xref: /freebsd/sys/net/iflib.c (revision c9a49a4fd80541d3bd567fcb3023d7ab389df3bf)
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 static void iflib_timer(void *arg);
150 
151 typedef struct iflib_filter_info {
152 	driver_filter_t *ifi_filter;
153 	void *ifi_filter_arg;
154 	struct grouptask *ifi_task;
155 	void *ifi_ctx;
156 } *iflib_filter_info_t;
157 
158 struct iflib_ctx {
159 	KOBJ_FIELDS;
160    /*
161    * Pointer to hardware driver's softc
162    */
163 	void *ifc_softc;
164 	device_t ifc_dev;
165 	if_t ifc_ifp;
166 
167 	cpuset_t ifc_cpus;
168 	if_shared_ctx_t ifc_sctx;
169 	struct if_softc_ctx ifc_softc_ctx;
170 
171 	struct sx ifc_ctx_sx;
172 	struct mtx ifc_state_mtx;
173 
174 	uint16_t ifc_nhwtxqs;
175 
176 	iflib_txq_t ifc_txqs;
177 	iflib_rxq_t ifc_rxqs;
178 	uint32_t ifc_if_flags;
179 	uint32_t ifc_flags;
180 	uint32_t ifc_max_fl_buf_size;
181 	int ifc_in_detach;
182 
183 	int ifc_link_state;
184 	int ifc_link_irq;
185 	int ifc_watchdog_events;
186 	struct cdev *ifc_led_dev;
187 	struct resource *ifc_msix_mem;
188 
189 	struct if_irq ifc_legacy_irq;
190 	struct grouptask ifc_admin_task;
191 	struct grouptask ifc_vflr_task;
192 	struct iflib_filter_info ifc_filter_info;
193 	struct ifmedia	ifc_media;
194 
195 	struct sysctl_oid *ifc_sysctl_node;
196 	uint16_t ifc_sysctl_ntxqs;
197 	uint16_t ifc_sysctl_nrxqs;
198 	uint16_t ifc_sysctl_qs_eq_override;
199 	uint16_t ifc_sysctl_rx_budget;
200 	uint16_t ifc_sysctl_tx_abdicate;
201 
202 	qidx_t ifc_sysctl_ntxds[8];
203 	qidx_t ifc_sysctl_nrxds[8];
204 	struct if_txrx ifc_txrx;
205 #define isc_txd_encap  ifc_txrx.ift_txd_encap
206 #define isc_txd_flush  ifc_txrx.ift_txd_flush
207 #define isc_txd_credits_update  ifc_txrx.ift_txd_credits_update
208 #define isc_rxd_available ifc_txrx.ift_rxd_available
209 #define isc_rxd_pkt_get ifc_txrx.ift_rxd_pkt_get
210 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
211 #define isc_rxd_flush ifc_txrx.ift_rxd_flush
212 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
213 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
214 #define isc_legacy_intr ifc_txrx.ift_legacy_intr
215 	eventhandler_tag ifc_vlan_attach_event;
216 	eventhandler_tag ifc_vlan_detach_event;
217 	uint8_t ifc_mac[ETHER_ADDR_LEN];
218 	char ifc_mtx_name[16];
219 };
220 
221 
222 void *
223 iflib_get_softc(if_ctx_t ctx)
224 {
225 
226 	return (ctx->ifc_softc);
227 }
228 
229 device_t
230 iflib_get_dev(if_ctx_t ctx)
231 {
232 
233 	return (ctx->ifc_dev);
234 }
235 
236 if_t
237 iflib_get_ifp(if_ctx_t ctx)
238 {
239 
240 	return (ctx->ifc_ifp);
241 }
242 
243 struct ifmedia *
244 iflib_get_media(if_ctx_t ctx)
245 {
246 
247 	return (&ctx->ifc_media);
248 }
249 
250 uint32_t
251 iflib_get_flags(if_ctx_t ctx)
252 {
253 	return (ctx->ifc_flags);
254 }
255 
256 void
257 iflib_set_detach(if_ctx_t ctx)
258 {
259 	ctx->ifc_in_detach = 1;
260 }
261 
262 void
263 iflib_set_mac(if_ctx_t ctx, uint8_t mac[ETHER_ADDR_LEN])
264 {
265 
266 	bcopy(mac, ctx->ifc_mac, ETHER_ADDR_LEN);
267 }
268 
269 if_softc_ctx_t
270 iflib_get_softc_ctx(if_ctx_t ctx)
271 {
272 
273 	return (&ctx->ifc_softc_ctx);
274 }
275 
276 if_shared_ctx_t
277 iflib_get_sctx(if_ctx_t ctx)
278 {
279 
280 	return (ctx->ifc_sctx);
281 }
282 
283 #define IP_ALIGNED(m) ((((uintptr_t)(m)->m_data) & 0x3) == 0x2)
284 #define CACHE_PTR_INCREMENT (CACHE_LINE_SIZE/sizeof(void*))
285 #define CACHE_PTR_NEXT(ptr) ((void *)(((uintptr_t)(ptr)+CACHE_LINE_SIZE-1) & (CACHE_LINE_SIZE-1)))
286 
287 #define LINK_ACTIVE(ctx) ((ctx)->ifc_link_state == LINK_STATE_UP)
288 #define CTX_IS_VF(ctx) ((ctx)->ifc_sctx->isc_flags & IFLIB_IS_VF)
289 
290 #define RX_SW_DESC_MAP_CREATED	(1 << 0)
291 #define TX_SW_DESC_MAP_CREATED	(1 << 1)
292 #define RX_SW_DESC_INUSE        (1 << 3)
293 #define TX_SW_DESC_MAPPED       (1 << 4)
294 
295 #define	M_TOOBIG		M_PROTO1
296 
297 typedef struct iflib_sw_rx_desc_array {
298 	bus_dmamap_t	*ifsd_map;         /* bus_dma maps for packet */
299 	struct mbuf	**ifsd_m;           /* pkthdr mbufs */
300 	caddr_t		*ifsd_cl;          /* direct cluster pointer for rx */
301 	uint8_t		*ifsd_flags;
302 } iflib_rxsd_array_t;
303 
304 typedef struct iflib_sw_tx_desc_array {
305 	bus_dmamap_t    *ifsd_map;         /* bus_dma maps for packet */
306 	struct mbuf    **ifsd_m;           /* pkthdr mbufs */
307 	uint8_t		*ifsd_flags;
308 } if_txsd_vec_t;
309 
310 
311 /* magic number that should be high enough for any hardware */
312 #define IFLIB_MAX_TX_SEGS		128
313 /* bnxt supports 64 with hardware LRO enabled */
314 #define IFLIB_MAX_RX_SEGS		64
315 #define IFLIB_RX_COPY_THRESH		128
316 #define IFLIB_MAX_RX_REFRESH		32
317 /* The minimum descriptors per second before we start coalescing */
318 #define IFLIB_MIN_DESC_SEC		16384
319 #define IFLIB_DEFAULT_TX_UPDATE_FREQ	16
320 #define IFLIB_QUEUE_IDLE		0
321 #define IFLIB_QUEUE_HUNG		1
322 #define IFLIB_QUEUE_WORKING		2
323 /* maximum number of txqs that can share an rx interrupt */
324 #define IFLIB_MAX_TX_SHARED_INTR	4
325 
326 /* this should really scale with ring size - this is a fairly arbitrary value */
327 #define TX_BATCH_SIZE			32
328 
329 #define IFLIB_RESTART_BUDGET		8
330 
331 
332 #define CSUM_OFFLOAD		(CSUM_IP_TSO|CSUM_IP6_TSO|CSUM_IP| \
333 				 CSUM_IP_UDP|CSUM_IP_TCP|CSUM_IP_SCTP| \
334 				 CSUM_IP6_UDP|CSUM_IP6_TCP|CSUM_IP6_SCTP)
335 struct iflib_txq {
336 	qidx_t		ift_in_use;
337 	qidx_t		ift_cidx;
338 	qidx_t		ift_cidx_processed;
339 	qidx_t		ift_pidx;
340 	uint8_t		ift_gen;
341 	uint8_t		ift_br_offset;
342 	uint16_t	ift_npending;
343 	uint16_t	ift_db_pending;
344 	uint16_t	ift_rs_pending;
345 	/* implicit pad */
346 	uint8_t		ift_txd_size[8];
347 	uint64_t	ift_processed;
348 	uint64_t	ift_cleaned;
349 	uint64_t	ift_cleaned_prev;
350 #if MEMORY_LOGGING
351 	uint64_t	ift_enqueued;
352 	uint64_t	ift_dequeued;
353 #endif
354 	uint64_t	ift_no_tx_dma_setup;
355 	uint64_t	ift_no_desc_avail;
356 	uint64_t	ift_mbuf_defrag_failed;
357 	uint64_t	ift_mbuf_defrag;
358 	uint64_t	ift_map_failed;
359 	uint64_t	ift_txd_encap_efbig;
360 	uint64_t	ift_pullups;
361 	uint64_t	ift_last_timer_tick;
362 
363 	struct mtx	ift_mtx;
364 	struct mtx	ift_db_mtx;
365 
366 	/* constant values */
367 	if_ctx_t	ift_ctx;
368 	struct ifmp_ring        *ift_br;
369 	struct grouptask	ift_task;
370 	qidx_t		ift_size;
371 	uint16_t	ift_id;
372 	struct callout	ift_timer;
373 
374 	if_txsd_vec_t	ift_sds;
375 	uint8_t		ift_qstatus;
376 	uint8_t		ift_closed;
377 	uint8_t		ift_update_freq;
378 	struct iflib_filter_info ift_filter_info;
379 	bus_dma_tag_t		ift_desc_tag;
380 	bus_dma_tag_t		ift_tso_desc_tag;
381 	iflib_dma_info_t	ift_ifdi;
382 #define MTX_NAME_LEN 16
383 	char                    ift_mtx_name[MTX_NAME_LEN];
384 	char                    ift_db_mtx_name[MTX_NAME_LEN];
385 	bus_dma_segment_t	ift_segs[IFLIB_MAX_TX_SEGS]  __aligned(CACHE_LINE_SIZE);
386 #ifdef IFLIB_DIAGNOSTICS
387 	uint64_t ift_cpu_exec_count[256];
388 #endif
389 } __aligned(CACHE_LINE_SIZE);
390 
391 struct iflib_fl {
392 	qidx_t		ifl_cidx;
393 	qidx_t		ifl_pidx;
394 	qidx_t		ifl_credits;
395 	uint8_t		ifl_gen;
396 	uint8_t		ifl_rxd_size;
397 #if MEMORY_LOGGING
398 	uint64_t	ifl_m_enqueued;
399 	uint64_t	ifl_m_dequeued;
400 	uint64_t	ifl_cl_enqueued;
401 	uint64_t	ifl_cl_dequeued;
402 #endif
403 	/* implicit pad */
404 
405 	bitstr_t 	*ifl_rx_bitmap;
406 	qidx_t		ifl_fragidx;
407 	/* constant */
408 	qidx_t		ifl_size;
409 	uint16_t	ifl_buf_size;
410 	uint16_t	ifl_cltype;
411 	uma_zone_t	ifl_zone;
412 	iflib_rxsd_array_t	ifl_sds;
413 	iflib_rxq_t	ifl_rxq;
414 	uint8_t		ifl_id;
415 	bus_dma_tag_t           ifl_desc_tag;
416 	iflib_dma_info_t	ifl_ifdi;
417 	uint64_t	ifl_bus_addrs[IFLIB_MAX_RX_REFRESH] __aligned(CACHE_LINE_SIZE);
418 	caddr_t		ifl_vm_addrs[IFLIB_MAX_RX_REFRESH];
419 	qidx_t	ifl_rxd_idxs[IFLIB_MAX_RX_REFRESH];
420 }  __aligned(CACHE_LINE_SIZE);
421 
422 static inline qidx_t
423 get_inuse(int size, qidx_t cidx, qidx_t pidx, uint8_t gen)
424 {
425 	qidx_t used;
426 
427 	if (pidx > cidx)
428 		used = pidx - cidx;
429 	else if (pidx < cidx)
430 		used = size - cidx + pidx;
431 	else if (gen == 0 && pidx == cidx)
432 		used = 0;
433 	else if (gen == 1 && pidx == cidx)
434 		used = size;
435 	else
436 		panic("bad state");
437 
438 	return (used);
439 }
440 
441 #define TXQ_AVAIL(txq) (txq->ift_size - get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx, txq->ift_gen))
442 
443 #define IDXDIFF(head, tail, wrap) \
444 	((head) >= (tail) ? (head) - (tail) : (wrap) - (tail) + (head))
445 
446 struct iflib_rxq {
447 	/* If there is a separate completion queue -
448 	 * these are the cq cidx and pidx. Otherwise
449 	 * these are unused.
450 	 */
451 	qidx_t		ifr_size;
452 	qidx_t		ifr_cq_cidx;
453 	qidx_t		ifr_cq_pidx;
454 	uint8_t		ifr_cq_gen;
455 	uint8_t		ifr_fl_offset;
456 
457 	if_ctx_t	ifr_ctx;
458 	iflib_fl_t	ifr_fl;
459 	uint64_t	ifr_rx_irq;
460 	uint16_t	ifr_id;
461 	uint8_t		ifr_lro_enabled;
462 	uint8_t		ifr_nfl;
463 	uint8_t		ifr_ntxqirq;
464 	uint8_t		ifr_txqid[IFLIB_MAX_TX_SHARED_INTR];
465 	struct lro_ctrl			ifr_lc;
466 	struct grouptask        ifr_task;
467 	struct iflib_filter_info ifr_filter_info;
468 	iflib_dma_info_t		ifr_ifdi;
469 
470 	/* dynamically allocate if any drivers need a value substantially larger than this */
471 	struct if_rxd_frag	ifr_frags[IFLIB_MAX_RX_SEGS] __aligned(CACHE_LINE_SIZE);
472 #ifdef IFLIB_DIAGNOSTICS
473 	uint64_t ifr_cpu_exec_count[256];
474 #endif
475 }  __aligned(CACHE_LINE_SIZE);
476 
477 typedef struct if_rxsd {
478 	caddr_t *ifsd_cl;
479 	struct mbuf **ifsd_m;
480 	iflib_fl_t ifsd_fl;
481 	qidx_t ifsd_cidx;
482 } *if_rxsd_t;
483 
484 /* multiple of word size */
485 #ifdef __LP64__
486 #define PKT_INFO_SIZE	6
487 #define RXD_INFO_SIZE	5
488 #define PKT_TYPE uint64_t
489 #else
490 #define PKT_INFO_SIZE	11
491 #define RXD_INFO_SIZE	8
492 #define PKT_TYPE uint32_t
493 #endif
494 #define PKT_LOOP_BOUND  ((PKT_INFO_SIZE/3)*3)
495 #define RXD_LOOP_BOUND  ((RXD_INFO_SIZE/4)*4)
496 
497 typedef struct if_pkt_info_pad {
498 	PKT_TYPE pkt_val[PKT_INFO_SIZE];
499 } *if_pkt_info_pad_t;
500 typedef struct if_rxd_info_pad {
501 	PKT_TYPE rxd_val[RXD_INFO_SIZE];
502 } *if_rxd_info_pad_t;
503 
504 CTASSERT(sizeof(struct if_pkt_info_pad) == sizeof(struct if_pkt_info));
505 CTASSERT(sizeof(struct if_rxd_info_pad) == sizeof(struct if_rxd_info));
506 
507 
508 static inline void
509 pkt_info_zero(if_pkt_info_t pi)
510 {
511 	if_pkt_info_pad_t pi_pad;
512 
513 	pi_pad = (if_pkt_info_pad_t)pi;
514 	pi_pad->pkt_val[0] = 0; pi_pad->pkt_val[1] = 0; pi_pad->pkt_val[2] = 0;
515 	pi_pad->pkt_val[3] = 0; pi_pad->pkt_val[4] = 0; pi_pad->pkt_val[5] = 0;
516 #ifndef __LP64__
517 	pi_pad->pkt_val[6] = 0; pi_pad->pkt_val[7] = 0; pi_pad->pkt_val[8] = 0;
518 	pi_pad->pkt_val[9] = 0; pi_pad->pkt_val[10] = 0;
519 #endif
520 }
521 
522 static device_method_t iflib_pseudo_methods[] = {
523 	DEVMETHOD(device_attach, noop_attach),
524 	DEVMETHOD(device_detach, iflib_pseudo_detach),
525 	DEVMETHOD_END
526 };
527 
528 driver_t iflib_pseudodriver = {
529 	"iflib_pseudo", iflib_pseudo_methods, sizeof(struct iflib_ctx),
530 };
531 
532 static inline void
533 rxd_info_zero(if_rxd_info_t ri)
534 {
535 	if_rxd_info_pad_t ri_pad;
536 	int i;
537 
538 	ri_pad = (if_rxd_info_pad_t)ri;
539 	for (i = 0; i < RXD_LOOP_BOUND; i += 4) {
540 		ri_pad->rxd_val[i] = 0;
541 		ri_pad->rxd_val[i+1] = 0;
542 		ri_pad->rxd_val[i+2] = 0;
543 		ri_pad->rxd_val[i+3] = 0;
544 	}
545 #ifdef __LP64__
546 	ri_pad->rxd_val[RXD_INFO_SIZE-1] = 0;
547 #endif
548 }
549 
550 /*
551  * Only allow a single packet to take up most 1/nth of the tx ring
552  */
553 #define MAX_SINGLE_PACKET_FRACTION 12
554 #define IF_BAD_DMA (bus_addr_t)-1
555 
556 #define CTX_ACTIVE(ctx) ((if_getdrvflags((ctx)->ifc_ifp) & IFF_DRV_RUNNING))
557 
558 #define CTX_LOCK_INIT(_sc)  sx_init(&(_sc)->ifc_ctx_sx, "iflib ctx lock")
559 #define CTX_LOCK(ctx) sx_xlock(&(ctx)->ifc_ctx_sx)
560 #define CTX_UNLOCK(ctx) sx_xunlock(&(ctx)->ifc_ctx_sx)
561 #define CTX_LOCK_DESTROY(ctx) sx_destroy(&(ctx)->ifc_ctx_sx)
562 
563 
564 #define STATE_LOCK_INIT(_sc, _name)  mtx_init(&(_sc)->ifc_state_mtx, _name, "iflib state lock", MTX_DEF)
565 #define STATE_LOCK(ctx) mtx_lock(&(ctx)->ifc_state_mtx)
566 #define STATE_UNLOCK(ctx) mtx_unlock(&(ctx)->ifc_state_mtx)
567 #define STATE_LOCK_DESTROY(ctx) mtx_destroy(&(ctx)->ifc_state_mtx)
568 
569 
570 
571 #define CALLOUT_LOCK(txq)	mtx_lock(&txq->ift_mtx)
572 #define CALLOUT_UNLOCK(txq) 	mtx_unlock(&txq->ift_mtx)
573 
574 
575 /* Our boot-time initialization hook */
576 static int	iflib_module_event_handler(module_t, int, void *);
577 
578 static moduledata_t iflib_moduledata = {
579 	"iflib",
580 	iflib_module_event_handler,
581 	NULL
582 };
583 
584 DECLARE_MODULE(iflib, iflib_moduledata, SI_SUB_INIT_IF, SI_ORDER_ANY);
585 MODULE_VERSION(iflib, 1);
586 
587 MODULE_DEPEND(iflib, pci, 1, 1, 1);
588 MODULE_DEPEND(iflib, ether, 1, 1, 1);
589 
590 TASKQGROUP_DEFINE(if_io_tqg, mp_ncpus, 1);
591 TASKQGROUP_DEFINE(if_config_tqg, 1, 1);
592 
593 #ifndef IFLIB_DEBUG_COUNTERS
594 #ifdef INVARIANTS
595 #define IFLIB_DEBUG_COUNTERS 1
596 #else
597 #define IFLIB_DEBUG_COUNTERS 0
598 #endif /* !INVARIANTS */
599 #endif
600 
601 static SYSCTL_NODE(_net, OID_AUTO, iflib, CTLFLAG_RD, 0,
602                    "iflib driver parameters");
603 
604 /*
605  * XXX need to ensure that this can't accidentally cause the head to be moved backwards
606  */
607 static int iflib_min_tx_latency = 0;
608 SYSCTL_INT(_net_iflib, OID_AUTO, min_tx_latency, CTLFLAG_RW,
609 		   &iflib_min_tx_latency, 0, "minimize transmit latency at the possible expense of throughput");
610 static int iflib_no_tx_batch = 0;
611 SYSCTL_INT(_net_iflib, OID_AUTO, no_tx_batch, CTLFLAG_RW,
612 		   &iflib_no_tx_batch, 0, "minimize transmit latency at the possible expense of throughput");
613 
614 
615 #if IFLIB_DEBUG_COUNTERS
616 
617 static int iflib_tx_seen;
618 static int iflib_tx_sent;
619 static int iflib_tx_encap;
620 static int iflib_rx_allocs;
621 static int iflib_fl_refills;
622 static int iflib_fl_refills_large;
623 static int iflib_tx_frees;
624 
625 SYSCTL_INT(_net_iflib, OID_AUTO, tx_seen, CTLFLAG_RD,
626 		   &iflib_tx_seen, 0, "# tx mbufs seen");
627 SYSCTL_INT(_net_iflib, OID_AUTO, tx_sent, CTLFLAG_RD,
628 		   &iflib_tx_sent, 0, "# tx mbufs sent");
629 SYSCTL_INT(_net_iflib, OID_AUTO, tx_encap, CTLFLAG_RD,
630 		   &iflib_tx_encap, 0, "# tx mbufs encapped");
631 SYSCTL_INT(_net_iflib, OID_AUTO, tx_frees, CTLFLAG_RD,
632 		   &iflib_tx_frees, 0, "# tx frees");
633 SYSCTL_INT(_net_iflib, OID_AUTO, rx_allocs, CTLFLAG_RD,
634 		   &iflib_rx_allocs, 0, "# rx allocations");
635 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills, CTLFLAG_RD,
636 		   &iflib_fl_refills, 0, "# refills");
637 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills_large, CTLFLAG_RD,
638 		   &iflib_fl_refills_large, 0, "# large refills");
639 
640 
641 static int iflib_txq_drain_flushing;
642 static int iflib_txq_drain_oactive;
643 static int iflib_txq_drain_notready;
644 static int iflib_txq_drain_encapfail;
645 
646 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_flushing, CTLFLAG_RD,
647 		   &iflib_txq_drain_flushing, 0, "# drain flushes");
648 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_oactive, CTLFLAG_RD,
649 		   &iflib_txq_drain_oactive, 0, "# drain oactives");
650 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_notready, CTLFLAG_RD,
651 		   &iflib_txq_drain_notready, 0, "# drain notready");
652 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_encapfail, CTLFLAG_RD,
653 		   &iflib_txq_drain_encapfail, 0, "# drain encap fails");
654 
655 
656 static int iflib_encap_load_mbuf_fail;
657 static int iflib_encap_pad_mbuf_fail;
658 static int iflib_encap_txq_avail_fail;
659 static int iflib_encap_txd_encap_fail;
660 
661 SYSCTL_INT(_net_iflib, OID_AUTO, encap_load_mbuf_fail, CTLFLAG_RD,
662 		   &iflib_encap_load_mbuf_fail, 0, "# busdma load failures");
663 SYSCTL_INT(_net_iflib, OID_AUTO, encap_pad_mbuf_fail, CTLFLAG_RD,
664 		   &iflib_encap_pad_mbuf_fail, 0, "# runt frame pad failures");
665 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txq_avail_fail, CTLFLAG_RD,
666 		   &iflib_encap_txq_avail_fail, 0, "# txq avail failures");
667 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txd_encap_fail, CTLFLAG_RD,
668 		   &iflib_encap_txd_encap_fail, 0, "# driver encap failures");
669 
670 static int iflib_task_fn_rxs;
671 static int iflib_rx_intr_enables;
672 static int iflib_fast_intrs;
673 static int iflib_intr_link;
674 static int iflib_intr_msix;
675 static int iflib_rx_unavail;
676 static int iflib_rx_ctx_inactive;
677 static int iflib_rx_zero_len;
678 static int iflib_rx_if_input;
679 static int iflib_rx_mbuf_null;
680 static int iflib_rxd_flush;
681 
682 static int iflib_verbose_debug;
683 
684 SYSCTL_INT(_net_iflib, OID_AUTO, intr_link, CTLFLAG_RD,
685 		   &iflib_intr_link, 0, "# intr link calls");
686 SYSCTL_INT(_net_iflib, OID_AUTO, intr_msix, CTLFLAG_RD,
687 		   &iflib_intr_msix, 0, "# intr msix calls");
688 SYSCTL_INT(_net_iflib, OID_AUTO, task_fn_rx, CTLFLAG_RD,
689 		   &iflib_task_fn_rxs, 0, "# task_fn_rx calls");
690 SYSCTL_INT(_net_iflib, OID_AUTO, rx_intr_enables, CTLFLAG_RD,
691 		   &iflib_rx_intr_enables, 0, "# rx intr enables");
692 SYSCTL_INT(_net_iflib, OID_AUTO, fast_intrs, CTLFLAG_RD,
693 		   &iflib_fast_intrs, 0, "# fast_intr calls");
694 SYSCTL_INT(_net_iflib, OID_AUTO, rx_unavail, CTLFLAG_RD,
695 		   &iflib_rx_unavail, 0, "# times rxeof called with no available data");
696 SYSCTL_INT(_net_iflib, OID_AUTO, rx_ctx_inactive, CTLFLAG_RD,
697 		   &iflib_rx_ctx_inactive, 0, "# times rxeof called with inactive context");
698 SYSCTL_INT(_net_iflib, OID_AUTO, rx_zero_len, CTLFLAG_RD,
699 		   &iflib_rx_zero_len, 0, "# times rxeof saw zero len mbuf");
700 SYSCTL_INT(_net_iflib, OID_AUTO, rx_if_input, CTLFLAG_RD,
701 		   &iflib_rx_if_input, 0, "# times rxeof called if_input");
702 SYSCTL_INT(_net_iflib, OID_AUTO, rx_mbuf_null, CTLFLAG_RD,
703 		   &iflib_rx_mbuf_null, 0, "# times rxeof got null mbuf");
704 SYSCTL_INT(_net_iflib, OID_AUTO, rxd_flush, CTLFLAG_RD,
705 	         &iflib_rxd_flush, 0, "# times rxd_flush called");
706 SYSCTL_INT(_net_iflib, OID_AUTO, verbose_debug, CTLFLAG_RW,
707 		   &iflib_verbose_debug, 0, "enable verbose debugging");
708 
709 #define DBG_COUNTER_INC(name) atomic_add_int(&(iflib_ ## name), 1)
710 static void
711 iflib_debug_reset(void)
712 {
713 	iflib_tx_seen = iflib_tx_sent = iflib_tx_encap = iflib_rx_allocs =
714 		iflib_fl_refills = iflib_fl_refills_large = iflib_tx_frees =
715 		iflib_txq_drain_flushing = iflib_txq_drain_oactive =
716 		iflib_txq_drain_notready = iflib_txq_drain_encapfail =
717 		iflib_encap_load_mbuf_fail = iflib_encap_pad_mbuf_fail =
718 		iflib_encap_txq_avail_fail = iflib_encap_txd_encap_fail =
719 		iflib_task_fn_rxs = iflib_rx_intr_enables = iflib_fast_intrs =
720 		iflib_intr_link = iflib_intr_msix = iflib_rx_unavail =
721 		iflib_rx_ctx_inactive = iflib_rx_zero_len = iflib_rx_if_input =
722 		iflib_rx_mbuf_null = iflib_rxd_flush = 0;
723 }
724 
725 #else
726 #define DBG_COUNTER_INC(name)
727 static void iflib_debug_reset(void) {}
728 #endif
729 
730 #define IFLIB_DEBUG 0
731 
732 static void iflib_tx_structures_free(if_ctx_t ctx);
733 static void iflib_rx_structures_free(if_ctx_t ctx);
734 static int iflib_queues_alloc(if_ctx_t ctx);
735 static int iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq);
736 static int iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget);
737 static int iflib_qset_structures_setup(if_ctx_t ctx);
738 static int iflib_msix_init(if_ctx_t ctx);
739 static int iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filterarg, int *rid, const char *str);
740 static void iflib_txq_check_drain(iflib_txq_t txq, int budget);
741 static uint32_t iflib_txq_can_drain(struct ifmp_ring *);
742 static int iflib_register(if_ctx_t);
743 static void iflib_init_locked(if_ctx_t ctx);
744 static void iflib_add_device_sysctl_pre(if_ctx_t ctx);
745 static void iflib_add_device_sysctl_post(if_ctx_t ctx);
746 static void iflib_ifmp_purge(iflib_txq_t txq);
747 static void _iflib_pre_assert(if_softc_ctx_t scctx);
748 static void iflib_if_init_locked(if_ctx_t ctx);
749 #ifndef __NO_STRICT_ALIGNMENT
750 static struct mbuf * iflib_fixup_rx(struct mbuf *m);
751 #endif
752 
753 NETDUMP_DEFINE(iflib);
754 
755 #ifdef DEV_NETMAP
756 #include <sys/selinfo.h>
757 #include <net/netmap.h>
758 #include <dev/netmap/netmap_kern.h>
759 
760 MODULE_DEPEND(iflib, netmap, 1, 1, 1);
761 
762 static int netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, uint32_t nm_i, bool init);
763 
764 /*
765  * device-specific sysctl variables:
766  *
767  * iflib_crcstrip: 0: keep CRC in rx frames (default), 1: strip it.
768  *	During regular operations the CRC is stripped, but on some
769  *	hardware reception of frames not multiple of 64 is slower,
770  *	so using crcstrip=0 helps in benchmarks.
771  *
772  * iflib_rx_miss, iflib_rx_miss_bufs:
773  *	count packets that might be missed due to lost interrupts.
774  */
775 SYSCTL_DECL(_dev_netmap);
776 /*
777  * The xl driver by default strips CRCs and we do not override it.
778  */
779 
780 int iflib_crcstrip = 1;
781 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_crcstrip,
782     CTLFLAG_RW, &iflib_crcstrip, 1, "strip CRC on rx frames");
783 
784 int iflib_rx_miss, iflib_rx_miss_bufs;
785 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss,
786     CTLFLAG_RW, &iflib_rx_miss, 0, "potentially missed rx intr");
787 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss_bufs,
788     CTLFLAG_RW, &iflib_rx_miss_bufs, 0, "potentially missed rx intr bufs");
789 
790 /*
791  * Register/unregister. We are already under netmap lock.
792  * Only called on the first register or the last unregister.
