xref: /freebsd/sys/net/iflib.c (revision b89a7cc2ed6e4398d5be502f5bb5885d1ec6ff0f)
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/jail.h>
41 #include <sys/kernel.h>
42 #include <sys/lock.h>
43 #include <sys/md5.h>
44 #include <sys/mutex.h>
45 #include <sys/module.h>
46 #include <sys/kobj.h>
47 #include <sys/rman.h>
48 #include <sys/proc.h>
49 #include <sys/sbuf.h>
50 #include <sys/smp.h>
51 #include <sys/socket.h>
52 #include <sys/sockio.h>
53 #include <sys/sysctl.h>
54 #include <sys/syslog.h>
55 #include <sys/taskqueue.h>
56 #include <sys/limits.h>
57 
58 #include <net/if.h>
59 #include <net/if_var.h>
60 #include <net/if_types.h>
61 #include <net/if_media.h>
62 #include <net/bpf.h>
63 #include <net/ethernet.h>
64 #include <net/mp_ring.h>
65 #include <net/vnet.h>
66 
67 #include <netinet/in.h>
68 #include <netinet/in_pcb.h>
69 #include <netinet/tcp_lro.h>
70 #include <netinet/in_systm.h>
71 #include <netinet/if_ether.h>
72 #include <netinet/ip.h>
73 #include <netinet/ip6.h>
74 #include <netinet/tcp.h>
75 #include <netinet/ip_var.h>
76 #include <netinet/netdump/netdump.h>
77 #include <netinet6/ip6_var.h>
78 
79 #include <machine/bus.h>
80 #include <machine/in_cksum.h>
81 
82 #include <vm/vm.h>
83 #include <vm/pmap.h>
84 
85 #include <dev/led/led.h>
86 #include <dev/pci/pcireg.h>
87 #include <dev/pci/pcivar.h>
88 #include <dev/pci/pci_private.h>
89 
90 #include <net/iflib.h>
91 #include <net/iflib_private.h>
92 
93 #include "ifdi_if.h"
94 
95 #ifdef PCI_IOV
96 #include <dev/pci/pci_iov.h>
97 #endif
98 
99 #include <sys/bitstring.h>
100 /*
101  * enable accounting of every mbuf as it comes in to and goes out of
102  * iflib's software descriptor references
103  */
104 #define MEMORY_LOGGING 0
105 /*
106  * Enable mbuf vectors for compressing long mbuf chains
107  */
108 
109 /*
110  * NB:
111  * - Prefetching in tx cleaning should perhaps be a tunable. The distance ahead
112  *   we prefetch needs to be determined by the time spent in m_free vis a vis
113  *   the cost of a prefetch. This will of course vary based on the workload:
114  *      - NFLX's m_free path is dominated by vm-based M_EXT manipulation which
115  *        is quite expensive, thus suggesting very little prefetch.
116  *      - small packet forwarding which is just returning a single mbuf to
117  *        UMA will typically be very fast vis a vis the cost of a memory
118  *        access.
119  */
120 
121 
122 /*
123  * File organization:
124  *  - private structures
125  *  - iflib private utility functions
126  *  - ifnet functions
127  *  - vlan registry and other exported functions
128  *  - iflib public core functions
129  *
130  *
131  */
132 MALLOC_DEFINE(M_IFLIB, "iflib", "ifnet library");
133 
134 struct iflib_txq;
135 typedef struct iflib_txq *iflib_txq_t;
136 struct iflib_rxq;
137 typedef struct iflib_rxq *iflib_rxq_t;
138 struct iflib_fl;
139 typedef struct iflib_fl *iflib_fl_t;
140 
141 struct iflib_ctx;
142 
143 static void iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid);
144 static void iflib_timer(void *arg);
145 
146 typedef struct iflib_filter_info {
147 	driver_filter_t *ifi_filter;
148 	void *ifi_filter_arg;
149 	struct grouptask *ifi_task;
150 	void *ifi_ctx;
151 } *iflib_filter_info_t;
152 
153 struct iflib_ctx {
154 	KOBJ_FIELDS;
155 	/*
156 	 * Pointer to hardware driver's softc
157 	 */
158 	void *ifc_softc;
159 	device_t ifc_dev;
160 	if_t ifc_ifp;
161 
162 	cpuset_t ifc_cpus;
163 	if_shared_ctx_t ifc_sctx;
164 	struct if_softc_ctx ifc_softc_ctx;
165 
166 	struct sx ifc_ctx_sx;
167 	struct mtx ifc_state_mtx;
168 
169 	iflib_txq_t ifc_txqs;
170 	iflib_rxq_t ifc_rxqs;
171 	uint32_t ifc_if_flags;
172 	uint32_t ifc_flags;
173 	uint32_t ifc_max_fl_buf_size;
174 
175 	int ifc_link_state;
176 	int ifc_link_irq;
177 	int ifc_watchdog_events;
178 	struct cdev *ifc_led_dev;
179 	struct resource *ifc_msix_mem;
180 
181 	struct if_irq ifc_legacy_irq;
182 	struct grouptask ifc_admin_task;
183 	struct grouptask ifc_vflr_task;
184 	struct iflib_filter_info ifc_filter_info;
185 	struct ifmedia	ifc_media;
186 
187 	struct sysctl_oid *ifc_sysctl_node;
188 	uint16_t ifc_sysctl_ntxqs;
189 	uint16_t ifc_sysctl_nrxqs;
190 	uint16_t ifc_sysctl_qs_eq_override;
191 	uint16_t ifc_sysctl_rx_budget;
192 	uint16_t ifc_sysctl_tx_abdicate;
193 
194 	qidx_t ifc_sysctl_ntxds[8];
195 	qidx_t ifc_sysctl_nrxds[8];
196 	struct if_txrx ifc_txrx;
197 #define isc_txd_encap  ifc_txrx.ift_txd_encap
198 #define isc_txd_flush  ifc_txrx.ift_txd_flush
199 #define isc_txd_credits_update  ifc_txrx.ift_txd_credits_update
200 #define isc_rxd_available ifc_txrx.ift_rxd_available
201 #define isc_rxd_pkt_get ifc_txrx.ift_rxd_pkt_get
202 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
203 #define isc_rxd_flush ifc_txrx.ift_rxd_flush
204 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
205 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
206 #define isc_legacy_intr ifc_txrx.ift_legacy_intr
207 	eventhandler_tag ifc_vlan_attach_event;
208 	eventhandler_tag ifc_vlan_detach_event;
209 	uint8_t ifc_mac[ETHER_ADDR_LEN];
210 	char ifc_mtx_name[16];
211 };
212 
213 
214 void *
215 iflib_get_softc(if_ctx_t ctx)
216 {
217 
218 	return (ctx->ifc_softc);
219 }
220 
221 device_t
222 iflib_get_dev(if_ctx_t ctx)
223 {
224 
225 	return (ctx->ifc_dev);
226 }
227 
228 if_t
229 iflib_get_ifp(if_ctx_t ctx)
230 {
231 
232 	return (ctx->ifc_ifp);
233 }
234 
235 struct ifmedia *
236 iflib_get_media(if_ctx_t ctx)
237 {
238 
239 	return (&ctx->ifc_media);
240 }
241 
242 uint32_t
243 iflib_get_flags(if_ctx_t ctx)
244 {
245 	return (ctx->ifc_flags);
246 }
247 
248 void
249 iflib_set_mac(if_ctx_t ctx, uint8_t mac[ETHER_ADDR_LEN])
250 {
251 
252 	bcopy(mac, ctx->ifc_mac, ETHER_ADDR_LEN);
253 }
254 
255 if_softc_ctx_t
256 iflib_get_softc_ctx(if_ctx_t ctx)
257 {
258 
259 	return (&ctx->ifc_softc_ctx);
260 }
261 
262 if_shared_ctx_t
263 iflib_get_sctx(if_ctx_t ctx)
264 {
265 
266 	return (ctx->ifc_sctx);
267 }
268 
269 #define IP_ALIGNED(m) ((((uintptr_t)(m)->m_data) & 0x3) == 0x2)
270 #define CACHE_PTR_INCREMENT (CACHE_LINE_SIZE/sizeof(void*))
271 #define CACHE_PTR_NEXT(ptr) ((void *)(((uintptr_t)(ptr)+CACHE_LINE_SIZE-1) & (CACHE_LINE_SIZE-1)))
272 
273 #define LINK_ACTIVE(ctx) ((ctx)->ifc_link_state == LINK_STATE_UP)
274 #define CTX_IS_VF(ctx) ((ctx)->ifc_sctx->isc_flags & IFLIB_IS_VF)
275 
276 typedef struct iflib_sw_rx_desc_array {
277 	bus_dmamap_t	*ifsd_map;         /* bus_dma maps for packet */
278 	struct mbuf	**ifsd_m;           /* pkthdr mbufs */
279 	caddr_t		*ifsd_cl;          /* direct cluster pointer for rx */
280 	bus_addr_t	*ifsd_ba;          /* bus addr of cluster for rx */
281 } iflib_rxsd_array_t;
282 
283 typedef struct iflib_sw_tx_desc_array {
284 	bus_dmamap_t    *ifsd_map;         /* bus_dma maps for packet */
285 	bus_dmamap_t	*ifsd_tso_map;     /* bus_dma maps for TSO packet */
286 	struct mbuf    **ifsd_m;           /* pkthdr mbufs */
287 } if_txsd_vec_t;
288 
289 
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 
309 #define CSUM_OFFLOAD		(CSUM_IP_TSO|CSUM_IP6_TSO|CSUM_IP| \
310 				 CSUM_IP_UDP|CSUM_IP_TCP|CSUM_IP_SCTP| \
311 				 CSUM_IP6_UDP|CSUM_IP6_TCP|CSUM_IP6_SCTP)
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 16
360 	char                    ift_mtx_name[MTX_NAME_LEN];
361 	char                    ift_db_mtx_name[MTX_NAME_LEN];
362 	bus_dma_segment_t	ift_segs[IFLIB_MAX_TX_SEGS]  __aligned(CACHE_LINE_SIZE);
363 #ifdef IFLIB_DIAGNOSTICS
364 	uint64_t ift_cpu_exec_count[256];
365 #endif
366 } __aligned(CACHE_LINE_SIZE);
367 
368 struct iflib_fl {
369 	qidx_t		ifl_cidx;
370 	qidx_t		ifl_pidx;
371 	qidx_t		ifl_credits;
372 	uint8_t		ifl_gen;
373 	uint8_t		ifl_rxd_size;
374 #if MEMORY_LOGGING
375 	uint64_t	ifl_m_enqueued;
376 	uint64_t	ifl_m_dequeued;
377 	uint64_t	ifl_cl_enqueued;
378 	uint64_t	ifl_cl_dequeued;
379 #endif
380 	/* implicit pad */
381 
382 	bitstr_t 	*ifl_rx_bitmap;
383 	qidx_t		ifl_fragidx;
384 	/* constant */
385 	qidx_t		ifl_size;
386 	uint16_t	ifl_buf_size;
387 	uint16_t	ifl_cltype;
388 	uma_zone_t	ifl_zone;
389 	iflib_rxsd_array_t	ifl_sds;
390 	iflib_rxq_t	ifl_rxq;
391 	uint8_t		ifl_id;
392 	bus_dma_tag_t	ifl_buf_tag;
393 	iflib_dma_info_t	ifl_ifdi;
394 	uint64_t	ifl_bus_addrs[IFLIB_MAX_RX_REFRESH] __aligned(CACHE_LINE_SIZE);
395 	caddr_t		ifl_vm_addrs[IFLIB_MAX_RX_REFRESH];
396 	qidx_t	ifl_rxd_idxs[IFLIB_MAX_RX_REFRESH];
397 }  __aligned(CACHE_LINE_SIZE);
398 
399 static inline qidx_t
400 get_inuse(int size, qidx_t cidx, qidx_t pidx, uint8_t gen)
401 {
402 	qidx_t used;
403 
404 	if (pidx > cidx)
405 		used = pidx - cidx;
406 	else if (pidx < cidx)
407 		used = size - cidx + pidx;
408 	else if (gen == 0 && pidx == cidx)
409 		used = 0;
410 	else if (gen == 1 && pidx == cidx)
411 		used = size;
412 	else
413 		panic("bad state");
414 
415 	return (used);
416 }
417 
418 #define TXQ_AVAIL(txq) (txq->ift_size - get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx, txq->ift_gen))
419 
420 #define IDXDIFF(head, tail, wrap) \
421 	((head) >= (tail) ? (head) - (tail) : (wrap) - (tail) + (head))
422 
423 struct iflib_rxq {
424 	/* If there is a separate completion queue -
425 	 * these are the cq cidx and pidx. Otherwise
426 	 * these are unused.
427 	 */
428 	qidx_t		ifr_size;
429 	qidx_t		ifr_cq_cidx;
430 	qidx_t		ifr_cq_pidx;
431 	uint8_t		ifr_cq_gen;
432 	uint8_t		ifr_fl_offset;
433 
434 	if_ctx_t	ifr_ctx;
435 	iflib_fl_t	ifr_fl;
436 	uint64_t	ifr_rx_irq;
437 	uint16_t	ifr_id;
438 	uint8_t		ifr_lro_enabled;
439 	uint8_t		ifr_nfl;
440 	uint8_t		ifr_ntxqirq;
441 	uint8_t		ifr_txqid[IFLIB_MAX_TX_SHARED_INTR];
442 	struct lro_ctrl			ifr_lc;
443 	struct grouptask        ifr_task;
444 	struct iflib_filter_info ifr_filter_info;
445 	iflib_dma_info_t		ifr_ifdi;
446 
447 	/* dynamically allocate if any drivers need a value substantially larger than this */
448 	struct if_rxd_frag	ifr_frags[IFLIB_MAX_RX_SEGS] __aligned(CACHE_LINE_SIZE);
449 #ifdef IFLIB_DIAGNOSTICS
450 	uint64_t ifr_cpu_exec_count[256];
451 #endif
452 }  __aligned(CACHE_LINE_SIZE);
453 
454 typedef struct if_rxsd {
455 	caddr_t *ifsd_cl;
456 	struct mbuf **ifsd_m;
457 	iflib_fl_t ifsd_fl;
458 	qidx_t ifsd_cidx;
459 } *if_rxsd_t;
460 
461 /* multiple of word size */
462 #ifdef __LP64__
463 #define PKT_INFO_SIZE	6
464 #define RXD_INFO_SIZE	5
465 #define PKT_TYPE uint64_t
466 #else
467 #define PKT_INFO_SIZE	11
468 #define RXD_INFO_SIZE	8
469 #define PKT_TYPE uint32_t
470 #endif
471 #define PKT_LOOP_BOUND  ((PKT_INFO_SIZE/3)*3)
472 #define RXD_LOOP_BOUND  ((RXD_INFO_SIZE/4)*4)
473 
474 typedef struct if_pkt_info_pad {
475 	PKT_TYPE pkt_val[PKT_INFO_SIZE];
476 } *if_pkt_info_pad_t;
477 typedef struct if_rxd_info_pad {
478 	PKT_TYPE rxd_val[RXD_INFO_SIZE];
479 } *if_rxd_info_pad_t;
480 
481 CTASSERT(sizeof(struct if_pkt_info_pad) == sizeof(struct if_pkt_info));
482 CTASSERT(sizeof(struct if_rxd_info_pad) == sizeof(struct if_rxd_info));
483 
484 
485 static inline void
486 pkt_info_zero(if_pkt_info_t pi)
487 {
488 	if_pkt_info_pad_t pi_pad;
489 
490 	pi_pad = (if_pkt_info_pad_t)pi;
491 	pi_pad->pkt_val[0] = 0; pi_pad->pkt_val[1] = 0; pi_pad->pkt_val[2] = 0;
492 	pi_pad->pkt_val[3] = 0; pi_pad->pkt_val[4] = 0; pi_pad->pkt_val[5] = 0;
493 #ifndef __LP64__
494 	pi_pad->pkt_val[6] = 0; pi_pad->pkt_val[7] = 0; pi_pad->pkt_val[8] = 0;
495 	pi_pad->pkt_val[9] = 0; pi_pad->pkt_val[10] = 0;
496 #endif
497 }
498 
499 static device_method_t iflib_pseudo_methods[] = {
500 	DEVMETHOD(device_attach, noop_attach),
501 	DEVMETHOD(device_detach, iflib_pseudo_detach),
502 	DEVMETHOD_END
503 };
504 
505 driver_t iflib_pseudodriver = {
506 	"iflib_pseudo", iflib_pseudo_methods, sizeof(struct iflib_ctx),
507 };
508 
509 static inline void
510 rxd_info_zero(if_rxd_info_t ri)
511 {
512 	if_rxd_info_pad_t ri_pad;
513 	int i;
514 
515 	ri_pad = (if_rxd_info_pad_t)ri;
516 	for (i = 0; i < RXD_LOOP_BOUND; i += 4) {
517 		ri_pad->rxd_val[i] = 0;
518 		ri_pad->rxd_val[i+1] = 0;
519 		ri_pad->rxd_val[i+2] = 0;
520 		ri_pad->rxd_val[i+3] = 0;
521 	}
522 #ifdef __LP64__
523 	ri_pad->rxd_val[RXD_INFO_SIZE-1] = 0;
524 #endif
525 }
526 
527 /*
528  * Only allow a single packet to take up most 1/nth of the tx ring
529  */
530 #define MAX_SINGLE_PACKET_FRACTION 12
531 #define IF_BAD_DMA (bus_addr_t)-1
532 
533 #define CTX_ACTIVE(ctx) ((if_getdrvflags((ctx)->ifc_ifp) & IFF_DRV_RUNNING))
534 
535 #define CTX_LOCK_INIT(_sc)  sx_init(&(_sc)->ifc_ctx_sx, "iflib ctx lock")
536 #define CTX_LOCK(ctx) sx_xlock(&(ctx)->ifc_ctx_sx)
537 #define CTX_UNLOCK(ctx) sx_xunlock(&(ctx)->ifc_ctx_sx)
538 #define CTX_LOCK_DESTROY(ctx) sx_destroy(&(ctx)->ifc_ctx_sx)
539 
540 
541 #define STATE_LOCK_INIT(_sc, _name)  mtx_init(&(_sc)->ifc_state_mtx, _name, "iflib state lock", MTX_DEF)
542 #define STATE_LOCK(ctx) mtx_lock(&(ctx)->ifc_state_mtx)
543 #define STATE_UNLOCK(ctx) mtx_unlock(&(ctx)->ifc_state_mtx)
544 #define STATE_LOCK_DESTROY(ctx) mtx_destroy(&(ctx)->ifc_state_mtx)
545 
546 
547 
548 #define CALLOUT_LOCK(txq)	mtx_lock(&txq->ift_mtx)
549 #define CALLOUT_UNLOCK(txq) 	mtx_unlock(&txq->ift_mtx)
550 
551 void
552 iflib_set_detach(if_ctx_t ctx)
553 {
554 	STATE_LOCK(ctx);
555 	ctx->ifc_flags |= IFC_IN_DETACH;
556 	STATE_UNLOCK(ctx);
557 }
558 
559 /* Our boot-time initialization hook */
560 static int	iflib_module_event_handler(module_t, int, void *);
561 
562 static moduledata_t iflib_moduledata = {
563 	"iflib",
564 	iflib_module_event_handler,
565 	NULL
566 };
567 
568 DECLARE_MODULE(iflib, iflib_moduledata, SI_SUB_INIT_IF, SI_ORDER_ANY);
569 MODULE_VERSION(iflib, 1);
570 
571 MODULE_DEPEND(iflib, pci, 1, 1, 1);
572 MODULE_DEPEND(iflib, ether, 1, 1, 1);
573 
574 TASKQGROUP_DEFINE(if_io_tqg, mp_ncpus, 1);
575 TASKQGROUP_DEFINE(if_config_tqg, 1, 1);
576 
577 #ifndef IFLIB_DEBUG_COUNTERS
578 #ifdef INVARIANTS
579 #define IFLIB_DEBUG_COUNTERS 1
580 #else
581 #define IFLIB_DEBUG_COUNTERS 0
582 #endif /* !INVARIANTS */
583 #endif
584 
585 static SYSCTL_NODE(_net, OID_AUTO, iflib, CTLFLAG_RD, 0,
586                    "iflib driver parameters");
587 
588 /*
589  * XXX need to ensure that this can't accidentally cause the head to be moved backwards
590  */
591 static int iflib_min_tx_latency = 0;
592 SYSCTL_INT(_net_iflib, OID_AUTO, min_tx_latency, CTLFLAG_RW,
593 		   &iflib_min_tx_latency, 0, "minimize transmit latency at the possible expense of throughput");
594 static int iflib_no_tx_batch = 0;
595 SYSCTL_INT(_net_iflib, OID_AUTO, no_tx_batch, CTLFLAG_RW,
596 		   &iflib_no_tx_batch, 0, "minimize transmit latency at the possible expense of throughput");
597 
598 
599 #if IFLIB_DEBUG_COUNTERS
600 
601 static int iflib_tx_seen;
602 static int iflib_tx_sent;
603 static int iflib_tx_encap;
604 static int iflib_rx_allocs;
605 static int iflib_fl_refills;
606 static int iflib_fl_refills_large;
607 static int iflib_tx_frees;
608 
609 SYSCTL_INT(_net_iflib, OID_AUTO, tx_seen, CTLFLAG_RD,
610 		   &iflib_tx_seen, 0, "# tx mbufs seen");
611 SYSCTL_INT(_net_iflib, OID_AUTO, tx_sent, CTLFLAG_RD,
612 		   &iflib_tx_sent, 0, "# tx mbufs sent");
613 SYSCTL_INT(_net_iflib, OID_AUTO, tx_encap, CTLFLAG_RD,
614 		   &iflib_tx_encap, 0, "# tx mbufs encapped");
615 SYSCTL_INT(_net_iflib, OID_AUTO, tx_frees, CTLFLAG_RD,
616 		   &iflib_tx_frees, 0, "# tx frees");
617 SYSCTL_INT(_net_iflib, OID_AUTO, rx_allocs, CTLFLAG_RD,
618 		   &iflib_rx_allocs, 0, "# rx allocations");
619 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills, CTLFLAG_RD,
620 		   &iflib_fl_refills, 0, "# refills");
621 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills_large, CTLFLAG_RD,
622 		   &iflib_fl_refills_large, 0, "# large refills");
623 
624 
625 static int iflib_txq_drain_flushing;
626 static int iflib_txq_drain_oactive;
627 static int iflib_txq_drain_notready;
628 
629 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_flushing, CTLFLAG_RD,
630 		   &iflib_txq_drain_flushing, 0, "# drain flushes");
631 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_oactive, CTLFLAG_RD,
632 		   &iflib_txq_drain_oactive, 0, "# drain oactives");
633 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_notready, CTLFLAG_RD,
634 		   &iflib_txq_drain_notready, 0, "# drain notready");
635 
636 
637 static int iflib_encap_load_mbuf_fail;
638 static int iflib_encap_pad_mbuf_fail;
639 static int iflib_encap_txq_avail_fail;
640 static int iflib_encap_txd_encap_fail;
641 
642 SYSCTL_INT(_net_iflib, OID_AUTO, encap_load_mbuf_fail, CTLFLAG_RD,
643 		   &iflib_encap_load_mbuf_fail, 0, "# busdma load failures");
644 SYSCTL_INT(_net_iflib, OID_AUTO, encap_pad_mbuf_fail, CTLFLAG_RD,
645 		   &iflib_encap_pad_mbuf_fail, 0, "# runt frame pad failures");
646 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txq_avail_fail, CTLFLAG_RD,
647 		   &iflib_encap_txq_avail_fail, 0, "# txq avail failures");
648 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txd_encap_fail, CTLFLAG_RD,
649 		   &iflib_encap_txd_encap_fail, 0, "# driver encap failures");
650 
651 static int iflib_task_fn_rxs;
652 static int iflib_rx_intr_enables;
653 static int iflib_fast_intrs;
654 static int iflib_rx_unavail;
655 static int iflib_rx_ctx_inactive;
656 static int iflib_rx_if_input;
657 static int iflib_rx_mbuf_null;
658 static int iflib_rxd_flush;
659 
660 static int iflib_verbose_debug;
661 
662 SYSCTL_INT(_net_iflib, OID_AUTO, task_fn_rx, CTLFLAG_RD,
663 		   &iflib_task_fn_rxs, 0, "# task_fn_rx calls");
664 SYSCTL_INT(_net_iflib, OID_AUTO, rx_intr_enables, CTLFLAG_RD,
665 		   &iflib_rx_intr_enables, 0, "# rx intr enables");
666 SYSCTL_INT(_net_iflib, OID_AUTO, fast_intrs, CTLFLAG_RD,
667 		   &iflib_fast_intrs, 0, "# fast_intr calls");
668 SYSCTL_INT(_net_iflib, OID_AUTO, rx_unavail, CTLFLAG_RD,
669 		   &iflib_rx_unavail, 0, "# times rxeof called with no available data");
670 SYSCTL_INT(_net_iflib, OID_AUTO, rx_ctx_inactive, CTLFLAG_RD,
671 		   &iflib_rx_ctx_inactive, 0, "# times rxeof called with inactive context");
672 SYSCTL_INT(_net_iflib, OID_AUTO, rx_if_input, CTLFLAG_RD,
673 		   &iflib_rx_if_input, 0, "# times rxeof called if_input");
674 SYSCTL_INT(_net_iflib, OID_AUTO, rx_mbuf_null, CTLFLAG_RD,
675 		   &iflib_rx_mbuf_null, 0, "# times rxeof got null mbuf");
676 SYSCTL_INT(_net_iflib, OID_AUTO, rxd_flush, CTLFLAG_RD,
677 	         &iflib_rxd_flush, 0, "# times rxd_flush called");
678 SYSCTL_INT(_net_iflib, OID_AUTO, verbose_debug, CTLFLAG_RW,
679 		   &iflib_verbose_debug, 0, "enable verbose debugging");
680 
681 #define DBG_COUNTER_INC(name) atomic_add_int(&(iflib_ ## name), 1)
682 static void
683 iflib_debug_reset(void)
684 {
685 	iflib_tx_seen = iflib_tx_sent = iflib_tx_encap = iflib_rx_allocs =
686 		iflib_fl_refills = iflib_fl_refills_large = iflib_tx_frees =
687 		iflib_txq_drain_flushing = iflib_txq_drain_oactive =
688 		iflib_txq_drain_notready =
689 		iflib_encap_load_mbuf_fail = iflib_encap_pad_mbuf_fail =
690 		iflib_encap_txq_avail_fail = iflib_encap_txd_encap_fail =
691 		iflib_task_fn_rxs = iflib_rx_intr_enables = iflib_fast_intrs =
692 		iflib_rx_unavail =
693 		iflib_rx_ctx_inactive = iflib_rx_if_input =
694 		iflib_rx_mbuf_null = iflib_rxd_flush = 0;
695 }
696 
697 #else
698 #define DBG_COUNTER_INC(name)
699 static void iflib_debug_reset(void) {}
700 #endif
701 
702 #define IFLIB_DEBUG 0
703 
704 static void iflib_tx_structures_free(if_ctx_t ctx);
705 static void iflib_rx_structures_free(if_ctx_t ctx);
706 static int iflib_queues_alloc(if_ctx_t ctx);
707 static int iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq);
708 static int iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget);
709 static int iflib_qset_structures_setup(if_ctx_t ctx);
710 static int iflib_msix_init(if_ctx_t ctx);
711 static int iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filterarg, int *rid, const char *str);
712 static void iflib_txq_check_drain(iflib_txq_t txq, int budget);
713 static uint32_t iflib_txq_can_drain(struct ifmp_ring *);
714 #ifdef ALTQ
715 static void iflib_altq_if_start(if_t ifp);
716 static int iflib_altq_if_transmit(if_t ifp, struct mbuf *m);
717 #endif
718 static int iflib_register(if_ctx_t);
719 static void iflib_init_locked(if_ctx_t ctx);
720 static void iflib_add_device_sysctl_pre(if_ctx_t ctx);
721 static void iflib_add_device_sysctl_post(if_ctx_t ctx);
722 static void iflib_ifmp_purge(iflib_txq_t txq);
723 static void _iflib_pre_assert(if_softc_ctx_t scctx);
724 static void iflib_if_init_locked(if_ctx_t ctx);
725 static void iflib_free_intr_mem(if_ctx_t ctx);
726 #ifndef __NO_STRICT_ALIGNMENT
727 static struct mbuf * iflib_fixup_rx(struct mbuf *m);
728 #endif
729 
730 NETDUMP_DEFINE(iflib);
731 
732 #ifdef DEV_NETMAP
733 #include <sys/selinfo.h>
734 #include <net/netmap.h>
735 #include <dev/netmap/netmap_kern.h>
736 
737 MODULE_DEPEND(iflib, netmap, 1, 1, 1);
738 
739 static int netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, uint32_t nm_i, bool init);
740 
741 /*
742  * device-specific sysctl variables:
743  *
744  * iflib_crcstrip: 0: keep CRC in rx frames (default), 1: strip it.
745  *	During regular operations the CRC is stripped, but on some
746  *	hardware reception of frames not multiple of 64 is slower,
747  *	so using crcstrip=0 helps in benchmarks.
748  *
749  * iflib_rx_miss, iflib_rx_miss_bufs:
750  *	count packets that might be missed due to lost interrupts.
751  */
752 SYSCTL_DECL(_dev_netmap);
753 /*
754  * The xl driver by default strips CRCs and we do not override it.
755  */
756 
757 int iflib_crcstrip = 1;
758 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_crcstrip,
759     CTLFLAG_RW, &iflib_crcstrip, 1, "strip CRC on rx frames");
760 
761 int iflib_rx_miss, iflib_rx_miss_bufs;
762 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss,
763     CTLFLAG_RW, &iflib_rx_miss, 0, "potentially missed rx intr");
764 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss_bufs,
765     CTLFLAG_RW, &iflib_rx_miss_bufs, 0, "potentially missed rx intr bufs");
766 
767 /*
768  * Register/unregister. We are already under netmap lock.
769  * Only called on the first register or the last unregister.
