m);
703 #endif
704 static int iflib_register(if_ctx_t);
705 static void iflib_deregister(if_ctx_t);
706 static void iflib_unregister_vlan_handlers(if_ctx_t ctx);
707 static uint16_t iflib_get_mbuf_size_for(unsigned int size);
708 static void iflib_init_locked(if_ctx_t ctx);
709 static void iflib_add_device_sysctl_pre(if_ctx_t ctx);
710 static void iflib_add_device_sysctl_post(if_ctx_t ctx);
711 static void iflib_ifmp_purge(iflib_txq_t txq);
712 static void _iflib_pre_assert(if_softc_ctx_t scctx);
713 static void iflib_if_init_locked(if_ctx_t ctx);
714 static void iflib_free_intr_mem(if_ctx_t ctx);
715 #ifndef __NO_STRICT_ALIGNMENT
716 static struct mbuf * iflib_fixup_rx(struct mbuf *m);
717 #endif
718 
719 static SLIST_HEAD(cpu_offset_list, cpu_offset) cpu_offsets =
720     SLIST_HEAD_INITIALIZER(cpu_offsets);
721 struct cpu_offset {
722 	SLIST_ENTRY(cpu_offset) entries;
723 	cpuset_t	set;
724 	unsigned int	refcount;
725 	uint16_t	offset;
726 };
727 static struct mtx cpu_offset_mtx;
728 MTX_SYSINIT(iflib_cpu_offset, &cpu_offset_mtx, "iflib_cpu_offset lock",
729     MTX_DEF);
730 
731 DEBUGNET_DEFINE(iflib);
732 
733 static int
734 iflib_num_rx_descs(if_ctx_t ctx)
735 {
736 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
737 	if_shared_ctx_t sctx = ctx->ifc_sctx;
738 	uint16_t first_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
739 
740 	return scctx->isc_nrxd[first_rxq];
741 }
742 
743 static int
744 iflib_num_tx_descs(if_ctx_t ctx)
745 {
746 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
747 	if_shared_ctx_t sctx = ctx->ifc_sctx;
748 	uint16_t first_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
749 
750 	return scctx->isc_ntxd[first_txq];
751 }
752 
753 #ifdef DEV_NETMAP
754 #include <sys/selinfo.h>
755 #include <net/netmap.h>
756 #include <dev/netmap/netmap_kern.h>
757 
758 MODULE_DEPEND(iflib, netmap, 1, 1, 1);
759 
760 static int netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, bool init);
761 
762 /*
763  * device-specific sysctl variables:
764  *
765  * iflib_crcstrip: 0: keep CRC in rx frames (default), 1: strip it.
766  *	During regular operations the CRC is stripped, but on some
767  *	hardware reception of frames not multiple of 64 is slower,
768  *	so using crcstrip=0 helps in benchmarks.
769  *
770  * iflib_rx_miss, iflib_rx_miss_bufs:
771  *	count packets that might be missed due to lost interrupts.
772  */
773 SYSCTL_DECL(_dev_netmap);
774 /*
775  * The xl driver by default strips CRCs and we do not override it.
776  */
777 
778 int iflib_crcstrip = 1;
779 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_crcstrip,
780     CTLFLAG_RW, &iflib_crcstrip, 1, "strip CRC on RX frames");
781 
782 int iflib_rx_miss, iflib_rx_miss_bufs;
783 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss,
784     CTLFLAG_RW, &iflib_rx_miss, 0, "potentially missed RX intr");
785 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss_bufs,
786     CTLFLAG_RW, &iflib_rx_miss_bufs, 0, "potentially missed RX intr bufs");
787 
788 /*
789  * Register/unregister. We are already under netmap lock.
790  * Only called on the first register or the last unregister.
791  */
792 static int
793 iflib_netmap_register(struct netmap_adapter *na, int onoff)
794 {
795 	if_t ifp = na->ifp;
796 	if_ctx_t ctx = ifp->if_softc;
797 	int status;
798 
799 	CTX_LOCK(ctx);
800 	IFDI_INTR_DISABLE(ctx);
801 
802 	/* Tell the stack that the interface is no longer active */
803 	ifp->if_drv_flags &= ~(IFF_DRV_RUNNING | IFF_DRV_OACTIVE);
804 
805 	if (!CTX_IS_VF(ctx))
806 		IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip);
807 
808 	iflib_stop(ctx);
809 
810 	/*
811 	 * Enable (or disable) netmap flags, and intercept (or restore)
812 	 * ifp->if_transmit. This is done once the device has been stopped
813 	 * to prevent race conditions.
814 	 */
815 	if (onoff) {
816 		nm_set_native_flags(na);
817 	} else {
818 		nm_clear_native_flags(na);
819 	}
820 
821 	iflib_init_locked(ctx);
822 	IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip); // XXX why twice ?
823 	status = ifp->if_drv_flags & IFF_DRV_RUNNING ? 0 : 1;
824 	if (status)
825 		nm_clear_native_flags(na);
826 	CTX_UNLOCK(ctx);
827 	return (status);
828 }
829 
830 static int
831 netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, bool init)
832 {
833 	struct netmap_adapter *na = kring->na;
834 	u_int const lim = kring->nkr_num_slots - 1;
835 	u_int head = kring->rhead;
836 	u_int nm_i = kring->nr_hwcur;
837 	struct netmap_ring *ring = kring->ring;
838 	bus_dmamap_t *map;
839 	struct if_rxd_update iru;
840 	if_ctx_t ctx = rxq->ifr_ctx;
841 	iflib_fl_t fl = &rxq->ifr_fl[0];
842 	u_int nic_i_first, nic_i;
843 	int i;
844 #if IFLIB_DEBUG_COUNTERS
845 	int rf_count = 0;
846 #endif
847 
848 	/*
849 	 * Netmap requires that we leave (at least) one free slot
850 	 * in the ring, so that it can distinguish between an empty
851 	 * ring (nr_hwcur == nr_hwtail, i.e. all the buffers owned by the
852 	 * user) and a full ring (nr_hwtail == (nr_hwcur - 1) mod N, i.e.
853 	 * all the buffers owned by the kernel).
854 	 * We thus set head (the refill limit) to nr_hwcur - 1
855 	 * at initialization. The rest of the code will then make sure
856 	 * than nr_hwtail never overcomes nr_hwcur.
857 	 */
858 	if (__predict_false(init)) {
859 		head = nm_prev(nm_i, lim);
860 	} else if (nm_i == head) {
861 		/* Nothing to do. We can leave early. */
862 		return (0);
863 	}
864 
865 	iru_init(&iru, rxq, 0 /* flid */);
866 	map = fl->ifl_sds.ifsd_map;
867 	nic_i = netmap_idx_k2n(kring, nm_i);
868 	DBG_COUNTER_INC(fl_refills);
869 	while (nm_i != head) {
870 #if IFLIB_DEBUG_COUNTERS
871 		if (++rf_count == 9)
872 			DBG_COUNTER_INC(fl_refills_large);
873 #endif
874 		nic_i_first = nic_i;
875 		for (i = 0; i < IFLIB_MAX_RX_REFRESH && nm_i != head; i++) {
876 			struct netmap_slot *slot = &ring->slot[nm_i];
877 			void *addr = PNMB(na, slot, &fl->ifl_bus_addrs[i]);
878 
879 			MPASS(i < IFLIB_MAX_RX_REFRESH);
880 
881 			if (addr == NETMAP_BUF_BASE(na)) /* bad buf */
882 			        return netmap_ring_reinit(kring);
883 
884 			fl->ifl_rxd_idxs[i] = nic_i;
885 
886 			if (__predict_false(init)) {
887 				netmap_load_map(na, fl->ifl_buf_tag,
888 				    map[nic_i], addr);
889 			} else if (slot->flags & NS_BUF_CHANGED) {
890 				/* buffer has changed, reload map */
891 				netmap_reload_map(na, fl->ifl_buf_tag,
892 				    map[nic_i], addr);
893 			}
894 			bus_dmamap_sync(fl->ifl_buf_tag, map[nic_i],
895 			    BUS_DMASYNC_PREREAD);
896 			slot->flags &= ~NS_BUF_CHANGED;
897 
898 			nm_i = nm_next(nm_i, lim);
899 			nic_i = nm_next(nic_i, lim);
900 		}
901 
902 		iru.iru_pidx = nic_i_first;
903 		iru.iru_count = i;
904 		ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
905 	}
906 	kring->nr_hwcur = head;
907 
908 	/* The pidx argument of isc_rxd_flush() is the index of the last valid
909 	 * slot in the free list ring. We need therefore to decrement nic_i,
910 	 * similarly to what happens in iflib_fl_refill() for ifl_pidx. */
911 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
912 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
913 	ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id,
914 	    nm_prev(nic_i, lim));
915 	DBG_COUNTER_INC(rxd_flush);
916 
917 	return (0);
918 }
919 
920 /*
921  * Reconcile kernel and user view of the transmit ring.
922  *
923  * All information is in the kring.
924  * Userspace wants to send packets up to the one before kring->rhead,
925  * kernel knows kring->nr_hwcur is the first unsent packet.
926  *
927  * Here we push packets out (as many as possible), and possibly
928  * reclaim buffers from previously completed transmission.
929  *
930  * The caller (netmap) guarantees that there is only one instance
931  * running at any time. Any interference with other driver
932  * methods should be handled by the individual drivers.
933  */
934 static int
935 iflib_netmap_txsync(struct netmap_kring *kring, int flags)
936 {
937 	struct netmap_adapter *na = kring->na;
938 	if_t ifp = na->ifp;
939 	struct netmap_ring *ring = kring->ring;
940 	u_int nm_i;	/* index into the netmap kring */
941 	u_int nic_i;	/* index into the NIC ring */
942 	u_int n;
943 	u_int const lim = kring->nkr_num_slots - 1;
944 	u_int const head = kring->rhead;
945 	struct if_pkt_info pi;
946 
947 	/*
948 	 * interrupts on every tx packet are expensive so request
949 	 * them every half ring, or where NS_REPORT is set
950 	 */
951 	u_int report_frequency = kring->nkr_num_slots >> 1;
952 	/* device-specific */
953 	if_ctx_t ctx = ifp->if_softc;
954 	iflib_txq_t txq = &ctx->ifc_txqs[kring->ring_id];
955 
956 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
957 	    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
958 
959 	/*
960 	 * First part: process new packets to send.
961 	 * nm_i is the current index in the netmap kring,
962 	 * nic_i is the corresponding index in the NIC ring.
963 	 *
964 	 * If we have packets to send (nm_i != head)
965 	 * iterate over the netmap ring, fetch length and update
966 	 * the corresponding slot in the NIC ring. Some drivers also
967 	 * need to update the buffer's physical address in the NIC slot
968 	 * even NS_BUF_CHANGED is not set (PNMB computes the addresses).
969 	 *
970 	 * The netmap_reload_map() calls is especially expensive,
971 	 * even when (as in this case) the tag is 0, so do only
972 	 * when the buffer has actually changed.
973 	 *
974 	 * If possible do not set the report/intr bit on all slots,
975 	 * but only a few times per ring or when NS_REPORT is set.
976 	 *
977 	 * Finally, on 10G and faster drivers, it might be useful
978 	 * to prefetch the next slot and txr entry.
979 	 */
980 
981 	nm_i = kring->nr_hwcur;
982 	if (nm_i != head) {	/* we have new packets to send */
983 		pkt_info_zero(&pi);
984 		pi.ipi_segs = txq->ift_segs;
985 		pi.ipi_qsidx = kring->ring_id;
986 		nic_i = netmap_idx_k2n(kring, nm_i);
987 
988 		__builtin_prefetch(&ring->slot[nm_i]);
989 		__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i]);
990 		__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i]);
991 
992 		for (n = 0; nm_i != head; n++) {
993 			struct netmap_slot *slot = &ring->slot[nm_i];
994 			u_int len = slot->len;
995 			uint64_t paddr;
996 			void *addr = PNMB(na, slot, &paddr);
997 			int flags = (slot->flags & NS_REPORT ||
998 				nic_i == 0 || nic_i == report_frequency) ?
999 				IPI_TX_INTR : 0;
1000 
1001 			/* device-specific */
1002 			pi.ipi_len = len;
1003 			pi.ipi_segs[0].ds_addr = paddr;
1004 			pi.ipi_segs[0].ds_len = len;
1005 			pi.ipi_nsegs = 1;
1006 			pi.ipi_ndescs = 0;
1007 			pi.ipi_pidx = nic_i;
1008 			pi.ipi_flags = flags;
1009 
1010 			/* Fill the slot in the NIC ring. */
1011 			ctx->isc_txd_encap(ctx->ifc_softc, &pi);
1012 			DBG_COUNTER_INC(tx_encap);
1013 
1014 			/* prefetch for next round */
1015 			__builtin_prefetch(&ring->slot[nm_i + 1]);
1016 			__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i + 1]);
1017 			__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i + 1]);
1018 
1019 			NM_CHECK_ADDR_LEN(na, addr, len);
1020 
1021 			if (slot->flags & NS_BUF_CHANGED) {
1022 				/* buffer has changed, reload map */
1023 				netmap_reload_map(na, txq->ift_buf_tag,
1024 				    txq->ift_sds.ifsd_map[nic_i], addr);
1025 			}
1026 			/* make sure changes to the buffer are synced */
1027 			bus_dmamap_sync(txq->ift_buf_tag,
1028 			    txq->ift_sds.ifsd_map[nic_i],
1029 			    BUS_DMASYNC_PREWRITE);
1030 
1031 			slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED);
1032 			nm_i = nm_next(nm_i, lim);
1033 			nic_i = nm_next(nic_i, lim);
1034 		}
1035 		kring->nr_hwcur = nm_i;
1036 
1037 		/* synchronize the NIC ring */
1038 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1039 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1040 
1041 		/* (re)start the tx unit up to slot nic_i (excluded) */
1042 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, nic_i);
1043 	}
1044 
1045 	/*
1046 	 * Second part: reclaim buffers for completed transmissions.
1047 	 *
1048 	 * If there are unclaimed buffers, attempt to reclaim them.
1049 	 * If none are reclaimed, and TX IRQs are not in use, do an initial
1050 	 * minimal delay, then trigger the tx handler which will spin in the
1051 	 * group task queue.
1052 	 */
1053 	if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1054 		if (iflib_tx_credits_update(ctx, txq)) {
1055 			/* some tx completed, increment avail */
1056 			nic_i = txq->ift_cidx_processed;
1057 			kring->nr_hwtail = nm_prev(netmap_idx_n2k(kring, nic_i), lim);
1058 		}
1059 	}
1060 	if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ))
1061 		if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1062 			callout_reset_on(&txq->ift_timer, hz < 2000 ? 1 : hz / 1000,
1063 			    iflib_timer, txq, txq->ift_timer.c_cpu);
1064 	}
1065 	return (0);
1066 }
1067 
1068 /*
1069  * Reconcile kernel and user view of the receive ring.
1070  * Same as for the txsync, this routine must be efficient.
1071  * The caller guarantees a single invocations, but races against
1072  * the rest of the driver should be handled here.
1073  *
1074  * On call, kring->rhead is the first packet that userspace wants
1075  * to keep, and kring->rcur is the wakeup point.
1076  * The kernel has previously reported packets up to kring->rtail.
1077  *
1078  * If (flags & NAF_FORCE_READ) also check for incoming packets irrespective
1079  * of whether or not we received an interrupt.
1080  */
1081 static int
1082 iflib_netmap_rxsync(struct netmap_kring *kring, int flags)
1083 {
1084 	struct netmap_adapter *na = kring->na;
1085 	struct netmap_ring *ring = kring->ring;
1086 	if_t ifp = na->ifp;
1087 	uint32_t nm_i;	/* index into the netmap ring */
1088 	uint32_t nic_i;	/* index into the NIC ring */
1089 	u_int n;
1090 	u_int const lim = kring->nkr_num_slots - 1;
1091 	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
1092 
1093 	if_ctx_t ctx = ifp->if_softc;
1094 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1095 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1096 	iflib_rxq_t rxq = &ctx->ifc_rxqs[kring->ring_id];
1097 	iflib_fl_t fl = &rxq->ifr_fl[0];
1098 	struct if_rxd_info ri;
1099 	qidx_t *cidxp;
1100 
1101 	/*
1102 	 * netmap only uses free list 0, to avoid out of order consumption
1103 	 * of receive buffers
1104 	 */
1105 
1106 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
1107 	    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1108 
1109 	/*
1110 	 * First part: import newly received packets.
1111 	 *
1112 	 * nm_i is the index of the next free slot in the netmap ring,
1113 	 * nic_i is the index of the next received packet in the NIC ring
1114 	 * (or in the free list 0 if IFLIB_HAS_RXCQ is set), and they may
1115 	 * differ in case if_init() has been called while
1116 	 * in netmap mode. For the receive ring we have
1117 	 *
1118 	 *	nic_i = fl->ifl_cidx;
1119 	 *	nm_i = kring->nr_hwtail (previous)
1120 	 * and
1121 	 *	nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1122 	 *
1123 	 * fl->ifl_cidx is set to 0 on a ring reinit
1124 	 */
1125 	if (netmap_no_pendintr || force_update) {
1126 		uint32_t hwtail_lim = nm_prev(kring->nr_hwcur, lim);
1127 		bool have_rxcq = sctx->isc_flags & IFLIB_HAS_RXCQ;
1128 		int crclen = iflib_crcstrip ? 0 : 4;
1129 		int error, avail;
1130 
1131 		/*
1132 		 * For the free list consumer index, we use the same
1133 		 * logic as in iflib_rxeof().
1134 		 */
1135 		if (have_rxcq)
1136 			cidxp = &rxq->ifr_cq_cidx;
1137 		else
1138 			cidxp = &fl->ifl_cidx;
1139 		avail = ctx->isc_rxd_available(ctx->ifc_softc,
1140 		    rxq->ifr_id, *cidxp, USHRT_MAX);
1141 
1142 		nic_i = fl->ifl_cidx;
1143 		nm_i = netmap_idx_n2k(kring, nic_i);
1144 		MPASS(nm_i == kring->nr_hwtail);
1145 		for (n = 0; avail > 0 && nm_i != hwtail_lim; n++, avail--) {
1146 			rxd_info_zero(&ri);
1147 			ri.iri_frags = rxq->ifr_frags;
1148 			ri.iri_qsidx = kring->ring_id;
1149 			ri.iri_ifp = ctx->ifc_ifp;
1150 			ri.iri_cidx = *cidxp;
1151 
1152 			error = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
1153 			ring->slot[nm_i].len = error ? 0 : ri.iri_len - crclen;
1154 			ring->slot[nm_i].flags = 0;
1155 			if (have_rxcq) {
1156 				*cidxp = ri.iri_cidx;
1157 				while (*cidxp >= scctx->isc_nrxd[0])
1158 					*cidxp -= scctx->isc_nrxd[0];
1159 			}
1160 			bus_dmamap_sync(fl->ifl_buf_tag,
1161 			    fl->ifl_sds.ifsd_map[nic_i], BUS_DMASYNC_POSTREAD);
1162 			nm_i = nm_next(nm_i, lim);
1163 			fl->ifl_cidx = nic_i = nm_next(nic_i, lim);
1164 		}
1165 		if (n) { /* update the state variables */
1166 			if (netmap_no_pendintr && !force_update) {
1167 				/* diagnostics */
1168 				iflib_rx_miss ++;
1169 				iflib_rx_miss_bufs += n;
1170 			}
1171 			kring->nr_hwtail = nm_i;
1172 		}
1173 		kring->nr_kflags &= ~NKR_PENDINTR;
1174 	}
1175 	/*
1176 	 * Second part: skip past packets that userspace has released.
1177 	 * (kring->nr_hwcur to head excluded),
1178 	 * and make the buffers available for reception.
1179 	 * As usual nm_i is the index in the netmap ring,
1180 	 * nic_i is the index in the NIC ring, and
1181 	 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1182 	 */
1183 	netmap_fl_refill(rxq, kring, false);
1184 
1185 	return (0);
1186 }
1187 
1188 static void
1189 iflib_netmap_intr(struct netmap_adapter *na, int onoff)
1190 {
1191 	if_ctx_t ctx = na->ifp->if_softc;
1192 
1193 	CTX_LOCK(ctx);
1194 	if (onoff) {
1195 		IFDI_INTR_ENABLE(ctx);
1196 	} else {
1197 		IFDI_INTR_DISABLE(ctx);
1198 	}
1199 	CTX_UNLOCK(ctx);
1200 }
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 		/*
1238 		 * In netmap mode, set the map for the packet buffer.
1239 		 * NOTE: Some drivers (not this one) also need to set
1240 		 * the physical buffer address in the NIC ring.
1241 		 * netmap_idx_n2k() maps a nic index, i, into the corresponding
1242 		 * netmap slot index, si
1243 		 */
1244 		int si = netmap_idx_n2k(na->tx_rings[txq->ift_id], i);
1245 		netmap_load_map(na, txq->ift_buf_tag, txq->ift_sds.ifsd_map[i],
1246 		    NMB(na, slot + si));
1247 	}
1248 	return (1);
1249 }
1250 
1251 static int
1252 iflib_netmap_rxq_init(if_ctx_t ctx, iflib_rxq_t rxq)
1253 {
1254 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1255 	struct netmap_kring *kring;
1256 	struct netmap_slot *slot;
1257 
1258 	slot = netmap_reset(na, NR_RX, rxq->ifr_id, 0);
1259 	if (slot == NULL)
1260 		return (0);
1261 	kring = na->rx_rings[rxq->ifr_id];
1262 	netmap_fl_refill(rxq, kring, true);
1263 	return (1);
1264 }
1265 
1266 static void
1267 iflib_netmap_timer_adjust(if_ctx_t ctx, iflib_txq_t txq, uint32_t *reset_on)
1268 {
1269 	struct netmap_kring *kring;
1270 	uint16_t txqid;
1271 
1272 	txqid = txq->ift_id;
1273 	kring = netmap_kring_on(NA(ctx->ifc_ifp), txqid, NR_TX);
1274 	if (kring == NULL)
1275 		return;
1276 
1277 	if (kring->nr_hwcur != nm_next(kring->nr_hwtail, kring->nkr_num_slots - 1)) {
1278 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1279 		    BUS_DMASYNC_POSTREAD);
1280 		if (ctx->isc_txd_credits_update(ctx->ifc_softc, txqid, false))
1281 			netmap_tx_irq(ctx->ifc_ifp, txqid);
1282 		if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ)) {
1283 			if (hz < 2000)
1284 				*reset_on = 1;
1285 			else
1286 				*reset_on = hz / 1000;
1287 		}
1288 	}
1289 }
1290 
1291 #define iflib_netmap_detach(ifp) netmap_detach(ifp)
1292 
1293 #else
1294 #define iflib_netmap_txq_init(ctx, txq) (0)
1295 #define iflib_netmap_rxq_init(ctx, rxq) (0)
1296 #define iflib_netmap_detach(ifp)
1297 
1298 #define iflib_netmap_attach(ctx) (0)
1299 #define netmap_rx_irq(ifp, qid, budget) (0)
1300 #define netmap_tx_irq(ifp, qid) do {} while (0)
1301 #define iflib_netmap_timer_adjust(ctx, txq, reset_on)
1302 #endif
1303 
1304 #if defined(__i386__) || defined(__amd64__)
1305 static __inline void
1306 prefetch(void *x)
1307 {
1308 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1309 }
1310 static __inline void
1311 prefetch2cachelines(void *x)
1312 {
1313 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1314 #if (CACHE_LINE_SIZE < 128)
1315 	__asm volatile("prefetcht0 %0" :: "m" (*(((unsigned long *)x)+CACHE_LINE_SIZE/(sizeof(unsigned long)))));
1316 #endif
1317 }
1318 #else
1319 #define prefetch(x)
1320 #define prefetch2cachelines(x)
1321 #endif
1322 
1323 static void
1324 iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid)
1325 {
1326 	iflib_fl_t fl;
1327 
1328 	fl = &rxq->ifr_fl[flid];
1329 	iru->iru_paddrs = fl->ifl_bus_addrs;
1330 	iru->iru_idxs = fl->ifl_rxd_idxs;
1331 	iru->iru_qsidx = rxq->ifr_id;
1332 	iru->iru_buf_size = fl->ifl_buf_size;
1333 	iru->iru_flidx = fl->ifl_id;
1334 }
1335 
1336 static void
1337 _iflib_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int err)
1338 {
1339 	if (err)
1340 		return;
1341 	*(bus_addr_t *) arg = segs[0].ds_addr;
1342 }
1343 
1344 int
1345 iflib_dma_alloc_align(if_ctx_t ctx, int size, int align, iflib_dma_info_t dma, int mapflags)
1346 {
1347 	int err;
1348 	device_t dev = ctx->ifc_dev;
1349 
1350 	err = bus_dma_tag_create(bus_get_dma_tag(dev),	/* parent */
1351 				align, 0,		/* alignment, bounds */
1352 				BUS_SPACE_MAXADDR,	/* lowaddr */
1353 				BUS_SPACE_MAXADDR,	/* highaddr */
1354 				NULL, NULL,		/* filter, filterarg */
1355 				size,			/* maxsize */
1356 				1,			/* nsegments */
1357 				size,			/* maxsegsize */
1358 				BUS_DMA_ALLOCNOW,	/* flags */
1359 				NULL,			/* lockfunc */
1360 				NULL,			/* lockarg */
1361 				&dma->idi_tag);
1362 	if (err) {
1363 		device_printf(dev,
1364 		    "%s: bus_dma_tag_create failed: %d\n",
1365 		    __func__, err);
1366 		goto fail_0;
1367 	}
1368 
1369 	err = bus_dmamem_alloc(dma->idi_tag, (void**) &dma->idi_vaddr,
1370 	    BUS_DMA_NOWAIT | BUS_DMA_COHERENT | BUS_DMA_ZERO, &dma->idi_map);
1371 	if (err) {
1372 		device_printf(dev,
1373 		    "%s: bus_dmamem_alloc(%ju) failed: %d\n",
1374 		    __func__, (uintmax_t)size, err);
1375 		goto fail_1;
1376 	}
1377 
1378 	dma->idi_paddr = IF_BAD_DMA;
1379 	err = bus_dmamap_load(dma->idi_tag, dma->idi_map, dma->idi_vaddr,
1380 	    size, _iflib_dmamap_cb, &dma->idi_paddr, mapflags | BUS_DMA_NOWAIT);
1381 	if (err || dma->idi_paddr == IF_BAD_DMA) {
1382 		device_printf(dev,
1383 		    "%s: bus_dmamap_load failed: %d\n",
1384 		    __func__, err);
1385 		goto fail_2;
1386 	}
1387 
1388 	dma->idi_size = size;
1389 	return (0);
1390 
1391 fail_2:
1392 	bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1393 fail_1:
1394 	bus_dma_tag_destroy(dma->idi_tag);
1395 fail_0:
1396 	dma->idi_tag = NULL;
1397 
1398 	return (err);
1399 }
1400 
1401 int
1402 iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags)
1403 {
1404 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1405 
1406 	KASSERT(sctx->isc_q_align != 0, ("alignment value not initialized"));
1407 
1408 	return (iflib_dma_alloc_align(ctx, size, sctx->isc_q_align, dma, mapflags));
1409 }
1410 
1411 int
1412 iflib_dma_alloc_multi(if_ctx_t ctx, int *sizes, iflib_dma_info_t *dmalist, int mapflags, int count)
1413 {
1414 	int i, err;
1415 	iflib_dma_info_t *dmaiter;
1416 
1417 	dmaiter = dmalist;
1418 	for (i = 0; i < count; i++, dmaiter++) {
1419 		if ((err = iflib_dma_alloc(ctx, sizes[i], *dmaiter, mapflags)) != 0)
1420 			break;
1421 	}
1422 	if (err)
1423 		iflib_dma_free_multi(dmalist, i);
1424 	return (err);
1425 }
1426 
1427 void
1428 iflib_dma_free(iflib_dma_info_t dma)
1429 {
1430 	if (dma->idi_tag == NULL)
1431 		return;
1432 	if (dma->idi_paddr != IF_BAD_DMA) {
1433 		bus_dmamap_sync(dma->idi_tag, dma->idi_map,
1434 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1435 		bus_dmamap_unload(dma->idi_tag, dma->idi_map);
1436 		dma->idi_paddr = IF_BAD_DMA;
1437 	}
1438 	if (dma->idi_vaddr != NULL) {
1439 		bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1440 		dma->idi_vaddr = NULL;
1441 	}
1442 	bus_dma_tag_destroy(dma->idi_tag);
1443 	dma->idi_tag = NULL;
1444 }
1445 
1446 void
1447 iflib_dma_free_multi(iflib_dma_info_t *dmalist, int count)
1448 {
1449 	int i;
1450 	iflib_dma_info_t *dmaiter = dmalist;
1451 
1452 	for (i = 0; i < count; i++, dmaiter++)
1453 		iflib_dma_free(*dmaiter);
1454 }
1455 
1456 static int
1457 iflib_fast_intr(void *arg)
1458 {
1459 	iflib_filter_info_t info = arg;
1460 	struct grouptask *gtask = info->ifi_task;
1461 	int result;
1462 
1463 	DBG_COUNTER_INC(fast_intrs);
1464 	if (info->ifi_filter != NULL) {
1465 		result = info->ifi_filter(info->ifi_filter_arg);
1466 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1467 			return (result);
1468 	}
1469 
1470 	GROUPTASK_ENQUEUE(gtask);
1471 	return (FILTER_HANDLED);
1472 }
1473 
1474 static int
1475 iflib_fast_intr_rxtx(void *arg)
1476 {
1477 	iflib_filter_info_t info = arg;
1478 	struct grouptask *gtask = info->ifi_task;
1479 	if_ctx_t ctx;
1480 	iflib_rxq_t rxq = (iflib_rxq_t)info->ifi_ctx;
1481 	iflib_txq_t txq;
1482 	void *sc;
1483 	int i, cidx, result;
1484 	qidx_t txqid;
1485 	bool intr_enable, intr_legacy;
1486 
1487 	DBG_COUNTER_INC(fast_intrs);
1488 	if (info->ifi_filter != NULL) {
1489 		result = info->ifi_filter(info->ifi_filter_arg);
1490 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1491 			return (result);
1492 	}
1493 
1494 	ctx = rxq->ifr_ctx;
1495 	sc = ctx->ifc_softc;
1496 	intr_enable = false;
1497 	intr_legacy = !!(ctx->ifc_flags & IFC_LEGACY);
1498 	MPASS(rxq->ifr_ntxqirq);
1499 	for (i = 0; i < rxq->ifr_ntxqirq; i++) {
1500 		txqid = rxq->ifr_txqid[i];
1501 		txq = &ctx->ifc_txqs[txqid];
1502 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1503 		    BUS_DMASYNC_POSTREAD);
1504 		if (!ctx->isc_txd_credits_update(sc, txqid, false)) {
1505 			if (intr_legacy)
1506 				intr_enable = true;
1507 			else
1508 				IFDI_TX_QUEUE_INTR_ENABLE(ctx, txqid);
1509 			continue;
1510 		}
1511 		GROUPTASK_ENQUEUE(&txq->ift_task);
1512 	}
1513 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_RXCQ)
1514 		cidx = rxq->ifr_cq_cidx;
1515 	else
1516 		cidx = rxq->ifr_fl[0].ifl_cidx;
1517 	if (iflib_rxd_avail(ctx, rxq, cidx, 1))
1518 		GROUPTASK_ENQUEUE(gtask);
1519 	else {
1520 		if (intr_legacy)
1521 			intr_enable = true;
1522 		else
1523 			IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
1524 		DBG_COUNTER_INC(rx_intr_enables);
1525 	}
1526 	if (intr_enable)
1527 		IFDI_INTR_ENABLE(ctx);
1528 	return (FILTER_HANDLED);
1529 }
1530 
1531 
1532 static int
1533 iflib_fast_intr_ctx(void *arg)
1534 {
1535 	iflib_filter_info_t info = arg;
1536 	struct grouptask *gtask = info->ifi_task;
1537 	int result;
1538 
1539 	DBG_COUNTER_INC(fast_intrs);
1540 	if (info->ifi_filter != NULL) {
1541 		result = info->ifi_filter(info->ifi_filter_arg);
1542 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1543 			return (result);
1544 	}
1545 
1546 	GROUPTASK_ENQUEUE(gtask);
1547 	return (FILTER_HANDLED);
1548 }
1549 
1550 static int
1551 _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
1552 		 driver_filter_t filter, driver_intr_t handler, void *arg,
1553 		 const char *name)
1554 {
1555 	struct resource *res;
1556 	void *tag = NULL;
1557 	device_t dev = ctx->ifc_dev;
1558 	int flags, i, rc;
1559 
1560 	flags = RF_ACTIVE;
1561 	if (ctx->ifc_flags & IFC_LEGACY)
1562 		flags |= RF_SHAREABLE;
1563 	MPASS(rid < 512);
1564 	i = rid;
1565 	res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &i, flags);
1566 	if (res == NULL) {
1567 		device_printf(dev,
1568 		    "failed to allocate IRQ for rid %d, name %s.\n", rid, name);
1569 		return (ENOMEM);
1570 	}
1571 	irq->ii_res = res;
1572 	KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL"));
1573 	rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET,
1574 						filter, handler, arg, &tag);
1575 	if (rc != 0) {
1576 		device_printf(dev,
1577 		    "failed to setup interrupt for rid %d, name %s: %d\n",
1578 					  rid, name ? name : "unknown", rc);
1579 		return (rc);
1580 	} else if (name)
1581 		bus_describe_intr(dev, res, tag, "%s", name);
1582 
1583 	irq->ii_tag = tag;
1584 	return (0);
1585 }
1586 
1587 /*********************************************************************
1588  *
1589  *  Allocate DMA resources for TX buffers as well as memory for the TX
1590  *  mbuf map.  TX DMA maps (non-TSO/TSO) and TX mbuf map are kept in a
1591  *  iflib_sw_tx_desc_array structure, storing all the information that
1592  *  is needed to transmit a packet on the wire.  This is called only
1593  *  once at attach, setup is done every reset.
