xref: /freebsd/sys/net/iflib.c (revision 7f87c0406df8e22d6a76b2d1245043619b40e983)
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, const 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 	if (nm_i != head) {	/* we have new packets to send */
964 		pkt_info_zero(&pi);
965 		pi.ipi_segs = txq->ift_segs;
966 		pi.ipi_qsidx = kring->ring_id;
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 there are unclaimed buffers, attempt to reclaim them.
1030 	 * If none are reclaimed, and TX IRQs are not in use, do an initial
1031 	 * minimal delay, then trigger the tx handler which will spin in the
1032 	 * group task queue.
1033 	 */
1034 	if (kring->nr_hwtail != nm_prev(head, lim)) {
1035 		if (iflib_tx_credits_update(ctx, txq)) {
1036 			/* some tx completed, increment avail */
1037 			nic_i = txq->ift_cidx_processed;
1038 			kring->nr_hwtail = nm_prev(netmap_idx_n2k(kring, nic_i), lim);
1039 		}
1040 		else {
1041 			if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ)) {
1042 				DELAY(1);
1043 				GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txq->ift_id].ift_task);
1044 			}
1045 		}
1046 	}
1047 	return (0);
1048 }
1049 
1050 /*
1051  * Reconcile kernel and user view of the receive ring.
1052  * Same as for the txsync, this routine must be efficient.
1053  * The caller guarantees a single invocations, but races against
1054  * the rest of the driver should be handled here.
1055  *
1056  * On call, kring->rhead is the first packet that userspace wants
1057  * to keep, and kring->rcur is the wakeup point.
1058  * The kernel has previously reported packets up to kring->rtail.
1059  *
1060  * If (flags & NAF_FORCE_READ) also check for incoming packets irrespective
1061  * of whether or not we received an interrupt.
1062  */
1063 static int
1064 iflib_netmap_rxsync(struct netmap_kring *kring, int flags)
1065 {
1066 	struct netmap_adapter *na = kring->na;
1067 	struct netmap_ring *ring = kring->ring;
1068 	uint32_t nm_i;	/* index into the netmap ring */
1069 	uint32_t nic_i;	/* index into the NIC ring */
1070 	u_int i, n;
1071 	u_int const lim = kring->nkr_num_slots - 1;
1072 	u_int const head = netmap_idx_n2k(kring, kring->rhead);
1073 	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
1074 	struct if_rxd_info ri;
1075 
1076 	struct ifnet *ifp = na->ifp;
1077 	if_ctx_t ctx = ifp->if_softc;
1078 	iflib_rxq_t rxq = &ctx->ifc_rxqs[kring->ring_id];
1079 	iflib_fl_t fl = rxq->ifr_fl;
1080 	if (head > lim)
1081 		return netmap_ring_reinit(kring);
1082 
1083 	/* XXX check sync modes */
1084 	for (i = 0, fl = rxq->ifr_fl; i < rxq->ifr_nfl; i++, fl++) {
1085 		if (fl->ifl_sds.ifsd_map == NULL)
1086 			continue;
1087 		bus_dmamap_sync(rxq->ifr_fl[i].ifl_desc_tag, fl->ifl_ifdi->idi_map,
1088 				BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1089 	}
1090 	/*
1091 	 * First part: import newly received packets.
1092 	 *
1093 	 * nm_i is the index of the next free slot in the netmap ring,
1094 	 * nic_i is the index of the next received packet in the NIC ring,
1095 	 * and they may differ in case if_init() has been called while
1096 	 * in netmap mode. For the receive ring we have
1097 	 *
1098 	 *	nic_i = rxr->next_check;
1099 	 *	nm_i = kring->nr_hwtail (previous)
1100 	 * and
1101 	 *	nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1102 	 *
1103 	 * rxr->next_check is set to 0 on a ring reinit
1104 	 */
1105 	if (netmap_no_pendintr || force_update) {
1106 		int crclen = iflib_crcstrip ? 0 : 4;
1107 		int error, avail;
1108 
1109 		for (i = 0; i < rxq->ifr_nfl; i++) {
1110 			fl = &rxq->ifr_fl[i];
1111 			nic_i = fl->ifl_cidx;
1112 			nm_i = netmap_idx_n2k(kring, nic_i);
1113 			avail = iflib_rxd_avail(ctx, rxq, nic_i, USHRT_MAX);
1114 			for (n = 0; avail > 0; n++, avail--) {
1115 				rxd_info_zero(&ri);
1116 				ri.iri_frags = rxq->ifr_frags;
1117 				ri.iri_qsidx = kring->ring_id;
1118 				ri.iri_ifp = ctx->ifc_ifp;
1119 				ri.iri_cidx = nic_i;
1120 
1121 				error = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
1122 				ring->slot[nm_i].len = error ? 0 : ri.iri_len - crclen;
1123 				ring->slot[nm_i].flags = 0;
1124 				if (fl->ifl_sds.ifsd_map)
1125 					bus_dmamap_sync(fl->ifl_ifdi->idi_tag,
1126 							fl->ifl_sds.ifsd_map[nic_i], BUS_DMASYNC_POSTREAD);
1127 				nm_i = nm_next(nm_i, lim);
1128 				nic_i = nm_next(nic_i, lim);
1129 			}
1130 			if (n) { /* update the state variables */
1131 				if (netmap_no_pendintr && !force_update) {
1132 					/* diagnostics */
1133 					iflib_rx_miss ++;
1134 					iflib_rx_miss_bufs += n;
1135 				}
1136 				fl->ifl_cidx = nic_i;
1137 				kring->nr_hwtail = netmap_idx_k2n(kring, nm_i);
1138 			}
1139 			kring->nr_kflags &= ~NKR_PENDINTR;
1140 		}
1141 	}
1142 	/*
1143 	 * Second part: skip past packets that userspace has released.
1144 	 * (kring->nr_hwcur to head excluded),
1145 	 * and make the buffers available for reception.
1146 	 * As usual nm_i is the index in the netmap ring,
1147 	 * nic_i is the index in the NIC ring, and
1148 	 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1149 	 */
1150 	/* XXX not sure how this will work with multiple free lists */
1151 	nm_i = netmap_idx_n2k(kring, kring->nr_hwcur);
1152 
1153 	return (netmap_fl_refill(rxq, kring, nm_i, false));
1154 }
1155 
1156 static void
1157 iflib_netmap_intr(struct netmap_adapter *na, int onoff)
1158 {
1159 	struct ifnet *ifp = na->ifp;
1160 	if_ctx_t ctx = ifp->if_softc;
1161 
1162 	CTX_LOCK(ctx);
1163 	if (onoff) {
1164 		IFDI_INTR_ENABLE(ctx);
1165 	} else {
1166 		IFDI_INTR_DISABLE(ctx);
1167 	}
1168 	CTX_UNLOCK(ctx);
1169 }
1170 
1171 
1172 static int
1173 iflib_netmap_attach(if_ctx_t ctx)
1174 {
1175 	struct netmap_adapter na;
1176 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1177 
1178 	bzero(&na, sizeof(na));
1179 
1180 	na.ifp = ctx->ifc_ifp;
1181 	na.na_flags = NAF_BDG_MAYSLEEP;
1182 	MPASS(ctx->ifc_softc_ctx.isc_ntxqsets);
1183 	MPASS(ctx->ifc_softc_ctx.isc_nrxqsets);
1184 
1185 	na.num_tx_desc = scctx->isc_ntxd[0];
1186 	na.num_rx_desc = scctx->isc_nrxd[0];
1187 	na.nm_txsync = iflib_netmap_txsync;
1188 	na.nm_rxsync = iflib_netmap_rxsync;
1189 	na.nm_register = iflib_netmap_register;
1190 	na.nm_intr = iflib_netmap_intr;
1191 	na.num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
1192 	na.num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
1193 	return (netmap_attach(&na));
1194 }
1195 
1196 static void
1197 iflib_netmap_txq_init(if_ctx_t ctx, iflib_txq_t txq)
1198 {
1199 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1200 	struct netmap_slot *slot;
1201 
1202 	slot = netmap_reset(na, NR_TX, txq->ift_id, 0);
1203 	if (slot == NULL)
1204 		return;
1205 	if (txq->ift_sds.ifsd_map == NULL)
1206 		return;
1207 
1208 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxd[0]; i++) {
1209 
1210 		/*
1211 		 * In netmap mode, set the map for the packet buffer.
1212 		 * NOTE: Some drivers (not this one) also need to set
1213 		 * the physical buffer address in the NIC ring.
1214 		 * netmap_idx_n2k() maps a nic index, i, into the corresponding
1215 		 * netmap slot index, si
1216 		 */
1217 		int si = netmap_idx_n2k(na->tx_rings[txq->ift_id], i);
1218 		netmap_load_map(na, txq->ift_desc_tag, txq->ift_sds.ifsd_map[i], NMB(na, slot + si));
1219 	}
1220 }
1221 
1222 static void
1223 iflib_netmap_rxq_init(if_ctx_t ctx, iflib_rxq_t rxq)
1224 {
1225 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1226 	struct netmap_kring *kring = na->rx_rings[rxq->ifr_id];
1227 	struct netmap_slot *slot;
1228 	uint32_t nm_i;
1229 
1230 	slot = netmap_reset(na, NR_RX, rxq->ifr_id, 0);
1231 	if (slot == NULL)
1232 		return;
1233 	nm_i = netmap_idx_n2k(kring, 0);
1234 	netmap_fl_refill(rxq, kring, nm_i, true);
1235 }
1236 
1237 #define iflib_netmap_detach(ifp) netmap_detach(ifp)
1238 
1239 #else
1240 #define iflib_netmap_txq_init(ctx, txq)
1241 #define iflib_netmap_rxq_init(ctx, rxq)
1242 #define iflib_netmap_detach(ifp)
1243 
1244 #define iflib_netmap_attach(ctx) (0)
1245 #define netmap_rx_irq(ifp, qid, budget) (0)
1246 #define netmap_tx_irq(ifp, qid) do {} while (0)
1247 
1248 #endif
1249 
1250 #if defined(__i386__) || defined(__amd64__)
1251 static __inline void
1252 prefetch(void *x)
1253 {
1254 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1255 }
1256 static __inline void
1257 prefetch2cachelines(void *x)
1258 {
1259 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1260 #if (CACHE_LINE_SIZE < 128)
1261 	__asm volatile("prefetcht0 %0" :: "m" (*(((unsigned long *)x)+CACHE_LINE_SIZE/(sizeof(unsigned long)))));
1262 #endif
1263 }
1264 #else
1265 #define prefetch(x)
1266 #define prefetch2cachelines(x)
1267 #endif
1268 
1269 static void
1270 iflib_gen_mac(if_ctx_t ctx)
1271 {
1272 	struct thread *td;
1273 	MD5_CTX mdctx;
1274 	char uuid[HOSTUUIDLEN+1];
1275 	char buf[HOSTUUIDLEN+16];
1276 	uint8_t *mac;
1277 	unsigned char digest[16];
1278 
1279 	td = curthread;
1280 	mac = ctx->ifc_mac;
1281 	uuid[HOSTUUIDLEN] = 0;
1282 	bcopy(td->td_ucred->cr_prison->pr_hostuuid, uuid, HOSTUUIDLEN);
1283 	snprintf(buf, HOSTUUIDLEN+16, "%s-%s", uuid, device_get_nameunit(ctx->ifc_dev));
1284 	/*
1285 	 * Generate a pseudo-random, deterministic MAC
1286 	 * address based on the UUID and unit number.
1287 	 * The FreeBSD Foundation OUI of 58-9C-FC is used.
1288 	 */
1289 	MD5Init(&mdctx);
1290 	MD5Update(&mdctx, buf, strlen(buf));
1291 	MD5Final(digest, &mdctx);
1292 
1293 	mac[0] = 0x58;
1294 	mac[1] = 0x9C;
1295 	mac[2] = 0xFC;
1296 	mac[3] = digest[0];
1297 	mac[4] = digest[1];
1298 	mac[5] = digest[2];
1299 }
1300 
1301 static void
1302 iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid)
1303 {
1304 	iflib_fl_t fl;
1305 
1306 	fl = &rxq->ifr_fl[flid];
1307 	iru->iru_paddrs = fl->ifl_bus_addrs;
1308 	iru->iru_vaddrs = &fl->ifl_vm_addrs[0];
1309 	iru->iru_idxs = fl->ifl_rxd_idxs;
1310 	iru->iru_qsidx = rxq->ifr_id;
1311 	iru->iru_buf_size = fl->ifl_buf_size;
1312 	iru->iru_flidx = fl->ifl_id;
1313 }
1314 
1315 static void
1316 _iflib_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int err)
1317 {
1318 	if (err)
1319 		return;
1320 	*(bus_addr_t *) arg = segs[0].ds_addr;
1321 }
1322 
1323 int
1324 iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags)
1325 {
1326 	int err;
1327 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1328 	device_t dev = ctx->ifc_dev;
1329 
1330 	KASSERT(sctx->isc_q_align != 0, ("alignment value not initialized"));
1331 
1332 	err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1333 				sctx->isc_q_align, 0,	/* alignment, bounds */
1334 				BUS_SPACE_MAXADDR,	/* lowaddr */
1335 				BUS_SPACE_MAXADDR,	/* highaddr */
1336 				NULL, NULL,		/* filter, filterarg */
1337 				size,			/* maxsize */
1338 				1,			/* nsegments */
1339 				size,			/* maxsegsize */
1340 				BUS_DMA_ALLOCNOW,	/* flags */
1341 				NULL,			/* lockfunc */
1342 				NULL,			/* lockarg */
1343 				&dma->idi_tag);
1344 	if (err) {
1345 		device_printf(dev,
1346 		    "%s: bus_dma_tag_create failed: %d\n",
1347 		    __func__, err);
1348 		goto fail_0;
1349 	}
1350 
1351 	err = bus_dmamem_alloc(dma->idi_tag, (void**) &dma->idi_vaddr,
1352 	    BUS_DMA_NOWAIT | BUS_DMA_COHERENT | BUS_DMA_ZERO, &dma->idi_map);
1353 	if (err) {
1354 		device_printf(dev,
1355 		    "%s: bus_dmamem_alloc(%ju) failed: %d\n",
1356 		    __func__, (uintmax_t)size, err);
1357 		goto fail_1;
1358 	}
1359 
1360 	dma->idi_paddr = IF_BAD_DMA;
1361 	err = bus_dmamap_load(dma->idi_tag, dma->idi_map, dma->idi_vaddr,
1362 	    size, _iflib_dmamap_cb, &dma->idi_paddr, mapflags | BUS_DMA_NOWAIT);
1363 	if (err || dma->idi_paddr == IF_BAD_DMA) {
1364 		device_printf(dev,
1365 		    "%s: bus_dmamap_load failed: %d\n",
1366 		    __func__, err);
1367 		goto fail_2;
1368 	}
1369 
1370 	dma->idi_size = size;
1371 	return (0);
1372 
1373 fail_2:
1374 	bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1375 fail_1:
1376 	bus_dma_tag_destroy(dma->idi_tag);
1377 fail_0:
1378 	dma->idi_tag = NULL;
1379 
1380 	return (err);
1381 }
1382 
1383 int
1384 iflib_dma_alloc_multi(if_ctx_t ctx, int *sizes, iflib_dma_info_t *dmalist, int mapflags, int count)
1385 {
1386 	int i, err;
1387 	iflib_dma_info_t *dmaiter;
1388 
1389 	dmaiter = dmalist;
1390 	for (i = 0; i < count; i++, dmaiter++) {
1391 		if ((err = iflib_dma_alloc(ctx, sizes[i], *dmaiter, mapflags)) != 0)
1392 			break;
1393 	}
1394 	if (err)
1395 		iflib_dma_free_multi(dmalist, i);
1396 	return (err);
1397 }
1398 
1399 void
1400 iflib_dma_free(iflib_dma_info_t dma)
1401 {
1402 	if (dma->idi_tag == NULL)
1403 		return;
1404 	if (dma->idi_paddr != IF_BAD_DMA) {
1405 		bus_dmamap_sync(dma->idi_tag, dma->idi_map,
1406 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1407 		bus_dmamap_unload(dma->idi_tag, dma->idi_map);
1408 		dma->idi_paddr = IF_BAD_DMA;
1409 	}
1410 	if (dma->idi_vaddr != NULL) {
1411 		bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1412 		dma->idi_vaddr = NULL;
1413 	}
1414 	bus_dma_tag_destroy(dma->idi_tag);
1415 	dma->idi_tag = NULL;
1416 }
1417 
1418 void
1419 iflib_dma_free_multi(iflib_dma_info_t *dmalist, int count)
1420 {
1421 	int i;
1422 	iflib_dma_info_t *dmaiter = dmalist;
1423 
1424 	for (i = 0; i < count; i++, dmaiter++)
1425 		iflib_dma_free(*dmaiter);
1426 }
1427 
1428 #ifdef EARLY_AP_STARTUP
1429 static const int iflib_started = 1;
1430 #else
1431 /*
1432  * We used to abuse the smp_started flag to decide if the queues have been
1433  * fully initialized (by late taskqgroup_adjust() calls in a SYSINIT()).
1434  * That gave bad races, since the SYSINIT() runs strictly after smp_started
1435  * is set.  Run a SYSINIT() strictly after that to just set a usable
1436  * completion flag.
1437  */
1438 
1439 static int iflib_started;
1440 
1441 static void
1442 iflib_record_started(void *arg)
1443 {
1444 	iflib_started = 1;
1445 }
1446 
1447 SYSINIT(iflib_record_started, SI_SUB_SMP + 1, SI_ORDER_FIRST,
1448 	iflib_record_started, NULL);
1449 #endif
1450 
1451 static int
1452 iflib_fast_intr(void *arg)
1453 {
1454 	iflib_filter_info_t info = arg;
1455 	struct grouptask *gtask = info->ifi_task;
1456 	if (!iflib_started)
1457 		return (FILTER_HANDLED);
1458 
1459 	DBG_COUNTER_INC(fast_intrs);
1460 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1461 		return (FILTER_HANDLED);
1462 
1463 	GROUPTASK_ENQUEUE(gtask);
1464 	return (FILTER_HANDLED);
1465 }
1466 
1467 static int
1468 iflib_fast_intr_rxtx(void *arg)
1469 {
1470 	iflib_filter_info_t info = arg;
1471 	struct grouptask *gtask = info->ifi_task;
1472 	iflib_rxq_t rxq = (iflib_rxq_t)info->ifi_ctx;
1473 	if_ctx_t ctx = NULL;;
1474 	int i, cidx;
1475 
1476 	if (!iflib_started)
1477 		return (FILTER_HANDLED);
1478 
1479 	DBG_COUNTER_INC(fast_intrs);
1480 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1481 		return (FILTER_HANDLED);
1482 
1483 	MPASS(rxq->ifr_ntxqirq);
1484 	for (i = 0; i < rxq->ifr_ntxqirq; i++) {
1485 		qidx_t txqid = rxq->ifr_txqid[i];
1486 
1487 		ctx = rxq->ifr_ctx;
1488 
1489 		if (!ctx->isc_txd_credits_update(ctx->ifc_softc, txqid, false)) {
1490 			IFDI_TX_QUEUE_INTR_ENABLE(ctx, txqid);
1491 			continue;
1492 		}
1493 		GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
1494 	}
1495 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_RXCQ)
1496 		cidx = rxq->ifr_cq_cidx;
1497 	else
1498 		cidx = rxq->ifr_fl[0].ifl_cidx;
1499 	if (iflib_rxd_avail(ctx, rxq, cidx, 1))
1500 		GROUPTASK_ENQUEUE(gtask);
1501 	else
1502 		IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
1503 	return (FILTER_HANDLED);
1504 }
1505 
1506 
1507 static int
1508 iflib_fast_intr_ctx(void *arg)
1509 {
1510 	iflib_filter_info_t info = arg;
1511 	struct grouptask *gtask = info->ifi_task;
1512 
1513 	if (!iflib_started)
1514 		return (FILTER_HANDLED);
1515 
1516 	DBG_COUNTER_INC(fast_intrs);
1517 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1518 		return (FILTER_HANDLED);
1519 
1520 	GROUPTASK_ENQUEUE(gtask);
1521 	return (FILTER_HANDLED);
1522 }
1523 
1524 static int
1525 _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
1526 		 driver_filter_t filter, driver_intr_t handler, void *arg,
1527 		 const char *name)
1528 {
1529 	int rc, flags;
1530 	struct resource *res;
1531 	void *tag = NULL;
1532 	device_t dev = ctx->ifc_dev;
1533 
1534 	flags = RF_ACTIVE;
1535 	if (ctx->ifc_flags & IFC_LEGACY)
1536 		flags |= RF_SHAREABLE;
1537 	MPASS(rid < 512);
1538 	irq->ii_rid = rid;
1539 	res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &irq->ii_rid, flags);
1540 	if (res == NULL) {
1541 		device_printf(dev,
1542 		    "failed to allocate IRQ for rid %d, name %s.\n", rid, name);
1543 		return (ENOMEM);
1544 	}
1545 	irq->ii_res = res;
1546 	KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL"));
1547 	rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET,
1548 						filter, handler, arg, &tag);
1549 	if (rc != 0) {
1550 		device_printf(dev,
1551 		    "failed to setup interrupt for rid %d, name %s: %d\n",
1552 					  rid, name ? name : "unknown", rc);
1553 		return (rc);
1554 	} else if (name)
1555 		bus_describe_intr(dev, res, tag, "%s", name);
1556 
1557 	irq->ii_tag = tag;
1558 	return (0);
1559 }
1560 
1561 
1562 /*********************************************************************
1563  *
1564  *  Allocate memory for tx_buffer structures. The tx_buffer stores all
1565  *  the information needed to transmit a packet on the wire. This is
1566  *  called only once at attach, setup is done every reset.
1567  *
1568  **********************************************************************/
1569 
1570 static int
1571 iflib_txsd_alloc(iflib_txq_t txq)
1572 {
1573 	if_ctx_t ctx = txq->ift_ctx;
1574 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1575 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1576 	device_t dev = ctx->ifc_dev;
1577 	bus_size_t tsomaxsize;
1578 	int err, nsegments, ntsosegments;
1579 
1580 	nsegments = scctx->isc_tx_nsegments;
1581 	ntsosegments = scctx->isc_tx_tso_segments_max;
1582 	tsomaxsize = scctx->isc_tx_tso_size_max;
1583 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_VLAN_MTU)
1584 		tsomaxsize += sizeof(struct ether_vlan_header);
1585 	MPASS(scctx->isc_ntxd[0] > 0);
1586 	MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0);
1587 	MPASS(nsegments > 0);
1588 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) {
1589 		MPASS(ntsosegments > 0);
1590 		MPASS(sctx->isc_tso_maxsize >= tsomaxsize);
1591 	}
1592 
1593 	/*
1594 	 * Setup DMA descriptor areas.
