xref: /freebsd/sys/net/iflib.c (revision 77ebcc05eac2658a68b447e654cfdf7ff3e703b8)
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 	struct mbuf    **ifsd_m;           /* pkthdr mbufs */
286 } if_txsd_vec_t;
287 
288 
289 /* magic number that should be high enough for any hardware */
290 #define IFLIB_MAX_TX_SEGS		128
291 /* bnxt supports 64 with hardware LRO enabled */
292 #define IFLIB_MAX_RX_SEGS		64
293 #define IFLIB_RX_COPY_THRESH		128
294 #define IFLIB_MAX_RX_REFRESH		32
295 /* The minimum descriptors per second before we start coalescing */
296 #define IFLIB_MIN_DESC_SEC		16384
297 #define IFLIB_DEFAULT_TX_UPDATE_FREQ	16
298 #define IFLIB_QUEUE_IDLE		0
299 #define IFLIB_QUEUE_HUNG		1
300 #define IFLIB_QUEUE_WORKING		2
301 /* maximum number of txqs that can share an rx interrupt */
302 #define IFLIB_MAX_TX_SHARED_INTR	4
303 
304 /* this should really scale with ring size - this is a fairly arbitrary value */
305 #define TX_BATCH_SIZE			32
306 
307 #define IFLIB_RESTART_BUDGET		8
308 
309 
310 #define CSUM_OFFLOAD		(CSUM_IP_TSO|CSUM_IP6_TSO|CSUM_IP| \
311 				 CSUM_IP_UDP|CSUM_IP_TCP|CSUM_IP_SCTP| \
312 				 CSUM_IP6_UDP|CSUM_IP6_TCP|CSUM_IP6_SCTP)
313 struct iflib_txq {
314 	qidx_t		ift_in_use;
315 	qidx_t		ift_cidx;
316 	qidx_t		ift_cidx_processed;
317 	qidx_t		ift_pidx;
318 	uint8_t		ift_gen;
319 	uint8_t		ift_br_offset;
320 	uint16_t	ift_npending;
321 	uint16_t	ift_db_pending;
322 	uint16_t	ift_rs_pending;
323 	/* implicit pad */
324 	uint8_t		ift_txd_size[8];
325 	uint64_t	ift_processed;
326 	uint64_t	ift_cleaned;
327 	uint64_t	ift_cleaned_prev;
328 #if MEMORY_LOGGING
329 	uint64_t	ift_enqueued;
330 	uint64_t	ift_dequeued;
331 #endif
332 	uint64_t	ift_no_tx_dma_setup;
333 	uint64_t	ift_no_desc_avail;
334 	uint64_t	ift_mbuf_defrag_failed;
335 	uint64_t	ift_mbuf_defrag;
336 	uint64_t	ift_map_failed;
337 	uint64_t	ift_txd_encap_efbig;
338 	uint64_t	ift_pullups;
339 	uint64_t	ift_last_timer_tick;
340 
341 	struct mtx	ift_mtx;
342 	struct mtx	ift_db_mtx;
343 
344 	/* constant values */
345 	if_ctx_t	ift_ctx;
346 	struct ifmp_ring        *ift_br;
347 	struct grouptask	ift_task;
348 	qidx_t		ift_size;
349 	uint16_t	ift_id;
350 	struct callout	ift_timer;
351 
352 	if_txsd_vec_t	ift_sds;
353 	uint8_t		ift_qstatus;
354 	uint8_t		ift_closed;
355 	uint8_t		ift_update_freq;
356 	struct iflib_filter_info ift_filter_info;
357 	bus_dma_tag_t		ift_desc_tag;
358 	bus_dma_tag_t		ift_tso_desc_tag;
359 	iflib_dma_info_t	ift_ifdi;
360 #define MTX_NAME_LEN 16
361 	char                    ift_mtx_name[MTX_NAME_LEN];
362 	char                    ift_db_mtx_name[MTX_NAME_LEN];
363 	bus_dma_segment_t	ift_segs[IFLIB_MAX_TX_SEGS]  __aligned(CACHE_LINE_SIZE);
364 #ifdef IFLIB_DIAGNOSTICS
365 	uint64_t ift_cpu_exec_count[256];
366 #endif
367 } __aligned(CACHE_LINE_SIZE);
368 
369 struct iflib_fl {
370 	qidx_t		ifl_cidx;
371 	qidx_t		ifl_pidx;
372 	qidx_t		ifl_credits;
373 	uint8_t		ifl_gen;
374 	uint8_t		ifl_rxd_size;
375 #if MEMORY_LOGGING
376 	uint64_t	ifl_m_enqueued;
377 	uint64_t	ifl_m_dequeued;
378 	uint64_t	ifl_cl_enqueued;
379 	uint64_t	ifl_cl_dequeued;
380 #endif
381 	/* implicit pad */
382 
383 	bitstr_t 	*ifl_rx_bitmap;
384 	qidx_t		ifl_fragidx;
385 	/* constant */
386 	qidx_t		ifl_size;
387 	uint16_t	ifl_buf_size;
388 	uint16_t	ifl_cltype;
389 	uma_zone_t	ifl_zone;
390 	iflib_rxsd_array_t	ifl_sds;
391 	iflib_rxq_t	ifl_rxq;
392 	uint8_t		ifl_id;
393 	bus_dma_tag_t           ifl_desc_tag;
394 	iflib_dma_info_t	ifl_ifdi;
395 	uint64_t	ifl_bus_addrs[IFLIB_MAX_RX_REFRESH] __aligned(CACHE_LINE_SIZE);
396 	caddr_t		ifl_vm_addrs[IFLIB_MAX_RX_REFRESH];
397 	qidx_t	ifl_rxd_idxs[IFLIB_MAX_RX_REFRESH];
398 }  __aligned(CACHE_LINE_SIZE);
399 
400 static inline qidx_t
401 get_inuse(int size, qidx_t cidx, qidx_t pidx, uint8_t gen)
402 {
403 	qidx_t used;
404 
405 	if (pidx > cidx)
406 		used = pidx - cidx;
407 	else if (pidx < cidx)
408 		used = size - cidx + pidx;
409 	else if (gen == 0 && pidx == cidx)
410 		used = 0;
411 	else if (gen == 1 && pidx == cidx)
412 		used = size;
413 	else
414 		panic("bad state");
415 
416 	return (used);
417 }
418 
419 #define TXQ_AVAIL(txq) (txq->ift_size - get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx, txq->ift_gen))
420 
421 #define IDXDIFF(head, tail, wrap) \
422 	((head) >= (tail) ? (head) - (tail) : (wrap) - (tail) + (head))
423 
424 struct iflib_rxq {
425 	/* If there is a separate completion queue -
426 	 * these are the cq cidx and pidx. Otherwise
427 	 * these are unused.
428 	 */
429 	qidx_t		ifr_size;
430 	qidx_t		ifr_cq_cidx;
431 	qidx_t		ifr_cq_pidx;
432 	uint8_t		ifr_cq_gen;
433 	uint8_t		ifr_fl_offset;
434 
435 	if_ctx_t	ifr_ctx;
436 	iflib_fl_t	ifr_fl;
437 	uint64_t	ifr_rx_irq;
438 	uint16_t	ifr_id;
439 	uint8_t		ifr_lro_enabled;
440 	uint8_t		ifr_nfl;
441 	uint8_t		ifr_ntxqirq;
442 	uint8_t		ifr_txqid[IFLIB_MAX_TX_SHARED_INTR];
443 	struct lro_ctrl			ifr_lc;
444 	struct grouptask        ifr_task;
445 	struct iflib_filter_info ifr_filter_info;
446 	iflib_dma_info_t		ifr_ifdi;
447 
448 	/* dynamically allocate if any drivers need a value substantially larger than this */
449 	struct if_rxd_frag	ifr_frags[IFLIB_MAX_RX_SEGS] __aligned(CACHE_LINE_SIZE);
450 #ifdef IFLIB_DIAGNOSTICS
451 	uint64_t ifr_cpu_exec_count[256];
452 #endif
453 }  __aligned(CACHE_LINE_SIZE);
454 
455 typedef struct if_rxsd {
456 	caddr_t *ifsd_cl;
457 	struct mbuf **ifsd_m;
458 	iflib_fl_t ifsd_fl;
459 	qidx_t ifsd_cidx;
460 } *if_rxsd_t;
461 
462 /* multiple of word size */
463 #ifdef __LP64__
464 #define PKT_INFO_SIZE	6
465 #define RXD_INFO_SIZE	5
466 #define PKT_TYPE uint64_t
467 #else
468 #define PKT_INFO_SIZE	11
469 #define RXD_INFO_SIZE	8
470 #define PKT_TYPE uint32_t
471 #endif
472 #define PKT_LOOP_BOUND  ((PKT_INFO_SIZE/3)*3)
473 #define RXD_LOOP_BOUND  ((RXD_INFO_SIZE/4)*4)
474 
475 typedef struct if_pkt_info_pad {
476 	PKT_TYPE pkt_val[PKT_INFO_SIZE];
477 } *if_pkt_info_pad_t;
478 typedef struct if_rxd_info_pad {
479 	PKT_TYPE rxd_val[RXD_INFO_SIZE];
480 } *if_rxd_info_pad_t;
481 
482 CTASSERT(sizeof(struct if_pkt_info_pad) == sizeof(struct if_pkt_info));
483 CTASSERT(sizeof(struct if_rxd_info_pad) == sizeof(struct if_rxd_info));
484 
485 
486 static inline void
487 pkt_info_zero(if_pkt_info_t pi)
488 {
489 	if_pkt_info_pad_t pi_pad;
490 
491 	pi_pad = (if_pkt_info_pad_t)pi;
492 	pi_pad->pkt_val[0] = 0; pi_pad->pkt_val[1] = 0; pi_pad->pkt_val[2] = 0;
493 	pi_pad->pkt_val[3] = 0; pi_pad->pkt_val[4] = 0; pi_pad->pkt_val[5] = 0;
494 #ifndef __LP64__
495 	pi_pad->pkt_val[6] = 0; pi_pad->pkt_val[7] = 0; pi_pad->pkt_val[8] = 0;
496 	pi_pad->pkt_val[9] = 0; pi_pad->pkt_val[10] = 0;
497 #endif
498 }
499 
500 static device_method_t iflib_pseudo_methods[] = {
501 	DEVMETHOD(device_attach, noop_attach),
502 	DEVMETHOD(device_detach, iflib_pseudo_detach),
503 	DEVMETHOD_END
504 };
505 
506 driver_t iflib_pseudodriver = {
507 	"iflib_pseudo", iflib_pseudo_methods, sizeof(struct iflib_ctx),
508 };
509 
510 static inline void
511 rxd_info_zero(if_rxd_info_t ri)
512 {
513 	if_rxd_info_pad_t ri_pad;
514 	int i;
515 
516 	ri_pad = (if_rxd_info_pad_t)ri;
517 	for (i = 0; i < RXD_LOOP_BOUND; i += 4) {
518 		ri_pad->rxd_val[i] = 0;
519 		ri_pad->rxd_val[i+1] = 0;
520 		ri_pad->rxd_val[i+2] = 0;
521 		ri_pad->rxd_val[i+3] = 0;
522 	}
523 #ifdef __LP64__
524 	ri_pad->rxd_val[RXD_INFO_SIZE-1] = 0;
525 #endif
526 }
527 
528 /*
529  * Only allow a single packet to take up most 1/nth of the tx ring
530  */
531 #define MAX_SINGLE_PACKET_FRACTION 12
532 #define IF_BAD_DMA (bus_addr_t)-1
533 
534 #define CTX_ACTIVE(ctx) ((if_getdrvflags((ctx)->ifc_ifp) & IFF_DRV_RUNNING))
535 
536 #define CTX_LOCK_INIT(_sc)  sx_init(&(_sc)->ifc_ctx_sx, "iflib ctx lock")
537 #define CTX_LOCK(ctx) sx_xlock(&(ctx)->ifc_ctx_sx)
538 #define CTX_UNLOCK(ctx) sx_xunlock(&(ctx)->ifc_ctx_sx)
539 #define CTX_LOCK_DESTROY(ctx) sx_destroy(&(ctx)->ifc_ctx_sx)
540 
541 
542 #define STATE_LOCK_INIT(_sc, _name)  mtx_init(&(_sc)->ifc_state_mtx, _name, "iflib state lock", MTX_DEF)
543 #define STATE_LOCK(ctx) mtx_lock(&(ctx)->ifc_state_mtx)
544 #define STATE_UNLOCK(ctx) mtx_unlock(&(ctx)->ifc_state_mtx)
545 #define STATE_LOCK_DESTROY(ctx) mtx_destroy(&(ctx)->ifc_state_mtx)
546 
547 
548 
549 #define CALLOUT_LOCK(txq)	mtx_lock(&txq->ift_mtx)
550 #define CALLOUT_UNLOCK(txq) 	mtx_unlock(&txq->ift_mtx)
551 
552 void
553 iflib_set_detach(if_ctx_t ctx)
554 {
555 	STATE_LOCK(ctx);
556 	ctx->ifc_flags |= IFC_IN_DETACH;
557 	STATE_UNLOCK(ctx);
558 }
559 
560 /* Our boot-time initialization hook */
561 static int	iflib_module_event_handler(module_t, int, void *);
562 
563 static moduledata_t iflib_moduledata = {
564 	"iflib",
565 	iflib_module_event_handler,
566 	NULL
567 };
568 
569 DECLARE_MODULE(iflib, iflib_moduledata, SI_SUB_INIT_IF, SI_ORDER_ANY);
570 MODULE_VERSION(iflib, 1);
571 
572 MODULE_DEPEND(iflib, pci, 1, 1, 1);
573 MODULE_DEPEND(iflib, ether, 1, 1, 1);
574 
575 TASKQGROUP_DEFINE(if_io_tqg, mp_ncpus, 1);
576 TASKQGROUP_DEFINE(if_config_tqg, 1, 1);
577 
578 #ifndef IFLIB_DEBUG_COUNTERS
579 #ifdef INVARIANTS
580 #define IFLIB_DEBUG_COUNTERS 1
581 #else
582 #define IFLIB_DEBUG_COUNTERS 0
583 #endif /* !INVARIANTS */
584 #endif
585 
586 static SYSCTL_NODE(_net, OID_AUTO, iflib, CTLFLAG_RD, 0,
587                    "iflib driver parameters");
588 
589 /*
590  * XXX need to ensure that this can't accidentally cause the head to be moved backwards
591  */
592 static int iflib_min_tx_latency = 0;
593 SYSCTL_INT(_net_iflib, OID_AUTO, min_tx_latency, CTLFLAG_RW,
594 		   &iflib_min_tx_latency, 0, "minimize transmit latency at the possible expense of throughput");
595 static int iflib_no_tx_batch = 0;
596 SYSCTL_INT(_net_iflib, OID_AUTO, no_tx_batch, CTLFLAG_RW,
597 		   &iflib_no_tx_batch, 0, "minimize transmit latency at the possible expense of throughput");
598 
599 
600 #if IFLIB_DEBUG_COUNTERS
601 
602 static int iflib_tx_seen;
603 static int iflib_tx_sent;
604 static int iflib_tx_encap;
605 static int iflib_rx_allocs;
606 static int iflib_fl_refills;
607 static int iflib_fl_refills_large;
608 static int iflib_tx_frees;
609 
610 SYSCTL_INT(_net_iflib, OID_AUTO, tx_seen, CTLFLAG_RD,
611 		   &iflib_tx_seen, 0, "# tx mbufs seen");
612 SYSCTL_INT(_net_iflib, OID_AUTO, tx_sent, CTLFLAG_RD,
613 		   &iflib_tx_sent, 0, "# tx mbufs sent");
614 SYSCTL_INT(_net_iflib, OID_AUTO, tx_encap, CTLFLAG_RD,
615 		   &iflib_tx_encap, 0, "# tx mbufs encapped");
616 SYSCTL_INT(_net_iflib, OID_AUTO, tx_frees, CTLFLAG_RD,
617 		   &iflib_tx_frees, 0, "# tx frees");
618 SYSCTL_INT(_net_iflib, OID_AUTO, rx_allocs, CTLFLAG_RD,
619 		   &iflib_rx_allocs, 0, "# rx allocations");
620 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills, CTLFLAG_RD,
621 		   &iflib_fl_refills, 0, "# refills");
622 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills_large, CTLFLAG_RD,
623 		   &iflib_fl_refills_large, 0, "# large refills");
624 
625 
626 static int iflib_txq_drain_flushing;
627 static int iflib_txq_drain_oactive;
628 static int iflib_txq_drain_notready;
629 
630 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_flushing, CTLFLAG_RD,
631 		   &iflib_txq_drain_flushing, 0, "# drain flushes");
632 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_oactive, CTLFLAG_RD,
633 		   &iflib_txq_drain_oactive, 0, "# drain oactives");
634 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_notready, CTLFLAG_RD,
635 		   &iflib_txq_drain_notready, 0, "# drain notready");
636 
637 
638 static int iflib_encap_load_mbuf_fail;
639 static int iflib_encap_pad_mbuf_fail;
640 static int iflib_encap_txq_avail_fail;
641 static int iflib_encap_txd_encap_fail;
642 
643 SYSCTL_INT(_net_iflib, OID_AUTO, encap_load_mbuf_fail, CTLFLAG_RD,
644 		   &iflib_encap_load_mbuf_fail, 0, "# busdma load failures");
645 SYSCTL_INT(_net_iflib, OID_AUTO, encap_pad_mbuf_fail, CTLFLAG_RD,
646 		   &iflib_encap_pad_mbuf_fail, 0, "# runt frame pad failures");
647 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txq_avail_fail, CTLFLAG_RD,
648 		   &iflib_encap_txq_avail_fail, 0, "# txq avail failures");
649 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txd_encap_fail, CTLFLAG_RD,
650 		   &iflib_encap_txd_encap_fail, 0, "# driver encap failures");
651 
652 static int iflib_task_fn_rxs;
653 static int iflib_rx_intr_enables;
654 static int iflib_fast_intrs;
655 static int iflib_rx_unavail;
656 static int iflib_rx_ctx_inactive;
657 static int iflib_rx_if_input;
658 static int iflib_rx_mbuf_null;
659 static int iflib_rxd_flush;
660 
661 static int iflib_verbose_debug;
662 
663 SYSCTL_INT(_net_iflib, OID_AUTO, task_fn_rx, CTLFLAG_RD,
664 		   &iflib_task_fn_rxs, 0, "# task_fn_rx calls");
665 SYSCTL_INT(_net_iflib, OID_AUTO, rx_intr_enables, CTLFLAG_RD,
666 		   &iflib_rx_intr_enables, 0, "# rx intr enables");
667 SYSCTL_INT(_net_iflib, OID_AUTO, fast_intrs, CTLFLAG_RD,
668 		   &iflib_fast_intrs, 0, "# fast_intr calls");
669 SYSCTL_INT(_net_iflib, OID_AUTO, rx_unavail, CTLFLAG_RD,
670 		   &iflib_rx_unavail, 0, "# times rxeof called with no available data");
671 SYSCTL_INT(_net_iflib, OID_AUTO, rx_ctx_inactive, CTLFLAG_RD,
672 		   &iflib_rx_ctx_inactive, 0, "# times rxeof called with inactive context");
673 SYSCTL_INT(_net_iflib, OID_AUTO, rx_if_input, CTLFLAG_RD,
674 		   &iflib_rx_if_input, 0, "# times rxeof called if_input");
675 SYSCTL_INT(_net_iflib, OID_AUTO, rx_mbuf_null, CTLFLAG_RD,
676 		   &iflib_rx_mbuf_null, 0, "# times rxeof got null mbuf");
677 SYSCTL_INT(_net_iflib, OID_AUTO, rxd_flush, CTLFLAG_RD,
678 	         &iflib_rxd_flush, 0, "# times rxd_flush called");
679 SYSCTL_INT(_net_iflib, OID_AUTO, verbose_debug, CTLFLAG_RW,
680 		   &iflib_verbose_debug, 0, "enable verbose debugging");
681 
682 #define DBG_COUNTER_INC(name) atomic_add_int(&(iflib_ ## name), 1)
683 static void
684 iflib_debug_reset(void)
685 {
686 	iflib_tx_seen = iflib_tx_sent = iflib_tx_encap = iflib_rx_allocs =
687 		iflib_fl_refills = iflib_fl_refills_large = iflib_tx_frees =
688 		iflib_txq_drain_flushing = iflib_txq_drain_oactive =
689 		iflib_txq_drain_notready =
690 		iflib_encap_load_mbuf_fail = iflib_encap_pad_mbuf_fail =
691 		iflib_encap_txq_avail_fail = iflib_encap_txd_encap_fail =
692 		iflib_task_fn_rxs = iflib_rx_intr_enables = iflib_fast_intrs =
693 		iflib_rx_unavail =
694 		iflib_rx_ctx_inactive = iflib_rx_if_input =
695 		iflib_rx_mbuf_null = iflib_rxd_flush = 0;
696 }
697 
698 #else
699 #define DBG_COUNTER_INC(name)
700 static void iflib_debug_reset(void) {}
701 #endif
702 
703 #define IFLIB_DEBUG 0
704 
705 static void iflib_tx_structures_free(if_ctx_t ctx);
706 static void iflib_rx_structures_free(if_ctx_t ctx);
707 static int iflib_queues_alloc(if_ctx_t ctx);
708 static int iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq);
709 static int iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget);
710 static int iflib_qset_structures_setup(if_ctx_t ctx);
711 static int iflib_msix_init(if_ctx_t ctx);
712 static int iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filterarg, int *rid, const char *str);
713 static void iflib_txq_check_drain(iflib_txq_t txq, int budget);
714 static uint32_t iflib_txq_can_drain(struct ifmp_ring *);
715 #ifdef ALTQ
716 static void iflib_altq_if_start(if_t ifp);
717 static int iflib_altq_if_transmit(if_t ifp, struct mbuf *m);
718 #endif
719 static int iflib_register(if_ctx_t);
720 static void iflib_init_locked(if_ctx_t ctx);
721 static void iflib_add_device_sysctl_pre(if_ctx_t ctx);
722 static void iflib_add_device_sysctl_post(if_ctx_t ctx);
723 static void iflib_ifmp_purge(iflib_txq_t txq);
724 static void _iflib_pre_assert(if_softc_ctx_t scctx);
725 static void iflib_if_init_locked(if_ctx_t ctx);
726 static void iflib_free_intr_mem(if_ctx_t ctx);
727 #ifndef __NO_STRICT_ALIGNMENT
728 static struct mbuf * iflib_fixup_rx(struct mbuf *m);
729 #endif
730 
731 NETDUMP_DEFINE(iflib);
732 
733 #ifdef DEV_NETMAP
734 #include <sys/selinfo.h>
735 #include <net/netmap.h>
736 #include <dev/netmap/netmap_kern.h>
737 
738 MODULE_DEPEND(iflib, netmap, 1, 1, 1);
739 
740 static int netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, uint32_t nm_i, bool init);
741 
742 /*
743  * device-specific sysctl variables:
744  *
745  * iflib_crcstrip: 0: keep CRC in rx frames (default), 1: strip it.
746  *	During regular operations the CRC is stripped, but on some
747  *	hardware reception of frames not multiple of 64 is slower,
748  *	so using crcstrip=0 helps in benchmarks.
749  *
750  * iflib_rx_miss, iflib_rx_miss_bufs:
751  *	count packets that might be missed due to lost interrupts.
752  */
753 SYSCTL_DECL(_dev_netmap);
754 /*
755  * The xl driver by default strips CRCs and we do not override it.
756  */
757 
758 int iflib_crcstrip = 1;
759 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_crcstrip,
760     CTLFLAG_RW, &iflib_crcstrip, 1, "strip CRC on rx frames");
761 
762 int iflib_rx_miss, iflib_rx_miss_bufs;
763 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss,
764     CTLFLAG_RW, &iflib_rx_miss, 0, "potentially missed rx intr");
765 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss_bufs,
766     CTLFLAG_RW, &iflib_rx_miss_bufs, 0, "potentially missed rx intr bufs");
767 
768 /*
769  * Register/unregister. We are already under netmap lock.
770  * Only called on the first register or the last unregister.
771  */
772 static int
773 iflib_netmap_register(struct netmap_adapter *na, int onoff)
774 {
775 	struct ifnet *ifp = na->ifp;
776 	if_ctx_t ctx = ifp->if_softc;
777 	int status;
778 
779 	CTX_LOCK(ctx);
780 	IFDI_INTR_DISABLE(ctx);
781 
782 	/* Tell the stack that the interface is no longer active */
783 	ifp->if_drv_flags &= ~(IFF_DRV_RUNNING | IFF_DRV_OACTIVE);
784 
785 	if (!CTX_IS_VF(ctx))
786 		IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip);
787 
788 	/* enable or disable flags and callbacks in na and ifp */
789 	if (onoff) {
790 		nm_set_native_flags(na);
791 	} else {
792 		nm_clear_native_flags(na);
793 	}
794 	iflib_stop(ctx);
795 	iflib_init_locked(ctx);
796 	IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip); // XXX why twice ?
797 	status = ifp->if_drv_flags & IFF_DRV_RUNNING ? 0 : 1;
798 	if (status)
799 		nm_clear_native_flags(na);
800 	CTX_UNLOCK(ctx);
801 	return (status);
802 }
803 
804 static int
805 netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, uint32_t nm_i, bool init)
806 {
807 	struct netmap_adapter *na = kring->na;
808 	u_int const lim = kring->nkr_num_slots - 1;
809 	u_int head = kring->rhead;
810 	struct netmap_ring *ring = kring->ring;
811 	bus_dmamap_t *map;
812 	struct if_rxd_update iru;
813 	if_ctx_t ctx = rxq->ifr_ctx;
814 	iflib_fl_t fl = &rxq->ifr_fl[0];
815 	uint32_t refill_pidx, nic_i;
816 #if IFLIB_DEBUG_COUNTERS
817 	int rf_count = 0;
818 #endif
819 
820 	if (nm_i == head && __predict_true(!init))
821 		return 0;
822 	iru_init(&iru, rxq, 0 /* flid */);
823 	map = fl->ifl_sds.ifsd_map;
824 	refill_pidx = netmap_idx_k2n(kring, nm_i);
825 	/*
826 	 * IMPORTANT: we must leave one free slot in the ring,
827 	 * so move head back by one unit
828 	 */
829 	head = nm_prev(head, lim);
830 	nic_i = UINT_MAX;
831 	DBG_COUNTER_INC(fl_refills);
832 	while (nm_i != head) {
833 #if IFLIB_DEBUG_COUNTERS
834 		if (++rf_count == 9)
835 			DBG_COUNTER_INC(fl_refills_large);
836 #endif
837 		for (int tmp_pidx = 0; tmp_pidx < IFLIB_MAX_RX_REFRESH && nm_i != head; tmp_pidx++) {
838 			struct netmap_slot *slot = &ring->slot[nm_i];
839 			void *addr = PNMB(na, slot, &fl->ifl_bus_addrs[tmp_pidx]);
840 			uint32_t nic_i_dma = refill_pidx;
841 			nic_i = netmap_idx_k2n(kring, nm_i);
842 
843 			MPASS(tmp_pidx < IFLIB_MAX_RX_REFRESH);
844 
845 			if (addr == NETMAP_BUF_BASE(na)) /* bad buf */
846 			        return netmap_ring_reinit(kring);
847 
848 			fl->ifl_vm_addrs[tmp_pidx] = addr;
849 			if (__predict_false(init) && map) {
850 				netmap_load_map(na, fl->ifl_ifdi->idi_tag, map[nic_i], addr);
851 			} else if (map && (slot->flags & NS_BUF_CHANGED)) {
852 				/* buffer has changed, reload map */
853 				netmap_reload_map(na, fl->ifl_ifdi->idi_tag, map[nic_i], addr);
854 			}
855 			slot->flags &= ~NS_BUF_CHANGED;
856 
857 			nm_i = nm_next(nm_i, lim);
858 			fl->ifl_rxd_idxs[tmp_pidx] = nic_i = nm_next(nic_i, lim);
859 			if (nm_i != head && tmp_pidx < IFLIB_MAX_RX_REFRESH-1)
860 				continue;
861 
862 			iru.iru_pidx = refill_pidx;
863 			iru.iru_count = tmp_pidx+1;
864 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
865 
866 			refill_pidx = nic_i;
867 			if (map == NULL)
868 				continue;
869 
870 			for (int n = 0; n < iru.iru_count; n++) {
871 				bus_dmamap_sync(fl->ifl_ifdi->idi_tag, map[nic_i_dma],
872 						BUS_DMASYNC_PREREAD);
873 				/* XXX - change this to not use the netmap func*/
874 				nic_i_dma = nm_next(nic_i_dma, lim);
875 			}
876 		}
877 	}
878 	kring->nr_hwcur = head;
879 
880 	if (map)
881 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
882 				BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
883 	if (__predict_true(nic_i != UINT_MAX)) {
884 		ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id, nic_i);
885 		DBG_COUNTER_INC(rxd_flush);
886 	}
887 	return (0);
888 }
889 
890 /*
891  * Reconcile kernel and user view of the transmit ring.
892  *
893  * All information is in the kring.
894  * Userspace wants to send packets up to the one before kring->rhead,
895  * kernel knows kring->nr_hwcur is the first unsent packet.
896  *
897  * Here we push packets out (as many as possible), and possibly
898  * reclaim buffers from previously completed transmission.
899  *
900  * The caller (netmap) guarantees that there is only one instance
901  * running at any time. Any interference with other driver
902  * methods should be handled by the individual drivers.