1594  *
1595  **********************************************************************/
1596 static int
1597 iflib_txsd_alloc(iflib_txq_t txq)
1598 {
1599 	if_ctx_t ctx = txq->ift_ctx;
1600 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1601 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1602 	device_t dev = ctx->ifc_dev;
1603 	bus_size_t tsomaxsize;
1604 	int err, nsegments, ntsosegments;
1605 	bool tso;
1606 
1607 	nsegments = scctx->isc_tx_nsegments;
1608 	ntsosegments = scctx->isc_tx_tso_segments_max;
1609 	tsomaxsize = scctx->isc_tx_tso_size_max;
1610 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_VLAN_MTU)
1611 		tsomaxsize += sizeof(struct ether_vlan_header);
1612 	MPASS(scctx->isc_ntxd[0] > 0);
1613 	MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0);
1614 	MPASS(nsegments > 0);
1615 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) {
1616 		MPASS(ntsosegments > 0);
1617 		MPASS(sctx->isc_tso_maxsize >= tsomaxsize);
1618 	}
1619 
1620 	/*
1621 	 * Set up DMA tags for TX buffers.
1622 	 */
1623 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1624 			       1, 0,			/* alignment, bounds */
1625 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1626 			       BUS_SPACE_MAXADDR,	/* highaddr */
1627 			       NULL, NULL,		/* filter, filterarg */
1628 			       sctx->isc_tx_maxsize,		/* maxsize */
1629 			       nsegments,	/* nsegments */
1630 			       sctx->isc_tx_maxsegsize,	/* maxsegsize */
1631 			       0,			/* flags */
1632 			       NULL,			/* lockfunc */
1633 			       NULL,			/* lockfuncarg */
1634 			       &txq->ift_buf_tag))) {
1635 		device_printf(dev,"Unable to allocate TX DMA tag: %d\n", err);
1636 		device_printf(dev,"maxsize: %ju nsegments: %d maxsegsize: %ju\n",
1637 		    (uintmax_t)sctx->isc_tx_maxsize, nsegments, (uintmax_t)sctx->isc_tx_maxsegsize);
1638 		goto fail;
1639 	}
1640 	tso = (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) != 0;
1641 	if (tso && (err = bus_dma_tag_create(bus_get_dma_tag(dev),
1642 			       1, 0,			/* alignment, bounds */
1643 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1644 			       BUS_SPACE_MAXADDR,	/* highaddr */
1645 			       NULL, NULL,		/* filter, filterarg */
1646 			       tsomaxsize,		/* maxsize */
1647 			       ntsosegments,	/* nsegments */
1648 			       sctx->isc_tso_maxsegsize,/* maxsegsize */
1649 			       0,			/* flags */
1650 			       NULL,			/* lockfunc */
1651 			       NULL,			/* lockfuncarg */
1652 			       &txq->ift_tso_buf_tag))) {
1653 		device_printf(dev, "Unable to allocate TSO TX DMA tag: %d\n",
1654 		    err);
1655 		goto fail;
1656 	}
1657 
1658 	/* Allocate memory for the TX mbuf map. */
1659 	if (!(txq->ift_sds.ifsd_m =
1660 	    (struct mbuf **) malloc(sizeof(struct mbuf *) *
1661 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1662 		device_printf(dev, "Unable to allocate TX mbuf map memory\n");
1663 		err = ENOMEM;
1664 		goto fail;
1665 	}
1666 
1667 	/*
1668 	 * Create the DMA maps for TX buffers.
1669 	 */
1670 	if ((txq->ift_sds.ifsd_map = (bus_dmamap_t *)malloc(
1671 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1672 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1673 		device_printf(dev,
1674 		    "Unable to allocate TX buffer DMA map memory\n");
1675 		err = ENOMEM;
1676 		goto fail;
1677 	}
1678 	if (tso && (txq->ift_sds.ifsd_tso_map = (bus_dmamap_t *)malloc(
1679 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1680 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1681 		device_printf(dev,
1682 		    "Unable to allocate TSO TX buffer map memory\n");
1683 		err = ENOMEM;
1684 		goto fail;
1685 	}
1686 	for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) {
1687 		err = bus_dmamap_create(txq->ift_buf_tag, 0,
1688 		    &txq->ift_sds.ifsd_map[i]);
1689 		if (err != 0) {
1690 			device_printf(dev, "Unable to create TX DMA map\n");
1691 			goto fail;
1692 		}
1693 		if (!tso)
1694 			continue;
1695 		err = bus_dmamap_create(txq->ift_tso_buf_tag, 0,
1696 		    &txq->ift_sds.ifsd_tso_map[i]);
1697 		if (err != 0) {
1698 			device_printf(dev, "Unable to create TSO TX DMA map\n");
1699 			goto fail;
1700 		}
1701 	}
1702 	return (0);
1703 fail:
1704 	/* We free all, it handles case where we are in the middle */
1705 	iflib_tx_structures_free(ctx);
1706 	return (err);
1707 }
1708 
1709 static void
1710 iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i)
1711 {
1712 	bus_dmamap_t map;
1713 
1714 	if (txq->ift_sds.ifsd_map != NULL) {
1715 		map = txq->ift_sds.ifsd_map[i];
1716 		bus_dmamap_sync(txq->ift_buf_tag, map, BUS_DMASYNC_POSTWRITE);
1717 		bus_dmamap_unload(txq->ift_buf_tag, map);
1718 		bus_dmamap_destroy(txq->ift_buf_tag, map);
1719 		txq->ift_sds.ifsd_map[i] = NULL;
1720 	}
1721 
1722 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1723 		map = txq->ift_sds.ifsd_tso_map[i];
1724 		bus_dmamap_sync(txq->ift_tso_buf_tag, map,
1725 		    BUS_DMASYNC_POSTWRITE);
1726 		bus_dmamap_unload(txq->ift_tso_buf_tag, map);
1727 		bus_dmamap_destroy(txq->ift_tso_buf_tag, map);
1728 		txq->ift_sds.ifsd_tso_map[i] = NULL;
1729 	}
1730 }
1731 
1732 static void
1733 iflib_txq_destroy(iflib_txq_t txq)
1734 {
1735 	if_ctx_t ctx = txq->ift_ctx;
1736 
1737 	for (int i = 0; i < txq->ift_size; i++)
1738 		iflib_txsd_destroy(ctx, txq, i);
1739 
1740 	if (txq->ift_br != NULL) {
1741 		ifmp_ring_free(txq->ift_br);
1742 		txq->ift_br = NULL;
1743 	}
1744 
1745 	mtx_destroy(&txq->ift_mtx);
1746 
1747 	if (txq->ift_sds.ifsd_map != NULL) {
1748 		free(txq->ift_sds.ifsd_map, M_IFLIB);
1749 		txq->ift_sds.ifsd_map = NULL;
1750 	}
1751 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1752 		free(txq->ift_sds.ifsd_tso_map, M_IFLIB);
1753 		txq->ift_sds.ifsd_tso_map = NULL;
1754 	}
1755 	if (txq->ift_sds.ifsd_m != NULL) {
1756 		free(txq->ift_sds.ifsd_m, M_IFLIB);
1757 		txq->ift_sds.ifsd_m = NULL;
1758 	}
1759 	if (txq->ift_buf_tag != NULL) {
1760 		bus_dma_tag_destroy(txq->ift_buf_tag);
1761 		txq->ift_buf_tag = NULL;
1762 	}
1763 	if (txq->ift_tso_buf_tag != NULL) {
1764 		bus_dma_tag_destroy(txq->ift_tso_buf_tag);
1765 		txq->ift_tso_buf_tag = NULL;
1766 	}
1767 	if (txq->ift_ifdi != NULL) {
1768 		free(txq->ift_ifdi, M_IFLIB);
1769 	}
1770 }
1771 
1772 static void
1773 iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i)
1774 {
1775 	struct mbuf **mp;
1776 
1777 	mp = &txq->ift_sds.ifsd_m[i];
1778 	if (*mp == NULL)
1779 		return;
1780 
1781 	if (txq->ift_sds.ifsd_map != NULL) {
1782 		bus_dmamap_sync(txq->ift_buf_tag,
1783 		    txq->ift_sds.ifsd_map[i], BUS_DMASYNC_POSTWRITE);
1784 		bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[i]);
1785 	}
1786 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1787 		bus_dmamap_sync(txq->ift_tso_buf_tag,
1788 		    txq->ift_sds.ifsd_tso_map[i], BUS_DMASYNC_POSTWRITE);
1789 		bus_dmamap_unload(txq->ift_tso_buf_tag,
1790 		    txq->ift_sds.ifsd_tso_map[i]);
1791 	}
1792 	m_free(*mp);
1793 	DBG_COUNTER_INC(tx_frees);
1794 	*mp = NULL;
1795 }
1796 
1797 static int
1798 iflib_txq_setup(iflib_txq_t txq)
1799 {
1800 	if_ctx_t ctx = txq->ift_ctx;
1801 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1802 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1803 	iflib_dma_info_t di;
1804 	int i;
1805 
1806 	/* Set number of descriptors available */
1807 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
1808 	/* XXX make configurable */
1809 	txq->ift_update_freq = IFLIB_DEFAULT_TX_UPDATE_FREQ;
1810 
1811 	/* Reset indices */
1812 	txq->ift_cidx_processed = 0;
1813 	txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0;
1814 	txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset];
1815 
1816 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1817 		bzero((void *)di->idi_vaddr, di->idi_size);
1818 
1819 	IFDI_TXQ_SETUP(ctx, txq->ift_id);
1820 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1821 		bus_dmamap_sync(di->idi_tag, di->idi_map,
1822 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1823 	return (0);
1824 }
1825 
1826 /*********************************************************************
1827  *
1828  *  Allocate DMA resources for RX buffers as well as memory for the RX
1829  *  mbuf map, direct RX cluster pointer map and RX cluster bus address
1830  *  map.  RX DMA map, RX mbuf map, direct RX cluster pointer map and
1831  *  RX cluster map are kept in a iflib_sw_rx_desc_array structure.
1832  *  Since we use use one entry in iflib_sw_rx_desc_array per received
1833  *  packet, the maximum number of entries we'll need is equal to the
1834  *  number of hardware receive descriptors that we've allocated.
1835  *
1836  **********************************************************************/
1837 static int
1838 iflib_rxsd_alloc(iflib_rxq_t rxq)
1839 {
1840 	if_ctx_t ctx = rxq->ifr_ctx;
1841 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1842 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1843 	device_t dev = ctx->ifc_dev;
1844 	iflib_fl_t fl;
1845 	int			err;
1846 
1847 	MPASS(scctx->isc_nrxd[0] > 0);
1848 	MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0);
1849 
1850 	fl = rxq->ifr_fl;
1851 	for (int i = 0; i <  rxq->ifr_nfl; i++, fl++) {
1852 		fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */
1853 		/* Set up DMA tag for RX buffers. */
1854 		err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1855 					 1, 0,			/* alignment, bounds */
1856 					 BUS_SPACE_MAXADDR,	/* lowaddr */
1857 					 BUS_SPACE_MAXADDR,	/* highaddr */
1858 					 NULL, NULL,		/* filter, filterarg */
1859 					 sctx->isc_rx_maxsize,	/* maxsize */
1860 					 sctx->isc_rx_nsegments,	/* nsegments */
1861 					 sctx->isc_rx_maxsegsize,	/* maxsegsize */
1862 					 0,			/* flags */
1863 					 NULL,			/* lockfunc */
1864 					 NULL,			/* lockarg */
1865 					 &fl->ifl_buf_tag);
1866 		if (err) {
1867 			device_printf(dev,
1868 			    "Unable to allocate RX DMA tag: %d\n", err);
1869 			goto fail;
1870 		}
1871 
1872 		/* Allocate memory for the RX mbuf map. */
1873 		if (!(fl->ifl_sds.ifsd_m =
1874 		      (struct mbuf **) malloc(sizeof(struct mbuf *) *
1875 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1876 			device_printf(dev,
1877 			    "Unable to allocate RX mbuf map memory\n");
1878 			err = ENOMEM;
1879 			goto fail;
1880 		}
1881 
1882 		/* Allocate memory for the direct RX cluster pointer map. */
1883 		if (!(fl->ifl_sds.ifsd_cl =
1884 		      (caddr_t *) malloc(sizeof(caddr_t) *
1885 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1886 			device_printf(dev,
1887 			    "Unable to allocate RX cluster map memory\n");
1888 			err = ENOMEM;
1889 			goto fail;
1890 		}
1891 
1892 		/* Allocate memory for the RX cluster bus address map. */
1893 		if (!(fl->ifl_sds.ifsd_ba =
1894 		      (bus_addr_t *) malloc(sizeof(bus_addr_t) *
1895 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1896 			device_printf(dev,
1897 			    "Unable to allocate RX bus address map memory\n");
1898 			err = ENOMEM;
1899 			goto fail;
1900 		}
1901 
1902 		/*
1903 		 * Create the DMA maps for RX buffers.
1904 		 */
1905 		if (!(fl->ifl_sds.ifsd_map =
1906 		      (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1907 			device_printf(dev,
1908 			    "Unable to allocate RX buffer DMA map memory\n");
1909 			err = ENOMEM;
1910 			goto fail;
1911 		}
1912 		for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++) {
1913 			err = bus_dmamap_create(fl->ifl_buf_tag, 0,
1914 			    &fl->ifl_sds.ifsd_map[i]);
1915 			if (err != 0) {
1916 				device_printf(dev, "Unable to create RX buffer DMA map\n");
1917 				goto fail;
1918 			}
1919 		}
1920 	}
1921 	return (0);
1922 
1923 fail:
1924 	iflib_rx_structures_free(ctx);
1925 	return (err);
1926 }
1927 
1928 
1929 /*
1930  * Internal service routines
1931  */
1932 
1933 struct rxq_refill_cb_arg {
1934 	int               error;
1935 	bus_dma_segment_t seg;
1936 	int               nseg;
1937 };
1938 
1939 static void
1940 _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
1941 {
1942 	struct rxq_refill_cb_arg *cb_arg = arg;
1943 
1944 	cb_arg->error = error;
1945 	cb_arg->seg = segs[0];
1946 	cb_arg->nseg = nseg;
1947 }
1948 
1949 /**
1950  * iflib_fl_refill - refill an rxq free-buffer list
1951  * @ctx: the iflib context
1952  * @fl: the free list to refill
1953  * @count: the number of new buffers to allocate
1954  *
1955  * (Re)populate an rxq free-buffer list with up to @count new packet buffers.
1956  * The caller must assure that @count does not exceed the queue's capacity.
1957  */
1958 static uint8_t
1959 iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
1960 {
1961 	struct if_rxd_update iru;
1962 	struct rxq_refill_cb_arg cb_arg;
1963 	struct mbuf *m;
1964 	caddr_t cl, *sd_cl;
1965 	struct mbuf **sd_m;
1966 	bus_dmamap_t *sd_map;
1967 	bus_addr_t bus_addr, *sd_ba;
1968 	int err, frag_idx, i, idx, n, pidx;
1969 	qidx_t credits;
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 		if (fl->ifl_pidx == 0)
2079 			pidx = fl->ifl_size - 1;
2080 		else
2081 			pidx = fl->ifl_pidx - 1;
2082 
2083 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2084 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2085 		ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id,
2086 		    fl->ifl_id, pidx);
2087 		if (__predict_true(bit_test(fl->ifl_rx_bitmap, frag_idx))) {
2088 			fl->ifl_fragidx = frag_idx + 1;
2089 			if (fl->ifl_fragidx == fl->ifl_size)
2090 				fl->ifl_fragidx = 0;
2091 		} else {
2092 			fl->ifl_fragidx = frag_idx;
2093 		}
2094 	}
2095 
2096 	return (n == -1 ? 0 : IFLIB_RXEOF_EMPTY);
2097 }
2098 
2099 static inline uint8_t
2100 iflib_fl_refill_all(if_ctx_t ctx, iflib_fl_t fl)
2101 {
2102 	/* we avoid allowing pidx to catch up with cidx as it confuses ixl */
2103 	int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1;
2104 #ifdef INVARIANTS
2105 	int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1;
2106 #endif
2107 
2108 	MPASS(fl->ifl_credits <= fl->ifl_size);
2109 	MPASS(reclaimable == delta);
2110 
2111 	if (reclaimable > 0)
2112 		return (iflib_fl_refill(ctx, fl, reclaimable));
2113 	return (0);
2114 }
2115 
2116 uint8_t
2117 iflib_in_detach(if_ctx_t ctx)
2118 {
2119 	bool in_detach;
2120 
2121 	STATE_LOCK(ctx);
2122 	in_detach = !!(ctx->ifc_flags & IFC_IN_DETACH);
2123 	STATE_UNLOCK(ctx);
2124 	return (in_detach);
2125 }
2126 
2127 static void
2128 iflib_fl_bufs_free(iflib_fl_t fl)
2129 {
2130 	iflib_dma_info_t idi = fl->ifl_ifdi;
2131 	bus_dmamap_t sd_map;
2132 	uint32_t i;
2133 
2134 	for (i = 0; i < fl->ifl_size; i++) {
2135 		struct mbuf **sd_m = &fl->ifl_sds.ifsd_m[i];
2136 		caddr_t *sd_cl = &fl->ifl_sds.ifsd_cl[i];
2137 
2138 		if (*sd_cl != NULL) {
2139 			sd_map = fl->ifl_sds.ifsd_map[i];
2140 			bus_dmamap_sync(fl->ifl_buf_tag, sd_map,
2141 			    BUS_DMASYNC_POSTREAD);
2142 			bus_dmamap_unload(fl->ifl_buf_tag, sd_map);
2143 			uma_zfree(fl->ifl_zone, *sd_cl);
2144 			*sd_cl = NULL;
2145 			if (*sd_m != NULL) {
2146 				m_init(*sd_m, M_NOWAIT, MT_DATA, 0);
2147 				uma_zfree(zone_mbuf, *sd_m);
2148 				*sd_m = NULL;
2149 			}
2150 		} else {
2151 			MPASS(*sd_m == NULL);
2152 		}
2153 #if MEMORY_LOGGING
2154 		fl->ifl_m_dequeued++;
2155 		fl->ifl_cl_dequeued++;
2156 #endif
2157 	}
2158 #ifdef INVARIANTS
2159 	for (i = 0; i < fl->ifl_size; i++) {
2160 		MPASS(fl->ifl_sds.ifsd_cl[i] == NULL);
2161 		MPASS(fl->ifl_sds.ifsd_m[i] == NULL);
2162 	}
2163 #endif
2164 	/*
2165 	 * Reset free list values
2166 	 */
2167 	fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = fl->ifl_fragidx = 0;
2168 	bzero(idi->idi_vaddr, idi->idi_size);
2169 }
2170 
2171 /*********************************************************************
2172  *
2173  *  Initialize a free list and its buffers.
2174  *
2175  **********************************************************************/
2176 static int
2177 iflib_fl_setup(iflib_fl_t fl)
2178 {
2179 	iflib_rxq_t rxq = fl->ifl_rxq;
2180 	if_ctx_t ctx = rxq->ifr_ctx;
2181 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2182 	int qidx;
2183 
2184 	bit_nclear(fl->ifl_rx_bitmap, 0, fl->ifl_size - 1);
2185 	/*
2186 	** Free current RX buffer structs and their mbufs
2187 	*/
2188 	iflib_fl_bufs_free(fl);
2189 	/* Now replenish the mbufs */
2190 	MPASS(fl->ifl_credits == 0);
2191 	qidx = rxq->ifr_fl_offset + fl->ifl_id;
2192 	if (scctx->isc_rxd_buf_size[qidx] != 0)
2193 		fl->ifl_buf_size = scctx->isc_rxd_buf_size[qidx];
2194 	else
2195 		fl->ifl_buf_size = ctx->ifc_rx_mbuf_sz;
2196 	/*
2197 	 * ifl_buf_size may be a driver-supplied value, so pull it up
2198 	 * to the selected mbuf size.
2199 	 */
2200 	fl->ifl_buf_size = iflib_get_mbuf_size_for(fl->ifl_buf_size);
2201 	if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size)
2202 		ctx->ifc_max_fl_buf_size = fl->ifl_buf_size;
2203 	fl->ifl_cltype = m_gettype(fl->ifl_buf_size);
2204 	fl->ifl_zone = m_getzone(fl->ifl_buf_size);
2205 
2206 
2207 	/* avoid pre-allocating zillions of clusters to an idle card
2208 	 * potentially speeding up attach
2209 	 */
2210 	(void)iflib_fl_refill(ctx, fl, min(128, fl->ifl_size));
2211 	MPASS(min(128, fl->ifl_size) == fl->ifl_credits);
2212 	if (min(128, fl->ifl_size) != fl->ifl_credits)
2213 		return (ENOBUFS);
2214 	/*
2215 	 * handle failure
2216 	 */
2217 	MPASS(rxq != NULL);
2218 	MPASS(fl->ifl_ifdi != NULL);
2219 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2220 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2221 	return (0);
2222 }
2223 
2224 /*********************************************************************
2225  *
2226  *  Free receive ring data structures
2227  *
2228  **********************************************************************/
2229 static void
2230 iflib_rx_sds_free(iflib_rxq_t rxq)
2231 {
2232 	iflib_fl_t fl;
2233 	int i, j;
2234 
2235 	if (rxq->ifr_fl != NULL) {
2236 		for (i = 0; i < rxq->ifr_nfl; i++) {
2237 			fl = &rxq->ifr_fl[i];
2238 			if (fl->ifl_buf_tag != NULL) {
2239 				if (fl->ifl_sds.ifsd_map != NULL) {
2240 					for (j = 0; j < fl->ifl_size; j++) {
2241 						bus_dmamap_sync(
2242 						    fl->ifl_buf_tag,
2243 						    fl->ifl_sds.ifsd_map[j],
2244 						    BUS_DMASYNC_POSTREAD);
2245 						bus_dmamap_unload(
2246 						    fl->ifl_buf_tag,
2247 						    fl->ifl_sds.ifsd_map[j]);
2248 						bus_dmamap_destroy(
2249 						    fl->ifl_buf_tag,
2250 						    fl->ifl_sds.ifsd_map[j]);
2251 					}
2252 				}
2253 				bus_dma_tag_destroy(fl->ifl_buf_tag);
2254 				fl->ifl_buf_tag = NULL;
2255 			}
2256 			free(fl->ifl_sds.ifsd_m, M_IFLIB);
2257 			free(fl->ifl_sds.ifsd_cl, M_IFLIB);
2258 			free(fl->ifl_sds.ifsd_ba, M_IFLIB);
2259 			free(fl->ifl_sds.ifsd_map, M_IFLIB);
2260 			fl->ifl_sds.ifsd_m = NULL;
2261 			fl->ifl_sds.ifsd_cl = NULL;
2262 			fl->ifl_sds.ifsd_ba = NULL;
2263 			fl->ifl_sds.ifsd_map = NULL;
2264 		}
2265 		free(rxq->ifr_fl, M_IFLIB);
2266 		rxq->ifr_fl = NULL;
2267 		free(rxq->ifr_ifdi, M_IFLIB);
2268 		rxq->ifr_ifdi = NULL;
2269 		rxq->ifr_cq_cidx = 0;
2270 	}
2271 }
2272 
2273 /*
2274  * Timer routine
2275  */
2276 static void
2277 iflib_timer(void *arg)
2278 {
2279 	iflib_txq_t txq = arg;
2280 	if_ctx_t ctx = txq->ift_ctx;
2281 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2282 	uint64_t this_tick = ticks;
2283 	uint32_t reset_on = hz / 2;
2284 
2285 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
2286 		return;
2287 
2288 	/*
2289 	** Check on the state of the TX queue(s), this
2290 	** can be done without the lock because its RO
2291 	** and the HUNG state will be static if set.