1595 	 */
1596 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1597 			       1, 0,			/* alignment, bounds */
1598 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1599 			       BUS_SPACE_MAXADDR,	/* highaddr */
1600 			       NULL, NULL,		/* filter, filterarg */
1601 			       sctx->isc_tx_maxsize,		/* maxsize */
1602 			       nsegments,	/* nsegments */
1603 			       sctx->isc_tx_maxsegsize,	/* maxsegsize */
1604 			       0,			/* flags */
1605 			       NULL,			/* lockfunc */
1606 			       NULL,			/* lockfuncarg */
1607 			       &txq->ift_desc_tag))) {
1608 		device_printf(dev,"Unable to allocate TX DMA tag: %d\n", err);
1609 		device_printf(dev,"maxsize: %ju nsegments: %d maxsegsize: %ju\n",
1610 		    (uintmax_t)sctx->isc_tx_maxsize, nsegments, (uintmax_t)sctx->isc_tx_maxsegsize);
1611 		goto fail;
1612 	}
1613 	if ((if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) &
1614 	    (err = bus_dma_tag_create(bus_get_dma_tag(dev),
1615 			       1, 0,			/* alignment, bounds */
1616 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1617 			       BUS_SPACE_MAXADDR,	/* highaddr */
1618 			       NULL, NULL,		/* filter, filterarg */
1619 			       tsomaxsize,		/* maxsize */
1620 			       ntsosegments,	/* nsegments */
1621 			       sctx->isc_tso_maxsegsize,/* maxsegsize */
1622 			       0,			/* flags */
1623 			       NULL,			/* lockfunc */
1624 			       NULL,			/* lockfuncarg */
1625 			       &txq->ift_tso_desc_tag))) {
1626 		device_printf(dev,"Unable to allocate TX TSO DMA tag: %d\n", err);
1627 
1628 		goto fail;
1629 	}
1630 	if (!(txq->ift_sds.ifsd_flags =
1631 	    (uint8_t *) malloc(sizeof(uint8_t) *
1632 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1633 		device_printf(dev, "Unable to allocate tx_buffer memory\n");
1634 		err = ENOMEM;
1635 		goto fail;
1636 	}
1637 	if (!(txq->ift_sds.ifsd_m =
1638 	    (struct mbuf **) malloc(sizeof(struct mbuf *) *
1639 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1640 		device_printf(dev, "Unable to allocate tx_buffer memory\n");
1641 		err = ENOMEM;
1642 		goto fail;
1643 	}
1644 
1645         /* Create the descriptor buffer dma maps */
1646 #if defined(ACPI_DMAR) || (! (defined(__i386__) || defined(__amd64__)))
1647 	if ((ctx->ifc_flags & IFC_DMAR) == 0)
1648 		return (0);
1649 
1650 	if (!(txq->ift_sds.ifsd_map =
1651 	    (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1652 		device_printf(dev, "Unable to allocate tx_buffer map memory\n");
1653 		err = ENOMEM;
1654 		goto fail;
1655 	}
1656 
1657 	for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) {
1658 		err = bus_dmamap_create(txq->ift_desc_tag, 0, &txq->ift_sds.ifsd_map[i]);
1659 		if (err != 0) {
1660 			device_printf(dev, "Unable to create TX DMA map\n");
1661 			goto fail;
1662 		}
1663 	}
1664 #endif
1665 	return (0);
1666 fail:
1667 	/* We free all, it handles case where we are in the middle */
1668 	iflib_tx_structures_free(ctx);
1669 	return (err);
1670 }
1671 
1672 static void
1673 iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i)
1674 {
1675 	bus_dmamap_t map;
1676 
1677 	map = NULL;
1678 	if (txq->ift_sds.ifsd_map != NULL)
1679 		map = txq->ift_sds.ifsd_map[i];
1680 	if (map != NULL) {
1681 		bus_dmamap_unload(txq->ift_desc_tag, map);
1682 		bus_dmamap_destroy(txq->ift_desc_tag, map);
1683 		txq->ift_sds.ifsd_map[i] = NULL;
1684 	}
1685 }
1686 
1687 static void
1688 iflib_txq_destroy(iflib_txq_t txq)
1689 {
1690 	if_ctx_t ctx = txq->ift_ctx;
1691 
1692 	for (int i = 0; i < txq->ift_size; i++)
1693 		iflib_txsd_destroy(ctx, txq, i);
1694 	if (txq->ift_sds.ifsd_map != NULL) {
1695 		free(txq->ift_sds.ifsd_map, M_IFLIB);
1696 		txq->ift_sds.ifsd_map = NULL;
1697 	}
1698 	if (txq->ift_sds.ifsd_m != NULL) {
1699 		free(txq->ift_sds.ifsd_m, M_IFLIB);
1700 		txq->ift_sds.ifsd_m = NULL;
1701 	}
1702 	if (txq->ift_sds.ifsd_flags != NULL) {
1703 		free(txq->ift_sds.ifsd_flags, M_IFLIB);
1704 		txq->ift_sds.ifsd_flags = NULL;
1705 	}
1706 	if (txq->ift_desc_tag != NULL) {
1707 		bus_dma_tag_destroy(txq->ift_desc_tag);
1708 		txq->ift_desc_tag = NULL;
1709 	}
1710 	if (txq->ift_tso_desc_tag != NULL) {
1711 		bus_dma_tag_destroy(txq->ift_tso_desc_tag);
1712 		txq->ift_tso_desc_tag = NULL;
1713 	}
1714 }
1715 
1716 static void
1717 iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i)
1718 {
1719 	struct mbuf **mp;
1720 
1721 	mp = &txq->ift_sds.ifsd_m[i];
1722 	if (*mp == NULL)
1723 		return;
1724 
1725 	if (txq->ift_sds.ifsd_map != NULL) {
1726 		bus_dmamap_sync(txq->ift_desc_tag,
1727 				txq->ift_sds.ifsd_map[i],
1728 				BUS_DMASYNC_POSTWRITE);
1729 		bus_dmamap_unload(txq->ift_desc_tag,
1730 				  txq->ift_sds.ifsd_map[i]);
1731 	}
1732 	m_free(*mp);
1733 	DBG_COUNTER_INC(tx_frees);
1734 	*mp = NULL;
1735 }
1736 
1737 static int
1738 iflib_txq_setup(iflib_txq_t txq)
1739 {
1740 	if_ctx_t ctx = txq->ift_ctx;
1741 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1742 	iflib_dma_info_t di;
1743 	int i;
1744 
1745 	/* Set number of descriptors available */
1746 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
1747 	/* XXX make configurable */
1748 	txq->ift_update_freq = IFLIB_DEFAULT_TX_UPDATE_FREQ;
1749 
1750 	/* Reset indices */
1751 	txq->ift_cidx_processed = 0;
1752 	txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0;
1753 	txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset];
1754 
1755 	for (i = 0, di = txq->ift_ifdi; i < ctx->ifc_nhwtxqs; i++, di++)
1756 		bzero((void *)di->idi_vaddr, di->idi_size);
1757 
1758 	IFDI_TXQ_SETUP(ctx, txq->ift_id);
1759 	for (i = 0, di = txq->ift_ifdi; i < ctx->ifc_nhwtxqs; i++, di++)
1760 		bus_dmamap_sync(di->idi_tag, di->idi_map,
1761 						BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1762 	return (0);
1763 }
1764 
1765 /*********************************************************************
1766  *
1767  *  Allocate memory for rx_buffer structures. Since we use one
1768  *  rx_buffer per received packet, the maximum number of rx_buffer's
1769  *  that we'll need is equal to the number of receive descriptors
1770  *  that we've allocated.
1771  *
1772  **********************************************************************/
1773 static int
1774 iflib_rxsd_alloc(iflib_rxq_t rxq)
1775 {
1776 	if_ctx_t ctx = rxq->ifr_ctx;
1777 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1778 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1779 	device_t dev = ctx->ifc_dev;
1780 	iflib_fl_t fl;
1781 	int			err;
1782 
1783 	MPASS(scctx->isc_nrxd[0] > 0);
1784 	MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0);
1785 
1786 	fl = rxq->ifr_fl;
1787 	for (int i = 0; i <  rxq->ifr_nfl; i++, fl++) {
1788 		fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */
1789 		err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1790 					 1, 0,			/* alignment, bounds */
1791 					 BUS_SPACE_MAXADDR,	/* lowaddr */
1792 					 BUS_SPACE_MAXADDR,	/* highaddr */
1793 					 NULL, NULL,		/* filter, filterarg */
1794 					 sctx->isc_rx_maxsize,	/* maxsize */
1795 					 sctx->isc_rx_nsegments,	/* nsegments */
1796 					 sctx->isc_rx_maxsegsize,	/* maxsegsize */
1797 					 0,			/* flags */
1798 					 NULL,			/* lockfunc */
1799 					 NULL,			/* lockarg */
1800 					 &fl->ifl_desc_tag);
1801 		if (err) {
1802 			device_printf(dev, "%s: bus_dma_tag_create failed %d\n",
1803 				__func__, err);
1804 			goto fail;
1805 		}
1806 		if (!(fl->ifl_sds.ifsd_flags =
1807 		      (uint8_t *) malloc(sizeof(uint8_t) *
1808 					 scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1809 			device_printf(dev, "Unable to allocate tx_buffer memory\n");
1810 			err = ENOMEM;
1811 			goto fail;
1812 		}
1813 		if (!(fl->ifl_sds.ifsd_m =
1814 		      (struct mbuf **) malloc(sizeof(struct mbuf *) *
1815 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1816 			device_printf(dev, "Unable to allocate tx_buffer memory\n");
1817 			err = ENOMEM;
1818 			goto fail;
1819 		}
1820 		if (!(fl->ifl_sds.ifsd_cl =
1821 		      (caddr_t *) malloc(sizeof(caddr_t) *
1822 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1823 			device_printf(dev, "Unable to allocate tx_buffer memory\n");
1824 			err = ENOMEM;
1825 			goto fail;
1826 		}
1827 
1828 		/* Create the descriptor buffer dma maps */
1829 #if defined(ACPI_DMAR) || (! (defined(__i386__) || defined(__amd64__)))
1830 		if ((ctx->ifc_flags & IFC_DMAR) == 0)
1831 			continue;
1832 
1833 		if (!(fl->ifl_sds.ifsd_map =
1834 		      (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1835 			device_printf(dev, "Unable to allocate tx_buffer map memory\n");
1836 			err = ENOMEM;
1837 			goto fail;
1838 		}
1839 
1840 		for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++) {
1841 			err = bus_dmamap_create(fl->ifl_desc_tag, 0, &fl->ifl_sds.ifsd_map[i]);
1842 			if (err != 0) {
1843 				device_printf(dev, "Unable to create RX buffer DMA map\n");
1844 				goto fail;
1845 			}
1846 		}
1847 #endif
1848 	}
1849 	return (0);
1850 
1851 fail:
1852 	iflib_rx_structures_free(ctx);
1853 	return (err);
1854 }
1855 
1856 
1857 /*
1858  * Internal service routines
1859  */
1860 
1861 struct rxq_refill_cb_arg {
1862 	int               error;
1863 	bus_dma_segment_t seg;
1864 	int               nseg;
1865 };
1866 
1867 static void
1868 _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
1869 {
1870 	struct rxq_refill_cb_arg *cb_arg = arg;
1871 
1872 	cb_arg->error = error;
1873 	cb_arg->seg = segs[0];
1874 	cb_arg->nseg = nseg;
1875 }
1876 
1877 
1878 #ifdef ACPI_DMAR
1879 #define IS_DMAR(ctx) (ctx->ifc_flags & IFC_DMAR)
1880 #else
1881 #define IS_DMAR(ctx) (0)
1882 #endif
1883 
1884 /**
1885  *	rxq_refill - refill an rxq  free-buffer list
1886  *	@ctx: the iflib context
1887  *	@rxq: the free-list to refill
1888  *	@n: the number of new buffers to allocate
1889  *
1890  *	(Re)populate an rxq free-buffer list with up to @n new packet buffers.
1891  *	The caller must assure that @n does not exceed the queue's capacity.
1892  */
1893 static void
1894 _iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
1895 {
1896 	struct mbuf *m;
1897 	int idx, frag_idx = fl->ifl_fragidx;
1898         int pidx = fl->ifl_pidx;
1899 	caddr_t cl, *sd_cl;
1900 	struct mbuf **sd_m;
1901 	uint8_t *sd_flags;
1902 	struct if_rxd_update iru;
1903 	bus_dmamap_t *sd_map;
1904 	int n, i = 0;
1905 	uint64_t bus_addr;
1906 	int err;
1907 	qidx_t credits;
1908 
1909 	sd_m = fl->ifl_sds.ifsd_m;
1910 	sd_map = fl->ifl_sds.ifsd_map;
1911 	sd_cl = fl->ifl_sds.ifsd_cl;
1912 	sd_flags = fl->ifl_sds.ifsd_flags;
1913 	idx = pidx;
1914 	credits = fl->ifl_credits;
1915 
1916 	n  = count;
1917 	MPASS(n > 0);
1918 	MPASS(credits + n <= fl->ifl_size);
1919 
1920 	if (pidx < fl->ifl_cidx)
1921 		MPASS(pidx + n <= fl->ifl_cidx);
1922 	if (pidx == fl->ifl_cidx && (credits < fl->ifl_size))
1923 		MPASS(fl->ifl_gen == 0);
1924 	if (pidx > fl->ifl_cidx)
1925 		MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx);
1926 
1927 	DBG_COUNTER_INC(fl_refills);
1928 	if (n > 8)
1929 		DBG_COUNTER_INC(fl_refills_large);
1930 	iru_init(&iru, fl->ifl_rxq, fl->ifl_id);
1931 	while (n--) {
1932 		/*
1933 		 * We allocate an uninitialized mbuf + cluster, mbuf is
1934 		 * initialized after rx.
1935 		 *
1936 		 * If the cluster is still set then we know a minimum sized packet was received
1937 		 */
1938 		bit_ffc_at(fl->ifl_rx_bitmap, frag_idx, fl->ifl_size,  &frag_idx);
1939 		if ((frag_idx < 0) || (frag_idx >= fl->ifl_size))
1940                 	bit_ffc(fl->ifl_rx_bitmap, fl->ifl_size, &frag_idx);
1941 		if ((cl = sd_cl[frag_idx]) == NULL) {
1942                        if ((cl = sd_cl[frag_idx] = m_cljget(NULL, M_NOWAIT, fl->ifl_buf_size)) == NULL)
1943 				break;
1944 #if MEMORY_LOGGING
1945 			fl->ifl_cl_enqueued++;
1946 #endif
1947 		}
1948 		if ((m = m_gethdr(M_NOWAIT, MT_NOINIT)) == NULL) {
1949 			break;
1950 		}
1951 #if MEMORY_LOGGING
1952 		fl->ifl_m_enqueued++;
1953 #endif
1954 
1955 		DBG_COUNTER_INC(rx_allocs);
1956 #if defined(__i386__) || defined(__amd64__)
1957 		if (!IS_DMAR(ctx)) {
1958 			bus_addr = pmap_kextract((vm_offset_t)cl);
1959 		} else
1960 #endif
1961 		{
1962 			struct rxq_refill_cb_arg cb_arg;
1963 
1964 			cb_arg.error = 0;
1965 			MPASS(sd_map != NULL);
1966 			MPASS(sd_map[frag_idx] != NULL);
1967 			err = bus_dmamap_load(fl->ifl_desc_tag, sd_map[frag_idx],
1968 		         cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg, 0);
1969 			bus_dmamap_sync(fl->ifl_desc_tag, sd_map[frag_idx],
1970 					BUS_DMASYNC_PREREAD);
1971 
1972 			if (err != 0 || cb_arg.error) {
1973 				/*
1974 				 * !zone_pack ?
1975 				 */
1976 				if (fl->ifl_zone == zone_pack)
1977 					uma_zfree(fl->ifl_zone, cl);
1978 				m_free(m);
1979 				n = 0;
1980 				goto done;
1981 			}
1982 			bus_addr = cb_arg.seg.ds_addr;
1983 		}
1984                 bit_set(fl->ifl_rx_bitmap, frag_idx);
1985 		sd_flags[frag_idx] |= RX_SW_DESC_INUSE;
1986 
1987 		MPASS(sd_m[frag_idx] == NULL);
1988 		sd_cl[frag_idx] = cl;
1989 		sd_m[frag_idx] = m;
1990 		fl->ifl_rxd_idxs[i] = frag_idx;
1991 		fl->ifl_bus_addrs[i] = bus_addr;
1992 		fl->ifl_vm_addrs[i] = cl;
1993 		credits++;
1994 		i++;
1995 		MPASS(credits <= fl->ifl_size);
1996 		if (++idx == fl->ifl_size) {
1997 			fl->ifl_gen = 1;
1998 			idx = 0;
1999 		}
2000 		if (n == 0 || i == IFLIB_MAX_RX_REFRESH) {
2001 			iru.iru_pidx = pidx;
2002 			iru.iru_count = i;
2003 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2004 			i = 0;
2005 			pidx = idx;
2006 			fl->ifl_pidx = idx;
2007 			fl->ifl_credits = credits;
2008 		}
2009 
2010 	}
2011 done:
2012 	if (i) {
2013 		iru.iru_pidx = pidx;
2014 		iru.iru_count = i;
2015 		ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2016 		fl->ifl_pidx = idx;
2017 		fl->ifl_credits = credits;
2018 	}
2019 	DBG_COUNTER_INC(rxd_flush);
2020 	if (fl->ifl_pidx == 0)
2021 		pidx = fl->ifl_size - 1;
2022 	else
2023 		pidx = fl->ifl_pidx - 1;
2024 
2025 	if (sd_map)
2026 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2027 				BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2028 	ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id, fl->ifl_id, pidx);
2029 	fl->ifl_fragidx = frag_idx;
2030 }
2031 
2032 static __inline void
2033 __iflib_fl_refill_lt(if_ctx_t ctx, iflib_fl_t fl, int max)
2034 {
2035 	/* we avoid allowing pidx to catch up with cidx as it confuses ixl */
2036 	int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1;
2037 #ifdef INVARIANTS
2038 	int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1;
2039 #endif
2040 
2041 	MPASS(fl->ifl_credits <= fl->ifl_size);
2042 	MPASS(reclaimable == delta);
2043 
2044 	if (reclaimable > 0)
2045 		_iflib_fl_refill(ctx, fl, min(max, reclaimable));
2046 }
2047 
2048 static void
2049 iflib_fl_bufs_free(iflib_fl_t fl)
2050 {
2051 	iflib_dma_info_t idi = fl->ifl_ifdi;
2052 	uint32_t i;
2053 
2054 	for (i = 0; i < fl->ifl_size; i++) {
2055 		struct mbuf **sd_m = &fl->ifl_sds.ifsd_m[i];
2056 		uint8_t *sd_flags = &fl->ifl_sds.ifsd_flags[i];
2057 		caddr_t *sd_cl = &fl->ifl_sds.ifsd_cl[i];
2058 
2059 		if (*sd_flags & RX_SW_DESC_INUSE) {
2060 			if (fl->ifl_sds.ifsd_map != NULL) {
2061 				bus_dmamap_t sd_map = fl->ifl_sds.ifsd_map[i];
2062 				bus_dmamap_unload(fl->ifl_desc_tag, sd_map);
2063 				if (fl->ifl_rxq->ifr_ctx->ifc_in_detach)
2064 					bus_dmamap_destroy(fl->ifl_desc_tag, sd_map);
2065 			}
2066 			if (*sd_m != NULL) {
2067 				m_init(*sd_m, M_NOWAIT, MT_DATA, 0);
2068 				uma_zfree(zone_mbuf, *sd_m);
2069 			}
2070 			if (*sd_cl != NULL)
2071 				uma_zfree(fl->ifl_zone, *sd_cl);
2072 			*sd_flags = 0;
2073 		} else {
2074 			MPASS(*sd_cl == NULL);
2075 			MPASS(*sd_m == NULL);
2076 		}
2077 #if MEMORY_LOGGING
2078 		fl->ifl_m_dequeued++;
2079 		fl->ifl_cl_dequeued++;
2080 #endif
2081 		*sd_cl = NULL;
2082 		*sd_m = NULL;
2083 	}
2084 #ifdef INVARIANTS
2085 	for (i = 0; i < fl->ifl_size; i++) {
2086 		MPASS(fl->ifl_sds.ifsd_flags[i] == 0);
2087 		MPASS(fl->ifl_sds.ifsd_cl[i] == NULL);
2088 		MPASS(fl->ifl_sds.ifsd_m[i] == NULL);
2089 	}
2090 #endif
2091 	/*
2092 	 * Reset free list values
2093 	 */
2094 	fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = fl->ifl_fragidx = 0;
2095 	bzero(idi->idi_vaddr, idi->idi_size);
2096 }
2097 
2098 /*********************************************************************
2099  *
2100  *  Initialize a receive ring and its buffers.
2101  *
2102  **********************************************************************/
2103 static int
2104 iflib_fl_setup(iflib_fl_t fl)
2105 {
2106 	iflib_rxq_t rxq = fl->ifl_rxq;
2107 	if_ctx_t ctx = rxq->ifr_ctx;
2108 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2109 
2110 	bit_nclear(fl->ifl_rx_bitmap, 0, fl->ifl_size - 1);
2111 	/*
2112 	** Free current RX buffer structs and their mbufs
2113 	*/
2114 	iflib_fl_bufs_free(fl);
2115 	/* Now replenish the mbufs */
2116 	MPASS(fl->ifl_credits == 0);
2117 	/*
2118 	 * XXX don't set the max_frame_size to larger
2119 	 * than the hardware can handle
2120 	 */
2121 	if (sctx->isc_max_frame_size <= 2048)
2122 		fl->ifl_buf_size = MCLBYTES;
2123 #ifndef CONTIGMALLOC_WORKS
2124 	else
2125 		fl->ifl_buf_size = MJUMPAGESIZE;
2126 #else
2127 	else if (sctx->isc_max_frame_size <= 4096)
2128 		fl->ifl_buf_size = MJUMPAGESIZE;
2129 	else if (sctx->isc_max_frame_size <= 9216)
2130 		fl->ifl_buf_size = MJUM9BYTES;
2131 	else
2132 		fl->ifl_buf_size = MJUM16BYTES;
2133 #endif
2134 	if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size)
2135 		ctx->ifc_max_fl_buf_size = fl->ifl_buf_size;
2136 	fl->ifl_cltype = m_gettype(fl->ifl_buf_size);
2137 	fl->ifl_zone = m_getzone(fl->ifl_buf_size);
2138 
2139 
2140 	/* avoid pre-allocating zillions of clusters to an idle card
2141 	 * potentially speeding up attach
2142 	 */
2143 	_iflib_fl_refill(ctx, fl, min(128, fl->ifl_size));
2144 	MPASS(min(128, fl->ifl_size) == fl->ifl_credits);
2145 	if (min(128, fl->ifl_size) != fl->ifl_credits)
2146 		return (ENOBUFS);
2147 	/*
2148 	 * handle failure
2149 	 */
2150 	MPASS(rxq != NULL);
2151 	MPASS(fl->ifl_ifdi != NULL);
2152 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2153 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2154 	return (0);
2155 }
2156 
2157 /*********************************************************************
2158  *
2159  *  Free receive ring data structures
2160  *
2161  **********************************************************************/
2162 static void
2163 iflib_rx_sds_free(iflib_rxq_t rxq)
2164 {
2165 	iflib_fl_t fl;
2166 	int i;
2167 
2168 	if (rxq->ifr_fl != NULL) {
2169 		for (i = 0; i < rxq->ifr_nfl; i++) {
2170 			fl = &rxq->ifr_fl[i];
2171 			if (fl->ifl_desc_tag != NULL) {
2172 				bus_dma_tag_destroy(fl->ifl_desc_tag);
2173 				fl->ifl_desc_tag = NULL;
2174 			}
2175 			free(fl->ifl_sds.ifsd_m, M_IFLIB);
2176 			free(fl->ifl_sds.ifsd_cl, M_IFLIB);
2177 			/* XXX destroy maps first */
2178 			free(fl->ifl_sds.ifsd_map, M_IFLIB);
2179 			fl->ifl_sds.ifsd_m = NULL;
2180 			fl->ifl_sds.ifsd_cl = NULL;
2181 			fl->ifl_sds.ifsd_map = NULL;
2182 		}
2183 		free(rxq->ifr_fl, M_IFLIB);
2184 		rxq->ifr_fl = NULL;
2185 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
2186 	}
2187 }
2188 
2189 /*
2190  * MI independent logic
2191  *
2192  */
2193 static void
2194 iflib_timer(void *arg)
2195 {
2196 	iflib_txq_t txq = arg;
2197 	if_ctx_t ctx = txq->ift_ctx;
2198 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2199 
2200 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
2201 		return;
2202 	/*
2203 	** Check on the state of the TX queue(s), this
2204 	** can be done without the lock because its RO
2205 	** and the HUNG state will be static if set.