903  */
904 static int
905 iflib_netmap_txsync(struct netmap_kring *kring, int flags)
906 {
907 	struct netmap_adapter *na = kring->na;
908 	struct ifnet *ifp = na->ifp;
909 	struct netmap_ring *ring = kring->ring;
910 	u_int nm_i;	/* index into the netmap kring */
911 	u_int nic_i;	/* index into the NIC ring */
912 	u_int n;
913 	u_int const lim = kring->nkr_num_slots - 1;
914 	u_int const head = kring->rhead;
915 	struct if_pkt_info pi;
916 
917 	/*
918 	 * interrupts on every tx packet are expensive so request
919 	 * them every half ring, or where NS_REPORT is set
920 	 */
921 	u_int report_frequency = kring->nkr_num_slots >> 1;
922 	/* device-specific */
923 	if_ctx_t ctx = ifp->if_softc;
924 	iflib_txq_t txq = &ctx->ifc_txqs[kring->ring_id];
925 
926 	bus_dmamap_sync(txq->ift_desc_tag, txq->ift_ifdi->idi_map,
927 	    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
928 
929 
930 	/*
931 	 * First part: process new packets to send.
932 	 * nm_i is the current index in the netmap kring,
933 	 * nic_i is the corresponding index in the NIC ring.
934 	 *
935 	 * If we have packets to send (nm_i != head)
936 	 * iterate over the netmap ring, fetch length and update
937 	 * the corresponding slot in the NIC ring. Some drivers also
938 	 * need to update the buffer's physical address in the NIC slot
939 	 * even NS_BUF_CHANGED is not set (PNMB computes the addresses).
940 	 *
941 	 * The netmap_reload_map() calls is especially expensive,
942 	 * even when (as in this case) the tag is 0, so do only
943 	 * when the buffer has actually changed.
944 	 *
945 	 * If possible do not set the report/intr bit on all slots,
946 	 * but only a few times per ring or when NS_REPORT is set.
947 	 *
948 	 * Finally, on 10G and faster drivers, it might be useful
949 	 * to prefetch the next slot and txr entry.
950 	 */
951 
952 	nm_i = kring->nr_hwcur;
953 	if (nm_i != head) {	/* we have new packets to send */
954 		pkt_info_zero(&pi);
955 		pi.ipi_segs = txq->ift_segs;
956 		pi.ipi_qsidx = kring->ring_id;
957 		nic_i = netmap_idx_k2n(kring, nm_i);
958 
959 		__builtin_prefetch(&ring->slot[nm_i]);
960 		__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i]);
961 		if (txq->ift_sds.ifsd_map)
962 			__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i]);
963 
964 		for (n = 0; nm_i != head; n++) {
965 			struct netmap_slot *slot = &ring->slot[nm_i];
966 			u_int len = slot->len;
967 			uint64_t paddr;
968 			void *addr = PNMB(na, slot, &paddr);
969 			int flags = (slot->flags & NS_REPORT ||
970 				nic_i == 0 || nic_i == report_frequency) ?
971 				IPI_TX_INTR : 0;
972 
973 			/* device-specific */
974 			pi.ipi_len = len;
975 			pi.ipi_segs[0].ds_addr = paddr;
976 			pi.ipi_segs[0].ds_len = len;
977 			pi.ipi_nsegs = 1;
978 			pi.ipi_ndescs = 0;
979 			pi.ipi_pidx = nic_i;
980 			pi.ipi_flags = flags;
981 
982 			/* Fill the slot in the NIC ring. */
983 			ctx->isc_txd_encap(ctx->ifc_softc, &pi);
984 			DBG_COUNTER_INC(tx_encap);
985 
986 			/* prefetch for next round */
987 			__builtin_prefetch(&ring->slot[nm_i + 1]);
988 			__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i + 1]);
989 			if (txq->ift_sds.ifsd_map) {
990 				__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i + 1]);
991 
992 				NM_CHECK_ADDR_LEN(na, addr, len);
993 
994 				if (slot->flags & NS_BUF_CHANGED) {
995 					/* buffer has changed, reload map */
996 					netmap_reload_map(na, txq->ift_desc_tag, txq->ift_sds.ifsd_map[nic_i], addr);
997 				}
998 				/* make sure changes to the buffer are synced */
999 				bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_sds.ifsd_map[nic_i],
1000 						BUS_DMASYNC_PREWRITE);
1001 			}
1002 			slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED);
1003 			nm_i = nm_next(nm_i, lim);
1004 			nic_i = nm_next(nic_i, lim);
1005 		}
1006 		kring->nr_hwcur = nm_i;
1007 
1008 		/* synchronize the NIC ring */
1009 		bus_dmamap_sync(txq->ift_desc_tag, txq->ift_ifdi->idi_map,
1010 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1011 
1012 		/* (re)start the tx unit up to slot nic_i (excluded) */
1013 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, nic_i);
1014 	}
1015 
1016 	/*
1017 	 * Second part: reclaim buffers for completed transmissions.
1018 	 *
1019 	 * If there are unclaimed buffers, attempt to reclaim them.
1020 	 * If none are reclaimed, and TX IRQs are not in use, do an initial
1021 	 * minimal delay, then trigger the tx handler which will spin in the
1022 	 * group task queue.
1023 	 */
1024 	if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1025 		if (iflib_tx_credits_update(ctx, txq)) {
1026 			/* some tx completed, increment avail */
1027 			nic_i = txq->ift_cidx_processed;
1028 			kring->nr_hwtail = nm_prev(netmap_idx_n2k(kring, nic_i), lim);
1029 		}
1030 	}
1031 	if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ))
1032 		if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1033 			callout_reset_on(&txq->ift_timer, hz < 2000 ? 1 : hz / 1000,
1034 			    iflib_timer, txq, txq->ift_timer.c_cpu);
1035 	}
1036 	return (0);
1037 }
1038 
1039 /*
1040  * Reconcile kernel and user view of the receive ring.
1041  * Same as for the txsync, this routine must be efficient.
1042  * The caller guarantees a single invocations, but races against
1043  * the rest of the driver should be handled here.
1044  *
1045  * On call, kring->rhead is the first packet that userspace wants
1046  * to keep, and kring->rcur is the wakeup point.
1047  * The kernel has previously reported packets up to kring->rtail.
1048  *
1049  * If (flags & NAF_FORCE_READ) also check for incoming packets irrespective
1050  * of whether or not we received an interrupt.
1051  */
1052 static int
1053 iflib_netmap_rxsync(struct netmap_kring *kring, int flags)
1054 {
1055 	struct netmap_adapter *na = kring->na;
1056 	struct netmap_ring *ring = kring->ring;
1057 	uint32_t nm_i;	/* index into the netmap ring */
1058 	uint32_t nic_i;	/* index into the NIC ring */
1059 	u_int i, n;
1060 	u_int const lim = kring->nkr_num_slots - 1;
1061 	u_int const head = kring->rhead;
1062 	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
1063 	struct if_rxd_info ri;
1064 
1065 	struct ifnet *ifp = na->ifp;
1066 	if_ctx_t ctx = ifp->if_softc;
1067 	iflib_rxq_t rxq = &ctx->ifc_rxqs[kring->ring_id];
1068 	iflib_fl_t fl = rxq->ifr_fl;
1069 	if (head > lim)
1070 		return netmap_ring_reinit(kring);
1071 
1072 	/* XXX check sync modes */
1073 	for (i = 0, fl = rxq->ifr_fl; i < rxq->ifr_nfl; i++, fl++) {
1074 		if (fl->ifl_sds.ifsd_map == NULL)
1075 			continue;
1076 		bus_dmamap_sync(rxq->ifr_fl[i].ifl_desc_tag, fl->ifl_ifdi->idi_map,
1077 				BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1078 	}
1079 	/*
1080 	 * First part: import newly received packets.
1081 	 *
1082 	 * nm_i is the index of the next free slot in the netmap ring,
1083 	 * nic_i is the index of the next received packet in the NIC ring,
1084 	 * and they may differ in case if_init() has been called while
1085 	 * in netmap mode. For the receive ring we have
1086 	 *
1087 	 *	nic_i = rxr->next_check;
1088 	 *	nm_i = kring->nr_hwtail (previous)
1089 	 * and
1090 	 *	nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1091 	 *
1092 	 * rxr->next_check is set to 0 on a ring reinit
1093 	 */
1094 	if (netmap_no_pendintr || force_update) {
1095 		int crclen = iflib_crcstrip ? 0 : 4;
1096 		int error, avail;
1097 
1098 		for (i = 0; i < rxq->ifr_nfl; i++) {
1099 			fl = &rxq->ifr_fl[i];
1100 			nic_i = fl->ifl_cidx;
1101 			nm_i = netmap_idx_n2k(kring, nic_i);
1102 			avail = iflib_rxd_avail(ctx, rxq, nic_i, USHRT_MAX);
1103 			for (n = 0; avail > 0; n++, avail--) {
1104 				rxd_info_zero(&ri);
1105 				ri.iri_frags = rxq->ifr_frags;
1106 				ri.iri_qsidx = kring->ring_id;
1107 				ri.iri_ifp = ctx->ifc_ifp;
1108 				ri.iri_cidx = nic_i;
1109 
1110 				error = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
1111 				ring->slot[nm_i].len = error ? 0 : ri.iri_len - crclen;
1112 				ring->slot[nm_i].flags = 0;
1113 				bus_dmamap_sync(fl->ifl_ifdi->idi_tag,
1114 				    fl->ifl_sds.ifsd_map[nic_i], BUS_DMASYNC_POSTREAD);
1115 				nm_i = nm_next(nm_i, lim);
1116 				nic_i = nm_next(nic_i, lim);
1117 			}
1118 			if (n) { /* update the state variables */
1119 				if (netmap_no_pendintr && !force_update) {
1120 					/* diagnostics */
1121 					iflib_rx_miss ++;
1122 					iflib_rx_miss_bufs += n;
1123 				}
1124 				fl->ifl_cidx = nic_i;
1125 				kring->nr_hwtail = nm_i;
1126 			}
1127 			kring->nr_kflags &= ~NKR_PENDINTR;
1128 		}
1129 	}
1130 	/*
1131 	 * Second part: skip past packets that userspace has released.
1132 	 * (kring->nr_hwcur to head excluded),
1133 	 * and make the buffers available for reception.
1134 	 * As usual nm_i is the index in the netmap ring,
1135 	 * nic_i is the index in the NIC ring, and
1136 	 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1137 	 */
1138 	/* XXX not sure how this will work with multiple free lists */
1139 	nm_i = kring->nr_hwcur;
1140 
1141 	return (netmap_fl_refill(rxq, kring, nm_i, false));
1142 }
1143 
1144 static void
1145 iflib_netmap_intr(struct netmap_adapter *na, int onoff)
1146 {
1147 	struct ifnet *ifp = na->ifp;
1148 	if_ctx_t ctx = ifp->if_softc;
1149 
1150 	CTX_LOCK(ctx);
1151 	if (onoff) {
1152 		IFDI_INTR_ENABLE(ctx);
1153 	} else {
1154 		IFDI_INTR_DISABLE(ctx);
1155 	}
1156 	CTX_UNLOCK(ctx);
1157 }
1158 
1159 
1160 static int
1161 iflib_netmap_attach(if_ctx_t ctx)
1162 {
1163 	struct netmap_adapter na;
1164 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1165 
1166 	bzero(&na, sizeof(na));
1167 
1168 	na.ifp = ctx->ifc_ifp;
1169 	na.na_flags = NAF_BDG_MAYSLEEP;
1170 	MPASS(ctx->ifc_softc_ctx.isc_ntxqsets);
1171 	MPASS(ctx->ifc_softc_ctx.isc_nrxqsets);
1172 
1173 	na.num_tx_desc = scctx->isc_ntxd[0];
1174 	na.num_rx_desc = scctx->isc_nrxd[0];
1175 	na.nm_txsync = iflib_netmap_txsync;
1176 	na.nm_rxsync = iflib_netmap_rxsync;
1177 	na.nm_register = iflib_netmap_register;
1178 	na.nm_intr = iflib_netmap_intr;
1179 	na.num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
1180 	na.num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
1181 	return (netmap_attach(&na));
1182 }
1183 
1184 static void
1185 iflib_netmap_txq_init(if_ctx_t ctx, iflib_txq_t txq)
1186 {
1187 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1188 	struct netmap_slot *slot;
1189 
1190 	slot = netmap_reset(na, NR_TX, txq->ift_id, 0);
1191 	if (slot == NULL)
1192 		return;
1193 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxd[0]; i++) {
1194 
1195 		/*
1196 		 * In netmap mode, set the map for the packet buffer.
1197 		 * NOTE: Some drivers (not this one) also need to set
1198 		 * the physical buffer address in the NIC ring.
1199 		 * netmap_idx_n2k() maps a nic index, i, into the corresponding
1200 		 * netmap slot index, si
1201 		 */
1202 		int si = netmap_idx_n2k(na->tx_rings[txq->ift_id], i);
1203 		netmap_load_map(na, txq->ift_desc_tag, txq->ift_sds.ifsd_map[i], NMB(na, slot + si));
1204 	}
1205 }
1206 
1207 static void
1208 iflib_netmap_rxq_init(if_ctx_t ctx, iflib_rxq_t rxq)
1209 {
1210 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1211 	struct netmap_kring *kring = na->rx_rings[rxq->ifr_id];
1212 	struct netmap_slot *slot;
1213 	uint32_t nm_i;
1214 
1215 	slot = netmap_reset(na, NR_RX, rxq->ifr_id, 0);
1216 	if (slot == NULL)
1217 		return;
1218 	nm_i = netmap_idx_n2k(kring, 0);
1219 	netmap_fl_refill(rxq, kring, nm_i, true);
1220 }
1221 
1222 static void
1223 iflib_netmap_timer_adjust(if_ctx_t ctx, uint16_t txqid, uint32_t *reset_on)
1224 {
1225 	struct netmap_kring *kring;
1226 
1227 	kring = NA(ctx->ifc_ifp)->tx_rings[txqid];
1228 
1229 	if (kring->nr_hwcur != nm_next(kring->nr_hwtail, kring->nkr_num_slots - 1)) {
1230 		if (ctx->isc_txd_credits_update(ctx->ifc_softc, txqid, false))
1231 			netmap_tx_irq(ctx->ifc_ifp, txqid);
1232 		if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ)) {
1233 			if (hz < 2000)
1234 				*reset_on = 1;
1235 			else
1236 				*reset_on = hz / 1000;
1237 		}
1238 	}
1239 }
1240 
1241 #define iflib_netmap_detach(ifp) netmap_detach(ifp)
1242 
1243 #else
1244 #define iflib_netmap_txq_init(ctx, txq)
1245 #define iflib_netmap_rxq_init(ctx, rxq)
1246 #define iflib_netmap_detach(ifp)
1247 
1248 #define iflib_netmap_attach(ctx) (0)
1249 #define netmap_rx_irq(ifp, qid, budget) (0)
1250 #define netmap_tx_irq(ifp, qid) do {} while (0)
1251 #define iflib_netmap_timer_adjust(ctx, txqid, reset_on)
1252 
1253 #endif
1254 
1255 #if defined(__i386__) || defined(__amd64__)
1256 static __inline void
1257 prefetch(void *x)
1258 {
1259 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1260 }
1261 static __inline void
1262 prefetch2cachelines(void *x)
1263 {
1264 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1265 #if (CACHE_LINE_SIZE < 128)
1266 	__asm volatile("prefetcht0 %0" :: "m" (*(((unsigned long *)x)+CACHE_LINE_SIZE/(sizeof(unsigned long)))));
1267 #endif
1268 }
1269 #else
1270 #define prefetch(x)
1271 #define prefetch2cachelines(x)
1272 #endif
1273 
1274 static void
1275 iflib_gen_mac(if_ctx_t ctx)
1276 {
1277 	struct thread *td;
1278 	MD5_CTX mdctx;
1279 	char uuid[HOSTUUIDLEN+1];
1280 	char buf[HOSTUUIDLEN+16];
1281 	uint8_t *mac;
1282 	unsigned char digest[16];
1283 
1284 	td = curthread;
1285 	mac = ctx->ifc_mac;
1286 	uuid[HOSTUUIDLEN] = 0;
1287 	bcopy(td->td_ucred->cr_prison->pr_hostuuid, uuid, HOSTUUIDLEN);
1288 	snprintf(buf, HOSTUUIDLEN+16, "%s-%s", uuid, device_get_nameunit(ctx->ifc_dev));
1289 	/*
1290 	 * Generate a pseudo-random, deterministic MAC
1291 	 * address based on the UUID and unit number.
1292 	 * The FreeBSD Foundation OUI of 58-9C-FC is used.
1293 	 */
1294 	MD5Init(&mdctx);
1295 	MD5Update(&mdctx, buf, strlen(buf));
1296 	MD5Final(digest, &mdctx);
1297 
1298 	mac[0] = 0x58;
1299 	mac[1] = 0x9C;
1300 	mac[2] = 0xFC;
1301 	mac[3] = digest[0];
1302 	mac[4] = digest[1];
1303 	mac[5] = digest[2];
1304 }
1305 
1306 static void
1307 iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid)
1308 {
1309 	iflib_fl_t fl;
1310 
1311 	fl = &rxq->ifr_fl[flid];
1312 	iru->iru_paddrs = fl->ifl_bus_addrs;
1313 	iru->iru_vaddrs = &fl->ifl_vm_addrs[0];
1314 	iru->iru_idxs = fl->ifl_rxd_idxs;
1315 	iru->iru_qsidx = rxq->ifr_id;
1316 	iru->iru_buf_size = fl->ifl_buf_size;
1317 	iru->iru_flidx = fl->ifl_id;
1318 }
1319 
1320 static void
1321 _iflib_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int err)
1322 {
1323 	if (err)
1324 		return;
1325 	*(bus_addr_t *) arg = segs[0].ds_addr;
1326 }
1327 
1328 int
1329 iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags)
1330 {
1331 	int err;
1332 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1333 	device_t dev = ctx->ifc_dev;
1334 
1335 	KASSERT(sctx->isc_q_align != 0, ("alignment value not initialized"));
1336 
1337 	err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1338 				sctx->isc_q_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_multi(if_ctx_t ctx, int *sizes, iflib_dma_info_t *dmalist, int mapflags, int count)
1390 {
1391 	int i, err;
1392 	iflib_dma_info_t *dmaiter;
1393 
1394 	dmaiter = dmalist;
1395 	for (i = 0; i < count; i++, dmaiter++) {
1396 		if ((err = iflib_dma_alloc(ctx, sizes[i], *dmaiter, mapflags)) != 0)
1397 			break;
1398 	}
1399 	if (err)
1400 		iflib_dma_free_multi(dmalist, i);
1401 	return (err);
1402 }
1403 
1404 void
1405 iflib_dma_free(iflib_dma_info_t dma)
1406 {
1407 	if (dma->idi_tag == NULL)
1408 		return;
1409 	if (dma->idi_paddr != IF_BAD_DMA) {
1410 		bus_dmamap_sync(dma->idi_tag, dma->idi_map,
1411 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1412 		bus_dmamap_unload(dma->idi_tag, dma->idi_map);
1413 		dma->idi_paddr = IF_BAD_DMA;
1414 	}
1415 	if (dma->idi_vaddr != NULL) {
1416 		bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1417 		dma->idi_vaddr = NULL;
1418 	}
1419 	bus_dma_tag_destroy(dma->idi_tag);
1420 	dma->idi_tag = NULL;
1421 }
1422 
1423 void
1424 iflib_dma_free_multi(iflib_dma_info_t *dmalist, int count)
1425 {
1426 	int i;
1427 	iflib_dma_info_t *dmaiter = dmalist;
1428 
1429 	for (i = 0; i < count; i++, dmaiter++)
1430 		iflib_dma_free(*dmaiter);
1431 }
1432 
1433 #ifdef EARLY_AP_STARTUP
1434 static const int iflib_started = 1;
1435 #else
1436 /*
1437  * We used to abuse the smp_started flag to decide if the queues have been
1438  * fully initialized (by late taskqgroup_adjust() calls in a SYSINIT()).
1439  * That gave bad races, since the SYSINIT() runs strictly after smp_started
1440  * is set.  Run a SYSINIT() strictly after that to just set a usable
1441  * completion flag.
1442  */
1443 
1444 static int iflib_started;
1445 
1446 static void
1447 iflib_record_started(void *arg)
1448 {
1449 	iflib_started = 1;
1450 }
1451 
1452 SYSINIT(iflib_record_started, SI_SUB_SMP + 1, SI_ORDER_FIRST,
1453 	iflib_record_started, NULL);
1454 #endif
1455 
1456 static int
1457 iflib_fast_intr(void *arg)
1458 {
1459 	iflib_filter_info_t info = arg;
1460 	struct grouptask *gtask = info->ifi_task;
1461 	if (!iflib_started)
1462 		return (FILTER_HANDLED);
1463 
1464 	DBG_COUNTER_INC(fast_intrs);
1465 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1466 		return (FILTER_HANDLED);
1467 
1468 	GROUPTASK_ENQUEUE(gtask);
1469 	return (FILTER_HANDLED);
1470 }
1471 
1472 static int
1473 iflib_fast_intr_rxtx(void *arg)
1474 {
1475 	iflib_filter_info_t info = arg;
1476 	struct grouptask *gtask = info->ifi_task;
1477 	iflib_rxq_t rxq = (iflib_rxq_t)info->ifi_ctx;
1478 	if_ctx_t ctx = NULL;;
1479 	int i, cidx;
1480 
1481 	if (!iflib_started)
1482 		return (FILTER_HANDLED);
1483 
1484 	DBG_COUNTER_INC(fast_intrs);
1485 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1486 		return (FILTER_HANDLED);
1487 
1488 	MPASS(rxq->ifr_ntxqirq);
1489 	for (i = 0; i < rxq->ifr_ntxqirq; i++) {
1490 		qidx_t txqid = rxq->ifr_txqid[i];
1491 
1492 		ctx = rxq->ifr_ctx;
1493 
1494 		if (!ctx->isc_txd_credits_update(ctx->ifc_softc, txqid, false)) {
1495 			IFDI_TX_QUEUE_INTR_ENABLE(ctx, txqid);
1496 			continue;
1497 		}
1498 		GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
1499 	}
1500 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_RXCQ)
1501 		cidx = rxq->ifr_cq_cidx;
1502 	else
1503 		cidx = rxq->ifr_fl[0].ifl_cidx;
1504 	if (iflib_rxd_avail(ctx, rxq, cidx, 1))
1505 		GROUPTASK_ENQUEUE(gtask);
1506 	else {
1507 		IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
1508 		DBG_COUNTER_INC(rx_intr_enables);
1509 	}
1510 	return (FILTER_HANDLED);
1511 }
1512 
1513 
1514 static int
1515 iflib_fast_intr_ctx(void *arg)
1516 {
1517 	iflib_filter_info_t info = arg;
1518 	struct grouptask *gtask = info->ifi_task;
1519 
1520 	if (!iflib_started)
1521 		return (FILTER_HANDLED);
1522 
1523 	DBG_COUNTER_INC(fast_intrs);
1524 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1525 		return (FILTER_HANDLED);
1526 
1527 	GROUPTASK_ENQUEUE(gtask);
1528 	return (FILTER_HANDLED);
1529 }
1530 
1531 static int
1532 _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
1533 		 driver_filter_t filter, driver_intr_t handler, void *arg,
1534 		 const char *name)
1535 {
1536 	int rc, flags;
1537 	struct resource *res;
1538 	void *tag = NULL;
1539 	device_t dev = ctx->ifc_dev;
1540 
1541 	flags = RF_ACTIVE;
1542 	if (ctx->ifc_flags & IFC_LEGACY)
1543 		flags |= RF_SHAREABLE;
1544 	MPASS(rid < 512);
1545 	irq->ii_rid = rid;
1546 	res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &irq->ii_rid, flags);
1547 	if (res == NULL) {
1548 		device_printf(dev,
1549 		    "failed to allocate IRQ for rid %d, name %s.\n", rid, name);
1550 		return (ENOMEM);
1551 	}
1552 	irq->ii_res = res;
1553 	KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL"));
1554 	rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET,
1555 						filter, handler, arg, &tag);
1556 	if (rc != 0) {
1557 		device_printf(dev,
1558 		    "failed to setup interrupt for rid %d, name %s: %d\n",
1559 					  rid, name ? name : "unknown", rc);
1560 		return (rc);
1561 	} else if (name)
1562 		bus_describe_intr(dev, res, tag, "%s", name);
1563 
1564 	irq->ii_tag = tag;
1565 	return (0);
1566 }
1567 
1568 
1569 /*********************************************************************
1570  *
1571  *  Allocate memory for tx_buffer structures. The tx_buffer stores all
1572  *  the information needed to transmit a packet on the wire. This is
1573  *  called only once at attach, setup is done every reset.
1574  *
1575  **********************************************************************/
1576 
1577 static int
1578 iflib_txsd_alloc(iflib_txq_t txq)
1579 {
1580 	if_ctx_t ctx = txq->ift_ctx;
1581 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1582 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1583 	device_t dev = ctx->ifc_dev;
1584 	bus_size_t tsomaxsize;
1585 	int err, nsegments, ntsosegments;
1586 
1587 	nsegments = scctx->isc_tx_nsegments;
1588 	ntsosegments = scctx->isc_tx_tso_segments_max;
1589 	tsomaxsize = scctx->isc_tx_tso_size_max;
1590 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_VLAN_MTU)
1591 		tsomaxsize += sizeof(struct ether_vlan_header);
1592 	MPASS(scctx->isc_ntxd[0] > 0);
1593 	MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0);
1594 	MPASS(nsegments > 0);
1595 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) {
1596 		MPASS(ntsosegments > 0);
1597 		MPASS(sctx->isc_tso_maxsize >= tsomaxsize);
1598 	}
1599 
1600 	/*
1601 	 * Setup DMA descriptor areas.
1602 	 */
1603 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1604 			       1, 0,			/* alignment, bounds */
1605 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1606 			       BUS_SPACE_MAXADDR,	/* highaddr */
1607 			       NULL, NULL,		/* filter, filterarg */
1608 			       sctx->isc_tx_maxsize,		/* maxsize */
1609 			       nsegments,	/* nsegments */
1610 			       sctx->isc_tx_maxsegsize,	/* maxsegsize */
1611 			       0,			/* flags */
1612 			       NULL,			/* lockfunc */
1613 			       NULL,			/* lockfuncarg */
1614 			       &txq->ift_desc_tag))) {
1615 		device_printf(dev,"Unable to allocate TX DMA tag: %d\n", err);
1616 		device_printf(dev,"maxsize: %ju nsegments: %d maxsegsize: %ju\n",
1617 		    (uintmax_t)sctx->isc_tx_maxsize, nsegments, (uintmax_t)sctx->isc_tx_maxsegsize);
1618 		goto fail;
1619 	}
1620 	if ((if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) &
1621 	    (err = bus_dma_tag_create(bus_get_dma_tag(dev),
1622 			       1, 0,			/* alignment, bounds */
1623 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1624 			       BUS_SPACE_MAXADDR,	/* highaddr */
1625 			       NULL, NULL,		/* filter, filterarg */
1626 			       tsomaxsize,		/* maxsize */
1627 			       ntsosegments,	/* nsegments */
1628 			       sctx->isc_tso_maxsegsize,/* maxsegsize */
1629 			       0,			/* flags */
1630 			       NULL,			/* lockfunc */
1631 			       NULL,			/* lockfuncarg */
1632 			       &txq->ift_tso_desc_tag))) {
1633 		device_printf(dev,"Unable to allocate TX TSO DMA tag: %d\n", err);
1634 
1635 		goto fail;
1636 	}
1637 	if (!(txq->ift_sds.ifsd_m =
1638 	    (struct mbuf **) malloc(sizeof(struct mbuf *) *
1639 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1640 		device_printf(dev, "Unable to allocate tx_buffer memory\n");
1641 		err = ENOMEM;
1642 		goto fail;
1643 	}
1644 
1645         /* Create the descriptor buffer dma maps */
1646 	if (!(txq->ift_sds.ifsd_map =
1647 	    (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1648 		device_printf(dev, "Unable to allocate tx_buffer map memory\n");
1649 		err = ENOMEM;
1650 		goto fail;
1651 	}
1652 
1653 	for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) {
1654 		err = bus_dmamap_create(txq->ift_desc_tag, 0, &txq->ift_sds.ifsd_map[i]);
1655 		if (err != 0) {
1656 			device_printf(dev, "Unable to create TX DMA map\n");
1657 			goto fail;
1658 		}
1659 	}
1660 	return (0);
1661 fail:
1662 	/* We free all, it handles case where we are in the middle */
1663 	iflib_tx_structures_free(ctx);
1664 	return (err);
1665 }
1666 
1667 static void
1668 iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i)
1669 {
1670 	bus_dmamap_t map;
1671 
1672 	map = NULL;
1673 	if (txq->ift_sds.ifsd_map != NULL)
1674 		map = txq->ift_sds.ifsd_map[i];
1675 	if (map != NULL) {
1676 		bus_dmamap_unload(txq->ift_desc_tag, map);
1677 		bus_dmamap_destroy(txq->ift_desc_tag, map);
1678 		txq->ift_sds.ifsd_map[i] = NULL;
1679 	}
1680 }
1681 
1682 static void
1683 iflib_txq_destroy(iflib_txq_t txq)
1684 {
1685 	if_ctx_t ctx = txq->ift_ctx;
1686 
1687 	for (int i = 0; i < txq->ift_size; i++)
1688 		iflib_txsd_destroy(ctx, txq, i);
1689 	if (txq->ift_sds.ifsd_map != NULL) {
1690 		free(txq->ift_sds.ifsd_map, M_IFLIB);
1691 		txq->ift_sds.ifsd_map = NULL;
1692 	}
1693 	if (txq->ift_sds.ifsd_m != NULL) {
1694 		free(txq->ift_sds.ifsd_m, M_IFLIB);
1695 		txq->ift_sds.ifsd_m = NULL;
1696 	}
1697 	if (txq->ift_desc_tag != NULL) {
1698 		bus_dma_tag_destroy(txq->ift_desc_tag);
1699 		txq->ift_desc_tag = NULL;
1700 	}
1701 	if (txq->ift_tso_desc_tag != NULL) {
1702 		bus_dma_tag_destroy(txq->ift_tso_desc_tag);
1703 		txq->ift_tso_desc_tag = NULL;
1704 	}
1705 }
1706 
1707 static void
1708 iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i)
1709 {
1710 	struct mbuf **mp;
1711 
1712 	mp = &txq->ift_sds.ifsd_m[i];
1713 	if (*mp == NULL)
1714 		return;
1715 
1716 	if (txq->ift_sds.ifsd_map != NULL) {
1717 		bus_dmamap_sync(txq->ift_desc_tag,
1718 				txq->ift_sds.ifsd_map[i],
1719 				BUS_DMASYNC_POSTWRITE);
1720 		bus_dmamap_unload(txq->ift_desc_tag,
1721 				  txq->ift_sds.ifsd_map[i]);
1722 	}
1723 	m_free(*mp);
1724 	DBG_COUNTER_INC(tx_frees);
1725 	*mp = NULL;
1726 }
1727 
1728 static int
1729 iflib_txq_setup(iflib_txq_t txq)
1730 {
1731 	if_ctx_t ctx = txq->ift_ctx;
1732 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1733 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1734 	iflib_dma_info_t di;
1735 	int i;
1736 
1737 	/* Set number of descriptors available */
1738 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
1739 	/* XXX make configurable */
1740 	txq->ift_update_freq = IFLIB_DEFAULT_TX_UPDATE_FREQ;
1741 
1742 	/* Reset indices */
1743 	txq->ift_cidx_processed = 0;
1744 	txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0;
1745 	txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset];
1746 
1747 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1748 		bzero((void *)di->idi_vaddr, di->idi_size);
1749 
1750 	IFDI_TXQ_SETUP(ctx, txq->ift_id);
1751 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1752 		bus_dmamap_sync(di->idi_tag, di->idi_map,
1753 						BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1754 	return (0);
1755 }
1756 
1757 /*********************************************************************
1758  *
1759  *  Allocate memory for rx_buffer structures. Since we use one
1760  *  rx_buffer per received packet, the maximum number of rx_buffer's
1761  *  that we'll need is equal to the number of receive descriptors
1762  *  that we've allocated.