2292 	*/
2293 	if (this_tick - txq->ift_last_timer_tick >= hz / 2) {
2294 		txq->ift_last_timer_tick = this_tick;
2295 		IFDI_TIMER(ctx, txq->ift_id);
2296 		if ((txq->ift_qstatus == IFLIB_QUEUE_HUNG) &&
2297 		    ((txq->ift_cleaned_prev == txq->ift_cleaned) ||
2298 		     (sctx->isc_pause_frames == 0)))
2299 			goto hung;
2300 
2301 		if (txq->ift_qstatus != IFLIB_QUEUE_IDLE &&
2302 		    ifmp_ring_is_stalled(txq->ift_br)) {
2303 			KASSERT(ctx->ifc_link_state == LINK_STATE_UP, ("queue can't be marked as hung if interface is down"));
2304 			txq->ift_qstatus = IFLIB_QUEUE_HUNG;
2305 		}
2306 		txq->ift_cleaned_prev = txq->ift_cleaned;
2307 	}
2308 #ifdef DEV_NETMAP
2309 	if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP)
2310 		iflib_netmap_timer_adjust(ctx, txq, &reset_on);
2311 #endif
2312 	/* handle any laggards */
2313 	if (txq->ift_db_pending)
2314 		GROUPTASK_ENQUEUE(&txq->ift_task);
2315 
2316 	sctx->isc_pause_frames = 0;
2317 	if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)
2318 		callout_reset_on(&txq->ift_timer, reset_on, iflib_timer, txq, txq->ift_timer.c_cpu);
2319 	return;
2320 
2321  hung:
2322 	device_printf(ctx->ifc_dev,
2323 	    "Watchdog timeout (TX: %d desc avail: %d pidx: %d) -- resetting\n",
2324 	    txq->ift_id, TXQ_AVAIL(txq), txq->ift_pidx);
2325 	STATE_LOCK(ctx);
2326 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2327 	ctx->ifc_flags |= (IFC_DO_WATCHDOG|IFC_DO_RESET);
2328 	iflib_admin_intr_deferred(ctx);
2329 	STATE_UNLOCK(ctx);
2330 }
2331 
2332 static uint16_t
2333 iflib_get_mbuf_size_for(unsigned int size)
2334 {
2335 
2336 	if (size <= MCLBYTES)
2337 		return (MCLBYTES);
2338 	else
2339 		return (MJUMPAGESIZE);
2340 }
2341 
2342 static void
2343 iflib_calc_rx_mbuf_sz(if_ctx_t ctx)
2344 {
2345 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2346 
2347 	/*
2348 	 * XXX don't set the max_frame_size to larger
2349 	 * than the hardware can handle
2350 	 */
2351 	ctx->ifc_rx_mbuf_sz =
2352 	    iflib_get_mbuf_size_for(sctx->isc_max_frame_size);
2353 }
2354 
2355 uint32_t
2356 iflib_get_rx_mbuf_sz(if_ctx_t ctx)
2357 {
2358 
2359 	return (ctx->ifc_rx_mbuf_sz);
2360 }
2361 
2362 static void
2363 iflib_init_locked(if_ctx_t ctx)
2364 {
2365 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2366 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2367 	if_t ifp = ctx->ifc_ifp;
2368 	iflib_fl_t fl;
2369 	iflib_txq_t txq;
2370 	iflib_rxq_t rxq;
2371 	int i, j, tx_ip_csum_flags, tx_ip6_csum_flags;
2372 
2373 	if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2374 	IFDI_INTR_DISABLE(ctx);
2375 
2376 	tx_ip_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP);
2377 	tx_ip6_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | CSUM_IP6_SCTP);
2378 	/* Set hardware offload abilities */
2379 	if_clearhwassist(ifp);
2380 	if (if_getcapenable(ifp) & IFCAP_TXCSUM)
2381 		if_sethwassistbits(ifp, tx_ip_csum_flags, 0);
2382 	if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6)
2383 		if_sethwassistbits(ifp,  tx_ip6_csum_flags, 0);
2384 	if (if_getcapenable(ifp) & IFCAP_TSO4)
2385 		if_sethwassistbits(ifp, CSUM_IP_TSO, 0);
2386 	if (if_getcapenable(ifp) & IFCAP_TSO6)
2387 		if_sethwassistbits(ifp, CSUM_IP6_TSO, 0);
2388 
2389 	for (i = 0, txq = ctx->ifc_txqs; i < sctx->isc_ntxqsets; i++, txq++) {
2390 		CALLOUT_LOCK(txq);
2391 		callout_stop(&txq->ift_timer);
2392 		CALLOUT_UNLOCK(txq);
2393 		iflib_netmap_txq_init(ctx, txq);
2394 	}
2395 
2396 	/*
2397 	 * Calculate a suitable Rx mbuf size prior to calling IFDI_INIT, so
2398 	 * that drivers can use the value when setting up the hardware receive
2399 	 * buffers.
2400 	 */
2401 	iflib_calc_rx_mbuf_sz(ctx);
2402 
2403 #ifdef INVARIANTS
2404 	i = if_getdrvflags(ifp);
2405 #endif
2406 	IFDI_INIT(ctx);
2407 	MPASS(if_getdrvflags(ifp) == i);
2408 	for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
2409 		if (iflib_netmap_rxq_init(ctx, rxq) > 0) {
2410 			/* This rxq is in netmap mode. Skip normal init. */
2411 			continue;
2412 		}
2413 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
2414 			if (iflib_fl_setup(fl)) {
2415 				device_printf(ctx->ifc_dev,
2416 				    "setting up free list %d failed - "
2417 				    "check cluster settings\n", j);
2418 				goto done;
2419 			}
2420 		}
2421 	}
2422 done:
2423 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
2424 	IFDI_INTR_ENABLE(ctx);
2425 	txq = ctx->ifc_txqs;
2426 	for (i = 0; i < sctx->isc_ntxqsets; i++, txq++)
2427 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq,
2428 			txq->ift_timer.c_cpu);
2429 }
2430 
2431 static int
2432 iflib_media_change(if_t ifp)
2433 {
2434 	if_ctx_t ctx = if_getsoftc(ifp);
2435 	int err;
2436 
2437 	CTX_LOCK(ctx);
2438 	if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0)
2439 		iflib_init_locked(ctx);
2440 	CTX_UNLOCK(ctx);
2441 	return (err);
2442 }
2443 
2444 static void
2445 iflib_media_status(if_t ifp, struct ifmediareq *ifmr)
2446 {
2447 	if_ctx_t ctx = if_getsoftc(ifp);
2448 
2449 	CTX_LOCK(ctx);
2450 	IFDI_UPDATE_ADMIN_STATUS(ctx);
2451 	IFDI_MEDIA_STATUS(ctx, ifmr);
2452 	CTX_UNLOCK(ctx);
2453 }
2454 
2455 void
2456 iflib_stop(if_ctx_t ctx)
2457 {
2458 	iflib_txq_t txq = ctx->ifc_txqs;
2459 	iflib_rxq_t rxq = ctx->ifc_rxqs;
2460 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2461 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2462 	iflib_dma_info_t di;
2463 	iflib_fl_t fl;
2464 	int i, j;
2465 
2466 	/* Tell the stack that the interface is no longer active */
2467 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2468 
2469 	IFDI_INTR_DISABLE(ctx);
2470 	DELAY(1000);
2471 	IFDI_STOP(ctx);
2472 	DELAY(1000);
2473 
2474 	iflib_debug_reset();
2475 	/* Wait for current tx queue users to exit to disarm watchdog timer. */
2476 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) {
2477 		/* make sure all transmitters have completed before proceeding XXX */
2478 
2479 		CALLOUT_LOCK(txq);
2480 		callout_stop(&txq->ift_timer);
2481 		CALLOUT_UNLOCK(txq);
2482 
2483 		/* clean any enqueued buffers */
2484 		iflib_ifmp_purge(txq);
2485 		/* Free any existing tx buffers. */
2486 		for (j = 0; j < txq->ift_size; j++) {
2487 			iflib_txsd_free(ctx, txq, j);
2488 		}
2489 		txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0;
2490 		txq->ift_in_use = txq->ift_gen = txq->ift_cidx = txq->ift_pidx = txq->ift_no_desc_avail = 0;
2491 		txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0;
2492 		txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0;
2493 		txq->ift_pullups = 0;
2494 		ifmp_ring_reset_stats(txq->ift_br);
2495 		for (j = 0, di = txq->ift_ifdi; j < sctx->isc_ntxqs; j++, di++)
2496 			bzero((void *)di->idi_vaddr, di->idi_size);
2497 	}
2498 	for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) {
2499 		/* make sure all transmitters have completed before proceeding XXX */
2500 
2501 		rxq->ifr_cq_cidx = 0;
2502 		for (j = 0, di = rxq->ifr_ifdi; j < sctx->isc_nrxqs; j++, di++)
2503 			bzero((void *)di->idi_vaddr, di->idi_size);
2504 		/* also resets the free lists pidx/cidx */
2505 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
2506 			iflib_fl_bufs_free(fl);
2507 	}
2508 }
2509 
2510 static inline caddr_t
2511 calc_next_rxd(iflib_fl_t fl, int cidx)
2512 {
2513 	qidx_t size;
2514 	int nrxd;
2515 	caddr_t start, end, cur, next;
2516 
2517 	nrxd = fl->ifl_size;
2518 	size = fl->ifl_rxd_size;
2519 	start = fl->ifl_ifdi->idi_vaddr;
2520 
2521 	if (__predict_false(size == 0))
2522 		return (start);
2523 	cur = start + size*cidx;
2524 	end = start + size*nrxd;
2525 	next = CACHE_PTR_NEXT(cur);
2526 	return (next < end ? next : start);
2527 }
2528 
2529 static inline void
2530 prefetch_pkts(iflib_fl_t fl, int cidx)
2531 {
2532 	int nextptr;
2533 	int nrxd = fl->ifl_size;
2534 	caddr_t next_rxd;
2535 
2536 
2537 	nextptr = (cidx + CACHE_PTR_INCREMENT) & (nrxd-1);
2538 	prefetch(&fl->ifl_sds.ifsd_m[nextptr]);
2539 	prefetch(&fl->ifl_sds.ifsd_cl[nextptr]);
2540 	next_rxd = calc_next_rxd(fl, cidx);
2541 	prefetch(next_rxd);
2542 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 1) & (nrxd-1)]);
2543 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 2) & (nrxd-1)]);
2544 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 3) & (nrxd-1)]);
2545 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 4) & (nrxd-1)]);
2546 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 1) & (nrxd-1)]);
2547 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 2) & (nrxd-1)]);
2548 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 3) & (nrxd-1)]);
2549 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 4) & (nrxd-1)]);
2550 }
2551 
2552 static struct mbuf *
2553 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, bool unload, if_rxsd_t sd,
2554     int *pf_rv, if_rxd_info_t ri)
2555 {
2556 	bus_dmamap_t map;
2557 	iflib_fl_t fl;
2558 	caddr_t payload;
2559 	struct mbuf *m;
2560 	int flid, cidx, len, next;
2561 
2562 	map = NULL;
2563 	flid = irf->irf_flid;
2564 	cidx = irf->irf_idx;
2565 	fl = &rxq->ifr_fl[flid];
2566 	sd->ifsd_fl = fl;
2567 	m = fl->ifl_sds.ifsd_m[cidx];
2568 	sd->ifsd_cl = &fl->ifl_sds.ifsd_cl[cidx];
2569 	fl->ifl_credits--;
2570 #if MEMORY_LOGGING
2571 	fl->ifl_m_dequeued++;
2572 #endif
2573 	if (rxq->ifr_ctx->ifc_flags & IFC_PREFETCH)
2574 		prefetch_pkts(fl, cidx);
2575 	next = (cidx + CACHE_PTR_INCREMENT) & (fl->ifl_size-1);
2576 	prefetch(&fl->ifl_sds.ifsd_map[next]);
2577 	map = fl->ifl_sds.ifsd_map[cidx];
2578 
2579 	bus_dmamap_sync(fl->ifl_buf_tag, map, BUS_DMASYNC_POSTREAD);
2580 
2581 	if (rxq->pfil != NULL && PFIL_HOOKED_IN(rxq->pfil) && pf_rv != NULL &&
2582 	    irf->irf_len != 0) {
2583 		payload  = *sd->ifsd_cl;
2584 		payload +=  ri->iri_pad;
2585 		len = ri->iri_len - ri->iri_pad;
2586 		*pf_rv = pfil_run_hooks(rxq->pfil, payload, ri->iri_ifp,
2587 		    len | PFIL_MEMPTR | PFIL_IN, NULL);
2588 		switch (*pf_rv) {
2589 		case PFIL_DROPPED:
2590 		case PFIL_CONSUMED:
2591 			/*
2592 			 * The filter ate it.  Everything is recycled.
2593 			 */
2594 			m = NULL;
2595 			unload = 0;
2596 			break;
2597 		case PFIL_REALLOCED:
2598 			/*
2599 			 * The filter copied it.  Everything is recycled.
2600 			 */
2601 			m = pfil_mem2mbuf(payload);
2602 			unload = 0;
2603 			break;
2604 		case PFIL_PASS:
2605 			/*
2606 			 * Filter said it was OK, so receive like
2607 			 * normal
2608 			 */
2609 			fl->ifl_sds.ifsd_m[cidx] = NULL;
2610 			break;
2611 		default:
2612 			MPASS(0);
2613 		}
2614 	} else {
2615 		fl->ifl_sds.ifsd_m[cidx] = NULL;
2616 		*pf_rv = PFIL_PASS;
2617 	}
2618 
2619 	if (unload && irf->irf_len != 0)
2620 		bus_dmamap_unload(fl->ifl_buf_tag, map);
2621 	fl->ifl_cidx = (fl->ifl_cidx + 1) & (fl->ifl_size-1);
2622 	if (__predict_false(fl->ifl_cidx == 0))
2623 		fl->ifl_gen = 0;
2624 	bit_clear(fl->ifl_rx_bitmap, cidx);
2625 	return (m);
2626 }
2627 
2628 static struct mbuf *
2629 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri, if_rxsd_t sd, int *pf_rv)
2630 {
2631 	struct mbuf *m, *mh, *mt;
2632 	caddr_t cl;
2633 	int  *pf_rv_ptr, flags, i, padlen;
2634 	bool consumed;
2635 
2636 	i = 0;
2637 	mh = NULL;
2638 	consumed = false;
2639 	*pf_rv = PFIL_PASS;
2640 	pf_rv_ptr = pf_rv;
2641 	do {
2642 		m = rxd_frag_to_sd(rxq, &ri->iri_frags[i], !consumed, sd,
2643 		    pf_rv_ptr, ri);
2644 
2645 		MPASS(*sd->ifsd_cl != NULL);
2646 
2647 		/*
2648 		 * Exclude zero-length frags & frags from
2649 		 * packets the filter has consumed or dropped
2650 		 */
2651 		if (ri->iri_frags[i].irf_len == 0 || consumed ||
2652 		    *pf_rv == PFIL_CONSUMED || *pf_rv == PFIL_DROPPED) {
2653 			if (mh == NULL) {
2654 				/* everything saved here */
2655 				consumed = true;
2656 				pf_rv_ptr = NULL;
2657 				continue;
2658 			}
2659 			/* XXX we can save the cluster here, but not the mbuf */
2660 			m_init(m, M_NOWAIT, MT_DATA, 0);
2661 			m_free(m);
2662 			continue;
2663 		}
2664 		if (mh == NULL) {
2665 			flags = M_PKTHDR|M_EXT;
2666 			mh = mt = m;
2667 			padlen = ri->iri_pad;
2668 		} else {
2669 			flags = M_EXT;
2670 			mt->m_next = m;
2671 			mt = m;
2672 			/* assuming padding is only on the first fragment */
2673 			padlen = 0;
2674 		}
2675 		cl = *sd->ifsd_cl;
2676 		*sd->ifsd_cl = NULL;
2677 
2678 		/* Can these two be made one ? */
2679 		m_init(m, M_NOWAIT, MT_DATA, flags);
2680 		m_cljset(m, cl, sd->ifsd_fl->ifl_cltype);
2681 		/*
2682 		 * These must follow m_init and m_cljset
2683 		 */
2684 		m->m_data += padlen;
2685 		ri->iri_len -= padlen;
2686 		m->m_len = ri->iri_frags[i].irf_len;
2687 	} while (++i < ri->iri_nfrags);
2688 
2689 	return (mh);
2690 }
2691 
2692 /*
2693  * Process one software descriptor
2694  */
2695 static struct mbuf *
2696 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri)
2697 {
2698 	struct if_rxsd sd;
2699 	struct mbuf *m;
2700 	int pf_rv;
2701 
2702 	/* should I merge this back in now that the two paths are basically duplicated? */
2703 	if (ri->iri_nfrags == 1 &&
2704 	    ri->iri_frags[0].irf_len != 0 &&
2705 	    ri->iri_frags[0].irf_len <= MIN(IFLIB_RX_COPY_THRESH, MHLEN)) {
2706 		m = rxd_frag_to_sd(rxq, &ri->iri_frags[0], false, &sd,
2707 		    &pf_rv, ri);
2708 		if (pf_rv != PFIL_PASS && pf_rv != PFIL_REALLOCED)
2709 			return (m);
2710 		if (pf_rv == PFIL_PASS) {
2711 			m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR);
2712 #ifndef __NO_STRICT_ALIGNMENT
2713 			if (!IP_ALIGNED(m))
2714 				m->m_data += 2;
2715 #endif
2716 			memcpy(m->m_data, *sd.ifsd_cl, ri->iri_len);
2717 			m->m_len = ri->iri_frags[0].irf_len;
2718 		}
2719 	} else {
2720 		m = assemble_segments(rxq, ri, &sd, &pf_rv);
2721 		if (m == NULL)
2722 			return (NULL);
2723 		if (pf_rv != PFIL_PASS && pf_rv != PFIL_REALLOCED)
2724 			return (m);
2725 	}
2726 	m->m_pkthdr.len = ri->iri_len;
2727 	m->m_pkthdr.rcvif = ri->iri_ifp;
2728 	m->m_flags |= ri->iri_flags;
2729 	m->m_pkthdr.ether_vtag = ri->iri_vtag;
2730 	m->m_pkthdr.flowid = ri->iri_flowid;
2731 	M_HASHTYPE_SET(m, ri->iri_rsstype);
2732 	m->m_pkthdr.csum_flags = ri->iri_csum_flags;
2733 	m->m_pkthdr.csum_data = ri->iri_csum_data;
2734 	return (m);
2735 }
2736 
2737 #if defined(INET6) || defined(INET)
2738 static void
2739 iflib_get_ip_forwarding(struct lro_ctrl *lc, bool *v4, bool *v6)
2740 {
2741 	CURVNET_SET(lc->ifp->if_vnet);
2742 #if defined(INET6)
2743 	*v6 = V_ip6_forwarding;
2744 #endif
2745 #if defined(INET)
2746 	*v4 = V_ipforwarding;
2747 #endif
2748 	CURVNET_RESTORE();
2749 }
2750 
2751 /*
2752  * Returns true if it's possible this packet could be LROed.
2753  * if it returns false, it is guaranteed that tcp_lro_rx()
2754  * would not return zero.
2755  */
2756 static bool
2757 iflib_check_lro_possible(struct mbuf *m, bool v4_forwarding, bool v6_forwarding)
2758 {
2759 	struct ether_header *eh;
2760 
2761 	eh = mtod(m, struct ether_header *);
2762 	switch (eh->ether_type) {
2763 #if defined(INET6)
2764 		case htons(ETHERTYPE_IPV6):
2765 			return (!v6_forwarding);
2766 #endif
2767 #if defined (INET)
2768 		case htons(ETHERTYPE_IP):
2769 			return (!v4_forwarding);
2770 #endif
2771 	}
2772 
2773 	return false;
2774 }
2775 #else
2776 static void
2777 iflib_get_ip_forwarding(struct lro_ctrl *lc __unused, bool *v4 __unused, bool *v6 __unused)
2778 {
2779 }
2780 #endif
2781 
2782 static void
2783 _task_fn_rx_watchdog(void *context)
2784 {
2785 	iflib_rxq_t rxq = context;
2786 
2787 	GROUPTASK_ENQUEUE(&rxq->ifr_task);
2788 }
2789 
2790 static uint8_t
2791 iflib_rxeof(iflib_rxq_t rxq, qidx_t budget)
2792 {
2793 	if_t ifp;
2794 	if_ctx_t ctx = rxq->ifr_ctx;
2795 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2796 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2797 	int avail, i;
2798 	qidx_t *cidxp;
2799 	struct if_rxd_info ri;
2800 	int err, budget_left, rx_bytes, rx_pkts;
2801 	iflib_fl_t fl;
2802 	int lro_enabled;
2803 	bool v4_forwarding, v6_forwarding, lro_possible;
2804 	uint8_t retval = 0;
2805 
2806 	/*
2807 	 * XXX early demux data packets so that if_input processing only handles
2808 	 * acks in interrupt context
2809 	 */
2810 	struct mbuf *m, *mh, *mt, *mf;
2811 
2812 	NET_EPOCH_ASSERT();
2813 
2814 	lro_possible = v4_forwarding = v6_forwarding = false;
2815 	ifp = ctx->ifc_ifp;
2816 	mh = mt = NULL;
2817 	MPASS(budget > 0);
2818 	rx_pkts	= rx_bytes = 0;
2819 	if (sctx->isc_flags & IFLIB_HAS_RXCQ)
2820 		cidxp = &rxq->ifr_cq_cidx;
2821 	else
2822 		cidxp = &rxq->ifr_fl[0].ifl_cidx;
2823 	if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) {
2824 		for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2825 			retval |= iflib_fl_refill_all(ctx, fl);
2826 		DBG_COUNTER_INC(rx_unavail);
2827 		return (retval);
2828 	}
2829 
2830 	/* pfil needs the vnet to be set */
2831 	CURVNET_SET_QUIET(ifp->if_vnet);
2832 	for (budget_left = budget; budget_left > 0 && avail > 0;) {
2833 		if (__predict_false(!CTX_ACTIVE(ctx))) {
2834 			DBG_COUNTER_INC(rx_ctx_inactive);
2835 			break;
2836 		}
2837 		/*
2838 		 * Reset client set fields to their default values
2839 		 */
2840 		rxd_info_zero(&ri);
2841 		ri.iri_qsidx = rxq->ifr_id;
2842 		ri.iri_cidx = *cidxp;
2843 		ri.iri_ifp = ifp;
2844 		ri.iri_frags = rxq->ifr_frags;
2845 		err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
2846 
2847 		if (err)
2848 			goto err;
2849 		rx_pkts += 1;
2850 		rx_bytes += ri.iri_len;
2851 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
2852 			*cidxp = ri.iri_cidx;
2853 			/* Update our consumer index */
2854 			/* XXX NB: shurd - check if this is still safe */
2855 			while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0])
2856 				rxq->ifr_cq_cidx -= scctx->isc_nrxd[0];
2857 			/* was this only a completion queue message? */
2858 			if (__predict_false(ri.iri_nfrags == 0))
2859 				continue;
2860 		}
2861 		MPASS(ri.iri_nfrags != 0);
2862 		MPASS(ri.iri_len != 0);
2863 
2864 		/* will advance the cidx on the corresponding free lists */
2865 		m = iflib_rxd_pkt_get(rxq, &ri);
2866 		avail--;
2867 		budget_left--;
2868 		if (avail == 0 && budget_left)
2869 			avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left);
2870 
2871 		if (__predict_false(m == NULL))
2872 			continue;
2873 
2874 		/* imm_pkt: -- cxgb */
2875 		if (mh == NULL)
2876 			mh = mt = m;
2877 		else {
2878 			mt->m_nextpkt = m;
2879 			mt = m;
2880 		}
2881 	}
2882 	CURVNET_RESTORE();
2883 	/* make sure that we can refill faster than drain */
2884 	for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2885 		retval |= iflib_fl_refill_all(ctx, fl);
2886 
2887 	lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO);
2888 	if (lro_enabled)
2889 		iflib_get_ip_forwarding(&rxq->ifr_lc, &v4_forwarding, &v6_forwarding);
2890 	mt = mf = NULL;
2891 	while (mh != NULL) {
2892 		m = mh;
2893 		mh = mh->m_nextpkt;
2894 		m->m_nextpkt = NULL;
2895 #ifndef __NO_STRICT_ALIGNMENT
2896 		if (!IP_ALIGNED(m) && (m = iflib_fixup_rx(m)) == NULL)
2897 			continue;
2898 #endif
2899 		rx_bytes += m->m_pkthdr.len;
2900 		rx_pkts++;
2901 #if defined(INET6) || defined(INET)
2902 		if (lro_enabled) {
2903 			if (!lro_possible) {
2904 				lro_possible = iflib_check_lro_possible(m, v4_forwarding, v6_forwarding);
2905 				if (lro_possible && mf != NULL) {
2906 					ifp->if_input(ifp, mf);
2907 					DBG_COUNTER_INC(rx_if_input);
2908 					mt = mf = NULL;
2909 				}
2910 			}
2911 			if ((m->m_pkthdr.csum_flags & (CSUM_L4_CALC|CSUM_L4_VALID)) ==
2912 			    (CSUM_L4_CALC|CSUM_L4_VALID)) {
2913 				if (lro_possible && tcp_lro_rx(&rxq->ifr_lc, m, 0) == 0)
2914 					continue;
2915 			}
2916 		}
2917 #endif
2918 		if (lro_possible) {
2919 			ifp->if_input(ifp, m);
2920 			DBG_COUNTER_INC(rx_if_input);
2921 			continue;
2922 		}
2923 
2924 		if (mf == NULL)
2925 			mf = m;
2926 		if (mt != NULL)
2927 			mt->m_nextpkt = m;
2928 		mt = m;
2929 	}
2930 	if (mf != NULL) {
2931 		ifp->if_input(ifp, mf);
2932 		DBG_COUNTER_INC(rx_if_input);
2933 	}
2934 
2935 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
2936 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
2937 
2938 	/*
2939 	 * Flush any outstanding LRO work
2940 	 */
2941 #if defined(INET6) || defined(INET)
2942 	tcp_lro_flush_all(&rxq->ifr_lc);
2943 #endif
2944 	if (avail != 0 || iflib_rxd_avail(ctx, rxq, *cidxp, 1) != 0)
2945 		retval |= IFLIB_RXEOF_MORE;
2946 	return (retval);
2947 err:
2948 	STATE_LOCK(ctx);
2949 	ctx->ifc_flags |= IFC_DO_RESET;
2950 	iflib_admin_intr_deferred(ctx);
2951 	STATE_UNLOCK(ctx);
2952 	return (0);
2953 }
2954 
2955 #define TXD_NOTIFY_COUNT(txq) (((txq)->ift_size / (txq)->ift_update_freq)-1)
2956 static inline qidx_t
2957 txq_max_db_deferred(iflib_txq_t txq, qidx_t in_use)
2958 {
2959 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2960 	qidx_t minthresh = txq->ift_size / 8;
2961 	if (in_use > 4*minthresh)
2962 		return (notify_count);
2963 	if (in_use > 2*minthresh)
2964 		return (notify_count >> 1);
2965 	if (in_use > minthresh)
2966 		return (notify_count >> 3);
2967 	return (0);
2968 }
2969 
2970 static inline qidx_t
2971 txq_max_rs_deferred(iflib_txq_t txq)
2972 {
2973 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2974 	qidx_t minthresh = txq->ift_size / 8;
2975 	if (txq->ift_in_use > 4*minthresh)
2976 		return (notify_count);
2977 	if (txq->ift_in_use > 2*minthresh)
2978 		return (notify_count >> 1);
2979 	if (txq->ift_in_use > minthresh)
2980 		return (notify_count >> 2);
2981 	return (2);
2982 }
2983 
2984 #define M_CSUM_FLAGS(m) ((m)->m_pkthdr.csum_flags)
2985 #define M_HAS_VLANTAG(m) (m->m_flags & M_VLANTAG)
2986 
2987 #define TXQ_MAX_DB_DEFERRED(txq, in_use) txq_max_db_deferred((txq), (in_use))
2988 #define TXQ_MAX_RS_DEFERRED(txq) txq_max_rs_deferred(txq)
2989 #define TXQ_MAX_DB_CONSUMED(size) (size >> 4)
2990 
2991 /* forward compatibility for cxgb */
2992 #define FIRST_QSET(ctx) 0
2993 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets)
2994 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets)
2995 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx))
2996 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments))
2997 
2998 /* XXX we should be setting this to something other than zero */
2999 #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh)
3000 #define	MAX_TX_DESC(ctx) max((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max, \
3001     (ctx)->ifc_softc_ctx.isc_tx_nsegments)
3002 
3003 static inline bool
3004 iflib_txd_db_check(if_ctx_t ctx, iflib_txq_t txq, int ring, qidx_t in_use)
3005 {
3006 	qidx_t dbval, max;
3007 	bool rang;
3008 
3009 	rang = false;
3010 	max = TXQ_MAX_DB_DEFERRED(txq, in_use);
3011 	if (ring || txq->ift_db_pending >= max) {
3012 		dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx;
3013 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3014 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
3015 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval);
3016 		txq->ift_db_pending = txq->ift_npending = 0;
3017 		rang = true;
3018 	}
3019 	return (rang);
3020 }
3021 
3022 #ifdef PKT_DEBUG
3023 static void
3024 print_pkt(if_pkt_info_t pi)
3025 {
3026 	printf("pi len:  %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n",
3027 	       pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx);
3028 	printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n",
3029 	       pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag);
3030 	printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n",
3031 	       pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto);
3032 }
3033 #endif
3034 
3035 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO)
3036 #define IS_TX_OFFLOAD4(pi) ((pi)->ipi_csum_flags & (CSUM_IP_TCP | CSUM_IP_TSO))
3037 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO)
3038 #define IS_TX_OFFLOAD6(pi) ((pi)->ipi_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_TSO))
3039 
3040 static int
3041 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp)
3042 {
3043 	if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx;
3044 	struct ether_vlan_header *eh;
3045 	struct mbuf *m;
3046 
3047 	m = *mp;
3048 	if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) &&
3049 	    M_WRITABLE(m) == 0) {
3050 		if ((m = m_dup(m, M_NOWAIT)) == NULL) {
3051 			return (ENOMEM);
3052 		} else {
3053 			m_freem(*mp);
3054 			DBG_COUNTER_INC(tx_frees);
3055 			*mp = m;
3056 		}
3057 	}
3058 
3059 	/*
3060 	 * Determine where frame payload starts.