2206 	*/
2207 	IFDI_TIMER(ctx, txq->ift_id);
2208 	if ((txq->ift_qstatus == IFLIB_QUEUE_HUNG) &&
2209 	    ((txq->ift_cleaned_prev == txq->ift_cleaned) ||
2210 	     (sctx->isc_pause_frames == 0)))
2211 		goto hung;
2212 
2213 	if (ifmp_ring_is_stalled(txq->ift_br))
2214 		txq->ift_qstatus = IFLIB_QUEUE_HUNG;
2215 	txq->ift_cleaned_prev = txq->ift_cleaned;
2216 	/* handle any laggards */
2217 	if (txq->ift_db_pending)
2218 		GROUPTASK_ENQUEUE(&txq->ift_task);
2219 
2220 	sctx->isc_pause_frames = 0;
2221 	if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)
2222 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq, txq->ift_timer.c_cpu);
2223 	return;
2224  hung:
2225 	device_printf(ctx->ifc_dev,  "TX(%d) desc avail = %d, pidx = %d\n",
2226 				  txq->ift_id, TXQ_AVAIL(txq), txq->ift_pidx);
2227 	STATE_LOCK(ctx);
2228 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2229 	ctx->ifc_flags |= (IFC_DO_WATCHDOG|IFC_DO_RESET);
2230 	iflib_admin_intr_deferred(ctx);
2231 	STATE_UNLOCK(ctx);
2232 }
2233 
2234 static void
2235 iflib_init_locked(if_ctx_t ctx)
2236 {
2237 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2238 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2239 	if_t ifp = ctx->ifc_ifp;
2240 	iflib_fl_t fl;
2241 	iflib_txq_t txq;
2242 	iflib_rxq_t rxq;
2243 	int i, j, tx_ip_csum_flags, tx_ip6_csum_flags;
2244 
2245 
2246 	if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2247 	IFDI_INTR_DISABLE(ctx);
2248 
2249 	tx_ip_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP);
2250 	tx_ip6_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | CSUM_IP6_SCTP);
2251 	/* Set hardware offload abilities */
2252 	if_clearhwassist(ifp);
2253 	if (if_getcapenable(ifp) & IFCAP_TXCSUM)
2254 		if_sethwassistbits(ifp, tx_ip_csum_flags, 0);
2255 	if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6)
2256 		if_sethwassistbits(ifp,  tx_ip6_csum_flags, 0);
2257 	if (if_getcapenable(ifp) & IFCAP_TSO4)
2258 		if_sethwassistbits(ifp, CSUM_IP_TSO, 0);
2259 	if (if_getcapenable(ifp) & IFCAP_TSO6)
2260 		if_sethwassistbits(ifp, CSUM_IP6_TSO, 0);
2261 
2262 	for (i = 0, txq = ctx->ifc_txqs; i < sctx->isc_ntxqsets; i++, txq++) {
2263 		CALLOUT_LOCK(txq);
2264 		callout_stop(&txq->ift_timer);
2265 		CALLOUT_UNLOCK(txq);
2266 		iflib_netmap_txq_init(ctx, txq);
2267 	}
2268 #ifdef INVARIANTS
2269 	i = if_getdrvflags(ifp);
2270 #endif
2271 	IFDI_INIT(ctx);
2272 	MPASS(if_getdrvflags(ifp) == i);
2273 	for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
2274 		/* XXX this should really be done on a per-queue basis */
2275 		if (if_getcapenable(ifp) & IFCAP_NETMAP) {
2276 			MPASS(rxq->ifr_id == i);
2277 			iflib_netmap_rxq_init(ctx, rxq);
2278 			continue;
2279 		}
2280 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
2281 			if (iflib_fl_setup(fl)) {
2282 				device_printf(ctx->ifc_dev, "freelist setup failed - check cluster settings\n");
2283 				goto done;
2284 			}
2285 		}
2286 	}
2287 done:
2288 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
2289 	IFDI_INTR_ENABLE(ctx);
2290 	txq = ctx->ifc_txqs;
2291 	for (i = 0; i < sctx->isc_ntxqsets; i++, txq++)
2292 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq,
2293 			txq->ift_timer.c_cpu);
2294 }
2295 
2296 static int
2297 iflib_media_change(if_t ifp)
2298 {
2299 	if_ctx_t ctx = if_getsoftc(ifp);
2300 	int err;
2301 
2302 	CTX_LOCK(ctx);
2303 	if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0)
2304 		iflib_init_locked(ctx);
2305 	CTX_UNLOCK(ctx);
2306 	return (err);
2307 }
2308 
2309 static void
2310 iflib_media_status(if_t ifp, struct ifmediareq *ifmr)
2311 {
2312 	if_ctx_t ctx = if_getsoftc(ifp);
2313 
2314 	CTX_LOCK(ctx);
2315 	IFDI_UPDATE_ADMIN_STATUS(ctx);
2316 	IFDI_MEDIA_STATUS(ctx, ifmr);
2317 	CTX_UNLOCK(ctx);
2318 }
2319 
2320 void
2321 iflib_stop(if_ctx_t ctx)
2322 {
2323 	iflib_txq_t txq = ctx->ifc_txqs;
2324 	iflib_rxq_t rxq = ctx->ifc_rxqs;
2325 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2326 	iflib_dma_info_t di;
2327 	iflib_fl_t fl;
2328 	int i, j;
2329 
2330 	/* Tell the stack that the interface is no longer active */
2331 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2332 
2333 	IFDI_INTR_DISABLE(ctx);
2334 	DELAY(1000);
2335 	IFDI_STOP(ctx);
2336 	DELAY(1000);
2337 
2338 	iflib_debug_reset();
2339 	/* Wait for current tx queue users to exit to disarm watchdog timer. */
2340 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) {
2341 		/* make sure all transmitters have completed before proceeding XXX */
2342 
2343 		CALLOUT_LOCK(txq);
2344 		callout_stop(&txq->ift_timer);
2345 		CALLOUT_UNLOCK(txq);
2346 
2347 		/* clean any enqueued buffers */
2348 		iflib_ifmp_purge(txq);
2349 		/* Free any existing tx buffers. */
2350 		for (j = 0; j < txq->ift_size; j++) {
2351 			iflib_txsd_free(ctx, txq, j);
2352 		}
2353 		txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0;
2354 		txq->ift_in_use = txq->ift_gen = txq->ift_cidx = txq->ift_pidx = txq->ift_no_desc_avail = 0;
2355 		txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0;
2356 		txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0;
2357 		txq->ift_pullups = 0;
2358 		ifmp_ring_reset_stats(txq->ift_br);
2359 		for (j = 0, di = txq->ift_ifdi; j < ctx->ifc_nhwtxqs; j++, di++)
2360 			bzero((void *)di->idi_vaddr, di->idi_size);
2361 	}
2362 	for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) {
2363 		/* make sure all transmitters have completed before proceeding XXX */
2364 
2365 		for (j = 0, di = rxq->ifr_ifdi; j < rxq->ifr_nfl; j++, di++)
2366 			bzero((void *)di->idi_vaddr, di->idi_size);
2367 		/* also resets the free lists pidx/cidx */
2368 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
2369 			iflib_fl_bufs_free(fl);
2370 	}
2371 }
2372 
2373 static inline caddr_t
2374 calc_next_rxd(iflib_fl_t fl, int cidx)
2375 {
2376 	qidx_t size;
2377 	int nrxd;
2378 	caddr_t start, end, cur, next;
2379 
2380 	nrxd = fl->ifl_size;
2381 	size = fl->ifl_rxd_size;
2382 	start = fl->ifl_ifdi->idi_vaddr;
2383 
2384 	if (__predict_false(size == 0))
2385 		return (start);
2386 	cur = start + size*cidx;
2387 	end = start + size*nrxd;
2388 	next = CACHE_PTR_NEXT(cur);
2389 	return (next < end ? next : start);
2390 }
2391 
2392 static inline void
2393 prefetch_pkts(iflib_fl_t fl, int cidx)
2394 {
2395 	int nextptr;
2396 	int nrxd = fl->ifl_size;
2397 	caddr_t next_rxd;
2398 
2399 
2400 	nextptr = (cidx + CACHE_PTR_INCREMENT) & (nrxd-1);
2401 	prefetch(&fl->ifl_sds.ifsd_m[nextptr]);
2402 	prefetch(&fl->ifl_sds.ifsd_cl[nextptr]);
2403 	next_rxd = calc_next_rxd(fl, cidx);
2404 	prefetch(next_rxd);
2405 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 1) & (nrxd-1)]);
2406 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 2) & (nrxd-1)]);
2407 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 3) & (nrxd-1)]);
2408 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 4) & (nrxd-1)]);
2409 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 1) & (nrxd-1)]);
2410 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 2) & (nrxd-1)]);
2411 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 3) & (nrxd-1)]);
2412 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 4) & (nrxd-1)]);
2413 }
2414 
2415 static void
2416 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, int unload, if_rxsd_t sd)
2417 {
2418 	int flid, cidx;
2419 	bus_dmamap_t map;
2420 	iflib_fl_t fl;
2421 	iflib_dma_info_t di;
2422 	int next;
2423 
2424 	map = NULL;
2425 	flid = irf->irf_flid;
2426 	cidx = irf->irf_idx;
2427 	fl = &rxq->ifr_fl[flid];
2428 	sd->ifsd_fl = fl;
2429 	sd->ifsd_cidx = cidx;
2430 	sd->ifsd_m = &fl->ifl_sds.ifsd_m[cidx];
2431 	sd->ifsd_cl = &fl->ifl_sds.ifsd_cl[cidx];
2432 	fl->ifl_credits--;
2433 #if MEMORY_LOGGING
2434 	fl->ifl_m_dequeued++;
2435 #endif
2436 	if (rxq->ifr_ctx->ifc_flags & IFC_PREFETCH)
2437 		prefetch_pkts(fl, cidx);
2438 	if (fl->ifl_sds.ifsd_map != NULL) {
2439 		next = (cidx + CACHE_PTR_INCREMENT) & (fl->ifl_size-1);
2440 		prefetch(&fl->ifl_sds.ifsd_map[next]);
2441 		map = fl->ifl_sds.ifsd_map[cidx];
2442 		di = fl->ifl_ifdi;
2443 		next = (cidx + CACHE_LINE_SIZE) & (fl->ifl_size-1);
2444 		prefetch(&fl->ifl_sds.ifsd_flags[next]);
2445 		bus_dmamap_sync(di->idi_tag, di->idi_map,
2446 				BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
2447 
2448 	/* not valid assert if bxe really does SGE from non-contiguous elements */
2449 		MPASS(fl->ifl_cidx == cidx);
2450 		if (unload)
2451 			bus_dmamap_unload(fl->ifl_desc_tag, map);
2452 	}
2453 	fl->ifl_cidx = (fl->ifl_cidx + 1) & (fl->ifl_size-1);
2454 	if (__predict_false(fl->ifl_cidx == 0))
2455 		fl->ifl_gen = 0;
2456 	if (map != NULL)
2457 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2458 			BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2459         bit_clear(fl->ifl_rx_bitmap, cidx);
2460 }
2461 
2462 static struct mbuf *
2463 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri, if_rxsd_t sd)
2464 {
2465 	int i, padlen , flags;
2466 	struct mbuf *m, *mh, *mt;
2467 	caddr_t cl;
2468 
2469 	i = 0;
2470 	mh = NULL;
2471 	do {
2472 		rxd_frag_to_sd(rxq, &ri->iri_frags[i], TRUE, sd);
2473 
2474 		MPASS(*sd->ifsd_cl != NULL);
2475 		MPASS(*sd->ifsd_m != NULL);
2476 
2477 		/* Don't include zero-length frags */
2478 		if (ri->iri_frags[i].irf_len == 0) {
2479 			/* XXX we can save the cluster here, but not the mbuf */
2480 			m_init(*sd->ifsd_m, M_NOWAIT, MT_DATA, 0);
2481 			m_free(*sd->ifsd_m);
2482 			*sd->ifsd_m = NULL;
2483 			continue;
2484 		}
2485 		m = *sd->ifsd_m;
2486 		*sd->ifsd_m = NULL;
2487 		if (mh == NULL) {
2488 			flags = M_PKTHDR|M_EXT;
2489 			mh = mt = m;
2490 			padlen = ri->iri_pad;
2491 		} else {
2492 			flags = M_EXT;
2493 			mt->m_next = m;
2494 			mt = m;
2495 			/* assuming padding is only on the first fragment */
2496 			padlen = 0;
2497 		}
2498 		cl = *sd->ifsd_cl;
2499 		*sd->ifsd_cl = NULL;
2500 
2501 		/* Can these two be made one ? */
2502 		m_init(m, M_NOWAIT, MT_DATA, flags);
2503 		m_cljset(m, cl, sd->ifsd_fl->ifl_cltype);
2504 		/*
2505 		 * These must follow m_init and m_cljset
2506 		 */
2507 		m->m_data += padlen;
2508 		ri->iri_len -= padlen;
2509 		m->m_len = ri->iri_frags[i].irf_len;
2510 	} while (++i < ri->iri_nfrags);
2511 
2512 	return (mh);
2513 }
2514 
2515 /*
2516  * Process one software descriptor
2517  */
2518 static struct mbuf *
2519 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri)
2520 {
2521 	struct if_rxsd sd;
2522 	struct mbuf *m;
2523 
2524 	/* should I merge this back in now that the two paths are basically duplicated? */
2525 	if (ri->iri_nfrags == 1 &&
2526 	    ri->iri_frags[0].irf_len <= MIN(IFLIB_RX_COPY_THRESH, MHLEN)) {
2527 		rxd_frag_to_sd(rxq, &ri->iri_frags[0], FALSE, &sd);
2528 		m = *sd.ifsd_m;
2529 		*sd.ifsd_m = NULL;
2530 		m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR);
2531 #ifndef __NO_STRICT_ALIGNMENT
2532 		if (!IP_ALIGNED(m))
2533 			m->m_data += 2;
2534 #endif
2535 		memcpy(m->m_data, *sd.ifsd_cl, ri->iri_len);
2536 		m->m_len = ri->iri_frags[0].irf_len;
2537        } else {
2538 		m = assemble_segments(rxq, ri, &sd);
2539 	}
2540 	m->m_pkthdr.len = ri->iri_len;
2541 	m->m_pkthdr.rcvif = ri->iri_ifp;
2542 	m->m_flags |= ri->iri_flags;
2543 	m->m_pkthdr.ether_vtag = ri->iri_vtag;
2544 	m->m_pkthdr.flowid = ri->iri_flowid;
2545 	M_HASHTYPE_SET(m, ri->iri_rsstype);
2546 	m->m_pkthdr.csum_flags = ri->iri_csum_flags;
2547 	m->m_pkthdr.csum_data = ri->iri_csum_data;
2548 	return (m);
2549 }
2550 
2551 #if defined(INET6) || defined(INET)
2552 static void
2553 iflib_get_ip_forwarding(struct lro_ctrl *lc, bool *v4, bool *v6)
2554 {
2555 	CURVNET_SET(lc->ifp->if_vnet);
2556 #if defined(INET6)
2557 	*v6 = VNET(ip6_forwarding);
2558 #endif
2559 #if defined(INET)
2560 	*v4 = VNET(ipforwarding);
2561 #endif
2562 	CURVNET_RESTORE();
2563 }
2564 
2565 /*
2566  * Returns true if it's possible this packet could be LROed.
2567  * if it returns false, it is guaranteed that tcp_lro_rx()
2568  * would not return zero.
2569  */
2570 static bool
2571 iflib_check_lro_possible(struct mbuf *m, bool v4_forwarding, bool v6_forwarding)
2572 {
2573 	struct ether_header *eh;
2574 	uint16_t eh_type;
2575 
2576 	eh = mtod(m, struct ether_header *);
2577 	eh_type = ntohs(eh->ether_type);
2578 	switch (eh_type) {
2579 #if defined(INET6)
2580 		case ETHERTYPE_IPV6:
2581 			return !v6_forwarding;
2582 #endif
2583 #if defined (INET)
2584 		case ETHERTYPE_IP:
2585 			return !v4_forwarding;
2586 #endif
2587 	}
2588 
2589 	return false;
2590 }
2591 #else
2592 static void
2593 iflib_get_ip_forwarding(struct lro_ctrl *lc __unused, bool *v4 __unused, bool *v6 __unused)
2594 {
2595 }
2596 #endif
2597 
2598 static bool
2599 iflib_rxeof(iflib_rxq_t rxq, qidx_t budget)
2600 {
2601 	if_ctx_t ctx = rxq->ifr_ctx;
2602 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2603 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2604 	int avail, i;
2605 	qidx_t *cidxp;
2606 	struct if_rxd_info ri;
2607 	int err, budget_left, rx_bytes, rx_pkts;
2608 	iflib_fl_t fl;
2609 	struct ifnet *ifp;
2610 	int lro_enabled;
2611 	bool v4_forwarding, v6_forwarding, lro_possible;
2612 
2613 	/*
2614 	 * XXX early demux data packets so that if_input processing only handles
2615 	 * acks in interrupt context
2616 	 */
2617 	struct mbuf *m, *mh, *mt, *mf;
2618 
2619 	lro_possible = v4_forwarding = v6_forwarding = false;
2620 	ifp = ctx->ifc_ifp;
2621 	mh = mt = NULL;
2622 	MPASS(budget > 0);
2623 	rx_pkts	= rx_bytes = 0;
2624 	if (sctx->isc_flags & IFLIB_HAS_RXCQ)
2625 		cidxp = &rxq->ifr_cq_cidx;
2626 	else
2627 		cidxp = &rxq->ifr_fl[0].ifl_cidx;
2628 	if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) {
2629 		for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2630 			__iflib_fl_refill_lt(ctx, fl, budget + 8);
2631 		DBG_COUNTER_INC(rx_unavail);
2632 		return (false);
2633 	}
2634 
2635 	for (budget_left = budget; (budget_left > 0) && (avail > 0); budget_left--, avail--) {
2636 		if (__predict_false(!CTX_ACTIVE(ctx))) {
2637 			DBG_COUNTER_INC(rx_ctx_inactive);
2638 			break;
2639 		}
2640 		/*
2641 		 * Reset client set fields to their default values
2642 		 */
2643 		rxd_info_zero(&ri);
2644 		ri.iri_qsidx = rxq->ifr_id;
2645 		ri.iri_cidx = *cidxp;
2646 		ri.iri_ifp = ifp;
2647 		ri.iri_frags = rxq->ifr_frags;
2648 		err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
2649 
2650 		if (err)
2651 			goto err;
2652 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
2653 			*cidxp = ri.iri_cidx;
2654 			/* Update our consumer index */
2655 			/* XXX NB: shurd - check if this is still safe */
2656 			while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0]) {
2657 				rxq->ifr_cq_cidx -= scctx->isc_nrxd[0];
2658 				rxq->ifr_cq_gen = 0;
2659 			}
2660 			/* was this only a completion queue message? */
2661 			if (__predict_false(ri.iri_nfrags == 0))
2662 				continue;
2663 		}
2664 		MPASS(ri.iri_nfrags != 0);
2665 		MPASS(ri.iri_len != 0);
2666 
2667 		/* will advance the cidx on the corresponding free lists */
2668 		m = iflib_rxd_pkt_get(rxq, &ri);
2669 		if (avail == 0 && budget_left)
2670 			avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left);
2671 
2672 		if (__predict_false(m == NULL)) {
2673 			DBG_COUNTER_INC(rx_mbuf_null);
2674 			continue;
2675 		}
2676 		/* imm_pkt: -- cxgb */
2677 		if (mh == NULL)
2678 			mh = mt = m;
2679 		else {
2680 			mt->m_nextpkt = m;
2681 			mt = m;
2682 		}
2683 	}
2684 	/* make sure that we can refill faster than drain */
2685 	for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2686 		__iflib_fl_refill_lt(ctx, fl, budget + 8);
2687 
2688 	lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO);
2689 	if (lro_enabled)
2690 		iflib_get_ip_forwarding(&rxq->ifr_lc, &v4_forwarding, &v6_forwarding);
2691 	mt = mf = NULL;
2692 	while (mh != NULL) {
2693 		m = mh;
2694 		mh = mh->m_nextpkt;
2695 		m->m_nextpkt = NULL;
2696 #ifndef __NO_STRICT_ALIGNMENT
2697 		if (!IP_ALIGNED(m) && (m = iflib_fixup_rx(m)) == NULL)
2698 			continue;
2699 #endif
2700 		rx_bytes += m->m_pkthdr.len;
2701 		rx_pkts++;
2702 #if defined(INET6) || defined(INET)
2703 		if (lro_enabled) {
2704 			if (!lro_possible) {
2705 				lro_possible = iflib_check_lro_possible(m, v4_forwarding, v6_forwarding);
2706 				if (lro_possible && mf != NULL) {
2707 					ifp->if_input(ifp, mf);
2708 					DBG_COUNTER_INC(rx_if_input);
2709 					mt = mf = NULL;
2710 				}
2711 			}
2712 			if ((m->m_pkthdr.csum_flags & (CSUM_L4_CALC|CSUM_L4_VALID)) ==
2713 			    (CSUM_L4_CALC|CSUM_L4_VALID)) {
2714 				if (lro_possible && tcp_lro_rx(&rxq->ifr_lc, m, 0) == 0)
2715 					continue;
2716 			}
2717 		}
2718 #endif
2719 		if (lro_possible) {
2720 			ifp->if_input(ifp, m);
2721 			DBG_COUNTER_INC(rx_if_input);
2722 			continue;
2723 		}
2724 
2725 		if (mf == NULL)
2726 			mf = m;
2727 		if (mt != NULL)
2728 			mt->m_nextpkt = m;
2729 		mt = m;
2730 	}
2731 	if (mf != NULL) {
2732 		ifp->if_input(ifp, mf);
2733 		DBG_COUNTER_INC(rx_if_input);
2734 	}
2735 
2736 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
2737 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
2738 
2739 	/*
2740 	 * Flush any outstanding LRO work
2741 	 */
2742 #if defined(INET6) || defined(INET)
2743 	tcp_lro_flush_all(&rxq->ifr_lc);
2744 #endif
2745 	if (avail)
2746 		return true;
2747 	return (iflib_rxd_avail(ctx, rxq, *cidxp, 1));
2748 err:
2749 	STATE_LOCK(ctx);
2750 	ctx->ifc_flags |= IFC_DO_RESET;
2751 	iflib_admin_intr_deferred(ctx);
2752 	STATE_UNLOCK(ctx);
2753 	return (false);
2754 }
2755 
2756 #define TXD_NOTIFY_COUNT(txq) (((txq)->ift_size / (txq)->ift_update_freq)-1)
2757 static inline qidx_t
2758 txq_max_db_deferred(iflib_txq_t txq, qidx_t in_use)
2759 {
2760 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2761 	qidx_t minthresh = txq->ift_size / 8;
2762 	if (in_use > 4*minthresh)
2763 		return (notify_count);
2764 	if (in_use > 2*minthresh)
2765 		return (notify_count >> 1);
2766 	if (in_use > minthresh)
2767 		return (notify_count >> 3);
2768 	return (0);
2769 }
2770 
2771 static inline qidx_t
2772 txq_max_rs_deferred(iflib_txq_t txq)
2773 {
2774 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2775 	qidx_t minthresh = txq->ift_size / 8;
2776 	if (txq->ift_in_use > 4*minthresh)
2777 		return (notify_count);
2778 	if (txq->ift_in_use > 2*minthresh)
2779 		return (notify_count >> 1);
2780 	if (txq->ift_in_use > minthresh)
2781 		return (notify_count >> 2);
2782 	return (2);
2783 }
2784 
2785 #define M_CSUM_FLAGS(m) ((m)->m_pkthdr.csum_flags)
2786 #define M_HAS_VLANTAG(m) (m->m_flags & M_VLANTAG)
2787 
2788 #define TXQ_MAX_DB_DEFERRED(txq, in_use) txq_max_db_deferred((txq), (in_use))
2789 #define TXQ_MAX_RS_DEFERRED(txq) txq_max_rs_deferred(txq)
2790 #define TXQ_MAX_DB_CONSUMED(size) (size >> 4)
2791 
2792 /* forward compatibility for cxgb */
2793 #define FIRST_QSET(ctx) 0
2794 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets)
2795 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets)
2796 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx))
2797 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments))
2798 
2799 /* XXX we should be setting this to something other than zero */
2800 #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh)
2801 #define MAX_TX_DESC(ctx) ((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max)
2802 
2803 static inline bool
2804 iflib_txd_db_check(if_ctx_t ctx, iflib_txq_t txq, int ring, qidx_t in_use)
2805 {
2806 	qidx_t dbval, max;
2807 	bool rang;
2808 
2809 	rang = false;
2810 	max = TXQ_MAX_DB_DEFERRED(txq, in_use);
2811 	if (ring || txq->ift_db_pending >= max) {
2812 		dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx;
2813 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval);
2814 		txq->ift_db_pending = txq->ift_npending = 0;
2815 		rang = true;
2816 	}
2817 	return (rang);
2818 }
2819 
2820 #ifdef PKT_DEBUG
2821 static void
2822 print_pkt(if_pkt_info_t pi)
2823 {
2824 	printf("pi len:  %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n",
2825 	       pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx);
2826 	printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n",
2827 	       pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag);
2828 	printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n",
2829 	       pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto);
2830 }
2831 #endif
2832 
2833 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO)
2834 #define IS_TX_OFFLOAD4(pi) ((pi)->ipi_csum_flags & (CSUM_IP_TCP | CSUM_IP_TSO))
2835 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO)
2836 #define IS_TX_OFFLOAD6(pi) ((pi)->ipi_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_TSO))
2837 
2838 static int
2839 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp)
2840 {
2841 	if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx;
2842 	struct ether_vlan_header *eh;
2843 	struct mbuf *m, *n;
2844 
2845 	n = m = *mp;
2846 	if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) &&
2847 	    M_WRITABLE(m) == 0) {
2848 		if ((m = m_dup(m, M_NOWAIT)) == NULL) {
2849 			return (ENOMEM);
2850 		} else {
2851 			m_freem(*mp);
2852 			n = *mp = m;
2853 		}
2854 	}
2855 
2856 	/*
2857 	 * Determine where frame payload starts.
2858 	 * Jump over vlan headers if already present,
2859 	 * helpful for QinQ too.
2860 	 */
2861 	if (__predict_false(m->m_len < sizeof(*eh))) {
2862 		txq->ift_pullups++;
2863 		if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL))
2864 			return (ENOMEM);
2865 	}
2866 	eh = mtod(m, struct ether_vlan_header *);
2867 	if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
2868 		pi->ipi_etype = ntohs(eh->evl_proto);
2869 		pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
2870 	} else {
2871 		pi->ipi_etype = ntohs(eh->evl_encap_proto);
2872 		pi->ipi_ehdrlen = ETHER_HDR_LEN;
2873 	}
2874 
2875 	switch (pi->ipi_etype) {
2876 #ifdef INET
2877 	case ETHERTYPE_IP:
2878 	{
2879 		struct ip *ip = NULL;
2880 		struct tcphdr *th = NULL;
2881 		int minthlen;
2882 
2883 		minthlen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip) + sizeof(*th));
2884 		if (__predict_false(m->m_len < minthlen)) {
2885 			/*
2886 			 * if this code bloat is causing too much of a hit
2887 			 * move it to a separate function and mark it noinline
2888 			 */
2889 			if (m->m_len == pi->ipi_ehdrlen) {
2890 				n = m->m_next;
2891 				MPASS(n);
2892 				if (n->m_len >= sizeof(*ip))  {
2893 					ip = (struct ip *)n->m_data;
2894 					if (n->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2895 						th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2896 				} else {
2897 					txq->ift_pullups++;
2898 					if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
2899 						return (ENOMEM);
2900 					ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2901 				}
2902 			} else {
2903 				txq->ift_pullups++;
2904 				if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
2905 					return (ENOMEM);
2906 				ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2907 				if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2908 					th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2909 			}
2910 		} else {
2911 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2912 			if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2913 				th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2914 		}
2915 		pi->ipi_ip_hlen = ip->ip_hl << 2;
2916 		pi->ipi_ipproto = ip->ip_p;
2917 		pi->ipi_flags |= IPI_TX_IPV4;
2918 
2919 		if ((sctx->isc_flags & IFLIB_NEED_ZERO_CSUM) && (pi->ipi_csum_flags & CSUM_IP))
2920                        ip->ip_sum = 0;
2921 
2922 		/* TCP checksum offload may require TCP header length */
2923 		if (IS_TX_OFFLOAD4(pi)) {
2924 			if (__predict_true(pi->ipi_ipproto == IPPROTO_TCP)) {
2925 				if (__predict_false(th == NULL)) {
2926 					txq->ift_pullups++;
2927 					if (__predict_false((m = m_pullup(m, (ip->ip_hl << 2) + sizeof(*th))) == NULL))
2928 						return (ENOMEM);
2929 					th = (struct tcphdr *)((caddr_t)ip + pi->ipi_ip_hlen);
2930 				}
2931 				pi->ipi_tcp_hflags = th->th_flags;
2932 				pi->ipi_tcp_hlen = th->th_off << 2;
2933 				pi->ipi_tcp_seq = th->th_seq;
2934 			}
2935 			if (IS_TSO4(pi)) {
2936 				if (__predict_false(ip->ip_p != IPPROTO_TCP))
2937 					return (ENXIO);
2938 				th->th_sum = in_pseudo(ip->ip_src.s_addr,
2939 						       ip->ip_dst.s_addr, htons(IPPROTO_TCP));
2940 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
2941 				if (sctx->isc_flags & IFLIB_TSO_INIT_IP) {
2942 					ip->ip_sum = 0;
2943 					ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz);
2944 				}
2945 			}
2946 		}
2947 		break;
2948 	}
2949 #endif
2950 #ifdef INET6
2951 	case ETHERTYPE_IPV6:
2952 	{
2953 		struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
2954 		struct tcphdr *th;
2955 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
2956 
2957 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
2958 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
2959 				return (ENOMEM);
2960 		}
2961 		th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
2962 
2963 		/* XXX-BZ this will go badly in case of ext hdrs. */
2964 		pi->ipi_ipproto = ip6->ip6_nxt;
2965 		pi->ipi_flags |= IPI_TX_IPV6;
2966 
2967 		/* TCP checksum offload may require TCP header length */
2968 		if (IS_TX_OFFLOAD6(pi)) {
2969 			if (pi->ipi_ipproto == IPPROTO_TCP) {
2970 				if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) {
2971 					txq->ift_pullups++;
2972 					if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL))
2973 						return (ENOMEM);
2974 				}
2975 				pi->ipi_tcp_hflags = th->th_flags;
2976 				pi->ipi_tcp_hlen = th->th_off << 2;
2977 				pi->ipi_tcp_seq = th->th_seq;
2978 			}
2979 			if (IS_TSO6(pi)) {
2980 				if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP))
2981 					return (ENXIO);
2982 				/*
2983 				 * The corresponding flag is set by the stack in the IPv4
2984 				 * TSO case, but not in IPv6 (at least in FreeBSD 10.2).