1763  *
1764  **********************************************************************/
1765 static int
1766 iflib_rxsd_alloc(iflib_rxq_t rxq)
1767 {
1768 	if_ctx_t ctx = rxq->ifr_ctx;
1769 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1770 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1771 	device_t dev = ctx->ifc_dev;
1772 	iflib_fl_t fl;
1773 	int			err;
1774 
1775 	MPASS(scctx->isc_nrxd[0] > 0);
1776 	MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0);
1777 
1778 	fl = rxq->ifr_fl;
1779 	for (int i = 0; i <  rxq->ifr_nfl; i++, fl++) {
1780 		fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */
1781 		err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1782 					 1, 0,			/* alignment, bounds */
1783 					 BUS_SPACE_MAXADDR,	/* lowaddr */
1784 					 BUS_SPACE_MAXADDR,	/* highaddr */
1785 					 NULL, NULL,		/* filter, filterarg */
1786 					 sctx->isc_rx_maxsize,	/* maxsize */
1787 					 sctx->isc_rx_nsegments,	/* nsegments */
1788 					 sctx->isc_rx_maxsegsize,	/* maxsegsize */
1789 					 0,			/* flags */
1790 					 NULL,			/* lockfunc */
1791 					 NULL,			/* lockarg */
1792 					 &fl->ifl_desc_tag);
1793 		if (err) {
1794 			device_printf(dev, "%s: bus_dma_tag_create failed %d\n",
1795 				__func__, err);
1796 			goto fail;
1797 		}
1798 		if (!(fl->ifl_sds.ifsd_m =
1799 		      (struct mbuf **) malloc(sizeof(struct mbuf *) *
1800 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1801 			device_printf(dev, "Unable to allocate tx_buffer memory\n");
1802 			err = ENOMEM;
1803 			goto fail;
1804 		}
1805 		if (!(fl->ifl_sds.ifsd_cl =
1806 		      (caddr_t *) malloc(sizeof(caddr_t) *
1807 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1808 			device_printf(dev, "Unable to allocate tx_buffer memory\n");
1809 			err = ENOMEM;
1810 			goto fail;
1811 		}
1812 
1813 		if (!(fl->ifl_sds.ifsd_ba =
1814 		      (bus_addr_t *) malloc(sizeof(bus_addr_t) *
1815 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1816 			device_printf(dev, "Unable to allocate rx bus addr memory\n");
1817 			err = ENOMEM;
1818 			goto fail;
1819 		}
1820 
1821 		/* Create the descriptor buffer dma maps */
1822 		if (!(fl->ifl_sds.ifsd_map =
1823 		      (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1824 			device_printf(dev, "Unable to allocate tx_buffer map memory\n");
1825 			err = ENOMEM;
1826 			goto fail;
1827 		}
1828 
1829 		for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++) {
1830 			err = bus_dmamap_create(fl->ifl_desc_tag, 0, &fl->ifl_sds.ifsd_map[i]);
1831 			if (err != 0) {
1832 				device_printf(dev, "Unable to create RX buffer DMA map\n");
1833 				goto fail;
1834 			}
1835 		}
1836 	}
1837 	return (0);
1838 
1839 fail:
1840 	iflib_rx_structures_free(ctx);
1841 	return (err);
1842 }
1843 
1844 
1845 /*
1846  * Internal service routines
1847  */
1848 
1849 struct rxq_refill_cb_arg {
1850 	int               error;
1851 	bus_dma_segment_t seg;
1852 	int               nseg;
1853 };
1854 
1855 static void
1856 _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
1857 {
1858 	struct rxq_refill_cb_arg *cb_arg = arg;
1859 
1860 	cb_arg->error = error;
1861 	cb_arg->seg = segs[0];
1862 	cb_arg->nseg = nseg;
1863 }
1864 
1865 /**
1866  *	rxq_refill - refill an rxq  free-buffer list
1867  *	@ctx: the iflib context
1868  *	@rxq: the free-list to refill
1869  *	@n: the number of new buffers to allocate
1870  *
1871  *	(Re)populate an rxq free-buffer list with up to @n new packet buffers.
1872  *	The caller must assure that @n does not exceed the queue's capacity.
1873  */
1874 static void
1875 _iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
1876 {
1877 	struct mbuf *m;
1878 	int idx, frag_idx = fl->ifl_fragidx;
1879         int pidx = fl->ifl_pidx;
1880 	caddr_t cl, *sd_cl;
1881 	struct mbuf **sd_m;
1882 	struct if_rxd_update iru;
1883 	struct rxq_refill_cb_arg cb_arg;
1884 	bus_dmamap_t *sd_map;
1885 	int n, i = 0;
1886 	bus_addr_t bus_addr, *sd_ba;
1887 	int err;
1888 	qidx_t credits;
1889 
1890 	sd_m = fl->ifl_sds.ifsd_m;
1891 	sd_map = fl->ifl_sds.ifsd_map;
1892 	sd_cl = fl->ifl_sds.ifsd_cl;
1893 	sd_ba = fl->ifl_sds.ifsd_ba;
1894 	idx = pidx;
1895 	credits = fl->ifl_credits;
1896 
1897 	n  = count;
1898 	MPASS(n > 0);
1899 	MPASS(credits + n <= fl->ifl_size);
1900 
1901 	if (pidx < fl->ifl_cidx)
1902 		MPASS(pidx + n <= fl->ifl_cidx);
1903 	if (pidx == fl->ifl_cidx && (credits < fl->ifl_size))
1904 		MPASS(fl->ifl_gen == 0);
1905 	if (pidx > fl->ifl_cidx)
1906 		MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx);
1907 
1908 	DBG_COUNTER_INC(fl_refills);
1909 	if (n > 8)
1910 		DBG_COUNTER_INC(fl_refills_large);
1911 	iru_init(&iru, fl->ifl_rxq, fl->ifl_id);
1912 	while (n--) {
1913 		/*
1914 		 * We allocate an uninitialized mbuf + cluster, mbuf is
1915 		 * initialized after rx.
1916 		 *
1917 		 * If the cluster is still set then we know a minimum sized packet was received
1918 		 */
1919 		bit_ffc_at(fl->ifl_rx_bitmap, frag_idx, fl->ifl_size,  &frag_idx);
1920 		if ((frag_idx < 0) || (frag_idx >= fl->ifl_size))
1921                 	bit_ffc(fl->ifl_rx_bitmap, fl->ifl_size, &frag_idx);
1922 		if ((cl = sd_cl[frag_idx]) == NULL) {
1923 			if ((cl = m_cljget(NULL, M_NOWAIT, fl->ifl_buf_size)) == NULL)
1924 				break;
1925 
1926 			cb_arg.error = 0;
1927 			MPASS(sd_map != NULL);
1928 			err = bus_dmamap_load(fl->ifl_desc_tag, sd_map[frag_idx],
1929 			    cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg, 0);
1930 			bus_dmamap_sync(fl->ifl_desc_tag, sd_map[frag_idx],
1931 			    BUS_DMASYNC_PREREAD);
1932 
1933 			if (err != 0 || cb_arg.error) {
1934 				/*
1935 				 * !zone_pack ?
1936 				 */
1937 				if (fl->ifl_zone == zone_pack)
1938 					uma_zfree(fl->ifl_zone, cl);
1939 				break;
1940 			}
1941 
1942 			sd_ba[frag_idx] =  bus_addr = cb_arg.seg.ds_addr;
1943 			sd_cl[frag_idx] = cl;
1944 #if MEMORY_LOGGING
1945 			fl->ifl_cl_enqueued++;
1946 #endif
1947 		} else {
1948 			bus_addr = sd_ba[frag_idx];
1949 		}
1950 
1951 		bit_set(fl->ifl_rx_bitmap, frag_idx);
1952 		MPASS(sd_m[frag_idx] == NULL);
1953 		if ((m = m_gethdr(M_NOWAIT, MT_NOINIT)) == NULL) {
1954 			break;
1955 		}
1956 		sd_m[frag_idx] = m;
1957 #if MEMORY_LOGGING
1958 		fl->ifl_m_enqueued++;
1959 #endif
1960 
1961 		DBG_COUNTER_INC(rx_allocs);
1962 		fl->ifl_rxd_idxs[i] = frag_idx;
1963 		fl->ifl_bus_addrs[i] = bus_addr;
1964 		fl->ifl_vm_addrs[i] = cl;
1965 		credits++;
1966 		i++;
1967 		MPASS(credits <= fl->ifl_size);
1968 		if (++idx == fl->ifl_size) {
1969 			fl->ifl_gen = 1;
1970 			idx = 0;
1971 		}
1972 		if (n == 0 || i == IFLIB_MAX_RX_REFRESH) {
1973 			iru.iru_pidx = pidx;
1974 			iru.iru_count = i;
1975 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
1976 			i = 0;
1977 			pidx = idx;
1978 			fl->ifl_pidx = idx;
1979 			fl->ifl_credits = credits;
1980 		}
1981 
1982 	}
1983 
1984 	if (i) {
1985 		iru.iru_pidx = pidx;
1986 		iru.iru_count = i;
1987 		ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
1988 		fl->ifl_pidx = idx;
1989 		fl->ifl_credits = credits;
1990 	}
1991 	DBG_COUNTER_INC(rxd_flush);
1992 	if (fl->ifl_pidx == 0)
1993 		pidx = fl->ifl_size - 1;
1994 	else
1995 		pidx = fl->ifl_pidx - 1;
1996 
1997 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
1998 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1999 	ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id, fl->ifl_id, pidx);
2000 	fl->ifl_fragidx = frag_idx;
2001 }
2002 
2003 static __inline void
2004 __iflib_fl_refill_lt(if_ctx_t ctx, iflib_fl_t fl, int max)
2005 {
2006 	/* we avoid allowing pidx to catch up with cidx as it confuses ixl */
2007 	int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1;
2008 #ifdef INVARIANTS
2009 	int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1;
2010 #endif
2011 
2012 	MPASS(fl->ifl_credits <= fl->ifl_size);
2013 	MPASS(reclaimable == delta);
2014 
2015 	if (reclaimable > 0)
2016 		_iflib_fl_refill(ctx, fl, min(max, reclaimable));
2017 }
2018 
2019 uint8_t
2020 iflib_in_detach(if_ctx_t ctx)
2021 {
2022 	bool in_detach;
2023 	STATE_LOCK(ctx);
2024 	in_detach = !!(ctx->ifc_flags & IFC_IN_DETACH);
2025 	STATE_UNLOCK(ctx);
2026 	return (in_detach);
2027 }
2028 
2029 static void
2030 iflib_fl_bufs_free(iflib_fl_t fl)
2031 {
2032 	iflib_dma_info_t idi = fl->ifl_ifdi;
2033 	uint32_t i;
2034 
2035 	for (i = 0; i < fl->ifl_size; i++) {
2036 		struct mbuf **sd_m = &fl->ifl_sds.ifsd_m[i];
2037 		caddr_t *sd_cl = &fl->ifl_sds.ifsd_cl[i];
2038 
2039 		if (*sd_cl != NULL) {
2040 			bus_dmamap_t sd_map = fl->ifl_sds.ifsd_map[i];
2041 			bus_dmamap_unload(fl->ifl_desc_tag, sd_map);
2042 			if (*sd_cl != NULL)
2043 				uma_zfree(fl->ifl_zone, *sd_cl);
2044 			// XXX: Should this get moved out?
2045 			if (iflib_in_detach(fl->ifl_rxq->ifr_ctx))
2046 				bus_dmamap_destroy(fl->ifl_desc_tag, sd_map);
2047 			if (*sd_m != NULL) {
2048 				m_init(*sd_m, M_NOWAIT, MT_DATA, 0);
2049 				uma_zfree(zone_mbuf, *sd_m);
2050 			}
2051 		} else {
2052 			MPASS(*sd_cl == NULL);
2053 			MPASS(*sd_m == NULL);
2054 		}
2055 #if MEMORY_LOGGING
2056 		fl->ifl_m_dequeued++;
2057 		fl->ifl_cl_dequeued++;
2058 #endif
2059 		*sd_cl = NULL;
2060 		*sd_m = NULL;
2061 	}
2062 #ifdef INVARIANTS
2063 	for (i = 0; i < fl->ifl_size; i++) {
2064 		MPASS(fl->ifl_sds.ifsd_cl[i] == NULL);
2065 		MPASS(fl->ifl_sds.ifsd_m[i] == NULL);
2066 	}
2067 #endif
2068 	/*
2069 	 * Reset free list values
2070 	 */
2071 	fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = fl->ifl_fragidx = 0;
2072 	bzero(idi->idi_vaddr, idi->idi_size);
2073 }
2074 
2075 /*********************************************************************
2076  *
2077  *  Initialize a receive ring and its buffers.
2078  *
2079  **********************************************************************/
2080 static int
2081 iflib_fl_setup(iflib_fl_t fl)
2082 {
2083 	iflib_rxq_t rxq = fl->ifl_rxq;
2084 	if_ctx_t ctx = rxq->ifr_ctx;
2085 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2086 
2087 	bit_nclear(fl->ifl_rx_bitmap, 0, fl->ifl_size - 1);
2088 	/*
2089 	** Free current RX buffer structs and their mbufs
2090 	*/
2091 	iflib_fl_bufs_free(fl);
2092 	/* Now replenish the mbufs */
2093 	MPASS(fl->ifl_credits == 0);
2094 	/*
2095 	 * XXX don't set the max_frame_size to larger
2096 	 * than the hardware can handle
2097 	 */
2098 	if (sctx->isc_max_frame_size <= 2048)
2099 		fl->ifl_buf_size = MCLBYTES;
2100 #ifndef CONTIGMALLOC_WORKS
2101 	else
2102 		fl->ifl_buf_size = MJUMPAGESIZE;
2103 #else
2104 	else if (sctx->isc_max_frame_size <= 4096)
2105 		fl->ifl_buf_size = MJUMPAGESIZE;
2106 	else if (sctx->isc_max_frame_size <= 9216)
2107 		fl->ifl_buf_size = MJUM9BYTES;
2108 	else
2109 		fl->ifl_buf_size = MJUM16BYTES;
2110 #endif
2111 	if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size)
2112 		ctx->ifc_max_fl_buf_size = fl->ifl_buf_size;
2113 	fl->ifl_cltype = m_gettype(fl->ifl_buf_size);
2114 	fl->ifl_zone = m_getzone(fl->ifl_buf_size);
2115 
2116 
2117 	/* avoid pre-allocating zillions of clusters to an idle card
2118 	 * potentially speeding up attach
2119 	 */
2120 	_iflib_fl_refill(ctx, fl, min(128, fl->ifl_size));
2121 	MPASS(min(128, fl->ifl_size) == fl->ifl_credits);
2122 	if (min(128, fl->ifl_size) != fl->ifl_credits)
2123 		return (ENOBUFS);
2124 	/*
2125 	 * handle failure
2126 	 */
2127 	MPASS(rxq != NULL);
2128 	MPASS(fl->ifl_ifdi != NULL);
2129 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2130 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2131 	return (0);
2132 }
2133 
2134 /*********************************************************************
2135  *
2136  *  Free receive ring data structures
2137  *
2138  **********************************************************************/
2139 static void
2140 iflib_rx_sds_free(iflib_rxq_t rxq)
2141 {
2142 	iflib_fl_t fl;
2143 	int i;
2144 
2145 	if (rxq->ifr_fl != NULL) {
2146 		for (i = 0; i < rxq->ifr_nfl; i++) {
2147 			fl = &rxq->ifr_fl[i];
2148 			if (fl->ifl_desc_tag != NULL) {
2149 				bus_dma_tag_destroy(fl->ifl_desc_tag);
2150 				fl->ifl_desc_tag = NULL;
2151 			}
2152 			free(fl->ifl_sds.ifsd_m, M_IFLIB);
2153 			free(fl->ifl_sds.ifsd_cl, M_IFLIB);
2154 			free(fl->ifl_sds.ifsd_ba, M_IFLIB);
2155 			/* XXX destroy maps first */
2156 			free(fl->ifl_sds.ifsd_map, M_IFLIB);
2157 			fl->ifl_sds.ifsd_m = NULL;
2158 			fl->ifl_sds.ifsd_cl = NULL;
2159 			fl->ifl_sds.ifsd_ba = NULL;
2160 			fl->ifl_sds.ifsd_map = NULL;
2161 		}
2162 		free(rxq->ifr_fl, M_IFLIB);
2163 		rxq->ifr_fl = NULL;
2164 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
2165 	}
2166 }
2167 
2168 /*
2169  * MI independent logic
2170  *
2171  */
2172 static void
2173 iflib_timer(void *arg)
2174 {
2175 	iflib_txq_t txq = arg;
2176 	if_ctx_t ctx = txq->ift_ctx;
2177 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2178 	uint64_t this_tick = ticks;
2179 	uint32_t reset_on = hz / 2;
2180 
2181 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
2182 		return;
2183 	/*
2184 	** Check on the state of the TX queue(s), this
2185 	** can be done without the lock because its RO
2186 	** and the HUNG state will be static if set.
2187 	*/
2188 	if (this_tick - txq->ift_last_timer_tick >= hz / 2) {
2189 		txq->ift_last_timer_tick = this_tick;
2190 		IFDI_TIMER(ctx, txq->ift_id);
2191 		if ((txq->ift_qstatus == IFLIB_QUEUE_HUNG) &&
2192 		    ((txq->ift_cleaned_prev == txq->ift_cleaned) ||
2193 		     (sctx->isc_pause_frames == 0)))
2194 			goto hung;
2195 
2196 		if (ifmp_ring_is_stalled(txq->ift_br))
2197 			txq->ift_qstatus = IFLIB_QUEUE_HUNG;
2198 		txq->ift_cleaned_prev = txq->ift_cleaned;
2199 	}
2200 #ifdef DEV_NETMAP
2201 	if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP)
2202 		iflib_netmap_timer_adjust(ctx, txq->ift_id, &reset_on);
2203 #endif
2204 	/* handle any laggards */
2205 	if (txq->ift_db_pending)
2206 		GROUPTASK_ENQUEUE(&txq->ift_task);
2207 
2208 	sctx->isc_pause_frames = 0;
2209 	if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)
2210 		callout_reset_on(&txq->ift_timer, reset_on, iflib_timer, txq, txq->ift_timer.c_cpu);
2211 	return;
2212  hung:
2213 	device_printf(ctx->ifc_dev,  "TX(%d) desc avail = %d, pidx = %d\n",
2214 				  txq->ift_id, TXQ_AVAIL(txq), txq->ift_pidx);
2215 	STATE_LOCK(ctx);
2216 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2217 	ctx->ifc_flags |= (IFC_DO_WATCHDOG|IFC_DO_RESET);
2218 	iflib_admin_intr_deferred(ctx);
2219 	STATE_UNLOCK(ctx);
2220 }
2221 
2222 static void
2223 iflib_init_locked(if_ctx_t ctx)
2224 {
2225 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2226 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2227 	if_t ifp = ctx->ifc_ifp;
2228 	iflib_fl_t fl;
2229 	iflib_txq_t txq;
2230 	iflib_rxq_t rxq;
2231 	int i, j, tx_ip_csum_flags, tx_ip6_csum_flags;
2232 
2233 
2234 	if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2235 	IFDI_INTR_DISABLE(ctx);
2236 
2237 	tx_ip_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP);
2238 	tx_ip6_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | CSUM_IP6_SCTP);
2239 	/* Set hardware offload abilities */
2240 	if_clearhwassist(ifp);
2241 	if (if_getcapenable(ifp) & IFCAP_TXCSUM)
2242 		if_sethwassistbits(ifp, tx_ip_csum_flags, 0);
2243 	if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6)
2244 		if_sethwassistbits(ifp,  tx_ip6_csum_flags, 0);
2245 	if (if_getcapenable(ifp) & IFCAP_TSO4)
2246 		if_sethwassistbits(ifp, CSUM_IP_TSO, 0);
2247 	if (if_getcapenable(ifp) & IFCAP_TSO6)
2248 		if_sethwassistbits(ifp, CSUM_IP6_TSO, 0);
2249 
2250 	for (i = 0, txq = ctx->ifc_txqs; i < sctx->isc_ntxqsets; i++, txq++) {
2251 		CALLOUT_LOCK(txq);
2252 		callout_stop(&txq->ift_timer);
2253 		CALLOUT_UNLOCK(txq);
2254 		iflib_netmap_txq_init(ctx, txq);
2255 	}
2256 #ifdef INVARIANTS
2257 	i = if_getdrvflags(ifp);
2258 #endif
2259 	IFDI_INIT(ctx);
2260 	MPASS(if_getdrvflags(ifp) == i);
2261 	for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
2262 		/* XXX this should really be done on a per-queue basis */
2263 		if (if_getcapenable(ifp) & IFCAP_NETMAP) {
2264 			MPASS(rxq->ifr_id == i);
2265 			iflib_netmap_rxq_init(ctx, rxq);
2266 			continue;
2267 		}
2268 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
2269 			if (iflib_fl_setup(fl)) {
2270 				device_printf(ctx->ifc_dev, "freelist setup failed - check cluster settings\n");
2271 				goto done;
2272 			}
2273 		}
2274 	}
2275 done:
2276 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
2277 	IFDI_INTR_ENABLE(ctx);
2278 	txq = ctx->ifc_txqs;
2279 	for (i = 0; i < sctx->isc_ntxqsets; i++, txq++)
2280 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq,
2281 			txq->ift_timer.c_cpu);
2282 }
2283 
2284 static int
2285 iflib_media_change(if_t ifp)
2286 {
2287 	if_ctx_t ctx = if_getsoftc(ifp);
2288 	int err;
2289 
2290 	CTX_LOCK(ctx);
2291 	if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0)
2292 		iflib_init_locked(ctx);
2293 	CTX_UNLOCK(ctx);
2294 	return (err);
2295 }
2296 
2297 static void
2298 iflib_media_status(if_t ifp, struct ifmediareq *ifmr)
2299 {
2300 	if_ctx_t ctx = if_getsoftc(ifp);
2301 
2302 	CTX_LOCK(ctx);
2303 	IFDI_UPDATE_ADMIN_STATUS(ctx);
2304 	IFDI_MEDIA_STATUS(ctx, ifmr);
2305 	CTX_UNLOCK(ctx);
2306 }
2307 
2308 void
2309 iflib_stop(if_ctx_t ctx)
2310 {
2311 	iflib_txq_t txq = ctx->ifc_txqs;
2312 	iflib_rxq_t rxq = ctx->ifc_rxqs;
2313 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2314 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2315 	iflib_dma_info_t di;
2316 	iflib_fl_t fl;
2317 	int i, j;
2318 
2319 	/* Tell the stack that the interface is no longer active */
2320 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2321 
2322 	IFDI_INTR_DISABLE(ctx);
2323 	DELAY(1000);
2324 	IFDI_STOP(ctx);
2325 	DELAY(1000);
2326 
2327 	iflib_debug_reset();
2328 	/* Wait for current tx queue users to exit to disarm watchdog timer. */
2329 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) {
2330 		/* make sure all transmitters have completed before proceeding XXX */
2331 
2332 		CALLOUT_LOCK(txq);
2333 		callout_stop(&txq->ift_timer);
2334 		CALLOUT_UNLOCK(txq);
2335 
2336 		/* clean any enqueued buffers */
2337 		iflib_ifmp_purge(txq);
2338 		/* Free any existing tx buffers. */
2339 		for (j = 0; j < txq->ift_size; j++) {
2340 			iflib_txsd_free(ctx, txq, j);
2341 		}
2342 		txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0;
2343 		txq->ift_in_use = txq->ift_gen = txq->ift_cidx = txq->ift_pidx = txq->ift_no_desc_avail = 0;
2344 		txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0;
2345 		txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0;
2346 		txq->ift_pullups = 0;
2347 		ifmp_ring_reset_stats(txq->ift_br);
2348 		for (j = 0, di = txq->ift_ifdi; j < sctx->isc_ntxqs; j++, di++)
2349 			bzero((void *)di->idi_vaddr, di->idi_size);
2350 	}
2351 	for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) {
2352 		/* make sure all transmitters have completed before proceeding XXX */
2353 
2354 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
2355 		for (j = 0, di = rxq->ifr_ifdi; j < sctx->isc_nrxqs; j++, di++)
2356 			bzero((void *)di->idi_vaddr, di->idi_size);
2357 		/* also resets the free lists pidx/cidx */
2358 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
2359 			iflib_fl_bufs_free(fl);
2360 	}
2361 }
2362 
2363 static inline caddr_t
2364 calc_next_rxd(iflib_fl_t fl, int cidx)
2365 {
2366 	qidx_t size;
2367 	int nrxd;
2368 	caddr_t start, end, cur, next;
2369 
2370 	nrxd = fl->ifl_size;
2371 	size = fl->ifl_rxd_size;
2372 	start = fl->ifl_ifdi->idi_vaddr;
2373 
2374 	if (__predict_false(size == 0))
2375 		return (start);
2376 	cur = start + size*cidx;
2377 	end = start + size*nrxd;
2378 	next = CACHE_PTR_NEXT(cur);
2379 	return (next < end ? next : start);
2380 }
2381 
2382 static inline void
2383 prefetch_pkts(iflib_fl_t fl, int cidx)
2384 {
2385 	int nextptr;
2386 	int nrxd = fl->ifl_size;
2387 	caddr_t next_rxd;
2388 
2389 
2390 	nextptr = (cidx + CACHE_PTR_INCREMENT) & (nrxd-1);
2391 	prefetch(&fl->ifl_sds.ifsd_m[nextptr]);
2392 	prefetch(&fl->ifl_sds.ifsd_cl[nextptr]);
2393 	next_rxd = calc_next_rxd(fl, cidx);
2394 	prefetch(next_rxd);
2395 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 1) & (nrxd-1)]);
2396 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 2) & (nrxd-1)]);
2397 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 3) & (nrxd-1)]);
2398 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 4) & (nrxd-1)]);
2399 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 1) & (nrxd-1)]);
2400 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 2) & (nrxd-1)]);
2401 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 3) & (nrxd-1)]);
2402 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 4) & (nrxd-1)]);
2403 }
2404 
2405 static void
2406 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, int unload, if_rxsd_t sd)
2407 {
2408 	int flid, cidx;
2409 	bus_dmamap_t map;
2410 	iflib_fl_t fl;
2411 	iflib_dma_info_t di;
2412 	int next;
2413 
2414 	map = NULL;
2415 	flid = irf->irf_flid;
2416 	cidx = irf->irf_idx;
2417 	fl = &rxq->ifr_fl[flid];
2418 	sd->ifsd_fl = fl;
2419 	sd->ifsd_cidx = cidx;
2420 	sd->ifsd_m = &fl->ifl_sds.ifsd_m[cidx];
2421 	sd->ifsd_cl = &fl->ifl_sds.ifsd_cl[cidx];
2422 	fl->ifl_credits--;
2423 #if MEMORY_LOGGING
2424 	fl->ifl_m_dequeued++;
2425 #endif
2426 	if (rxq->ifr_ctx->ifc_flags & IFC_PREFETCH)
2427 		prefetch_pkts(fl, cidx);
2428 	next = (cidx + CACHE_PTR_INCREMENT) & (fl->ifl_size-1);
2429 	prefetch(&fl->ifl_sds.ifsd_map[next]);
2430 	map = fl->ifl_sds.ifsd_map[cidx];
2431 	di = fl->ifl_ifdi;
2432 	next = (cidx + CACHE_LINE_SIZE) & (fl->ifl_size-1);
2433 	bus_dmamap_sync(di->idi_tag, di->idi_map,
2434 	    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
2435 
2436 	/* not valid assert if bxe really does SGE from non-contiguous elements */
2437 	MPASS(fl->ifl_cidx == cidx);
2438 	if (unload)
2439 		bus_dmamap_unload(fl->ifl_desc_tag, map);
2440 	fl->ifl_cidx = (fl->ifl_cidx + 1) & (fl->ifl_size-1);
2441 	if (__predict_false(fl->ifl_cidx == 0))
2442 		fl->ifl_gen = 0;
2443 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2444 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2445         bit_clear(fl->ifl_rx_bitmap, cidx);
2446 }
2447 
2448 static struct mbuf *
2449 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri, if_rxsd_t sd)
2450 {
2451 	int i, padlen , flags;
2452 	struct mbuf *m, *mh, *mt;
2453 	caddr_t cl;
2454 
2455 	i = 0;
2456 	mh = NULL;
2457 	do {
2458 		rxd_frag_to_sd(rxq, &ri->iri_frags[i], TRUE, sd);
2459 
2460 		MPASS(*sd->ifsd_cl != NULL);
2461 		MPASS(*sd->ifsd_m != NULL);
2462 
2463 		/* Don't include zero-length frags */
2464 		if (ri->iri_frags[i].irf_len == 0) {
2465 			/* XXX we can save the cluster here, but not the mbuf */
2466 			m_init(*sd->ifsd_m, M_NOWAIT, MT_DATA, 0);
2467 			m_free(*sd->ifsd_m);
2468 			*sd->ifsd_m = NULL;
2469 			continue;
2470 		}
2471 		m = *sd->ifsd_m;
2472 		*sd->ifsd_m = NULL;
2473 		if (mh == NULL) {
2474 			flags = M_PKTHDR|M_EXT;
2475 			mh = mt = m;
2476 			padlen = ri->iri_pad;
2477 		} else {
2478 			flags = M_EXT;
2479 			mt->m_next = m;
2480 			mt = m;
2481 			/* assuming padding is only on the first fragment */
2482 			padlen = 0;
2483 		}
2484 		cl = *sd->ifsd_cl;
2485 		*sd->ifsd_cl = NULL;
2486 
2487 		/* Can these two be made one ? */
2488 		m_init(m, M_NOWAIT, MT_DATA, flags);
2489 		m_cljset(m, cl, sd->ifsd_fl->ifl_cltype);
2490 		/*
2491 		 * These must follow m_init and m_cljset
2492 		 */
2493 		m->m_data += padlen;
2494 		ri->iri_len -= padlen;
2495 		m->m_len = ri->iri_frags[i].irf_len;
2496 	} while (++i < ri->iri_nfrags);
2497 
2498 	return (mh);
2499 }
2500 
2501 /*
2502  * Process one software descriptor
2503  */
2504 static struct mbuf *
2505 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri)
2506 {
2507 	struct if_rxsd sd;
2508 	struct mbuf *m;
2509 
2510 	/* should I merge this back in now that the two paths are basically duplicated? */
2511 	if (ri->iri_nfrags == 1 &&
2512 	    ri->iri_frags[0].irf_len <= MIN(IFLIB_RX_COPY_THRESH, MHLEN)) {
2513 		rxd_frag_to_sd(rxq, &ri->iri_frags[0], FALSE, &sd);
2514 		m = *sd.ifsd_m;
2515 		*sd.ifsd_m = NULL;
2516 		m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR);
2517 #ifndef __NO_STRICT_ALIGNMENT
2518 		if (!IP_ALIGNED(m))
2519 			m->m_data += 2;
2520 #endif
2521 		memcpy(m->m_data, *sd.ifsd_cl, ri->iri_len);
2522 		m->m_len = ri->iri_frags[0].irf_len;
2523        } else {
2524 		m = assemble_segments(rxq, ri, &sd);
2525 	}
2526 	m->m_pkthdr.len = ri->iri_len;
2527 	m->m_pkthdr.rcvif = ri->iri_ifp;
2528 	m->m_flags |= ri->iri_flags;
2529 	m->m_pkthdr.ether_vtag = ri->iri_vtag;
2530 	m->m_pkthdr.flowid = ri->iri_flowid;
2531 	M_HASHTYPE_SET(m, ri->iri_rsstype);
2532 	m->m_pkthdr.csum_flags = ri->iri_csum_flags;
2533 	m->m_pkthdr.csum_data = ri->iri_csum_data;
2534 	return (m);
2535 }
2536 
2537 #if defined(INET6) || defined(INET)
2538 static void
2539 iflib_get_ip_forwarding(struct lro_ctrl *lc, bool *v4, bool *v6)
2540 {
2541 	CURVNET_SET(lc->ifp->if_vnet);
2542 #if defined(INET6)
2543 	*v6 = VNET(ip6_forwarding);
2544 #endif
2545 #if defined(INET)
2546 	*v4 = VNET(ipforwarding);
2547 #endif
2548 	CURVNET_RESTORE();
2549 }
2550 
2551 /*
2552  * Returns true if it's possible this packet could be LROed.