793  */
794 static int
795 iflib_netmap_register(struct netmap_adapter *na, int onoff)
796 {
797 	struct ifnet *ifp = na->ifp;
798 	if_ctx_t ctx = ifp->if_softc;
799 	int status;
800 
801 	CTX_LOCK(ctx);
802 	IFDI_INTR_DISABLE(ctx);
803 
804 	/* Tell the stack that the interface is no longer active */
805 	ifp->if_drv_flags &= ~(IFF_DRV_RUNNING | IFF_DRV_OACTIVE);
806 
807 	if (!CTX_IS_VF(ctx))
808 		IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip);
809 
810 	/* enable or disable flags and callbacks in na and ifp */
811 	if (onoff) {
812 		nm_set_native_flags(na);
813 	} else {
814 		nm_clear_native_flags(na);
815 	}
816 	iflib_stop(ctx);
817 	iflib_init_locked(ctx);
818 	IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip); // XXX why twice ?
819 	status = ifp->if_drv_flags & IFF_DRV_RUNNING ? 0 : 1;
820 	if (status)
821 		nm_clear_native_flags(na);
822 	CTX_UNLOCK(ctx);
823 	return (status);
824 }
825 
826 static int
827 netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, uint32_t nm_i, bool init)
828 {
829 	struct netmap_adapter *na = kring->na;
830 	u_int const lim = kring->nkr_num_slots - 1;
831 	u_int head = kring->rhead;
832 	struct netmap_ring *ring = kring->ring;
833 	bus_dmamap_t *map;
834 	struct if_rxd_update iru;
835 	if_ctx_t ctx = rxq->ifr_ctx;
836 	iflib_fl_t fl = &rxq->ifr_fl[0];
837 	uint32_t refill_pidx, nic_i;
838 
839 	if (nm_i == head && __predict_true(!init))
840 		return 0;
841 	iru_init(&iru, rxq, 0 /* flid */);
842 	map = fl->ifl_sds.ifsd_map;
843 	refill_pidx = netmap_idx_k2n(kring, nm_i);
844 	/*
845 	 * IMPORTANT: we must leave one free slot in the ring,
846 	 * so move head back by one unit
847 	 */
848 	head = nm_prev(head, lim);
849 	nic_i = UINT_MAX;
850 	while (nm_i != head) {
851 		for (int tmp_pidx = 0; tmp_pidx < IFLIB_MAX_RX_REFRESH && nm_i != head; tmp_pidx++) {
852 			struct netmap_slot *slot = &ring->slot[nm_i];
853 			void *addr = PNMB(na, slot, &fl->ifl_bus_addrs[tmp_pidx]);
854 			uint32_t nic_i_dma = refill_pidx;
855 			nic_i = netmap_idx_k2n(kring, nm_i);
856 
857 			MPASS(tmp_pidx < IFLIB_MAX_RX_REFRESH);
858 
859 			if (addr == NETMAP_BUF_BASE(na)) /* bad buf */
860 			        return netmap_ring_reinit(kring);
861 
862 			fl->ifl_vm_addrs[tmp_pidx] = addr;
863 			if (__predict_false(init) && map) {
864 				netmap_load_map(na, fl->ifl_ifdi->idi_tag, map[nic_i], addr);
865 			} else if (map && (slot->flags & NS_BUF_CHANGED)) {
866 				/* buffer has changed, reload map */
867 				netmap_reload_map(na, fl->ifl_ifdi->idi_tag, map[nic_i], addr);
868 			}
869 			slot->flags &= ~NS_BUF_CHANGED;
870 
871 			nm_i = nm_next(nm_i, lim);
872 			fl->ifl_rxd_idxs[tmp_pidx] = nic_i = nm_next(nic_i, lim);
873 			if (nm_i != head && tmp_pidx < IFLIB_MAX_RX_REFRESH-1)
874 				continue;
875 
876 			iru.iru_pidx = refill_pidx;
877 			iru.iru_count = tmp_pidx+1;
878 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
879 
880 			refill_pidx = nic_i;
881 			if (map == NULL)
882 				continue;
883 
884 			for (int n = 0; n < iru.iru_count; n++) {
885 				bus_dmamap_sync(fl->ifl_ifdi->idi_tag, map[nic_i_dma],
886 						BUS_DMASYNC_PREREAD);
887 				/* XXX - change this to not use the netmap func*/
888 				nic_i_dma = nm_next(nic_i_dma, lim);
889 			}
890 		}
891 	}
892 	kring->nr_hwcur = head;
893 
894 	if (map)
895 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
896 				BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
897 	if (__predict_true(nic_i != UINT_MAX))
898 		ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id, nic_i);
899 	return (0);
900 }
901 
902 /*
903  * Reconcile kernel and user view of the transmit ring.
904  *
905  * All information is in the kring.
906  * Userspace wants to send packets up to the one before kring->rhead,
907  * kernel knows kring->nr_hwcur is the first unsent packet.
908  *
909  * Here we push packets out (as many as possible), and possibly
910  * reclaim buffers from previously completed transmission.
911  *
912  * The caller (netmap) guarantees that there is only one instance
913  * running at any time. Any interference with other driver
914  * methods should be handled by the individual drivers.
915  */
916 static int
917 iflib_netmap_txsync(struct netmap_kring *kring, int flags)
918 {
919 	struct netmap_adapter *na = kring->na;
920 	struct ifnet *ifp = na->ifp;
921 	struct netmap_ring *ring = kring->ring;
922 	u_int nm_i;	/* index into the netmap kring */
923 	u_int nic_i;	/* index into the NIC ring */
924 	u_int n;
925 	u_int const lim = kring->nkr_num_slots - 1;
926 	u_int const head = kring->rhead;
927 	struct if_pkt_info pi;
928 
929 	/*
930 	 * interrupts on every tx packet are expensive so request
931 	 * them every half ring, or where NS_REPORT is set
932 	 */
933 	u_int report_frequency = kring->nkr_num_slots >> 1;
934 	/* device-specific */
935 	if_ctx_t ctx = ifp->if_softc;
936 	iflib_txq_t txq = &ctx->ifc_txqs[kring->ring_id];
937 
938 	if (txq->ift_sds.ifsd_map)
939 		bus_dmamap_sync(txq->ift_desc_tag, txq->ift_ifdi->idi_map,
940 				BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
941 
942 
943 	/*
944 	 * First part: process new packets to send.
945 	 * nm_i is the current index in the netmap kring,
946 	 * nic_i is the corresponding index in the NIC ring.
947 	 *
948 	 * If we have packets to send (nm_i != head)
949 	 * iterate over the netmap ring, fetch length and update
950 	 * the corresponding slot in the NIC ring. Some drivers also
951 	 * need to update the buffer's physical address in the NIC slot
952 	 * even NS_BUF_CHANGED is not set (PNMB computes the addresses).
953 	 *
954 	 * The netmap_reload_map() calls is especially expensive,
955 	 * even when (as in this case) the tag is 0, so do only
956 	 * when the buffer has actually changed.
957 	 *
958 	 * If possible do not set the report/intr bit on all slots,
959 	 * but only a few times per ring or when NS_REPORT is set.
960 	 *
961 	 * Finally, on 10G and faster drivers, it might be useful
962 	 * to prefetch the next slot and txr entry.
963 	 */
964 
965 	nm_i = kring->nr_hwcur;
966 	if (nm_i != head) {	/* we have new packets to send */
967 		pkt_info_zero(&pi);
968 		pi.ipi_segs = txq->ift_segs;
969 		pi.ipi_qsidx = kring->ring_id;
970 		nic_i = netmap_idx_k2n(kring, nm_i);
971 
972 		__builtin_prefetch(&ring->slot[nm_i]);
973 		__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i]);
974 		if (txq->ift_sds.ifsd_map)
975 			__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i]);
976 
977 		for (n = 0; nm_i != head; n++) {
978 			struct netmap_slot *slot = &ring->slot[nm_i];
979 			u_int len = slot->len;
980 			uint64_t paddr;
981 			void *addr = PNMB(na, slot, &paddr);
982 			int flags = (slot->flags & NS_REPORT ||
983 				nic_i == 0 || nic_i == report_frequency) ?
984 				IPI_TX_INTR : 0;
985 
986 			/* device-specific */
987 			pi.ipi_len = len;
988 			pi.ipi_segs[0].ds_addr = paddr;
989 			pi.ipi_segs[0].ds_len = len;
990 			pi.ipi_nsegs = 1;
991 			pi.ipi_ndescs = 0;
992 			pi.ipi_pidx = nic_i;
993 			pi.ipi_flags = flags;
994 
995 			/* Fill the slot in the NIC ring. */
996 			ctx->isc_txd_encap(ctx->ifc_softc, &pi);
997 
998 			/* prefetch for next round */
999 			__builtin_prefetch(&ring->slot[nm_i + 1]);
1000 			__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i + 1]);
1001 			if (txq->ift_sds.ifsd_map) {
1002 				__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i + 1]);
1003 
1004 				NM_CHECK_ADDR_LEN(na, addr, len);
1005 
1006 				if (slot->flags & NS_BUF_CHANGED) {
1007 					/* buffer has changed, reload map */
1008 					netmap_reload_map(na, txq->ift_desc_tag, txq->ift_sds.ifsd_map[nic_i], addr);
1009 				}
1010 				/* make sure changes to the buffer are synced */
1011 				bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_sds.ifsd_map[nic_i],
1012 						BUS_DMASYNC_PREWRITE);
1013 			}
1014 			slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED);
1015 			nm_i = nm_next(nm_i, lim);
1016 			nic_i = nm_next(nic_i, lim);
1017 		}
1018 		kring->nr_hwcur = nm_i;
1019 
1020 		/* synchronize the NIC ring */
1021 		if (txq->ift_sds.ifsd_map)
1022 			bus_dmamap_sync(txq->ift_desc_tag, txq->ift_ifdi->idi_map,
1023 						BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1024 
1025 		/* (re)start the tx unit up to slot nic_i (excluded) */
1026 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, nic_i);
1027 	}
1028 
1029 	/*
1030 	 * Second part: reclaim buffers for completed transmissions.
1031 	 *
1032 	 * If there are unclaimed buffers, attempt to reclaim them.
1033 	 * If none are reclaimed, and TX IRQs are not in use, do an initial
1034 	 * minimal delay, then trigger the tx handler which will spin in the
1035 	 * group task queue.
1036 	 */
1037 	if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1038 		if (iflib_tx_credits_update(ctx, txq)) {
1039 			/* some tx completed, increment avail */
1040 			nic_i = txq->ift_cidx_processed;
1041 			kring->nr_hwtail = nm_prev(netmap_idx_n2k(kring, nic_i), lim);
1042 		}
1043 	}
1044 	if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ))
1045 		if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1046 			callout_reset_on(&txq->ift_timer, hz < 2000 ? 1 : hz / 1000,
1047 			    iflib_timer, txq, txq->ift_timer.c_cpu);
1048 	}
1049 	return (0);
1050 }
1051 
1052 /*
1053  * Reconcile kernel and user view of the receive ring.
1054  * Same as for the txsync, this routine must be efficient.
1055  * The caller guarantees a single invocations, but races against
1056  * the rest of the driver should be handled here.
1057  *
1058  * On call, kring->rhead is the first packet that userspace wants
1059  * to keep, and kring->rcur is the wakeup point.
1060  * The kernel has previously reported packets up to kring->rtail.
1061  *
1062  * If (flags & NAF_FORCE_READ) also check for incoming packets irrespective
1063  * of whether or not we received an interrupt.
1064  */
1065 static int
1066 iflib_netmap_rxsync(struct netmap_kring *kring, int flags)
1067 {
1068 	struct netmap_adapter *na = kring->na;
1069 	struct netmap_ring *ring = kring->ring;
1070 	uint32_t nm_i;	/* index into the netmap ring */
1071 	uint32_t nic_i;	/* index into the NIC ring */
1072 	u_int i, n;
1073 	u_int const lim = kring->nkr_num_slots - 1;
1074 	u_int const head = kring->rhead;
1075 	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
1076 	struct if_rxd_info ri;
1077 
1078 	struct ifnet *ifp = na->ifp;
1079 	if_ctx_t ctx = ifp->if_softc;
1080 	iflib_rxq_t rxq = &ctx->ifc_rxqs[kring->ring_id];
1081 	iflib_fl_t fl = rxq->ifr_fl;
1082 	if (head > lim)
1083 		return netmap_ring_reinit(kring);
1084 
1085 	/* XXX check sync modes */
1086 	for (i = 0, fl = rxq->ifr_fl; i < rxq->ifr_nfl; i++, fl++) {
1087 		if (fl->ifl_sds.ifsd_map == NULL)
1088 			continue;
1089 		bus_dmamap_sync(rxq->ifr_fl[i].ifl_desc_tag, fl->ifl_ifdi->idi_map,
1090 				BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1091 	}
1092 	/*
1093 	 * First part: import newly received packets.
1094 	 *
1095 	 * nm_i is the index of the next free slot in the netmap ring,
1096 	 * nic_i is the index of the next received packet in the NIC ring,
1097 	 * and they may differ in case if_init() has been called while
1098 	 * in netmap mode. For the receive ring we have
1099 	 *
1100 	 *	nic_i = rxr->next_check;
1101 	 *	nm_i = kring->nr_hwtail (previous)
1102 	 * and
1103 	 *	nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1104 	 *
1105 	 * rxr->next_check is set to 0 on a ring reinit
1106 	 */
1107 	if (netmap_no_pendintr || force_update) {
1108 		int crclen = iflib_crcstrip ? 0 : 4;
1109 		int error, avail;
1110 
1111 		for (i = 0; i < rxq->ifr_nfl; i++) {
1112 			fl = &rxq->ifr_fl[i];
1113 			nic_i = fl->ifl_cidx;
1114 			nm_i = netmap_idx_n2k(kring, nic_i);
1115 			avail = iflib_rxd_avail(ctx, rxq, nic_i, USHRT_MAX);
1116 			for (n = 0; avail > 0; n++, avail--) {
1117 				rxd_info_zero(&ri);
1118 				ri.iri_frags = rxq->ifr_frags;
1119 				ri.iri_qsidx = kring->ring_id;
1120 				ri.iri_ifp = ctx->ifc_ifp;
1121 				ri.iri_cidx = nic_i;
1122 
1123 				error = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
1124 				ring->slot[nm_i].len = error ? 0 : ri.iri_len - crclen;
1125 				ring->slot[nm_i].flags = 0;
1126 				if (fl->ifl_sds.ifsd_map)
1127 					bus_dmamap_sync(fl->ifl_ifdi->idi_tag,
1128 							fl->ifl_sds.ifsd_map[nic_i], BUS_DMASYNC_POSTREAD);
1129 				nm_i = nm_next(nm_i, lim);
1130 				nic_i = nm_next(nic_i, lim);
1131 			}
1132 			if (n) { /* update the state variables */
1133 				if (netmap_no_pendintr && !force_update) {
1134 					/* diagnostics */
1135 					iflib_rx_miss ++;
1136 					iflib_rx_miss_bufs += n;
1137 				}
1138 				fl->ifl_cidx = nic_i;
1139 				kring->nr_hwtail = nm_i;
1140 			}
1141 			kring->nr_kflags &= ~NKR_PENDINTR;
1142 		}
1143 	}
1144 	/*
1145 	 * Second part: skip past packets that userspace has released.
1146 	 * (kring->nr_hwcur to head excluded),
1147 	 * and make the buffers available for reception.
1148 	 * As usual nm_i is the index in the netmap ring,
1149 	 * nic_i is the index in the NIC ring, and
1150 	 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1151 	 */
1152 	/* XXX not sure how this will work with multiple free lists */
1153 	nm_i = kring->nr_hwcur;
1154 
1155 	return (netmap_fl_refill(rxq, kring, nm_i, false));
1156 }
1157 
1158 static void
1159 iflib_netmap_intr(struct netmap_adapter *na, int onoff)
1160 {
1161 	struct ifnet *ifp = na->ifp;
1162 	if_ctx_t ctx = ifp->if_softc;
1163 
1164 	CTX_LOCK(ctx);
1165 	if (onoff) {
1166 		IFDI_INTR_ENABLE(ctx);
1167 	} else {
1168 		IFDI_INTR_DISABLE(ctx);
1169 	}
1170 	CTX_UNLOCK(ctx);
1171 }
1172 
1173 
1174 static int
1175 iflib_netmap_attach(if_ctx_t ctx)
1176 {
1177 	struct netmap_adapter na;
1178 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1179 
1180 	bzero(&na, sizeof(na));
1181 
1182 	na.ifp = ctx->ifc_ifp;
1183 	na.na_flags = NAF_BDG_MAYSLEEP;
1184 	MPASS(ctx->ifc_softc_ctx.isc_ntxqsets);
1185 	MPASS(ctx->ifc_softc_ctx.isc_nrxqsets);
1186 
1187 	na.num_tx_desc = scctx->isc_ntxd[0];
1188 	na.num_rx_desc = scctx->isc_nrxd[0];
1189 	na.nm_txsync = iflib_netmap_txsync;
1190 	na.nm_rxsync = iflib_netmap_rxsync;
1191 	na.nm_register = iflib_netmap_register;
1192 	na.nm_intr = iflib_netmap_intr;
1193 	na.num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
1194 	na.num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
1195 	return (netmap_attach(&na));
1196 }
1197 
1198 static void
1199 iflib_netmap_txq_init(if_ctx_t ctx, iflib_txq_t txq)
1200 {
1201 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1202 	struct netmap_slot *slot;
1203 
1204 	slot = netmap_reset(na, NR_TX, txq->ift_id, 0);
1205 	if (slot == NULL)
1206 		return;
1207 	if (txq->ift_sds.ifsd_map == NULL)
1208 		return;
1209 
1210 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxd[0]; i++) {
1211 
1212 		/*
1213 		 * In netmap mode, set the map for the packet buffer.
1214 		 * NOTE: Some drivers (not this one) also need to set
1215 		 * the physical buffer address in the NIC ring.
1216 		 * netmap_idx_n2k() maps a nic index, i, into the corresponding
1217 		 * netmap slot index, si
1218 		 */
1219 		int si = netmap_idx_n2k(na->tx_rings[txq->ift_id], i);
1220 		netmap_load_map(na, txq->ift_desc_tag, txq->ift_sds.ifsd_map[i], NMB(na, slot + si));
1221 	}
1222 }
1223 
1224 static void
1225 iflib_netmap_rxq_init(if_ctx_t ctx, iflib_rxq_t rxq)
1226 {
1227 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1228 	struct netmap_kring *kring = na->rx_rings[rxq->ifr_id];
1229 	struct netmap_slot *slot;
1230 	uint32_t nm_i;
1231 
1232 	slot = netmap_reset(na, NR_RX, rxq->ifr_id, 0);
1233 	if (slot == NULL)
1234 		return;
1235 	nm_i = netmap_idx_n2k(kring, 0);
1236 	netmap_fl_refill(rxq, kring, nm_i, true);
1237 }
1238 
1239 static void
1240 iflib_netmap_timer_adjust(if_ctx_t ctx, uint16_t txqid, uint32_t *reset_on)
1241 {
1242 	struct netmap_kring *kring;
1243 
1244 	kring = NA(ctx->ifc_ifp)->tx_rings[txqid];
1245 
1246 	if (kring->nr_hwcur != nm_next(kring->nr_hwtail, kring->nkr_num_slots - 1)) {
1247 		if (ctx->isc_txd_credits_update(ctx->ifc_softc, txqid, false))
1248 			netmap_tx_irq(ctx->ifc_ifp, txqid);
1249 		if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ)) {
1250 			if (hz < 2000)
1251 				*reset_on = 1;
1252 			else
1253 				*reset_on = hz / 1000;
1254 		}
1255 	}
1256 }
1257 
1258 #define iflib_netmap_detach(ifp) netmap_detach(ifp)
1259 
1260 #else
1261 #define iflib_netmap_txq_init(ctx, txq)
1262 #define iflib_netmap_rxq_init(ctx, rxq)
1263 #define iflib_netmap_detach(ifp)
1264 
1265 #define iflib_netmap_attach(ctx) (0)
1266 #define netmap_rx_irq(ifp, qid, budget) (0)
1267 #define netmap_tx_irq(ifp, qid) do {} while (0)
1268 #define iflib_netmap_timer_adjust(ctx, txqid, reset_on)
1269 
1270 #endif
1271 
1272 #if defined(__i386__) || defined(__amd64__)
1273 static __inline void
1274 prefetch(void *x)
1275 {
1276 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1277 }
1278 static __inline void
1279 prefetch2cachelines(void *x)
1280 {
1281 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1282 #if (CACHE_LINE_SIZE < 128)
1283 	__asm volatile("prefetcht0 %0" :: "m" (*(((unsigned long *)x)+CACHE_LINE_SIZE/(sizeof(unsigned long)))));
1284 #endif
1285 }
1286 #else
1287 #define prefetch(x)
1288 #define prefetch2cachelines(x)
1289 #endif
1290 
1291 static void
1292 iflib_gen_mac(if_ctx_t ctx)
1293 {
1294 	struct thread *td;
1295 	MD5_CTX mdctx;
1296 	char uuid[HOSTUUIDLEN+1];
1297 	char buf[HOSTUUIDLEN+16];
1298 	uint8_t *mac;
1299 	unsigned char digest[16];
1300 
1301 	td = curthread;
1302 	mac = ctx->ifc_mac;
1303 	uuid[HOSTUUIDLEN] = 0;
1304 	bcopy(td->td_ucred->cr_prison->pr_hostuuid, uuid, HOSTUUIDLEN);
1305 	snprintf(buf, HOSTUUIDLEN+16, "%s-%s", uuid, device_get_nameunit(ctx->ifc_dev));
1306 	/*
1307 	 * Generate a pseudo-random, deterministic MAC
1308 	 * address based on the UUID and unit number.
1309 	 * The FreeBSD Foundation OUI of 58-9C-FC is used.
1310 	 */
1311 	MD5Init(&mdctx);
1312 	MD5Update(&mdctx, buf, strlen(buf));
1313 	MD5Final(digest, &mdctx);
1314 
1315 	mac[0] = 0x58;
1316 	mac[1] = 0x9C;
1317 	mac[2] = 0xFC;
1318 	mac[3] = digest[0];
1319 	mac[4] = digest[1];
1320 	mac[5] = digest[2];
1321 }
1322 
1323 static void
1324 iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid)
1325 {
1326 	iflib_fl_t fl;
1327 
1328 	fl = &rxq->ifr_fl[flid];
1329 	iru->iru_paddrs = fl->ifl_bus_addrs;
1330 	iru->iru_vaddrs = &fl->ifl_vm_addrs[0];
1331 	iru->iru_idxs = fl->ifl_rxd_idxs;
1332 	iru->iru_qsidx = rxq->ifr_id;
1333 	iru->iru_buf_size = fl->ifl_buf_size;
1334 	iru->iru_flidx = fl->ifl_id;
1335 }
1336 
1337 static void
1338 _iflib_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int err)
1339 {
1340 	if (err)
1341 		return;
1342 	*(bus_addr_t *) arg = segs[0].ds_addr;
1343 }
1344 
1345 int
1346 iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags)
1347 {
1348 	int err;
1349 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1350 	device_t dev = ctx->ifc_dev;
1351 
1352 	KASSERT(sctx->isc_q_align != 0, ("alignment value not initialized"));
1353 
1354 	err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1355 				sctx->isc_q_align, 0,	/* alignment, bounds */
1356 				BUS_SPACE_MAXADDR,	/* lowaddr */
1357 				BUS_SPACE_MAXADDR,	/* highaddr */
1358 				NULL, NULL,		/* filter, filterarg */
1359 				size,			/* maxsize */
1360 				1,			/* nsegments */
1361 				size,			/* maxsegsize */
1362 				BUS_DMA_ALLOCNOW,	/* flags */
1363 				NULL,			/* lockfunc */
1364 				NULL,			/* lockarg */
1365 				&dma->idi_tag);
1366 	if (err) {
1367 		device_printf(dev,
1368 		    "%s: bus_dma_tag_create failed: %d\n",
1369 		    __func__, err);
1370 		goto fail_0;
1371 	}
1372 
1373 	err = bus_dmamem_alloc(dma->idi_tag, (void**) &dma->idi_vaddr,
1374 	    BUS_DMA_NOWAIT | BUS_DMA_COHERENT | BUS_DMA_ZERO, &dma->idi_map);
1375 	if (err) {
1376 		device_printf(dev,
1377 		    "%s: bus_dmamem_alloc(%ju) failed: %d\n",
1378 		    __func__, (uintmax_t)size, err);
1379 		goto fail_1;
1380 	}
1381 
1382 	dma->idi_paddr = IF_BAD_DMA;
1383 	err = bus_dmamap_load(dma->idi_tag, dma->idi_map, dma->idi_vaddr,
1384 	    size, _iflib_dmamap_cb, &dma->idi_paddr, mapflags | BUS_DMA_NOWAIT);
1385 	if (err || dma->idi_paddr == IF_BAD_DMA) {
1386 		device_printf(dev,
1387 		    "%s: bus_dmamap_load failed: %d\n",
1388 		    __func__, err);
1389 		goto fail_2;
1390 	}
1391 
1392 	dma->idi_size = size;
1393 	return (0);
1394 
1395 fail_2:
1396 	bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1397 fail_1:
1398 	bus_dma_tag_destroy(dma->idi_tag);
1399 fail_0:
1400 	dma->idi_tag = NULL;
1401 
1402 	return (err);
1403 }
1404 
1405 int
1406 iflib_dma_alloc_multi(if_ctx_t ctx, int *sizes, iflib_dma_info_t *dmalist, int mapflags, int count)
1407 {
1408 	int i, err;
1409 	iflib_dma_info_t *dmaiter;
1410 
1411 	dmaiter = dmalist;
1412 	for (i = 0; i < count; i++, dmaiter++) {
1413 		if ((err = iflib_dma_alloc(ctx, sizes[i], *dmaiter, mapflags)) != 0)
1414 			break;
1415 	}
1416 	if (err)
1417 		iflib_dma_free_multi(dmalist, i);
1418 	return (err);
1419 }
1420 
1421 void
1422 iflib_dma_free(iflib_dma_info_t dma)
1423 {
1424 	if (dma->idi_tag == NULL)
1425 		return;
1426 	if (dma->idi_paddr != IF_BAD_DMA) {
1427 		bus_dmamap_sync(dma->idi_tag, dma->idi_map,
1428 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1429 		bus_dmamap_unload(dma->idi_tag, dma->idi_map);
1430 		dma->idi_paddr = IF_BAD_DMA;
1431 	}
1432 	if (dma->idi_vaddr != NULL) {
1433 		bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1434 		dma->idi_vaddr = NULL;
1435 	}
1436 	bus_dma_tag_destroy(dma->idi_tag);
1437 	dma->idi_tag = NULL;
1438 }
1439 
1440 void
1441 iflib_dma_free_multi(iflib_dma_info_t *dmalist, int count)
1442 {
1443 	int i;
1444 	iflib_dma_info_t *dmaiter = dmalist;
1445 
1446 	for (i = 0; i < count; i++, dmaiter++)
1447 		iflib_dma_free(*dmaiter);
1448 }
1449 
1450 #ifdef EARLY_AP_STARTUP
1451 static const int iflib_started = 1;
1452 #else
1453 /*
1454  * We used to abuse the smp_started flag to decide if the queues have been
1455  * fully initialized (by late taskqgroup_adjust() calls in a SYSINIT()).
1456  * That gave bad races, since the SYSINIT() runs strictly after smp_started
1457  * is set.  Run a SYSINIT() strictly after that to just set a usable
1458  * completion flag.
1459  */
1460 
1461 static int iflib_started;
1462 
1463 static void
1464 iflib_record_started(void *arg)
1465 {
1466 	iflib_started = 1;
1467 }
1468 
1469 SYSINIT(iflib_record_started, SI_SUB_SMP + 1, SI_ORDER_FIRST,
1470 	iflib_record_started, NULL);
1471 #endif
1472 
1473 static int
1474 iflib_fast_intr(void *arg)
1475 {
1476 	iflib_filter_info_t info = arg;
1477 	struct grouptask *gtask = info->ifi_task;
1478 	if (!iflib_started)
1479 		return (FILTER_HANDLED);
1480 
1481 	DBG_COUNTER_INC(fast_intrs);
1482 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1483 		return (FILTER_HANDLED);
1484 
1485 	GROUPTASK_ENQUEUE(gtask);
1486 	return (FILTER_HANDLED);
1487 }
1488 
1489 static int
1490 iflib_fast_intr_rxtx(void *arg)
1491 {
1492 	iflib_filter_info_t info = arg;
1493 	struct grouptask *gtask = info->ifi_task;
1494 	iflib_rxq_t rxq = (iflib_rxq_t)info->ifi_ctx;
1495 	if_ctx_t ctx = NULL;;
1496 	int i, cidx;
1497 
1498 	if (!iflib_started)
1499 		return (FILTER_HANDLED);
1500 
1501 	DBG_COUNTER_INC(fast_intrs);
1502 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1503 		return (FILTER_HANDLED);
1504 
1505 	MPASS(rxq->ifr_ntxqirq);
1506 	for (i = 0; i < rxq->ifr_ntxqirq; i++) {
1507 		qidx_t txqid = rxq->ifr_txqid[i];
1508 
1509 		ctx = rxq->ifr_ctx;
1510 
1511 		if (!ctx->isc_txd_credits_update(ctx->ifc_softc, txqid, false)) {
1512 			IFDI_TX_QUEUE_INTR_ENABLE(ctx, txqid);
1513 			continue;
1514 		}
1515 		GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
1516 	}
1517 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_RXCQ)
1518 		cidx = rxq->ifr_cq_cidx;
1519 	else
1520 		cidx = rxq->ifr_fl[0].ifl_cidx;
1521 	if (iflib_rxd_avail(ctx, rxq, cidx, 1))
1522 		GROUPTASK_ENQUEUE(gtask);
1523 	else
1524 		IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
1525 	return (FILTER_HANDLED);
1526 }
1527 
1528 
1529 static int
1530 iflib_fast_intr_ctx(void *arg)
1531 {
1532 	iflib_filter_info_t info = arg;
1533 	struct grouptask *gtask = info->ifi_task;
1534 
1535 	if (!iflib_started)
1536 		return (FILTER_HANDLED);
1537 
1538 	DBG_COUNTER_INC(fast_intrs);
1539 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1540 		return (FILTER_HANDLED);
1541 
1542 	GROUPTASK_ENQUEUE(gtask);
1543 	return (FILTER_HANDLED);
1544 }
1545 
1546 static int
1547 _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
1548 		 driver_filter_t filter, driver_intr_t handler, void *arg,
1549 		 const char *name)
1550 {
1551 	int rc, flags;
1552 	struct resource *res;
1553 	void *tag = NULL;
1554 	device_t dev = ctx->ifc_dev;
1555 
1556 	flags = RF_ACTIVE;
1557 	if (ctx->ifc_flags & IFC_LEGACY)
1558 		flags |= RF_SHAREABLE;
1559 	MPASS(rid < 512);
1560 	irq->ii_rid = rid;
1561 	res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &irq->ii_rid, flags);
1562 	if (res == NULL) {
1563 		device_printf(dev,
1564 		    "failed to allocate IRQ for rid %d, name %s.\n", rid, name);
1565 		return (ENOMEM);
1566 	}
1567 	irq->ii_res = res;
1568 	KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL"));
1569 	rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET,
1570 						filter, handler, arg, &tag);
1571 	if (rc != 0) {
1572 		device_printf(dev,
1573 		    "failed to setup interrupt for rid %d, name %s: %d\n",
1574 					  rid, name ? name : "unknown", rc);
1575 		return (rc);
1576 	} else if (name)
1577 		bus_describe_intr(dev, res, tag, "%s", name);
1578 
1579 	irq->ii_tag = tag;
1580 	return (0);
1581 }
1582 
1583 
1584 /*********************************************************************
1585  *
1586  *  Allocate memory for tx_buffer structures. The tx_buffer stores all
1587  *  the information needed to transmit a packet on the wire. This is
1588  *  called only once at attach, setup is done every reset.