770  */
771 static int
772 iflib_netmap_register(struct netmap_adapter *na, int onoff)
773 {
774 	struct ifnet *ifp = na->ifp;
775 	if_ctx_t ctx = ifp->if_softc;
776 	int status;
777 
778 	CTX_LOCK(ctx);
779 	IFDI_INTR_DISABLE(ctx);
780 
781 	/* Tell the stack that the interface is no longer active */
782 	ifp->if_drv_flags &= ~(IFF_DRV_RUNNING | IFF_DRV_OACTIVE);
783 
784 	if (!CTX_IS_VF(ctx))
785 		IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip);
786 
787 	/* enable or disable flags and callbacks in na and ifp */
788 	if (onoff) {
789 		nm_set_native_flags(na);
790 	} else {
791 		nm_clear_native_flags(na);
792 	}
793 	iflib_stop(ctx);
794 	iflib_init_locked(ctx);
795 	IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip); // XXX why twice ?
796 	status = ifp->if_drv_flags & IFF_DRV_RUNNING ? 0 : 1;
797 	if (status)
798 		nm_clear_native_flags(na);
799 	CTX_UNLOCK(ctx);
800 	return (status);
801 }
802 
803 static int
804 netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, uint32_t nm_i, bool init)
805 {
806 	struct netmap_adapter *na = kring->na;
807 	u_int const lim = kring->nkr_num_slots - 1;
808 	u_int head = kring->rhead;
809 	struct netmap_ring *ring = kring->ring;
810 	bus_dmamap_t *map;
811 	struct if_rxd_update iru;
812 	if_ctx_t ctx = rxq->ifr_ctx;
813 	iflib_fl_t fl = &rxq->ifr_fl[0];
814 	uint32_t refill_pidx, nic_i;
815 #if IFLIB_DEBUG_COUNTERS
816 	int rf_count = 0;
817 #endif
818 
819 	if (nm_i == head && __predict_true(!init))
820 		return 0;
821 	iru_init(&iru, rxq, 0 /* flid */);
822 	map = fl->ifl_sds.ifsd_map;
823 	refill_pidx = netmap_idx_k2n(kring, nm_i);
824 	/*
825 	 * IMPORTANT: we must leave one free slot in the ring,
826 	 * so move head back by one unit
827 	 */
828 	head = nm_prev(head, lim);
829 	nic_i = UINT_MAX;
830 	DBG_COUNTER_INC(fl_refills);
831 	while (nm_i != head) {
832 #if IFLIB_DEBUG_COUNTERS
833 		if (++rf_count == 9)
834 			DBG_COUNTER_INC(fl_refills_large);
835 #endif
836 		for (int tmp_pidx = 0; tmp_pidx < IFLIB_MAX_RX_REFRESH && nm_i != head; tmp_pidx++) {
837 			struct netmap_slot *slot = &ring->slot[nm_i];
838 			void *addr = PNMB(na, slot, &fl->ifl_bus_addrs[tmp_pidx]);
839 			uint32_t nic_i_dma = refill_pidx;
840 			nic_i = netmap_idx_k2n(kring, nm_i);
841 
842 			MPASS(tmp_pidx < IFLIB_MAX_RX_REFRESH);
843 
844 			if (addr == NETMAP_BUF_BASE(na)) /* bad buf */
845 			        return netmap_ring_reinit(kring);
846 
847 			fl->ifl_vm_addrs[tmp_pidx] = addr;
848 			if (__predict_false(init)) {
849 				netmap_load_map(na, fl->ifl_buf_tag,
850 				    map[nic_i], addr);
851 			} else if (slot->flags & NS_BUF_CHANGED) {
852 				/* buffer has changed, reload map */
853 				netmap_reload_map(na, fl->ifl_buf_tag,
854 				    map[nic_i], addr);
855 			}
856 			slot->flags &= ~NS_BUF_CHANGED;
857 
858 			nm_i = nm_next(nm_i, lim);
859 			fl->ifl_rxd_idxs[tmp_pidx] = nic_i = nm_next(nic_i, lim);
860 			if (nm_i != head && tmp_pidx < IFLIB_MAX_RX_REFRESH-1)
861 				continue;
862 
863 			iru.iru_pidx = refill_pidx;
864 			iru.iru_count = tmp_pidx+1;
865 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
866 			refill_pidx = nic_i;
867 			for (int n = 0; n < iru.iru_count; n++) {
868 				bus_dmamap_sync(fl->ifl_buf_tag, map[nic_i_dma],
869 						BUS_DMASYNC_PREREAD);
870 				/* XXX - change this to not use the netmap func*/
871 				nic_i_dma = nm_next(nic_i_dma, lim);
872 			}
873 		}
874 	}
875 	kring->nr_hwcur = head;
876 
877 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
878 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
879 	if (__predict_true(nic_i != UINT_MAX)) {
880 		ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id, nic_i);
881 		DBG_COUNTER_INC(rxd_flush);
882 	}
883 	return (0);
884 }
885 
886 /*
887  * Reconcile kernel and user view of the transmit ring.
888  *
889  * All information is in the kring.
890  * Userspace wants to send packets up to the one before kring->rhead,
891  * kernel knows kring->nr_hwcur is the first unsent packet.
892  *
893  * Here we push packets out (as many as possible), and possibly
894  * reclaim buffers from previously completed transmission.
895  *
896  * The caller (netmap) guarantees that there is only one instance
897  * running at any time. Any interference with other driver
898  * methods should be handled by the individual drivers.
899  */
900 static int
901 iflib_netmap_txsync(struct netmap_kring *kring, int flags)
902 {
903 	struct netmap_adapter *na = kring->na;
904 	struct ifnet *ifp = na->ifp;
905 	struct netmap_ring *ring = kring->ring;
906 	u_int nm_i;	/* index into the netmap kring */
907 	u_int nic_i;	/* index into the NIC ring */
908 	u_int n;
909 	u_int const lim = kring->nkr_num_slots - 1;
910 	u_int const head = kring->rhead;
911 	struct if_pkt_info pi;
912 
913 	/*
914 	 * interrupts on every tx packet are expensive so request
915 	 * them every half ring, or where NS_REPORT is set
916 	 */
917 	u_int report_frequency = kring->nkr_num_slots >> 1;
918 	/* device-specific */
919 	if_ctx_t ctx = ifp->if_softc;
920 	iflib_txq_t txq = &ctx->ifc_txqs[kring->ring_id];
921 
922 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
923 	    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
924 
925 	/*
926 	 * First part: process new packets to send.
927 	 * nm_i is the current index in the netmap kring,
928 	 * nic_i is the corresponding index in the NIC ring.
929 	 *
930 	 * If we have packets to send (nm_i != head)
931 	 * iterate over the netmap ring, fetch length and update
932 	 * the corresponding slot in the NIC ring. Some drivers also
933 	 * need to update the buffer's physical address in the NIC slot
934 	 * even NS_BUF_CHANGED is not set (PNMB computes the addresses).
935 	 *
936 	 * The netmap_reload_map() calls is especially expensive,
937 	 * even when (as in this case) the tag is 0, so do only
938 	 * when the buffer has actually changed.
939 	 *
940 	 * If possible do not set the report/intr bit on all slots,
941 	 * but only a few times per ring or when NS_REPORT is set.
942 	 *
943 	 * Finally, on 10G and faster drivers, it might be useful
944 	 * to prefetch the next slot and txr entry.
945 	 */
946 
947 	nm_i = kring->nr_hwcur;
948 	if (nm_i != head) {	/* we have new packets to send */
949 		pkt_info_zero(&pi);
950 		pi.ipi_segs = txq->ift_segs;
951 		pi.ipi_qsidx = kring->ring_id;
952 		nic_i = netmap_idx_k2n(kring, nm_i);
953 
954 		__builtin_prefetch(&ring->slot[nm_i]);
955 		__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i]);
956 		__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i]);
957 
958 		for (n = 0; nm_i != head; n++) {
959 			struct netmap_slot *slot = &ring->slot[nm_i];
960 			u_int len = slot->len;
961 			uint64_t paddr;
962 			void *addr = PNMB(na, slot, &paddr);
963 			int flags = (slot->flags & NS_REPORT ||
964 				nic_i == 0 || nic_i == report_frequency) ?
965 				IPI_TX_INTR : 0;
966 
967 			/* device-specific */
968 			pi.ipi_len = len;
969 			pi.ipi_segs[0].ds_addr = paddr;
970 			pi.ipi_segs[0].ds_len = len;
971 			pi.ipi_nsegs = 1;
972 			pi.ipi_ndescs = 0;
973 			pi.ipi_pidx = nic_i;
974 			pi.ipi_flags = flags;
975 
976 			/* Fill the slot in the NIC ring. */
977 			ctx->isc_txd_encap(ctx->ifc_softc, &pi);
978 			DBG_COUNTER_INC(tx_encap);
979 
980 			/* prefetch for next round */
981 			__builtin_prefetch(&ring->slot[nm_i + 1]);
982 			__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i + 1]);
983 			__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i + 1]);
984 
985 			NM_CHECK_ADDR_LEN(na, addr, len);
986 
987 			if (slot->flags & NS_BUF_CHANGED) {
988 				/* buffer has changed, reload map */
989 				netmap_reload_map(na, txq->ift_buf_tag,
990 				    txq->ift_sds.ifsd_map[nic_i], addr);
991 			}
992 			/* make sure changes to the buffer are synced */
993 			bus_dmamap_sync(txq->ift_buf_tag,
994 			    txq->ift_sds.ifsd_map[nic_i],
995 			    BUS_DMASYNC_PREWRITE);
996 
997 			slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED);
998 			nm_i = nm_next(nm_i, lim);
999 			nic_i = nm_next(nic_i, lim);
1000 		}
1001 		kring->nr_hwcur = nm_i;
1002 
1003 		/* synchronize the NIC ring */
1004 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1005 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1006 
1007 		/* (re)start the tx unit up to slot nic_i (excluded) */
1008 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, nic_i);
1009 	}
1010 
1011 	/*
1012 	 * Second part: reclaim buffers for completed transmissions.
1013 	 *
1014 	 * If there are unclaimed buffers, attempt to reclaim them.
1015 	 * If none are reclaimed, and TX IRQs are not in use, do an initial
1016 	 * minimal delay, then trigger the tx handler which will spin in the
1017 	 * group task queue.
1018 	 */
1019 	if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1020 		if (iflib_tx_credits_update(ctx, txq)) {
1021 			/* some tx completed, increment avail */
1022 			nic_i = txq->ift_cidx_processed;
1023 			kring->nr_hwtail = nm_prev(netmap_idx_n2k(kring, nic_i), lim);
1024 		}
1025 	}
1026 	if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ))
1027 		if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1028 			callout_reset_on(&txq->ift_timer, hz < 2000 ? 1 : hz / 1000,
1029 			    iflib_timer, txq, txq->ift_timer.c_cpu);
1030 	}
1031 	return (0);
1032 }
1033 
1034 /*
1035  * Reconcile kernel and user view of the receive ring.
1036  * Same as for the txsync, this routine must be efficient.
1037  * The caller guarantees a single invocations, but races against
1038  * the rest of the driver should be handled here.
1039  *
1040  * On call, kring->rhead is the first packet that userspace wants
1041  * to keep, and kring->rcur is the wakeup point.
1042  * The kernel has previously reported packets up to kring->rtail.
1043  *
1044  * If (flags & NAF_FORCE_READ) also check for incoming packets irrespective
1045  * of whether or not we received an interrupt.
1046  */
1047 static int
1048 iflib_netmap_rxsync(struct netmap_kring *kring, int flags)
1049 {
1050 	struct netmap_adapter *na = kring->na;
1051 	struct netmap_ring *ring = kring->ring;
1052 	iflib_fl_t fl;
1053 	uint32_t nm_i;	/* index into the netmap ring */
1054 	uint32_t nic_i;	/* index into the NIC ring */
1055 	u_int i, n;
1056 	u_int const lim = kring->nkr_num_slots - 1;
1057 	u_int const head = kring->rhead;
1058 	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
1059 	struct if_rxd_info ri;
1060 
1061 	struct ifnet *ifp = na->ifp;
1062 	if_ctx_t ctx = ifp->if_softc;
1063 	iflib_rxq_t rxq = &ctx->ifc_rxqs[kring->ring_id];
1064 	if (head > lim)
1065 		return netmap_ring_reinit(kring);
1066 
1067 	/*
1068 	 * XXX netmap_fl_refill() only ever (re)fills free list 0 so far.
1069 	 */
1070 
1071 	for (i = 0, fl = rxq->ifr_fl; i < rxq->ifr_nfl; i++, fl++) {
1072 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
1073 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1074 	}
1075 
1076 	/*
1077 	 * First part: import newly received packets.
1078 	 *
1079 	 * nm_i is the index of the next free slot in the netmap ring,
1080 	 * nic_i is the index of the next received packet in the NIC ring,
1081 	 * and they may differ in case if_init() has been called while
1082 	 * in netmap mode. For the receive ring we have
1083 	 *
1084 	 *	nic_i = rxr->next_check;
1085 	 *	nm_i = kring->nr_hwtail (previous)
1086 	 * and
1087 	 *	nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1088 	 *
1089 	 * rxr->next_check is set to 0 on a ring reinit
1090 	 */
1091 	if (netmap_no_pendintr || force_update) {
1092 		int crclen = iflib_crcstrip ? 0 : 4;
1093 		int error, avail;
1094 
1095 		for (i = 0; i < rxq->ifr_nfl; i++) {
1096 			fl = &rxq->ifr_fl[i];
1097 			nic_i = fl->ifl_cidx;
1098 			nm_i = netmap_idx_n2k(kring, nic_i);
1099 			avail = ctx->isc_rxd_available(ctx->ifc_softc,
1100 			    rxq->ifr_id, nic_i, USHRT_MAX);
1101 			for (n = 0; avail > 0; n++, avail--) {
1102 				rxd_info_zero(&ri);
1103 				ri.iri_frags = rxq->ifr_frags;
1104 				ri.iri_qsidx = kring->ring_id;
1105 				ri.iri_ifp = ctx->ifc_ifp;
1106 				ri.iri_cidx = nic_i;
1107 
1108 				error = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
1109 				ring->slot[nm_i].len = error ? 0 : ri.iri_len - crclen;
1110 				ring->slot[nm_i].flags = 0;
1111 				bus_dmamap_sync(fl->ifl_buf_tag,
1112 				    fl->ifl_sds.ifsd_map[nic_i], BUS_DMASYNC_POSTREAD);
1113 				nm_i = nm_next(nm_i, lim);
1114 				nic_i = nm_next(nic_i, lim);
1115 			}
1116 			if (n) { /* update the state variables */
1117 				if (netmap_no_pendintr && !force_update) {
1118 					/* diagnostics */
1119 					iflib_rx_miss ++;
1120 					iflib_rx_miss_bufs += n;
1121 				}
1122 				fl->ifl_cidx = nic_i;
1123 				kring->nr_hwtail = nm_i;
1124 			}
1125 			kring->nr_kflags &= ~NKR_PENDINTR;
1126 		}
1127 	}
1128 	/*
1129 	 * Second part: skip past packets that userspace has released.
1130 	 * (kring->nr_hwcur to head excluded),
1131 	 * and make the buffers available for reception.
1132 	 * As usual nm_i is the index in the netmap ring,
1133 	 * nic_i is the index in the NIC ring, and
1134 	 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1135 	 */
1136 	/* XXX not sure how this will work with multiple free lists */
1137 	nm_i = kring->nr_hwcur;
1138 
1139 	return (netmap_fl_refill(rxq, kring, nm_i, false));
1140 }
1141 
1142 static void
1143 iflib_netmap_intr(struct netmap_adapter *na, int onoff)
1144 {
1145 	struct ifnet *ifp = na->ifp;
1146 	if_ctx_t ctx = ifp->if_softc;
1147 
1148 	CTX_LOCK(ctx);
1149 	if (onoff) {
1150 		IFDI_INTR_ENABLE(ctx);
1151 	} else {
1152 		IFDI_INTR_DISABLE(ctx);
1153 	}
1154 	CTX_UNLOCK(ctx);
1155 }
1156 
1157 
1158 static int
1159 iflib_netmap_attach(if_ctx_t ctx)
1160 {
1161 	struct netmap_adapter na;
1162 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1163 
1164 	bzero(&na, sizeof(na));
1165 
1166 	na.ifp = ctx->ifc_ifp;
1167 	na.na_flags = NAF_BDG_MAYSLEEP;
1168 	MPASS(ctx->ifc_softc_ctx.isc_ntxqsets);
1169 	MPASS(ctx->ifc_softc_ctx.isc_nrxqsets);
1170 
1171 	na.num_tx_desc = scctx->isc_ntxd[0];
1172 	na.num_rx_desc = scctx->isc_nrxd[0];
1173 	na.nm_txsync = iflib_netmap_txsync;
1174 	na.nm_rxsync = iflib_netmap_rxsync;
1175 	na.nm_register = iflib_netmap_register;
1176 	na.nm_intr = iflib_netmap_intr;
1177 	na.num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
1178 	na.num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
1179 	return (netmap_attach(&na));
1180 }
1181 
1182 static void
1183 iflib_netmap_txq_init(if_ctx_t ctx, iflib_txq_t txq)
1184 {
1185 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1186 	struct netmap_slot *slot;
1187 
1188 	slot = netmap_reset(na, NR_TX, txq->ift_id, 0);
1189 	if (slot == NULL)
1190 		return;
1191 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxd[0]; i++) {
1192 
1193 		/*
1194 		 * In netmap mode, set the map for the packet buffer.
1195 		 * NOTE: Some drivers (not this one) also need to set
1196 		 * the physical buffer address in the NIC ring.
1197 		 * netmap_idx_n2k() maps a nic index, i, into the corresponding
1198 		 * netmap slot index, si
1199 		 */
1200 		int si = netmap_idx_n2k(na->tx_rings[txq->ift_id], i);
1201 		netmap_load_map(na, txq->ift_buf_tag, txq->ift_sds.ifsd_map[i],
1202 		    NMB(na, slot + si));
1203 	}
1204 }
1205 
1206 static void
1207 iflib_netmap_rxq_init(if_ctx_t ctx, iflib_rxq_t rxq)
1208 {
1209 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1210 	struct netmap_kring *kring = na->rx_rings[rxq->ifr_id];
1211 	struct netmap_slot *slot;
1212 	uint32_t nm_i;
1213 
1214 	slot = netmap_reset(na, NR_RX, rxq->ifr_id, 0);
1215 	if (slot == NULL)
1216 		return;
1217 	nm_i = netmap_idx_n2k(kring, 0);
1218 	netmap_fl_refill(rxq, kring, nm_i, true);
1219 }
1220 
1221 static void
1222 iflib_netmap_timer_adjust(if_ctx_t ctx, iflib_txq_t txq, uint32_t *reset_on)
1223 {
1224 	struct netmap_kring *kring;
1225 	uint16_t txqid;
1226 
1227 	txqid = txq->ift_id;
1228 	kring = NA(ctx->ifc_ifp)->tx_rings[txqid];
1229 
1230 	if (kring->nr_hwcur != nm_next(kring->nr_hwtail, kring->nkr_num_slots - 1)) {
1231 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1232 		    BUS_DMASYNC_POSTREAD);
1233 		if (ctx->isc_txd_credits_update(ctx->ifc_softc, txqid, false))
1234 			netmap_tx_irq(ctx->ifc_ifp, txqid);
1235 		if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ)) {
1236 			if (hz < 2000)
1237 				*reset_on = 1;
1238 			else
1239 				*reset_on = hz / 1000;
1240 		}
1241 	}
1242 }
1243 
1244 #define iflib_netmap_detach(ifp) netmap_detach(ifp)
1245 
1246 #else
1247 #define iflib_netmap_txq_init(ctx, txq)
1248 #define iflib_netmap_rxq_init(ctx, rxq)
1249 #define iflib_netmap_detach(ifp)
1250 
1251 #define iflib_netmap_attach(ctx) (0)
1252 #define netmap_rx_irq(ifp, qid, budget) (0)
1253 #define netmap_tx_irq(ifp, qid) do {} while (0)
1254 #define iflib_netmap_timer_adjust(ctx, txq, reset_on)
1255 
1256 #endif
1257 
1258 #if defined(__i386__) || defined(__amd64__)
1259 static __inline void
1260 prefetch(void *x)
1261 {
1262 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1263 }
1264 static __inline void
1265 prefetch2cachelines(void *x)
1266 {
1267 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1268 #if (CACHE_LINE_SIZE < 128)
1269 	__asm volatile("prefetcht0 %0" :: "m" (*(((unsigned long *)x)+CACHE_LINE_SIZE/(sizeof(unsigned long)))));
1270 #endif
1271 }
1272 #else
1273 #define prefetch(x)
1274 #define prefetch2cachelines(x)
1275 #endif
1276 
1277 static void
1278 iflib_gen_mac(if_ctx_t ctx)
1279 {
1280 	struct thread *td;
1281 	MD5_CTX mdctx;
1282 	char uuid[HOSTUUIDLEN+1];
1283 	char buf[HOSTUUIDLEN+16];
1284 	uint8_t *mac;
1285 	unsigned char digest[16];
1286 
1287 	td = curthread;
1288 	mac = ctx->ifc_mac;
1289 	uuid[HOSTUUIDLEN] = 0;
1290 	bcopy(td->td_ucred->cr_prison->pr_hostuuid, uuid, HOSTUUIDLEN);
1291 	snprintf(buf, HOSTUUIDLEN+16, "%s-%s", uuid, device_get_nameunit(ctx->ifc_dev));
1292 	/*
1293 	 * Generate a pseudo-random, deterministic MAC
1294 	 * address based on the UUID and unit number.
1295 	 * The FreeBSD Foundation OUI of 58-9C-FC is used.
1296 	 */
1297 	MD5Init(&mdctx);
1298 	MD5Update(&mdctx, buf, strlen(buf));
1299 	MD5Final(digest, &mdctx);
1300 
1301 	mac[0] = 0x58;
1302 	mac[1] = 0x9C;
1303 	mac[2] = 0xFC;
1304 	mac[3] = digest[0];
1305 	mac[4] = digest[1];
1306 	mac[5] = digest[2];
1307 }
1308 
1309 static void
1310 iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid)
1311 {
1312 	iflib_fl_t fl;
1313 
1314 	fl = &rxq->ifr_fl[flid];
1315 	iru->iru_paddrs = fl->ifl_bus_addrs;
1316 	iru->iru_vaddrs = &fl->ifl_vm_addrs[0];
1317 	iru->iru_idxs = fl->ifl_rxd_idxs;
1318 	iru->iru_qsidx = rxq->ifr_id;
1319 	iru->iru_buf_size = fl->ifl_buf_size;
1320 	iru->iru_flidx = fl->ifl_id;
1321 }
1322 
1323 static void
1324 _iflib_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int err)
1325 {
1326 	if (err)
1327 		return;
1328 	*(bus_addr_t *) arg = segs[0].ds_addr;
1329 }
1330 
1331 int
1332 iflib_dma_alloc_align(if_ctx_t ctx, int size, int align, iflib_dma_info_t dma, int mapflags)
1333 {
1334 	int err;
1335 	device_t dev = ctx->ifc_dev;
1336 
1337 	err = bus_dma_tag_create(bus_get_dma_tag(dev),	/* parent */
1338 				align, 0,		/* alignment, bounds */
1339 				BUS_SPACE_MAXADDR,	/* lowaddr */
1340 				BUS_SPACE_MAXADDR,	/* highaddr */
1341 				NULL, NULL,		/* filter, filterarg */
1342 				size,			/* maxsize */
1343 				1,			/* nsegments */
1344 				size,			/* maxsegsize */
1345 				BUS_DMA_ALLOCNOW,	/* flags */
1346 				NULL,			/* lockfunc */
1347 				NULL,			/* lockarg */
1348 				&dma->idi_tag);
1349 	if (err) {
1350 		device_printf(dev,
1351 		    "%s: bus_dma_tag_create failed: %d\n",
1352 		    __func__, err);
1353 		goto fail_0;
1354 	}
1355 
1356 	err = bus_dmamem_alloc(dma->idi_tag, (void**) &dma->idi_vaddr,
1357 	    BUS_DMA_NOWAIT | BUS_DMA_COHERENT | BUS_DMA_ZERO, &dma->idi_map);
1358 	if (err) {
1359 		device_printf(dev,
1360 		    "%s: bus_dmamem_alloc(%ju) failed: %d\n",
1361 		    __func__, (uintmax_t)size, err);
1362 		goto fail_1;
1363 	}
1364 
1365 	dma->idi_paddr = IF_BAD_DMA;
1366 	err = bus_dmamap_load(dma->idi_tag, dma->idi_map, dma->idi_vaddr,
1367 	    size, _iflib_dmamap_cb, &dma->idi_paddr, mapflags | BUS_DMA_NOWAIT);
1368 	if (err || dma->idi_paddr == IF_BAD_DMA) {
1369 		device_printf(dev,
1370 		    "%s: bus_dmamap_load failed: %d\n",
1371 		    __func__, err);
1372 		goto fail_2;
1373 	}
1374 
1375 	dma->idi_size = size;
1376 	return (0);
1377 
1378 fail_2:
1379 	bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1380 fail_1:
1381 	bus_dma_tag_destroy(dma->idi_tag);
1382 fail_0:
1383 	dma->idi_tag = NULL;
1384 
1385 	return (err);
1386 }
1387 
1388 int
1389 iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags)
1390 {
1391 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1392 
1393 	KASSERT(sctx->isc_q_align != 0, ("alignment value not initialized"));
1394 
1395 	return (iflib_dma_alloc_align(ctx, size, sctx->isc_q_align, dma, mapflags));
1396 }
1397 
1398 int
1399 iflib_dma_alloc_multi(if_ctx_t ctx, int *sizes, iflib_dma_info_t *dmalist, int mapflags, int count)
1400 {
1401 	int i, err;
1402 	iflib_dma_info_t *dmaiter;
1403 
1404 	dmaiter = dmalist;
1405 	for (i = 0; i < count; i++, dmaiter++) {
1406 		if ((err = iflib_dma_alloc(ctx, sizes[i], *dmaiter, mapflags)) != 0)
1407 			break;
1408 	}
1409 	if (err)
1410 		iflib_dma_free_multi(dmalist, i);
1411 	return (err);
1412 }
1413 
1414 void
1415 iflib_dma_free(iflib_dma_info_t dma)
1416 {
1417 	if (dma->idi_tag == NULL)
1418 		return;
1419 	if (dma->idi_paddr != IF_BAD_DMA) {
1420 		bus_dmamap_sync(dma->idi_tag, dma->idi_map,
1421 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1422 		bus_dmamap_unload(dma->idi_tag, dma->idi_map);
1423 		dma->idi_paddr = IF_BAD_DMA;
1424 	}
1425 	if (dma->idi_vaddr != NULL) {
1426 		bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1427 		dma->idi_vaddr = NULL;
1428 	}
1429 	bus_dma_tag_destroy(dma->idi_tag);
1430 	dma->idi_tag = NULL;
1431 }
1432 
1433 void
1434 iflib_dma_free_multi(iflib_dma_info_t *dmalist, int count)
1435 {
1436 	int i;
1437 	iflib_dma_info_t *dmaiter = dmalist;
1438 
1439 	for (i = 0; i < count; i++, dmaiter++)
1440 		iflib_dma_free(*dmaiter);
1441 }
1442 
1443 #ifdef EARLY_AP_STARTUP
1444 static const int iflib_started = 1;
1445 #else
1446 /*
1447  * We used to abuse the smp_started flag to decide if the queues have been
1448  * fully initialized (by late taskqgroup_adjust() calls in a SYSINIT()).
1449  * That gave bad races, since the SYSINIT() runs strictly after smp_started
1450  * is set.  Run a SYSINIT() strictly after that to just set a usable
1451  * completion flag.
1452  */
1453 
1454 static int iflib_started;
1455 
1456 static void
1457 iflib_record_started(void *arg)
1458 {
1459 	iflib_started = 1;
1460 }
1461 
1462 SYSINIT(iflib_record_started, SI_SUB_SMP + 1, SI_ORDER_FIRST,
1463 	iflib_record_started, NULL);
1464 #endif
1465 
1466 static int
1467 iflib_fast_intr(void *arg)
1468 {
1469 	iflib_filter_info_t info = arg;
1470 	struct grouptask *gtask = info->ifi_task;
1471 	if (!iflib_started)
1472 		return (FILTER_HANDLED);
1473 
1474 	DBG_COUNTER_INC(fast_intrs);
1475 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1476 		return (FILTER_HANDLED);
1477 
1478 	GROUPTASK_ENQUEUE(gtask);
1479 	return (FILTER_HANDLED);
1480 }
1481 
1482 static int
1483 iflib_fast_intr_rxtx(void *arg)
1484 {
1485 	iflib_filter_info_t info = arg;
1486 	struct grouptask *gtask = info->ifi_task;
1487 	if_ctx_t ctx;
1488 	iflib_rxq_t rxq = (iflib_rxq_t)info->ifi_ctx;
1489 	iflib_txq_t txq;
1490 	void *sc;
1491 	int i, cidx;
1492 	qidx_t txqid;
1493 
1494 	if (!iflib_started)
1495 		return (FILTER_HANDLED);
1496 
1497 	DBG_COUNTER_INC(fast_intrs);
1498 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1499 		return (FILTER_HANDLED);
1500 
1501 	ctx = rxq->ifr_ctx;
1502 	sc = ctx->ifc_softc;
1503 	MPASS(rxq->ifr_ntxqirq);
1504 	for (i = 0; i < rxq->ifr_ntxqirq; i++) {
1505 		txqid = rxq->ifr_txqid[i];
1506 		txq = &ctx->ifc_txqs[txqid];
1507 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1508 		    BUS_DMASYNC_POSTREAD);
1509 		if (!ctx->isc_txd_credits_update(sc, txqid, false)) {
1510 			IFDI_TX_QUEUE_INTR_ENABLE(ctx, txqid);
1511 			continue;
1512 		}
1513 		GROUPTASK_ENQUEUE(&txq->ift_task);
1514 	}
1515 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_RXCQ)
1516 		cidx = rxq->ifr_cq_cidx;
1517 	else
1518 		cidx = rxq->ifr_fl[0].ifl_cidx;
1519 	if (iflib_rxd_avail(ctx, rxq, cidx, 1))
1520 		GROUPTASK_ENQUEUE(gtask);
1521 	else {
1522 		IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
1523 		DBG_COUNTER_INC(rx_intr_enables);
1524 	}
1525 	return (FILTER_HANDLED);
1526 }
1527 
1528 
1529 static int
1530 iflib_fast_intr_ctx(void *arg)
1531 {
1532 	iflib_filter_info_t info = arg;
1533 	struct grouptask *gtask = info->ifi_task;
1534 
1535 	if (!iflib_started)
1536 		return (FILTER_HANDLED);
1537 
1538 	DBG_COUNTER_INC(fast_intrs);
1539 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1540 		return (FILTER_HANDLED);
1541 
1542 	GROUPTASK_ENQUEUE(gtask);
1543 	return (FILTER_HANDLED);
1544 }
1545 
1546 static int
1547 _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
1548 		 driver_filter_t filter, driver_intr_t handler, void *arg,
1549 		 const char *name)
1550 {
1551 	int rc, flags;
1552 	struct resource *res;
1553 	void *tag = NULL;
1554 	device_t dev = ctx->ifc_dev;
1555 
1556 	flags = RF_ACTIVE;
1557 	if (ctx->ifc_flags & IFC_LEGACY)
1558 		flags |= RF_SHAREABLE;
1559 	MPASS(rid < 512);
1560 	irq->ii_rid = rid;
1561 	res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &irq->ii_rid, flags);
1562 	if (res == NULL) {
1563 		device_printf(dev,
1564 		    "failed to allocate IRQ for rid %d, name %s.\n", rid, name);
1565 		return (ENOMEM);
1566 	}
1567 	irq->ii_res = res;
1568 	KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL"));
1569 	rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET,
1570 						filter, handler, arg, &tag);
1571 	if (rc != 0) {
1572 		device_printf(dev,
1573 		    "failed to setup interrupt for rid %d, name %s: %d\n",
1574 					  rid, name ? name : "unknown", rc);
1575 		return (rc);
1576 	} else if (name)
1577 		bus_describe_intr(dev, res, tag, "%s", name);
1578 
1579 	irq->ii_tag = tag;
1580 	return (0);
1581 }
1582 
1583 
1584 /*********************************************************************
1585  *
1586  *  Allocate DMA resources for TX buffers as well as memory for the TX
1587  *  mbuf map.  TX DMA maps (non-TSO/TSO) and TX mbuf map are kept in a
1588  *  iflib_sw_tx_desc_array structure, storing all the information that
1589  *  is needed to transmit a packet on the wire.  This is called only
1590  *  once at attach, setup is done every reset.