2553  * if it returns false, it is guaranteed that tcp_lro_rx()
2554  * would not return zero.
2555  */
2556 static bool
2557 iflib_check_lro_possible(struct mbuf *m, bool v4_forwarding, bool v6_forwarding)
2558 {
2559 	struct ether_header *eh;
2560 	uint16_t eh_type;
2561 
2562 	eh = mtod(m, struct ether_header *);
2563 	eh_type = ntohs(eh->ether_type);
2564 	switch (eh_type) {
2565 #if defined(INET6)
2566 		case ETHERTYPE_IPV6:
2567 			return !v6_forwarding;
2568 #endif
2569 #if defined (INET)
2570 		case ETHERTYPE_IP:
2571 			return !v4_forwarding;
2572 #endif
2573 	}
2574 
2575 	return false;
2576 }
2577 #else
2578 static void
2579 iflib_get_ip_forwarding(struct lro_ctrl *lc __unused, bool *v4 __unused, bool *v6 __unused)
2580 {
2581 }
2582 #endif
2583 
2584 static bool
2585 iflib_rxeof(iflib_rxq_t rxq, qidx_t budget)
2586 {
2587 	if_ctx_t ctx = rxq->ifr_ctx;
2588 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2589 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2590 	int avail, i;
2591 	qidx_t *cidxp;
2592 	struct if_rxd_info ri;
2593 	int err, budget_left, rx_bytes, rx_pkts;
2594 	iflib_fl_t fl;
2595 	struct ifnet *ifp;
2596 	int lro_enabled;
2597 	bool v4_forwarding, v6_forwarding, lro_possible;
2598 
2599 	/*
2600 	 * XXX early demux data packets so that if_input processing only handles
2601 	 * acks in interrupt context
2602 	 */
2603 	struct mbuf *m, *mh, *mt, *mf;
2604 
2605 	lro_possible = v4_forwarding = v6_forwarding = false;
2606 	ifp = ctx->ifc_ifp;
2607 	mh = mt = NULL;
2608 	MPASS(budget > 0);
2609 	rx_pkts	= rx_bytes = 0;
2610 	if (sctx->isc_flags & IFLIB_HAS_RXCQ)
2611 		cidxp = &rxq->ifr_cq_cidx;
2612 	else
2613 		cidxp = &rxq->ifr_fl[0].ifl_cidx;
2614 	if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) {
2615 		for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2616 			__iflib_fl_refill_lt(ctx, fl, budget + 8);
2617 		DBG_COUNTER_INC(rx_unavail);
2618 		return (false);
2619 	}
2620 
2621 	for (budget_left = budget; budget_left > 0 && avail > 0;) {
2622 		if (__predict_false(!CTX_ACTIVE(ctx))) {
2623 			DBG_COUNTER_INC(rx_ctx_inactive);
2624 			break;
2625 		}
2626 		/*
2627 		 * Reset client set fields to their default values
2628 		 */
2629 		rxd_info_zero(&ri);
2630 		ri.iri_qsidx = rxq->ifr_id;
2631 		ri.iri_cidx = *cidxp;
2632 		ri.iri_ifp = ifp;
2633 		ri.iri_frags = rxq->ifr_frags;
2634 		err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
2635 
2636 		if (err)
2637 			goto err;
2638 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
2639 			*cidxp = ri.iri_cidx;
2640 			/* Update our consumer index */
2641 			/* XXX NB: shurd - check if this is still safe */
2642 			while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0]) {
2643 				rxq->ifr_cq_cidx -= scctx->isc_nrxd[0];
2644 				rxq->ifr_cq_gen = 0;
2645 			}
2646 			/* was this only a completion queue message? */
2647 			if (__predict_false(ri.iri_nfrags == 0))
2648 				continue;
2649 		}
2650 		MPASS(ri.iri_nfrags != 0);
2651 		MPASS(ri.iri_len != 0);
2652 
2653 		/* will advance the cidx on the corresponding free lists */
2654 		m = iflib_rxd_pkt_get(rxq, &ri);
2655 		avail--;
2656 		budget_left--;
2657 		if (avail == 0 && budget_left)
2658 			avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left);
2659 
2660 		if (__predict_false(m == NULL)) {
2661 			DBG_COUNTER_INC(rx_mbuf_null);
2662 			continue;
2663 		}
2664 		/* imm_pkt: -- cxgb */
2665 		if (mh == NULL)
2666 			mh = mt = m;
2667 		else {
2668 			mt->m_nextpkt = m;
2669 			mt = m;
2670 		}
2671 	}
2672 	/* make sure that we can refill faster than drain */
2673 	for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2674 		__iflib_fl_refill_lt(ctx, fl, budget + 8);
2675 
2676 	lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO);
2677 	if (lro_enabled)
2678 		iflib_get_ip_forwarding(&rxq->ifr_lc, &v4_forwarding, &v6_forwarding);
2679 	mt = mf = NULL;
2680 	while (mh != NULL) {
2681 		m = mh;
2682 		mh = mh->m_nextpkt;
2683 		m->m_nextpkt = NULL;
2684 #ifndef __NO_STRICT_ALIGNMENT
2685 		if (!IP_ALIGNED(m) && (m = iflib_fixup_rx(m)) == NULL)
2686 			continue;
2687 #endif
2688 		rx_bytes += m->m_pkthdr.len;
2689 		rx_pkts++;
2690 #if defined(INET6) || defined(INET)
2691 		if (lro_enabled) {
2692 			if (!lro_possible) {
2693 				lro_possible = iflib_check_lro_possible(m, v4_forwarding, v6_forwarding);
2694 				if (lro_possible && mf != NULL) {
2695 					ifp->if_input(ifp, mf);
2696 					DBG_COUNTER_INC(rx_if_input);
2697 					mt = mf = NULL;
2698 				}
2699 			}
2700 			if ((m->m_pkthdr.csum_flags & (CSUM_L4_CALC|CSUM_L4_VALID)) ==
2701 			    (CSUM_L4_CALC|CSUM_L4_VALID)) {
2702 				if (lro_possible && tcp_lro_rx(&rxq->ifr_lc, m, 0) == 0)
2703 					continue;
2704 			}
2705 		}
2706 #endif
2707 		if (lro_possible) {
2708 			ifp->if_input(ifp, m);
2709 			DBG_COUNTER_INC(rx_if_input);
2710 			continue;
2711 		}
2712 
2713 		if (mf == NULL)
2714 			mf = m;
2715 		if (mt != NULL)
2716 			mt->m_nextpkt = m;
2717 		mt = m;
2718 	}
2719 	if (mf != NULL) {
2720 		ifp->if_input(ifp, mf);
2721 		DBG_COUNTER_INC(rx_if_input);
2722 	}
2723 
2724 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
2725 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
2726 
2727 	/*
2728 	 * Flush any outstanding LRO work
2729 	 */
2730 #if defined(INET6) || defined(INET)
2731 	tcp_lro_flush_all(&rxq->ifr_lc);
2732 #endif
2733 	if (avail)
2734 		return true;
2735 	return (iflib_rxd_avail(ctx, rxq, *cidxp, 1));
2736 err:
2737 	STATE_LOCK(ctx);
2738 	ctx->ifc_flags |= IFC_DO_RESET;
2739 	iflib_admin_intr_deferred(ctx);
2740 	STATE_UNLOCK(ctx);
2741 	return (false);
2742 }
2743 
2744 #define TXD_NOTIFY_COUNT(txq) (((txq)->ift_size / (txq)->ift_update_freq)-1)
2745 static inline qidx_t
2746 txq_max_db_deferred(iflib_txq_t txq, qidx_t in_use)
2747 {
2748 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2749 	qidx_t minthresh = txq->ift_size / 8;
2750 	if (in_use > 4*minthresh)
2751 		return (notify_count);
2752 	if (in_use > 2*minthresh)
2753 		return (notify_count >> 1);
2754 	if (in_use > minthresh)
2755 		return (notify_count >> 3);
2756 	return (0);
2757 }
2758 
2759 static inline qidx_t
2760 txq_max_rs_deferred(iflib_txq_t txq)
2761 {
2762 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2763 	qidx_t minthresh = txq->ift_size / 8;
2764 	if (txq->ift_in_use > 4*minthresh)
2765 		return (notify_count);
2766 	if (txq->ift_in_use > 2*minthresh)
2767 		return (notify_count >> 1);
2768 	if (txq->ift_in_use > minthresh)
2769 		return (notify_count >> 2);
2770 	return (2);
2771 }
2772 
2773 #define M_CSUM_FLAGS(m) ((m)->m_pkthdr.csum_flags)
2774 #define M_HAS_VLANTAG(m) (m->m_flags & M_VLANTAG)
2775 
2776 #define TXQ_MAX_DB_DEFERRED(txq, in_use) txq_max_db_deferred((txq), (in_use))
2777 #define TXQ_MAX_RS_DEFERRED(txq) txq_max_rs_deferred(txq)
2778 #define TXQ_MAX_DB_CONSUMED(size) (size >> 4)
2779 
2780 /* forward compatibility for cxgb */
2781 #define FIRST_QSET(ctx) 0
2782 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets)
2783 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets)
2784 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx))
2785 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments))
2786 
2787 /* XXX we should be setting this to something other than zero */
2788 #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh)
2789 #define	MAX_TX_DESC(ctx) max((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max, \
2790     (ctx)->ifc_softc_ctx.isc_tx_nsegments)
2791 
2792 static inline bool
2793 iflib_txd_db_check(if_ctx_t ctx, iflib_txq_t txq, int ring, qidx_t in_use)
2794 {
2795 	qidx_t dbval, max;
2796 	bool rang;
2797 
2798 	rang = false;
2799 	max = TXQ_MAX_DB_DEFERRED(txq, in_use);
2800 	if (ring || txq->ift_db_pending >= max) {
2801 		dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx;
2802 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval);
2803 		txq->ift_db_pending = txq->ift_npending = 0;
2804 		rang = true;
2805 	}
2806 	return (rang);
2807 }
2808 
2809 #ifdef PKT_DEBUG
2810 static void
2811 print_pkt(if_pkt_info_t pi)
2812 {
2813 	printf("pi len:  %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n",
2814 	       pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx);
2815 	printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n",
2816 	       pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag);
2817 	printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n",
2818 	       pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto);
2819 }
2820 #endif
2821 
2822 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO)
2823 #define IS_TX_OFFLOAD4(pi) ((pi)->ipi_csum_flags & (CSUM_IP_TCP | CSUM_IP_TSO))
2824 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO)
2825 #define IS_TX_OFFLOAD6(pi) ((pi)->ipi_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_TSO))
2826 
2827 static int
2828 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp)
2829 {
2830 	if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx;
2831 	struct ether_vlan_header *eh;
2832 	struct mbuf *m;
2833 
2834 	m = *mp;
2835 	if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) &&
2836 	    M_WRITABLE(m) == 0) {
2837 		if ((m = m_dup(m, M_NOWAIT)) == NULL) {
2838 			return (ENOMEM);
2839 		} else {
2840 			m_freem(*mp);
2841 			DBG_COUNTER_INC(tx_frees);
2842 			*mp = m;
2843 		}
2844 	}
2845 
2846 	/*
2847 	 * Determine where frame payload starts.
2848 	 * Jump over vlan headers if already present,
2849 	 * helpful for QinQ too.
2850 	 */
2851 	if (__predict_false(m->m_len < sizeof(*eh))) {
2852 		txq->ift_pullups++;
2853 		if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL))
2854 			return (ENOMEM);
2855 	}
2856 	eh = mtod(m, struct ether_vlan_header *);
2857 	if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
2858 		pi->ipi_etype = ntohs(eh->evl_proto);
2859 		pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
2860 	} else {
2861 		pi->ipi_etype = ntohs(eh->evl_encap_proto);
2862 		pi->ipi_ehdrlen = ETHER_HDR_LEN;
2863 	}
2864 
2865 	switch (pi->ipi_etype) {
2866 #ifdef INET
2867 	case ETHERTYPE_IP:
2868 	{
2869 		struct mbuf *n;
2870 		struct ip *ip = NULL;
2871 		struct tcphdr *th = NULL;
2872 		int minthlen;
2873 
2874 		minthlen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip) + sizeof(*th));
2875 		if (__predict_false(m->m_len < minthlen)) {
2876 			/*
2877 			 * if this code bloat is causing too much of a hit
2878 			 * move it to a separate function and mark it noinline
2879 			 */
2880 			if (m->m_len == pi->ipi_ehdrlen) {
2881 				n = m->m_next;
2882 				MPASS(n);
2883 				if (n->m_len >= sizeof(*ip))  {
2884 					ip = (struct ip *)n->m_data;
2885 					if (n->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2886 						th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2887 				} else {
2888 					txq->ift_pullups++;
2889 					if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
2890 						return (ENOMEM);
2891 					ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2892 				}
2893 			} else {
2894 				txq->ift_pullups++;
2895 				if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
2896 					return (ENOMEM);
2897 				ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2898 				if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2899 					th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2900 			}
2901 		} else {
2902 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2903 			if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2904 				th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2905 		}
2906 		pi->ipi_ip_hlen = ip->ip_hl << 2;
2907 		pi->ipi_ipproto = ip->ip_p;
2908 		pi->ipi_flags |= IPI_TX_IPV4;
2909 
2910 		/* TCP checksum offload may require TCP header length */
2911 		if (IS_TX_OFFLOAD4(pi)) {
2912 			if (__predict_true(pi->ipi_ipproto == IPPROTO_TCP)) {
2913 				if (__predict_false(th == NULL)) {
2914 					txq->ift_pullups++;
2915 					if (__predict_false((m = m_pullup(m, (ip->ip_hl << 2) + sizeof(*th))) == NULL))
2916 						return (ENOMEM);
2917 					th = (struct tcphdr *)((caddr_t)ip + pi->ipi_ip_hlen);
2918 				}
2919 				pi->ipi_tcp_hflags = th->th_flags;
2920 				pi->ipi_tcp_hlen = th->th_off << 2;
2921 				pi->ipi_tcp_seq = th->th_seq;
2922 			}
2923 			if (IS_TSO4(pi)) {
2924 				if (__predict_false(ip->ip_p != IPPROTO_TCP))
2925 					return (ENXIO);
2926 				/*
2927 				 * TSO always requires hardware checksum offload.
2928 				 */
2929 				pi->ipi_csum_flags |= (CSUM_IP_TCP | CSUM_IP);
2930 				th->th_sum = in_pseudo(ip->ip_src.s_addr,
2931 						       ip->ip_dst.s_addr, htons(IPPROTO_TCP));
2932 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
2933 				if (sctx->isc_flags & IFLIB_TSO_INIT_IP) {
2934 					ip->ip_sum = 0;
2935 					ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz);
2936 				}
2937 			}
2938 		}
2939 		if ((sctx->isc_flags & IFLIB_NEED_ZERO_CSUM) && (pi->ipi_csum_flags & CSUM_IP))
2940                        ip->ip_sum = 0;
2941 
2942 		break;
2943 	}
2944 #endif
2945 #ifdef INET6
2946 	case ETHERTYPE_IPV6:
2947 	{
2948 		struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
2949 		struct tcphdr *th;
2950 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
2951 
2952 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
2953 			txq->ift_pullups++;
2954 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
2955 				return (ENOMEM);
2956 		}
2957 		th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
2958 
2959 		/* XXX-BZ this will go badly in case of ext hdrs. */
2960 		pi->ipi_ipproto = ip6->ip6_nxt;
2961 		pi->ipi_flags |= IPI_TX_IPV6;
2962 
2963 		/* TCP checksum offload may require TCP header length */
2964 		if (IS_TX_OFFLOAD6(pi)) {
2965 			if (pi->ipi_ipproto == IPPROTO_TCP) {
2966 				if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) {
2967 					txq->ift_pullups++;
2968 					if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL))
2969 						return (ENOMEM);
2970 				}
2971 				pi->ipi_tcp_hflags = th->th_flags;
2972 				pi->ipi_tcp_hlen = th->th_off << 2;
2973 				pi->ipi_tcp_seq = th->th_seq;
2974 			}
2975 			if (IS_TSO6(pi)) {
2976 				if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP))
2977 					return (ENXIO);
2978 				/*
2979 				 * TSO always requires hardware checksum offload.
2980 				 */
2981 				pi->ipi_csum_flags |= CSUM_IP6_TCP;
2982 				th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0);
2983 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
2984 			}
2985 		}
2986 		break;
2987 	}
2988 #endif
2989 	default:
2990 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
2991 		pi->ipi_ip_hlen = 0;
2992 		break;
2993 	}
2994 	*mp = m;
2995 
2996 	return (0);
2997 }
2998 
2999 /*
3000  * If dodgy hardware rejects the scatter gather chain we've handed it
3001  * we'll need to remove the mbuf chain from ifsg_m[] before we can add the
3002  * m_defrag'd mbufs
3003  */
3004 static __noinline struct mbuf *
3005 iflib_remove_mbuf(iflib_txq_t txq)
3006 {
3007 	int ntxd, pidx;
3008 	struct mbuf *m, **ifsd_m;
3009 	bus_dmamap_t *ifsd_map;
3010 
3011 	ifsd_m = txq->ift_sds.ifsd_m;
3012 	ntxd = txq->ift_size;
3013 	pidx = txq->ift_pidx & (ntxd - 1);
3014 	ifsd_m = txq->ift_sds.ifsd_m;
3015 	ifsd_map = txq->ift_sds.ifsd_map;
3016 	m = ifsd_m[pidx];
3017 	ifsd_m[pidx] = NULL;
3018 	bus_dmamap_unload(txq->ift_desc_tag, ifsd_map[pidx]);
3019 #if MEMORY_LOGGING
3020 	txq->ift_dequeued++;
3021 #endif
3022 	return (m);
3023 }
3024 
3025 static inline caddr_t
3026 calc_next_txd(iflib_txq_t txq, int cidx, uint8_t qid)
3027 {
3028 	qidx_t size;
3029 	int ntxd;
3030 	caddr_t start, end, cur, next;
3031 
3032 	ntxd = txq->ift_size;
3033 	size = txq->ift_txd_size[qid];
3034 	start = txq->ift_ifdi[qid].idi_vaddr;
3035 
3036 	if (__predict_false(size == 0))
3037 		return (start);
3038 	cur = start + size*cidx;
3039 	end = start + size*ntxd;
3040 	next = CACHE_PTR_NEXT(cur);
3041 	return (next < end ? next : start);
3042 }
3043 
3044 /*
3045  * Pad an mbuf to ensure a minimum ethernet frame size.
3046  * min_frame_size is the frame size (less CRC) to pad the mbuf to
3047  */
3048 static __noinline int
3049 iflib_ether_pad(device_t dev, struct mbuf **m_head, uint16_t min_frame_size)
3050 {
3051 	/*
3052 	 * 18 is enough bytes to pad an ARP packet to 46 bytes, and
3053 	 * and ARP message is the smallest common payload I can think of
3054 	 */
3055 	static char pad[18];	/* just zeros */
3056 	int n;
3057 	struct mbuf *new_head;
3058 
3059 	if (!M_WRITABLE(*m_head)) {
3060 		new_head = m_dup(*m_head, M_NOWAIT);
3061 		if (new_head == NULL) {
3062 			m_freem(*m_head);
3063 			device_printf(dev, "cannot pad short frame, m_dup() failed");
3064 			DBG_COUNTER_INC(encap_pad_mbuf_fail);
3065 			DBG_COUNTER_INC(tx_frees);
3066 			return ENOMEM;
3067 		}
3068 		m_freem(*m_head);
3069 		*m_head = new_head;
3070 	}
3071 
3072 	for (n = min_frame_size - (*m_head)->m_pkthdr.len;
3073 	     n > 0; n -= sizeof(pad))
3074 		if (!m_append(*m_head, min(n, sizeof(pad)), pad))
3075 			break;
3076 
3077 	if (n > 0) {
3078 		m_freem(*m_head);
3079 		device_printf(dev, "cannot pad short frame\n");
3080 		DBG_COUNTER_INC(encap_pad_mbuf_fail);
3081 		DBG_COUNTER_INC(tx_frees);
3082 		return (ENOBUFS);
3083 	}
3084 
3085 	return 0;
3086 }
3087 
3088 static int
3089 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp)
3090 {
3091 	if_ctx_t		ctx;
3092 	if_shared_ctx_t		sctx;
3093 	if_softc_ctx_t		scctx;
3094 	bus_dma_segment_t	*segs;
3095 	struct mbuf		*m_head, **ifsd_m;
3096 	void			*next_txd;
3097 	bus_dmamap_t		map;
3098 	struct if_pkt_info	pi;
3099 	int remap = 0;
3100 	int err, nsegs, ndesc, max_segs, pidx, cidx, next, ntxd;
3101 	bus_dma_tag_t desc_tag;
3102 
3103 	ctx = txq->ift_ctx;
3104 	sctx = ctx->ifc_sctx;
3105 	scctx = &ctx->ifc_softc_ctx;
3106 	segs = txq->ift_segs;
3107 	ntxd = txq->ift_size;
3108 	m_head = *m_headp;
3109 	map = NULL;
3110 
3111 	/*
3112 	 * If we're doing TSO the next descriptor to clean may be quite far ahead
3113 	 */
3114 	cidx = txq->ift_cidx;
3115 	pidx = txq->ift_pidx;
3116 	if (ctx->ifc_flags & IFC_PREFETCH) {
3117 		next = (cidx + CACHE_PTR_INCREMENT) & (ntxd-1);
3118 		if (!(ctx->ifc_flags & IFLIB_HAS_TXCQ)) {
3119 			next_txd = calc_next_txd(txq, cidx, 0);
3120 			prefetch(next_txd);
3121 		}
3122 
3123 		/* prefetch the next cache line of mbuf pointers and flags */
3124 		prefetch(&txq->ift_sds.ifsd_m[next]);
3125 		prefetch(&txq->ift_sds.ifsd_map[next]);
3126 		next = (cidx + CACHE_LINE_SIZE) & (ntxd-1);
3127 	}
3128 	map = txq->ift_sds.ifsd_map[pidx];
3129 	ifsd_m = txq->ift_sds.ifsd_m;
3130 
3131 	if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
3132 		desc_tag = txq->ift_tso_desc_tag;
3133 		max_segs = scctx->isc_tx_tso_segments_max;
3134 		MPASS(desc_tag != NULL);
3135 		MPASS(max_segs > 0);
3136 	} else {
3137 		desc_tag = txq->ift_desc_tag;
3138 		max_segs = scctx->isc_tx_nsegments;
3139 	}
3140 	if ((sctx->isc_flags & IFLIB_NEED_ETHER_PAD) &&
3141 	    __predict_false(m_head->m_pkthdr.len < scctx->isc_min_frame_size)) {
3142 		err = iflib_ether_pad(ctx->ifc_dev, m_headp, scctx->isc_min_frame_size);
3143 		if (err) {
3144 			DBG_COUNTER_INC(encap_txd_encap_fail);
3145 			return err;
3146 		}
3147 	}
3148 	m_head = *m_headp;
3149 
3150 	pkt_info_zero(&pi);
3151 	pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG|M_BCAST|M_MCAST));
3152 	pi.ipi_pidx = pidx;
3153 	pi.ipi_qsidx = txq->ift_id;
3154 	pi.ipi_len = m_head->m_pkthdr.len;
3155 	pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags;
3156 	pi.ipi_vtag = (m_head->m_flags & M_VLANTAG) ? m_head->m_pkthdr.ether_vtag : 0;
3157 
3158 	/* deliberate bitwise OR to make one condition */
3159 	if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) {
3160 		if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0)) {
3161 			DBG_COUNTER_INC(encap_txd_encap_fail);
3162 			return (err);
3163 		}
3164 		m_head = *m_headp;
3165 	}
3166 
3167 retry:
3168 	err = bus_dmamap_load_mbuf_sg(desc_tag, map, m_head, segs, &nsegs,
3169 	    BUS_DMA_NOWAIT);
3170 defrag:
3171 	if (__predict_false(err)) {
3172 		switch (err) {
3173 		case EFBIG:
3174 			/* try collapse once and defrag once */
3175 			if (remap == 0) {
3176 				m_head = m_collapse(*m_headp, M_NOWAIT, max_segs);
3177 				/* try defrag if collapsing fails */
3178 				if (m_head == NULL)
3179 					remap++;
3180 			}
3181 			if (remap == 1) {
3182 				txq->ift_mbuf_defrag++;
3183 				m_head = m_defrag(*m_headp, M_NOWAIT);
3184 			}
3185 			remap++;
3186 			if (__predict_false(m_head == NULL))
3187 				goto defrag_failed;
3188 			*m_headp = m_head;
3189 			goto retry;
3190 			break;
3191 		case ENOMEM:
3192 			txq->ift_no_tx_dma_setup++;
3193 			break;
3194 		default:
3195 			txq->ift_no_tx_dma_setup++;
3196 			m_freem(*m_headp);
3197 			DBG_COUNTER_INC(tx_frees);
3198 			*m_headp = NULL;
3199 			break;
3200 		}
3201 		txq->ift_map_failed++;
3202 		DBG_COUNTER_INC(encap_load_mbuf_fail);
3203 		DBG_COUNTER_INC(encap_txd_encap_fail);
3204 		return (err);
3205 	}
3206 	ifsd_m[pidx] = m_head;
3207 	/*
3208 	 * XXX assumes a 1 to 1 relationship between segments and
3209 	 *        descriptors - this does not hold true on all drivers, e.g.