1589  *
1590  **********************************************************************/
1591 
1592 static int
1593 iflib_txsd_alloc(iflib_txq_t txq)
1594 {
1595 	if_ctx_t ctx = txq->ift_ctx;
1596 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1597 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1598 	device_t dev = ctx->ifc_dev;
1599 	bus_size_t tsomaxsize;
1600 	int err, nsegments, ntsosegments;
1601 
1602 	nsegments = scctx->isc_tx_nsegments;
1603 	ntsosegments = scctx->isc_tx_tso_segments_max;
1604 	tsomaxsize = scctx->isc_tx_tso_size_max;
1605 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_VLAN_MTU)
1606 		tsomaxsize += sizeof(struct ether_vlan_header);
1607 	MPASS(scctx->isc_ntxd[0] > 0);
1608 	MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0);
1609 	MPASS(nsegments > 0);
1610 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) {
1611 		MPASS(ntsosegments > 0);
1612 		MPASS(sctx->isc_tso_maxsize >= tsomaxsize);
1613 	}
1614 
1615 	/*
1616 	 * Setup DMA descriptor areas.
1617 	 */
1618 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1619 			       1, 0,			/* alignment, bounds */
1620 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1621 			       BUS_SPACE_MAXADDR,	/* highaddr */
1622 			       NULL, NULL,		/* filter, filterarg */
1623 			       sctx->isc_tx_maxsize,		/* maxsize */
1624 			       nsegments,	/* nsegments */
1625 			       sctx->isc_tx_maxsegsize,	/* maxsegsize */
1626 			       0,			/* flags */
1627 			       NULL,			/* lockfunc */
1628 			       NULL,			/* lockfuncarg */
1629 			       &txq->ift_desc_tag))) {
1630 		device_printf(dev,"Unable to allocate TX DMA tag: %d\n", err);
1631 		device_printf(dev,"maxsize: %ju nsegments: %d maxsegsize: %ju\n",
1632 		    (uintmax_t)sctx->isc_tx_maxsize, nsegments, (uintmax_t)sctx->isc_tx_maxsegsize);
1633 		goto fail;
1634 	}
1635 	if ((if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) &
1636 	    (err = bus_dma_tag_create(bus_get_dma_tag(dev),
1637 			       1, 0,			/* alignment, bounds */
1638 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1639 			       BUS_SPACE_MAXADDR,	/* highaddr */
1640 			       NULL, NULL,		/* filter, filterarg */
1641 			       tsomaxsize,		/* maxsize */
1642 			       ntsosegments,	/* nsegments */
1643 			       sctx->isc_tso_maxsegsize,/* maxsegsize */
1644 			       0,			/* flags */
1645 			       NULL,			/* lockfunc */
1646 			       NULL,			/* lockfuncarg */
1647 			       &txq->ift_tso_desc_tag))) {
1648 		device_printf(dev,"Unable to allocate TX TSO DMA tag: %d\n", err);
1649 
1650 		goto fail;
1651 	}
1652 	if (!(txq->ift_sds.ifsd_flags =
1653 	    (uint8_t *) malloc(sizeof(uint8_t) *
1654 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1655 		device_printf(dev, "Unable to allocate tx_buffer memory\n");
1656 		err = ENOMEM;
1657 		goto fail;
1658 	}
1659 	if (!(txq->ift_sds.ifsd_m =
1660 	    (struct mbuf **) malloc(sizeof(struct mbuf *) *
1661 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1662 		device_printf(dev, "Unable to allocate tx_buffer memory\n");
1663 		err = ENOMEM;
1664 		goto fail;
1665 	}
1666 
1667         /* Create the descriptor buffer dma maps */
1668 #if defined(ACPI_DMAR) || (! (defined(__i386__) || defined(__amd64__)))
1669 	if ((ctx->ifc_flags & IFC_DMAR) == 0)
1670 		return (0);
1671 
1672 	if (!(txq->ift_sds.ifsd_map =
1673 	    (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1674 		device_printf(dev, "Unable to allocate tx_buffer map memory\n");
1675 		err = ENOMEM;
1676 		goto fail;
1677 	}
1678 
1679 	for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) {
1680 		err = bus_dmamap_create(txq->ift_desc_tag, 0, &txq->ift_sds.ifsd_map[i]);
1681 		if (err != 0) {
1682 			device_printf(dev, "Unable to create TX DMA map\n");
1683 			goto fail;
1684 		}
1685 	}
1686 #endif
1687 	return (0);
1688 fail:
1689 	/* We free all, it handles case where we are in the middle */
1690 	iflib_tx_structures_free(ctx);
1691 	return (err);
1692 }
1693 
1694 static void
1695 iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i)
1696 {
1697 	bus_dmamap_t map;
1698 
1699 	map = NULL;
1700 	if (txq->ift_sds.ifsd_map != NULL)
1701 		map = txq->ift_sds.ifsd_map[i];
1702 	if (map != NULL) {
1703 		bus_dmamap_unload(txq->ift_desc_tag, map);
1704 		bus_dmamap_destroy(txq->ift_desc_tag, map);
1705 		txq->ift_sds.ifsd_map[i] = NULL;
1706 	}
1707 }
1708 
1709 static void
1710 iflib_txq_destroy(iflib_txq_t txq)
1711 {
1712 	if_ctx_t ctx = txq->ift_ctx;
1713 
1714 	for (int i = 0; i < txq->ift_size; i++)
1715 		iflib_txsd_destroy(ctx, txq, i);
1716 	if (txq->ift_sds.ifsd_map != NULL) {
1717 		free(txq->ift_sds.ifsd_map, M_IFLIB);
1718 		txq->ift_sds.ifsd_map = NULL;
1719 	}
1720 	if (txq->ift_sds.ifsd_m != NULL) {
1721 		free(txq->ift_sds.ifsd_m, M_IFLIB);
1722 		txq->ift_sds.ifsd_m = NULL;
1723 	}
1724 	if (txq->ift_sds.ifsd_flags != NULL) {
1725 		free(txq->ift_sds.ifsd_flags, M_IFLIB);
1726 		txq->ift_sds.ifsd_flags = NULL;
1727 	}
1728 	if (txq->ift_desc_tag != NULL) {
1729 		bus_dma_tag_destroy(txq->ift_desc_tag);
1730 		txq->ift_desc_tag = NULL;
1731 	}
1732 	if (txq->ift_tso_desc_tag != NULL) {
1733 		bus_dma_tag_destroy(txq->ift_tso_desc_tag);
1734 		txq->ift_tso_desc_tag = NULL;
1735 	}
1736 }
1737 
1738 static void
1739 iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i)
1740 {
1741 	struct mbuf **mp;
1742 
1743 	mp = &txq->ift_sds.ifsd_m[i];
1744 	if (*mp == NULL)
1745 		return;
1746 
1747 	if (txq->ift_sds.ifsd_map != NULL) {
1748 		bus_dmamap_sync(txq->ift_desc_tag,
1749 				txq->ift_sds.ifsd_map[i],
1750 				BUS_DMASYNC_POSTWRITE);
1751 		bus_dmamap_unload(txq->ift_desc_tag,
1752 				  txq->ift_sds.ifsd_map[i]);
1753 	}
1754 	m_free(*mp);
1755 	DBG_COUNTER_INC(tx_frees);
1756 	*mp = NULL;
1757 }
1758 
1759 static int
1760 iflib_txq_setup(iflib_txq_t txq)
1761 {
1762 	if_ctx_t ctx = txq->ift_ctx;
1763 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1764 	iflib_dma_info_t di;
1765 	int i;
1766 
1767 	/* Set number of descriptors available */
1768 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
1769 	/* XXX make configurable */
1770 	txq->ift_update_freq = IFLIB_DEFAULT_TX_UPDATE_FREQ;
1771 
1772 	/* Reset indices */
1773 	txq->ift_cidx_processed = 0;
1774 	txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0;
1775 	txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset];
1776 
1777 	for (i = 0, di = txq->ift_ifdi; i < ctx->ifc_nhwtxqs; i++, di++)
1778 		bzero((void *)di->idi_vaddr, di->idi_size);
1779 
1780 	IFDI_TXQ_SETUP(ctx, txq->ift_id);
1781 	for (i = 0, di = txq->ift_ifdi; i < ctx->ifc_nhwtxqs; i++, di++)
1782 		bus_dmamap_sync(di->idi_tag, di->idi_map,
1783 						BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1784 	return (0);
1785 }
1786 
1787 /*********************************************************************
1788  *
1789  *  Allocate memory for rx_buffer structures. Since we use one
1790  *  rx_buffer per received packet, the maximum number of rx_buffer's
1791  *  that we'll need is equal to the number of receive descriptors
1792  *  that we've allocated.
1793  *
1794  **********************************************************************/
1795 static int
1796 iflib_rxsd_alloc(iflib_rxq_t rxq)
1797 {
1798 	if_ctx_t ctx = rxq->ifr_ctx;
1799 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1800 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1801 	device_t dev = ctx->ifc_dev;
1802 	iflib_fl_t fl;
1803 	int			err;
1804 
1805 	MPASS(scctx->isc_nrxd[0] > 0);
1806 	MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0);
1807 
1808 	fl = rxq->ifr_fl;
1809 	for (int i = 0; i <  rxq->ifr_nfl; i++, fl++) {
1810 		fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */
1811 		err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1812 					 1, 0,			/* alignment, bounds */
1813 					 BUS_SPACE_MAXADDR,	/* lowaddr */
1814 					 BUS_SPACE_MAXADDR,	/* highaddr */
1815 					 NULL, NULL,		/* filter, filterarg */
1816 					 sctx->isc_rx_maxsize,	/* maxsize */
1817 					 sctx->isc_rx_nsegments,	/* nsegments */
1818 					 sctx->isc_rx_maxsegsize,	/* maxsegsize */
1819 					 0,			/* flags */
1820 					 NULL,			/* lockfunc */
1821 					 NULL,			/* lockarg */
1822 					 &fl->ifl_desc_tag);
1823 		if (err) {
1824 			device_printf(dev, "%s: bus_dma_tag_create failed %d\n",
1825 				__func__, err);
1826 			goto fail;
1827 		}
1828 		if (!(fl->ifl_sds.ifsd_flags =
1829 		      (uint8_t *) malloc(sizeof(uint8_t) *
1830 					 scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1831 			device_printf(dev, "Unable to allocate tx_buffer memory\n");
1832 			err = ENOMEM;
1833 			goto fail;
1834 		}
1835 		if (!(fl->ifl_sds.ifsd_m =
1836 		      (struct mbuf **) malloc(sizeof(struct mbuf *) *
1837 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1838 			device_printf(dev, "Unable to allocate tx_buffer memory\n");
1839 			err = ENOMEM;
1840 			goto fail;
1841 		}
1842 		if (!(fl->ifl_sds.ifsd_cl =
1843 		      (caddr_t *) malloc(sizeof(caddr_t) *
1844 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1845 			device_printf(dev, "Unable to allocate tx_buffer memory\n");
1846 			err = ENOMEM;
1847 			goto fail;
1848 		}
1849 
1850 		/* Create the descriptor buffer dma maps */
1851 #if defined(ACPI_DMAR) || (! (defined(__i386__) || defined(__amd64__)))
1852 		if ((ctx->ifc_flags & IFC_DMAR) == 0)
1853 			continue;
1854 
1855 		if (!(fl->ifl_sds.ifsd_map =
1856 		      (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1857 			device_printf(dev, "Unable to allocate tx_buffer map memory\n");
1858 			err = ENOMEM;
1859 			goto fail;
1860 		}
1861 
1862 		for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++) {
1863 			err = bus_dmamap_create(fl->ifl_desc_tag, 0, &fl->ifl_sds.ifsd_map[i]);
1864 			if (err != 0) {
1865 				device_printf(dev, "Unable to create RX buffer DMA map\n");
1866 				goto fail;
1867 			}
1868 		}
1869 #endif
1870 	}
1871 	return (0);
1872 
1873 fail:
1874 	iflib_rx_structures_free(ctx);
1875 	return (err);
1876 }
1877 
1878 
1879 /*
1880  * Internal service routines
1881  */
1882 
1883 struct rxq_refill_cb_arg {
1884 	int               error;
1885 	bus_dma_segment_t seg;
1886 	int               nseg;
1887 };
1888 
1889 static void
1890 _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
1891 {
1892 	struct rxq_refill_cb_arg *cb_arg = arg;
1893 
1894 	cb_arg->error = error;
1895 	cb_arg->seg = segs[0];
1896 	cb_arg->nseg = nseg;
1897 }
1898 
1899 
1900 #ifdef ACPI_DMAR
1901 #define IS_DMAR(ctx) (ctx->ifc_flags & IFC_DMAR)
1902 #else
1903 #define IS_DMAR(ctx) (0)
1904 #endif
1905 
1906 /**
1907  *	rxq_refill - refill an rxq  free-buffer list
1908  *	@ctx: the iflib context
1909  *	@rxq: the free-list to refill
1910  *	@n: the number of new buffers to allocate
1911  *
1912  *	(Re)populate an rxq free-buffer list with up to @n new packet buffers.
1913  *	The caller must assure that @n does not exceed the queue's capacity.
1914  */
1915 static void
1916 _iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
1917 {
1918 	struct mbuf *m;
1919 	int idx, frag_idx = fl->ifl_fragidx;
1920         int pidx = fl->ifl_pidx;
1921 	caddr_t cl, *sd_cl;
1922 	struct mbuf **sd_m;
1923 	uint8_t *sd_flags;
1924 	struct if_rxd_update iru;
1925 	bus_dmamap_t *sd_map;
1926 	int n, i = 0;
1927 	uint64_t bus_addr;
1928 	int err;
1929 	qidx_t credits;
1930 
1931 	sd_m = fl->ifl_sds.ifsd_m;
1932 	sd_map = fl->ifl_sds.ifsd_map;
1933 	sd_cl = fl->ifl_sds.ifsd_cl;
1934 	sd_flags = fl->ifl_sds.ifsd_flags;
1935 	idx = pidx;
1936 	credits = fl->ifl_credits;
1937 
1938 	n  = count;
1939 	MPASS(n > 0);
1940 	MPASS(credits + n <= fl->ifl_size);
1941 
1942 	if (pidx < fl->ifl_cidx)
1943 		MPASS(pidx + n <= fl->ifl_cidx);
1944 	if (pidx == fl->ifl_cidx && (credits < fl->ifl_size))
1945 		MPASS(fl->ifl_gen == 0);
1946 	if (pidx > fl->ifl_cidx)
1947 		MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx);
1948 
1949 	DBG_COUNTER_INC(fl_refills);
1950 	if (n > 8)
1951 		DBG_COUNTER_INC(fl_refills_large);
1952 	iru_init(&iru, fl->ifl_rxq, fl->ifl_id);
1953 	while (n--) {
1954 		/*
1955 		 * We allocate an uninitialized mbuf + cluster, mbuf is
1956 		 * initialized after rx.
1957 		 *
1958 		 * If the cluster is still set then we know a minimum sized packet was received
1959 		 */
1960 		bit_ffc_at(fl->ifl_rx_bitmap, frag_idx, fl->ifl_size,  &frag_idx);
1961 		if ((frag_idx < 0) || (frag_idx >= fl->ifl_size))
1962                 	bit_ffc(fl->ifl_rx_bitmap, fl->ifl_size, &frag_idx);
1963 		if ((cl = sd_cl[frag_idx]) == NULL) {
1964                        if ((cl = sd_cl[frag_idx] = m_cljget(NULL, M_NOWAIT, fl->ifl_buf_size)) == NULL)
1965 				break;
1966 #if MEMORY_LOGGING
1967 			fl->ifl_cl_enqueued++;
1968 #endif
1969 		}
1970 		if ((m = m_gethdr(M_NOWAIT, MT_NOINIT)) == NULL) {
1971 			break;
1972 		}
1973 #if MEMORY_LOGGING
1974 		fl->ifl_m_enqueued++;
1975 #endif
1976 
1977 		DBG_COUNTER_INC(rx_allocs);
1978 #if defined(__i386__) || defined(__amd64__)
1979 		if (!IS_DMAR(ctx)) {
1980 			bus_addr = pmap_kextract((vm_offset_t)cl);
1981 		} else
1982 #endif
1983 		{
1984 			struct rxq_refill_cb_arg cb_arg;
1985 
1986 			cb_arg.error = 0;
1987 			MPASS(sd_map != NULL);
1988 			MPASS(sd_map[frag_idx] != NULL);
1989 			err = bus_dmamap_load(fl->ifl_desc_tag, sd_map[frag_idx],
1990 		         cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg, 0);
1991 			bus_dmamap_sync(fl->ifl_desc_tag, sd_map[frag_idx],
1992 					BUS_DMASYNC_PREREAD);
1993 
1994 			if (err != 0 || cb_arg.error) {
1995 				/*
1996 				 * !zone_pack ?
1997 				 */
1998 				if (fl->ifl_zone == zone_pack)
1999 					uma_zfree(fl->ifl_zone, cl);
2000 				m_free(m);
2001 				n = 0;
2002 				goto done;
2003 			}
2004 			bus_addr = cb_arg.seg.ds_addr;
2005 		}
2006                 bit_set(fl->ifl_rx_bitmap, frag_idx);
2007 		sd_flags[frag_idx] |= RX_SW_DESC_INUSE;
2008 
2009 		MPASS(sd_m[frag_idx] == NULL);
2010 		sd_cl[frag_idx] = cl;
2011 		sd_m[frag_idx] = m;
2012 		fl->ifl_rxd_idxs[i] = frag_idx;
2013 		fl->ifl_bus_addrs[i] = bus_addr;
2014 		fl->ifl_vm_addrs[i] = cl;
2015 		credits++;
2016 		i++;
2017 		MPASS(credits <= fl->ifl_size);
2018 		if (++idx == fl->ifl_size) {
2019 			fl->ifl_gen = 1;
2020 			idx = 0;
2021 		}
2022 		if (n == 0 || i == IFLIB_MAX_RX_REFRESH) {
2023 			iru.iru_pidx = pidx;
2024 			iru.iru_count = i;
2025 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2026 			i = 0;
2027 			pidx = idx;
2028 			fl->ifl_pidx = idx;
2029 			fl->ifl_credits = credits;
2030 		}
2031 
2032 	}
2033 done:
2034 	if (i) {
2035 		iru.iru_pidx = pidx;
2036 		iru.iru_count = i;
2037 		ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2038 		fl->ifl_pidx = idx;
2039 		fl->ifl_credits = credits;
2040 	}
2041 	DBG_COUNTER_INC(rxd_flush);
2042 	if (fl->ifl_pidx == 0)
2043 		pidx = fl->ifl_size - 1;
2044 	else
2045 		pidx = fl->ifl_pidx - 1;
2046 
2047 	if (sd_map)
2048 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2049 				BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2050 	ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id, fl->ifl_id, pidx);
2051 	fl->ifl_fragidx = frag_idx;
2052 }
2053 
2054 static __inline void
2055 __iflib_fl_refill_lt(if_ctx_t ctx, iflib_fl_t fl, int max)
2056 {
2057 	/* we avoid allowing pidx to catch up with cidx as it confuses ixl */
2058 	int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1;
2059 #ifdef INVARIANTS
2060 	int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1;
2061 #endif
2062 
2063 	MPASS(fl->ifl_credits <= fl->ifl_size);
2064 	MPASS(reclaimable == delta);
2065 
2066 	if (reclaimable > 0)
2067 		_iflib_fl_refill(ctx, fl, min(max, reclaimable));
2068 }
2069 
2070 static void
2071 iflib_fl_bufs_free(iflib_fl_t fl)
2072 {
2073 	iflib_dma_info_t idi = fl->ifl_ifdi;
2074 	uint32_t i;
2075 
2076 	for (i = 0; i < fl->ifl_size; i++) {
2077 		struct mbuf **sd_m = &fl->ifl_sds.ifsd_m[i];
2078 		uint8_t *sd_flags = &fl->ifl_sds.ifsd_flags[i];
2079 		caddr_t *sd_cl = &fl->ifl_sds.ifsd_cl[i];
2080 
2081 		if (*sd_flags & RX_SW_DESC_INUSE) {
2082 			if (fl->ifl_sds.ifsd_map != NULL) {
2083 				bus_dmamap_t sd_map = fl->ifl_sds.ifsd_map[i];
2084 				bus_dmamap_unload(fl->ifl_desc_tag, sd_map);
2085 				if (fl->ifl_rxq->ifr_ctx->ifc_in_detach)
2086 					bus_dmamap_destroy(fl->ifl_desc_tag, sd_map);
2087 			}
2088 			if (*sd_m != NULL) {
2089 				m_init(*sd_m, M_NOWAIT, MT_DATA, 0);
2090 				uma_zfree(zone_mbuf, *sd_m);
2091 			}
2092 			if (*sd_cl != NULL)
2093 				uma_zfree(fl->ifl_zone, *sd_cl);
2094 			*sd_flags = 0;
2095 		} else {
2096 			MPASS(*sd_cl == NULL);
2097 			MPASS(*sd_m == NULL);
2098 		}
2099 #if MEMORY_LOGGING
2100 		fl->ifl_m_dequeued++;
2101 		fl->ifl_cl_dequeued++;
2102 #endif
2103 		*sd_cl = NULL;
2104 		*sd_m = NULL;
2105 	}
2106 #ifdef INVARIANTS
2107 	for (i = 0; i < fl->ifl_size; i++) {
2108 		MPASS(fl->ifl_sds.ifsd_flags[i] == 0);
2109 		MPASS(fl->ifl_sds.ifsd_cl[i] == NULL);
2110 		MPASS(fl->ifl_sds.ifsd_m[i] == NULL);
2111 	}
2112 #endif
2113 	/*
2114 	 * Reset free list values
2115 	 */
2116 	fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = fl->ifl_fragidx = 0;
2117 	bzero(idi->idi_vaddr, idi->idi_size);
2118 }
2119 
2120 /*********************************************************************
2121  *
2122  *  Initialize a receive ring and its buffers.
2123  *
2124  **********************************************************************/
2125 static int
2126 iflib_fl_setup(iflib_fl_t fl)
2127 {
2128 	iflib_rxq_t rxq = fl->ifl_rxq;
2129 	if_ctx_t ctx = rxq->ifr_ctx;
2130 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2131 
2132 	bit_nclear(fl->ifl_rx_bitmap, 0, fl->ifl_size - 1);
2133 	/*
2134 	** Free current RX buffer structs and their mbufs
2135 	*/
2136 	iflib_fl_bufs_free(fl);
2137 	/* Now replenish the mbufs */
2138 	MPASS(fl->ifl_credits == 0);
2139 	/*
2140 	 * XXX don't set the max_frame_size to larger
2141 	 * than the hardware can handle
2142 	 */
2143 	if (sctx->isc_max_frame_size <= 2048)
2144 		fl->ifl_buf_size = MCLBYTES;
2145 #ifndef CONTIGMALLOC_WORKS
2146 	else
2147 		fl->ifl_buf_size = MJUMPAGESIZE;
2148 #else
2149 	else if (sctx->isc_max_frame_size <= 4096)
2150 		fl->ifl_buf_size = MJUMPAGESIZE;
2151 	else if (sctx->isc_max_frame_size <= 9216)
2152 		fl->ifl_buf_size = MJUM9BYTES;
2153 	else
2154 		fl->ifl_buf_size = MJUM16BYTES;
2155 #endif
2156 	if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size)
2157 		ctx->ifc_max_fl_buf_size = fl->ifl_buf_size;
2158 	fl->ifl_cltype = m_gettype(fl->ifl_buf_size);
2159 	fl->ifl_zone = m_getzone(fl->ifl_buf_size);
2160 
2161 
2162 	/* avoid pre-allocating zillions of clusters to an idle card
2163 	 * potentially speeding up attach
2164 	 */
2165 	_iflib_fl_refill(ctx, fl, min(128, fl->ifl_size));
2166 	MPASS(min(128, fl->ifl_size) == fl->ifl_credits);
2167 	if (min(128, fl->ifl_size) != fl->ifl_credits)
2168 		return (ENOBUFS);
2169 	/*
2170 	 * handle failure
2171 	 */
2172 	MPASS(rxq != NULL);
2173 	MPASS(fl->ifl_ifdi != NULL);
2174 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2175 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2176 	return (0);
2177 }
2178 
2179 /*********************************************************************
2180  *
2181  *  Free receive ring data structures
2182  *
2183  **********************************************************************/
2184 static void
2185 iflib_rx_sds_free(iflib_rxq_t rxq)
2186 {
2187 	iflib_fl_t fl;
2188 	int i;
2189 
2190 	if (rxq->ifr_fl != NULL) {
2191 		for (i = 0; i < rxq->ifr_nfl; i++) {
2192 			fl = &rxq->ifr_fl[i];
2193 			if (fl->ifl_desc_tag != NULL) {
2194 				bus_dma_tag_destroy(fl->ifl_desc_tag);
2195 				fl->ifl_desc_tag = NULL;
2196 			}
2197 			free(fl->ifl_sds.ifsd_m, M_IFLIB);
2198 			free(fl->ifl_sds.ifsd_cl, M_IFLIB);
2199 			/* XXX destroy maps first */
2200 			free(fl->ifl_sds.ifsd_map, M_IFLIB);
2201 			fl->ifl_sds.ifsd_m = NULL;
2202 			fl->ifl_sds.ifsd_cl = NULL;
2203 			fl->ifl_sds.ifsd_map = NULL;
2204 		}
2205 		free(rxq->ifr_fl, M_IFLIB);
2206 		rxq->ifr_fl = NULL;
2207 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
2208 	}
2209 }
2210 
2211 /*
2212  * MI independent logic
2213  *
2214  */
2215 static void
2216 iflib_timer(void *arg)
2217 {
2218 	iflib_txq_t txq = arg;
2219 	if_ctx_t ctx = txq->ift_ctx;
2220 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2221 	uint64_t this_tick = ticks;
2222 	uint32_t reset_on = hz / 2;
2223 
2224 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
2225 		return;
2226 	/*
2227 	** Check on the state of the TX queue(s), this
2228 	** can be done without the lock because its RO
2229 	** and the HUNG state will be static if set.