1591  *
1592  **********************************************************************/
1593 static int
1594 iflib_txsd_alloc(iflib_txq_t txq)
1595 {
1596 	if_ctx_t ctx = txq->ift_ctx;
1597 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1598 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1599 	device_t dev = ctx->ifc_dev;
1600 	bus_size_t tsomaxsize;
1601 	int err, nsegments, ntsosegments;
1602 	bool tso;
1603 
1604 	nsegments = scctx->isc_tx_nsegments;
1605 	ntsosegments = scctx->isc_tx_tso_segments_max;
1606 	tsomaxsize = scctx->isc_tx_tso_size_max;
1607 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_VLAN_MTU)
1608 		tsomaxsize += sizeof(struct ether_vlan_header);
1609 	MPASS(scctx->isc_ntxd[0] > 0);
1610 	MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0);
1611 	MPASS(nsegments > 0);
1612 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) {
1613 		MPASS(ntsosegments > 0);
1614 		MPASS(sctx->isc_tso_maxsize >= tsomaxsize);
1615 	}
1616 
1617 	/*
1618 	 * Set up DMA tags for TX buffers.
1619 	 */
1620 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1621 			       1, 0,			/* alignment, bounds */
1622 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1623 			       BUS_SPACE_MAXADDR,	/* highaddr */
1624 			       NULL, NULL,		/* filter, filterarg */
1625 			       sctx->isc_tx_maxsize,		/* maxsize */
1626 			       nsegments,	/* nsegments */
1627 			       sctx->isc_tx_maxsegsize,	/* maxsegsize */
1628 			       0,			/* flags */
1629 			       NULL,			/* lockfunc */
1630 			       NULL,			/* lockfuncarg */
1631 			       &txq->ift_buf_tag))) {
1632 		device_printf(dev,"Unable to allocate TX DMA tag: %d\n", err);
1633 		device_printf(dev,"maxsize: %ju nsegments: %d maxsegsize: %ju\n",
1634 		    (uintmax_t)sctx->isc_tx_maxsize, nsegments, (uintmax_t)sctx->isc_tx_maxsegsize);
1635 		goto fail;
1636 	}
1637 	tso = (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) != 0;
1638 	if (tso && (err = bus_dma_tag_create(bus_get_dma_tag(dev),
1639 			       1, 0,			/* alignment, bounds */
1640 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1641 			       BUS_SPACE_MAXADDR,	/* highaddr */
1642 			       NULL, NULL,		/* filter, filterarg */
1643 			       tsomaxsize,		/* maxsize */
1644 			       ntsosegments,	/* nsegments */
1645 			       sctx->isc_tso_maxsegsize,/* maxsegsize */
1646 			       0,			/* flags */
1647 			       NULL,			/* lockfunc */
1648 			       NULL,			/* lockfuncarg */
1649 			       &txq->ift_tso_buf_tag))) {
1650 		device_printf(dev, "Unable to allocate TSO TX DMA tag: %d\n",
1651 		    err);
1652 		goto fail;
1653 	}
1654 
1655 	/* Allocate memory for the TX mbuf map. */
1656 	if (!(txq->ift_sds.ifsd_m =
1657 	    (struct mbuf **) malloc(sizeof(struct mbuf *) *
1658 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1659 		device_printf(dev, "Unable to allocate TX mbuf map memory\n");
1660 		err = ENOMEM;
1661 		goto fail;
1662 	}
1663 
1664 	/*
1665 	 * Create the DMA maps for TX buffers.
1666 	 */
1667 	if ((txq->ift_sds.ifsd_map = (bus_dmamap_t *)malloc(
1668 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1669 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1670 		device_printf(dev,
1671 		    "Unable to allocate TX buffer DMA map memory\n");
1672 		err = ENOMEM;
1673 		goto fail;
1674 	}
1675 	if (tso && (txq->ift_sds.ifsd_tso_map = (bus_dmamap_t *)malloc(
1676 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1677 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1678 		device_printf(dev,
1679 		    "Unable to allocate TSO TX buffer map memory\n");
1680 		err = ENOMEM;
1681 		goto fail;
1682 	}
1683 	for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) {
1684 		err = bus_dmamap_create(txq->ift_buf_tag, 0,
1685 		    &txq->ift_sds.ifsd_map[i]);
1686 		if (err != 0) {
1687 			device_printf(dev, "Unable to create TX DMA map\n");
1688 			goto fail;
1689 		}
1690 		if (!tso)
1691 			continue;
1692 		err = bus_dmamap_create(txq->ift_tso_buf_tag, 0,
1693 		    &txq->ift_sds.ifsd_tso_map[i]);
1694 		if (err != 0) {
1695 			device_printf(dev, "Unable to create TSO TX DMA map\n");
1696 			goto fail;
1697 		}
1698 	}
1699 	return (0);
1700 fail:
1701 	/* We free all, it handles case where we are in the middle */
1702 	iflib_tx_structures_free(ctx);
1703 	return (err);
1704 }
1705 
1706 static void
1707 iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i)
1708 {
1709 	bus_dmamap_t map;
1710 
1711 	map = NULL;
1712 	if (txq->ift_sds.ifsd_map != NULL)
1713 		map = txq->ift_sds.ifsd_map[i];
1714 	if (map != NULL) {
1715 		bus_dmamap_sync(txq->ift_buf_tag, map, BUS_DMASYNC_POSTWRITE);
1716 		bus_dmamap_unload(txq->ift_buf_tag, map);
1717 		bus_dmamap_destroy(txq->ift_buf_tag, map);
1718 		txq->ift_sds.ifsd_map[i] = NULL;
1719 	}
1720 
1721 	map = NULL;
1722 	if (txq->ift_sds.ifsd_tso_map != NULL)
1723 		map = txq->ift_sds.ifsd_tso_map[i];
1724 	if (map != NULL) {
1725 		bus_dmamap_sync(txq->ift_tso_buf_tag, map,
1726 		    BUS_DMASYNC_POSTWRITE);
1727 		bus_dmamap_unload(txq->ift_tso_buf_tag, map);
1728 		bus_dmamap_destroy(txq->ift_tso_buf_tag, map);
1729 		txq->ift_sds.ifsd_tso_map[i] = NULL;
1730 	}
1731 }
1732 
1733 static void
1734 iflib_txq_destroy(iflib_txq_t txq)
1735 {
1736 	if_ctx_t ctx = txq->ift_ctx;
1737 
1738 	for (int i = 0; i < txq->ift_size; i++)
1739 		iflib_txsd_destroy(ctx, txq, i);
1740 	if (txq->ift_sds.ifsd_map != NULL) {
1741 		free(txq->ift_sds.ifsd_map, M_IFLIB);
1742 		txq->ift_sds.ifsd_map = NULL;
1743 	}
1744 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1745 		free(txq->ift_sds.ifsd_tso_map, M_IFLIB);
1746 		txq->ift_sds.ifsd_tso_map = NULL;
1747 	}
1748 	if (txq->ift_sds.ifsd_m != NULL) {
1749 		free(txq->ift_sds.ifsd_m, M_IFLIB);
1750 		txq->ift_sds.ifsd_m = NULL;
1751 	}
1752 	if (txq->ift_buf_tag != NULL) {
1753 		bus_dma_tag_destroy(txq->ift_buf_tag);
1754 		txq->ift_buf_tag = NULL;
1755 	}
1756 	if (txq->ift_tso_buf_tag != NULL) {
1757 		bus_dma_tag_destroy(txq->ift_tso_buf_tag);
1758 		txq->ift_tso_buf_tag = NULL;
1759 	}
1760 }
1761 
1762 static void
1763 iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i)
1764 {
1765 	struct mbuf **mp;
1766 
1767 	mp = &txq->ift_sds.ifsd_m[i];
1768 	if (*mp == NULL)
1769 		return;
1770 
1771 	if (txq->ift_sds.ifsd_map != NULL) {
1772 		bus_dmamap_sync(txq->ift_buf_tag,
1773 		    txq->ift_sds.ifsd_map[i], BUS_DMASYNC_POSTWRITE);
1774 		bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[i]);
1775 	}
1776 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1777 		bus_dmamap_sync(txq->ift_tso_buf_tag,
1778 		    txq->ift_sds.ifsd_tso_map[i], BUS_DMASYNC_POSTWRITE);
1779 		bus_dmamap_unload(txq->ift_tso_buf_tag,
1780 		    txq->ift_sds.ifsd_tso_map[i]);
1781 	}
1782 	m_free(*mp);
1783 	DBG_COUNTER_INC(tx_frees);
1784 	*mp = NULL;
1785 }
1786 
1787 static int
1788 iflib_txq_setup(iflib_txq_t txq)
1789 {
1790 	if_ctx_t ctx = txq->ift_ctx;
1791 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1792 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1793 	iflib_dma_info_t di;
1794 	int i;
1795 
1796 	/* Set number of descriptors available */
1797 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
1798 	/* XXX make configurable */
1799 	txq->ift_update_freq = IFLIB_DEFAULT_TX_UPDATE_FREQ;
1800 
1801 	/* Reset indices */
1802 	txq->ift_cidx_processed = 0;
1803 	txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0;
1804 	txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset];
1805 
1806 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1807 		bzero((void *)di->idi_vaddr, di->idi_size);
1808 
1809 	IFDI_TXQ_SETUP(ctx, txq->ift_id);
1810 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1811 		bus_dmamap_sync(di->idi_tag, di->idi_map,
1812 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1813 	return (0);
1814 }
1815 
1816 /*********************************************************************
1817  *
1818  *  Allocate DMA resources for RX buffers as well as memory for the RX
1819  *  mbuf map, direct RX cluster pointer map and RX cluster bus address
1820  *  map.  RX DMA map, RX mbuf map, direct RX cluster pointer map and
1821  *  RX cluster map are kept in a iflib_sw_rx_desc_array structure.
1822  *  Since we use use one entry in iflib_sw_rx_desc_array per received
1823  *  packet, the maximum number of entries we'll need is equal to the
1824  *  number of hardware receive descriptors that we've allocated.
1825  *
1826  **********************************************************************/
1827 static int
1828 iflib_rxsd_alloc(iflib_rxq_t rxq)
1829 {
1830 	if_ctx_t ctx = rxq->ifr_ctx;
1831 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1832 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1833 	device_t dev = ctx->ifc_dev;
1834 	iflib_fl_t fl;
1835 	int			err;
1836 
1837 	MPASS(scctx->isc_nrxd[0] > 0);
1838 	MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0);
1839 
1840 	fl = rxq->ifr_fl;
1841 	for (int i = 0; i <  rxq->ifr_nfl; i++, fl++) {
1842 		fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */
1843 		/* Set up DMA tag for RX buffers. */
1844 		err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1845 					 1, 0,			/* alignment, bounds */
1846 					 BUS_SPACE_MAXADDR,	/* lowaddr */
1847 					 BUS_SPACE_MAXADDR,	/* highaddr */
1848 					 NULL, NULL,		/* filter, filterarg */
1849 					 sctx->isc_rx_maxsize,	/* maxsize */
1850 					 sctx->isc_rx_nsegments,	/* nsegments */
1851 					 sctx->isc_rx_maxsegsize,	/* maxsegsize */
1852 					 0,			/* flags */
1853 					 NULL,			/* lockfunc */
1854 					 NULL,			/* lockarg */
1855 					 &fl->ifl_buf_tag);
1856 		if (err) {
1857 			device_printf(dev,
1858 			    "Unable to allocate RX DMA tag: %d\n", err);
1859 			goto fail;
1860 		}
1861 
1862 		/* Allocate memory for the RX mbuf map. */
1863 		if (!(fl->ifl_sds.ifsd_m =
1864 		      (struct mbuf **) malloc(sizeof(struct mbuf *) *
1865 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1866 			device_printf(dev,
1867 			    "Unable to allocate RX mbuf map memory\n");
1868 			err = ENOMEM;
1869 			goto fail;
1870 		}
1871 
1872 		/* Allocate memory for the direct RX cluster pointer map. */
1873 		if (!(fl->ifl_sds.ifsd_cl =
1874 		      (caddr_t *) malloc(sizeof(caddr_t) *
1875 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1876 			device_printf(dev,
1877 			    "Unable to allocate RX cluster map memory\n");
1878 			err = ENOMEM;
1879 			goto fail;
1880 		}
1881 
1882 		/* Allocate memory for the RX cluster bus address map. */
1883 		if (!(fl->ifl_sds.ifsd_ba =
1884 		      (bus_addr_t *) malloc(sizeof(bus_addr_t) *
1885 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1886 			device_printf(dev,
1887 			    "Unable to allocate RX bus address map memory\n");
1888 			err = ENOMEM;
1889 			goto fail;
1890 		}
1891 
1892 		/*
1893 		 * Create the DMA maps for RX buffers.
1894 		 */
1895 		if (!(fl->ifl_sds.ifsd_map =
1896 		      (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1897 			device_printf(dev,
1898 			    "Unable to allocate RX buffer DMA map memory\n");
1899 			err = ENOMEM;
1900 			goto fail;
1901 		}
1902 		for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++) {
1903 			err = bus_dmamap_create(fl->ifl_buf_tag, 0,
1904 			    &fl->ifl_sds.ifsd_map[i]);
1905 			if (err != 0) {
1906 				device_printf(dev, "Unable to create RX buffer DMA map\n");
1907 				goto fail;
1908 			}
1909 		}
1910 	}
1911 	return (0);
1912 
1913 fail:
1914 	iflib_rx_structures_free(ctx);
1915 	return (err);
1916 }
1917 
1918 
1919 /*
1920  * Internal service routines
1921  */
1922 
1923 struct rxq_refill_cb_arg {
1924 	int               error;
1925 	bus_dma_segment_t seg;
1926 	int               nseg;
1927 };
1928 
1929 static void
1930 _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
1931 {
1932 	struct rxq_refill_cb_arg *cb_arg = arg;
1933 
1934 	cb_arg->error = error;
1935 	cb_arg->seg = segs[0];
1936 	cb_arg->nseg = nseg;
1937 }
1938 
1939 /**
1940  *	rxq_refill - refill an rxq  free-buffer list
1941  *	@ctx: the iflib context
1942  *	@rxq: the free-list to refill
1943  *	@n: the number of new buffers to allocate
1944  *
1945  *	(Re)populate an rxq free-buffer list with up to @n new packet buffers.
1946  *	The caller must assure that @n does not exceed the queue's capacity.
1947  */
1948 static void
1949 _iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
1950 {
1951 	struct if_rxd_update iru;
1952 	struct rxq_refill_cb_arg cb_arg;
1953 	struct mbuf *m;
1954 	caddr_t cl, *sd_cl;
1955 	struct mbuf **sd_m;
1956 	bus_dmamap_t *sd_map;
1957 	bus_addr_t bus_addr, *sd_ba;
1958 	int err, frag_idx, i, idx, n, pidx;
1959 	qidx_t credits;
1960 
1961 	sd_m = fl->ifl_sds.ifsd_m;
1962 	sd_map = fl->ifl_sds.ifsd_map;
1963 	sd_cl = fl->ifl_sds.ifsd_cl;
1964 	sd_ba = fl->ifl_sds.ifsd_ba;
1965 	pidx = fl->ifl_pidx;
1966 	idx = pidx;
1967 	frag_idx = fl->ifl_fragidx;
1968 	credits = fl->ifl_credits;
1969 
1970 	i = 0;
1971 	n = count;
1972 	MPASS(n > 0);
1973 	MPASS(credits + n <= fl->ifl_size);
1974 
1975 	if (pidx < fl->ifl_cidx)
1976 		MPASS(pidx + n <= fl->ifl_cidx);
1977 	if (pidx == fl->ifl_cidx && (credits < fl->ifl_size))
1978 		MPASS(fl->ifl_gen == 0);
1979 	if (pidx > fl->ifl_cidx)
1980 		MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx);
1981 
1982 	DBG_COUNTER_INC(fl_refills);
1983 	if (n > 8)
1984 		DBG_COUNTER_INC(fl_refills_large);
1985 	iru_init(&iru, fl->ifl_rxq, fl->ifl_id);
1986 	while (n--) {
1987 		/*
1988 		 * We allocate an uninitialized mbuf + cluster, mbuf is
1989 		 * initialized after rx.
1990 		 *
1991 		 * If the cluster is still set then we know a minimum sized packet was received
1992 		 */
1993 		bit_ffc_at(fl->ifl_rx_bitmap, frag_idx, fl->ifl_size,
1994 		    &frag_idx);
1995 		if (frag_idx < 0)
1996 			bit_ffc(fl->ifl_rx_bitmap, fl->ifl_size, &frag_idx);
1997 		MPASS(frag_idx >= 0);
1998 		if ((cl = sd_cl[frag_idx]) == NULL) {
1999 			if ((cl = m_cljget(NULL, M_NOWAIT, fl->ifl_buf_size)) == NULL)
2000 				break;
2001 
2002 			cb_arg.error = 0;
2003 			MPASS(sd_map != NULL);
2004 			err = bus_dmamap_load(fl->ifl_buf_tag, sd_map[frag_idx],
2005 			    cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg,
2006 			    BUS_DMA_NOWAIT);
2007 			if (err != 0 || cb_arg.error) {
2008 				/*
2009 				 * !zone_pack ?
2010 				 */
2011 				if (fl->ifl_zone == zone_pack)
2012 					uma_zfree(fl->ifl_zone, cl);
2013 				break;
2014 			}
2015 
2016 			sd_ba[frag_idx] =  bus_addr = cb_arg.seg.ds_addr;
2017 			sd_cl[frag_idx] = cl;
2018 #if MEMORY_LOGGING
2019 			fl->ifl_cl_enqueued++;
2020 #endif
2021 		} else {
2022 			bus_addr = sd_ba[frag_idx];
2023 		}
2024 		bus_dmamap_sync(fl->ifl_buf_tag, sd_map[frag_idx],
2025 		    BUS_DMASYNC_PREREAD);
2026 
2027 		MPASS(sd_m[frag_idx] == NULL);
2028 		if ((m = m_gethdr(M_NOWAIT, MT_NOINIT)) == NULL) {
2029 			break;
2030 		}
2031 		sd_m[frag_idx] = m;
2032 		bit_set(fl->ifl_rx_bitmap, frag_idx);
2033 #if MEMORY_LOGGING
2034 		fl->ifl_m_enqueued++;
2035 #endif
2036 
2037 		DBG_COUNTER_INC(rx_allocs);
2038 		fl->ifl_rxd_idxs[i] = frag_idx;
2039 		fl->ifl_bus_addrs[i] = bus_addr;
2040 		fl->ifl_vm_addrs[i] = cl;
2041 		credits++;
2042 		i++;
2043 		MPASS(credits <= fl->ifl_size);
2044 		if (++idx == fl->ifl_size) {
2045 			fl->ifl_gen = 1;
2046 			idx = 0;
2047 		}
2048 		if (n == 0 || i == IFLIB_MAX_RX_REFRESH) {
2049 			iru.iru_pidx = pidx;
2050 			iru.iru_count = i;
2051 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2052 			i = 0;
2053 			pidx = idx;
2054 			fl->ifl_pidx = idx;
2055 			fl->ifl_credits = credits;
2056 		}
2057 	}
2058 
2059 	if (i) {
2060 		iru.iru_pidx = pidx;
2061 		iru.iru_count = i;
2062 		ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2063 		fl->ifl_pidx = idx;
2064 		fl->ifl_credits = credits;
2065 	}
2066 	DBG_COUNTER_INC(rxd_flush);
2067 	if (fl->ifl_pidx == 0)
2068 		pidx = fl->ifl_size - 1;
2069 	else
2070 		pidx = fl->ifl_pidx - 1;
2071 
2072 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2073 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2074 	ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id, fl->ifl_id, pidx);
2075 	fl->ifl_fragidx = frag_idx;
2076 }
2077 
2078 static __inline void
2079 __iflib_fl_refill_lt(if_ctx_t ctx, iflib_fl_t fl, int max)
2080 {
2081 	/* we avoid allowing pidx to catch up with cidx as it confuses ixl */
2082 	int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1;
2083 #ifdef INVARIANTS
2084 	int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1;
2085 #endif
2086 
2087 	MPASS(fl->ifl_credits <= fl->ifl_size);
2088 	MPASS(reclaimable == delta);
2089 
2090 	if (reclaimable > 0)
2091 		_iflib_fl_refill(ctx, fl, min(max, reclaimable));
2092 }
2093 
2094 uint8_t
2095 iflib_in_detach(if_ctx_t ctx)
2096 {
2097 	bool in_detach;
2098 	STATE_LOCK(ctx);
2099 	in_detach = !!(ctx->ifc_flags & IFC_IN_DETACH);
2100 	STATE_UNLOCK(ctx);
2101 	return (in_detach);
2102 }
2103 
2104 static void
2105 iflib_fl_bufs_free(iflib_fl_t fl)
2106 {
2107 	iflib_dma_info_t idi = fl->ifl_ifdi;
2108 	bus_dmamap_t sd_map;
2109 	uint32_t i;
2110 
2111 	for (i = 0; i < fl->ifl_size; i++) {
2112 		struct mbuf **sd_m = &fl->ifl_sds.ifsd_m[i];
2113 		caddr_t *sd_cl = &fl->ifl_sds.ifsd_cl[i];
2114 
2115 		if (*sd_cl != NULL) {
2116 			sd_map = fl->ifl_sds.ifsd_map[i];
2117 			bus_dmamap_sync(fl->ifl_buf_tag, sd_map,
2118 			    BUS_DMASYNC_POSTREAD);
2119 			bus_dmamap_unload(fl->ifl_buf_tag, sd_map);
2120 			if (*sd_cl != NULL)
2121 				uma_zfree(fl->ifl_zone, *sd_cl);
2122 			// XXX: Should this get moved out?
2123 			if (iflib_in_detach(fl->ifl_rxq->ifr_ctx))
2124 				bus_dmamap_destroy(fl->ifl_buf_tag, sd_map);
2125 			if (*sd_m != NULL) {
2126 				m_init(*sd_m, M_NOWAIT, MT_DATA, 0);
2127 				uma_zfree(zone_mbuf, *sd_m);
2128 			}
2129 		} else {
2130 			MPASS(*sd_cl == NULL);
2131 			MPASS(*sd_m == NULL);
2132 		}
2133 #if MEMORY_LOGGING
2134 		fl->ifl_m_dequeued++;
2135 		fl->ifl_cl_dequeued++;
2136 #endif
2137 		*sd_cl = NULL;
2138 		*sd_m = NULL;
2139 	}
2140 #ifdef INVARIANTS
2141 	for (i = 0; i < fl->ifl_size; i++) {
2142 		MPASS(fl->ifl_sds.ifsd_cl[i] == NULL);
2143 		MPASS(fl->ifl_sds.ifsd_m[i] == NULL);
2144 	}
2145 #endif
2146 	/*
2147 	 * Reset free list values
2148 	 */
2149 	fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = fl->ifl_fragidx = 0;
2150 	bzero(idi->idi_vaddr, idi->idi_size);
2151 }
2152 
2153 /*********************************************************************
2154  *
2155  *  Initialize a receive ring and its buffers.
2156  *
2157  **********************************************************************/
2158 static int
2159 iflib_fl_setup(iflib_fl_t fl)
2160 {
2161 	iflib_rxq_t rxq = fl->ifl_rxq;
2162 	if_ctx_t ctx = rxq->ifr_ctx;
2163 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2164 
2165 	bit_nclear(fl->ifl_rx_bitmap, 0, fl->ifl_size - 1);
2166 	/*
2167 	** Free current RX buffer structs and their mbufs
2168 	*/
2169 	iflib_fl_bufs_free(fl);
2170 	/* Now replenish the mbufs */
2171 	MPASS(fl->ifl_credits == 0);
2172 	/*
2173 	 * XXX don't set the max_frame_size to larger
2174 	 * than the hardware can handle
2175 	 */
2176 	if (sctx->isc_max_frame_size <= 2048)
2177 		fl->ifl_buf_size = MCLBYTES;
2178 #ifndef CONTIGMALLOC_WORKS
2179 	else
2180 		fl->ifl_buf_size = MJUMPAGESIZE;
2181 #else
2182 	else if (sctx->isc_max_frame_size <= 4096)
2183 		fl->ifl_buf_size = MJUMPAGESIZE;
2184 	else if (sctx->isc_max_frame_size <= 9216)
2185 		fl->ifl_buf_size = MJUM9BYTES;
2186 	else
2187 		fl->ifl_buf_size = MJUM16BYTES;
2188 #endif
2189 	if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size)
2190 		ctx->ifc_max_fl_buf_size = fl->ifl_buf_size;
2191 	fl->ifl_cltype = m_gettype(fl->ifl_buf_size);
2192 	fl->ifl_zone = m_getzone(fl->ifl_buf_size);
2193 
2194 
2195 	/* avoid pre-allocating zillions of clusters to an idle card
2196 	 * potentially speeding up attach
2197 	 */
2198 	_iflib_fl_refill(ctx, fl, min(128, fl->ifl_size));
2199 	MPASS(min(128, fl->ifl_size) == fl->ifl_credits);
2200 	if (min(128, fl->ifl_size) != fl->ifl_credits)
2201 		return (ENOBUFS);
2202 	/*
2203 	 * handle failure
2204 	 */
2205 	MPASS(rxq != NULL);
2206 	MPASS(fl->ifl_ifdi != NULL);
2207 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2208 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2209 	return (0);
2210 }
2211 
2212 /*********************************************************************
2213  *
2214  *  Free receive ring data structures
2215  *
2216  **********************************************************************/
2217 static void
2218 iflib_rx_sds_free(iflib_rxq_t rxq)
2219 {
2220 	iflib_fl_t fl;
2221 	int i, j;
2222 
2223 	if (rxq->ifr_fl != NULL) {
2224 		for (i = 0; i < rxq->ifr_nfl; i++) {
2225 			fl = &rxq->ifr_fl[i];
2226 			if (fl->ifl_buf_tag != NULL) {
2227 				if (fl->ifl_sds.ifsd_map != NULL) {
2228 					for (j = 0; j < fl->ifl_size; j++) {
2229 						if (fl->ifl_sds.ifsd_map[j] ==
2230 						    NULL)
2231 							continue;
2232 						bus_dmamap_sync(
2233 						    fl->ifl_buf_tag,
2234 						    fl->ifl_sds.ifsd_map[j],
2235 						    BUS_DMASYNC_POSTREAD);
2236 						bus_dmamap_unload(
2237 						    fl->ifl_buf_tag,
2238 						    fl->ifl_sds.ifsd_map[j]);
2239 					}
2240 				}
2241 				bus_dma_tag_destroy(fl->ifl_buf_tag);
2242 				fl->ifl_buf_tag = NULL;
2243 			}
2244 			free(fl->ifl_sds.ifsd_m, M_IFLIB);
2245 			free(fl->ifl_sds.ifsd_cl, M_IFLIB);
2246 			free(fl->ifl_sds.ifsd_ba, M_IFLIB);
2247 			free(fl->ifl_sds.ifsd_map, M_IFLIB);
2248 			fl->ifl_sds.ifsd_m = NULL;
2249 			fl->ifl_sds.ifsd_cl = NULL;
2250 			fl->ifl_sds.ifsd_ba = NULL;
2251 			fl->ifl_sds.ifsd_map = NULL;
2252 		}
2253 		free(rxq->ifr_fl, M_IFLIB);
2254 		rxq->ifr_fl = NULL;
2255 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
2256 	}
2257 }
2258 
2259 /*
2260  * MI independent logic
2261  *
2262  */
2263 static void
2264 iflib_timer(void *arg)
2265 {
2266 	iflib_txq_t txq = arg;
2267 	if_ctx_t ctx = txq->ift_ctx;
2268 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2269 	uint64_t this_tick = ticks;
2270 	uint32_t reset_on = hz / 2;
2271 
2272 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
2273 		return;
2274 	/*
2275 	** Check on the state of the TX queue(s), this
2276 	** can be done without the lock because its RO
2277 	** and the HUNG state will be static if set.