3210 	 *        cxgb
3211 	 */
3212 	if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) {
3213 		txq->ift_no_desc_avail++;
3214 		bus_dmamap_unload(desc_tag, map);
3215 		DBG_COUNTER_INC(encap_txq_avail_fail);
3216 		DBG_COUNTER_INC(encap_txd_encap_fail);
3217 		if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0)
3218 			GROUPTASK_ENQUEUE(&txq->ift_task);
3219 		return (ENOBUFS);
3220 	}
3221 	/*
3222 	 * On Intel cards we can greatly reduce the number of TX interrupts
3223 	 * we see by only setting report status on every Nth descriptor.
3224 	 * However, this also means that the driver will need to keep track
3225 	 * of the descriptors that RS was set on to check them for the DD bit.
3226 	 */
3227 	txq->ift_rs_pending += nsegs + 1;
3228 	if (txq->ift_rs_pending > TXQ_MAX_RS_DEFERRED(txq) ||
3229 	     iflib_no_tx_batch || (TXQ_AVAIL(txq) - nsegs) <= MAX_TX_DESC(ctx) + 2) {
3230 		pi.ipi_flags |= IPI_TX_INTR;
3231 		txq->ift_rs_pending = 0;
3232 	}
3233 
3234 	pi.ipi_segs = segs;
3235 	pi.ipi_nsegs = nsegs;
3236 
3237 	MPASS(pidx >= 0 && pidx < txq->ift_size);
3238 #ifdef PKT_DEBUG
3239 	print_pkt(&pi);
3240 #endif
3241 	bus_dmamap_sync(desc_tag, map, BUS_DMASYNC_PREWRITE);
3242 	if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) {
3243 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3244 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
3245 		DBG_COUNTER_INC(tx_encap);
3246 		MPASS(pi.ipi_new_pidx < txq->ift_size);
3247 
3248 		ndesc = pi.ipi_new_pidx - pi.ipi_pidx;
3249 		if (pi.ipi_new_pidx < pi.ipi_pidx) {
3250 			ndesc += txq->ift_size;
3251 			txq->ift_gen = 1;
3252 		}
3253 		/*
3254 		 * drivers can need as many as
3255 		 * two sentinels
3256 		 */
3257 		MPASS(ndesc <= pi.ipi_nsegs + 2);
3258 		MPASS(pi.ipi_new_pidx != pidx);
3259 		MPASS(ndesc > 0);
3260 		txq->ift_in_use += ndesc;
3261 
3262 		/*
3263 		 * We update the last software descriptor again here because there may
3264 		 * be a sentinel and/or there may be more mbufs than segments
3265 		 */
3266 		txq->ift_pidx = pi.ipi_new_pidx;
3267 		txq->ift_npending += pi.ipi_ndescs;
3268 	} else {
3269 		*m_headp = m_head = iflib_remove_mbuf(txq);
3270 		if (err == EFBIG) {
3271 			txq->ift_txd_encap_efbig++;
3272 			if (remap < 2) {
3273 				remap = 1;
3274 				goto defrag;
3275 			}
3276 		}
3277 		goto defrag_failed;
3278 	}
3279 	/*
3280 	 * err can't possibly be non-zero here, so we don't neet to test it
3281 	 * to see if we need to DBG_COUNTER_INC(encap_txd_encap_fail).
3282 	 */
3283 	return (err);
3284 
3285 defrag_failed:
3286 	txq->ift_mbuf_defrag_failed++;
3287 	txq->ift_map_failed++;
3288 	m_freem(*m_headp);
3289 	DBG_COUNTER_INC(tx_frees);
3290 	*m_headp = NULL;
3291 	DBG_COUNTER_INC(encap_txd_encap_fail);
3292 	return (ENOMEM);
3293 }
3294 
3295 static void
3296 iflib_tx_desc_free(iflib_txq_t txq, int n)
3297 {
3298 	uint32_t qsize, cidx, mask, gen;
3299 	struct mbuf *m, **ifsd_m;
3300 	bus_dmamap_t *ifsd_map;
3301 	bool do_prefetch;
3302 
3303 	cidx = txq->ift_cidx;
3304 	gen = txq->ift_gen;
3305 	qsize = txq->ift_size;
3306 	mask = qsize-1;
3307 	ifsd_m = txq->ift_sds.ifsd_m;
3308 	ifsd_map = txq->ift_sds.ifsd_map;
3309 	do_prefetch = (txq->ift_ctx->ifc_flags & IFC_PREFETCH);
3310 
3311 	while (n-- > 0) {
3312 		if (do_prefetch) {
3313 			prefetch(ifsd_m[(cidx + 3) & mask]);
3314 			prefetch(ifsd_m[(cidx + 4) & mask]);
3315 		}
3316 		if ((m = ifsd_m[cidx]) != NULL) {
3317 			prefetch(&ifsd_m[(cidx + CACHE_PTR_INCREMENT) & mask]);
3318 			bus_dmamap_unload(txq->ift_desc_tag, ifsd_map[cidx]);
3319 			/* XXX we don't support any drivers that batch packets yet */
3320 			MPASS(m->m_nextpkt == NULL);
3321 			m_freem(m);
3322 			ifsd_m[cidx] = NULL;
3323 #if MEMORY_LOGGING
3324 			txq->ift_dequeued++;
3325 #endif
3326 			DBG_COUNTER_INC(tx_frees);
3327 		}
3328 		if (__predict_false(++cidx == qsize)) {
3329 			cidx = 0;
3330 			gen = 0;
3331 		}
3332 	}
3333 	txq->ift_cidx = cidx;
3334 	txq->ift_gen = gen;
3335 }
3336 
3337 static __inline int
3338 iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh)
3339 {
3340 	int reclaim;
3341 	if_ctx_t ctx = txq->ift_ctx;
3342 
3343 	KASSERT(thresh >= 0, ("invalid threshold to reclaim"));
3344 	MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size);
3345 
3346 	/*
3347 	 * Need a rate-limiting check so that this isn't called every time
3348 	 */
3349 	iflib_tx_credits_update(ctx, txq);
3350 	reclaim = DESC_RECLAIMABLE(txq);
3351 
3352 	if (reclaim <= thresh /* + MAX_TX_DESC(txq->ift_ctx) */) {
3353 #ifdef INVARIANTS
3354 		if (iflib_verbose_debug) {
3355 			printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __FUNCTION__,
3356 			       txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments,
3357 			       reclaim, thresh);
3358 
3359 		}
3360 #endif
3361 		return (0);
3362 	}
3363 	iflib_tx_desc_free(txq, reclaim);
3364 	txq->ift_cleaned += reclaim;
3365 	txq->ift_in_use -= reclaim;
3366 
3367 	return (reclaim);
3368 }
3369 
3370 static struct mbuf **
3371 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset, int remaining)
3372 {
3373 	int next, size;
3374 	struct mbuf **items;
3375 
3376 	size = r->size;
3377 	next = (cidx + CACHE_PTR_INCREMENT) & (size-1);
3378 	items = __DEVOLATILE(struct mbuf **, &r->items[0]);
3379 
3380 	prefetch(items[(cidx + offset) & (size-1)]);
3381 	if (remaining > 1) {
3382 		prefetch2cachelines(&items[next]);
3383 		prefetch2cachelines(items[(cidx + offset + 1) & (size-1)]);
3384 		prefetch2cachelines(items[(cidx + offset + 2) & (size-1)]);
3385 		prefetch2cachelines(items[(cidx + offset + 3) & (size-1)]);
3386 	}
3387 	return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (size-1)]));
3388 }
3389 
3390 static void
3391 iflib_txq_check_drain(iflib_txq_t txq, int budget)
3392 {
3393 
3394 	ifmp_ring_check_drainage(txq->ift_br, budget);
3395 }
3396 
3397 static uint32_t
3398 iflib_txq_can_drain(struct ifmp_ring *r)
3399 {
3400 	iflib_txq_t txq = r->cookie;
3401 	if_ctx_t ctx = txq->ift_ctx;
3402 
3403 	return ((TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2) ||
3404 		ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false));
3405 }
3406 
3407 static uint32_t
3408 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3409 {
3410 	iflib_txq_t txq = r->cookie;
3411 	if_ctx_t ctx = txq->ift_ctx;
3412 	struct ifnet *ifp = ctx->ifc_ifp;
3413 	struct mbuf **mp, *m;
3414 	int i, count, consumed, pkt_sent, bytes_sent, mcast_sent, avail;
3415 	int reclaimed, err, in_use_prev, desc_used;
3416 	bool do_prefetch, ring, rang;
3417 
3418 	if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) ||
3419 			    !LINK_ACTIVE(ctx))) {
3420 		DBG_COUNTER_INC(txq_drain_notready);
3421 		return (0);
3422 	}
3423 	reclaimed = iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
3424 	rang = iflib_txd_db_check(ctx, txq, reclaimed, txq->ift_in_use);
3425 	avail = IDXDIFF(pidx, cidx, r->size);
3426 	if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) {
3427 		DBG_COUNTER_INC(txq_drain_flushing);
3428 		for (i = 0; i < avail; i++) {
3429 			if (__predict_true(r->items[(cidx + i) & (r->size-1)] != (void *)txq))
3430 				m_free(r->items[(cidx + i) & (r->size-1)]);
3431 			r->items[(cidx + i) & (r->size-1)] = NULL;
3432 		}
3433 		return (avail);
3434 	}
3435 
3436 	if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3437 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3438 		CALLOUT_LOCK(txq);
3439 		callout_stop(&txq->ift_timer);
3440 		CALLOUT_UNLOCK(txq);
3441 		DBG_COUNTER_INC(txq_drain_oactive);
3442 		return (0);
3443 	}
3444 	if (reclaimed)
3445 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3446 	consumed = mcast_sent = bytes_sent = pkt_sent = 0;
3447 	count = MIN(avail, TX_BATCH_SIZE);
3448 #ifdef INVARIANTS
3449 	if (iflib_verbose_debug)
3450 		printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __FUNCTION__,
3451 		       avail, ctx->ifc_flags, TXQ_AVAIL(txq));
3452 #endif
3453 	do_prefetch = (ctx->ifc_flags & IFC_PREFETCH);
3454 	avail = TXQ_AVAIL(txq);
3455 	err = 0;
3456 	for (desc_used = i = 0; i < count && avail > MAX_TX_DESC(ctx) + 2; i++) {
3457 		int rem = do_prefetch ? count - i : 0;
3458 
3459 		mp = _ring_peek_one(r, cidx, i, rem);
3460 		MPASS(mp != NULL && *mp != NULL);
3461 		if (__predict_false(*mp == (struct mbuf *)txq)) {
3462 			consumed++;
3463 			reclaimed++;
3464 			continue;
3465 		}
3466 		in_use_prev = txq->ift_in_use;
3467 		err = iflib_encap(txq, mp);
3468 		if (__predict_false(err)) {
3469 			/* no room - bail out */
3470 			if (err == ENOBUFS)
3471 				break;
3472 			consumed++;
3473 			/* we can't send this packet - skip it */
3474 			continue;
3475 		}
3476 		consumed++;
3477 		pkt_sent++;
3478 		m = *mp;
3479 		DBG_COUNTER_INC(tx_sent);
3480 		bytes_sent += m->m_pkthdr.len;
3481 		mcast_sent += !!(m->m_flags & M_MCAST);
3482 		avail = TXQ_AVAIL(txq);
3483 
3484 		txq->ift_db_pending += (txq->ift_in_use - in_use_prev);
3485 		desc_used += (txq->ift_in_use - in_use_prev);
3486 		ETHER_BPF_MTAP(ifp, m);
3487 		if (__predict_false(!(ifp->if_drv_flags & IFF_DRV_RUNNING)))
3488 			break;
3489 		rang = iflib_txd_db_check(ctx, txq, false, in_use_prev);
3490 	}
3491 
3492 	/* deliberate use of bitwise or to avoid gratuitous short-circuit */
3493 	ring = rang ? false  : (iflib_min_tx_latency | err) || (TXQ_AVAIL(txq) < MAX_TX_DESC(ctx));
3494 	iflib_txd_db_check(ctx, txq, ring, txq->ift_in_use);
3495 	if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
3496 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
3497 	if (mcast_sent)
3498 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
3499 #ifdef INVARIANTS
3500 	if (iflib_verbose_debug)
3501 		printf("consumed=%d\n", consumed);
3502 #endif
3503 	return (consumed);
3504 }
3505 
3506 static uint32_t
3507 iflib_txq_drain_always(struct ifmp_ring *r)
3508 {
3509 	return (1);
3510 }
3511 
3512 static uint32_t
3513 iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3514 {
3515 	int i, avail;
3516 	struct mbuf **mp;
3517 	iflib_txq_t txq;
3518 
3519 	txq = r->cookie;
3520 
3521 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3522 	CALLOUT_LOCK(txq);
3523 	callout_stop(&txq->ift_timer);
3524 	CALLOUT_UNLOCK(txq);
3525 
3526 	avail = IDXDIFF(pidx, cidx, r->size);
3527 	for (i = 0; i < avail; i++) {
3528 		mp = _ring_peek_one(r, cidx, i, avail - i);
3529 		if (__predict_false(*mp == (struct mbuf *)txq))
3530 			continue;
3531 		m_freem(*mp);
3532 		DBG_COUNTER_INC(tx_frees);
3533 	}
3534 	MPASS(ifmp_ring_is_stalled(r) == 0);
3535 	return (avail);
3536 }
3537 
3538 static void
3539 iflib_ifmp_purge(iflib_txq_t txq)
3540 {
3541 	struct ifmp_ring *r;
3542 
3543 	r = txq->ift_br;
3544 	r->drain = iflib_txq_drain_free;
3545 	r->can_drain = iflib_txq_drain_always;
3546 
3547 	ifmp_ring_check_drainage(r, r->size);
3548 
3549 	r->drain = iflib_txq_drain;
3550 	r->can_drain = iflib_txq_can_drain;
3551 }
3552 
3553 static void
3554 _task_fn_tx(void *context)
3555 {
3556 	iflib_txq_t txq = context;
3557 	if_ctx_t ctx = txq->ift_ctx;
3558 	struct ifnet *ifp = ctx->ifc_ifp;
3559 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
3560 
3561 #ifdef IFLIB_DIAGNOSTICS
3562 	txq->ift_cpu_exec_count[curcpu]++;
3563 #endif
3564 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3565 		return;
3566 	if (if_getcapenable(ifp) & IFCAP_NETMAP) {
3567 		if (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false))
3568 			netmap_tx_irq(ifp, txq->ift_id);
3569 		IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3570 		return;
3571 	}
3572 #ifdef ALTQ
3573 	if (ALTQ_IS_ENABLED(&ifp->if_snd))
3574 		iflib_altq_if_start(ifp);
3575 #endif
3576 	if (txq->ift_db_pending)
3577 		ifmp_ring_enqueue(txq->ift_br, (void **)&txq, 1, TX_BATCH_SIZE, abdicate);
3578 	else if (!abdicate)
3579 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3580 	/*
3581 	 * When abdicating, we always need to check drainage, not just when we don't enqueue
3582 	 */
3583 	if (abdicate)
3584 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3585 	ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3586 	if (ctx->ifc_flags & IFC_LEGACY)
3587 		IFDI_INTR_ENABLE(ctx);
3588 	else {
3589 #ifdef INVARIANTS
3590 		int rc =
3591 #endif
3592 			IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3593 			KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver"));
3594 	}
3595 }
3596 
3597 static void
3598 _task_fn_rx(void *context)
3599 {
3600 	iflib_rxq_t rxq = context;
3601 	if_ctx_t ctx = rxq->ifr_ctx;
3602 	bool more;
3603 	uint16_t budget;
3604 
3605 #ifdef IFLIB_DIAGNOSTICS
3606 	rxq->ifr_cpu_exec_count[curcpu]++;
3607 #endif
3608 	DBG_COUNTER_INC(task_fn_rxs);
3609 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3610 		return;
3611 	more = true;
3612 #ifdef DEV_NETMAP
3613 	if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP) {
3614 		u_int work = 0;
3615 		if (netmap_rx_irq(ctx->ifc_ifp, rxq->ifr_id, &work)) {
3616 			more = false;
3617 		}
3618 	}
3619 #endif
3620 	budget = ctx->ifc_sysctl_rx_budget;
3621 	if (budget == 0)
3622 		budget = 16;	/* XXX */
3623 	if (more == false || (more = iflib_rxeof(rxq, budget)) == false) {
3624 		if (ctx->ifc_flags & IFC_LEGACY)
3625 			IFDI_INTR_ENABLE(ctx);
3626 		else {
3627 #ifdef INVARIANTS
3628 			int rc =
3629 #endif
3630 				IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
3631 			KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver"));
3632 			DBG_COUNTER_INC(rx_intr_enables);
3633 		}
3634 	}
3635 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3636 		return;
3637 	if (more)
3638 		GROUPTASK_ENQUEUE(&rxq->ifr_task);
3639 }
3640 
3641 static void
3642 _task_fn_admin(void *context)
3643 {
3644 	if_ctx_t ctx = context;
3645 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
3646 	iflib_txq_t txq;
3647 	int i;
3648 	bool oactive, running, do_reset, do_watchdog, in_detach;
3649 	uint32_t reset_on = hz / 2;
3650 
3651 	STATE_LOCK(ctx);
3652 	running = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING);
3653 	oactive = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE);
3654 	do_reset = (ctx->ifc_flags & IFC_DO_RESET);
3655 	do_watchdog = (ctx->ifc_flags & IFC_DO_WATCHDOG);
3656 	in_detach = (ctx->ifc_flags & IFC_IN_DETACH);
3657 	ctx->ifc_flags &= ~(IFC_DO_RESET|IFC_DO_WATCHDOG);
3658 	STATE_UNLOCK(ctx);
3659 
3660 	if ((!running && !oactive) && !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
3661 		return;
3662 	if (in_detach)
3663 		return;
3664 
3665 	CTX_LOCK(ctx);
3666 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3667 		CALLOUT_LOCK(txq);
3668 		callout_stop(&txq->ift_timer);
3669 		CALLOUT_UNLOCK(txq);
3670 	}
3671 	if (do_watchdog) {
3672 		ctx->ifc_watchdog_events++;
3673 		IFDI_WATCHDOG_RESET(ctx);
3674 	}
3675 	IFDI_UPDATE_ADMIN_STATUS(ctx);
3676 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3677 #ifdef DEV_NETMAP
3678 		reset_on = hz / 2;
3679 		if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP)
3680 			iflib_netmap_timer_adjust(ctx, txq->ift_id, &reset_on);
3681 #endif
3682 		callout_reset_on(&txq->ift_timer, reset_on, iflib_timer, txq, txq->ift_timer.c_cpu);
3683 	}
3684 	IFDI_LINK_INTR_ENABLE(ctx);
3685 	if (do_reset)
3686 		iflib_if_init_locked(ctx);
3687 	CTX_UNLOCK(ctx);
3688 
3689 	if (LINK_ACTIVE(ctx) == 0)
3690 		return;
3691 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3692 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
3693 }
3694 
3695 
3696 static void
3697 _task_fn_iov(void *context)
3698 {
3699 	if_ctx_t ctx = context;
3700 
3701 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING) &&
3702 	    !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
3703 		return;
3704 
3705 	CTX_LOCK(ctx);
3706 	IFDI_VFLR_HANDLE(ctx);
3707 	CTX_UNLOCK(ctx);
3708 }
3709 
3710 static int
3711 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)
3712 {
3713 	int err;
3714 	if_int_delay_info_t info;
3715 	if_ctx_t ctx;
3716 
3717 	info = (if_int_delay_info_t)arg1;
3718 	ctx = info->iidi_ctx;
3719 	info->iidi_req = req;
3720 	info->iidi_oidp = oidp;
3721 	CTX_LOCK(ctx);
3722 	err = IFDI_SYSCTL_INT_DELAY(ctx, info);
3723 	CTX_UNLOCK(ctx);
3724 	return (err);
3725 }
3726 
3727 /*********************************************************************
3728  *
3729  *  IFNET FUNCTIONS
3730  *
3731  **********************************************************************/
3732 
3733 static void
3734 iflib_if_init_locked(if_ctx_t ctx)
3735 {
3736 	iflib_stop(ctx);
3737 	iflib_init_locked(ctx);
3738 }
3739 
3740 
3741 static void
3742 iflib_if_init(void *arg)
3743 {
3744 	if_ctx_t ctx = arg;
3745 
3746 	CTX_LOCK(ctx);
3747 	iflib_if_init_locked(ctx);
3748 	CTX_UNLOCK(ctx);
3749 }
3750 
3751 static int
3752 iflib_if_transmit(if_t ifp, struct mbuf *m)
3753 {
3754 	if_ctx_t	ctx = if_getsoftc(ifp);
3755 
3756 	iflib_txq_t txq;
3757 	int err, qidx;
3758 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
3759 
3760 	if (__predict_false((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) {
3761 		DBG_COUNTER_INC(tx_frees);
3762 		m_freem(m);
3763 		return (ENOBUFS);
3764 	}
3765 
3766 	MPASS(m->m_nextpkt == NULL);
3767 	/* ALTQ-enabled interfaces always use queue 0. */
3768 	qidx = 0;
3769 	if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m) && !ALTQ_IS_ENABLED(&ifp->if_snd))
3770 		qidx = QIDX(ctx, m);
3771 	/*
3772 	 * XXX calculate buf_ring based on flowid (divvy up bits?)
3773 	 */
3774 	txq = &ctx->ifc_txqs[qidx];
3775 
3776 #ifdef DRIVER_BACKPRESSURE
3777 	if (txq->ift_closed) {
3778 		while (m != NULL) {
3779 			next = m->m_nextpkt;
3780 			m->m_nextpkt = NULL;
3781 			m_freem(m);
3782 			DBG_COUNTER_INC(tx_frees);
3783 			m = next;
3784 		}
3785 		return (ENOBUFS);
3786 	}
3787 #endif
3788 #ifdef notyet
3789 	qidx = count = 0;
3790 	mp = marr;
3791 	next = m;
3792 	do {
3793 		count++;
3794 		next = next->m_nextpkt;
3795 	} while (next != NULL);
3796 
3797 	if (count > nitems(marr))
3798 		if ((mp = malloc(count*sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) {
3799 			/* XXX check nextpkt */
3800 			m_freem(m);
3801 			/* XXX simplify for now */
3802 			DBG_COUNTER_INC(tx_frees);
3803 			return (ENOBUFS);
3804 		}
3805 	for (next = m, i = 0; next != NULL; i++) {
3806 		mp[i] = next;
3807 		next = next->m_nextpkt;
3808 		mp[i]->m_nextpkt = NULL;
3809 	}
3810 #endif
3811 	DBG_COUNTER_INC(tx_seen);
3812 	err = ifmp_ring_enqueue(txq->ift_br, (void **)&m, 1, TX_BATCH_SIZE, abdicate);
3813 
3814 	if (abdicate)
3815 		GROUPTASK_ENQUEUE(&txq->ift_task);
3816  	if (err) {
3817 		if (!abdicate)
3818 			GROUPTASK_ENQUEUE(&txq->ift_task);
3819 		/* support forthcoming later */
3820 #ifdef DRIVER_BACKPRESSURE
3821 		txq->ift_closed = TRUE;
3822 #endif
3823 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3824 		m_freem(m);
3825 		DBG_COUNTER_INC(tx_frees);
3826 	}
3827 
3828 	return (err);
3829 }
3830 
3831 #ifdef ALTQ
3832 /*
3833  * The overall approach to integrating iflib with ALTQ is to continue to use
3834  * the iflib mp_ring machinery between the ALTQ queue(s) and the hardware
3835  * ring.  Technically, when using ALTQ, queueing to an intermediate mp_ring
3836  * is redundant/unnecessary, but doing so minimizes the amount of
3837  * ALTQ-specific code required in iflib.  It is assumed that the overhead of
3838  * redundantly queueing to an intermediate mp_ring is swamped by the
3839  * performance limitations inherent in using ALTQ.
3840  *
3841  * When ALTQ support is compiled in, all iflib drivers will use a transmit
3842  * routine, iflib_altq_if_transmit(), that checks if ALTQ is enabled for the
3843  * given interface.  If ALTQ is enabled for an interface, then all
3844  * transmitted packets for that interface will be submitted to the ALTQ
3845  * subsystem via IFQ_ENQUEUE().  We don't use the legacy if_transmit()
3846  * implementation because it uses IFQ_HANDOFF(), which will duplicatively
3847  * update stats that the iflib machinery handles, and which is sensitve to
3848  * the disused IFF_DRV_OACTIVE flag.  Additionally, iflib_altq_if_start()
3849  * will be installed as the start routine for use by ALTQ facilities that
3850  * need to trigger queue drains on a scheduled basis.
3851  *
3852  */
3853 static void
3854 iflib_altq_if_start(if_t ifp)
3855 {
3856 	struct ifaltq *ifq = &ifp->if_snd;
3857 	struct mbuf *m;
3858 
3859 	IFQ_LOCK(ifq);
3860 	IFQ_DEQUEUE_NOLOCK(ifq, m);
3861 	while (m != NULL) {
3862 		iflib_if_transmit(ifp, m);
3863 		IFQ_DEQUEUE_NOLOCK(ifq, m);
3864 	}
3865 	IFQ_UNLOCK(ifq);
3866 }
3867 
3868 static int
3869 iflib_altq_if_transmit(if_t ifp, struct mbuf *m)
3870 {
3871 	int err;
3872 
3873 	if (ALTQ_IS_ENABLED(&ifp->if_snd)) {
3874 		IFQ_ENQUEUE(&ifp->if_snd, m, err);
3875 		if (err == 0)
3876 			iflib_altq_if_start(ifp);
3877 	} else
3878 		err = iflib_if_transmit(ifp, m);
3879 
3880 	return (err);
3881 }
3882 #endif /* ALTQ */
3883 
3884 static void
3885 iflib_if_qflush(if_t ifp)
3886 {
3887 	if_ctx_t ctx = if_getsoftc(ifp);
3888 	iflib_txq_t txq = ctx->ifc_txqs;
3889 	int i;
3890 
3891 	STATE_LOCK(ctx);
3892 	ctx->ifc_flags |= IFC_QFLUSH;
3893 	STATE_UNLOCK(ctx);
3894 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
3895 		while (!(ifmp_ring_is_idle(txq->ift_br) || ifmp_ring_is_stalled(txq->ift_br)))
3896 			iflib_txq_check_drain(txq, 0);
3897 	STATE_LOCK(ctx);
3898 	ctx->ifc_flags &= ~IFC_QFLUSH;
3899 	STATE_UNLOCK(ctx);
3900 
3901 	/*
3902 	 * When ALTQ is enabled, this will also take care of purging the
3903 	 * ALTQ queue(s).
3904 	 */
3905 	if_qflush(ifp);
3906 }
3907 
3908 
3909 #define IFCAP_FLAGS (IFCAP_HWCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \
3910 		     IFCAP_TSO | IFCAP_VLAN_HWTAGGING | IFCAP_HWSTATS | \
3911 		     IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | \
3912 		     IFCAP_VLAN_HWTSO | IFCAP_VLAN_HWCSUM)
3913 
3914 static int
3915 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data)
3916 {
3917 	if_ctx_t ctx = if_getsoftc(ifp);
3918 	struct ifreq	*ifr = (struct ifreq *)data;
3919 #if defined(INET) || defined(INET6)
3920 	struct ifaddr	*ifa = (struct ifaddr *)data;
3921 #endif
3922 	bool		avoid_reset = FALSE;
3923 	int		err = 0, reinit = 0, bits;
3924 
3925 	switch (command) {
3926 	case SIOCSIFADDR:
3927 #ifdef INET
3928 		if (ifa->ifa_addr->sa_family == AF_INET)
3929 			avoid_reset = TRUE;
3930 #endif
3931 #ifdef INET6
3932 		if (ifa->ifa_addr->sa_family == AF_INET6)
3933 			avoid_reset = TRUE;
3934 #endif
3935 		/*
3936 		** Calling init results in link renegotiation,
3937 		** so we avoid doing it when possible.