2230 	*/
2231 	if (this_tick - txq->ift_last_timer_tick >= hz / 2) {
2232 		txq->ift_last_timer_tick = this_tick;
2233 		IFDI_TIMER(ctx, txq->ift_id);
2234 		if ((txq->ift_qstatus == IFLIB_QUEUE_HUNG) &&
2235 		    ((txq->ift_cleaned_prev == txq->ift_cleaned) ||
2236 		     (sctx->isc_pause_frames == 0)))
2237 			goto hung;
2238 
2239 		if (ifmp_ring_is_stalled(txq->ift_br))
2240 			txq->ift_qstatus = IFLIB_QUEUE_HUNG;
2241 		txq->ift_cleaned_prev = txq->ift_cleaned;
2242 	}
2243 #ifdef DEV_NETMAP
2244 	if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP)
2245 		iflib_netmap_timer_adjust(ctx, txq->ift_id, &reset_on);
2246 #endif
2247 	/* handle any laggards */
2248 	if (txq->ift_db_pending)
2249 		GROUPTASK_ENQUEUE(&txq->ift_task);
2250 
2251 	sctx->isc_pause_frames = 0;
2252 	if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)
2253 		callout_reset_on(&txq->ift_timer, reset_on, iflib_timer, txq, txq->ift_timer.c_cpu);
2254 	return;
2255  hung:
2256 	device_printf(ctx->ifc_dev,  "TX(%d) desc avail = %d, pidx = %d\n",
2257 				  txq->ift_id, TXQ_AVAIL(txq), txq->ift_pidx);
2258 	STATE_LOCK(ctx);
2259 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2260 	ctx->ifc_flags |= (IFC_DO_WATCHDOG|IFC_DO_RESET);
2261 	iflib_admin_intr_deferred(ctx);
2262 	STATE_UNLOCK(ctx);
2263 }
2264 
2265 static void
2266 iflib_init_locked(if_ctx_t ctx)
2267 {
2268 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2269 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2270 	if_t ifp = ctx->ifc_ifp;
2271 	iflib_fl_t fl;
2272 	iflib_txq_t txq;
2273 	iflib_rxq_t rxq;
2274 	int i, j, tx_ip_csum_flags, tx_ip6_csum_flags;
2275 
2276 
2277 	if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2278 	IFDI_INTR_DISABLE(ctx);
2279 
2280 	tx_ip_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP);
2281 	tx_ip6_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | CSUM_IP6_SCTP);
2282 	/* Set hardware offload abilities */
2283 	if_clearhwassist(ifp);
2284 	if (if_getcapenable(ifp) & IFCAP_TXCSUM)
2285 		if_sethwassistbits(ifp, tx_ip_csum_flags, 0);
2286 	if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6)
2287 		if_sethwassistbits(ifp,  tx_ip6_csum_flags, 0);
2288 	if (if_getcapenable(ifp) & IFCAP_TSO4)
2289 		if_sethwassistbits(ifp, CSUM_IP_TSO, 0);
2290 	if (if_getcapenable(ifp) & IFCAP_TSO6)
2291 		if_sethwassistbits(ifp, CSUM_IP6_TSO, 0);
2292 
2293 	for (i = 0, txq = ctx->ifc_txqs; i < sctx->isc_ntxqsets; i++, txq++) {
2294 		CALLOUT_LOCK(txq);
2295 		callout_stop(&txq->ift_timer);
2296 		CALLOUT_UNLOCK(txq);
2297 		iflib_netmap_txq_init(ctx, txq);
2298 	}
2299 #ifdef INVARIANTS
2300 	i = if_getdrvflags(ifp);
2301 #endif
2302 	IFDI_INIT(ctx);
2303 	MPASS(if_getdrvflags(ifp) == i);
2304 	for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
2305 		/* XXX this should really be done on a per-queue basis */
2306 		if (if_getcapenable(ifp) & IFCAP_NETMAP) {
2307 			MPASS(rxq->ifr_id == i);
2308 			iflib_netmap_rxq_init(ctx, rxq);
2309 			continue;
2310 		}
2311 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
2312 			if (iflib_fl_setup(fl)) {
2313 				device_printf(ctx->ifc_dev, "freelist setup failed - check cluster settings\n");
2314 				goto done;
2315 			}
2316 		}
2317 	}
2318 done:
2319 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
2320 	IFDI_INTR_ENABLE(ctx);
2321 	txq = ctx->ifc_txqs;
2322 	for (i = 0; i < sctx->isc_ntxqsets; i++, txq++)
2323 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq,
2324 			txq->ift_timer.c_cpu);
2325 }
2326 
2327 static int
2328 iflib_media_change(if_t ifp)
2329 {
2330 	if_ctx_t ctx = if_getsoftc(ifp);
2331 	int err;
2332 
2333 	CTX_LOCK(ctx);
2334 	if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0)
2335 		iflib_init_locked(ctx);
2336 	CTX_UNLOCK(ctx);
2337 	return (err);
2338 }
2339 
2340 static void
2341 iflib_media_status(if_t ifp, struct ifmediareq *ifmr)
2342 {
2343 	if_ctx_t ctx = if_getsoftc(ifp);
2344 
2345 	CTX_LOCK(ctx);
2346 	IFDI_UPDATE_ADMIN_STATUS(ctx);
2347 	IFDI_MEDIA_STATUS(ctx, ifmr);
2348 	CTX_UNLOCK(ctx);
2349 }
2350 
2351 void
2352 iflib_stop(if_ctx_t ctx)
2353 {
2354 	iflib_txq_t txq = ctx->ifc_txqs;
2355 	iflib_rxq_t rxq = ctx->ifc_rxqs;
2356 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2357 	iflib_dma_info_t di;
2358 	iflib_fl_t fl;
2359 	int i, j;
2360 
2361 	/* Tell the stack that the interface is no longer active */
2362 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2363 
2364 	IFDI_INTR_DISABLE(ctx);
2365 	DELAY(1000);
2366 	IFDI_STOP(ctx);
2367 	DELAY(1000);
2368 
2369 	iflib_debug_reset();
2370 	/* Wait for current tx queue users to exit to disarm watchdog timer. */
2371 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) {
2372 		/* make sure all transmitters have completed before proceeding XXX */
2373 
2374 		CALLOUT_LOCK(txq);
2375 		callout_stop(&txq->ift_timer);
2376 		CALLOUT_UNLOCK(txq);
2377 
2378 		/* clean any enqueued buffers */
2379 		iflib_ifmp_purge(txq);
2380 		/* Free any existing tx buffers. */
2381 		for (j = 0; j < txq->ift_size; j++) {
2382 			iflib_txsd_free(ctx, txq, j);
2383 		}
2384 		txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0;
2385 		txq->ift_in_use = txq->ift_gen = txq->ift_cidx = txq->ift_pidx = txq->ift_no_desc_avail = 0;
2386 		txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0;
2387 		txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0;
2388 		txq->ift_pullups = 0;
2389 		ifmp_ring_reset_stats(txq->ift_br);
2390 		for (j = 0, di = txq->ift_ifdi; j < ctx->ifc_nhwtxqs; j++, di++)
2391 			bzero((void *)di->idi_vaddr, di->idi_size);
2392 	}
2393 	for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) {
2394 		/* make sure all transmitters have completed before proceeding XXX */
2395 
2396 		for (j = 0, di = rxq->ifr_ifdi; j < rxq->ifr_nfl; j++, di++)
2397 			bzero((void *)di->idi_vaddr, di->idi_size);
2398 		/* also resets the free lists pidx/cidx */
2399 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
2400 			iflib_fl_bufs_free(fl);
2401 	}
2402 }
2403 
2404 static inline caddr_t
2405 calc_next_rxd(iflib_fl_t fl, int cidx)
2406 {
2407 	qidx_t size;
2408 	int nrxd;
2409 	caddr_t start, end, cur, next;
2410 
2411 	nrxd = fl->ifl_size;
2412 	size = fl->ifl_rxd_size;
2413 	start = fl->ifl_ifdi->idi_vaddr;
2414 
2415 	if (__predict_false(size == 0))
2416 		return (start);
2417 	cur = start + size*cidx;
2418 	end = start + size*nrxd;
2419 	next = CACHE_PTR_NEXT(cur);
2420 	return (next < end ? next : start);
2421 }
2422 
2423 static inline void
2424 prefetch_pkts(iflib_fl_t fl, int cidx)
2425 {
2426 	int nextptr;
2427 	int nrxd = fl->ifl_size;
2428 	caddr_t next_rxd;
2429 
2430 
2431 	nextptr = (cidx + CACHE_PTR_INCREMENT) & (nrxd-1);
2432 	prefetch(&fl->ifl_sds.ifsd_m[nextptr]);
2433 	prefetch(&fl->ifl_sds.ifsd_cl[nextptr]);
2434 	next_rxd = calc_next_rxd(fl, cidx);
2435 	prefetch(next_rxd);
2436 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 1) & (nrxd-1)]);
2437 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 2) & (nrxd-1)]);
2438 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 3) & (nrxd-1)]);
2439 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 4) & (nrxd-1)]);
2440 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 1) & (nrxd-1)]);
2441 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 2) & (nrxd-1)]);
2442 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 3) & (nrxd-1)]);
2443 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 4) & (nrxd-1)]);
2444 }
2445 
2446 static void
2447 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, int unload, if_rxsd_t sd)
2448 {
2449 	int flid, cidx;
2450 	bus_dmamap_t map;
2451 	iflib_fl_t fl;
2452 	iflib_dma_info_t di;
2453 	int next;
2454 
2455 	map = NULL;
2456 	flid = irf->irf_flid;
2457 	cidx = irf->irf_idx;
2458 	fl = &rxq->ifr_fl[flid];
2459 	sd->ifsd_fl = fl;
2460 	sd->ifsd_cidx = cidx;
2461 	sd->ifsd_m = &fl->ifl_sds.ifsd_m[cidx];
2462 	sd->ifsd_cl = &fl->ifl_sds.ifsd_cl[cidx];
2463 	fl->ifl_credits--;
2464 #if MEMORY_LOGGING
2465 	fl->ifl_m_dequeued++;
2466 #endif
2467 	if (rxq->ifr_ctx->ifc_flags & IFC_PREFETCH)
2468 		prefetch_pkts(fl, cidx);
2469 	if (fl->ifl_sds.ifsd_map != NULL) {
2470 		next = (cidx + CACHE_PTR_INCREMENT) & (fl->ifl_size-1);
2471 		prefetch(&fl->ifl_sds.ifsd_map[next]);
2472 		map = fl->ifl_sds.ifsd_map[cidx];
2473 		di = fl->ifl_ifdi;
2474 		next = (cidx + CACHE_LINE_SIZE) & (fl->ifl_size-1);
2475 		prefetch(&fl->ifl_sds.ifsd_flags[next]);
2476 		bus_dmamap_sync(di->idi_tag, di->idi_map,
2477 				BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
2478 
2479 	/* not valid assert if bxe really does SGE from non-contiguous elements */
2480 		MPASS(fl->ifl_cidx == cidx);
2481 		if (unload)
2482 			bus_dmamap_unload(fl->ifl_desc_tag, map);
2483 	}
2484 	fl->ifl_cidx = (fl->ifl_cidx + 1) & (fl->ifl_size-1);
2485 	if (__predict_false(fl->ifl_cidx == 0))
2486 		fl->ifl_gen = 0;
2487 	if (map != NULL)
2488 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2489 			BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2490         bit_clear(fl->ifl_rx_bitmap, cidx);
2491 }
2492 
2493 static struct mbuf *
2494 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri, if_rxsd_t sd)
2495 {
2496 	int i, padlen , flags;
2497 	struct mbuf *m, *mh, *mt;
2498 	caddr_t cl;
2499 
2500 	i = 0;
2501 	mh = NULL;
2502 	do {
2503 		rxd_frag_to_sd(rxq, &ri->iri_frags[i], TRUE, sd);
2504 
2505 		MPASS(*sd->ifsd_cl != NULL);
2506 		MPASS(*sd->ifsd_m != NULL);
2507 
2508 		/* Don't include zero-length frags */
2509 		if (ri->iri_frags[i].irf_len == 0) {
2510 			/* XXX we can save the cluster here, but not the mbuf */
2511 			m_init(*sd->ifsd_m, M_NOWAIT, MT_DATA, 0);
2512 			m_free(*sd->ifsd_m);
2513 			*sd->ifsd_m = NULL;
2514 			continue;
2515 		}
2516 		m = *sd->ifsd_m;
2517 		*sd->ifsd_m = NULL;
2518 		if (mh == NULL) {
2519 			flags = M_PKTHDR|M_EXT;
2520 			mh = mt = m;
2521 			padlen = ri->iri_pad;
2522 		} else {
2523 			flags = M_EXT;
2524 			mt->m_next = m;
2525 			mt = m;
2526 			/* assuming padding is only on the first fragment */
2527 			padlen = 0;
2528 		}
2529 		cl = *sd->ifsd_cl;
2530 		*sd->ifsd_cl = NULL;
2531 
2532 		/* Can these two be made one ? */
2533 		m_init(m, M_NOWAIT, MT_DATA, flags);
2534 		m_cljset(m, cl, sd->ifsd_fl->ifl_cltype);
2535 		/*
2536 		 * These must follow m_init and m_cljset
2537 		 */
2538 		m->m_data += padlen;
2539 		ri->iri_len -= padlen;
2540 		m->m_len = ri->iri_frags[i].irf_len;
2541 	} while (++i < ri->iri_nfrags);
2542 
2543 	return (mh);
2544 }
2545 
2546 /*
2547  * Process one software descriptor
2548  */
2549 static struct mbuf *
2550 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri)
2551 {
2552 	struct if_rxsd sd;
2553 	struct mbuf *m;
2554 
2555 	/* should I merge this back in now that the two paths are basically duplicated? */
2556 	if (ri->iri_nfrags == 1 &&
2557 	    ri->iri_frags[0].irf_len <= MIN(IFLIB_RX_COPY_THRESH, MHLEN)) {
2558 		rxd_frag_to_sd(rxq, &ri->iri_frags[0], FALSE, &sd);
2559 		m = *sd.ifsd_m;
2560 		*sd.ifsd_m = NULL;
2561 		m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR);
2562 #ifndef __NO_STRICT_ALIGNMENT
2563 		if (!IP_ALIGNED(m))
2564 			m->m_data += 2;
2565 #endif
2566 		memcpy(m->m_data, *sd.ifsd_cl, ri->iri_len);
2567 		m->m_len = ri->iri_frags[0].irf_len;
2568        } else {
2569 		m = assemble_segments(rxq, ri, &sd);
2570 	}
2571 	m->m_pkthdr.len = ri->iri_len;
2572 	m->m_pkthdr.rcvif = ri->iri_ifp;
2573 	m->m_flags |= ri->iri_flags;
2574 	m->m_pkthdr.ether_vtag = ri->iri_vtag;
2575 	m->m_pkthdr.flowid = ri->iri_flowid;
2576 	M_HASHTYPE_SET(m, ri->iri_rsstype);
2577 	m->m_pkthdr.csum_flags = ri->iri_csum_flags;
2578 	m->m_pkthdr.csum_data = ri->iri_csum_data;
2579 	return (m);
2580 }
2581 
2582 #if defined(INET6) || defined(INET)
2583 static void
2584 iflib_get_ip_forwarding(struct lro_ctrl *lc, bool *v4, bool *v6)
2585 {
2586 	CURVNET_SET(lc->ifp->if_vnet);
2587 #if defined(INET6)
2588 	*v6 = VNET(ip6_forwarding);
2589 #endif
2590 #if defined(INET)
2591 	*v4 = VNET(ipforwarding);
2592 #endif
2593 	CURVNET_RESTORE();
2594 }
2595 
2596 /*
2597  * Returns true if it's possible this packet could be LROed.
2598  * if it returns false, it is guaranteed that tcp_lro_rx()
2599  * would not return zero.
2600  */
2601 static bool
2602 iflib_check_lro_possible(struct mbuf *m, bool v4_forwarding, bool v6_forwarding)
2603 {
2604 	struct ether_header *eh;
2605 	uint16_t eh_type;
2606 
2607 	eh = mtod(m, struct ether_header *);
2608 	eh_type = ntohs(eh->ether_type);
2609 	switch (eh_type) {
2610 #if defined(INET6)
2611 		case ETHERTYPE_IPV6:
2612 			return !v6_forwarding;
2613 #endif
2614 #if defined (INET)
2615 		case ETHERTYPE_IP:
2616 			return !v4_forwarding;
2617 #endif
2618 	}
2619 
2620 	return false;
2621 }
2622 #else
2623 static void
2624 iflib_get_ip_forwarding(struct lro_ctrl *lc __unused, bool *v4 __unused, bool *v6 __unused)
2625 {
2626 }
2627 #endif
2628 
2629 static bool
2630 iflib_rxeof(iflib_rxq_t rxq, qidx_t budget)
2631 {
2632 	if_ctx_t ctx = rxq->ifr_ctx;
2633 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2634 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2635 	int avail, i;
2636 	qidx_t *cidxp;
2637 	struct if_rxd_info ri;
2638 	int err, budget_left, rx_bytes, rx_pkts;
2639 	iflib_fl_t fl;
2640 	struct ifnet *ifp;
2641 	int lro_enabled;
2642 	bool v4_forwarding, v6_forwarding, lro_possible;
2643 
2644 	/*
2645 	 * XXX early demux data packets so that if_input processing only handles
2646 	 * acks in interrupt context
2647 	 */
2648 	struct mbuf *m, *mh, *mt, *mf;
2649 
2650 	lro_possible = v4_forwarding = v6_forwarding = false;
2651 	ifp = ctx->ifc_ifp;
2652 	mh = mt = NULL;
2653 	MPASS(budget > 0);
2654 	rx_pkts	= rx_bytes = 0;
2655 	if (sctx->isc_flags & IFLIB_HAS_RXCQ)
2656 		cidxp = &rxq->ifr_cq_cidx;
2657 	else
2658 		cidxp = &rxq->ifr_fl[0].ifl_cidx;
2659 	if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) {
2660 		for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2661 			__iflib_fl_refill_lt(ctx, fl, budget + 8);
2662 		DBG_COUNTER_INC(rx_unavail);
2663 		return (false);
2664 	}
2665 
2666 	for (budget_left = budget; budget_left > 0 && avail > 0;) {
2667 		if (__predict_false(!CTX_ACTIVE(ctx))) {
2668 			DBG_COUNTER_INC(rx_ctx_inactive);
2669 			break;
2670 		}
2671 		/*
2672 		 * Reset client set fields to their default values
2673 		 */
2674 		rxd_info_zero(&ri);
2675 		ri.iri_qsidx = rxq->ifr_id;
2676 		ri.iri_cidx = *cidxp;
2677 		ri.iri_ifp = ifp;
2678 		ri.iri_frags = rxq->ifr_frags;
2679 		err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
2680 
2681 		if (err)
2682 			goto err;
2683 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
2684 			*cidxp = ri.iri_cidx;
2685 			/* Update our consumer index */
2686 			/* XXX NB: shurd - check if this is still safe */
2687 			while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0]) {
2688 				rxq->ifr_cq_cidx -= scctx->isc_nrxd[0];
2689 				rxq->ifr_cq_gen = 0;
2690 			}
2691 			/* was this only a completion queue message? */
2692 			if (__predict_false(ri.iri_nfrags == 0))
2693 				continue;
2694 		}
2695 		MPASS(ri.iri_nfrags != 0);
2696 		MPASS(ri.iri_len != 0);
2697 
2698 		/* will advance the cidx on the corresponding free lists */
2699 		m = iflib_rxd_pkt_get(rxq, &ri);
2700 		avail--;
2701 		budget_left--;
2702 		if (avail == 0 && budget_left)
2703 			avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left);
2704 
2705 		if (__predict_false(m == NULL)) {
2706 			DBG_COUNTER_INC(rx_mbuf_null);
2707 			continue;
2708 		}
2709 		/* imm_pkt: -- cxgb */
2710 		if (mh == NULL)
2711 			mh = mt = m;
2712 		else {
2713 			mt->m_nextpkt = m;
2714 			mt = m;
2715 		}
2716 	}
2717 	/* make sure that we can refill faster than drain */
2718 	for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2719 		__iflib_fl_refill_lt(ctx, fl, budget + 8);
2720 
2721 	lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO);
2722 	if (lro_enabled)
2723 		iflib_get_ip_forwarding(&rxq->ifr_lc, &v4_forwarding, &v6_forwarding);
2724 	mt = mf = NULL;
2725 	while (mh != NULL) {
2726 		m = mh;
2727 		mh = mh->m_nextpkt;
2728 		m->m_nextpkt = NULL;
2729 #ifndef __NO_STRICT_ALIGNMENT
2730 		if (!IP_ALIGNED(m) && (m = iflib_fixup_rx(m)) == NULL)
2731 			continue;
2732 #endif
2733 		rx_bytes += m->m_pkthdr.len;
2734 		rx_pkts++;
2735 #if defined(INET6) || defined(INET)
2736 		if (lro_enabled) {
2737 			if (!lro_possible) {
2738 				lro_possible = iflib_check_lro_possible(m, v4_forwarding, v6_forwarding);
2739 				if (lro_possible && mf != NULL) {
2740 					ifp->if_input(ifp, mf);
2741 					DBG_COUNTER_INC(rx_if_input);
2742 					mt = mf = NULL;
2743 				}
2744 			}
2745 			if ((m->m_pkthdr.csum_flags & (CSUM_L4_CALC|CSUM_L4_VALID)) ==
2746 			    (CSUM_L4_CALC|CSUM_L4_VALID)) {
2747 				if (lro_possible && tcp_lro_rx(&rxq->ifr_lc, m, 0) == 0)
2748 					continue;
2749 			}
2750 		}
2751 #endif
2752 		if (lro_possible) {
2753 			ifp->if_input(ifp, m);
2754 			DBG_COUNTER_INC(rx_if_input);
2755 			continue;
2756 		}
2757 
2758 		if (mf == NULL)
2759 			mf = m;
2760 		if (mt != NULL)
2761 			mt->m_nextpkt = m;
2762 		mt = m;
2763 	}
2764 	if (mf != NULL) {
2765 		ifp->if_input(ifp, mf);
2766 		DBG_COUNTER_INC(rx_if_input);
2767 	}
2768 
2769 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
2770 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
2771 
2772 	/*
2773 	 * Flush any outstanding LRO work
2774 	 */
2775 #if defined(INET6) || defined(INET)
2776 	tcp_lro_flush_all(&rxq->ifr_lc);
2777 #endif
2778 	if (avail)
2779 		return true;
2780 	return (iflib_rxd_avail(ctx, rxq, *cidxp, 1));
2781 err:
2782 	STATE_LOCK(ctx);
2783 	ctx->ifc_flags |= IFC_DO_RESET;
2784 	iflib_admin_intr_deferred(ctx);
2785 	STATE_UNLOCK(ctx);
2786 	return (false);
2787 }
2788 
2789 #define TXD_NOTIFY_COUNT(txq) (((txq)->ift_size / (txq)->ift_update_freq)-1)
2790 static inline qidx_t
2791 txq_max_db_deferred(iflib_txq_t txq, qidx_t in_use)
2792 {
2793 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2794 	qidx_t minthresh = txq->ift_size / 8;
2795 	if (in_use > 4*minthresh)
2796 		return (notify_count);
2797 	if (in_use > 2*minthresh)
2798 		return (notify_count >> 1);
2799 	if (in_use > minthresh)
2800 		return (notify_count >> 3);
2801 	return (0);
2802 }
2803 
2804 static inline qidx_t
2805 txq_max_rs_deferred(iflib_txq_t txq)
2806 {
2807 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2808 	qidx_t minthresh = txq->ift_size / 8;
2809 	if (txq->ift_in_use > 4*minthresh)
2810 		return (notify_count);
2811 	if (txq->ift_in_use > 2*minthresh)
2812 		return (notify_count >> 1);
2813 	if (txq->ift_in_use > minthresh)
2814 		return (notify_count >> 2);
2815 	return (2);
2816 }
2817 
2818 #define M_CSUM_FLAGS(m) ((m)->m_pkthdr.csum_flags)
2819 #define M_HAS_VLANTAG(m) (m->m_flags & M_VLANTAG)
2820 
2821 #define TXQ_MAX_DB_DEFERRED(txq, in_use) txq_max_db_deferred((txq), (in_use))
2822 #define TXQ_MAX_RS_DEFERRED(txq) txq_max_rs_deferred(txq)
2823 #define TXQ_MAX_DB_CONSUMED(size) (size >> 4)
2824 
2825 /* forward compatibility for cxgb */
2826 #define FIRST_QSET(ctx) 0
2827 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets)
2828 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets)
2829 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx))
2830 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments))
2831 
2832 /* XXX we should be setting this to something other than zero */
2833 #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh)
2834 #define	MAX_TX_DESC(ctx) max((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max, \
2835     (ctx)->ifc_softc_ctx.isc_tx_nsegments)
2836 
2837 static inline bool
2838 iflib_txd_db_check(if_ctx_t ctx, iflib_txq_t txq, int ring, qidx_t in_use)
2839 {
2840 	qidx_t dbval, max;
2841 	bool rang;
2842 
2843 	rang = false;
2844 	max = TXQ_MAX_DB_DEFERRED(txq, in_use);
2845 	if (ring || txq->ift_db_pending >= max) {
2846 		dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx;
2847 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval);
2848 		txq->ift_db_pending = txq->ift_npending = 0;
2849 		rang = true;
2850 	}
2851 	return (rang);
2852 }
2853 
2854 #ifdef PKT_DEBUG
2855 static void
2856 print_pkt(if_pkt_info_t pi)
2857 {
2858 	printf("pi len:  %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n",
2859 	       pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx);
2860 	printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n",
2861 	       pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag);
2862 	printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n",
2863 	       pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto);
2864 }
2865 #endif
2866 
2867 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO)
2868 #define IS_TX_OFFLOAD4(pi) ((pi)->ipi_csum_flags & (CSUM_IP_TCP | CSUM_IP_TSO))
2869 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO)
2870 #define IS_TX_OFFLOAD6(pi) ((pi)->ipi_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_TSO))
2871 
2872 static int
2873 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp)
2874 {
2875 	if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx;
2876 	struct ether_vlan_header *eh;
2877 	struct mbuf *m;
2878 
2879 	m = *mp;
2880 	if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) &&
2881 	    M_WRITABLE(m) == 0) {
2882 		if ((m = m_dup(m, M_NOWAIT)) == NULL) {
2883 			return (ENOMEM);
2884 		} else {
2885 			m_freem(*mp);
2886 			*mp = m;
2887 		}
2888 	}
2889 
2890 	/*
2891 	 * Determine where frame payload starts.
2892 	 * Jump over vlan headers if already present,
2893 	 * helpful for QinQ too.
2894 	 */
2895 	if (__predict_false(m->m_len < sizeof(*eh))) {
2896 		txq->ift_pullups++;
2897 		if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL))
2898 			return (ENOMEM);
2899 	}
2900 	eh = mtod(m, struct ether_vlan_header *);
2901 	if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
2902 		pi->ipi_etype = ntohs(eh->evl_proto);
2903 		pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
2904 	} else {
2905 		pi->ipi_etype = ntohs(eh->evl_encap_proto);
2906 		pi->ipi_ehdrlen = ETHER_HDR_LEN;
2907 	}
2908 
2909 	switch (pi->ipi_etype) {
2910 #ifdef INET
2911 	case ETHERTYPE_IP:
2912 	{
2913 		struct mbuf *n;
2914 		struct ip *ip = NULL;
2915 		struct tcphdr *th = NULL;
2916 		int minthlen;
2917 
2918 		minthlen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip) + sizeof(*th));
2919 		if (__predict_false(m->m_len < minthlen)) {
2920 			/*
2921 			 * if this code bloat is causing too much of a hit
2922 			 * move it to a separate function and mark it noinline
2923 			 */
2924 			if (m->m_len == pi->ipi_ehdrlen) {
2925 				n = m->m_next;
2926 				MPASS(n);
2927 				if (n->m_len >= sizeof(*ip))  {
2928 					ip = (struct ip *)n->m_data;
2929 					if (n->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2930 						th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2931 				} else {
2932 					txq->ift_pullups++;
2933 					if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
2934 						return (ENOMEM);
2935 					ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2936 				}
2937 			} else {
2938 				txq->ift_pullups++;
2939 				if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
2940 					return (ENOMEM);
2941 				ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2942 				if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2943 					th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2944 			}
2945 		} else {
2946 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2947 			if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2948 				th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2949 		}
2950 		pi->ipi_ip_hlen = ip->ip_hl << 2;
2951 		pi->ipi_ipproto = ip->ip_p;
2952 		pi->ipi_flags |= IPI_TX_IPV4;
2953 
2954 		if ((sctx->isc_flags & IFLIB_NEED_ZERO_CSUM) && (pi->ipi_csum_flags & CSUM_IP))
2955                        ip->ip_sum = 0;
2956 
2957 		/* TCP checksum offload may require TCP header length */
2958 		if (IS_TX_OFFLOAD4(pi)) {
2959 			if (__predict_true(pi->ipi_ipproto == IPPROTO_TCP)) {
2960 				if (__predict_false(th == NULL)) {
2961 					txq->ift_pullups++;
2962 					if (__predict_false((m = m_pullup(m, (ip->ip_hl << 2) + sizeof(*th))) == NULL))
2963 						return (ENOMEM);
2964 					th = (struct tcphdr *)((caddr_t)ip + pi->ipi_ip_hlen);
2965 				}
2966 				pi->ipi_tcp_hflags = th->th_flags;
2967 				pi->ipi_tcp_hlen = th->th_off << 2;
2968 				pi->ipi_tcp_seq = th->th_seq;
2969 			}
2970 			if (IS_TSO4(pi)) {
2971 				if (__predict_false(ip->ip_p != IPPROTO_TCP))
2972 					return (ENXIO);
2973 				th->th_sum = in_pseudo(ip->ip_src.s_addr,
2974 						       ip->ip_dst.s_addr, htons(IPPROTO_TCP));
2975 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
2976 				if (sctx->isc_flags & IFLIB_TSO_INIT_IP) {
2977 					ip->ip_sum = 0;
2978 					ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz);
2979 				}
2980 			}
2981 		}
2982 		break;
2983 	}
2984 #endif
2985 #ifdef INET6
2986 	case ETHERTYPE_IPV6:
2987 	{
2988 		struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
2989 		struct tcphdr *th;
2990 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
2991 
2992 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
2993 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
2994 				return (ENOMEM);
2995 		}
2996 		th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
2997 
2998 		/* XXX-BZ this will go badly in case of ext hdrs. */
2999 		pi->ipi_ipproto = ip6->ip6_nxt;
3000 		pi->ipi_flags |= IPI_TX_IPV6;
3001 
3002 		/* TCP checksum offload may require TCP header length */
3003 		if (IS_TX_OFFLOAD6(pi)) {
3004 			if (pi->ipi_ipproto == IPPROTO_TCP) {
3005 				if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) {
3006 					txq->ift_pullups++;
3007 					if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL))
3008 						return (ENOMEM);
3009 				}
3010 				pi->ipi_tcp_hflags = th->th_flags;
3011 				pi->ipi_tcp_hlen = th->th_off << 2;
3012 				pi->ipi_tcp_seq = th->th_seq;
3013 			}
3014 			if (IS_TSO6(pi)) {
3015 				if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP))
3016 					return (ENXIO);
3017 				/*
3018 				 * The corresponding flag is set by the stack in the IPv4
3019 				 * TSO case, but not in IPv6 (at least in FreeBSD 10.2).
3020 				 * So, set it here because the rest of the flow requires it.
3021 				 */
3022 				pi->ipi_csum_flags |= CSUM_IP6_TCP;
3023 				th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0);
3024 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3025 			}
3026 		}
3027 		break;
3028 	}
3029 #endif
3030 	default:
3031 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
3032 		pi->ipi_ip_hlen = 0;
3033 		break;
3034 	}
3035 	*mp = m;
3036 
3037 	return (0);
3038 }
3039 
3040 static  __noinline  struct mbuf *
3041 collapse_pkthdr(struct mbuf *m0)
3042 {
3043 	struct mbuf *m, *m_next, *tmp;
3044 
3045 	m = m0;
3046 	m_next = m->m_next;
3047 	while (m_next != NULL && m_next->m_len == 0) {
3048 		m = m_next;
3049 		m->m_next = NULL;
3050 		m_free(m);
3051 		m_next = m_next->m_next;
3052 	}
3053 	m = m0;
3054 	m->m_next = m_next;
3055 	if (m_next == NULL)
3056 		return (m);
3057 	if ((m_next->m_flags & M_EXT) == 0) {
3058 		m = m_defrag(m, M_NOWAIT);
3059 	} else {
3060 		tmp = m_next->m_next;
3061 		memcpy(m_next, m, MPKTHSIZE);
3062 		m = m_next;
3063 		m->m_next = tmp;
3064 	}
3065 	return (m);
3066 }
3067 
3068 /*
3069  * If dodgy hardware rejects the scatter gather chain we've handed it
3070  * we'll need to remove the mbuf chain from ifsg_m[] before we can add the
3071  * m_defrag'd mbufs
3072  */
3073 static __noinline struct mbuf *
3074 iflib_remove_mbuf(iflib_txq_t txq)
3075 {
3076 	int ntxd, i, pidx;
3077 	struct mbuf *m, *mh, **ifsd_m;
3078 
3079 	pidx = txq->ift_pidx;
3080 	ifsd_m = txq->ift_sds.ifsd_m;
3081 	ntxd = txq->ift_size;
3082 	mh = m = ifsd_m[pidx];
3083 	ifsd_m[pidx] = NULL;
3084 #if MEMORY_LOGGING
3085 	txq->ift_dequeued++;
3086 #endif
3087 	i = 1;
3088 
3089 	while (m) {
3090 		ifsd_m[(pidx + i) & (ntxd -1)] = NULL;
3091 #if MEMORY_LOGGING
3092 		txq->ift_dequeued++;
3093 #endif
3094 		m = m->m_next;
3095 		i++;
3096 	}
3097 	return (mh);
3098 }
3099 
3100 static int
3101 iflib_busdma_load_mbuf_sg(iflib_txq_t txq, bus_dma_tag_t tag, bus_dmamap_t map,
3102 			  struct mbuf **m0, bus_dma_segment_t *segs, int *nsegs,
3103 			  int max_segs, int flags)
3104 {
3105 	if_ctx_t ctx;
3106 	if_shared_ctx_t		sctx;
3107 	if_softc_ctx_t		scctx;
3108 	int i, next, pidx, err, ntxd, count;
3109 	struct mbuf *m, *tmp, **ifsd_m;
3110 
3111 	m = *m0;
3112 
3113 	/*
3114 	 * Please don't ever do this
3115 	 */
3116 	if (__predict_false(m->m_len == 0))
3117 		*m0 = collapse_pkthdr(m);
3118 
3119 	ctx = txq->ift_ctx;
3120 	sctx = ctx->ifc_sctx;
3121 	scctx = &ctx->ifc_softc_ctx;
3122 	ifsd_m = txq->ift_sds.ifsd_m;
3123 	ntxd = txq->ift_size;
3124 	pidx = txq->ift_pidx;
3125 	if (map != NULL) {
3126 		uint8_t *ifsd_flags = txq->ift_sds.ifsd_flags;
3127 
3128 		err = bus_dmamap_load_mbuf_sg(tag, map,
3129 					      *m0, segs, nsegs, BUS_DMA_NOWAIT);
3130 		if (err)
3131 			return (err);
3132 		ifsd_flags[pidx] |= TX_SW_DESC_MAPPED;
3133 		count = 0;
3134 		m = *m0;
3135 		do {
3136 			if (__predict_false(m->m_len <= 0)) {
3137 				tmp = m;
3138 				m = m->m_next;
3139 				tmp->m_next = NULL;
3140 				m_free(tmp);
3141 				continue;
3142 			}
3143 			m = m->m_next;
3144 			count++;
3145 		} while (m != NULL);
3146 		if (count > *nsegs) {
3147 			ifsd_m[pidx] = *m0;
3148 			ifsd_m[pidx]->m_flags |= M_TOOBIG;
3149 			return (0);
3150 		}
3151 		m = *m0;
3152 		count = 0;
3153 		do {
3154 			next = (pidx + count) & (ntxd-1);
3155 			MPASS(ifsd_m[next] == NULL);
3156 			ifsd_m[next] = m;
3157 			count++;
3158 			tmp = m;
3159 			m = m->m_next;
3160 		} while (m != NULL);
3161 	} else {
3162 		int buflen, sgsize, maxsegsz, max_sgsize;
3163 		vm_offset_t vaddr;
3164 		vm_paddr_t curaddr;
3165 
3166 		count = i = 0;
3167 		m = *m0;
3168 		if (m->m_pkthdr.csum_flags & CSUM_TSO)
3169 			maxsegsz = scctx->isc_tx_tso_segsize_max;
3170 		else
3171 			maxsegsz = sctx->isc_tx_maxsegsize;
3172 
3173 		do {
3174 			if (__predict_false(m->m_len <= 0)) {
3175 				tmp = m;
3176 				m = m->m_next;
3177 				tmp->m_next = NULL;
3178 				m_free(tmp);
3179 				continue;
3180 			}
3181 			buflen = m->m_len;
3182 			vaddr = (vm_offset_t)m->m_data;
3183 			/*
3184 			 * see if we can't be smarter about physically
3185 			 * contiguous mappings
3186 			 */
3187 			next = (pidx + count) & (ntxd-1);
3188 			MPASS(ifsd_m[next] == NULL);
3189 #if MEMORY_LOGGING
3190 			txq->ift_enqueued++;
3191 #endif
3192 			ifsd_m[next] = m;
3193 			while (buflen > 0) {
3194 				if (i >= max_segs)
3195 					goto err;
3196 				max_sgsize = MIN(buflen, maxsegsz);
3197 				curaddr = pmap_kextract(vaddr);
3198 				sgsize = PAGE_SIZE - (curaddr & PAGE_MASK);
3199 				sgsize = MIN(sgsize, max_sgsize);
3200 				segs[i].ds_addr = curaddr;
3201 				segs[i].ds_len = sgsize;
3202 				vaddr += sgsize;
3203 				buflen -= sgsize;
3204 				i++;
3205 			}
3206 			count++;
3207 			tmp = m;
3208 			m = m->m_next;
3209 		} while (m != NULL);
3210 		*nsegs = i;
3211 	}
3212 	return (0);
3213 err:
3214 	*m0 = iflib_remove_mbuf(txq);
3215 	return (EFBIG);
3216 }
3217 
3218 static inline caddr_t
3219 calc_next_txd(iflib_txq_t txq, int cidx, uint8_t qid)
3220 {
3221 	qidx_t size;
3222 	int ntxd;
3223 	caddr_t start, end, cur, next;
3224 
3225 	ntxd = txq->ift_size;
3226 	size = txq->ift_txd_size[qid];
3227 	start = txq->ift_ifdi[qid].idi_vaddr;
3228 
3229 	if (__predict_false(size == 0))
3230 		return (start);
3231 	cur = start + size*cidx;
3232 	end = start + size*ntxd;
3233 	next = CACHE_PTR_NEXT(cur);
3234 	return (next < end ? next : start);
3235 }
3236 
3237 /*
3238  * Pad an mbuf to ensure a minimum ethernet frame size.