2985 				 * So, set it here because the rest of the flow requires it.
2986 				 */
2987 				pi->ipi_csum_flags |= CSUM_IP6_TCP;
2988 				th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0);
2989 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
2990 			}
2991 		}
2992 		break;
2993 	}
2994 #endif
2995 	default:
2996 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
2997 		pi->ipi_ip_hlen = 0;
2998 		break;
2999 	}
3000 	*mp = m;
3001 
3002 	return (0);
3003 }
3004 
3005 static  __noinline  struct mbuf *
3006 collapse_pkthdr(struct mbuf *m0)
3007 {
3008 	struct mbuf *m, *m_next, *tmp;
3009 
3010 	m = m0;
3011 	m_next = m->m_next;
3012 	while (m_next != NULL && m_next->m_len == 0) {
3013 		m = m_next;
3014 		m->m_next = NULL;
3015 		m_free(m);
3016 		m_next = m_next->m_next;
3017 	}
3018 	m = m0;
3019 	m->m_next = m_next;
3020 	if ((m_next->m_flags & M_EXT) == 0) {
3021 		m = m_defrag(m, M_NOWAIT);
3022 	} else {
3023 		tmp = m_next->m_next;
3024 		memcpy(m_next, m, MPKTHSIZE);
3025 		m = m_next;
3026 		m->m_next = tmp;
3027 	}
3028 	return (m);
3029 }
3030 
3031 /*
3032  * If dodgy hardware rejects the scatter gather chain we've handed it
3033  * we'll need to remove the mbuf chain from ifsg_m[] before we can add the
3034  * m_defrag'd mbufs
3035  */
3036 static __noinline struct mbuf *
3037 iflib_remove_mbuf(iflib_txq_t txq)
3038 {
3039 	int ntxd, i, pidx;
3040 	struct mbuf *m, *mh, **ifsd_m;
3041 
3042 	pidx = txq->ift_pidx;
3043 	ifsd_m = txq->ift_sds.ifsd_m;
3044 	ntxd = txq->ift_size;
3045 	mh = m = ifsd_m[pidx];
3046 	ifsd_m[pidx] = NULL;
3047 #if MEMORY_LOGGING
3048 	txq->ift_dequeued++;
3049 #endif
3050 	i = 1;
3051 
3052 	while (m) {
3053 		ifsd_m[(pidx + i) & (ntxd -1)] = NULL;
3054 #if MEMORY_LOGGING
3055 		txq->ift_dequeued++;
3056 #endif
3057 		m = m->m_next;
3058 		i++;
3059 	}
3060 	return (mh);
3061 }
3062 
3063 static int
3064 iflib_busdma_load_mbuf_sg(iflib_txq_t txq, bus_dma_tag_t tag, bus_dmamap_t map,
3065 			  struct mbuf **m0, bus_dma_segment_t *segs, int *nsegs,
3066 			  int max_segs, int flags)
3067 {
3068 	if_ctx_t ctx;
3069 	if_shared_ctx_t		sctx;
3070 	if_softc_ctx_t		scctx;
3071 	int i, next, pidx, err, ntxd, count;
3072 	struct mbuf *m, *tmp, **ifsd_m;
3073 
3074 	m = *m0;
3075 
3076 	/*
3077 	 * Please don't ever do this
3078 	 */
3079 	if (__predict_false(m->m_len == 0))
3080 		*m0 = m = collapse_pkthdr(m);
3081 
3082 	ctx = txq->ift_ctx;
3083 	sctx = ctx->ifc_sctx;
3084 	scctx = &ctx->ifc_softc_ctx;
3085 	ifsd_m = txq->ift_sds.ifsd_m;
3086 	ntxd = txq->ift_size;
3087 	pidx = txq->ift_pidx;
3088 	if (map != NULL) {
3089 		uint8_t *ifsd_flags = txq->ift_sds.ifsd_flags;
3090 
3091 		err = bus_dmamap_load_mbuf_sg(tag, map,
3092 					      *m0, segs, nsegs, BUS_DMA_NOWAIT);
3093 		if (err)
3094 			return (err);
3095 		ifsd_flags[pidx] |= TX_SW_DESC_MAPPED;
3096 		count = 0;
3097 		m = *m0;
3098 		do {
3099 			if (__predict_false(m->m_len <= 0)) {
3100 				tmp = m;
3101 				m = m->m_next;
3102 				tmp->m_next = NULL;
3103 				m_free(tmp);
3104 				continue;
3105 			}
3106 			m = m->m_next;
3107 			count++;
3108 		} while (m != NULL);
3109 		if (count > *nsegs) {
3110 			ifsd_m[pidx] = *m0;
3111 			ifsd_m[pidx]->m_flags |= M_TOOBIG;
3112 			return (0);
3113 		}
3114 		m = *m0;
3115 		count = 0;
3116 		do {
3117 			next = (pidx + count) & (ntxd-1);
3118 			MPASS(ifsd_m[next] == NULL);
3119 			ifsd_m[next] = m;
3120 			count++;
3121 			tmp = m;
3122 			m = m->m_next;
3123 		} while (m != NULL);
3124 	} else {
3125 		int buflen, sgsize, maxsegsz, max_sgsize;
3126 		vm_offset_t vaddr;
3127 		vm_paddr_t curaddr;
3128 
3129 		count = i = 0;
3130 		m = *m0;
3131 		if (m->m_pkthdr.csum_flags & CSUM_TSO)
3132 			maxsegsz = scctx->isc_tx_tso_segsize_max;
3133 		else
3134 			maxsegsz = sctx->isc_tx_maxsegsize;
3135 
3136 		do {
3137 			if (__predict_false(m->m_len <= 0)) {
3138 				tmp = m;
3139 				m = m->m_next;
3140 				tmp->m_next = NULL;
3141 				m_free(tmp);
3142 				continue;
3143 			}
3144 			buflen = m->m_len;
3145 			vaddr = (vm_offset_t)m->m_data;
3146 			/*
3147 			 * see if we can't be smarter about physically
3148 			 * contiguous mappings
3149 			 */
3150 			next = (pidx + count) & (ntxd-1);
3151 			MPASS(ifsd_m[next] == NULL);
3152 #if MEMORY_LOGGING
3153 			txq->ift_enqueued++;
3154 #endif
3155 			ifsd_m[next] = m;
3156 			while (buflen > 0) {
3157 				if (i >= max_segs)
3158 					goto err;
3159 				max_sgsize = MIN(buflen, maxsegsz);
3160 				curaddr = pmap_kextract(vaddr);
3161 				sgsize = PAGE_SIZE - (curaddr & PAGE_MASK);
3162 				sgsize = MIN(sgsize, max_sgsize);
3163 				segs[i].ds_addr = curaddr;
3164 				segs[i].ds_len = sgsize;
3165 				vaddr += sgsize;
3166 				buflen -= sgsize;
3167 				i++;
3168 			}
3169 			count++;
3170 			tmp = m;
3171 			m = m->m_next;
3172 		} while (m != NULL);
3173 		*nsegs = i;
3174 	}
3175 	return (0);
3176 err:
3177 	*m0 = iflib_remove_mbuf(txq);
3178 	return (EFBIG);
3179 }
3180 
3181 static inline caddr_t
3182 calc_next_txd(iflib_txq_t txq, int cidx, uint8_t qid)
3183 {
3184 	qidx_t size;
3185 	int ntxd;
3186 	caddr_t start, end, cur, next;
3187 
3188 	ntxd = txq->ift_size;
3189 	size = txq->ift_txd_size[qid];
3190 	start = txq->ift_ifdi[qid].idi_vaddr;
3191 
3192 	if (__predict_false(size == 0))
3193 		return (start);
3194 	cur = start + size*cidx;
3195 	end = start + size*ntxd;
3196 	next = CACHE_PTR_NEXT(cur);
3197 	return (next < end ? next : start);
3198 }
3199 
3200 /*
3201  * Pad an mbuf to ensure a minimum ethernet frame size.
3202  * min_frame_size is the frame size (less CRC) to pad the mbuf to
3203  */
3204 static __noinline int
3205 iflib_ether_pad(device_t dev, struct mbuf **m_head, uint16_t min_frame_size)
3206 {
3207 	/*
3208 	 * 18 is enough bytes to pad an ARP packet to 46 bytes, and
3209 	 * and ARP message is the smallest common payload I can think of
3210 	 */
3211 	static char pad[18];	/* just zeros */
3212 	int n;
3213 	struct mbuf *new_head;
3214 
3215 	if (!M_WRITABLE(*m_head)) {
3216 		new_head = m_dup(*m_head, M_NOWAIT);
3217 		if (new_head == NULL) {
3218 			m_freem(*m_head);
3219 			device_printf(dev, "cannot pad short frame, m_dup() failed");
3220 			DBG_COUNTER_INC(encap_pad_mbuf_fail);
3221 			return ENOMEM;
3222 		}
3223 		m_freem(*m_head);
3224 		*m_head = new_head;
3225 	}
3226 
3227 	for (n = min_frame_size - (*m_head)->m_pkthdr.len;
3228 	     n > 0; n -= sizeof(pad))
3229 		if (!m_append(*m_head, min(n, sizeof(pad)), pad))
3230 			break;
3231 
3232 	if (n > 0) {
3233 		m_freem(*m_head);
3234 		device_printf(dev, "cannot pad short frame\n");
3235 		DBG_COUNTER_INC(encap_pad_mbuf_fail);
3236 		return (ENOBUFS);
3237 	}
3238 
3239 	return 0;
3240 }
3241 
3242 static int
3243 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp)
3244 {
3245 	if_ctx_t		ctx;
3246 	if_shared_ctx_t		sctx;
3247 	if_softc_ctx_t		scctx;
3248 	bus_dma_segment_t	*segs;
3249 	struct mbuf		*m_head;
3250 	void			*next_txd;
3251 	bus_dmamap_t		map;
3252 	struct if_pkt_info	pi;
3253 	int remap = 0;
3254 	int err, nsegs, ndesc, max_segs, pidx, cidx, next, ntxd;
3255 	bus_dma_tag_t desc_tag;
3256 
3257 	segs = txq->ift_segs;
3258 	ctx = txq->ift_ctx;
3259 	sctx = ctx->ifc_sctx;
3260 	scctx = &ctx->ifc_softc_ctx;
3261 	segs = txq->ift_segs;
3262 	ntxd = txq->ift_size;
3263 	m_head = *m_headp;
3264 	map = NULL;
3265 
3266 	/*
3267 	 * If we're doing TSO the next descriptor to clean may be quite far ahead
3268 	 */
3269 	cidx = txq->ift_cidx;
3270 	pidx = txq->ift_pidx;
3271 	if (ctx->ifc_flags & IFC_PREFETCH) {
3272 		next = (cidx + CACHE_PTR_INCREMENT) & (ntxd-1);
3273 		if (!(ctx->ifc_flags & IFLIB_HAS_TXCQ)) {
3274 			next_txd = calc_next_txd(txq, cidx, 0);
3275 			prefetch(next_txd);
3276 		}
3277 
3278 		/* prefetch the next cache line of mbuf pointers and flags */
3279 		prefetch(&txq->ift_sds.ifsd_m[next]);
3280 		if (txq->ift_sds.ifsd_map != NULL) {
3281 			prefetch(&txq->ift_sds.ifsd_map[next]);
3282 			next = (cidx + CACHE_LINE_SIZE) & (ntxd-1);
3283 			prefetch(&txq->ift_sds.ifsd_flags[next]);
3284 		}
3285 	} else if (txq->ift_sds.ifsd_map != NULL)
3286 		map = txq->ift_sds.ifsd_map[pidx];
3287 
3288 	if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
3289 		desc_tag = txq->ift_tso_desc_tag;
3290 		max_segs = scctx->isc_tx_tso_segments_max;
3291 		MPASS(desc_tag != NULL);
3292 		MPASS(max_segs > 0);
3293 	} else {
3294 		desc_tag = txq->ift_desc_tag;
3295 		max_segs = scctx->isc_tx_nsegments;
3296 	}
3297 	if ((sctx->isc_flags & IFLIB_NEED_ETHER_PAD) &&
3298 	    __predict_false(m_head->m_pkthdr.len < scctx->isc_min_frame_size)) {
3299 		err = iflib_ether_pad(ctx->ifc_dev, m_headp, scctx->isc_min_frame_size);
3300 		if (err)
3301 			return err;
3302 	}
3303 	m_head = *m_headp;
3304 
3305 	pkt_info_zero(&pi);
3306 	pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG|M_BCAST|M_MCAST));
3307 	pi.ipi_pidx = pidx;
3308 	pi.ipi_qsidx = txq->ift_id;
3309 	pi.ipi_len = m_head->m_pkthdr.len;
3310 	pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags;
3311 	pi.ipi_vtag = (m_head->m_flags & M_VLANTAG) ? m_head->m_pkthdr.ether_vtag : 0;
3312 
3313 	/* deliberate bitwise OR to make one condition */
3314 	if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) {
3315 		if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0))
3316 			return (err);
3317 		m_head = *m_headp;
3318 	}
3319 
3320 retry:
3321 	err = iflib_busdma_load_mbuf_sg(txq, desc_tag, map, m_headp, segs, &nsegs, max_segs, BUS_DMA_NOWAIT);
3322 defrag:
3323 	if (__predict_false(err)) {
3324 		switch (err) {
3325 		case EFBIG:
3326 			/* try collapse once and defrag once */
3327 			if (remap == 0) {
3328 				m_head = m_collapse(*m_headp, M_NOWAIT, max_segs);
3329 				/* try defrag if collapsing fails */
3330 				if (m_head == NULL)
3331 					remap++;
3332 			}
3333 			if (remap == 1)
3334 				m_head = m_defrag(*m_headp, M_NOWAIT);
3335 			remap++;
3336 			if (__predict_false(m_head == NULL))
3337 				goto defrag_failed;
3338 			txq->ift_mbuf_defrag++;
3339 			*m_headp = m_head;
3340 			goto retry;
3341 			break;
3342 		case ENOMEM:
3343 			txq->ift_no_tx_dma_setup++;
3344 			break;
3345 		default:
3346 			txq->ift_no_tx_dma_setup++;
3347 			m_freem(*m_headp);
3348 			DBG_COUNTER_INC(tx_frees);
3349 			*m_headp = NULL;
3350 			break;
3351 		}
3352 		txq->ift_map_failed++;
3353 		DBG_COUNTER_INC(encap_load_mbuf_fail);
3354 		return (err);
3355 	}
3356 
3357 	/*
3358 	 * XXX assumes a 1 to 1 relationship between segments and
3359 	 *        descriptors - this does not hold true on all drivers, e.g.
3360 	 *        cxgb
3361 	 */
3362 	if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) {
3363 		txq->ift_no_desc_avail++;
3364 		if (map != NULL)
3365 			bus_dmamap_unload(desc_tag, map);
3366 		DBG_COUNTER_INC(encap_txq_avail_fail);
3367 		if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0)
3368 			GROUPTASK_ENQUEUE(&txq->ift_task);
3369 		return (ENOBUFS);
3370 	}
3371 	/*
3372 	 * On Intel cards we can greatly reduce the number of TX interrupts
3373 	 * we see by only setting report status on every Nth descriptor.
3374 	 * However, this also means that the driver will need to keep track
3375 	 * of the descriptors that RS was set on to check them for the DD bit.
3376 	 */
3377 	txq->ift_rs_pending += nsegs + 1;
3378 	if (txq->ift_rs_pending > TXQ_MAX_RS_DEFERRED(txq) ||
3379 	     iflib_no_tx_batch || (TXQ_AVAIL(txq) - nsegs) <= MAX_TX_DESC(ctx) + 2) {
3380 		pi.ipi_flags |= IPI_TX_INTR;
3381 		txq->ift_rs_pending = 0;
3382 	}
3383 
3384 	pi.ipi_segs = segs;
3385 	pi.ipi_nsegs = nsegs;
3386 
3387 	MPASS(pidx >= 0 && pidx < txq->ift_size);
3388 #ifdef PKT_DEBUG
3389 	print_pkt(&pi);
3390 #endif
3391 	if (map != NULL)
3392 		bus_dmamap_sync(desc_tag, map, BUS_DMASYNC_PREWRITE);
3393 	if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) {
3394 		if (map != NULL)
3395 			bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3396 					BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
3397 		DBG_COUNTER_INC(tx_encap);
3398 		MPASS(pi.ipi_new_pidx < txq->ift_size);
3399 
3400 		ndesc = pi.ipi_new_pidx - pi.ipi_pidx;
3401 		if (pi.ipi_new_pidx < pi.ipi_pidx) {
3402 			ndesc += txq->ift_size;
3403 			txq->ift_gen = 1;
3404 		}
3405 		/*
3406 		 * drivers can need as many as
3407 		 * two sentinels
3408 		 */
3409 		MPASS(ndesc <= pi.ipi_nsegs + 2);
3410 		MPASS(pi.ipi_new_pidx != pidx);
3411 		MPASS(ndesc > 0);
3412 		txq->ift_in_use += ndesc;
3413 
3414 		/*
3415 		 * We update the last software descriptor again here because there may
3416 		 * be a sentinel and/or there may be more mbufs than segments
3417 		 */
3418 		txq->ift_pidx = pi.ipi_new_pidx;
3419 		txq->ift_npending += pi.ipi_ndescs;
3420 	} else {
3421 		*m_headp = m_head = iflib_remove_mbuf(txq);
3422 		if (err == EFBIG) {
3423 			txq->ift_txd_encap_efbig++;
3424 			if (remap < 2) {
3425 				remap = 1;
3426 				goto defrag;
3427 			}
3428 		}
3429 		DBG_COUNTER_INC(encap_txd_encap_fail);
3430 		goto defrag_failed;
3431 	}
3432 	return (err);
3433 
3434 defrag_failed:
3435 	txq->ift_mbuf_defrag_failed++;
3436 	txq->ift_map_failed++;
3437 	m_freem(*m_headp);
3438 	DBG_COUNTER_INC(tx_frees);
3439 	*m_headp = NULL;
3440 	return (ENOMEM);
3441 }
3442 
3443 static void
3444 iflib_tx_desc_free(iflib_txq_t txq, int n)
3445 {
3446 	int hasmap;
3447 	uint32_t qsize, cidx, mask, gen;
3448 	struct mbuf *m, **ifsd_m;
3449 	uint8_t *ifsd_flags;
3450 	bus_dmamap_t *ifsd_map;
3451 	bool do_prefetch;
3452 
3453 	cidx = txq->ift_cidx;
3454 	gen = txq->ift_gen;
3455 	qsize = txq->ift_size;
3456 	mask = qsize-1;
3457 	hasmap = txq->ift_sds.ifsd_map != NULL;
3458 	ifsd_flags = txq->ift_sds.ifsd_flags;
3459 	ifsd_m = txq->ift_sds.ifsd_m;
3460 	ifsd_map = txq->ift_sds.ifsd_map;
3461 	do_prefetch = (txq->ift_ctx->ifc_flags & IFC_PREFETCH);
3462 
3463 	while (n-- > 0) {
3464 		if (do_prefetch) {
3465 			prefetch(ifsd_m[(cidx + 3) & mask]);
3466 			prefetch(ifsd_m[(cidx + 4) & mask]);
3467 		}
3468 		if (ifsd_m[cidx] != NULL) {
3469 			prefetch(&ifsd_m[(cidx + CACHE_PTR_INCREMENT) & mask]);
3470 			prefetch(&ifsd_flags[(cidx + CACHE_PTR_INCREMENT) & mask]);
3471 			if (hasmap && (ifsd_flags[cidx] & TX_SW_DESC_MAPPED)) {
3472 				/*
3473 				 * does it matter if it's not the TSO tag? If so we'll
3474 				 * have to add the type to flags
3475 				 */
3476 				bus_dmamap_unload(txq->ift_desc_tag, ifsd_map[cidx]);
3477 				ifsd_flags[cidx] &= ~TX_SW_DESC_MAPPED;
3478 			}
3479 			if ((m = ifsd_m[cidx]) != NULL) {
3480 				/* XXX we don't support any drivers that batch packets yet */
3481 				MPASS(m->m_nextpkt == NULL);
3482 				/* if the number of clusters exceeds the number of segments
3483 				 * there won't be space on the ring to save a pointer to each
3484 				 * cluster so we simply free the list here
3485 				 */
3486 				if (m->m_flags & M_TOOBIG) {
3487 					m_freem(m);
3488 				} else {
3489 					m_free(m);
3490 				}
3491 				ifsd_m[cidx] = NULL;
3492 #if MEMORY_LOGGING
3493 				txq->ift_dequeued++;
3494 #endif
3495 				DBG_COUNTER_INC(tx_frees);
3496 			}
3497 		}
3498 		if (__predict_false(++cidx == qsize)) {
3499 			cidx = 0;
3500 			gen = 0;
3501 		}
3502 	}
3503 	txq->ift_cidx = cidx;
3504 	txq->ift_gen = gen;
3505 }
3506 
3507 static __inline int
3508 iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh)
3509 {
3510 	int reclaim;
3511 	if_ctx_t ctx = txq->ift_ctx;
3512 
3513 	KASSERT(thresh >= 0, ("invalid threshold to reclaim"));
3514 	MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size);
3515 
3516 	/*
3517 	 * Need a rate-limiting check so that this isn't called every time
3518 	 */
3519 	iflib_tx_credits_update(ctx, txq);
3520 	reclaim = DESC_RECLAIMABLE(txq);
3521 
3522 	if (reclaim <= thresh /* + MAX_TX_DESC(txq->ift_ctx) */) {
3523 #ifdef INVARIANTS
3524 		if (iflib_verbose_debug) {
3525 			printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __FUNCTION__,
3526 			       txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments,
3527 			       reclaim, thresh);
3528 
3529 		}
3530 #endif
3531 		return (0);
3532 	}
3533 	iflib_tx_desc_free(txq, reclaim);
3534 	txq->ift_cleaned += reclaim;
3535 	txq->ift_in_use -= reclaim;
3536 
3537 	return (reclaim);
3538 }
3539 
3540 static struct mbuf **
3541 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset, int remaining)
3542 {
3543 	int next, size;
3544 	struct mbuf **items;
3545 
3546 	size = r->size;
3547 	next = (cidx + CACHE_PTR_INCREMENT) & (size-1);
3548 	items = __DEVOLATILE(struct mbuf **, &r->items[0]);
3549 
3550 	prefetch(items[(cidx + offset) & (size-1)]);
3551 	if (remaining > 1) {
3552 		prefetch2cachelines(&items[next]);
3553 		prefetch2cachelines(items[(cidx + offset + 1) & (size-1)]);
3554 		prefetch2cachelines(items[(cidx + offset + 2) & (size-1)]);
3555 		prefetch2cachelines(items[(cidx + offset + 3) & (size-1)]);
3556 	}
3557 	return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (size-1)]));
3558 }
3559 
3560 static void
3561 iflib_txq_check_drain(iflib_txq_t txq, int budget)
3562 {
3563 
3564 	ifmp_ring_check_drainage(txq->ift_br, budget);
3565 }
3566 
3567 static uint32_t
3568 iflib_txq_can_drain(struct ifmp_ring *r)
3569 {
3570 	iflib_txq_t txq = r->cookie;
3571 	if_ctx_t ctx = txq->ift_ctx;
3572 
3573 	return ((TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2) ||
3574 		ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false));
3575 }
3576 
3577 static uint32_t
3578 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3579 {
3580 	iflib_txq_t txq = r->cookie;
3581 	if_ctx_t ctx = txq->ift_ctx;
3582 	struct ifnet *ifp = ctx->ifc_ifp;
3583 	struct mbuf **mp, *m;
3584 	int i, count, consumed, pkt_sent, bytes_sent, mcast_sent, avail;
3585 	int reclaimed, err, in_use_prev, desc_used;
3586 	bool do_prefetch, ring, rang;
3587 
3588 	if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) ||
3589 			    !LINK_ACTIVE(ctx))) {
3590 		DBG_COUNTER_INC(txq_drain_notready);
3591 		return (0);
3592 	}
3593 	reclaimed = iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
3594 	rang = iflib_txd_db_check(ctx, txq, reclaimed, txq->ift_in_use);
3595 	avail = IDXDIFF(pidx, cidx, r->size);
3596 	if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) {
3597 		DBG_COUNTER_INC(txq_drain_flushing);
3598 		for (i = 0; i < avail; i++) {
3599 			m_free(r->items[(cidx + i) & (r->size-1)]);
3600 			r->items[(cidx + i) & (r->size-1)] = NULL;
3601 		}
3602 		return (avail);
3603 	}
3604 
3605 	if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3606 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3607 		CALLOUT_LOCK(txq);
3608 		callout_stop(&txq->ift_timer);
3609 		CALLOUT_UNLOCK(txq);
3610 		DBG_COUNTER_INC(txq_drain_oactive);
3611 		return (0);
3612 	}
3613 	if (reclaimed)
3614 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3615 	consumed = mcast_sent = bytes_sent = pkt_sent = 0;
3616 	count = MIN(avail, TX_BATCH_SIZE);
3617 #ifdef INVARIANTS
3618 	if (iflib_verbose_debug)
3619 		printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __FUNCTION__,
3620 		       avail, ctx->ifc_flags, TXQ_AVAIL(txq));
3621 #endif
3622 	do_prefetch = (ctx->ifc_flags & IFC_PREFETCH);
3623 	avail = TXQ_AVAIL(txq);
3624 	err = 0;
3625 	for (desc_used = i = 0; i < count && avail > MAX_TX_DESC(ctx) + 2; i++) {
3626 		int rem = do_prefetch ? count - i : 0;
3627 
3628 		mp = _ring_peek_one(r, cidx, i, rem);
3629 		MPASS(mp != NULL && *mp != NULL);
3630 		if (__predict_false(*mp == (struct mbuf *)txq)) {
3631 			consumed++;
3632 			reclaimed++;
3633 			continue;
3634 		}
3635 		in_use_prev = txq->ift_in_use;
3636 		err = iflib_encap(txq, mp);
3637 		if (__predict_false(err)) {
3638 			DBG_COUNTER_INC(txq_drain_encapfail);
3639 			/* no room - bail out */
3640 			if (err == ENOBUFS)
3641 				break;
3642 			consumed++;
3643 			DBG_COUNTER_INC(txq_drain_encapfail);
3644 			/* we can't send this packet - skip it */
3645 			continue;
3646 		}
3647 		consumed++;
3648 		pkt_sent++;
3649 		m = *mp;
3650 		DBG_COUNTER_INC(tx_sent);
3651 		bytes_sent += m->m_pkthdr.len;
3652 		mcast_sent += !!(m->m_flags & M_MCAST);
3653 		avail = TXQ_AVAIL(txq);
3654 
3655 		txq->ift_db_pending += (txq->ift_in_use - in_use_prev);
3656 		desc_used += (txq->ift_in_use - in_use_prev);
3657 		ETHER_BPF_MTAP(ifp, m);
3658 		if (__predict_false(!(ifp->if_drv_flags & IFF_DRV_RUNNING)))
3659 			break;
3660 		rang = iflib_txd_db_check(ctx, txq, false, in_use_prev);
3661 	}
3662 
3663 	/* deliberate use of bitwise or to avoid gratuitous short-circuit */
3664 	ring = rang ? false  : (iflib_min_tx_latency | err) || (TXQ_AVAIL(txq) < MAX_TX_DESC(ctx));
3665 	iflib_txd_db_check(ctx, txq, ring, txq->ift_in_use);
3666 	if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
3667 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
3668 	if (mcast_sent)
3669 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
3670 #ifdef INVARIANTS
3671 	if (iflib_verbose_debug)
3672 		printf("consumed=%d\n", consumed);
3673 #endif
3674 	return (consumed);
3675 }
3676 
3677 static uint32_t
3678 iflib_txq_drain_always(struct ifmp_ring *r)
3679 {
3680 	return (1);
3681 }
3682 
3683 static uint32_t
3684 iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3685 {
3686 	int i, avail;
3687 	struct mbuf **mp;
3688 	iflib_txq_t txq;
3689 
3690 	txq = r->cookie;
3691 
3692 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3693 	CALLOUT_LOCK(txq);
3694 	callout_stop(&txq->ift_timer);
3695 	CALLOUT_UNLOCK(txq);
3696 
3697 	avail = IDXDIFF(pidx, cidx, r->size);
3698 	for (i = 0; i < avail; i++) {
3699 		mp = _ring_peek_one(r, cidx, i, avail - i);
3700 		if (__predict_false(*mp == (struct mbuf *)txq))
3701 			continue;
3702 		m_freem(*mp);
3703 	}
3704 	MPASS(ifmp_ring_is_stalled(r) == 0);
3705 	return (avail);
3706 }
3707 
3708 static void
3709 iflib_ifmp_purge(iflib_txq_t txq)
3710 {
3711 	struct ifmp_ring *r;
3712 
3713 	r = txq->ift_br;
3714 	r->drain = iflib_txq_drain_free;
3715 	r->can_drain = iflib_txq_drain_always;
3716 
3717 	ifmp_ring_check_drainage(r, r->size);
3718 
3719 	r->drain = iflib_txq_drain;
3720 	r->can_drain = iflib_txq_can_drain;
3721 }
3722 
3723 static void
3724 _task_fn_tx(void *context)
3725 {
3726 	iflib_txq_t txq = context;
3727 	if_ctx_t ctx = txq->ift_ctx;
3728 	struct ifnet *ifp = ctx->ifc_ifp;
3729 
3730 #ifdef IFLIB_DIAGNOSTICS
3731 	txq->ift_cpu_exec_count[curcpu]++;
3732 #endif
3733 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3734 		return;
3735 	if (if_getcapenable(ifp) & IFCAP_NETMAP) {
3736 		/*
3737 		 * If there are no available credits, and TX IRQs are not in use,
3738 		 * re-schedule the task immediately.
