xref: /freebsd/sys/net/iflib.c (revision 48d41ef0fb2a73cd29a75544e164a59829c29351)
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 #if defined(__i386__) || defined(__amd64__)
96 #include <sys/memdesc.h>
97 #include <machine/bus.h>
98 #include <machine/md_var.h>
99 #include <machine/specialreg.h>
100 #include <x86/include/busdma_impl.h>
101 #include <x86/iommu/busdma_dmar.h>
102 #endif
103 
104 #include <sys/bitstring.h>
105 /*
106  * enable accounting of every mbuf as it comes in to and goes out of
107  * iflib's software descriptor references
108  */
109 #define MEMORY_LOGGING 0
110 /*
111  * Enable mbuf vectors for compressing long mbuf chains
112  */
113 
114 /*
115  * NB:
116  * - Prefetching in tx cleaning should perhaps be a tunable. The distance ahead
117  *   we prefetch needs to be determined by the time spent in m_free vis a vis
118  *   the cost of a prefetch. This will of course vary based on the workload:
119  *      - NFLX's m_free path is dominated by vm-based M_EXT manipulation which
120  *        is quite expensive, thus suggesting very little prefetch.
121  *      - small packet forwarding which is just returning a single mbuf to
122  *        UMA will typically be very fast vis a vis the cost of a memory
123  *        access.
124  */
125 
126 
127 /*
128  * File organization:
129  *  - private structures
130  *  - iflib private utility functions
131  *  - ifnet functions
132  *  - vlan registry and other exported functions
133  *  - iflib public core functions
134  *
135  *
136  */
137 MALLOC_DEFINE(M_IFLIB, "iflib", "ifnet library");
138 
139 struct iflib_txq;
140 typedef struct iflib_txq *iflib_txq_t;
141 struct iflib_rxq;
142 typedef struct iflib_rxq *iflib_rxq_t;
143 struct iflib_fl;
144 typedef struct iflib_fl *iflib_fl_t;
145 
146 struct iflib_ctx;
147 
148 static void iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid);
149 
150 typedef struct iflib_filter_info {
151 	driver_filter_t *ifi_filter;
152 	void *ifi_filter_arg;
153 	struct grouptask *ifi_task;
154 	void *ifi_ctx;
155 } *iflib_filter_info_t;
156 
157 struct iflib_ctx {
158 	KOBJ_FIELDS;
159    /*
160    * Pointer to hardware driver's softc
161    */
162 	void *ifc_softc;
163 	device_t ifc_dev;
164 	if_t ifc_ifp;
165 
166 	cpuset_t ifc_cpus;
167 	if_shared_ctx_t ifc_sctx;
168 	struct if_softc_ctx ifc_softc_ctx;
169 
170 	struct sx ifc_ctx_sx;
171 	struct mtx ifc_state_mtx;
172 
173 	uint16_t ifc_nhwtxqs;
174 
175 	iflib_txq_t ifc_txqs;
176 	iflib_rxq_t ifc_rxqs;
177 	uint32_t ifc_if_flags;
178 	uint32_t ifc_flags;
179 	uint32_t ifc_max_fl_buf_size;
180 	int ifc_in_detach;
181 
182 	int ifc_link_state;
183 	int ifc_link_irq;
184 	int ifc_watchdog_events;
185 	struct cdev *ifc_led_dev;
186 	struct resource *ifc_msix_mem;
187 
188 	struct if_irq ifc_legacy_irq;
189 	struct grouptask ifc_admin_task;
190 	struct grouptask ifc_vflr_task;
191 	struct iflib_filter_info ifc_filter_info;
192 	struct ifmedia	ifc_media;
193 
194 	struct sysctl_oid *ifc_sysctl_node;
195 	uint16_t ifc_sysctl_ntxqs;
196 	uint16_t ifc_sysctl_nrxqs;
197 	uint16_t ifc_sysctl_qs_eq_override;
198 	uint16_t ifc_sysctl_rx_budget;
199 
200 	qidx_t ifc_sysctl_ntxds[8];
201 	qidx_t ifc_sysctl_nrxds[8];
202 	struct if_txrx ifc_txrx;
203 #define isc_txd_encap  ifc_txrx.ift_txd_encap
204 #define isc_txd_flush  ifc_txrx.ift_txd_flush
205 #define isc_txd_credits_update  ifc_txrx.ift_txd_credits_update
206 #define isc_rxd_available ifc_txrx.ift_rxd_available
207 #define isc_rxd_pkt_get ifc_txrx.ift_rxd_pkt_get
208 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
209 #define isc_rxd_flush ifc_txrx.ift_rxd_flush
210 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
211 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
212 #define isc_legacy_intr ifc_txrx.ift_legacy_intr
213 	eventhandler_tag ifc_vlan_attach_event;
214 	eventhandler_tag ifc_vlan_detach_event;
215 	uint8_t ifc_mac[ETHER_ADDR_LEN];
216 	char ifc_mtx_name[16];
217 };
218 
219 
220 void *
221 iflib_get_softc(if_ctx_t ctx)
222 {
223 
224 	return (ctx->ifc_softc);
225 }
226 
227 device_t
228 iflib_get_dev(if_ctx_t ctx)
229 {
230 
231 	return (ctx->ifc_dev);
232 }
233 
234 if_t
235 iflib_get_ifp(if_ctx_t ctx)
236 {
237 
238 	return (ctx->ifc_ifp);
239 }
240 
241 struct ifmedia *
242 iflib_get_media(if_ctx_t ctx)
243 {
244 
245 	return (&ctx->ifc_media);
246 }
247 
248 uint32_t
249 iflib_get_flags(if_ctx_t ctx)
250 {
251 	return (ctx->ifc_flags);
252 }
253 
254 void
255 iflib_set_detach(if_ctx_t ctx)
256 {
257 	ctx->ifc_in_detach = 1;
258 }
259 
260 void
261 iflib_set_mac(if_ctx_t ctx, uint8_t mac[ETHER_ADDR_LEN])
262 {
263 
264 	bcopy(mac, ctx->ifc_mac, ETHER_ADDR_LEN);
265 }
266 
267 if_softc_ctx_t
268 iflib_get_softc_ctx(if_ctx_t ctx)
269 {
270 
271 	return (&ctx->ifc_softc_ctx);
272 }
273 
274 if_shared_ctx_t
275 iflib_get_sctx(if_ctx_t ctx)
276 {
277 
278 	return (ctx->ifc_sctx);
279 }
280 
281 #define IP_ALIGNED(m) ((((uintptr_t)(m)->m_data) & 0x3) == 0x2)
282 #define CACHE_PTR_INCREMENT (CACHE_LINE_SIZE/sizeof(void*))
283 #define CACHE_PTR_NEXT(ptr) ((void *)(((uintptr_t)(ptr)+CACHE_LINE_SIZE-1) & (CACHE_LINE_SIZE-1)))
284 
285 #define LINK_ACTIVE(ctx) ((ctx)->ifc_link_state == LINK_STATE_UP)
286 #define CTX_IS_VF(ctx) ((ctx)->ifc_sctx->isc_flags & IFLIB_IS_VF)
287 
288 #define RX_SW_DESC_MAP_CREATED	(1 << 0)
289 #define TX_SW_DESC_MAP_CREATED	(1 << 1)
290 #define RX_SW_DESC_INUSE        (1 << 3)
291 #define TX_SW_DESC_MAPPED       (1 << 4)
292 
293 #define	M_TOOBIG		M_PROTO1
294 
295 typedef struct iflib_sw_rx_desc_array {
296 	bus_dmamap_t	*ifsd_map;         /* bus_dma maps for packet */
297 	struct mbuf	**ifsd_m;           /* pkthdr mbufs */
298 	caddr_t		*ifsd_cl;          /* direct cluster pointer for rx */
299 	uint8_t		*ifsd_flags;
300 } iflib_rxsd_array_t;
301 
302 typedef struct iflib_sw_tx_desc_array {
303 	bus_dmamap_t    *ifsd_map;         /* bus_dma maps for packet */
304 	struct mbuf    **ifsd_m;           /* pkthdr mbufs */
305 	uint8_t		*ifsd_flags;
306 } if_txsd_vec_t;
307 
308 
309 /* magic number that should be high enough for any hardware */
310 #define IFLIB_MAX_TX_SEGS		128
311 /* bnxt supports 64 with hardware LRO enabled */
312 #define IFLIB_MAX_RX_SEGS		64
313 #define IFLIB_RX_COPY_THRESH		128
314 #define IFLIB_MAX_RX_REFRESH		32
315 /* The minimum descriptors per second before we start coalescing */
316 #define IFLIB_MIN_DESC_SEC		16384
317 #define IFLIB_DEFAULT_TX_UPDATE_FREQ	16
318 #define IFLIB_QUEUE_IDLE		0
319 #define IFLIB_QUEUE_HUNG		1
320 #define IFLIB_QUEUE_WORKING		2
321 /* maximum number of txqs that can share an rx interrupt */
322 #define IFLIB_MAX_TX_SHARED_INTR	4
323 
324 /* this should really scale with ring size - this is a fairly arbitrary value */
325 #define TX_BATCH_SIZE			32
326 
327 #define IFLIB_RESTART_BUDGET		8
328 
329 
330 #define CSUM_OFFLOAD		(CSUM_IP_TSO|CSUM_IP6_TSO|CSUM_IP| \
331 				 CSUM_IP_UDP|CSUM_IP_TCP|CSUM_IP_SCTP| \
332 				 CSUM_IP6_UDP|CSUM_IP6_TCP|CSUM_IP6_SCTP)
333 struct iflib_txq {
334 	qidx_t		ift_in_use;
335 	qidx_t		ift_cidx;
336 	qidx_t		ift_cidx_processed;
337 	qidx_t		ift_pidx;
338 	uint8_t		ift_gen;
339 	uint8_t		ift_br_offset;
340 	uint16_t	ift_npending;
341 	uint16_t	ift_db_pending;
342 	uint16_t	ift_rs_pending;
343 	/* implicit pad */
344 	uint8_t		ift_txd_size[8];
345 	uint64_t	ift_processed;
346 	uint64_t	ift_cleaned;
347 	uint64_t	ift_cleaned_prev;
348 #if MEMORY_LOGGING
349 	uint64_t	ift_enqueued;
350 	uint64_t	ift_dequeued;
351 #endif
352 	uint64_t	ift_no_tx_dma_setup;
353 	uint64_t	ift_no_desc_avail;
354 	uint64_t	ift_mbuf_defrag_failed;
355 	uint64_t	ift_mbuf_defrag;
356 	uint64_t	ift_map_failed;
357 	uint64_t	ift_txd_encap_efbig;
358 	uint64_t	ift_pullups;
359 
360 	struct mtx	ift_mtx;
361 	struct mtx	ift_db_mtx;
362 
363 	/* constant values */
364 	if_ctx_t	ift_ctx;
365 	struct ifmp_ring        *ift_br;
366 	struct grouptask	ift_task;
367 	qidx_t		ift_size;
368 	uint16_t	ift_id;
369 	struct callout	ift_timer;
370 
371 	if_txsd_vec_t	ift_sds;
372 	uint8_t		ift_qstatus;
373 	uint8_t		ift_closed;
374 	uint8_t		ift_update_freq;
375 	struct iflib_filter_info ift_filter_info;
376 	bus_dma_tag_t		ift_desc_tag;
377 	bus_dma_tag_t		ift_tso_desc_tag;
378 	iflib_dma_info_t	ift_ifdi;
379 #define MTX_NAME_LEN 16
380 	char                    ift_mtx_name[MTX_NAME_LEN];
381 	char                    ift_db_mtx_name[MTX_NAME_LEN];
382 	bus_dma_segment_t	ift_segs[IFLIB_MAX_TX_SEGS]  __aligned(CACHE_LINE_SIZE);
383 #ifdef IFLIB_DIAGNOSTICS
384 	uint64_t ift_cpu_exec_count[256];
385 #endif
386 } __aligned(CACHE_LINE_SIZE);
387 
388 struct iflib_fl {
389 	qidx_t		ifl_cidx;
390 	qidx_t		ifl_pidx;
391 	qidx_t		ifl_credits;
392 	uint8_t		ifl_gen;
393 	uint8_t		ifl_rxd_size;
394 #if MEMORY_LOGGING
395 	uint64_t	ifl_m_enqueued;
396 	uint64_t	ifl_m_dequeued;
397 	uint64_t	ifl_cl_enqueued;
398 	uint64_t	ifl_cl_dequeued;
399 #endif
400 	/* implicit pad */
401 
402 	bitstr_t 	*ifl_rx_bitmap;
403 	qidx_t		ifl_fragidx;
404 	/* constant */
405 	qidx_t		ifl_size;
406 	uint16_t	ifl_buf_size;
407 	uint16_t	ifl_cltype;
408 	uma_zone_t	ifl_zone;
409 	iflib_rxsd_array_t	ifl_sds;
410 	iflib_rxq_t	ifl_rxq;
411 	uint8_t		ifl_id;
412 	bus_dma_tag_t           ifl_desc_tag;
413 	iflib_dma_info_t	ifl_ifdi;
414 	uint64_t	ifl_bus_addrs[IFLIB_MAX_RX_REFRESH] __aligned(CACHE_LINE_SIZE);
415 	caddr_t		ifl_vm_addrs[IFLIB_MAX_RX_REFRESH];
416 	qidx_t	ifl_rxd_idxs[IFLIB_MAX_RX_REFRESH];
417 }  __aligned(CACHE_LINE_SIZE);
418 
419 static inline qidx_t
420 get_inuse(int size, qidx_t cidx, qidx_t pidx, uint8_t gen)
421 {
422 	qidx_t used;
423 
424 	if (pidx > cidx)
425 		used = pidx - cidx;
426 	else if (pidx < cidx)
427 		used = size - cidx + pidx;
428 	else if (gen == 0 && pidx == cidx)
429 		used = 0;
430 	else if (gen == 1 && pidx == cidx)
431 		used = size;
432 	else
433 		panic("bad state");
434 
435 	return (used);
436 }
437 
438 #define TXQ_AVAIL(txq) (txq->ift_size - get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx, txq->ift_gen))
439 
440 #define IDXDIFF(head, tail, wrap) \
441 	((head) >= (tail) ? (head) - (tail) : (wrap) - (tail) + (head))
442 
443 struct iflib_rxq {
444 	/* If there is a separate completion queue -
445 	 * these are the cq cidx and pidx. Otherwise
446 	 * these are unused.
447 	 */
448 	qidx_t		ifr_size;
449 	qidx_t		ifr_cq_cidx;
450 	qidx_t		ifr_cq_pidx;
451 	uint8_t		ifr_cq_gen;
452 	uint8_t		ifr_fl_offset;
453 
454 	if_ctx_t	ifr_ctx;
455 	iflib_fl_t	ifr_fl;
456 	uint64_t	ifr_rx_irq;
457 	uint16_t	ifr_id;
458 	uint8_t		ifr_lro_enabled;
459 	uint8_t		ifr_nfl;
460 	uint8_t		ifr_ntxqirq;
461 	uint8_t		ifr_txqid[IFLIB_MAX_TX_SHARED_INTR];
462 	struct lro_ctrl			ifr_lc;
463 	struct grouptask        ifr_task;
464 	struct iflib_filter_info ifr_filter_info;
465 	iflib_dma_info_t		ifr_ifdi;
466 
467 	/* dynamically allocate if any drivers need a value substantially larger than this */
468 	struct if_rxd_frag	ifr_frags[IFLIB_MAX_RX_SEGS] __aligned(CACHE_LINE_SIZE);
469 #ifdef IFLIB_DIAGNOSTICS
470 	uint64_t ifr_cpu_exec_count[256];
471 #endif
472 }  __aligned(CACHE_LINE_SIZE);
473 
474 typedef struct if_rxsd {
475 	caddr_t *ifsd_cl;
476 	struct mbuf **ifsd_m;
477 	iflib_fl_t ifsd_fl;
478 	qidx_t ifsd_cidx;
479 } *if_rxsd_t;
480 
481 /* multiple of word size */
482 #ifdef __LP64__
483 #define PKT_INFO_SIZE	6
484 #define RXD_INFO_SIZE	5
485 #define PKT_TYPE uint64_t
486 #else
487 #define PKT_INFO_SIZE	11
488 #define RXD_INFO_SIZE	8
489 #define PKT_TYPE uint32_t
490 #endif
491 #define PKT_LOOP_BOUND  ((PKT_INFO_SIZE/3)*3)
492 #define RXD_LOOP_BOUND  ((RXD_INFO_SIZE/4)*4)
493 
494 typedef struct if_pkt_info_pad {
495 	PKT_TYPE pkt_val[PKT_INFO_SIZE];
496 } *if_pkt_info_pad_t;
497 typedef struct if_rxd_info_pad {
498 	PKT_TYPE rxd_val[RXD_INFO_SIZE];
499 } *if_rxd_info_pad_t;
500 
501 CTASSERT(sizeof(struct if_pkt_info_pad) == sizeof(struct if_pkt_info));
502 CTASSERT(sizeof(struct if_rxd_info_pad) == sizeof(struct if_rxd_info));
503 
504 
505 static inline void
506 pkt_info_zero(if_pkt_info_t pi)
507 {
508 	if_pkt_info_pad_t pi_pad;
509 
510 	pi_pad = (if_pkt_info_pad_t)pi;
511 	pi_pad->pkt_val[0] = 0; pi_pad->pkt_val[1] = 0; pi_pad->pkt_val[2] = 0;
512 	pi_pad->pkt_val[3] = 0; pi_pad->pkt_val[4] = 0; pi_pad->pkt_val[5] = 0;
513 #ifndef __LP64__
514 	pi_pad->pkt_val[6] = 0; pi_pad->pkt_val[7] = 0; pi_pad->pkt_val[8] = 0;
515 	pi_pad->pkt_val[9] = 0; pi_pad->pkt_val[10] = 0;
516 #endif
517 }
518 
519 static device_method_t iflib_pseudo_methods[] = {
520 	DEVMETHOD(device_attach, noop_attach),
521 	DEVMETHOD(device_detach, iflib_pseudo_detach),
522 	DEVMETHOD_END
523 };
524 
525 driver_t iflib_pseudodriver = {
526 	"iflib_pseudo", iflib_pseudo_methods, sizeof(struct iflib_ctx),
527 };
528 
529 static inline void
530 rxd_info_zero(if_rxd_info_t ri)
531 {
532 	if_rxd_info_pad_t ri_pad;
533 	int i;
534 
535 	ri_pad = (if_rxd_info_pad_t)ri;
536 	for (i = 0; i < RXD_LOOP_BOUND; i += 4) {
537 		ri_pad->rxd_val[i] = 0;
538 		ri_pad->rxd_val[i+1] = 0;
539 		ri_pad->rxd_val[i+2] = 0;
540 		ri_pad->rxd_val[i+3] = 0;
541 	}
542 #ifdef __LP64__
543 	ri_pad->rxd_val[RXD_INFO_SIZE-1] = 0;
544 #endif
545 }
546 
547 /*
548  * Only allow a single packet to take up most 1/nth of the tx ring
549  */
550 #define MAX_SINGLE_PACKET_FRACTION 12
551 #define IF_BAD_DMA (bus_addr_t)-1
552 
553 #define CTX_ACTIVE(ctx) ((if_getdrvflags((ctx)->ifc_ifp) & IFF_DRV_RUNNING))
554 
555 #define CTX_LOCK_INIT(_sc)  sx_init(&(_sc)->ifc_ctx_sx, "iflib ctx lock")
556 #define CTX_LOCK(ctx) sx_xlock(&(ctx)->ifc_ctx_sx)
557 #define CTX_UNLOCK(ctx) sx_xunlock(&(ctx)->ifc_ctx_sx)
558 #define CTX_LOCK_DESTROY(ctx) sx_destroy(&(ctx)->ifc_ctx_sx)
559 
560 
561 #define STATE_LOCK_INIT(_sc, _name)  mtx_init(&(_sc)->ifc_state_mtx, _name, "iflib state lock", MTX_DEF)
562 #define STATE_LOCK(ctx) mtx_lock(&(ctx)->ifc_state_mtx)
563 #define STATE_UNLOCK(ctx) mtx_unlock(&(ctx)->ifc_state_mtx)
564 #define STATE_LOCK_DESTROY(ctx) mtx_destroy(&(ctx)->ifc_state_mtx)
565 
566 
567 
568 #define CALLOUT_LOCK(txq)	mtx_lock(&txq->ift_mtx)
569 #define CALLOUT_UNLOCK(txq) 	mtx_unlock(&txq->ift_mtx)
570 
571 
572 /* Our boot-time initialization hook */
573 static int	iflib_module_event_handler(module_t, int, void *);
574 
575 static moduledata_t iflib_moduledata = {
576 	"iflib",
577 	iflib_module_event_handler,
578 	NULL
579 };
580 
581 DECLARE_MODULE(iflib, iflib_moduledata, SI_SUB_INIT_IF, SI_ORDER_ANY);
582 MODULE_VERSION(iflib, 1);
583 
584 MODULE_DEPEND(iflib, pci, 1, 1, 1);
585 MODULE_DEPEND(iflib, ether, 1, 1, 1);
586 
587 TASKQGROUP_DEFINE(if_io_tqg, mp_ncpus, 1);
588 TASKQGROUP_DEFINE(if_config_tqg, 1, 1);
589 
590 #ifndef IFLIB_DEBUG_COUNTERS
591 #ifdef INVARIANTS
592 #define IFLIB_DEBUG_COUNTERS 1
593 #else
594 #define IFLIB_DEBUG_COUNTERS 0
595 #endif /* !INVARIANTS */
596 #endif
597 
598 static SYSCTL_NODE(_net, OID_AUTO, iflib, CTLFLAG_RD, 0,
599                    "iflib driver parameters");
600 
601 /*
602  * XXX need to ensure that this can't accidentally cause the head to be moved backwards
603  */
604 static int iflib_min_tx_latency = 0;
605 SYSCTL_INT(_net_iflib, OID_AUTO, min_tx_latency, CTLFLAG_RW,
606 		   &iflib_min_tx_latency, 0, "minimize transmit latency at the possible expense of throughput");
607 static int iflib_no_tx_batch = 0;
608 SYSCTL_INT(_net_iflib, OID_AUTO, no_tx_batch, CTLFLAG_RW,
609 		   &iflib_no_tx_batch, 0, "minimize transmit latency at the possible expense of throughput");
610 
611 
612 #if IFLIB_DEBUG_COUNTERS
613 
614 static int iflib_tx_seen;
615 static int iflib_tx_sent;
616 static int iflib_tx_encap;
617 static int iflib_rx_allocs;
618 static int iflib_fl_refills;
619 static int iflib_fl_refills_large;
620 static int iflib_tx_frees;
621 
622 SYSCTL_INT(_net_iflib, OID_AUTO, tx_seen, CTLFLAG_RD,
623 		   &iflib_tx_seen, 0, "# tx mbufs seen");
624 SYSCTL_INT(_net_iflib, OID_AUTO, tx_sent, CTLFLAG_RD,
625 		   &iflib_tx_sent, 0, "# tx mbufs sent");
626 SYSCTL_INT(_net_iflib, OID_AUTO, tx_encap, CTLFLAG_RD,
627 		   &iflib_tx_encap, 0, "# tx mbufs encapped");
628 SYSCTL_INT(_net_iflib, OID_AUTO, tx_frees, CTLFLAG_RD,
629 		   &iflib_tx_frees, 0, "# tx frees");
630 SYSCTL_INT(_net_iflib, OID_AUTO, rx_allocs, CTLFLAG_RD,
631 		   &iflib_rx_allocs, 0, "# rx allocations");
632 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills, CTLFLAG_RD,
633 		   &iflib_fl_refills, 0, "# refills");
634 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills_large, CTLFLAG_RD,
635 		   &iflib_fl_refills_large, 0, "# large refills");
636 
637 
638 static int iflib_txq_drain_flushing;
639 static int iflib_txq_drain_oactive;
640 static int iflib_txq_drain_notready;
641 static int iflib_txq_drain_encapfail;
642 
643 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_flushing, CTLFLAG_RD,
644 		   &iflib_txq_drain_flushing, 0, "# drain flushes");
645 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_oactive, CTLFLAG_RD,
646 		   &iflib_txq_drain_oactive, 0, "# drain oactives");
647 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_notready, CTLFLAG_RD,
648 		   &iflib_txq_drain_notready, 0, "# drain notready");
649 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_encapfail, CTLFLAG_RD,
650 		   &iflib_txq_drain_encapfail, 0, "# drain encap fails");
651 
652 
653 static int iflib_encap_load_mbuf_fail;
654 static int iflib_encap_pad_mbuf_fail;
655 static int iflib_encap_txq_avail_fail;
656 static int iflib_encap_txd_encap_fail;
657 
658 SYSCTL_INT(_net_iflib, OID_AUTO, encap_load_mbuf_fail, CTLFLAG_RD,
659 		   &iflib_encap_load_mbuf_fail, 0, "# busdma load failures");
660 SYSCTL_INT(_net_iflib, OID_AUTO, encap_pad_mbuf_fail, CTLFLAG_RD,
661 		   &iflib_encap_pad_mbuf_fail, 0, "# runt frame pad failures");
662 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txq_avail_fail, CTLFLAG_RD,
663 		   &iflib_encap_txq_avail_fail, 0, "# txq avail failures");
664 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txd_encap_fail, CTLFLAG_RD,
665 		   &iflib_encap_txd_encap_fail, 0, "# driver encap failures");
666 
667 static int iflib_task_fn_rxs;
668 static int iflib_rx_intr_enables;
669 static int iflib_fast_intrs;
670 static int iflib_intr_link;
671 static int iflib_intr_msix;
672 static int iflib_rx_unavail;
673 static int iflib_rx_ctx_inactive;
674 static int iflib_rx_zero_len;
675 static int iflib_rx_if_input;
676 static int iflib_rx_mbuf_null;
677 static int iflib_rxd_flush;
678 
679 static int iflib_verbose_debug;
680 
681 SYSCTL_INT(_net_iflib, OID_AUTO, intr_link, CTLFLAG_RD,
682 		   &iflib_intr_link, 0, "# intr link calls");
683 SYSCTL_INT(_net_iflib, OID_AUTO, intr_msix, CTLFLAG_RD,
684 		   &iflib_intr_msix, 0, "# intr msix calls");
685 SYSCTL_INT(_net_iflib, OID_AUTO, task_fn_rx, CTLFLAG_RD,
686 		   &iflib_task_fn_rxs, 0, "# task_fn_rx calls");
687 SYSCTL_INT(_net_iflib, OID_AUTO, rx_intr_enables, CTLFLAG_RD,
688 		   &iflib_rx_intr_enables, 0, "# rx intr enables");
689 SYSCTL_INT(_net_iflib, OID_AUTO, fast_intrs, CTLFLAG_RD,
690 		   &iflib_fast_intrs, 0, "# fast_intr calls");
691 SYSCTL_INT(_net_iflib, OID_AUTO, rx_unavail, CTLFLAG_RD,
692 		   &iflib_rx_unavail, 0, "# times rxeof called with no available data");
693 SYSCTL_INT(_net_iflib, OID_AUTO, rx_ctx_inactive, CTLFLAG_RD,
694 		   &iflib_rx_ctx_inactive, 0, "# times rxeof called with inactive context");
695 SYSCTL_INT(_net_iflib, OID_AUTO, rx_zero_len, CTLFLAG_RD,
696 		   &iflib_rx_zero_len, 0, "# times rxeof saw zero len mbuf");
697 SYSCTL_INT(_net_iflib, OID_AUTO, rx_if_input, CTLFLAG_RD,
698 		   &iflib_rx_if_input, 0, "# times rxeof called if_input");
699 SYSCTL_INT(_net_iflib, OID_AUTO, rx_mbuf_null, CTLFLAG_RD,
700 		   &iflib_rx_mbuf_null, 0, "# times rxeof got null mbuf");
701 SYSCTL_INT(_net_iflib, OID_AUTO, rxd_flush, CTLFLAG_RD,
702 	         &iflib_rxd_flush, 0, "# times rxd_flush called");
703 SYSCTL_INT(_net_iflib, OID_AUTO, verbose_debug, CTLFLAG_RW,
704 		   &iflib_verbose_debug, 0, "enable verbose debugging");
705 
706 #define DBG_COUNTER_INC(name) atomic_add_int(&(iflib_ ## name), 1)
707 static void
708 iflib_debug_reset(void)
709 {
710 	iflib_tx_seen = iflib_tx_sent = iflib_tx_encap = iflib_rx_allocs =
711 		iflib_fl_refills = iflib_fl_refills_large = iflib_tx_frees =
712 		iflib_txq_drain_flushing = iflib_txq_drain_oactive =
713 		iflib_txq_drain_notready = iflib_txq_drain_encapfail =
714 		iflib_encap_load_mbuf_fail = iflib_encap_pad_mbuf_fail =
715 		iflib_encap_txq_avail_fail = iflib_encap_txd_encap_fail =
716 		iflib_task_fn_rxs = iflib_rx_intr_enables = iflib_fast_intrs =
717 		iflib_intr_link = iflib_intr_msix = iflib_rx_unavail =
718 		iflib_rx_ctx_inactive = iflib_rx_zero_len = iflib_rx_if_input =
719 		iflib_rx_mbuf_null = iflib_rxd_flush = 0;
720 }
721 
722 #else
723 #define DBG_COUNTER_INC(name)
724 static void iflib_debug_reset(void) {}
725 #endif
726 
727 #define IFLIB_DEBUG 0
728 
729 static void iflib_tx_structures_free(if_ctx_t ctx);
730 static void iflib_rx_structures_free(if_ctx_t ctx);
731 static int iflib_queues_alloc(if_ctx_t ctx);
732 static int iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq);
733 static int iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget);
734 static int iflib_qset_structures_setup(if_ctx_t ctx);
735 static int iflib_msix_init(if_ctx_t ctx);
736 static int iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filterarg, int *rid, char *str);
737 static void iflib_txq_check_drain(iflib_txq_t txq, int budget);
738 static uint32_t iflib_txq_can_drain(struct ifmp_ring *);
739 static int iflib_register(if_ctx_t);
740 static void iflib_init_locked(if_ctx_t ctx);
741 static void iflib_add_device_sysctl_pre(if_ctx_t ctx);
742 static void iflib_add_device_sysctl_post(if_ctx_t ctx);
743 static void iflib_ifmp_purge(iflib_txq_t txq);
744 static void _iflib_pre_assert(if_softc_ctx_t scctx);
745 static void iflib_if_init_locked(if_ctx_t ctx);
746 #ifndef __NO_STRICT_ALIGNMENT
747 static struct mbuf * iflib_fixup_rx(struct mbuf *m);
748 #endif
749 
750 NETDUMP_DEFINE(iflib);
751 
752 #ifdef DEV_NETMAP
753 #include <sys/selinfo.h>
754 #include <net/netmap.h>
755 #include <dev/netmap/netmap_kern.h>
756 
757 MODULE_DEPEND(iflib, netmap, 1, 1, 1);
758 
759 static int netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, uint32_t nm_i, bool init);
760 
761 /*
762  * device-specific sysctl variables:
763  *
764  * iflib_crcstrip: 0: keep CRC in rx frames (default), 1: strip it.