3239  * min_frame_size is the frame size (less CRC) to pad the mbuf to
3240  */
3241 static __noinline int
3242 iflib_ether_pad(device_t dev, struct mbuf **m_head, uint16_t min_frame_size)
3243 {
3244 	/*
3245 	 * 18 is enough bytes to pad an ARP packet to 46 bytes, and
3246 	 * and ARP message is the smallest common payload I can think of
3247 	 */
3248 	static char pad[18];	/* just zeros */
3249 	int n;
3250 	struct mbuf *new_head;
3251 
3252 	if (!M_WRITABLE(*m_head)) {
3253 		new_head = m_dup(*m_head, M_NOWAIT);
3254 		if (new_head == NULL) {
3255 			m_freem(*m_head);
3256 			device_printf(dev, "cannot pad short frame, m_dup() failed");
3257 			DBG_COUNTER_INC(encap_pad_mbuf_fail);
3258 			return ENOMEM;
3259 		}
3260 		m_freem(*m_head);
3261 		*m_head = new_head;
3262 	}
3263 
3264 	for (n = min_frame_size - (*m_head)->m_pkthdr.len;
3265 	     n > 0; n -= sizeof(pad))
3266 		if (!m_append(*m_head, min(n, sizeof(pad)), pad))
3267 			break;
3268 
3269 	if (n > 0) {
3270 		m_freem(*m_head);
3271 		device_printf(dev, "cannot pad short frame\n");
3272 		DBG_COUNTER_INC(encap_pad_mbuf_fail);
3273 		return (ENOBUFS);
3274 	}
3275 
3276 	return 0;
3277 }
3278 
3279 static int
3280 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp)
3281 {
3282 	if_ctx_t		ctx;
3283 	if_shared_ctx_t		sctx;
3284 	if_softc_ctx_t		scctx;
3285 	bus_dma_segment_t	*segs;
3286 	struct mbuf		*m_head;
3287 	void			*next_txd;
3288 	bus_dmamap_t		map;
3289 	struct if_pkt_info	pi;
3290 	int remap = 0;
3291 	int err, nsegs, ndesc, max_segs, pidx, cidx, next, ntxd;
3292 	bus_dma_tag_t desc_tag;
3293 
3294 	ctx = txq->ift_ctx;
3295 	sctx = ctx->ifc_sctx;
3296 	scctx = &ctx->ifc_softc_ctx;
3297 	segs = txq->ift_segs;
3298 	ntxd = txq->ift_size;
3299 	m_head = *m_headp;
3300 	map = NULL;
3301 
3302 	/*
3303 	 * If we're doing TSO the next descriptor to clean may be quite far ahead
3304 	 */
3305 	cidx = txq->ift_cidx;
3306 	pidx = txq->ift_pidx;
3307 	if (ctx->ifc_flags & IFC_PREFETCH) {
3308 		next = (cidx + CACHE_PTR_INCREMENT) & (ntxd-1);
3309 		if (!(ctx->ifc_flags & IFLIB_HAS_TXCQ)) {
3310 			next_txd = calc_next_txd(txq, cidx, 0);
3311 			prefetch(next_txd);
3312 		}
3313 
3314 		/* prefetch the next cache line of mbuf pointers and flags */
3315 		prefetch(&txq->ift_sds.ifsd_m[next]);
3316 		if (txq->ift_sds.ifsd_map != NULL) {
3317 			prefetch(&txq->ift_sds.ifsd_map[next]);
3318 			next = (cidx + CACHE_LINE_SIZE) & (ntxd-1);
3319 			prefetch(&txq->ift_sds.ifsd_flags[next]);
3320 		}
3321 	} else if (txq->ift_sds.ifsd_map != NULL)
3322 		map = txq->ift_sds.ifsd_map[pidx];
3323 
3324 	if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
3325 		desc_tag = txq->ift_tso_desc_tag;
3326 		max_segs = scctx->isc_tx_tso_segments_max;
3327 		MPASS(desc_tag != NULL);
3328 		MPASS(max_segs > 0);
3329 	} else {
3330 		desc_tag = txq->ift_desc_tag;
3331 		max_segs = scctx->isc_tx_nsegments;
3332 	}
3333 	if ((sctx->isc_flags & IFLIB_NEED_ETHER_PAD) &&
3334 	    __predict_false(m_head->m_pkthdr.len < scctx->isc_min_frame_size)) {
3335 		err = iflib_ether_pad(ctx->ifc_dev, m_headp, scctx->isc_min_frame_size);
3336 		if (err)
3337 			return err;
3338 	}
3339 	m_head = *m_headp;
3340 
3341 	pkt_info_zero(&pi);
3342 	pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG|M_BCAST|M_MCAST));
3343 	pi.ipi_pidx = pidx;
3344 	pi.ipi_qsidx = txq->ift_id;
3345 	pi.ipi_len = m_head->m_pkthdr.len;
3346 	pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags;
3347 	pi.ipi_vtag = (m_head->m_flags & M_VLANTAG) ? m_head->m_pkthdr.ether_vtag : 0;
3348 
3349 	/* deliberate bitwise OR to make one condition */
3350 	if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) {
3351 		if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0))
3352 			return (err);
3353 		m_head = *m_headp;
3354 	}
3355 
3356 retry:
3357 	err = iflib_busdma_load_mbuf_sg(txq, desc_tag, map, m_headp, segs, &nsegs, max_segs, BUS_DMA_NOWAIT);
3358 defrag:
3359 	if (__predict_false(err)) {
3360 		switch (err) {
3361 		case EFBIG:
3362 			/* try collapse once and defrag once */
3363 			if (remap == 0) {
3364 				m_head = m_collapse(*m_headp, M_NOWAIT, max_segs);
3365 				/* try defrag if collapsing fails */
3366 				if (m_head == NULL)
3367 					remap++;
3368 			}
3369 			if (remap == 1)
3370 				m_head = m_defrag(*m_headp, M_NOWAIT);
3371 			remap++;
3372 			if (__predict_false(m_head == NULL))
3373 				goto defrag_failed;
3374 			txq->ift_mbuf_defrag++;
3375 			*m_headp = m_head;
3376 			goto retry;
3377 			break;
3378 		case ENOMEM:
3379 			txq->ift_no_tx_dma_setup++;
3380 			break;
3381 		default:
3382 			txq->ift_no_tx_dma_setup++;
3383 			m_freem(*m_headp);
3384 			DBG_COUNTER_INC(tx_frees);
3385 			*m_headp = NULL;
3386 			break;
3387 		}
3388 		txq->ift_map_failed++;
3389 		DBG_COUNTER_INC(encap_load_mbuf_fail);
3390 		return (err);
3391 	}
3392 
3393 	/*
3394 	 * XXX assumes a 1 to 1 relationship between segments and
3395 	 *        descriptors - this does not hold true on all drivers, e.g.
3396 	 *        cxgb
3397 	 */
3398 	if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) {
3399 		txq->ift_no_desc_avail++;
3400 		if (map != NULL)
3401 			bus_dmamap_unload(desc_tag, map);
3402 		DBG_COUNTER_INC(encap_txq_avail_fail);
3403 		if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0)
3404 			GROUPTASK_ENQUEUE(&txq->ift_task);
3405 		return (ENOBUFS);
3406 	}
3407 	/*
3408 	 * On Intel cards we can greatly reduce the number of TX interrupts
3409 	 * we see by only setting report status on every Nth descriptor.
3410 	 * However, this also means that the driver will need to keep track
3411 	 * of the descriptors that RS was set on to check them for the DD bit.
3412 	 */
3413 	txq->ift_rs_pending += nsegs + 1;
3414 	if (txq->ift_rs_pending > TXQ_MAX_RS_DEFERRED(txq) ||
3415 	     iflib_no_tx_batch || (TXQ_AVAIL(txq) - nsegs) <= MAX_TX_DESC(ctx) + 2) {
3416 		pi.ipi_flags |= IPI_TX_INTR;
3417 		txq->ift_rs_pending = 0;
3418 	}
3419 
3420 	pi.ipi_segs = segs;
3421 	pi.ipi_nsegs = nsegs;
3422 
3423 	MPASS(pidx >= 0 && pidx < txq->ift_size);
3424 #ifdef PKT_DEBUG
3425 	print_pkt(&pi);
3426 #endif
3427 	if (map != NULL)
3428 		bus_dmamap_sync(desc_tag, map, BUS_DMASYNC_PREWRITE);
3429 	if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) {
3430 		if (map != NULL)
3431 			bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3432 					BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
3433 		DBG_COUNTER_INC(tx_encap);
3434 		MPASS(pi.ipi_new_pidx < txq->ift_size);
3435 
3436 		ndesc = pi.ipi_new_pidx - pi.ipi_pidx;
3437 		if (pi.ipi_new_pidx < pi.ipi_pidx) {
3438 			ndesc += txq->ift_size;
3439 			txq->ift_gen = 1;
3440 		}
3441 		/*
3442 		 * drivers can need as many as
3443 		 * two sentinels
3444 		 */
3445 		MPASS(ndesc <= pi.ipi_nsegs + 2);
3446 		MPASS(pi.ipi_new_pidx != pidx);
3447 		MPASS(ndesc > 0);
3448 		txq->ift_in_use += ndesc;
3449 
3450 		/*
3451 		 * We update the last software descriptor again here because there may
3452 		 * be a sentinel and/or there may be more mbufs than segments
3453 		 */
3454 		txq->ift_pidx = pi.ipi_new_pidx;
3455 		txq->ift_npending += pi.ipi_ndescs;
3456 	} else {
3457 		*m_headp = m_head = iflib_remove_mbuf(txq);
3458 		if (err == EFBIG) {
3459 			txq->ift_txd_encap_efbig++;
3460 			if (remap < 2) {
3461 				remap = 1;
3462 				goto defrag;
3463 			}
3464 		}
3465 		DBG_COUNTER_INC(encap_txd_encap_fail);
3466 		goto defrag_failed;
3467 	}
3468 	return (err);
3469 
3470 defrag_failed:
3471 	txq->ift_mbuf_defrag_failed++;
3472 	txq->ift_map_failed++;
3473 	m_freem(*m_headp);
3474 	DBG_COUNTER_INC(tx_frees);
3475 	*m_headp = NULL;
3476 	return (ENOMEM);
3477 }
3478 
3479 static void
3480 iflib_tx_desc_free(iflib_txq_t txq, int n)
3481 {
3482 	int hasmap;
3483 	uint32_t qsize, cidx, mask, gen;
3484 	struct mbuf *m, **ifsd_m;
3485 	uint8_t *ifsd_flags;
3486 	bus_dmamap_t *ifsd_map;
3487 	bool do_prefetch;
3488 
3489 	cidx = txq->ift_cidx;
3490 	gen = txq->ift_gen;
3491 	qsize = txq->ift_size;
3492 	mask = qsize-1;
3493 	hasmap = txq->ift_sds.ifsd_map != NULL;
3494 	ifsd_flags = txq->ift_sds.ifsd_flags;
3495 	ifsd_m = txq->ift_sds.ifsd_m;
3496 	ifsd_map = txq->ift_sds.ifsd_map;
3497 	do_prefetch = (txq->ift_ctx->ifc_flags & IFC_PREFETCH);
3498 
3499 	while (n-- > 0) {
3500 		if (do_prefetch) {
3501 			prefetch(ifsd_m[(cidx + 3) & mask]);
3502 			prefetch(ifsd_m[(cidx + 4) & mask]);
3503 		}
3504 		if (ifsd_m[cidx] != NULL) {
3505 			prefetch(&ifsd_m[(cidx + CACHE_PTR_INCREMENT) & mask]);
3506 			prefetch(&ifsd_flags[(cidx + CACHE_PTR_INCREMENT) & mask]);
3507 			if (hasmap && (ifsd_flags[cidx] & TX_SW_DESC_MAPPED)) {
3508 				/*
3509 				 * does it matter if it's not the TSO tag? If so we'll
3510 				 * have to add the type to flags
3511 				 */
3512 				bus_dmamap_unload(txq->ift_desc_tag, ifsd_map[cidx]);
3513 				ifsd_flags[cidx] &= ~TX_SW_DESC_MAPPED;
3514 			}
3515 			if ((m = ifsd_m[cidx]) != NULL) {
3516 				/* XXX we don't support any drivers that batch packets yet */
3517 				MPASS(m->m_nextpkt == NULL);
3518 				/* if the number of clusters exceeds the number of segments
3519 				 * there won't be space on the ring to save a pointer to each
3520 				 * cluster so we simply free the list here
3521 				 */
3522 				if (m->m_flags & M_TOOBIG) {
3523 					m_freem(m);
3524 				} else {
3525 					m_free(m);
3526 				}
3527 				ifsd_m[cidx] = NULL;
3528 #if MEMORY_LOGGING
3529 				txq->ift_dequeued++;
3530 #endif
3531 				DBG_COUNTER_INC(tx_frees);
3532 			}
3533 		}
3534 		if (__predict_false(++cidx == qsize)) {
3535 			cidx = 0;
3536 			gen = 0;
3537 		}
3538 	}
3539 	txq->ift_cidx = cidx;
3540 	txq->ift_gen = gen;
3541 }
3542 
3543 static __inline int
3544 iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh)
3545 {
3546 	int reclaim;
3547 	if_ctx_t ctx = txq->ift_ctx;
3548 
3549 	KASSERT(thresh >= 0, ("invalid threshold to reclaim"));
3550 	MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size);
3551 
3552 	/*
3553 	 * Need a rate-limiting check so that this isn't called every time
3554 	 */
3555 	iflib_tx_credits_update(ctx, txq);
3556 	reclaim = DESC_RECLAIMABLE(txq);
3557 
3558 	if (reclaim <= thresh /* + MAX_TX_DESC(txq->ift_ctx) */) {
3559 #ifdef INVARIANTS
3560 		if (iflib_verbose_debug) {
3561 			printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __FUNCTION__,
3562 			       txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments,
3563 			       reclaim, thresh);
3564 
3565 		}
3566 #endif
3567 		return (0);
3568 	}
3569 	iflib_tx_desc_free(txq, reclaim);
3570 	txq->ift_cleaned += reclaim;
3571 	txq->ift_in_use -= reclaim;
3572 
3573 	return (reclaim);
3574 }
3575 
3576 static struct mbuf **
3577 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset, int remaining)
3578 {
3579 	int next, size;
3580 	struct mbuf **items;
3581 
3582 	size = r->size;
3583 	next = (cidx + CACHE_PTR_INCREMENT) & (size-1);
3584 	items = __DEVOLATILE(struct mbuf **, &r->items[0]);
3585 
3586 	prefetch(items[(cidx + offset) & (size-1)]);
3587 	if (remaining > 1) {
3588 		prefetch2cachelines(&items[next]);
3589 		prefetch2cachelines(items[(cidx + offset + 1) & (size-1)]);
3590 		prefetch2cachelines(items[(cidx + offset + 2) & (size-1)]);
3591 		prefetch2cachelines(items[(cidx + offset + 3) & (size-1)]);
3592 	}
3593 	return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (size-1)]));
3594 }
3595 
3596 static void
3597 iflib_txq_check_drain(iflib_txq_t txq, int budget)
3598 {
3599 
3600 	ifmp_ring_check_drainage(txq->ift_br, budget);
3601 }
3602 
3603 static uint32_t
3604 iflib_txq_can_drain(struct ifmp_ring *r)
3605 {
3606 	iflib_txq_t txq = r->cookie;
3607 	if_ctx_t ctx = txq->ift_ctx;
3608 
3609 	return ((TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2) ||
3610 		ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false));
3611 }
3612 
3613 static uint32_t
3614 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3615 {
3616 	iflib_txq_t txq = r->cookie;
3617 	if_ctx_t ctx = txq->ift_ctx;
3618 	struct ifnet *ifp = ctx->ifc_ifp;
3619 	struct mbuf **mp, *m;
3620 	int i, count, consumed, pkt_sent, bytes_sent, mcast_sent, avail;
3621 	int reclaimed, err, in_use_prev, desc_used;
3622 	bool do_prefetch, ring, rang;
3623 
3624 	if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) ||
3625 			    !LINK_ACTIVE(ctx))) {
3626 		DBG_COUNTER_INC(txq_drain_notready);
3627 		return (0);
3628 	}
3629 	reclaimed = iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
3630 	rang = iflib_txd_db_check(ctx, txq, reclaimed, txq->ift_in_use);
3631 	avail = IDXDIFF(pidx, cidx, r->size);
3632 	if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) {
3633 		DBG_COUNTER_INC(txq_drain_flushing);
3634 		for (i = 0; i < avail; i++) {
3635 			m_free(r->items[(cidx + i) & (r->size-1)]);
3636 			r->items[(cidx + i) & (r->size-1)] = NULL;
3637 		}
3638 		return (avail);
3639 	}
3640 
3641 	if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3642 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3643 		CALLOUT_LOCK(txq);
3644 		callout_stop(&txq->ift_timer);
3645 		CALLOUT_UNLOCK(txq);
3646 		DBG_COUNTER_INC(txq_drain_oactive);
3647 		return (0);
3648 	}
3649 	if (reclaimed)
3650 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3651 	consumed = mcast_sent = bytes_sent = pkt_sent = 0;
3652 	count = MIN(avail, TX_BATCH_SIZE);
3653 #ifdef INVARIANTS
3654 	if (iflib_verbose_debug)
3655 		printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __FUNCTION__,
3656 		       avail, ctx->ifc_flags, TXQ_AVAIL(txq));
3657 #endif
3658 	do_prefetch = (ctx->ifc_flags & IFC_PREFETCH);
3659 	avail = TXQ_AVAIL(txq);
3660 	err = 0;
3661 	for (desc_used = i = 0; i < count && avail > MAX_TX_DESC(ctx) + 2; i++) {
3662 		int rem = do_prefetch ? count - i : 0;
3663 
3664 		mp = _ring_peek_one(r, cidx, i, rem);
3665 		MPASS(mp != NULL && *mp != NULL);
3666 		if (__predict_false(*mp == (struct mbuf *)txq)) {
3667 			consumed++;
3668 			reclaimed++;
3669 			continue;
3670 		}
3671 		in_use_prev = txq->ift_in_use;
3672 		err = iflib_encap(txq, mp);
3673 		if (__predict_false(err)) {
3674 			DBG_COUNTER_INC(txq_drain_encapfail);
3675 			/* no room - bail out */
3676 			if (err == ENOBUFS)
3677 				break;
3678 			consumed++;
3679 			DBG_COUNTER_INC(txq_drain_encapfail);
3680 			/* we can't send this packet - skip it */
3681 			continue;
3682 		}
3683 		consumed++;
3684 		pkt_sent++;
3685 		m = *mp;
3686 		DBG_COUNTER_INC(tx_sent);
3687 		bytes_sent += m->m_pkthdr.len;
3688 		mcast_sent += !!(m->m_flags & M_MCAST);
3689 		avail = TXQ_AVAIL(txq);
3690 
3691 		txq->ift_db_pending += (txq->ift_in_use - in_use_prev);
3692 		desc_used += (txq->ift_in_use - in_use_prev);
3693 		ETHER_BPF_MTAP(ifp, m);
3694 		if (__predict_false(!(ifp->if_drv_flags & IFF_DRV_RUNNING)))
3695 			break;
3696 		rang = iflib_txd_db_check(ctx, txq, false, in_use_prev);
3697 	}
3698 
3699 	/* deliberate use of bitwise or to avoid gratuitous short-circuit */
3700 	ring = rang ? false  : (iflib_min_tx_latency | err) || (TXQ_AVAIL(txq) < MAX_TX_DESC(ctx));
3701 	iflib_txd_db_check(ctx, txq, ring, txq->ift_in_use);
3702 	if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
3703 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
3704 	if (mcast_sent)
3705 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
3706 #ifdef INVARIANTS
3707 	if (iflib_verbose_debug)
3708 		printf("consumed=%d\n", consumed);
3709 #endif
3710 	return (consumed);
3711 }
3712 
3713 static uint32_t
3714 iflib_txq_drain_always(struct ifmp_ring *r)
3715 {
3716 	return (1);
3717 }
3718 
3719 static uint32_t
3720 iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3721 {
3722 	int i, avail;
3723 	struct mbuf **mp;
3724 	iflib_txq_t txq;
3725 
3726 	txq = r->cookie;
3727 
3728 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3729 	CALLOUT_LOCK(txq);
3730 	callout_stop(&txq->ift_timer);
3731 	CALLOUT_UNLOCK(txq);
3732 
3733 	avail = IDXDIFF(pidx, cidx, r->size);
3734 	for (i = 0; i < avail; i++) {
3735 		mp = _ring_peek_one(r, cidx, i, avail - i);
3736 		if (__predict_false(*mp == (struct mbuf *)txq))
3737 			continue;
3738 		m_freem(*mp);
3739 	}
3740 	MPASS(ifmp_ring_is_stalled(r) == 0);
3741 	return (avail);
3742 }
3743 
3744 static void
3745 iflib_ifmp_purge(iflib_txq_t txq)
3746 {
3747 	struct ifmp_ring *r;
3748 
3749 	r = txq->ift_br;
3750 	r->drain = iflib_txq_drain_free;
3751 	r->can_drain = iflib_txq_drain_always;
3752 
3753 	ifmp_ring_check_drainage(r, r->size);
3754 
3755 	r->drain = iflib_txq_drain;
3756 	r->can_drain = iflib_txq_can_drain;
3757 }
3758 
3759 static void
3760 _task_fn_tx(void *context)
3761 {
3762 	iflib_txq_t txq = context;
3763 	if_ctx_t ctx = txq->ift_ctx;
3764 	struct ifnet *ifp = ctx->ifc_ifp;
3765 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
3766 
3767 #ifdef IFLIB_DIAGNOSTICS
3768 	txq->ift_cpu_exec_count[curcpu]++;
3769 #endif
3770 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3771 		return;
3772 	if (if_getcapenable(ifp) & IFCAP_NETMAP) {
3773 		if (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false))
3774 			netmap_tx_irq(ifp, txq->ift_id);
3775 		IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3776 		return;
3777 	}
3778 	if (txq->ift_db_pending)
3779 		ifmp_ring_enqueue(txq->ift_br, (void **)&txq, 1, TX_BATCH_SIZE, abdicate);
3780 	else if (!abdicate)
3781 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3782 	/*
3783 	 * When abdicating, we always need to check drainage, not just when we don't enqueue
3784 	 */
3785 	if (abdicate)
3786 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3787 	ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3788 	if (ctx->ifc_flags & IFC_LEGACY)
3789 		IFDI_INTR_ENABLE(ctx);
3790 	else {
3791 #ifdef INVARIANTS
3792 		int rc =
3793 #endif
3794 			IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3795 			KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver"));
3796 	}
3797 }
3798 
3799 static void
3800 _task_fn_rx(void *context)
3801 {
3802 	iflib_rxq_t rxq = context;
3803 	if_ctx_t ctx = rxq->ifr_ctx;
3804 	bool more;
3805 	uint16_t budget;
3806 
3807 #ifdef IFLIB_DIAGNOSTICS
3808 	rxq->ifr_cpu_exec_count[curcpu]++;
3809 #endif
3810 	DBG_COUNTER_INC(task_fn_rxs);
3811 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3812 		return;
3813 	more = true;
3814 #ifdef DEV_NETMAP
3815 	if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP) {
3816 		u_int work = 0;
3817 		if (netmap_rx_irq(ctx->ifc_ifp, rxq->ifr_id, &work)) {
3818 			more = false;
3819 		}
3820 	}
3821 #endif
3822 	budget = ctx->ifc_sysctl_rx_budget;
3823 	if (budget == 0)
3824 		budget = 16;	/* XXX */
3825 	if (more == false || (more = iflib_rxeof(rxq, budget)) == false) {
3826 		if (ctx->ifc_flags & IFC_LEGACY)
3827 			IFDI_INTR_ENABLE(ctx);
3828 		else {
3829 #ifdef INVARIANTS
3830 			int rc =
3831 #endif
3832 				IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
3833 			KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver"));
3834 			DBG_COUNTER_INC(rx_intr_enables);
3835 		}
3836 	}
3837 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3838 		return;
3839 	if (more)
3840 		GROUPTASK_ENQUEUE(&rxq->ifr_task);
3841 }
3842 
3843 static void
3844 _task_fn_admin(void *context)
3845 {
3846 	if_ctx_t ctx = context;
3847 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
3848 	iflib_txq_t txq;
3849 	int i;
3850 	bool oactive, running, do_reset, do_watchdog;
3851 	uint32_t reset_on = hz / 2;
3852 
3853 	STATE_LOCK(ctx);
3854 	running = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING);
3855 	oactive = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE);
3856 	do_reset = (ctx->ifc_flags & IFC_DO_RESET);
3857 	do_watchdog = (ctx->ifc_flags & IFC_DO_WATCHDOG);
3858 	ctx->ifc_flags &= ~(IFC_DO_RESET|IFC_DO_WATCHDOG);
3859 	STATE_UNLOCK(ctx);
3860 
3861 	if ((!running & !oactive) &&
3862 	    !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
3863 		return;
3864 
3865 	CTX_LOCK(ctx);
3866 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3867 		CALLOUT_LOCK(txq);
3868 		callout_stop(&txq->ift_timer);
3869 		CALLOUT_UNLOCK(txq);
3870 	}
3871 	if (do_watchdog) {
3872 		ctx->ifc_watchdog_events++;
3873 		IFDI_WATCHDOG_RESET(ctx);
3874 	}
3875 	IFDI_UPDATE_ADMIN_STATUS(ctx);
3876 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3877 #ifdef DEV_NETMAP
3878 		reset_on = hz / 2;
3879 		if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP)
3880 			iflib_netmap_timer_adjust(ctx, txq->ift_id, &reset_on);
3881 #endif
3882 		callout_reset_on(&txq->ift_timer, reset_on, iflib_timer, txq, txq->ift_timer.c_cpu);
3883 	}
3884 	IFDI_LINK_INTR_ENABLE(ctx);
3885 	if (do_reset)
3886 		iflib_if_init_locked(ctx);
3887 	CTX_UNLOCK(ctx);
3888 
3889 	if (LINK_ACTIVE(ctx) == 0)
3890 		return;
3891 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3892 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
3893 }
3894 
3895 
3896 static void
3897 _task_fn_iov(void *context)
3898 {
3899 	if_ctx_t ctx = context;
3900 
3901 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3902 		return;
3903 
3904 	CTX_LOCK(ctx);
3905 	IFDI_VFLR_HANDLE(ctx);
3906 	CTX_UNLOCK(ctx);
3907 }
3908 
3909 static int
3910 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)
3911 {
3912 	int err;
3913 	if_int_delay_info_t info;
3914 	if_ctx_t ctx;
3915 
3916 	info = (if_int_delay_info_t)arg1;
3917 	ctx = info->iidi_ctx;
3918 	info->iidi_req = req;
3919 	info->iidi_oidp = oidp;
3920 	CTX_LOCK(ctx);
3921 	err = IFDI_SYSCTL_INT_DELAY(ctx, info);
3922 	CTX_UNLOCK(ctx);
3923 	return (err);
3924 }
3925 
3926 /*********************************************************************
3927  *
3928  *  IFNET FUNCTIONS
3929  *
3930  **********************************************************************/
3931 
3932 static void
3933 iflib_if_init_locked(if_ctx_t ctx)
3934 {
3935 	iflib_stop(ctx);
3936 	iflib_init_locked(ctx);
3937 }
3938 
3939 
3940 static void
3941 iflib_if_init(void *arg)
3942 {
3943 	if_ctx_t ctx = arg;
3944 
3945 	CTX_LOCK(ctx);
3946 	iflib_if_init_locked(ctx);
3947 	CTX_UNLOCK(ctx);
3948 }
3949 
3950 static int
3951 iflib_if_transmit(if_t ifp, struct mbuf *m)
3952 {
3953 	if_ctx_t	ctx = if_getsoftc(ifp);
3954 
3955 	iflib_txq_t txq;
3956 	int err, qidx;
3957 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
3958 
3959 	if (__predict_false((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) {
3960 		DBG_COUNTER_INC(tx_frees);
3961 		m_freem(m);
3962 		return (ENOBUFS);
3963 	}
3964 
3965 	MPASS(m->m_nextpkt == NULL);
3966 	qidx = 0;
3967 	if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m))
3968 		qidx = QIDX(ctx, m);
3969 	/*
3970 	 * XXX calculate buf_ring based on flowid (divvy up bits?)