3739 		 */
3740 		if (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false))
3741 			netmap_tx_irq(ifp, txq->ift_id);
3742 		IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3743 		return;
3744 	}
3745 	if (txq->ift_db_pending)
3746 		ifmp_ring_enqueue(txq->ift_br, (void **)&txq, 1, TX_BATCH_SIZE);
3747 	ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3748 	if (ctx->ifc_flags & IFC_LEGACY)
3749 		IFDI_INTR_ENABLE(ctx);
3750 	else {
3751 #ifdef INVARIANTS
3752 		int rc =
3753 #endif
3754 			IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3755 			KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver"));
3756 	}
3757 }
3758 
3759 static void
3760 _task_fn_rx(void *context)
3761 {
3762 	iflib_rxq_t rxq = context;
3763 	if_ctx_t ctx = rxq->ifr_ctx;
3764 	bool more;
3765 	uint16_t budget;
3766 
3767 #ifdef IFLIB_DIAGNOSTICS
3768 	rxq->ifr_cpu_exec_count[curcpu]++;
3769 #endif
3770 	DBG_COUNTER_INC(task_fn_rxs);
3771 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3772 		return;
3773 	more = true;
3774 #ifdef DEV_NETMAP
3775 	if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP) {
3776 		u_int work = 0;
3777 		if (netmap_rx_irq(ctx->ifc_ifp, rxq->ifr_id, &work)) {
3778 			more = false;
3779 		}
3780 	}
3781 #endif
3782 	budget = ctx->ifc_sysctl_rx_budget;
3783 	if (budget == 0)
3784 		budget = 16;	/* XXX */
3785 	if (more == false || (more = iflib_rxeof(rxq, budget)) == false) {
3786 		if (ctx->ifc_flags & IFC_LEGACY)
3787 			IFDI_INTR_ENABLE(ctx);
3788 		else {
3789 #ifdef INVARIANTS
3790 			int rc =
3791 #endif
3792 				IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
3793 			KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver"));
3794 			DBG_COUNTER_INC(rx_intr_enables);
3795 		}
3796 	}
3797 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3798 		return;
3799 	if (more)
3800 		GROUPTASK_ENQUEUE(&rxq->ifr_task);
3801 }
3802 
3803 static void
3804 _task_fn_admin(void *context)
3805 {
3806 	if_ctx_t ctx = context;
3807 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
3808 	iflib_txq_t txq;
3809 	int i;
3810 	bool oactive, running, do_reset, do_watchdog;
3811 
3812 	STATE_LOCK(ctx);
3813 	running = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING);
3814 	oactive = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE);
3815 	do_reset = (ctx->ifc_flags & IFC_DO_RESET);
3816 	do_watchdog = (ctx->ifc_flags & IFC_DO_WATCHDOG);
3817 	ctx->ifc_flags &= ~(IFC_DO_RESET|IFC_DO_WATCHDOG);
3818 	STATE_UNLOCK(ctx);
3819 
3820 	if ((!running & !oactive) &&
3821 	    !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
3822 		return;
3823 
3824 	CTX_LOCK(ctx);
3825 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3826 		CALLOUT_LOCK(txq);
3827 		callout_stop(&txq->ift_timer);
3828 		CALLOUT_UNLOCK(txq);
3829 	}
3830 	if (do_watchdog) {
3831 		ctx->ifc_watchdog_events++;
3832 		IFDI_WATCHDOG_RESET(ctx);
3833 	}
3834 	IFDI_UPDATE_ADMIN_STATUS(ctx);
3835 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3836 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq, txq->ift_timer.c_cpu);
3837 	IFDI_LINK_INTR_ENABLE(ctx);
3838 	if (do_reset)
3839 		iflib_if_init_locked(ctx);
3840 	CTX_UNLOCK(ctx);
3841 
3842 	if (LINK_ACTIVE(ctx) == 0)
3843 		return;
3844 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3845 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
3846 }
3847 
3848 
3849 static void
3850 _task_fn_iov(void *context)
3851 {
3852 	if_ctx_t ctx = context;
3853 
3854 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3855 		return;
3856 
3857 	CTX_LOCK(ctx);
3858 	IFDI_VFLR_HANDLE(ctx);
3859 	CTX_UNLOCK(ctx);
3860 }
3861 
3862 static int
3863 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)
3864 {
3865 	int err;
3866 	if_int_delay_info_t info;
3867 	if_ctx_t ctx;
3868 
3869 	info = (if_int_delay_info_t)arg1;
3870 	ctx = info->iidi_ctx;
3871 	info->iidi_req = req;
3872 	info->iidi_oidp = oidp;
3873 	CTX_LOCK(ctx);
3874 	err = IFDI_SYSCTL_INT_DELAY(ctx, info);
3875 	CTX_UNLOCK(ctx);
3876 	return (err);
3877 }
3878 
3879 /*********************************************************************
3880  *
3881  *  IFNET FUNCTIONS
3882  *
3883  **********************************************************************/
3884 
3885 static void
3886 iflib_if_init_locked(if_ctx_t ctx)
3887 {
3888 	iflib_stop(ctx);
3889 	iflib_init_locked(ctx);
3890 }
3891 
3892 
3893 static void
3894 iflib_if_init(void *arg)
3895 {
3896 	if_ctx_t ctx = arg;
3897 
3898 	CTX_LOCK(ctx);
3899 	iflib_if_init_locked(ctx);
3900 	CTX_UNLOCK(ctx);
3901 }
3902 
3903 static int
3904 iflib_if_transmit(if_t ifp, struct mbuf *m)
3905 {
3906 	if_ctx_t	ctx = if_getsoftc(ifp);
3907 
3908 	iflib_txq_t txq;
3909 	int err, qidx;
3910 
3911 	if (__predict_false((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) {
3912 		DBG_COUNTER_INC(tx_frees);
3913 		m_freem(m);
3914 		return (ENOBUFS);
3915 	}
3916 
3917 	MPASS(m->m_nextpkt == NULL);
3918 	qidx = 0;
3919 	if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m))
3920 		qidx = QIDX(ctx, m);
3921 	/*
3922 	 * XXX calculate buf_ring based on flowid (divvy up bits?)
3923 	 */
3924 	txq = &ctx->ifc_txqs[qidx];
3925 
3926 #ifdef DRIVER_BACKPRESSURE
3927 	if (txq->ift_closed) {
3928 		while (m != NULL) {
3929 			next = m->m_nextpkt;
3930 			m->m_nextpkt = NULL;
3931 			m_freem(m);
3932 			m = next;
3933 		}
3934 		return (ENOBUFS);
3935 	}
3936 #endif
3937 #ifdef notyet
3938 	qidx = count = 0;
3939 	mp = marr;
3940 	next = m;
3941 	do {
3942 		count++;
3943 		next = next->m_nextpkt;
3944 	} while (next != NULL);
3945 
3946 	if (count > nitems(marr))
3947 		if ((mp = malloc(count*sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) {
3948 			/* XXX check nextpkt */
3949 			m_freem(m);
3950 			/* XXX simplify for now */
3951 			DBG_COUNTER_INC(tx_frees);
3952 			return (ENOBUFS);
3953 		}
3954 	for (next = m, i = 0; next != NULL; i++) {
3955 		mp[i] = next;
3956 		next = next->m_nextpkt;
3957 		mp[i]->m_nextpkt = NULL;
3958 	}
3959 #endif
3960 	DBG_COUNTER_INC(tx_seen);
3961 	err = ifmp_ring_enqueue(txq->ift_br, (void **)&m, 1, TX_BATCH_SIZE);
3962 
3963 	GROUPTASK_ENQUEUE(&txq->ift_task);
3964 	if (err) {
3965 		/* support forthcoming later */
3966 #ifdef DRIVER_BACKPRESSURE
3967 		txq->ift_closed = TRUE;
3968 #endif
3969 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3970 		m_freem(m);
3971 	}
3972 
3973 	return (err);
3974 }
3975 
3976 static void
3977 iflib_if_qflush(if_t ifp)
3978 {
3979 	if_ctx_t ctx = if_getsoftc(ifp);
3980 	iflib_txq_t txq = ctx->ifc_txqs;
3981 	int i;
3982 
3983 	STATE_LOCK(ctx);
3984 	ctx->ifc_flags |= IFC_QFLUSH;
3985 	STATE_UNLOCK(ctx);
3986 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
3987 		while (!(ifmp_ring_is_idle(txq->ift_br) || ifmp_ring_is_stalled(txq->ift_br)))
3988 			iflib_txq_check_drain(txq, 0);
3989 	STATE_LOCK(ctx);
3990 	ctx->ifc_flags &= ~IFC_QFLUSH;
3991 	STATE_UNLOCK(ctx);
3992 
3993 	if_qflush(ifp);
3994 }
3995 
3996 
3997 #define IFCAP_FLAGS (IFCAP_TXCSUM_IPV6 | IFCAP_RXCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \
3998 		     IFCAP_TSO4 | IFCAP_TSO6 | IFCAP_VLAN_HWTAGGING | IFCAP_HWSTATS | \
3999 		     IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | IFCAP_VLAN_HWTSO)
4000 
4001 static int
4002 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data)
4003 {
4004 	if_ctx_t ctx = if_getsoftc(ifp);
4005 	struct ifreq	*ifr = (struct ifreq *)data;
4006 #if defined(INET) || defined(INET6)
4007 	struct ifaddr	*ifa = (struct ifaddr *)data;
4008 #endif
4009 	bool		avoid_reset = FALSE;
4010 	int		err = 0, reinit = 0, bits;
4011 
4012 	switch (command) {
4013 	case SIOCSIFADDR:
4014 #ifdef INET
4015 		if (ifa->ifa_addr->sa_family == AF_INET)
4016 			avoid_reset = TRUE;
4017 #endif
4018 #ifdef INET6
4019 		if (ifa->ifa_addr->sa_family == AF_INET6)
4020 			avoid_reset = TRUE;
4021 #endif
4022 		/*
4023 		** Calling init results in link renegotiation,
4024 		** so we avoid doing it when possible.
4025 		*/
4026 		if (avoid_reset) {
4027 			if_setflagbits(ifp, IFF_UP,0);
4028 			if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
4029 				reinit = 1;
4030 #ifdef INET
4031 			if (!(if_getflags(ifp) & IFF_NOARP))
4032 				arp_ifinit(ifp, ifa);
4033 #endif
4034 		} else
4035 			err = ether_ioctl(ifp, command, data);
4036 		break;
4037 	case SIOCSIFMTU:
4038 		CTX_LOCK(ctx);
4039 		if (ifr->ifr_mtu == if_getmtu(ifp)) {
4040 			CTX_UNLOCK(ctx);
4041 			break;
4042 		}
4043 		bits = if_getdrvflags(ifp);
4044 		/* stop the driver and free any clusters before proceeding */
4045 		iflib_stop(ctx);
4046 
4047 		if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) {
4048 			STATE_LOCK(ctx);
4049 			if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size)
4050 				ctx->ifc_flags |= IFC_MULTISEG;
4051 			else
4052 				ctx->ifc_flags &= ~IFC_MULTISEG;
4053 			STATE_UNLOCK(ctx);
4054 			err = if_setmtu(ifp, ifr->ifr_mtu);
4055 		}
4056 		iflib_init_locked(ctx);
4057 		STATE_LOCK(ctx);
4058 		if_setdrvflags(ifp, bits);
4059 		STATE_UNLOCK(ctx);
4060 		CTX_UNLOCK(ctx);
4061 		break;
4062 	case SIOCSIFFLAGS:
4063 		CTX_LOCK(ctx);
4064 		if (if_getflags(ifp) & IFF_UP) {
4065 			if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4066 				if ((if_getflags(ifp) ^ ctx->ifc_if_flags) &
4067 				    (IFF_PROMISC | IFF_ALLMULTI)) {
4068 					err = IFDI_PROMISC_SET(ctx, if_getflags(ifp));
4069 				}
4070 			} else
4071 				reinit = 1;
4072 		} else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4073 			iflib_stop(ctx);
4074 		}
4075 		ctx->ifc_if_flags = if_getflags(ifp);
4076 		CTX_UNLOCK(ctx);
4077 		break;
4078 	case SIOCADDMULTI:
4079 	case SIOCDELMULTI:
4080 		if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4081 			CTX_LOCK(ctx);
4082 			IFDI_INTR_DISABLE(ctx);
4083 			IFDI_MULTI_SET(ctx);
4084 			IFDI_INTR_ENABLE(ctx);
4085 			CTX_UNLOCK(ctx);
4086 		}
4087 		break;
4088 	case SIOCSIFMEDIA:
4089 		CTX_LOCK(ctx);
4090 		IFDI_MEDIA_SET(ctx);
4091 		CTX_UNLOCK(ctx);
4092 		/* falls thru */
4093 	case SIOCGIFMEDIA:
4094 	case SIOCGIFXMEDIA:
4095 		err = ifmedia_ioctl(ifp, ifr, &ctx->ifc_media, command);
4096 		break;
4097 	case SIOCGI2C:
4098 	{
4099 		struct ifi2creq i2c;
4100 
4101 		err = copyin(ifr_data_get_ptr(ifr), &i2c, sizeof(i2c));
4102 		if (err != 0)
4103 			break;
4104 		if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) {
4105 			err = EINVAL;
4106 			break;
4107 		}
4108 		if (i2c.len > sizeof(i2c.data)) {
4109 			err = EINVAL;
4110 			break;
4111 		}
4112 
4113 		if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0)
4114 			err = copyout(&i2c, ifr_data_get_ptr(ifr),
4115 			    sizeof(i2c));
4116 		break;
4117 	}
4118 	case SIOCSIFCAP:
4119 	{
4120 		int mask, setmask;
4121 
4122 		mask = ifr->ifr_reqcap ^ if_getcapenable(ifp);
4123 		setmask = 0;
4124 #ifdef TCP_OFFLOAD
4125 		setmask |= mask & (IFCAP_TOE4|IFCAP_TOE6);
4126 #endif
4127 		setmask |= (mask & IFCAP_FLAGS);
4128 
4129 		if (setmask & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6))
4130 			setmask |= (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6);
4131 		if ((mask & IFCAP_WOL) &&
4132 		    (if_getcapabilities(ifp) & IFCAP_WOL) != 0)
4133 			setmask |= (mask & (IFCAP_WOL_MCAST|IFCAP_WOL_MAGIC));
4134 		if_vlancap(ifp);
4135 		/*
4136 		 * want to ensure that traffic has stopped before we change any of the flags
4137 		 */
4138 		if (setmask) {
4139 			CTX_LOCK(ctx);
4140 			bits = if_getdrvflags(ifp);
4141 			if (bits & IFF_DRV_RUNNING)
4142 				iflib_stop(ctx);
4143 			STATE_LOCK(ctx);
4144 			if_togglecapenable(ifp, setmask);
4145 			STATE_UNLOCK(ctx);
4146 			if (bits & IFF_DRV_RUNNING)
4147 				iflib_init_locked(ctx);
4148 			STATE_LOCK(ctx);
4149 			if_setdrvflags(ifp, bits);
4150 			STATE_UNLOCK(ctx);
4151 			CTX_UNLOCK(ctx);
4152 		}
4153 		break;
4154 	}
4155 	case SIOCGPRIVATE_0:
4156 	case SIOCSDRVSPEC:
4157 	case SIOCGDRVSPEC:
4158 		CTX_LOCK(ctx);
4159 		err = IFDI_PRIV_IOCTL(ctx, command, data);
4160 		CTX_UNLOCK(ctx);
4161 		break;
4162 	default:
4163 		err = ether_ioctl(ifp, command, data);
4164 		break;
4165 	}
4166 	if (reinit)
4167 		iflib_if_init(ctx);
4168 	return (err);
4169 }
4170 
4171 static uint64_t
4172 iflib_if_get_counter(if_t ifp, ift_counter cnt)
4173 {
4174 	if_ctx_t ctx = if_getsoftc(ifp);
4175 
4176 	return (IFDI_GET_COUNTER(ctx, cnt));
4177 }
4178 
4179 /*********************************************************************
4180  *
4181  *  OTHER FUNCTIONS EXPORTED TO THE STACK
4182  *
4183  **********************************************************************/
4184 
4185 static void
4186 iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag)
4187 {
4188 	if_ctx_t ctx = if_getsoftc(ifp);
4189 
4190 	if ((void *)ctx != arg)
4191 		return;
4192 
4193 	if ((vtag == 0) || (vtag > 4095))
4194 		return;
4195 
4196 	CTX_LOCK(ctx);
4197 	IFDI_VLAN_REGISTER(ctx, vtag);
4198 	/* Re-init to load the changes */
4199 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
4200 		iflib_if_init_locked(ctx);
4201 	CTX_UNLOCK(ctx);
4202 }
4203 
4204 static void
4205 iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag)
4206 {
4207 	if_ctx_t ctx = if_getsoftc(ifp);
4208 
4209 	if ((void *)ctx != arg)
4210 		return;
4211 
4212 	if ((vtag == 0) || (vtag > 4095))
4213 		return;
4214 
4215 	CTX_LOCK(ctx);
4216 	IFDI_VLAN_UNREGISTER(ctx, vtag);
4217 	/* Re-init to load the changes */
4218 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
4219 		iflib_if_init_locked(ctx);
4220 	CTX_UNLOCK(ctx);
4221 }
4222 
4223 static void
4224 iflib_led_func(void *arg, int onoff)
4225 {
4226 	if_ctx_t ctx = arg;
4227 
4228 	CTX_LOCK(ctx);
4229 	IFDI_LED_FUNC(ctx, onoff);
4230 	CTX_UNLOCK(ctx);
4231 }
4232 
4233 /*********************************************************************
4234  *
4235  *  BUS FUNCTION DEFINITIONS
4236  *
4237  **********************************************************************/
4238 
4239 int
4240 iflib_device_probe(device_t dev)
4241 {
4242 	pci_vendor_info_t *ent;
4243 
4244 	uint16_t	pci_vendor_id, pci_device_id;
4245 	uint16_t	pci_subvendor_id, pci_subdevice_id;
4246 	uint16_t	pci_rev_id;
4247 	if_shared_ctx_t sctx;
4248 
4249 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4250 		return (ENOTSUP);
4251 
4252 	pci_vendor_id = pci_get_vendor(dev);
4253 	pci_device_id = pci_get_device(dev);
4254 	pci_subvendor_id = pci_get_subvendor(dev);
4255 	pci_subdevice_id = pci_get_subdevice(dev);
4256 	pci_rev_id = pci_get_revid(dev);
4257 	if (sctx->isc_parse_devinfo != NULL)
4258 		sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id);
4259 
4260 	ent = sctx->isc_vendor_info;
4261 	while (ent->pvi_vendor_id != 0) {
4262 		if (pci_vendor_id != ent->pvi_vendor_id) {
4263 			ent++;
4264 			continue;
4265 		}
4266 		if ((pci_device_id == ent->pvi_device_id) &&
4267 		    ((pci_subvendor_id == ent->pvi_subvendor_id) ||
4268 		     (ent->pvi_subvendor_id == 0)) &&
4269 		    ((pci_subdevice_id == ent->pvi_subdevice_id) ||
4270 		     (ent->pvi_subdevice_id == 0)) &&
4271 		    ((pci_rev_id == ent->pvi_rev_id) ||
4272 		     (ent->pvi_rev_id == 0))) {
4273 
4274 			device_set_desc_copy(dev, ent->pvi_name);
4275 			/* this needs to be changed to zero if the bus probing code
4276 			 * ever stops re-probing on best match because the sctx
4277 			 * may have its values over written by register calls
4278 			 * in subsequent probes
4279 			 */
4280 			return (BUS_PROBE_DEFAULT);
4281 		}
4282 		ent++;
4283 	}
4284 	return (ENXIO);
4285 }
4286 
4287 static void
4288 iflib_reset_qvalues(if_ctx_t ctx)
4289 {
4290 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4291 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4292 	device_t dev = ctx->ifc_dev;
4293 	int i;
4294 
4295 	scctx->isc_txrx_budget_bytes_max = IFLIB_MAX_TX_BYTES;
4296 	scctx->isc_tx_qdepth = IFLIB_DEFAULT_TX_QDEPTH;
4297 	/*
4298 	 * XXX sanity check that ntxd & nrxd are a power of 2
4299 	 */
4300 	if (ctx->ifc_sysctl_ntxqs != 0)
4301 		scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs;
4302 	if (ctx->ifc_sysctl_nrxqs != 0)
4303 		scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs;
4304 
4305 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4306 		if (ctx->ifc_sysctl_ntxds[i] != 0)
4307 			scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i];
4308 		else
4309 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4310 	}
4311 
4312 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4313 		if (ctx->ifc_sysctl_nrxds[i] != 0)
4314 			scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i];
4315 		else
4316 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4317 	}
4318 
4319 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4320 		if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) {
4321 			device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n",
4322 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]);
4323 			scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i];
4324 		}
4325 		if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) {
4326 			device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n",
4327 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]);
4328 			scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i];
4329 		}
4330 	}
4331 
4332 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4333 		if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) {
4334 			device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n",
4335 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]);
4336 			scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i];
4337 		}
4338 		if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) {
4339 			device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n",
4340 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]);
4341 			scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i];
4342 		}
4343 	}
4344 }
4345 
4346 int
4347 iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp)
4348 {
4349 	int err, rid, msix;
4350 	if_ctx_t ctx;
4351 	if_t ifp;
4352 	if_softc_ctx_t scctx;
4353 	int i;
4354 	uint16_t main_txq;
4355 	uint16_t main_rxq;
4356 
4357 
4358 	ctx = malloc(sizeof(* ctx), M_IFLIB, M_WAITOK|M_ZERO);
4359 
4360 	if (sc == NULL) {
4361 		sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4362 		device_set_softc(dev, ctx);
4363 		ctx->ifc_flags |= IFC_SC_ALLOCATED;
4364 	}
4365 
4366 	ctx->ifc_sctx = sctx;
4367 	ctx->ifc_dev = dev;
4368 	ctx->ifc_softc = sc;
4369 
4370 	if ((err = iflib_register(ctx)) != 0) {
4371 		if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4372 			free(sc, M_IFLIB);
4373 		free(ctx, M_IFLIB);
4374 		device_printf(dev, "iflib_register failed %d\n", err);
4375 		return (err);
4376 	}
4377 	iflib_add_device_sysctl_pre(ctx);
4378 
4379 	scctx = &ctx->ifc_softc_ctx;
4380 	ifp = ctx->ifc_ifp;
4381 
4382 	iflib_reset_qvalues(ctx);
4383 	CTX_LOCK(ctx);
4384 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4385 		CTX_UNLOCK(ctx);
4386 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4387 		return (err);
4388 	}
4389 	_iflib_pre_assert(scctx);
4390 	ctx->ifc_txrx = *scctx->isc_txrx;
4391 
4392 #ifdef INVARIANTS
4393 	MPASS(scctx->isc_capabilities);
4394 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
4395 		MPASS(scctx->isc_tx_csum_flags);
4396 #endif
4397 
4398 	if_setcapabilities(ifp, scctx->isc_capabilities | IFCAP_HWSTATS);
4399 	if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS);
4400 
4401 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4402 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4403 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4404 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4405 
4406 #ifdef ACPI_DMAR
4407 	if (dmar_get_dma_tag(device_get_parent(dev), dev) != NULL)
4408 		ctx->ifc_flags |= IFC_DMAR;
4409 #elif !(defined(__i386__) || defined(__amd64__))
4410 	/* set unconditionally for !x86 */
4411 	ctx->ifc_flags |= IFC_DMAR;
4412 #endif
4413 
4414 	main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4415 	main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4416 
4417 	/* XXX change for per-queue sizes */
4418 	device_printf(dev, "using %d tx descriptors and %d rx descriptors\n",
4419 		      scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4420 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4421 		if (!powerof2(scctx->isc_nrxd[i])) {
4422 			/* round down instead? */
4423 			device_printf(dev, "# rx descriptors must be a power of 2\n");
4424 			err = EINVAL;
4425 			goto fail;
4426 		}
4427 	}
4428 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4429 		if (!powerof2(scctx->isc_ntxd[i])) {
4430 			device_printf(dev,
4431 			    "# tx descriptors must be a power of 2");
4432 			err = EINVAL;
4433 			goto fail;
4434 		}
4435 	}
4436 
4437 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4438 	    MAX_SINGLE_PACKET_FRACTION)
4439 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4440 		    MAX_SINGLE_PACKET_FRACTION);
4441 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4442 	    MAX_SINGLE_PACKET_FRACTION)
4443 		scctx->isc_tx_tso_segments_max = max(1,
4444 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4445 
4446 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4447 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
4448 		/*
4449 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
4450 		 * but some MACs do.