2278 	*/
2279 	if (this_tick - txq->ift_last_timer_tick >= hz / 2) {
2280 		txq->ift_last_timer_tick = this_tick;
2281 		IFDI_TIMER(ctx, txq->ift_id);
2282 		if ((txq->ift_qstatus == IFLIB_QUEUE_HUNG) &&
2283 		    ((txq->ift_cleaned_prev == txq->ift_cleaned) ||
2284 		     (sctx->isc_pause_frames == 0)))
2285 			goto hung;
2286 
2287 		if (ifmp_ring_is_stalled(txq->ift_br))
2288 			txq->ift_qstatus = IFLIB_QUEUE_HUNG;
2289 		txq->ift_cleaned_prev = txq->ift_cleaned;
2290 	}
2291 #ifdef DEV_NETMAP
2292 	if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP)
2293 		iflib_netmap_timer_adjust(ctx, txq, &reset_on);
2294 #endif
2295 	/* handle any laggards */
2296 	if (txq->ift_db_pending)
2297 		GROUPTASK_ENQUEUE(&txq->ift_task);
2298 
2299 	sctx->isc_pause_frames = 0;
2300 	if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)
2301 		callout_reset_on(&txq->ift_timer, reset_on, iflib_timer, txq, txq->ift_timer.c_cpu);
2302 	return;
2303  hung:
2304 	device_printf(ctx->ifc_dev,  "TX(%d) desc avail = %d, pidx = %d\n",
2305 				  txq->ift_id, TXQ_AVAIL(txq), txq->ift_pidx);
2306 	STATE_LOCK(ctx);
2307 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2308 	ctx->ifc_flags |= (IFC_DO_WATCHDOG|IFC_DO_RESET);
2309 	iflib_admin_intr_deferred(ctx);
2310 	STATE_UNLOCK(ctx);
2311 }
2312 
2313 static void
2314 iflib_init_locked(if_ctx_t ctx)
2315 {
2316 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2317 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2318 	if_t ifp = ctx->ifc_ifp;
2319 	iflib_fl_t fl;
2320 	iflib_txq_t txq;
2321 	iflib_rxq_t rxq;
2322 	int i, j, tx_ip_csum_flags, tx_ip6_csum_flags;
2323 
2324 
2325 	if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2326 	IFDI_INTR_DISABLE(ctx);
2327 
2328 	tx_ip_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP);
2329 	tx_ip6_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | CSUM_IP6_SCTP);
2330 	/* Set hardware offload abilities */
2331 	if_clearhwassist(ifp);
2332 	if (if_getcapenable(ifp) & IFCAP_TXCSUM)
2333 		if_sethwassistbits(ifp, tx_ip_csum_flags, 0);
2334 	if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6)
2335 		if_sethwassistbits(ifp,  tx_ip6_csum_flags, 0);
2336 	if (if_getcapenable(ifp) & IFCAP_TSO4)
2337 		if_sethwassistbits(ifp, CSUM_IP_TSO, 0);
2338 	if (if_getcapenable(ifp) & IFCAP_TSO6)
2339 		if_sethwassistbits(ifp, CSUM_IP6_TSO, 0);
2340 
2341 	for (i = 0, txq = ctx->ifc_txqs; i < sctx->isc_ntxqsets; i++, txq++) {
2342 		CALLOUT_LOCK(txq);
2343 		callout_stop(&txq->ift_timer);
2344 		CALLOUT_UNLOCK(txq);
2345 		iflib_netmap_txq_init(ctx, txq);
2346 	}
2347 #ifdef INVARIANTS
2348 	i = if_getdrvflags(ifp);
2349 #endif
2350 	IFDI_INIT(ctx);
2351 	MPASS(if_getdrvflags(ifp) == i);
2352 	for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
2353 		/* XXX this should really be done on a per-queue basis */
2354 		if (if_getcapenable(ifp) & IFCAP_NETMAP) {
2355 			MPASS(rxq->ifr_id == i);
2356 			iflib_netmap_rxq_init(ctx, rxq);
2357 			continue;
2358 		}
2359 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
2360 			if (iflib_fl_setup(fl)) {
2361 				device_printf(ctx->ifc_dev, "freelist setup failed - check cluster settings\n");
2362 				goto done;
2363 			}
2364 		}
2365 	}
2366 done:
2367 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
2368 	IFDI_INTR_ENABLE(ctx);
2369 	txq = ctx->ifc_txqs;
2370 	for (i = 0; i < sctx->isc_ntxqsets; i++, txq++)
2371 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq,
2372 			txq->ift_timer.c_cpu);
2373 }
2374 
2375 static int
2376 iflib_media_change(if_t ifp)
2377 {
2378 	if_ctx_t ctx = if_getsoftc(ifp);
2379 	int err;
2380 
2381 	CTX_LOCK(ctx);
2382 	if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0)
2383 		iflib_init_locked(ctx);
2384 	CTX_UNLOCK(ctx);
2385 	return (err);
2386 }
2387 
2388 static void
2389 iflib_media_status(if_t ifp, struct ifmediareq *ifmr)
2390 {
2391 	if_ctx_t ctx = if_getsoftc(ifp);
2392 
2393 	CTX_LOCK(ctx);
2394 	IFDI_UPDATE_ADMIN_STATUS(ctx);
2395 	IFDI_MEDIA_STATUS(ctx, ifmr);
2396 	CTX_UNLOCK(ctx);
2397 }
2398 
2399 void
2400 iflib_stop(if_ctx_t ctx)
2401 {
2402 	iflib_txq_t txq = ctx->ifc_txqs;
2403 	iflib_rxq_t rxq = ctx->ifc_rxqs;
2404 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2405 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2406 	iflib_dma_info_t di;
2407 	iflib_fl_t fl;
2408 	int i, j;
2409 
2410 	/* Tell the stack that the interface is no longer active */
2411 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2412 
2413 	IFDI_INTR_DISABLE(ctx);
2414 	DELAY(1000);
2415 	IFDI_STOP(ctx);
2416 	DELAY(1000);
2417 
2418 	iflib_debug_reset();
2419 	/* Wait for current tx queue users to exit to disarm watchdog timer. */
2420 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) {
2421 		/* make sure all transmitters have completed before proceeding XXX */
2422 
2423 		CALLOUT_LOCK(txq);
2424 		callout_stop(&txq->ift_timer);
2425 		CALLOUT_UNLOCK(txq);
2426 
2427 		/* clean any enqueued buffers */
2428 		iflib_ifmp_purge(txq);
2429 		/* Free any existing tx buffers. */
2430 		for (j = 0; j < txq->ift_size; j++) {
2431 			iflib_txsd_free(ctx, txq, j);
2432 		}
2433 		txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0;
2434 		txq->ift_in_use = txq->ift_gen = txq->ift_cidx = txq->ift_pidx = txq->ift_no_desc_avail = 0;
2435 		txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0;
2436 		txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0;
2437 		txq->ift_pullups = 0;
2438 		ifmp_ring_reset_stats(txq->ift_br);
2439 		for (j = 0, di = txq->ift_ifdi; j < sctx->isc_ntxqs; j++, di++)
2440 			bzero((void *)di->idi_vaddr, di->idi_size);
2441 	}
2442 	for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) {
2443 		/* make sure all transmitters have completed before proceeding XXX */
2444 
2445 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
2446 		for (j = 0, di = rxq->ifr_ifdi; j < sctx->isc_nrxqs; j++, di++)
2447 			bzero((void *)di->idi_vaddr, di->idi_size);
2448 		/* also resets the free lists pidx/cidx */
2449 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
2450 			iflib_fl_bufs_free(fl);
2451 	}
2452 }
2453 
2454 static inline caddr_t
2455 calc_next_rxd(iflib_fl_t fl, int cidx)
2456 {
2457 	qidx_t size;
2458 	int nrxd;
2459 	caddr_t start, end, cur, next;
2460 
2461 	nrxd = fl->ifl_size;
2462 	size = fl->ifl_rxd_size;
2463 	start = fl->ifl_ifdi->idi_vaddr;
2464 
2465 	if (__predict_false(size == 0))
2466 		return (start);
2467 	cur = start + size*cidx;
2468 	end = start + size*nrxd;
2469 	next = CACHE_PTR_NEXT(cur);
2470 	return (next < end ? next : start);
2471 }
2472 
2473 static inline void
2474 prefetch_pkts(iflib_fl_t fl, int cidx)
2475 {
2476 	int nextptr;
2477 	int nrxd = fl->ifl_size;
2478 	caddr_t next_rxd;
2479 
2480 
2481 	nextptr = (cidx + CACHE_PTR_INCREMENT) & (nrxd-1);
2482 	prefetch(&fl->ifl_sds.ifsd_m[nextptr]);
2483 	prefetch(&fl->ifl_sds.ifsd_cl[nextptr]);
2484 	next_rxd = calc_next_rxd(fl, cidx);
2485 	prefetch(next_rxd);
2486 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 1) & (nrxd-1)]);
2487 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 2) & (nrxd-1)]);
2488 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 3) & (nrxd-1)]);
2489 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 4) & (nrxd-1)]);
2490 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 1) & (nrxd-1)]);
2491 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 2) & (nrxd-1)]);
2492 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 3) & (nrxd-1)]);
2493 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 4) & (nrxd-1)]);
2494 }
2495 
2496 static void
2497 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, int unload, if_rxsd_t sd)
2498 {
2499 	int flid, cidx;
2500 	bus_dmamap_t map;
2501 	iflib_fl_t fl;
2502 	int next;
2503 
2504 	map = NULL;
2505 	flid = irf->irf_flid;
2506 	cidx = irf->irf_idx;
2507 	fl = &rxq->ifr_fl[flid];
2508 	sd->ifsd_fl = fl;
2509 	sd->ifsd_cidx = cidx;
2510 	sd->ifsd_m = &fl->ifl_sds.ifsd_m[cidx];
2511 	sd->ifsd_cl = &fl->ifl_sds.ifsd_cl[cidx];
2512 	fl->ifl_credits--;
2513 #if MEMORY_LOGGING
2514 	fl->ifl_m_dequeued++;
2515 #endif
2516 	if (rxq->ifr_ctx->ifc_flags & IFC_PREFETCH)
2517 		prefetch_pkts(fl, cidx);
2518 	next = (cidx + CACHE_PTR_INCREMENT) & (fl->ifl_size-1);
2519 	prefetch(&fl->ifl_sds.ifsd_map[next]);
2520 	map = fl->ifl_sds.ifsd_map[cidx];
2521 	next = (cidx + CACHE_LINE_SIZE) & (fl->ifl_size-1);
2522 
2523 	/* not valid assert if bxe really does SGE from non-contiguous elements */
2524 	MPASS(fl->ifl_cidx == cidx);
2525 	bus_dmamap_sync(fl->ifl_buf_tag, map, BUS_DMASYNC_POSTREAD);
2526 	if (unload)
2527 		bus_dmamap_unload(fl->ifl_buf_tag, map);
2528 	fl->ifl_cidx = (fl->ifl_cidx + 1) & (fl->ifl_size-1);
2529 	if (__predict_false(fl->ifl_cidx == 0))
2530 		fl->ifl_gen = 0;
2531 	bit_clear(fl->ifl_rx_bitmap, cidx);
2532 }
2533 
2534 static struct mbuf *
2535 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri, if_rxsd_t sd)
2536 {
2537 	int i, padlen , flags;
2538 	struct mbuf *m, *mh, *mt;
2539 	caddr_t cl;
2540 
2541 	i = 0;
2542 	mh = NULL;
2543 	do {
2544 		rxd_frag_to_sd(rxq, &ri->iri_frags[i], TRUE, sd);
2545 
2546 		MPASS(*sd->ifsd_cl != NULL);
2547 		MPASS(*sd->ifsd_m != NULL);
2548 
2549 		/* Don't include zero-length frags */
2550 		if (ri->iri_frags[i].irf_len == 0) {
2551 			/* XXX we can save the cluster here, but not the mbuf */
2552 			m_init(*sd->ifsd_m, M_NOWAIT, MT_DATA, 0);
2553 			m_free(*sd->ifsd_m);
2554 			*sd->ifsd_m = NULL;
2555 			continue;
2556 		}
2557 		m = *sd->ifsd_m;
2558 		*sd->ifsd_m = NULL;
2559 		if (mh == NULL) {
2560 			flags = M_PKTHDR|M_EXT;
2561 			mh = mt = m;
2562 			padlen = ri->iri_pad;
2563 		} else {
2564 			flags = M_EXT;
2565 			mt->m_next = m;
2566 			mt = m;
2567 			/* assuming padding is only on the first fragment */
2568 			padlen = 0;
2569 		}
2570 		cl = *sd->ifsd_cl;
2571 		*sd->ifsd_cl = NULL;
2572 
2573 		/* Can these two be made one ? */
2574 		m_init(m, M_NOWAIT, MT_DATA, flags);
2575 		m_cljset(m, cl, sd->ifsd_fl->ifl_cltype);
2576 		/*
2577 		 * These must follow m_init and m_cljset
2578 		 */
2579 		m->m_data += padlen;
2580 		ri->iri_len -= padlen;
2581 		m->m_len = ri->iri_frags[i].irf_len;
2582 	} while (++i < ri->iri_nfrags);
2583 
2584 	return (mh);
2585 }
2586 
2587 /*
2588  * Process one software descriptor
2589  */
2590 static struct mbuf *
2591 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri)
2592 {
2593 	struct if_rxsd sd;
2594 	struct mbuf *m;
2595 
2596 	/* should I merge this back in now that the two paths are basically duplicated? */
2597 	if (ri->iri_nfrags == 1 &&
2598 	    ri->iri_frags[0].irf_len <= MIN(IFLIB_RX_COPY_THRESH, MHLEN)) {
2599 		rxd_frag_to_sd(rxq, &ri->iri_frags[0], FALSE, &sd);
2600 		m = *sd.ifsd_m;
2601 		*sd.ifsd_m = NULL;
2602 		m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR);
2603 #ifndef __NO_STRICT_ALIGNMENT
2604 		if (!IP_ALIGNED(m))
2605 			m->m_data += 2;
2606 #endif
2607 		memcpy(m->m_data, *sd.ifsd_cl, ri->iri_len);
2608 		m->m_len = ri->iri_frags[0].irf_len;
2609        } else {
2610 		m = assemble_segments(rxq, ri, &sd);
2611 	}
2612 	m->m_pkthdr.len = ri->iri_len;
2613 	m->m_pkthdr.rcvif = ri->iri_ifp;
2614 	m->m_flags |= ri->iri_flags;
2615 	m->m_pkthdr.ether_vtag = ri->iri_vtag;
2616 	m->m_pkthdr.flowid = ri->iri_flowid;
2617 	M_HASHTYPE_SET(m, ri->iri_rsstype);
2618 	m->m_pkthdr.csum_flags = ri->iri_csum_flags;
2619 	m->m_pkthdr.csum_data = ri->iri_csum_data;
2620 	return (m);
2621 }
2622 
2623 #if defined(INET6) || defined(INET)
2624 static void
2625 iflib_get_ip_forwarding(struct lro_ctrl *lc, bool *v4, bool *v6)
2626 {
2627 	CURVNET_SET(lc->ifp->if_vnet);
2628 #if defined(INET6)
2629 	*v6 = VNET(ip6_forwarding);
2630 #endif
2631 #if defined(INET)
2632 	*v4 = VNET(ipforwarding);
2633 #endif
2634 	CURVNET_RESTORE();
2635 }
2636 
2637 /*
2638  * Returns true if it's possible this packet could be LROed.
2639  * if it returns false, it is guaranteed that tcp_lro_rx()
2640  * would not return zero.
2641  */
2642 static bool
2643 iflib_check_lro_possible(struct mbuf *m, bool v4_forwarding, bool v6_forwarding)
2644 {
2645 	struct ether_header *eh;
2646 	uint16_t eh_type;
2647 
2648 	eh = mtod(m, struct ether_header *);
2649 	eh_type = ntohs(eh->ether_type);
2650 	switch (eh_type) {
2651 #if defined(INET6)
2652 		case ETHERTYPE_IPV6:
2653 			return !v6_forwarding;
2654 #endif
2655 #if defined (INET)
2656 		case ETHERTYPE_IP:
2657 			return !v4_forwarding;
2658 #endif
2659 	}
2660 
2661 	return false;
2662 }
2663 #else
2664 static void
2665 iflib_get_ip_forwarding(struct lro_ctrl *lc __unused, bool *v4 __unused, bool *v6 __unused)
2666 {
2667 }
2668 #endif
2669 
2670 static bool
2671 iflib_rxeof(iflib_rxq_t rxq, qidx_t budget)
2672 {
2673 	if_ctx_t ctx = rxq->ifr_ctx;
2674 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2675 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2676 	int avail, i;
2677 	qidx_t *cidxp;
2678 	struct if_rxd_info ri;
2679 	int err, budget_left, rx_bytes, rx_pkts;
2680 	iflib_fl_t fl;
2681 	struct ifnet *ifp;
2682 	int lro_enabled;
2683 	bool v4_forwarding, v6_forwarding, lro_possible;
2684 
2685 	/*
2686 	 * XXX early demux data packets so that if_input processing only handles
2687 	 * acks in interrupt context
2688 	 */
2689 	struct mbuf *m, *mh, *mt, *mf;
2690 
2691 	lro_possible = v4_forwarding = v6_forwarding = false;
2692 	ifp = ctx->ifc_ifp;
2693 	mh = mt = NULL;
2694 	MPASS(budget > 0);
2695 	rx_pkts	= rx_bytes = 0;
2696 	if (sctx->isc_flags & IFLIB_HAS_RXCQ)
2697 		cidxp = &rxq->ifr_cq_cidx;
2698 	else
2699 		cidxp = &rxq->ifr_fl[0].ifl_cidx;
2700 	if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) {
2701 		for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2702 			__iflib_fl_refill_lt(ctx, fl, budget + 8);
2703 		DBG_COUNTER_INC(rx_unavail);
2704 		return (false);
2705 	}
2706 
2707 	for (budget_left = budget; budget_left > 0 && avail > 0;) {
2708 		if (__predict_false(!CTX_ACTIVE(ctx))) {
2709 			DBG_COUNTER_INC(rx_ctx_inactive);
2710 			break;
2711 		}
2712 		/*
2713 		 * Reset client set fields to their default values
2714 		 */
2715 		rxd_info_zero(&ri);
2716 		ri.iri_qsidx = rxq->ifr_id;
2717 		ri.iri_cidx = *cidxp;
2718 		ri.iri_ifp = ifp;
2719 		ri.iri_frags = rxq->ifr_frags;
2720 		err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
2721 
2722 		if (err)
2723 			goto err;
2724 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
2725 			*cidxp = ri.iri_cidx;
2726 			/* Update our consumer index */
2727 			/* XXX NB: shurd - check if this is still safe */
2728 			while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0]) {
2729 				rxq->ifr_cq_cidx -= scctx->isc_nrxd[0];
2730 				rxq->ifr_cq_gen = 0;
2731 			}
2732 			/* was this only a completion queue message? */
2733 			if (__predict_false(ri.iri_nfrags == 0))
2734 				continue;
2735 		}
2736 		MPASS(ri.iri_nfrags != 0);
2737 		MPASS(ri.iri_len != 0);
2738 
2739 		/* will advance the cidx on the corresponding free lists */
2740 		m = iflib_rxd_pkt_get(rxq, &ri);
2741 		avail--;
2742 		budget_left--;
2743 		if (avail == 0 && budget_left)
2744 			avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left);
2745 
2746 		if (__predict_false(m == NULL)) {
2747 			DBG_COUNTER_INC(rx_mbuf_null);
2748 			continue;
2749 		}
2750 		/* imm_pkt: -- cxgb */
2751 		if (mh == NULL)
2752 			mh = mt = m;
2753 		else {
2754 			mt->m_nextpkt = m;
2755 			mt = m;
2756 		}
2757 	}
2758 	/* make sure that we can refill faster than drain */
2759 	for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2760 		__iflib_fl_refill_lt(ctx, fl, budget + 8);
2761 
2762 	lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO);
2763 	if (lro_enabled)
2764 		iflib_get_ip_forwarding(&rxq->ifr_lc, &v4_forwarding, &v6_forwarding);
2765 	mt = mf = NULL;
2766 	while (mh != NULL) {
2767 		m = mh;
2768 		mh = mh->m_nextpkt;
2769 		m->m_nextpkt = NULL;
2770 #ifndef __NO_STRICT_ALIGNMENT
2771 		if (!IP_ALIGNED(m) && (m = iflib_fixup_rx(m)) == NULL)
2772 			continue;
2773 #endif
2774 		rx_bytes += m->m_pkthdr.len;
2775 		rx_pkts++;
2776 #if defined(INET6) || defined(INET)
2777 		if (lro_enabled) {
2778 			if (!lro_possible) {
2779 				lro_possible = iflib_check_lro_possible(m, v4_forwarding, v6_forwarding);
2780 				if (lro_possible && mf != NULL) {
2781 					ifp->if_input(ifp, mf);
2782 					DBG_COUNTER_INC(rx_if_input);
2783 					mt = mf = NULL;
2784 				}
2785 			}
2786 			if ((m->m_pkthdr.csum_flags & (CSUM_L4_CALC|CSUM_L4_VALID)) ==
2787 			    (CSUM_L4_CALC|CSUM_L4_VALID)) {
2788 				if (lro_possible && tcp_lro_rx(&rxq->ifr_lc, m, 0) == 0)
2789 					continue;
2790 			}
2791 		}
2792 #endif
2793 		if (lro_possible) {
2794 			ifp->if_input(ifp, m);
2795 			DBG_COUNTER_INC(rx_if_input);
2796 			continue;
2797 		}
2798 
2799 		if (mf == NULL)
2800 			mf = m;
2801 		if (mt != NULL)
2802 			mt->m_nextpkt = m;
2803 		mt = m;
2804 	}
2805 	if (mf != NULL) {
2806 		ifp->if_input(ifp, mf);
2807 		DBG_COUNTER_INC(rx_if_input);
2808 	}
2809 
2810 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
2811 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
2812 
2813 	/*
2814 	 * Flush any outstanding LRO work
2815 	 */
2816 #if defined(INET6) || defined(INET)
2817 	tcp_lro_flush_all(&rxq->ifr_lc);
2818 #endif
2819 	if (avail)
2820 		return true;
2821 	return (iflib_rxd_avail(ctx, rxq, *cidxp, 1));
2822 err:
2823 	STATE_LOCK(ctx);
2824 	ctx->ifc_flags |= IFC_DO_RESET;
2825 	iflib_admin_intr_deferred(ctx);
2826 	STATE_UNLOCK(ctx);
2827 	return (false);
2828 }
2829 
2830 #define TXD_NOTIFY_COUNT(txq) (((txq)->ift_size / (txq)->ift_update_freq)-1)
2831 static inline qidx_t
2832 txq_max_db_deferred(iflib_txq_t txq, qidx_t in_use)
2833 {
2834 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2835 	qidx_t minthresh = txq->ift_size / 8;
2836 	if (in_use > 4*minthresh)
2837 		return (notify_count);
2838 	if (in_use > 2*minthresh)
2839 		return (notify_count >> 1);
2840 	if (in_use > minthresh)
2841 		return (notify_count >> 3);
2842 	return (0);
2843 }
2844 
2845 static inline qidx_t
2846 txq_max_rs_deferred(iflib_txq_t txq)
2847 {
2848 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2849 	qidx_t minthresh = txq->ift_size / 8;
2850 	if (txq->ift_in_use > 4*minthresh)
2851 		return (notify_count);
2852 	if (txq->ift_in_use > 2*minthresh)
2853 		return (notify_count >> 1);
2854 	if (txq->ift_in_use > minthresh)
2855 		return (notify_count >> 2);
2856 	return (2);
2857 }
2858 
2859 #define M_CSUM_FLAGS(m) ((m)->m_pkthdr.csum_flags)
2860 #define M_HAS_VLANTAG(m) (m->m_flags & M_VLANTAG)
2861 
2862 #define TXQ_MAX_DB_DEFERRED(txq, in_use) txq_max_db_deferred((txq), (in_use))
2863 #define TXQ_MAX_RS_DEFERRED(txq) txq_max_rs_deferred(txq)
2864 #define TXQ_MAX_DB_CONSUMED(size) (size >> 4)
2865 
2866 /* forward compatibility for cxgb */
2867 #define FIRST_QSET(ctx) 0
2868 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets)
2869 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets)
2870 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx))
2871 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments))
2872 
2873 /* XXX we should be setting this to something other than zero */
2874 #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh)
2875 #define	MAX_TX_DESC(ctx) max((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max, \
2876     (ctx)->ifc_softc_ctx.isc_tx_nsegments)
2877 
2878 static inline bool
2879 iflib_txd_db_check(if_ctx_t ctx, iflib_txq_t txq, int ring, qidx_t in_use)
2880 {
2881 	qidx_t dbval, max;
2882 	bool rang;
2883 
2884 	rang = false;
2885 	max = TXQ_MAX_DB_DEFERRED(txq, in_use);
2886 	if (ring || txq->ift_db_pending >= max) {
2887 		dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx;
2888 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
2889 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2890 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval);
2891 		txq->ift_db_pending = txq->ift_npending = 0;
2892 		rang = true;
2893 	}
2894 	return (rang);
2895 }
2896 
2897 #ifdef PKT_DEBUG
2898 static void
2899 print_pkt(if_pkt_info_t pi)
2900 {
2901 	printf("pi len:  %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n",
2902 	       pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx);
2903 	printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n",
2904 	       pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag);
2905 	printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n",
2906 	       pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto);
2907 }
2908 #endif
2909 
2910 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO)
2911 #define IS_TX_OFFLOAD4(pi) ((pi)->ipi_csum_flags & (CSUM_IP_TCP | CSUM_IP_TSO))
2912 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO)
2913 #define IS_TX_OFFLOAD6(pi) ((pi)->ipi_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_TSO))
2914 
2915 static int
2916 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp)
2917 {
2918 	if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx;
2919 	struct ether_vlan_header *eh;
2920 	struct mbuf *m;
2921 
2922 	m = *mp;
2923 	if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) &&
2924 	    M_WRITABLE(m) == 0) {
2925 		if ((m = m_dup(m, M_NOWAIT)) == NULL) {
2926 			return (ENOMEM);
2927 		} else {
2928 			m_freem(*mp);
2929 			DBG_COUNTER_INC(tx_frees);
2930 			*mp = m;
2931 		}
2932 	}
2933 
2934 	/*
2935 	 * Determine where frame payload starts.
2936 	 * Jump over vlan headers if already present,
2937 	 * helpful for QinQ too.
2938 	 */
2939 	if (__predict_false(m->m_len < sizeof(*eh))) {
2940 		txq->ift_pullups++;
2941 		if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL))
2942 			return (ENOMEM);
2943 	}
2944 	eh = mtod(m, struct ether_vlan_header *);
2945 	if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
2946 		pi->ipi_etype = ntohs(eh->evl_proto);
2947 		pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
2948 	} else {
2949 		pi->ipi_etype = ntohs(eh->evl_encap_proto);
2950 		pi->ipi_ehdrlen = ETHER_HDR_LEN;
2951 	}
2952 
2953 	switch (pi->ipi_etype) {
2954 #ifdef INET
2955 	case ETHERTYPE_IP:
2956 	{
2957 		struct mbuf *n;
2958 		struct ip *ip = NULL;
2959 		struct tcphdr *th = NULL;
2960 		int minthlen;
2961 
2962 		minthlen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip) + sizeof(*th));
2963 		if (__predict_false(m->m_len < minthlen)) {
2964 			/*
2965 			 * if this code bloat is causing too much of a hit
2966 			 * move it to a separate function and mark it noinline
2967 			 */
2968 			if (m->m_len == pi->ipi_ehdrlen) {
2969 				n = m->m_next;
2970 				MPASS(n);
2971 				if (n->m_len >= sizeof(*ip))  {
2972 					ip = (struct ip *)n->m_data;
2973 					if (n->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2974 						th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2975 				} else {
2976 					txq->ift_pullups++;
2977 					if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
2978 						return (ENOMEM);
2979 					ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2980 				}
2981 			} else {
2982 				txq->ift_pullups++;
2983 				if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
2984 					return (ENOMEM);
2985 				ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2986 				if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2987 					th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2988 			}
2989 		} else {
2990 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2991 			if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2992 				th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2993 		}
2994 		pi->ipi_ip_hlen = ip->ip_hl << 2;
2995 		pi->ipi_ipproto = ip->ip_p;
2996 		pi->ipi_flags |= IPI_TX_IPV4;
2997 
2998 		/* TCP checksum offload may require TCP header length */
2999 		if (IS_TX_OFFLOAD4(pi)) {
3000 			if (__predict_true(pi->ipi_ipproto == IPPROTO_TCP)) {
3001 				if (__predict_false(th == NULL)) {
3002 					txq->ift_pullups++;
3003 					if (__predict_false((m = m_pullup(m, (ip->ip_hl << 2) + sizeof(*th))) == NULL))
3004 						return (ENOMEM);
3005 					th = (struct tcphdr *)((caddr_t)ip + pi->ipi_ip_hlen);
3006 				}
3007 				pi->ipi_tcp_hflags = th->th_flags;
3008 				pi->ipi_tcp_hlen = th->th_off << 2;
3009 				pi->ipi_tcp_seq = th->th_seq;
3010 			}
3011 			if (IS_TSO4(pi)) {
3012 				if (__predict_false(ip->ip_p != IPPROTO_TCP))
3013 					return (ENXIO);
3014 				/*
3015 				 * TSO always requires hardware checksum offload.
3016 				 */
3017 				pi->ipi_csum_flags |= (CSUM_IP_TCP | CSUM_IP);
3018 				th->th_sum = in_pseudo(ip->ip_src.s_addr,
3019 						       ip->ip_dst.s_addr, htons(IPPROTO_TCP));
3020 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3021 				if (sctx->isc_flags & IFLIB_TSO_INIT_IP) {
3022 					ip->ip_sum = 0;
3023 					ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz);
3024 				}
3025 			}
3026 		}
3027 		if ((sctx->isc_flags & IFLIB_NEED_ZERO_CSUM) && (pi->ipi_csum_flags & CSUM_IP))
3028                        ip->ip_sum = 0;
3029 
3030 		break;
3031 	}
3032 #endif
3033 #ifdef INET6
3034 	case ETHERTYPE_IPV6:
3035 	{
3036 		struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
3037 		struct tcphdr *th;
3038 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
3039 
3040 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
3041 			txq->ift_pullups++;
3042 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
3043 				return (ENOMEM);
3044 		}
3045 		th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
3046 
3047 		/* XXX-BZ this will go badly in case of ext hdrs. */
3048 		pi->ipi_ipproto = ip6->ip6_nxt;
3049 		pi->ipi_flags |= IPI_TX_IPV6;
3050 
3051 		/* TCP checksum offload may require TCP header length */
3052 		if (IS_TX_OFFLOAD6(pi)) {
3053 			if (pi->ipi_ipproto == IPPROTO_TCP) {
3054 				if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) {
3055 					txq->ift_pullups++;
3056 					if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL))
3057 						return (ENOMEM);
3058 				}
3059 				pi->ipi_tcp_hflags = th->th_flags;
3060 				pi->ipi_tcp_hlen = th->th_off << 2;
3061 				pi->ipi_tcp_seq = th->th_seq;
3062 			}
3063 			if (IS_TSO6(pi)) {
3064 				if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP))
3065 					return (ENXIO);
3066 				/*
3067 				 * TSO always requires hardware checksum offload.
3068 				 */
3069 				pi->ipi_csum_flags |= CSUM_IP6_TCP;
3070 				th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0);
3071 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3072 			}
3073 		}
3074 		break;
3075 	}
3076 #endif
3077 	default:
3078 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
3079 		pi->ipi_ip_hlen = 0;
3080 		break;
3081 	}
3082 	*mp = m;
3083 
3084 	return (0);
3085 }
3086 
3087 /*
3088  * If dodgy hardware rejects the scatter gather chain we've handed it
3089  * we'll need to remove the mbuf chain from ifsg_m[] before we can add the
3090  * m_defrag'd mbufs
3091  */
3092 static __noinline struct mbuf *
3093 iflib_remove_mbuf(iflib_txq_t txq)
3094 {
3095 	int ntxd, pidx;
3096 	struct mbuf *m, **ifsd_m;
3097 
3098 	ifsd_m = txq->ift_sds.ifsd_m;
3099 	ntxd = txq->ift_size;
3100 	pidx = txq->ift_pidx & (ntxd - 1);
3101 	ifsd_m = txq->ift_sds.ifsd_m;
3102 	m = ifsd_m[pidx];
3103 	ifsd_m[pidx] = NULL;
3104 	bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[pidx]);
3105 	if (txq->ift_sds.ifsd_tso_map != NULL)
3106 		bus_dmamap_unload(txq->ift_tso_buf_tag,
3107 		    txq->ift_sds.ifsd_tso_map[pidx]);
3108 #if MEMORY_LOGGING
3109 	txq->ift_dequeued++;
3110 #endif
3111 	return (m);
3112 }
3113 
3114 static inline caddr_t
3115 calc_next_txd(iflib_txq_t txq, int cidx, uint8_t qid)
3116 {
3117 	qidx_t size;
3118 	int ntxd;
3119 	caddr_t start, end, cur, next;
3120 
3121 	ntxd = txq->ift_size;
3122 	size = txq->ift_txd_size[qid];
3123 	start = txq->ift_ifdi[qid].idi_vaddr;
3124 
3125 	if (__predict_false(size == 0))
3126 		return (start);
3127 	cur = start + size*cidx;
3128 	end = start + size*ntxd;
3129 	next = CACHE_PTR_NEXT(cur);
3130 	return (next < end ? next : start);
3131 }
3132 
3133 /*
3134  * Pad an mbuf to ensure a minimum ethernet frame size.