3938 		*/
3939 		if (avoid_reset) {
3940 			if_setflagbits(ifp, IFF_UP,0);
3941 			if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
3942 				reinit = 1;
3943 #ifdef INET
3944 			if (!(if_getflags(ifp) & IFF_NOARP))
3945 				arp_ifinit(ifp, ifa);
3946 #endif
3947 		} else
3948 			err = ether_ioctl(ifp, command, data);
3949 		break;
3950 	case SIOCSIFMTU:
3951 		CTX_LOCK(ctx);
3952 		if (ifr->ifr_mtu == if_getmtu(ifp)) {
3953 			CTX_UNLOCK(ctx);
3954 			break;
3955 		}
3956 		bits = if_getdrvflags(ifp);
3957 		/* stop the driver and free any clusters before proceeding */
3958 		iflib_stop(ctx);
3959 
3960 		if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) {
3961 			STATE_LOCK(ctx);
3962 			if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size)
3963 				ctx->ifc_flags |= IFC_MULTISEG;
3964 			else
3965 				ctx->ifc_flags &= ~IFC_MULTISEG;
3966 			STATE_UNLOCK(ctx);
3967 			err = if_setmtu(ifp, ifr->ifr_mtu);
3968 		}
3969 		iflib_init_locked(ctx);
3970 		STATE_LOCK(ctx);
3971 		if_setdrvflags(ifp, bits);
3972 		STATE_UNLOCK(ctx);
3973 		CTX_UNLOCK(ctx);
3974 		break;
3975 	case SIOCSIFFLAGS:
3976 		CTX_LOCK(ctx);
3977 		if (if_getflags(ifp) & IFF_UP) {
3978 			if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
3979 				if ((if_getflags(ifp) ^ ctx->ifc_if_flags) &
3980 				    (IFF_PROMISC | IFF_ALLMULTI)) {
3981 					err = IFDI_PROMISC_SET(ctx, if_getflags(ifp));
3982 				}
3983 			} else
3984 				reinit = 1;
3985 		} else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
3986 			iflib_stop(ctx);
3987 		}
3988 		ctx->ifc_if_flags = if_getflags(ifp);
3989 		CTX_UNLOCK(ctx);
3990 		break;
3991 	case SIOCADDMULTI:
3992 	case SIOCDELMULTI:
3993 		if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
3994 			CTX_LOCK(ctx);
3995 			IFDI_INTR_DISABLE(ctx);
3996 			IFDI_MULTI_SET(ctx);
3997 			IFDI_INTR_ENABLE(ctx);
3998 			CTX_UNLOCK(ctx);
3999 		}
4000 		break;
4001 	case SIOCSIFMEDIA:
4002 		CTX_LOCK(ctx);
4003 		IFDI_MEDIA_SET(ctx);
4004 		CTX_UNLOCK(ctx);
4005 		/* falls thru */
4006 	case SIOCGIFMEDIA:
4007 	case SIOCGIFXMEDIA:
4008 		err = ifmedia_ioctl(ifp, ifr, &ctx->ifc_media, command);
4009 		break;
4010 	case SIOCGI2C:
4011 	{
4012 		struct ifi2creq i2c;
4013 
4014 		err = copyin(ifr_data_get_ptr(ifr), &i2c, sizeof(i2c));
4015 		if (err != 0)
4016 			break;
4017 		if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) {
4018 			err = EINVAL;
4019 			break;
4020 		}
4021 		if (i2c.len > sizeof(i2c.data)) {
4022 			err = EINVAL;
4023 			break;
4024 		}
4025 
4026 		if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0)
4027 			err = copyout(&i2c, ifr_data_get_ptr(ifr),
4028 			    sizeof(i2c));
4029 		break;
4030 	}
4031 	case SIOCSIFCAP:
4032 	{
4033 		int mask, setmask, oldmask;
4034 
4035 		oldmask = if_getcapenable(ifp);
4036 		mask = ifr->ifr_reqcap ^ oldmask;
4037 		mask &= ctx->ifc_softc_ctx.isc_capabilities;
4038 		setmask = 0;
4039 #ifdef TCP_OFFLOAD
4040 		setmask |= mask & (IFCAP_TOE4|IFCAP_TOE6);
4041 #endif
4042 		setmask |= (mask & IFCAP_FLAGS);
4043 		setmask |= (mask & IFCAP_WOL);
4044 
4045 		/*
4046 		 * If any RX csum has changed, change all the ones that
4047 		 * are supported by the driver.
4048 		 */
4049 		if (setmask & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6)) {
4050 			setmask |= ctx->ifc_softc_ctx.isc_capabilities &
4051 			    (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6);
4052 		}
4053 
4054 		/*
4055 		 * want to ensure that traffic has stopped before we change any of the flags
4056 		 */
4057 		if (setmask) {
4058 			CTX_LOCK(ctx);
4059 			bits = if_getdrvflags(ifp);
4060 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4061 				iflib_stop(ctx);
4062 			STATE_LOCK(ctx);
4063 			if_togglecapenable(ifp, setmask);
4064 			STATE_UNLOCK(ctx);
4065 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4066 				iflib_init_locked(ctx);
4067 			STATE_LOCK(ctx);
4068 			if_setdrvflags(ifp, bits);
4069 			STATE_UNLOCK(ctx);
4070 			CTX_UNLOCK(ctx);
4071 		}
4072 		if_vlancap(ifp);
4073 		break;
4074 	}
4075 	case SIOCGPRIVATE_0:
4076 	case SIOCSDRVSPEC:
4077 	case SIOCGDRVSPEC:
4078 		CTX_LOCK(ctx);
4079 		err = IFDI_PRIV_IOCTL(ctx, command, data);
4080 		CTX_UNLOCK(ctx);
4081 		break;
4082 	default:
4083 		err = ether_ioctl(ifp, command, data);
4084 		break;
4085 	}
4086 	if (reinit)
4087 		iflib_if_init(ctx);
4088 	return (err);
4089 }
4090 
4091 static uint64_t
4092 iflib_if_get_counter(if_t ifp, ift_counter cnt)
4093 {
4094 	if_ctx_t ctx = if_getsoftc(ifp);
4095 
4096 	return (IFDI_GET_COUNTER(ctx, cnt));
4097 }
4098 
4099 /*********************************************************************
4100  *
4101  *  OTHER FUNCTIONS EXPORTED TO THE STACK
4102  *
4103  **********************************************************************/
4104 
4105 static void
4106 iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag)
4107 {
4108 	if_ctx_t ctx = if_getsoftc(ifp);
4109 
4110 	if ((void *)ctx != arg)
4111 		return;
4112 
4113 	if ((vtag == 0) || (vtag > 4095))
4114 		return;
4115 
4116 	CTX_LOCK(ctx);
4117 	IFDI_VLAN_REGISTER(ctx, vtag);
4118 	/* Re-init to load the changes */
4119 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
4120 		iflib_if_init_locked(ctx);
4121 	CTX_UNLOCK(ctx);
4122 }
4123 
4124 static void
4125 iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag)
4126 {
4127 	if_ctx_t ctx = if_getsoftc(ifp);
4128 
4129 	if ((void *)ctx != arg)
4130 		return;
4131 
4132 	if ((vtag == 0) || (vtag > 4095))
4133 		return;
4134 
4135 	CTX_LOCK(ctx);
4136 	IFDI_VLAN_UNREGISTER(ctx, vtag);
4137 	/* Re-init to load the changes */
4138 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
4139 		iflib_if_init_locked(ctx);
4140 	CTX_UNLOCK(ctx);
4141 }
4142 
4143 static void
4144 iflib_led_func(void *arg, int onoff)
4145 {
4146 	if_ctx_t ctx = arg;
4147 
4148 	CTX_LOCK(ctx);
4149 	IFDI_LED_FUNC(ctx, onoff);
4150 	CTX_UNLOCK(ctx);
4151 }
4152 
4153 /*********************************************************************
4154  *
4155  *  BUS FUNCTION DEFINITIONS
4156  *
4157  **********************************************************************/
4158 
4159 int
4160 iflib_device_probe(device_t dev)
4161 {
4162 	pci_vendor_info_t *ent;
4163 
4164 	uint16_t	pci_vendor_id, pci_device_id;
4165 	uint16_t	pci_subvendor_id, pci_subdevice_id;
4166 	uint16_t	pci_rev_id;
4167 	if_shared_ctx_t sctx;
4168 
4169 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4170 		return (ENOTSUP);
4171 
4172 	pci_vendor_id = pci_get_vendor(dev);
4173 	pci_device_id = pci_get_device(dev);
4174 	pci_subvendor_id = pci_get_subvendor(dev);
4175 	pci_subdevice_id = pci_get_subdevice(dev);
4176 	pci_rev_id = pci_get_revid(dev);
4177 	if (sctx->isc_parse_devinfo != NULL)
4178 		sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id);
4179 
4180 	ent = sctx->isc_vendor_info;
4181 	while (ent->pvi_vendor_id != 0) {
4182 		if (pci_vendor_id != ent->pvi_vendor_id) {
4183 			ent++;
4184 			continue;
4185 		}
4186 		if ((pci_device_id == ent->pvi_device_id) &&
4187 		    ((pci_subvendor_id == ent->pvi_subvendor_id) ||
4188 		     (ent->pvi_subvendor_id == 0)) &&
4189 		    ((pci_subdevice_id == ent->pvi_subdevice_id) ||
4190 		     (ent->pvi_subdevice_id == 0)) &&
4191 		    ((pci_rev_id == ent->pvi_rev_id) ||
4192 		     (ent->pvi_rev_id == 0))) {
4193 
4194 			device_set_desc_copy(dev, ent->pvi_name);
4195 			/* this needs to be changed to zero if the bus probing code
4196 			 * ever stops re-probing on best match because the sctx
4197 			 * may have its values over written by register calls
4198 			 * in subsequent probes
4199 			 */
4200 			return (BUS_PROBE_DEFAULT);
4201 		}
4202 		ent++;
4203 	}
4204 	return (ENXIO);
4205 }
4206 
4207 static void
4208 iflib_reset_qvalues(if_ctx_t ctx)
4209 {
4210 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4211 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4212 	device_t dev = ctx->ifc_dev;
4213 	int i;
4214 
4215 	scctx->isc_txrx_budget_bytes_max = IFLIB_MAX_TX_BYTES;
4216 	scctx->isc_tx_qdepth = IFLIB_DEFAULT_TX_QDEPTH;
4217 	/*
4218 	 * XXX sanity check that ntxd & nrxd are a power of 2
4219 	 */
4220 	if (ctx->ifc_sysctl_ntxqs != 0)
4221 		scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs;
4222 	if (ctx->ifc_sysctl_nrxqs != 0)
4223 		scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs;
4224 
4225 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4226 		if (ctx->ifc_sysctl_ntxds[i] != 0)
4227 			scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i];
4228 		else
4229 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4230 	}
4231 
4232 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4233 		if (ctx->ifc_sysctl_nrxds[i] != 0)
4234 			scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i];
4235 		else
4236 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4237 	}
4238 
4239 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4240 		if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) {
4241 			device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n",
4242 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]);
4243 			scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i];
4244 		}
4245 		if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) {
4246 			device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n",
4247 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]);
4248 			scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i];
4249 		}
4250 	}
4251 
4252 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4253 		if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) {
4254 			device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n",
4255 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]);
4256 			scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i];
4257 		}
4258 		if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) {
4259 			device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n",
4260 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]);
4261 			scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i];
4262 		}
4263 	}
4264 }
4265 
4266 int
4267 iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp)
4268 {
4269 	int err, rid, msix;
4270 	if_ctx_t ctx;
4271 	if_t ifp;
4272 	if_softc_ctx_t scctx;
4273 	int i;
4274 	uint16_t main_txq;
4275 	uint16_t main_rxq;
4276 
4277 
4278 	ctx = malloc(sizeof(* ctx), M_IFLIB, M_WAITOK|M_ZERO);
4279 
4280 	if (sc == NULL) {
4281 		sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4282 		device_set_softc(dev, ctx);
4283 		ctx->ifc_flags |= IFC_SC_ALLOCATED;
4284 	}
4285 
4286 	ctx->ifc_sctx = sctx;
4287 	ctx->ifc_dev = dev;
4288 	ctx->ifc_softc = sc;
4289 
4290 	if ((err = iflib_register(ctx)) != 0) {
4291 		if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4292 			free(sc, M_IFLIB);
4293 		free(ctx, M_IFLIB);
4294 		device_printf(dev, "iflib_register failed %d\n", err);
4295 		return (err);
4296 	}
4297 	iflib_add_device_sysctl_pre(ctx);
4298 
4299 	scctx = &ctx->ifc_softc_ctx;
4300 	ifp = ctx->ifc_ifp;
4301 
4302 	iflib_reset_qvalues(ctx);
4303 	CTX_LOCK(ctx);
4304 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4305 		CTX_UNLOCK(ctx);
4306 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4307 		return (err);
4308 	}
4309 	_iflib_pre_assert(scctx);
4310 	ctx->ifc_txrx = *scctx->isc_txrx;
4311 
4312 #ifdef INVARIANTS
4313 	MPASS(scctx->isc_capabilities);
4314 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
4315 		MPASS(scctx->isc_tx_csum_flags);
4316 #endif
4317 
4318 	if_setcapabilities(ifp, scctx->isc_capabilities | IFCAP_HWSTATS);
4319 	if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS);
4320 
4321 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4322 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4323 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4324 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4325 
4326 	main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4327 	main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4328 
4329 	/* XXX change for per-queue sizes */
4330 	device_printf(dev, "using %d tx descriptors and %d rx descriptors\n",
4331 		      scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4332 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4333 		if (!powerof2(scctx->isc_nrxd[i])) {
4334 			/* round down instead? */
4335 			device_printf(dev, "# rx descriptors must be a power of 2\n");
4336 			err = EINVAL;
4337 			goto fail;
4338 		}
4339 	}
4340 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4341 		if (!powerof2(scctx->isc_ntxd[i])) {
4342 			device_printf(dev,
4343 			    "# tx descriptors must be a power of 2");
4344 			err = EINVAL;
4345 			goto fail;
4346 		}
4347 	}
4348 
4349 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4350 	    MAX_SINGLE_PACKET_FRACTION)
4351 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4352 		    MAX_SINGLE_PACKET_FRACTION);
4353 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4354 	    MAX_SINGLE_PACKET_FRACTION)
4355 		scctx->isc_tx_tso_segments_max = max(1,
4356 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4357 
4358 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4359 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
4360 		/*
4361 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
4362 		 * but some MACs do.
4363 		 */
4364 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
4365 		    IP_MAXPACKET));
4366 		/*
4367 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
4368 		 * into account.  In the worst case, each of these calls will
4369 		 * add another mbuf and, thus, the requirement for another DMA
4370 		 * segment.  So for best performance, it doesn't make sense to
4371 		 * advertize a maximum of TSO segments that typically will
4372 		 * require defragmentation in iflib_encap().
4373 		 */
4374 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
4375 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
4376 	}
4377 	if (scctx->isc_rss_table_size == 0)
4378 		scctx->isc_rss_table_size = 64;
4379 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4380 
4381 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4382 	/* XXX format name */
4383 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx, -1, "admin");
4384 
4385 	/* Set up cpu set.  If it fails, use the set of all CPUs. */
4386 	if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) != 0) {
4387 		device_printf(dev, "Unable to fetch CPU list\n");
4388 		CPU_COPY(&all_cpus, &ctx->ifc_cpus);
4389 	}
4390 	MPASS(CPU_COUNT(&ctx->ifc_cpus) > 0);
4391 
4392 	/*
4393 	** Now setup MSI or MSI/X, should
4394 	** return us the number of supported
4395 	** vectors. (Will be 1 for MSI)
4396 	*/
4397 	if (sctx->isc_flags & IFLIB_SKIP_MSIX) {
4398 		msix = scctx->isc_vectors;
4399 	} else if (scctx->isc_msix_bar != 0)
4400 	       /*
4401 		* The simple fact that isc_msix_bar is not 0 does not mean we
4402 		* we have a good value there that is known to work.
4403 		*/
4404 		msix = iflib_msix_init(ctx);
4405 	else {
4406 		scctx->isc_vectors = 1;
4407 		scctx->isc_ntxqsets = 1;
4408 		scctx->isc_nrxqsets = 1;
4409 		scctx->isc_intr = IFLIB_INTR_LEGACY;
4410 		msix = 0;
4411 	}
4412 	/* Get memory for the station queues */
4413 	if ((err = iflib_queues_alloc(ctx))) {
4414 		device_printf(dev, "Unable to allocate queue memory\n");
4415 		goto fail;
4416 	}
4417 
4418 	if ((err = iflib_qset_structures_setup(ctx)))
4419 		goto fail_queues;
4420 
4421 	/*
4422 	 * Group taskqueues aren't properly set up until SMP is started,
4423 	 * so we disable interrupts until we can handle them post
4424 	 * SI_SUB_SMP.
4425 	 *
4426 	 * XXX: disabling interrupts doesn't actually work, at least for
4427 	 * the non-MSI case.  When they occur before SI_SUB_SMP completes,
4428 	 * we do null handling and depend on this not causing too large an
4429 	 * interrupt storm.
4430 	 */
4431 	IFDI_INTR_DISABLE(ctx);
4432 	if (msix > 1 && (err = IFDI_MSIX_INTR_ASSIGN(ctx, msix)) != 0) {
4433 		device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n", err);
4434 		goto fail_intr_free;
4435 	}
4436 	if (msix <= 1) {
4437 		rid = 0;
4438 		if (scctx->isc_intr == IFLIB_INTR_MSI) {
4439 			MPASS(msix == 1);
4440 			rid = 1;
4441 		}
4442 		if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) {
4443 			device_printf(dev, "iflib_legacy_setup failed %d\n", err);
4444 			goto fail_intr_free;
4445 		}
4446 	}
4447 
4448 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac);
4449 
4450 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4451 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4452 		goto fail_detach;
4453 	}
4454 
4455 	/*
4456 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4457 	 * This must appear after the call to ether_ifattach() because
4458 	 * ether_ifattach() sets if_hdrlen to the default value.
4459 	 */
4460 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4461 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
4462 
4463 	if ((err = iflib_netmap_attach(ctx))) {
4464 		device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err);
4465 		goto fail_detach;
4466 	}
4467 	*ctxp = ctx;
4468 
4469 	NETDUMP_SET(ctx->ifc_ifp, iflib);
4470 
4471 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4472 	iflib_add_device_sysctl_post(ctx);
4473 	ctx->ifc_flags |= IFC_INIT_DONE;
4474 	CTX_UNLOCK(ctx);
4475 	return (0);
4476 
4477 fail_detach:
4478 	ether_ifdetach(ctx->ifc_ifp);
4479 fail_intr_free:
4480 fail_queues:
4481 	iflib_tx_structures_free(ctx);
4482 	iflib_rx_structures_free(ctx);
4483 fail:
4484 	iflib_free_intr_mem(ctx);
4485 	IFDI_DETACH(ctx);
4486 	CTX_UNLOCK(ctx);
4487 
4488 	return (err);
4489 }
4490 
4491 int
4492 iflib_pseudo_register(device_t dev, if_shared_ctx_t sctx, if_ctx_t *ctxp,
4493 					  struct iflib_cloneattach_ctx *clctx)
4494 {
4495 	int err;
4496 	if_ctx_t ctx;
4497 	if_t ifp;
4498 	if_softc_ctx_t scctx;
4499 	int i;
4500 	void *sc;
4501 	uint16_t main_txq;
4502 	uint16_t main_rxq;
4503 
4504 	ctx = malloc(sizeof(*ctx), M_IFLIB, M_WAITOK|M_ZERO);
4505 	sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4506 	ctx->ifc_flags |= IFC_SC_ALLOCATED;
4507 	if (sctx->isc_flags & (IFLIB_PSEUDO|IFLIB_VIRTUAL))
4508 		ctx->ifc_flags |= IFC_PSEUDO;
4509 
4510 	ctx->ifc_sctx = sctx;
4511 	ctx->ifc_softc = sc;
4512 	ctx->ifc_dev = dev;
4513 
4514 	if ((err = iflib_register(ctx)) != 0) {
4515 		device_printf(dev, "%s: iflib_register failed %d\n", __func__, err);
4516 		free(sc, M_IFLIB);
4517 		free(ctx, M_IFLIB);
4518 		return (err);
4519 	}
4520 	iflib_add_device_sysctl_pre(ctx);
4521 
4522 	scctx = &ctx->ifc_softc_ctx;
4523 	ifp = ctx->ifc_ifp;
4524 
4525 	/*
4526 	 * XXX sanity check that ntxd & nrxd are a power of 2
4527 	 */
4528 	iflib_reset_qvalues(ctx);
4529 
4530 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4531 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4532 		return (err);
4533 	}
4534 	if (sctx->isc_flags & IFLIB_GEN_MAC)
4535 		iflib_gen_mac(ctx);
4536 	if ((err = IFDI_CLONEATTACH(ctx, clctx->cc_ifc, clctx->cc_name,
4537 								clctx->cc_params)) != 0) {
4538 		device_printf(dev, "IFDI_CLONEATTACH failed %d\n", err);
4539 		return (err);
4540 	}
4541 	ifmedia_add(&ctx->ifc_media, IFM_ETHER | IFM_1000_T | IFM_FDX, 0, NULL);
4542 	ifmedia_add(&ctx->ifc_media, IFM_ETHER | IFM_AUTO, 0, NULL);
4543 	ifmedia_set(&ctx->ifc_media, IFM_ETHER | IFM_AUTO);
4544 
4545 #ifdef INVARIANTS
4546 	MPASS(scctx->isc_capabilities);
4547 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
4548 		MPASS(scctx->isc_tx_csum_flags);
4549 #endif
4550 
4551 	if_setcapabilities(ifp, scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_LINKSTATE);
4552 	if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_LINKSTATE);
4553 
4554 	ifp->if_flags |= IFF_NOGROUP;
4555 	if (sctx->isc_flags & IFLIB_PSEUDO) {
4556 		ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac);
4557 
4558 		if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4559 			device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4560 			goto fail_detach;
4561 		}
4562 		*ctxp = ctx;
4563 
4564 		/*
4565 		 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4566 		 * This must appear after the call to ether_ifattach() because
4567 		 * ether_ifattach() sets if_hdrlen to the default value.
4568 		 */
4569 		if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4570 			if_setifheaderlen(ifp,
4571 			    sizeof(struct ether_vlan_header));
4572 
4573 		if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4574 		iflib_add_device_sysctl_post(ctx);
4575 		ctx->ifc_flags |= IFC_INIT_DONE;
4576 		return (0);
4577 	}
4578 	_iflib_pre_assert(scctx);
4579 	ctx->ifc_txrx = *scctx->isc_txrx;
4580 
4581 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4582 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4583 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4584 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4585 
4586 	main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4587 	main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4588 
4589 	/* XXX change for per-queue sizes */
4590 	device_printf(dev, "using %d tx descriptors and %d rx descriptors\n",
4591 		      scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4592 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4593 		if (!powerof2(scctx->isc_nrxd[i])) {
4594 			/* round down instead? */
4595 			device_printf(dev, "# rx descriptors must be a power of 2\n");
4596 			err = EINVAL;
4597 			goto fail;
4598 		}
4599 	}
4600 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4601 		if (!powerof2(scctx->isc_ntxd[i])) {
4602 			device_printf(dev,
4603 			    "# tx descriptors must be a power of 2");
4604 			err = EINVAL;
4605 			goto fail;
4606 		}
4607 	}
4608 
4609 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4610 	    MAX_SINGLE_PACKET_FRACTION)
4611 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4612 		    MAX_SINGLE_PACKET_FRACTION);
4613 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4614 	    MAX_SINGLE_PACKET_FRACTION)
4615 		scctx->isc_tx_tso_segments_max = max(1,
4616 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4617 
4618 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4619 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
4620 		/*
4621 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
4622 		 * but some MACs do.
4623 		 */
4624 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
4625 		    IP_MAXPACKET));
4626 		/*
4627 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
4628 		 * into account.  In the worst case, each of these calls will
4629 		 * add another mbuf and, thus, the requirement for another DMA
4630 		 * segment.  So for best performance, it doesn't make sense to
4631 		 * advertize a maximum of TSO segments that typically will
4632 		 * require defragmentation in iflib_encap().
4633 		 */
4634 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
4635 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
4636 	}
4637 	if (scctx->isc_rss_table_size == 0)
4638 		scctx->isc_rss_table_size = 64;
4639 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4640 
4641 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4642 	/* XXX format name */
4643 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx, -1, "admin");
4644 
4645 	/* XXX --- can support > 1 -- but keep it simple for now */
4646 	scctx->isc_intr = IFLIB_INTR_LEGACY;
4647 
4648 	/* Get memory for the station queues */
4649 	if ((err = iflib_queues_alloc(ctx))) {
4650 		device_printf(dev, "Unable to allocate queue memory\n");
4651 		goto fail;
4652 	}
4653 
4654 	if ((err = iflib_qset_structures_setup(ctx))) {
4655 		device_printf(dev, "qset structure setup failed %d\n", err);
4656 		goto fail_queues;
4657 	}
4658 
4659 	/*
4660 	 * XXX What if anything do we want to do about interrupts?
4661 	 */
4662 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac);
4663 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4664 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4665 		goto fail_detach;
4666 	}
4667 
4668 	/*
4669 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4670 	 * This must appear after the call to ether_ifattach() because
4671 	 * ether_ifattach() sets if_hdrlen to the default value.