4451 		 */
4452 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
4453 		    IP_MAXPACKET));
4454 		/*
4455 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
4456 		 * into account.  In the worst case, each of these calls will
4457 		 * add another mbuf and, thus, the requirement for another DMA
4458 		 * segment.  So for best performance, it doesn't make sense to
4459 		 * advertize a maximum of TSO segments that typically will
4460 		 * require defragmentation in iflib_encap().
4461 		 */
4462 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
4463 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
4464 	}
4465 	if (scctx->isc_rss_table_size == 0)
4466 		scctx->isc_rss_table_size = 64;
4467 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4468 
4469 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4470 	/* XXX format name */
4471 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx, -1, "admin");
4472 
4473 	/* Set up cpu set.  If it fails, use the set of all CPUs. */
4474 	if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) != 0) {
4475 		device_printf(dev, "Unable to fetch CPU list\n");
4476 		CPU_COPY(&all_cpus, &ctx->ifc_cpus);
4477 	}
4478 	MPASS(CPU_COUNT(&ctx->ifc_cpus) > 0);
4479 
4480 	/*
4481 	** Now setup MSI or MSI/X, should
4482 	** return us the number of supported
4483 	** vectors. (Will be 1 for MSI)
4484 	*/
4485 	if (sctx->isc_flags & IFLIB_SKIP_MSIX) {
4486 		msix = scctx->isc_vectors;
4487 	} else if (scctx->isc_msix_bar != 0)
4488 	       /*
4489 		* The simple fact that isc_msix_bar is not 0 does not mean we
4490 		* we have a good value there that is known to work.
4491 		*/
4492 		msix = iflib_msix_init(ctx);
4493 	else {
4494 		scctx->isc_vectors = 1;
4495 		scctx->isc_ntxqsets = 1;
4496 		scctx->isc_nrxqsets = 1;
4497 		scctx->isc_intr = IFLIB_INTR_LEGACY;
4498 		msix = 0;
4499 	}
4500 	/* Get memory for the station queues */
4501 	if ((err = iflib_queues_alloc(ctx))) {
4502 		device_printf(dev, "Unable to allocate queue memory\n");
4503 		goto fail;
4504 	}
4505 
4506 	if ((err = iflib_qset_structures_setup(ctx)))
4507 		goto fail_queues;
4508 
4509 	/*
4510 	 * Group taskqueues aren't properly set up until SMP is started,
4511 	 * so we disable interrupts until we can handle them post
4512 	 * SI_SUB_SMP.
4513 	 *
4514 	 * XXX: disabling interrupts doesn't actually work, at least for
4515 	 * the non-MSI case.  When they occur before SI_SUB_SMP completes,
4516 	 * we do null handling and depend on this not causing too large an
4517 	 * interrupt storm.
4518 	 */
4519 	IFDI_INTR_DISABLE(ctx);
4520 	if (msix > 1 && (err = IFDI_MSIX_INTR_ASSIGN(ctx, msix)) != 0) {
4521 		device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n", err);
4522 		goto fail_intr_free;
4523 	}
4524 	if (msix <= 1) {
4525 		rid = 0;
4526 		if (scctx->isc_intr == IFLIB_INTR_MSI) {
4527 			MPASS(msix == 1);
4528 			rid = 1;
4529 		}
4530 		if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) {
4531 			device_printf(dev, "iflib_legacy_setup failed %d\n", err);
4532 			goto fail_intr_free;
4533 		}
4534 	}
4535 
4536 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac);
4537 
4538 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4539 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4540 		goto fail_detach;
4541 	}
4542 
4543 	/*
4544 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4545 	 * This must appear after the call to ether_ifattach() because
4546 	 * ether_ifattach() sets if_hdrlen to the default value.
4547 	 */
4548 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4549 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
4550 
4551 	if ((err = iflib_netmap_attach(ctx))) {
4552 		device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err);
4553 		goto fail_detach;
4554 	}
4555 	*ctxp = ctx;
4556 
4557 	NETDUMP_SET(ctx->ifc_ifp, iflib);
4558 
4559 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4560 	iflib_add_device_sysctl_post(ctx);
4561 	ctx->ifc_flags |= IFC_INIT_DONE;
4562 	CTX_UNLOCK(ctx);
4563 	return (0);
4564 fail_detach:
4565 	ether_ifdetach(ctx->ifc_ifp);
4566 fail_intr_free:
4567 	if (scctx->isc_intr == IFLIB_INTR_MSIX || scctx->isc_intr == IFLIB_INTR_MSI)
4568 		pci_release_msi(ctx->ifc_dev);
4569 fail_queues:
4570 	iflib_tx_structures_free(ctx);
4571 	iflib_rx_structures_free(ctx);
4572 fail:
4573 	IFDI_DETACH(ctx);
4574 	CTX_UNLOCK(ctx);
4575 	return (err);
4576 }
4577 
4578 int
4579 iflib_pseudo_register(device_t dev, if_shared_ctx_t sctx, if_ctx_t *ctxp,
4580 					  struct iflib_cloneattach_ctx *clctx)
4581 {
4582 	int err;
4583 	if_ctx_t ctx;
4584 	if_t ifp;
4585 	if_softc_ctx_t scctx;
4586 	int i;
4587 	void *sc;
4588 	uint16_t main_txq;
4589 	uint16_t main_rxq;
4590 
4591 	ctx = malloc(sizeof(*ctx), M_IFLIB, M_WAITOK|M_ZERO);
4592 	sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4593 	ctx->ifc_flags |= IFC_SC_ALLOCATED;
4594 	if (sctx->isc_flags & (IFLIB_PSEUDO|IFLIB_VIRTUAL))
4595 		ctx->ifc_flags |= IFC_PSEUDO;
4596 
4597 	ctx->ifc_sctx = sctx;
4598 	ctx->ifc_softc = sc;
4599 	ctx->ifc_dev = dev;
4600 
4601 	if ((err = iflib_register(ctx)) != 0) {
4602 		device_printf(dev, "%s: iflib_register failed %d\n", __func__, err);
4603 		free(sc, M_IFLIB);
4604 		free(ctx, M_IFLIB);
4605 		return (err);
4606 	}
4607 	iflib_add_device_sysctl_pre(ctx);
4608 
4609 	scctx = &ctx->ifc_softc_ctx;
4610 	ifp = ctx->ifc_ifp;
4611 
4612 	/*
4613 	 * XXX sanity check that ntxd & nrxd are a power of 2
4614 	 */
4615 	iflib_reset_qvalues(ctx);
4616 
4617 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4618 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4619 		return (err);
4620 	}
4621 	if (sctx->isc_flags & IFLIB_GEN_MAC)
4622 		iflib_gen_mac(ctx);
4623 	if ((err = IFDI_CLONEATTACH(ctx, clctx->cc_ifc, clctx->cc_name,
4624 								clctx->cc_params)) != 0) {
4625 		device_printf(dev, "IFDI_CLONEATTACH failed %d\n", err);
4626 		return (err);
4627 	}
4628 	ifmedia_add(&ctx->ifc_media, IFM_ETHER | IFM_1000_T | IFM_FDX, 0, NULL);
4629 	ifmedia_add(&ctx->ifc_media, IFM_ETHER | IFM_AUTO, 0, NULL);
4630 	ifmedia_set(&ctx->ifc_media, IFM_ETHER | IFM_AUTO);
4631 
4632 #ifdef INVARIANTS
4633 	MPASS(scctx->isc_capabilities);
4634 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
4635 		MPASS(scctx->isc_tx_csum_flags);
4636 #endif
4637 
4638 	if_setcapabilities(ifp, scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_LINKSTATE);
4639 	if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_LINKSTATE);
4640 
4641 	ifp->if_flags |= IFF_NOGROUP;
4642 	if (sctx->isc_flags & IFLIB_PSEUDO) {
4643 		ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac);
4644 
4645 		if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4646 			device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4647 			goto fail_detach;
4648 		}
4649 		*ctxp = ctx;
4650 
4651 		/*
4652 		 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4653 		 * This must appear after the call to ether_ifattach() because
4654 		 * ether_ifattach() sets if_hdrlen to the default value.
4655 		 */
4656 		if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4657 			if_setifheaderlen(ifp,
4658 			    sizeof(struct ether_vlan_header));
4659 
4660 		if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4661 		iflib_add_device_sysctl_post(ctx);
4662 		ctx->ifc_flags |= IFC_INIT_DONE;
4663 		return (0);
4664 	}
4665 	_iflib_pre_assert(scctx);
4666 	ctx->ifc_txrx = *scctx->isc_txrx;
4667 
4668 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4669 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4670 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4671 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4672 
4673 	main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4674 	main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4675 
4676 	/* XXX change for per-queue sizes */
4677 	device_printf(dev, "using %d tx descriptors and %d rx descriptors\n",
4678 		      scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4679 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4680 		if (!powerof2(scctx->isc_nrxd[i])) {
4681 			/* round down instead? */
4682 			device_printf(dev, "# rx descriptors must be a power of 2\n");
4683 			err = EINVAL;
4684 			goto fail;
4685 		}
4686 	}
4687 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4688 		if (!powerof2(scctx->isc_ntxd[i])) {
4689 			device_printf(dev,
4690 			    "# tx descriptors must be a power of 2");
4691 			err = EINVAL;
4692 			goto fail;
4693 		}
4694 	}
4695 
4696 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4697 	    MAX_SINGLE_PACKET_FRACTION)
4698 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4699 		    MAX_SINGLE_PACKET_FRACTION);
4700 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4701 	    MAX_SINGLE_PACKET_FRACTION)
4702 		scctx->isc_tx_tso_segments_max = max(1,
4703 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4704 
4705 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4706 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
4707 		/*
4708 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
4709 		 * but some MACs do.
4710 		 */
4711 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
4712 		    IP_MAXPACKET));
4713 		/*
4714 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
4715 		 * into account.  In the worst case, each of these calls will
4716 		 * add another mbuf and, thus, the requirement for another DMA
4717 		 * segment.  So for best performance, it doesn't make sense to
4718 		 * advertize a maximum of TSO segments that typically will
4719 		 * require defragmentation in iflib_encap().
4720 		 */
4721 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
4722 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
4723 	}
4724 	if (scctx->isc_rss_table_size == 0)
4725 		scctx->isc_rss_table_size = 64;
4726 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4727 
4728 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4729 	/* XXX format name */
4730 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx, -1, "admin");
4731 
4732 	/* XXX --- can support > 1 -- but keep it simple for now */
4733 	scctx->isc_intr = IFLIB_INTR_LEGACY;
4734 
4735 	/* Get memory for the station queues */
4736 	if ((err = iflib_queues_alloc(ctx))) {
4737 		device_printf(dev, "Unable to allocate queue memory\n");
4738 		goto fail;
4739 	}
4740 
4741 	if ((err = iflib_qset_structures_setup(ctx))) {
4742 		device_printf(dev, "qset structure setup failed %d\n", err);
4743 		goto fail_queues;
4744 	}
4745 
4746 	/*
4747 	 * XXX What if anything do we want to do about interrupts?
4748 	 */
4749 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac);
4750 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4751 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4752 		goto fail_detach;
4753 	}
4754 
4755 	/*
4756 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4757 	 * This must appear after the call to ether_ifattach() because
4758 	 * ether_ifattach() sets if_hdrlen to the default value.
4759 	 */
4760 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4761 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
4762 
4763 	/* XXX handle more than one queue */
4764 	for (i = 0; i < scctx->isc_nrxqsets; i++)
4765 		IFDI_RX_CLSET(ctx, 0, i, ctx->ifc_rxqs[i].ifr_fl[0].ifl_sds.ifsd_cl);
4766 
4767 	*ctxp = ctx;
4768 
4769 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4770 	iflib_add_device_sysctl_post(ctx);
4771 	ctx->ifc_flags |= IFC_INIT_DONE;
4772 	return (0);
4773 fail_detach:
4774 	ether_ifdetach(ctx->ifc_ifp);
4775 fail_queues:
4776 	iflib_tx_structures_free(ctx);
4777 	iflib_rx_structures_free(ctx);
4778 fail:
4779 	IFDI_DETACH(ctx);
4780 	return (err);
4781 }
4782 
4783 int
4784 iflib_pseudo_deregister(if_ctx_t ctx)
4785 {
4786 	if_t ifp = ctx->ifc_ifp;
4787 	iflib_txq_t txq;
4788 	iflib_rxq_t rxq;
4789 	int i, j;
4790 	struct taskqgroup *tqg;
4791 	iflib_fl_t fl;
4792 
4793 	/* Unregister VLAN events */
4794 	if (ctx->ifc_vlan_attach_event != NULL)
4795 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
4796 	if (ctx->ifc_vlan_detach_event != NULL)
4797 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
4798 
4799 	ether_ifdetach(ifp);
4800 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
4801 	CTX_LOCK_DESTROY(ctx);
4802 	/* XXX drain any dependent tasks */
4803 	tqg = qgroup_if_io_tqg;
4804 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
4805 		callout_drain(&txq->ift_timer);
4806 		if (txq->ift_task.gt_uniq != NULL)
4807 			taskqgroup_detach(tqg, &txq->ift_task);
4808 	}
4809 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4810 		if (rxq->ifr_task.gt_uniq != NULL)
4811 			taskqgroup_detach(tqg, &rxq->ifr_task);
4812 
4813 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
4814 			free(fl->ifl_rx_bitmap, M_IFLIB);
4815 	}
4816 	tqg = qgroup_if_config_tqg;
4817 	if (ctx->ifc_admin_task.gt_uniq != NULL)
4818 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
4819 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
4820 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
4821 
4822 	if_free(ifp);
4823 
4824 	iflib_tx_structures_free(ctx);
4825 	iflib_rx_structures_free(ctx);
4826 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4827 		free(ctx->ifc_softc, M_IFLIB);
4828 	free(ctx, M_IFLIB);
4829 	return (0);
4830 }
4831 
4832 int
4833 iflib_device_attach(device_t dev)
4834 {
4835 	if_ctx_t ctx;
4836 	if_shared_ctx_t sctx;
4837 
4838 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4839 		return (ENOTSUP);
4840 
4841 	pci_enable_busmaster(dev);
4842 
4843 	return (iflib_device_register(dev, NULL, sctx, &ctx));
4844 }
4845 
4846 int
4847 iflib_device_deregister(if_ctx_t ctx)
4848 {
4849 	if_t ifp = ctx->ifc_ifp;
4850 	iflib_txq_t txq;
4851 	iflib_rxq_t rxq;
4852 	device_t dev = ctx->ifc_dev;
4853 	int i, j;
4854 	struct taskqgroup *tqg;
4855 	iflib_fl_t fl;
4856 
4857 	/* Make sure VLANS are not using driver */
4858 	if (if_vlantrunkinuse(ifp)) {
4859 		device_printf(dev,"Vlan in use, detach first\n");
4860 		return (EBUSY);
4861 	}
4862 
4863 	CTX_LOCK(ctx);
4864 	ctx->ifc_in_detach = 1;
4865 	iflib_stop(ctx);
4866 	CTX_UNLOCK(ctx);
4867 
4868 	/* Unregister VLAN events */
4869 	if (ctx->ifc_vlan_attach_event != NULL)
4870 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
4871 	if (ctx->ifc_vlan_detach_event != NULL)
4872 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
4873 
4874 	iflib_netmap_detach(ifp);
4875 	ether_ifdetach(ifp);
4876 	if (ctx->ifc_led_dev != NULL)
4877 		led_destroy(ctx->ifc_led_dev);
4878 	/* XXX drain any dependent tasks */
4879 	tqg = qgroup_if_io_tqg;
4880 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
4881 		callout_drain(&txq->ift_timer);
4882 		if (txq->ift_task.gt_uniq != NULL)
4883 			taskqgroup_detach(tqg, &txq->ift_task);
4884 	}
4885 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4886 		if (rxq->ifr_task.gt_uniq != NULL)
4887 			taskqgroup_detach(tqg, &rxq->ifr_task);
4888 
4889 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
4890 			free(fl->ifl_rx_bitmap, M_IFLIB);
4891 
4892 	}
4893 	tqg = qgroup_if_config_tqg;
4894 	if (ctx->ifc_admin_task.gt_uniq != NULL)
4895 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
4896 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
4897 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
4898 	CTX_LOCK(ctx);
4899 	IFDI_DETACH(ctx);
4900 	CTX_UNLOCK(ctx);
4901 
4902 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
4903 	CTX_LOCK_DESTROY(ctx);
4904 	device_set_softc(ctx->ifc_dev, NULL);
4905 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) {
4906 		pci_release_msi(dev);
4907 	}
4908 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) {
4909 		iflib_irq_free(ctx, &ctx->ifc_legacy_irq);
4910 	}
4911 	if (ctx->ifc_msix_mem != NULL) {
4912 		bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY,
4913 			ctx->ifc_softc_ctx.isc_msix_bar, ctx->ifc_msix_mem);
4914 		ctx->ifc_msix_mem = NULL;
4915 	}
4916 
4917 	bus_generic_detach(dev);
4918 	if_free(ifp);
4919 
4920 	iflib_tx_structures_free(ctx);
4921 	iflib_rx_structures_free(ctx);
4922 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4923 		free(ctx->ifc_softc, M_IFLIB);
4924 	free(ctx, M_IFLIB);
4925 	return (0);
4926 }
4927 
4928 
4929 int
4930 iflib_device_detach(device_t dev)
4931 {
4932 	if_ctx_t ctx = device_get_softc(dev);
4933 
4934 	return (iflib_device_deregister(ctx));
4935 }
4936 
4937 int
4938 iflib_device_suspend(device_t dev)
4939 {
4940 	if_ctx_t ctx = device_get_softc(dev);
4941 
4942 	CTX_LOCK(ctx);
4943 	IFDI_SUSPEND(ctx);
4944 	CTX_UNLOCK(ctx);
4945 
4946 	return bus_generic_suspend(dev);
4947 }
4948 int
4949 iflib_device_shutdown(device_t dev)
4950 {
4951 	if_ctx_t ctx = device_get_softc(dev);
4952 
4953 	CTX_LOCK(ctx);
4954 	IFDI_SHUTDOWN(ctx);
4955 	CTX_UNLOCK(ctx);
4956 
4957 	return bus_generic_suspend(dev);
4958 }
4959 
4960 
4961 int
4962 iflib_device_resume(device_t dev)
4963 {
4964 	if_ctx_t ctx = device_get_softc(dev);
4965 	iflib_txq_t txq = ctx->ifc_txqs;
4966 
4967 	CTX_LOCK(ctx);
4968 	IFDI_RESUME(ctx);
4969 	iflib_init_locked(ctx);
4970 	CTX_UNLOCK(ctx);
4971 	for (int i = 0; i < NTXQSETS(ctx); i++, txq++)
4972 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
4973 
4974 	return (bus_generic_resume(dev));
4975 }
4976 
4977 int
4978 iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params)
4979 {
4980 	int error;
4981 	if_ctx_t ctx = device_get_softc(dev);
4982 
4983 	CTX_LOCK(ctx);
4984 	error = IFDI_IOV_INIT(ctx, num_vfs, params);
4985 	CTX_UNLOCK(ctx);
4986 
4987 	return (error);
4988 }
4989 
4990 void
4991 iflib_device_iov_uninit(device_t dev)
4992 {
4993 	if_ctx_t ctx = device_get_softc(dev);
4994 
4995 	CTX_LOCK(ctx);
4996 	IFDI_IOV_UNINIT(ctx);
4997 	CTX_UNLOCK(ctx);
4998 }
4999 
5000 int
5001 iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params)
5002 {
5003 	int error;
5004 	if_ctx_t ctx = device_get_softc(dev);
5005 
5006 	CTX_LOCK(ctx);
5007 	error = IFDI_IOV_VF_ADD(ctx, vfnum, params);
5008 	CTX_UNLOCK(ctx);
5009 
5010 	return (error);
5011 }
5012 
5013 /*********************************************************************
5014  *
5015  *  MODULE FUNCTION DEFINITIONS
5016  *
5017  **********************************************************************/
5018 
5019 /*
5020  * - Start a fast taskqueue thread for each core
5021  * - Start a taskqueue for control operations
5022  */
5023 static int
5024 iflib_module_init(void)
5025 {
5026 	return (0);
5027 }
5028 
5029 static int
5030 iflib_module_event_handler(module_t mod, int what, void *arg)
5031 {
5032 	int err;
5033 
5034 	switch (what) {
5035 	case MOD_LOAD:
5036 		if ((err = iflib_module_init()) != 0)
5037 			return (err);
5038 		break;
5039 	case MOD_UNLOAD:
5040 		return (EBUSY);
5041 	default:
5042 		return (EOPNOTSUPP);
5043 	}
5044 
5045 	return (0);
5046 }
5047 
5048 /*********************************************************************
5049  *
5050  *  PUBLIC FUNCTION DEFINITIONS
5051  *     ordered as in iflib.h
5052  *
5053  **********************************************************************/
5054 
5055 
5056 static void
5057 _iflib_assert(if_shared_ctx_t sctx)
5058 {
5059 	MPASS(sctx->isc_tx_maxsize);
5060 	MPASS(sctx->isc_tx_maxsegsize);
5061 
5062 	MPASS(sctx->isc_rx_maxsize);
5063 	MPASS(sctx->isc_rx_nsegments);
5064 	MPASS(sctx->isc_rx_maxsegsize);
5065 
5066 	MPASS(sctx->isc_nrxd_min[0]);
5067 	MPASS(sctx->isc_nrxd_max[0]);
5068 	MPASS(sctx->isc_nrxd_default[0]);
5069 	MPASS(sctx->isc_ntxd_min[0]);
5070 	MPASS(sctx->isc_ntxd_max[0]);
5071 	MPASS(sctx->isc_ntxd_default[0]);
5072 }
5073 
5074 static void
5075 _iflib_pre_assert(if_softc_ctx_t scctx)
5076 {
5077 
5078 	MPASS(scctx->isc_txrx->ift_txd_encap);
5079 	MPASS(scctx->isc_txrx->ift_txd_flush);
5080 	MPASS(scctx->isc_txrx->ift_txd_credits_update);
5081 	MPASS(scctx->isc_txrx->ift_rxd_available);
5082 	MPASS(scctx->isc_txrx->ift_rxd_pkt_get);
5083 	MPASS(scctx->isc_txrx->ift_rxd_refill);
5084 	MPASS(scctx->isc_txrx->ift_rxd_flush);
5085 }
5086 
5087 static int
5088 iflib_register(if_ctx_t ctx)
5089 {
5090 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5091 	driver_t *driver = sctx->isc_driver;
5092 	device_t dev = ctx->ifc_dev;
5093 	if_t ifp;
5094 
5095 	_iflib_assert(sctx);
5096 
5097 	CTX_LOCK_INIT(ctx);
5098 	STATE_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev));
5099 	ifp = ctx->ifc_ifp = if_gethandle(IFT_ETHER);
5100 	if (ifp == NULL) {
5101 		device_printf(dev, "can not allocate ifnet structure\n");
5102 		return (ENOMEM);
5103 	}
5104 
5105 	/*
5106 	 * Initialize our context's device specific methods
5107 	 */
5108 	kobj_init((kobj_t) ctx, (kobj_class_t) driver);
5109 	kobj_class_compile((kobj_class_t) driver);
5110 	driver->refs++;
5111 
5112 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
5113 	if_setsoftc(ifp, ctx);
5114 	if_setdev(ifp, dev);
5115 	if_setinitfn(ifp, iflib_if_init);
5116 	if_setioctlfn(ifp, iflib_if_ioctl);
5117 	if_settransmitfn(ifp, iflib_if_transmit);
5118 	if_setqflushfn(ifp, iflib_if_qflush);
5119 	if_setflags(ifp, IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST);
5120 
5121 	ctx->ifc_vlan_attach_event =
5122 		EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx,
5123 							  EVENTHANDLER_PRI_FIRST);
5124 	ctx->ifc_vlan_detach_event =
5125 		EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx,
5126 							  EVENTHANDLER_PRI_FIRST);
5127 
5128 	ifmedia_init(&ctx->ifc_media, IFM_IMASK,
5129 					 iflib_media_change, iflib_media_status);
5130 
5131 	return (0);
5132 }
5133 
5134 
5135 static int
5136 iflib_queues_alloc(if_ctx_t ctx)
5137 {
5138 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5139 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5140 	device_t dev = ctx->ifc_dev;
5141 	int nrxqsets = scctx->isc_nrxqsets;
5142 	int ntxqsets = scctx->isc_ntxqsets;
5143 	iflib_txq_t txq;
5144 	iflib_rxq_t rxq;
5145 	iflib_fl_t fl = NULL;
5146 	int i, j, cpu, err, txconf, rxconf;
5147 	iflib_dma_info_t ifdip;
5148 	uint32_t *rxqsizes = scctx->isc_rxqsizes;
5149 	uint32_t *txqsizes = scctx->isc_txqsizes;
5150 	uint8_t nrxqs = sctx->isc_nrxqs;
5151 	uint8_t ntxqs = sctx->isc_ntxqs;
5152 	int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1;
5153 	caddr_t *vaddrs;
5154 	uint64_t *paddrs;
5155 
5156 	KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1"));
5157 	KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1"));
5158 
5159 	/* Allocate the TX ring struct memory */
5160 	if (!(ctx->ifc_txqs =
5161 	    (iflib_txq_t) malloc(sizeof(struct iflib_txq) *
5162 	    ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5163 		device_printf(dev, "Unable to allocate TX ring memory\n");
5164 		err = ENOMEM;
5165 		goto fail;
5166 	}
5167 
5168 	/* Now allocate the RX */
5169 	if (!(ctx->ifc_rxqs =
5170 	    (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) *
5171 	    nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5172 		device_printf(dev, "Unable to allocate RX ring memory\n");
5173 		err = ENOMEM;
5174 		goto rx_fail;
5175 	}
5176 
5177 	txq = ctx->ifc_txqs;
5178 	rxq = ctx->ifc_rxqs;
5179 
5180 	/*
5181 	 * XXX handle allocation failure
5182 	 */