3061 	 * Jump over vlan headers if already present,
3062 	 * helpful for QinQ too.
3063 	 */
3064 	if (__predict_false(m->m_len < sizeof(*eh))) {
3065 		txq->ift_pullups++;
3066 		if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL))
3067 			return (ENOMEM);
3068 	}
3069 	eh = mtod(m, struct ether_vlan_header *);
3070 	if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
3071 		pi->ipi_etype = ntohs(eh->evl_proto);
3072 		pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
3073 	} else {
3074 		pi->ipi_etype = ntohs(eh->evl_encap_proto);
3075 		pi->ipi_ehdrlen = ETHER_HDR_LEN;
3076 	}
3077 
3078 	switch (pi->ipi_etype) {
3079 #ifdef INET
3080 	case ETHERTYPE_IP:
3081 	{
3082 		struct mbuf *n;
3083 		struct ip *ip = NULL;
3084 		struct tcphdr *th = NULL;
3085 		int minthlen;
3086 
3087 		minthlen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip) + sizeof(*th));
3088 		if (__predict_false(m->m_len < minthlen)) {
3089 			/*
3090 			 * if this code bloat is causing too much of a hit
3091 			 * move it to a separate function and mark it noinline
3092 			 */
3093 			if (m->m_len == pi->ipi_ehdrlen) {
3094 				n = m->m_next;
3095 				MPASS(n);
3096 				if (n->m_len >= sizeof(*ip))  {
3097 					ip = (struct ip *)n->m_data;
3098 					if (n->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3099 						th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3100 				} else {
3101 					txq->ift_pullups++;
3102 					if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
3103 						return (ENOMEM);
3104 					ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3105 				}
3106 			} else {
3107 				txq->ift_pullups++;
3108 				if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
3109 					return (ENOMEM);
3110 				ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3111 				if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3112 					th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3113 			}
3114 		} else {
3115 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3116 			if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3117 				th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3118 		}
3119 		pi->ipi_ip_hlen = ip->ip_hl << 2;
3120 		pi->ipi_ipproto = ip->ip_p;
3121 		pi->ipi_flags |= IPI_TX_IPV4;
3122 
3123 		/* TCP checksum offload may require TCP header length */
3124 		if (IS_TX_OFFLOAD4(pi)) {
3125 			if (__predict_true(pi->ipi_ipproto == IPPROTO_TCP)) {
3126 				if (__predict_false(th == NULL)) {
3127 					txq->ift_pullups++;
3128 					if (__predict_false((m = m_pullup(m, (ip->ip_hl << 2) + sizeof(*th))) == NULL))
3129 						return (ENOMEM);
3130 					th = (struct tcphdr *)((caddr_t)ip + pi->ipi_ip_hlen);
3131 				}
3132 				pi->ipi_tcp_hflags = th->th_flags;
3133 				pi->ipi_tcp_hlen = th->th_off << 2;
3134 				pi->ipi_tcp_seq = th->th_seq;
3135 			}
3136 			if (IS_TSO4(pi)) {
3137 				if (__predict_false(ip->ip_p != IPPROTO_TCP))
3138 					return (ENXIO);
3139 				/*
3140 				 * TSO always requires hardware checksum offload.
3141 				 */
3142 				pi->ipi_csum_flags |= (CSUM_IP_TCP | CSUM_IP);
3143 				th->th_sum = in_pseudo(ip->ip_src.s_addr,
3144 						       ip->ip_dst.s_addr, htons(IPPROTO_TCP));
3145 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3146 				if (sctx->isc_flags & IFLIB_TSO_INIT_IP) {
3147 					ip->ip_sum = 0;
3148 					ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz);
3149 				}
3150 			}
3151 		}
3152 		if ((sctx->isc_flags & IFLIB_NEED_ZERO_CSUM) && (pi->ipi_csum_flags & CSUM_IP))
3153                        ip->ip_sum = 0;
3154 
3155 		break;
3156 	}
3157 #endif
3158 #ifdef INET6
3159 	case ETHERTYPE_IPV6:
3160 	{
3161 		struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
3162 		struct tcphdr *th;
3163 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
3164 
3165 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
3166 			txq->ift_pullups++;
3167 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
3168 				return (ENOMEM);
3169 		}
3170 		th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
3171 
3172 		/* XXX-BZ this will go badly in case of ext hdrs. */
3173 		pi->ipi_ipproto = ip6->ip6_nxt;
3174 		pi->ipi_flags |= IPI_TX_IPV6;
3175 
3176 		/* TCP checksum offload may require TCP header length */
3177 		if (IS_TX_OFFLOAD6(pi)) {
3178 			if (pi->ipi_ipproto == IPPROTO_TCP) {
3179 				if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) {
3180 					txq->ift_pullups++;
3181 					if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL))
3182 						return (ENOMEM);
3183 				}
3184 				pi->ipi_tcp_hflags = th->th_flags;
3185 				pi->ipi_tcp_hlen = th->th_off << 2;
3186 				pi->ipi_tcp_seq = th->th_seq;
3187 			}
3188 			if (IS_TSO6(pi)) {
3189 				if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP))
3190 					return (ENXIO);
3191 				/*
3192 				 * TSO always requires hardware checksum offload.
3193 				 */
3194 				pi->ipi_csum_flags |= CSUM_IP6_TCP;
3195 				th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0);
3196 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3197 			}
3198 		}
3199 		break;
3200 	}
3201 #endif
3202 	default:
3203 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
3204 		pi->ipi_ip_hlen = 0;
3205 		break;
3206 	}
3207 	*mp = m;
3208 
3209 	return (0);
3210 }
3211 
3212 /*
3213  * If dodgy hardware rejects the scatter gather chain we've handed it
3214  * we'll need to remove the mbuf chain from ifsg_m[] before we can add the
3215  * m_defrag'd mbufs
3216  */
3217 static __noinline struct mbuf *
3218 iflib_remove_mbuf(iflib_txq_t txq)
3219 {
3220 	int ntxd, pidx;
3221 	struct mbuf *m, **ifsd_m;
3222 
3223 	ifsd_m = txq->ift_sds.ifsd_m;
3224 	ntxd = txq->ift_size;
3225 	pidx = txq->ift_pidx & (ntxd - 1);
3226 	ifsd_m = txq->ift_sds.ifsd_m;
3227 	m = ifsd_m[pidx];
3228 	ifsd_m[pidx] = NULL;
3229 	bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[pidx]);
3230 	if (txq->ift_sds.ifsd_tso_map != NULL)
3231 		bus_dmamap_unload(txq->ift_tso_buf_tag,
3232 		    txq->ift_sds.ifsd_tso_map[pidx]);
3233 #if MEMORY_LOGGING
3234 	txq->ift_dequeued++;
3235 #endif
3236 	return (m);
3237 }
3238 
3239 static inline caddr_t
3240 calc_next_txd(iflib_txq_t txq, int cidx, uint8_t qid)
3241 {
3242 	qidx_t size;
3243 	int ntxd;
3244 	caddr_t start, end, cur, next;
3245 
3246 	ntxd = txq->ift_size;
3247 	size = txq->ift_txd_size[qid];
3248 	start = txq->ift_ifdi[qid].idi_vaddr;
3249 
3250 	if (__predict_false(size == 0))
3251 		return (start);
3252 	cur = start + size*cidx;
3253 	end = start + size*ntxd;
3254 	next = CACHE_PTR_NEXT(cur);
3255 	return (next < end ? next : start);
3256 }
3257 
3258 /*
3259  * Pad an mbuf to ensure a minimum ethernet frame size.
3260  * min_frame_size is the frame size (less CRC) to pad the mbuf to
3261  */
3262 static __noinline int
3263 iflib_ether_pad(device_t dev, struct mbuf **m_head, uint16_t min_frame_size)
3264 {
3265 	/*
3266 	 * 18 is enough bytes to pad an ARP packet to 46 bytes, and
3267 	 * and ARP message is the smallest common payload I can think of
3268 	 */
3269 	static char pad[18];	/* just zeros */
3270 	int n;
3271 	struct mbuf *new_head;
3272 
3273 	if (!M_WRITABLE(*m_head)) {
3274 		new_head = m_dup(*m_head, M_NOWAIT);
3275 		if (new_head == NULL) {
3276 			m_freem(*m_head);
3277 			device_printf(dev, "cannot pad short frame, m_dup() failed");
3278 			DBG_COUNTER_INC(encap_pad_mbuf_fail);
3279 			DBG_COUNTER_INC(tx_frees);
3280 			return ENOMEM;
3281 		}
3282 		m_freem(*m_head);
3283 		*m_head = new_head;
3284 	}
3285 
3286 	for (n = min_frame_size - (*m_head)->m_pkthdr.len;
3287 	     n > 0; n -= sizeof(pad))
3288 		if (!m_append(*m_head, min(n, sizeof(pad)), pad))
3289 			break;
3290 
3291 	if (n > 0) {
3292 		m_freem(*m_head);
3293 		device_printf(dev, "cannot pad short frame\n");
3294 		DBG_COUNTER_INC(encap_pad_mbuf_fail);
3295 		DBG_COUNTER_INC(tx_frees);
3296 		return (ENOBUFS);
3297 	}
3298 
3299 	return 0;
3300 }
3301 
3302 static int
3303 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp)
3304 {
3305 	if_ctx_t		ctx;
3306 	if_shared_ctx_t		sctx;
3307 	if_softc_ctx_t		scctx;
3308 	bus_dma_tag_t		buf_tag;
3309 	bus_dma_segment_t	*segs;
3310 	struct mbuf		*m_head, **ifsd_m;
3311 	void			*next_txd;
3312 	bus_dmamap_t		map;
3313 	struct if_pkt_info	pi;
3314 	int remap = 0;
3315 	int err, nsegs, ndesc, max_segs, pidx, cidx, next, ntxd;
3316 
3317 	ctx = txq->ift_ctx;
3318 	sctx = ctx->ifc_sctx;
3319 	scctx = &ctx->ifc_softc_ctx;
3320 	segs = txq->ift_segs;
3321 	ntxd = txq->ift_size;
3322 	m_head = *m_headp;
3323 	map = NULL;
3324 
3325 	/*
3326 	 * If we're doing TSO the next descriptor to clean may be quite far ahead
3327 	 */
3328 	cidx = txq->ift_cidx;
3329 	pidx = txq->ift_pidx;
3330 	if (ctx->ifc_flags & IFC_PREFETCH) {
3331 		next = (cidx + CACHE_PTR_INCREMENT) & (ntxd-1);
3332 		if (!(ctx->ifc_flags & IFLIB_HAS_TXCQ)) {
3333 			next_txd = calc_next_txd(txq, cidx, 0);
3334 			prefetch(next_txd);
3335 		}
3336 
3337 		/* prefetch the next cache line of mbuf pointers and flags */
3338 		prefetch(&txq->ift_sds.ifsd_m[next]);
3339 		prefetch(&txq->ift_sds.ifsd_map[next]);
3340 		next = (cidx + CACHE_LINE_SIZE) & (ntxd-1);
3341 	}
3342 	map = txq->ift_sds.ifsd_map[pidx];
3343 	ifsd_m = txq->ift_sds.ifsd_m;
3344 
3345 	if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
3346 		buf_tag = txq->ift_tso_buf_tag;
3347 		max_segs = scctx->isc_tx_tso_segments_max;
3348 		map = txq->ift_sds.ifsd_tso_map[pidx];
3349 		MPASS(buf_tag != NULL);
3350 		MPASS(max_segs > 0);
3351 	} else {
3352 		buf_tag = txq->ift_buf_tag;
3353 		max_segs = scctx->isc_tx_nsegments;
3354 		map = txq->ift_sds.ifsd_map[pidx];
3355 	}
3356 	if ((sctx->isc_flags & IFLIB_NEED_ETHER_PAD) &&
3357 	    __predict_false(m_head->m_pkthdr.len < scctx->isc_min_frame_size)) {
3358 		err = iflib_ether_pad(ctx->ifc_dev, m_headp, scctx->isc_min_frame_size);
3359 		if (err) {
3360 			DBG_COUNTER_INC(encap_txd_encap_fail);
3361 			return err;
3362 		}
3363 	}
3364 	m_head = *m_headp;
3365 
3366 	pkt_info_zero(&pi);
3367 	pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG|M_BCAST|M_MCAST));
3368 	pi.ipi_pidx = pidx;
3369 	pi.ipi_qsidx = txq->ift_id;
3370 	pi.ipi_len = m_head->m_pkthdr.len;
3371 	pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags;
3372 	pi.ipi_vtag = M_HAS_VLANTAG(m_head) ? m_head->m_pkthdr.ether_vtag : 0;
3373 
3374 	/* deliberate bitwise OR to make one condition */
3375 	if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) {
3376 		if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0)) {
3377 			DBG_COUNTER_INC(encap_txd_encap_fail);
3378 			return (err);
3379 		}
3380 		m_head = *m_headp;
3381 	}
3382 
3383 retry:
3384 	err = bus_dmamap_load_mbuf_sg(buf_tag, map, m_head, segs, &nsegs,
3385 	    BUS_DMA_NOWAIT);
3386 defrag:
3387 	if (__predict_false(err)) {
3388 		switch (err) {
3389 		case EFBIG:
3390 			/* try collapse once and defrag once */
3391 			if (remap == 0) {
3392 				m_head = m_collapse(*m_headp, M_NOWAIT, max_segs);
3393 				/* try defrag if collapsing fails */
3394 				if (m_head == NULL)
3395 					remap++;
3396 			}
3397 			if (remap == 1) {
3398 				txq->ift_mbuf_defrag++;
3399 				m_head = m_defrag(*m_headp, M_NOWAIT);
3400 			}
3401 			/*
3402 			 * remap should never be >1 unless bus_dmamap_load_mbuf_sg
3403 			 * failed to map an mbuf that was run through m_defrag
3404 			 */
3405 			MPASS(remap <= 1);
3406 			if (__predict_false(m_head == NULL || remap > 1))
3407 				goto defrag_failed;
3408 			remap++;
3409 			*m_headp = m_head;
3410 			goto retry;
3411 			break;
3412 		case ENOMEM:
3413 			txq->ift_no_tx_dma_setup++;
3414 			break;
3415 		default:
3416 			txq->ift_no_tx_dma_setup++;
3417 			m_freem(*m_headp);
3418 			DBG_COUNTER_INC(tx_frees);
3419 			*m_headp = NULL;
3420 			break;
3421 		}
3422 		txq->ift_map_failed++;
3423 		DBG_COUNTER_INC(encap_load_mbuf_fail);
3424 		DBG_COUNTER_INC(encap_txd_encap_fail);
3425 		return (err);
3426 	}
3427 	ifsd_m[pidx] = m_head;
3428 	/*
3429 	 * XXX assumes a 1 to 1 relationship between segments and
3430 	 *        descriptors - this does not hold true on all drivers, e.g.
3431 	 *        cxgb
3432 	 */
3433 	if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) {
3434 		txq->ift_no_desc_avail++;
3435 		bus_dmamap_unload(buf_tag, map);
3436 		DBG_COUNTER_INC(encap_txq_avail_fail);
3437 		DBG_COUNTER_INC(encap_txd_encap_fail);
3438 		if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0)
3439 			GROUPTASK_ENQUEUE(&txq->ift_task);
3440 		return (ENOBUFS);
3441 	}
3442 	/*
3443 	 * On Intel cards we can greatly reduce the number of TX interrupts
3444 	 * we see by only setting report status on every Nth descriptor.
3445 	 * However, this also means that the driver will need to keep track
3446 	 * of the descriptors that RS was set on to check them for the DD bit.
3447 	 */
3448 	txq->ift_rs_pending += nsegs + 1;
3449 	if (txq->ift_rs_pending > TXQ_MAX_RS_DEFERRED(txq) ||
3450 	     iflib_no_tx_batch || (TXQ_AVAIL(txq) - nsegs) <= MAX_TX_DESC(ctx) + 2) {
3451 		pi.ipi_flags |= IPI_TX_INTR;
3452 		txq->ift_rs_pending = 0;
3453 	}
3454 
3455 	pi.ipi_segs = segs;
3456 	pi.ipi_nsegs = nsegs;
3457 
3458 	MPASS(pidx >= 0 && pidx < txq->ift_size);
3459 #ifdef PKT_DEBUG
3460 	print_pkt(&pi);
3461 #endif
3462 	if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) {
3463 		bus_dmamap_sync(buf_tag, map, BUS_DMASYNC_PREWRITE);
3464 		DBG_COUNTER_INC(tx_encap);
3465 		MPASS(pi.ipi_new_pidx < txq->ift_size);
3466 
3467 		ndesc = pi.ipi_new_pidx - pi.ipi_pidx;
3468 		if (pi.ipi_new_pidx < pi.ipi_pidx) {
3469 			ndesc += txq->ift_size;
3470 			txq->ift_gen = 1;
3471 		}
3472 		/*
3473 		 * drivers can need as many as
3474 		 * two sentinels
3475 		 */
3476 		MPASS(ndesc <= pi.ipi_nsegs + 2);
3477 		MPASS(pi.ipi_new_pidx != pidx);
3478 		MPASS(ndesc > 0);
3479 		txq->ift_in_use += ndesc;
3480 
3481 		/*
3482 		 * We update the last software descriptor again here because there may
3483 		 * be a sentinel and/or there may be more mbufs than segments
3484 		 */
3485 		txq->ift_pidx = pi.ipi_new_pidx;
3486 		txq->ift_npending += pi.ipi_ndescs;
3487 	} else {
3488 		*m_headp = m_head = iflib_remove_mbuf(txq);
3489 		if (err == EFBIG) {
3490 			txq->ift_txd_encap_efbig++;
3491 			if (remap < 2) {
3492 				remap = 1;
3493 				goto defrag;
3494 			}
3495 		}
3496 		goto defrag_failed;
3497 	}
3498 	/*
3499 	 * err can't possibly be non-zero here, so we don't neet to test it
3500 	 * to see if we need to DBG_COUNTER_INC(encap_txd_encap_fail).
3501 	 */
3502 	return (err);
3503 
3504 defrag_failed:
3505 	txq->ift_mbuf_defrag_failed++;
3506 	txq->ift_map_failed++;
3507 	m_freem(*m_headp);
3508 	DBG_COUNTER_INC(tx_frees);
3509 	*m_headp = NULL;
3510 	DBG_COUNTER_INC(encap_txd_encap_fail);
3511 	return (ENOMEM);
3512 }
3513 
3514 static void
3515 iflib_tx_desc_free(iflib_txq_t txq, int n)
3516 {
3517 	uint32_t qsize, cidx, mask, gen;
3518 	struct mbuf *m, **ifsd_m;
3519 	bool do_prefetch;
3520 
3521 	cidx = txq->ift_cidx;
3522 	gen = txq->ift_gen;
3523 	qsize = txq->ift_size;
3524 	mask = qsize-1;
3525 	ifsd_m = txq->ift_sds.ifsd_m;
3526 	do_prefetch = (txq->ift_ctx->ifc_flags & IFC_PREFETCH);
3527 
3528 	while (n-- > 0) {
3529 		if (do_prefetch) {
3530 			prefetch(ifsd_m[(cidx + 3) & mask]);
3531 			prefetch(ifsd_m[(cidx + 4) & mask]);
3532 		}
3533 		if ((m = ifsd_m[cidx]) != NULL) {
3534 			prefetch(&ifsd_m[(cidx + CACHE_PTR_INCREMENT) & mask]);
3535 			if (m->m_pkthdr.csum_flags & CSUM_TSO) {
3536 				bus_dmamap_sync(txq->ift_tso_buf_tag,
3537 				    txq->ift_sds.ifsd_tso_map[cidx],
3538 				    BUS_DMASYNC_POSTWRITE);
3539 				bus_dmamap_unload(txq->ift_tso_buf_tag,
3540 				    txq->ift_sds.ifsd_tso_map[cidx]);
3541 			} else {
3542 				bus_dmamap_sync(txq->ift_buf_tag,
3543 				    txq->ift_sds.ifsd_map[cidx],
3544 				    BUS_DMASYNC_POSTWRITE);
3545 				bus_dmamap_unload(txq->ift_buf_tag,
3546 				    txq->ift_sds.ifsd_map[cidx]);
3547 			}
3548 			/* XXX we don't support any drivers that batch packets yet */
3549 			MPASS(m->m_nextpkt == NULL);
3550 			m_freem(m);
3551 			ifsd_m[cidx] = NULL;
3552 #if MEMORY_LOGGING
3553 			txq->ift_dequeued++;
3554 #endif
3555 			DBG_COUNTER_INC(tx_frees);
3556 		}
3557 		if (__predict_false(++cidx == qsize)) {
3558 			cidx = 0;
3559 			gen = 0;
3560 		}
3561 	}
3562 	txq->ift_cidx = cidx;
3563 	txq->ift_gen = gen;
3564 }
3565 
3566 static __inline int
3567 iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh)
3568 {
3569 	int reclaim;
3570 	if_ctx_t ctx = txq->ift_ctx;
3571 
3572 	KASSERT(thresh >= 0, ("invalid threshold to reclaim"));
3573 	MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size);
3574 
3575 	/*
3576 	 * Need a rate-limiting check so that this isn't called every time
3577 	 */
3578 	iflib_tx_credits_update(ctx, txq);
3579 	reclaim = DESC_RECLAIMABLE(txq);
3580 
3581 	if (reclaim <= thresh /* + MAX_TX_DESC(txq->ift_ctx) */) {
3582 #ifdef INVARIANTS
3583 		if (iflib_verbose_debug) {
3584 			printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __FUNCTION__,
3585 			       txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments,
3586 			       reclaim, thresh);
3587 
3588 		}
3589 #endif
3590 		return (0);
3591 	}
3592 	iflib_tx_desc_free(txq, reclaim);
3593 	txq->ift_cleaned += reclaim;
3594 	txq->ift_in_use -= reclaim;
3595 
3596 	return (reclaim);
3597 }
3598 
3599 static struct mbuf **
3600 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset, int remaining)
3601 {
3602 	int next, size;
3603 	struct mbuf **items;
3604 
3605 	size = r->size;
3606 	next = (cidx + CACHE_PTR_INCREMENT) & (size-1);
3607 	items = __DEVOLATILE(struct mbuf **, &r->items[0]);
3608 
3609 	prefetch(items[(cidx + offset) & (size-1)]);
3610 	if (remaining > 1) {
3611 		prefetch2cachelines(&items[next]);
3612 		prefetch2cachelines(items[(cidx + offset + 1) & (size-1)]);
3613 		prefetch2cachelines(items[(cidx + offset + 2) & (size-1)]);
3614 		prefetch2cachelines(items[(cidx + offset + 3) & (size-1)]);
3615 	}
3616 	return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (size-1)]));
3617 }
3618 
3619 static void
3620 iflib_txq_check_drain(iflib_txq_t txq, int budget)
3621 {
3622 
3623 	ifmp_ring_check_drainage(txq->ift_br, budget);
3624 }
3625 
3626 static uint32_t
3627 iflib_txq_can_drain(struct ifmp_ring *r)
3628 {
3629 	iflib_txq_t txq = r->cookie;
3630 	if_ctx_t ctx = txq->ift_ctx;
3631 
3632 	if (TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2)
3633 		return (1);
3634 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3635 	    BUS_DMASYNC_POSTREAD);
3636 	return (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id,
3637 	    false));
3638 }
3639 
3640 static uint32_t
3641 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3642 {
3643 	iflib_txq_t txq = r->cookie;
3644 	if_ctx_t ctx = txq->ift_ctx;
3645 	if_t ifp = ctx->ifc_ifp;
3646 	struct mbuf *m, **mp;
3647 	int avail, bytes_sent, consumed, count, err, i, in_use_prev;
3648 	int mcast_sent, pkt_sent, reclaimed, txq_avail;
3649 	bool do_prefetch, rang, ring;
3650 
3651 	if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) ||
3652 			    !LINK_ACTIVE(ctx))) {
3653 		DBG_COUNTER_INC(txq_drain_notready);
3654 		return (0);
3655 	}
3656 	reclaimed = iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
3657 	rang = iflib_txd_db_check(ctx, txq, reclaimed, txq->ift_in_use);
3658 	avail = IDXDIFF(pidx, cidx, r->size);
3659 	if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) {
3660 		DBG_COUNTER_INC(txq_drain_flushing);
3661 		for (i = 0; i < avail; i++) {
3662 			if (__predict_true(r->items[(cidx + i) & (r->size-1)] != (void *)txq))
3663 				m_free(r->items[(cidx + i) & (r->size-1)]);
3664 			r->items[(cidx + i) & (r->size-1)] = NULL;
3665 		}
3666 		return (avail);
3667 	}
3668 
3669 	if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3670 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3671 		CALLOUT_LOCK(txq);
3672 		callout_stop(&txq->ift_timer);
3673 		CALLOUT_UNLOCK(txq);
3674 		DBG_COUNTER_INC(txq_drain_oactive);
3675 		return (0);
3676 	}
3677 	if (reclaimed)
3678 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3679 	consumed = mcast_sent = bytes_sent = pkt_sent = 0;
3680 	count = MIN(avail, TX_BATCH_SIZE);
3681 #ifdef INVARIANTS
3682 	if (iflib_verbose_debug)
3683 		printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __FUNCTION__,
3684 		       avail, ctx->ifc_flags, TXQ_AVAIL(txq));
3685 #endif
3686 	do_prefetch = (ctx->ifc_flags & IFC_PREFETCH);
3687 	txq_avail = TXQ_AVAIL(txq);
3688 	err = 0;
3689 	for (i = 0; i < count && txq_avail > MAX_TX_DESC(ctx) + 2; i++) {
3690 		int rem = do_prefetch ? count - i : 0;
3691 
3692 		mp = _ring_peek_one(r, cidx, i, rem);
3693 		MPASS(mp != NULL && *mp != NULL);
3694 		if (__predict_false(*mp == (struct mbuf *)txq)) {
3695 			consumed++;
3696 			continue;
3697 		}
3698 		in_use_prev = txq->ift_in_use;
3699 		err = iflib_encap(txq, mp);
3700 		if (__predict_false(err)) {
3701 			/* no room - bail out */
3702 			if (err == ENOBUFS)
3703 				break;
3704 			consumed++;
3705 			/* we can't send this packet - skip it */
3706 			continue;
3707 		}
3708 		consumed++;
3709 		pkt_sent++;
3710 		m = *mp;
3711 		DBG_COUNTER_INC(tx_sent);
3712 		bytes_sent += m->m_pkthdr.len;
3713 		mcast_sent += !!(m->m_flags & M_MCAST);
3714 		txq_avail = TXQ_AVAIL(txq);
3715 
3716 		txq->ift_db_pending += (txq->ift_in_use - in_use_prev);
3717 		ETHER_BPF_MTAP(ifp, m);
3718 		if (__predict_false(!(ifp->if_drv_flags & IFF_DRV_RUNNING)))
3719 			break;
3720 		rang = iflib_txd_db_check(ctx, txq, false, in_use_prev);
3721 	}
3722 
3723 	/* deliberate use of bitwise or to avoid gratuitous short-circuit */
3724 	ring = rang ? false  : (iflib_min_tx_latency | err) || (TXQ_AVAIL(txq) < MAX_TX_DESC(ctx));
3725 	iflib_txd_db_check(ctx, txq, ring, txq->ift_in_use);
3726 	if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
3727 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
3728 	if (mcast_sent)
3729 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
3730 #ifdef INVARIANTS
3731 	if (iflib_verbose_debug)
3732 		printf("consumed=%d\n", consumed);
3733 #endif
3734 	return (consumed);
3735 }
3736 
3737 static uint32_t
3738 iflib_txq_drain_always(struct ifmp_ring *r)
3739 {
3740 	return (1);
3741 }
3742 
3743 static uint32_t
3744 iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3745 {
3746 	int i, avail;
3747 	struct mbuf **mp;
3748 	iflib_txq_t txq;
3749 
3750 	txq = r->cookie;
3751 
3752 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3753 	CALLOUT_LOCK(txq);
3754 	callout_stop(&txq->ift_timer);
3755 	CALLOUT_UNLOCK(txq);
3756 
3757 	avail = IDXDIFF(pidx, cidx, r->size);
3758 	for (i = 0; i < avail; i++) {
3759 		mp = _ring_peek_one(r, cidx, i, avail - i);
3760 		if (__predict_false(*mp == (struct mbuf *)txq))
3761 			continue;
3762 		m_freem(*mp);
3763 		DBG_COUNTER_INC(tx_frees);
3764 	}
3765 	MPASS(ifmp_ring_is_stalled(r) == 0);
3766 	return (avail);
3767 }
3768 
3769 static void
3770 iflib_ifmp_purge(iflib_txq_t txq)
3771 {
3772 	struct ifmp_ring *r;
3773 
3774 	r = txq->ift_br;
3775 	r->drain = iflib_txq_drain_free;
3776 	r->can_drain = iflib_txq_drain_always;
3777 
3778 	ifmp_ring_check_drainage(r, r->size);
3779 
3780 	r->drain = iflib_txq_drain;
3781 	r->can_drain = iflib_txq_can_drain;
3782 }
3783 
3784 static void
3785 _task_fn_tx(void *context)
3786 {
3787 	iflib_txq_t txq = context;
3788 	if_ctx_t ctx = txq->ift_ctx;
3789 	if_t ifp = ctx->ifc_ifp;
3790 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
3791 
3792 #ifdef IFLIB_DIAGNOSTICS
3793 	txq->ift_cpu_exec_count[curcpu]++;
3794 #endif
3795 	if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
3796 		return;
3797 #ifdef DEV_NETMAP
3798 	if ((if_getcapenable(ifp) & IFCAP_NETMAP) &&
3799 	    netmap_tx_irq(ifp, txq->ift_id))
3800 		goto skip_ifmp;
3801 #endif
3802 #ifdef ALTQ
3803 	if (ALTQ_IS_ENABLED(&ifp->if_snd))
3804 		iflib_altq_if_start(ifp);
3805 #endif
3806 	if (txq->ift_db_pending)
3807 		ifmp_ring_enqueue(txq->ift_br, (void **)&txq, 1, TX_BATCH_SIZE, abdicate);
3808 	else if (!abdicate)
3809 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3810 	/*
3811 	 * When abdicating, we always need to check drainage, not just when we don't enqueue
3812 	 */
3813 	if (abdicate)
3814 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3815 #ifdef DEV_NETMAP
3816 skip_ifmp:
3817 #endif
3818 	if (ctx->ifc_flags & IFC_LEGACY)
3819 		IFDI_INTR_ENABLE(ctx);
3820 	else
3821 		IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3822 }
3823 
3824 static void
3825 _task_fn_rx(void *context)
3826 {
3827 	iflib_rxq_t rxq = context;
3828 	if_ctx_t ctx = rxq->ifr_ctx;
3829 	uint8_t more;
3830 	uint16_t budget;
3831 #ifdef DEV_NETMAP
3832 	u_int work = 0;
3833 	int nmirq;
3834 #endif
3835 
3836 #ifdef IFLIB_DIAGNOSTICS
3837 	rxq->ifr_cpu_exec_count[curcpu]++;
3838 #endif
3839 	DBG_COUNTER_INC(task_fn_rxs);
3840 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3841 		return;
3842 #ifdef DEV_NETMAP
3843 	nmirq = netmap_rx_irq(ctx->ifc_ifp, rxq->ifr_id, &work);
3844 	if (nmirq != NM_IRQ_PASS) {
3845 		more = (nmirq == NM_IRQ_RESCHED) ? IFLIB_RXEOF_MORE : 0;
3846 		goto skip_rxeof;
3847 	}
3848 #endif
3849 	budget = ctx->ifc_sysctl_rx_budget;
3850 	if (budget == 0)
3851 		budget = 16;	/* XXX */
3852 	more = iflib_rxeof(rxq, budget);
3853 #ifdef DEV_NETMAP
3854 skip_rxeof:
3855 #endif
3856 	if ((more & IFLIB_RXEOF_MORE) == 0) {
3857 		if (ctx->ifc_flags & IFC_LEGACY)
3858 			IFDI_INTR_ENABLE(ctx);
3859 		else
3860 			IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