765  *	During regular operations the CRC is stripped, but on some
766  *	hardware reception of frames not multiple of 64 is slower,
767  *	so using crcstrip=0 helps in benchmarks.
768  *
769  * iflib_rx_miss, iflib_rx_miss_bufs:
770  *	count packets that might be missed due to lost interrupts.
771  */
772 SYSCTL_DECL(_dev_netmap);
773 /*
774  * The xl driver by default strips CRCs and we do not override it.
775  */
776 
777 int iflib_crcstrip = 1;
778 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_crcstrip,
779     CTLFLAG_RW, &iflib_crcstrip, 1, "strip CRC on rx frames");
780 
781 int iflib_rx_miss, iflib_rx_miss_bufs;
782 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss,
783     CTLFLAG_RW, &iflib_rx_miss, 0, "potentially missed rx intr");
784 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss_bufs,
785     CTLFLAG_RW, &iflib_rx_miss_bufs, 0, "potentially missed rx intr bufs");
786 
787 /*
788  * Register/unregister. We are already under netmap lock.
789  * Only called on the first register or the last unregister.
790  */
791 static int
792 iflib_netmap_register(struct netmap_adapter *na, int onoff)
793 {
794 	struct ifnet *ifp = na->ifp;
795 	if_ctx_t ctx = ifp->if_softc;
796 	int status;
797 
798 	CTX_LOCK(ctx);
799 	IFDI_INTR_DISABLE(ctx);
800 
801 	/* Tell the stack that the interface is no longer active */
802 	ifp->if_drv_flags &= ~(IFF_DRV_RUNNING | IFF_DRV_OACTIVE);
803 
804 	if (!CTX_IS_VF(ctx))
805 		IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip);
806 
807 	/* enable or disable flags and callbacks in na and ifp */
808 	if (onoff) {
809 		nm_set_native_flags(na);
810 	} else {
811 		nm_clear_native_flags(na);
812 	}
813 	iflib_stop(ctx);
814 	iflib_init_locked(ctx);
815 	IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip); // XXX why twice ?
816 	status = ifp->if_drv_flags & IFF_DRV_RUNNING ? 0 : 1;
817 	if (status)
818 		nm_clear_native_flags(na);
819 	CTX_UNLOCK(ctx);
820 	return (status);
821 }
822 
823 static int
824 netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, uint32_t nm_i, bool init)
825 {
826 	struct netmap_adapter *na = kring->na;
827 	u_int const lim = kring->nkr_num_slots - 1;
828 	u_int head = kring->rhead;
829 	struct netmap_ring *ring = kring->ring;
830 	bus_dmamap_t *map;
831 	struct if_rxd_update iru;
832 	if_ctx_t ctx = rxq->ifr_ctx;
833 	iflib_fl_t fl = &rxq->ifr_fl[0];
834 	uint32_t refill_pidx, nic_i;
835 
836 	if (nm_i == head && __predict_true(!init))
837 		return 0;
838 	iru_init(&iru, rxq, 0 /* flid */);
839 	map = fl->ifl_sds.ifsd_map;
840 	refill_pidx = netmap_idx_k2n(kring, nm_i);
841 	/*
842 	 * IMPORTANT: we must leave one free slot in the ring,
843 	 * so move head back by one unit
844 	 */
845 	head = nm_prev(head, lim);
846 	nic_i = UINT_MAX;
847 	while (nm_i != head) {
848 		for (int tmp_pidx = 0; tmp_pidx < IFLIB_MAX_RX_REFRESH && nm_i != head; tmp_pidx++) {
849 			struct netmap_slot *slot = &ring->slot[nm_i];
850 			void *addr = PNMB(na, slot, &fl->ifl_bus_addrs[tmp_pidx]);
851 			uint32_t nic_i_dma = refill_pidx;
852 			nic_i = netmap_idx_k2n(kring, nm_i);
853 
854 			MPASS(tmp_pidx < IFLIB_MAX_RX_REFRESH);
855 
856 			if (addr == NETMAP_BUF_BASE(na)) /* bad buf */
857 			        return netmap_ring_reinit(kring);
858 
859 			fl->ifl_vm_addrs[tmp_pidx] = addr;
860 			if (__predict_false(init) && map) {
861 				netmap_load_map(na, fl->ifl_ifdi->idi_tag, map[nic_i], addr);
862 			} else if (map && (slot->flags & NS_BUF_CHANGED)) {
863 				/* buffer has changed, reload map */
864 				netmap_reload_map(na, fl->ifl_ifdi->idi_tag, map[nic_i], addr);
865 			}
866 			slot->flags &= ~NS_BUF_CHANGED;
867 
868 			nm_i = nm_next(nm_i, lim);
869 			fl->ifl_rxd_idxs[tmp_pidx] = nic_i = nm_next(nic_i, lim);
870 			if (nm_i != head && tmp_pidx < IFLIB_MAX_RX_REFRESH-1)
871 				continue;
872 
873 			iru.iru_pidx = refill_pidx;
874 			iru.iru_count = tmp_pidx+1;
875 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
876 
877 			refill_pidx = nic_i;
878 			if (map == NULL)
879 				continue;
880 
881 			for (int n = 0; n < iru.iru_count; n++) {
882 				bus_dmamap_sync(fl->ifl_ifdi->idi_tag, map[nic_i_dma],
883 						BUS_DMASYNC_PREREAD);
884 				/* XXX - change this to not use the netmap func*/
885 				nic_i_dma = nm_next(nic_i_dma, lim);
886 			}
887 		}
888 	}
889 	kring->nr_hwcur = head;
890 
891 	if (map)
892 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
893 				BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
894 	if (__predict_true(nic_i != UINT_MAX))
895 		ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id, nic_i);
896 	return (0);
897 }
898 
899 /*
900  * Reconcile kernel and user view of the transmit ring.
901  *
902  * All information is in the kring.
903  * Userspace wants to send packets up to the one before kring->rhead,
904  * kernel knows kring->nr_hwcur is the first unsent packet.
905  *
906  * Here we push packets out (as many as possible), and possibly
907  * reclaim buffers from previously completed transmission.
908  *
909  * The caller (netmap) guarantees that there is only one instance
910  * running at any time. Any interference with other driver
911  * methods should be handled by the individual drivers.
912  */
913 static int
914 iflib_netmap_txsync(struct netmap_kring *kring, int flags)
915 {
916 	struct netmap_adapter *na = kring->na;
917 	struct ifnet *ifp = na->ifp;
918 	struct netmap_ring *ring = kring->ring;
919 	u_int nm_i;	/* index into the netmap ring */
920 	u_int nic_i;	/* index into the NIC ring */
921 	u_int n;
922 	u_int const lim = kring->nkr_num_slots - 1;
923 	u_int const head = kring->rhead;
924 	struct if_pkt_info pi;
925 
926 	/*
927 	 * interrupts on every tx packet are expensive so request
928 	 * them every half ring, or where NS_REPORT is set
929 	 */
930 	u_int report_frequency = kring->nkr_num_slots >> 1;
931 	/* device-specific */
932 	if_ctx_t ctx = ifp->if_softc;
933 	iflib_txq_t txq = &ctx->ifc_txqs[kring->ring_id];
934 
935 	if (txq->ift_sds.ifsd_map)
936 		bus_dmamap_sync(txq->ift_desc_tag, txq->ift_ifdi->idi_map,
937 				BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
938 
939 
940 	/*
941 	 * First part: process new packets to send.
942 	 * nm_i is the current index in the netmap ring,
943 	 * nic_i is the corresponding index in the NIC ring.
944 	 *
945 	 * If we have packets to send (nm_i != head)
946 	 * iterate over the netmap ring, fetch length and update
947 	 * the corresponding slot in the NIC ring. Some drivers also
948 	 * need to update the buffer's physical address in the NIC slot
949 	 * even NS_BUF_CHANGED is not set (PNMB computes the addresses).
950 	 *
951 	 * The netmap_reload_map() calls is especially expensive,
952 	 * even when (as in this case) the tag is 0, so do only
953 	 * when the buffer has actually changed.
954 	 *
955 	 * If possible do not set the report/intr bit on all slots,
956 	 * but only a few times per ring or when NS_REPORT is set.
957 	 *
958 	 * Finally, on 10G and faster drivers, it might be useful
959 	 * to prefetch the next slot and txr entry.
960 	 */
961 
962 	nm_i = netmap_idx_n2k(kring, kring->nr_hwcur);
963 	pkt_info_zero(&pi);
964 	pi.ipi_segs = txq->ift_segs;
965 	pi.ipi_qsidx = kring->ring_id;
966 	if (nm_i != head) {	/* we have new packets to send */
967 		nic_i = netmap_idx_k2n(kring, nm_i);
968 
969 		__builtin_prefetch(&ring->slot[nm_i]);
970 		__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i]);
971 		if (txq->ift_sds.ifsd_map)
972 			__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i]);
973 
974 		for (n = 0; nm_i != head; n++) {
975 			struct netmap_slot *slot = &ring->slot[nm_i];
976 			u_int len = slot->len;
977 			uint64_t paddr;
978 			void *addr = PNMB(na, slot, &paddr);
979 			int flags = (slot->flags & NS_REPORT ||
980 				nic_i == 0 || nic_i == report_frequency) ?
981 				IPI_TX_INTR : 0;
982 
983 			/* device-specific */
984 			pi.ipi_len = len;
985 			pi.ipi_segs[0].ds_addr = paddr;
986 			pi.ipi_segs[0].ds_len = len;
987 			pi.ipi_nsegs = 1;
988 			pi.ipi_ndescs = 0;
989 			pi.ipi_pidx = nic_i;
990 			pi.ipi_flags = flags;
991 
992 			/* Fill the slot in the NIC ring. */
993 			ctx->isc_txd_encap(ctx->ifc_softc, &pi);
994 
995 			/* prefetch for next round */
996 			__builtin_prefetch(&ring->slot[nm_i + 1]);
997 			__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i + 1]);
998 			if (txq->ift_sds.ifsd_map) {
999 				__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i + 1]);
1000 
1001 				NM_CHECK_ADDR_LEN(na, addr, len);
1002 
1003 				if (slot->flags & NS_BUF_CHANGED) {
1004 					/* buffer has changed, reload map */
1005 					netmap_reload_map(na, txq->ift_desc_tag, txq->ift_sds.ifsd_map[nic_i], addr);
1006 				}
1007 				/* make sure changes to the buffer are synced */
1008 				bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_sds.ifsd_map[nic_i],
1009 						BUS_DMASYNC_PREWRITE);
1010 			}
1011 			slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED);
1012 			nm_i = nm_next(nm_i, lim);
1013 			nic_i = nm_next(nic_i, lim);
1014 		}
1015 		kring->nr_hwcur = head;
1016 
1017 		/* synchronize the NIC ring */
1018 		if (txq->ift_sds.ifsd_map)
1019 			bus_dmamap_sync(txq->ift_desc_tag, txq->ift_ifdi->idi_map,
1020 						BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1021 
1022 		/* (re)start the tx unit up to slot nic_i (excluded) */
1023 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, nic_i);
1024 	}
1025 
1026 	/*
1027 	 * Second part: reclaim buffers for completed transmissions.
1028 	 */
1029 	if (iflib_tx_credits_update(ctx, txq)) {
1030 		/* some tx completed, increment avail */
1031 		nic_i = txq->ift_cidx_processed;
1032 		kring->nr_hwtail = nm_prev(netmap_idx_n2k(kring, nic_i), lim);
1033 	}
1034 	return (0);
1035 }
1036 
1037 /*
1038  * Reconcile kernel and user view of the receive ring.
1039  * Same as for the txsync, this routine must be efficient.
1040  * The caller guarantees a single invocations, but races against
1041  * the rest of the driver should be handled here.
1042  *
1043  * On call, kring->rhead is the first packet that userspace wants
1044  * to keep, and kring->rcur is the wakeup point.
1045  * The kernel has previously reported packets up to kring->rtail.
1046  *
1047  * If (flags & NAF_FORCE_READ) also check for incoming packets irrespective
1048  * of whether or not we received an interrupt.
1049  */
1050 static int
1051 iflib_netmap_rxsync(struct netmap_kring *kring, int flags)
1052 {
1053 	struct netmap_adapter *na = kring->na;
1054 	struct netmap_ring *ring = kring->ring;
1055 	uint32_t nm_i;	/* index into the netmap ring */
1056 	uint32_t nic_i;	/* index into the NIC ring */
1057 	u_int i, n;
1058 	u_int const lim = kring->nkr_num_slots - 1;
1059 	u_int const head = netmap_idx_n2k(kring, kring->rhead);
1060 	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
1061 	struct if_rxd_info ri;
1062 
1063 	struct ifnet *ifp = na->ifp;
1064 	if_ctx_t ctx = ifp->if_softc;
1065 	iflib_rxq_t rxq = &ctx->ifc_rxqs[kring->ring_id];
1066 	iflib_fl_t fl = rxq->ifr_fl;
1067 	if (head > lim)
1068 		return netmap_ring_reinit(kring);
1069 
1070 	/* XXX check sync modes */
1071 	for (i = 0, fl = rxq->ifr_fl; i < rxq->ifr_nfl; i++, fl++) {
1072 		if (fl->ifl_sds.ifsd_map == NULL)
1073 			continue;
1074 		bus_dmamap_sync(rxq->ifr_fl[i].ifl_desc_tag, fl->ifl_ifdi->idi_map,
1075 				BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1076 	}
1077 	/*
1078 	 * First part: import newly received packets.
1079 	 *
1080 	 * nm_i is the index of the next free slot in the netmap ring,
1081 	 * nic_i is the index of the next received packet in the NIC ring,
1082 	 * and they may differ in case if_init() has been called while
1083 	 * in netmap mode. For the receive ring we have
1084 	 *
1085 	 *	nic_i = rxr->next_check;
1086 	 *	nm_i = kring->nr_hwtail (previous)
1087 	 * and
1088 	 *	nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1089 	 *
1090 	 * rxr->next_check is set to 0 on a ring reinit
1091 	 */
1092 	if (netmap_no_pendintr || force_update) {
1093 		int crclen = iflib_crcstrip ? 0 : 4;
1094 		int error, avail;
1095 
1096 		for (i = 0; i < rxq->ifr_nfl; i++) {
1097 			fl = &rxq->ifr_fl[i];
1098 			nic_i = fl->ifl_cidx;
1099 			nm_i = netmap_idx_n2k(kring, nic_i);
1100 			avail = iflib_rxd_avail(ctx, rxq, nic_i, USHRT_MAX);
1101 			for (n = 0; avail > 0; n++, avail--) {
1102 				rxd_info_zero(&ri);
1103 				ri.iri_frags = rxq->ifr_frags;
1104 				ri.iri_qsidx = kring->ring_id;
1105 				ri.iri_ifp = ctx->ifc_ifp;
1106 				ri.iri_cidx = nic_i;
1107 
1108 				error = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
1109 				ring->slot[nm_i].len = error ? 0 : ri.iri_len - crclen;
1110 				ring->slot[nm_i].flags = 0;
1111 				if (fl->ifl_sds.ifsd_map)
1112 					bus_dmamap_sync(fl->ifl_ifdi->idi_tag,
1113 							fl->ifl_sds.ifsd_map[nic_i], BUS_DMASYNC_POSTREAD);
1114 				nm_i = nm_next(nm_i, lim);
1115 				nic_i = nm_next(nic_i, lim);
1116 			}
1117 			if (n) { /* update the state variables */
1118 				if (netmap_no_pendintr && !force_update) {
1119 					/* diagnostics */
1120 					iflib_rx_miss ++;
1121 					iflib_rx_miss_bufs += n;
1122 				}
1123 				fl->ifl_cidx = nic_i;
1124 				kring->nr_hwtail = netmap_idx_k2n(kring, nm_i);
1125 			}
1126 			kring->nr_kflags &= ~NKR_PENDINTR;
1127 		}
1128 	}
1129 	/*
1130 	 * Second part: skip past packets that userspace has released.
1131 	 * (kring->nr_hwcur to head excluded),
1132 	 * and make the buffers available for reception.
1133 	 * As usual nm_i is the index in the netmap ring,
1134 	 * nic_i is the index in the NIC ring, and
1135 	 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1136 	 */
1137 	/* XXX not sure how this will work with multiple free lists */
1138 	nm_i = netmap_idx_n2k(kring, kring->nr_hwcur);
1139 
1140 	return (netmap_fl_refill(rxq, kring, nm_i, false));
1141 }
1142 
1143 static void
1144 iflib_netmap_intr(struct netmap_adapter *na, int onoff)
1145 {
1146 	struct ifnet *ifp = na->ifp;
1147 	if_ctx_t ctx = ifp->if_softc;
1148 
1149 	CTX_LOCK(ctx);
1150 	if (onoff) {
1151 		IFDI_INTR_ENABLE(ctx);
1152 	} else {
1153 		IFDI_INTR_DISABLE(ctx);
1154 	}
1155 	CTX_UNLOCK(ctx);
1156 }
1157 
1158 
1159 static int
1160 iflib_netmap_attach(if_ctx_t ctx)
1161 {
1162 	struct netmap_adapter na;
1163 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1164 
1165 	bzero(&na, sizeof(na));
1166 
1167 	na.ifp = ctx->ifc_ifp;
1168 	na.na_flags = NAF_BDG_MAYSLEEP;
1169 	MPASS(ctx->ifc_softc_ctx.isc_ntxqsets);
1170 	MPASS(ctx->ifc_softc_ctx.isc_nrxqsets);
1171 
1172 	na.num_tx_desc = scctx->isc_ntxd[0];
1173 	na.num_rx_desc = scctx->isc_nrxd[0];
1174 	na.nm_txsync = iflib_netmap_txsync;
1175 	na.nm_rxsync = iflib_netmap_rxsync;
1176 	na.nm_register = iflib_netmap_register;
1177 	na.nm_intr = iflib_netmap_intr;
1178 	na.num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
1179 	na.num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
1180 	return (netmap_attach(&na));
1181 }
1182 
1183 static void
1184 iflib_netmap_txq_init(if_ctx_t ctx, iflib_txq_t txq)
1185 {
1186 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1187 	struct netmap_slot *slot;
1188 
1189 	slot = netmap_reset(na, NR_TX, txq->ift_id, 0);
1190 	if (slot == NULL)
1191 		return;
1192 	if (txq->ift_sds.ifsd_map == NULL)
1193 		return;
1194 
1195 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxd[0]; i++) {
1196 
1197 		/*
1198 		 * In netmap mode, set the map for the packet buffer.
1199 		 * NOTE: Some drivers (not this one) also need to set
1200 		 * the physical buffer address in the NIC ring.
1201 		 * netmap_idx_n2k() maps a nic index, i, into the corresponding
1202 		 * netmap slot index, si
1203 		 */
1204 		int si = netmap_idx_n2k(na->tx_rings[txq->ift_id], i);
1205 		netmap_load_map(na, txq->ift_desc_tag, txq->ift_sds.ifsd_map[i], NMB(na, slot + si));
1206 	}
1207 }
1208 
1209 static void
1210 iflib_netmap_rxq_init(if_ctx_t ctx, iflib_rxq_t rxq)
1211 {
1212 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1213 	struct netmap_kring *kring = na->rx_rings[rxq->ifr_id];
1214 	struct netmap_slot *slot;
1215 	uint32_t nm_i;
1216 
1217 	slot = netmap_reset(na, NR_RX, rxq->ifr_id, 0);
1218 	if (slot == NULL)
1219 		return;
1220 	nm_i = netmap_idx_n2k(kring, 0);
1221 	netmap_fl_refill(rxq, kring, nm_i, true);
1222 }
1223 
1224 #define iflib_netmap_detach(ifp) netmap_detach(ifp)
1225 
1226 #else
1227 #define iflib_netmap_txq_init(ctx, txq)
1228 #define iflib_netmap_rxq_init(ctx, rxq)
1229 #define iflib_netmap_detach(ifp)
1230 
1231 #define iflib_netmap_attach(ctx) (0)
1232 #define netmap_rx_irq(ifp, qid, budget) (0)
1233 #define netmap_tx_irq(ifp, qid) do {} while (0)
1234 
1235 #endif
1236 
1237 #if defined(__i386__) || defined(__amd64__)
1238 static __inline void
1239 prefetch(void *x)
1240 {
1241 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1242 }
1243 static __inline void
1244 prefetch2cachelines(void *x)
1245 {
1246 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1247 #if (CACHE_LINE_SIZE < 128)
1248 	__asm volatile("prefetcht0 %0" :: "m" (*(((unsigned long *)x)+CACHE_LINE_SIZE/(sizeof(unsigned long)))));
1249 #endif
1250 }
1251 #else
1252 #define prefetch(x)
1253 #define prefetch2cachelines(x)
1254 #endif
1255 
1256 static void
1257 iflib_gen_mac(if_ctx_t ctx)
1258 {
1259 	struct thread *td;
1260 	struct ifnet *ifp;
1261 	MD5_CTX mdctx;
1262 	char uuid[HOSTUUIDLEN+1];
1263 	char buf[HOSTUUIDLEN+16];
1264 	uint8_t *mac;
1265 	unsigned char digest[16];
1266 
1267 	td = curthread;
1268 	ifp = ctx->ifc_ifp;
1269 	mac = ctx->ifc_mac;
1270 	uuid[HOSTUUIDLEN] = 0;
1271 	bcopy(td->td_ucred->cr_prison->pr_hostuuid, uuid, HOSTUUIDLEN);
1272 	snprintf(buf, HOSTUUIDLEN+16, "%s-%s", uuid, device_get_nameunit(ctx->ifc_dev));
1273 	/*
1274 	 * Generate a pseudo-random, deterministic MAC
1275 	 * address based on the UUID and unit number.
1276 	 * The FreeBSD Foundation OUI of 58-9C-FC is used.
1277 	 */
1278 	MD5Init(&mdctx);
1279 	MD5Update(&mdctx, buf, strlen(buf));
1280 	MD5Final(digest, &mdctx);
1281 
1282 	mac[0] = 0x58;
1283 	mac[1] = 0x9C;
1284 	mac[2] = 0xFC;
1285 	mac[3] = digest[0];
1286 	mac[4] = digest[1];
1287 	mac[5] = digest[2];
1288 }
1289 
1290 static void
1291 iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid)
1292 {
1293 	iflib_fl_t fl;
1294 
1295 	fl = &rxq->ifr_fl[flid];
1296 	iru->iru_paddrs = fl->ifl_bus_addrs;
1297 	iru->iru_vaddrs = &fl->ifl_vm_addrs[0];
1298 	iru->iru_idxs = fl->ifl_rxd_idxs;
1299 	iru->iru_qsidx = rxq->ifr_id;
1300 	iru->iru_buf_size = fl->ifl_buf_size;
1301 	iru->iru_flidx = fl->ifl_id;
1302 }
1303 
1304 static void
1305 _iflib_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int err)
1306 {
1307 	if (err)
1308 		return;
1309 	*(bus_addr_t *) arg = segs[0].ds_addr;
1310 }
1311 
1312 int
1313 iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags)
1314 {
1315 	int err;
1316 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1317 	device_t dev = ctx->ifc_dev;
1318 
1319 	KASSERT(sctx->isc_q_align != 0, ("alignment value not initialized"));
1320 
1321 	err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1322 				sctx->isc_q_align, 0,	/* alignment, bounds */
1323 				BUS_SPACE_MAXADDR,	/* lowaddr */
1324 				BUS_SPACE_MAXADDR,	/* highaddr */
1325 				NULL, NULL,		/* filter, filterarg */
1326 				size,			/* maxsize */
1327 				1,			/* nsegments */
1328 				size,			/* maxsegsize */
1329 				BUS_DMA_ALLOCNOW,	/* flags */
1330 				NULL,			/* lockfunc */
1331 				NULL,			/* lockarg */
1332 				&dma->idi_tag);
1333 	if (err) {
1334 		device_printf(dev,
1335 		    "%s: bus_dma_tag_create failed: %d\n",
1336 		    __func__, err);
1337 		goto fail_0;
1338 	}
1339 
1340 	err = bus_dmamem_alloc(dma->idi_tag, (void**) &dma->idi_vaddr,
1341 	    BUS_DMA_NOWAIT | BUS_DMA_COHERENT | BUS_DMA_ZERO, &dma->idi_map);
1342 	if (err) {
1343 		device_printf(dev,
1344 		    "%s: bus_dmamem_alloc(%ju) failed: %d\n",
1345 		    __func__, (uintmax_t)size, err);
1346 		goto fail_1;
1347 	}
1348 
1349 	dma->idi_paddr = IF_BAD_DMA;
1350 	err = bus_dmamap_load(dma->idi_tag, dma->idi_map, dma->idi_vaddr,
1351 	    size, _iflib_dmamap_cb, &dma->idi_paddr, mapflags | BUS_DMA_NOWAIT);
1352 	if (err || dma->idi_paddr == IF_BAD_DMA) {
1353 		device_printf(dev,
1354 		    "%s: bus_dmamap_load failed: %d\n",
1355 		    __func__, err);
1356 		goto fail_2;
1357 	}
1358 
1359 	dma->idi_size = size;
1360 	return (0);
1361 
1362 fail_2:
1363 	bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1364 fail_1:
1365 	bus_dma_tag_destroy(dma->idi_tag);
1366 fail_0:
1367 	dma->idi_tag = NULL;
1368 
1369 	return (err);
1370 }
1371 
1372 int
1373 iflib_dma_alloc_multi(if_ctx_t ctx, int *sizes, iflib_dma_info_t *dmalist, int mapflags, int count)
1374 {
1375 	int i, err;
1376 	iflib_dma_info_t *dmaiter;
1377 
1378 	dmaiter = dmalist;
1379 	for (i = 0; i < count; i++, dmaiter++) {
1380 		if ((err = iflib_dma_alloc(ctx, sizes[i], *dmaiter, mapflags)) != 0)
1381 			break;
1382 	}
1383 	if (err)
1384 		iflib_dma_free_multi(dmalist, i);
1385 	return (err);
1386 }
1387 
1388 void
1389 iflib_dma_free(iflib_dma_info_t dma)
1390 {
1391 	if (dma->idi_tag == NULL)
1392 		return;
1393 	if (dma->idi_paddr != IF_BAD_DMA) {
1394 		bus_dmamap_sync(dma->idi_tag, dma->idi_map,
1395 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1396 		bus_dmamap_unload(dma->idi_tag, dma->idi_map);
1397 		dma->idi_paddr = IF_BAD_DMA;
1398 	}
1399 	if (dma->idi_vaddr != NULL) {
1400 		bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1401 		dma->idi_vaddr = NULL;
1402 	}
1403 	bus_dma_tag_destroy(dma->idi_tag);
1404 	dma->idi_tag = NULL;
1405 }
1406 
1407 void
1408 iflib_dma_free_multi(iflib_dma_info_t *dmalist, int count)
1409 {
1410 	int i;
1411 	iflib_dma_info_t *dmaiter = dmalist;
1412 
1413 	for (i = 0; i < count; i++, dmaiter++)
1414 		iflib_dma_free(*dmaiter);
1415 }
1416 
1417 #ifdef EARLY_AP_STARTUP
1418 static const int iflib_started = 1;
1419 #else
1420 /*
1421  * We used to abuse the smp_started flag to decide if the queues have been
1422  * fully initialized (by late taskqgroup_adjust() calls in a SYSINIT()).
1423  * That gave bad races, since the SYSINIT() runs strictly after smp_started
1424  * is set.  Run a SYSINIT() strictly after that to just set a usable
1425  * completion flag.
1426  */
1427 
1428 static int iflib_started;
1429 
1430 static void
1431 iflib_record_started(void *arg)
1432 {
1433 	iflib_started = 1;
1434 }
1435 
1436 SYSINIT(iflib_record_started, SI_SUB_SMP + 1, SI_ORDER_FIRST,
1437 	iflib_record_started, NULL);
1438 #endif
1439 
1440 static int
1441 iflib_fast_intr(void *arg)
1442 {
1443 	iflib_filter_info_t info = arg;
1444 	struct grouptask *gtask = info->ifi_task;
1445 	if (!iflib_started)
1446 		return (FILTER_HANDLED);
1447 
1448 	DBG_COUNTER_INC(fast_intrs);
1449 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1450 		return (FILTER_HANDLED);
1451 
1452 	GROUPTASK_ENQUEUE(gtask);
1453 	return (FILTER_HANDLED);
1454 }
1455 
1456 static int
1457 iflib_fast_intr_rxtx(void *arg)
1458 {
1459 	iflib_filter_info_t info = arg;
1460 	struct grouptask *gtask = info->ifi_task;
1461 	iflib_rxq_t rxq = (iflib_rxq_t)info->ifi_ctx;
1462 	if_ctx_t ctx = NULL;;
1463 	int i, cidx;
1464 
1465 	if (!iflib_started)
1466 		return (FILTER_HANDLED);
1467 
1468 	DBG_COUNTER_INC(fast_intrs);
1469 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1470 		return (FILTER_HANDLED);
1471 
1472 	MPASS(rxq->ifr_ntxqirq);
1473 	for (i = 0; i < rxq->ifr_ntxqirq; i++) {
1474 		qidx_t txqid = rxq->ifr_txqid[i];
1475 
1476 		ctx = rxq->ifr_ctx;
1477 
1478 		if (!ctx->isc_txd_credits_update(ctx->ifc_softc, txqid, false)) {
1479 			IFDI_TX_QUEUE_INTR_ENABLE(ctx, txqid);
1480 			continue;
1481 		}
1482 		GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
1483 	}
1484 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_RXCQ)
1485 		cidx = rxq->ifr_cq_cidx;
1486 	else
1487 		cidx = rxq->ifr_fl[0].ifl_cidx;
1488 	if (iflib_rxd_avail(ctx, rxq, cidx, 1))
1489 		GROUPTASK_ENQUEUE(gtask);
1490 	else
1491 		IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
1492 	return (FILTER_HANDLED);
1493 }
1494 
1495 
1496 static int
1497 iflib_fast_intr_ctx(void *arg)
1498 {
1499 	iflib_filter_info_t info = arg;
1500 	struct grouptask *gtask = info->ifi_task;
1501 
1502 	if (!iflib_started)
1503 		return (FILTER_HANDLED);
1504 
1505 	DBG_COUNTER_INC(fast_intrs);
1506 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1507 		return (FILTER_HANDLED);
1508 
1509 	GROUPTASK_ENQUEUE(gtask);
1510 	return (FILTER_HANDLED);
1511 }
1512 
1513 static int
1514 _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
1515 	driver_filter_t filter, driver_intr_t handler, void *arg,
1516 				 char *name)
1517 {
1518 	int rc, flags;
1519 	struct resource *res;
1520 	void *tag = NULL;
1521 	device_t dev = ctx->ifc_dev;
1522 
1523 	flags = RF_ACTIVE;
1524 	if (ctx->ifc_flags & IFC_LEGACY)
1525 		flags |= RF_SHAREABLE;
1526 	MPASS(rid < 512);
1527 	irq->ii_rid = rid;
1528 	res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &irq->ii_rid, flags);
1529 	if (res == NULL) {
1530 		device_printf(dev,
1531 		    "failed to allocate IRQ for rid %d, name %s.\n", rid, name);
1532 		return (ENOMEM);
1533 	}
1534 	irq->ii_res = res;
1535 	KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL"));
1536 	rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET,
1537 						filter, handler, arg, &tag);
1538 	if (rc != 0) {
1539 		device_printf(dev,
1540 		    "failed to setup interrupt for rid %d, name %s: %d\n",
1541 					  rid, name ? name : "unknown", rc);
1542 		return (rc);
1543 	} else if (name)
1544 		bus_describe_intr(dev, res, tag, "%s", name);
1545 
1546 	irq->ii_tag = tag;
1547 	return (0);
1548 }
1549 
1550 
1551 /*********************************************************************
1552  *
1553  *  Allocate memory for tx_buffer structures. The tx_buffer stores all
1554  *  the information needed to transmit a packet on the wire. This is
1555  *  called only once at attach, setup is done every reset.