5183 	for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) {
5184 		/* Set up some basics */
5185 
5186 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs, M_IFLIB, M_WAITOK|M_ZERO)) == NULL) {
5187 			device_printf(dev, "failed to allocate iflib_dma_info\n");
5188 			err = ENOMEM;
5189 			goto err_tx_desc;
5190 		}
5191 		txq->ift_ifdi = ifdip;
5192 		for (j = 0; j < ntxqs; j++, ifdip++) {
5193 			if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, BUS_DMA_NOWAIT)) {
5194 				device_printf(dev, "Unable to allocate Descriptor memory\n");
5195 				err = ENOMEM;
5196 				goto err_tx_desc;
5197 			}
5198 			txq->ift_txd_size[j] = scctx->isc_txd_size[j];
5199 			bzero((void *)ifdip->idi_vaddr, txqsizes[j]);
5200 		}
5201 		txq->ift_ctx = ctx;
5202 		txq->ift_id = i;
5203 		if (sctx->isc_flags & IFLIB_HAS_TXCQ) {
5204 			txq->ift_br_offset = 1;
5205 		} else {
5206 			txq->ift_br_offset = 0;
5207 		}
5208 		/* XXX fix this */
5209 		txq->ift_timer.c_cpu = cpu;
5210 
5211 		if (iflib_txsd_alloc(txq)) {
5212 			device_printf(dev, "Critical Failure setting up TX buffers\n");
5213 			err = ENOMEM;
5214 			goto err_tx_desc;
5215 		}
5216 
5217 		/* Initialize the TX lock */
5218 		snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:tx(%d):callout",
5219 		    device_get_nameunit(dev), txq->ift_id);
5220 		mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF);
5221 		callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0);
5222 
5223 		snprintf(txq->ift_db_mtx_name, MTX_NAME_LEN, "%s:tx(%d):db",
5224 			 device_get_nameunit(dev), txq->ift_id);
5225 
5226 		err = ifmp_ring_alloc(&txq->ift_br, 2048, txq, iflib_txq_drain,
5227 				      iflib_txq_can_drain, M_IFLIB, M_WAITOK);
5228 		if (err) {
5229 			/* XXX free any allocated rings */
5230 			device_printf(dev, "Unable to allocate buf_ring\n");
5231 			goto err_tx_desc;
5232 		}
5233 	}
5234 
5235 	for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) {
5236 		/* Set up some basics */
5237 
5238 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs, M_IFLIB, M_WAITOK|M_ZERO)) == NULL) {
5239 			device_printf(dev, "failed to allocate iflib_dma_info\n");
5240 			err = ENOMEM;
5241 			goto err_tx_desc;
5242 		}
5243 
5244 		rxq->ifr_ifdi = ifdip;
5245 		/* XXX this needs to be changed if #rx queues != #tx queues */
5246 		rxq->ifr_ntxqirq = 1;
5247 		rxq->ifr_txqid[0] = i;
5248 		for (j = 0; j < nrxqs; j++, ifdip++) {
5249 			if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, BUS_DMA_NOWAIT)) {
5250 				device_printf(dev, "Unable to allocate Descriptor memory\n");
5251 				err = ENOMEM;
5252 				goto err_tx_desc;
5253 			}
5254 			bzero((void *)ifdip->idi_vaddr, rxqsizes[j]);
5255 		}
5256 		rxq->ifr_ctx = ctx;
5257 		rxq->ifr_id = i;
5258 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
5259 			rxq->ifr_fl_offset = 1;
5260 		} else {
5261 			rxq->ifr_fl_offset = 0;
5262 		}
5263 		rxq->ifr_nfl = nfree_lists;
5264 		if (!(fl =
5265 			  (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) {
5266 			device_printf(dev, "Unable to allocate free list memory\n");
5267 			err = ENOMEM;
5268 			goto err_tx_desc;
5269 		}
5270 		rxq->ifr_fl = fl;
5271 		for (j = 0; j < nfree_lists; j++) {
5272 			fl[j].ifl_rxq = rxq;
5273 			fl[j].ifl_id = j;
5274 			fl[j].ifl_ifdi = &rxq->ifr_ifdi[j + rxq->ifr_fl_offset];
5275 			fl[j].ifl_rxd_size = scctx->isc_rxd_size[j];
5276 		}
5277         /* Allocate receive buffers for the ring*/
5278 		if (iflib_rxsd_alloc(rxq)) {
5279 			device_printf(dev,
5280 			    "Critical Failure setting up receive buffers\n");
5281 			err = ENOMEM;
5282 			goto err_rx_desc;
5283 		}
5284 
5285 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5286 			fl->ifl_rx_bitmap = bit_alloc(fl->ifl_size, M_IFLIB, M_WAITOK|M_ZERO);
5287 	}
5288 
5289 	/* TXQs */
5290 	vaddrs = malloc(sizeof(caddr_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5291 	paddrs = malloc(sizeof(uint64_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5292 	for (i = 0; i < ntxqsets; i++) {
5293 		iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi;
5294 
5295 		for (j = 0; j < ntxqs; j++, di++) {
5296 			vaddrs[i*ntxqs + j] = di->idi_vaddr;
5297 			paddrs[i*ntxqs + j] = di->idi_paddr;
5298 		}
5299 	}
5300 	if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) {
5301 		device_printf(ctx->ifc_dev, "device queue allocation failed\n");
5302 		iflib_tx_structures_free(ctx);
5303 		free(vaddrs, M_IFLIB);
5304 		free(paddrs, M_IFLIB);
5305 		goto err_rx_desc;
5306 	}
5307 	free(vaddrs, M_IFLIB);
5308 	free(paddrs, M_IFLIB);
5309 
5310 	/* RXQs */
5311 	vaddrs = malloc(sizeof(caddr_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5312 	paddrs = malloc(sizeof(uint64_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5313 	for (i = 0; i < nrxqsets; i++) {
5314 		iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi;
5315 
5316 		for (j = 0; j < nrxqs; j++, di++) {
5317 			vaddrs[i*nrxqs + j] = di->idi_vaddr;
5318 			paddrs[i*nrxqs + j] = di->idi_paddr;
5319 		}
5320 	}
5321 	if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) {
5322 		device_printf(ctx->ifc_dev, "device queue allocation failed\n");
5323 		iflib_tx_structures_free(ctx);
5324 		free(vaddrs, M_IFLIB);
5325 		free(paddrs, M_IFLIB);
5326 		goto err_rx_desc;
5327 	}
5328 	free(vaddrs, M_IFLIB);
5329 	free(paddrs, M_IFLIB);
5330 
5331 	return (0);
5332 
5333 /* XXX handle allocation failure changes */
5334 err_rx_desc:
5335 err_tx_desc:
5336 rx_fail:
5337 	if (ctx->ifc_rxqs != NULL)
5338 		free(ctx->ifc_rxqs, M_IFLIB);
5339 	ctx->ifc_rxqs = NULL;
5340 	if (ctx->ifc_txqs != NULL)
5341 		free(ctx->ifc_txqs, M_IFLIB);
5342 	ctx->ifc_txqs = NULL;
5343 fail:
5344 	return (err);
5345 }
5346 
5347 static int
5348 iflib_tx_structures_setup(if_ctx_t ctx)
5349 {
5350 	iflib_txq_t txq = ctx->ifc_txqs;
5351 	int i;
5352 
5353 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
5354 		iflib_txq_setup(txq);
5355 
5356 	return (0);
5357 }
5358 
5359 static void
5360 iflib_tx_structures_free(if_ctx_t ctx)
5361 {
5362 	iflib_txq_t txq = ctx->ifc_txqs;
5363 	int i, j;
5364 
5365 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
5366 		iflib_txq_destroy(txq);
5367 		for (j = 0; j < ctx->ifc_nhwtxqs; j++)
5368 			iflib_dma_free(&txq->ift_ifdi[j]);
5369 	}
5370 	free(ctx->ifc_txqs, M_IFLIB);
5371 	ctx->ifc_txqs = NULL;
5372 	IFDI_QUEUES_FREE(ctx);
5373 }
5374 
5375 /*********************************************************************
5376  *
5377  *  Initialize all receive rings.
5378  *
5379  **********************************************************************/
5380 static int
5381 iflib_rx_structures_setup(if_ctx_t ctx)
5382 {
5383 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5384 	int q;
5385 #if defined(INET6) || defined(INET)
5386 	int i, err;
5387 #endif
5388 
5389 	for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) {
5390 #if defined(INET6) || defined(INET)
5391 		tcp_lro_free(&rxq->ifr_lc);
5392 		if ((err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp,
5393 		    TCP_LRO_ENTRIES, min(1024,
5394 		    ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset]))) != 0) {
5395 			device_printf(ctx->ifc_dev, "LRO Initialization failed!\n");
5396 			goto fail;
5397 		}
5398 		rxq->ifr_lro_enabled = TRUE;
5399 #endif
5400 		IFDI_RXQ_SETUP(ctx, rxq->ifr_id);
5401 	}
5402 	return (0);
5403 #if defined(INET6) || defined(INET)
5404 fail:
5405 	/*
5406 	 * Free RX software descriptors allocated so far, we will only handle
5407 	 * the rings that completed, the failing case will have
5408 	 * cleaned up for itself. 'q' failed, so its the terminus.
5409 	 */
5410 	rxq = ctx->ifc_rxqs;
5411 	for (i = 0; i < q; ++i, rxq++) {
5412 		iflib_rx_sds_free(rxq);
5413 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
5414 	}
5415 	return (err);
5416 #endif
5417 }
5418 
5419 /*********************************************************************
5420  *
5421  *  Free all receive rings.
5422  *
5423  **********************************************************************/
5424 static void
5425 iflib_rx_structures_free(if_ctx_t ctx)
5426 {
5427 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5428 
5429 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) {
5430 		iflib_rx_sds_free(rxq);
5431 	}
5432 }
5433 
5434 static int
5435 iflib_qset_structures_setup(if_ctx_t ctx)
5436 {
5437 	int err;
5438 
5439 	/*
5440 	 * It is expected that the caller takes care of freeing queues if this
5441 	 * fails.
5442 	 */
5443 	if ((err = iflib_tx_structures_setup(ctx)) != 0) {
5444 		device_printf(ctx->ifc_dev, "iflib_tx_structures_setup failed: %d\n", err);
5445 		return (err);
5446 	}
5447 
5448 	if ((err = iflib_rx_structures_setup(ctx)) != 0)
5449 		device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err);
5450 
5451 	return (err);
5452 }
5453 
5454 int
5455 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
5456 		driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, const char *name)
5457 {
5458 
5459 	return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name));
5460 }
5461 
5462 #ifdef SMP
5463 static int
5464 find_nth(if_ctx_t ctx, int qid)
5465 {
5466 	cpuset_t cpus;
5467 	int i, cpuid, eqid, count;
5468 
5469 	CPU_COPY(&ctx->ifc_cpus, &cpus);
5470 	count = CPU_COUNT(&cpus);
5471 	eqid = qid % count;
5472 	/* clear up to the qid'th bit */
5473 	for (i = 0; i < eqid; i++) {
5474 		cpuid = CPU_FFS(&cpus);
5475 		MPASS(cpuid != 0);
5476 		CPU_CLR(cpuid-1, &cpus);
5477 	}
5478 	cpuid = CPU_FFS(&cpus);
5479 	MPASS(cpuid != 0);
5480 	return (cpuid-1);
5481 }
5482 
5483 #ifdef SCHED_ULE
5484 extern struct cpu_group *cpu_top;              /* CPU topology */
5485 
5486 static int
5487 find_child_with_core(int cpu, struct cpu_group *grp)
5488 {
5489 	int i;
5490 
5491 	if (grp->cg_children == 0)
5492 		return -1;
5493 
5494 	MPASS(grp->cg_child);
5495 	for (i = 0; i < grp->cg_children; i++) {
5496 		if (CPU_ISSET(cpu, &grp->cg_child[i].cg_mask))
5497 			return i;
5498 	}
5499 
5500 	return -1;
5501 }
5502 
5503 /*
5504  * Find the nth "close" core to the specified core
5505  * "close" is defined as the deepest level that shares
5506  * at least an L2 cache.  With threads, this will be
5507  * threads on the same core.  If the sahred cache is L3
5508  * or higher, simply returns the same core.
5509  */
5510 static int
5511 find_close_core(int cpu, int core_offset)
5512 {
5513 	struct cpu_group *grp;
5514 	int i;
5515 	int fcpu;
5516 	cpuset_t cs;
5517 
5518 	grp = cpu_top;
5519 	if (grp == NULL)
5520 		return cpu;
5521 	i = 0;
5522 	while ((i = find_child_with_core(cpu, grp)) != -1) {
5523 		/* If the child only has one cpu, don't descend */
5524 		if (grp->cg_child[i].cg_count <= 1)
5525 			break;
5526 		grp = &grp->cg_child[i];
5527 	}
5528 
5529 	/* If they don't share at least an L2 cache, use the same CPU */
5530 	if (grp->cg_level > CG_SHARE_L2 || grp->cg_level == CG_SHARE_NONE)
5531 		return cpu;
5532 
5533 	/* Now pick one */
5534 	CPU_COPY(&grp->cg_mask, &cs);
5535 
5536 	/* Add the selected CPU offset to core offset. */
5537 	for (i = 0; (fcpu = CPU_FFS(&cs)) != 0; i++) {
5538 		if (fcpu - 1 == cpu)
5539 			break;
5540 		CPU_CLR(fcpu - 1, &cs);
5541 	}
5542 	MPASS(fcpu);
5543 
5544 	core_offset += i;
5545 
5546 	CPU_COPY(&grp->cg_mask, &cs);
5547 	for (i = core_offset % grp->cg_count; i > 0; i--) {
5548 		MPASS(CPU_FFS(&cs));
5549 		CPU_CLR(CPU_FFS(&cs) - 1, &cs);
5550 	}
5551 	MPASS(CPU_FFS(&cs));
5552 	return CPU_FFS(&cs) - 1;
5553 }
5554 #else
5555 static int
5556 find_close_core(int cpu, int core_offset __unused)
5557 {
5558 	return cpu;
5559 }
5560 #endif
5561 
5562 static int
5563 get_core_offset(if_ctx_t ctx, iflib_intr_type_t type, int qid)
5564 {
5565 	switch (type) {
5566 	case IFLIB_INTR_TX:
5567 		/* TX queues get cores which share at least an L2 cache with the corresponding RX queue */
5568 		/* XXX handle multiple RX threads per core and more than two core per L2 group */
5569 		return qid / CPU_COUNT(&ctx->ifc_cpus) + 1;
5570 	case IFLIB_INTR_RX:
5571 	case IFLIB_INTR_RXTX:
5572 		/* RX queues get the specified core */
5573 		return qid / CPU_COUNT(&ctx->ifc_cpus);
5574 	default:
5575 		return -1;
5576 	}
5577 }
5578 #else
5579 #define get_core_offset(ctx, type, qid)	CPU_FIRST()
5580 #define find_close_core(cpuid, tid)	CPU_FIRST()
5581 #define find_nth(ctx, gid)		CPU_FIRST()
5582 #endif
5583 
5584 /* Just to avoid copy/paste */
5585 static inline int
5586 iflib_irq_set_affinity(if_ctx_t ctx, int irq, iflib_intr_type_t type, int qid,
5587     struct grouptask *gtask, struct taskqgroup *tqg, void *uniq, const char *name)
5588 {
5589 	int cpuid;
5590 	int err, tid;
5591 
5592 	cpuid = find_nth(ctx, qid);
5593 	tid = get_core_offset(ctx, type, qid);
5594 	MPASS(tid >= 0);
5595 	cpuid = find_close_core(cpuid, tid);
5596 	err = taskqgroup_attach_cpu(tqg, gtask, uniq, cpuid, irq, name);
5597 	if (err) {
5598 		device_printf(ctx->ifc_dev, "taskqgroup_attach_cpu failed %d\n", err);
5599 		return (err);
5600 	}
5601 #ifdef notyet
5602 	if (cpuid > ctx->ifc_cpuid_highest)
5603 		ctx->ifc_cpuid_highest = cpuid;
5604 #endif
5605 	return 0;
5606 }
5607 
5608 int
5609 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid,
5610 			iflib_intr_type_t type, driver_filter_t *filter,
5611 			void *filter_arg, int qid, const char *name)
5612 {
5613 	struct grouptask *gtask;
5614 	struct taskqgroup *tqg;
5615 	iflib_filter_info_t info;
5616 	gtask_fn_t *fn;
5617 	int tqrid, err;
5618 	driver_filter_t *intr_fast;
5619 	void *q;
5620 
5621 	info = &ctx->ifc_filter_info;
5622 	tqrid = rid;
5623 
5624 	switch (type) {
5625 	/* XXX merge tx/rx for netmap? */
5626 	case IFLIB_INTR_TX:
5627 		q = &ctx->ifc_txqs[qid];
5628 		info = &ctx->ifc_txqs[qid].ift_filter_info;
5629 		gtask = &ctx->ifc_txqs[qid].ift_task;
5630 		tqg = qgroup_if_io_tqg;
5631 		fn = _task_fn_tx;
5632 		intr_fast = iflib_fast_intr;
5633 		GROUPTASK_INIT(gtask, 0, fn, q);
5634 		ctx->ifc_flags |= IFC_NETMAP_TX_IRQ;
5635 		break;
5636 	case IFLIB_INTR_RX:
5637 		q = &ctx->ifc_rxqs[qid];
5638 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
5639 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5640 		tqg = qgroup_if_io_tqg;
5641 		fn = _task_fn_rx;
5642 		intr_fast = iflib_fast_intr;
5643 		GROUPTASK_INIT(gtask, 0, fn, q);
5644 		break;
5645 	case IFLIB_INTR_RXTX:
5646 		q = &ctx->ifc_rxqs[qid];
5647 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
5648 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5649 		tqg = qgroup_if_io_tqg;
5650 		fn = _task_fn_rx;
5651 		intr_fast = iflib_fast_intr_rxtx;
5652 		GROUPTASK_INIT(gtask, 0, fn, q);
5653 		break;
5654 	case IFLIB_INTR_ADMIN:
5655 		q = ctx;
5656 		tqrid = -1;
5657 		info = &ctx->ifc_filter_info;
5658 		gtask = &ctx->ifc_admin_task;
5659 		tqg = qgroup_if_config_tqg;
5660 		fn = _task_fn_admin;
5661 		intr_fast = iflib_fast_intr_ctx;
5662 		break;
5663 	default:
5664 		panic("unknown net intr type");
5665 	}
5666 
5667 	info->ifi_filter = filter;
5668 	info->ifi_filter_arg = filter_arg;
5669 	info->ifi_task = gtask;
5670 	info->ifi_ctx = q;
5671 
5672 	err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info,  name);
5673 	if (err != 0) {
5674 		device_printf(ctx->ifc_dev, "_iflib_irq_alloc failed %d\n", err);
5675 		return (err);
5676 	}
5677 	if (type == IFLIB_INTR_ADMIN)
5678 		return (0);
5679 
5680 	if (tqrid != -1) {
5681 		err = iflib_irq_set_affinity(ctx, rman_get_start(irq->ii_res), type, qid, gtask, tqg, q, name);
5682 		if (err)
5683 			return (err);
5684 	} else {
5685 		taskqgroup_attach(tqg, gtask, q, rman_get_start(irq->ii_res), name);
5686 	}
5687 
5688 	return (0);
5689 }
5690 
5691 void
5692 iflib_softirq_alloc_generic(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type,  void *arg, int qid, const char *name)
5693 {
5694 	struct grouptask *gtask;
5695 	struct taskqgroup *tqg;
5696 	gtask_fn_t *fn;
5697 	void *q;
5698 	int irq_num = -1;
5699 	int err;
5700 
5701 	switch (type) {
5702 	case IFLIB_INTR_TX:
5703 		q = &ctx->ifc_txqs[qid];
5704 		gtask = &ctx->ifc_txqs[qid].ift_task;
5705 		tqg = qgroup_if_io_tqg;
5706 		fn = _task_fn_tx;
5707 		if (irq != NULL)
5708 			irq_num = rman_get_start(irq->ii_res);
5709 		break;
5710 	case IFLIB_INTR_RX:
5711 		q = &ctx->ifc_rxqs[qid];
5712 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5713 		tqg = qgroup_if_io_tqg;
5714 		fn = _task_fn_rx;
5715 		if (irq != NULL)
5716 			irq_num = rman_get_start(irq->ii_res);
5717 		break;
5718 	case IFLIB_INTR_IOV:
5719 		q = ctx;
5720 		gtask = &ctx->ifc_vflr_task;
5721 		tqg = qgroup_if_config_tqg;
5722 		fn = _task_fn_iov;
5723 		break;
5724 	default:
5725 		panic("unknown net intr type");
5726 	}
5727 	GROUPTASK_INIT(gtask, 0, fn, q);
5728 	if (irq_num != -1) {
5729 		err = iflib_irq_set_affinity(ctx, irq_num, type, qid, gtask, tqg, q, name);
5730 		if (err)
5731 			taskqgroup_attach(tqg, gtask, q, irq_num, name);
5732 	}
5733 	else {
5734 		taskqgroup_attach(tqg, gtask, q, irq_num, name);
5735 	}
5736 }
5737 
5738 void
5739 iflib_irq_free(if_ctx_t ctx, if_irq_t irq)
5740 {
5741 	if (irq->ii_tag)
5742 		bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag);
5743 
5744 	if (irq->ii_res)
5745 		bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ, irq->ii_rid, irq->ii_res);
5746 }
5747 
5748 static int
5749 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, const char *name)
5750 {
5751 	iflib_txq_t txq = ctx->ifc_txqs;
5752 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5753 	if_irq_t irq = &ctx->ifc_legacy_irq;
5754 	iflib_filter_info_t info;
5755 	struct grouptask *gtask;
5756 	struct taskqgroup *tqg;
5757 	gtask_fn_t *fn;
5758 	int tqrid;
5759 	void *q;
5760 	int err;
5761 
5762 	q = &ctx->ifc_rxqs[0];
5763 	info = &rxq[0].ifr_filter_info;
5764 	gtask = &rxq[0].ifr_task;
5765 	tqg = qgroup_if_io_tqg;
5766 	tqrid = irq->ii_rid = *rid;
5767 	fn = _task_fn_rx;
5768 
5769 	ctx->ifc_flags |= IFC_LEGACY;
5770 	info->ifi_filter = filter;
5771 	info->ifi_filter_arg = filter_arg;
5772 	info->ifi_task = gtask;
5773 	info->ifi_ctx = ctx;
5774 
5775 	/* We allocate a single interrupt resource */
5776 	if ((err = _iflib_irq_alloc(ctx, irq, tqrid, iflib_fast_intr_ctx, NULL, info, name)) != 0)
5777 		return (err);
5778 	GROUPTASK_INIT(gtask, 0, fn, q);
5779 	taskqgroup_attach(tqg, gtask, q, rman_get_start(irq->ii_res), name);
5780 
5781 	GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq);
5782 	taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, rman_get_start(irq->ii_res), "tx");
5783 	return (0);
5784 }
5785 
5786 void
5787 iflib_led_create(if_ctx_t ctx)
5788 {
5789 
5790 	ctx->ifc_led_dev = led_create(iflib_led_func, ctx,
5791 	    device_get_nameunit(ctx->ifc_dev));
5792 }
5793 
5794 void
5795 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid)
5796 {
5797 
5798 	GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
5799 }
5800 
5801 void
5802 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid)
5803 {
5804 
5805 	GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task);
5806 }
5807 
5808 void
5809 iflib_admin_intr_deferred(if_ctx_t ctx)
5810 {
5811 #ifdef INVARIANTS
5812 	struct grouptask *gtask;
5813 
5814 	gtask = &ctx->ifc_admin_task;
5815 	MPASS(gtask != NULL && gtask->gt_taskqueue != NULL);
5816 #endif
5817 
5818 	GROUPTASK_ENQUEUE(&ctx->ifc_admin_task);
5819 }
5820 
5821 void
5822 iflib_iov_intr_deferred(if_ctx_t ctx)
5823 {
5824 
5825 	GROUPTASK_ENQUEUE(&ctx->ifc_vflr_task);
5826 }
5827 
5828 void
5829 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, char *name)
5830 {
5831 
5832 	taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, -1, name);
5833 }
5834 
5835 void
5836 iflib_config_gtask_init(void *ctx, struct grouptask *gtask, gtask_fn_t *fn,
5837 	const char *name)
5838 {
5839 
5840 	GROUPTASK_INIT(gtask, 0, fn, ctx);
5841 	taskqgroup_attach(qgroup_if_config_tqg, gtask, gtask, -1, name);
5842 }
5843 
5844 void
5845 iflib_config_gtask_deinit(struct grouptask *gtask)
5846 {
5847 
5848 	taskqgroup_detach(qgroup_if_config_tqg, gtask);
5849 }
5850 
5851 void
5852 iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate)
5853 {
5854 	if_t ifp = ctx->ifc_ifp;
5855 	iflib_txq_t txq = ctx->ifc_txqs;
5856 
5857 	if_setbaudrate(ifp, baudrate);
5858 	if (baudrate >= IF_Gbps(10)) {
5859 		STATE_LOCK(ctx);
5860 		ctx->ifc_flags |= IFC_PREFETCH;
5861 		STATE_UNLOCK(ctx);
5862 	}
5863 	/* If link down, disable watchdog */
5864 	if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) {
5865 		for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++)
5866 			txq->ift_qstatus = IFLIB_QUEUE_IDLE;
5867 	}
5868 	ctx->ifc_link_state = link_state;
5869 	if_link_state_change(ifp, link_state);
5870 }
5871 
5872 static int
5873 iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq)
5874 {
5875 	int credits;
5876 #ifdef INVARIANTS
5877 	int credits_pre = txq->ift_cidx_processed;
5878 #endif
5879 
5880 	if (ctx->isc_txd_credits_update == NULL)
5881 		return (0);
5882 
5883 	if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, true)) == 0)
5884 		return (0);
5885 
5886 	txq->ift_processed += credits;
5887 	txq->ift_cidx_processed += credits;
5888 
5889 	MPASS(credits_pre + credits == txq->ift_cidx_processed);
5890 	if (txq->ift_cidx_processed >= txq->ift_size)
5891 		txq->ift_cidx_processed -= txq->ift_size;
5892 	return (credits);
5893 }
5894 
5895 static int
5896 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget)
5897 {
5898 
5899 	return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx,
5900 	    budget));
5901 }
5902 
5903 void
5904 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name,
5905 	const char *description, if_int_delay_info_t info,
5906 	int offset, int value)
5907 {
5908 	info->iidi_ctx = ctx;
5909 	info->iidi_offset = offset;
5910 	info->iidi_value = value;
5911 	SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev),
5912 	    SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)),
5913 	    OID_AUTO, name, CTLTYPE_INT|CTLFLAG_RW,
5914 	    info, 0, iflib_sysctl_int_delay, "I", description);
5915 }
5916 
5917 struct sx *
5918 iflib_ctx_lock_get(if_ctx_t ctx)
5919 {
5920 
5921 	return (&ctx->ifc_ctx_sx);
5922 }
5923 
5924 static int
5925 iflib_msix_init(if_ctx_t ctx)
5926 {
5927 	device_t dev = ctx->ifc_dev;
5928 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5929 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5930 	int vectors, queues, rx_queues, tx_queues, queuemsgs, msgs;
5931 	int iflib_num_tx_queues, iflib_num_rx_queues;
5932 	int err, admincnt, bar;
5933 
5934 	iflib_num_tx_queues = ctx->ifc_sysctl_ntxqs;
5935 	iflib_num_rx_queues = ctx->ifc_sysctl_nrxqs;
5936 
5937 	device_printf(dev, "msix_init qsets capped at %d\n", imax(scctx->isc_ntxqsets, scctx->isc_nrxqsets));
5938 
5939 	bar = ctx->ifc_softc_ctx.isc_msix_bar;
5940 	admincnt = sctx->isc_admin_intrcnt;
5941 	/* Override by tuneable */
5942 	if (scctx->isc_disable_msix)
5943 		goto msi;
5944 
5945 	/*
5946 	 * bar == -1 => "trust me I know what I'm doing"
5947 	 * Some drivers are for hardware that is so shoddily
5948 	 * documented that no one knows which bars are which
5949 	 * so the developer has to map all bars. This hack
5950 	 * allows shoddy garbage to use msix in this framework.