4672 	 */
4673 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4674 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
4675 
4676 	/* XXX handle more than one queue */
4677 	for (i = 0; i < scctx->isc_nrxqsets; i++)
4678 		IFDI_RX_CLSET(ctx, 0, i, ctx->ifc_rxqs[i].ifr_fl[0].ifl_sds.ifsd_cl);
4679 
4680 	*ctxp = ctx;
4681 
4682 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4683 	iflib_add_device_sysctl_post(ctx);
4684 	ctx->ifc_flags |= IFC_INIT_DONE;
4685 	return (0);
4686 fail_detach:
4687 	ether_ifdetach(ctx->ifc_ifp);
4688 fail_queues:
4689 	iflib_tx_structures_free(ctx);
4690 	iflib_rx_structures_free(ctx);
4691 fail:
4692 	IFDI_DETACH(ctx);
4693 	return (err);
4694 }
4695 
4696 int
4697 iflib_pseudo_deregister(if_ctx_t ctx)
4698 {
4699 	if_t ifp = ctx->ifc_ifp;
4700 	iflib_txq_t txq;
4701 	iflib_rxq_t rxq;
4702 	int i, j;
4703 	struct taskqgroup *tqg;
4704 	iflib_fl_t fl;
4705 
4706 	/* Unregister VLAN events */
4707 	if (ctx->ifc_vlan_attach_event != NULL)
4708 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
4709 	if (ctx->ifc_vlan_detach_event != NULL)
4710 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
4711 
4712 	ether_ifdetach(ifp);
4713 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
4714 	CTX_LOCK_DESTROY(ctx);
4715 	/* XXX drain any dependent tasks */
4716 	tqg = qgroup_if_io_tqg;
4717 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
4718 		callout_drain(&txq->ift_timer);
4719 		if (txq->ift_task.gt_uniq != NULL)
4720 			taskqgroup_detach(tqg, &txq->ift_task);
4721 	}
4722 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4723 		if (rxq->ifr_task.gt_uniq != NULL)
4724 			taskqgroup_detach(tqg, &rxq->ifr_task);
4725 
4726 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
4727 			free(fl->ifl_rx_bitmap, M_IFLIB);
4728 	}
4729 	tqg = qgroup_if_config_tqg;
4730 	if (ctx->ifc_admin_task.gt_uniq != NULL)
4731 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
4732 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
4733 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
4734 
4735 	if_free(ifp);
4736 
4737 	iflib_tx_structures_free(ctx);
4738 	iflib_rx_structures_free(ctx);
4739 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4740 		free(ctx->ifc_softc, M_IFLIB);
4741 	free(ctx, M_IFLIB);
4742 	return (0);
4743 }
4744 
4745 int
4746 iflib_device_attach(device_t dev)
4747 {
4748 	if_ctx_t ctx;
4749 	if_shared_ctx_t sctx;
4750 
4751 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4752 		return (ENOTSUP);
4753 
4754 	pci_enable_busmaster(dev);
4755 
4756 	return (iflib_device_register(dev, NULL, sctx, &ctx));
4757 }
4758 
4759 int
4760 iflib_device_deregister(if_ctx_t ctx)
4761 {
4762 	if_t ifp = ctx->ifc_ifp;
4763 	iflib_txq_t txq;
4764 	iflib_rxq_t rxq;
4765 	device_t dev = ctx->ifc_dev;
4766 	int i, j;
4767 	struct taskqgroup *tqg;
4768 	iflib_fl_t fl;
4769 
4770 	/* Make sure VLANS are not using driver */
4771 	if (if_vlantrunkinuse(ifp)) {
4772 		device_printf(dev, "Vlan in use, detach first\n");
4773 		return (EBUSY);
4774 	}
4775 #ifdef PCI_IOV
4776 	if (!CTX_IS_VF(ctx) && pci_iov_detach(dev) != 0) {
4777 		device_printf(dev, "SR-IOV in use; detach first.\n");
4778 		return (EBUSY);
4779 	}
4780 #endif
4781 
4782 	STATE_LOCK(ctx);
4783 	ctx->ifc_flags |= IFC_IN_DETACH;
4784 	STATE_UNLOCK(ctx);
4785 
4786 	CTX_LOCK(ctx);
4787 	iflib_stop(ctx);
4788 	CTX_UNLOCK(ctx);
4789 
4790 	/* Unregister VLAN events */
4791 	if (ctx->ifc_vlan_attach_event != NULL)
4792 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
4793 	if (ctx->ifc_vlan_detach_event != NULL)
4794 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
4795 
4796 	iflib_netmap_detach(ifp);
4797 	ether_ifdetach(ifp);
4798 	if (ctx->ifc_led_dev != NULL)
4799 		led_destroy(ctx->ifc_led_dev);
4800 	/* XXX drain any dependent tasks */
4801 	tqg = qgroup_if_io_tqg;
4802 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
4803 		callout_drain(&txq->ift_timer);
4804 		if (txq->ift_task.gt_uniq != NULL)
4805 			taskqgroup_detach(tqg, &txq->ift_task);
4806 	}
4807 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4808 		if (rxq->ifr_task.gt_uniq != NULL)
4809 			taskqgroup_detach(tqg, &rxq->ifr_task);
4810 
4811 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
4812 			free(fl->ifl_rx_bitmap, M_IFLIB);
4813 
4814 	}
4815 	tqg = qgroup_if_config_tqg;
4816 	if (ctx->ifc_admin_task.gt_uniq != NULL)
4817 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
4818 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
4819 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
4820 	CTX_LOCK(ctx);
4821 	IFDI_DETACH(ctx);
4822 	CTX_UNLOCK(ctx);
4823 
4824 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
4825 	CTX_LOCK_DESTROY(ctx);
4826 	device_set_softc(ctx->ifc_dev, NULL);
4827 	iflib_free_intr_mem(ctx);
4828 
4829 	bus_generic_detach(dev);
4830 	if_free(ifp);
4831 
4832 	iflib_tx_structures_free(ctx);
4833 	iflib_rx_structures_free(ctx);
4834 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4835 		free(ctx->ifc_softc, M_IFLIB);
4836 	STATE_LOCK_DESTROY(ctx);
4837 	free(ctx, M_IFLIB);
4838 	return (0);
4839 }
4840 
4841 static void
4842 iflib_free_intr_mem(if_ctx_t ctx)
4843 {
4844 
4845 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) {
4846 		pci_release_msi(ctx->ifc_dev);
4847 	}
4848 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) {
4849 		iflib_irq_free(ctx, &ctx->ifc_legacy_irq);
4850 	}
4851 	if (ctx->ifc_msix_mem != NULL) {
4852 		bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY,
4853 			ctx->ifc_softc_ctx.isc_msix_bar, ctx->ifc_msix_mem);
4854 		ctx->ifc_msix_mem = NULL;
4855 	}
4856 }
4857 
4858 int
4859 iflib_device_detach(device_t dev)
4860 {
4861 	if_ctx_t ctx = device_get_softc(dev);
4862 
4863 	return (iflib_device_deregister(ctx));
4864 }
4865 
4866 int
4867 iflib_device_suspend(device_t dev)
4868 {
4869 	if_ctx_t ctx = device_get_softc(dev);
4870 
4871 	CTX_LOCK(ctx);
4872 	IFDI_SUSPEND(ctx);
4873 	CTX_UNLOCK(ctx);
4874 
4875 	return bus_generic_suspend(dev);
4876 }
4877 int
4878 iflib_device_shutdown(device_t dev)
4879 {
4880 	if_ctx_t ctx = device_get_softc(dev);
4881 
4882 	CTX_LOCK(ctx);
4883 	IFDI_SHUTDOWN(ctx);
4884 	CTX_UNLOCK(ctx);
4885 
4886 	return bus_generic_suspend(dev);
4887 }
4888 
4889 
4890 int
4891 iflib_device_resume(device_t dev)
4892 {
4893 	if_ctx_t ctx = device_get_softc(dev);
4894 	iflib_txq_t txq = ctx->ifc_txqs;
4895 
4896 	CTX_LOCK(ctx);
4897 	IFDI_RESUME(ctx);
4898 	iflib_init_locked(ctx);
4899 	CTX_UNLOCK(ctx);
4900 	for (int i = 0; i < NTXQSETS(ctx); i++, txq++)
4901 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
4902 
4903 	return (bus_generic_resume(dev));
4904 }
4905 
4906 int
4907 iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params)
4908 {
4909 	int error;
4910 	if_ctx_t ctx = device_get_softc(dev);
4911 
4912 	CTX_LOCK(ctx);
4913 	error = IFDI_IOV_INIT(ctx, num_vfs, params);
4914 	CTX_UNLOCK(ctx);
4915 
4916 	return (error);
4917 }
4918 
4919 void
4920 iflib_device_iov_uninit(device_t dev)
4921 {
4922 	if_ctx_t ctx = device_get_softc(dev);
4923 
4924 	CTX_LOCK(ctx);
4925 	IFDI_IOV_UNINIT(ctx);
4926 	CTX_UNLOCK(ctx);
4927 }
4928 
4929 int
4930 iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params)
4931 {
4932 	int error;
4933 	if_ctx_t ctx = device_get_softc(dev);
4934 
4935 	CTX_LOCK(ctx);
4936 	error = IFDI_IOV_VF_ADD(ctx, vfnum, params);
4937 	CTX_UNLOCK(ctx);
4938 
4939 	return (error);
4940 }
4941 
4942 /*********************************************************************
4943  *
4944  *  MODULE FUNCTION DEFINITIONS
4945  *
4946  **********************************************************************/
4947 
4948 /*
4949  * - Start a fast taskqueue thread for each core
4950  * - Start a taskqueue for control operations
4951  */
4952 static int
4953 iflib_module_init(void)
4954 {
4955 	return (0);
4956 }
4957 
4958 static int
4959 iflib_module_event_handler(module_t mod, int what, void *arg)
4960 {
4961 	int err;
4962 
4963 	switch (what) {
4964 	case MOD_LOAD:
4965 		if ((err = iflib_module_init()) != 0)
4966 			return (err);
4967 		break;
4968 	case MOD_UNLOAD:
4969 		return (EBUSY);
4970 	default:
4971 		return (EOPNOTSUPP);
4972 	}
4973 
4974 	return (0);
4975 }
4976 
4977 /*********************************************************************
4978  *
4979  *  PUBLIC FUNCTION DEFINITIONS
4980  *     ordered as in iflib.h
4981  *
4982  **********************************************************************/
4983 
4984 
4985 static void
4986 _iflib_assert(if_shared_ctx_t sctx)
4987 {
4988 	MPASS(sctx->isc_tx_maxsize);
4989 	MPASS(sctx->isc_tx_maxsegsize);
4990 
4991 	MPASS(sctx->isc_rx_maxsize);
4992 	MPASS(sctx->isc_rx_nsegments);
4993 	MPASS(sctx->isc_rx_maxsegsize);
4994 
4995 	MPASS(sctx->isc_nrxd_min[0]);
4996 	MPASS(sctx->isc_nrxd_max[0]);
4997 	MPASS(sctx->isc_nrxd_default[0]);
4998 	MPASS(sctx->isc_ntxd_min[0]);
4999 	MPASS(sctx->isc_ntxd_max[0]);
5000 	MPASS(sctx->isc_ntxd_default[0]);
5001 }
5002 
5003 static void
5004 _iflib_pre_assert(if_softc_ctx_t scctx)
5005 {
5006 
5007 	MPASS(scctx->isc_txrx->ift_txd_encap);
5008 	MPASS(scctx->isc_txrx->ift_txd_flush);
5009 	MPASS(scctx->isc_txrx->ift_txd_credits_update);
5010 	MPASS(scctx->isc_txrx->ift_rxd_available);
5011 	MPASS(scctx->isc_txrx->ift_rxd_pkt_get);
5012 	MPASS(scctx->isc_txrx->ift_rxd_refill);
5013 	MPASS(scctx->isc_txrx->ift_rxd_flush);
5014 }
5015 
5016 static int
5017 iflib_register(if_ctx_t ctx)
5018 {
5019 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5020 	driver_t *driver = sctx->isc_driver;
5021 	device_t dev = ctx->ifc_dev;
5022 	if_t ifp;
5023 
5024 	_iflib_assert(sctx);
5025 
5026 	CTX_LOCK_INIT(ctx);
5027 	STATE_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev));
5028 	ifp = ctx->ifc_ifp = if_alloc(IFT_ETHER);
5029 	if (ifp == NULL) {
5030 		device_printf(dev, "can not allocate ifnet structure\n");
5031 		return (ENOMEM);
5032 	}
5033 
5034 	/*
5035 	 * Initialize our context's device specific methods
5036 	 */
5037 	kobj_init((kobj_t) ctx, (kobj_class_t) driver);
5038 	kobj_class_compile((kobj_class_t) driver);
5039 	driver->refs++;
5040 
5041 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
5042 	if_setsoftc(ifp, ctx);
5043 	if_setdev(ifp, dev);
5044 	if_setinitfn(ifp, iflib_if_init);
5045 	if_setioctlfn(ifp, iflib_if_ioctl);
5046 #ifdef ALTQ
5047 	if_setstartfn(ifp, iflib_altq_if_start);
5048 	if_settransmitfn(ifp, iflib_altq_if_transmit);
5049 	if_setsendqready(ifp);
5050 #else
5051 	if_settransmitfn(ifp, iflib_if_transmit);
5052 #endif
5053 	if_setqflushfn(ifp, iflib_if_qflush);
5054 	if_setflags(ifp, IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST);
5055 
5056 	ctx->ifc_vlan_attach_event =
5057 		EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx,
5058 							  EVENTHANDLER_PRI_FIRST);
5059 	ctx->ifc_vlan_detach_event =
5060 		EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx,
5061 							  EVENTHANDLER_PRI_FIRST);
5062 
5063 	ifmedia_init(&ctx->ifc_media, IFM_IMASK,
5064 					 iflib_media_change, iflib_media_status);
5065 
5066 	return (0);
5067 }
5068 
5069 
5070 static int
5071 iflib_queues_alloc(if_ctx_t ctx)
5072 {
5073 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5074 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5075 	device_t dev = ctx->ifc_dev;
5076 	int nrxqsets = scctx->isc_nrxqsets;
5077 	int ntxqsets = scctx->isc_ntxqsets;
5078 	iflib_txq_t txq;
5079 	iflib_rxq_t rxq;
5080 	iflib_fl_t fl = NULL;
5081 	int i, j, cpu, err, txconf, rxconf;
5082 	iflib_dma_info_t ifdip;
5083 	uint32_t *rxqsizes = scctx->isc_rxqsizes;
5084 	uint32_t *txqsizes = scctx->isc_txqsizes;
5085 	uint8_t nrxqs = sctx->isc_nrxqs;
5086 	uint8_t ntxqs = sctx->isc_ntxqs;
5087 	int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1;
5088 	caddr_t *vaddrs;
5089 	uint64_t *paddrs;
5090 
5091 	KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1"));
5092 	KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1"));
5093 
5094 	/* Allocate the TX ring struct memory */
5095 	if (!(ctx->ifc_txqs =
5096 	    (iflib_txq_t) malloc(sizeof(struct iflib_txq) *
5097 	    ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5098 		device_printf(dev, "Unable to allocate TX ring memory\n");
5099 		err = ENOMEM;
5100 		goto fail;
5101 	}
5102 
5103 	/* Now allocate the RX */
5104 	if (!(ctx->ifc_rxqs =
5105 	    (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) *
5106 	    nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5107 		device_printf(dev, "Unable to allocate RX ring memory\n");
5108 		err = ENOMEM;
5109 		goto rx_fail;
5110 	}
5111 
5112 	txq = ctx->ifc_txqs;
5113 	rxq = ctx->ifc_rxqs;
5114 
5115 	/*
5116 	 * XXX handle allocation failure
5117 	 */
5118 	for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) {
5119 		/* Set up some basics */
5120 
5121 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs, M_IFLIB, M_WAITOK|M_ZERO)) == NULL) {
5122 			device_printf(dev, "failed to allocate iflib_dma_info\n");
5123 			err = ENOMEM;
5124 			goto err_tx_desc;
5125 		}
5126 		txq->ift_ifdi = ifdip;
5127 		for (j = 0; j < ntxqs; j++, ifdip++) {
5128 			if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, BUS_DMA_NOWAIT)) {
5129 				device_printf(dev, "Unable to allocate Descriptor memory\n");
5130 				err = ENOMEM;
5131 				goto err_tx_desc;
5132 			}
5133 			txq->ift_txd_size[j] = scctx->isc_txd_size[j];
5134 			bzero((void *)ifdip->idi_vaddr, txqsizes[j]);
5135 		}
5136 		txq->ift_ctx = ctx;
5137 		txq->ift_id = i;
5138 		if (sctx->isc_flags & IFLIB_HAS_TXCQ) {
5139 			txq->ift_br_offset = 1;
5140 		} else {
5141 			txq->ift_br_offset = 0;
5142 		}
5143 		/* XXX fix this */
5144 		txq->ift_timer.c_cpu = cpu;
5145 
5146 		if (iflib_txsd_alloc(txq)) {
5147 			device_printf(dev, "Critical Failure setting up TX buffers\n");
5148 			err = ENOMEM;
5149 			goto err_tx_desc;
5150 		}
5151 
5152 		/* Initialize the TX lock */
5153 		snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:tx(%d):callout",
5154 		    device_get_nameunit(dev), txq->ift_id);
5155 		mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF);
5156 		callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0);
5157 
5158 		snprintf(txq->ift_db_mtx_name, MTX_NAME_LEN, "%s:tx(%d):db",
5159 			 device_get_nameunit(dev), txq->ift_id);
5160 
5161 		err = ifmp_ring_alloc(&txq->ift_br, 2048, txq, iflib_txq_drain,
5162 				      iflib_txq_can_drain, M_IFLIB, M_WAITOK);
5163 		if (err) {
5164 			/* XXX free any allocated rings */
5165 			device_printf(dev, "Unable to allocate buf_ring\n");
5166 			goto err_tx_desc;
5167 		}
5168 	}
5169 
5170 	for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) {
5171 		/* Set up some basics */
5172 
5173 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs, M_IFLIB, M_WAITOK|M_ZERO)) == NULL) {
5174 			device_printf(dev, "failed to allocate iflib_dma_info\n");
5175 			err = ENOMEM;
5176 			goto err_tx_desc;
5177 		}
5178 
5179 		rxq->ifr_ifdi = ifdip;
5180 		/* XXX this needs to be changed if #rx queues != #tx queues */
5181 		rxq->ifr_ntxqirq = 1;
5182 		rxq->ifr_txqid[0] = i;
5183 		for (j = 0; j < nrxqs; j++, ifdip++) {
5184 			if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, BUS_DMA_NOWAIT)) {
5185 				device_printf(dev, "Unable to allocate Descriptor memory\n");
5186 				err = ENOMEM;
5187 				goto err_tx_desc;
5188 			}
5189 			bzero((void *)ifdip->idi_vaddr, rxqsizes[j]);
5190 		}
5191 		rxq->ifr_ctx = ctx;
5192 		rxq->ifr_id = i;
5193 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
5194 			rxq->ifr_fl_offset = 1;
5195 		} else {
5196 			rxq->ifr_fl_offset = 0;
5197 		}
5198 		rxq->ifr_nfl = nfree_lists;
5199 		if (!(fl =
5200 			  (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) {
5201 			device_printf(dev, "Unable to allocate free list memory\n");
5202 			err = ENOMEM;
5203 			goto err_tx_desc;
5204 		}
5205 		rxq->ifr_fl = fl;
5206 		for (j = 0; j < nfree_lists; j++) {
5207 			fl[j].ifl_rxq = rxq;
5208 			fl[j].ifl_id = j;
5209 			fl[j].ifl_ifdi = &rxq->ifr_ifdi[j + rxq->ifr_fl_offset];
5210 			fl[j].ifl_rxd_size = scctx->isc_rxd_size[j];
5211 		}
5212 		/* Allocate receive buffers for the ring */
5213 		if (iflib_rxsd_alloc(rxq)) {
5214 			device_printf(dev,
5215 			    "Critical Failure setting up receive buffers\n");
5216 			err = ENOMEM;
5217 			goto err_rx_desc;
5218 		}
5219 
5220 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5221 			fl->ifl_rx_bitmap = bit_alloc(fl->ifl_size, M_IFLIB, M_WAITOK|M_ZERO);
5222 	}
5223 
5224 	/* TXQs */
5225 	vaddrs = malloc(sizeof(caddr_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5226 	paddrs = malloc(sizeof(uint64_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5227 	for (i = 0; i < ntxqsets; i++) {
5228 		iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi;
5229 
5230 		for (j = 0; j < ntxqs; j++, di++) {
5231 			vaddrs[i*ntxqs + j] = di->idi_vaddr;
5232 			paddrs[i*ntxqs + j] = di->idi_paddr;
5233 		}
5234 	}
5235 	if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) {
5236 		device_printf(ctx->ifc_dev, "device queue allocation failed\n");
5237 		iflib_tx_structures_free(ctx);
5238 		free(vaddrs, M_IFLIB);
5239 		free(paddrs, M_IFLIB);
5240 		goto err_rx_desc;
5241 	}
5242 	free(vaddrs, M_IFLIB);
5243 	free(paddrs, M_IFLIB);
5244 
5245 	/* RXQs */
5246 	vaddrs = malloc(sizeof(caddr_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5247 	paddrs = malloc(sizeof(uint64_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5248 	for (i = 0; i < nrxqsets; i++) {
5249 		iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi;
5250 
5251 		for (j = 0; j < nrxqs; j++, di++) {
5252 			vaddrs[i*nrxqs + j] = di->idi_vaddr;
5253 			paddrs[i*nrxqs + j] = di->idi_paddr;
5254 		}
5255 	}
5256 	if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) {
5257 		device_printf(ctx->ifc_dev, "device queue allocation failed\n");
5258 		iflib_tx_structures_free(ctx);
5259 		free(vaddrs, M_IFLIB);
5260 		free(paddrs, M_IFLIB);
5261 		goto err_rx_desc;
5262 	}
5263 	free(vaddrs, M_IFLIB);
5264 	free(paddrs, M_IFLIB);
5265 
5266 	return (0);
5267 
5268 /* XXX handle allocation failure changes */
5269 err_rx_desc:
5270 err_tx_desc:
5271 rx_fail:
5272 	if (ctx->ifc_rxqs != NULL)
5273 		free(ctx->ifc_rxqs, M_IFLIB);
5274 	ctx->ifc_rxqs = NULL;
5275 	if (ctx->ifc_txqs != NULL)
5276 		free(ctx->ifc_txqs, M_IFLIB);
5277 	ctx->ifc_txqs = NULL;
5278 fail:
5279 	return (err);
5280 }
5281 
5282 static int
5283 iflib_tx_structures_setup(if_ctx_t ctx)
5284 {
5285 	iflib_txq_t txq = ctx->ifc_txqs;
5286 	int i;
5287 
5288 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
5289 		iflib_txq_setup(txq);
5290 
5291 	return (0);
5292 }
5293 
5294 static void
5295 iflib_tx_structures_free(if_ctx_t ctx)
5296 {
5297 	iflib_txq_t txq = ctx->ifc_txqs;
5298 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5299 	int i, j;
5300 
5301 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
5302 		iflib_txq_destroy(txq);
5303 		for (j = 0; j < sctx->isc_ntxqs; j++)
5304 			iflib_dma_free(&txq->ift_ifdi[j]);
5305 	}
5306 	free(ctx->ifc_txqs, M_IFLIB);
5307 	ctx->ifc_txqs = NULL;
5308 	IFDI_QUEUES_FREE(ctx);
5309 }
5310 
5311 /*********************************************************************
5312  *
5313  *  Initialize all receive rings.
5314  *
5315  **********************************************************************/
5316 static int
5317 iflib_rx_structures_setup(if_ctx_t ctx)
5318 {
5319 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5320 	int q;
5321 #if defined(INET6) || defined(INET)
5322 	int i, err;
5323 #endif
5324 
5325 	for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) {
5326 #if defined(INET6) || defined(INET)
5327 		tcp_lro_free(&rxq->ifr_lc);
5328 		if ((err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp,
5329 		    TCP_LRO_ENTRIES, min(1024,
5330 		    ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset]))) != 0) {
5331 			device_printf(ctx->ifc_dev, "LRO Initialization failed!\n");
5332 			goto fail;
5333 		}
5334 		rxq->ifr_lro_enabled = TRUE;
5335 #endif
5336 		IFDI_RXQ_SETUP(ctx, rxq->ifr_id);
5337 	}
5338 	return (0);
5339 #if defined(INET6) || defined(INET)
5340 fail:
5341 	/*
5342 	 * Free RX software descriptors allocated so far, we will only handle
5343 	 * the rings that completed, the failing case will have
5344 	 * cleaned up for itself. 'q' failed, so its the terminus.
5345 	 */
5346 	rxq = ctx->ifc_rxqs;
5347 	for (i = 0; i < q; ++i, rxq++) {
5348 		iflib_rx_sds_free(rxq);
5349 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
5350 	}
5351 	return (err);
5352 #endif
5353 }
5354 
5355 /*********************************************************************
5356  *
5357  *  Free all receive rings.
5358  *
5359  **********************************************************************/
5360 static void
5361 iflib_rx_structures_free(if_ctx_t ctx)
5362 {
5363 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5364 
5365 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) {
5366 		iflib_rx_sds_free(rxq);
5367 	}
5368 	free(ctx->ifc_rxqs, M_IFLIB);
5369 	ctx->ifc_rxqs = NULL;
5370 }
5371 
5372 static int
5373 iflib_qset_structures_setup(if_ctx_t ctx)
5374 {
5375 	int err;
5376 
5377 	/*
5378 	 * It is expected that the caller takes care of freeing queues if this
5379 	 * fails.
5380 	 */
5381 	if ((err = iflib_tx_structures_setup(ctx)) != 0) {
5382 		device_printf(ctx->ifc_dev, "iflib_tx_structures_setup failed: %d\n", err);
5383 		return (err);
5384 	}
5385 
5386 	if ((err = iflib_rx_structures_setup(ctx)) != 0)
5387 		device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err);
5388 
5389 	return (err);
5390 }
5391 
5392 int
5393 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
5394 		driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, const char *name)
5395 {
5396 
5397 	return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name));
5398 }
5399 
5400 #ifdef SMP
5401 static int
5402 find_nth(if_ctx_t ctx, int qid)
5403 {
5404 	cpuset_t cpus;
5405 	int i, cpuid, eqid, count;
5406 
5407 	CPU_COPY(&ctx->ifc_cpus, &cpus);
5408 	count = CPU_COUNT(&cpus);
5409 	eqid = qid % count;
5410 	/* clear up to the qid'th bit */
5411 	for (i = 0; i < eqid; i++) {
5412 		cpuid = CPU_FFS(&cpus);
5413 		MPASS(cpuid != 0);
5414 		CPU_CLR(cpuid-1, &cpus);
5415 	}
5416 	cpuid = CPU_FFS(&cpus);
5417 	MPASS(cpuid != 0);
5418 	return (cpuid-1);
5419 }
5420 
5421 #ifdef SCHED_ULE
5422 extern struct cpu_group *cpu_top;              /* CPU topology */
5423 
5424 static int
5425 find_child_with_core(int cpu, struct cpu_group *grp)
5426 {
5427 	int i;
5428 
5429 	if (grp->cg_children == 0)
5430 		return -1;
5431 
5432 	MPASS(grp->cg_child);
5433 	for (i = 0; i < grp->cg_children; i++) {
5434 		if (CPU_ISSET(cpu, &grp->cg_child[i].cg_mask))
5435 			return i;
5436 	}
5437 
5438 	return -1;
5439 }
5440 
5441 /*
5442  * Find the nth "close" core to the specified core
5443  * "close" is defined as the deepest level that shares
5444  * at least an L2 cache.  With threads, this will be
5445  * threads on the same core.  If the sahred cache is L3
5446  * or higher, simply returns the same core.