3861 		DBG_COUNTER_INC(rx_intr_enables);
3862 	}
3863 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3864 		return;
3865 
3866 	if (more & IFLIB_RXEOF_MORE)
3867 		GROUPTASK_ENQUEUE(&rxq->ifr_task);
3868 	else if (more & IFLIB_RXEOF_EMPTY)
3869 		callout_reset_curcpu(&rxq->ifr_watchdog, 1, &_task_fn_rx_watchdog, rxq);
3870 }
3871 
3872 static void
3873 _task_fn_admin(void *context)
3874 {
3875 	if_ctx_t ctx = context;
3876 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
3877 	iflib_txq_t txq;
3878 	int i;
3879 	bool oactive, running, do_reset, do_watchdog, in_detach;
3880 	uint32_t reset_on = hz / 2;
3881 
3882 	STATE_LOCK(ctx);
3883 	running = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING);
3884 	oactive = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE);
3885 	do_reset = (ctx->ifc_flags & IFC_DO_RESET);
3886 	do_watchdog = (ctx->ifc_flags & IFC_DO_WATCHDOG);
3887 	in_detach = (ctx->ifc_flags & IFC_IN_DETACH);
3888 	ctx->ifc_flags &= ~(IFC_DO_RESET|IFC_DO_WATCHDOG);
3889 	STATE_UNLOCK(ctx);
3890 
3891 	if ((!running && !oactive) && !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
3892 		return;
3893 	if (in_detach)
3894 		return;
3895 
3896 	CTX_LOCK(ctx);
3897 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3898 		CALLOUT_LOCK(txq);
3899 		callout_stop(&txq->ift_timer);
3900 		CALLOUT_UNLOCK(txq);
3901 	}
3902 	if (do_watchdog) {
3903 		ctx->ifc_watchdog_events++;
3904 		IFDI_WATCHDOG_RESET(ctx);
3905 	}
3906 	IFDI_UPDATE_ADMIN_STATUS(ctx);
3907 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3908 #ifdef DEV_NETMAP
3909 		reset_on = hz / 2;
3910 		if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP)
3911 			iflib_netmap_timer_adjust(ctx, txq, &reset_on);
3912 #endif
3913 		callout_reset_on(&txq->ift_timer, reset_on, iflib_timer, txq, txq->ift_timer.c_cpu);
3914 	}
3915 	IFDI_LINK_INTR_ENABLE(ctx);
3916 	if (do_reset)
3917 		iflib_if_init_locked(ctx);
3918 	CTX_UNLOCK(ctx);
3919 
3920 	if (LINK_ACTIVE(ctx) == 0)
3921 		return;
3922 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3923 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
3924 }
3925 
3926 
3927 static void
3928 _task_fn_iov(void *context)
3929 {
3930 	if_ctx_t ctx = context;
3931 
3932 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING) &&
3933 	    !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
3934 		return;
3935 
3936 	CTX_LOCK(ctx);
3937 	IFDI_VFLR_HANDLE(ctx);
3938 	CTX_UNLOCK(ctx);
3939 }
3940 
3941 static int
3942 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)
3943 {
3944 	int err;
3945 	if_int_delay_info_t info;
3946 	if_ctx_t ctx;
3947 
3948 	info = (if_int_delay_info_t)arg1;
3949 	ctx = info->iidi_ctx;
3950 	info->iidi_req = req;
3951 	info->iidi_oidp = oidp;
3952 	CTX_LOCK(ctx);
3953 	err = IFDI_SYSCTL_INT_DELAY(ctx, info);
3954 	CTX_UNLOCK(ctx);
3955 	return (err);
3956 }
3957 
3958 /*********************************************************************
3959  *
3960  *  IFNET FUNCTIONS
3961  *
3962  **********************************************************************/
3963 
3964 static void
3965 iflib_if_init_locked(if_ctx_t ctx)
3966 {
3967 	iflib_stop(ctx);
3968 	iflib_init_locked(ctx);
3969 }
3970 
3971 
3972 static void
3973 iflib_if_init(void *arg)
3974 {
3975 	if_ctx_t ctx = arg;
3976 
3977 	CTX_LOCK(ctx);
3978 	iflib_if_init_locked(ctx);
3979 	CTX_UNLOCK(ctx);
3980 }
3981 
3982 static int
3983 iflib_if_transmit(if_t ifp, struct mbuf *m)
3984 {
3985 	if_ctx_t	ctx = if_getsoftc(ifp);
3986 
3987 	iflib_txq_t txq;
3988 	int err, qidx;
3989 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
3990 
3991 	if (__predict_false((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) {
3992 		DBG_COUNTER_INC(tx_frees);
3993 		m_freem(m);
3994 		return (ENETDOWN);
3995 	}
3996 
3997 	MPASS(m->m_nextpkt == NULL);
3998 	/* ALTQ-enabled interfaces always use queue 0. */
3999 	qidx = 0;
4000 	if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m) && !ALTQ_IS_ENABLED(&ifp->if_snd))
4001 		qidx = QIDX(ctx, m);
4002 	/*
4003 	 * XXX calculate buf_ring based on flowid (divvy up bits?)
4004 	 */
4005 	txq = &ctx->ifc_txqs[qidx];
4006 
4007 #ifdef DRIVER_BACKPRESSURE
4008 	if (txq->ift_closed) {
4009 		while (m != NULL) {
4010 			next = m->m_nextpkt;
4011 			m->m_nextpkt = NULL;
4012 			m_freem(m);
4013 			DBG_COUNTER_INC(tx_frees);
4014 			m = next;
4015 		}
4016 		return (ENOBUFS);
4017 	}
4018 #endif
4019 #ifdef notyet
4020 	qidx = count = 0;
4021 	mp = marr;
4022 	next = m;
4023 	do {
4024 		count++;
4025 		next = next->m_nextpkt;
4026 	} while (next != NULL);
4027 
4028 	if (count > nitems(marr))
4029 		if ((mp = malloc(count*sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) {
4030 			/* XXX check nextpkt */
4031 			m_freem(m);
4032 			/* XXX simplify for now */
4033 			DBG_COUNTER_INC(tx_frees);
4034 			return (ENOBUFS);
4035 		}
4036 	for (next = m, i = 0; next != NULL; i++) {
4037 		mp[i] = next;
4038 		next = next->m_nextpkt;
4039 		mp[i]->m_nextpkt = NULL;
4040 	}
4041 #endif
4042 	DBG_COUNTER_INC(tx_seen);
4043 	err = ifmp_ring_enqueue(txq->ift_br, (void **)&m, 1, TX_BATCH_SIZE, abdicate);
4044 
4045 	if (abdicate)
4046 		GROUPTASK_ENQUEUE(&txq->ift_task);
4047  	if (err) {
4048 		if (!abdicate)
4049 			GROUPTASK_ENQUEUE(&txq->ift_task);
4050 		/* support forthcoming later */
4051 #ifdef DRIVER_BACKPRESSURE
4052 		txq->ift_closed = TRUE;
4053 #endif
4054 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
4055 		m_freem(m);
4056 		DBG_COUNTER_INC(tx_frees);
4057 	}
4058 
4059 	return (err);
4060 }
4061 
4062 #ifdef ALTQ
4063 /*
4064  * The overall approach to integrating iflib with ALTQ is to continue to use
4065  * the iflib mp_ring machinery between the ALTQ queue(s) and the hardware
4066  * ring.  Technically, when using ALTQ, queueing to an intermediate mp_ring
4067  * is redundant/unnecessary, but doing so minimizes the amount of
4068  * ALTQ-specific code required in iflib.  It is assumed that the overhead of
4069  * redundantly queueing to an intermediate mp_ring is swamped by the
4070  * performance limitations inherent in using ALTQ.
4071  *
4072  * When ALTQ support is compiled in, all iflib drivers will use a transmit
4073  * routine, iflib_altq_if_transmit(), that checks if ALTQ is enabled for the
4074  * given interface.  If ALTQ is enabled for an interface, then all
4075  * transmitted packets for that interface will be submitted to the ALTQ
4076  * subsystem via IFQ_ENQUEUE().  We don't use the legacy if_transmit()
4077  * implementation because it uses IFQ_HANDOFF(), which will duplicatively
4078  * update stats that the iflib machinery handles, and which is sensitve to
4079  * the disused IFF_DRV_OACTIVE flag.  Additionally, iflib_altq_if_start()
4080  * will be installed as the start routine for use by ALTQ facilities that
4081  * need to trigger queue drains on a scheduled basis.
4082  *
4083  */
4084 static void
4085 iflib_altq_if_start(if_t ifp)
4086 {
4087 	struct ifaltq *ifq = &ifp->if_snd;
4088 	struct mbuf *m;
4089 
4090 	IFQ_LOCK(ifq);
4091 	IFQ_DEQUEUE_NOLOCK(ifq, m);
4092 	while (m != NULL) {
4093 		iflib_if_transmit(ifp, m);
4094 		IFQ_DEQUEUE_NOLOCK(ifq, m);
4095 	}
4096 	IFQ_UNLOCK(ifq);
4097 }
4098 
4099 static int
4100 iflib_altq_if_transmit(if_t ifp, struct mbuf *m)
4101 {
4102 	int err;
4103 
4104 	if (ALTQ_IS_ENABLED(&ifp->if_snd)) {
4105 		IFQ_ENQUEUE(&ifp->if_snd, m, err);
4106 		if (err == 0)
4107 			iflib_altq_if_start(ifp);
4108 	} else
4109 		err = iflib_if_transmit(ifp, m);
4110 
4111 	return (err);
4112 }
4113 #endif /* ALTQ */
4114 
4115 static void
4116 iflib_if_qflush(if_t ifp)
4117 {
4118 	if_ctx_t ctx = if_getsoftc(ifp);
4119 	iflib_txq_t txq = ctx->ifc_txqs;
4120 	int i;
4121 
4122 	STATE_LOCK(ctx);
4123 	ctx->ifc_flags |= IFC_QFLUSH;
4124 	STATE_UNLOCK(ctx);
4125 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
4126 		while (!(ifmp_ring_is_idle(txq->ift_br) || ifmp_ring_is_stalled(txq->ift_br)))
4127 			iflib_txq_check_drain(txq, 0);
4128 	STATE_LOCK(ctx);
4129 	ctx->ifc_flags &= ~IFC_QFLUSH;
4130 	STATE_UNLOCK(ctx);
4131 
4132 	/*
4133 	 * When ALTQ is enabled, this will also take care of purging the
4134 	 * ALTQ queue(s).
4135 	 */
4136 	if_qflush(ifp);
4137 }
4138 
4139 
4140 #define IFCAP_FLAGS (IFCAP_HWCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \
4141 		     IFCAP_TSO | IFCAP_VLAN_HWTAGGING | IFCAP_HWSTATS | \
4142 		     IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | \
4143 		     IFCAP_VLAN_HWTSO | IFCAP_VLAN_HWCSUM | IFCAP_NOMAP)
4144 
4145 static int
4146 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data)
4147 {
4148 	if_ctx_t ctx = if_getsoftc(ifp);
4149 	struct ifreq	*ifr = (struct ifreq *)data;
4150 #if defined(INET) || defined(INET6)
4151 	struct ifaddr	*ifa = (struct ifaddr *)data;
4152 #endif
4153 	bool		avoid_reset = false;
4154 	int		err = 0, reinit = 0, bits;
4155 
4156 	switch (command) {
4157 	case SIOCSIFADDR:
4158 #ifdef INET
4159 		if (ifa->ifa_addr->sa_family == AF_INET)
4160 			avoid_reset = true;
4161 #endif
4162 #ifdef INET6
4163 		if (ifa->ifa_addr->sa_family == AF_INET6)
4164 			avoid_reset = true;
4165 #endif
4166 		/*
4167 		** Calling init results in link renegotiation,
4168 		** so we avoid doing it when possible.
4169 		*/
4170 		if (avoid_reset) {
4171 			if_setflagbits(ifp, IFF_UP,0);
4172 			if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
4173 				reinit = 1;
4174 #ifdef INET
4175 			if (!(if_getflags(ifp) & IFF_NOARP))
4176 				arp_ifinit(ifp, ifa);
4177 #endif
4178 		} else
4179 			err = ether_ioctl(ifp, command, data);
4180 		break;
4181 	case SIOCSIFMTU:
4182 		CTX_LOCK(ctx);
4183 		if (ifr->ifr_mtu == if_getmtu(ifp)) {
4184 			CTX_UNLOCK(ctx);
4185 			break;
4186 		}
4187 		bits = if_getdrvflags(ifp);
4188 		/* stop the driver and free any clusters before proceeding */
4189 		iflib_stop(ctx);
4190 
4191 		if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) {
4192 			STATE_LOCK(ctx);
4193 			if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size)
4194 				ctx->ifc_flags |= IFC_MULTISEG;
4195 			else
4196 				ctx->ifc_flags &= ~IFC_MULTISEG;
4197 			STATE_UNLOCK(ctx);
4198 			err = if_setmtu(ifp, ifr->ifr_mtu);
4199 		}
4200 		iflib_init_locked(ctx);
4201 		STATE_LOCK(ctx);
4202 		if_setdrvflags(ifp, bits);
4203 		STATE_UNLOCK(ctx);
4204 		CTX_UNLOCK(ctx);
4205 		break;
4206 	case SIOCSIFFLAGS:
4207 		CTX_LOCK(ctx);
4208 		if (if_getflags(ifp) & IFF_UP) {
4209 			if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4210 				if ((if_getflags(ifp) ^ ctx->ifc_if_flags) &
4211 				    (IFF_PROMISC | IFF_ALLMULTI)) {
4212 					CTX_UNLOCK(ctx);
4213 					err = IFDI_PROMISC_SET(ctx, if_getflags(ifp));
4214 					CTX_LOCK(ctx);
4215 				}
4216 			} else
4217 				reinit = 1;
4218 		} else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4219 			iflib_stop(ctx);
4220 		}
4221 		ctx->ifc_if_flags = if_getflags(ifp);
4222 		CTX_UNLOCK(ctx);
4223 		break;
4224 	case SIOCADDMULTI:
4225 	case SIOCDELMULTI:
4226 		if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4227 			CTX_LOCK(ctx);
4228 			IFDI_INTR_DISABLE(ctx);
4229 			IFDI_MULTI_SET(ctx);
4230 			IFDI_INTR_ENABLE(ctx);
4231 			CTX_UNLOCK(ctx);
4232 		}
4233 		break;
4234 	case SIOCSIFMEDIA:
4235 		CTX_LOCK(ctx);
4236 		IFDI_MEDIA_SET(ctx);
4237 		CTX_UNLOCK(ctx);
4238 		/* FALLTHROUGH */
4239 	case SIOCGIFMEDIA:
4240 	case SIOCGIFXMEDIA:
4241 		err = ifmedia_ioctl(ifp, ifr, ctx->ifc_mediap, command);
4242 		break;
4243 	case SIOCGI2C:
4244 	{
4245 		struct ifi2creq i2c;
4246 
4247 		err = copyin(ifr_data_get_ptr(ifr), &i2c, sizeof(i2c));
4248 		if (err != 0)
4249 			break;
4250 		if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) {
4251 			err = EINVAL;
4252 			break;
4253 		}
4254 		if (i2c.len > sizeof(i2c.data)) {
4255 			err = EINVAL;
4256 			break;
4257 		}
4258 
4259 		if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0)
4260 			err = copyout(&i2c, ifr_data_get_ptr(ifr),
4261 			    sizeof(i2c));
4262 		break;
4263 	}
4264 	case SIOCSIFCAP:
4265 	{
4266 		int mask, setmask, oldmask;
4267 
4268 		oldmask = if_getcapenable(ifp);
4269 		mask = ifr->ifr_reqcap ^ oldmask;
4270 		mask &= ctx->ifc_softc_ctx.isc_capabilities | IFCAP_NOMAP;
4271 		setmask = 0;
4272 #ifdef TCP_OFFLOAD
4273 		setmask |= mask & (IFCAP_TOE4|IFCAP_TOE6);
4274 #endif
4275 		setmask |= (mask & IFCAP_FLAGS);
4276 		setmask |= (mask & IFCAP_WOL);
4277 
4278 		/*
4279 		 * If any RX csum has changed, change all the ones that
4280 		 * are supported by the driver.
4281 		 */
4282 		if (setmask & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6)) {
4283 			setmask |= ctx->ifc_softc_ctx.isc_capabilities &
4284 			    (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6);
4285 		}
4286 
4287 		/*
4288 		 * want to ensure that traffic has stopped before we change any of the flags
4289 		 */
4290 		if (setmask) {
4291 			CTX_LOCK(ctx);
4292 			bits = if_getdrvflags(ifp);
4293 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4294 				iflib_stop(ctx);
4295 			STATE_LOCK(ctx);
4296 			if_togglecapenable(ifp, setmask);
4297 			STATE_UNLOCK(ctx);
4298 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4299 				iflib_init_locked(ctx);
4300 			STATE_LOCK(ctx);
4301 			if_setdrvflags(ifp, bits);
4302 			STATE_UNLOCK(ctx);
4303 			CTX_UNLOCK(ctx);
4304 		}
4305 		if_vlancap(ifp);
4306 		break;
4307 	}
4308 	case SIOCGPRIVATE_0:
4309 	case SIOCSDRVSPEC:
4310 	case SIOCGDRVSPEC:
4311 		CTX_LOCK(ctx);
4312 		err = IFDI_PRIV_IOCTL(ctx, command, data);
4313 		CTX_UNLOCK(ctx);
4314 		break;
4315 	default:
4316 		err = ether_ioctl(ifp, command, data);
4317 		break;
4318 	}
4319 	if (reinit)
4320 		iflib_if_init(ctx);
4321 	return (err);
4322 }
4323 
4324 static uint64_t
4325 iflib_if_get_counter(if_t ifp, ift_counter cnt)
4326 {
4327 	if_ctx_t ctx = if_getsoftc(ifp);
4328 
4329 	return (IFDI_GET_COUNTER(ctx, cnt));
4330 }
4331 
4332 /*********************************************************************
4333  *
4334  *  OTHER FUNCTIONS EXPORTED TO THE STACK
4335  *
4336  **********************************************************************/
4337 
4338 static void
4339 iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag)
4340 {
4341 	if_ctx_t ctx = if_getsoftc(ifp);
4342 
4343 	if ((void *)ctx != arg)
4344 		return;
4345 
4346 	if ((vtag == 0) || (vtag > 4095))
4347 		return;
4348 
4349 	if (iflib_in_detach(ctx))
4350 		return;
4351 
4352 	CTX_LOCK(ctx);
4353 	/* Driver may need all untagged packets to be flushed */
4354 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4355 		iflib_stop(ctx);
4356 	IFDI_VLAN_REGISTER(ctx, vtag);
4357 	/* Re-init to load the changes, if required */
4358 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4359 		iflib_init_locked(ctx);
4360 	CTX_UNLOCK(ctx);
4361 }
4362 
4363 static void
4364 iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag)
4365 {
4366 	if_ctx_t ctx = if_getsoftc(ifp);
4367 
4368 	if ((void *)ctx != arg)
4369 		return;
4370 
4371 	if ((vtag == 0) || (vtag > 4095))
4372 		return;
4373 
4374 	CTX_LOCK(ctx);
4375 	/* Driver may need all tagged packets to be flushed */
4376 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4377 		iflib_stop(ctx);
4378 	IFDI_VLAN_UNREGISTER(ctx, vtag);
4379 	/* Re-init to load the changes, if required */
4380 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4381 		iflib_init_locked(ctx);
4382 	CTX_UNLOCK(ctx);
4383 }
4384 
4385 static void
4386 iflib_led_func(void *arg, int onoff)
4387 {
4388 	if_ctx_t ctx = arg;
4389 
4390 	CTX_LOCK(ctx);
4391 	IFDI_LED_FUNC(ctx, onoff);
4392 	CTX_UNLOCK(ctx);
4393 }
4394 
4395 /*********************************************************************
4396  *
4397  *  BUS FUNCTION DEFINITIONS
4398  *
4399  **********************************************************************/
4400 
4401 int
4402 iflib_device_probe(device_t dev)
4403 {
4404 	const pci_vendor_info_t *ent;
4405 	if_shared_ctx_t sctx;
4406 	uint16_t pci_device_id, pci_rev_id, pci_subdevice_id, pci_subvendor_id;
4407 	uint16_t pci_vendor_id;
4408 
4409 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4410 		return (ENOTSUP);
4411 
4412 	pci_vendor_id = pci_get_vendor(dev);
4413 	pci_device_id = pci_get_device(dev);
4414 	pci_subvendor_id = pci_get_subvendor(dev);
4415 	pci_subdevice_id = pci_get_subdevice(dev);
4416 	pci_rev_id = pci_get_revid(dev);
4417 	if (sctx->isc_parse_devinfo != NULL)
4418 		sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id);
4419 
4420 	ent = sctx->isc_vendor_info;
4421 	while (ent->pvi_vendor_id != 0) {
4422 		if (pci_vendor_id != ent->pvi_vendor_id) {
4423 			ent++;
4424 			continue;
4425 		}
4426 		if ((pci_device_id == ent->pvi_device_id) &&
4427 		    ((pci_subvendor_id == ent->pvi_subvendor_id) ||
4428 		     (ent->pvi_subvendor_id == 0)) &&
4429 		    ((pci_subdevice_id == ent->pvi_subdevice_id) ||
4430 		     (ent->pvi_subdevice_id == 0)) &&
4431 		    ((pci_rev_id == ent->pvi_rev_id) ||
4432 		     (ent->pvi_rev_id == 0))) {
4433 
4434 			device_set_desc_copy(dev, ent->pvi_name);
4435 			/* this needs to be changed to zero if the bus probing code
4436 			 * ever stops re-probing on best match because the sctx
4437 			 * may have its values over written by register calls
4438 			 * in subsequent probes
4439 			 */
4440 			return (BUS_PROBE_DEFAULT);
4441 		}
4442 		ent++;
4443 	}
4444 	return (ENXIO);
4445 }
4446 
4447 int
4448 iflib_device_probe_vendor(device_t dev)
4449 {
4450 	int probe;
4451 
4452 	probe = iflib_device_probe(dev);
4453 	if (probe == BUS_PROBE_DEFAULT)
4454 		return (BUS_PROBE_VENDOR);
4455 	else
4456 		return (probe);
4457 }
4458 
4459 static void
4460 iflib_reset_qvalues(if_ctx_t ctx)
4461 {
4462 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4463 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4464 	device_t dev = ctx->ifc_dev;
4465 	int i;
4466 
4467 	if (ctx->ifc_sysctl_ntxqs != 0)
4468 		scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs;
4469 	if (ctx->ifc_sysctl_nrxqs != 0)
4470 		scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs;
4471 
4472 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4473 		if (ctx->ifc_sysctl_ntxds[i] != 0)
4474 			scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i];
4475 		else
4476 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4477 	}
4478 
4479 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4480 		if (ctx->ifc_sysctl_nrxds[i] != 0)
4481 			scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i];
4482 		else
4483 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4484 	}
4485 
4486 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4487 		if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) {
4488 			device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n",
4489 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]);
4490 			scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i];
4491 		}
4492 		if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) {
4493 			device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n",
4494 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]);
4495 			scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i];
4496 		}
4497 		if (!powerof2(scctx->isc_nrxd[i])) {
4498 			device_printf(dev, "nrxd%d: %d is not a power of 2 - using default value of %d\n",
4499 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_default[i]);
4500 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4501 		}
4502 	}
4503 
4504 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4505 		if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) {
4506 			device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n",
4507 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]);
4508 			scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i];
4509 		}
4510 		if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) {
4511 			device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n",
4512 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]);
4513 			scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i];
4514 		}
4515 		if (!powerof2(scctx->isc_ntxd[i])) {
4516 			device_printf(dev, "ntxd%d: %d is not a power of 2 - using default value of %d\n",
4517 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_default[i]);
4518 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4519 		}
4520 	}
4521 }
4522 
4523 static void
4524 iflib_add_pfil(if_ctx_t ctx)
4525 {
4526 	struct pfil_head *pfil;
4527 	struct pfil_head_args pa;
4528 	iflib_rxq_t rxq;
4529 	int i;
4530 
4531 	pa.pa_version = PFIL_VERSION;
4532 	pa.pa_flags = PFIL_IN;
4533 	pa.pa_type = PFIL_TYPE_ETHERNET;
4534 	pa.pa_headname = ctx->ifc_ifp->if_xname;
4535 	pfil = pfil_head_register(&pa);
4536 
4537 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4538 		rxq->pfil = pfil;
4539 	}
4540 }
4541 
4542 static void
4543 iflib_rem_pfil(if_ctx_t ctx)
4544 {
4545 	struct pfil_head *pfil;
4546 	iflib_rxq_t rxq;
4547 	int i;
4548 
4549 	rxq = ctx->ifc_rxqs;
4550 	pfil = rxq->pfil;
4551 	for (i = 0; i < NRXQSETS(ctx); i++, rxq++) {
4552 		rxq->pfil = NULL;
4553 	}
4554 	pfil_head_unregister(pfil);
4555 }
4556 
4557 static uint16_t
4558 get_ctx_core_offset(if_ctx_t ctx)
4559 {
4560 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4561 	struct cpu_offset *op;
4562 	uint16_t qc;
4563 	uint16_t ret = ctx->ifc_sysctl_core_offset;
4564 
4565 	if (ret != CORE_OFFSET_UNSPECIFIED)
4566 		return (ret);
4567 
4568 	if (ctx->ifc_sysctl_separate_txrx)
4569 		qc = scctx->isc_ntxqsets + scctx->isc_nrxqsets;
4570 	else
4571 		qc = max(scctx->isc_ntxqsets, scctx->isc_nrxqsets);
4572 
4573 	mtx_lock(&cpu_offset_mtx);
4574 	SLIST_FOREACH(op, &cpu_offsets, entries) {
4575 		if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) {
4576 			ret = op->offset;
4577 			op->offset += qc;
4578 			MPASS(op->refcount < UINT_MAX);
4579 			op->refcount++;
4580 			break;
4581 		}
4582 	}
4583 	if (ret == CORE_OFFSET_UNSPECIFIED) {
4584 		ret = 0;
4585 		op = malloc(sizeof(struct cpu_offset), M_IFLIB,
4586 		    M_NOWAIT | M_ZERO);
4587 		if (op == NULL) {
4588 			device_printf(ctx->ifc_dev,
4589 			    "allocation for cpu offset failed.\n");
4590 		} else {
4591 			op->offset = qc;
4592 			op->refcount = 1;
4593 			CPU_COPY(&ctx->ifc_cpus, &op->set);
4594 			SLIST_INSERT_HEAD(&cpu_offsets, op, entries);
4595 		}
4596 	}
4597 	mtx_unlock(&cpu_offset_mtx);
4598 
4599 	return (ret);
4600 }
4601 
4602 static void
4603 unref_ctx_core_offset(if_ctx_t ctx)
4604 {
4605 	struct cpu_offset *op, *top;
4606 
4607 	mtx_lock(&cpu_offset_mtx);
4608 	SLIST_FOREACH_SAFE(op, &cpu_offsets, entries, top) {
4609 		if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) {
4610 			MPASS(op->refcount > 0);
4611 			op->refcount--;
4612 			if (op->refcount == 0) {
4613 				SLIST_REMOVE(&cpu_offsets, op, cpu_offset, entries);
4614 				free(op, M_IFLIB);
4615 			}
4616 			break;
4617 		}
4618 	}
4619 	mtx_unlock(&cpu_offset_mtx);
4620 }
4621 
4622 int
4623 iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp)
4624 {
4625 	if_ctx_t ctx;
4626 	if_t ifp;
4627 	if_softc_ctx_t scctx;
4628 	kobjop_desc_t kobj_desc;
4629 	kobj_method_t *kobj_method;
4630 	int err, msix, rid;
4631 	int num_txd, num_rxd;
4632 
4633 	ctx = malloc(sizeof(* ctx), M_IFLIB, M_WAITOK|M_ZERO);
4634 
4635 	if (sc == NULL) {
4636 		sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4637 		device_set_softc(dev, ctx);
4638 		ctx->ifc_flags |= IFC_SC_ALLOCATED;
4639 	}
4640 
4641 	ctx->ifc_sctx = sctx;
4642 	ctx->ifc_dev = dev;
4643 	ctx->ifc_softc = sc;
4644 
4645 	if ((err = iflib_register(ctx)) != 0) {
4646 		device_printf(dev, "iflib_register failed %d\n", err);
4647 		goto fail_ctx_free;
4648 	}
4649 	iflib_add_device_sysctl_pre(ctx);
4650 
4651 	scctx = &ctx->ifc_softc_ctx;
4652 	ifp = ctx->ifc_ifp;
4653 
4654 	iflib_reset_qvalues(ctx);
4655 	CTX_LOCK(ctx);
4656 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4657 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4658 		goto fail_unlock;
4659 	}
4660 	_iflib_pre_assert(scctx);
4661 	ctx->ifc_txrx = *scctx->isc_txrx;
4662 
4663 	if (sctx->isc_flags & IFLIB_DRIVER_MEDIA)
4664 		ctx->ifc_mediap = scctx->isc_media;
4665 
4666 #ifdef INVARIANTS
4667 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
4668 		MPASS(scctx->isc_tx_csum_flags);
4669 #endif
4670 
4671 	if_setcapabilities(ifp,
4672 	    scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_NOMAP);
4673 	if_setcapenable(ifp,
4674 	    scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_NOMAP);
4675 
4676 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4677 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4678 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4679 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4680 
4681 	num_txd = iflib_num_tx_descs(ctx);
4682 	num_rxd = iflib_num_rx_descs(ctx);
4683 
4684 	/* XXX change for per-queue sizes */
4685 	device_printf(dev, "Using %d TX descriptors and %d RX descriptors\n",
4686 	    num_txd, num_rxd);
4687 
4688 	if (scctx->isc_tx_nsegments > num_txd / MAX_SINGLE_PACKET_FRACTION)
4689 		scctx->isc_tx_nsegments = max(1, num_txd /
4690 		    MAX_SINGLE_PACKET_FRACTION);
4691 	if (scctx->isc_tx_tso_segments_max > num_txd /
4692 	    MAX_SINGLE_PACKET_FRACTION)
4693 		scctx->isc_tx_tso_segments_max = max(1,
4694 		    num_txd / MAX_SINGLE_PACKET_FRACTION);
4695 
4696 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4697 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
4698 		/*
4699 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
4700 		 * but some MACs do.