3135  * min_frame_size is the frame size (less CRC) to pad the mbuf to
3136  */
3137 static __noinline int
3138 iflib_ether_pad(device_t dev, struct mbuf **m_head, uint16_t min_frame_size)
3139 {
3140 	/*
3141 	 * 18 is enough bytes to pad an ARP packet to 46 bytes, and
3142 	 * and ARP message is the smallest common payload I can think of
3143 	 */
3144 	static char pad[18];	/* just zeros */
3145 	int n;
3146 	struct mbuf *new_head;
3147 
3148 	if (!M_WRITABLE(*m_head)) {
3149 		new_head = m_dup(*m_head, M_NOWAIT);
3150 		if (new_head == NULL) {
3151 			m_freem(*m_head);
3152 			device_printf(dev, "cannot pad short frame, m_dup() failed");
3153 			DBG_COUNTER_INC(encap_pad_mbuf_fail);
3154 			DBG_COUNTER_INC(tx_frees);
3155 			return ENOMEM;
3156 		}
3157 		m_freem(*m_head);
3158 		*m_head = new_head;
3159 	}
3160 
3161 	for (n = min_frame_size - (*m_head)->m_pkthdr.len;
3162 	     n > 0; n -= sizeof(pad))
3163 		if (!m_append(*m_head, min(n, sizeof(pad)), pad))
3164 			break;
3165 
3166 	if (n > 0) {
3167 		m_freem(*m_head);
3168 		device_printf(dev, "cannot pad short frame\n");
3169 		DBG_COUNTER_INC(encap_pad_mbuf_fail);
3170 		DBG_COUNTER_INC(tx_frees);
3171 		return (ENOBUFS);
3172 	}
3173 
3174 	return 0;
3175 }
3176 
3177 static int
3178 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp)
3179 {
3180 	if_ctx_t		ctx;
3181 	if_shared_ctx_t		sctx;
3182 	if_softc_ctx_t		scctx;
3183 	bus_dma_tag_t		buf_tag;
3184 	bus_dma_segment_t	*segs;
3185 	struct mbuf		*m_head, **ifsd_m;
3186 	void			*next_txd;
3187 	bus_dmamap_t		map;
3188 	struct if_pkt_info	pi;
3189 	int remap = 0;
3190 	int err, nsegs, ndesc, max_segs, pidx, cidx, next, ntxd;
3191 
3192 	ctx = txq->ift_ctx;
3193 	sctx = ctx->ifc_sctx;
3194 	scctx = &ctx->ifc_softc_ctx;
3195 	segs = txq->ift_segs;
3196 	ntxd = txq->ift_size;
3197 	m_head = *m_headp;
3198 	map = NULL;
3199 
3200 	/*
3201 	 * If we're doing TSO the next descriptor to clean may be quite far ahead
3202 	 */
3203 	cidx = txq->ift_cidx;
3204 	pidx = txq->ift_pidx;
3205 	if (ctx->ifc_flags & IFC_PREFETCH) {
3206 		next = (cidx + CACHE_PTR_INCREMENT) & (ntxd-1);
3207 		if (!(ctx->ifc_flags & IFLIB_HAS_TXCQ)) {
3208 			next_txd = calc_next_txd(txq, cidx, 0);
3209 			prefetch(next_txd);
3210 		}
3211 
3212 		/* prefetch the next cache line of mbuf pointers and flags */
3213 		prefetch(&txq->ift_sds.ifsd_m[next]);
3214 		prefetch(&txq->ift_sds.ifsd_map[next]);
3215 		next = (cidx + CACHE_LINE_SIZE) & (ntxd-1);
3216 	}
3217 	map = txq->ift_sds.ifsd_map[pidx];
3218 	ifsd_m = txq->ift_sds.ifsd_m;
3219 
3220 	if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
3221 		buf_tag = txq->ift_tso_buf_tag;
3222 		max_segs = scctx->isc_tx_tso_segments_max;
3223 		map = txq->ift_sds.ifsd_tso_map[pidx];
3224 		MPASS(buf_tag != NULL);
3225 		MPASS(max_segs > 0);
3226 	} else {
3227 		buf_tag = txq->ift_buf_tag;
3228 		max_segs = scctx->isc_tx_nsegments;
3229 		map = txq->ift_sds.ifsd_map[pidx];
3230 	}
3231 	if ((sctx->isc_flags & IFLIB_NEED_ETHER_PAD) &&
3232 	    __predict_false(m_head->m_pkthdr.len < scctx->isc_min_frame_size)) {
3233 		err = iflib_ether_pad(ctx->ifc_dev, m_headp, scctx->isc_min_frame_size);
3234 		if (err) {
3235 			DBG_COUNTER_INC(encap_txd_encap_fail);
3236 			return err;
3237 		}
3238 	}
3239 	m_head = *m_headp;
3240 
3241 	pkt_info_zero(&pi);
3242 	pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG|M_BCAST|M_MCAST));
3243 	pi.ipi_pidx = pidx;
3244 	pi.ipi_qsidx = txq->ift_id;
3245 	pi.ipi_len = m_head->m_pkthdr.len;
3246 	pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags;
3247 	pi.ipi_vtag = (m_head->m_flags & M_VLANTAG) ? m_head->m_pkthdr.ether_vtag : 0;
3248 
3249 	/* deliberate bitwise OR to make one condition */
3250 	if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) {
3251 		if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0)) {
3252 			DBG_COUNTER_INC(encap_txd_encap_fail);
3253 			return (err);
3254 		}
3255 		m_head = *m_headp;
3256 	}
3257 
3258 retry:
3259 	err = bus_dmamap_load_mbuf_sg(buf_tag, map, m_head, segs, &nsegs,
3260 	    BUS_DMA_NOWAIT);
3261 defrag:
3262 	if (__predict_false(err)) {
3263 		switch (err) {
3264 		case EFBIG:
3265 			/* try collapse once and defrag once */
3266 			if (remap == 0) {
3267 				m_head = m_collapse(*m_headp, M_NOWAIT, max_segs);
3268 				/* try defrag if collapsing fails */
3269 				if (m_head == NULL)
3270 					remap++;
3271 			}
3272 			if (remap == 1) {
3273 				txq->ift_mbuf_defrag++;
3274 				m_head = m_defrag(*m_headp, M_NOWAIT);
3275 			}
3276 			remap++;
3277 			if (__predict_false(m_head == NULL))
3278 				goto defrag_failed;
3279 			*m_headp = m_head;
3280 			goto retry;
3281 			break;
3282 		case ENOMEM:
3283 			txq->ift_no_tx_dma_setup++;
3284 			break;
3285 		default:
3286 			txq->ift_no_tx_dma_setup++;
3287 			m_freem(*m_headp);
3288 			DBG_COUNTER_INC(tx_frees);
3289 			*m_headp = NULL;
3290 			break;
3291 		}
3292 		txq->ift_map_failed++;
3293 		DBG_COUNTER_INC(encap_load_mbuf_fail);
3294 		DBG_COUNTER_INC(encap_txd_encap_fail);
3295 		return (err);
3296 	}
3297 	ifsd_m[pidx] = m_head;
3298 	/*
3299 	 * XXX assumes a 1 to 1 relationship between segments and
3300 	 *        descriptors - this does not hold true on all drivers, e.g.
3301 	 *        cxgb
3302 	 */
3303 	if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) {
3304 		txq->ift_no_desc_avail++;
3305 		bus_dmamap_unload(buf_tag, map);
3306 		DBG_COUNTER_INC(encap_txq_avail_fail);
3307 		DBG_COUNTER_INC(encap_txd_encap_fail);
3308 		if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0)
3309 			GROUPTASK_ENQUEUE(&txq->ift_task);
3310 		return (ENOBUFS);
3311 	}
3312 	/*
3313 	 * On Intel cards we can greatly reduce the number of TX interrupts
3314 	 * we see by only setting report status on every Nth descriptor.
3315 	 * However, this also means that the driver will need to keep track
3316 	 * of the descriptors that RS was set on to check them for the DD bit.
3317 	 */
3318 	txq->ift_rs_pending += nsegs + 1;
3319 	if (txq->ift_rs_pending > TXQ_MAX_RS_DEFERRED(txq) ||
3320 	     iflib_no_tx_batch || (TXQ_AVAIL(txq) - nsegs) <= MAX_TX_DESC(ctx) + 2) {
3321 		pi.ipi_flags |= IPI_TX_INTR;
3322 		txq->ift_rs_pending = 0;
3323 	}
3324 
3325 	pi.ipi_segs = segs;
3326 	pi.ipi_nsegs = nsegs;
3327 
3328 	MPASS(pidx >= 0 && pidx < txq->ift_size);
3329 #ifdef PKT_DEBUG
3330 	print_pkt(&pi);
3331 #endif
3332 	if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) {
3333 		bus_dmamap_sync(buf_tag, map, BUS_DMASYNC_PREWRITE);
3334 		DBG_COUNTER_INC(tx_encap);
3335 		MPASS(pi.ipi_new_pidx < txq->ift_size);
3336 
3337 		ndesc = pi.ipi_new_pidx - pi.ipi_pidx;
3338 		if (pi.ipi_new_pidx < pi.ipi_pidx) {
3339 			ndesc += txq->ift_size;
3340 			txq->ift_gen = 1;
3341 		}
3342 		/*
3343 		 * drivers can need as many as
3344 		 * two sentinels
3345 		 */
3346 		MPASS(ndesc <= pi.ipi_nsegs + 2);
3347 		MPASS(pi.ipi_new_pidx != pidx);
3348 		MPASS(ndesc > 0);
3349 		txq->ift_in_use += ndesc;
3350 
3351 		/*
3352 		 * We update the last software descriptor again here because there may
3353 		 * be a sentinel and/or there may be more mbufs than segments
3354 		 */
3355 		txq->ift_pidx = pi.ipi_new_pidx;
3356 		txq->ift_npending += pi.ipi_ndescs;
3357 	} else {
3358 		*m_headp = m_head = iflib_remove_mbuf(txq);
3359 		if (err == EFBIG) {
3360 			txq->ift_txd_encap_efbig++;
3361 			if (remap < 2) {
3362 				remap = 1;
3363 				goto defrag;
3364 			}
3365 		}
3366 		goto defrag_failed;
3367 	}
3368 	/*
3369 	 * err can't possibly be non-zero here, so we don't neet to test it
3370 	 * to see if we need to DBG_COUNTER_INC(encap_txd_encap_fail).
3371 	 */
3372 	return (err);
3373 
3374 defrag_failed:
3375 	txq->ift_mbuf_defrag_failed++;
3376 	txq->ift_map_failed++;
3377 	m_freem(*m_headp);
3378 	DBG_COUNTER_INC(tx_frees);
3379 	*m_headp = NULL;
3380 	DBG_COUNTER_INC(encap_txd_encap_fail);
3381 	return (ENOMEM);
3382 }
3383 
3384 static void
3385 iflib_tx_desc_free(iflib_txq_t txq, int n)
3386 {
3387 	uint32_t qsize, cidx, mask, gen;
3388 	struct mbuf *m, **ifsd_m;
3389 	bool do_prefetch;
3390 
3391 	cidx = txq->ift_cidx;
3392 	gen = txq->ift_gen;
3393 	qsize = txq->ift_size;
3394 	mask = qsize-1;
3395 	ifsd_m = txq->ift_sds.ifsd_m;
3396 	do_prefetch = (txq->ift_ctx->ifc_flags & IFC_PREFETCH);
3397 
3398 	while (n-- > 0) {
3399 		if (do_prefetch) {
3400 			prefetch(ifsd_m[(cidx + 3) & mask]);
3401 			prefetch(ifsd_m[(cidx + 4) & mask]);
3402 		}
3403 		if ((m = ifsd_m[cidx]) != NULL) {
3404 			prefetch(&ifsd_m[(cidx + CACHE_PTR_INCREMENT) & mask]);
3405 			if (m->m_pkthdr.csum_flags & CSUM_TSO) {
3406 				bus_dmamap_sync(txq->ift_tso_buf_tag,
3407 				    txq->ift_sds.ifsd_tso_map[cidx],
3408 				    BUS_DMASYNC_POSTWRITE);
3409 				bus_dmamap_unload(txq->ift_tso_buf_tag,
3410 				    txq->ift_sds.ifsd_tso_map[cidx]);
3411 			} else {
3412 				bus_dmamap_sync(txq->ift_buf_tag,
3413 				    txq->ift_sds.ifsd_map[cidx],
3414 				    BUS_DMASYNC_POSTWRITE);
3415 				bus_dmamap_unload(txq->ift_buf_tag,
3416 				    txq->ift_sds.ifsd_map[cidx]);
3417 			}
3418 			/* XXX we don't support any drivers that batch packets yet */
3419 			MPASS(m->m_nextpkt == NULL);
3420 			m_freem(m);
3421 			ifsd_m[cidx] = NULL;
3422 #if MEMORY_LOGGING
3423 			txq->ift_dequeued++;
3424 #endif
3425 			DBG_COUNTER_INC(tx_frees);
3426 		}
3427 		if (__predict_false(++cidx == qsize)) {
3428 			cidx = 0;
3429 			gen = 0;
3430 		}
3431 	}
3432 	txq->ift_cidx = cidx;
3433 	txq->ift_gen = gen;
3434 }
3435 
3436 static __inline int
3437 iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh)
3438 {
3439 	int reclaim;
3440 	if_ctx_t ctx = txq->ift_ctx;
3441 
3442 	KASSERT(thresh >= 0, ("invalid threshold to reclaim"));
3443 	MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size);
3444 
3445 	/*
3446 	 * Need a rate-limiting check so that this isn't called every time
3447 	 */
3448 	iflib_tx_credits_update(ctx, txq);
3449 	reclaim = DESC_RECLAIMABLE(txq);
3450 
3451 	if (reclaim <= thresh /* + MAX_TX_DESC(txq->ift_ctx) */) {
3452 #ifdef INVARIANTS
3453 		if (iflib_verbose_debug) {
3454 			printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __FUNCTION__,
3455 			       txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments,
3456 			       reclaim, thresh);
3457 
3458 		}
3459 #endif
3460 		return (0);
3461 	}
3462 	iflib_tx_desc_free(txq, reclaim);
3463 	txq->ift_cleaned += reclaim;
3464 	txq->ift_in_use -= reclaim;
3465 
3466 	return (reclaim);
3467 }
3468 
3469 static struct mbuf **
3470 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset, int remaining)
3471 {
3472 	int next, size;
3473 	struct mbuf **items;
3474 
3475 	size = r->size;
3476 	next = (cidx + CACHE_PTR_INCREMENT) & (size-1);
3477 	items = __DEVOLATILE(struct mbuf **, &r->items[0]);
3478 
3479 	prefetch(items[(cidx + offset) & (size-1)]);
3480 	if (remaining > 1) {
3481 		prefetch2cachelines(&items[next]);
3482 		prefetch2cachelines(items[(cidx + offset + 1) & (size-1)]);
3483 		prefetch2cachelines(items[(cidx + offset + 2) & (size-1)]);
3484 		prefetch2cachelines(items[(cidx + offset + 3) & (size-1)]);
3485 	}
3486 	return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (size-1)]));
3487 }
3488 
3489 static void
3490 iflib_txq_check_drain(iflib_txq_t txq, int budget)
3491 {
3492 
3493 	ifmp_ring_check_drainage(txq->ift_br, budget);
3494 }
3495 
3496 static uint32_t
3497 iflib_txq_can_drain(struct ifmp_ring *r)
3498 {
3499 	iflib_txq_t txq = r->cookie;
3500 	if_ctx_t ctx = txq->ift_ctx;
3501 
3502 	if (TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2)
3503 		return (1);
3504 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3505 	    BUS_DMASYNC_POSTREAD);
3506 	return (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id,
3507 	    false));
3508 }
3509 
3510 static uint32_t
3511 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3512 {
3513 	iflib_txq_t txq = r->cookie;
3514 	if_ctx_t ctx = txq->ift_ctx;
3515 	struct ifnet *ifp = ctx->ifc_ifp;
3516 	struct mbuf **mp, *m;
3517 	int i, count, consumed, pkt_sent, bytes_sent, mcast_sent, avail;
3518 	int reclaimed, err, in_use_prev, desc_used;
3519 	bool do_prefetch, ring, rang;
3520 
3521 	if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) ||
3522 			    !LINK_ACTIVE(ctx))) {
3523 		DBG_COUNTER_INC(txq_drain_notready);
3524 		return (0);
3525 	}
3526 	reclaimed = iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
3527 	rang = iflib_txd_db_check(ctx, txq, reclaimed, txq->ift_in_use);
3528 	avail = IDXDIFF(pidx, cidx, r->size);
3529 	if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) {
3530 		DBG_COUNTER_INC(txq_drain_flushing);
3531 		for (i = 0; i < avail; i++) {
3532 			if (__predict_true(r->items[(cidx + i) & (r->size-1)] != (void *)txq))
3533 				m_free(r->items[(cidx + i) & (r->size-1)]);
3534 			r->items[(cidx + i) & (r->size-1)] = NULL;
3535 		}
3536 		return (avail);
3537 	}
3538 
3539 	if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3540 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3541 		CALLOUT_LOCK(txq);
3542 		callout_stop(&txq->ift_timer);
3543 		CALLOUT_UNLOCK(txq);
3544 		DBG_COUNTER_INC(txq_drain_oactive);
3545 		return (0);
3546 	}
3547 	if (reclaimed)
3548 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3549 	consumed = mcast_sent = bytes_sent = pkt_sent = 0;
3550 	count = MIN(avail, TX_BATCH_SIZE);
3551 #ifdef INVARIANTS
3552 	if (iflib_verbose_debug)
3553 		printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __FUNCTION__,
3554 		       avail, ctx->ifc_flags, TXQ_AVAIL(txq));
3555 #endif
3556 	do_prefetch = (ctx->ifc_flags & IFC_PREFETCH);
3557 	avail = TXQ_AVAIL(txq);
3558 	err = 0;
3559 	for (desc_used = i = 0; i < count && avail > MAX_TX_DESC(ctx) + 2; i++) {
3560 		int rem = do_prefetch ? count - i : 0;
3561 
3562 		mp = _ring_peek_one(r, cidx, i, rem);
3563 		MPASS(mp != NULL && *mp != NULL);
3564 		if (__predict_false(*mp == (struct mbuf *)txq)) {
3565 			consumed++;
3566 			reclaimed++;
3567 			continue;
3568 		}
3569 		in_use_prev = txq->ift_in_use;
3570 		err = iflib_encap(txq, mp);
3571 		if (__predict_false(err)) {
3572 			/* no room - bail out */
3573 			if (err == ENOBUFS)
3574 				break;
3575 			consumed++;
3576 			/* we can't send this packet - skip it */
3577 			continue;
3578 		}
3579 		consumed++;
3580 		pkt_sent++;
3581 		m = *mp;
3582 		DBG_COUNTER_INC(tx_sent);
3583 		bytes_sent += m->m_pkthdr.len;
3584 		mcast_sent += !!(m->m_flags & M_MCAST);
3585 		avail = TXQ_AVAIL(txq);
3586 
3587 		txq->ift_db_pending += (txq->ift_in_use - in_use_prev);
3588 		desc_used += (txq->ift_in_use - in_use_prev);
3589 		ETHER_BPF_MTAP(ifp, m);
3590 		if (__predict_false(!(ifp->if_drv_flags & IFF_DRV_RUNNING)))
3591 			break;
3592 		rang = iflib_txd_db_check(ctx, txq, false, in_use_prev);
3593 	}
3594 
3595 	/* deliberate use of bitwise or to avoid gratuitous short-circuit */
3596 	ring = rang ? false  : (iflib_min_tx_latency | err) || (TXQ_AVAIL(txq) < MAX_TX_DESC(ctx));
3597 	iflib_txd_db_check(ctx, txq, ring, txq->ift_in_use);
3598 	if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
3599 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
3600 	if (mcast_sent)
3601 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
3602 #ifdef INVARIANTS
3603 	if (iflib_verbose_debug)
3604 		printf("consumed=%d\n", consumed);
3605 #endif
3606 	return (consumed);
3607 }
3608 
3609 static uint32_t
3610 iflib_txq_drain_always(struct ifmp_ring *r)
3611 {
3612 	return (1);
3613 }
3614 
3615 static uint32_t
3616 iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3617 {
3618 	int i, avail;
3619 	struct mbuf **mp;
3620 	iflib_txq_t txq;
3621 
3622 	txq = r->cookie;
3623 
3624 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3625 	CALLOUT_LOCK(txq);
3626 	callout_stop(&txq->ift_timer);
3627 	CALLOUT_UNLOCK(txq);
3628 
3629 	avail = IDXDIFF(pidx, cidx, r->size);
3630 	for (i = 0; i < avail; i++) {
3631 		mp = _ring_peek_one(r, cidx, i, avail - i);
3632 		if (__predict_false(*mp == (struct mbuf *)txq))
3633 			continue;
3634 		m_freem(*mp);
3635 		DBG_COUNTER_INC(tx_frees);
3636 	}
3637 	MPASS(ifmp_ring_is_stalled(r) == 0);
3638 	return (avail);
3639 }
3640 
3641 static void
3642 iflib_ifmp_purge(iflib_txq_t txq)
3643 {
3644 	struct ifmp_ring *r;
3645 
3646 	r = txq->ift_br;
3647 	r->drain = iflib_txq_drain_free;
3648 	r->can_drain = iflib_txq_drain_always;
3649 
3650 	ifmp_ring_check_drainage(r, r->size);
3651 
3652 	r->drain = iflib_txq_drain;
3653 	r->can_drain = iflib_txq_can_drain;
3654 }
3655 
3656 static void
3657 _task_fn_tx(void *context)
3658 {
3659 	iflib_txq_t txq = context;
3660 	if_ctx_t ctx = txq->ift_ctx;
3661 #if defined(ALTQ) || defined(DEV_NETMAP)
3662 	if_t ifp = ctx->ifc_ifp;
3663 #endif
3664 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
3665 
3666 #ifdef IFLIB_DIAGNOSTICS
3667 	txq->ift_cpu_exec_count[curcpu]++;
3668 #endif
3669 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3670 		return;
3671 #ifdef DEV_NETMAP
3672 	if (if_getcapenable(ifp) & IFCAP_NETMAP) {
3673 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3674 		    BUS_DMASYNC_POSTREAD);
3675 		if (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false))
3676 			netmap_tx_irq(ifp, txq->ift_id);
3677 		IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3678 		return;
3679 	}
3680 #endif
3681 #ifdef ALTQ
3682 	if (ALTQ_IS_ENABLED(&ifp->if_snd))
3683 		iflib_altq_if_start(ifp);
3684 #endif
3685 	if (txq->ift_db_pending)
3686 		ifmp_ring_enqueue(txq->ift_br, (void **)&txq, 1, TX_BATCH_SIZE, abdicate);
3687 	else if (!abdicate)
3688 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3689 	/*
3690 	 * When abdicating, we always need to check drainage, not just when we don't enqueue
3691 	 */
3692 	if (abdicate)
3693 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3694 	if (ctx->ifc_flags & IFC_LEGACY)
3695 		IFDI_INTR_ENABLE(ctx);
3696 	else {
3697 #ifdef INVARIANTS
3698 		int rc =
3699 #endif
3700 			IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3701 			KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver"));
3702 	}
3703 }
3704 
3705 static void
3706 _task_fn_rx(void *context)
3707 {
3708 	iflib_rxq_t rxq = context;
3709 	if_ctx_t ctx = rxq->ifr_ctx;
3710 	bool more;
3711 	uint16_t budget;
3712 
3713 #ifdef IFLIB_DIAGNOSTICS
3714 	rxq->ifr_cpu_exec_count[curcpu]++;
3715 #endif
3716 	DBG_COUNTER_INC(task_fn_rxs);
3717 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3718 		return;
3719 	more = true;
3720 #ifdef DEV_NETMAP
3721 	if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP) {
3722 		u_int work = 0;
3723 		if (netmap_rx_irq(ctx->ifc_ifp, rxq->ifr_id, &work)) {
3724 			more = false;
3725 		}
3726 	}
3727 #endif
3728 	budget = ctx->ifc_sysctl_rx_budget;
3729 	if (budget == 0)
3730 		budget = 16;	/* XXX */
3731 	if (more == false || (more = iflib_rxeof(rxq, budget)) == false) {
3732 		if (ctx->ifc_flags & IFC_LEGACY)
3733 			IFDI_INTR_ENABLE(ctx);
3734 		else {
3735 #ifdef INVARIANTS
3736 			int rc =
3737 #endif
3738 				IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
3739 			KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver"));
3740 			DBG_COUNTER_INC(rx_intr_enables);
3741 		}
3742 	}
3743 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3744 		return;
3745 	if (more)
3746 		GROUPTASK_ENQUEUE(&rxq->ifr_task);
3747 }
3748 
3749 static void
3750 _task_fn_admin(void *context)
3751 {
3752 	if_ctx_t ctx = context;
3753 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
3754 	iflib_txq_t txq;
3755 	int i;
3756 	bool oactive, running, do_reset, do_watchdog, in_detach;
3757 	uint32_t reset_on = hz / 2;
3758 
3759 	STATE_LOCK(ctx);
3760 	running = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING);
3761 	oactive = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE);
3762 	do_reset = (ctx->ifc_flags & IFC_DO_RESET);
3763 	do_watchdog = (ctx->ifc_flags & IFC_DO_WATCHDOG);
3764 	in_detach = (ctx->ifc_flags & IFC_IN_DETACH);
3765 	ctx->ifc_flags &= ~(IFC_DO_RESET|IFC_DO_WATCHDOG);
3766 	STATE_UNLOCK(ctx);
3767 
3768 	if ((!running && !oactive) && !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
3769 		return;
3770 	if (in_detach)
3771 		return;
3772 
3773 	CTX_LOCK(ctx);
3774 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3775 		CALLOUT_LOCK(txq);
3776 		callout_stop(&txq->ift_timer);
3777 		CALLOUT_UNLOCK(txq);
3778 	}
3779 	if (do_watchdog) {
3780 		ctx->ifc_watchdog_events++;
3781 		IFDI_WATCHDOG_RESET(ctx);
3782 	}
3783 	IFDI_UPDATE_ADMIN_STATUS(ctx);
3784 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3785 #ifdef DEV_NETMAP
3786 		reset_on = hz / 2;
3787 		if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP)
3788 			iflib_netmap_timer_adjust(ctx, txq, &reset_on);
3789 #endif
3790 		callout_reset_on(&txq->ift_timer, reset_on, iflib_timer, txq, txq->ift_timer.c_cpu);
3791 	}
3792 	IFDI_LINK_INTR_ENABLE(ctx);
3793 	if (do_reset)
3794 		iflib_if_init_locked(ctx);
3795 	CTX_UNLOCK(ctx);
3796 
3797 	if (LINK_ACTIVE(ctx) == 0)
3798 		return;
3799 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3800 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
3801 }
3802 
3803 
3804 static void
3805 _task_fn_iov(void *context)
3806 {
3807 	if_ctx_t ctx = context;
3808 
3809 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING) &&
3810 	    !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
3811 		return;
3812 
3813 	CTX_LOCK(ctx);
3814 	IFDI_VFLR_HANDLE(ctx);
3815 	CTX_UNLOCK(ctx);
3816 }
3817 
3818 static int
3819 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)
3820 {
3821 	int err;
3822 	if_int_delay_info_t info;
3823 	if_ctx_t ctx;
3824 
3825 	info = (if_int_delay_info_t)arg1;
3826 	ctx = info->iidi_ctx;
3827 	info->iidi_req = req;
3828 	info->iidi_oidp = oidp;
3829 	CTX_LOCK(ctx);
3830 	err = IFDI_SYSCTL_INT_DELAY(ctx, info);
3831 	CTX_UNLOCK(ctx);
3832 	return (err);
3833 }
3834 
3835 /*********************************************************************
3836  *
3837  *  IFNET FUNCTIONS
3838  *
3839  **********************************************************************/
3840 
3841 static void
3842 iflib_if_init_locked(if_ctx_t ctx)
3843 {
3844 	iflib_stop(ctx);
3845 	iflib_init_locked(ctx);
3846 }
3847 
3848 
3849 static void
3850 iflib_if_init(void *arg)
3851 {
3852 	if_ctx_t ctx = arg;
3853 
3854 	CTX_LOCK(ctx);
3855 	iflib_if_init_locked(ctx);
3856 	CTX_UNLOCK(ctx);
3857 }
3858 
3859 static int
3860 iflib_if_transmit(if_t ifp, struct mbuf *m)
3861 {
3862 	if_ctx_t	ctx = if_getsoftc(ifp);
3863 
3864 	iflib_txq_t txq;
3865 	int err, qidx;
3866 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
3867 
3868 	if (__predict_false((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) {
3869 		DBG_COUNTER_INC(tx_frees);
3870 		m_freem(m);
3871 		return (ENOBUFS);
3872 	}
3873 
3874 	MPASS(m->m_nextpkt == NULL);
3875 	/* ALTQ-enabled interfaces always use queue 0. */
3876 	qidx = 0;
3877 	if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m) && !ALTQ_IS_ENABLED(&ifp->if_snd))
3878 		qidx = QIDX(ctx, m);
3879 	/*
3880 	 * XXX calculate buf_ring based on flowid (divvy up bits?)
3881 	 */
3882 	txq = &ctx->ifc_txqs[qidx];
3883 
3884 #ifdef DRIVER_BACKPRESSURE
3885 	if (txq->ift_closed) {
3886 		while (m != NULL) {
3887 			next = m->m_nextpkt;
3888 			m->m_nextpkt = NULL;
3889 			m_freem(m);
3890 			DBG_COUNTER_INC(tx_frees);
3891 			m = next;
3892 		}
3893 		return (ENOBUFS);
3894 	}
3895 #endif
3896 #ifdef notyet
3897 	qidx = count = 0;
3898 	mp = marr;
3899 	next = m;
3900 	do {
3901 		count++;
3902 		next = next->m_nextpkt;
3903 	} while (next != NULL);
3904 
3905 	if (count > nitems(marr))
3906 		if ((mp = malloc(count*sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) {
3907 			/* XXX check nextpkt */
3908 			m_freem(m);
3909 			/* XXX simplify for now */
3910 			DBG_COUNTER_INC(tx_frees);
3911 			return (ENOBUFS);
3912 		}
3913 	for (next = m, i = 0; next != NULL; i++) {
3914 		mp[i] = next;
3915 		next = next->m_nextpkt;
3916 		mp[i]->m_nextpkt = NULL;
3917 	}
3918 #endif
3919 	DBG_COUNTER_INC(tx_seen);
3920 	err = ifmp_ring_enqueue(txq->ift_br, (void **)&m, 1, TX_BATCH_SIZE, abdicate);
3921 
3922 	if (abdicate)
3923 		GROUPTASK_ENQUEUE(&txq->ift_task);
3924  	if (err) {
3925 		if (!abdicate)
3926 			GROUPTASK_ENQUEUE(&txq->ift_task);
3927 		/* support forthcoming later */
3928 #ifdef DRIVER_BACKPRESSURE
3929 		txq->ift_closed = TRUE;
3930 #endif
3931 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3932 		m_freem(m);
3933 		DBG_COUNTER_INC(tx_frees);
3934 	}
3935 
3936 	return (err);
3937 }
3938 
3939 #ifdef ALTQ
3940 /*
3941  * The overall approach to integrating iflib with ALTQ is to continue to use
3942  * the iflib mp_ring machinery between the ALTQ queue(s) and the hardware
3943  * ring.  Technically, when using ALTQ, queueing to an intermediate mp_ring
3944  * is redundant/unnecessary, but doing so minimizes the amount of
3945  * ALTQ-specific code required in iflib.  It is assumed that the overhead of
3946  * redundantly queueing to an intermediate mp_ring is swamped by the
3947  * performance limitations inherent in using ALTQ.
3948  *
3949  * When ALTQ support is compiled in, all iflib drivers will use a transmit
3950  * routine, iflib_altq_if_transmit(), that checks if ALTQ is enabled for the
3951  * given interface.  If ALTQ is enabled for an interface, then all
3952  * transmitted packets for that interface will be submitted to the ALTQ
3953  * subsystem via IFQ_ENQUEUE().  We don't use the legacy if_transmit()
3954  * implementation because it uses IFQ_HANDOFF(), which will duplicatively
3955  * update stats that the iflib machinery handles, and which is sensitve to
3956  * the disused IFF_DRV_OACTIVE flag.  Additionally, iflib_altq_if_start()
3957  * will be installed as the start routine for use by ALTQ facilities that
3958  * need to trigger queue drains on a scheduled basis.