1556  *
1557  **********************************************************************/
1558 
1559 static int
1560 iflib_txsd_alloc(iflib_txq_t txq)
1561 {
1562 	if_ctx_t ctx = txq->ift_ctx;
1563 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1564 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1565 	device_t dev = ctx->ifc_dev;
1566 	int err, nsegments, ntsosegments;
1567 
1568 	nsegments = scctx->isc_tx_nsegments;
1569 	ntsosegments = scctx->isc_tx_tso_segments_max;
1570 	MPASS(scctx->isc_ntxd[0] > 0);
1571 	MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0);
1572 	MPASS(nsegments > 0);
1573 	MPASS(ntsosegments > 0);
1574 	/*
1575 	 * Setup DMA descriptor areas.
1576 	 */
1577 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1578 			       1, 0,			/* alignment, bounds */
1579 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1580 			       BUS_SPACE_MAXADDR,	/* highaddr */
1581 			       NULL, NULL,		/* filter, filterarg */
1582 			       sctx->isc_tx_maxsize,		/* maxsize */
1583 			       nsegments,	/* nsegments */
1584 			       sctx->isc_tx_maxsegsize,	/* maxsegsize */
1585 			       0,			/* flags */
1586 			       NULL,			/* lockfunc */
1587 			       NULL,			/* lockfuncarg */
1588 			       &txq->ift_desc_tag))) {
1589 		device_printf(dev,"Unable to allocate TX DMA tag: %d\n", err);
1590 		device_printf(dev,"maxsize: %ju nsegments: %d maxsegsize: %ju\n",
1591 		    (uintmax_t)sctx->isc_tx_maxsize, nsegments, (uintmax_t)sctx->isc_tx_maxsegsize);
1592 		goto fail;
1593 	}
1594 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1595 			       1, 0,			/* alignment, bounds */
1596 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1597 			       BUS_SPACE_MAXADDR,	/* highaddr */
1598 			       NULL, NULL,		/* filter, filterarg */
1599 			       scctx->isc_tx_tso_size_max,		/* maxsize */
1600 			       ntsosegments,	/* nsegments */
1601 			       scctx->isc_tx_tso_segsize_max,	/* maxsegsize */
1602 			       0,			/* flags */
1603 			       NULL,			/* lockfunc */
1604 			       NULL,			/* lockfuncarg */
1605 			       &txq->ift_tso_desc_tag))) {
1606 		device_printf(dev,"Unable to allocate TX TSO DMA tag: %d\n", err);
1607 
1608 		goto fail;
1609 	}
1610 	if (!(txq->ift_sds.ifsd_flags =
1611 	    (uint8_t *) malloc(sizeof(uint8_t) *
1612 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1613 		device_printf(dev, "Unable to allocate tx_buffer memory\n");
1614 		err = ENOMEM;
1615 		goto fail;
1616 	}
1617 	if (!(txq->ift_sds.ifsd_m =
1618 	    (struct mbuf **) malloc(sizeof(struct mbuf *) *
1619 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1620 		device_printf(dev, "Unable to allocate tx_buffer memory\n");
1621 		err = ENOMEM;
1622 		goto fail;
1623 	}
1624 
1625         /* Create the descriptor buffer dma maps */
1626 #if defined(ACPI_DMAR) || (! (defined(__i386__) || defined(__amd64__)))
1627 	if ((ctx->ifc_flags & IFC_DMAR) == 0)
1628 		return (0);
1629 
1630 	if (!(txq->ift_sds.ifsd_map =
1631 	    (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1632 		device_printf(dev, "Unable to allocate tx_buffer map memory\n");
1633 		err = ENOMEM;
1634 		goto fail;
1635 	}
1636 
1637 	for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) {
1638 		err = bus_dmamap_create(txq->ift_desc_tag, 0, &txq->ift_sds.ifsd_map[i]);
1639 		if (err != 0) {
1640 			device_printf(dev, "Unable to create TX DMA map\n");
1641 			goto fail;
1642 		}
1643 	}
1644 #endif
1645 	return (0);
1646 fail:
1647 	/* We free all, it handles case where we are in the middle */
1648 	iflib_tx_structures_free(ctx);
1649 	return (err);
1650 }
1651 
1652 static void
1653 iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i)
1654 {
1655 	bus_dmamap_t map;
1656 
1657 	map = NULL;
1658 	if (txq->ift_sds.ifsd_map != NULL)
1659 		map = txq->ift_sds.ifsd_map[i];
1660 	if (map != NULL) {
1661 		bus_dmamap_unload(txq->ift_desc_tag, map);
1662 		bus_dmamap_destroy(txq->ift_desc_tag, map);
1663 		txq->ift_sds.ifsd_map[i] = NULL;
1664 	}
1665 }
1666 
1667 static void
1668 iflib_txq_destroy(iflib_txq_t txq)
1669 {
1670 	if_ctx_t ctx = txq->ift_ctx;
1671 
1672 	for (int i = 0; i < txq->ift_size; i++)
1673 		iflib_txsd_destroy(ctx, txq, i);
1674 	if (txq->ift_sds.ifsd_map != NULL) {
1675 		free(txq->ift_sds.ifsd_map, M_IFLIB);
1676 		txq->ift_sds.ifsd_map = NULL;
1677 	}
1678 	if (txq->ift_sds.ifsd_m != NULL) {
1679 		free(txq->ift_sds.ifsd_m, M_IFLIB);
1680 		txq->ift_sds.ifsd_m = NULL;
1681 	}
1682 	if (txq->ift_sds.ifsd_flags != NULL) {
1683 		free(txq->ift_sds.ifsd_flags, M_IFLIB);
1684 		txq->ift_sds.ifsd_flags = NULL;
1685 	}
1686 	if (txq->ift_desc_tag != NULL) {
1687 		bus_dma_tag_destroy(txq->ift_desc_tag);
1688 		txq->ift_desc_tag = NULL;
1689 	}
1690 	if (txq->ift_tso_desc_tag != NULL) {
1691 		bus_dma_tag_destroy(txq->ift_tso_desc_tag);
1692 		txq->ift_tso_desc_tag = NULL;
1693 	}
1694 }
1695 
1696 static void
1697 iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i)
1698 {
1699 	struct mbuf **mp;
1700 
1701 	mp = &txq->ift_sds.ifsd_m[i];
1702 	if (*mp == NULL)
1703 		return;
1704 
1705 	if (txq->ift_sds.ifsd_map != NULL) {
1706 		bus_dmamap_sync(txq->ift_desc_tag,
1707 				txq->ift_sds.ifsd_map[i],
1708 				BUS_DMASYNC_POSTWRITE);
1709 		bus_dmamap_unload(txq->ift_desc_tag,
1710 				  txq->ift_sds.ifsd_map[i]);
1711 	}
1712 	m_free(*mp);
1713 	DBG_COUNTER_INC(tx_frees);
1714 	*mp = NULL;
1715 }
1716 
1717 static int
1718 iflib_txq_setup(iflib_txq_t txq)
1719 {
1720 	if_ctx_t ctx = txq->ift_ctx;
1721 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1722 	iflib_dma_info_t di;
1723 	int i;
1724 
1725 	/* Set number of descriptors available */
1726 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
1727 	/* XXX make configurable */
1728 	txq->ift_update_freq = IFLIB_DEFAULT_TX_UPDATE_FREQ;
1729 
1730 	/* Reset indices */
1731 	txq->ift_cidx_processed = 0;
1732 	txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0;
1733 	txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset];
1734 
1735 	for (i = 0, di = txq->ift_ifdi; i < ctx->ifc_nhwtxqs; i++, di++)
1736 		bzero((void *)di->idi_vaddr, di->idi_size);
1737 
1738 	IFDI_TXQ_SETUP(ctx, txq->ift_id);
1739 	for (i = 0, di = txq->ift_ifdi; i < ctx->ifc_nhwtxqs; i++, di++)
1740 		bus_dmamap_sync(di->idi_tag, di->idi_map,
1741 						BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1742 	return (0);
1743 }
1744 
1745 /*********************************************************************
1746  *
1747  *  Allocate memory for rx_buffer structures. Since we use one
1748  *  rx_buffer per received packet, the maximum number of rx_buffer's
1749  *  that we'll need is equal to the number of receive descriptors
1750  *  that we've allocated.
1751  *
1752  **********************************************************************/
1753 static int
1754 iflib_rxsd_alloc(iflib_rxq_t rxq)
1755 {
1756 	if_ctx_t ctx = rxq->ifr_ctx;
1757 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1758 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1759 	device_t dev = ctx->ifc_dev;
1760 	iflib_fl_t fl;
1761 	int			err;
1762 
1763 	MPASS(scctx->isc_nrxd[0] > 0);
1764 	MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0);
1765 
1766 	fl = rxq->ifr_fl;
1767 	for (int i = 0; i <  rxq->ifr_nfl; i++, fl++) {
1768 		fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */
1769 		err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1770 					 1, 0,			/* alignment, bounds */
1771 					 BUS_SPACE_MAXADDR,	/* lowaddr */
1772 					 BUS_SPACE_MAXADDR,	/* highaddr */
1773 					 NULL, NULL,		/* filter, filterarg */
1774 					 sctx->isc_rx_maxsize,	/* maxsize */
1775 					 sctx->isc_rx_nsegments,	/* nsegments */
1776 					 sctx->isc_rx_maxsegsize,	/* maxsegsize */
1777 					 0,			/* flags */
1778 					 NULL,			/* lockfunc */
1779 					 NULL,			/* lockarg */
1780 					 &fl->ifl_desc_tag);
1781 		if (err) {
1782 			device_printf(dev, "%s: bus_dma_tag_create failed %d\n",
1783 				__func__, err);
1784 			goto fail;
1785 		}
1786 		if (!(fl->ifl_sds.ifsd_flags =
1787 		      (uint8_t *) malloc(sizeof(uint8_t) *
1788 					 scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1789 			device_printf(dev, "Unable to allocate tx_buffer memory\n");
1790 			err = ENOMEM;
1791 			goto fail;
1792 		}
1793 		if (!(fl->ifl_sds.ifsd_m =
1794 		      (struct mbuf **) malloc(sizeof(struct mbuf *) *
1795 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1796 			device_printf(dev, "Unable to allocate tx_buffer memory\n");
1797 			err = ENOMEM;
1798 			goto fail;
1799 		}
1800 		if (!(fl->ifl_sds.ifsd_cl =
1801 		      (caddr_t *) malloc(sizeof(caddr_t) *
1802 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1803 			device_printf(dev, "Unable to allocate tx_buffer memory\n");
1804 			err = ENOMEM;
1805 			goto fail;
1806 		}
1807 
1808 		/* Create the descriptor buffer dma maps */
1809 #if defined(ACPI_DMAR) || (! (defined(__i386__) || defined(__amd64__)))
1810 		if ((ctx->ifc_flags & IFC_DMAR) == 0)
1811 			continue;
1812 
1813 		if (!(fl->ifl_sds.ifsd_map =
1814 		      (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1815 			device_printf(dev, "Unable to allocate tx_buffer map memory\n");
1816 			err = ENOMEM;
1817 			goto fail;
1818 		}
1819 
1820 		for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++) {
1821 			err = bus_dmamap_create(fl->ifl_desc_tag, 0, &fl->ifl_sds.ifsd_map[i]);
1822 			if (err != 0) {
1823 				device_printf(dev, "Unable to create RX buffer DMA map\n");
1824 				goto fail;
1825 			}
1826 		}
1827 #endif
1828 	}
1829 	return (0);
1830 
1831 fail:
1832 	iflib_rx_structures_free(ctx);
1833 	return (err);
1834 }
1835 
1836 
1837 /*
1838  * Internal service routines
1839  */
1840 
1841 struct rxq_refill_cb_arg {
1842 	int               error;
1843 	bus_dma_segment_t seg;
1844 	int               nseg;
1845 };
1846 
1847 static void
1848 _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
1849 {
1850 	struct rxq_refill_cb_arg *cb_arg = arg;
1851 
1852 	cb_arg->error = error;
1853 	cb_arg->seg = segs[0];
1854 	cb_arg->nseg = nseg;
1855 }
1856 
1857 
1858 #ifdef ACPI_DMAR
1859 #define IS_DMAR(ctx) (ctx->ifc_flags & IFC_DMAR)
1860 #else
1861 #define IS_DMAR(ctx) (0)
1862 #endif
1863 
1864 /**
1865  *	rxq_refill - refill an rxq  free-buffer list
1866  *	@ctx: the iflib context
1867  *	@rxq: the free-list to refill
1868  *	@n: the number of new buffers to allocate
1869  *
1870  *	(Re)populate an rxq free-buffer list with up to @n new packet buffers.
1871  *	The caller must assure that @n does not exceed the queue's capacity.
1872  */
1873 static void
1874 _iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
1875 {
1876 	struct mbuf *m;
1877 	int idx, frag_idx = fl->ifl_fragidx;
1878         int pidx = fl->ifl_pidx;
1879 	caddr_t cl, *sd_cl;
1880 	struct mbuf **sd_m;
1881 	uint8_t *sd_flags;
1882 	struct if_rxd_update iru;
1883 	bus_dmamap_t *sd_map;
1884 	int n, i = 0;
1885 	uint64_t bus_addr;
1886 	int err;
1887 	qidx_t credits;
1888 
1889 	sd_m = fl->ifl_sds.ifsd_m;
1890 	sd_map = fl->ifl_sds.ifsd_map;
1891 	sd_cl = fl->ifl_sds.ifsd_cl;
1892 	sd_flags = fl->ifl_sds.ifsd_flags;
1893 	idx = pidx;
1894 	credits = fl->ifl_credits;
1895 
1896 	n  = count;
1897 	MPASS(n > 0);
1898 	MPASS(credits + n <= fl->ifl_size);
1899 
1900 	if (pidx < fl->ifl_cidx)
1901 		MPASS(pidx + n <= fl->ifl_cidx);
1902 	if (pidx == fl->ifl_cidx && (credits < fl->ifl_size))
1903 		MPASS(fl->ifl_gen == 0);
1904 	if (pidx > fl->ifl_cidx)
1905 		MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx);
1906 
1907 	DBG_COUNTER_INC(fl_refills);
1908 	if (n > 8)
1909 		DBG_COUNTER_INC(fl_refills_large);
1910 	iru_init(&iru, fl->ifl_rxq, fl->ifl_id);
1911 	while (n--) {
1912 		/*
1913 		 * We allocate an uninitialized mbuf + cluster, mbuf is
1914 		 * initialized after rx.
1915 		 *
1916 		 * If the cluster is still set then we know a minimum sized packet was received
1917 		 */
1918 		bit_ffc_at(fl->ifl_rx_bitmap, frag_idx, fl->ifl_size,  &frag_idx);
1919 		if ((frag_idx < 0) || (frag_idx >= fl->ifl_size))
1920                 	bit_ffc(fl->ifl_rx_bitmap, fl->ifl_size, &frag_idx);
1921 		if ((cl = sd_cl[frag_idx]) == NULL) {
1922                        if ((cl = sd_cl[frag_idx] = m_cljget(NULL, M_NOWAIT, fl->ifl_buf_size)) == NULL)
1923 				break;
1924 #if MEMORY_LOGGING
1925 			fl->ifl_cl_enqueued++;
1926 #endif
1927 		}
1928 		if ((m = m_gethdr(M_NOWAIT, MT_NOINIT)) == NULL) {
1929 			break;
1930 		}
1931 #if MEMORY_LOGGING
1932 		fl->ifl_m_enqueued++;
1933 #endif
1934 
1935 		DBG_COUNTER_INC(rx_allocs);
1936 #if defined(__i386__) || defined(__amd64__)
1937 		if (!IS_DMAR(ctx)) {
1938 			bus_addr = pmap_kextract((vm_offset_t)cl);
1939 		} else
1940 #endif
1941 		{
1942 			struct rxq_refill_cb_arg cb_arg;
1943 
1944 			cb_arg.error = 0;
1945 			MPASS(sd_map != NULL);
1946 			MPASS(sd_map[frag_idx] != NULL);
1947 			err = bus_dmamap_load(fl->ifl_desc_tag, sd_map[frag_idx],
1948 		         cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg, 0);
1949 			bus_dmamap_sync(fl->ifl_desc_tag, sd_map[frag_idx],
1950 					BUS_DMASYNC_PREREAD);
1951 
1952 			if (err != 0 || cb_arg.error) {
1953 				/*
1954 				 * !zone_pack ?
1955 				 */
1956 				if (fl->ifl_zone == zone_pack)
1957 					uma_zfree(fl->ifl_zone, cl);
1958 				m_free(m);
1959 				n = 0;
1960 				goto done;
1961 			}
1962 			bus_addr = cb_arg.seg.ds_addr;
1963 		}
1964                 bit_set(fl->ifl_rx_bitmap, frag_idx);
1965 		sd_flags[frag_idx] |= RX_SW_DESC_INUSE;
1966 
1967 		MPASS(sd_m[frag_idx] == NULL);
1968 		sd_cl[frag_idx] = cl;
1969 		sd_m[frag_idx] = m;
1970 		fl->ifl_rxd_idxs[i] = frag_idx;
1971 		fl->ifl_bus_addrs[i] = bus_addr;
1972 		fl->ifl_vm_addrs[i] = cl;
1973 		credits++;
1974 		i++;
1975 		MPASS(credits <= fl->ifl_size);
1976 		if (++idx == fl->ifl_size) {
1977 			fl->ifl_gen = 1;
1978 			idx = 0;
1979 		}
1980 		if (n == 0 || i == IFLIB_MAX_RX_REFRESH) {
1981 			iru.iru_pidx = pidx;
1982 			iru.iru_count = i;
1983 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
1984 			i = 0;
1985 			pidx = idx;
1986 			fl->ifl_pidx = idx;
1987 			fl->ifl_credits = credits;
1988 		}
1989 
1990 	}
1991 done:
1992 	if (i) {
1993 		iru.iru_pidx = pidx;
1994 		iru.iru_count = i;
1995 		ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
1996 		fl->ifl_pidx = idx;
1997 		fl->ifl_credits = credits;
1998 	}
1999 	DBG_COUNTER_INC(rxd_flush);
2000 	if (fl->ifl_pidx == 0)
2001 		pidx = fl->ifl_size - 1;
2002 	else
2003 		pidx = fl->ifl_pidx - 1;
2004 
2005 	if (sd_map)
2006 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2007 				BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2008 	ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id, fl->ifl_id, pidx);
2009 	fl->ifl_fragidx = frag_idx;
2010 }
2011 
2012 static __inline void
2013 __iflib_fl_refill_lt(if_ctx_t ctx, iflib_fl_t fl, int max)
2014 {
2015 	/* we avoid allowing pidx to catch up with cidx as it confuses ixl */
2016 	int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1;
2017 #ifdef INVARIANTS
2018 	int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1;
2019 #endif
2020 
2021 	MPASS(fl->ifl_credits <= fl->ifl_size);
2022 	MPASS(reclaimable == delta);
2023 
2024 	if (reclaimable > 0)
2025 		_iflib_fl_refill(ctx, fl, min(max, reclaimable));
2026 }
2027 
2028 static void
2029 iflib_fl_bufs_free(iflib_fl_t fl)
2030 {
2031 	iflib_dma_info_t idi = fl->ifl_ifdi;
2032 	uint32_t i;
2033 
2034 	for (i = 0; i < fl->ifl_size; i++) {
2035 		struct mbuf **sd_m = &fl->ifl_sds.ifsd_m[i];
2036 		uint8_t *sd_flags = &fl->ifl_sds.ifsd_flags[i];
2037 		caddr_t *sd_cl = &fl->ifl_sds.ifsd_cl[i];
2038 
2039 		if (*sd_flags & RX_SW_DESC_INUSE) {
2040 			if (fl->ifl_sds.ifsd_map != NULL) {
2041 				bus_dmamap_t sd_map = fl->ifl_sds.ifsd_map[i];
2042 				bus_dmamap_unload(fl->ifl_desc_tag, sd_map);
2043 				if (fl->ifl_rxq->ifr_ctx->ifc_in_detach)
2044 					bus_dmamap_destroy(fl->ifl_desc_tag, sd_map);
2045 			}
2046 			if (*sd_m != NULL) {
2047 				m_init(*sd_m, M_NOWAIT, MT_DATA, 0);
2048 				uma_zfree(zone_mbuf, *sd_m);
2049 			}
2050 			if (*sd_cl != NULL)
2051 				uma_zfree(fl->ifl_zone, *sd_cl);
2052 			*sd_flags = 0;
2053 		} else {
2054 			MPASS(*sd_cl == NULL);
2055 			MPASS(*sd_m == NULL);
2056 		}
2057 #if MEMORY_LOGGING
2058 		fl->ifl_m_dequeued++;
2059 		fl->ifl_cl_dequeued++;
2060 #endif
2061 		*sd_cl = NULL;
2062 		*sd_m = NULL;
2063 	}
2064 #ifdef INVARIANTS
2065 	for (i = 0; i < fl->ifl_size; i++) {
2066 		MPASS(fl->ifl_sds.ifsd_flags[i] == 0);
2067 		MPASS(fl->ifl_sds.ifsd_cl[i] == NULL);
2068 		MPASS(fl->ifl_sds.ifsd_m[i] == NULL);
2069 	}
2070 #endif
2071 	/*
2072 	 * Reset free list values
2073 	 */
2074 	fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = fl->ifl_fragidx = 0;
2075 	bzero(idi->idi_vaddr, idi->idi_size);
2076 }
2077 
2078 /*********************************************************************
2079  *
2080  *  Initialize a receive ring and its buffers.
2081  *
2082  **********************************************************************/
2083 static int
2084 iflib_fl_setup(iflib_fl_t fl)
2085 {
2086 	iflib_rxq_t rxq = fl->ifl_rxq;
2087 	if_ctx_t ctx = rxq->ifr_ctx;
2088 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2089 
2090 	bit_nclear(fl->ifl_rx_bitmap, 0, fl->ifl_size - 1);
2091 	/*
2092 	** Free current RX buffer structs and their mbufs
2093 	*/
2094 	iflib_fl_bufs_free(fl);
2095 	/* Now replenish the mbufs */
2096 	MPASS(fl->ifl_credits == 0);
2097 	/*
2098 	 * XXX don't set the max_frame_size to larger
2099 	 * than the hardware can handle
2100 	 */
2101 	if (sctx->isc_max_frame_size <= 2048)
2102 		fl->ifl_buf_size = MCLBYTES;
2103 #ifndef CONTIGMALLOC_WORKS
2104 	else
2105 		fl->ifl_buf_size = MJUMPAGESIZE;
2106 #else
2107 	else if (sctx->isc_max_frame_size <= 4096)
2108 		fl->ifl_buf_size = MJUMPAGESIZE;
2109 	else if (sctx->isc_max_frame_size <= 9216)
2110 		fl->ifl_buf_size = MJUM9BYTES;
2111 	else
2112 		fl->ifl_buf_size = MJUM16BYTES;
2113 #endif
2114 	if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size)
2115 		ctx->ifc_max_fl_buf_size = fl->ifl_buf_size;
2116 	fl->ifl_cltype = m_gettype(fl->ifl_buf_size);
2117 	fl->ifl_zone = m_getzone(fl->ifl_buf_size);
2118 
2119 
2120 	/* avoid pre-allocating zillions of clusters to an idle card
2121 	 * potentially speeding up attach
2122 	 */
2123 	_iflib_fl_refill(ctx, fl, min(128, fl->ifl_size));
2124 	MPASS(min(128, fl->ifl_size) == fl->ifl_credits);
2125 	if (min(128, fl->ifl_size) != fl->ifl_credits)
2126 		return (ENOBUFS);
2127 	/*
2128 	 * handle failure
2129 	 */
2130 	MPASS(rxq != NULL);
2131 	MPASS(fl->ifl_ifdi != NULL);
2132 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2133 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2134 	return (0);
2135 }
2136 
2137 /*********************************************************************
2138  *
2139  *  Free receive ring data structures
2140  *
2141  **********************************************************************/
2142 static void
2143 iflib_rx_sds_free(iflib_rxq_t rxq)
2144 {
2145 	iflib_fl_t fl;
2146 	int i;
2147 
2148 	if (rxq->ifr_fl != NULL) {
2149 		for (i = 0; i < rxq->ifr_nfl; i++) {
2150 			fl = &rxq->ifr_fl[i];
2151 			if (fl->ifl_desc_tag != NULL) {
2152 				bus_dma_tag_destroy(fl->ifl_desc_tag);
2153 				fl->ifl_desc_tag = NULL;
2154 			}
2155 			free(fl->ifl_sds.ifsd_m, M_IFLIB);
2156 			free(fl->ifl_sds.ifsd_cl, M_IFLIB);
2157 			/* XXX destroy maps first */
2158 			free(fl->ifl_sds.ifsd_map, M_IFLIB);
2159 			fl->ifl_sds.ifsd_m = NULL;
2160 			fl->ifl_sds.ifsd_cl = NULL;
2161 			fl->ifl_sds.ifsd_map = NULL;
2162 		}
2163 		free(rxq->ifr_fl, M_IFLIB);
2164 		rxq->ifr_fl = NULL;
2165 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
2166 	}
2167 }
2168 
2169 /*
2170  * MI independent logic
2171  *
2172  */
2173 static void
2174 iflib_timer(void *arg)
2175 {
2176 	iflib_txq_t txq = arg;
2177 	if_ctx_t ctx = txq->ift_ctx;
2178 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2179 
2180 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
2181 		return;
2182 	/*
2183 	** Check on the state of the TX queue(s), this
2184 	** can be done without the lock because its RO
2185 	** and the HUNG state will be static if set.