3971 	 */
3972 	txq = &ctx->ifc_txqs[qidx];
3973 
3974 #ifdef DRIVER_BACKPRESSURE
3975 	if (txq->ift_closed) {
3976 		while (m != NULL) {
3977 			next = m->m_nextpkt;
3978 			m->m_nextpkt = NULL;
3979 			m_freem(m);
3980 			m = next;
3981 		}
3982 		return (ENOBUFS);
3983 	}
3984 #endif
3985 #ifdef notyet
3986 	qidx = count = 0;
3987 	mp = marr;
3988 	next = m;
3989 	do {
3990 		count++;
3991 		next = next->m_nextpkt;
3992 	} while (next != NULL);
3993 
3994 	if (count > nitems(marr))
3995 		if ((mp = malloc(count*sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) {
3996 			/* XXX check nextpkt */
3997 			m_freem(m);
3998 			/* XXX simplify for now */
3999 			DBG_COUNTER_INC(tx_frees);
4000 			return (ENOBUFS);
4001 		}
4002 	for (next = m, i = 0; next != NULL; i++) {
4003 		mp[i] = next;
4004 		next = next->m_nextpkt;
4005 		mp[i]->m_nextpkt = NULL;
4006 	}
4007 #endif
4008 	DBG_COUNTER_INC(tx_seen);
4009 	err = ifmp_ring_enqueue(txq->ift_br, (void **)&m, 1, TX_BATCH_SIZE, abdicate);
4010 
4011 	if (abdicate)
4012 		GROUPTASK_ENQUEUE(&txq->ift_task);
4013  	if (err) {
4014 		if (!abdicate)
4015 			GROUPTASK_ENQUEUE(&txq->ift_task);
4016 		/* support forthcoming later */
4017 #ifdef DRIVER_BACKPRESSURE
4018 		txq->ift_closed = TRUE;
4019 #endif
4020 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
4021 		m_freem(m);
4022 	}
4023 
4024 	return (err);
4025 }
4026 
4027 static void
4028 iflib_if_qflush(if_t ifp)
4029 {
4030 	if_ctx_t ctx = if_getsoftc(ifp);
4031 	iflib_txq_t txq = ctx->ifc_txqs;
4032 	int i;
4033 
4034 	STATE_LOCK(ctx);
4035 	ctx->ifc_flags |= IFC_QFLUSH;
4036 	STATE_UNLOCK(ctx);
4037 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
4038 		while (!(ifmp_ring_is_idle(txq->ift_br) || ifmp_ring_is_stalled(txq->ift_br)))
4039 			iflib_txq_check_drain(txq, 0);
4040 	STATE_LOCK(ctx);
4041 	ctx->ifc_flags &= ~IFC_QFLUSH;
4042 	STATE_UNLOCK(ctx);
4043 
4044 	if_qflush(ifp);
4045 }
4046 
4047 
4048 #define IFCAP_FLAGS (IFCAP_TXCSUM_IPV6 | IFCAP_RXCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \
4049 		     IFCAP_TSO4 | IFCAP_TSO6 | IFCAP_VLAN_HWTAGGING | IFCAP_HWSTATS | \
4050 		     IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | IFCAP_VLAN_HWTSO)
4051 
4052 static int
4053 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data)
4054 {
4055 	if_ctx_t ctx = if_getsoftc(ifp);
4056 	struct ifreq	*ifr = (struct ifreq *)data;
4057 #if defined(INET) || defined(INET6)
4058 	struct ifaddr	*ifa = (struct ifaddr *)data;
4059 #endif
4060 	bool		avoid_reset = FALSE;
4061 	int		err = 0, reinit = 0, bits;
4062 
4063 	switch (command) {
4064 	case SIOCSIFADDR:
4065 #ifdef INET
4066 		if (ifa->ifa_addr->sa_family == AF_INET)
4067 			avoid_reset = TRUE;
4068 #endif
4069 #ifdef INET6
4070 		if (ifa->ifa_addr->sa_family == AF_INET6)
4071 			avoid_reset = TRUE;
4072 #endif
4073 		/*
4074 		** Calling init results in link renegotiation,
4075 		** so we avoid doing it when possible.
4076 		*/
4077 		if (avoid_reset) {
4078 			if_setflagbits(ifp, IFF_UP,0);
4079 			if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
4080 				reinit = 1;
4081 #ifdef INET
4082 			if (!(if_getflags(ifp) & IFF_NOARP))
4083 				arp_ifinit(ifp, ifa);
4084 #endif
4085 		} else
4086 			err = ether_ioctl(ifp, command, data);
4087 		break;
4088 	case SIOCSIFMTU:
4089 		CTX_LOCK(ctx);
4090 		if (ifr->ifr_mtu == if_getmtu(ifp)) {
4091 			CTX_UNLOCK(ctx);
4092 			break;
4093 		}
4094 		bits = if_getdrvflags(ifp);
4095 		/* stop the driver and free any clusters before proceeding */
4096 		iflib_stop(ctx);
4097 
4098 		if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) {
4099 			STATE_LOCK(ctx);
4100 			if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size)
4101 				ctx->ifc_flags |= IFC_MULTISEG;
4102 			else
4103 				ctx->ifc_flags &= ~IFC_MULTISEG;
4104 			STATE_UNLOCK(ctx);
4105 			err = if_setmtu(ifp, ifr->ifr_mtu);
4106 		}
4107 		iflib_init_locked(ctx);
4108 		STATE_LOCK(ctx);
4109 		if_setdrvflags(ifp, bits);
4110 		STATE_UNLOCK(ctx);
4111 		CTX_UNLOCK(ctx);
4112 		break;
4113 	case SIOCSIFFLAGS:
4114 		CTX_LOCK(ctx);
4115 		if (if_getflags(ifp) & IFF_UP) {
4116 			if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4117 				if ((if_getflags(ifp) ^ ctx->ifc_if_flags) &
4118 				    (IFF_PROMISC | IFF_ALLMULTI)) {
4119 					err = IFDI_PROMISC_SET(ctx, if_getflags(ifp));
4120 				}
4121 			} else
4122 				reinit = 1;
4123 		} else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4124 			iflib_stop(ctx);
4125 		}
4126 		ctx->ifc_if_flags = if_getflags(ifp);
4127 		CTX_UNLOCK(ctx);
4128 		break;
4129 	case SIOCADDMULTI:
4130 	case SIOCDELMULTI:
4131 		if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4132 			CTX_LOCK(ctx);
4133 			IFDI_INTR_DISABLE(ctx);
4134 			IFDI_MULTI_SET(ctx);
4135 			IFDI_INTR_ENABLE(ctx);
4136 			CTX_UNLOCK(ctx);
4137 		}
4138 		break;
4139 	case SIOCSIFMEDIA:
4140 		CTX_LOCK(ctx);
4141 		IFDI_MEDIA_SET(ctx);
4142 		CTX_UNLOCK(ctx);
4143 		/* falls thru */
4144 	case SIOCGIFMEDIA:
4145 	case SIOCGIFXMEDIA:
4146 		err = ifmedia_ioctl(ifp, ifr, &ctx->ifc_media, command);
4147 		break;
4148 	case SIOCGI2C:
4149 	{
4150 		struct ifi2creq i2c;
4151 
4152 		err = copyin(ifr_data_get_ptr(ifr), &i2c, sizeof(i2c));
4153 		if (err != 0)
4154 			break;
4155 		if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) {
4156 			err = EINVAL;
4157 			break;
4158 		}
4159 		if (i2c.len > sizeof(i2c.data)) {
4160 			err = EINVAL;
4161 			break;
4162 		}
4163 
4164 		if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0)
4165 			err = copyout(&i2c, ifr_data_get_ptr(ifr),
4166 			    sizeof(i2c));
4167 		break;
4168 	}
4169 	case SIOCSIFCAP:
4170 	{
4171 		int mask, setmask;
4172 
4173 		mask = ifr->ifr_reqcap ^ if_getcapenable(ifp);
4174 		setmask = 0;
4175 #ifdef TCP_OFFLOAD
4176 		setmask |= mask & (IFCAP_TOE4|IFCAP_TOE6);
4177 #endif
4178 		setmask |= (mask & IFCAP_FLAGS);
4179 
4180 		if (setmask & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6))
4181 			setmask |= (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6);
4182 		if ((mask & IFCAP_WOL) &&
4183 		    (if_getcapabilities(ifp) & IFCAP_WOL) != 0)
4184 			setmask |= (mask & (IFCAP_WOL_MCAST|IFCAP_WOL_MAGIC));
4185 		if_vlancap(ifp);
4186 		/*
4187 		 * want to ensure that traffic has stopped before we change any of the flags
4188 		 */
4189 		if (setmask) {
4190 			CTX_LOCK(ctx);
4191 			bits = if_getdrvflags(ifp);
4192 			if (bits & IFF_DRV_RUNNING)
4193 				iflib_stop(ctx);
4194 			STATE_LOCK(ctx);
4195 			if_togglecapenable(ifp, setmask);
4196 			STATE_UNLOCK(ctx);
4197 			if (bits & IFF_DRV_RUNNING)
4198 				iflib_init_locked(ctx);
4199 			STATE_LOCK(ctx);
4200 			if_setdrvflags(ifp, bits);
4201 			STATE_UNLOCK(ctx);
4202 			CTX_UNLOCK(ctx);
4203 		}
4204 		break;
4205 	}
4206 	case SIOCGPRIVATE_0:
4207 	case SIOCSDRVSPEC:
4208 	case SIOCGDRVSPEC:
4209 		CTX_LOCK(ctx);
4210 		err = IFDI_PRIV_IOCTL(ctx, command, data);
4211 		CTX_UNLOCK(ctx);
4212 		break;
4213 	default:
4214 		err = ether_ioctl(ifp, command, data);
4215 		break;
4216 	}
4217 	if (reinit)
4218 		iflib_if_init(ctx);
4219 	return (err);
4220 }
4221 
4222 static uint64_t
4223 iflib_if_get_counter(if_t ifp, ift_counter cnt)
4224 {
4225 	if_ctx_t ctx = if_getsoftc(ifp);
4226 
4227 	return (IFDI_GET_COUNTER(ctx, cnt));
4228 }
4229 
4230 /*********************************************************************
4231  *
4232  *  OTHER FUNCTIONS EXPORTED TO THE STACK
4233  *
4234  **********************************************************************/
4235 
4236 static void
4237 iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag)
4238 {
4239 	if_ctx_t ctx = if_getsoftc(ifp);
4240 
4241 	if ((void *)ctx != arg)
4242 		return;
4243 
4244 	if ((vtag == 0) || (vtag > 4095))
4245 		return;
4246 
4247 	CTX_LOCK(ctx);
4248 	IFDI_VLAN_REGISTER(ctx, vtag);
4249 	/* Re-init to load the changes */
4250 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
4251 		iflib_if_init_locked(ctx);
4252 	CTX_UNLOCK(ctx);
4253 }
4254 
4255 static void
4256 iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag)
4257 {
4258 	if_ctx_t ctx = if_getsoftc(ifp);
4259 
4260 	if ((void *)ctx != arg)
4261 		return;
4262 
4263 	if ((vtag == 0) || (vtag > 4095))
4264 		return;
4265 
4266 	CTX_LOCK(ctx);
4267 	IFDI_VLAN_UNREGISTER(ctx, vtag);
4268 	/* Re-init to load the changes */
4269 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
4270 		iflib_if_init_locked(ctx);
4271 	CTX_UNLOCK(ctx);
4272 }
4273 
4274 static void
4275 iflib_led_func(void *arg, int onoff)
4276 {
4277 	if_ctx_t ctx = arg;
4278 
4279 	CTX_LOCK(ctx);
4280 	IFDI_LED_FUNC(ctx, onoff);
4281 	CTX_UNLOCK(ctx);
4282 }
4283 
4284 /*********************************************************************
4285  *
4286  *  BUS FUNCTION DEFINITIONS
4287  *
4288  **********************************************************************/
4289 
4290 int
4291 iflib_device_probe(device_t dev)
4292 {
4293 	pci_vendor_info_t *ent;
4294 
4295 	uint16_t	pci_vendor_id, pci_device_id;
4296 	uint16_t	pci_subvendor_id, pci_subdevice_id;
4297 	uint16_t	pci_rev_id;
4298 	if_shared_ctx_t sctx;
4299 
4300 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4301 		return (ENOTSUP);
4302 
4303 	pci_vendor_id = pci_get_vendor(dev);
4304 	pci_device_id = pci_get_device(dev);
4305 	pci_subvendor_id = pci_get_subvendor(dev);
4306 	pci_subdevice_id = pci_get_subdevice(dev);
4307 	pci_rev_id = pci_get_revid(dev);
4308 	if (sctx->isc_parse_devinfo != NULL)
4309 		sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id);
4310 
4311 	ent = sctx->isc_vendor_info;
4312 	while (ent->pvi_vendor_id != 0) {
4313 		if (pci_vendor_id != ent->pvi_vendor_id) {
4314 			ent++;
4315 			continue;
4316 		}
4317 		if ((pci_device_id == ent->pvi_device_id) &&
4318 		    ((pci_subvendor_id == ent->pvi_subvendor_id) ||
4319 		     (ent->pvi_subvendor_id == 0)) &&
4320 		    ((pci_subdevice_id == ent->pvi_subdevice_id) ||
4321 		     (ent->pvi_subdevice_id == 0)) &&
4322 		    ((pci_rev_id == ent->pvi_rev_id) ||
4323 		     (ent->pvi_rev_id == 0))) {
4324 
4325 			device_set_desc_copy(dev, ent->pvi_name);
4326 			/* this needs to be changed to zero if the bus probing code
4327 			 * ever stops re-probing on best match because the sctx
4328 			 * may have its values over written by register calls
4329 			 * in subsequent probes
4330 			 */
4331 			return (BUS_PROBE_DEFAULT);
4332 		}
4333 		ent++;
4334 	}
4335 	return (ENXIO);
4336 }
4337 
4338 static void
4339 iflib_reset_qvalues(if_ctx_t ctx)
4340 {
4341 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4342 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4343 	device_t dev = ctx->ifc_dev;
4344 	int i;
4345 
4346 	scctx->isc_txrx_budget_bytes_max = IFLIB_MAX_TX_BYTES;
4347 	scctx->isc_tx_qdepth = IFLIB_DEFAULT_TX_QDEPTH;
4348 	/*
4349 	 * XXX sanity check that ntxd & nrxd are a power of 2
4350 	 */
4351 	if (ctx->ifc_sysctl_ntxqs != 0)
4352 		scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs;
4353 	if (ctx->ifc_sysctl_nrxqs != 0)
4354 		scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs;
4355 
4356 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4357 		if (ctx->ifc_sysctl_ntxds[i] != 0)
4358 			scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i];
4359 		else
4360 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4361 	}
4362 
4363 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4364 		if (ctx->ifc_sysctl_nrxds[i] != 0)
4365 			scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i];
4366 		else
4367 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4368 	}
4369 
4370 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4371 		if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) {
4372 			device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n",
4373 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]);
4374 			scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i];
4375 		}
4376 		if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) {
4377 			device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n",
4378 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]);
4379 			scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i];
4380 		}
4381 	}
4382 
4383 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4384 		if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) {
4385 			device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n",
4386 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]);
4387 			scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i];
4388 		}
4389 		if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) {
4390 			device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n",
4391 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]);
4392 			scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i];
4393 		}
4394 	}
4395 }
4396 
4397 int
4398 iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp)
4399 {
4400 	int err, rid, msix;
4401 	if_ctx_t ctx;
4402 	if_t ifp;
4403 	if_softc_ctx_t scctx;
4404 	int i;
4405 	uint16_t main_txq;
4406 	uint16_t main_rxq;
4407 
4408 
4409 	ctx = malloc(sizeof(* ctx), M_IFLIB, M_WAITOK|M_ZERO);
4410 
4411 	if (sc == NULL) {
4412 		sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4413 		device_set_softc(dev, ctx);
4414 		ctx->ifc_flags |= IFC_SC_ALLOCATED;
4415 	}
4416 
4417 	ctx->ifc_sctx = sctx;
4418 	ctx->ifc_dev = dev;
4419 	ctx->ifc_softc = sc;
4420 
4421 	if ((err = iflib_register(ctx)) != 0) {
4422 		if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4423 			free(sc, M_IFLIB);
4424 		free(ctx, M_IFLIB);
4425 		device_printf(dev, "iflib_register failed %d\n", err);
4426 		return (err);
4427 	}
4428 	iflib_add_device_sysctl_pre(ctx);
4429 
4430 	scctx = &ctx->ifc_softc_ctx;
4431 	ifp = ctx->ifc_ifp;
4432 
4433 	iflib_reset_qvalues(ctx);
4434 	CTX_LOCK(ctx);
4435 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4436 		CTX_UNLOCK(ctx);
4437 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4438 		return (err);
4439 	}
4440 	_iflib_pre_assert(scctx);
4441 	ctx->ifc_txrx = *scctx->isc_txrx;
4442 
4443 #ifdef INVARIANTS
4444 	MPASS(scctx->isc_capabilities);
4445 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
4446 		MPASS(scctx->isc_tx_csum_flags);
4447 #endif
4448 
4449 	if_setcapabilities(ifp, scctx->isc_capabilities | IFCAP_HWSTATS);
4450 	if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS);
4451 
4452 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4453 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4454 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4455 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4456 
4457 #ifdef ACPI_DMAR
4458 	if (dmar_get_dma_tag(device_get_parent(dev), dev) != NULL)
4459 		ctx->ifc_flags |= IFC_DMAR;
4460 #elif !(defined(__i386__) || defined(__amd64__))
4461 	/* set unconditionally for !x86 */
4462 	ctx->ifc_flags |= IFC_DMAR;
4463 #endif
4464 
4465 	main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4466 	main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4467 
4468 	/* XXX change for per-queue sizes */
4469 	device_printf(dev, "using %d tx descriptors and %d rx descriptors\n",
4470 		      scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4471 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4472 		if (!powerof2(scctx->isc_nrxd[i])) {
4473 			/* round down instead? */
4474 			device_printf(dev, "# rx descriptors must be a power of 2\n");
4475 			err = EINVAL;
4476 			goto fail;
4477 		}
4478 	}
4479 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4480 		if (!powerof2(scctx->isc_ntxd[i])) {
4481 			device_printf(dev,
4482 			    "# tx descriptors must be a power of 2");
4483 			err = EINVAL;
4484 			goto fail;
4485 		}
4486 	}
4487 
4488 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4489 	    MAX_SINGLE_PACKET_FRACTION)
4490 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4491 		    MAX_SINGLE_PACKET_FRACTION);
4492 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4493 	    MAX_SINGLE_PACKET_FRACTION)
4494 		scctx->isc_tx_tso_segments_max = max(1,
4495 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4496 
4497 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4498 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
4499 		/*
4500 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
4501 		 * but some MACs do.
4502 		 */
4503 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
4504 		    IP_MAXPACKET));
4505 		/*
4506 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
4507 		 * into account.  In the worst case, each of these calls will
4508 		 * add another mbuf and, thus, the requirement for another DMA
4509 		 * segment.  So for best performance, it doesn't make sense to
4510 		 * advertize a maximum of TSO segments that typically will
4511 		 * require defragmentation in iflib_encap().
4512 		 */
4513 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
4514 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
4515 	}
4516 	if (scctx->isc_rss_table_size == 0)
4517 		scctx->isc_rss_table_size = 64;
4518 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4519 
4520 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4521 	/* XXX format name */
4522 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx, -1, "admin");
4523 
4524 	/* Set up cpu set.  If it fails, use the set of all CPUs. */
4525 	if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) != 0) {
4526 		device_printf(dev, "Unable to fetch CPU list\n");
4527 		CPU_COPY(&all_cpus, &ctx->ifc_cpus);
4528 	}
4529 	MPASS(CPU_COUNT(&ctx->ifc_cpus) > 0);
4530 
4531 	/*
4532 	** Now setup MSI or MSI/X, should
4533 	** return us the number of supported
4534 	** vectors. (Will be 1 for MSI)
4535 	*/
4536 	if (sctx->isc_flags & IFLIB_SKIP_MSIX) {
4537 		msix = scctx->isc_vectors;
4538 	} else if (scctx->isc_msix_bar != 0)
4539 	       /*
4540 		* The simple fact that isc_msix_bar is not 0 does not mean we
4541 		* we have a good value there that is known to work.
4542 		*/
4543 		msix = iflib_msix_init(ctx);
4544 	else {
4545 		scctx->isc_vectors = 1;
4546 		scctx->isc_ntxqsets = 1;
4547 		scctx->isc_nrxqsets = 1;
4548 		scctx->isc_intr = IFLIB_INTR_LEGACY;
4549 		msix = 0;
4550 	}
4551 	/* Get memory for the station queues */
4552 	if ((err = iflib_queues_alloc(ctx))) {
4553 		device_printf(dev, "Unable to allocate queue memory\n");
4554 		goto fail;
4555 	}
4556 
4557 	if ((err = iflib_qset_structures_setup(ctx)))
4558 		goto fail_queues;
4559 
4560 	/*
4561 	 * Group taskqueues aren't properly set up until SMP is started,
4562 	 * so we disable interrupts until we can handle them post
4563 	 * SI_SUB_SMP.
4564 	 *
4565 	 * XXX: disabling interrupts doesn't actually work, at least for
4566 	 * the non-MSI case.  When they occur before SI_SUB_SMP completes,
4567 	 * we do null handling and depend on this not causing too large an
4568 	 * interrupt storm.
4569 	 */
4570 	IFDI_INTR_DISABLE(ctx);
4571 	if (msix > 1 && (err = IFDI_MSIX_INTR_ASSIGN(ctx, msix)) != 0) {
4572 		device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n", err);
4573 		goto fail_intr_free;
4574 	}
4575 	if (msix <= 1) {
4576 		rid = 0;
4577 		if (scctx->isc_intr == IFLIB_INTR_MSI) {
4578 			MPASS(msix == 1);
4579 			rid = 1;
4580 		}
4581 		if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) {
4582 			device_printf(dev, "iflib_legacy_setup failed %d\n", err);
4583 			goto fail_intr_free;
4584 		}
4585 	}
4586 
4587 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac);
4588 
4589 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4590 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4591 		goto fail_detach;
4592 	}
4593 
4594 	/*
4595 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4596 	 * This must appear after the call to ether_ifattach() because
4597 	 * ether_ifattach() sets if_hdrlen to the default value.
4598 	 */
4599 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4600 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
4601 
4602 	if ((err = iflib_netmap_attach(ctx))) {
4603 		device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err);
4604 		goto fail_detach;
4605 	}
4606 	*ctxp = ctx;
4607 
4608 	NETDUMP_SET(ctx->ifc_ifp, iflib);
4609 
4610 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4611 	iflib_add_device_sysctl_post(ctx);
4612 	ctx->ifc_flags |= IFC_INIT_DONE;
4613 	CTX_UNLOCK(ctx);
4614 	return (0);
4615 fail_detach:
4616 	ether_ifdetach(ctx->ifc_ifp);
4617 fail_intr_free:
4618 	if (scctx->isc_intr == IFLIB_INTR_MSIX || scctx->isc_intr == IFLIB_INTR_MSI)
4619 		pci_release_msi(ctx->ifc_dev);
4620 fail_queues:
4621 	iflib_tx_structures_free(ctx);
4622 	iflib_rx_structures_free(ctx);
4623 fail:
4624 	IFDI_DETACH(ctx);
4625 	CTX_UNLOCK(ctx);
4626 	return (err);
4627 }
4628 
4629 int
4630 iflib_pseudo_register(device_t dev, if_shared_ctx_t sctx, if_ctx_t *ctxp,
4631 					  struct iflib_cloneattach_ctx *clctx)
4632 {
4633 	int err;
4634 	if_ctx_t ctx;
4635 	if_t ifp;
4636 	if_softc_ctx_t scctx;
4637 	int i;
4638 	void *sc;
4639 	uint16_t main_txq;
4640 	uint16_t main_rxq;
4641 
4642 	ctx = malloc(sizeof(*ctx), M_IFLIB, M_WAITOK|M_ZERO);
4643 	sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4644 	ctx->ifc_flags |= IFC_SC_ALLOCATED;
4645 	if (sctx->isc_flags & (IFLIB_PSEUDO|IFLIB_VIRTUAL))
4646 		ctx->ifc_flags |= IFC_PSEUDO;
4647 
4648 	ctx->ifc_sctx = sctx;
4649 	ctx->ifc_softc = sc;
4650 	ctx->ifc_dev = dev;
4651 
4652 	if ((err = iflib_register(ctx)) != 0) {
4653 		device_printf(dev, "%s: iflib_register failed %d\n", __func__, err);
4654 		free(sc, M_IFLIB);
4655 		free(ctx, M_IFLIB);
4656 		return (err);
4657 	}
4658 	iflib_add_device_sysctl_pre(ctx);
4659 
4660 	scctx = &ctx->ifc_softc_ctx;
4661 	ifp = ctx->ifc_ifp;
4662 
4663 	/*
4664 	 * XXX sanity check that ntxd & nrxd are a power of 2
4665 	 */
4666 	iflib_reset_qvalues(ctx);
4667 
4668 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4669 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4670 		return (err);
4671 	}
4672 	if (sctx->isc_flags & IFLIB_GEN_MAC)
4673 		iflib_gen_mac(ctx);
4674 	if ((err = IFDI_CLONEATTACH(ctx, clctx->cc_ifc, clctx->cc_name,
4675 								clctx->cc_params)) != 0) {
4676 		device_printf(dev, "IFDI_CLONEATTACH failed %d\n", err);
4677 		return (err);
4678 	}
4679 	ifmedia_add(&ctx->ifc_media, IFM_ETHER | IFM_1000_T | IFM_FDX, 0, NULL);
4680 	ifmedia_add(&ctx->ifc_media, IFM_ETHER | IFM_AUTO, 0, NULL);
4681 	ifmedia_set(&ctx->ifc_media, IFM_ETHER | IFM_AUTO);
4682 
4683 #ifdef INVARIANTS
4684 	MPASS(scctx->isc_capabilities);
4685 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
4686 		MPASS(scctx->isc_tx_csum_flags);
4687 #endif
4688 
4689 	if_setcapabilities(ifp, scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_LINKSTATE);
4690 	if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_LINKSTATE);
4691 
4692 	ifp->if_flags |= IFF_NOGROUP;
4693 	if (sctx->isc_flags & IFLIB_PSEUDO) {
4694 		ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac);
4695 
4696 		if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4697 			device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4698 			goto fail_detach;
4699 		}
4700 		*ctxp = ctx;
4701 
4702 		/*
4703 		 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4704 		 * This must appear after the call to ether_ifattach() because
4705 		 * ether_ifattach() sets if_hdrlen to the default value.
4706 		 */
4707 		if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4708 			if_setifheaderlen(ifp,
4709 			    sizeof(struct ether_vlan_header));
4710 
4711 		if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4712 		iflib_add_device_sysctl_post(ctx);
4713 		ctx->ifc_flags |= IFC_INIT_DONE;
4714 		return (0);
4715 	}
4716 	_iflib_pre_assert(scctx);
4717 	ctx->ifc_txrx = *scctx->isc_txrx;
4718 
4719 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4720 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4721 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4722 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4723 
4724 	main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4725 	main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4726 
4727 	/* XXX change for per-queue sizes */
4728 	device_printf(dev, "using %d tx descriptors and %d rx descriptors\n",
4729 		      scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4730 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4731 		if (!powerof2(scctx->isc_nrxd[i])) {
4732 			/* round down instead? */
4733 			device_printf(dev, "# rx descriptors must be a power of 2\n");
4734 			err = EINVAL;
4735 			goto fail;
4736 		}
4737 	}
4738 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4739 		if (!powerof2(scctx->isc_ntxd[i])) {
4740 			device_printf(dev,
4741 			    "# tx descriptors must be a power of 2");
4742 			err = EINVAL;
4743 			goto fail;
4744 		}
4745 	}
4746 
4747 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4748 	    MAX_SINGLE_PACKET_FRACTION)
4749 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4750 		    MAX_SINGLE_PACKET_FRACTION);
4751 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4752 	    MAX_SINGLE_PACKET_FRACTION)
4753 		scctx->isc_tx_tso_segments_max = max(1,
4754 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4755 
4756 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4757 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
4758 		/*
4759 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
4760 		 * but some MACs do.
4761 		 */
4762 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
4763 		    IP_MAXPACKET));
4764 		/*
4765 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
4766 		 * into account.  In the worst case, each of these calls will
4767 		 * add another mbuf and, thus, the requirement for another DMA
4768 		 * segment.  So for best performance, it doesn't make sense to
4769 		 * advertize a maximum of TSO segments that typically will
4770 		 * require defragmentation in iflib_encap().
4771 		 */
4772 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
4773 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
4774 	}
4775 	if (scctx->isc_rss_table_size == 0)
4776 		scctx->isc_rss_table_size = 64;
4777 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4778 
4779 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4780 	/* XXX format name */
4781 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx, -1, "admin");
4782 
4783 	/* XXX --- can support > 1 -- but keep it simple for now */
4784 	scctx->isc_intr = IFLIB_INTR_LEGACY;
4785 
4786 	/* Get memory for the station queues */
4787 	if ((err = iflib_queues_alloc(ctx))) {
4788 		device_printf(dev, "Unable to allocate queue memory\n");
4789 		goto fail;
4790 	}
4791 
4792 	if ((err = iflib_qset_structures_setup(ctx))) {
4793 		device_printf(dev, "qset structure setup failed %d\n", err);
4794 		goto fail_queues;
4795 	}
4796 
4797 	/*
4798 	 * XXX What if anything do we want to do about interrupts?
4799 	 */
4800 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac);
4801 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4802 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4803 		goto fail_detach;
4804 	}
4805 
4806 	/*
4807 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4808 	 * This must appear after the call to ether_ifattach() because
4809 	 * ether_ifattach() sets if_hdrlen to the default value.