5447  */
5448 static int
5449 find_close_core(int cpu, int core_offset)
5450 {
5451 	struct cpu_group *grp;
5452 	int i;
5453 	int fcpu;
5454 	cpuset_t cs;
5455 
5456 	grp = cpu_top;
5457 	if (grp == NULL)
5458 		return cpu;
5459 	i = 0;
5460 	while ((i = find_child_with_core(cpu, grp)) != -1) {
5461 		/* If the child only has one cpu, don't descend */
5462 		if (grp->cg_child[i].cg_count <= 1)
5463 			break;
5464 		grp = &grp->cg_child[i];
5465 	}
5466 
5467 	/* If they don't share at least an L2 cache, use the same CPU */
5468 	if (grp->cg_level > CG_SHARE_L2 || grp->cg_level == CG_SHARE_NONE)
5469 		return cpu;
5470 
5471 	/* Now pick one */
5472 	CPU_COPY(&grp->cg_mask, &cs);
5473 
5474 	/* Add the selected CPU offset to core offset. */
5475 	for (i = 0; (fcpu = CPU_FFS(&cs)) != 0; i++) {
5476 		if (fcpu - 1 == cpu)
5477 			break;
5478 		CPU_CLR(fcpu - 1, &cs);
5479 	}
5480 	MPASS(fcpu);
5481 
5482 	core_offset += i;
5483 
5484 	CPU_COPY(&grp->cg_mask, &cs);
5485 	for (i = core_offset % grp->cg_count; i > 0; i--) {
5486 		MPASS(CPU_FFS(&cs));
5487 		CPU_CLR(CPU_FFS(&cs) - 1, &cs);
5488 	}
5489 	MPASS(CPU_FFS(&cs));
5490 	return CPU_FFS(&cs) - 1;
5491 }
5492 #else
5493 static int
5494 find_close_core(int cpu, int core_offset __unused)
5495 {
5496 	return cpu;
5497 }
5498 #endif
5499 
5500 static int
5501 get_core_offset(if_ctx_t ctx, iflib_intr_type_t type, int qid)
5502 {
5503 	switch (type) {
5504 	case IFLIB_INTR_TX:
5505 		/* TX queues get cores which share at least an L2 cache with the corresponding RX queue */
5506 		/* XXX handle multiple RX threads per core and more than two core per L2 group */
5507 		return qid / CPU_COUNT(&ctx->ifc_cpus) + 1;
5508 	case IFLIB_INTR_RX:
5509 	case IFLIB_INTR_RXTX:
5510 		/* RX queues get the specified core */
5511 		return qid / CPU_COUNT(&ctx->ifc_cpus);
5512 	default:
5513 		return -1;
5514 	}
5515 }
5516 #else
5517 #define get_core_offset(ctx, type, qid)	CPU_FIRST()
5518 #define find_close_core(cpuid, tid)	CPU_FIRST()
5519 #define find_nth(ctx, gid)		CPU_FIRST()
5520 #endif
5521 
5522 /* Just to avoid copy/paste */
5523 static inline int
5524 iflib_irq_set_affinity(if_ctx_t ctx, int irq, iflib_intr_type_t type, int qid,
5525     struct grouptask *gtask, struct taskqgroup *tqg, void *uniq, const char *name)
5526 {
5527 	int cpuid;
5528 	int err, tid;
5529 
5530 	cpuid = find_nth(ctx, qid);
5531 	tid = get_core_offset(ctx, type, qid);
5532 	MPASS(tid >= 0);
5533 	cpuid = find_close_core(cpuid, tid);
5534 	err = taskqgroup_attach_cpu(tqg, gtask, uniq, cpuid, irq, name);
5535 	if (err) {
5536 		device_printf(ctx->ifc_dev, "taskqgroup_attach_cpu failed %d\n", err);
5537 		return (err);
5538 	}
5539 #ifdef notyet
5540 	if (cpuid > ctx->ifc_cpuid_highest)
5541 		ctx->ifc_cpuid_highest = cpuid;
5542 #endif
5543 	return 0;
5544 }
5545 
5546 int
5547 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid,
5548 			iflib_intr_type_t type, driver_filter_t *filter,
5549 			void *filter_arg, int qid, const char *name)
5550 {
5551 	struct grouptask *gtask;
5552 	struct taskqgroup *tqg;
5553 	iflib_filter_info_t info;
5554 	gtask_fn_t *fn;
5555 	int tqrid, err;
5556 	driver_filter_t *intr_fast;
5557 	void *q;
5558 
5559 	info = &ctx->ifc_filter_info;
5560 	tqrid = rid;
5561 
5562 	switch (type) {
5563 	/* XXX merge tx/rx for netmap? */
5564 	case IFLIB_INTR_TX:
5565 		q = &ctx->ifc_txqs[qid];
5566 		info = &ctx->ifc_txqs[qid].ift_filter_info;
5567 		gtask = &ctx->ifc_txqs[qid].ift_task;
5568 		tqg = qgroup_if_io_tqg;
5569 		fn = _task_fn_tx;
5570 		intr_fast = iflib_fast_intr;
5571 		GROUPTASK_INIT(gtask, 0, fn, q);
5572 		ctx->ifc_flags |= IFC_NETMAP_TX_IRQ;
5573 		break;
5574 	case IFLIB_INTR_RX:
5575 		q = &ctx->ifc_rxqs[qid];
5576 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
5577 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5578 		tqg = qgroup_if_io_tqg;
5579 		fn = _task_fn_rx;
5580 		intr_fast = iflib_fast_intr;
5581 		GROUPTASK_INIT(gtask, 0, fn, q);
5582 		break;
5583 	case IFLIB_INTR_RXTX:
5584 		q = &ctx->ifc_rxqs[qid];
5585 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
5586 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5587 		tqg = qgroup_if_io_tqg;
5588 		fn = _task_fn_rx;
5589 		intr_fast = iflib_fast_intr_rxtx;
5590 		GROUPTASK_INIT(gtask, 0, fn, q);
5591 		break;
5592 	case IFLIB_INTR_ADMIN:
5593 		q = ctx;
5594 		tqrid = -1;
5595 		info = &ctx->ifc_filter_info;
5596 		gtask = &ctx->ifc_admin_task;
5597 		tqg = qgroup_if_config_tqg;
5598 		fn = _task_fn_admin;
5599 		intr_fast = iflib_fast_intr_ctx;
5600 		break;
5601 	default:
5602 		panic("unknown net intr type");
5603 	}
5604 
5605 	info->ifi_filter = filter;
5606 	info->ifi_filter_arg = filter_arg;
5607 	info->ifi_task = gtask;
5608 	info->ifi_ctx = q;
5609 
5610 	err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info,  name);
5611 	if (err != 0) {
5612 		device_printf(ctx->ifc_dev, "_iflib_irq_alloc failed %d\n", err);
5613 		return (err);
5614 	}
5615 	if (type == IFLIB_INTR_ADMIN)
5616 		return (0);
5617 
5618 	if (tqrid != -1) {
5619 		err = iflib_irq_set_affinity(ctx, rman_get_start(irq->ii_res), type, qid, gtask, tqg, q, name);
5620 		if (err)
5621 			return (err);
5622 	} else {
5623 		taskqgroup_attach(tqg, gtask, q, rman_get_start(irq->ii_res), name);
5624 	}
5625 
5626 	return (0);
5627 }
5628 
5629 void
5630 iflib_softirq_alloc_generic(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type, void *arg, int qid, const char *name)
5631 {
5632 	struct grouptask *gtask;
5633 	struct taskqgroup *tqg;
5634 	gtask_fn_t *fn;
5635 	void *q;
5636 	int irq_num = -1;
5637 	int err;
5638 
5639 	switch (type) {
5640 	case IFLIB_INTR_TX:
5641 		q = &ctx->ifc_txqs[qid];
5642 		gtask = &ctx->ifc_txqs[qid].ift_task;
5643 		tqg = qgroup_if_io_tqg;
5644 		fn = _task_fn_tx;
5645 		if (irq != NULL)
5646 			irq_num = rman_get_start(irq->ii_res);
5647 		break;
5648 	case IFLIB_INTR_RX:
5649 		q = &ctx->ifc_rxqs[qid];
5650 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5651 		tqg = qgroup_if_io_tqg;
5652 		fn = _task_fn_rx;
5653 		if (irq != NULL)
5654 			irq_num = rman_get_start(irq->ii_res);
5655 		break;
5656 	case IFLIB_INTR_IOV:
5657 		q = ctx;
5658 		gtask = &ctx->ifc_vflr_task;
5659 		tqg = qgroup_if_config_tqg;
5660 		fn = _task_fn_iov;
5661 		break;
5662 	default:
5663 		panic("unknown net intr type");
5664 	}
5665 	GROUPTASK_INIT(gtask, 0, fn, q);
5666 	if (irq_num != -1) {
5667 		err = iflib_irq_set_affinity(ctx, irq_num, type, qid, gtask, tqg, q, name);
5668 		if (err)
5669 			taskqgroup_attach(tqg, gtask, q, irq_num, name);
5670 	}
5671 	else {
5672 		taskqgroup_attach(tqg, gtask, q, irq_num, name);
5673 	}
5674 }
5675 
5676 void
5677 iflib_irq_free(if_ctx_t ctx, if_irq_t irq)
5678 {
5679 	if (irq->ii_tag)
5680 		bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag);
5681 
5682 	if (irq->ii_res)
5683 		bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ, irq->ii_rid, irq->ii_res);
5684 }
5685 
5686 static int
5687 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, const char *name)
5688 {
5689 	iflib_txq_t txq = ctx->ifc_txqs;
5690 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5691 	if_irq_t irq = &ctx->ifc_legacy_irq;
5692 	iflib_filter_info_t info;
5693 	struct grouptask *gtask;
5694 	struct taskqgroup *tqg;
5695 	gtask_fn_t *fn;
5696 	int tqrid;
5697 	void *q;
5698 	int err;
5699 
5700 	q = &ctx->ifc_rxqs[0];
5701 	info = &rxq[0].ifr_filter_info;
5702 	gtask = &rxq[0].ifr_task;
5703 	tqg = qgroup_if_io_tqg;
5704 	tqrid = irq->ii_rid = *rid;
5705 	fn = _task_fn_rx;
5706 
5707 	ctx->ifc_flags |= IFC_LEGACY;
5708 	info->ifi_filter = filter;
5709 	info->ifi_filter_arg = filter_arg;
5710 	info->ifi_task = gtask;
5711 	info->ifi_ctx = ctx;
5712 
5713 	/* We allocate a single interrupt resource */
5714 	if ((err = _iflib_irq_alloc(ctx, irq, tqrid, iflib_fast_intr_ctx, NULL, info, name)) != 0)
5715 		return (err);
5716 	GROUPTASK_INIT(gtask, 0, fn, q);
5717 	taskqgroup_attach(tqg, gtask, q, rman_get_start(irq->ii_res), name);
5718 
5719 	GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq);
5720 	taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, rman_get_start(irq->ii_res), "tx");
5721 	return (0);
5722 }
5723 
5724 void
5725 iflib_led_create(if_ctx_t ctx)
5726 {
5727 
5728 	ctx->ifc_led_dev = led_create(iflib_led_func, ctx,
5729 	    device_get_nameunit(ctx->ifc_dev));
5730 }
5731 
5732 void
5733 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid)
5734 {
5735 
5736 	GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
5737 }
5738 
5739 void
5740 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid)
5741 {
5742 
5743 	GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task);
5744 }
5745 
5746 void
5747 iflib_admin_intr_deferred(if_ctx_t ctx)
5748 {
5749 #ifdef INVARIANTS
5750 	struct grouptask *gtask;
5751 
5752 	gtask = &ctx->ifc_admin_task;
5753 	MPASS(gtask != NULL && gtask->gt_taskqueue != NULL);
5754 #endif
5755 
5756 	GROUPTASK_ENQUEUE(&ctx->ifc_admin_task);
5757 }
5758 
5759 void
5760 iflib_iov_intr_deferred(if_ctx_t ctx)
5761 {
5762 
5763 	GROUPTASK_ENQUEUE(&ctx->ifc_vflr_task);
5764 }
5765 
5766 void
5767 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, char *name)
5768 {
5769 
5770 	taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, -1, name);
5771 }
5772 
5773 void
5774 iflib_config_gtask_init(void *ctx, struct grouptask *gtask, gtask_fn_t *fn,
5775 	const char *name)
5776 {
5777 
5778 	GROUPTASK_INIT(gtask, 0, fn, ctx);
5779 	taskqgroup_attach(qgroup_if_config_tqg, gtask, gtask, -1, name);
5780 }
5781 
5782 void
5783 iflib_config_gtask_deinit(struct grouptask *gtask)
5784 {
5785 
5786 	taskqgroup_detach(qgroup_if_config_tqg, gtask);
5787 }
5788 
5789 void
5790 iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate)
5791 {
5792 	if_t ifp = ctx->ifc_ifp;
5793 	iflib_txq_t txq = ctx->ifc_txqs;
5794 
5795 	if_setbaudrate(ifp, baudrate);
5796 	if (baudrate >= IF_Gbps(10)) {
5797 		STATE_LOCK(ctx);
5798 		ctx->ifc_flags |= IFC_PREFETCH;
5799 		STATE_UNLOCK(ctx);
5800 	}
5801 	/* If link down, disable watchdog */
5802 	if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) {
5803 		for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++)
5804 			txq->ift_qstatus = IFLIB_QUEUE_IDLE;
5805 	}
5806 	ctx->ifc_link_state = link_state;
5807 	if_link_state_change(ifp, link_state);
5808 }
5809 
5810 static int
5811 iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq)
5812 {
5813 	int credits;
5814 #ifdef INVARIANTS
5815 	int credits_pre = txq->ift_cidx_processed;
5816 #endif
5817 
5818 	if (ctx->isc_txd_credits_update == NULL)
5819 		return (0);
5820 
5821 	if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, true)) == 0)
5822 		return (0);
5823 
5824 	txq->ift_processed += credits;
5825 	txq->ift_cidx_processed += credits;
5826 
5827 	MPASS(credits_pre + credits == txq->ift_cidx_processed);
5828 	if (txq->ift_cidx_processed >= txq->ift_size)
5829 		txq->ift_cidx_processed -= txq->ift_size;
5830 	return (credits);
5831 }
5832 
5833 static int
5834 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget)
5835 {
5836 
5837 	return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx,
5838 	    budget));
5839 }
5840 
5841 void
5842 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name,
5843 	const char *description, if_int_delay_info_t info,
5844 	int offset, int value)
5845 {
5846 	info->iidi_ctx = ctx;
5847 	info->iidi_offset = offset;
5848 	info->iidi_value = value;
5849 	SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev),
5850 	    SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)),
5851 	    OID_AUTO, name, CTLTYPE_INT|CTLFLAG_RW,
5852 	    info, 0, iflib_sysctl_int_delay, "I", description);
5853 }
5854 
5855 struct sx *
5856 iflib_ctx_lock_get(if_ctx_t ctx)
5857 {
5858 
5859 	return (&ctx->ifc_ctx_sx);
5860 }
5861 
5862 static int
5863 iflib_msix_init(if_ctx_t ctx)
5864 {
5865 	device_t dev = ctx->ifc_dev;
5866 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5867 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5868 	int vectors, queues, rx_queues, tx_queues, queuemsgs, msgs;
5869 	int iflib_num_tx_queues, iflib_num_rx_queues;
5870 	int err, admincnt, bar;
5871 
5872 	iflib_num_tx_queues = ctx->ifc_sysctl_ntxqs;
5873 	iflib_num_rx_queues = ctx->ifc_sysctl_nrxqs;
5874 
5875 	device_printf(dev, "msix_init qsets capped at %d\n", imax(scctx->isc_ntxqsets, scctx->isc_nrxqsets));
5876 
5877 	bar = ctx->ifc_softc_ctx.isc_msix_bar;
5878 	admincnt = sctx->isc_admin_intrcnt;
5879 	/* Override by tuneable */
5880 	if (scctx->isc_disable_msix)
5881 		goto msi;
5882 
5883 	/*
5884 	 * bar == -1 => "trust me I know what I'm doing"
5885 	 * Some drivers are for hardware that is so shoddily
5886 	 * documented that no one knows which bars are which
5887 	 * so the developer has to map all bars. This hack
5888 	 * allows shoddy garbage to use msix in this framework.
5889 	 */
5890 	if (bar != -1) {
5891 		ctx->ifc_msix_mem = bus_alloc_resource_any(dev,
5892 	            SYS_RES_MEMORY, &bar, RF_ACTIVE);
5893 		if (ctx->ifc_msix_mem == NULL) {
5894 			/* May not be enabled */
5895 			device_printf(dev, "Unable to map MSIX table \n");
5896 			goto msi;
5897 		}
5898 	}
5899 	/* First try MSI/X */
5900 	if ((msgs = pci_msix_count(dev)) == 0) { /* system has msix disabled */
5901 		device_printf(dev, "System has MSIX disabled \n");
5902 		bus_release_resource(dev, SYS_RES_MEMORY,
5903 		    bar, ctx->ifc_msix_mem);
5904 		ctx->ifc_msix_mem = NULL;
5905 		goto msi;
5906 	}
5907 #if IFLIB_DEBUG
5908 	/* use only 1 qset in debug mode */
5909 	queuemsgs = min(msgs - admincnt, 1);
5910 #else
5911 	queuemsgs = msgs - admincnt;
5912 #endif
5913 #ifdef RSS
5914 	queues = imin(queuemsgs, rss_getnumbuckets());
5915 #else
5916 	queues = queuemsgs;
5917 #endif
5918 	queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues);
5919 	device_printf(dev, "pxm cpus: %d queue msgs: %d admincnt: %d\n",
5920 				  CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt);
5921 #ifdef  RSS
5922 	/* If we're doing RSS, clamp at the number of RSS buckets */
5923 	if (queues > rss_getnumbuckets())
5924 		queues = rss_getnumbuckets();
5925 #endif
5926 	if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt)
5927 		rx_queues = iflib_num_rx_queues;
5928 	else
5929 		rx_queues = queues;
5930 
5931 	if (rx_queues > scctx->isc_nrxqsets)
5932 		rx_queues = scctx->isc_nrxqsets;
5933 
5934 	/*
5935 	 * We want this to be all logical CPUs by default
5936 	 */
5937 	if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues)
5938 		tx_queues = iflib_num_tx_queues;
5939 	else
5940 		tx_queues = mp_ncpus;
5941 
5942 	if (tx_queues > scctx->isc_ntxqsets)
5943 		tx_queues = scctx->isc_ntxqsets;
5944 
5945 	if (ctx->ifc_sysctl_qs_eq_override == 0) {
5946 #ifdef INVARIANTS
5947 		if (tx_queues != rx_queues)
5948 			device_printf(dev,
5949 			    "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n",
5950 			    min(rx_queues, tx_queues), min(rx_queues, tx_queues));
5951 #endif
5952 		tx_queues = min(rx_queues, tx_queues);
5953 		rx_queues = min(rx_queues, tx_queues);
5954 	}
5955 
5956 	device_printf(dev, "using %d rx queues %d tx queues \n", rx_queues, tx_queues);
5957 
5958 	vectors = rx_queues + admincnt;
5959 	if ((err = pci_alloc_msix(dev, &vectors)) == 0) {
5960 		device_printf(dev, "Using MSIX interrupts with %d vectors\n", vectors);
5961 		scctx->isc_vectors = vectors;
5962 		scctx->isc_nrxqsets = rx_queues;
5963 		scctx->isc_ntxqsets = tx_queues;
5964 		scctx->isc_intr = IFLIB_INTR_MSIX;
5965 
5966 		return (vectors);
5967 	} else {
5968 		device_printf(dev,
5969 		    "failed to allocate %d msix vectors, err: %d - using MSI\n", vectors, err);
5970 		bus_release_resource(dev, SYS_RES_MEMORY, bar,
5971 		    ctx->ifc_msix_mem);
5972 		ctx->ifc_msix_mem = NULL;
5973 	}
5974 msi:
5975 	vectors = pci_msi_count(dev);
5976 	scctx->isc_nrxqsets = 1;
5977 	scctx->isc_ntxqsets = 1;
5978 	scctx->isc_vectors = vectors;
5979 	if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) {
5980 		device_printf(dev,"Using an MSI interrupt\n");
5981 		scctx->isc_intr = IFLIB_INTR_MSI;
5982 	} else {
5983 		scctx->isc_vectors = 1;
5984 		device_printf(dev,"Using a Legacy interrupt\n");
5985 		scctx->isc_intr = IFLIB_INTR_LEGACY;
5986 	}
5987 
5988 	return (vectors);
5989 }
5990 
5991 static const char *ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" };
5992 
5993 static int
5994 mp_ring_state_handler(SYSCTL_HANDLER_ARGS)
5995 {
5996 	int rc;
5997 	uint16_t *state = ((uint16_t *)oidp->oid_arg1);
5998 	struct sbuf *sb;
5999 	const char *ring_state = "UNKNOWN";
6000 
6001 	/* XXX needed ? */
6002 	rc = sysctl_wire_old_buffer(req, 0);
6003 	MPASS(rc == 0);
6004 	if (rc != 0)
6005 		return (rc);
6006 	sb = sbuf_new_for_sysctl(NULL, NULL, 80, req);
6007 	MPASS(sb != NULL);
6008 	if (sb == NULL)
6009 		return (ENOMEM);
6010 	if (state[3] <= 3)
6011 		ring_state = ring_states[state[3]];
6012 
6013 	sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s",
6014 		    state[0], state[1], state[2], ring_state);
6015 	rc = sbuf_finish(sb);
6016 	sbuf_delete(sb);
6017         return(rc);
6018 }
6019 
6020 enum iflib_ndesc_handler {
6021 	IFLIB_NTXD_HANDLER,
6022 	IFLIB_NRXD_HANDLER,
6023 };
6024 
6025 static int
6026 mp_ndesc_handler(SYSCTL_HANDLER_ARGS)
6027 {
6028 	if_ctx_t ctx = (void *)arg1;
6029 	enum iflib_ndesc_handler type = arg2;
6030 	char buf[256] = {0};
6031 	qidx_t *ndesc;
6032 	char *p, *next;
6033 	int nqs, rc, i;
6034 
6035 	MPASS(type == IFLIB_NTXD_HANDLER || type == IFLIB_NRXD_HANDLER);
6036 
6037 	nqs = 8;
6038 	switch(type) {
6039 	case IFLIB_NTXD_HANDLER:
6040 		ndesc = ctx->ifc_sysctl_ntxds;
6041 		if (ctx->ifc_sctx)
6042 			nqs = ctx->ifc_sctx->isc_ntxqs;
6043 		break;
6044 	case IFLIB_NRXD_HANDLER:
6045 		ndesc = ctx->ifc_sysctl_nrxds;
6046 		if (ctx->ifc_sctx)
6047 			nqs = ctx->ifc_sctx->isc_nrxqs;
6048 		break;
6049 	default:
6050 			panic("unhandled type");
6051 	}
6052 	if (nqs == 0)
6053 		nqs = 8;
6054 
6055 	for (i=0; i<8; i++) {
6056 		if (i >= nqs)
6057 			break;
6058 		if (i)
6059 			strcat(buf, ",");
6060 		sprintf(strchr(buf, 0), "%d", ndesc[i]);
6061 	}
6062 
6063 	rc = sysctl_handle_string(oidp, buf, sizeof(buf), req);
6064 	if (rc || req->newptr == NULL)
6065 		return rc;
6066 
6067 	for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p;
6068 	    i++, p = strsep(&next, " ,")) {
6069 		ndesc[i] = strtoul(p, NULL, 10);
6070 	}
6071 
6072 	return(rc);
6073 }
6074 
6075 #define NAME_BUFLEN 32
6076 static void
6077 iflib_add_device_sysctl_pre(if_ctx_t ctx)
6078 {
6079         device_t dev = iflib_get_dev(ctx);
6080 	struct sysctl_oid_list *child, *oid_list;
6081 	struct sysctl_ctx_list *ctx_list;
6082 	struct sysctl_oid *node;
6083 
6084 	ctx_list = device_get_sysctl_ctx(dev);
6085 	child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev));
6086 	ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, "iflib",
6087 						      CTLFLAG_RD, NULL, "IFLIB fields");
6088 	oid_list = SYSCTL_CHILDREN(node);
6089 
6090 	SYSCTL_ADD_STRING(ctx_list, oid_list, OID_AUTO, "driver_version",
6091 		       CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version, 0,
6092 		       "driver version");
6093 
6094 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs",
6095 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0,
6096 			"# of txqs to use, 0 => use default #");
6097 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs",
6098 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0,
6099 			"# of rxqs to use, 0 => use default #");
6100 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_qs_enable",
6101 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0,
6102                        "permit #txq != #rxq");
6103 	SYSCTL_ADD_INT(ctx_list, oid_list, OID_AUTO, "disable_msix",
6104                       CTLFLAG_RWTUN, &ctx->ifc_softc_ctx.isc_disable_msix, 0,
6105                       "disable MSIX (default 0)");
6106 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "rx_budget",
6107 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_rx_budget, 0,
6108                        "set the rx budget");
6109 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "tx_abdicate",
6110 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_tx_abdicate, 0,
6111 		       "cause tx to abdicate instead of running to completion");
6112 
6113 	/* XXX change for per-queue sizes */
6114 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_ntxds",
6115 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NTXD_HANDLER,
6116                        mp_ndesc_handler, "A",
6117                        "list of # of tx descriptors to use, 0 = use default #");
6118 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_nrxds",
6119 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NRXD_HANDLER,
6120                        mp_ndesc_handler, "A",
6121                        "list of # of rx descriptors to use, 0 = use default #");
6122 }
6123 
6124 static void
6125 iflib_add_device_sysctl_post(if_ctx_t ctx)
6126 {
6127 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6128 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6129         device_t dev = iflib_get_dev(ctx);
6130 	struct sysctl_oid_list *child;
6131 	struct sysctl_ctx_list *ctx_list;
6132 	iflib_fl_t fl;
6133 	iflib_txq_t txq;
6134 	iflib_rxq_t rxq;
6135 	int i, j;
6136 	char namebuf[NAME_BUFLEN];
6137 	char *qfmt;
6138 	struct sysctl_oid *queue_node, *fl_node, *node;
6139 	struct sysctl_oid_list *queue_list, *fl_list;
6140 	ctx_list = device_get_sysctl_ctx(dev);
6141 
6142 	node = ctx->ifc_sysctl_node;
6143 	child = SYSCTL_CHILDREN(node);
6144 
6145 	if (scctx->isc_ntxqsets > 100)
6146 		qfmt = "txq%03d";
6147 	else if (scctx->isc_ntxqsets > 10)
6148 		qfmt = "txq%02d";
6149 	else
6150 		qfmt = "txq%d";
6151 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
6152 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6153 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6154 					     CTLFLAG_RD, NULL, "Queue Name");
6155 		queue_list = SYSCTL_CHILDREN(queue_node);
6156 #if MEMORY_LOGGING
6157 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued",
6158 				CTLFLAG_RD,
6159 				&txq->ift_dequeued, "total mbufs freed");
6160 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued",
6161 				CTLFLAG_RD,
6162 				&txq->ift_enqueued, "total mbufs enqueued");
6163 #endif
6164 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag",
6165 				   CTLFLAG_RD,
6166 				   &txq->ift_mbuf_defrag, "# of times m_defrag was called");
6167 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "m_pullups",
6168 				   CTLFLAG_RD,
6169 				   &txq->ift_pullups, "# of times m_pullup was called");
6170 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag_failed",
6171 				   CTLFLAG_RD,
6172 				   &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed");
6173 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_desc_avail",
6174 				   CTLFLAG_RD,
6175 				   &txq->ift_no_desc_avail, "# of times no descriptors were available");
6176 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "tx_map_failed",
6177 				   CTLFLAG_RD,
6178 				   &txq->ift_map_failed, "# of times dma map failed");
6179 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txd_encap_efbig",
6180 				   CTLFLAG_RD,
6181 				   &txq->ift_txd_encap_efbig, "# of times txd_encap returned EFBIG");
6182 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_tx_dma_setup",
6183 				   CTLFLAG_RD,
6184 				   &txq->ift_no_tx_dma_setup, "# of times map failed for other than EFBIG");
6185 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx",
6186 				   CTLFLAG_RD,
6187 				   &txq->ift_pidx, 1, "Producer Index");
6188 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx",
6189 				   CTLFLAG_RD,
6190 				   &txq->ift_cidx, 1, "Consumer Index");
6191 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx_processed",
6192 				   CTLFLAG_RD,
6193 				   &txq->ift_cidx_processed, 1, "Consumer Index seen by credit update");
6194 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use",
6195 				   CTLFLAG_RD,
6196 				   &txq->ift_in_use, 1, "descriptors in use");
6197 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_processed",
6198 				   CTLFLAG_RD,
6199 				   &txq->ift_processed, "descriptors procesed for clean");
6200 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned",
6201 				   CTLFLAG_RD,
6202 				   &txq->ift_cleaned, "total cleaned");
6203 		SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state",
6204 				CTLTYPE_STRING | CTLFLAG_RD, __DEVOLATILE(uint64_t *, &txq->ift_br->state),
6205 				0, mp_ring_state_handler, "A", "soft ring state");
6206 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_enqueues",
6207 				       CTLFLAG_RD, &txq->ift_br->enqueues,
6208 				       "# of enqueues to the mp_ring for this queue");
6209 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_drops",
6210 				       CTLFLAG_RD, &txq->ift_br->drops,
6211 				       "# of drops in the mp_ring for this queue");
6212 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_starts",
6213 				       CTLFLAG_RD, &txq->ift_br->starts,
6214 				       "# of normal consumer starts in the mp_ring for this queue");
6215 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_stalls",
6216 				       CTLFLAG_RD, &txq->ift_br->stalls,
6217 					       "# of consumer stalls in the mp_ring for this queue");
6218 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_restarts",
6219 			       CTLFLAG_RD, &txq->ift_br->restarts,
6220 				       "# of consumer restarts in the mp_ring for this queue");
6221 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_abdications",
6222 				       CTLFLAG_RD, &txq->ift_br->abdications,
6223 				       "# of consumer abdications in the mp_ring for this queue");
6224 	}
6225 
6226 	if (scctx->isc_nrxqsets > 100)
6227 		qfmt = "rxq%03d";
6228 	else if (scctx->isc_nrxqsets > 10)
6229 		qfmt = "rxq%02d";
6230 	else
6231 		qfmt = "rxq%d";
6232 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
6233 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6234 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6235 					     CTLFLAG_RD, NULL, "Queue Name");
6236 		queue_list = SYSCTL_CHILDREN(queue_node);
6237 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
6238 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_pidx",
6239 				       CTLFLAG_RD,
6240 				       &rxq->ifr_cq_pidx, 1, "Producer Index");
6241 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_cidx",
6242 				       CTLFLAG_RD,
6243 				       &rxq->ifr_cq_cidx, 1, "Consumer Index");
6244 		}
6245 
6246 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
6247 			snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j);
6248 			fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list, OID_AUTO, namebuf,
6249 						     CTLFLAG_RD, NULL, "freelist Name");
6250 			fl_list = SYSCTL_CHILDREN(fl_node);
6251 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx",
6252 				       CTLFLAG_RD,
6253 				       &fl->ifl_pidx, 1, "Producer Index");
6254 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx",
6255 				       CTLFLAG_RD,
6256 				       &fl->ifl_cidx, 1, "Consumer Index");
6257 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits",
6258 				       CTLFLAG_RD,
6259 				       &fl->ifl_credits, 1, "credits available");
6260 #if MEMORY_LOGGING
6261 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_enqueued",
6262 					CTLFLAG_RD,
6263 					&fl->ifl_m_enqueued, "mbufs allocated");
6264 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_dequeued",
6265 					CTLFLAG_RD,
6266 					&fl->ifl_m_dequeued, "mbufs freed");
6267 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_enqueued",
6268 					CTLFLAG_RD,
6269 					&fl->ifl_cl_enqueued, "clusters allocated");
6270 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_dequeued",
6271 					CTLFLAG_RD,
6272 					&fl->ifl_cl_dequeued, "clusters freed");
6273 #endif
6274 
6275 		}
6276 	}
6277 
6278 }
6279 
6280 void
6281 iflib_request_reset(if_ctx_t ctx)
6282 {
6283 
6284 	STATE_LOCK(ctx);
6285 	ctx->ifc_flags |= IFC_DO_RESET;
6286 	STATE_UNLOCK(ctx);
6287 }
6288 
6289 #ifndef __NO_STRICT_ALIGNMENT
6290 static struct mbuf *
6291 iflib_fixup_rx(struct mbuf *m)
6292 {
6293 	struct mbuf *n;
6294 
6295 	if (m->m_len <= (MCLBYTES - ETHER_HDR_LEN)) {
6296 		bcopy(m->m_data, m->m_data + ETHER_HDR_LEN, m->m_len);
6297 		m->m_data += ETHER_HDR_LEN;
6298 		n = m;
6299 	} else {
6300 		MGETHDR(n, M_NOWAIT, MT_DATA);
6301 		if (n == NULL) {
6302 			m_freem(m);
6303 			return (NULL);
6304 		}
6305 		bcopy(m->m_data, n->m_data, ETHER_HDR_LEN);
6306 		m->m_data += ETHER_HDR_LEN;
6307 		m->m_len -= ETHER_HDR_LEN;
6308 		n->m_len = ETHER_HDR_LEN;
6309 		M_MOVE_PKTHDR(n, m);
6310 		n->m_next = m;
6311 	}
6312 	return (n);
6313 }
6314 #endif
6315 
6316 #ifdef NETDUMP
6317 static void
6318 iflib_netdump_init(struct ifnet *ifp, int *nrxr, int *ncl, int *clsize)
6319 {
6320 	if_ctx_t ctx;
6321 
6322 	ctx = if_getsoftc(ifp);
6323 	CTX_LOCK(ctx);
6324 	*nrxr = NRXQSETS(ctx);
6325 	*ncl = ctx->ifc_rxqs[0].ifr_fl->ifl_size;
6326 	*clsize = ctx->ifc_rxqs[0].ifr_fl->ifl_buf_size;
6327 	CTX_UNLOCK(ctx);
6328 }
6329 
6330 static void
6331 iflib_netdump_event(struct ifnet *ifp, enum netdump_ev event)
6332 {
6333 	if_ctx_t ctx;
6334 	if_softc_ctx_t scctx;
6335 	iflib_fl_t fl;
6336 	iflib_rxq_t rxq;
6337 	int i, j;
6338 
6339 	ctx = if_getsoftc(ifp);
6340 	scctx = &ctx->ifc_softc_ctx;
6341 
6342 	switch (event) {
6343 	case NETDUMP_START:
6344 		for (i = 0; i < scctx->isc_nrxqsets; i++) {
6345 			rxq = &ctx->ifc_rxqs[i];
6346 			for (j = 0; j < rxq->ifr_nfl; j++) {
6347 				fl = rxq->ifr_fl;
6348 				fl->ifl_zone = m_getzone(fl->ifl_buf_size);
6349 			}
6350 		}
6351 		iflib_no_tx_batch = 1;
6352 		break;
6353 	default:
6354 		break;
6355 	}
6356 }
6357 
6358 static int
6359 iflib_netdump_transmit(struct ifnet *ifp, struct mbuf *m)
6360 {
6361 	if_ctx_t ctx;
6362 	iflib_txq_t txq;
6363 	int error;
6364 
6365 	ctx = if_getsoftc(ifp);
6366 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
6367 	    IFF_DRV_RUNNING)
6368 		return (EBUSY);
6369 
6370 	txq = &ctx->ifc_txqs[0];
6371 	error = iflib_encap(txq, &m);
6372 	if (error == 0)
6373 		(void)iflib_txd_db_check(ctx, txq, true, txq->ift_in_use);
6374 	return (error);
6375 }
6376 
6377 static int
6378 iflib_netdump_poll(struct ifnet *ifp, int count)
6379 {
6380 	if_ctx_t ctx;
6381 	if_softc_ctx_t scctx;
6382 	iflib_txq_t txq;
6383 	int i;
6384 
6385 	ctx = if_getsoftc(ifp);
6386 	scctx = &ctx->ifc_softc_ctx;
6387 
6388 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
6389 	    IFF_DRV_RUNNING)
6390 		return (EBUSY);
6391 
6392 	txq = &ctx->ifc_txqs[0];
6393 	(void)iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
6394 
6395 	for (i = 0; i < scctx->isc_nrxqsets; i++)
6396 		(void)iflib_rxeof(&ctx->ifc_rxqs[i], 16 /* XXX */);
6397 	return (0);
6398 }
6399 #endif /* NETDUMP */
6400