3959  *
3960  */
3961 static void
3962 iflib_altq_if_start(if_t ifp)
3963 {
3964 	struct ifaltq *ifq = &ifp->if_snd;
3965 	struct mbuf *m;
3966 
3967 	IFQ_LOCK(ifq);
3968 	IFQ_DEQUEUE_NOLOCK(ifq, m);
3969 	while (m != NULL) {
3970 		iflib_if_transmit(ifp, m);
3971 		IFQ_DEQUEUE_NOLOCK(ifq, m);
3972 	}
3973 	IFQ_UNLOCK(ifq);
3974 }
3975 
3976 static int
3977 iflib_altq_if_transmit(if_t ifp, struct mbuf *m)
3978 {
3979 	int err;
3980 
3981 	if (ALTQ_IS_ENABLED(&ifp->if_snd)) {
3982 		IFQ_ENQUEUE(&ifp->if_snd, m, err);
3983 		if (err == 0)
3984 			iflib_altq_if_start(ifp);
3985 	} else
3986 		err = iflib_if_transmit(ifp, m);
3987 
3988 	return (err);
3989 }
3990 #endif /* ALTQ */
3991 
3992 static void
3993 iflib_if_qflush(if_t ifp)
3994 {
3995 	if_ctx_t ctx = if_getsoftc(ifp);
3996 	iflib_txq_t txq = ctx->ifc_txqs;
3997 	int i;
3998 
3999 	STATE_LOCK(ctx);
4000 	ctx->ifc_flags |= IFC_QFLUSH;
4001 	STATE_UNLOCK(ctx);
4002 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
4003 		while (!(ifmp_ring_is_idle(txq->ift_br) || ifmp_ring_is_stalled(txq->ift_br)))
4004 			iflib_txq_check_drain(txq, 0);
4005 	STATE_LOCK(ctx);
4006 	ctx->ifc_flags &= ~IFC_QFLUSH;
4007 	STATE_UNLOCK(ctx);
4008 
4009 	/*
4010 	 * When ALTQ is enabled, this will also take care of purging the
4011 	 * ALTQ queue(s).
4012 	 */
4013 	if_qflush(ifp);
4014 }
4015 
4016 
4017 #define IFCAP_FLAGS (IFCAP_HWCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \
4018 		     IFCAP_TSO | IFCAP_VLAN_HWTAGGING | IFCAP_HWSTATS | \
4019 		     IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | \
4020 		     IFCAP_VLAN_HWTSO | IFCAP_VLAN_HWCSUM)
4021 
4022 static int
4023 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data)
4024 {
4025 	if_ctx_t ctx = if_getsoftc(ifp);
4026 	struct ifreq	*ifr = (struct ifreq *)data;
4027 #if defined(INET) || defined(INET6)
4028 	struct ifaddr	*ifa = (struct ifaddr *)data;
4029 #endif
4030 	bool		avoid_reset = FALSE;
4031 	int		err = 0, reinit = 0, bits;
4032 
4033 	switch (command) {
4034 	case SIOCSIFADDR:
4035 #ifdef INET
4036 		if (ifa->ifa_addr->sa_family == AF_INET)
4037 			avoid_reset = TRUE;
4038 #endif
4039 #ifdef INET6
4040 		if (ifa->ifa_addr->sa_family == AF_INET6)
4041 			avoid_reset = TRUE;
4042 #endif
4043 		/*
4044 		** Calling init results in link renegotiation,
4045 		** so we avoid doing it when possible.
4046 		*/
4047 		if (avoid_reset) {
4048 			if_setflagbits(ifp, IFF_UP,0);
4049 			if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
4050 				reinit = 1;
4051 #ifdef INET
4052 			if (!(if_getflags(ifp) & IFF_NOARP))
4053 				arp_ifinit(ifp, ifa);
4054 #endif
4055 		} else
4056 			err = ether_ioctl(ifp, command, data);
4057 		break;
4058 	case SIOCSIFMTU:
4059 		CTX_LOCK(ctx);
4060 		if (ifr->ifr_mtu == if_getmtu(ifp)) {
4061 			CTX_UNLOCK(ctx);
4062 			break;
4063 		}
4064 		bits = if_getdrvflags(ifp);
4065 		/* stop the driver and free any clusters before proceeding */
4066 		iflib_stop(ctx);
4067 
4068 		if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) {
4069 			STATE_LOCK(ctx);
4070 			if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size)
4071 				ctx->ifc_flags |= IFC_MULTISEG;
4072 			else
4073 				ctx->ifc_flags &= ~IFC_MULTISEG;
4074 			STATE_UNLOCK(ctx);
4075 			err = if_setmtu(ifp, ifr->ifr_mtu);
4076 		}
4077 		iflib_init_locked(ctx);
4078 		STATE_LOCK(ctx);
4079 		if_setdrvflags(ifp, bits);
4080 		STATE_UNLOCK(ctx);
4081 		CTX_UNLOCK(ctx);
4082 		break;
4083 	case SIOCSIFFLAGS:
4084 		CTX_LOCK(ctx);
4085 		if (if_getflags(ifp) & IFF_UP) {
4086 			if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4087 				if ((if_getflags(ifp) ^ ctx->ifc_if_flags) &
4088 				    (IFF_PROMISC | IFF_ALLMULTI)) {
4089 					err = IFDI_PROMISC_SET(ctx, if_getflags(ifp));
4090 				}
4091 			} else
4092 				reinit = 1;
4093 		} else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4094 			iflib_stop(ctx);
4095 		}
4096 		ctx->ifc_if_flags = if_getflags(ifp);
4097 		CTX_UNLOCK(ctx);
4098 		break;
4099 	case SIOCADDMULTI:
4100 	case SIOCDELMULTI:
4101 		if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4102 			CTX_LOCK(ctx);
4103 			IFDI_INTR_DISABLE(ctx);
4104 			IFDI_MULTI_SET(ctx);
4105 			IFDI_INTR_ENABLE(ctx);
4106 			CTX_UNLOCK(ctx);
4107 		}
4108 		break;
4109 	case SIOCSIFMEDIA:
4110 		CTX_LOCK(ctx);
4111 		IFDI_MEDIA_SET(ctx);
4112 		CTX_UNLOCK(ctx);
4113 		/* falls thru */
4114 	case SIOCGIFMEDIA:
4115 	case SIOCGIFXMEDIA:
4116 		err = ifmedia_ioctl(ifp, ifr, &ctx->ifc_media, command);
4117 		break;
4118 	case SIOCGI2C:
4119 	{
4120 		struct ifi2creq i2c;
4121 
4122 		err = copyin(ifr_data_get_ptr(ifr), &i2c, sizeof(i2c));
4123 		if (err != 0)
4124 			break;
4125 		if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) {
4126 			err = EINVAL;
4127 			break;
4128 		}
4129 		if (i2c.len > sizeof(i2c.data)) {
4130 			err = EINVAL;
4131 			break;
4132 		}
4133 
4134 		if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0)
4135 			err = copyout(&i2c, ifr_data_get_ptr(ifr),
4136 			    sizeof(i2c));
4137 		break;
4138 	}
4139 	case SIOCSIFCAP:
4140 	{
4141 		int mask, setmask, oldmask;
4142 
4143 		oldmask = if_getcapenable(ifp);
4144 		mask = ifr->ifr_reqcap ^ oldmask;
4145 		mask &= ctx->ifc_softc_ctx.isc_capabilities;
4146 		setmask = 0;
4147 #ifdef TCP_OFFLOAD
4148 		setmask |= mask & (IFCAP_TOE4|IFCAP_TOE6);
4149 #endif
4150 		setmask |= (mask & IFCAP_FLAGS);
4151 		setmask |= (mask & IFCAP_WOL);
4152 
4153 		/*
4154 		 * If any RX csum has changed, change all the ones that
4155 		 * are supported by the driver.
4156 		 */
4157 		if (setmask & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6)) {
4158 			setmask |= ctx->ifc_softc_ctx.isc_capabilities &
4159 			    (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6);
4160 		}
4161 
4162 		/*
4163 		 * want to ensure that traffic has stopped before we change any of the flags
4164 		 */
4165 		if (setmask) {
4166 			CTX_LOCK(ctx);
4167 			bits = if_getdrvflags(ifp);
4168 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4169 				iflib_stop(ctx);
4170 			STATE_LOCK(ctx);
4171 			if_togglecapenable(ifp, setmask);
4172 			STATE_UNLOCK(ctx);
4173 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4174 				iflib_init_locked(ctx);
4175 			STATE_LOCK(ctx);
4176 			if_setdrvflags(ifp, bits);
4177 			STATE_UNLOCK(ctx);
4178 			CTX_UNLOCK(ctx);
4179 		}
4180 		if_vlancap(ifp);
4181 		break;
4182 	}
4183 	case SIOCGPRIVATE_0:
4184 	case SIOCSDRVSPEC:
4185 	case SIOCGDRVSPEC:
4186 		CTX_LOCK(ctx);
4187 		err = IFDI_PRIV_IOCTL(ctx, command, data);
4188 		CTX_UNLOCK(ctx);
4189 		break;
4190 	default:
4191 		err = ether_ioctl(ifp, command, data);
4192 		break;
4193 	}
4194 	if (reinit)
4195 		iflib_if_init(ctx);
4196 	return (err);
4197 }
4198 
4199 static uint64_t
4200 iflib_if_get_counter(if_t ifp, ift_counter cnt)
4201 {
4202 	if_ctx_t ctx = if_getsoftc(ifp);
4203 
4204 	return (IFDI_GET_COUNTER(ctx, cnt));
4205 }
4206 
4207 /*********************************************************************
4208  *
4209  *  OTHER FUNCTIONS EXPORTED TO THE STACK
4210  *
4211  **********************************************************************/
4212 
4213 static void
4214 iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag)
4215 {
4216 	if_ctx_t ctx = if_getsoftc(ifp);
4217 
4218 	if ((void *)ctx != arg)
4219 		return;
4220 
4221 	if ((vtag == 0) || (vtag > 4095))
4222 		return;
4223 
4224 	CTX_LOCK(ctx);
4225 	IFDI_VLAN_REGISTER(ctx, vtag);
4226 	/* Re-init to load the changes */
4227 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
4228 		iflib_if_init_locked(ctx);
4229 	CTX_UNLOCK(ctx);
4230 }
4231 
4232 static void
4233 iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag)
4234 {
4235 	if_ctx_t ctx = if_getsoftc(ifp);
4236 
4237 	if ((void *)ctx != arg)
4238 		return;
4239 
4240 	if ((vtag == 0) || (vtag > 4095))
4241 		return;
4242 
4243 	CTX_LOCK(ctx);
4244 	IFDI_VLAN_UNREGISTER(ctx, vtag);
4245 	/* Re-init to load the changes */
4246 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
4247 		iflib_if_init_locked(ctx);
4248 	CTX_UNLOCK(ctx);
4249 }
4250 
4251 static void
4252 iflib_led_func(void *arg, int onoff)
4253 {
4254 	if_ctx_t ctx = arg;
4255 
4256 	CTX_LOCK(ctx);
4257 	IFDI_LED_FUNC(ctx, onoff);
4258 	CTX_UNLOCK(ctx);
4259 }
4260 
4261 /*********************************************************************
4262  *
4263  *  BUS FUNCTION DEFINITIONS
4264  *
4265  **********************************************************************/
4266 
4267 int
4268 iflib_device_probe(device_t dev)
4269 {
4270 	pci_vendor_info_t *ent;
4271 
4272 	uint16_t	pci_vendor_id, pci_device_id;
4273 	uint16_t	pci_subvendor_id, pci_subdevice_id;
4274 	uint16_t	pci_rev_id;
4275 	if_shared_ctx_t sctx;
4276 
4277 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4278 		return (ENOTSUP);
4279 
4280 	pci_vendor_id = pci_get_vendor(dev);
4281 	pci_device_id = pci_get_device(dev);
4282 	pci_subvendor_id = pci_get_subvendor(dev);
4283 	pci_subdevice_id = pci_get_subdevice(dev);
4284 	pci_rev_id = pci_get_revid(dev);
4285 	if (sctx->isc_parse_devinfo != NULL)
4286 		sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id);
4287 
4288 	ent = sctx->isc_vendor_info;
4289 	while (ent->pvi_vendor_id != 0) {
4290 		if (pci_vendor_id != ent->pvi_vendor_id) {
4291 			ent++;
4292 			continue;
4293 		}
4294 		if ((pci_device_id == ent->pvi_device_id) &&
4295 		    ((pci_subvendor_id == ent->pvi_subvendor_id) ||
4296 		     (ent->pvi_subvendor_id == 0)) &&
4297 		    ((pci_subdevice_id == ent->pvi_subdevice_id) ||
4298 		     (ent->pvi_subdevice_id == 0)) &&
4299 		    ((pci_rev_id == ent->pvi_rev_id) ||
4300 		     (ent->pvi_rev_id == 0))) {
4301 
4302 			device_set_desc_copy(dev, ent->pvi_name);
4303 			/* this needs to be changed to zero if the bus probing code
4304 			 * ever stops re-probing on best match because the sctx
4305 			 * may have its values over written by register calls
4306 			 * in subsequent probes
4307 			 */
4308 			return (BUS_PROBE_DEFAULT);
4309 		}
4310 		ent++;
4311 	}
4312 	return (ENXIO);
4313 }
4314 
4315 static void
4316 iflib_reset_qvalues(if_ctx_t ctx)
4317 {
4318 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4319 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4320 	device_t dev = ctx->ifc_dev;
4321 	int i;
4322 
4323 	scctx->isc_txrx_budget_bytes_max = IFLIB_MAX_TX_BYTES;
4324 	scctx->isc_tx_qdepth = IFLIB_DEFAULT_TX_QDEPTH;
4325 	/*
4326 	 * XXX sanity check that ntxd & nrxd are a power of 2
4327 	 */
4328 	if (ctx->ifc_sysctl_ntxqs != 0)
4329 		scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs;
4330 	if (ctx->ifc_sysctl_nrxqs != 0)
4331 		scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs;
4332 
4333 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4334 		if (ctx->ifc_sysctl_ntxds[i] != 0)
4335 			scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i];
4336 		else
4337 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4338 	}
4339 
4340 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4341 		if (ctx->ifc_sysctl_nrxds[i] != 0)
4342 			scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i];
4343 		else
4344 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4345 	}
4346 
4347 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4348 		if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) {
4349 			device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n",
4350 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]);
4351 			scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i];
4352 		}
4353 		if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) {
4354 			device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n",
4355 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]);
4356 			scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i];
4357 		}
4358 	}
4359 
4360 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4361 		if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) {
4362 			device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n",
4363 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]);
4364 			scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i];
4365 		}
4366 		if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) {
4367 			device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n",
4368 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]);
4369 			scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i];
4370 		}
4371 	}
4372 }
4373 
4374 int
4375 iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp)
4376 {
4377 	int err, rid, msix;
4378 	if_ctx_t ctx;
4379 	if_t ifp;
4380 	if_softc_ctx_t scctx;
4381 	int i;
4382 	uint16_t main_txq;
4383 	uint16_t main_rxq;
4384 
4385 
4386 	ctx = malloc(sizeof(* ctx), M_IFLIB, M_WAITOK|M_ZERO);
4387 
4388 	if (sc == NULL) {
4389 		sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4390 		device_set_softc(dev, ctx);
4391 		ctx->ifc_flags |= IFC_SC_ALLOCATED;
4392 	}
4393 
4394 	ctx->ifc_sctx = sctx;
4395 	ctx->ifc_dev = dev;
4396 	ctx->ifc_softc = sc;
4397 
4398 	if ((err = iflib_register(ctx)) != 0) {
4399 		device_printf(dev, "iflib_register failed %d\n", err);
4400 		goto fail_ctx_free;
4401 	}
4402 	iflib_add_device_sysctl_pre(ctx);
4403 
4404 	scctx = &ctx->ifc_softc_ctx;
4405 	ifp = ctx->ifc_ifp;
4406 
4407 	iflib_reset_qvalues(ctx);
4408 	CTX_LOCK(ctx);
4409 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4410 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4411 		goto fail_unlock;
4412 	}
4413 	_iflib_pre_assert(scctx);
4414 	ctx->ifc_txrx = *scctx->isc_txrx;
4415 
4416 #ifdef INVARIANTS
4417 	MPASS(scctx->isc_capabilities);
4418 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
4419 		MPASS(scctx->isc_tx_csum_flags);
4420 #endif
4421 
4422 	if_setcapabilities(ifp, scctx->isc_capabilities | IFCAP_HWSTATS);
4423 	if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS);
4424 
4425 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4426 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4427 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4428 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4429 
4430 	main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4431 	main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4432 
4433 	/* XXX change for per-queue sizes */
4434 	device_printf(dev, "Using %d tx descriptors and %d rx descriptors\n",
4435 	    scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4436 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4437 		if (!powerof2(scctx->isc_nrxd[i])) {
4438 			/* round down instead? */
4439 			device_printf(dev, "# rx descriptors must be a power of 2\n");
4440 			err = EINVAL;
4441 			goto fail_iflib_detach;
4442 		}
4443 	}
4444 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4445 		if (!powerof2(scctx->isc_ntxd[i])) {
4446 			device_printf(dev,
4447 			    "# tx descriptors must be a power of 2");
4448 			err = EINVAL;
4449 			goto fail_iflib_detach;
4450 		}
4451 	}
4452 
4453 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4454 	    MAX_SINGLE_PACKET_FRACTION)
4455 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4456 		    MAX_SINGLE_PACKET_FRACTION);
4457 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4458 	    MAX_SINGLE_PACKET_FRACTION)
4459 		scctx->isc_tx_tso_segments_max = max(1,
4460 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4461 
4462 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4463 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
4464 		/*
4465 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
4466 		 * but some MACs do.
4467 		 */
4468 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
4469 		    IP_MAXPACKET));
4470 		/*
4471 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
4472 		 * into account.  In the worst case, each of these calls will
4473 		 * add another mbuf and, thus, the requirement for another DMA
4474 		 * segment.  So for best performance, it doesn't make sense to
4475 		 * advertize a maximum of TSO segments that typically will
4476 		 * require defragmentation in iflib_encap().
4477 		 */
4478 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
4479 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
4480 	}
4481 	if (scctx->isc_rss_table_size == 0)
4482 		scctx->isc_rss_table_size = 64;
4483 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4484 
4485 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4486 	/* XXX format name */
4487 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx,
4488 	    NULL, NULL, "admin");
4489 
4490 	/* Set up cpu set.  If it fails, use the set of all CPUs. */
4491 	if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) != 0) {
4492 		device_printf(dev, "Unable to fetch CPU list\n");
4493 		CPU_COPY(&all_cpus, &ctx->ifc_cpus);
4494 	}
4495 	MPASS(CPU_COUNT(&ctx->ifc_cpus) > 0);
4496 
4497 	/*
4498 	** Now set up MSI or MSI-X, should return us the number of supported
4499 	** vectors (will be 1 for a legacy interrupt and MSI).
4500 	*/
4501 	if (sctx->isc_flags & IFLIB_SKIP_MSIX) {
4502 		msix = scctx->isc_vectors;
4503 	} else if (scctx->isc_msix_bar != 0)
4504 	       /*
4505 		* The simple fact that isc_msix_bar is not 0 does not mean we
4506 		* we have a good value there that is known to work.
4507 		*/
4508 		msix = iflib_msix_init(ctx);
4509 	else {
4510 		scctx->isc_vectors = 1;
4511 		scctx->isc_ntxqsets = 1;
4512 		scctx->isc_nrxqsets = 1;
4513 		scctx->isc_intr = IFLIB_INTR_LEGACY;
4514 		msix = 0;
4515 	}
4516 	/* Get memory for the station queues */
4517 	if ((err = iflib_queues_alloc(ctx))) {
4518 		device_printf(dev, "Unable to allocate queue memory\n");
4519 		goto fail_intr_free;
4520 	}
4521 
4522 	if ((err = iflib_qset_structures_setup(ctx)))
4523 		goto fail_queues;
4524 
4525 	/*
4526 	 * Group taskqueues aren't properly set up until SMP is started,
4527 	 * so we disable interrupts until we can handle them post
4528 	 * SI_SUB_SMP.
4529 	 *
4530 	 * XXX: disabling interrupts doesn't actually work, at least for
4531 	 * the non-MSI case.  When they occur before SI_SUB_SMP completes,
4532 	 * we do null handling and depend on this not causing too large an
4533 	 * interrupt storm.
4534 	 */
4535 	IFDI_INTR_DISABLE(ctx);
4536 	if (msix > 1 && (err = IFDI_MSIX_INTR_ASSIGN(ctx, msix)) != 0) {
4537 		device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n", err);
4538 		goto fail_queues;
4539 	}
4540 	if (msix <= 1) {
4541 		rid = 0;
4542 		if (scctx->isc_intr == IFLIB_INTR_MSI) {
4543 			MPASS(msix == 1);
4544 			rid = 1;
4545 		}
4546 		if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) {
4547 			device_printf(dev, "iflib_legacy_setup failed %d\n", err);
4548 			goto fail_queues;
4549 		}
4550 	}
4551 
4552 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac);
4553 
4554 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4555 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4556 		goto fail_detach;
4557 	}
4558 
4559 	/*
4560 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4561 	 * This must appear after the call to ether_ifattach() because
4562 	 * ether_ifattach() sets if_hdrlen to the default value.
4563 	 */
4564 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4565 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
4566 
4567 	if ((err = iflib_netmap_attach(ctx))) {
4568 		device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err);
4569 		goto fail_detach;
4570 	}
4571 	*ctxp = ctx;
4572 
4573 	NETDUMP_SET(ctx->ifc_ifp, iflib);
4574 
4575 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4576 	iflib_add_device_sysctl_post(ctx);
4577 	ctx->ifc_flags |= IFC_INIT_DONE;
4578 	CTX_UNLOCK(ctx);
4579 	return (0);
4580 
4581 fail_detach:
4582 	ether_ifdetach(ctx->ifc_ifp);
4583 fail_intr_free:
4584 	iflib_free_intr_mem(ctx);
4585 fail_queues:
4586 	iflib_tx_structures_free(ctx);
4587 	iflib_rx_structures_free(ctx);
4588 fail_iflib_detach:
4589 	IFDI_DETACH(ctx);
4590 fail_unlock:
4591 	CTX_UNLOCK(ctx);
4592 fail_ctx_free:
4593         if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4594                 free(ctx->ifc_softc, M_IFLIB);
4595         free(ctx, M_IFLIB);
4596 	return (err);
4597 }
4598 
4599 int
4600 iflib_pseudo_register(device_t dev, if_shared_ctx_t sctx, if_ctx_t *ctxp,
4601 					  struct iflib_cloneattach_ctx *clctx)
4602 {
4603 	int err;
4604 	if_ctx_t ctx;
4605 	if_t ifp;
4606 	if_softc_ctx_t scctx;
4607 	int i;
4608 	void *sc;
4609 	uint16_t main_txq;
4610 	uint16_t main_rxq;
4611 
4612 	ctx = malloc(sizeof(*ctx), M_IFLIB, M_WAITOK|M_ZERO);
4613 	sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4614 	ctx->ifc_flags |= IFC_SC_ALLOCATED;
4615 	if (sctx->isc_flags & (IFLIB_PSEUDO|IFLIB_VIRTUAL))
4616 		ctx->ifc_flags |= IFC_PSEUDO;
4617 
4618 	ctx->ifc_sctx = sctx;
4619 	ctx->ifc_softc = sc;
4620 	ctx->ifc_dev = dev;
4621 
4622 	if ((err = iflib_register(ctx)) != 0) {
4623 		device_printf(dev, "%s: iflib_register failed %d\n", __func__, err);
4624 		goto fail_ctx_free;
4625 	}
4626 	iflib_add_device_sysctl_pre(ctx);
4627 
4628 	scctx = &ctx->ifc_softc_ctx;
4629 	ifp = ctx->ifc_ifp;
4630 
4631 	/*
4632 	 * XXX sanity check that ntxd & nrxd are a power of 2
4633 	 */
4634 	iflib_reset_qvalues(ctx);
4635 
4636 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4637 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4638 		goto fail_ctx_free;
4639 	}
4640 	if (sctx->isc_flags & IFLIB_GEN_MAC)
4641 		iflib_gen_mac(ctx);
4642 	if ((err = IFDI_CLONEATTACH(ctx, clctx->cc_ifc, clctx->cc_name,
4643 								clctx->cc_params)) != 0) {
4644 		device_printf(dev, "IFDI_CLONEATTACH failed %d\n", err);
4645 		goto fail_ctx_free;
4646 	}
4647 	ifmedia_add(&ctx->ifc_media, IFM_ETHER | IFM_1000_T | IFM_FDX, 0, NULL);
4648 	ifmedia_add(&ctx->ifc_media, IFM_ETHER | IFM_AUTO, 0, NULL);
4649 	ifmedia_set(&ctx->ifc_media, IFM_ETHER | IFM_AUTO);
4650 
4651 #ifdef INVARIANTS
4652 	MPASS(scctx->isc_capabilities);
4653 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
4654 		MPASS(scctx->isc_tx_csum_flags);
4655 #endif
4656 
4657 	if_setcapabilities(ifp, scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_LINKSTATE);
4658 	if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_LINKSTATE);
4659 
4660 	ifp->if_flags |= IFF_NOGROUP;
4661 	if (sctx->isc_flags & IFLIB_PSEUDO) {
4662 		ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac);
4663 
4664 		if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4665 			device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4666 			goto fail_detach;
4667 		}
4668 		*ctxp = ctx;
4669 
4670 		/*
4671 		 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4672 		 * This must appear after the call to ether_ifattach() because
4673 		 * ether_ifattach() sets if_hdrlen to the default value.
4674 		 */
4675 		if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4676 			if_setifheaderlen(ifp,
4677 			    sizeof(struct ether_vlan_header));
4678 
4679 		if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4680 		iflib_add_device_sysctl_post(ctx);
4681 		ctx->ifc_flags |= IFC_INIT_DONE;
4682 		return (0);
4683 	}
4684 	_iflib_pre_assert(scctx);
4685 	ctx->ifc_txrx = *scctx->isc_txrx;
4686 
4687 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4688 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4689 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4690 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4691 
4692 	main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4693 	main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4694 
4695 	/* XXX change for per-queue sizes */
4696 	device_printf(dev, "Using %d tx descriptors and %d rx descriptors\n",
4697 	    scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4698 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4699 		if (!powerof2(scctx->isc_nrxd[i])) {
4700 			/* round down instead? */
4701 			device_printf(dev, "# rx descriptors must be a power of 2\n");
4702 			err = EINVAL;
4703 			goto fail_iflib_detach;
4704 		}
4705 	}
4706 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4707 		if (!powerof2(scctx->isc_ntxd[i])) {
4708 			device_printf(dev,
4709 			    "# tx descriptors must be a power of 2");
4710 			err = EINVAL;
4711 			goto fail_iflib_detach;
4712 		}
4713 	}
4714 
4715 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4716 	    MAX_SINGLE_PACKET_FRACTION)
4717 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4718 		    MAX_SINGLE_PACKET_FRACTION);
4719 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4720 	    MAX_SINGLE_PACKET_FRACTION)
4721 		scctx->isc_tx_tso_segments_max = max(1,
4722 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4723 
4724 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4725 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
4726 		/*
4727 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
4728 		 * but some MACs do.
4729 		 */
4730 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
4731 		    IP_MAXPACKET));
4732 		/*
4733 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
4734 		 * into account.  In the worst case, each of these calls will
4735 		 * add another mbuf and, thus, the requirement for another DMA
4736 		 * segment.  So for best performance, it doesn't make sense to
4737 		 * advertize a maximum of TSO segments that typically will
4738 		 * require defragmentation in iflib_encap().
4739 		 */
4740 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
4741 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
4742 	}
4743 	if (scctx->isc_rss_table_size == 0)
4744 		scctx->isc_rss_table_size = 64;
4745 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4746 
4747 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4748 	/* XXX format name */
4749 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx,
4750 	    NULL, NULL, "admin");
4751 
4752 	/* XXX --- can support > 1 -- but keep it simple for now */
4753 	scctx->isc_intr = IFLIB_INTR_LEGACY;
4754 
4755 	/* Get memory for the station queues */
4756 	if ((err = iflib_queues_alloc(ctx))) {
4757 		device_printf(dev, "Unable to allocate queue memory\n");
4758 		goto fail_iflib_detach;
4759 	}
4760 
4761 	if ((err = iflib_qset_structures_setup(ctx))) {
4762 		device_printf(dev, "qset structure setup failed %d\n", err);
4763 		goto fail_queues;
4764 	}
4765 
4766 	/*
4767 	 * XXX What if anything do we want to do about interrupts?
4768 	 */
4769 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac);
4770 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4771 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4772 		goto fail_detach;
4773 	}
4774 
4775 	/*
4776 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4777 	 * This must appear after the call to ether_ifattach() because
4778 	 * ether_ifattach() sets if_hdrlen to the default value.