4810 	 */
4811 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4812 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
4813 
4814 	/* XXX handle more than one queue */
4815 	for (i = 0; i < scctx->isc_nrxqsets; i++)
4816 		IFDI_RX_CLSET(ctx, 0, i, ctx->ifc_rxqs[i].ifr_fl[0].ifl_sds.ifsd_cl);
4817 
4818 	*ctxp = ctx;
4819 
4820 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4821 	iflib_add_device_sysctl_post(ctx);
4822 	ctx->ifc_flags |= IFC_INIT_DONE;
4823 	return (0);
4824 fail_detach:
4825 	ether_ifdetach(ctx->ifc_ifp);
4826 fail_queues:
4827 	iflib_tx_structures_free(ctx);
4828 	iflib_rx_structures_free(ctx);
4829 fail:
4830 	IFDI_DETACH(ctx);
4831 	return (err);
4832 }
4833 
4834 int
4835 iflib_pseudo_deregister(if_ctx_t ctx)
4836 {
4837 	if_t ifp = ctx->ifc_ifp;
4838 	iflib_txq_t txq;
4839 	iflib_rxq_t rxq;
4840 	int i, j;
4841 	struct taskqgroup *tqg;
4842 	iflib_fl_t fl;
4843 
4844 	/* Unregister VLAN events */
4845 	if (ctx->ifc_vlan_attach_event != NULL)
4846 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
4847 	if (ctx->ifc_vlan_detach_event != NULL)
4848 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
4849 
4850 	ether_ifdetach(ifp);
4851 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
4852 	CTX_LOCK_DESTROY(ctx);
4853 	/* XXX drain any dependent tasks */
4854 	tqg = qgroup_if_io_tqg;
4855 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
4856 		callout_drain(&txq->ift_timer);
4857 		if (txq->ift_task.gt_uniq != NULL)
4858 			taskqgroup_detach(tqg, &txq->ift_task);
4859 	}
4860 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4861 		if (rxq->ifr_task.gt_uniq != NULL)
4862 			taskqgroup_detach(tqg, &rxq->ifr_task);
4863 
4864 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
4865 			free(fl->ifl_rx_bitmap, M_IFLIB);
4866 	}
4867 	tqg = qgroup_if_config_tqg;
4868 	if (ctx->ifc_admin_task.gt_uniq != NULL)
4869 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
4870 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
4871 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
4872 
4873 	if_free(ifp);
4874 
4875 	iflib_tx_structures_free(ctx);
4876 	iflib_rx_structures_free(ctx);
4877 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4878 		free(ctx->ifc_softc, M_IFLIB);
4879 	free(ctx, M_IFLIB);
4880 	return (0);
4881 }
4882 
4883 int
4884 iflib_device_attach(device_t dev)
4885 {
4886 	if_ctx_t ctx;
4887 	if_shared_ctx_t sctx;
4888 
4889 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4890 		return (ENOTSUP);
4891 
4892 	pci_enable_busmaster(dev);
4893 
4894 	return (iflib_device_register(dev, NULL, sctx, &ctx));
4895 }
4896 
4897 int
4898 iflib_device_deregister(if_ctx_t ctx)
4899 {
4900 	if_t ifp = ctx->ifc_ifp;
4901 	iflib_txq_t txq;
4902 	iflib_rxq_t rxq;
4903 	device_t dev = ctx->ifc_dev;
4904 	int i, j;
4905 	struct taskqgroup *tqg;
4906 	iflib_fl_t fl;
4907 
4908 	/* Make sure VLANS are not using driver */
4909 	if (if_vlantrunkinuse(ifp)) {
4910 		device_printf(dev,"Vlan in use, detach first\n");
4911 		return (EBUSY);
4912 	}
4913 
4914 	CTX_LOCK(ctx);
4915 	ctx->ifc_in_detach = 1;
4916 	iflib_stop(ctx);
4917 	CTX_UNLOCK(ctx);
4918 
4919 	/* Unregister VLAN events */
4920 	if (ctx->ifc_vlan_attach_event != NULL)
4921 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
4922 	if (ctx->ifc_vlan_detach_event != NULL)
4923 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
4924 
4925 	iflib_netmap_detach(ifp);
4926 	ether_ifdetach(ifp);
4927 	if (ctx->ifc_led_dev != NULL)
4928 		led_destroy(ctx->ifc_led_dev);
4929 	/* XXX drain any dependent tasks */
4930 	tqg = qgroup_if_io_tqg;
4931 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
4932 		callout_drain(&txq->ift_timer);
4933 		if (txq->ift_task.gt_uniq != NULL)
4934 			taskqgroup_detach(tqg, &txq->ift_task);
4935 	}
4936 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4937 		if (rxq->ifr_task.gt_uniq != NULL)
4938 			taskqgroup_detach(tqg, &rxq->ifr_task);
4939 
4940 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
4941 			free(fl->ifl_rx_bitmap, M_IFLIB);
4942 
4943 	}
4944 	tqg = qgroup_if_config_tqg;
4945 	if (ctx->ifc_admin_task.gt_uniq != NULL)
4946 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
4947 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
4948 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
4949 	CTX_LOCK(ctx);
4950 	IFDI_DETACH(ctx);
4951 	CTX_UNLOCK(ctx);
4952 
4953 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
4954 	CTX_LOCK_DESTROY(ctx);
4955 	device_set_softc(ctx->ifc_dev, NULL);
4956 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) {
4957 		pci_release_msi(dev);
4958 	}
4959 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) {
4960 		iflib_irq_free(ctx, &ctx->ifc_legacy_irq);
4961 	}
4962 	if (ctx->ifc_msix_mem != NULL) {
4963 		bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY,
4964 			ctx->ifc_softc_ctx.isc_msix_bar, ctx->ifc_msix_mem);
4965 		ctx->ifc_msix_mem = NULL;
4966 	}
4967 
4968 	bus_generic_detach(dev);
4969 	if_free(ifp);
4970 
4971 	iflib_tx_structures_free(ctx);
4972 	iflib_rx_structures_free(ctx);
4973 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4974 		free(ctx->ifc_softc, M_IFLIB);
4975 	free(ctx, M_IFLIB);
4976 	return (0);
4977 }
4978 
4979 
4980 int
4981 iflib_device_detach(device_t dev)
4982 {
4983 	if_ctx_t ctx = device_get_softc(dev);
4984 
4985 	return (iflib_device_deregister(ctx));
4986 }
4987 
4988 int
4989 iflib_device_suspend(device_t dev)
4990 {
4991 	if_ctx_t ctx = device_get_softc(dev);
4992 
4993 	CTX_LOCK(ctx);
4994 	IFDI_SUSPEND(ctx);
4995 	CTX_UNLOCK(ctx);
4996 
4997 	return bus_generic_suspend(dev);
4998 }
4999 int
5000 iflib_device_shutdown(device_t dev)
5001 {
5002 	if_ctx_t ctx = device_get_softc(dev);
5003 
5004 	CTX_LOCK(ctx);
5005 	IFDI_SHUTDOWN(ctx);
5006 	CTX_UNLOCK(ctx);
5007 
5008 	return bus_generic_suspend(dev);
5009 }
5010 
5011 
5012 int
5013 iflib_device_resume(device_t dev)
5014 {
5015 	if_ctx_t ctx = device_get_softc(dev);
5016 	iflib_txq_t txq = ctx->ifc_txqs;
5017 
5018 	CTX_LOCK(ctx);
5019 	IFDI_RESUME(ctx);
5020 	iflib_init_locked(ctx);
5021 	CTX_UNLOCK(ctx);
5022 	for (int i = 0; i < NTXQSETS(ctx); i++, txq++)
5023 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
5024 
5025 	return (bus_generic_resume(dev));
5026 }
5027 
5028 int
5029 iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params)
5030 {
5031 	int error;
5032 	if_ctx_t ctx = device_get_softc(dev);
5033 
5034 	CTX_LOCK(ctx);
5035 	error = IFDI_IOV_INIT(ctx, num_vfs, params);
5036 	CTX_UNLOCK(ctx);
5037 
5038 	return (error);
5039 }
5040 
5041 void
5042 iflib_device_iov_uninit(device_t dev)
5043 {
5044 	if_ctx_t ctx = device_get_softc(dev);
5045 
5046 	CTX_LOCK(ctx);
5047 	IFDI_IOV_UNINIT(ctx);
5048 	CTX_UNLOCK(ctx);
5049 }
5050 
5051 int
5052 iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params)
5053 {
5054 	int error;
5055 	if_ctx_t ctx = device_get_softc(dev);
5056 
5057 	CTX_LOCK(ctx);
5058 	error = IFDI_IOV_VF_ADD(ctx, vfnum, params);
5059 	CTX_UNLOCK(ctx);
5060 
5061 	return (error);
5062 }
5063 
5064 /*********************************************************************
5065  *
5066  *  MODULE FUNCTION DEFINITIONS
5067  *
5068  **********************************************************************/
5069 
5070 /*
5071  * - Start a fast taskqueue thread for each core
5072  * - Start a taskqueue for control operations
5073  */
5074 static int
5075 iflib_module_init(void)
5076 {
5077 	return (0);
5078 }
5079 
5080 static int
5081 iflib_module_event_handler(module_t mod, int what, void *arg)
5082 {
5083 	int err;
5084 
5085 	switch (what) {
5086 	case MOD_LOAD:
5087 		if ((err = iflib_module_init()) != 0)
5088 			return (err);
5089 		break;
5090 	case MOD_UNLOAD:
5091 		return (EBUSY);
5092 	default:
5093 		return (EOPNOTSUPP);
5094 	}
5095 
5096 	return (0);
5097 }
5098 
5099 /*********************************************************************
5100  *
5101  *  PUBLIC FUNCTION DEFINITIONS
5102  *     ordered as in iflib.h
5103  *
5104  **********************************************************************/
5105 
5106 
5107 static void
5108 _iflib_assert(if_shared_ctx_t sctx)
5109 {
5110 	MPASS(sctx->isc_tx_maxsize);
5111 	MPASS(sctx->isc_tx_maxsegsize);
5112 
5113 	MPASS(sctx->isc_rx_maxsize);
5114 	MPASS(sctx->isc_rx_nsegments);
5115 	MPASS(sctx->isc_rx_maxsegsize);
5116 
5117 	MPASS(sctx->isc_nrxd_min[0]);
5118 	MPASS(sctx->isc_nrxd_max[0]);
5119 	MPASS(sctx->isc_nrxd_default[0]);
5120 	MPASS(sctx->isc_ntxd_min[0]);
5121 	MPASS(sctx->isc_ntxd_max[0]);
5122 	MPASS(sctx->isc_ntxd_default[0]);
5123 }
5124 
5125 static void
5126 _iflib_pre_assert(if_softc_ctx_t scctx)
5127 {
5128 
5129 	MPASS(scctx->isc_txrx->ift_txd_encap);
5130 	MPASS(scctx->isc_txrx->ift_txd_flush);
5131 	MPASS(scctx->isc_txrx->ift_txd_credits_update);
5132 	MPASS(scctx->isc_txrx->ift_rxd_available);
5133 	MPASS(scctx->isc_txrx->ift_rxd_pkt_get);
5134 	MPASS(scctx->isc_txrx->ift_rxd_refill);
5135 	MPASS(scctx->isc_txrx->ift_rxd_flush);
5136 }
5137 
5138 static int
5139 iflib_register(if_ctx_t ctx)
5140 {
5141 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5142 	driver_t *driver = sctx->isc_driver;
5143 	device_t dev = ctx->ifc_dev;
5144 	if_t ifp;
5145 
5146 	_iflib_assert(sctx);
5147 
5148 	CTX_LOCK_INIT(ctx);
5149 	STATE_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev));
5150 	ifp = ctx->ifc_ifp = if_gethandle(IFT_ETHER);
5151 	if (ifp == NULL) {
5152 		device_printf(dev, "can not allocate ifnet structure\n");
5153 		return (ENOMEM);
5154 	}
5155 
5156 	/*
5157 	 * Initialize our context's device specific methods
5158 	 */
5159 	kobj_init((kobj_t) ctx, (kobj_class_t) driver);
5160 	kobj_class_compile((kobj_class_t) driver);
5161 	driver->refs++;
5162 
5163 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
5164 	if_setsoftc(ifp, ctx);
5165 	if_setdev(ifp, dev);
5166 	if_setinitfn(ifp, iflib_if_init);
5167 	if_setioctlfn(ifp, iflib_if_ioctl);
5168 	if_settransmitfn(ifp, iflib_if_transmit);
5169 	if_setqflushfn(ifp, iflib_if_qflush);
5170 	if_setflags(ifp, IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST);
5171 
5172 	ctx->ifc_vlan_attach_event =
5173 		EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx,
5174 							  EVENTHANDLER_PRI_FIRST);
5175 	ctx->ifc_vlan_detach_event =
5176 		EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx,
5177 							  EVENTHANDLER_PRI_FIRST);
5178 
5179 	ifmedia_init(&ctx->ifc_media, IFM_IMASK,
5180 					 iflib_media_change, iflib_media_status);
5181 
5182 	return (0);
5183 }
5184 
5185 
5186 static int
5187 iflib_queues_alloc(if_ctx_t ctx)
5188 {
5189 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5190 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5191 	device_t dev = ctx->ifc_dev;
5192 	int nrxqsets = scctx->isc_nrxqsets;
5193 	int ntxqsets = scctx->isc_ntxqsets;
5194 	iflib_txq_t txq;
5195 	iflib_rxq_t rxq;
5196 	iflib_fl_t fl = NULL;
5197 	int i, j, cpu, err, txconf, rxconf;
5198 	iflib_dma_info_t ifdip;
5199 	uint32_t *rxqsizes = scctx->isc_rxqsizes;
5200 	uint32_t *txqsizes = scctx->isc_txqsizes;
5201 	uint8_t nrxqs = sctx->isc_nrxqs;
5202 	uint8_t ntxqs = sctx->isc_ntxqs;
5203 	int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1;
5204 	caddr_t *vaddrs;
5205 	uint64_t *paddrs;
5206 
5207 	KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1"));
5208 	KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1"));
5209 
5210 	/* Allocate the TX ring struct memory */
5211 	if (!(ctx->ifc_txqs =
5212 	    (iflib_txq_t) malloc(sizeof(struct iflib_txq) *
5213 	    ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5214 		device_printf(dev, "Unable to allocate TX ring memory\n");
5215 		err = ENOMEM;
5216 		goto fail;
5217 	}
5218 
5219 	/* Now allocate the RX */
5220 	if (!(ctx->ifc_rxqs =
5221 	    (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) *
5222 	    nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5223 		device_printf(dev, "Unable to allocate RX ring memory\n");
5224 		err = ENOMEM;
5225 		goto rx_fail;
5226 	}
5227 
5228 	txq = ctx->ifc_txqs;
5229 	rxq = ctx->ifc_rxqs;
5230 
5231 	/*
5232 	 * XXX handle allocation failure
5233 	 */
5234 	for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) {
5235 		/* Set up some basics */
5236 
5237 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs, M_IFLIB, M_WAITOK|M_ZERO)) == NULL) {
5238 			device_printf(dev, "failed to allocate iflib_dma_info\n");
5239 			err = ENOMEM;
5240 			goto err_tx_desc;
5241 		}
5242 		txq->ift_ifdi = ifdip;
5243 		for (j = 0; j < ntxqs; j++, ifdip++) {
5244 			if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, BUS_DMA_NOWAIT)) {
5245 				device_printf(dev, "Unable to allocate Descriptor memory\n");
5246 				err = ENOMEM;
5247 				goto err_tx_desc;
5248 			}
5249 			txq->ift_txd_size[j] = scctx->isc_txd_size[j];
5250 			bzero((void *)ifdip->idi_vaddr, txqsizes[j]);
5251 		}
5252 		txq->ift_ctx = ctx;
5253 		txq->ift_id = i;
5254 		if (sctx->isc_flags & IFLIB_HAS_TXCQ) {
5255 			txq->ift_br_offset = 1;
5256 		} else {
5257 			txq->ift_br_offset = 0;
5258 		}
5259 		/* XXX fix this */
5260 		txq->ift_timer.c_cpu = cpu;
5261 
5262 		if (iflib_txsd_alloc(txq)) {
5263 			device_printf(dev, "Critical Failure setting up TX buffers\n");
5264 			err = ENOMEM;
5265 			goto err_tx_desc;
5266 		}
5267 
5268 		/* Initialize the TX lock */
5269 		snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:tx(%d):callout",
5270 		    device_get_nameunit(dev), txq->ift_id);
5271 		mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF);
5272 		callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0);
5273 
5274 		snprintf(txq->ift_db_mtx_name, MTX_NAME_LEN, "%s:tx(%d):db",
5275 			 device_get_nameunit(dev), txq->ift_id);
5276 
5277 		err = ifmp_ring_alloc(&txq->ift_br, 2048, txq, iflib_txq_drain,
5278 				      iflib_txq_can_drain, M_IFLIB, M_WAITOK);
5279 		if (err) {
5280 			/* XXX free any allocated rings */
5281 			device_printf(dev, "Unable to allocate buf_ring\n");
5282 			goto err_tx_desc;
5283 		}
5284 	}
5285 
5286 	for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) {
5287 		/* Set up some basics */
5288 
5289 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs, M_IFLIB, M_WAITOK|M_ZERO)) == NULL) {
5290 			device_printf(dev, "failed to allocate iflib_dma_info\n");
5291 			err = ENOMEM;
5292 			goto err_tx_desc;
5293 		}
5294 
5295 		rxq->ifr_ifdi = ifdip;
5296 		/* XXX this needs to be changed if #rx queues != #tx queues */
5297 		rxq->ifr_ntxqirq = 1;
5298 		rxq->ifr_txqid[0] = i;
5299 		for (j = 0; j < nrxqs; j++, ifdip++) {
5300 			if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, BUS_DMA_NOWAIT)) {
5301 				device_printf(dev, "Unable to allocate Descriptor memory\n");
5302 				err = ENOMEM;
5303 				goto err_tx_desc;
5304 			}
5305 			bzero((void *)ifdip->idi_vaddr, rxqsizes[j]);
5306 		}
5307 		rxq->ifr_ctx = ctx;
5308 		rxq->ifr_id = i;
5309 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
5310 			rxq->ifr_fl_offset = 1;
5311 		} else {
5312 			rxq->ifr_fl_offset = 0;
5313 		}
5314 		rxq->ifr_nfl = nfree_lists;
5315 		if (!(fl =
5316 			  (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) {
5317 			device_printf(dev, "Unable to allocate free list memory\n");
5318 			err = ENOMEM;
5319 			goto err_tx_desc;
5320 		}
5321 		rxq->ifr_fl = fl;
5322 		for (j = 0; j < nfree_lists; j++) {
5323 			fl[j].ifl_rxq = rxq;
5324 			fl[j].ifl_id = j;
5325 			fl[j].ifl_ifdi = &rxq->ifr_ifdi[j + rxq->ifr_fl_offset];
5326 			fl[j].ifl_rxd_size = scctx->isc_rxd_size[j];
5327 		}
5328         /* Allocate receive buffers for the ring*/
5329 		if (iflib_rxsd_alloc(rxq)) {
5330 			device_printf(dev,
5331 			    "Critical Failure setting up receive buffers\n");
5332 			err = ENOMEM;
5333 			goto err_rx_desc;
5334 		}
5335 
5336 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5337 			fl->ifl_rx_bitmap = bit_alloc(fl->ifl_size, M_IFLIB, M_WAITOK|M_ZERO);
5338 	}
5339 
5340 	/* TXQs */
5341 	vaddrs = malloc(sizeof(caddr_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5342 	paddrs = malloc(sizeof(uint64_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5343 	for (i = 0; i < ntxqsets; i++) {
5344 		iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi;
5345 
5346 		for (j = 0; j < ntxqs; j++, di++) {
5347 			vaddrs[i*ntxqs + j] = di->idi_vaddr;
5348 			paddrs[i*ntxqs + j] = di->idi_paddr;
5349 		}
5350 	}
5351 	if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) {
5352 		device_printf(ctx->ifc_dev, "device queue allocation failed\n");
5353 		iflib_tx_structures_free(ctx);
5354 		free(vaddrs, M_IFLIB);
5355 		free(paddrs, M_IFLIB);
5356 		goto err_rx_desc;
5357 	}
5358 	free(vaddrs, M_IFLIB);
5359 	free(paddrs, M_IFLIB);
5360 
5361 	/* RXQs */
5362 	vaddrs = malloc(sizeof(caddr_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5363 	paddrs = malloc(sizeof(uint64_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5364 	for (i = 0; i < nrxqsets; i++) {
5365 		iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi;
5366 
5367 		for (j = 0; j < nrxqs; j++, di++) {
5368 			vaddrs[i*nrxqs + j] = di->idi_vaddr;
5369 			paddrs[i*nrxqs + j] = di->idi_paddr;
5370 		}
5371 	}
5372 	if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) {
5373 		device_printf(ctx->ifc_dev, "device queue allocation failed\n");
5374 		iflib_tx_structures_free(ctx);
5375 		free(vaddrs, M_IFLIB);
5376 		free(paddrs, M_IFLIB);
5377 		goto err_rx_desc;
5378 	}
5379 	free(vaddrs, M_IFLIB);
5380 	free(paddrs, M_IFLIB);
5381 
5382 	return (0);
5383 
5384 /* XXX handle allocation failure changes */
5385 err_rx_desc:
5386 err_tx_desc:
5387 rx_fail:
5388 	if (ctx->ifc_rxqs != NULL)
5389 		free(ctx->ifc_rxqs, M_IFLIB);
5390 	ctx->ifc_rxqs = NULL;
5391 	if (ctx->ifc_txqs != NULL)
5392 		free(ctx->ifc_txqs, M_IFLIB);
5393 	ctx->ifc_txqs = NULL;
5394 fail:
5395 	return (err);
5396 }
5397 
5398 static int
5399 iflib_tx_structures_setup(if_ctx_t ctx)
5400 {
5401 	iflib_txq_t txq = ctx->ifc_txqs;
5402 	int i;
5403 
5404 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
5405 		iflib_txq_setup(txq);
5406 
5407 	return (0);
5408 }
5409 
5410 static void
5411 iflib_tx_structures_free(if_ctx_t ctx)
5412 {
5413 	iflib_txq_t txq = ctx->ifc_txqs;
5414 	int i, j;
5415 
5416 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
5417 		iflib_txq_destroy(txq);
5418 		for (j = 0; j < ctx->ifc_nhwtxqs; j++)
5419 			iflib_dma_free(&txq->ift_ifdi[j]);
5420 	}
5421 	free(ctx->ifc_txqs, M_IFLIB);
5422 	ctx->ifc_txqs = NULL;
5423 	IFDI_QUEUES_FREE(ctx);
5424 }
5425 
5426 /*********************************************************************
5427  *
5428  *  Initialize all receive rings.
5429  *
5430  **********************************************************************/
5431 static int
5432 iflib_rx_structures_setup(if_ctx_t ctx)
5433 {
5434 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5435 	int q;
5436 #if defined(INET6) || defined(INET)
5437 	int i, err;
5438 #endif
5439 
5440 	for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) {
5441 #if defined(INET6) || defined(INET)
5442 		tcp_lro_free(&rxq->ifr_lc);
5443 		if ((err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp,
5444 		    TCP_LRO_ENTRIES, min(1024,
5445 		    ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset]))) != 0) {
5446 			device_printf(ctx->ifc_dev, "LRO Initialization failed!\n");
5447 			goto fail;
5448 		}
5449 		rxq->ifr_lro_enabled = TRUE;
5450 #endif
5451 		IFDI_RXQ_SETUP(ctx, rxq->ifr_id);
5452 	}
5453 	return (0);
5454 #if defined(INET6) || defined(INET)
5455 fail:
5456 	/*
5457 	 * Free RX software descriptors allocated so far, we will only handle
5458 	 * the rings that completed, the failing case will have
5459 	 * cleaned up for itself. 'q' failed, so its the terminus.
5460 	 */
5461 	rxq = ctx->ifc_rxqs;
5462 	for (i = 0; i < q; ++i, rxq++) {
5463 		iflib_rx_sds_free(rxq);
5464 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
5465 	}
5466 	return (err);
5467 #endif
5468 }
5469 
5470 /*********************************************************************
5471  *
5472  *  Free all receive rings.
5473  *
5474  **********************************************************************/
5475 static void
5476 iflib_rx_structures_free(if_ctx_t ctx)
5477 {
5478 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5479 
5480 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) {
5481 		iflib_rx_sds_free(rxq);
5482 	}
5483 }
5484 
5485 static int
5486 iflib_qset_structures_setup(if_ctx_t ctx)
5487 {
5488 	int err;
5489 
5490 	/*
5491 	 * It is expected that the caller takes care of freeing queues if this
5492 	 * fails.
5493 	 */
5494 	if ((err = iflib_tx_structures_setup(ctx)) != 0) {
5495 		device_printf(ctx->ifc_dev, "iflib_tx_structures_setup failed: %d\n", err);
5496 		return (err);
5497 	}
5498 
5499 	if ((err = iflib_rx_structures_setup(ctx)) != 0)
5500 		device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err);
5501 
5502 	return (err);
5503 }
5504 
5505 int
5506 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
5507 		driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, const char *name)
5508 {
5509 
5510 	return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name));
5511 }
5512 
5513 #ifdef SMP
5514 static int
5515 find_nth(if_ctx_t ctx, int qid)
5516 {
5517 	cpuset_t cpus;
5518 	int i, cpuid, eqid, count;
5519 
5520 	CPU_COPY(&ctx->ifc_cpus, &cpus);
5521 	count = CPU_COUNT(&cpus);
5522 	eqid = qid % count;
5523 	/* clear up to the qid'th bit */
5524 	for (i = 0; i < eqid; i++) {
5525 		cpuid = CPU_FFS(&cpus);
5526 		MPASS(cpuid != 0);
5527 		CPU_CLR(cpuid-1, &cpus);
5528 	}
5529 	cpuid = CPU_FFS(&cpus);
5530 	MPASS(cpuid != 0);
5531 	return (cpuid-1);
5532 }
5533 
5534 #ifdef SCHED_ULE
5535 extern struct cpu_group *cpu_top;              /* CPU topology */
5536 
5537 static int
5538 find_child_with_core(int cpu, struct cpu_group *grp)
5539 {
5540 	int i;
5541 
5542 	if (grp->cg_children == 0)
5543 		return -1;
5544 
5545 	MPASS(grp->cg_child);
5546 	for (i = 0; i < grp->cg_children; i++) {
5547 		if (CPU_ISSET(cpu, &grp->cg_child[i].cg_mask))
5548 			return i;
5549 	}
5550 
5551 	return -1;
5552 }
5553 
5554 /*
5555  * Find the nth "close" core to the specified core
5556  * "close" is defined as the deepest level that shares
5557  * at least an L2 cache.  With threads, this will be
5558  * threads on the same core.  If the sahred cache is L3
5559  * or higher, simply returns the same core.