4701 		 */
4702 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
4703 		    IP_MAXPACKET));
4704 		/*
4705 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
4706 		 * into account.  In the worst case, each of these calls will
4707 		 * add another mbuf and, thus, the requirement for another DMA
4708 		 * segment.  So for best performance, it doesn't make sense to
4709 		 * advertize a maximum of TSO segments that typically will
4710 		 * require defragmentation in iflib_encap().
4711 		 */
4712 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
4713 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
4714 	}
4715 	if (scctx->isc_rss_table_size == 0)
4716 		scctx->isc_rss_table_size = 64;
4717 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4718 
4719 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4720 	/* XXX format name */
4721 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx,
4722 	    NULL, NULL, "admin");
4723 
4724 	/* Set up cpu set.  If it fails, use the set of all CPUs. */
4725 	if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) != 0) {
4726 		device_printf(dev, "Unable to fetch CPU list\n");
4727 		CPU_COPY(&all_cpus, &ctx->ifc_cpus);
4728 	}
4729 	MPASS(CPU_COUNT(&ctx->ifc_cpus) > 0);
4730 
4731 	/*
4732 	** Now set up MSI or MSI-X, should return us the number of supported
4733 	** vectors (will be 1 for a legacy interrupt and MSI).
4734 	*/
4735 	if (sctx->isc_flags & IFLIB_SKIP_MSIX) {
4736 		msix = scctx->isc_vectors;
4737 	} else if (scctx->isc_msix_bar != 0)
4738 	       /*
4739 		* The simple fact that isc_msix_bar is not 0 does not mean we
4740 		* we have a good value there that is known to work.
4741 		*/
4742 		msix = iflib_msix_init(ctx);
4743 	else {
4744 		scctx->isc_vectors = 1;
4745 		scctx->isc_ntxqsets = 1;
4746 		scctx->isc_nrxqsets = 1;
4747 		scctx->isc_intr = IFLIB_INTR_LEGACY;
4748 		msix = 0;
4749 	}
4750 	/* Get memory for the station queues */
4751 	if ((err = iflib_queues_alloc(ctx))) {
4752 		device_printf(dev, "Unable to allocate queue memory\n");
4753 		goto fail_intr_free;
4754 	}
4755 
4756 	if ((err = iflib_qset_structures_setup(ctx)))
4757 		goto fail_queues;
4758 
4759 	/*
4760 	 * Now that we know how many queues there are, get the core offset.
4761 	 */
4762 	ctx->ifc_sysctl_core_offset = get_ctx_core_offset(ctx);
4763 
4764 	if (msix > 1) {
4765 		/*
4766 		 * When using MSI-X, ensure that ifdi_{r,t}x_queue_intr_enable
4767 		 * aren't the default NULL implementation.
4768 		 */
4769 		kobj_desc = &ifdi_rx_queue_intr_enable_desc;
4770 		kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL,
4771 		    kobj_desc);
4772 		if (kobj_method == &kobj_desc->deflt) {
4773 			device_printf(dev,
4774 			    "MSI-X requires ifdi_rx_queue_intr_enable method");
4775 			err = EOPNOTSUPP;
4776 			goto fail_queues;
4777 		}
4778 		kobj_desc = &ifdi_tx_queue_intr_enable_desc;
4779 		kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL,
4780 		    kobj_desc);
4781 		if (kobj_method == &kobj_desc->deflt) {
4782 			device_printf(dev,
4783 			    "MSI-X requires ifdi_tx_queue_intr_enable method");
4784 			err = EOPNOTSUPP;
4785 			goto fail_queues;
4786 		}
4787 
4788 		/*
4789 		 * Assign the MSI-X vectors.
4790 		 * Note that the default NULL ifdi_msix_intr_assign method will
4791 		 * fail here, too.
4792 		 */
4793 		err = IFDI_MSIX_INTR_ASSIGN(ctx, msix);
4794 		if (err != 0) {
4795 			device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n",
4796 			    err);
4797 			goto fail_queues;
4798 		}
4799 	} else if (scctx->isc_intr != IFLIB_INTR_MSIX) {
4800 		rid = 0;
4801 		if (scctx->isc_intr == IFLIB_INTR_MSI) {
4802 			MPASS(msix == 1);
4803 			rid = 1;
4804 		}
4805 		if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) {
4806 			device_printf(dev, "iflib_legacy_setup failed %d\n", err);
4807 			goto fail_queues;
4808 		}
4809 	} else {
4810 		device_printf(dev,
4811 		    "Cannot use iflib with only 1 MSI-X interrupt!\n");
4812 		err = ENODEV;
4813 		goto fail_intr_free;
4814 	}
4815 
4816 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
4817 
4818 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4819 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4820 		goto fail_detach;
4821 	}
4822 
4823 	/*
4824 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4825 	 * This must appear after the call to ether_ifattach() because
4826 	 * ether_ifattach() sets if_hdrlen to the default value.
4827 	 */
4828 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4829 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
4830 
4831 	if ((err = iflib_netmap_attach(ctx))) {
4832 		device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err);
4833 		goto fail_detach;
4834 	}
4835 	*ctxp = ctx;
4836 
4837 	DEBUGNET_SET(ctx->ifc_ifp, iflib);
4838 
4839 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4840 	iflib_add_device_sysctl_post(ctx);
4841 	iflib_add_pfil(ctx);
4842 	ctx->ifc_flags |= IFC_INIT_DONE;
4843 	CTX_UNLOCK(ctx);
4844 
4845 	return (0);
4846 
4847 fail_detach:
4848 	ether_ifdetach(ctx->ifc_ifp);
4849 fail_intr_free:
4850 	iflib_free_intr_mem(ctx);
4851 fail_queues:
4852 	iflib_tx_structures_free(ctx);
4853 	iflib_rx_structures_free(ctx);
4854 	taskqgroup_detach(qgroup_if_config_tqg, &ctx->ifc_admin_task);
4855 	IFDI_DETACH(ctx);
4856 fail_unlock:
4857 	CTX_UNLOCK(ctx);
4858 	iflib_deregister(ctx);
4859 fail_ctx_free:
4860 	device_set_softc(ctx->ifc_dev, NULL);
4861         if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4862                 free(ctx->ifc_softc, M_IFLIB);
4863         free(ctx, M_IFLIB);
4864 	return (err);
4865 }
4866 
4867 int
4868 iflib_pseudo_register(device_t dev, if_shared_ctx_t sctx, if_ctx_t *ctxp,
4869 					  struct iflib_cloneattach_ctx *clctx)
4870 {
4871 	int num_txd, num_rxd;
4872 	int err;
4873 	if_ctx_t ctx;
4874 	if_t ifp;
4875 	if_softc_ctx_t scctx;
4876 	int i;
4877 	void *sc;
4878 
4879 	ctx = malloc(sizeof(*ctx), M_IFLIB, M_WAITOK|M_ZERO);
4880 	sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4881 	ctx->ifc_flags |= IFC_SC_ALLOCATED;
4882 	if (sctx->isc_flags & (IFLIB_PSEUDO|IFLIB_VIRTUAL))
4883 		ctx->ifc_flags |= IFC_PSEUDO;
4884 
4885 	ctx->ifc_sctx = sctx;
4886 	ctx->ifc_softc = sc;
4887 	ctx->ifc_dev = dev;
4888 
4889 	if ((err = iflib_register(ctx)) != 0) {
4890 		device_printf(dev, "%s: iflib_register failed %d\n", __func__, err);
4891 		goto fail_ctx_free;
4892 	}
4893 	iflib_add_device_sysctl_pre(ctx);
4894 
4895 	scctx = &ctx->ifc_softc_ctx;
4896 	ifp = ctx->ifc_ifp;
4897 
4898 	iflib_reset_qvalues(ctx);
4899 	CTX_LOCK(ctx);
4900 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4901 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4902 		goto fail_unlock;
4903 	}
4904 	if (sctx->isc_flags & IFLIB_GEN_MAC)
4905 		ether_gen_addr(ifp, &ctx->ifc_mac);
4906 	if ((err = IFDI_CLONEATTACH(ctx, clctx->cc_ifc, clctx->cc_name,
4907 								clctx->cc_params)) != 0) {
4908 		device_printf(dev, "IFDI_CLONEATTACH failed %d\n", err);
4909 		goto fail_unlock;
4910 	}
4911 #ifdef INVARIANTS
4912 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
4913 		MPASS(scctx->isc_tx_csum_flags);
4914 #endif
4915 
4916 	if_setcapabilities(ifp, scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_LINKSTATE);
4917 	if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_LINKSTATE);
4918 
4919 	ifp->if_flags |= IFF_NOGROUP;
4920 	if (sctx->isc_flags & IFLIB_PSEUDO) {
4921 		ifmedia_add(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO, 0, NULL);
4922 		ifmedia_set(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO);
4923 		if (sctx->isc_flags & IFLIB_PSEUDO_ETHER) {
4924 			ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
4925 		} else {
4926 			if_attach(ctx->ifc_ifp);
4927 			bpfattach(ctx->ifc_ifp, DLT_NULL, sizeof(u_int32_t));
4928 		}
4929 
4930 		if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4931 			device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4932 			goto fail_detach;
4933 		}
4934 		*ctxp = ctx;
4935 
4936 		/*
4937 		 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4938 		 * This must appear after the call to ether_ifattach() because
4939 		 * ether_ifattach() sets if_hdrlen to the default value.
4940 		 */
4941 		if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4942 			if_setifheaderlen(ifp,
4943 			    sizeof(struct ether_vlan_header));
4944 
4945 		if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4946 		iflib_add_device_sysctl_post(ctx);
4947 		ctx->ifc_flags |= IFC_INIT_DONE;
4948 		CTX_UNLOCK(ctx);
4949 		return (0);
4950 	}
4951 	ifmedia_add(ctx->ifc_mediap, IFM_ETHER | IFM_1000_T | IFM_FDX, 0, NULL);
4952 	ifmedia_add(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO, 0, NULL);
4953 	ifmedia_set(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO);
4954 
4955 	_iflib_pre_assert(scctx);
4956 	ctx->ifc_txrx = *scctx->isc_txrx;
4957 
4958 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4959 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4960 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4961 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4962 
4963 	num_txd = iflib_num_tx_descs(ctx);
4964 	num_rxd = iflib_num_rx_descs(ctx);
4965 
4966 	/* XXX change for per-queue sizes */
4967 	device_printf(dev, "Using %d TX descriptors and %d RX descriptors\n",
4968 	    num_txd, num_rxd);
4969 
4970 	if (scctx->isc_tx_nsegments > num_txd / MAX_SINGLE_PACKET_FRACTION)
4971 		scctx->isc_tx_nsegments = max(1, num_txd /
4972 		    MAX_SINGLE_PACKET_FRACTION);
4973 	if (scctx->isc_tx_tso_segments_max > num_txd /
4974 	    MAX_SINGLE_PACKET_FRACTION)
4975 		scctx->isc_tx_tso_segments_max = max(1,
4976 		    num_txd / MAX_SINGLE_PACKET_FRACTION);
4977 
4978 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4979 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
4980 		/*
4981 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
4982 		 * but some MACs do.
4983 		 */
4984 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
4985 		    IP_MAXPACKET));
4986 		/*
4987 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
4988 		 * into account.  In the worst case, each of these calls will
4989 		 * add another mbuf and, thus, the requirement for another DMA
4990 		 * segment.  So for best performance, it doesn't make sense to
4991 		 * advertize a maximum of TSO segments that typically will
4992 		 * require defragmentation in iflib_encap().
4993 		 */
4994 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
4995 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
4996 	}
4997 	if (scctx->isc_rss_table_size == 0)
4998 		scctx->isc_rss_table_size = 64;
4999 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
5000 
5001 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
5002 	/* XXX format name */
5003 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx,
5004 	    NULL, NULL, "admin");
5005 
5006 	/* XXX --- can support > 1 -- but keep it simple for now */
5007 	scctx->isc_intr = IFLIB_INTR_LEGACY;
5008 
5009 	/* Get memory for the station queues */
5010 	if ((err = iflib_queues_alloc(ctx))) {
5011 		device_printf(dev, "Unable to allocate queue memory\n");
5012 		goto fail_iflib_detach;
5013 	}
5014 
5015 	if ((err = iflib_qset_structures_setup(ctx))) {
5016 		device_printf(dev, "qset structure setup failed %d\n", err);
5017 		goto fail_queues;
5018 	}
5019 
5020 	/*
5021 	 * XXX What if anything do we want to do about interrupts?
5022 	 */
5023 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
5024 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
5025 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
5026 		goto fail_detach;
5027 	}
5028 
5029 	/*
5030 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
5031 	 * This must appear after the call to ether_ifattach() because
5032 	 * ether_ifattach() sets if_hdrlen to the default value.
5033 	 */
5034 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
5035 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
5036 
5037 	/* XXX handle more than one queue */
5038 	for (i = 0; i < scctx->isc_nrxqsets; i++)
5039 		IFDI_RX_CLSET(ctx, 0, i, ctx->ifc_rxqs[i].ifr_fl[0].ifl_sds.ifsd_cl);
5040 
5041 	*ctxp = ctx;
5042 
5043 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
5044 	iflib_add_device_sysctl_post(ctx);
5045 	ctx->ifc_flags |= IFC_INIT_DONE;
5046 	CTX_UNLOCK(ctx);
5047 
5048 	return (0);
5049 fail_detach:
5050 	ether_ifdetach(ctx->ifc_ifp);
5051 fail_queues:
5052 	iflib_tx_structures_free(ctx);
5053 	iflib_rx_structures_free(ctx);
5054 fail_iflib_detach:
5055 	IFDI_DETACH(ctx);
5056 fail_unlock:
5057 	CTX_UNLOCK(ctx);
5058 	iflib_deregister(ctx);
5059 fail_ctx_free:
5060 	free(ctx->ifc_softc, M_IFLIB);
5061 	free(ctx, M_IFLIB);
5062 	return (err);
5063 }
5064 
5065 int
5066 iflib_pseudo_deregister(if_ctx_t ctx)
5067 {
5068 	if_t ifp = ctx->ifc_ifp;
5069 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5070 	iflib_txq_t txq;
5071 	iflib_rxq_t rxq;
5072 	int i, j;
5073 	struct taskqgroup *tqg;
5074 	iflib_fl_t fl;
5075 
5076 	/* Unregister VLAN event handlers early */
5077 	iflib_unregister_vlan_handlers(ctx);
5078 
5079 	if ((sctx->isc_flags & IFLIB_PSEUDO)  &&
5080 		(sctx->isc_flags & IFLIB_PSEUDO_ETHER) == 0) {
5081 		bpfdetach(ifp);
5082 		if_detach(ifp);
5083 	} else {
5084 		ether_ifdetach(ifp);
5085 	}
5086 	/* XXX drain any dependent tasks */
5087 	tqg = qgroup_if_io_tqg;
5088 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
5089 		callout_drain(&txq->ift_timer);
5090 		if (txq->ift_task.gt_uniq != NULL)
5091 			taskqgroup_detach(tqg, &txq->ift_task);
5092 	}
5093 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
5094 		callout_drain(&rxq->ifr_watchdog);
5095 		if (rxq->ifr_task.gt_uniq != NULL)
5096 			taskqgroup_detach(tqg, &rxq->ifr_task);
5097 
5098 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5099 			free(fl->ifl_rx_bitmap, M_IFLIB);
5100 	}
5101 	tqg = qgroup_if_config_tqg;
5102 	if (ctx->ifc_admin_task.gt_uniq != NULL)
5103 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
5104 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
5105 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
5106 
5107 	iflib_tx_structures_free(ctx);
5108 	iflib_rx_structures_free(ctx);
5109 
5110 	iflib_deregister(ctx);
5111 
5112 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5113 		free(ctx->ifc_softc, M_IFLIB);
5114 	free(ctx, M_IFLIB);
5115 	return (0);
5116 }
5117 
5118 int
5119 iflib_device_attach(device_t dev)
5120 {
5121 	if_ctx_t ctx;
5122 	if_shared_ctx_t sctx;
5123 
5124 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
5125 		return (ENOTSUP);
5126 
5127 	pci_enable_busmaster(dev);
5128 
5129 	return (iflib_device_register(dev, NULL, sctx, &ctx));
5130 }
5131 
5132 int
5133 iflib_device_deregister(if_ctx_t ctx)
5134 {
5135 	if_t ifp = ctx->ifc_ifp;
5136 	iflib_txq_t txq;
5137 	iflib_rxq_t rxq;
5138 	device_t dev = ctx->ifc_dev;
5139 	int i, j;
5140 	struct taskqgroup *tqg;
5141 	iflib_fl_t fl;
5142 
5143 	/* Make sure VLANS are not using driver */
5144 	if (if_vlantrunkinuse(ifp)) {
5145 		device_printf(dev, "Vlan in use, detach first\n");
5146 		return (EBUSY);
5147 	}
5148 #ifdef PCI_IOV
5149 	if (!CTX_IS_VF(ctx) && pci_iov_detach(dev) != 0) {
5150 		device_printf(dev, "SR-IOV in use; detach first.\n");
5151 		return (EBUSY);
5152 	}
5153 #endif
5154 
5155 	STATE_LOCK(ctx);
5156 	ctx->ifc_flags |= IFC_IN_DETACH;
5157 	STATE_UNLOCK(ctx);
5158 
5159 	/* Unregister VLAN handlers before calling iflib_stop() */
5160 	iflib_unregister_vlan_handlers(ctx);
5161 
5162 	iflib_netmap_detach(ifp);
5163 	ether_ifdetach(ifp);
5164 
5165 	CTX_LOCK(ctx);
5166 	iflib_stop(ctx);
5167 	CTX_UNLOCK(ctx);
5168 
5169 	iflib_rem_pfil(ctx);
5170 	if (ctx->ifc_led_dev != NULL)
5171 		led_destroy(ctx->ifc_led_dev);
5172 	/* XXX drain any dependent tasks */
5173 	tqg = qgroup_if_io_tqg;
5174 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
5175 		callout_drain(&txq->ift_timer);
5176 		if (txq->ift_task.gt_uniq != NULL)
5177 			taskqgroup_detach(tqg, &txq->ift_task);
5178 	}
5179 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
5180 		if (rxq->ifr_task.gt_uniq != NULL)
5181 			taskqgroup_detach(tqg, &rxq->ifr_task);
5182 
5183 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5184 			free(fl->ifl_rx_bitmap, M_IFLIB);
5185 	}
5186 	tqg = qgroup_if_config_tqg;
5187 	if (ctx->ifc_admin_task.gt_uniq != NULL)
5188 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
5189 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
5190 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
5191 	CTX_LOCK(ctx);
5192 	IFDI_DETACH(ctx);
5193 	CTX_UNLOCK(ctx);
5194 
5195 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
5196 	iflib_free_intr_mem(ctx);
5197 
5198 	bus_generic_detach(dev);
5199 
5200 	iflib_tx_structures_free(ctx);
5201 	iflib_rx_structures_free(ctx);
5202 
5203 	iflib_deregister(ctx);
5204 
5205 	device_set_softc(ctx->ifc_dev, NULL);
5206 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5207 		free(ctx->ifc_softc, M_IFLIB);
5208 	unref_ctx_core_offset(ctx);
5209 	free(ctx, M_IFLIB);
5210 	return (0);
5211 }
5212 
5213 static void
5214 iflib_free_intr_mem(if_ctx_t ctx)
5215 {
5216 
5217 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) {
5218 		iflib_irq_free(ctx, &ctx->ifc_legacy_irq);
5219 	}
5220 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) {
5221 		pci_release_msi(ctx->ifc_dev);
5222 	}
5223 	if (ctx->ifc_msix_mem != NULL) {
5224 		bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY,
5225 		    rman_get_rid(ctx->ifc_msix_mem), ctx->ifc_msix_mem);
5226 		ctx->ifc_msix_mem = NULL;
5227 	}
5228 }
5229 
5230 int
5231 iflib_device_detach(device_t dev)
5232 {
5233 	if_ctx_t ctx = device_get_softc(dev);
5234 
5235 	return (iflib_device_deregister(ctx));
5236 }
5237 
5238 int
5239 iflib_device_suspend(device_t dev)
5240 {
5241 	if_ctx_t ctx = device_get_softc(dev);
5242 
5243 	CTX_LOCK(ctx);
5244 	IFDI_SUSPEND(ctx);
5245 	CTX_UNLOCK(ctx);
5246 
5247 	return bus_generic_suspend(dev);
5248 }
5249 int
5250 iflib_device_shutdown(device_t dev)
5251 {
5252 	if_ctx_t ctx = device_get_softc(dev);
5253 
5254 	CTX_LOCK(ctx);
5255 	IFDI_SHUTDOWN(ctx);
5256 	CTX_UNLOCK(ctx);
5257 
5258 	return bus_generic_suspend(dev);
5259 }
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 
5357 static void
5358 _iflib_assert(if_shared_ctx_t sctx)
5359 {
5360 	int i;
5361 
5362 	MPASS(sctx->isc_tx_maxsize);
5363 	MPASS(sctx->isc_tx_maxsegsize);
5364 
5365 	MPASS(sctx->isc_rx_maxsize);
5366 	MPASS(sctx->isc_rx_nsegments);
5367 	MPASS(sctx->isc_rx_maxsegsize);
5368 
5369 	MPASS(sctx->isc_nrxqs >= 1 && sctx->isc_nrxqs <= 8);
5370 	for (i = 0; i < sctx->isc_nrxqs; i++) {
5371 		MPASS(sctx->isc_nrxd_min[i]);
5372 		MPASS(powerof2(sctx->isc_nrxd_min[i]));
5373 		MPASS(sctx->isc_nrxd_max[i]);
5374 		MPASS(powerof2(sctx->isc_nrxd_max[i]));
5375 		MPASS(sctx->isc_nrxd_default[i]);
5376 		MPASS(powerof2(sctx->isc_nrxd_default[i]));
5377 	}
5378 
5379 	MPASS(sctx->isc_ntxqs >= 1 && sctx->isc_ntxqs <= 8);
5380 	for (i = 0; i < sctx->isc_ntxqs; i++) {
5381 		MPASS(sctx->isc_ntxd_min[i]);
5382 		MPASS(powerof2(sctx->isc_ntxd_min[i]));
5383 		MPASS(sctx->isc_ntxd_max[i]);
5384 		MPASS(powerof2(sctx->isc_ntxd_max[i]));
5385 		MPASS(sctx->isc_ntxd_default[i]);
5386 		MPASS(powerof2(sctx->isc_ntxd_default[i]));
5387 	}
5388 }
5389 
5390 static void
5391 _iflib_pre_assert(if_softc_ctx_t scctx)
5392 {
5393 
5394 	MPASS(scctx->isc_txrx->ift_txd_encap);
5395 	MPASS(scctx->isc_txrx->ift_txd_flush);
5396 	MPASS(scctx->isc_txrx->ift_txd_credits_update);
5397 	MPASS(scctx->isc_txrx->ift_rxd_available);
5398 	MPASS(scctx->isc_txrx->ift_rxd_pkt_get);
5399 	MPASS(scctx->isc_txrx->ift_rxd_refill);
5400 	MPASS(scctx->isc_txrx->ift_rxd_flush);
5401 }
5402 
5403 static int
5404 iflib_register(if_ctx_t ctx)
5405 {
5406 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5407 	driver_t *driver = sctx->isc_driver;
5408 	device_t dev = ctx->ifc_dev;
5409 	if_t ifp;
5410 	u_char type;
5411 	int iflags;
5412 
5413 	if ((sctx->isc_flags & IFLIB_PSEUDO) == 0)
5414 		_iflib_assert(sctx);
5415 
5416 	CTX_LOCK_INIT(ctx);
5417 	STATE_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev));
5418 	if (sctx->isc_flags & IFLIB_PSEUDO) {
5419 		if (sctx->isc_flags & IFLIB_PSEUDO_ETHER)
5420 			type = IFT_ETHER;
5421 		else
5422 			type = IFT_PPP;
5423 	} else
5424 		type = IFT_ETHER;
5425 	ifp = ctx->ifc_ifp = if_alloc(type);
5426 	if (ifp == NULL) {
5427 		device_printf(dev, "can not allocate ifnet structure\n");
5428 		return (ENOMEM);
5429 	}
5430 
5431 	/*
5432 	 * Initialize our context's device specific methods
5433 	 */
5434 	kobj_init((kobj_t) ctx, (kobj_class_t) driver);
5435 	kobj_class_compile((kobj_class_t) driver);
5436 
5437 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
5438 	if_setsoftc(ifp, ctx);
5439 	if_setdev(ifp, dev);
5440 	if_setinitfn(ifp, iflib_if_init);
5441 	if_setioctlfn(ifp, iflib_if_ioctl);
5442 #ifdef ALTQ
5443 	if_setstartfn(ifp, iflib_altq_if_start);
5444 	if_settransmitfn(ifp, iflib_altq_if_transmit);
5445 	if_setsendqready(ifp);
5446 #else
5447 	if_settransmitfn(ifp, iflib_if_transmit);
5448 #endif
5449 	if_setqflushfn(ifp, iflib_if_qflush);
5450 	iflags = IFF_MULTICAST | IFF_KNOWSEPOCH;
5451 
5452 	if ((sctx->isc_flags & IFLIB_PSEUDO) &&
5453 		(sctx->isc_flags & IFLIB_PSEUDO_ETHER) == 0)
5454 		iflags |= IFF_POINTOPOINT;
5455 	else
5456 		iflags |= IFF_BROADCAST | IFF_SIMPLEX;
5457 	if_setflags(ifp, iflags);
5458 	ctx->ifc_vlan_attach_event =
5459 		EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx,
5460 							  EVENTHANDLER_PRI_FIRST);
5461 	ctx->ifc_vlan_detach_event =
5462 		EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx,
5463 							  EVENTHANDLER_PRI_FIRST);
5464 
5465 	if ((sctx->isc_flags & IFLIB_DRIVER_MEDIA) == 0) {
5466 		ctx->ifc_mediap = &ctx->ifc_media;
5467 		ifmedia_init(ctx->ifc_mediap, IFM_IMASK,
5468 		    iflib_media_change, iflib_media_status);
5469 	}
5470 	return (0);
5471 }
5472 
5473 static void
5474 iflib_unregister_vlan_handlers(if_ctx_t ctx)
5475 {
5476 	/* Unregister VLAN events */
5477 	if (ctx->ifc_vlan_attach_event != NULL) {
5478 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
5479 		ctx->ifc_vlan_attach_event = NULL;
5480 	}
5481 	if (ctx->ifc_vlan_detach_event != NULL) {
5482 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
5483 		ctx->ifc_vlan_detach_event = NULL;
5484 	}
5485 
5486 }
5487 
5488 static void
5489 iflib_deregister(if_ctx_t ctx)
5490 {
5491 	if_t ifp = ctx->ifc_ifp;
5492 
5493 	/* Remove all media */
5494 	ifmedia_removeall(&ctx->ifc_media);
5495 
5496 	/* Ensure that VLAN event handlers are unregistered */
5497 	iflib_unregister_vlan_handlers(ctx);
5498 
5499 	/* Release kobject reference */
5500 	kobj_delete((kobj_t) ctx, NULL);
5501 
5502 	/* Free the ifnet structure */
5503 	if_free(ifp);
5504 
5505 	STATE_LOCK_DESTROY(ctx);
5506 
5507 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
5508 	CTX_LOCK_DESTROY(ctx);
5509 }
5510 
5511 static int
5512 iflib_queues_alloc(if_ctx_t ctx)
5513 {
5514 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5515 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5516 	device_t dev = ctx->ifc_dev;
5517 	int nrxqsets = scctx->isc_nrxqsets;
5518 	int ntxqsets = scctx->isc_ntxqsets;
5519 	iflib_txq_t txq;
5520 	iflib_rxq_t rxq;
5521 	iflib_fl_t fl = NULL;
5522 	int i, j, cpu, err, txconf, rxconf;
5523 	iflib_dma_info_t ifdip;
5524 	uint32_t *rxqsizes = scctx->isc_rxqsizes;
5525 	uint32_t *txqsizes = scctx->isc_txqsizes;
5526 	uint8_t nrxqs = sctx->isc_nrxqs;
5527 	uint8_t ntxqs = sctx->isc_ntxqs;
5528 	int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1;
5529 	caddr_t *vaddrs;
5530 	uint64_t *paddrs;
5531 
5532 	KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1"));
5533 	KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1"));
5534 
5535 	/* Allocate the TX ring struct memory */
5536 	if (!(ctx->ifc_txqs =
5537 	    (iflib_txq_t) malloc(sizeof(struct iflib_txq) *
5538 	    ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5539 		device_printf(dev, "Unable to allocate TX ring memory\n");
5540 		err = ENOMEM;
5541 		goto fail;
5542 	}
5543 
5544 	/* Now allocate the RX */
5545 	if (!(ctx->ifc_rxqs =
5546 	    (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) *
5547 	    nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5548 		device_printf(dev, "Unable to allocate RX ring memory\n");
5549 		err = ENOMEM;
5550 		goto rx_fail;
5551 	}
5552 
5553 	txq = ctx->ifc_txqs;
5554 	rxq = ctx->ifc_rxqs;
5555 
5556 	/*
5557 	 * XXX handle allocation failure
5558 	 */