4779 	 */
4780 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4781 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
4782 
4783 	/* XXX handle more than one queue */
4784 	for (i = 0; i < scctx->isc_nrxqsets; i++)
4785 		IFDI_RX_CLSET(ctx, 0, i, ctx->ifc_rxqs[i].ifr_fl[0].ifl_sds.ifsd_cl);
4786 
4787 	*ctxp = ctx;
4788 
4789 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4790 	iflib_add_device_sysctl_post(ctx);
4791 	ctx->ifc_flags |= IFC_INIT_DONE;
4792 	return (0);
4793 fail_detach:
4794 	ether_ifdetach(ctx->ifc_ifp);
4795 fail_queues:
4796 	iflib_tx_structures_free(ctx);
4797 	iflib_rx_structures_free(ctx);
4798 fail_iflib_detach:
4799 	IFDI_DETACH(ctx);
4800 fail_ctx_free:
4801 	free(ctx->ifc_softc, M_IFLIB);
4802 	free(ctx, M_IFLIB);
4803 	return (err);
4804 }
4805 
4806 int
4807 iflib_pseudo_deregister(if_ctx_t ctx)
4808 {
4809 	if_t ifp = ctx->ifc_ifp;
4810 	iflib_txq_t txq;
4811 	iflib_rxq_t rxq;
4812 	int i, j;
4813 	struct taskqgroup *tqg;
4814 	iflib_fl_t fl;
4815 
4816 	/* Unregister VLAN events */
4817 	if (ctx->ifc_vlan_attach_event != NULL)
4818 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
4819 	if (ctx->ifc_vlan_detach_event != NULL)
4820 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
4821 
4822 	ether_ifdetach(ifp);
4823 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
4824 	CTX_LOCK_DESTROY(ctx);
4825 	/* XXX drain any dependent tasks */
4826 	tqg = qgroup_if_io_tqg;
4827 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
4828 		callout_drain(&txq->ift_timer);
4829 		if (txq->ift_task.gt_uniq != NULL)
4830 			taskqgroup_detach(tqg, &txq->ift_task);
4831 	}
4832 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4833 		if (rxq->ifr_task.gt_uniq != NULL)
4834 			taskqgroup_detach(tqg, &rxq->ifr_task);
4835 
4836 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
4837 			free(fl->ifl_rx_bitmap, M_IFLIB);
4838 	}
4839 	tqg = qgroup_if_config_tqg;
4840 	if (ctx->ifc_admin_task.gt_uniq != NULL)
4841 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
4842 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
4843 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
4844 
4845 	if_free(ifp);
4846 
4847 	iflib_tx_structures_free(ctx);
4848 	iflib_rx_structures_free(ctx);
4849 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4850 		free(ctx->ifc_softc, M_IFLIB);
4851 	free(ctx, M_IFLIB);
4852 	return (0);
4853 }
4854 
4855 int
4856 iflib_device_attach(device_t dev)
4857 {
4858 	if_ctx_t ctx;
4859 	if_shared_ctx_t sctx;
4860 
4861 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4862 		return (ENOTSUP);
4863 
4864 	pci_enable_busmaster(dev);
4865 
4866 	return (iflib_device_register(dev, NULL, sctx, &ctx));
4867 }
4868 
4869 int
4870 iflib_device_deregister(if_ctx_t ctx)
4871 {
4872 	if_t ifp = ctx->ifc_ifp;
4873 	iflib_txq_t txq;
4874 	iflib_rxq_t rxq;
4875 	device_t dev = ctx->ifc_dev;
4876 	int i, j;
4877 	struct taskqgroup *tqg;
4878 	iflib_fl_t fl;
4879 
4880 	/* Make sure VLANS are not using driver */
4881 	if (if_vlantrunkinuse(ifp)) {
4882 		device_printf(dev, "Vlan in use, detach first\n");
4883 		return (EBUSY);
4884 	}
4885 #ifdef PCI_IOV
4886 	if (!CTX_IS_VF(ctx) && pci_iov_detach(dev) != 0) {
4887 		device_printf(dev, "SR-IOV in use; detach first.\n");
4888 		return (EBUSY);
4889 	}
4890 #endif
4891 
4892 	STATE_LOCK(ctx);
4893 	ctx->ifc_flags |= IFC_IN_DETACH;
4894 	STATE_UNLOCK(ctx);
4895 
4896 	CTX_LOCK(ctx);
4897 	iflib_stop(ctx);
4898 	CTX_UNLOCK(ctx);
4899 
4900 	/* Unregister VLAN events */
4901 	if (ctx->ifc_vlan_attach_event != NULL)
4902 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
4903 	if (ctx->ifc_vlan_detach_event != NULL)
4904 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
4905 
4906 	iflib_netmap_detach(ifp);
4907 	ether_ifdetach(ifp);
4908 	if (ctx->ifc_led_dev != NULL)
4909 		led_destroy(ctx->ifc_led_dev);
4910 	/* XXX drain any dependent tasks */
4911 	tqg = qgroup_if_io_tqg;
4912 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
4913 		callout_drain(&txq->ift_timer);
4914 		if (txq->ift_task.gt_uniq != NULL)
4915 			taskqgroup_detach(tqg, &txq->ift_task);
4916 	}
4917 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4918 		if (rxq->ifr_task.gt_uniq != NULL)
4919 			taskqgroup_detach(tqg, &rxq->ifr_task);
4920 
4921 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
4922 			free(fl->ifl_rx_bitmap, M_IFLIB);
4923 	}
4924 	tqg = qgroup_if_config_tqg;
4925 	if (ctx->ifc_admin_task.gt_uniq != NULL)
4926 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
4927 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
4928 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
4929 	CTX_LOCK(ctx);
4930 	IFDI_DETACH(ctx);
4931 	CTX_UNLOCK(ctx);
4932 
4933 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
4934 	CTX_LOCK_DESTROY(ctx);
4935 	device_set_softc(ctx->ifc_dev, NULL);
4936 	iflib_free_intr_mem(ctx);
4937 
4938 	bus_generic_detach(dev);
4939 	if_free(ifp);
4940 
4941 	iflib_tx_structures_free(ctx);
4942 	iflib_rx_structures_free(ctx);
4943 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4944 		free(ctx->ifc_softc, M_IFLIB);
4945 	STATE_LOCK_DESTROY(ctx);
4946 	free(ctx, M_IFLIB);
4947 	return (0);
4948 }
4949 
4950 static void
4951 iflib_free_intr_mem(if_ctx_t ctx)
4952 {
4953 
4954 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) {
4955 		iflib_irq_free(ctx, &ctx->ifc_legacy_irq);
4956 	}
4957 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) {
4958 		pci_release_msi(ctx->ifc_dev);
4959 	}
4960 	if (ctx->ifc_msix_mem != NULL) {
4961 		bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY,
4962 		    rman_get_rid(ctx->ifc_msix_mem), ctx->ifc_msix_mem);
4963 		ctx->ifc_msix_mem = NULL;
4964 	}
4965 }
4966 
4967 int
4968 iflib_device_detach(device_t dev)
4969 {
4970 	if_ctx_t ctx = device_get_softc(dev);
4971 
4972 	return (iflib_device_deregister(ctx));
4973 }
4974 
4975 int
4976 iflib_device_suspend(device_t dev)
4977 {
4978 	if_ctx_t ctx = device_get_softc(dev);
4979 
4980 	CTX_LOCK(ctx);
4981 	IFDI_SUSPEND(ctx);
4982 	CTX_UNLOCK(ctx);
4983 
4984 	return bus_generic_suspend(dev);
4985 }
4986 int
4987 iflib_device_shutdown(device_t dev)
4988 {
4989 	if_ctx_t ctx = device_get_softc(dev);
4990 
4991 	CTX_LOCK(ctx);
4992 	IFDI_SHUTDOWN(ctx);
4993 	CTX_UNLOCK(ctx);
4994 
4995 	return bus_generic_suspend(dev);
4996 }
4997 
4998 
4999 int
5000 iflib_device_resume(device_t dev)
5001 {
5002 	if_ctx_t ctx = device_get_softc(dev);
5003 	iflib_txq_t txq = ctx->ifc_txqs;
5004 
5005 	CTX_LOCK(ctx);
5006 	IFDI_RESUME(ctx);
5007 	iflib_if_init_locked(ctx);
5008 	CTX_UNLOCK(ctx);
5009 	for (int i = 0; i < NTXQSETS(ctx); i++, txq++)
5010 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
5011 
5012 	return (bus_generic_resume(dev));
5013 }
5014 
5015 int
5016 iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params)
5017 {
5018 	int error;
5019 	if_ctx_t ctx = device_get_softc(dev);
5020 
5021 	CTX_LOCK(ctx);
5022 	error = IFDI_IOV_INIT(ctx, num_vfs, params);
5023 	CTX_UNLOCK(ctx);
5024 
5025 	return (error);
5026 }
5027 
5028 void
5029 iflib_device_iov_uninit(device_t dev)
5030 {
5031 	if_ctx_t ctx = device_get_softc(dev);
5032 
5033 	CTX_LOCK(ctx);
5034 	IFDI_IOV_UNINIT(ctx);
5035 	CTX_UNLOCK(ctx);
5036 }
5037 
5038 int
5039 iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params)
5040 {
5041 	int error;
5042 	if_ctx_t ctx = device_get_softc(dev);
5043 
5044 	CTX_LOCK(ctx);
5045 	error = IFDI_IOV_VF_ADD(ctx, vfnum, params);
5046 	CTX_UNLOCK(ctx);
5047 
5048 	return (error);
5049 }
5050 
5051 /*********************************************************************
5052  *
5053  *  MODULE FUNCTION DEFINITIONS
5054  *
5055  **********************************************************************/
5056 
5057 /*
5058  * - Start a fast taskqueue thread for each core
5059  * - Start a taskqueue for control operations
5060  */
5061 static int
5062 iflib_module_init(void)
5063 {
5064 	return (0);
5065 }
5066 
5067 static int
5068 iflib_module_event_handler(module_t mod, int what, void *arg)
5069 {
5070 	int err;
5071 
5072 	switch (what) {
5073 	case MOD_LOAD:
5074 		if ((err = iflib_module_init()) != 0)
5075 			return (err);
5076 		break;
5077 	case MOD_UNLOAD:
5078 		return (EBUSY);
5079 	default:
5080 		return (EOPNOTSUPP);
5081 	}
5082 
5083 	return (0);
5084 }
5085 
5086 /*********************************************************************
5087  *
5088  *  PUBLIC FUNCTION DEFINITIONS
5089  *     ordered as in iflib.h
5090  *
5091  **********************************************************************/
5092 
5093 
5094 static void
5095 _iflib_assert(if_shared_ctx_t sctx)
5096 {
5097 	MPASS(sctx->isc_tx_maxsize);
5098 	MPASS(sctx->isc_tx_maxsegsize);
5099 
5100 	MPASS(sctx->isc_rx_maxsize);
5101 	MPASS(sctx->isc_rx_nsegments);
5102 	MPASS(sctx->isc_rx_maxsegsize);
5103 
5104 	MPASS(sctx->isc_nrxd_min[0]);
5105 	MPASS(sctx->isc_nrxd_max[0]);
5106 	MPASS(sctx->isc_nrxd_default[0]);
5107 	MPASS(sctx->isc_ntxd_min[0]);
5108 	MPASS(sctx->isc_ntxd_max[0]);
5109 	MPASS(sctx->isc_ntxd_default[0]);
5110 }
5111 
5112 static void
5113 _iflib_pre_assert(if_softc_ctx_t scctx)
5114 {
5115 
5116 	MPASS(scctx->isc_txrx->ift_txd_encap);
5117 	MPASS(scctx->isc_txrx->ift_txd_flush);
5118 	MPASS(scctx->isc_txrx->ift_txd_credits_update);
5119 	MPASS(scctx->isc_txrx->ift_rxd_available);
5120 	MPASS(scctx->isc_txrx->ift_rxd_pkt_get);
5121 	MPASS(scctx->isc_txrx->ift_rxd_refill);
5122 	MPASS(scctx->isc_txrx->ift_rxd_flush);
5123 }
5124 
5125 static int
5126 iflib_register(if_ctx_t ctx)
5127 {
5128 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5129 	driver_t *driver = sctx->isc_driver;
5130 	device_t dev = ctx->ifc_dev;
5131 	if_t ifp;
5132 
5133 	_iflib_assert(sctx);
5134 
5135 	CTX_LOCK_INIT(ctx);
5136 	STATE_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev));
5137 	ifp = ctx->ifc_ifp = if_alloc(IFT_ETHER);
5138 	if (ifp == NULL) {
5139 		device_printf(dev, "can not allocate ifnet structure\n");
5140 		return (ENOMEM);
5141 	}
5142 
5143 	/*
5144 	 * Initialize our context's device specific methods
5145 	 */
5146 	kobj_init((kobj_t) ctx, (kobj_class_t) driver);
5147 	kobj_class_compile((kobj_class_t) driver);
5148 	driver->refs++;
5149 
5150 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
5151 	if_setsoftc(ifp, ctx);
5152 	if_setdev(ifp, dev);
5153 	if_setinitfn(ifp, iflib_if_init);
5154 	if_setioctlfn(ifp, iflib_if_ioctl);
5155 #ifdef ALTQ
5156 	if_setstartfn(ifp, iflib_altq_if_start);
5157 	if_settransmitfn(ifp, iflib_altq_if_transmit);
5158 	if_setsendqready(ifp);
5159 #else
5160 	if_settransmitfn(ifp, iflib_if_transmit);
5161 #endif
5162 	if_setqflushfn(ifp, iflib_if_qflush);
5163 	if_setflags(ifp, IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST);
5164 
5165 	ctx->ifc_vlan_attach_event =
5166 		EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx,
5167 							  EVENTHANDLER_PRI_FIRST);
5168 	ctx->ifc_vlan_detach_event =
5169 		EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx,
5170 							  EVENTHANDLER_PRI_FIRST);
5171 
5172 	ifmedia_init(&ctx->ifc_media, IFM_IMASK,
5173 					 iflib_media_change, iflib_media_status);
5174 
5175 	return (0);
5176 }
5177 
5178 
5179 static int
5180 iflib_queues_alloc(if_ctx_t ctx)
5181 {
5182 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5183 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5184 	device_t dev = ctx->ifc_dev;
5185 	int nrxqsets = scctx->isc_nrxqsets;
5186 	int ntxqsets = scctx->isc_ntxqsets;
5187 	iflib_txq_t txq;
5188 	iflib_rxq_t rxq;
5189 	iflib_fl_t fl = NULL;
5190 	int i, j, cpu, err, txconf, rxconf;
5191 	iflib_dma_info_t ifdip;
5192 	uint32_t *rxqsizes = scctx->isc_rxqsizes;
5193 	uint32_t *txqsizes = scctx->isc_txqsizes;
5194 	uint8_t nrxqs = sctx->isc_nrxqs;
5195 	uint8_t ntxqs = sctx->isc_ntxqs;
5196 	int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1;
5197 	caddr_t *vaddrs;
5198 	uint64_t *paddrs;
5199 
5200 	KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1"));
5201 	KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1"));
5202 
5203 	/* Allocate the TX ring struct memory */
5204 	if (!(ctx->ifc_txqs =
5205 	    (iflib_txq_t) malloc(sizeof(struct iflib_txq) *
5206 	    ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5207 		device_printf(dev, "Unable to allocate TX ring memory\n");
5208 		err = ENOMEM;
5209 		goto fail;
5210 	}
5211 
5212 	/* Now allocate the RX */
5213 	if (!(ctx->ifc_rxqs =
5214 	    (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) *
5215 	    nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5216 		device_printf(dev, "Unable to allocate RX ring memory\n");
5217 		err = ENOMEM;
5218 		goto rx_fail;
5219 	}
5220 
5221 	txq = ctx->ifc_txqs;
5222 	rxq = ctx->ifc_rxqs;
5223 
5224 	/*
5225 	 * XXX handle allocation failure
5226 	 */
5227 	for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) {
5228 		/* Set up some basics */
5229 
5230 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs,
5231 		    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
5232 			device_printf(dev,
5233 			    "Unable to allocate TX DMA info memory\n");
5234 			err = ENOMEM;
5235 			goto err_tx_desc;
5236 		}
5237 		txq->ift_ifdi = ifdip;
5238 		for (j = 0; j < ntxqs; j++, ifdip++) {
5239 			if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, 0)) {
5240 				device_printf(dev,
5241 				    "Unable to allocate TX descriptors\n");
5242 				err = ENOMEM;
5243 				goto err_tx_desc;
5244 			}
5245 			txq->ift_txd_size[j] = scctx->isc_txd_size[j];
5246 			bzero((void *)ifdip->idi_vaddr, txqsizes[j]);
5247 		}
5248 		txq->ift_ctx = ctx;
5249 		txq->ift_id = i;
5250 		if (sctx->isc_flags & IFLIB_HAS_TXCQ) {
5251 			txq->ift_br_offset = 1;
5252 		} else {
5253 			txq->ift_br_offset = 0;
5254 		}
5255 		/* XXX fix this */
5256 		txq->ift_timer.c_cpu = cpu;
5257 
5258 		if (iflib_txsd_alloc(txq)) {
5259 			device_printf(dev, "Critical Failure setting up TX buffers\n");
5260 			err = ENOMEM;
5261 			goto err_tx_desc;
5262 		}
5263 
5264 		/* Initialize the TX lock */
5265 		snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:tx(%d):callout",
5266 		    device_get_nameunit(dev), txq->ift_id);
5267 		mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF);
5268 		callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0);
5269 
5270 		snprintf(txq->ift_db_mtx_name, MTX_NAME_LEN, "%s:tx(%d):db",
5271 			 device_get_nameunit(dev), txq->ift_id);
5272 
5273 		err = ifmp_ring_alloc(&txq->ift_br, 2048, txq, iflib_txq_drain,
5274 				      iflib_txq_can_drain, M_IFLIB, M_WAITOK);
5275 		if (err) {
5276 			/* XXX free any allocated rings */
5277 			device_printf(dev, "Unable to allocate buf_ring\n");
5278 			goto err_tx_desc;
5279 		}
5280 	}
5281 
5282 	for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) {
5283 		/* Set up some basics */
5284 
5285 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs,
5286 		   M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
5287 			device_printf(dev,
5288 			    "Unable to allocate RX DMA info memory\n");
5289 			err = ENOMEM;
5290 			goto err_tx_desc;
5291 		}
5292 
5293 		rxq->ifr_ifdi = ifdip;
5294 		/* XXX this needs to be changed if #rx queues != #tx queues */
5295 		rxq->ifr_ntxqirq = 1;
5296 		rxq->ifr_txqid[0] = i;
5297 		for (j = 0; j < nrxqs; j++, ifdip++) {
5298 			if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, 0)) {
5299 				device_printf(dev,
5300 				    "Unable to allocate RX descriptors\n");
5301 				err = ENOMEM;
5302 				goto err_tx_desc;
5303 			}
5304 			bzero((void *)ifdip->idi_vaddr, rxqsizes[j]);
5305 		}
5306 		rxq->ifr_ctx = ctx;
5307 		rxq->ifr_id = i;
5308 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
5309 			rxq->ifr_fl_offset = 1;
5310 		} else {
5311 			rxq->ifr_fl_offset = 0;
5312 		}
5313 		rxq->ifr_nfl = nfree_lists;
5314 		if (!(fl =
5315 			  (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) {
5316 			device_printf(dev, "Unable to allocate free list memory\n");
5317 			err = ENOMEM;
5318 			goto err_tx_desc;
5319 		}
5320 		rxq->ifr_fl = fl;
5321 		for (j = 0; j < nfree_lists; j++) {
5322 			fl[j].ifl_rxq = rxq;
5323 			fl[j].ifl_id = j;
5324 			fl[j].ifl_ifdi = &rxq->ifr_ifdi[j + rxq->ifr_fl_offset];
5325 			fl[j].ifl_rxd_size = scctx->isc_rxd_size[j];
5326 		}
5327 		/* Allocate receive buffers for the ring */
5328 		if (iflib_rxsd_alloc(rxq)) {
5329 			device_printf(dev,
5330 			    "Critical Failure setting up receive buffers\n");
5331 			err = ENOMEM;
5332 			goto err_rx_desc;
5333 		}
5334 
5335 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5336 			fl->ifl_rx_bitmap = bit_alloc(fl->ifl_size, M_IFLIB,
5337 			    M_WAITOK);
5338 	}
5339 
5340 	/* TXQs */
5341 	vaddrs = malloc(sizeof(caddr_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5342 	paddrs = malloc(sizeof(uint64_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5343 	for (i = 0; i < ntxqsets; i++) {
5344 		iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi;
5345 
5346 		for (j = 0; j < ntxqs; j++, di++) {
5347 			vaddrs[i*ntxqs + j] = di->idi_vaddr;
5348 			paddrs[i*ntxqs + j] = di->idi_paddr;
5349 		}
5350 	}
5351 	if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) {
5352 		device_printf(ctx->ifc_dev,
5353 		    "Unable to allocate device TX queue\n");
5354 		iflib_tx_structures_free(ctx);
5355 		free(vaddrs, M_IFLIB);
5356 		free(paddrs, M_IFLIB);
5357 		goto err_rx_desc;
5358 	}
5359 	free(vaddrs, M_IFLIB);
5360 	free(paddrs, M_IFLIB);
5361 
5362 	/* RXQs */
5363 	vaddrs = malloc(sizeof(caddr_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5364 	paddrs = malloc(sizeof(uint64_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5365 	for (i = 0; i < nrxqsets; i++) {
5366 		iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi;
5367 
5368 		for (j = 0; j < nrxqs; j++, di++) {
5369 			vaddrs[i*nrxqs + j] = di->idi_vaddr;
5370 			paddrs[i*nrxqs + j] = di->idi_paddr;
5371 		}
5372 	}
5373 	if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) {
5374 		device_printf(ctx->ifc_dev,
5375 		    "Unable to allocate device RX queue\n");
5376 		iflib_tx_structures_free(ctx);
5377 		free(vaddrs, M_IFLIB);
5378 		free(paddrs, M_IFLIB);
5379 		goto err_rx_desc;
5380 	}
5381 	free(vaddrs, M_IFLIB);
5382 	free(paddrs, M_IFLIB);
5383 
5384 	return (0);
5385 
5386 /* XXX handle allocation failure changes */
5387 err_rx_desc:
5388 err_tx_desc:
5389 rx_fail:
5390 	if (ctx->ifc_rxqs != NULL)
5391 		free(ctx->ifc_rxqs, M_IFLIB);
5392 	ctx->ifc_rxqs = NULL;
5393 	if (ctx->ifc_txqs != NULL)
5394 		free(ctx->ifc_txqs, M_IFLIB);
5395 	ctx->ifc_txqs = NULL;
5396 fail:
5397 	return (err);
5398 }
5399 
5400 static int
5401 iflib_tx_structures_setup(if_ctx_t ctx)
5402 {
5403 	iflib_txq_t txq = ctx->ifc_txqs;
5404 	int i;
5405 
5406 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
5407 		iflib_txq_setup(txq);
5408 
5409 	return (0);
5410 }
5411 
5412 static void
5413 iflib_tx_structures_free(if_ctx_t ctx)
5414 {
5415 	iflib_txq_t txq = ctx->ifc_txqs;
5416 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5417 	int i, j;
5418 
5419 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
5420 		iflib_txq_destroy(txq);
5421 		for (j = 0; j < sctx->isc_ntxqs; j++)
5422 			iflib_dma_free(&txq->ift_ifdi[j]);
5423 	}
5424 	free(ctx->ifc_txqs, M_IFLIB);
5425 	ctx->ifc_txqs = NULL;
5426 	IFDI_QUEUES_FREE(ctx);
5427 }
5428 
5429 /*********************************************************************
5430  *
5431  *  Initialize all receive rings.
5432  *
5433  **********************************************************************/
5434 static int
5435 iflib_rx_structures_setup(if_ctx_t ctx)
5436 {
5437 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5438 	int q;
5439 #if defined(INET6) || defined(INET)
5440 	int i, err;
5441 #endif
5442 
5443 	for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) {
5444 #if defined(INET6) || defined(INET)
5445 		tcp_lro_free(&rxq->ifr_lc);
5446 		if ((err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp,
5447 		    TCP_LRO_ENTRIES, min(1024,
5448 		    ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset]))) != 0) {
5449 			device_printf(ctx->ifc_dev, "LRO Initialization failed!\n");
5450 			goto fail;
5451 		}
5452 		rxq->ifr_lro_enabled = TRUE;
5453 #endif
5454 		IFDI_RXQ_SETUP(ctx, rxq->ifr_id);
5455 	}
5456 	return (0);
5457 #if defined(INET6) || defined(INET)
5458 fail:
5459 	/*
5460 	 * Free RX software descriptors allocated so far, we will only handle
5461 	 * the rings that completed, the failing case will have
5462 	 * cleaned up for itself. 'q' failed, so its the terminus.
5463 	 */
5464 	rxq = ctx->ifc_rxqs;
5465 	for (i = 0; i < q; ++i, rxq++) {
5466 		iflib_rx_sds_free(rxq);
5467 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
5468 	}
5469 	return (err);
5470 #endif
5471 }
5472 
5473 /*********************************************************************
5474  *
5475  *  Free all receive rings.
5476  *
5477  **********************************************************************/
5478 static void
5479 iflib_rx_structures_free(if_ctx_t ctx)
5480 {
5481 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5482 
5483 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) {
5484 		iflib_rx_sds_free(rxq);
5485 	}
5486 	free(ctx->ifc_rxqs, M_IFLIB);
5487 	ctx->ifc_rxqs = NULL;
5488 }
5489 
5490 static int
5491 iflib_qset_structures_setup(if_ctx_t ctx)
5492 {
5493 	int err;
5494 
5495 	/*
5496 	 * It is expected that the caller takes care of freeing queues if this
5497 	 * fails.
5498 	 */
5499 	if ((err = iflib_tx_structures_setup(ctx)) != 0) {
5500 		device_printf(ctx->ifc_dev, "iflib_tx_structures_setup failed: %d\n", err);
5501 		return (err);
5502 	}
5503 
5504 	if ((err = iflib_rx_structures_setup(ctx)) != 0)
5505 		device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err);
5506 
5507 	return (err);
5508 }
5509 
5510 int
5511 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
5512 		driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, const char *name)
5513 {
5514 
5515 	return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name));
5516 }
5517 
5518 #ifdef SMP
5519 static int
5520 find_nth(if_ctx_t ctx, int qid)
5521 {
5522 	cpuset_t cpus;
5523 	int i, cpuid, eqid, count;
5524 
5525 	CPU_COPY(&ctx->ifc_cpus, &cpus);
5526 	count = CPU_COUNT(&cpus);
5527 	eqid = qid % count;
5528 	/* clear up to the qid'th bit */
5529 	for (i = 0; i < eqid; i++) {
5530 		cpuid = CPU_FFS(&cpus);
5531 		MPASS(cpuid != 0);
5532 		CPU_CLR(cpuid-1, &cpus);
5533 	}
5534 	cpuid = CPU_FFS(&cpus);
5535 	MPASS(cpuid != 0);
5536 	return (cpuid-1);
5537 }
5538 
5539 #ifdef SCHED_ULE
5540 extern struct cpu_group *cpu_top;              /* CPU topology */
5541 
5542 static int
5543 find_child_with_core(int cpu, struct cpu_group *grp)
5544 {
5545 	int i;
5546 
5547 	if (grp->cg_children == 0)
5548 		return -1;
5549 
5550 	MPASS(grp->cg_child);
5551 	for (i = 0; i < grp->cg_children; i++) {
5552 		if (CPU_ISSET(cpu, &grp->cg_child[i].cg_mask))
5553 			return i;
5554 	}
5555 
5556 	return -1;
5557 }
5558 
5559 /*
5560  * Find the nth "close" core to the specified core
5561  * "close" is defined as the deepest level that shares
5562  * at least an L2 cache.  With threads, this will be
5563  * threads on the same core.  If the sahred cache is L3
5564  * or higher, simply returns the same core.