5560  */
5561 static int
5562 find_close_core(int cpu, int core_offset)
5563 {
5564 	struct cpu_group *grp;
5565 	int i;
5566 	int fcpu;
5567 	cpuset_t cs;
5568 
5569 	grp = cpu_top;
5570 	if (grp == NULL)
5571 		return cpu;
5572 	i = 0;
5573 	while ((i = find_child_with_core(cpu, grp)) != -1) {
5574 		/* If the child only has one cpu, don't descend */
5575 		if (grp->cg_child[i].cg_count <= 1)
5576 			break;
5577 		grp = &grp->cg_child[i];
5578 	}
5579 
5580 	/* If they don't share at least an L2 cache, use the same CPU */
5581 	if (grp->cg_level > CG_SHARE_L2 || grp->cg_level == CG_SHARE_NONE)
5582 		return cpu;
5583 
5584 	/* Now pick one */
5585 	CPU_COPY(&grp->cg_mask, &cs);
5586 
5587 	/* Add the selected CPU offset to core offset. */
5588 	for (i = 0; (fcpu = CPU_FFS(&cs)) != 0; i++) {
5589 		if (fcpu - 1 == cpu)
5590 			break;
5591 		CPU_CLR(fcpu - 1, &cs);
5592 	}
5593 	MPASS(fcpu);
5594 
5595 	core_offset += i;
5596 
5597 	CPU_COPY(&grp->cg_mask, &cs);
5598 	for (i = core_offset % grp->cg_count; i > 0; i--) {
5599 		MPASS(CPU_FFS(&cs));
5600 		CPU_CLR(CPU_FFS(&cs) - 1, &cs);
5601 	}
5602 	MPASS(CPU_FFS(&cs));
5603 	return CPU_FFS(&cs) - 1;
5604 }
5605 #else
5606 static int
5607 find_close_core(int cpu, int core_offset __unused)
5608 {
5609 	return cpu;
5610 }
5611 #endif
5612 
5613 static int
5614 get_core_offset(if_ctx_t ctx, iflib_intr_type_t type, int qid)
5615 {
5616 	switch (type) {
5617 	case IFLIB_INTR_TX:
5618 		/* TX queues get cores which share at least an L2 cache with the corresponding RX queue */
5619 		/* XXX handle multiple RX threads per core and more than two core per L2 group */
5620 		return qid / CPU_COUNT(&ctx->ifc_cpus) + 1;
5621 	case IFLIB_INTR_RX:
5622 	case IFLIB_INTR_RXTX:
5623 		/* RX queues get the specified core */
5624 		return qid / CPU_COUNT(&ctx->ifc_cpus);
5625 	default:
5626 		return -1;
5627 	}
5628 }
5629 #else
5630 #define get_core_offset(ctx, type, qid)	CPU_FIRST()
5631 #define find_close_core(cpuid, tid)	CPU_FIRST()
5632 #define find_nth(ctx, gid)		CPU_FIRST()
5633 #endif
5634 
5635 /* Just to avoid copy/paste */
5636 static inline int
5637 iflib_irq_set_affinity(if_ctx_t ctx, int irq, iflib_intr_type_t type, int qid,
5638     struct grouptask *gtask, struct taskqgroup *tqg, void *uniq, const char *name)
5639 {
5640 	int cpuid;
5641 	int err, tid;
5642 
5643 	cpuid = find_nth(ctx, qid);
5644 	tid = get_core_offset(ctx, type, qid);
5645 	MPASS(tid >= 0);
5646 	cpuid = find_close_core(cpuid, tid);
5647 	err = taskqgroup_attach_cpu(tqg, gtask, uniq, cpuid, irq, name);
5648 	if (err) {
5649 		device_printf(ctx->ifc_dev, "taskqgroup_attach_cpu failed %d\n", err);
5650 		return (err);
5651 	}
5652 #ifdef notyet
5653 	if (cpuid > ctx->ifc_cpuid_highest)
5654 		ctx->ifc_cpuid_highest = cpuid;
5655 #endif
5656 	return 0;
5657 }
5658 
5659 int
5660 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid,
5661 			iflib_intr_type_t type, driver_filter_t *filter,
5662 			void *filter_arg, int qid, const char *name)
5663 {
5664 	struct grouptask *gtask;
5665 	struct taskqgroup *tqg;
5666 	iflib_filter_info_t info;
5667 	gtask_fn_t *fn;
5668 	int tqrid, err;
5669 	driver_filter_t *intr_fast;
5670 	void *q;
5671 
5672 	info = &ctx->ifc_filter_info;
5673 	tqrid = rid;
5674 
5675 	switch (type) {
5676 	/* XXX merge tx/rx for netmap? */
5677 	case IFLIB_INTR_TX:
5678 		q = &ctx->ifc_txqs[qid];
5679 		info = &ctx->ifc_txqs[qid].ift_filter_info;
5680 		gtask = &ctx->ifc_txqs[qid].ift_task;
5681 		tqg = qgroup_if_io_tqg;
5682 		fn = _task_fn_tx;
5683 		intr_fast = iflib_fast_intr;
5684 		GROUPTASK_INIT(gtask, 0, fn, q);
5685 		ctx->ifc_flags |= IFC_NETMAP_TX_IRQ;
5686 		break;
5687 	case IFLIB_INTR_RX:
5688 		q = &ctx->ifc_rxqs[qid];
5689 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
5690 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5691 		tqg = qgroup_if_io_tqg;
5692 		fn = _task_fn_rx;
5693 		intr_fast = iflib_fast_intr;
5694 		GROUPTASK_INIT(gtask, 0, fn, q);
5695 		break;
5696 	case IFLIB_INTR_RXTX:
5697 		q = &ctx->ifc_rxqs[qid];
5698 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
5699 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5700 		tqg = qgroup_if_io_tqg;
5701 		fn = _task_fn_rx;
5702 		intr_fast = iflib_fast_intr_rxtx;
5703 		GROUPTASK_INIT(gtask, 0, fn, q);
5704 		break;
5705 	case IFLIB_INTR_ADMIN:
5706 		q = ctx;
5707 		tqrid = -1;
5708 		info = &ctx->ifc_filter_info;
5709 		gtask = &ctx->ifc_admin_task;
5710 		tqg = qgroup_if_config_tqg;
5711 		fn = _task_fn_admin;
5712 		intr_fast = iflib_fast_intr_ctx;
5713 		break;
5714 	default:
5715 		panic("unknown net intr type");
5716 	}
5717 
5718 	info->ifi_filter = filter;
5719 	info->ifi_filter_arg = filter_arg;
5720 	info->ifi_task = gtask;
5721 	info->ifi_ctx = q;
5722 
5723 	err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info,  name);
5724 	if (err != 0) {
5725 		device_printf(ctx->ifc_dev, "_iflib_irq_alloc failed %d\n", err);
5726 		return (err);
5727 	}
5728 	if (type == IFLIB_INTR_ADMIN)
5729 		return (0);
5730 
5731 	if (tqrid != -1) {
5732 		err = iflib_irq_set_affinity(ctx, rman_get_start(irq->ii_res), type, qid, gtask, tqg, q, name);
5733 		if (err)
5734 			return (err);
5735 	} else {
5736 		taskqgroup_attach(tqg, gtask, q, rman_get_start(irq->ii_res), name);
5737 	}
5738 
5739 	return (0);
5740 }
5741 
5742 void
5743 iflib_softirq_alloc_generic(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type,  void *arg, int qid, const char *name)
5744 {
5745 	struct grouptask *gtask;
5746 	struct taskqgroup *tqg;
5747 	gtask_fn_t *fn;
5748 	void *q;
5749 	int irq_num = -1;
5750 	int err;
5751 
5752 	switch (type) {
5753 	case IFLIB_INTR_TX:
5754 		q = &ctx->ifc_txqs[qid];
5755 		gtask = &ctx->ifc_txqs[qid].ift_task;
5756 		tqg = qgroup_if_io_tqg;
5757 		fn = _task_fn_tx;
5758 		if (irq != NULL)
5759 			irq_num = rman_get_start(irq->ii_res);
5760 		break;
5761 	case IFLIB_INTR_RX:
5762 		q = &ctx->ifc_rxqs[qid];
5763 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5764 		tqg = qgroup_if_io_tqg;
5765 		fn = _task_fn_rx;
5766 		if (irq != NULL)
5767 			irq_num = rman_get_start(irq->ii_res);
5768 		break;
5769 	case IFLIB_INTR_IOV:
5770 		q = ctx;
5771 		gtask = &ctx->ifc_vflr_task;
5772 		tqg = qgroup_if_config_tqg;
5773 		fn = _task_fn_iov;
5774 		break;
5775 	default:
5776 		panic("unknown net intr type");
5777 	}
5778 	GROUPTASK_INIT(gtask, 0, fn, q);
5779 	if (irq_num != -1) {
5780 		err = iflib_irq_set_affinity(ctx, irq_num, type, qid, gtask, tqg, q, name);
5781 		if (err)
5782 			taskqgroup_attach(tqg, gtask, q, irq_num, name);
5783 	}
5784 	else {
5785 		taskqgroup_attach(tqg, gtask, q, irq_num, name);
5786 	}
5787 }
5788 
5789 void
5790 iflib_irq_free(if_ctx_t ctx, if_irq_t irq)
5791 {
5792 	if (irq->ii_tag)
5793 		bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag);
5794 
5795 	if (irq->ii_res)
5796 		bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ, irq->ii_rid, irq->ii_res);
5797 }
5798 
5799 static int
5800 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, const char *name)
5801 {
5802 	iflib_txq_t txq = ctx->ifc_txqs;
5803 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5804 	if_irq_t irq = &ctx->ifc_legacy_irq;
5805 	iflib_filter_info_t info;
5806 	struct grouptask *gtask;
5807 	struct taskqgroup *tqg;
5808 	gtask_fn_t *fn;
5809 	int tqrid;
5810 	void *q;
5811 	int err;
5812 
5813 	q = &ctx->ifc_rxqs[0];
5814 	info = &rxq[0].ifr_filter_info;
5815 	gtask = &rxq[0].ifr_task;
5816 	tqg = qgroup_if_io_tqg;
5817 	tqrid = irq->ii_rid = *rid;
5818 	fn = _task_fn_rx;
5819 
5820 	ctx->ifc_flags |= IFC_LEGACY;
5821 	info->ifi_filter = filter;
5822 	info->ifi_filter_arg = filter_arg;
5823 	info->ifi_task = gtask;
5824 	info->ifi_ctx = ctx;
5825 
5826 	/* We allocate a single interrupt resource */
5827 	if ((err = _iflib_irq_alloc(ctx, irq, tqrid, iflib_fast_intr_ctx, NULL, info, name)) != 0)
5828 		return (err);
5829 	GROUPTASK_INIT(gtask, 0, fn, q);
5830 	taskqgroup_attach(tqg, gtask, q, rman_get_start(irq->ii_res), name);
5831 
5832 	GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq);
5833 	taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, rman_get_start(irq->ii_res), "tx");
5834 	return (0);
5835 }
5836 
5837 void
5838 iflib_led_create(if_ctx_t ctx)
5839 {
5840 
5841 	ctx->ifc_led_dev = led_create(iflib_led_func, ctx,
5842 	    device_get_nameunit(ctx->ifc_dev));
5843 }
5844 
5845 void
5846 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid)
5847 {
5848 
5849 	GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
5850 }
5851 
5852 void
5853 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid)
5854 {
5855 
5856 	GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task);
5857 }
5858 
5859 void
5860 iflib_admin_intr_deferred(if_ctx_t ctx)
5861 {
5862 #ifdef INVARIANTS
5863 	struct grouptask *gtask;
5864 
5865 	gtask = &ctx->ifc_admin_task;
5866 	MPASS(gtask != NULL && gtask->gt_taskqueue != NULL);
5867 #endif
5868 
5869 	GROUPTASK_ENQUEUE(&ctx->ifc_admin_task);
5870 }
5871 
5872 void
5873 iflib_iov_intr_deferred(if_ctx_t ctx)
5874 {
5875 
5876 	GROUPTASK_ENQUEUE(&ctx->ifc_vflr_task);
5877 }
5878 
5879 void
5880 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, char *name)
5881 {
5882 
5883 	taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, -1, name);
5884 }
5885 
5886 void
5887 iflib_config_gtask_init(void *ctx, struct grouptask *gtask, gtask_fn_t *fn,
5888 	const char *name)
5889 {
5890 
5891 	GROUPTASK_INIT(gtask, 0, fn, ctx);
5892 	taskqgroup_attach(qgroup_if_config_tqg, gtask, gtask, -1, name);
5893 }
5894 
5895 void
5896 iflib_config_gtask_deinit(struct grouptask *gtask)
5897 {
5898 
5899 	taskqgroup_detach(qgroup_if_config_tqg, gtask);
5900 }
5901 
5902 void
5903 iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate)
5904 {
5905 	if_t ifp = ctx->ifc_ifp;
5906 	iflib_txq_t txq = ctx->ifc_txqs;
5907 
5908 	if_setbaudrate(ifp, baudrate);
5909 	if (baudrate >= IF_Gbps(10)) {
5910 		STATE_LOCK(ctx);
5911 		ctx->ifc_flags |= IFC_PREFETCH;
5912 		STATE_UNLOCK(ctx);
5913 	}
5914 	/* If link down, disable watchdog */
5915 	if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) {
5916 		for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++)
5917 			txq->ift_qstatus = IFLIB_QUEUE_IDLE;
5918 	}
5919 	ctx->ifc_link_state = link_state;
5920 	if_link_state_change(ifp, link_state);
5921 }
5922 
5923 static int
5924 iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq)
5925 {
5926 	int credits;
5927 #ifdef INVARIANTS
5928 	int credits_pre = txq->ift_cidx_processed;
5929 #endif
5930 
5931 	if (ctx->isc_txd_credits_update == NULL)
5932 		return (0);
5933 
5934 	if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, true)) == 0)
5935 		return (0);
5936 
5937 	txq->ift_processed += credits;
5938 	txq->ift_cidx_processed += credits;
5939 
5940 	MPASS(credits_pre + credits == txq->ift_cidx_processed);
5941 	if (txq->ift_cidx_processed >= txq->ift_size)
5942 		txq->ift_cidx_processed -= txq->ift_size;
5943 	return (credits);
5944 }
5945 
5946 static int
5947 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget)
5948 {
5949 
5950 	return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx,
5951 	    budget));
5952 }
5953 
5954 void
5955 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name,
5956 	const char *description, if_int_delay_info_t info,
5957 	int offset, int value)
5958 {
5959 	info->iidi_ctx = ctx;
5960 	info->iidi_offset = offset;
5961 	info->iidi_value = value;
5962 	SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev),
5963 	    SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)),
5964 	    OID_AUTO, name, CTLTYPE_INT|CTLFLAG_RW,
5965 	    info, 0, iflib_sysctl_int_delay, "I", description);
5966 }
5967 
5968 struct sx *
5969 iflib_ctx_lock_get(if_ctx_t ctx)
5970 {
5971 
5972 	return (&ctx->ifc_ctx_sx);
5973 }
5974 
5975 static int
5976 iflib_msix_init(if_ctx_t ctx)
5977 {
5978 	device_t dev = ctx->ifc_dev;
5979 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5980 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5981 	int vectors, queues, rx_queues, tx_queues, queuemsgs, msgs;
5982 	int iflib_num_tx_queues, iflib_num_rx_queues;
5983 	int err, admincnt, bar;
5984 
5985 	iflib_num_tx_queues = ctx->ifc_sysctl_ntxqs;
5986 	iflib_num_rx_queues = ctx->ifc_sysctl_nrxqs;
5987 
5988 	device_printf(dev, "msix_init qsets capped at %d\n", imax(scctx->isc_ntxqsets, scctx->isc_nrxqsets));
5989 
5990 	bar = ctx->ifc_softc_ctx.isc_msix_bar;
5991 	admincnt = sctx->isc_admin_intrcnt;
5992 	/* Override by tuneable */
5993 	if (scctx->isc_disable_msix)
5994 		goto msi;
5995 
5996 	/*
5997 	 * bar == -1 => "trust me I know what I'm doing"
5998 	 * Some drivers are for hardware that is so shoddily
5999 	 * documented that no one knows which bars are which
6000 	 * so the developer has to map all bars. This hack
6001 	 * allows shoddy garbage to use msix in this framework.
6002 	 */
6003 	if (bar != -1) {
6004 		ctx->ifc_msix_mem = bus_alloc_resource_any(dev,
6005 	            SYS_RES_MEMORY, &bar, RF_ACTIVE);
6006 		if (ctx->ifc_msix_mem == NULL) {
6007 			/* May not be enabled */
6008 			device_printf(dev, "Unable to map MSIX table \n");
6009 			goto msi;
6010 		}
6011 	}
6012 	/* First try MSI/X */
6013 	if ((msgs = pci_msix_count(dev)) == 0) { /* system has msix disabled */
6014 		device_printf(dev, "System has MSIX disabled \n");
6015 		bus_release_resource(dev, SYS_RES_MEMORY,
6016 		    bar, ctx->ifc_msix_mem);
6017 		ctx->ifc_msix_mem = NULL;
6018 		goto msi;
6019 	}
6020 #if IFLIB_DEBUG
6021 	/* use only 1 qset in debug mode */
6022 	queuemsgs = min(msgs - admincnt, 1);
6023 #else
6024 	queuemsgs = msgs - admincnt;
6025 #endif
6026 #ifdef RSS
6027 	queues = imin(queuemsgs, rss_getnumbuckets());
6028 #else
6029 	queues = queuemsgs;
6030 #endif
6031 	queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues);
6032 	device_printf(dev, "pxm cpus: %d queue msgs: %d admincnt: %d\n",
6033 				  CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt);
6034 #ifdef  RSS
6035 	/* If we're doing RSS, clamp at the number of RSS buckets */
6036 	if (queues > rss_getnumbuckets())
6037 		queues = rss_getnumbuckets();
6038 #endif
6039 	if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt)
6040 		rx_queues = iflib_num_rx_queues;
6041 	else
6042 		rx_queues = queues;
6043 
6044 	if (rx_queues > scctx->isc_nrxqsets)
6045 		rx_queues = scctx->isc_nrxqsets;
6046 
6047 	/*
6048 	 * We want this to be all logical CPUs by default
6049 	 */
6050 	if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues)
6051 		tx_queues = iflib_num_tx_queues;
6052 	else
6053 		tx_queues = mp_ncpus;
6054 
6055 	if (tx_queues > scctx->isc_ntxqsets)
6056 		tx_queues = scctx->isc_ntxqsets;
6057 
6058 	if (ctx->ifc_sysctl_qs_eq_override == 0) {
6059 #ifdef INVARIANTS
6060 		if (tx_queues != rx_queues)
6061 			device_printf(dev, "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n",
6062 				      min(rx_queues, tx_queues), min(rx_queues, tx_queues));
6063 #endif
6064 		tx_queues = min(rx_queues, tx_queues);
6065 		rx_queues = min(rx_queues, tx_queues);
6066 	}
6067 
6068 	device_printf(dev, "using %d rx queues %d tx queues \n", rx_queues, tx_queues);
6069 
6070 	vectors = rx_queues + admincnt;
6071 	if ((err = pci_alloc_msix(dev, &vectors)) == 0) {
6072 		device_printf(dev,
6073 					  "Using MSIX interrupts with %d vectors\n", vectors);
6074 		scctx->isc_vectors = vectors;
6075 		scctx->isc_nrxqsets = rx_queues;
6076 		scctx->isc_ntxqsets = tx_queues;
6077 		scctx->isc_intr = IFLIB_INTR_MSIX;
6078 
6079 		return (vectors);
6080 	} else {
6081 		device_printf(dev, "failed to allocate %d msix vectors, err: %d - using MSI\n", vectors, err);
6082 		bus_release_resource(dev, SYS_RES_MEMORY, bar,
6083 		    ctx->ifc_msix_mem);
6084 		ctx->ifc_msix_mem = NULL;
6085 	}
6086 msi:
6087 	vectors = pci_msi_count(dev);
6088 	scctx->isc_nrxqsets = 1;
6089 	scctx->isc_ntxqsets = 1;
6090 	scctx->isc_vectors = vectors;
6091 	if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) {
6092 		device_printf(dev,"Using an MSI interrupt\n");
6093 		scctx->isc_intr = IFLIB_INTR_MSI;
6094 	} else {
6095 		scctx->isc_vectors = 1;
6096 		device_printf(dev,"Using a Legacy interrupt\n");
6097 		scctx->isc_intr = IFLIB_INTR_LEGACY;
6098 	}
6099 
6100 	return (vectors);
6101 }
6102 
6103 static const char *ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" };
6104 
6105 static int
6106 mp_ring_state_handler(SYSCTL_HANDLER_ARGS)
6107 {
6108 	int rc;
6109 	uint16_t *state = ((uint16_t *)oidp->oid_arg1);
6110 	struct sbuf *sb;
6111 	const char *ring_state = "UNKNOWN";
6112 
6113 	/* XXX needed ? */
6114 	rc = sysctl_wire_old_buffer(req, 0);
6115 	MPASS(rc == 0);
6116 	if (rc != 0)
6117 		return (rc);
6118 	sb = sbuf_new_for_sysctl(NULL, NULL, 80, req);
6119 	MPASS(sb != NULL);
6120 	if (sb == NULL)
6121 		return (ENOMEM);
6122 	if (state[3] <= 3)
6123 		ring_state = ring_states[state[3]];
6124 
6125 	sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s",
6126 		    state[0], state[1], state[2], ring_state);
6127 	rc = sbuf_finish(sb);
6128 	sbuf_delete(sb);
6129         return(rc);
6130 }
6131 
6132 enum iflib_ndesc_handler {
6133 	IFLIB_NTXD_HANDLER,
6134 	IFLIB_NRXD_HANDLER,
6135 };
6136 
6137 static int
6138 mp_ndesc_handler(SYSCTL_HANDLER_ARGS)
6139 {
6140 	if_ctx_t ctx = (void *)arg1;
6141 	enum iflib_ndesc_handler type = arg2;
6142 	char buf[256] = {0};
6143 	qidx_t *ndesc;
6144 	char *p, *next;
6145 	int nqs, rc, i;
6146 
6147 	MPASS(type == IFLIB_NTXD_HANDLER || type == IFLIB_NRXD_HANDLER);
6148 
6149 	nqs = 8;
6150 	switch(type) {
6151 	case IFLIB_NTXD_HANDLER:
6152 		ndesc = ctx->ifc_sysctl_ntxds;
6153 		if (ctx->ifc_sctx)
6154 			nqs = ctx->ifc_sctx->isc_ntxqs;
6155 		break;
6156 	case IFLIB_NRXD_HANDLER:
6157 		ndesc = ctx->ifc_sysctl_nrxds;
6158 		if (ctx->ifc_sctx)
6159 			nqs = ctx->ifc_sctx->isc_nrxqs;
6160 		break;
6161 	default:
6162 			panic("unhandled type");
6163 	}
6164 	if (nqs == 0)
6165 		nqs = 8;
6166 
6167 	for (i=0; i<8; i++) {
6168 		if (i >= nqs)
6169 			break;
6170 		if (i)
6171 			strcat(buf, ",");
6172 		sprintf(strchr(buf, 0), "%d", ndesc[i]);
6173 	}
6174 
6175 	rc = sysctl_handle_string(oidp, buf, sizeof(buf), req);
6176 	if (rc || req->newptr == NULL)
6177 		return rc;
6178 
6179 	for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p;
6180 	    i++, p = strsep(&next, " ,")) {
6181 		ndesc[i] = strtoul(p, NULL, 10);
6182 	}
6183 
6184 	return(rc);
6185 }
6186 
6187 #define NAME_BUFLEN 32
6188 static void
6189 iflib_add_device_sysctl_pre(if_ctx_t ctx)
6190 {
6191         device_t dev = iflib_get_dev(ctx);
6192 	struct sysctl_oid_list *child, *oid_list;
6193 	struct sysctl_ctx_list *ctx_list;
6194 	struct sysctl_oid *node;
6195 
6196 	ctx_list = device_get_sysctl_ctx(dev);
6197 	child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev));
6198 	ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, "iflib",
6199 						      CTLFLAG_RD, NULL, "IFLIB fields");
6200 	oid_list = SYSCTL_CHILDREN(node);
6201 
6202 	SYSCTL_ADD_STRING(ctx_list, oid_list, OID_AUTO, "driver_version",
6203 		       CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version, 0,
6204 		       "driver version");
6205 
6206 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs",
6207 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0,
6208 			"# of txqs to use, 0 => use default #");
6209 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs",
6210 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0,
6211 			"# of rxqs to use, 0 => use default #");
6212 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_qs_enable",
6213 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0,
6214                        "permit #txq != #rxq");
6215 	SYSCTL_ADD_INT(ctx_list, oid_list, OID_AUTO, "disable_msix",
6216                       CTLFLAG_RWTUN, &ctx->ifc_softc_ctx.isc_disable_msix, 0,
6217                       "disable MSIX (default 0)");
6218 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "rx_budget",
6219 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_rx_budget, 0,
6220                        "set the rx budget");
6221 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "tx_abdicate",
6222 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_tx_abdicate, 0,
6223 		       "cause tx to abdicate instead of running to completion");
6224 
6225 	/* XXX change for per-queue sizes */
6226 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_ntxds",
6227 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NTXD_HANDLER,
6228                        mp_ndesc_handler, "A",
6229                        "list of # of tx descriptors to use, 0 = use default #");
6230 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_nrxds",
6231 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NRXD_HANDLER,
6232                        mp_ndesc_handler, "A",
6233                        "list of # of rx descriptors to use, 0 = use default #");
6234 }
6235 
6236 static void
6237 iflib_add_device_sysctl_post(if_ctx_t ctx)
6238 {
6239 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6240 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6241         device_t dev = iflib_get_dev(ctx);
6242 	struct sysctl_oid_list *child;
6243 	struct sysctl_ctx_list *ctx_list;
6244 	iflib_fl_t fl;
6245 	iflib_txq_t txq;
6246 	iflib_rxq_t rxq;
6247 	int i, j;
6248 	char namebuf[NAME_BUFLEN];
6249 	char *qfmt;
6250 	struct sysctl_oid *queue_node, *fl_node, *node;
6251 	struct sysctl_oid_list *queue_list, *fl_list;
6252 	ctx_list = device_get_sysctl_ctx(dev);
6253 
6254 	node = ctx->ifc_sysctl_node;
6255 	child = SYSCTL_CHILDREN(node);
6256 
6257 	if (scctx->isc_ntxqsets > 100)
6258 		qfmt = "txq%03d";
6259 	else if (scctx->isc_ntxqsets > 10)
6260 		qfmt = "txq%02d";
6261 	else
6262 		qfmt = "txq%d";
6263 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
6264 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6265 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6266 					     CTLFLAG_RD, NULL, "Queue Name");
6267 		queue_list = SYSCTL_CHILDREN(queue_node);
6268 #if MEMORY_LOGGING
6269 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued",
6270 				CTLFLAG_RD,
6271 				&txq->ift_dequeued, "total mbufs freed");
6272 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued",
6273 				CTLFLAG_RD,
6274 				&txq->ift_enqueued, "total mbufs enqueued");
6275 #endif
6276 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag",
6277 				   CTLFLAG_RD,
6278 				   &txq->ift_mbuf_defrag, "# of times m_defrag was called");
6279 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "m_pullups",
6280 				   CTLFLAG_RD,
6281 				   &txq->ift_pullups, "# of times m_pullup was called");
6282 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag_failed",
6283 				   CTLFLAG_RD,
6284 				   &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed");
6285 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_desc_avail",
6286 				   CTLFLAG_RD,
6287 				   &txq->ift_no_desc_avail, "# of times no descriptors were available");
6288 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "tx_map_failed",
6289 				   CTLFLAG_RD,
6290 				   &txq->ift_map_failed, "# of times dma map failed");
6291 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txd_encap_efbig",
6292 				   CTLFLAG_RD,
6293 				   &txq->ift_txd_encap_efbig, "# of times txd_encap returned EFBIG");
6294 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_tx_dma_setup",
6295 				   CTLFLAG_RD,
6296 				   &txq->ift_no_tx_dma_setup, "# of times map failed for other than EFBIG");
6297 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx",
6298 				   CTLFLAG_RD,
6299 				   &txq->ift_pidx, 1, "Producer Index");
6300 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx",
6301 				   CTLFLAG_RD,
6302 				   &txq->ift_cidx, 1, "Consumer Index");
6303 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx_processed",
6304 				   CTLFLAG_RD,
6305 				   &txq->ift_cidx_processed, 1, "Consumer Index seen by credit update");
6306 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use",
6307 				   CTLFLAG_RD,
6308 				   &txq->ift_in_use, 1, "descriptors in use");
6309 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_processed",
6310 				   CTLFLAG_RD,
6311 				   &txq->ift_processed, "descriptors procesed for clean");
6312 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned",
6313 				   CTLFLAG_RD,
6314 				   &txq->ift_cleaned, "total cleaned");
6315 		SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state",
6316 				CTLTYPE_STRING | CTLFLAG_RD, __DEVOLATILE(uint64_t *, &txq->ift_br->state),
6317 				0, mp_ring_state_handler, "A", "soft ring state");
6318 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_enqueues",
6319 				       CTLFLAG_RD, &txq->ift_br->enqueues,
6320 				       "# of enqueues to the mp_ring for this queue");
6321 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_drops",
6322 				       CTLFLAG_RD, &txq->ift_br->drops,
6323 				       "# of drops in the mp_ring for this queue");
6324 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_starts",
6325 				       CTLFLAG_RD, &txq->ift_br->starts,
6326 				       "# of normal consumer starts in the mp_ring for this queue");
6327 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_stalls",
6328 				       CTLFLAG_RD, &txq->ift_br->stalls,
6329 					       "# of consumer stalls in the mp_ring for this queue");
6330 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_restarts",
6331 			       CTLFLAG_RD, &txq->ift_br->restarts,
6332 				       "# of consumer restarts in the mp_ring for this queue");
6333 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_abdications",
6334 				       CTLFLAG_RD, &txq->ift_br->abdications,
6335 				       "# of consumer abdications in the mp_ring for this queue");
6336 	}
6337 
6338 	if (scctx->isc_nrxqsets > 100)
6339 		qfmt = "rxq%03d";
6340 	else if (scctx->isc_nrxqsets > 10)
6341 		qfmt = "rxq%02d";
6342 	else
6343 		qfmt = "rxq%d";
6344 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
6345 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6346 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6347 					     CTLFLAG_RD, NULL, "Queue Name");
6348 		queue_list = SYSCTL_CHILDREN(queue_node);
6349 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
6350 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_pidx",
6351 				       CTLFLAG_RD,
6352 				       &rxq->ifr_cq_pidx, 1, "Producer Index");
6353 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_cidx",
6354 				       CTLFLAG_RD,
6355 				       &rxq->ifr_cq_cidx, 1, "Consumer Index");
6356 		}
6357 
6358 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
6359 			snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j);
6360 			fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list, OID_AUTO, namebuf,
6361 						     CTLFLAG_RD, NULL, "freelist Name");
6362 			fl_list = SYSCTL_CHILDREN(fl_node);
6363 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx",
6364 				       CTLFLAG_RD,
6365 				       &fl->ifl_pidx, 1, "Producer Index");
6366 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx",
6367 				       CTLFLAG_RD,
6368 				       &fl->ifl_cidx, 1, "Consumer Index");
6369 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits",
6370 				       CTLFLAG_RD,
6371 				       &fl->ifl_credits, 1, "credits available");
6372 #if MEMORY_LOGGING
6373 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_enqueued",
6374 					CTLFLAG_RD,
6375 					&fl->ifl_m_enqueued, "mbufs allocated");
6376 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_dequeued",
6377 					CTLFLAG_RD,
6378 					&fl->ifl_m_dequeued, "mbufs freed");
6379 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_enqueued",
6380 					CTLFLAG_RD,
6381 					&fl->ifl_cl_enqueued, "clusters allocated");
6382 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_dequeued",
6383 					CTLFLAG_RD,
6384 					&fl->ifl_cl_dequeued, "clusters freed");
6385 #endif
6386 
6387 		}
6388 	}
6389 
6390 }
6391 
6392 #ifndef __NO_STRICT_ALIGNMENT
6393 static struct mbuf *
6394 iflib_fixup_rx(struct mbuf *m)
6395 {
6396 	struct mbuf *n;
6397 
6398 	if (m->m_len <= (MCLBYTES - ETHER_HDR_LEN)) {
6399 		bcopy(m->m_data, m->m_data + ETHER_HDR_LEN, m->m_len);
6400 		m->m_data += ETHER_HDR_LEN;
6401 		n = m;
6402 	} else {
6403 		MGETHDR(n, M_NOWAIT, MT_DATA);
6404 		if (n == NULL) {
6405 			m_freem(m);
6406 			return (NULL);
6407 		}
6408 		bcopy(m->m_data, n->m_data, ETHER_HDR_LEN);
6409 		m->m_data += ETHER_HDR_LEN;
6410 		m->m_len -= ETHER_HDR_LEN;
6411 		n->m_len = ETHER_HDR_LEN;
6412 		M_MOVE_PKTHDR(n, m);
6413 		n->m_next = m;
6414 	}
6415 	return (n);
6416 }
6417 #endif
6418 
6419 #ifdef NETDUMP
6420 static void
6421 iflib_netdump_init(struct ifnet *ifp, int *nrxr, int *ncl, int *clsize)
6422 {
6423 	if_ctx_t ctx;
6424 
6425 	ctx = if_getsoftc(ifp);
6426 	CTX_LOCK(ctx);
6427 	*nrxr = NRXQSETS(ctx);
6428 	*ncl = ctx->ifc_rxqs[0].ifr_fl->ifl_size;
6429 	*clsize = ctx->ifc_rxqs[0].ifr_fl->ifl_buf_size;
6430 	CTX_UNLOCK(ctx);
6431 }
6432 
6433 static void
6434 iflib_netdump_event(struct ifnet *ifp, enum netdump_ev event)
6435 {
6436 	if_ctx_t ctx;
6437 	if_softc_ctx_t scctx;
6438 	iflib_fl_t fl;
6439 	iflib_rxq_t rxq;
6440 	int i, j;
6441 
6442 	ctx = if_getsoftc(ifp);
6443 	scctx = &ctx->ifc_softc_ctx;
6444 
6445 	switch (event) {
6446 	case NETDUMP_START:
6447 		for (i = 0; i < scctx->isc_nrxqsets; i++) {
6448 			rxq = &ctx->ifc_rxqs[i];
6449 			for (j = 0; j < rxq->ifr_nfl; j++) {
6450 				fl = rxq->ifr_fl;
6451 				fl->ifl_zone = m_getzone(fl->ifl_buf_size);
6452 			}
6453 		}
6454 		iflib_no_tx_batch = 1;
6455 		break;
6456 	default:
6457 		break;
6458 	}
6459 }
6460 
6461 static int
6462 iflib_netdump_transmit(struct ifnet *ifp, struct mbuf *m)
6463 {
6464 	if_ctx_t ctx;
6465 	iflib_txq_t txq;
6466 	int error;
6467 
6468 	ctx = if_getsoftc(ifp);
6469 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
6470 	    IFF_DRV_RUNNING)
6471 		return (EBUSY);
6472 
6473 	txq = &ctx->ifc_txqs[0];
6474 	error = iflib_encap(txq, &m);
6475 	if (error == 0)
6476 		(void)iflib_txd_db_check(ctx, txq, true, txq->ift_in_use);
6477 	return (error);
6478 }
6479 
6480 static int
6481 iflib_netdump_poll(struct ifnet *ifp, int count)
6482 {
6483 	if_ctx_t ctx;
6484 	if_softc_ctx_t scctx;
6485 	iflib_txq_t txq;
6486 	int i;
6487 
6488 	ctx = if_getsoftc(ifp);
6489 	scctx = &ctx->ifc_softc_ctx;
6490 
6491 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
6492 	    IFF_DRV_RUNNING)
6493 		return (EBUSY);
6494 
6495 	txq = &ctx->ifc_txqs[0];
6496 	(void)iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
6497 
6498 	for (i = 0; i < scctx->isc_nrxqsets; i++)
6499 		(void)iflib_rxeof(&ctx->ifc_rxqs[i], 16 /* XXX */);
6500 	return (0);
6501 }
6502 #endif /* NETDUMP */
6503