5559 	for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) {
5560 		/* Set up some basics */
5561 
5562 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs,
5563 		    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
5564 			device_printf(dev,
5565 			    "Unable to allocate TX DMA info memory\n");
5566 			err = ENOMEM;
5567 			goto err_tx_desc;
5568 		}
5569 		txq->ift_ifdi = ifdip;
5570 		for (j = 0; j < ntxqs; j++, ifdip++) {
5571 			if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, 0)) {
5572 				device_printf(dev,
5573 				    "Unable to allocate TX descriptors\n");
5574 				err = ENOMEM;
5575 				goto err_tx_desc;
5576 			}
5577 			txq->ift_txd_size[j] = scctx->isc_txd_size[j];
5578 			bzero((void *)ifdip->idi_vaddr, txqsizes[j]);
5579 		}
5580 		txq->ift_ctx = ctx;
5581 		txq->ift_id = i;
5582 		if (sctx->isc_flags & IFLIB_HAS_TXCQ) {
5583 			txq->ift_br_offset = 1;
5584 		} else {
5585 			txq->ift_br_offset = 0;
5586 		}
5587 		/* XXX fix this */
5588 		txq->ift_timer.c_cpu = cpu;
5589 
5590 		if (iflib_txsd_alloc(txq)) {
5591 			device_printf(dev, "Critical Failure setting up TX buffers\n");
5592 			err = ENOMEM;
5593 			goto err_tx_desc;
5594 		}
5595 
5596 		/* Initialize the TX lock */
5597 		snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:TX(%d):callout",
5598 		    device_get_nameunit(dev), txq->ift_id);
5599 		mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF);
5600 		callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0);
5601 
5602 		err = ifmp_ring_alloc(&txq->ift_br, 2048, txq, iflib_txq_drain,
5603 				      iflib_txq_can_drain, M_IFLIB, M_WAITOK);
5604 		if (err) {
5605 			/* XXX free any allocated rings */
5606 			device_printf(dev, "Unable to allocate buf_ring\n");
5607 			goto err_tx_desc;
5608 		}
5609 	}
5610 
5611 	for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) {
5612 		/* Set up some basics */
5613 		callout_init(&rxq->ifr_watchdog, 1);
5614 
5615 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs,
5616 		   M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
5617 			device_printf(dev,
5618 			    "Unable to allocate RX DMA info memory\n");
5619 			err = ENOMEM;
5620 			goto err_tx_desc;
5621 		}
5622 
5623 		rxq->ifr_ifdi = ifdip;
5624 		/* XXX this needs to be changed if #rx queues != #tx queues */
5625 		rxq->ifr_ntxqirq = 1;
5626 		rxq->ifr_txqid[0] = i;
5627 		for (j = 0; j < nrxqs; j++, ifdip++) {
5628 			if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, 0)) {
5629 				device_printf(dev,
5630 				    "Unable to allocate RX descriptors\n");
5631 				err = ENOMEM;
5632 				goto err_tx_desc;
5633 			}
5634 			bzero((void *)ifdip->idi_vaddr, rxqsizes[j]);
5635 		}
5636 		rxq->ifr_ctx = ctx;
5637 		rxq->ifr_id = i;
5638 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
5639 			rxq->ifr_fl_offset = 1;
5640 		} else {
5641 			rxq->ifr_fl_offset = 0;
5642 		}
5643 		rxq->ifr_nfl = nfree_lists;
5644 		if (!(fl =
5645 			  (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) {
5646 			device_printf(dev, "Unable to allocate free list memory\n");
5647 			err = ENOMEM;
5648 			goto err_tx_desc;
5649 		}
5650 		rxq->ifr_fl = fl;
5651 		for (j = 0; j < nfree_lists; j++) {
5652 			fl[j].ifl_rxq = rxq;
5653 			fl[j].ifl_id = j;
5654 			fl[j].ifl_ifdi = &rxq->ifr_ifdi[j + rxq->ifr_fl_offset];
5655 			fl[j].ifl_rxd_size = scctx->isc_rxd_size[j];
5656 		}
5657 		/* Allocate receive buffers for the ring */
5658 		if (iflib_rxsd_alloc(rxq)) {
5659 			device_printf(dev,
5660 			    "Critical Failure setting up receive buffers\n");
5661 			err = ENOMEM;
5662 			goto err_rx_desc;
5663 		}
5664 
5665 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5666 			fl->ifl_rx_bitmap = bit_alloc(fl->ifl_size, M_IFLIB,
5667 			    M_WAITOK);
5668 	}
5669 
5670 	/* TXQs */
5671 	vaddrs = malloc(sizeof(caddr_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5672 	paddrs = malloc(sizeof(uint64_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5673 	for (i = 0; i < ntxqsets; i++) {
5674 		iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi;
5675 
5676 		for (j = 0; j < ntxqs; j++, di++) {
5677 			vaddrs[i*ntxqs + j] = di->idi_vaddr;
5678 			paddrs[i*ntxqs + j] = di->idi_paddr;
5679 		}
5680 	}
5681 	if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) {
5682 		device_printf(ctx->ifc_dev,
5683 		    "Unable to allocate device TX queue\n");
5684 		iflib_tx_structures_free(ctx);
5685 		free(vaddrs, M_IFLIB);
5686 		free(paddrs, M_IFLIB);
5687 		goto err_rx_desc;
5688 	}
5689 	free(vaddrs, M_IFLIB);
5690 	free(paddrs, M_IFLIB);
5691 
5692 	/* RXQs */
5693 	vaddrs = malloc(sizeof(caddr_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5694 	paddrs = malloc(sizeof(uint64_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5695 	for (i = 0; i < nrxqsets; i++) {
5696 		iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi;
5697 
5698 		for (j = 0; j < nrxqs; j++, di++) {
5699 			vaddrs[i*nrxqs + j] = di->idi_vaddr;
5700 			paddrs[i*nrxqs + j] = di->idi_paddr;
5701 		}
5702 	}
5703 	if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) {
5704 		device_printf(ctx->ifc_dev,
5705 		    "Unable to allocate device RX queue\n");
5706 		iflib_tx_structures_free(ctx);
5707 		free(vaddrs, M_IFLIB);
5708 		free(paddrs, M_IFLIB);
5709 		goto err_rx_desc;
5710 	}
5711 	free(vaddrs, M_IFLIB);
5712 	free(paddrs, M_IFLIB);
5713 
5714 	return (0);
5715 
5716 /* XXX handle allocation failure changes */
5717 err_rx_desc:
5718 err_tx_desc:
5719 rx_fail:
5720 	if (ctx->ifc_rxqs != NULL)
5721 		free(ctx->ifc_rxqs, M_IFLIB);
5722 	ctx->ifc_rxqs = NULL;
5723 	if (ctx->ifc_txqs != NULL)
5724 		free(ctx->ifc_txqs, M_IFLIB);
5725 	ctx->ifc_txqs = NULL;
5726 fail:
5727 	return (err);
5728 }
5729 
5730 static int
5731 iflib_tx_structures_setup(if_ctx_t ctx)
5732 {
5733 	iflib_txq_t txq = ctx->ifc_txqs;
5734 	int i;
5735 
5736 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
5737 		iflib_txq_setup(txq);
5738 
5739 	return (0);
5740 }
5741 
5742 static void
5743 iflib_tx_structures_free(if_ctx_t ctx)
5744 {
5745 	iflib_txq_t txq = ctx->ifc_txqs;
5746 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5747 	int i, j;
5748 
5749 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
5750 		for (j = 0; j < sctx->isc_ntxqs; j++)
5751 			iflib_dma_free(&txq->ift_ifdi[j]);
5752 		iflib_txq_destroy(txq);
5753 	}
5754 	free(ctx->ifc_txqs, M_IFLIB);
5755 	ctx->ifc_txqs = NULL;
5756 	IFDI_QUEUES_FREE(ctx);
5757 }
5758 
5759 /*********************************************************************
5760  *
5761  *  Initialize all receive rings.
5762  *
5763  **********************************************************************/
5764 static int
5765 iflib_rx_structures_setup(if_ctx_t ctx)
5766 {
5767 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5768 	int q;
5769 #if defined(INET6) || defined(INET)
5770 	int err, i;
5771 #endif
5772 
5773 	for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) {
5774 #if defined(INET6) || defined(INET)
5775 		if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_LRO) {
5776 			err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp,
5777 			    TCP_LRO_ENTRIES, min(1024,
5778 			    ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset]));
5779 			if (err != 0) {
5780 				device_printf(ctx->ifc_dev,
5781 				    "LRO Initialization failed!\n");
5782 				goto fail;
5783 			}
5784 		}
5785 #endif
5786 		IFDI_RXQ_SETUP(ctx, rxq->ifr_id);
5787 	}
5788 	return (0);
5789 #if defined(INET6) || defined(INET)
5790 fail:
5791 	/*
5792 	 * Free LRO resources allocated so far, we will only handle
5793 	 * the rings that completed, the failing case will have
5794 	 * cleaned up for itself.  'q' failed, so its the terminus.
5795 	 */
5796 	rxq = ctx->ifc_rxqs;
5797 	for (i = 0; i < q; ++i, rxq++) {
5798 		if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_LRO)
5799 			tcp_lro_free(&rxq->ifr_lc);
5800 	}
5801 	return (err);
5802 #endif
5803 }
5804 
5805 /*********************************************************************
5806  *
5807  *  Free all receive rings.
5808  *
5809  **********************************************************************/
5810 static void
5811 iflib_rx_structures_free(if_ctx_t ctx)
5812 {
5813 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5814 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5815 	int i, j;
5816 
5817 	for (i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) {
5818 		for (j = 0; j < sctx->isc_nrxqs; j++)
5819 			iflib_dma_free(&rxq->ifr_ifdi[j]);
5820 		iflib_rx_sds_free(rxq);
5821 #if defined(INET6) || defined(INET)
5822 		if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_LRO)
5823 			tcp_lro_free(&rxq->ifr_lc);
5824 #endif
5825 	}
5826 	free(ctx->ifc_rxqs, M_IFLIB);
5827 	ctx->ifc_rxqs = NULL;
5828 }
5829 
5830 static int
5831 iflib_qset_structures_setup(if_ctx_t ctx)
5832 {
5833 	int err;
5834 
5835 	/*
5836 	 * It is expected that the caller takes care of freeing queues if this
5837 	 * fails.
5838 	 */
5839 	if ((err = iflib_tx_structures_setup(ctx)) != 0) {
5840 		device_printf(ctx->ifc_dev, "iflib_tx_structures_setup failed: %d\n", err);
5841 		return (err);
5842 	}
5843 
5844 	if ((err = iflib_rx_structures_setup(ctx)) != 0)
5845 		device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err);
5846 
5847 	return (err);
5848 }
5849 
5850 int
5851 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
5852 		driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, const char *name)
5853 {
5854 
5855 	return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name));
5856 }
5857 
5858 #ifdef SMP
5859 static int
5860 find_nth(if_ctx_t ctx, int qid)
5861 {
5862 	cpuset_t cpus;
5863 	int i, cpuid, eqid, count;
5864 
5865 	CPU_COPY(&ctx->ifc_cpus, &cpus);
5866 	count = CPU_COUNT(&cpus);
5867 	eqid = qid % count;
5868 	/* clear up to the qid'th bit */
5869 	for (i = 0; i < eqid; i++) {
5870 		cpuid = CPU_FFS(&cpus);
5871 		MPASS(cpuid != 0);
5872 		CPU_CLR(cpuid-1, &cpus);
5873 	}
5874 	cpuid = CPU_FFS(&cpus);
5875 	MPASS(cpuid != 0);
5876 	return (cpuid-1);
5877 }
5878 
5879 #ifdef SCHED_ULE
5880 extern struct cpu_group *cpu_top;              /* CPU topology */
5881 
5882 static int
5883 find_child_with_core(int cpu, struct cpu_group *grp)
5884 {
5885 	int i;
5886 
5887 	if (grp->cg_children == 0)
5888 		return -1;
5889 
5890 	MPASS(grp->cg_child);
5891 	for (i = 0; i < grp->cg_children; i++) {
5892 		if (CPU_ISSET(cpu, &grp->cg_child[i].cg_mask))
5893 			return i;
5894 	}
5895 
5896 	return -1;
5897 }
5898 
5899 /*
5900  * Find the nth "close" core to the specified core
5901  * "close" is defined as the deepest level that shares
5902  * at least an L2 cache.  With threads, this will be
5903  * threads on the same core.  If the shared cache is L3
5904  * or higher, simply returns the same core.
5905  */
5906 static int
5907 find_close_core(int cpu, int core_offset)
5908 {
5909 	struct cpu_group *grp;
5910 	int i;
5911 	int fcpu;
5912 	cpuset_t cs;
5913 
5914 	grp = cpu_top;
5915 	if (grp == NULL)
5916 		return cpu;
5917 	i = 0;
5918 	while ((i = find_child_with_core(cpu, grp)) != -1) {
5919 		/* If the child only has one cpu, don't descend */
5920 		if (grp->cg_child[i].cg_count <= 1)
5921 			break;
5922 		grp = &grp->cg_child[i];
5923 	}
5924 
5925 	/* If they don't share at least an L2 cache, use the same CPU */
5926 	if (grp->cg_level > CG_SHARE_L2 || grp->cg_level == CG_SHARE_NONE)
5927 		return cpu;
5928 
5929 	/* Now pick one */
5930 	CPU_COPY(&grp->cg_mask, &cs);
5931 
5932 	/* Add the selected CPU offset to core offset. */
5933 	for (i = 0; (fcpu = CPU_FFS(&cs)) != 0; i++) {
5934 		if (fcpu - 1 == cpu)
5935 			break;
5936 		CPU_CLR(fcpu - 1, &cs);
5937 	}
5938 	MPASS(fcpu);
5939 
5940 	core_offset += i;
5941 
5942 	CPU_COPY(&grp->cg_mask, &cs);
5943 	for (i = core_offset % grp->cg_count; i > 0; i--) {
5944 		MPASS(CPU_FFS(&cs));
5945 		CPU_CLR(CPU_FFS(&cs) - 1, &cs);
5946 	}
5947 	MPASS(CPU_FFS(&cs));
5948 	return CPU_FFS(&cs) - 1;
5949 }
5950 #else
5951 static int
5952 find_close_core(int cpu, int core_offset __unused)
5953 {
5954 	return cpu;
5955 }
5956 #endif
5957 
5958 static int
5959 get_core_offset(if_ctx_t ctx, iflib_intr_type_t type, int qid)
5960 {
5961 	switch (type) {
5962 	case IFLIB_INTR_TX:
5963 		/* TX queues get cores which share at least an L2 cache with the corresponding RX queue */
5964 		/* XXX handle multiple RX threads per core and more than two core per L2 group */
5965 		return qid / CPU_COUNT(&ctx->ifc_cpus) + 1;
5966 	case IFLIB_INTR_RX:
5967 	case IFLIB_INTR_RXTX:
5968 		/* RX queues get the specified core */
5969 		return qid / CPU_COUNT(&ctx->ifc_cpus);
5970 	default:
5971 		return -1;
5972 	}
5973 }
5974 #else
5975 #define get_core_offset(ctx, type, qid)	CPU_FIRST()
5976 #define find_close_core(cpuid, tid)	CPU_FIRST()
5977 #define find_nth(ctx, gid)		CPU_FIRST()
5978 #endif
5979 
5980 /* Just to avoid copy/paste */
5981 static inline int
5982 iflib_irq_set_affinity(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type,
5983     int qid, struct grouptask *gtask, struct taskqgroup *tqg, void *uniq,
5984     const char *name)
5985 {
5986 	device_t dev;
5987 	int co, cpuid, err, tid;
5988 
5989 	dev = ctx->ifc_dev;
5990 	co = ctx->ifc_sysctl_core_offset;
5991 	if (ctx->ifc_sysctl_separate_txrx && type == IFLIB_INTR_TX)
5992 		co += ctx->ifc_softc_ctx.isc_nrxqsets;
5993 	cpuid = find_nth(ctx, qid + co);
5994 	tid = get_core_offset(ctx, type, qid);
5995 	if (tid < 0) {
5996 		device_printf(dev, "get_core_offset failed\n");
5997 		return (EOPNOTSUPP);
5998 	}
5999 	cpuid = find_close_core(cpuid, tid);
6000 	err = taskqgroup_attach_cpu(tqg, gtask, uniq, cpuid, dev, irq->ii_res,
6001 	    name);
6002 	if (err) {
6003 		device_printf(dev, "taskqgroup_attach_cpu failed %d\n", err);
6004 		return (err);
6005 	}
6006 #ifdef notyet
6007 	if (cpuid > ctx->ifc_cpuid_highest)
6008 		ctx->ifc_cpuid_highest = cpuid;
6009 #endif
6010 	return (0);
6011 }
6012 
6013 int
6014 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid,
6015 			iflib_intr_type_t type, driver_filter_t *filter,
6016 			void *filter_arg, int qid, const char *name)
6017 {
6018 	device_t dev;
6019 	struct grouptask *gtask;
6020 	struct taskqgroup *tqg;
6021 	iflib_filter_info_t info;
6022 	gtask_fn_t *fn;
6023 	int tqrid, err;
6024 	driver_filter_t *intr_fast;
6025 	void *q;
6026 
6027 	info = &ctx->ifc_filter_info;
6028 	tqrid = rid;
6029 
6030 	switch (type) {
6031 	/* XXX merge tx/rx for netmap? */
6032 	case IFLIB_INTR_TX:
6033 		q = &ctx->ifc_txqs[qid];
6034 		info = &ctx->ifc_txqs[qid].ift_filter_info;
6035 		gtask = &ctx->ifc_txqs[qid].ift_task;
6036 		tqg = qgroup_if_io_tqg;
6037 		fn = _task_fn_tx;
6038 		intr_fast = iflib_fast_intr;
6039 		GROUPTASK_INIT(gtask, 0, fn, q);
6040 		ctx->ifc_flags |= IFC_NETMAP_TX_IRQ;
6041 		break;
6042 	case IFLIB_INTR_RX:
6043 		q = &ctx->ifc_rxqs[qid];
6044 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
6045 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6046 		tqg = qgroup_if_io_tqg;
6047 		fn = _task_fn_rx;
6048 		intr_fast = iflib_fast_intr;
6049 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6050 		break;
6051 	case IFLIB_INTR_RXTX:
6052 		q = &ctx->ifc_rxqs[qid];
6053 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
6054 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6055 		tqg = qgroup_if_io_tqg;
6056 		fn = _task_fn_rx;
6057 		intr_fast = iflib_fast_intr_rxtx;
6058 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6059 		break;
6060 	case IFLIB_INTR_ADMIN:
6061 		q = ctx;
6062 		tqrid = -1;
6063 		info = &ctx->ifc_filter_info;
6064 		gtask = &ctx->ifc_admin_task;
6065 		tqg = qgroup_if_config_tqg;
6066 		fn = _task_fn_admin;
6067 		intr_fast = iflib_fast_intr_ctx;
6068 		break;
6069 	default:
6070 		device_printf(ctx->ifc_dev, "%s: unknown net intr type\n",
6071 		    __func__);
6072 		return (EINVAL);
6073 	}
6074 
6075 	info->ifi_filter = filter;
6076 	info->ifi_filter_arg = filter_arg;
6077 	info->ifi_task = gtask;
6078 	info->ifi_ctx = q;
6079 
6080 	dev = ctx->ifc_dev;
6081 	err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info,  name);
6082 	if (err != 0) {
6083 		device_printf(dev, "_iflib_irq_alloc failed %d\n", err);
6084 		return (err);
6085 	}
6086 	if (type == IFLIB_INTR_ADMIN)
6087 		return (0);
6088 
6089 	if (tqrid != -1) {
6090 		err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg,
6091 		    q, name);
6092 		if (err)
6093 			return (err);
6094 	} else {
6095 		taskqgroup_attach(tqg, gtask, q, dev, irq->ii_res, name);
6096 	}
6097 
6098 	return (0);
6099 }
6100 
6101 void
6102 iflib_softirq_alloc_generic(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type, void *arg, int qid, const char *name)
6103 {
6104 	struct grouptask *gtask;
6105 	struct taskqgroup *tqg;
6106 	gtask_fn_t *fn;
6107 	void *q;
6108 	int err;
6109 
6110 	switch (type) {
6111 	case IFLIB_INTR_TX:
6112 		q = &ctx->ifc_txqs[qid];
6113 		gtask = &ctx->ifc_txqs[qid].ift_task;
6114 		tqg = qgroup_if_io_tqg;
6115 		fn = _task_fn_tx;
6116 		GROUPTASK_INIT(gtask, 0, fn, q);
6117 		break;
6118 	case IFLIB_INTR_RX:
6119 		q = &ctx->ifc_rxqs[qid];
6120 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6121 		tqg = qgroup_if_io_tqg;
6122 		fn = _task_fn_rx;
6123 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6124 		break;
6125 	case IFLIB_INTR_IOV:
6126 		q = ctx;
6127 		gtask = &ctx->ifc_vflr_task;
6128 		tqg = qgroup_if_config_tqg;
6129 		fn = _task_fn_iov;
6130 		GROUPTASK_INIT(gtask, 0, fn, q);
6131 		break;
6132 	default:
6133 		panic("unknown net intr type");
6134 	}
6135 	if (irq != NULL) {
6136 		err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg,
6137 		    q, name);
6138 		if (err)
6139 			taskqgroup_attach(tqg, gtask, q, ctx->ifc_dev,
6140 			    irq->ii_res, name);
6141 	} else {
6142 		taskqgroup_attach(tqg, gtask, q, NULL, NULL, name);
6143 	}
6144 }
6145 
6146 void
6147 iflib_irq_free(if_ctx_t ctx, if_irq_t irq)
6148 {
6149 
6150 	if (irq->ii_tag)
6151 		bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag);
6152 
6153 	if (irq->ii_res)
6154 		bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ,
6155 		    rman_get_rid(irq->ii_res), irq->ii_res);
6156 }
6157 
6158 static int
6159 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, const char *name)
6160 {
6161 	iflib_txq_t txq = ctx->ifc_txqs;
6162 	iflib_rxq_t rxq = ctx->ifc_rxqs;
6163 	if_irq_t irq = &ctx->ifc_legacy_irq;
6164 	iflib_filter_info_t info;
6165 	device_t dev;
6166 	struct grouptask *gtask;
6167 	struct resource *res;
6168 	struct taskqgroup *tqg;
6169 	void *q;
6170 	int err, tqrid;
6171 	bool rx_only;
6172 
6173 	q = &ctx->ifc_rxqs[0];
6174 	info = &rxq[0].ifr_filter_info;
6175 	gtask = &rxq[0].ifr_task;
6176 	tqg = qgroup_if_io_tqg;
6177 	tqrid = *rid;
6178 	rx_only = (ctx->ifc_sctx->isc_flags & IFLIB_SINGLE_IRQ_RX_ONLY) != 0;
6179 
6180 	ctx->ifc_flags |= IFC_LEGACY;
6181 	info->ifi_filter = filter;
6182 	info->ifi_filter_arg = filter_arg;
6183 	info->ifi_task = gtask;
6184 	info->ifi_ctx = rx_only ? ctx : q;
6185 
6186 	dev = ctx->ifc_dev;
6187 	/* We allocate a single interrupt resource */
6188 	err = _iflib_irq_alloc(ctx, irq, tqrid, rx_only ? iflib_fast_intr_ctx :
6189 	    iflib_fast_intr_rxtx, NULL, info, name);
6190 	if (err != 0)
6191 		return (err);
6192 	NET_GROUPTASK_INIT(gtask, 0, _task_fn_rx, q);
6193 	res = irq->ii_res;
6194 	taskqgroup_attach(tqg, gtask, q, dev, res, name);
6195 
6196 	GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq);
6197 	taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, dev, res,
6198 	    "tx");
6199 	return (0);
6200 }
6201 
6202 void
6203 iflib_led_create(if_ctx_t ctx)
6204 {
6205 
6206 	ctx->ifc_led_dev = led_create(iflib_led_func, ctx,
6207 	    device_get_nameunit(ctx->ifc_dev));
6208 }
6209 
6210 void
6211 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid)
6212 {
6213 
6214 	GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
6215 }
6216 
6217 void
6218 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid)
6219 {
6220 
6221 	GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task);
6222 }
6223 
6224 void
6225 iflib_admin_intr_deferred(if_ctx_t ctx)
6226 {
6227 
6228 	MPASS(ctx->ifc_admin_task.gt_taskqueue != NULL);
6229 	GROUPTASK_ENQUEUE(&ctx->ifc_admin_task);
6230 }
6231 
6232 void
6233 iflib_iov_intr_deferred(if_ctx_t ctx)
6234 {
6235 
6236 	GROUPTASK_ENQUEUE(&ctx->ifc_vflr_task);
6237 }
6238 
6239 void
6240 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, const char *name)
6241 {
6242 
6243 	taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, NULL, NULL,
6244 	    name);
6245 }
6246 
6247 void
6248 iflib_config_gtask_init(void *ctx, struct grouptask *gtask, gtask_fn_t *fn,
6249 	const char *name)
6250 {
6251 
6252 	GROUPTASK_INIT(gtask, 0, fn, ctx);
6253 	taskqgroup_attach(qgroup_if_config_tqg, gtask, gtask, NULL, NULL,
6254 	    name);
6255 }
6256 
6257 void
6258 iflib_config_gtask_deinit(struct grouptask *gtask)
6259 {
6260 
6261 	taskqgroup_detach(qgroup_if_config_tqg, gtask);
6262 }
6263 
6264 void
6265 iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate)
6266 {
6267 	if_t ifp = ctx->ifc_ifp;
6268 	iflib_txq_t txq = ctx->ifc_txqs;
6269 
6270 	if_setbaudrate(ifp, baudrate);
6271 	if (baudrate >= IF_Gbps(10)) {
6272 		STATE_LOCK(ctx);
6273 		ctx->ifc_flags |= IFC_PREFETCH;
6274 		STATE_UNLOCK(ctx);
6275 	}
6276 	/* If link down, disable watchdog */
6277 	if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) {
6278 		for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++)
6279 			txq->ift_qstatus = IFLIB_QUEUE_IDLE;
6280 	}
6281 	ctx->ifc_link_state = link_state;
6282 	if_link_state_change(ifp, link_state);
6283 }
6284 
6285 static int
6286 iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq)
6287 {
6288 	int credits;
6289 #ifdef INVARIANTS
6290 	int credits_pre = txq->ift_cidx_processed;
6291 #endif
6292 
6293 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
6294 	    BUS_DMASYNC_POSTREAD);
6295 	if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, true)) == 0)
6296 		return (0);
6297 
6298 	txq->ift_processed += credits;
6299 	txq->ift_cidx_processed += credits;
6300 
6301 	MPASS(credits_pre + credits == txq->ift_cidx_processed);
6302 	if (txq->ift_cidx_processed >= txq->ift_size)
6303 		txq->ift_cidx_processed -= txq->ift_size;
6304 	return (credits);
6305 }
6306 
6307 static int
6308 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget)
6309 {
6310 	iflib_fl_t fl;
6311 	u_int i;
6312 
6313 	for (i = 0, fl = &rxq->ifr_fl[0]; i < rxq->ifr_nfl; i++, fl++)
6314 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
6315 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
6316 	return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx,
6317 	    budget));
6318 }
6319 
6320 void
6321 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name,
6322 	const char *description, if_int_delay_info_t info,
6323 	int offset, int value)
6324 {
6325 	info->iidi_ctx = ctx;
6326 	info->iidi_offset = offset;
6327 	info->iidi_value = value;
6328 	SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev),
6329 	    SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)),
6330 	    OID_AUTO, name, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE,
6331 	    info, 0, iflib_sysctl_int_delay, "I", description);
6332 }
6333 
6334 struct sx *
6335 iflib_ctx_lock_get(if_ctx_t ctx)
6336 {
6337 
6338 	return (&ctx->ifc_ctx_sx);
6339 }
6340 
6341 static int
6342 iflib_msix_init(if_ctx_t ctx)
6343 {
6344 	device_t dev = ctx->ifc_dev;
6345 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6346 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6347 	int admincnt, bar, err, iflib_num_rx_queues, iflib_num_tx_queues;
6348 	int msgs, queuemsgs, queues, rx_queues, tx_queues, vectors;
6349 
6350 	iflib_num_tx_queues = ctx->ifc_sysctl_ntxqs;
6351 	iflib_num_rx_queues = ctx->ifc_sysctl_nrxqs;
6352 
6353 	if (bootverbose)
6354 		device_printf(dev, "msix_init qsets capped at %d\n",
6355 		    imax(scctx->isc_ntxqsets, scctx->isc_nrxqsets));
6356 
6357 	/* Override by tuneable */
6358 	if (scctx->isc_disable_msix)
6359 		goto msi;
6360 
6361 	/* First try MSI-X */
6362 	if ((msgs = pci_msix_count(dev)) == 0) {
6363 		if (bootverbose)
6364 			device_printf(dev, "MSI-X not supported or disabled\n");
6365 		goto msi;
6366 	}
6367 
6368 	bar = ctx->ifc_softc_ctx.isc_msix_bar;
6369 	/*
6370 	 * bar == -1 => "trust me I know what I'm doing"
6371 	 * Some drivers are for hardware that is so shoddily
6372 	 * documented that no one knows which bars are which
6373 	 * so the developer has to map all bars. This hack
6374 	 * allows shoddy garbage to use MSI-X in this framework.