5565  */
5566 static int
5567 find_close_core(int cpu, int core_offset)
5568 {
5569 	struct cpu_group *grp;
5570 	int i;
5571 	int fcpu;
5572 	cpuset_t cs;
5573 
5574 	grp = cpu_top;
5575 	if (grp == NULL)
5576 		return cpu;
5577 	i = 0;
5578 	while ((i = find_child_with_core(cpu, grp)) != -1) {
5579 		/* If the child only has one cpu, don't descend */
5580 		if (grp->cg_child[i].cg_count <= 1)
5581 			break;
5582 		grp = &grp->cg_child[i];
5583 	}
5584 
5585 	/* If they don't share at least an L2 cache, use the same CPU */
5586 	if (grp->cg_level > CG_SHARE_L2 || grp->cg_level == CG_SHARE_NONE)
5587 		return cpu;
5588 
5589 	/* Now pick one */
5590 	CPU_COPY(&grp->cg_mask, &cs);
5591 
5592 	/* Add the selected CPU offset to core offset. */
5593 	for (i = 0; (fcpu = CPU_FFS(&cs)) != 0; i++) {
5594 		if (fcpu - 1 == cpu)
5595 			break;
5596 		CPU_CLR(fcpu - 1, &cs);
5597 	}
5598 	MPASS(fcpu);
5599 
5600 	core_offset += i;
5601 
5602 	CPU_COPY(&grp->cg_mask, &cs);
5603 	for (i = core_offset % grp->cg_count; i > 0; i--) {
5604 		MPASS(CPU_FFS(&cs));
5605 		CPU_CLR(CPU_FFS(&cs) - 1, &cs);
5606 	}
5607 	MPASS(CPU_FFS(&cs));
5608 	return CPU_FFS(&cs) - 1;
5609 }
5610 #else
5611 static int
5612 find_close_core(int cpu, int core_offset __unused)
5613 {
5614 	return cpu;
5615 }
5616 #endif
5617 
5618 static int
5619 get_core_offset(if_ctx_t ctx, iflib_intr_type_t type, int qid)
5620 {
5621 	switch (type) {
5622 	case IFLIB_INTR_TX:
5623 		/* TX queues get cores which share at least an L2 cache with the corresponding RX queue */
5624 		/* XXX handle multiple RX threads per core and more than two core per L2 group */
5625 		return qid / CPU_COUNT(&ctx->ifc_cpus) + 1;
5626 	case IFLIB_INTR_RX:
5627 	case IFLIB_INTR_RXTX:
5628 		/* RX queues get the specified core */
5629 		return qid / CPU_COUNT(&ctx->ifc_cpus);
5630 	default:
5631 		return -1;
5632 	}
5633 }
5634 #else
5635 #define get_core_offset(ctx, type, qid)	CPU_FIRST()
5636 #define find_close_core(cpuid, tid)	CPU_FIRST()
5637 #define find_nth(ctx, gid)		CPU_FIRST()
5638 #endif
5639 
5640 /* Just to avoid copy/paste */
5641 static inline int
5642 iflib_irq_set_affinity(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type,
5643     int qid, struct grouptask *gtask, struct taskqgroup *tqg, void *uniq,
5644     const char *name)
5645 {
5646 	device_t dev;
5647 	int err, cpuid, tid;
5648 
5649 	dev = ctx->ifc_dev;
5650 	cpuid = find_nth(ctx, qid);
5651 	tid = get_core_offset(ctx, type, qid);
5652 	MPASS(tid >= 0);
5653 	cpuid = find_close_core(cpuid, tid);
5654 	err = taskqgroup_attach_cpu(tqg, gtask, uniq, cpuid, dev, irq->ii_res,
5655 	    name);
5656 	if (err) {
5657 		device_printf(dev, "taskqgroup_attach_cpu failed %d\n", err);
5658 		return (err);
5659 	}
5660 #ifdef notyet
5661 	if (cpuid > ctx->ifc_cpuid_highest)
5662 		ctx->ifc_cpuid_highest = cpuid;
5663 #endif
5664 	return 0;
5665 }
5666 
5667 int
5668 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid,
5669 			iflib_intr_type_t type, driver_filter_t *filter,
5670 			void *filter_arg, int qid, const char *name)
5671 {
5672 	device_t dev;
5673 	struct grouptask *gtask;
5674 	struct taskqgroup *tqg;
5675 	iflib_filter_info_t info;
5676 	gtask_fn_t *fn;
5677 	int tqrid, err;
5678 	driver_filter_t *intr_fast;
5679 	void *q;
5680 
5681 	info = &ctx->ifc_filter_info;
5682 	tqrid = rid;
5683 
5684 	switch (type) {
5685 	/* XXX merge tx/rx for netmap? */
5686 	case IFLIB_INTR_TX:
5687 		q = &ctx->ifc_txqs[qid];
5688 		info = &ctx->ifc_txqs[qid].ift_filter_info;
5689 		gtask = &ctx->ifc_txqs[qid].ift_task;
5690 		tqg = qgroup_if_io_tqg;
5691 		fn = _task_fn_tx;
5692 		intr_fast = iflib_fast_intr;
5693 		GROUPTASK_INIT(gtask, 0, fn, q);
5694 		ctx->ifc_flags |= IFC_NETMAP_TX_IRQ;
5695 		break;
5696 	case IFLIB_INTR_RX:
5697 		q = &ctx->ifc_rxqs[qid];
5698 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
5699 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5700 		tqg = qgroup_if_io_tqg;
5701 		fn = _task_fn_rx;
5702 		intr_fast = iflib_fast_intr;
5703 		GROUPTASK_INIT(gtask, 0, fn, q);
5704 		break;
5705 	case IFLIB_INTR_RXTX:
5706 		q = &ctx->ifc_rxqs[qid];
5707 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
5708 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5709 		tqg = qgroup_if_io_tqg;
5710 		fn = _task_fn_rx;
5711 		intr_fast = iflib_fast_intr_rxtx;
5712 		GROUPTASK_INIT(gtask, 0, fn, q);
5713 		break;
5714 	case IFLIB_INTR_ADMIN:
5715 		q = ctx;
5716 		tqrid = -1;
5717 		info = &ctx->ifc_filter_info;
5718 		gtask = &ctx->ifc_admin_task;
5719 		tqg = qgroup_if_config_tqg;
5720 		fn = _task_fn_admin;
5721 		intr_fast = iflib_fast_intr_ctx;
5722 		break;
5723 	default:
5724 		panic("unknown net intr type");
5725 	}
5726 
5727 	info->ifi_filter = filter;
5728 	info->ifi_filter_arg = filter_arg;
5729 	info->ifi_task = gtask;
5730 	info->ifi_ctx = q;
5731 
5732 	dev = ctx->ifc_dev;
5733 	err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info,  name);
5734 	if (err != 0) {
5735 		device_printf(dev, "_iflib_irq_alloc failed %d\n", err);
5736 		return (err);
5737 	}
5738 	if (type == IFLIB_INTR_ADMIN)
5739 		return (0);
5740 
5741 	if (tqrid != -1) {
5742 		err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg,
5743 		    q, name);
5744 		if (err)
5745 			return (err);
5746 	} else {
5747 		taskqgroup_attach(tqg, gtask, q, dev, irq->ii_res, name);
5748 	}
5749 
5750 	return (0);
5751 }
5752 
5753 void
5754 iflib_softirq_alloc_generic(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type, void *arg, int qid, const char *name)
5755 {
5756 	struct grouptask *gtask;
5757 	struct taskqgroup *tqg;
5758 	gtask_fn_t *fn;
5759 	void *q;
5760 	int err;
5761 
5762 	switch (type) {
5763 	case IFLIB_INTR_TX:
5764 		q = &ctx->ifc_txqs[qid];
5765 		gtask = &ctx->ifc_txqs[qid].ift_task;
5766 		tqg = qgroup_if_io_tqg;
5767 		fn = _task_fn_tx;
5768 		break;
5769 	case IFLIB_INTR_RX:
5770 		q = &ctx->ifc_rxqs[qid];
5771 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5772 		tqg = qgroup_if_io_tqg;
5773 		fn = _task_fn_rx;
5774 		break;
5775 	case IFLIB_INTR_IOV:
5776 		q = ctx;
5777 		gtask = &ctx->ifc_vflr_task;
5778 		tqg = qgroup_if_config_tqg;
5779 		fn = _task_fn_iov;
5780 		break;
5781 	default:
5782 		panic("unknown net intr type");
5783 	}
5784 	GROUPTASK_INIT(gtask, 0, fn, q);
5785 	if (irq != NULL) {
5786 		err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg,
5787 		    q, name);
5788 		if (err)
5789 			taskqgroup_attach(tqg, gtask, q, ctx->ifc_dev,
5790 			    irq->ii_res, name);
5791 	} else {
5792 		taskqgroup_attach(tqg, gtask, q, NULL, NULL, name);
5793 	}
5794 }
5795 
5796 void
5797 iflib_irq_free(if_ctx_t ctx, if_irq_t irq)
5798 {
5799 
5800 	if (irq->ii_tag)
5801 		bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag);
5802 
5803 	if (irq->ii_res)
5804 		bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ,
5805 		    rman_get_rid(irq->ii_res), irq->ii_res);
5806 }
5807 
5808 static int
5809 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, const char *name)
5810 {
5811 	iflib_txq_t txq = ctx->ifc_txqs;
5812 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5813 	if_irq_t irq = &ctx->ifc_legacy_irq;
5814 	iflib_filter_info_t info;
5815 	device_t dev;
5816 	struct grouptask *gtask;
5817 	struct resource *res;
5818 	struct taskqgroup *tqg;
5819 	gtask_fn_t *fn;
5820 	int tqrid;
5821 	void *q;
5822 	int err;
5823 
5824 	q = &ctx->ifc_rxqs[0];
5825 	info = &rxq[0].ifr_filter_info;
5826 	gtask = &rxq[0].ifr_task;
5827 	tqg = qgroup_if_io_tqg;
5828 	tqrid = irq->ii_rid = *rid;
5829 	fn = _task_fn_rx;
5830 
5831 	ctx->ifc_flags |= IFC_LEGACY;
5832 	info->ifi_filter = filter;
5833 	info->ifi_filter_arg = filter_arg;
5834 	info->ifi_task = gtask;
5835 	info->ifi_ctx = ctx;
5836 
5837 	dev = ctx->ifc_dev;
5838 	/* We allocate a single interrupt resource */
5839 	if ((err = _iflib_irq_alloc(ctx, irq, tqrid, iflib_fast_intr_ctx, NULL, info, name)) != 0)
5840 		return (err);
5841 	GROUPTASK_INIT(gtask, 0, fn, q);
5842 	res = irq->ii_res;
5843 	taskqgroup_attach(tqg, gtask, q, dev, res, name);
5844 
5845 	GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq);
5846 	taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, dev, res,
5847 	    "tx");
5848 	return (0);
5849 }
5850 
5851 void
5852 iflib_led_create(if_ctx_t ctx)
5853 {
5854 
5855 	ctx->ifc_led_dev = led_create(iflib_led_func, ctx,
5856 	    device_get_nameunit(ctx->ifc_dev));
5857 }
5858 
5859 void
5860 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid)
5861 {
5862 
5863 	GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
5864 }
5865 
5866 void
5867 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid)
5868 {
5869 
5870 	GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task);
5871 }
5872 
5873 void
5874 iflib_admin_intr_deferred(if_ctx_t ctx)
5875 {
5876 #ifdef INVARIANTS
5877 	struct grouptask *gtask;
5878 
5879 	gtask = &ctx->ifc_admin_task;
5880 	MPASS(gtask != NULL && gtask->gt_taskqueue != NULL);
5881 #endif
5882 
5883 	GROUPTASK_ENQUEUE(&ctx->ifc_admin_task);
5884 }
5885 
5886 void
5887 iflib_iov_intr_deferred(if_ctx_t ctx)
5888 {
5889 
5890 	GROUPTASK_ENQUEUE(&ctx->ifc_vflr_task);
5891 }
5892 
5893 void
5894 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, char *name)
5895 {
5896 
5897 	taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, NULL, NULL,
5898 	    name);
5899 }
5900 
5901 void
5902 iflib_config_gtask_init(void *ctx, struct grouptask *gtask, gtask_fn_t *fn,
5903 	const char *name)
5904 {
5905 
5906 	GROUPTASK_INIT(gtask, 0, fn, ctx);
5907 	taskqgroup_attach(qgroup_if_config_tqg, gtask, gtask, NULL, NULL,
5908 	    name);
5909 }
5910 
5911 void
5912 iflib_config_gtask_deinit(struct grouptask *gtask)
5913 {
5914 
5915 	taskqgroup_detach(qgroup_if_config_tqg, gtask);
5916 }
5917 
5918 void
5919 iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate)
5920 {
5921 	if_t ifp = ctx->ifc_ifp;
5922 	iflib_txq_t txq = ctx->ifc_txqs;
5923 
5924 	if_setbaudrate(ifp, baudrate);
5925 	if (baudrate >= IF_Gbps(10)) {
5926 		STATE_LOCK(ctx);
5927 		ctx->ifc_flags |= IFC_PREFETCH;
5928 		STATE_UNLOCK(ctx);
5929 	}
5930 	/* If link down, disable watchdog */
5931 	if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) {
5932 		for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++)
5933 			txq->ift_qstatus = IFLIB_QUEUE_IDLE;
5934 	}
5935 	ctx->ifc_link_state = link_state;
5936 	if_link_state_change(ifp, link_state);
5937 }
5938 
5939 static int
5940 iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq)
5941 {
5942 	int credits;
5943 #ifdef INVARIANTS
5944 	int credits_pre = txq->ift_cidx_processed;
5945 #endif
5946 
5947 	if (ctx->isc_txd_credits_update == NULL)
5948 		return (0);
5949 
5950 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
5951 	    BUS_DMASYNC_POSTREAD);
5952 	if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, true)) == 0)
5953 		return (0);
5954 
5955 	txq->ift_processed += credits;
5956 	txq->ift_cidx_processed += credits;
5957 
5958 	MPASS(credits_pre + credits == txq->ift_cidx_processed);
5959 	if (txq->ift_cidx_processed >= txq->ift_size)
5960 		txq->ift_cidx_processed -= txq->ift_size;
5961 	return (credits);
5962 }
5963 
5964 static int
5965 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget)
5966 {
5967 	iflib_fl_t fl;
5968 	u_int i;
5969 
5970 	for (i = 0, fl = &rxq->ifr_fl[0]; i < rxq->ifr_nfl; i++, fl++)
5971 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
5972 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
5973 	return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx,
5974 	    budget));
5975 }
5976 
5977 void
5978 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name,
5979 	const char *description, if_int_delay_info_t info,
5980 	int offset, int value)
5981 {
5982 	info->iidi_ctx = ctx;
5983 	info->iidi_offset = offset;
5984 	info->iidi_value = value;
5985 	SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev),
5986 	    SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)),
5987 	    OID_AUTO, name, CTLTYPE_INT|CTLFLAG_RW,
5988 	    info, 0, iflib_sysctl_int_delay, "I", description);
5989 }
5990 
5991 struct sx *
5992 iflib_ctx_lock_get(if_ctx_t ctx)
5993 {
5994 
5995 	return (&ctx->ifc_ctx_sx);
5996 }
5997 
5998 static int
5999 iflib_msix_init(if_ctx_t ctx)
6000 {
6001 	device_t dev = ctx->ifc_dev;
6002 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6003 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6004 	int vectors, queues, rx_queues, tx_queues, queuemsgs, msgs;
6005 	int iflib_num_tx_queues, iflib_num_rx_queues;
6006 	int err, admincnt, bar;
6007 
6008 	iflib_num_tx_queues = ctx->ifc_sysctl_ntxqs;
6009 	iflib_num_rx_queues = ctx->ifc_sysctl_nrxqs;
6010 
6011 	if (bootverbose)
6012 		device_printf(dev, "msix_init qsets capped at %d\n",
6013 		    imax(scctx->isc_ntxqsets, scctx->isc_nrxqsets));
6014 
6015 	bar = ctx->ifc_softc_ctx.isc_msix_bar;
6016 	admincnt = sctx->isc_admin_intrcnt;
6017 	/* Override by tuneable */
6018 	if (scctx->isc_disable_msix)
6019 		goto msi;
6020 
6021 	/* First try MSI-X */
6022 	if ((msgs = pci_msix_count(dev)) == 0) {
6023 		if (bootverbose)
6024 			device_printf(dev, "MSI-X not supported or disabled\n");
6025 		goto msi;
6026 	}
6027 	/*
6028 	 * bar == -1 => "trust me I know what I'm doing"
6029 	 * Some drivers are for hardware that is so shoddily
6030 	 * documented that no one knows which bars are which
6031 	 * so the developer has to map all bars. This hack
6032 	 * allows shoddy garbage to use MSI-X in this framework.
6033 	 */
6034 	if (bar != -1) {
6035 		ctx->ifc_msix_mem = bus_alloc_resource_any(dev,
6036 	            SYS_RES_MEMORY, &bar, RF_ACTIVE);
6037 		if (ctx->ifc_msix_mem == NULL) {
6038 			device_printf(dev, "Unable to map MSI-X table\n");
6039 			goto msi;
6040 		}
6041 	}
6042 #if IFLIB_DEBUG
6043 	/* use only 1 qset in debug mode */
6044 	queuemsgs = min(msgs - admincnt, 1);
6045 #else
6046 	queuemsgs = msgs - admincnt;
6047 #endif
6048 #ifdef RSS
6049 	queues = imin(queuemsgs, rss_getnumbuckets());
6050 #else
6051 	queues = queuemsgs;
6052 #endif
6053 	queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues);
6054 	if (bootverbose)
6055 		device_printf(dev,
6056 		    "intr CPUs: %d queue msgs: %d admincnt: %d\n",
6057 		    CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt);
6058 #ifdef  RSS
6059 	/* If we're doing RSS, clamp at the number of RSS buckets */
6060 	if (queues > rss_getnumbuckets())
6061 		queues = rss_getnumbuckets();
6062 #endif
6063 	if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt)
6064 		rx_queues = iflib_num_rx_queues;
6065 	else
6066 		rx_queues = queues;
6067 
6068 	if (rx_queues > scctx->isc_nrxqsets)
6069 		rx_queues = scctx->isc_nrxqsets;
6070 
6071 	/*
6072 	 * We want this to be all logical CPUs by default
6073 	 */
6074 	if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues)
6075 		tx_queues = iflib_num_tx_queues;
6076 	else
6077 		tx_queues = mp_ncpus;
6078 
6079 	if (tx_queues > scctx->isc_ntxqsets)
6080 		tx_queues = scctx->isc_ntxqsets;
6081 
6082 	if (ctx->ifc_sysctl_qs_eq_override == 0) {
6083 #ifdef INVARIANTS
6084 		if (tx_queues != rx_queues)
6085 			device_printf(dev,
6086 			    "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n",
6087 			    min(rx_queues, tx_queues), min(rx_queues, tx_queues));
6088 #endif
6089 		tx_queues = min(rx_queues, tx_queues);
6090 		rx_queues = min(rx_queues, tx_queues);
6091 	}
6092 
6093 	device_printf(dev, "Using %d rx queues %d tx queues\n",
6094 	    rx_queues, tx_queues);
6095 
6096 	vectors = rx_queues + admincnt;
6097 	if ((err = pci_alloc_msix(dev, &vectors)) == 0) {
6098 		device_printf(dev, "Using MSI-X interrupts with %d vectors\n",
6099 		    vectors);
6100 		scctx->isc_vectors = vectors;
6101 		scctx->isc_nrxqsets = rx_queues;
6102 		scctx->isc_ntxqsets = tx_queues;
6103 		scctx->isc_intr = IFLIB_INTR_MSIX;
6104 
6105 		return (vectors);
6106 	} else {
6107 		device_printf(dev,
6108 		    "failed to allocate %d MSI-X vectors, err: %d - using MSI\n",
6109 		    vectors, err);
6110 		bus_release_resource(dev, SYS_RES_MEMORY, bar,
6111 		    ctx->ifc_msix_mem);
6112 		ctx->ifc_msix_mem = NULL;
6113 	}
6114 msi:
6115 	vectors = pci_msi_count(dev);
6116 	scctx->isc_nrxqsets = 1;
6117 	scctx->isc_ntxqsets = 1;
6118 	scctx->isc_vectors = vectors;
6119 	if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) {
6120 		device_printf(dev,"Using an MSI interrupt\n");
6121 		scctx->isc_intr = IFLIB_INTR_MSI;
6122 	} else {
6123 		scctx->isc_vectors = 1;
6124 		device_printf(dev,"Using a Legacy interrupt\n");
6125 		scctx->isc_intr = IFLIB_INTR_LEGACY;
6126 	}
6127 
6128 	return (vectors);
6129 }
6130 
6131 static const char *ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" };
6132 
6133 static int
6134 mp_ring_state_handler(SYSCTL_HANDLER_ARGS)
6135 {
6136 	int rc;
6137 	uint16_t *state = ((uint16_t *)oidp->oid_arg1);
6138 	struct sbuf *sb;
6139 	const char *ring_state = "UNKNOWN";
6140 
6141 	/* XXX needed ? */
6142 	rc = sysctl_wire_old_buffer(req, 0);
6143 	MPASS(rc == 0);
6144 	if (rc != 0)
6145 		return (rc);
6146 	sb = sbuf_new_for_sysctl(NULL, NULL, 80, req);
6147 	MPASS(sb != NULL);
6148 	if (sb == NULL)
6149 		return (ENOMEM);
6150 	if (state[3] <= 3)
6151 		ring_state = ring_states[state[3]];
6152 
6153 	sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s",
6154 		    state[0], state[1], state[2], ring_state);
6155 	rc = sbuf_finish(sb);
6156 	sbuf_delete(sb);
6157         return(rc);
6158 }
6159 
6160 enum iflib_ndesc_handler {
6161 	IFLIB_NTXD_HANDLER,
6162 	IFLIB_NRXD_HANDLER,
6163 };
6164 
6165 static int
6166 mp_ndesc_handler(SYSCTL_HANDLER_ARGS)
6167 {
6168 	if_ctx_t ctx = (void *)arg1;
6169 	enum iflib_ndesc_handler type = arg2;
6170 	char buf[256] = {0};
6171 	qidx_t *ndesc;
6172 	char *p, *next;
6173 	int nqs, rc, i;
6174 
6175 	MPASS(type == IFLIB_NTXD_HANDLER || type == IFLIB_NRXD_HANDLER);
6176 
6177 	nqs = 8;
6178 	switch(type) {
6179 	case IFLIB_NTXD_HANDLER:
6180 		ndesc = ctx->ifc_sysctl_ntxds;
6181 		if (ctx->ifc_sctx)
6182 			nqs = ctx->ifc_sctx->isc_ntxqs;
6183 		break;
6184 	case IFLIB_NRXD_HANDLER:
6185 		ndesc = ctx->ifc_sysctl_nrxds;
6186 		if (ctx->ifc_sctx)
6187 			nqs = ctx->ifc_sctx->isc_nrxqs;
6188 		break;
6189 	default:
6190 			panic("unhandled type");
6191 	}
6192 	if (nqs == 0)
6193 		nqs = 8;
6194 
6195 	for (i=0; i<8; i++) {
6196 		if (i >= nqs)
6197 			break;
6198 		if (i)
6199 			strcat(buf, ",");
6200 		sprintf(strchr(buf, 0), "%d", ndesc[i]);
6201 	}
6202 
6203 	rc = sysctl_handle_string(oidp, buf, sizeof(buf), req);
6204 	if (rc || req->newptr == NULL)
6205 		return rc;
6206 
6207 	for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p;
6208 	    i++, p = strsep(&next, " ,")) {
6209 		ndesc[i] = strtoul(p, NULL, 10);
6210 	}
6211 
6212 	return(rc);
6213 }
6214 
6215 #define NAME_BUFLEN 32
6216 static void
6217 iflib_add_device_sysctl_pre(if_ctx_t ctx)
6218 {
6219         device_t dev = iflib_get_dev(ctx);
6220 	struct sysctl_oid_list *child, *oid_list;
6221 	struct sysctl_ctx_list *ctx_list;
6222 	struct sysctl_oid *node;
6223 
6224 	ctx_list = device_get_sysctl_ctx(dev);
6225 	child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev));
6226 	ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, "iflib",
6227 						      CTLFLAG_RD, NULL, "IFLIB fields");
6228 	oid_list = SYSCTL_CHILDREN(node);
6229 
6230 	SYSCTL_ADD_STRING(ctx_list, oid_list, OID_AUTO, "driver_version",
6231 		       CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version, 0,
6232 		       "driver version");
6233 
6234 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs",
6235 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0,
6236 			"# of txqs to use, 0 => use default #");
6237 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs",
6238 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0,
6239 			"# of rxqs to use, 0 => use default #");
6240 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_qs_enable",
6241 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0,
6242                        "permit #txq != #rxq");
6243 	SYSCTL_ADD_INT(ctx_list, oid_list, OID_AUTO, "disable_msix",
6244                       CTLFLAG_RWTUN, &ctx->ifc_softc_ctx.isc_disable_msix, 0,
6245                       "disable MSI-X (default 0)");
6246 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "rx_budget",
6247 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_rx_budget, 0,
6248                        "set the rx budget");
6249 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "tx_abdicate",
6250 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_tx_abdicate, 0,
6251 		       "cause tx to abdicate instead of running to completion");
6252 
6253 	/* XXX change for per-queue sizes */
6254 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_ntxds",
6255 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NTXD_HANDLER,
6256                        mp_ndesc_handler, "A",
6257                        "list of # of tx descriptors to use, 0 = use default #");
6258 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_nrxds",
6259 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NRXD_HANDLER,
6260                        mp_ndesc_handler, "A",
6261                        "list of # of rx descriptors to use, 0 = use default #");
6262 }
6263 
6264 static void
6265 iflib_add_device_sysctl_post(if_ctx_t ctx)
6266 {
6267 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6268 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6269         device_t dev = iflib_get_dev(ctx);
6270 	struct sysctl_oid_list *child;
6271 	struct sysctl_ctx_list *ctx_list;
6272 	iflib_fl_t fl;
6273 	iflib_txq_t txq;
6274 	iflib_rxq_t rxq;
6275 	int i, j;
6276 	char namebuf[NAME_BUFLEN];
6277 	char *qfmt;
6278 	struct sysctl_oid *queue_node, *fl_node, *node;
6279 	struct sysctl_oid_list *queue_list, *fl_list;
6280 	ctx_list = device_get_sysctl_ctx(dev);
6281 
6282 	node = ctx->ifc_sysctl_node;
6283 	child = SYSCTL_CHILDREN(node);
6284 
6285 	if (scctx->isc_ntxqsets > 100)
6286 		qfmt = "txq%03d";
6287 	else if (scctx->isc_ntxqsets > 10)
6288 		qfmt = "txq%02d";
6289 	else
6290 		qfmt = "txq%d";
6291 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
6292 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6293 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6294 					     CTLFLAG_RD, NULL, "Queue Name");
6295 		queue_list = SYSCTL_CHILDREN(queue_node);
6296 #if MEMORY_LOGGING
6297 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued",
6298 				CTLFLAG_RD,
6299 				&txq->ift_dequeued, "total mbufs freed");
6300 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued",
6301 				CTLFLAG_RD,
6302 				&txq->ift_enqueued, "total mbufs enqueued");
6303 #endif
6304 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag",
6305 				   CTLFLAG_RD,
6306 				   &txq->ift_mbuf_defrag, "# of times m_defrag was called");
6307 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "m_pullups",
6308 				   CTLFLAG_RD,
6309 				   &txq->ift_pullups, "# of times m_pullup was called");
6310 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag_failed",
6311 				   CTLFLAG_RD,
6312 				   &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed");
6313 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_desc_avail",
6314 				   CTLFLAG_RD,
6315 				   &txq->ift_no_desc_avail, "# of times no descriptors were available");
6316 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "tx_map_failed",
6317 				   CTLFLAG_RD,
6318 				   &txq->ift_map_failed, "# of times dma map failed");
6319 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txd_encap_efbig",
6320 				   CTLFLAG_RD,
6321 				   &txq->ift_txd_encap_efbig, "# of times txd_encap returned EFBIG");
6322 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_tx_dma_setup",
6323 				   CTLFLAG_RD,
6324 				   &txq->ift_no_tx_dma_setup, "# of times map failed for other than EFBIG");
6325 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx",
6326 				   CTLFLAG_RD,
6327 				   &txq->ift_pidx, 1, "Producer Index");
6328 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx",
6329 				   CTLFLAG_RD,
6330 				   &txq->ift_cidx, 1, "Consumer Index");
6331 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx_processed",
6332 				   CTLFLAG_RD,
6333 				   &txq->ift_cidx_processed, 1, "Consumer Index seen by credit update");
6334 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use",
6335 				   CTLFLAG_RD,
6336 				   &txq->ift_in_use, 1, "descriptors in use");
6337 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_processed",
6338 				   CTLFLAG_RD,
6339 				   &txq->ift_processed, "descriptors procesed for clean");
6340 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned",
6341 				   CTLFLAG_RD,
6342 				   &txq->ift_cleaned, "total cleaned");
6343 		SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state",
6344 				CTLTYPE_STRING | CTLFLAG_RD, __DEVOLATILE(uint64_t *, &txq->ift_br->state),
6345 				0, mp_ring_state_handler, "A", "soft ring state");
6346 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_enqueues",
6347 				       CTLFLAG_RD, &txq->ift_br->enqueues,
6348 				       "# of enqueues to the mp_ring for this queue");
6349 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_drops",
6350 				       CTLFLAG_RD, &txq->ift_br->drops,
6351 				       "# of drops in the mp_ring for this queue");
6352 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_starts",
6353 				       CTLFLAG_RD, &txq->ift_br->starts,
6354 				       "# of normal consumer starts in the mp_ring for this queue");
6355 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_stalls",
6356 				       CTLFLAG_RD, &txq->ift_br->stalls,
6357 					       "# of consumer stalls in the mp_ring for this queue");
6358 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_restarts",
6359 			       CTLFLAG_RD, &txq->ift_br->restarts,
6360 				       "# of consumer restarts in the mp_ring for this queue");
6361 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_abdications",
6362 				       CTLFLAG_RD, &txq->ift_br->abdications,
6363 				       "# of consumer abdications in the mp_ring for this queue");
6364 	}
6365 
6366 	if (scctx->isc_nrxqsets > 100)
6367 		qfmt = "rxq%03d";
6368 	else if (scctx->isc_nrxqsets > 10)
6369 		qfmt = "rxq%02d";
6370 	else
6371 		qfmt = "rxq%d";
6372 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
6373 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6374 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6375 					     CTLFLAG_RD, NULL, "Queue Name");
6376 		queue_list = SYSCTL_CHILDREN(queue_node);
6377 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
6378 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_pidx",
6379 				       CTLFLAG_RD,
6380 				       &rxq->ifr_cq_pidx, 1, "Producer Index");
6381 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_cidx",
6382 				       CTLFLAG_RD,
6383 				       &rxq->ifr_cq_cidx, 1, "Consumer Index");
6384 		}
6385 
6386 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
6387 			snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j);
6388 			fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list, OID_AUTO, namebuf,
6389 						     CTLFLAG_RD, NULL, "freelist Name");
6390 			fl_list = SYSCTL_CHILDREN(fl_node);
6391 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx",
6392 				       CTLFLAG_RD,
6393 				       &fl->ifl_pidx, 1, "Producer Index");
6394 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx",
6395 				       CTLFLAG_RD,
6396 				       &fl->ifl_cidx, 1, "Consumer Index");
6397 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits",
6398 				       CTLFLAG_RD,
6399 				       &fl->ifl_credits, 1, "credits available");
6400 #if MEMORY_LOGGING
6401 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_enqueued",
6402 					CTLFLAG_RD,
6403 					&fl->ifl_m_enqueued, "mbufs allocated");
6404 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_dequeued",
6405 					CTLFLAG_RD,
6406 					&fl->ifl_m_dequeued, "mbufs freed");
6407 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_enqueued",
6408 					CTLFLAG_RD,
6409 					&fl->ifl_cl_enqueued, "clusters allocated");
6410 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_dequeued",
6411 					CTLFLAG_RD,
6412 					&fl->ifl_cl_dequeued, "clusters freed");
6413 #endif
6414 
6415 		}
6416 	}
6417 
6418 }
6419 
6420 void
6421 iflib_request_reset(if_ctx_t ctx)
6422 {
6423 
6424 	STATE_LOCK(ctx);
6425 	ctx->ifc_flags |= IFC_DO_RESET;
6426 	STATE_UNLOCK(ctx);
6427 }
6428 
6429 #ifndef __NO_STRICT_ALIGNMENT
6430 static struct mbuf *
6431 iflib_fixup_rx(struct mbuf *m)
6432 {
6433 	struct mbuf *n;
6434 
6435 	if (m->m_len <= (MCLBYTES - ETHER_HDR_LEN)) {
6436 		bcopy(m->m_data, m->m_data + ETHER_HDR_LEN, m->m_len);
6437 		m->m_data += ETHER_HDR_LEN;
6438 		n = m;
6439 	} else {
6440 		MGETHDR(n, M_NOWAIT, MT_DATA);
6441 		if (n == NULL) {
6442 			m_freem(m);
6443 			return (NULL);
6444 		}
6445 		bcopy(m->m_data, n->m_data, ETHER_HDR_LEN);
6446 		m->m_data += ETHER_HDR_LEN;
6447 		m->m_len -= ETHER_HDR_LEN;
6448 		n->m_len = ETHER_HDR_LEN;
6449 		M_MOVE_PKTHDR(n, m);
6450 		n->m_next = m;
6451 	}
6452 	return (n);
6453 }
6454 #endif
6455 
6456 #ifdef NETDUMP
6457 static void
6458 iflib_netdump_init(struct ifnet *ifp, int *nrxr, int *ncl, int *clsize)
6459 {
6460 	if_ctx_t ctx;
6461 
6462 	ctx = if_getsoftc(ifp);
6463 	CTX_LOCK(ctx);
6464 	*nrxr = NRXQSETS(ctx);
6465 	*ncl = ctx->ifc_rxqs[0].ifr_fl->ifl_size;
6466 	*clsize = ctx->ifc_rxqs[0].ifr_fl->ifl_buf_size;
6467 	CTX_UNLOCK(ctx);
6468 }
6469 
6470 static void
6471 iflib_netdump_event(struct ifnet *ifp, enum netdump_ev event)
6472 {
6473 	if_ctx_t ctx;
6474 	if_softc_ctx_t scctx;
6475 	iflib_fl_t fl;
6476 	iflib_rxq_t rxq;
6477 	int i, j;
6478 
6479 	ctx = if_getsoftc(ifp);
6480 	scctx = &ctx->ifc_softc_ctx;
6481 
6482 	switch (event) {
6483 	case NETDUMP_START:
6484 		for (i = 0; i < scctx->isc_nrxqsets; i++) {
6485 			rxq = &ctx->ifc_rxqs[i];
6486 			for (j = 0; j < rxq->ifr_nfl; j++) {
6487 				fl = rxq->ifr_fl;
6488 				fl->ifl_zone = m_getzone(fl->ifl_buf_size);
6489 			}
6490 		}
6491 		iflib_no_tx_batch = 1;
6492 		break;
6493 	default:
6494 		break;
6495 	}
6496 }
6497 
6498 static int
6499 iflib_netdump_transmit(struct ifnet *ifp, struct mbuf *m)
6500 {
6501 	if_ctx_t ctx;
6502 	iflib_txq_t txq;
6503 	int error;
6504 
6505 	ctx = if_getsoftc(ifp);
6506 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
6507 	    IFF_DRV_RUNNING)
6508 		return (EBUSY);
6509 
6510 	txq = &ctx->ifc_txqs[0];
6511 	error = iflib_encap(txq, &m);
6512 	if (error == 0)
6513 		(void)iflib_txd_db_check(ctx, txq, true, txq->ift_in_use);
6514 	return (error);
6515 }
6516 
6517 static int
6518 iflib_netdump_poll(struct ifnet *ifp, int count)
6519 {
6520 	if_ctx_t ctx;
6521 	if_softc_ctx_t scctx;
6522 	iflib_txq_t txq;
6523 	int i;
6524 
6525 	ctx = if_getsoftc(ifp);
6526 	scctx = &ctx->ifc_softc_ctx;
6527 
6528 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
6529 	    IFF_DRV_RUNNING)
6530 		return (EBUSY);
6531 
6532 	txq = &ctx->ifc_txqs[0];
6533 	(void)iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
6534 
6535 	for (i = 0; i < scctx->isc_nrxqsets; i++)
6536 		(void)iflib_rxeof(&ctx->ifc_rxqs[i], 16 /* XXX */);
6537 	return (0);
6538 }
6539 #endif /* NETDUMP */
6540