2186 	*/
2187 	IFDI_TIMER(ctx, txq->ift_id);
2188 	if ((txq->ift_qstatus == IFLIB_QUEUE_HUNG) &&
2189 	    ((txq->ift_cleaned_prev == txq->ift_cleaned) ||
2190 	     (sctx->isc_pause_frames == 0)))
2191 		goto hung;
2192 
2193 	if (ifmp_ring_is_stalled(txq->ift_br))
2194 		txq->ift_qstatus = IFLIB_QUEUE_HUNG;
2195 	txq->ift_cleaned_prev = txq->ift_cleaned;
2196 	/* handle any laggards */
2197 	if (txq->ift_db_pending)
2198 		GROUPTASK_ENQUEUE(&txq->ift_task);
2199 
2200 	sctx->isc_pause_frames = 0;
2201 	if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)
2202 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq, txq->ift_timer.c_cpu);
2203 	return;
2204  hung:
2205 	device_printf(ctx->ifc_dev,  "TX(%d) desc avail = %d, pidx = %d\n",
2206 				  txq->ift_id, TXQ_AVAIL(txq), txq->ift_pidx);
2207 	STATE_LOCK(ctx);
2208 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2209 	ctx->ifc_flags |= (IFC_DO_WATCHDOG|IFC_DO_RESET);
2210 	iflib_admin_intr_deferred(ctx);
2211 	STATE_UNLOCK(ctx);
2212 }
2213 
2214 static void
2215 iflib_init_locked(if_ctx_t ctx)
2216 {
2217 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2218 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2219 	if_t ifp = ctx->ifc_ifp;
2220 	iflib_fl_t fl;
2221 	iflib_txq_t txq;
2222 	iflib_rxq_t rxq;
2223 	int i, j, tx_ip_csum_flags, tx_ip6_csum_flags;
2224 
2225 
2226 	if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2227 	IFDI_INTR_DISABLE(ctx);
2228 
2229 	tx_ip_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP);
2230 	tx_ip6_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | CSUM_IP6_SCTP);
2231 	/* Set hardware offload abilities */
2232 	if_clearhwassist(ifp);
2233 	if (if_getcapenable(ifp) & IFCAP_TXCSUM)
2234 		if_sethwassistbits(ifp, tx_ip_csum_flags, 0);
2235 	if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6)
2236 		if_sethwassistbits(ifp,  tx_ip6_csum_flags, 0);
2237 	if (if_getcapenable(ifp) & IFCAP_TSO4)
2238 		if_sethwassistbits(ifp, CSUM_IP_TSO, 0);
2239 	if (if_getcapenable(ifp) & IFCAP_TSO6)
2240 		if_sethwassistbits(ifp, CSUM_IP6_TSO, 0);
2241 
2242 	for (i = 0, txq = ctx->ifc_txqs; i < sctx->isc_ntxqsets; i++, txq++) {
2243 		CALLOUT_LOCK(txq);
2244 		callout_stop(&txq->ift_timer);
2245 		CALLOUT_UNLOCK(txq);
2246 		iflib_netmap_txq_init(ctx, txq);
2247 	}
2248 #ifdef INVARIANTS
2249 	i = if_getdrvflags(ifp);
2250 #endif
2251 	IFDI_INIT(ctx);
2252 	MPASS(if_getdrvflags(ifp) == i);
2253 	for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
2254 		/* XXX this should really be done on a per-queue basis */
2255 		if (if_getcapenable(ifp) & IFCAP_NETMAP) {
2256 			MPASS(rxq->ifr_id == i);
2257 			iflib_netmap_rxq_init(ctx, rxq);
2258 			continue;
2259 		}
2260 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
2261 			if (iflib_fl_setup(fl)) {
2262 				device_printf(ctx->ifc_dev, "freelist setup failed - check cluster settings\n");
2263 				goto done;
2264 			}
2265 		}
2266 	}
2267 	done:
2268 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
2269 	IFDI_INTR_ENABLE(ctx);
2270 	txq = ctx->ifc_txqs;
2271 	for (i = 0; i < sctx->isc_ntxqsets; i++, txq++)
2272 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq,
2273 			txq->ift_timer.c_cpu);
2274 }
2275 
2276 static int
2277 iflib_media_change(if_t ifp)
2278 {
2279 	if_ctx_t ctx = if_getsoftc(ifp);
2280 	int err;
2281 
2282 	CTX_LOCK(ctx);
2283 	if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0)
2284 		iflib_init_locked(ctx);
2285 	CTX_UNLOCK(ctx);
2286 	return (err);
2287 }
2288 
2289 static void
2290 iflib_media_status(if_t ifp, struct ifmediareq *ifmr)
2291 {
2292 	if_ctx_t ctx = if_getsoftc(ifp);
2293 
2294 	CTX_LOCK(ctx);
2295 	IFDI_UPDATE_ADMIN_STATUS(ctx);
2296 	IFDI_MEDIA_STATUS(ctx, ifmr);
2297 	CTX_UNLOCK(ctx);
2298 }
2299 
2300 void
2301 iflib_stop(if_ctx_t ctx)
2302 {
2303 	iflib_txq_t txq = ctx->ifc_txqs;
2304 	iflib_rxq_t rxq = ctx->ifc_rxqs;
2305 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2306 	iflib_dma_info_t di;
2307 	iflib_fl_t fl;
2308 	int i, j;
2309 
2310 	/* Tell the stack that the interface is no longer active */
2311 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2312 
2313 	IFDI_INTR_DISABLE(ctx);
2314 	DELAY(1000);
2315 	IFDI_STOP(ctx);
2316 	DELAY(1000);
2317 
2318 	iflib_debug_reset();
2319 	/* Wait for current tx queue users to exit to disarm watchdog timer. */
2320 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) {
2321 		/* make sure all transmitters have completed before proceeding XXX */
2322 
2323 		CALLOUT_LOCK(txq);
2324 		callout_stop(&txq->ift_timer);
2325 		CALLOUT_UNLOCK(txq);
2326 
2327 		/* clean any enqueued buffers */
2328 		iflib_ifmp_purge(txq);
2329 		/* Free any existing tx buffers. */
2330 		for (j = 0; j < txq->ift_size; j++) {
2331 			iflib_txsd_free(ctx, txq, j);
2332 		}
2333 		txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0;
2334 		txq->ift_in_use = txq->ift_gen = txq->ift_cidx = txq->ift_pidx = txq->ift_no_desc_avail = 0;
2335 		txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0;
2336 		txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0;
2337 		txq->ift_pullups = 0;
2338 		ifmp_ring_reset_stats(txq->ift_br);
2339 		for (j = 0, di = txq->ift_ifdi; j < ctx->ifc_nhwtxqs; j++, di++)
2340 			bzero((void *)di->idi_vaddr, di->idi_size);
2341 	}
2342 	for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) {
2343 		/* make sure all transmitters have completed before proceeding XXX */
2344 
2345 		for (j = 0, di = rxq->ifr_ifdi; j < rxq->ifr_nfl; j++, di++)
2346 			bzero((void *)di->idi_vaddr, di->idi_size);
2347 		/* also resets the free lists pidx/cidx */
2348 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
2349 			iflib_fl_bufs_free(fl);
2350 	}
2351 }
2352 
2353 static inline caddr_t
2354 calc_next_rxd(iflib_fl_t fl, int cidx)
2355 {
2356 	qidx_t size;
2357 	int nrxd;
2358 	caddr_t start, end, cur, next;
2359 
2360 	nrxd = fl->ifl_size;
2361 	size = fl->ifl_rxd_size;
2362 	start = fl->ifl_ifdi->idi_vaddr;
2363 
2364 	if (__predict_false(size == 0))
2365 		return (start);
2366 	cur = start + size*cidx;
2367 	end = start + size*nrxd;
2368 	next = CACHE_PTR_NEXT(cur);
2369 	return (next < end ? next : start);
2370 }
2371 
2372 static inline void
2373 prefetch_pkts(iflib_fl_t fl, int cidx)
2374 {
2375 	int nextptr;
2376 	int nrxd = fl->ifl_size;
2377 	caddr_t next_rxd;
2378 
2379 
2380 	nextptr = (cidx + CACHE_PTR_INCREMENT) & (nrxd-1);
2381 	prefetch(&fl->ifl_sds.ifsd_m[nextptr]);
2382 	prefetch(&fl->ifl_sds.ifsd_cl[nextptr]);
2383 	next_rxd = calc_next_rxd(fl, cidx);
2384 	prefetch(next_rxd);
2385 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 1) & (nrxd-1)]);
2386 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 2) & (nrxd-1)]);
2387 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 3) & (nrxd-1)]);
2388 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 4) & (nrxd-1)]);
2389 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 1) & (nrxd-1)]);
2390 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 2) & (nrxd-1)]);
2391 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 3) & (nrxd-1)]);
2392 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 4) & (nrxd-1)]);
2393 }
2394 
2395 static void
2396 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, int unload, if_rxsd_t sd)
2397 {
2398 	int flid, cidx;
2399 	bus_dmamap_t map;
2400 	iflib_fl_t fl;
2401 	iflib_dma_info_t di;
2402 	int next;
2403 
2404 	map = NULL;
2405 	flid = irf->irf_flid;
2406 	cidx = irf->irf_idx;
2407 	fl = &rxq->ifr_fl[flid];
2408 	sd->ifsd_fl = fl;
2409 	sd->ifsd_cidx = cidx;
2410 	sd->ifsd_m = &fl->ifl_sds.ifsd_m[cidx];
2411 	sd->ifsd_cl = &fl->ifl_sds.ifsd_cl[cidx];
2412 	fl->ifl_credits--;
2413 #if MEMORY_LOGGING
2414 	fl->ifl_m_dequeued++;
2415 #endif
2416 	if (rxq->ifr_ctx->ifc_flags & IFC_PREFETCH)
2417 		prefetch_pkts(fl, cidx);
2418 	if (fl->ifl_sds.ifsd_map != NULL) {
2419 		next = (cidx + CACHE_PTR_INCREMENT) & (fl->ifl_size-1);
2420 		prefetch(&fl->ifl_sds.ifsd_map[next]);
2421 		map = fl->ifl_sds.ifsd_map[cidx];
2422 		di = fl->ifl_ifdi;
2423 		next = (cidx + CACHE_LINE_SIZE) & (fl->ifl_size-1);
2424 		prefetch(&fl->ifl_sds.ifsd_flags[next]);
2425 		bus_dmamap_sync(di->idi_tag, di->idi_map,
2426 				BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
2427 
2428 	/* not valid assert if bxe really does SGE from non-contiguous elements */
2429 		MPASS(fl->ifl_cidx == cidx);
2430 		if (unload)
2431 			bus_dmamap_unload(fl->ifl_desc_tag, map);
2432 	}
2433 	fl->ifl_cidx = (fl->ifl_cidx + 1) & (fl->ifl_size-1);
2434 	if (__predict_false(fl->ifl_cidx == 0))
2435 		fl->ifl_gen = 0;
2436 	if (map != NULL)
2437 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2438 			BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2439         bit_clear(fl->ifl_rx_bitmap, cidx);
2440 }
2441 
2442 static struct mbuf *
2443 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri, if_rxsd_t sd)
2444 {
2445 	int i, padlen , flags;
2446 	struct mbuf *m, *mh, *mt;
2447 	caddr_t cl;
2448 
2449 	i = 0;
2450 	mh = NULL;
2451 	do {
2452 		rxd_frag_to_sd(rxq, &ri->iri_frags[i], TRUE, sd);
2453 
2454 		MPASS(*sd->ifsd_cl != NULL);
2455 		MPASS(*sd->ifsd_m != NULL);
2456 
2457 		/* Don't include zero-length frags */
2458 		if (ri->iri_frags[i].irf_len == 0) {
2459 			/* XXX we can save the cluster here, but not the mbuf */
2460 			m_init(*sd->ifsd_m, M_NOWAIT, MT_DATA, 0);
2461 			m_free(*sd->ifsd_m);
2462 			*sd->ifsd_m = NULL;
2463 			continue;
2464 		}
2465 		m = *sd->ifsd_m;
2466 		*sd->ifsd_m = NULL;
2467 		if (mh == NULL) {
2468 			flags = M_PKTHDR|M_EXT;
2469 			mh = mt = m;
2470 			padlen = ri->iri_pad;
2471 		} else {
2472 			flags = M_EXT;
2473 			mt->m_next = m;
2474 			mt = m;
2475 			/* assuming padding is only on the first fragment */
2476 			padlen = 0;
2477 		}
2478 		cl = *sd->ifsd_cl;
2479 		*sd->ifsd_cl = NULL;
2480 
2481 		/* Can these two be made one ? */
2482 		m_init(m, M_NOWAIT, MT_DATA, flags);
2483 		m_cljset(m, cl, sd->ifsd_fl->ifl_cltype);
2484 		/*
2485 		 * These must follow m_init and m_cljset
2486 		 */
2487 		m->m_data += padlen;
2488 		ri->iri_len -= padlen;
2489 		m->m_len = ri->iri_frags[i].irf_len;
2490 	} while (++i < ri->iri_nfrags);
2491 
2492 	return (mh);
2493 }
2494 
2495 /*
2496  * Process one software descriptor
2497  */
2498 static struct mbuf *
2499 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri)
2500 {
2501 	struct if_rxsd sd;
2502 	struct mbuf *m;
2503 
2504 	/* should I merge this back in now that the two paths are basically duplicated? */
2505 	if (ri->iri_nfrags == 1 &&
2506 	    ri->iri_frags[0].irf_len <= MIN(IFLIB_RX_COPY_THRESH, MHLEN)) {
2507 		rxd_frag_to_sd(rxq, &ri->iri_frags[0], FALSE, &sd);
2508 		m = *sd.ifsd_m;
2509 		*sd.ifsd_m = NULL;
2510 		m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR);
2511 #ifndef __NO_STRICT_ALIGNMENT
2512 		if (!IP_ALIGNED(m))
2513 			m->m_data += 2;
2514 #endif
2515 		memcpy(m->m_data, *sd.ifsd_cl, ri->iri_len);
2516 		m->m_len = ri->iri_frags[0].irf_len;
2517        } else {
2518 		m = assemble_segments(rxq, ri, &sd);
2519 	}
2520 	m->m_pkthdr.len = ri->iri_len;
2521 	m->m_pkthdr.rcvif = ri->iri_ifp;
2522 	m->m_flags |= ri->iri_flags;
2523 	m->m_pkthdr.ether_vtag = ri->iri_vtag;
2524 	m->m_pkthdr.flowid = ri->iri_flowid;
2525 	M_HASHTYPE_SET(m, ri->iri_rsstype);
2526 	m->m_pkthdr.csum_flags = ri->iri_csum_flags;
2527 	m->m_pkthdr.csum_data = ri->iri_csum_data;
2528 	return (m);
2529 }
2530 
2531 #if defined(INET6) || defined(INET)
2532 static void
2533 iflib_get_ip_forwarding(struct lro_ctrl *lc, bool *v4, bool *v6)
2534 {
2535 	CURVNET_SET(lc->ifp->if_vnet);
2536 #if defined(INET6)
2537 	*v6 = VNET(ip6_forwarding);
2538 #endif
2539 #if defined(INET)
2540 	*v4 = VNET(ipforwarding);
2541 #endif
2542 	CURVNET_RESTORE();
2543 }
2544 
2545 /*
2546  * Returns true if it's possible this packet could be LROed.
2547  * if it returns false, it is guaranteed that tcp_lro_rx()
2548  * would not return zero.
2549  */
2550 static bool
2551 iflib_check_lro_possible(struct mbuf *m, bool v4_forwarding, bool v6_forwarding)
2552 {
2553 	struct ether_header *eh;
2554 	uint16_t eh_type;
2555 
2556 	eh = mtod(m, struct ether_header *);
2557 	eh_type = ntohs(eh->ether_type);
2558 	switch (eh_type) {
2559 #if defined(INET6)
2560 		case ETHERTYPE_IPV6:
2561 			return !v6_forwarding;
2562 #endif
2563 #if defined (INET)
2564 		case ETHERTYPE_IP:
2565 			return !v4_forwarding;
2566 #endif
2567 	}
2568 
2569 	return false;
2570 }
2571 #else
2572 static void
2573 iflib_get_ip_forwarding(struct lro_ctrl *lc __unused, bool *v4 __unused, bool *v6 __unused)
2574 {
2575 }
2576 #endif
2577 
2578 static bool
2579 iflib_rxeof(iflib_rxq_t rxq, qidx_t budget)
2580 {
2581 	if_ctx_t ctx = rxq->ifr_ctx;
2582 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2583 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2584 	int avail, i;
2585 	qidx_t *cidxp;
2586 	struct if_rxd_info ri;
2587 	int err, budget_left, rx_bytes, rx_pkts;
2588 	iflib_fl_t fl;
2589 	struct ifnet *ifp;
2590 	int lro_enabled;
2591 	bool lro_possible = false;
2592 	bool v4_forwarding, v6_forwarding;
2593 
2594 	/*
2595 	 * XXX early demux data packets so that if_input processing only handles
2596 	 * acks in interrupt context
2597 	 */
2598 	struct mbuf *m, *mh, *mt, *mf;
2599 
2600 	ifp = ctx->ifc_ifp;
2601 	mh = mt = NULL;
2602 	MPASS(budget > 0);
2603 	rx_pkts	= rx_bytes = 0;
2604 	if (sctx->isc_flags & IFLIB_HAS_RXCQ)
2605 		cidxp = &rxq->ifr_cq_cidx;
2606 	else
2607 		cidxp = &rxq->ifr_fl[0].ifl_cidx;
2608 	if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) {
2609 		for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2610 			__iflib_fl_refill_lt(ctx, fl, budget + 8);
2611 		DBG_COUNTER_INC(rx_unavail);
2612 		return (false);
2613 	}
2614 
2615 	for (budget_left = budget; (budget_left > 0) && (avail > 0); budget_left--, avail--) {
2616 		if (__predict_false(!CTX_ACTIVE(ctx))) {
2617 			DBG_COUNTER_INC(rx_ctx_inactive);
2618 			break;
2619 		}
2620 		/*
2621 		 * Reset client set fields to their default values
2622 		 */
2623 		rxd_info_zero(&ri);
2624 		ri.iri_qsidx = rxq->ifr_id;
2625 		ri.iri_cidx = *cidxp;
2626 		ri.iri_ifp = ifp;
2627 		ri.iri_frags = rxq->ifr_frags;
2628 		err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
2629 
2630 		if (err)
2631 			goto err;
2632 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
2633 			*cidxp = ri.iri_cidx;
2634 			/* Update our consumer index */
2635 			/* XXX NB: shurd - check if this is still safe */
2636 			while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0]) {
2637 				rxq->ifr_cq_cidx -= scctx->isc_nrxd[0];
2638 				rxq->ifr_cq_gen = 0;
2639 			}
2640 			/* was this only a completion queue message? */
2641 			if (__predict_false(ri.iri_nfrags == 0))
2642 				continue;
2643 		}
2644 		MPASS(ri.iri_nfrags != 0);
2645 		MPASS(ri.iri_len != 0);
2646 
2647 		/* will advance the cidx on the corresponding free lists */
2648 		m = iflib_rxd_pkt_get(rxq, &ri);
2649 		if (avail == 0 && budget_left)
2650 			avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left);
2651 
2652 		if (__predict_false(m == NULL)) {
2653 			DBG_COUNTER_INC(rx_mbuf_null);
2654 			continue;
2655 		}
2656 		/* imm_pkt: -- cxgb */
2657 		if (mh == NULL)
2658 			mh = mt = m;
2659 		else {
2660 			mt->m_nextpkt = m;
2661 			mt = m;
2662 		}
2663 	}
2664 	/* make sure that we can refill faster than drain */
2665 	for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2666 		__iflib_fl_refill_lt(ctx, fl, budget + 8);
2667 
2668 	lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO);
2669 	if (lro_enabled)
2670 		iflib_get_ip_forwarding(&rxq->ifr_lc, &v4_forwarding, &v6_forwarding);
2671 	mt = mf = NULL;
2672 	while (mh != NULL) {
2673 		m = mh;
2674 		mh = mh->m_nextpkt;
2675 		m->m_nextpkt = NULL;
2676 #ifndef __NO_STRICT_ALIGNMENT
2677 		if (!IP_ALIGNED(m) && (m = iflib_fixup_rx(m)) == NULL)
2678 			continue;
2679 #endif
2680 		rx_bytes += m->m_pkthdr.len;
2681 		rx_pkts++;
2682 #if defined(INET6) || defined(INET)
2683 		if (lro_enabled) {
2684 			if (!lro_possible) {
2685 				lro_possible = iflib_check_lro_possible(m, v4_forwarding, v6_forwarding);
2686 				if (lro_possible && mf != NULL) {
2687 					ifp->if_input(ifp, mf);
2688 					DBG_COUNTER_INC(rx_if_input);
2689 					mt = mf = NULL;
2690 				}
2691 			}
2692 			if ((m->m_pkthdr.csum_flags & (CSUM_L4_CALC|CSUM_L4_VALID)) ==
2693 			    (CSUM_L4_CALC|CSUM_L4_VALID)) {
2694 				if (lro_possible && tcp_lro_rx(&rxq->ifr_lc, m, 0) == 0)
2695 					continue;
2696 			}
2697 		}
2698 #endif
2699 		if (lro_possible) {
2700 			ifp->if_input(ifp, m);
2701 			DBG_COUNTER_INC(rx_if_input);
2702 			continue;
2703 		}
2704 
2705 		if (mf == NULL)
2706 			mf = m;
2707 		if (mt != NULL)
2708 			mt->m_nextpkt = m;
2709 		mt = m;
2710 	}
2711 	if (mf != NULL) {
2712 		ifp->if_input(ifp, mf);
2713 		DBG_COUNTER_INC(rx_if_input);
2714 	}
2715 
2716 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
2717 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
2718 
2719 	/*
2720 	 * Flush any outstanding LRO work
2721 	 */
2722 #if defined(INET6) || defined(INET)
2723 	tcp_lro_flush_all(&rxq->ifr_lc);
2724 #endif
2725 	if (avail)
2726 		return true;
2727 	return (iflib_rxd_avail(ctx, rxq, *cidxp, 1));
2728 err:
2729 	STATE_LOCK(ctx);
2730 	ctx->ifc_flags |= IFC_DO_RESET;
2731 	iflib_admin_intr_deferred(ctx);
2732 	STATE_UNLOCK(ctx);
2733 	return (false);
2734 }
2735 
2736 #define TXD_NOTIFY_COUNT(txq) (((txq)->ift_size / (txq)->ift_update_freq)-1)
2737 static inline qidx_t
2738 txq_max_db_deferred(iflib_txq_t txq, qidx_t in_use)
2739 {
2740 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2741 	qidx_t minthresh = txq->ift_size / 8;
2742 	if (in_use > 4*minthresh)
2743 		return (notify_count);
2744 	if (in_use > 2*minthresh)
2745 		return (notify_count >> 1);
2746 	if (in_use > minthresh)
2747 		return (notify_count >> 3);
2748 	return (0);
2749 }
2750 
2751 static inline qidx_t
2752 txq_max_rs_deferred(iflib_txq_t txq)
2753 {
2754 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2755 	qidx_t minthresh = txq->ift_size / 8;
2756 	if (txq->ift_in_use > 4*minthresh)
2757 		return (notify_count);
2758 	if (txq->ift_in_use > 2*minthresh)
2759 		return (notify_count >> 1);
2760 	if (txq->ift_in_use > minthresh)
2761 		return (notify_count >> 2);
2762 	return (2);
2763 }
2764 
2765 #define M_CSUM_FLAGS(m) ((m)->m_pkthdr.csum_flags)
2766 #define M_HAS_VLANTAG(m) (m->m_flags & M_VLANTAG)
2767 
2768 #define TXQ_MAX_DB_DEFERRED(txq, in_use) txq_max_db_deferred((txq), (in_use))
2769 #define TXQ_MAX_RS_DEFERRED(txq) txq_max_rs_deferred(txq)
2770 #define TXQ_MAX_DB_CONSUMED(size) (size >> 4)
2771 
2772 /* forward compatibility for cxgb */
2773 #define FIRST_QSET(ctx) 0
2774 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets)
2775 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets)
2776 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx))
2777 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments))
2778 
2779 /* XXX we should be setting this to something other than zero */
2780 #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh)
2781 #define MAX_TX_DESC(ctx) ((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max)
2782 
2783 static inline bool
2784 iflib_txd_db_check(if_ctx_t ctx, iflib_txq_t txq, int ring, qidx_t in_use)
2785 {
2786 	qidx_t dbval, max;
2787 	bool rang;
2788 
2789 	rang = false;
2790 	max = TXQ_MAX_DB_DEFERRED(txq, in_use);
2791 	if (ring || txq->ift_db_pending >= max) {
2792 		dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx;
2793 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval);
2794 		txq->ift_db_pending = txq->ift_npending = 0;
2795 		rang = true;
2796 	}
2797 	return (rang);
2798 }
2799 
2800 #ifdef PKT_DEBUG
2801 static void
2802 print_pkt(if_pkt_info_t pi)
2803 {
2804 	printf("pi len:  %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n",
2805 	       pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx);
2806 	printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n",
2807 	       pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag);
2808 	printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n",
2809 	       pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto);
2810 }
2811 #endif
2812 
2813 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO)
2814 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO)
2815 
2816 static int
2817 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp)
2818 {
2819 	if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx;
2820 	struct ether_vlan_header *eh;
2821 	struct mbuf *m, *n;
2822 
2823 	n = m = *mp;
2824 	if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) &&
2825 	    M_WRITABLE(m) == 0) {
2826 		if ((m = m_dup(m, M_NOWAIT)) == NULL) {
2827 			return (ENOMEM);
2828 		} else {
2829 			m_freem(*mp);
2830 			n = *mp = m;
2831 		}
2832 	}
2833 
2834 	/*
2835 	 * Determine where frame payload starts.
2836 	 * Jump over vlan headers if already present,
2837 	 * helpful for QinQ too.
2838 	 */
2839 	if (__predict_false(m->m_len < sizeof(*eh))) {
2840 		txq->ift_pullups++;
2841 		if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL))
2842 			return (ENOMEM);
2843 	}
2844 	eh = mtod(m, struct ether_vlan_header *);
2845 	if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
2846 		pi->ipi_etype = ntohs(eh->evl_proto);
2847 		pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
2848 	} else {
2849 		pi->ipi_etype = ntohs(eh->evl_encap_proto);
2850 		pi->ipi_ehdrlen = ETHER_HDR_LEN;
2851 	}
2852 
2853 	switch (pi->ipi_etype) {
2854 #ifdef INET
2855 	case ETHERTYPE_IP:
2856 	{
2857 		struct ip *ip = NULL;
2858 		struct tcphdr *th = NULL;
2859 		int minthlen;
2860 
2861 		minthlen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip) + sizeof(*th));
2862 		if (__predict_false(m->m_len < minthlen)) {
2863 			/*
2864 			 * if this code bloat is causing too much of a hit
2865 			 * move it to a separate function and mark it noinline
2866 			 */
2867 			if (m->m_len == pi->ipi_ehdrlen) {
2868 				n = m->m_next;
2869 				MPASS(n);
2870 				if (n->m_len >= sizeof(*ip))  {
2871 					ip = (struct ip *)n->m_data;
2872 					if (n->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2873 						th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2874 				} else {
2875 					txq->ift_pullups++;
2876 					if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
2877 						return (ENOMEM);
2878 					ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2879 				}
2880 			} else {
2881 				txq->ift_pullups++;
2882 				if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
2883 					return (ENOMEM);
2884 				ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2885 				if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2886 					th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2887 			}
2888 		} else {
2889 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2890 			if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2891 				th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2892 		}
2893 		pi->ipi_ip_hlen = ip->ip_hl << 2;
2894 		pi->ipi_ipproto = ip->ip_p;
2895 		pi->ipi_flags |= IPI_TX_IPV4;
2896 
2897 		if ((sctx->isc_flags & IFLIB_NEED_ZERO_CSUM) && (pi->ipi_csum_flags & CSUM_IP))
2898                        ip->ip_sum = 0;
2899 
2900 		if (IS_TSO4(pi)) {
2901 			if (pi->ipi_ipproto == IPPROTO_TCP) {
2902 				if (__predict_false(th == NULL)) {
2903 					txq->ift_pullups++;
2904 					if (__predict_false((m = m_pullup(m, (ip->ip_hl << 2) + sizeof(*th))) == NULL))
2905 						return (ENOMEM);
2906 					th = (struct tcphdr *)((caddr_t)ip + pi->ipi_ip_hlen);
2907 				}
2908 				pi->ipi_tcp_hflags = th->th_flags;
2909 				pi->ipi_tcp_hlen = th->th_off << 2;
2910 				pi->ipi_tcp_seq = th->th_seq;
2911 			}
2912 			if (__predict_false(ip->ip_p != IPPROTO_TCP))
2913 				return (ENXIO);
2914 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
2915 					       ip->ip_dst.s_addr, htons(IPPROTO_TCP));
2916 			pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
2917 			if (sctx->isc_flags & IFLIB_TSO_INIT_IP) {
2918 				ip->ip_sum = 0;
2919 				ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz);
2920 			}
2921 		}
2922 		break;
2923 	}
2924 #endif
2925 #ifdef INET6
2926 	case ETHERTYPE_IPV6:
2927 	{
2928 		struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
2929 		struct tcphdr *th;
2930 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
2931 
2932 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
2933 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
2934 				return (ENOMEM);
2935 		}
2936 		th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
2937 
2938 		/* XXX-BZ this will go badly in case of ext hdrs. */
2939 		pi->ipi_ipproto = ip6->ip6_nxt;
2940 		pi->ipi_flags |= IPI_TX_IPV6;
2941 
2942 		if (IS_TSO6(pi)) {
2943 			if (pi->ipi_ipproto == IPPROTO_TCP) {
2944 				if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) {
2945 					if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL))
2946 						return (ENOMEM);
2947 				}
2948 				pi->ipi_tcp_hflags = th->th_flags;
2949 				pi->ipi_tcp_hlen = th->th_off << 2;
2950 			}
2951 
2952 			if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP))
2953 				return (ENXIO);
2954 			/*
2955 			 * The corresponding flag is set by the stack in the IPv4
2956 			 * TSO case, but not in IPv6 (at least in FreeBSD 10.2).
2957 			 * So, set it here because the rest of the flow requires it.
2958 			 */
2959 			pi->ipi_csum_flags |= CSUM_TCP_IPV6;
2960 			th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0);
2961 			pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
2962 		}
2963 		break;
2964 	}
2965 #endif
2966 	default:
2967 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
2968 		pi->ipi_ip_hlen = 0;
2969 		break;
2970 	}
2971 	*mp = m;
2972 
2973 	return (0);
2974 }
2975 
2976 static  __noinline  struct mbuf *
2977 collapse_pkthdr(struct mbuf *m0)
2978 {
2979 	struct mbuf *m, *m_next, *tmp;
2980 
2981 	m = m0;
2982 	m_next = m->m_next;
2983 	while (m_next != NULL && m_next->m_len == 0) {
2984 		m = m_next;
2985 		m->m_next = NULL;
2986 		m_free(m);
2987 		m_next = m_next->m_next;
2988 	}
2989 	m = m0;
2990 	m->m_next = m_next;
2991 	if ((m_next->m_flags & M_EXT) == 0) {
2992 		m = m_defrag(m, M_NOWAIT);
2993 	} else {
2994 		tmp = m_next->m_next;
2995 		memcpy(m_next, m, MPKTHSIZE);
2996 		m = m_next;
2997 		m->m_next = tmp;
2998 	}
2999 	return (m);
3000 }
3001 
3002 /*
3003  * If dodgy hardware rejects the scatter gather chain we've handed it
3004  * we'll need to remove the mbuf chain from ifsg_m[] before we can add the
3005  * m_defrag'd mbufs
3006  */
3007 static __noinline struct mbuf *
3008 iflib_remove_mbuf(iflib_txq_t txq)
3009 {
3010 	int ntxd, i, pidx;
3011 	struct mbuf *m, *mh, **ifsd_m;
3012 
3013 	pidx = txq->ift_pidx;
3014 	ifsd_m = txq->ift_sds.ifsd_m;
3015 	ntxd = txq->ift_size;
3016 	mh = m = ifsd_m[pidx];
3017 	ifsd_m[pidx] = NULL;
3018 #if MEMORY_LOGGING
3019 	txq->ift_dequeued++;
3020 #endif
3021 	i = 1;
3022 
3023 	while (m) {
3024 		ifsd_m[(pidx + i) & (ntxd -1)] = NULL;
3025 #if MEMORY_LOGGING
3026 		txq->ift_dequeued++;
3027 #endif
3028 		m = m->m_next;
3029 		i++;
3030 	}
3031 	return (mh);
3032 }
3033 
3034 static int
3035 iflib_busdma_load_mbuf_sg(iflib_txq_t txq, bus_dma_tag_t tag, bus_dmamap_t map,
3036 			  struct mbuf **m0, bus_dma_segment_t *segs, int *nsegs,
3037 			  int max_segs, int flags)
3038 {
3039 	if_ctx_t ctx;
3040 	if_shared_ctx_t		sctx;
3041 	if_softc_ctx_t		scctx;
3042 	int i, next, pidx, err, ntxd, count;
3043 	struct mbuf *m, *tmp, **ifsd_m;
3044 
3045 	m = *m0;
3046 
3047 	/*
3048 	 * Please don't ever do this
3049 	 */
3050 	if (__predict_false(m->m_len == 0))
3051 		*m0 = m = collapse_pkthdr(m);
3052 
3053 	ctx = txq->ift_ctx;
3054 	sctx = ctx->ifc_sctx;
3055 	scctx = &ctx->ifc_softc_ctx;
3056 	ifsd_m = txq->ift_sds.ifsd_m;
3057 	ntxd = txq->ift_size;
3058 	pidx = txq->ift_pidx;
3059 	if (map != NULL) {
3060 		uint8_t *ifsd_flags = txq->ift_sds.ifsd_flags;
3061 
3062 		err = bus_dmamap_load_mbuf_sg(tag, map,
3063 					      *m0, segs, nsegs, BUS_DMA_NOWAIT);
3064 		if (err)
3065 			return (err);
3066 		ifsd_flags[pidx] |= TX_SW_DESC_MAPPED;
3067 		count = 0;
3068 		m = *m0;
3069 		do {
3070 			if (__predict_false(m->m_len <= 0)) {
3071 				tmp = m;
3072 				m = m->m_next;
3073 				tmp->m_next = NULL;
3074 				m_free(tmp);
3075 				continue;
3076 			}
3077 			m = m->m_next;
3078 			count++;
3079 		} while (m != NULL);
3080 		if (count > *nsegs) {
3081 			ifsd_m[pidx] = *m0;
3082 			ifsd_m[pidx]->m_flags |= M_TOOBIG;
3083 			return (0);
3084 		}
3085 		m = *m0;
3086 		count = 0;
3087 		do {
3088 			next = (pidx + count) & (ntxd-1);
3089 			MPASS(ifsd_m[next] == NULL);
3090 			ifsd_m[next] = m;
3091 			count++;
3092 			tmp = m;
3093 			m = m->m_next;
3094 		} while (m != NULL);
3095 	} else {
3096 		int buflen, sgsize, maxsegsz, max_sgsize;
3097 		vm_offset_t vaddr;
3098 		vm_paddr_t curaddr;
3099 
3100 		count = i = 0;
3101 		m = *m0;
3102 		if (m->m_pkthdr.csum_flags & CSUM_TSO)
3103 			maxsegsz = scctx->isc_tx_tso_segsize_max;
3104 		else
3105 			maxsegsz = sctx->isc_tx_maxsegsize;
3106 
3107 		do {
3108 			if (__predict_false(m->m_len <= 0)) {
3109 				tmp = m;
3110 				m = m->m_next;
3111 				tmp->m_next = NULL;
3112 				m_free(tmp);
3113 				continue;
3114 			}
3115 			buflen = m->m_len;
3116 			vaddr = (vm_offset_t)m->m_data;
3117 			/*
3118 			 * see if we can't be smarter about physically
3119 			 * contiguous mappings
3120 			 */
3121 			next = (pidx + count) & (ntxd-1);
3122 			MPASS(ifsd_m[next] == NULL);
3123 #if MEMORY_LOGGING
3124 			txq->ift_enqueued++;
3125 #endif
3126 			ifsd_m[next] = m;
3127 			while (buflen > 0) {
3128 				if (i >= max_segs)
3129 					goto err;
3130 				max_sgsize = MIN(buflen, maxsegsz);
3131 				curaddr = pmap_kextract(vaddr);
3132 				sgsize = PAGE_SIZE - (curaddr & PAGE_MASK);
3133 				sgsize = MIN(sgsize, max_sgsize);
3134 				segs[i].ds_addr = curaddr;
3135 				segs[i].ds_len = sgsize;
3136 				vaddr += sgsize;
3137 				buflen -= sgsize;
3138 				i++;
3139 			}
3140 			count++;
3141 			tmp = m;
3142 			m = m->m_next;
3143 		} while (m != NULL);
3144 		*nsegs = i;
3145 	}
3146 	return (0);
3147 err:
3148 	*m0 = iflib_remove_mbuf(txq);
3149 	return (EFBIG);
3150 }
3151 
3152 static inline caddr_t
3153 calc_next_txd(iflib_txq_t txq, int cidx, uint8_t qid)
3154 {
3155 	qidx_t size;
3156 	int ntxd;
3157 	caddr_t start, end, cur, next;
3158 
3159 	ntxd = txq->ift_size;
3160 	size = txq->ift_txd_size[qid];
3161 	start = txq->ift_ifdi[qid].idi_vaddr;
3162 
3163 	if (__predict_false(size == 0))
3164 		return (start);
3165 	cur = start + size*cidx;
3166 	end = start + size*ntxd;
3167 	next = CACHE_PTR_NEXT(cur);
3168 	return (next < end ? next : start);
3169 }
3170 
3171 /*
3172  * Pad an mbuf to ensure a minimum ethernet frame size.