6375 	 */
6376 	if (bar != -1) {
6377 		ctx->ifc_msix_mem = bus_alloc_resource_any(dev,
6378 	            SYS_RES_MEMORY, &bar, RF_ACTIVE);
6379 		if (ctx->ifc_msix_mem == NULL) {
6380 			device_printf(dev, "Unable to map MSI-X table\n");
6381 			goto msi;
6382 		}
6383 	}
6384 
6385 	admincnt = sctx->isc_admin_intrcnt;
6386 #if IFLIB_DEBUG
6387 	/* use only 1 qset in debug mode */
6388 	queuemsgs = min(msgs - admincnt, 1);
6389 #else
6390 	queuemsgs = msgs - admincnt;
6391 #endif
6392 #ifdef RSS
6393 	queues = imin(queuemsgs, rss_getnumbuckets());
6394 #else
6395 	queues = queuemsgs;
6396 #endif
6397 	queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues);
6398 	if (bootverbose)
6399 		device_printf(dev,
6400 		    "intr CPUs: %d queue msgs: %d admincnt: %d\n",
6401 		    CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt);
6402 #ifdef  RSS
6403 	/* If we're doing RSS, clamp at the number of RSS buckets */
6404 	if (queues > rss_getnumbuckets())
6405 		queues = rss_getnumbuckets();
6406 #endif
6407 	if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt)
6408 		rx_queues = iflib_num_rx_queues;
6409 	else
6410 		rx_queues = queues;
6411 
6412 	if (rx_queues > scctx->isc_nrxqsets)
6413 		rx_queues = scctx->isc_nrxqsets;
6414 
6415 	/*
6416 	 * We want this to be all logical CPUs by default
6417 	 */
6418 	if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues)
6419 		tx_queues = iflib_num_tx_queues;
6420 	else
6421 		tx_queues = mp_ncpus;
6422 
6423 	if (tx_queues > scctx->isc_ntxqsets)
6424 		tx_queues = scctx->isc_ntxqsets;
6425 
6426 	if (ctx->ifc_sysctl_qs_eq_override == 0) {
6427 #ifdef INVARIANTS
6428 		if (tx_queues != rx_queues)
6429 			device_printf(dev,
6430 			    "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n",
6431 			    min(rx_queues, tx_queues), min(rx_queues, tx_queues));
6432 #endif
6433 		tx_queues = min(rx_queues, tx_queues);
6434 		rx_queues = min(rx_queues, tx_queues);
6435 	}
6436 
6437 	vectors = rx_queues + admincnt;
6438 	if (msgs < vectors) {
6439 		device_printf(dev,
6440 		    "insufficient number of MSI-X vectors "
6441 		    "(supported %d, need %d)\n", msgs, vectors);
6442 		goto msi;
6443 	}
6444 
6445 	device_printf(dev, "Using %d RX queues %d TX queues\n", rx_queues,
6446 	    tx_queues);
6447 	msgs = vectors;
6448 	if ((err = pci_alloc_msix(dev, &vectors)) == 0) {
6449 		if (vectors != msgs) {
6450 			device_printf(dev,
6451 			    "Unable to allocate sufficient MSI-X vectors "
6452 			    "(got %d, need %d)\n", vectors, msgs);
6453 			pci_release_msi(dev);
6454 			if (bar != -1) {
6455 				bus_release_resource(dev, SYS_RES_MEMORY, bar,
6456 				    ctx->ifc_msix_mem);
6457 				ctx->ifc_msix_mem = NULL;
6458 			}
6459 			goto msi;
6460 		}
6461 		device_printf(dev, "Using MSI-X interrupts with %d vectors\n",
6462 		    vectors);
6463 		scctx->isc_vectors = vectors;
6464 		scctx->isc_nrxqsets = rx_queues;
6465 		scctx->isc_ntxqsets = tx_queues;
6466 		scctx->isc_intr = IFLIB_INTR_MSIX;
6467 
6468 		return (vectors);
6469 	} else {
6470 		device_printf(dev,
6471 		    "failed to allocate %d MSI-X vectors, err: %d\n", vectors,
6472 		    err);
6473 		if (bar != -1) {
6474 			bus_release_resource(dev, SYS_RES_MEMORY, bar,
6475 			    ctx->ifc_msix_mem);
6476 			ctx->ifc_msix_mem = NULL;
6477 		}
6478 	}
6479 
6480 msi:
6481 	vectors = pci_msi_count(dev);
6482 	scctx->isc_nrxqsets = 1;
6483 	scctx->isc_ntxqsets = 1;
6484 	scctx->isc_vectors = vectors;
6485 	if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) {
6486 		device_printf(dev,"Using an MSI interrupt\n");
6487 		scctx->isc_intr = IFLIB_INTR_MSI;
6488 	} else {
6489 		scctx->isc_vectors = 1;
6490 		device_printf(dev,"Using a Legacy interrupt\n");
6491 		scctx->isc_intr = IFLIB_INTR_LEGACY;
6492 	}
6493 
6494 	return (vectors);
6495 }
6496 
6497 static const char *ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" };
6498 
6499 static int
6500 mp_ring_state_handler(SYSCTL_HANDLER_ARGS)
6501 {
6502 	int rc;
6503 	uint16_t *state = ((uint16_t *)oidp->oid_arg1);
6504 	struct sbuf *sb;
6505 	const char *ring_state = "UNKNOWN";
6506 
6507 	/* XXX needed ? */
6508 	rc = sysctl_wire_old_buffer(req, 0);
6509 	MPASS(rc == 0);
6510 	if (rc != 0)
6511 		return (rc);
6512 	sb = sbuf_new_for_sysctl(NULL, NULL, 80, req);
6513 	MPASS(sb != NULL);
6514 	if (sb == NULL)
6515 		return (ENOMEM);
6516 	if (state[3] <= 3)
6517 		ring_state = ring_states[state[3]];
6518 
6519 	sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s",
6520 		    state[0], state[1], state[2], ring_state);
6521 	rc = sbuf_finish(sb);
6522 	sbuf_delete(sb);
6523         return(rc);
6524 }
6525 
6526 enum iflib_ndesc_handler {
6527 	IFLIB_NTXD_HANDLER,
6528 	IFLIB_NRXD_HANDLER,
6529 };
6530 
6531 static int
6532 mp_ndesc_handler(SYSCTL_HANDLER_ARGS)
6533 {
6534 	if_ctx_t ctx = (void *)arg1;
6535 	enum iflib_ndesc_handler type = arg2;
6536 	char buf[256] = {0};
6537 	qidx_t *ndesc;
6538 	char *p, *next;
6539 	int nqs, rc, i;
6540 
6541 	nqs = 8;
6542 	switch(type) {
6543 	case IFLIB_NTXD_HANDLER:
6544 		ndesc = ctx->ifc_sysctl_ntxds;
6545 		if (ctx->ifc_sctx)
6546 			nqs = ctx->ifc_sctx->isc_ntxqs;
6547 		break;
6548 	case IFLIB_NRXD_HANDLER:
6549 		ndesc = ctx->ifc_sysctl_nrxds;
6550 		if (ctx->ifc_sctx)
6551 			nqs = ctx->ifc_sctx->isc_nrxqs;
6552 		break;
6553 	default:
6554 		printf("%s: unhandled type\n", __func__);
6555 		return (EINVAL);
6556 	}
6557 	if (nqs == 0)
6558 		nqs = 8;
6559 
6560 	for (i=0; i<8; i++) {
6561 		if (i >= nqs)
6562 			break;
6563 		if (i)
6564 			strcat(buf, ",");
6565 		sprintf(strchr(buf, 0), "%d", ndesc[i]);
6566 	}
6567 
6568 	rc = sysctl_handle_string(oidp, buf, sizeof(buf), req);
6569 	if (rc || req->newptr == NULL)
6570 		return rc;
6571 
6572 	for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p;
6573 	    i++, p = strsep(&next, " ,")) {
6574 		ndesc[i] = strtoul(p, NULL, 10);
6575 	}
6576 
6577 	return(rc);
6578 }
6579 
6580 #define NAME_BUFLEN 32
6581 static void
6582 iflib_add_device_sysctl_pre(if_ctx_t ctx)
6583 {
6584         device_t dev = iflib_get_dev(ctx);
6585 	struct sysctl_oid_list *child, *oid_list;
6586 	struct sysctl_ctx_list *ctx_list;
6587 	struct sysctl_oid *node;
6588 
6589 	ctx_list = device_get_sysctl_ctx(dev);
6590 	child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev));
6591 	ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, "iflib",
6592 	    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "IFLIB fields");
6593 	oid_list = SYSCTL_CHILDREN(node);
6594 
6595 	SYSCTL_ADD_CONST_STRING(ctx_list, oid_list, OID_AUTO, "driver_version",
6596 		       CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version,
6597 		       "driver version");
6598 
6599 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs",
6600 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0,
6601 			"# of txqs to use, 0 => use default #");
6602 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs",
6603 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0,
6604 			"# of rxqs to use, 0 => use default #");
6605 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_qs_enable",
6606 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0,
6607                        "permit #txq != #rxq");
6608 	SYSCTL_ADD_INT(ctx_list, oid_list, OID_AUTO, "disable_msix",
6609                       CTLFLAG_RWTUN, &ctx->ifc_softc_ctx.isc_disable_msix, 0,
6610                       "disable MSI-X (default 0)");
6611 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "rx_budget",
6612 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_rx_budget, 0,
6613 		       "set the RX budget");
6614 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "tx_abdicate",
6615 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_tx_abdicate, 0,
6616 		       "cause TX to abdicate instead of running to completion");
6617 	ctx->ifc_sysctl_core_offset = CORE_OFFSET_UNSPECIFIED;
6618 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "core_offset",
6619 		       CTLFLAG_RDTUN, &ctx->ifc_sysctl_core_offset, 0,
6620 		       "offset to start using cores at");
6621 	SYSCTL_ADD_U8(ctx_list, oid_list, OID_AUTO, "separate_txrx",
6622 		       CTLFLAG_RDTUN, &ctx->ifc_sysctl_separate_txrx, 0,
6623 		       "use separate cores for TX and RX");
6624 
6625 	/* XXX change for per-queue sizes */
6626 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_ntxds",
6627 	    CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NEEDGIANT, ctx,
6628 	    IFLIB_NTXD_HANDLER, mp_ndesc_handler, "A",
6629 	    "list of # of TX descriptors to use, 0 = use default #");
6630 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_nrxds",
6631 	    CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NEEDGIANT, ctx,
6632 	    IFLIB_NRXD_HANDLER, mp_ndesc_handler, "A",
6633 	    "list of # of RX descriptors to use, 0 = use default #");
6634 }
6635 
6636 static void
6637 iflib_add_device_sysctl_post(if_ctx_t ctx)
6638 {
6639 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6640 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6641         device_t dev = iflib_get_dev(ctx);
6642 	struct sysctl_oid_list *child;
6643 	struct sysctl_ctx_list *ctx_list;
6644 	iflib_fl_t fl;
6645 	iflib_txq_t txq;
6646 	iflib_rxq_t rxq;
6647 	int i, j;
6648 	char namebuf[NAME_BUFLEN];
6649 	char *qfmt;
6650 	struct sysctl_oid *queue_node, *fl_node, *node;
6651 	struct sysctl_oid_list *queue_list, *fl_list;
6652 	ctx_list = device_get_sysctl_ctx(dev);
6653 
6654 	node = ctx->ifc_sysctl_node;
6655 	child = SYSCTL_CHILDREN(node);
6656 
6657 	if (scctx->isc_ntxqsets > 100)
6658 		qfmt = "txq%03d";
6659 	else if (scctx->isc_ntxqsets > 10)
6660 		qfmt = "txq%02d";
6661 	else
6662 		qfmt = "txq%d";
6663 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
6664 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6665 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6666 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "Queue Name");
6667 		queue_list = SYSCTL_CHILDREN(queue_node);
6668 #if MEMORY_LOGGING
6669 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued",
6670 				CTLFLAG_RD,
6671 				&txq->ift_dequeued, "total mbufs freed");
6672 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued",
6673 				CTLFLAG_RD,
6674 				&txq->ift_enqueued, "total mbufs enqueued");
6675 #endif
6676 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag",
6677 				   CTLFLAG_RD,
6678 				   &txq->ift_mbuf_defrag, "# of times m_defrag was called");
6679 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "m_pullups",
6680 				   CTLFLAG_RD,
6681 				   &txq->ift_pullups, "# of times m_pullup was called");
6682 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag_failed",
6683 				   CTLFLAG_RD,
6684 				   &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed");
6685 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_desc_avail",
6686 				   CTLFLAG_RD,
6687 				   &txq->ift_no_desc_avail, "# of times no descriptors were available");
6688 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "tx_map_failed",
6689 				   CTLFLAG_RD,
6690 				   &txq->ift_map_failed, "# of times DMA map failed");
6691 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txd_encap_efbig",
6692 				   CTLFLAG_RD,
6693 				   &txq->ift_txd_encap_efbig, "# of times txd_encap returned EFBIG");
6694 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_tx_dma_setup",
6695 				   CTLFLAG_RD,
6696 				   &txq->ift_no_tx_dma_setup, "# of times map failed for other than EFBIG");
6697 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx",
6698 				   CTLFLAG_RD,
6699 				   &txq->ift_pidx, 1, "Producer Index");
6700 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx",
6701 				   CTLFLAG_RD,
6702 				   &txq->ift_cidx, 1, "Consumer Index");
6703 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx_processed",
6704 				   CTLFLAG_RD,
6705 				   &txq->ift_cidx_processed, 1, "Consumer Index seen by credit update");
6706 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use",
6707 				   CTLFLAG_RD,
6708 				   &txq->ift_in_use, 1, "descriptors in use");
6709 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_processed",
6710 				   CTLFLAG_RD,
6711 				   &txq->ift_processed, "descriptors procesed for clean");
6712 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned",
6713 				   CTLFLAG_RD,
6714 				   &txq->ift_cleaned, "total cleaned");
6715 		SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state",
6716 		    CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_NEEDGIANT,
6717 		    __DEVOLATILE(uint64_t *, &txq->ift_br->state), 0,
6718 		    mp_ring_state_handler, "A", "soft ring state");
6719 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_enqueues",
6720 				       CTLFLAG_RD, &txq->ift_br->enqueues,
6721 				       "# of enqueues to the mp_ring for this queue");
6722 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_drops",
6723 				       CTLFLAG_RD, &txq->ift_br->drops,
6724 				       "# of drops in the mp_ring for this queue");
6725 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_starts",
6726 				       CTLFLAG_RD, &txq->ift_br->starts,
6727 				       "# of normal consumer starts in the mp_ring for this queue");
6728 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_stalls",
6729 				       CTLFLAG_RD, &txq->ift_br->stalls,
6730 					       "# of consumer stalls in the mp_ring for this queue");
6731 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_restarts",
6732 			       CTLFLAG_RD, &txq->ift_br->restarts,
6733 				       "# of consumer restarts in the mp_ring for this queue");
6734 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_abdications",
6735 				       CTLFLAG_RD, &txq->ift_br->abdications,
6736 				       "# of consumer abdications in the mp_ring for this queue");
6737 	}
6738 
6739 	if (scctx->isc_nrxqsets > 100)
6740 		qfmt = "rxq%03d";
6741 	else if (scctx->isc_nrxqsets > 10)
6742 		qfmt = "rxq%02d";
6743 	else
6744 		qfmt = "rxq%d";
6745 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
6746 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6747 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6748 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "Queue Name");
6749 		queue_list = SYSCTL_CHILDREN(queue_node);
6750 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
6751 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_cidx",
6752 				       CTLFLAG_RD,
6753 				       &rxq->ifr_cq_cidx, 1, "Consumer Index");
6754 		}
6755 
6756 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
6757 			snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j);
6758 			fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list, OID_AUTO, namebuf,
6759 			    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "freelist Name");
6760 			fl_list = SYSCTL_CHILDREN(fl_node);
6761 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx",
6762 				       CTLFLAG_RD,
6763 				       &fl->ifl_pidx, 1, "Producer Index");
6764 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx",
6765 				       CTLFLAG_RD,
6766 				       &fl->ifl_cidx, 1, "Consumer Index");
6767 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits",
6768 				       CTLFLAG_RD,
6769 				       &fl->ifl_credits, 1, "credits available");
6770 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "buf_size",
6771 				       CTLFLAG_RD,
6772 				       &fl->ifl_buf_size, 1, "buffer size");
6773 #if MEMORY_LOGGING
6774 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_enqueued",
6775 					CTLFLAG_RD,
6776 					&fl->ifl_m_enqueued, "mbufs allocated");
6777 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_dequeued",
6778 					CTLFLAG_RD,
6779 					&fl->ifl_m_dequeued, "mbufs freed");
6780 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_enqueued",
6781 					CTLFLAG_RD,
6782 					&fl->ifl_cl_enqueued, "clusters allocated");
6783 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_dequeued",
6784 					CTLFLAG_RD,
6785 					&fl->ifl_cl_dequeued, "clusters freed");
6786 #endif
6787 
6788 		}
6789 	}
6790 
6791 }
6792 
6793 void
6794 iflib_request_reset(if_ctx_t ctx)
6795 {
6796 
6797 	STATE_LOCK(ctx);
6798 	ctx->ifc_flags |= IFC_DO_RESET;
6799 	STATE_UNLOCK(ctx);
6800 }
6801 
6802 #ifndef __NO_STRICT_ALIGNMENT
6803 static struct mbuf *
6804 iflib_fixup_rx(struct mbuf *m)
6805 {
6806 	struct mbuf *n;
6807 
6808 	if (m->m_len <= (MCLBYTES - ETHER_HDR_LEN)) {
6809 		bcopy(m->m_data, m->m_data + ETHER_HDR_LEN, m->m_len);
6810 		m->m_data += ETHER_HDR_LEN;
6811 		n = m;
6812 	} else {
6813 		MGETHDR(n, M_NOWAIT, MT_DATA);
6814 		if (n == NULL) {
6815 			m_freem(m);
6816 			return (NULL);
6817 		}
6818 		bcopy(m->m_data, n->m_data, ETHER_HDR_LEN);
6819 		m->m_data += ETHER_HDR_LEN;
6820 		m->m_len -= ETHER_HDR_LEN;
6821 		n->m_len = ETHER_HDR_LEN;
6822 		M_MOVE_PKTHDR(n, m);
6823 		n->m_next = m;
6824 	}
6825 	return (n);
6826 }
6827 #endif
6828 
6829 #ifdef DEBUGNET
6830 static void
6831 iflib_debugnet_init(if_t ifp, int *nrxr, int *ncl, int *clsize)
6832 {
6833 	if_ctx_t ctx;
6834 
6835 	ctx = if_getsoftc(ifp);
6836 	CTX_LOCK(ctx);
6837 	*nrxr = NRXQSETS(ctx);
6838 	*ncl = ctx->ifc_rxqs[0].ifr_fl->ifl_size;
6839 	*clsize = ctx->ifc_rxqs[0].ifr_fl->ifl_buf_size;
6840 	CTX_UNLOCK(ctx);
6841 }
6842 
6843 static void
6844 iflib_debugnet_event(if_t ifp, enum debugnet_ev event)
6845 {
6846 	if_ctx_t ctx;
6847 	if_softc_ctx_t scctx;
6848 	iflib_fl_t fl;
6849 	iflib_rxq_t rxq;
6850 	int i, j;
6851 
6852 	ctx = if_getsoftc(ifp);
6853 	scctx = &ctx->ifc_softc_ctx;
6854 
6855 	switch (event) {
6856 	case DEBUGNET_START:
6857 		for (i = 0; i < scctx->isc_nrxqsets; i++) {
6858 			rxq = &ctx->ifc_rxqs[i];
6859 			for (j = 0; j < rxq->ifr_nfl; j++) {
6860 				fl = rxq->ifr_fl;
6861 				fl->ifl_zone = m_getzone(fl->ifl_buf_size);
6862 			}
6863 		}
6864 		iflib_no_tx_batch = 1;
6865 		break;
6866 	default:
6867 		break;
6868 	}
6869 }
6870 
6871 static int
6872 iflib_debugnet_transmit(if_t ifp, struct mbuf *m)
6873 {
6874 	if_ctx_t ctx;
6875 	iflib_txq_t txq;
6876 	int error;
6877 
6878 	ctx = if_getsoftc(ifp);
6879 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
6880 	    IFF_DRV_RUNNING)
6881 		return (EBUSY);
6882 
6883 	txq = &ctx->ifc_txqs[0];
6884 	error = iflib_encap(txq, &m);
6885 	if (error == 0)
6886 		(void)iflib_txd_db_check(ctx, txq, true, txq->ift_in_use);
6887 	return (error);
6888 }
6889 
6890 static int
6891 iflib_debugnet_poll(if_t ifp, int count)
6892 {
6893 	struct epoch_tracker et;
6894 	if_ctx_t ctx;
6895 	if_softc_ctx_t scctx;
6896 	iflib_txq_t txq;
6897 	int i;
6898 
6899 	ctx = if_getsoftc(ifp);
6900 	scctx = &ctx->ifc_softc_ctx;
6901 
6902 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
6903 	    IFF_DRV_RUNNING)
6904 		return (EBUSY);
6905 
6906 	txq = &ctx->ifc_txqs[0];
6907 	(void)iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
6908 
6909 	NET_EPOCH_ENTER(et);
6910 	for (i = 0; i < scctx->isc_nrxqsets; i++)
6911 		(void)iflib_rxeof(&ctx->ifc_rxqs[i], 16 /* XXX */);
6912 	NET_EPOCH_EXIT(et);
6913 	return (0);
6914 }
6915 #endif /* DEBUGNET */
6916