5951 	 */
5952 	if (bar != -1) {
5953 		ctx->ifc_msix_mem = bus_alloc_resource_any(dev,
5954 	            SYS_RES_MEMORY, &bar, RF_ACTIVE);
5955 		if (ctx->ifc_msix_mem == NULL) {
5956 			/* May not be enabled */
5957 			device_printf(dev, "Unable to map MSIX table \n");
5958 			goto msi;
5959 		}
5960 	}
5961 	/* First try MSI/X */
5962 	if ((msgs = pci_msix_count(dev)) == 0) { /* system has msix disabled */
5963 		device_printf(dev, "System has MSIX disabled \n");
5964 		bus_release_resource(dev, SYS_RES_MEMORY,
5965 		    bar, ctx->ifc_msix_mem);
5966 		ctx->ifc_msix_mem = NULL;
5967 		goto msi;
5968 	}
5969 #if IFLIB_DEBUG
5970 	/* use only 1 qset in debug mode */
5971 	queuemsgs = min(msgs - admincnt, 1);
5972 #else
5973 	queuemsgs = msgs - admincnt;
5974 #endif
5975 #ifdef RSS
5976 	queues = imin(queuemsgs, rss_getnumbuckets());
5977 #else
5978 	queues = queuemsgs;
5979 #endif
5980 	queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues);
5981 	device_printf(dev, "pxm cpus: %d queue msgs: %d admincnt: %d\n",
5982 				  CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt);
5983 #ifdef  RSS
5984 	/* If we're doing RSS, clamp at the number of RSS buckets */
5985 	if (queues > rss_getnumbuckets())
5986 		queues = rss_getnumbuckets();
5987 #endif
5988 	if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt)
5989 		rx_queues = iflib_num_rx_queues;
5990 	else
5991 		rx_queues = queues;
5992 
5993 	if (rx_queues > scctx->isc_nrxqsets)
5994 		rx_queues = scctx->isc_nrxqsets;
5995 
5996 	/*
5997 	 * We want this to be all logical CPUs by default
5998 	 */
5999 	if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues)
6000 		tx_queues = iflib_num_tx_queues;
6001 	else
6002 		tx_queues = mp_ncpus;
6003 
6004 	if (tx_queues > scctx->isc_ntxqsets)
6005 		tx_queues = scctx->isc_ntxqsets;
6006 
6007 	if (ctx->ifc_sysctl_qs_eq_override == 0) {
6008 #ifdef INVARIANTS
6009 		if (tx_queues != rx_queues)
6010 			device_printf(dev, "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n",
6011 				      min(rx_queues, tx_queues), min(rx_queues, tx_queues));
6012 #endif
6013 		tx_queues = min(rx_queues, tx_queues);
6014 		rx_queues = min(rx_queues, tx_queues);
6015 	}
6016 
6017 	device_printf(dev, "using %d rx queues %d tx queues \n", rx_queues, tx_queues);
6018 
6019 	vectors = rx_queues + admincnt;
6020 	if ((err = pci_alloc_msix(dev, &vectors)) == 0) {
6021 		device_printf(dev,
6022 					  "Using MSIX interrupts with %d vectors\n", vectors);
6023 		scctx->isc_vectors = vectors;
6024 		scctx->isc_nrxqsets = rx_queues;
6025 		scctx->isc_ntxqsets = tx_queues;
6026 		scctx->isc_intr = IFLIB_INTR_MSIX;
6027 
6028 		return (vectors);
6029 	} else {
6030 		device_printf(dev, "failed to allocate %d msix vectors, err: %d - using MSI\n", vectors, err);
6031 		bus_release_resource(dev, SYS_RES_MEMORY, bar,
6032 		    ctx->ifc_msix_mem);
6033 		ctx->ifc_msix_mem = NULL;
6034 	}
6035 msi:
6036 	vectors = pci_msi_count(dev);
6037 	scctx->isc_nrxqsets = 1;
6038 	scctx->isc_ntxqsets = 1;
6039 	scctx->isc_vectors = vectors;
6040 	if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) {
6041 		device_printf(dev,"Using an MSI interrupt\n");
6042 		scctx->isc_intr = IFLIB_INTR_MSI;
6043 	} else {
6044 		scctx->isc_vectors = 1;
6045 		device_printf(dev,"Using a Legacy interrupt\n");
6046 		scctx->isc_intr = IFLIB_INTR_LEGACY;
6047 	}
6048 
6049 	return (vectors);
6050 }
6051 
6052 static const char *ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" };
6053 
6054 static int
6055 mp_ring_state_handler(SYSCTL_HANDLER_ARGS)
6056 {
6057 	int rc;
6058 	uint16_t *state = ((uint16_t *)oidp->oid_arg1);
6059 	struct sbuf *sb;
6060 	const char *ring_state = "UNKNOWN";
6061 
6062 	/* XXX needed ? */
6063 	rc = sysctl_wire_old_buffer(req, 0);
6064 	MPASS(rc == 0);
6065 	if (rc != 0)
6066 		return (rc);
6067 	sb = sbuf_new_for_sysctl(NULL, NULL, 80, req);
6068 	MPASS(sb != NULL);
6069 	if (sb == NULL)
6070 		return (ENOMEM);
6071 	if (state[3] <= 3)
6072 		ring_state = ring_states[state[3]];
6073 
6074 	sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s",
6075 		    state[0], state[1], state[2], ring_state);
6076 	rc = sbuf_finish(sb);
6077 	sbuf_delete(sb);
6078         return(rc);
6079 }
6080 
6081 enum iflib_ndesc_handler {
6082 	IFLIB_NTXD_HANDLER,
6083 	IFLIB_NRXD_HANDLER,
6084 };
6085 
6086 static int
6087 mp_ndesc_handler(SYSCTL_HANDLER_ARGS)
6088 {
6089 	if_ctx_t ctx = (void *)arg1;
6090 	enum iflib_ndesc_handler type = arg2;
6091 	char buf[256] = {0};
6092 	qidx_t *ndesc;
6093 	char *p, *next;
6094 	int nqs, rc, i;
6095 
6096 	MPASS(type == IFLIB_NTXD_HANDLER || type == IFLIB_NRXD_HANDLER);
6097 
6098 	nqs = 8;
6099 	switch(type) {
6100 	case IFLIB_NTXD_HANDLER:
6101 		ndesc = ctx->ifc_sysctl_ntxds;
6102 		if (ctx->ifc_sctx)
6103 			nqs = ctx->ifc_sctx->isc_ntxqs;
6104 		break;
6105 	case IFLIB_NRXD_HANDLER:
6106 		ndesc = ctx->ifc_sysctl_nrxds;
6107 		if (ctx->ifc_sctx)
6108 			nqs = ctx->ifc_sctx->isc_nrxqs;
6109 		break;
6110 	default:
6111 			panic("unhandled type");
6112 	}
6113 	if (nqs == 0)
6114 		nqs = 8;
6115 
6116 	for (i=0; i<8; i++) {
6117 		if (i >= nqs)
6118 			break;
6119 		if (i)
6120 			strcat(buf, ",");
6121 		sprintf(strchr(buf, 0), "%d", ndesc[i]);
6122 	}
6123 
6124 	rc = sysctl_handle_string(oidp, buf, sizeof(buf), req);
6125 	if (rc || req->newptr == NULL)
6126 		return rc;
6127 
6128 	for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p;
6129 	    i++, p = strsep(&next, " ,")) {
6130 		ndesc[i] = strtoul(p, NULL, 10);
6131 	}
6132 
6133 	return(rc);
6134 }
6135 
6136 #define NAME_BUFLEN 32
6137 static void
6138 iflib_add_device_sysctl_pre(if_ctx_t ctx)
6139 {
6140         device_t dev = iflib_get_dev(ctx);
6141 	struct sysctl_oid_list *child, *oid_list;
6142 	struct sysctl_ctx_list *ctx_list;
6143 	struct sysctl_oid *node;
6144 
6145 	ctx_list = device_get_sysctl_ctx(dev);
6146 	child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev));
6147 	ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, "iflib",
6148 						      CTLFLAG_RD, NULL, "IFLIB fields");
6149 	oid_list = SYSCTL_CHILDREN(node);
6150 
6151 	SYSCTL_ADD_STRING(ctx_list, oid_list, OID_AUTO, "driver_version",
6152 		       CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version, 0,
6153 		       "driver version");
6154 
6155 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs",
6156 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0,
6157 			"# of txqs to use, 0 => use default #");
6158 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs",
6159 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0,
6160 			"# of rxqs to use, 0 => use default #");
6161 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_qs_enable",
6162 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0,
6163                        "permit #txq != #rxq");
6164 	SYSCTL_ADD_INT(ctx_list, oid_list, OID_AUTO, "disable_msix",
6165                       CTLFLAG_RWTUN, &ctx->ifc_softc_ctx.isc_disable_msix, 0,
6166                       "disable MSIX (default 0)");
6167 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "rx_budget",
6168 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_rx_budget, 0,
6169                        "set the rx budget");
6170 
6171 	/* XXX change for per-queue sizes */
6172 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_ntxds",
6173 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NTXD_HANDLER,
6174                        mp_ndesc_handler, "A",
6175                        "list of # of tx descriptors to use, 0 = use default #");
6176 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_nrxds",
6177 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NRXD_HANDLER,
6178                        mp_ndesc_handler, "A",
6179                        "list of # of rx descriptors to use, 0 = use default #");
6180 }
6181 
6182 static void
6183 iflib_add_device_sysctl_post(if_ctx_t ctx)
6184 {
6185 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6186 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6187         device_t dev = iflib_get_dev(ctx);
6188 	struct sysctl_oid_list *child;
6189 	struct sysctl_ctx_list *ctx_list;
6190 	iflib_fl_t fl;
6191 	iflib_txq_t txq;
6192 	iflib_rxq_t rxq;
6193 	int i, j;
6194 	char namebuf[NAME_BUFLEN];
6195 	char *qfmt;
6196 	struct sysctl_oid *queue_node, *fl_node, *node;
6197 	struct sysctl_oid_list *queue_list, *fl_list;
6198 	ctx_list = device_get_sysctl_ctx(dev);
6199 
6200 	node = ctx->ifc_sysctl_node;
6201 	child = SYSCTL_CHILDREN(node);
6202 
6203 	if (scctx->isc_ntxqsets > 100)
6204 		qfmt = "txq%03d";
6205 	else if (scctx->isc_ntxqsets > 10)
6206 		qfmt = "txq%02d";
6207 	else
6208 		qfmt = "txq%d";
6209 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
6210 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6211 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6212 					     CTLFLAG_RD, NULL, "Queue Name");
6213 		queue_list = SYSCTL_CHILDREN(queue_node);
6214 #if MEMORY_LOGGING
6215 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued",
6216 				CTLFLAG_RD,
6217 				&txq->ift_dequeued, "total mbufs freed");
6218 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued",
6219 				CTLFLAG_RD,
6220 				&txq->ift_enqueued, "total mbufs enqueued");
6221 #endif
6222 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag",
6223 				   CTLFLAG_RD,
6224 				   &txq->ift_mbuf_defrag, "# of times m_defrag was called");
6225 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "m_pullups",
6226 				   CTLFLAG_RD,
6227 				   &txq->ift_pullups, "# of times m_pullup was called");
6228 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag_failed",
6229 				   CTLFLAG_RD,
6230 				   &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed");
6231 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_desc_avail",
6232 				   CTLFLAG_RD,
6233 				   &txq->ift_no_desc_avail, "# of times no descriptors were available");
6234 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "tx_map_failed",
6235 				   CTLFLAG_RD,
6236 				   &txq->ift_map_failed, "# of times dma map failed");
6237 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txd_encap_efbig",
6238 				   CTLFLAG_RD,
6239 				   &txq->ift_txd_encap_efbig, "# of times txd_encap returned EFBIG");
6240 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_tx_dma_setup",
6241 				   CTLFLAG_RD,
6242 				   &txq->ift_no_tx_dma_setup, "# of times map failed for other than EFBIG");
6243 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx",
6244 				   CTLFLAG_RD,
6245 				   &txq->ift_pidx, 1, "Producer Index");
6246 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx",
6247 				   CTLFLAG_RD,
6248 				   &txq->ift_cidx, 1, "Consumer Index");
6249 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx_processed",
6250 				   CTLFLAG_RD,
6251 				   &txq->ift_cidx_processed, 1, "Consumer Index seen by credit update");
6252 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use",
6253 				   CTLFLAG_RD,
6254 				   &txq->ift_in_use, 1, "descriptors in use");
6255 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_processed",
6256 				   CTLFLAG_RD,
6257 				   &txq->ift_processed, "descriptors procesed for clean");
6258 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned",
6259 				   CTLFLAG_RD,
6260 				   &txq->ift_cleaned, "total cleaned");
6261 		SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state",
6262 				CTLTYPE_STRING | CTLFLAG_RD, __DEVOLATILE(uint64_t *, &txq->ift_br->state),
6263 				0, mp_ring_state_handler, "A", "soft ring state");
6264 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_enqueues",
6265 				       CTLFLAG_RD, &txq->ift_br->enqueues,
6266 				       "# of enqueues to the mp_ring for this queue");
6267 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_drops",
6268 				       CTLFLAG_RD, &txq->ift_br->drops,
6269 				       "# of drops in the mp_ring for this queue");
6270 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_starts",
6271 				       CTLFLAG_RD, &txq->ift_br->starts,
6272 				       "# of normal consumer starts in the mp_ring for this queue");
6273 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_stalls",
6274 				       CTLFLAG_RD, &txq->ift_br->stalls,
6275 					       "# of consumer stalls in the mp_ring for this queue");
6276 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_restarts",
6277 			       CTLFLAG_RD, &txq->ift_br->restarts,
6278 				       "# of consumer restarts in the mp_ring for this queue");
6279 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_abdications",
6280 				       CTLFLAG_RD, &txq->ift_br->abdications,
6281 				       "# of consumer abdications in the mp_ring for this queue");
6282 	}
6283 
6284 	if (scctx->isc_nrxqsets > 100)
6285 		qfmt = "rxq%03d";
6286 	else if (scctx->isc_nrxqsets > 10)
6287 		qfmt = "rxq%02d";
6288 	else
6289 		qfmt = "rxq%d";
6290 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
6291 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6292 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6293 					     CTLFLAG_RD, NULL, "Queue Name");
6294 		queue_list = SYSCTL_CHILDREN(queue_node);
6295 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
6296 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_pidx",
6297 				       CTLFLAG_RD,
6298 				       &rxq->ifr_cq_pidx, 1, "Producer Index");
6299 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_cidx",
6300 				       CTLFLAG_RD,
6301 				       &rxq->ifr_cq_cidx, 1, "Consumer Index");
6302 		}
6303 
6304 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
6305 			snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j);
6306 			fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list, OID_AUTO, namebuf,
6307 						     CTLFLAG_RD, NULL, "freelist Name");
6308 			fl_list = SYSCTL_CHILDREN(fl_node);
6309 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx",
6310 				       CTLFLAG_RD,
6311 				       &fl->ifl_pidx, 1, "Producer Index");
6312 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx",
6313 				       CTLFLAG_RD,
6314 				       &fl->ifl_cidx, 1, "Consumer Index");
6315 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits",
6316 				       CTLFLAG_RD,
6317 				       &fl->ifl_credits, 1, "credits available");
6318 #if MEMORY_LOGGING
6319 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_enqueued",
6320 					CTLFLAG_RD,
6321 					&fl->ifl_m_enqueued, "mbufs allocated");
6322 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_dequeued",
6323 					CTLFLAG_RD,
6324 					&fl->ifl_m_dequeued, "mbufs freed");
6325 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_enqueued",
6326 					CTLFLAG_RD,
6327 					&fl->ifl_cl_enqueued, "clusters allocated");
6328 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_dequeued",
6329 					CTLFLAG_RD,
6330 					&fl->ifl_cl_dequeued, "clusters freed");
6331 #endif
6332 
6333 		}
6334 	}
6335 
6336 }
6337 
6338 #ifndef __NO_STRICT_ALIGNMENT
6339 static struct mbuf *
6340 iflib_fixup_rx(struct mbuf *m)
6341 {
6342 	struct mbuf *n;
6343 
6344 	if (m->m_len <= (MCLBYTES - ETHER_HDR_LEN)) {
6345 		bcopy(m->m_data, m->m_data + ETHER_HDR_LEN, m->m_len);
6346 		m->m_data += ETHER_HDR_LEN;
6347 		n = m;
6348 	} else {
6349 		MGETHDR(n, M_NOWAIT, MT_DATA);
6350 		if (n == NULL) {
6351 			m_freem(m);
6352 			return (NULL);
6353 		}
6354 		bcopy(m->m_data, n->m_data, ETHER_HDR_LEN);
6355 		m->m_data += ETHER_HDR_LEN;
6356 		m->m_len -= ETHER_HDR_LEN;
6357 		n->m_len = ETHER_HDR_LEN;
6358 		M_MOVE_PKTHDR(n, m);
6359 		n->m_next = m;
6360 	}
6361 	return (n);
6362 }
6363 #endif
6364 
6365 #ifdef NETDUMP
6366 static void
6367 iflib_netdump_init(struct ifnet *ifp, int *nrxr, int *ncl, int *clsize)
6368 {
6369 	if_ctx_t ctx;
6370 
6371 	ctx = if_getsoftc(ifp);
6372 	CTX_LOCK(ctx);
6373 	*nrxr = NRXQSETS(ctx);
6374 	*ncl = ctx->ifc_rxqs[0].ifr_fl->ifl_size;
6375 	*clsize = ctx->ifc_rxqs[0].ifr_fl->ifl_buf_size;
6376 	CTX_UNLOCK(ctx);
6377 }
6378 
6379 static void
6380 iflib_netdump_event(struct ifnet *ifp, enum netdump_ev event)
6381 {
6382 	if_ctx_t ctx;
6383 	if_softc_ctx_t scctx;
6384 	iflib_fl_t fl;
6385 	iflib_rxq_t rxq;
6386 	int i, j;
6387 
6388 	ctx = if_getsoftc(ifp);
6389 	scctx = &ctx->ifc_softc_ctx;
6390 
6391 	switch (event) {
6392 	case NETDUMP_START:
6393 		for (i = 0; i < scctx->isc_nrxqsets; i++) {
6394 			rxq = &ctx->ifc_rxqs[i];
6395 			for (j = 0; j < rxq->ifr_nfl; j++) {
6396 				fl = rxq->ifr_fl;
6397 				fl->ifl_zone = m_getzone(fl->ifl_buf_size);
6398 			}
6399 		}
6400 		iflib_no_tx_batch = 1;
6401 		break;
6402 	default:
6403 		break;
6404 	}
6405 }
6406 
6407 static int
6408 iflib_netdump_transmit(struct ifnet *ifp, struct mbuf *m)
6409 {
6410 	if_ctx_t ctx;
6411 	iflib_txq_t txq;
6412 	int error;
6413 
6414 	ctx = if_getsoftc(ifp);
6415 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
6416 	    IFF_DRV_RUNNING)
6417 		return (EBUSY);
6418 
6419 	txq = &ctx->ifc_txqs[0];
6420 	error = iflib_encap(txq, &m);
6421 	if (error == 0)
6422 		(void)iflib_txd_db_check(ctx, txq, true, txq->ift_in_use);
6423 	return (error);
6424 }
6425 
6426 static int
6427 iflib_netdump_poll(struct ifnet *ifp, int count)
6428 {
6429 	if_ctx_t ctx;
6430 	if_softc_ctx_t scctx;
6431 	iflib_txq_t txq;
6432 	int i;
6433 
6434 	ctx = if_getsoftc(ifp);
6435 	scctx = &ctx->ifc_softc_ctx;
6436 
6437 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
6438 	    IFF_DRV_RUNNING)
6439 		return (EBUSY);
6440 
6441 	txq = &ctx->ifc_txqs[0];
6442 	(void)iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
6443 
6444 	for (i = 0; i < scctx->isc_nrxqsets; i++)
6445 		(void)iflib_rxeof(&ctx->ifc_rxqs[i], 16 /* XXX */);
6446 	return (0);
6447 }
6448 #endif /* NETDUMP */
6449