3173  * min_frame_size is the frame size (less CRC) to pad the mbuf to
3174  */
3175 static __noinline int
3176 iflib_ether_pad(device_t dev, struct mbuf **m_head, uint16_t min_frame_size)
3177 {
3178 	/*
3179 	 * 18 is enough bytes to pad an ARP packet to 46 bytes, and
3180 	 * and ARP message is the smallest common payload I can think of
3181 	 */
3182 	static char pad[18];	/* just zeros */
3183 	int n;
3184 	struct mbuf *new_head;
3185 
3186 	if (!M_WRITABLE(*m_head)) {
3187 		new_head = m_dup(*m_head, M_NOWAIT);
3188 		if (new_head == NULL) {
3189 			m_freem(*m_head);
3190 			device_printf(dev, "cannot pad short frame, m_dup() failed");
3191 			DBG_COUNTER_INC(encap_pad_mbuf_fail);
3192 			return ENOMEM;
3193 		}
3194 		m_freem(*m_head);
3195 		*m_head = new_head;
3196 	}
3197 
3198 	for (n = min_frame_size - (*m_head)->m_pkthdr.len;
3199 	     n > 0; n -= sizeof(pad))
3200 		if (!m_append(*m_head, min(n, sizeof(pad)), pad))
3201 			break;
3202 
3203 	if (n > 0) {
3204 		m_freem(*m_head);
3205 		device_printf(dev, "cannot pad short frame\n");
3206 		DBG_COUNTER_INC(encap_pad_mbuf_fail);
3207 		return (ENOBUFS);
3208 	}
3209 
3210 	return 0;
3211 }
3212 
3213 static int
3214 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp)
3215 {
3216 	if_ctx_t		ctx;
3217 	if_shared_ctx_t		sctx;
3218 	if_softc_ctx_t		scctx;
3219 	bus_dma_segment_t	*segs;
3220 	struct mbuf		*m_head;
3221 	void			*next_txd;
3222 	bus_dmamap_t		map;
3223 	struct if_pkt_info	pi;
3224 	int remap = 0;
3225 	int err, nsegs, ndesc, max_segs, pidx, cidx, next, ntxd;
3226 	bus_dma_tag_t desc_tag;
3227 
3228 	segs = txq->ift_segs;
3229 	ctx = txq->ift_ctx;
3230 	sctx = ctx->ifc_sctx;
3231 	scctx = &ctx->ifc_softc_ctx;
3232 	segs = txq->ift_segs;
3233 	ntxd = txq->ift_size;
3234 	m_head = *m_headp;
3235 	map = NULL;
3236 
3237 	/*
3238 	 * If we're doing TSO the next descriptor to clean may be quite far ahead
3239 	 */
3240 	cidx = txq->ift_cidx;
3241 	pidx = txq->ift_pidx;
3242 	if (ctx->ifc_flags & IFC_PREFETCH) {
3243 		next = (cidx + CACHE_PTR_INCREMENT) & (ntxd-1);
3244 		if (!(ctx->ifc_flags & IFLIB_HAS_TXCQ)) {
3245 			next_txd = calc_next_txd(txq, cidx, 0);
3246 			prefetch(next_txd);
3247 		}
3248 
3249 		/* prefetch the next cache line of mbuf pointers and flags */
3250 		prefetch(&txq->ift_sds.ifsd_m[next]);
3251 		if (txq->ift_sds.ifsd_map != NULL) {
3252 			prefetch(&txq->ift_sds.ifsd_map[next]);
3253 			next = (cidx + CACHE_LINE_SIZE) & (ntxd-1);
3254 			prefetch(&txq->ift_sds.ifsd_flags[next]);
3255 		}
3256 	} else if (txq->ift_sds.ifsd_map != NULL)
3257 		map = txq->ift_sds.ifsd_map[pidx];
3258 
3259 	if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
3260 		desc_tag = txq->ift_tso_desc_tag;
3261 		max_segs = scctx->isc_tx_tso_segments_max;
3262 	} else {
3263 		desc_tag = txq->ift_desc_tag;
3264 		max_segs = scctx->isc_tx_nsegments;
3265 	}
3266 	if ((sctx->isc_flags & IFLIB_NEED_ETHER_PAD) &&
3267 	    __predict_false(m_head->m_pkthdr.len < scctx->isc_min_frame_size)) {
3268 		err = iflib_ether_pad(ctx->ifc_dev, m_headp, scctx->isc_min_frame_size);
3269 		if (err)
3270 			return err;
3271 	}
3272 	m_head = *m_headp;
3273 
3274 	pkt_info_zero(&pi);
3275 	pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG|M_BCAST|M_MCAST));
3276 	pi.ipi_pidx = pidx;
3277 	pi.ipi_qsidx = txq->ift_id;
3278 	pi.ipi_len = m_head->m_pkthdr.len;
3279 	pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags;
3280 	pi.ipi_vtag = (m_head->m_flags & M_VLANTAG) ? m_head->m_pkthdr.ether_vtag : 0;
3281 
3282 	/* deliberate bitwise OR to make one condition */
3283 	if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) {
3284 		if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0))
3285 			return (err);
3286 		m_head = *m_headp;
3287 	}
3288 
3289 retry:
3290 	err = iflib_busdma_load_mbuf_sg(txq, desc_tag, map, m_headp, segs, &nsegs, max_segs, BUS_DMA_NOWAIT);
3291 defrag:
3292 	if (__predict_false(err)) {
3293 		switch (err) {
3294 		case EFBIG:
3295 			/* try collapse once and defrag once */
3296 			if (remap == 0) {
3297 				m_head = m_collapse(*m_headp, M_NOWAIT, max_segs);
3298 				/* try defrag if collapsing fails */
3299 				if (m_head == NULL)
3300 					remap++;
3301 			}
3302 			if (remap == 1)
3303 				m_head = m_defrag(*m_headp, M_NOWAIT);
3304 			remap++;
3305 			if (__predict_false(m_head == NULL))
3306 				goto defrag_failed;
3307 			txq->ift_mbuf_defrag++;
3308 			*m_headp = m_head;
3309 			goto retry;
3310 			break;
3311 		case ENOMEM:
3312 			txq->ift_no_tx_dma_setup++;
3313 			break;
3314 		default:
3315 			txq->ift_no_tx_dma_setup++;
3316 			m_freem(*m_headp);
3317 			DBG_COUNTER_INC(tx_frees);
3318 			*m_headp = NULL;
3319 			break;
3320 		}
3321 		txq->ift_map_failed++;
3322 		DBG_COUNTER_INC(encap_load_mbuf_fail);
3323 		return (err);
3324 	}
3325 
3326 	/*
3327 	 * XXX assumes a 1 to 1 relationship between segments and
3328 	 *        descriptors - this does not hold true on all drivers, e.g.
3329 	 *        cxgb
3330 	 */
3331 	if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) {
3332 		txq->ift_no_desc_avail++;
3333 		if (map != NULL)
3334 			bus_dmamap_unload(desc_tag, map);
3335 		DBG_COUNTER_INC(encap_txq_avail_fail);
3336 		if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0)
3337 			GROUPTASK_ENQUEUE(&txq->ift_task);
3338 		return (ENOBUFS);
3339 	}
3340 	/*
3341 	 * On Intel cards we can greatly reduce the number of TX interrupts
3342 	 * we see by only setting report status on every Nth descriptor.
3343 	 * However, this also means that the driver will need to keep track
3344 	 * of the descriptors that RS was set on to check them for the DD bit.
3345 	 */
3346 	txq->ift_rs_pending += nsegs + 1;
3347 	if (txq->ift_rs_pending > TXQ_MAX_RS_DEFERRED(txq) ||
3348 	     iflib_no_tx_batch || (TXQ_AVAIL(txq) - nsegs) <= MAX_TX_DESC(ctx) + 2) {
3349 		pi.ipi_flags |= IPI_TX_INTR;
3350 		txq->ift_rs_pending = 0;
3351 	}
3352 
3353 	pi.ipi_segs = segs;
3354 	pi.ipi_nsegs = nsegs;
3355 
3356 	MPASS(pidx >= 0 && pidx < txq->ift_size);
3357 #ifdef PKT_DEBUG
3358 	print_pkt(&pi);
3359 #endif
3360 	if (map != NULL)
3361 		bus_dmamap_sync(desc_tag, map, BUS_DMASYNC_PREWRITE);
3362 	if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) {
3363 		if (map != NULL)
3364 			bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3365 					BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
3366 		DBG_COUNTER_INC(tx_encap);
3367 		MPASS(pi.ipi_new_pidx < txq->ift_size);
3368 
3369 		ndesc = pi.ipi_new_pidx - pi.ipi_pidx;
3370 		if (pi.ipi_new_pidx < pi.ipi_pidx) {
3371 			ndesc += txq->ift_size;
3372 			txq->ift_gen = 1;
3373 		}
3374 		/*
3375 		 * drivers can need as many as
3376 		 * two sentinels
3377 		 */
3378 		MPASS(ndesc <= pi.ipi_nsegs + 2);
3379 		MPASS(pi.ipi_new_pidx != pidx);
3380 		MPASS(ndesc > 0);
3381 		txq->ift_in_use += ndesc;
3382 
3383 		/*
3384 		 * We update the last software descriptor again here because there may
3385 		 * be a sentinel and/or there may be more mbufs than segments
3386 		 */
3387 		txq->ift_pidx = pi.ipi_new_pidx;
3388 		txq->ift_npending += pi.ipi_ndescs;
3389 	} else {
3390 		*m_headp = m_head = iflib_remove_mbuf(txq);
3391 		if (err == EFBIG) {
3392 			txq->ift_txd_encap_efbig++;
3393 			if (remap < 2) {
3394 				remap = 1;
3395 				goto defrag;
3396 			}
3397 		}
3398 		DBG_COUNTER_INC(encap_txd_encap_fail);
3399 		goto defrag_failed;
3400 	}
3401 	return (err);
3402 
3403 defrag_failed:
3404 	txq->ift_mbuf_defrag_failed++;
3405 	txq->ift_map_failed++;
3406 	m_freem(*m_headp);
3407 	DBG_COUNTER_INC(tx_frees);
3408 	*m_headp = NULL;
3409 	return (ENOMEM);
3410 }
3411 
3412 static void
3413 iflib_tx_desc_free(iflib_txq_t txq, int n)
3414 {
3415 	int hasmap;
3416 	uint32_t qsize, cidx, mask, gen;
3417 	struct mbuf *m, **ifsd_m;
3418 	uint8_t *ifsd_flags;
3419 	bus_dmamap_t *ifsd_map;
3420 	bool do_prefetch;
3421 
3422 	cidx = txq->ift_cidx;
3423 	gen = txq->ift_gen;
3424 	qsize = txq->ift_size;
3425 	mask = qsize-1;
3426 	hasmap = txq->ift_sds.ifsd_map != NULL;
3427 	ifsd_flags = txq->ift_sds.ifsd_flags;
3428 	ifsd_m = txq->ift_sds.ifsd_m;
3429 	ifsd_map = txq->ift_sds.ifsd_map;
3430 	do_prefetch = (txq->ift_ctx->ifc_flags & IFC_PREFETCH);
3431 
3432 	while (n-- > 0) {
3433 		if (do_prefetch) {
3434 			prefetch(ifsd_m[(cidx + 3) & mask]);
3435 			prefetch(ifsd_m[(cidx + 4) & mask]);
3436 		}
3437 		if (ifsd_m[cidx] != NULL) {
3438 			prefetch(&ifsd_m[(cidx + CACHE_PTR_INCREMENT) & mask]);
3439 			prefetch(&ifsd_flags[(cidx + CACHE_PTR_INCREMENT) & mask]);
3440 			if (hasmap && (ifsd_flags[cidx] & TX_SW_DESC_MAPPED)) {
3441 				/*
3442 				 * does it matter if it's not the TSO tag? If so we'll
3443 				 * have to add the type to flags
3444 				 */
3445 				bus_dmamap_unload(txq->ift_desc_tag, ifsd_map[cidx]);
3446 				ifsd_flags[cidx] &= ~TX_SW_DESC_MAPPED;
3447 			}
3448 			if ((m = ifsd_m[cidx]) != NULL) {
3449 				/* XXX we don't support any drivers that batch packets yet */
3450 				MPASS(m->m_nextpkt == NULL);
3451 				/* if the number of clusters exceeds the number of segments
3452 				 * there won't be space on the ring to save a pointer to each
3453 				 * cluster so we simply free the list here
3454 				 */
3455 				if (m->m_flags & M_TOOBIG) {
3456 					m_freem(m);
3457 				} else {
3458 					m_free(m);
3459 				}
3460 				ifsd_m[cidx] = NULL;
3461 #if MEMORY_LOGGING
3462 				txq->ift_dequeued++;
3463 #endif
3464 				DBG_COUNTER_INC(tx_frees);
3465 			}
3466 		}
3467 		if (__predict_false(++cidx == qsize)) {
3468 			cidx = 0;
3469 			gen = 0;
3470 		}
3471 	}
3472 	txq->ift_cidx = cidx;
3473 	txq->ift_gen = gen;
3474 }
3475 
3476 static __inline int
3477 iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh)
3478 {
3479 	int reclaim;
3480 	if_ctx_t ctx = txq->ift_ctx;
3481 
3482 	KASSERT(thresh >= 0, ("invalid threshold to reclaim"));
3483 	MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size);
3484 
3485 	/*
3486 	 * Need a rate-limiting check so that this isn't called every time
3487 	 */
3488 	iflib_tx_credits_update(ctx, txq);
3489 	reclaim = DESC_RECLAIMABLE(txq);
3490 
3491 	if (reclaim <= thresh /* + MAX_TX_DESC(txq->ift_ctx) */) {
3492 #ifdef INVARIANTS
3493 		if (iflib_verbose_debug) {
3494 			printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __FUNCTION__,
3495 			       txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments,
3496 			       reclaim, thresh);
3497 
3498 		}
3499 #endif
3500 		return (0);
3501 	}
3502 	iflib_tx_desc_free(txq, reclaim);
3503 	txq->ift_cleaned += reclaim;
3504 	txq->ift_in_use -= reclaim;
3505 
3506 	return (reclaim);
3507 }
3508 
3509 static struct mbuf **
3510 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset, int remaining)
3511 {
3512 	int next, size;
3513 	struct mbuf **items;
3514 
3515 	size = r->size;
3516 	next = (cidx + CACHE_PTR_INCREMENT) & (size-1);
3517 	items = __DEVOLATILE(struct mbuf **, &r->items[0]);
3518 
3519 	prefetch(items[(cidx + offset) & (size-1)]);
3520 	if (remaining > 1) {
3521 		prefetch2cachelines(&items[next]);
3522 		prefetch2cachelines(items[(cidx + offset + 1) & (size-1)]);
3523 		prefetch2cachelines(items[(cidx + offset + 2) & (size-1)]);
3524 		prefetch2cachelines(items[(cidx + offset + 3) & (size-1)]);
3525 	}
3526 	return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (size-1)]));
3527 }
3528 
3529 static void
3530 iflib_txq_check_drain(iflib_txq_t txq, int budget)
3531 {
3532 
3533 	ifmp_ring_check_drainage(txq->ift_br, budget);
3534 }
3535 
3536 static uint32_t
3537 iflib_txq_can_drain(struct ifmp_ring *r)
3538 {
3539 	iflib_txq_t txq = r->cookie;
3540 	if_ctx_t ctx = txq->ift_ctx;
3541 
3542 	return ((TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2) ||
3543 		ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false));
3544 }
3545 
3546 static uint32_t
3547 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3548 {
3549 	iflib_txq_t txq = r->cookie;
3550 	if_ctx_t ctx = txq->ift_ctx;
3551 	struct ifnet *ifp = ctx->ifc_ifp;
3552 	struct mbuf **mp, *m;
3553 	int i, count, consumed, pkt_sent, bytes_sent, mcast_sent, avail;
3554 	int reclaimed, err, in_use_prev, desc_used;
3555 	bool do_prefetch, ring, rang;
3556 
3557 	if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) ||
3558 			    !LINK_ACTIVE(ctx))) {
3559 		DBG_COUNTER_INC(txq_drain_notready);
3560 		return (0);
3561 	}
3562 	reclaimed = iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
3563 	rang = iflib_txd_db_check(ctx, txq, reclaimed, txq->ift_in_use);
3564 	avail = IDXDIFF(pidx, cidx, r->size);
3565 	if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) {
3566 		DBG_COUNTER_INC(txq_drain_flushing);
3567 		for (i = 0; i < avail; i++) {
3568 			m_free(r->items[(cidx + i) & (r->size-1)]);
3569 			r->items[(cidx + i) & (r->size-1)] = NULL;
3570 		}
3571 		return (avail);
3572 	}
3573 
3574 	if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3575 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3576 		CALLOUT_LOCK(txq);
3577 		callout_stop(&txq->ift_timer);
3578 		CALLOUT_UNLOCK(txq);
3579 		DBG_COUNTER_INC(txq_drain_oactive);
3580 		return (0);
3581 	}
3582 	if (reclaimed)
3583 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3584 	consumed = mcast_sent = bytes_sent = pkt_sent = 0;
3585 	count = MIN(avail, TX_BATCH_SIZE);
3586 #ifdef INVARIANTS
3587 	if (iflib_verbose_debug)
3588 		printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __FUNCTION__,
3589 		       avail, ctx->ifc_flags, TXQ_AVAIL(txq));
3590 #endif
3591 	do_prefetch = (ctx->ifc_flags & IFC_PREFETCH);
3592 	avail = TXQ_AVAIL(txq);
3593 	err = 0;
3594 	for (desc_used = i = 0; i < count && avail > MAX_TX_DESC(ctx) + 2; i++) {
3595 		int rem = do_prefetch ? count - i : 0;
3596 
3597 		mp = _ring_peek_one(r, cidx, i, rem);
3598 		MPASS(mp != NULL && *mp != NULL);
3599 		if (__predict_false(*mp == (struct mbuf *)txq)) {
3600 			consumed++;
3601 			reclaimed++;
3602 			continue;
3603 		}
3604 		in_use_prev = txq->ift_in_use;
3605 		err = iflib_encap(txq, mp);
3606 		if (__predict_false(err)) {
3607 			DBG_COUNTER_INC(txq_drain_encapfail);
3608 			/* no room - bail out */
3609 			if (err == ENOBUFS)
3610 				break;
3611 			consumed++;
3612 			DBG_COUNTER_INC(txq_drain_encapfail);
3613 			/* we can't send this packet - skip it */
3614 			continue;
3615 		}
3616 		consumed++;
3617 		pkt_sent++;
3618 		m = *mp;
3619 		DBG_COUNTER_INC(tx_sent);
3620 		bytes_sent += m->m_pkthdr.len;
3621 		mcast_sent += !!(m->m_flags & M_MCAST);
3622 		avail = TXQ_AVAIL(txq);
3623 
3624 		txq->ift_db_pending += (txq->ift_in_use - in_use_prev);
3625 		desc_used += (txq->ift_in_use - in_use_prev);
3626 		ETHER_BPF_MTAP(ifp, m);
3627 		if (__predict_false(!(ifp->if_drv_flags & IFF_DRV_RUNNING)))
3628 			break;
3629 		rang = iflib_txd_db_check(ctx, txq, false, in_use_prev);
3630 	}
3631 
3632 	/* deliberate use of bitwise or to avoid gratuitous short-circuit */
3633 	ring = rang ? false  : (iflib_min_tx_latency | err) || (TXQ_AVAIL(txq) < MAX_TX_DESC(ctx));
3634 	iflib_txd_db_check(ctx, txq, ring, txq->ift_in_use);
3635 	if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
3636 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
3637 	if (mcast_sent)
3638 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
3639 #ifdef INVARIANTS
3640 	if (iflib_verbose_debug)
3641 		printf("consumed=%d\n", consumed);
3642 #endif
3643 	return (consumed);
3644 }
3645 
3646 static uint32_t
3647 iflib_txq_drain_always(struct ifmp_ring *r)
3648 {
3649 	return (1);
3650 }
3651 
3652 static uint32_t
3653 iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3654 {
3655 	int i, avail;
3656 	struct mbuf **mp;
3657 	iflib_txq_t txq;
3658 
3659 	txq = r->cookie;
3660 
3661 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3662 	CALLOUT_LOCK(txq);
3663 	callout_stop(&txq->ift_timer);
3664 	CALLOUT_UNLOCK(txq);
3665 
3666 	avail = IDXDIFF(pidx, cidx, r->size);
3667 	for (i = 0; i < avail; i++) {
3668 		mp = _ring_peek_one(r, cidx, i, avail - i);
3669 		if (__predict_false(*mp == (struct mbuf *)txq))
3670 			continue;
3671 		m_freem(*mp);
3672 	}
3673 	MPASS(ifmp_ring_is_stalled(r) == 0);
3674 	return (avail);
3675 }
3676 
3677 static void
3678 iflib_ifmp_purge(iflib_txq_t txq)
3679 {
3680 	struct ifmp_ring *r;
3681 
3682 	r = txq->ift_br;
3683 	r->drain = iflib_txq_drain_free;
3684 	r->can_drain = iflib_txq_drain_always;
3685 
3686 	ifmp_ring_check_drainage(r, r->size);
3687 
3688 	r->drain = iflib_txq_drain;
3689 	r->can_drain = iflib_txq_can_drain;
3690 }
3691 
3692 static void
3693 _task_fn_tx(void *context)
3694 {
3695 	iflib_txq_t txq = context;
3696 	if_ctx_t ctx = txq->ift_ctx;
3697 	struct ifnet *ifp = ctx->ifc_ifp;
3698 
3699 #ifdef IFLIB_DIAGNOSTICS
3700 	txq->ift_cpu_exec_count[curcpu]++;
3701 #endif
3702 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3703 		return;
3704 	if (if_getcapenable(ifp) & IFCAP_NETMAP) {
3705 		if (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false))
3706 			netmap_tx_irq(ifp, txq->ift_id);
3707 		IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3708 		return;
3709 	}
3710 	if (txq->ift_db_pending)
3711 		ifmp_ring_enqueue(txq->ift_br, (void **)&txq, 1, TX_BATCH_SIZE);
3712 	ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3713 	if (ctx->ifc_flags & IFC_LEGACY)
3714 		IFDI_INTR_ENABLE(ctx);
3715 	else {
3716 #ifdef INVARIANTS
3717 		int rc =
3718 #endif
3719 			IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3720 			KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver"));
3721 	}
3722 }
3723 
3724 static void
3725 _task_fn_rx(void *context)
3726 {
3727 	iflib_rxq_t rxq = context;
3728 	if_ctx_t ctx = rxq->ifr_ctx;
3729 	bool more;
3730 	uint16_t budget;
3731 
3732 #ifdef IFLIB_DIAGNOSTICS
3733 	rxq->ifr_cpu_exec_count[curcpu]++;
3734 #endif
3735 	DBG_COUNTER_INC(task_fn_rxs);
3736 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3737 		return;
3738 	more = true;
3739 #ifdef DEV_NETMAP
3740 	if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP) {
3741 		u_int work = 0;
3742 		if (netmap_rx_irq(ctx->ifc_ifp, rxq->ifr_id, &work)) {
3743 			more = false;
3744 		}
3745 	}
3746 #endif
3747 	budget = ctx->ifc_sysctl_rx_budget;
3748 	if (budget == 0)
3749 		budget = 16;	/* XXX */
3750 	if (more == false || (more = iflib_rxeof(rxq, budget)) == false) {
3751 		if (ctx->ifc_flags & IFC_LEGACY)
3752 			IFDI_INTR_ENABLE(ctx);
3753 		else {
3754 #ifdef INVARIANTS
3755 			int rc =
3756 #endif
3757 				IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
3758 			KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver"));
3759 			DBG_COUNTER_INC(rx_intr_enables);
3760 		}
3761 	}
3762 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3763 		return;
3764 	if (more)
3765 		GROUPTASK_ENQUEUE(&rxq->ifr_task);
3766 }
3767 
3768 static void
3769 _task_fn_admin(void *context)
3770 {
3771 	if_ctx_t ctx = context;
3772 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
3773 	iflib_txq_t txq;
3774 	int i;
3775 	bool oactive, running, do_reset, do_watchdog;
3776 
3777 	STATE_LOCK(ctx);
3778 	running = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING);
3779 	oactive = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE);
3780 	do_reset = (ctx->ifc_flags & IFC_DO_RESET);
3781 	do_watchdog = (ctx->ifc_flags & IFC_DO_WATCHDOG);
3782 	ctx->ifc_flags &= ~(IFC_DO_RESET|IFC_DO_WATCHDOG);
3783 	STATE_UNLOCK(ctx);
3784 
3785 	if (!running & !oactive)
3786 		return;
3787 
3788 	CTX_LOCK(ctx);
3789 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3790 		CALLOUT_LOCK(txq);
3791 		callout_stop(&txq->ift_timer);
3792 		CALLOUT_UNLOCK(txq);
3793 	}
3794 	if (do_watchdog) {
3795 		ctx->ifc_watchdog_events++;
3796 		IFDI_WATCHDOG_RESET(ctx);
3797 	}
3798 	IFDI_UPDATE_ADMIN_STATUS(ctx);
3799 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3800 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq, txq->ift_timer.c_cpu);
3801 	IFDI_LINK_INTR_ENABLE(ctx);
3802 	if (do_reset)
3803 		iflib_if_init_locked(ctx);
3804 	CTX_UNLOCK(ctx);
3805 
3806 	if (LINK_ACTIVE(ctx) == 0)
3807 		return;
3808 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3809 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
3810 }
3811 
3812 
3813 static void
3814 _task_fn_iov(void *context)
3815 {
3816 	if_ctx_t ctx = context;
3817 
3818 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3819 		return;
3820 
3821 	CTX_LOCK(ctx);
3822 	IFDI_VFLR_HANDLE(ctx);
3823 	CTX_UNLOCK(ctx);
3824 }
3825 
3826 static int
3827 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)
3828 {
3829 	int err;
3830 	if_int_delay_info_t info;
3831 	if_ctx_t ctx;
3832 
3833 	info = (if_int_delay_info_t)arg1;
3834 	ctx = info->iidi_ctx;
3835 	info->iidi_req = req;
3836 	info->iidi_oidp = oidp;
3837 	CTX_LOCK(ctx);
3838 	err = IFDI_SYSCTL_INT_DELAY(ctx, info);
3839 	CTX_UNLOCK(ctx);
3840 	return (err);
3841 }
3842 
3843 /*********************************************************************
3844  *
3845  *  IFNET FUNCTIONS
3846  *
3847  **********************************************************************/
3848 
3849 static void
3850 iflib_if_init_locked(if_ctx_t ctx)
3851 {
3852 	iflib_stop(ctx);
3853 	iflib_init_locked(ctx);
3854 }
3855 
3856 
3857 static void
3858 iflib_if_init(void *arg)
3859 {
3860 	if_ctx_t ctx = arg;
3861 
3862 	CTX_LOCK(ctx);
3863 	iflib_if_init_locked(ctx);
3864 	CTX_UNLOCK(ctx);
3865 }
3866 
3867 static int
3868 iflib_if_transmit(if_t ifp, struct mbuf *m)
3869 {
3870 	if_ctx_t	ctx = if_getsoftc(ifp);
3871 
3872 	iflib_txq_t txq;
3873 	int err, qidx;
3874 
3875 	if (__predict_false((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) {
3876 		DBG_COUNTER_INC(tx_frees);
3877 		m_freem(m);
3878 		return (ENOBUFS);
3879 	}
3880 
3881 	MPASS(m->m_nextpkt == NULL);
3882 	qidx = 0;
3883 	if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m))
3884 		qidx = QIDX(ctx, m);
3885 	/*
3886 	 * XXX calculate buf_ring based on flowid (divvy up bits?)
3887 	 */
3888 	txq = &ctx->ifc_txqs[qidx];
3889 
3890 #ifdef DRIVER_BACKPRESSURE
3891 	if (txq->ift_closed) {
3892 		while (m != NULL) {
3893 			next = m->m_nextpkt;
3894 			m->m_nextpkt = NULL;
3895 			m_freem(m);
3896 			m = next;
3897 		}
3898 		return (ENOBUFS);
3899 	}
3900 #endif
3901 #ifdef notyet
3902 	qidx = count = 0;
3903 	mp = marr;
3904 	next = m;
3905 	do {
3906 		count++;
3907 		next = next->m_nextpkt;
3908 	} while (next != NULL);
3909 
3910 	if (count > nitems(marr))
3911 		if ((mp = malloc(count*sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) {
3912 			/* XXX check nextpkt */
3913 			m_freem(m);
3914 			/* XXX simplify for now */
3915 			DBG_COUNTER_INC(tx_frees);
3916 			return (ENOBUFS);
3917 		}
3918 	for (next = m, i = 0; next != NULL; i++) {
3919 		mp[i] = next;
3920 		next = next->m_nextpkt;
3921 		mp[i]->m_nextpkt = NULL;
3922 	}
3923 #endif
3924 	DBG_COUNTER_INC(tx_seen);
3925 	err = ifmp_ring_enqueue(txq->ift_br, (void **)&m, 1, TX_BATCH_SIZE);
3926 
3927 	GROUPTASK_ENQUEUE(&txq->ift_task);
3928 	if (err) {
3929 		/* support forthcoming later */
3930 #ifdef DRIVER_BACKPRESSURE
3931 		txq->ift_closed = TRUE;
3932 #endif
3933 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3934 		m_freem(m);
3935 	}
3936 
3937 	return (err);
3938 }
3939 
3940 static void
3941 iflib_if_qflush(if_t ifp)
3942 {
3943 	if_ctx_t ctx = if_getsoftc(ifp);
3944 	iflib_txq_t txq = ctx->ifc_txqs;
3945 	int i;
3946 
3947 	STATE_LOCK(ctx);
3948 	ctx->ifc_flags |= IFC_QFLUSH;
3949 	STATE_UNLOCK(ctx);
3950 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
3951 		while (!(ifmp_ring_is_idle(txq->ift_br) || ifmp_ring_is_stalled(txq->ift_br)))
3952 			iflib_txq_check_drain(txq, 0);
3953 	STATE_LOCK(ctx);
3954 	ctx->ifc_flags &= ~IFC_QFLUSH;
3955 	STATE_UNLOCK(ctx);
3956 
3957 	if_qflush(ifp);
3958 }
3959 
3960 
3961 #define IFCAP_FLAGS (IFCAP_TXCSUM_IPV6 | IFCAP_RXCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \
3962 		     IFCAP_TSO4 | IFCAP_TSO6 | IFCAP_VLAN_HWTAGGING | IFCAP_HWSTATS | \
3963 		     IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | IFCAP_VLAN_HWTSO)
3964 
3965 static int
3966 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data)
3967 {
3968 	if_ctx_t ctx = if_getsoftc(ifp);
3969 	struct ifreq	*ifr = (struct ifreq *)data;
3970 #if defined(INET) || defined(INET6)
3971 	struct ifaddr	*ifa = (struct ifaddr *)data;
3972 #endif
3973 	bool		avoid_reset = FALSE;
3974 	int		err = 0, reinit = 0, bits;
3975 
3976 	switch (command) {
3977 	case SIOCSIFADDR:
3978 #ifdef INET
3979 		if (ifa->ifa_addr->sa_family == AF_INET)
3980 			avoid_reset = TRUE;
3981 #endif
3982 #ifdef INET6
3983 		if (ifa->ifa_addr->sa_family == AF_INET6)
3984 			avoid_reset = TRUE;
3985 #endif
3986 		/*
3987 		** Calling init results in link renegotiation,
3988 		** so we avoid doing it when possible.
3989 		*/
3990 		if (avoid_reset) {
3991 			if_setflagbits(ifp, IFF_UP,0);
3992 			if (!(if_getdrvflags(ifp)& IFF_DRV_RUNNING))
3993 				reinit = 1;
3994 #ifdef INET
3995 			if (!(if_getflags(ifp) & IFF_NOARP))
3996 				arp_ifinit(ifp, ifa);
3997 #endif
3998 		} else
3999 			err = ether_ioctl(ifp, command, data);
4000 		break;
4001 	case SIOCSIFMTU:
4002 		CTX_LOCK(ctx);
4003 		if (ifr->ifr_mtu == if_getmtu(ifp)) {
4004 			CTX_UNLOCK(ctx);
4005 			break;
4006 		}
4007 		bits = if_getdrvflags(ifp);
4008 		/* stop the driver and free any clusters before proceeding */
4009 		iflib_stop(ctx);
4010 
4011 		if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) {
4012 			STATE_LOCK(ctx);
4013 			if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size)
4014 				ctx->ifc_flags |= IFC_MULTISEG;
4015 			else
4016 				ctx->ifc_flags &= ~IFC_MULTISEG;
4017 			STATE_UNLOCK(ctx);
4018 			err = if_setmtu(ifp, ifr->ifr_mtu);
4019 		}
4020 		iflib_init_locked(ctx);
4021 		STATE_LOCK(ctx);
4022 		if_setdrvflags(ifp, bits);
4023 		STATE_UNLOCK(ctx);
4024 		CTX_UNLOCK(ctx);
4025 		break;
4026 	case SIOCSIFFLAGS:
4027 		CTX_LOCK(ctx);
4028 		if (if_getflags(ifp) & IFF_UP) {
4029 			if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4030 				if ((if_getflags(ifp) ^ ctx->ifc_if_flags) &
4031 				    (IFF_PROMISC | IFF_ALLMULTI)) {
4032 					err = IFDI_PROMISC_SET(ctx, if_getflags(ifp));
4033 				}
4034 			} else
4035 				reinit = 1;
4036 		} else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4037 			iflib_stop(ctx);
4038 		}
4039 		ctx->ifc_if_flags = if_getflags(ifp);
4040 		CTX_UNLOCK(ctx);
4041 		break;
4042 	case SIOCADDMULTI:
4043 	case SIOCDELMULTI:
4044 		if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4045 			CTX_LOCK(ctx);
4046 			IFDI_INTR_DISABLE(ctx);
4047 			IFDI_MULTI_SET(ctx);
4048 			IFDI_INTR_ENABLE(ctx);
4049 			CTX_UNLOCK(ctx);
4050 		}
4051 		break;
4052 	case SIOCSIFMEDIA:
4053 		CTX_LOCK(ctx);
4054 		IFDI_MEDIA_SET(ctx);
4055 		CTX_UNLOCK(ctx);
4056 		/* falls thru */
4057 	case SIOCGIFMEDIA:
4058 	case SIOCGIFXMEDIA:
4059 		err = ifmedia_ioctl(ifp, ifr, &ctx->ifc_media, command);
4060 		break;
4061 	case SIOCGI2C:
4062 	{
4063 		struct ifi2creq i2c;
4064 
4065 		err = copyin(ifr_data_get_ptr(ifr), &i2c, sizeof(i2c));
4066 		if (err != 0)
4067 			break;
4068 		if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) {
4069 			err = EINVAL;
4070 			break;
4071 		}
4072 		if (i2c.len > sizeof(i2c.data)) {
4073 			err = EINVAL;
4074 			break;
4075 		}
4076 
4077 		if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0)
4078 			err = copyout(&i2c, ifr_data_get_ptr(ifr),
4079 			    sizeof(i2c));
4080 		break;
4081 	}
4082 	case SIOCSIFCAP:
4083 	{
4084 		int mask, setmask;
4085 
4086 		mask = ifr->ifr_reqcap ^ if_getcapenable(ifp);
4087 		setmask = 0;
4088 #ifdef TCP_OFFLOAD
4089 		setmask |= mask & (IFCAP_TOE4|IFCAP_TOE6);
4090 #endif
4091 		setmask |= (mask & IFCAP_FLAGS);
4092 
4093 		if (setmask  & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6))
4094 			setmask |= (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6);
4095 		if ((mask & IFCAP_WOL) &&
4096 		    (if_getcapabilities(ifp) & IFCAP_WOL) != 0)
4097 			setmask |= (mask & (IFCAP_WOL_MCAST|IFCAP_WOL_MAGIC));
4098 		if_vlancap(ifp);
4099 		/*
4100 		 * want to ensure that traffic has stopped before we change any of the flags
4101 		 */
4102 		if (setmask) {
4103 			CTX_LOCK(ctx);
4104 			bits = if_getdrvflags(ifp);
4105 			if (bits & IFF_DRV_RUNNING)
4106 				iflib_stop(ctx);
4107 			STATE_LOCK(ctx);
4108 			if_togglecapenable(ifp, setmask);
4109 			STATE_UNLOCK(ctx);
4110 			if (bits & IFF_DRV_RUNNING)
4111 				iflib_init_locked(ctx);
4112 			STATE_LOCK(ctx);
4113 			if_setdrvflags(ifp, bits);
4114 			STATE_UNLOCK(ctx);
4115 			CTX_UNLOCK(ctx);
4116 		}
4117 		break;
4118 	    }
4119 	case SIOCGPRIVATE_0:
4120 	case SIOCSDRVSPEC:
4121 	case SIOCGDRVSPEC:
4122 		CTX_LOCK(ctx);
4123 		err = IFDI_PRIV_IOCTL(ctx, command, data);
4124 		CTX_UNLOCK(ctx);
4125 		break;
4126 	default:
4127 		err = ether_ioctl(ifp, command, data);
4128 		break;
4129 	}
4130 	if (reinit)
4131 		iflib_if_init(ctx);
4132 	return (err);
4133 }
4134 
4135 static uint64_t
4136 iflib_if_get_counter(if_t ifp, ift_counter cnt)
4137 {
4138 	if_ctx_t ctx = if_getsoftc(ifp);
4139 
4140 	return (IFDI_GET_COUNTER(ctx, cnt));
4141 }
4142 
4143 /*********************************************************************
4144  *
4145  *  OTHER FUNCTIONS EXPORTED TO THE STACK
4146  *
4147  **********************************************************************/
4148 
4149 static void
4150 iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag)
4151 {
4152 	if_ctx_t ctx = if_getsoftc(ifp);
4153 
4154 	if ((void *)ctx != arg)
4155 		return;
4156 
4157 	if ((vtag == 0) || (vtag > 4095))
4158 		return;
4159 
4160 	CTX_LOCK(ctx);
4161 	IFDI_VLAN_REGISTER(ctx, vtag);
4162 	/* Re-init to load the changes */
4163 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
4164 		iflib_if_init_locked(ctx);
4165 	CTX_UNLOCK(ctx);
4166 }
4167 
4168 static void
4169 iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag)
4170 {
4171 	if_ctx_t ctx = if_getsoftc(ifp);
4172 
4173 	if ((void *)ctx != arg)
4174 		return;
4175 
4176 	if ((vtag == 0) || (vtag > 4095))
4177 		return;
4178 
4179 	CTX_LOCK(ctx);
4180 	IFDI_VLAN_UNREGISTER(ctx, vtag);
4181 	/* Re-init to load the changes */
4182 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
4183 		iflib_if_init_locked(ctx);
4184 	CTX_UNLOCK(ctx);
4185 }
4186 
4187 static void
4188 iflib_led_func(void *arg, int onoff)
4189 {
4190 	if_ctx_t ctx = arg;
4191 
4192 	CTX_LOCK(ctx);
4193 	IFDI_LED_FUNC(ctx, onoff);
4194 	CTX_UNLOCK(ctx);
4195 }
4196 
4197 /*********************************************************************
4198  *
4199  *  BUS FUNCTION DEFINITIONS
4200  *
4201  **********************************************************************/
4202 
4203 int
4204 iflib_device_probe(device_t dev)
4205 {
4206 	pci_vendor_info_t *ent;
4207 
4208 	uint16_t	pci_vendor_id, pci_device_id;
4209 	uint16_t	pci_subvendor_id, pci_subdevice_id;
4210 	uint16_t	pci_rev_id;
4211 	if_shared_ctx_t sctx;
4212 
4213 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4214 		return (ENOTSUP);
4215 
4216 	pci_vendor_id = pci_get_vendor(dev);
4217 	pci_device_id = pci_get_device(dev);
4218 	pci_subvendor_id = pci_get_subvendor(dev);
4219 	pci_subdevice_id = pci_get_subdevice(dev);
4220 	pci_rev_id = pci_get_revid(dev);
4221 	if (sctx->isc_parse_devinfo != NULL)
4222 		sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id);
4223 
4224 	ent = sctx->isc_vendor_info;
4225 	while (ent->pvi_vendor_id != 0) {
4226 		if (pci_vendor_id != ent->pvi_vendor_id) {
4227 			ent++;
4228 			continue;
4229 		}
4230 		if ((pci_device_id == ent->pvi_device_id) &&
4231 		    ((pci_subvendor_id == ent->pvi_subvendor_id) ||
4232 		     (ent->pvi_subvendor_id == 0)) &&
4233 		    ((pci_subdevice_id == ent->pvi_subdevice_id) ||
4234 		     (ent->pvi_subdevice_id == 0)) &&
4235 		    ((pci_rev_id == ent->pvi_rev_id) ||
4236 		     (ent->pvi_rev_id == 0))) {
4237 
4238 			device_set_desc_copy(dev, ent->pvi_name);
4239 			/* this needs to be changed to zero if the bus probing code
4240 			 * ever stops re-probing on best match because the sctx
4241 			 * may have its values over written by register calls
4242 			 * in subsequent probes
4243 			 */
4244 			return (BUS_PROBE_DEFAULT);
4245 		}
4246 		ent++;
4247 	}
4248 	return (ENXIO);
4249 }
4250 
4251 static void
4252 iflib_reset_qvalues(if_ctx_t ctx)
4253 {
4254 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4255 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4256 	device_t dev = ctx->ifc_dev;
4257 	int i, main_txq, main_rxq;
4258 
4259 	main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4260 	main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4261 
4262 	scctx->isc_txrx_budget_bytes_max = IFLIB_MAX_TX_BYTES;
4263 	scctx->isc_tx_qdepth = IFLIB_DEFAULT_TX_QDEPTH;
4264 	/*
4265 	 * XXX sanity check that ntxd & nrxd are a power of 2
4266 	 */
4267 	if (ctx->ifc_sysctl_ntxqs != 0)
4268 		scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs;
4269 	if (ctx->ifc_sysctl_nrxqs != 0)
4270 		scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs;
4271 
4272 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4273 		if (ctx->ifc_sysctl_ntxds[i] != 0)
4274 			scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i];
4275 		else
4276 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4277 	}
4278 
4279 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4280 		if (ctx->ifc_sysctl_nrxds[i] != 0)
4281 			scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i];
4282 		else
4283 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4284 	}
4285 
4286 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4287 		if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) {
4288 			device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n",
4289 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]);
4290 			scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i];
4291 		}
4292 		if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) {
4293 			device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n",
4294 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]);
4295 			scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i];
4296 		}
4297 	}
4298 
4299 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4300 		if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) {
4301 			device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n",
4302 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]);
4303 			scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i];
4304 		}
4305 		if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) {
4306 			device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n",
4307 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]);
4308 			scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i];
4309 		}
4310 	}
4311 }
4312 
4313 int
4314 iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp)
4315 {
4316 	int err, rid, msix;
4317 	if_ctx_t ctx;
4318 	if_t ifp;
4319 	if_softc_ctx_t scctx;
4320 	int i;
4321 	uint16_t main_txq;
4322 	uint16_t main_rxq;
4323 
4324 
4325 	ctx = malloc(sizeof(* ctx), M_IFLIB, M_WAITOK|M_ZERO);
4326 
4327 	if (sc == NULL) {
4328 		sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4329 		device_set_softc(dev, ctx);
4330 		ctx->ifc_flags |= IFC_SC_ALLOCATED;
4331 	}
4332 
4333 	ctx->ifc_sctx = sctx;
4334 	ctx->ifc_dev = dev;
4335 	ctx->ifc_softc = sc;
4336 
4337 	if ((err = iflib_register(ctx)) != 0) {
4338 		if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4339 			free(sc, M_IFLIB);
4340 		free(ctx, M_IFLIB);
4341 		device_printf(dev, "iflib_register failed %d\n", err);
4342 		return (err);
4343 	}
4344 	iflib_add_device_sysctl_pre(ctx);
4345 
4346 	scctx = &ctx->ifc_softc_ctx;
4347 	ifp = ctx->ifc_ifp;
4348 
4349 	iflib_reset_qvalues(ctx);
4350 	CTX_LOCK(ctx);
4351 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4352 		CTX_UNLOCK(ctx);
4353 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4354 		return (err);
4355 	}
4356 	_iflib_pre_assert(scctx);
4357 	ctx->ifc_txrx = *scctx->isc_txrx;
4358 
4359 #ifdef INVARIANTS
4360 	MPASS(scctx->isc_capenable);
4361 	if (scctx->isc_capenable & IFCAP_TXCSUM)
4362 		MPASS(scctx->isc_tx_csum_flags);
4363 #endif
4364 
4365 	if_setcapabilities(ifp, scctx->isc_capenable | IFCAP_HWSTATS);
4366 	if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS);
4367 
4368 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4369 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4370 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4371 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4372 
4373 #ifdef ACPI_DMAR
4374 	if (dmar_get_dma_tag(device_get_parent(dev), dev) != NULL)
4375 		ctx->ifc_flags |= IFC_DMAR;
4376 #elif !(defined(__i386__) || defined(__amd64__))
4377 	/* set unconditionally for !x86 */
4378 	ctx->ifc_flags |= IFC_DMAR;
4379 #endif
4380 
4381 	main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4382 	main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4383 
4384 	/* XXX change for per-queue sizes */
4385 	device_printf(dev, "using %d tx descriptors and %d rx descriptors\n",
4386 		      scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4387 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4388 		if (!powerof2(scctx->isc_nrxd[i])) {
4389 			/* round down instead? */
4390 			device_printf(dev, "# rx descriptors must be a power of 2\n");
4391 			err = EINVAL;
4392 			goto fail;
4393 		}
4394 	}
4395 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4396 		if (!powerof2(scctx->isc_ntxd[i])) {
4397 			device_printf(dev,
4398 			    "# tx descriptors must be a power of 2");
4399 			err = EINVAL;
4400 			goto fail;
4401 		}
4402 	}
4403 
4404 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4405 	    MAX_SINGLE_PACKET_FRACTION)
4406 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4407 		    MAX_SINGLE_PACKET_FRACTION);
4408 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4409 	    MAX_SINGLE_PACKET_FRACTION)
4410 		scctx->isc_tx_tso_segments_max = max(1,
4411 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4412 
4413 	/*
4414 	 * Protect the stack against modern hardware
4415 	 */
4416 	if (scctx->isc_tx_tso_size_max > FREEBSD_TSO_SIZE_MAX)
4417 		scctx->isc_tx_tso_size_max = FREEBSD_TSO_SIZE_MAX;
4418 
4419 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4420 	ifp->if_hw_tsomaxsegcount = scctx->isc_tx_tso_segments_max;
4421 	ifp->if_hw_tsomax = scctx->isc_tx_tso_size_max;
4422 	ifp->if_hw_tsomaxsegsize = scctx->isc_tx_tso_segsize_max;
4423 	if (scctx->isc_rss_table_size == 0)
4424 		scctx->isc_rss_table_size = 64;
4425 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4426 
4427 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4428 	/* XXX format name */
4429 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx, -1, "admin");
4430 
4431 	/* Set up cpu set.  If it fails, use the set of all CPUs. */
4432 	if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) != 0) {
4433 		device_printf(dev, "Unable to fetch CPU list\n");
4434 		CPU_COPY(&all_cpus, &ctx->ifc_cpus);
4435 	}
4436 	MPASS(CPU_COUNT(&ctx->ifc_cpus) > 0);
4437 
4438 	/*
4439 	** Now setup MSI or MSI/X, should
4440 	** return us the number of supported
4441 	** vectors. (Will be 1 for MSI)
4442 	*/
4443 	if (sctx->isc_flags & IFLIB_SKIP_MSIX) {
4444 		msix = scctx->isc_vectors;
4445 	} else if (scctx->isc_msix_bar != 0)
4446 	       /*
4447 		* The simple fact that isc_msix_bar is not 0 does not mean we
4448 		* we have a good value there that is known to work.
4449 		*/
4450 		msix = iflib_msix_init(ctx);
4451 	else {
4452 		scctx->isc_vectors = 1;
4453 		scctx->isc_ntxqsets = 1;
4454 		scctx->isc_nrxqsets = 1;
4455 		scctx->isc_intr = IFLIB_INTR_LEGACY;
4456 		msix = 0;
4457 	}
4458 	/* Get memory for the station queues */
4459 	if ((err = iflib_queues_alloc(ctx))) {
4460 		device_printf(dev, "Unable to allocate queue memory\n");
4461 		goto fail;
4462 	}
4463 
4464 	if ((err = iflib_qset_structures_setup(ctx)))
4465 		goto fail_queues;
4466 
4467 	/*
4468 	 * Group taskqueues aren't properly set up until SMP is started,
4469 	 * so we disable interrupts until we can handle them post
4470 	 * SI_SUB_SMP.
4471 	 *
4472 	 * XXX: disabling interrupts doesn't actually work, at least for
4473 	 * the non-MSI case.  When they occur before SI_SUB_SMP completes,
4474 	 * we do null handling and depend on this not causing too large an
4475 	 * interrupt storm.
4476 	 */
4477 	IFDI_INTR_DISABLE(ctx);
4478 	if (msix > 1 && (err = IFDI_MSIX_INTR_ASSIGN(ctx, msix)) != 0) {
4479 		device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n", err);
4480 		goto fail_intr_free;
4481 	}
4482 	if (msix <= 1) {
4483 		rid = 0;
4484 		if (scctx->isc_intr == IFLIB_INTR_MSI) {
4485 			MPASS(msix == 1);
4486 			rid = 1;
4487 		}
4488 		if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) {
4489 			device_printf(dev, "iflib_legacy_setup failed %d\n", err);
4490 			goto fail_intr_free;
4491 		}
4492 	}
4493 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac);
4494 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4495 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4496 		goto fail_detach;
4497 	}
4498 	if ((err = iflib_netmap_attach(ctx))) {
4499 		device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err);
4500 		goto fail_detach;
4501 	}
4502 	*ctxp = ctx;
4503 
4504 	NETDUMP_SET(ctx->ifc_ifp, iflib);
4505 
4506 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4507 	iflib_add_device_sysctl_post(ctx);
4508 	ctx->ifc_flags |= IFC_INIT_DONE;
4509 	CTX_UNLOCK(ctx);
4510 	return (0);
4511 fail_detach:
4512 	ether_ifdetach(ctx->ifc_ifp);
4513 fail_intr_free:
4514 	if (scctx->isc_intr == IFLIB_INTR_MSIX || scctx->isc_intr == IFLIB_INTR_MSI)
4515 		pci_release_msi(ctx->ifc_dev);
4516 fail_queues:
4517 	iflib_tx_structures_free(ctx);
4518 	iflib_rx_structures_free(ctx);
4519 fail:
4520 	IFDI_DETACH(ctx);
4521 	CTX_UNLOCK(ctx);
4522 	return (err);
4523 }
4524 
4525 int
4526 iflib_pseudo_register(device_t dev, if_shared_ctx_t sctx, if_ctx_t *ctxp,
4527 					  struct iflib_cloneattach_ctx *clctx)
4528 {
4529 	int err;
4530 	if_ctx_t ctx;
4531 	if_t ifp;
4532 	if_softc_ctx_t scctx;
4533 	int i;
4534 	void *sc;
4535 	uint16_t main_txq;
4536 	uint16_t main_rxq;
4537 
4538 	ctx = malloc(sizeof(*ctx), M_IFLIB, M_WAITOK|M_ZERO);
4539 	sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4540 	ctx->ifc_flags |= IFC_SC_ALLOCATED;
4541 	if (sctx->isc_flags & (IFLIB_PSEUDO|IFLIB_VIRTUAL))
4542 		ctx->ifc_flags |= IFC_PSEUDO;
4543 
4544 	ctx->ifc_sctx = sctx;
4545 	ctx->ifc_softc = sc;
4546 	ctx->ifc_dev = dev;
4547 
4548 	if ((err = iflib_register(ctx)) != 0) {
4549 		device_printf(dev, "%s: iflib_register failed %d\n", __func__, err);
4550 		free(sc, M_IFLIB);
4551 		free(ctx, M_IFLIB);
4552 		return (err);
4553 	}
4554 	iflib_add_device_sysctl_pre(ctx);
4555 
4556 	scctx = &ctx->ifc_softc_ctx;
4557 	ifp = ctx->ifc_ifp;
4558 
4559 	/*
4560 	 * XXX sanity check that ntxd & nrxd are a power of 2
4561 	 */
4562 	iflib_reset_qvalues(ctx);
4563 
4564 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4565 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4566 		return (err);
4567 	}
4568 	if (sctx->isc_flags & IFLIB_GEN_MAC)
4569 		iflib_gen_mac(ctx);
4570 	if ((err = IFDI_CLONEATTACH(ctx, clctx->cc_ifc, clctx->cc_name,
4571 								clctx->cc_params)) != 0) {
4572 		device_printf(dev, "IFDI_CLONEATTACH failed %d\n", err);
4573 		return (err);
4574 	}
4575 	ifmedia_add(&ctx->ifc_media, IFM_ETHER | IFM_1000_T | IFM_FDX, 0, NULL);
4576 	ifmedia_add(&ctx->ifc_media, IFM_ETHER | IFM_AUTO, 0, NULL);
4577 	ifmedia_set(&ctx->ifc_media, IFM_ETHER | IFM_AUTO);
4578 
4579 #ifdef INVARIANTS
4580 	MPASS(scctx->isc_capenable);
4581 	if (scctx->isc_capenable & IFCAP_TXCSUM)
4582 		MPASS(scctx->isc_tx_csum_flags);
4583 #endif
4584 
4585 	if_setcapabilities(ifp, scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_LINKSTATE);
4586 	if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_LINKSTATE);
4587 
4588 	ifp->if_flags |= IFF_NOGROUP;
4589 	if (sctx->isc_flags & IFLIB_PSEUDO) {
4590 		ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac);
4591 
4592 		if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4593 			device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4594 			goto fail_detach;
4595 		}
4596 		*ctxp = ctx;
4597 
4598 		if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4599 		iflib_add_device_sysctl_post(ctx);
4600 		ctx->ifc_flags |= IFC_INIT_DONE;
4601 		return (0);
4602 	}
4603 	_iflib_pre_assert(scctx);
4604 	ctx->ifc_txrx = *scctx->isc_txrx;
4605 
4606 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4607 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4608 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4609 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4610 
4611 	main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4612 	main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4613 
4614 	/* XXX change for per-queue sizes */
4615 	device_printf(dev, "using %d tx descriptors and %d rx descriptors\n",
4616 		      scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4617 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4618 		if (!powerof2(scctx->isc_nrxd[i])) {
4619 			/* round down instead? */
4620 			device_printf(dev, "# rx descriptors must be a power of 2\n");
4621 			err = EINVAL;
4622 			goto fail;
4623 		}
4624 	}
4625 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4626 		if (!powerof2(scctx->isc_ntxd[i])) {
4627 			device_printf(dev,
4628 			    "# tx descriptors must be a power of 2");
4629 			err = EINVAL;
4630 			goto fail;
4631 		}
4632 	}
4633 
4634 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4635 	    MAX_SINGLE_PACKET_FRACTION)
4636 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4637 		    MAX_SINGLE_PACKET_FRACTION);
4638 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4639 	    MAX_SINGLE_PACKET_FRACTION)
4640 		scctx->isc_tx_tso_segments_max = max(1,
4641 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4642 
4643 	/*
4644 	 * Protect the stack against modern hardware
4645 	 */
4646 	if (scctx->isc_tx_tso_size_max > FREEBSD_TSO_SIZE_MAX)
4647 		scctx->isc_tx_tso_size_max = FREEBSD_TSO_SIZE_MAX;
4648 
4649 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4650 	ifp->if_hw_tsomaxsegcount = scctx->isc_tx_tso_segments_max;
4651 	ifp->if_hw_tsomax = scctx->isc_tx_tso_size_max;
4652 	ifp->if_hw_tsomaxsegsize = scctx->isc_tx_tso_segsize_max;
4653 	if (scctx->isc_rss_table_size == 0)
4654 		scctx->isc_rss_table_size = 64;
4655 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4656 
4657 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4658 	/* XXX format name */
4659 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx, -1, "admin");
4660 
4661 	/* XXX --- can support > 1 -- but keep it simple for now */
4662 	scctx->isc_intr = IFLIB_INTR_LEGACY;
4663 
4664 	/* Get memory for the station queues */
4665 	if ((err = iflib_queues_alloc(ctx))) {
4666 		device_printf(dev, "Unable to allocate queue memory\n");
4667 		goto fail;
4668 	}
4669 
4670 	if ((err = iflib_qset_structures_setup(ctx))) {
4671 		device_printf(dev, "qset structure setup failed %d\n", err);
4672 		goto fail_queues;
4673 	}
4674 	/*
4675 	 * XXX What if anything do we want to do about interrupts?
4676 	 */
4677 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac);
4678 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4679 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4680 		goto fail_detach;
4681 	}
4682 	/* XXX handle more than one queue */
4683 	for (i = 0; i < scctx->isc_nrxqsets; i++)
4684 		IFDI_RX_CLSET(ctx, 0, i, ctx->ifc_rxqs[i].ifr_fl[0].ifl_sds.ifsd_cl);
4685 
4686 	*ctxp = ctx;
4687 
4688 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4689 	iflib_add_device_sysctl_post(ctx);
4690 	ctx->ifc_flags |= IFC_INIT_DONE;
4691 	return (0);
4692 fail_detach:
4693 	ether_ifdetach(ctx->ifc_ifp);
4694 fail_queues:
4695 	iflib_tx_structures_free(ctx);
4696 	iflib_rx_structures_free(ctx);
4697 fail:
4698 	IFDI_DETACH(ctx);
4699 	return (err);
4700 }
4701 
4702 int
4703 iflib_pseudo_deregister(if_ctx_t ctx)
4704 {
4705 	if_t ifp = ctx->ifc_ifp;
4706 	iflib_txq_t txq;
4707 	iflib_rxq_t rxq;
4708 	int i, j;
4709 	struct taskqgroup *tqg;
4710 	iflib_fl_t fl;
4711 
4712 	/* Unregister VLAN events */
4713 	if (ctx->ifc_vlan_attach_event != NULL)
4714 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
4715 	if (ctx->ifc_vlan_detach_event != NULL)
4716 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
4717 
4718 	ether_ifdetach(ifp);
4719 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
4720 	CTX_LOCK_DESTROY(ctx);
4721 	/* XXX drain any dependent tasks */
4722 	tqg = qgroup_if_io_tqg;
4723 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
4724 		callout_drain(&txq->ift_timer);
4725 		if (txq->ift_task.gt_uniq != NULL)
4726 			taskqgroup_detach(tqg, &txq->ift_task);
4727 	}
4728 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4729 		if (rxq->ifr_task.gt_uniq != NULL)
4730 			taskqgroup_detach(tqg, &rxq->ifr_task);
4731 
4732 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
4733 			free(fl->ifl_rx_bitmap, M_IFLIB);
4734 	}
4735 	tqg = qgroup_if_config_tqg;
4736 	if (ctx->ifc_admin_task.gt_uniq != NULL)
4737 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
4738 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
4739 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
4740 
4741 	if_free(ifp);
4742 
4743 	iflib_tx_structures_free(ctx);
4744 	iflib_rx_structures_free(ctx);
4745 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4746 		free(ctx->ifc_softc, M_IFLIB);
4747 	free(ctx, M_IFLIB);
4748 	return (0);
4749 }
4750 
4751 int
4752 iflib_device_attach(device_t dev)
4753 {
4754 	if_ctx_t ctx;
4755 	if_shared_ctx_t sctx;
4756 
4757 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4758 		return (ENOTSUP);
4759 
4760 	pci_enable_busmaster(dev);
4761 
4762 	return (iflib_device_register(dev, NULL, sctx, &ctx));
4763 }
4764 
4765 int
4766 iflib_device_deregister(if_ctx_t ctx)
4767 {
4768 	if_t ifp = ctx->ifc_ifp;
4769 	iflib_txq_t txq;
4770 	iflib_rxq_t rxq;
4771 	device_t dev = ctx->ifc_dev;
4772 	int i, j;
4773 	struct taskqgroup *tqg;
4774 	iflib_fl_t fl;
4775 
4776 	/* Make sure VLANS are not using driver */
4777 	if (if_vlantrunkinuse(ifp)) {
4778 		device_printf(dev,"Vlan in use, detach first\n");
4779 		return (EBUSY);
4780 	}
4781 
4782 	CTX_LOCK(ctx);
4783 	ctx->ifc_in_detach = 1;
4784 	iflib_stop(ctx);
4785 	CTX_UNLOCK(ctx);
4786 
4787 	/* Unregister VLAN events */
4788 	if (ctx->ifc_vlan_attach_event != NULL)
4789 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
4790 	if (ctx->ifc_vlan_detach_event != NULL)
4791 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
4792 
4793 	iflib_netmap_detach(ifp);
4794 	ether_ifdetach(ifp);
4795 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
4796 	CTX_LOCK_DESTROY(ctx);
4797 	if (ctx->ifc_led_dev != NULL)
4798 		led_destroy(ctx->ifc_led_dev);
4799 	/* XXX drain any dependent tasks */
4800 	tqg = qgroup_if_io_tqg;
4801 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
4802 		callout_drain(&txq->ift_timer);
4803 		if (txq->ift_task.gt_uniq != NULL)
4804 			taskqgroup_detach(tqg, &txq->ift_task);
4805 	}
4806 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4807 		if (rxq->ifr_task.gt_uniq != NULL)
4808 			taskqgroup_detach(tqg, &rxq->ifr_task);
4809 
4810 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
4811 			free(fl->ifl_rx_bitmap, M_IFLIB);
4812 
4813 	}
4814 	tqg = qgroup_if_config_tqg;
4815 	if (ctx->ifc_admin_task.gt_uniq != NULL)
4816 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
4817 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
4818 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
4819 
4820 	IFDI_DETACH(ctx);
4821 	device_set_softc(ctx->ifc_dev, NULL);
4822 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) {
4823 		pci_release_msi(dev);
4824 	}
4825 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) {
4826 		iflib_irq_free(ctx, &ctx->ifc_legacy_irq);
4827 	}
4828 	if (ctx->ifc_msix_mem != NULL) {
4829 		bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY,
4830 			ctx->ifc_softc_ctx.isc_msix_bar, ctx->ifc_msix_mem);
4831 		ctx->ifc_msix_mem = NULL;
4832 	}
4833 
4834 	bus_generic_detach(dev);
4835 	if_free(ifp);
4836 
4837 	iflib_tx_structures_free(ctx);
4838 	iflib_rx_structures_free(ctx);
4839 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4840 		free(ctx->ifc_softc, M_IFLIB);
4841 	free(ctx, M_IFLIB);
4842 	return (0);
4843 }
4844 
4845 
4846 int
4847 iflib_device_detach(device_t dev)
4848 {
4849 	if_ctx_t ctx = device_get_softc(dev);
4850 
4851 	return (iflib_device_deregister(ctx));
4852 }
4853 
4854 int
4855 iflib_device_suspend(device_t dev)
4856 {
4857 	if_ctx_t ctx = device_get_softc(dev);
4858 
4859 	CTX_LOCK(ctx);
4860 	IFDI_SUSPEND(ctx);
4861 	CTX_UNLOCK(ctx);
4862 
4863 	return bus_generic_suspend(dev);
4864 }
4865 int
4866 iflib_device_shutdown(device_t dev)
4867 {
4868 	if_ctx_t ctx = device_get_softc(dev);
4869 
4870 	CTX_LOCK(ctx);
4871 	IFDI_SHUTDOWN(ctx);
4872 	CTX_UNLOCK(ctx);
4873 
4874 	return bus_generic_suspend(dev);
4875 }
4876 
4877 
4878 int
4879 iflib_device_resume(device_t dev)
4880 {
4881 	if_ctx_t ctx = device_get_softc(dev);
4882 	iflib_txq_t txq = ctx->ifc_txqs;
4883 
4884 	CTX_LOCK(ctx);
4885 	IFDI_RESUME(ctx);
4886 	iflib_init_locked(ctx);
4887 	CTX_UNLOCK(ctx);
4888 	for (int i = 0; i < NTXQSETS(ctx); i++, txq++)
4889 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
4890 
4891 	return (bus_generic_resume(dev));
4892 }
4893 
4894 int
4895 iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params)
4896 {
4897 	int error;
4898 	if_ctx_t ctx = device_get_softc(dev);
4899 
4900 	CTX_LOCK(ctx);
4901 	error = IFDI_IOV_INIT(ctx, num_vfs, params);
4902 	CTX_UNLOCK(ctx);
4903 
4904 	return (error);
4905 }
4906 
4907 void
4908 iflib_device_iov_uninit(device_t dev)
4909 {
4910 	if_ctx_t ctx = device_get_softc(dev);
4911 
4912 	CTX_LOCK(ctx);
4913 	IFDI_IOV_UNINIT(ctx);
4914 	CTX_UNLOCK(ctx);
4915 }
4916 
4917 int
4918 iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params)
4919 {
4920 	int error;
4921 	if_ctx_t ctx = device_get_softc(dev);
4922 
4923 	CTX_LOCK(ctx);
4924 	error = IFDI_IOV_VF_ADD(ctx, vfnum, params);
4925 	CTX_UNLOCK(ctx);
4926 
4927 	return (error);
4928 }
4929 
4930 /*********************************************************************
4931  *
4932  *  MODULE FUNCTION DEFINITIONS
4933  *
4934  **********************************************************************/
4935 
4936 /*
4937  * - Start a fast taskqueue thread for each core
4938  * - Start a taskqueue for control operations
4939  */
4940 static int
4941 iflib_module_init(void)
4942 {
4943 	return (0);
4944 }
4945 
4946 static int
4947 iflib_module_event_handler(module_t mod, int what, void *arg)
4948 {
4949 	int err;
4950 
4951 	switch (what) {
4952 	case MOD_LOAD:
4953 		if ((err = iflib_module_init()) != 0)
4954 			return (err);
4955 		break;
4956 	case MOD_UNLOAD:
4957 		return (EBUSY);
4958 	default:
4959 		return (EOPNOTSUPP);
4960 	}
4961 
4962 	return (0);
4963 }
4964 
4965 /*********************************************************************
4966  *
4967  *  PUBLIC FUNCTION DEFINITIONS
4968  *     ordered as in iflib.h
4969  *
4970  **********************************************************************/
4971 
4972 
4973 static void
4974 _iflib_assert(if_shared_ctx_t sctx)
4975 {
4976 	MPASS(sctx->isc_tx_maxsize);
4977 	MPASS(sctx->isc_tx_maxsegsize);
4978 
4979 	MPASS(sctx->isc_rx_maxsize);
4980 	MPASS(sctx->isc_rx_nsegments);
4981 	MPASS(sctx->isc_rx_maxsegsize);
4982 
4983 	MPASS(sctx->isc_nrxd_min[0]);
4984 	MPASS(sctx->isc_nrxd_max[0]);
4985 	MPASS(sctx->isc_nrxd_default[0]);
4986 	MPASS(sctx->isc_ntxd_min[0]);
4987 	MPASS(sctx->isc_ntxd_max[0]);
4988 	MPASS(sctx->isc_ntxd_default[0]);
4989 }
4990 
4991 static void
4992 _iflib_pre_assert(if_softc_ctx_t scctx)
4993 {
4994 
4995 	MPASS(scctx->isc_txrx->ift_txd_encap);
4996 	MPASS(scctx->isc_txrx->ift_txd_flush);
4997 	MPASS(scctx->isc_txrx->ift_txd_credits_update);
4998 	MPASS(scctx->isc_txrx->ift_rxd_available);
4999 	MPASS(scctx->isc_txrx->ift_rxd_pkt_get);
5000 	MPASS(scctx->isc_txrx->ift_rxd_refill);
5001 	MPASS(scctx->isc_txrx->ift_rxd_flush);
5002 }
5003 
5004 static int
5005 iflib_register(if_ctx_t ctx)
5006 {
5007 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5008 	driver_t *driver = sctx->isc_driver;
5009 	device_t dev = ctx->ifc_dev;
5010 	if_t ifp;
5011 
5012 	_iflib_assert(sctx);
5013 
5014 	CTX_LOCK_INIT(ctx);
5015 	STATE_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev));
5016 	ifp = ctx->ifc_ifp = if_gethandle(IFT_ETHER);
5017 	if (ifp == NULL) {
5018 		device_printf(dev, "can not allocate ifnet structure\n");
5019 		return (ENOMEM);
5020 	}
5021 
5022 	/*
5023 	 * Initialize our context's device specific methods
5024 	 */
5025 	kobj_init((kobj_t) ctx, (kobj_class_t) driver);
5026 	kobj_class_compile((kobj_class_t) driver);
5027 	driver->refs++;
5028 
5029 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
5030 	if_setsoftc(ifp, ctx);
5031 	if_setdev(ifp, dev);
5032 	if_setinitfn(ifp, iflib_if_init);
5033 	if_setioctlfn(ifp, iflib_if_ioctl);
5034 	if_settransmitfn(ifp, iflib_if_transmit);
5035 	if_setqflushfn(ifp, iflib_if_qflush);
5036 	if_setflags(ifp, IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST);
5037 
5038 	ctx->ifc_vlan_attach_event =
5039 		EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx,
5040 							  EVENTHANDLER_PRI_FIRST);
5041 	ctx->ifc_vlan_detach_event =
5042 		EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx,
5043 							  EVENTHANDLER_PRI_FIRST);
5044 
5045 	ifmedia_init(&ctx->ifc_media, IFM_IMASK,
5046 					 iflib_media_change, iflib_media_status);
5047 
5048 	return (0);
5049 }
5050 
5051 
5052 static int
5053 iflib_queues_alloc(if_ctx_t ctx)
5054 {
5055 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5056 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5057 	device_t dev = ctx->ifc_dev;
5058 	int nrxqsets = scctx->isc_nrxqsets;
5059 	int ntxqsets = scctx->isc_ntxqsets;
5060 	iflib_txq_t txq;
5061 	iflib_rxq_t rxq;
5062 	iflib_fl_t fl = NULL;
5063 	int i, j, cpu, err, txconf, rxconf;
5064 	iflib_dma_info_t ifdip;
5065 	uint32_t *rxqsizes = scctx->isc_rxqsizes;
5066 	uint32_t *txqsizes = scctx->isc_txqsizes;
5067 	uint8_t nrxqs = sctx->isc_nrxqs;
5068 	uint8_t ntxqs = sctx->isc_ntxqs;
5069 	int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1;
5070 	caddr_t *vaddrs;
5071 	uint64_t *paddrs;
5072 
5073 	KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1"));
5074 	KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1"));
5075 
5076 /* Allocate the TX ring struct memory */
5077 	if (!(ctx->ifc_txqs =
5078 	    (iflib_txq_t) malloc(sizeof(struct iflib_txq) *
5079 	    ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5080 		device_printf(dev, "Unable to allocate TX ring memory\n");
5081 		err = ENOMEM;
5082 		goto fail;
5083 	}
5084 
5085 	/* Now allocate the RX */
5086 	if (!(ctx->ifc_rxqs =
5087 	    (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) *
5088 	    nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5089 		device_printf(dev, "Unable to allocate RX ring memory\n");
5090 		err = ENOMEM;
5091 		goto rx_fail;
5092 	}
5093 
5094 	txq = ctx->ifc_txqs;
5095 	rxq = ctx->ifc_rxqs;
5096 
5097 	/*
5098 	 * XXX handle allocation failure
5099 	 */
5100 	for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) {
5101 		/* Set up some basics */
5102 
5103 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs, M_IFLIB, M_WAITOK|M_ZERO)) == NULL) {
5104 			device_printf(dev, "failed to allocate iflib_dma_info\n");
5105 			err = ENOMEM;
5106 			goto err_tx_desc;
5107 		}
5108 		txq->ift_ifdi = ifdip;
5109 		for (j = 0; j < ntxqs; j++, ifdip++) {
5110 			if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, BUS_DMA_NOWAIT)) {
5111 				device_printf(dev, "Unable to allocate Descriptor memory\n");
5112 				err = ENOMEM;
5113 				goto err_tx_desc;
5114 			}
5115 			txq->ift_txd_size[j] = scctx->isc_txd_size[j];
5116 			bzero((void *)ifdip->idi_vaddr, txqsizes[j]);
5117 		}
5118 		txq->ift_ctx = ctx;
5119 		txq->ift_id = i;
5120 		if (sctx->isc_flags & IFLIB_HAS_TXCQ) {
5121 			txq->ift_br_offset = 1;
5122 		} else {
5123 			txq->ift_br_offset = 0;
5124 		}
5125 		/* XXX fix this */
5126 		txq->ift_timer.c_cpu = cpu;
5127 
5128 		if (iflib_txsd_alloc(txq)) {
5129 			device_printf(dev, "Critical Failure setting up TX buffers\n");
5130 			err = ENOMEM;
5131 			goto err_tx_desc;
5132 		}
5133 
5134 		/* Initialize the TX lock */
5135 		snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:tx(%d):callout",
5136 		    device_get_nameunit(dev), txq->ift_id);
5137 		mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF);
5138 		callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0);
5139 
5140 		snprintf(txq->ift_db_mtx_name, MTX_NAME_LEN, "%s:tx(%d):db",
5141 			 device_get_nameunit(dev), txq->ift_id);
5142 
5143 		err = ifmp_ring_alloc(&txq->ift_br, 2048, txq, iflib_txq_drain,
5144 				      iflib_txq_can_drain, M_IFLIB, M_WAITOK);
5145 		if (err) {
5146 			/* XXX free any allocated rings */
5147 			device_printf(dev, "Unable to allocate buf_ring\n");
5148 			goto err_tx_desc;
5149 		}
5150 	}
5151 
5152 	for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) {
5153 		/* Set up some basics */
5154 
5155 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs, M_IFLIB, M_WAITOK|M_ZERO)) == NULL) {
5156 			device_printf(dev, "failed to allocate iflib_dma_info\n");
5157 			err = ENOMEM;
5158 			goto err_tx_desc;
5159 		}
5160 
5161 		rxq->ifr_ifdi = ifdip;
5162 		/* XXX this needs to be changed if #rx queues != #tx queues */
5163 		rxq->ifr_ntxqirq = 1;
5164 		rxq->ifr_txqid[0] = i;
5165 		for (j = 0; j < nrxqs; j++, ifdip++) {
5166 			if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, BUS_DMA_NOWAIT)) {
5167 				device_printf(dev, "Unable to allocate Descriptor memory\n");
5168 				err = ENOMEM;
5169 				goto err_tx_desc;
5170 			}
5171 			bzero((void *)ifdip->idi_vaddr, rxqsizes[j]);
5172 		}
5173 		rxq->ifr_ctx = ctx;
5174 		rxq->ifr_id = i;
5175 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
5176 			rxq->ifr_fl_offset = 1;
5177 		} else {
5178 			rxq->ifr_fl_offset = 0;
5179 		}
5180 		rxq->ifr_nfl = nfree_lists;
5181 		if (!(fl =
5182 			  (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) {
5183 			device_printf(dev, "Unable to allocate free list memory\n");
5184 			err = ENOMEM;
5185 			goto err_tx_desc;
5186 		}
5187 		rxq->ifr_fl = fl;
5188 		for (j = 0; j < nfree_lists; j++) {
5189 			fl[j].ifl_rxq = rxq;
5190 			fl[j].ifl_id = j;
5191 			fl[j].ifl_ifdi = &rxq->ifr_ifdi[j + rxq->ifr_fl_offset];
5192 			fl[j].ifl_rxd_size = scctx->isc_rxd_size[j];
5193 		}
5194         /* Allocate receive buffers for the ring*/
5195 		if (iflib_rxsd_alloc(rxq)) {
5196 			device_printf(dev,
5197 			    "Critical Failure setting up receive buffers\n");
5198 			err = ENOMEM;
5199 			goto err_rx_desc;
5200 		}
5201 
5202 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5203 			fl->ifl_rx_bitmap = bit_alloc(fl->ifl_size, M_IFLIB, M_WAITOK|M_ZERO);
5204 	}
5205 
5206 	/* TXQs */
5207 	vaddrs = malloc(sizeof(caddr_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5208 	paddrs = malloc(sizeof(uint64_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5209 	for (i = 0; i < ntxqsets; i++) {
5210 		iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi;
5211 
5212 		for (j = 0; j < ntxqs; j++, di++) {
5213 			vaddrs[i*ntxqs + j] = di->idi_vaddr;
5214 			paddrs[i*ntxqs + j] = di->idi_paddr;
5215 		}
5216 	}
5217 	if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) {
5218 		device_printf(ctx->ifc_dev, "device queue allocation failed\n");
5219 		iflib_tx_structures_free(ctx);
5220 		free(vaddrs, M_IFLIB);
5221 		free(paddrs, M_IFLIB);
5222 		goto err_rx_desc;
5223 	}
5224 	free(vaddrs, M_IFLIB);
5225 	free(paddrs, M_IFLIB);
5226 
5227 	/* RXQs */
5228 	vaddrs = malloc(sizeof(caddr_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5229 	paddrs = malloc(sizeof(uint64_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5230 	for (i = 0; i < nrxqsets; i++) {
5231 		iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi;
5232 
5233 		for (j = 0; j < nrxqs; j++, di++) {
5234 			vaddrs[i*nrxqs + j] = di->idi_vaddr;
5235 			paddrs[i*nrxqs + j] = di->idi_paddr;
5236 		}
5237 	}
5238 	if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) {
5239 		device_printf(ctx->ifc_dev, "device queue allocation failed\n");
5240 		iflib_tx_structures_free(ctx);
5241 		free(vaddrs, M_IFLIB);
5242 		free(paddrs, M_IFLIB);
5243 		goto err_rx_desc;
5244 	}
5245 	free(vaddrs, M_IFLIB);
5246 	free(paddrs, M_IFLIB);
5247 
5248 	return (0);
5249 
5250 /* XXX handle allocation failure changes */
5251 err_rx_desc:
5252 err_tx_desc:
5253 rx_fail:
5254 	if (ctx->ifc_rxqs != NULL)
5255 		free(ctx->ifc_rxqs, M_IFLIB);
5256 	ctx->ifc_rxqs = NULL;
5257 	if (ctx->ifc_txqs != NULL)
5258 		free(ctx->ifc_txqs, M_IFLIB);
5259 	ctx->ifc_txqs = NULL;
5260 fail:
5261 	return (err);
5262 }
5263 
5264 static int
5265 iflib_tx_structures_setup(if_ctx_t ctx)
5266 {
5267 	iflib_txq_t txq = ctx->ifc_txqs;
5268 	int i;
5269 
5270 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
5271 		iflib_txq_setup(txq);
5272 
5273 	return (0);
5274 }
5275 
5276 static void
5277 iflib_tx_structures_free(if_ctx_t ctx)
5278 {
5279 	iflib_txq_t txq = ctx->ifc_txqs;
5280 	int i, j;
5281 
5282 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
5283 		iflib_txq_destroy(txq);
5284 		for (j = 0; j < ctx->ifc_nhwtxqs; j++)
5285 			iflib_dma_free(&txq->ift_ifdi[j]);
5286 	}
5287 	free(ctx->ifc_txqs, M_IFLIB);
5288 	ctx->ifc_txqs = NULL;
5289 	IFDI_QUEUES_FREE(ctx);
5290 }
5291 
5292 /*********************************************************************
5293  *
5294  *  Initialize all receive rings.
5295  *
5296  **********************************************************************/
5297 static int
5298 iflib_rx_structures_setup(if_ctx_t ctx)
5299 {
5300 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5301 	int q;
5302 #if defined(INET6) || defined(INET)
5303 	int i, err;
5304 #endif
5305 
5306 	for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) {
5307 #if defined(INET6) || defined(INET)
5308 		tcp_lro_free(&rxq->ifr_lc);
5309 		if ((err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp,
5310 		    TCP_LRO_ENTRIES, min(1024,
5311 		    ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset]))) != 0) {
5312 			device_printf(ctx->ifc_dev, "LRO Initialization failed!\n");
5313 			goto fail;
5314 		}
5315 		rxq->ifr_lro_enabled = TRUE;
5316 #endif
5317 		IFDI_RXQ_SETUP(ctx, rxq->ifr_id);
5318 	}
5319 	return (0);
5320 #if defined(INET6) || defined(INET)
5321 fail:
5322 	/*
5323 	 * Free RX software descriptors allocated so far, we will only handle
5324 	 * the rings that completed, the failing case will have
5325 	 * cleaned up for itself. 'q' failed, so its the terminus.
5326 	 */
5327 	rxq = ctx->ifc_rxqs;
5328 	for (i = 0; i < q; ++i, rxq++) {
5329 		iflib_rx_sds_free(rxq);
5330 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
5331 	}
5332 	return (err);
5333 #endif
5334 }
5335 
5336 /*********************************************************************
5337  *
5338  *  Free all receive rings.
5339  *
5340  **********************************************************************/
5341 static void
5342 iflib_rx_structures_free(if_ctx_t ctx)
5343 {
5344 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5345 
5346 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) {
5347 		iflib_rx_sds_free(rxq);
5348 	}
5349 }
5350 
5351 static int
5352 iflib_qset_structures_setup(if_ctx_t ctx)
5353 {
5354 	int err;
5355 
5356 	/*
5357 	 * It is expected that the caller takes care of freeing queues if this
5358 	 * fails.
5359 	 */
5360 	if ((err = iflib_tx_structures_setup(ctx)) != 0) {
5361 		device_printf(ctx->ifc_dev, "iflib_tx_structures_setup failed: %d\n", err);
5362 		return (err);
5363 	}
5364 
5365 	if ((err = iflib_rx_structures_setup(ctx)) != 0)
5366 		device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err);
5367 
5368 	return (err);
5369 }
5370 
5371 int
5372 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
5373 				driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, char *name)
5374 {
5375 
5376 	return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name));
5377 }
5378 
5379 #ifdef SMP
5380 static int
5381 find_nth(if_ctx_t ctx, int qid)
5382 {
5383 	cpuset_t cpus;
5384 	int i, cpuid, eqid, count;
5385 
5386 	CPU_COPY(&ctx->ifc_cpus, &cpus);
5387 	count = CPU_COUNT(&cpus);
5388 	eqid = qid % count;
5389 	/* clear up to the qid'th bit */
5390 	for (i = 0; i < eqid; i++) {
5391 		cpuid = CPU_FFS(&cpus);
5392 		MPASS(cpuid != 0);
5393 		CPU_CLR(cpuid-1, &cpus);
5394 	}
5395 	cpuid = CPU_FFS(&cpus);
5396 	MPASS(cpuid != 0);
5397 	return (cpuid-1);
5398 }
5399 
5400 #ifdef SCHED_ULE
5401 extern struct cpu_group *cpu_top;              /* CPU topology */
5402 
5403 static int
5404 find_child_with_core(int cpu, struct cpu_group *grp)
5405 {
5406 	int i;
5407 
5408 	if (grp->cg_children == 0)
5409 		return -1;
5410 
5411 	MPASS(grp->cg_child);
5412 	for (i = 0; i < grp->cg_children; i++) {
5413 		if (CPU_ISSET(cpu, &grp->cg_child[i].cg_mask))
5414 			return i;
5415 	}
5416 
5417 	return -1;
5418 }
5419 
5420 /*
5421  * Find the nth "close" core to the specified core
5422  * "close" is defined as the deepest level that shares
5423  * at least an L2 cache.  With threads, this will be
5424  * threads on the same core.  If the sahred cache is L3
5425  * or higher, simply returns the same core.
5426  */
5427 static int
5428 find_close_core(int cpu, int core_offset)
5429 {
5430 	struct cpu_group *grp;
5431 	int i;
5432 	int fcpu;
5433 	cpuset_t cs;
5434 
5435 	grp = cpu_top;
5436 	if (grp == NULL)
5437 		return cpu;
5438 	i = 0;
5439 	while ((i = find_child_with_core(cpu, grp)) != -1) {
5440 		/* If the child only has one cpu, don't descend */
5441 		if (grp->cg_child[i].cg_count <= 1)
5442 			break;
5443 		grp = &grp->cg_child[i];
5444 	}
5445 
5446 	/* If they don't share at least an L2 cache, use the same CPU */
5447 	if (grp->cg_level > CG_SHARE_L2 || grp->cg_level == CG_SHARE_NONE)
5448 		return cpu;
5449 
5450 	/* Now pick one */
5451 	CPU_COPY(&grp->cg_mask, &cs);
5452 
5453 	/* Add the selected CPU offset to core offset. */
5454 	for (i = 0; (fcpu = CPU_FFS(&cs)) != 0; i++) {
5455 		if (fcpu - 1 == cpu)
5456 			break;
5457 		CPU_CLR(fcpu - 1, &cs);
5458 	}
5459 	MPASS(fcpu);
5460 
5461 	core_offset += i;
5462 
5463 	CPU_COPY(&grp->cg_mask, &cs);
5464 	for (i = core_offset % grp->cg_count; i > 0; i--) {
5465 		MPASS(CPU_FFS(&cs));
5466 		CPU_CLR(CPU_FFS(&cs) - 1, &cs);
5467 	}
5468 	MPASS(CPU_FFS(&cs));
5469 	return CPU_FFS(&cs) - 1;
5470 }
5471 #else
5472 static int
5473 find_close_core(int cpu, int core_offset __unused)
5474 {
5475 	return cpu;
5476 }
5477 #endif
5478 
5479 static int
5480 get_core_offset(if_ctx_t ctx, iflib_intr_type_t type, int qid)
5481 {
5482 	switch (type) {
5483 	case IFLIB_INTR_TX:
5484 		/* TX queues get cores which share at least an L2 cache with the corresponding RX queue */
5485 		/* XXX handle multiple RX threads per core and more than two core per L2 group */
5486 		return qid / CPU_COUNT(&ctx->ifc_cpus) + 1;
5487 	case IFLIB_INTR_RX:
5488 	case IFLIB_INTR_RXTX:
5489 		/* RX queues get the specified core */
5490 		return qid / CPU_COUNT(&ctx->ifc_cpus);
5491 	default:
5492 		return -1;
5493 	}
5494 }
5495 #else
5496 #define get_core_offset(ctx, type, qid)	CPU_FIRST()
5497 #define find_close_core(cpuid, tid)	CPU_FIRST()
5498 #define find_nth(ctx, gid)		CPU_FIRST()
5499 #endif
5500 
5501 /* Just to avoid copy/paste */
5502 static inline int
5503 iflib_irq_set_affinity(if_ctx_t ctx, int irq, iflib_intr_type_t type, int qid,
5504     struct grouptask *gtask, struct taskqgroup *tqg, void *uniq, char *name)
5505 {
5506 	int cpuid;
5507 	int err, tid;
5508 
5509 	cpuid = find_nth(ctx, qid);
5510 	tid = get_core_offset(ctx, type, qid);
5511 	MPASS(tid >= 0);
5512 	cpuid = find_close_core(cpuid, tid);
5513 	err = taskqgroup_attach_cpu(tqg, gtask, uniq, cpuid, irq, name);
5514 	if (err) {
5515 		device_printf(ctx->ifc_dev, "taskqgroup_attach_cpu failed %d\n", err);
5516 		return (err);
5517 	}
5518 #ifdef notyet
5519 	if (cpuid > ctx->ifc_cpuid_highest)
5520 		ctx->ifc_cpuid_highest = cpuid;
5521 #endif
5522 	return 0;
5523 }
5524 
5525 int
5526 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid,
5527 						iflib_intr_type_t type, driver_filter_t *filter,
5528 						void *filter_arg, int qid, char *name)
5529 {
5530 	struct grouptask *gtask;
5531 	struct taskqgroup *tqg;
5532 	iflib_filter_info_t info;
5533 	gtask_fn_t *fn;
5534 	int tqrid, err;
5535 	driver_filter_t *intr_fast;
5536 	void *q;
5537 
5538 	info = &ctx->ifc_filter_info;
5539 	tqrid = rid;
5540 
5541 	switch (type) {
5542 	/* XXX merge tx/rx for netmap? */
5543 	case IFLIB_INTR_TX:
5544 		q = &ctx->ifc_txqs[qid];
5545 		info = &ctx->ifc_txqs[qid].ift_filter_info;
5546 		gtask = &ctx->ifc_txqs[qid].ift_task;
5547 		tqg = qgroup_if_io_tqg;
5548 		fn = _task_fn_tx;
5549 		intr_fast = iflib_fast_intr;
5550 		GROUPTASK_INIT(gtask, 0, fn, q);
5551 		break;
5552 	case IFLIB_INTR_RX:
5553 		q = &ctx->ifc_rxqs[qid];
5554 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
5555 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5556 		tqg = qgroup_if_io_tqg;
5557 		fn = _task_fn_rx;
5558 		intr_fast = iflib_fast_intr;
5559 		GROUPTASK_INIT(gtask, 0, fn, q);
5560 		break;
5561 	case IFLIB_INTR_RXTX:
5562 		q = &ctx->ifc_rxqs[qid];
5563 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
5564 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5565 		tqg = qgroup_if_io_tqg;
5566 		fn = _task_fn_rx;
5567 		intr_fast = iflib_fast_intr_rxtx;
5568 		GROUPTASK_INIT(gtask, 0, fn, q);
5569 		break;
5570 	case IFLIB_INTR_ADMIN:
5571 		q = ctx;
5572 		tqrid = -1;
5573 		info = &ctx->ifc_filter_info;
5574 		gtask = &ctx->ifc_admin_task;
5575 		tqg = qgroup_if_config_tqg;
5576 		fn = _task_fn_admin;
5577 		intr_fast = iflib_fast_intr_ctx;
5578 		break;
5579 	default:
5580 		panic("unknown net intr type");
5581 	}
5582 
5583 	info->ifi_filter = filter;
5584 	info->ifi_filter_arg = filter_arg;
5585 	info->ifi_task = gtask;
5586 	info->ifi_ctx = q;
5587 
5588 	err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info,  name);
5589 	if (err != 0) {
5590 		device_printf(ctx->ifc_dev, "_iflib_irq_alloc failed %d\n", err);
5591 		return (err);
5592 	}
5593 	if (type == IFLIB_INTR_ADMIN)
5594 		return (0);
5595 
5596 	if (tqrid != -1) {
5597 		err = iflib_irq_set_affinity(ctx, rman_get_start(irq->ii_res), type, qid, gtask, tqg, q, name);
5598 		if (err)
5599 			return (err);
5600 	} else {
5601 		taskqgroup_attach(tqg, gtask, q, rman_get_start(irq->ii_res), name);
5602 	}
5603 
5604 	return (0);
5605 }
5606 
5607 void
5608 iflib_softirq_alloc_generic(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type,  void *arg, int qid, char *name)
5609 {
5610 	struct grouptask *gtask;
5611 	struct taskqgroup *tqg;
5612 	gtask_fn_t *fn;
5613 	void *q;
5614 	int irq_num = -1;
5615 	int err;
5616 
5617 	switch (type) {
5618 	case IFLIB_INTR_TX:
5619 		q = &ctx->ifc_txqs[qid];
5620 		gtask = &ctx->ifc_txqs[qid].ift_task;
5621 		tqg = qgroup_if_io_tqg;
5622 		fn = _task_fn_tx;
5623 		if (irq != NULL)
5624 			irq_num = rman_get_start(irq->ii_res);
5625 		break;
5626 	case IFLIB_INTR_RX:
5627 		q = &ctx->ifc_rxqs[qid];
5628 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5629 		tqg = qgroup_if_io_tqg;
5630 		fn = _task_fn_rx;
5631 		if (irq != NULL)
5632 			irq_num = rman_get_start(irq->ii_res);
5633 		break;
5634 	case IFLIB_INTR_IOV:
5635 		q = ctx;
5636 		gtask = &ctx->ifc_vflr_task;
5637 		tqg = qgroup_if_config_tqg;
5638 		fn = _task_fn_iov;
5639 		break;
5640 	default:
5641 		panic("unknown net intr type");
5642 	}
5643 	GROUPTASK_INIT(gtask, 0, fn, q);
5644 	if (irq_num != -1) {
5645 		err = iflib_irq_set_affinity(ctx, irq_num, type, qid, gtask, tqg, q, name);
5646 		if (err)
5647 			taskqgroup_attach(tqg, gtask, q, irq_num, name);
5648 	}
5649 	else {
5650 		taskqgroup_attach(tqg, gtask, q, irq_num, name);
5651 	}
5652 }
5653 
5654 void
5655 iflib_irq_free(if_ctx_t ctx, if_irq_t irq)
5656 {
5657 	if (irq->ii_tag)
5658 		bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag);
5659 
5660 	if (irq->ii_res)
5661 		bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ, irq->ii_rid, irq->ii_res);
5662 }
5663 
5664 static int
5665 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, char *name)
5666 {
5667 	iflib_txq_t txq = ctx->ifc_txqs;
5668 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5669 	if_irq_t irq = &ctx->ifc_legacy_irq;
5670 	iflib_filter_info_t info;
5671 	struct grouptask *gtask;
5672 	struct taskqgroup *tqg;
5673 	gtask_fn_t *fn;
5674 	int tqrid;
5675 	void *q;
5676 	int err;
5677 
5678 	q = &ctx->ifc_rxqs[0];
5679 	info = &rxq[0].ifr_filter_info;
5680 	gtask = &rxq[0].ifr_task;
5681 	tqg = qgroup_if_io_tqg;
5682 	tqrid = irq->ii_rid = *rid;
5683 	fn = _task_fn_rx;
5684 
5685 	ctx->ifc_flags |= IFC_LEGACY;
5686 	info->ifi_filter = filter;
5687 	info->ifi_filter_arg = filter_arg;
5688 	info->ifi_task = gtask;
5689 	info->ifi_ctx = ctx;
5690 
5691 	/* We allocate a single interrupt resource */
5692 	if ((err = _iflib_irq_alloc(ctx, irq, tqrid, iflib_fast_intr_ctx, NULL, info, name)) != 0)
5693 		return (err);
5694 	GROUPTASK_INIT(gtask, 0, fn, q);
5695 	taskqgroup_attach(tqg, gtask, q, rman_get_start(irq->ii_res), name);
5696 
5697 	GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq);
5698 	taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, rman_get_start(irq->ii_res), "tx");
5699 	return (0);
5700 }
5701 
5702 void
5703 iflib_led_create(if_ctx_t ctx)
5704 {
5705 
5706 	ctx->ifc_led_dev = led_create(iflib_led_func, ctx,
5707 	    device_get_nameunit(ctx->ifc_dev));
5708 }
5709 
5710 void
5711 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid)
5712 {
5713 
5714 	GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
5715 }
5716 
5717 void
5718 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid)
5719 {
5720 
5721 	GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task);
5722 }
5723 
5724 void
5725 iflib_admin_intr_deferred(if_ctx_t ctx)
5726 {
5727 #ifdef INVARIANTS
5728 	struct grouptask *gtask;
5729 
5730 	gtask = &ctx->ifc_admin_task;
5731 	MPASS(gtask != NULL && gtask->gt_taskqueue != NULL);
5732 #endif
5733 
5734 	GROUPTASK_ENQUEUE(&ctx->ifc_admin_task);
5735 }
5736 
5737 void
5738 iflib_iov_intr_deferred(if_ctx_t ctx)
5739 {
5740 
5741 	GROUPTASK_ENQUEUE(&ctx->ifc_vflr_task);
5742 }
5743 
5744 void
5745 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, char *name)
5746 {
5747 
5748 	taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, -1, name);
5749 }
5750 
5751 void
5752 iflib_config_gtask_init(void *ctx, struct grouptask *gtask, gtask_fn_t *fn,
5753 	const char *name)
5754 {
5755 
5756 	GROUPTASK_INIT(gtask, 0, fn, ctx);
5757 	taskqgroup_attach(qgroup_if_config_tqg, gtask, gtask, -1, name);
5758 }
5759 
5760 void
5761 iflib_config_gtask_deinit(struct grouptask *gtask)
5762 {
5763 
5764 	taskqgroup_detach(qgroup_if_config_tqg, gtask);
5765 }
5766 
5767 void
5768 iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate)
5769 {
5770 	if_t ifp = ctx->ifc_ifp;
5771 	iflib_txq_t txq = ctx->ifc_txqs;
5772 
5773 	if_setbaudrate(ifp, baudrate);
5774 	if (baudrate >= IF_Gbps(10)) {
5775 		STATE_LOCK(ctx);
5776 		ctx->ifc_flags |= IFC_PREFETCH;
5777 		STATE_UNLOCK(ctx);
5778 	}
5779 	/* If link down, disable watchdog */
5780 	if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) {
5781 		for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++)
5782 			txq->ift_qstatus = IFLIB_QUEUE_IDLE;
5783 	}
5784 	ctx->ifc_link_state = link_state;
5785 	if_link_state_change(ifp, link_state);
5786 }
5787 
5788 static int
5789 iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq)
5790 {
5791 	int credits;
5792 #ifdef INVARIANTS
5793 	int credits_pre = txq->ift_cidx_processed;
5794 #endif
5795 
5796 	if (ctx->isc_txd_credits_update == NULL)
5797 		return (0);
5798 
5799 	if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, true)) == 0)
5800 		return (0);
5801 
5802 	txq->ift_processed += credits;
5803 	txq->ift_cidx_processed += credits;
5804 
5805 	MPASS(credits_pre + credits == txq->ift_cidx_processed);
5806 	if (txq->ift_cidx_processed >= txq->ift_size)
5807 		txq->ift_cidx_processed -= txq->ift_size;
5808 	return (credits);
5809 }
5810 
5811 static int
5812 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget)
5813 {
5814 
5815 	return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx,
5816 	    budget));
5817 }
5818 
5819 void
5820 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name,
5821 	const char *description, if_int_delay_info_t info,
5822 	int offset, int value)
5823 {
5824 	info->iidi_ctx = ctx;
5825 	info->iidi_offset = offset;
5826 	info->iidi_value = value;
5827 	SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev),
5828 	    SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)),
5829 	    OID_AUTO, name, CTLTYPE_INT|CTLFLAG_RW,
5830 	    info, 0, iflib_sysctl_int_delay, "I", description);
5831 }
5832 
5833 struct sx *
5834 iflib_ctx_lock_get(if_ctx_t ctx)
5835 {
5836 
5837 	return (&ctx->ifc_ctx_sx);
5838 }
5839 
5840 static int
5841 iflib_msix_init(if_ctx_t ctx)
5842 {
5843 	device_t dev = ctx->ifc_dev;
5844 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5845 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5846 	int vectors, queues, rx_queues, tx_queues, queuemsgs, msgs;
5847 	int iflib_num_tx_queues, iflib_num_rx_queues;
5848 	int err, admincnt, bar;
5849 
5850 	iflib_num_tx_queues = ctx->ifc_sysctl_ntxqs;
5851 	iflib_num_rx_queues = ctx->ifc_sysctl_nrxqs;
5852 
5853 	device_printf(dev, "msix_init qsets capped at %d\n", imax(scctx->isc_ntxqsets, scctx->isc_nrxqsets));
5854 
5855 	bar = ctx->ifc_softc_ctx.isc_msix_bar;
5856 	admincnt = sctx->isc_admin_intrcnt;
5857 	/* Override by global tuneable */
5858 	{
5859 		int i;
5860 		size_t len = sizeof(i);
5861 		err = kernel_sysctlbyname(curthread, "hw.pci.enable_msix", &i, &len, NULL, 0, NULL, 0);
5862 		if (err == 0) {
5863 			if (i == 0)
5864 				goto msi;
5865 		}
5866 		else {
5867 			device_printf(dev, "unable to read hw.pci.enable_msix.");
5868 		}
5869 	}
5870 	/* Override by tuneable */
5871 	if (scctx->isc_disable_msix)
5872 		goto msi;
5873 
5874 	/*
5875 	** When used in a virtualized environment
5876 	** PCI BUSMASTER capability may not be set
5877 	** so explicity set it here and rewrite
5878 	** the ENABLE in the MSIX control register
5879 	** at this point to cause the host to
5880 	** successfully initialize us.
5881 	*/
5882 	{
5883 		int msix_ctrl, rid;
5884 
5885  		pci_enable_busmaster(dev);
5886 		rid = 0;
5887 		if (pci_find_cap(dev, PCIY_MSIX, &rid) == 0 && rid != 0) {
5888 			rid += PCIR_MSIX_CTRL;
5889 			msix_ctrl = pci_read_config(dev, rid, 2);
5890 			msix_ctrl |= PCIM_MSIXCTRL_MSIX_ENABLE;
5891 			pci_write_config(dev, rid, msix_ctrl, 2);
5892 		} else {
5893 			device_printf(dev, "PCIY_MSIX capability not found; "
5894 			                   "or rid %d == 0.\n", rid);
5895 			goto msi;
5896 		}
5897 	}
5898 
5899 	/*
5900 	 * bar == -1 => "trust me I know what I'm doing"
5901 	 * Some drivers are for hardware that is so shoddily
5902 	 * documented that no one knows which bars are which
5903 	 * so the developer has to map all bars. This hack
5904 	 * allows shoddy garbage to use msix in this framework.
5905 	 */
5906 	if (bar != -1) {
5907 		ctx->ifc_msix_mem = bus_alloc_resource_any(dev,
5908 	            SYS_RES_MEMORY, &bar, RF_ACTIVE);
5909 		if (ctx->ifc_msix_mem == NULL) {
5910 			/* May not be enabled */
5911 			device_printf(dev, "Unable to map MSIX table \n");
5912 			goto msi;
5913 		}
5914 	}
5915 	/* First try MSI/X */
5916 	if ((msgs = pci_msix_count(dev)) == 0) { /* system has msix disabled */
5917 		device_printf(dev, "System has MSIX disabled \n");
5918 		bus_release_resource(dev, SYS_RES_MEMORY,
5919 		    bar, ctx->ifc_msix_mem);
5920 		ctx->ifc_msix_mem = NULL;
5921 		goto msi;
5922 	}
5923 #if IFLIB_DEBUG
5924 	/* use only 1 qset in debug mode */
5925 	queuemsgs = min(msgs - admincnt, 1);
5926 #else
5927 	queuemsgs = msgs - admincnt;
5928 #endif
5929 #ifdef RSS
5930 	queues = imin(queuemsgs, rss_getnumbuckets());
5931 #else
5932 	queues = queuemsgs;
5933 #endif
5934 	queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues);
5935 	device_printf(dev, "pxm cpus: %d queue msgs: %d admincnt: %d\n",
5936 				  CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt);
5937 #ifdef  RSS
5938 	/* If we're doing RSS, clamp at the number of RSS buckets */
5939 	if (queues > rss_getnumbuckets())
5940 		queues = rss_getnumbuckets();
5941 #endif
5942 	if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt)
5943 		rx_queues = iflib_num_rx_queues;
5944 	else
5945 		rx_queues = queues;
5946 
5947 	if (rx_queues > scctx->isc_nrxqsets)
5948 		rx_queues = scctx->isc_nrxqsets;
5949 
5950 	/*
5951 	 * We want this to be all logical CPUs by default
5952 	 */
5953 	if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues)
5954 		tx_queues = iflib_num_tx_queues;
5955 	else
5956 		tx_queues = mp_ncpus;
5957 
5958 	if (tx_queues > scctx->isc_ntxqsets)
5959 		tx_queues = scctx->isc_ntxqsets;
5960 
5961 	if (ctx->ifc_sysctl_qs_eq_override == 0) {
5962 #ifdef INVARIANTS
5963 		if (tx_queues != rx_queues)
5964 			device_printf(dev, "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n",
5965 				      min(rx_queues, tx_queues), min(rx_queues, tx_queues));
5966 #endif
5967 		tx_queues = min(rx_queues, tx_queues);
5968 		rx_queues = min(rx_queues, tx_queues);
5969 	}
5970 
5971 	device_printf(dev, "using %d rx queues %d tx queues \n", rx_queues, tx_queues);
5972 
5973 	vectors = rx_queues + admincnt;
5974 	if ((err = pci_alloc_msix(dev, &vectors)) == 0) {
5975 		device_printf(dev,
5976 					  "Using MSIX interrupts with %d vectors\n", vectors);
5977 		scctx->isc_vectors = vectors;
5978 		scctx->isc_nrxqsets = rx_queues;
5979 		scctx->isc_ntxqsets = tx_queues;
5980 		scctx->isc_intr = IFLIB_INTR_MSIX;
5981 
5982 		return (vectors);
5983 	} else {
5984 		device_printf(dev, "failed to allocate %d msix vectors, err: %d - using MSI\n", vectors, err);
5985 	}
5986 msi:
5987 	vectors = pci_msi_count(dev);
5988 	scctx->isc_nrxqsets = 1;
5989 	scctx->isc_ntxqsets = 1;
5990 	scctx->isc_vectors = vectors;
5991 	if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) {
5992 		device_printf(dev,"Using an MSI interrupt\n");
5993 		scctx->isc_intr = IFLIB_INTR_MSI;
5994 	} else {
5995 		device_printf(dev,"Using a Legacy interrupt\n");
5996 		scctx->isc_intr = IFLIB_INTR_LEGACY;
5997 	}
5998 
5999 	return (vectors);
6000 }
6001 
6002 char * ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" };
6003 
6004 static int
6005 mp_ring_state_handler(SYSCTL_HANDLER_ARGS)
6006 {
6007 	int rc;
6008 	uint16_t *state = ((uint16_t *)oidp->oid_arg1);
6009 	struct sbuf *sb;
6010 	char *ring_state = "UNKNOWN";
6011 
6012 	/* XXX needed ? */
6013 	rc = sysctl_wire_old_buffer(req, 0);
6014 	MPASS(rc == 0);
6015 	if (rc != 0)
6016 		return (rc);
6017 	sb = sbuf_new_for_sysctl(NULL, NULL, 80, req);
6018 	MPASS(sb != NULL);
6019 	if (sb == NULL)
6020 		return (ENOMEM);
6021 	if (state[3] <= 3)
6022 		ring_state = ring_states[state[3]];
6023 
6024 	sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s",
6025 		    state[0], state[1], state[2], ring_state);
6026 	rc = sbuf_finish(sb);
6027 	sbuf_delete(sb);
6028         return(rc);
6029 }
6030 
6031 enum iflib_ndesc_handler {
6032 	IFLIB_NTXD_HANDLER,
6033 	IFLIB_NRXD_HANDLER,
6034 };
6035 
6036 static int
6037 mp_ndesc_handler(SYSCTL_HANDLER_ARGS)
6038 {
6039 	if_ctx_t ctx = (void *)arg1;
6040 	enum iflib_ndesc_handler type = arg2;
6041 	char buf[256] = {0};
6042 	qidx_t *ndesc;
6043 	char *p, *next;
6044 	int nqs, rc, i;
6045 
6046 	MPASS(type == IFLIB_NTXD_HANDLER || type == IFLIB_NRXD_HANDLER);
6047 
6048 	nqs = 8;
6049 	switch(type) {
6050 	case IFLIB_NTXD_HANDLER:
6051 		ndesc = ctx->ifc_sysctl_ntxds;
6052 		if (ctx->ifc_sctx)
6053 			nqs = ctx->ifc_sctx->isc_ntxqs;
6054 		break;
6055 	case IFLIB_NRXD_HANDLER:
6056 		ndesc = ctx->ifc_sysctl_nrxds;
6057 		if (ctx->ifc_sctx)
6058 			nqs = ctx->ifc_sctx->isc_nrxqs;
6059 		break;
6060 	default:
6061 			panic("unhandled type");
6062 	}
6063 	if (nqs == 0)
6064 		nqs = 8;
6065 
6066 	for (i=0; i<8; i++) {
6067 		if (i >= nqs)
6068 			break;
6069 		if (i)
6070 			strcat(buf, ",");
6071 		sprintf(strchr(buf, 0), "%d", ndesc[i]);
6072 	}
6073 
6074 	rc = sysctl_handle_string(oidp, buf, sizeof(buf), req);
6075 	if (rc || req->newptr == NULL)
6076 		return rc;
6077 
6078 	for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p;
6079 	    i++, p = strsep(&next, " ,")) {
6080 		ndesc[i] = strtoul(p, NULL, 10);
6081 	}
6082 
6083 	return(rc);
6084 }
6085 
6086 #define NAME_BUFLEN 32
6087 static void
6088 iflib_add_device_sysctl_pre(if_ctx_t ctx)
6089 {
6090         device_t dev = iflib_get_dev(ctx);
6091 	struct sysctl_oid_list *child, *oid_list;
6092 	struct sysctl_ctx_list *ctx_list;
6093 	struct sysctl_oid *node;
6094 
6095 	ctx_list = device_get_sysctl_ctx(dev);
6096 	child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev));
6097 	ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, "iflib",
6098 						      CTLFLAG_RD, NULL, "IFLIB fields");
6099 	oid_list = SYSCTL_CHILDREN(node);
6100 
6101 	SYSCTL_ADD_STRING(ctx_list, oid_list, OID_AUTO, "driver_version",
6102 		       CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version, 0,
6103 		       "driver version");
6104 
6105 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs",
6106 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0,
6107 			"# of txqs to use, 0 => use default #");
6108 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs",
6109 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0,
6110 			"# of rxqs to use, 0 => use default #");
6111 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_qs_enable",
6112 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0,
6113                        "permit #txq != #rxq");
6114 	SYSCTL_ADD_INT(ctx_list, oid_list, OID_AUTO, "disable_msix",
6115                       CTLFLAG_RWTUN, &ctx->ifc_softc_ctx.isc_disable_msix, 0,
6116                       "disable MSIX (default 0)");
6117 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "rx_budget",
6118 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_rx_budget, 0,
6119                        "set the rx budget");
6120 
6121 	/* XXX change for per-queue sizes */
6122 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_ntxds",
6123 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NTXD_HANDLER,
6124                        mp_ndesc_handler, "A",
6125                        "list of # of tx descriptors to use, 0 = use default #");
6126 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_nrxds",
6127 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NRXD_HANDLER,
6128                        mp_ndesc_handler, "A",
6129                        "list of # of rx descriptors to use, 0 = use default #");
6130 }
6131 
6132 static void
6133 iflib_add_device_sysctl_post(if_ctx_t ctx)
6134 {
6135 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6136 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6137         device_t dev = iflib_get_dev(ctx);
6138 	struct sysctl_oid_list *child;
6139 	struct sysctl_ctx_list *ctx_list;
6140 	iflib_fl_t fl;
6141 	iflib_txq_t txq;
6142 	iflib_rxq_t rxq;
6143 	int i, j;
6144 	char namebuf[NAME_BUFLEN];
6145 	char *qfmt;
6146 	struct sysctl_oid *queue_node, *fl_node, *node;
6147 	struct sysctl_oid_list *queue_list, *fl_list;
6148 	ctx_list = device_get_sysctl_ctx(dev);
6149 
6150 	node = ctx->ifc_sysctl_node;
6151 	child = SYSCTL_CHILDREN(node);
6152 
6153 	if (scctx->isc_ntxqsets > 100)
6154 		qfmt = "txq%03d";
6155 	else if (scctx->isc_ntxqsets > 10)
6156 		qfmt = "txq%02d";
6157 	else
6158 		qfmt = "txq%d";
6159 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
6160 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6161 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6162 					     CTLFLAG_RD, NULL, "Queue Name");
6163 		queue_list = SYSCTL_CHILDREN(queue_node);
6164 #if MEMORY_LOGGING
6165 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued",
6166 				CTLFLAG_RD,
6167 				&txq->ift_dequeued, "total mbufs freed");
6168 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued",
6169 				CTLFLAG_RD,
6170 				&txq->ift_enqueued, "total mbufs enqueued");
6171 #endif
6172 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag",
6173 				   CTLFLAG_RD,
6174 				   &txq->ift_mbuf_defrag, "# of times m_defrag was called");
6175 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "m_pullups",
6176 				   CTLFLAG_RD,
6177 				   &txq->ift_pullups, "# of times m_pullup was called");
6178 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag_failed",
6179 				   CTLFLAG_RD,
6180 				   &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed");
6181 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_desc_avail",
6182 				   CTLFLAG_RD,
6183 				   &txq->ift_no_desc_avail, "# of times no descriptors were available");
6184 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "tx_map_failed",
6185 				   CTLFLAG_RD,
6186 				   &txq->ift_map_failed, "# of times dma map failed");
6187 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txd_encap_efbig",
6188 				   CTLFLAG_RD,
6189 				   &txq->ift_txd_encap_efbig, "# of times txd_encap returned EFBIG");
6190 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_tx_dma_setup",
6191 				   CTLFLAG_RD,
6192 				   &txq->ift_no_tx_dma_setup, "# of times map failed for other than EFBIG");
6193 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx",
6194 				   CTLFLAG_RD,
6195 				   &txq->ift_pidx, 1, "Producer Index");
6196 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx",
6197 				   CTLFLAG_RD,
6198 				   &txq->ift_cidx, 1, "Consumer Index");
6199 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx_processed",
6200 				   CTLFLAG_RD,
6201 				   &txq->ift_cidx_processed, 1, "Consumer Index seen by credit update");
6202 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use",
6203 				   CTLFLAG_RD,
6204 				   &txq->ift_in_use, 1, "descriptors in use");
6205 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_processed",
6206 				   CTLFLAG_RD,
6207 				   &txq->ift_processed, "descriptors procesed for clean");
6208 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned",
6209 				   CTLFLAG_RD,
6210 				   &txq->ift_cleaned, "total cleaned");
6211 		SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state",
6212 				CTLTYPE_STRING | CTLFLAG_RD, __DEVOLATILE(uint64_t *, &txq->ift_br->state),
6213 				0, mp_ring_state_handler, "A", "soft ring state");
6214 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_enqueues",
6215 				       CTLFLAG_RD, &txq->ift_br->enqueues,
6216 				       "# of enqueues to the mp_ring for this queue");
6217 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_drops",
6218 				       CTLFLAG_RD, &txq->ift_br->drops,
6219 				       "# of drops in the mp_ring for this queue");
6220 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_starts",
6221 				       CTLFLAG_RD, &txq->ift_br->starts,
6222 				       "# of normal consumer starts in the mp_ring for this queue");
6223 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_stalls",
6224 				       CTLFLAG_RD, &txq->ift_br->stalls,
6225 					       "# of consumer stalls in the mp_ring for this queue");
6226 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_restarts",
6227 			       CTLFLAG_RD, &txq->ift_br->restarts,
6228 				       "# of consumer restarts in the mp_ring for this queue");
6229 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_abdications",
6230 				       CTLFLAG_RD, &txq->ift_br->abdications,
6231 				       "# of consumer abdications in the mp_ring for this queue");
6232 	}
6233 
6234 	if (scctx->isc_nrxqsets > 100)
6235 		qfmt = "rxq%03d";
6236 	else if (scctx->isc_nrxqsets > 10)
6237 		qfmt = "rxq%02d";
6238 	else
6239 		qfmt = "rxq%d";
6240 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
6241 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6242 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6243 					     CTLFLAG_RD, NULL, "Queue Name");
6244 		queue_list = SYSCTL_CHILDREN(queue_node);
6245 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
6246 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_pidx",
6247 				       CTLFLAG_RD,
6248 				       &rxq->ifr_cq_pidx, 1, "Producer Index");
6249 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_cidx",
6250 				       CTLFLAG_RD,
6251 				       &rxq->ifr_cq_cidx, 1, "Consumer Index");
6252 		}
6253 
6254 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
6255 			snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j);
6256 			fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list, OID_AUTO, namebuf,
6257 						     CTLFLAG_RD, NULL, "freelist Name");
6258 			fl_list = SYSCTL_CHILDREN(fl_node);
6259 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx",
6260 				       CTLFLAG_RD,
6261 				       &fl->ifl_pidx, 1, "Producer Index");
6262 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx",
6263 				       CTLFLAG_RD,
6264 				       &fl->ifl_cidx, 1, "Consumer Index");
6265 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits",
6266 				       CTLFLAG_RD,
6267 				       &fl->ifl_credits, 1, "credits available");
6268 #if MEMORY_LOGGING
6269 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_enqueued",
6270 					CTLFLAG_RD,
6271 					&fl->ifl_m_enqueued, "mbufs allocated");
6272 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_dequeued",
6273 					CTLFLAG_RD,
6274 					&fl->ifl_m_dequeued, "mbufs freed");
6275 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_enqueued",
6276 					CTLFLAG_RD,
6277 					&fl->ifl_cl_enqueued, "clusters allocated");
6278 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_dequeued",
6279 					CTLFLAG_RD,
6280 					&fl->ifl_cl_dequeued, "clusters freed");
6281 #endif
6282 
6283 		}
6284 	}
6285 
6286 }
6287 
6288 #ifndef __NO_STRICT_ALIGNMENT
6289 static struct mbuf *
6290 iflib_fixup_rx(struct mbuf *m)
6291 {
6292 	struct mbuf *n;
6293 
6294 	if (m->m_len <= (MCLBYTES - ETHER_HDR_LEN)) {
6295 		bcopy(m->m_data, m->m_data + ETHER_HDR_LEN, m->m_len);
6296 		m->m_data += ETHER_HDR_LEN;
6297 		n = m;
6298 	} else {
6299 		MGETHDR(n, M_NOWAIT, MT_DATA);
6300 		if (n == NULL) {
6301 			m_freem(m);
6302 			return (NULL);
6303 		}
6304 		bcopy(m->m_data, n->m_data, ETHER_HDR_LEN);
6305 		m->m_data += ETHER_HDR_LEN;
6306 		m->m_len -= ETHER_HDR_LEN;
6307 		n->m_len = ETHER_HDR_LEN;
6308 		M_MOVE_PKTHDR(n, m);
6309 		n->m_next = m;
6310 	}
6311 	return (n);
6312 }
6313 #endif
6314 
6315 #ifdef NETDUMP
6316 static void
6317 iflib_netdump_init(struct ifnet *ifp, int *nrxr, int *ncl, int *clsize)
6318 {
6319 	if_ctx_t ctx;
6320 
6321 	ctx = if_getsoftc(ifp);
6322 	CTX_LOCK(ctx);
6323 	*nrxr = NRXQSETS(ctx);
6324 	*ncl = ctx->ifc_rxqs[0].ifr_fl->ifl_size;
6325 	*clsize = ctx->ifc_rxqs[0].ifr_fl->ifl_buf_size;
6326 	CTX_UNLOCK(ctx);
6327 }
6328 
6329 static void
6330 iflib_netdump_event(struct ifnet *ifp, enum netdump_ev event)
6331 {
6332 	if_ctx_t ctx;
6333 	if_softc_ctx_t scctx;
6334 	iflib_fl_t fl;
6335 	iflib_rxq_t rxq;
6336 	int i, j;
6337 
6338 	ctx = if_getsoftc(ifp);
6339 	scctx = &ctx->ifc_softc_ctx;
6340 
6341 	switch (event) {
6342 	case NETDUMP_START:
6343 		for (i = 0; i < scctx->isc_nrxqsets; i++) {
6344 			rxq = &ctx->ifc_rxqs[i];
6345 			for (j = 0; j < rxq->ifr_nfl; j++) {
6346 				fl = rxq->ifr_fl;
6347 				fl->ifl_zone = m_getzone(fl->ifl_buf_size);
6348 			}
6349 		}
6350 		iflib_no_tx_batch = 1;
6351 		break;
6352 	default:
6353 		break;
6354 	}
6355 }
6356 
6357 static int
6358 iflib_netdump_transmit(struct ifnet *ifp, struct mbuf *m)
6359 {
6360 	if_ctx_t ctx;
6361 	iflib_txq_t txq;
6362 	int error;
6363 
6364 	ctx = if_getsoftc(ifp);
6365 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
6366 	    IFF_DRV_RUNNING)
6367 		return (EBUSY);
6368 
6369 	txq = &ctx->ifc_txqs[0];
6370 	error = iflib_encap(txq, &m);
6371 	if (error == 0)
6372 		(void)iflib_txd_db_check(ctx, txq, true, txq->ift_in_use);
6373 	return (error);
6374 }
6375 
6376 static int
6377 iflib_netdump_poll(struct ifnet *ifp, int count)
6378 {
6379 	if_ctx_t ctx;
6380 	if_softc_ctx_t scctx;
6381 	iflib_txq_t txq;
6382 	int i;
6383 
6384 	ctx = if_getsoftc(ifp);
6385 	scctx = &ctx->ifc_softc_ctx;
6386 
6387 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
6388 	    IFF_DRV_RUNNING)
6389 		return (EBUSY);
6390 
6391 	txq = &ctx->ifc_txqs[0];
6392 	(void)iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
6393 
6394 	for (i = 0; i < scctx->isc_nrxqsets; i++)
6395 		(void)iflib_rxeof(&ctx->ifc_rxqs[i], 16 /* XXX */);
6396 	return (0);
6397 }
6398 #endif /* NETDUMP */
6399