xref: /freebsd/sys/net/iflib.c (revision d34048812292b714a0bf99967270d18fe3097c62)
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 #ifdef PCI_IOV
105 #include <dev/pci/pci_iov.h>
106 #endif
107 
108 #include <sys/bitstring.h>
109 /*
110  * enable accounting of every mbuf as it comes in to and goes out of
111  * iflib's software descriptor references
112  */
113 #define MEMORY_LOGGING 0
114 /*
115  * Enable mbuf vectors for compressing long mbuf chains
116  */
117 
118 /*
119  * NB:
120  * - Prefetching in tx cleaning should perhaps be a tunable. The distance ahead
121  *   we prefetch needs to be determined by the time spent in m_free vis a vis
122  *   the cost of a prefetch. This will of course vary based on the workload:
123  *      - NFLX's m_free path is dominated by vm-based M_EXT manipulation which
124  *        is quite expensive, thus suggesting very little prefetch.
125  *      - small packet forwarding which is just returning a single mbuf to
126  *        UMA will typically be very fast vis a vis the cost of a memory
127  *        access.
128  */
129 
130 
131 /*
132  * File organization:
133  *  - private structures
134  *  - iflib private utility functions
135  *  - ifnet functions
136  *  - vlan registry and other exported functions
137  *  - iflib public core functions
138  *
139  *
140  */
141 MALLOC_DEFINE(M_IFLIB, "iflib", "ifnet library");
142 
143 struct iflib_txq;
144 typedef struct iflib_txq *iflib_txq_t;
145 struct iflib_rxq;
146 typedef struct iflib_rxq *iflib_rxq_t;
147 struct iflib_fl;
148 typedef struct iflib_fl *iflib_fl_t;
149 
150 struct iflib_ctx;
151 
152 static void iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid);
153 static void iflib_timer(void *arg);
154 
155 typedef struct iflib_filter_info {
156 	driver_filter_t *ifi_filter;
157 	void *ifi_filter_arg;
158 	struct grouptask *ifi_task;
159 	void *ifi_ctx;
160 } *iflib_filter_info_t;
161 
162 struct iflib_ctx {
163 	KOBJ_FIELDS;
164 	/*
165 	 * Pointer to hardware driver's softc
166 	 */
167 	void *ifc_softc;
168 	device_t ifc_dev;
169 	if_t ifc_ifp;
170 
171 	cpuset_t ifc_cpus;
172 	if_shared_ctx_t ifc_sctx;
173 	struct if_softc_ctx ifc_softc_ctx;
174 
175 	struct sx ifc_ctx_sx;
176 	struct mtx ifc_state_mtx;
177 
178 	iflib_txq_t ifc_txqs;
179 	iflib_rxq_t ifc_rxqs;
180 	uint32_t ifc_if_flags;
181 	uint32_t ifc_flags;
182 	uint32_t ifc_max_fl_buf_size;
183 
184 	int ifc_link_state;
185 	int ifc_link_irq;
186 	int ifc_watchdog_events;
187 	struct cdev *ifc_led_dev;
188 	struct resource *ifc_msix_mem;
189 
190 	struct if_irq ifc_legacy_irq;
191 	struct grouptask ifc_admin_task;
192 	struct grouptask ifc_vflr_task;
193 	struct iflib_filter_info ifc_filter_info;
194 	struct ifmedia	ifc_media;
195 
196 	struct sysctl_oid *ifc_sysctl_node;
197 	uint16_t ifc_sysctl_ntxqs;
198 	uint16_t ifc_sysctl_nrxqs;
199 	uint16_t ifc_sysctl_qs_eq_override;
200 	uint16_t ifc_sysctl_rx_budget;
201 	uint16_t ifc_sysctl_tx_abdicate;
202 
203 	qidx_t ifc_sysctl_ntxds[8];
204 	qidx_t ifc_sysctl_nrxds[8];
205 	struct if_txrx ifc_txrx;
206 #define isc_txd_encap  ifc_txrx.ift_txd_encap
207 #define isc_txd_flush  ifc_txrx.ift_txd_flush
208 #define isc_txd_credits_update  ifc_txrx.ift_txd_credits_update
209 #define isc_rxd_available ifc_txrx.ift_rxd_available
210 #define isc_rxd_pkt_get ifc_txrx.ift_rxd_pkt_get
211 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
212 #define isc_rxd_flush ifc_txrx.ift_rxd_flush
213 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
214 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
215 #define isc_legacy_intr ifc_txrx.ift_legacy_intr
216 	eventhandler_tag ifc_vlan_attach_event;
217 	eventhandler_tag ifc_vlan_detach_event;
218 	uint8_t ifc_mac[ETHER_ADDR_LEN];
219 	char ifc_mtx_name[16];
220 };
221 
222 
223 void *
224 iflib_get_softc(if_ctx_t ctx)
225 {
226 
227 	return (ctx->ifc_softc);
228 }
229 
230 device_t
231 iflib_get_dev(if_ctx_t ctx)
232 {
233 
234 	return (ctx->ifc_dev);
235 }
236 
237 if_t
238 iflib_get_ifp(if_ctx_t ctx)
239 {
240 
241 	return (ctx->ifc_ifp);
242 }
243 
244 struct ifmedia *
245 iflib_get_media(if_ctx_t ctx)
246 {
247 
248 	return (&ctx->ifc_media);
249 }
250 
251 uint32_t
252 iflib_get_flags(if_ctx_t ctx)
253 {
254 	return (ctx->ifc_flags);
255 }
256 
257 void
258 iflib_set_mac(if_ctx_t ctx, uint8_t mac[ETHER_ADDR_LEN])
259 {
260 
261 	bcopy(mac, ctx->ifc_mac, ETHER_ADDR_LEN);
262 }
263 
264 if_softc_ctx_t
265 iflib_get_softc_ctx(if_ctx_t ctx)
266 {
267 
268 	return (&ctx->ifc_softc_ctx);
269 }
270 
271 if_shared_ctx_t
272 iflib_get_sctx(if_ctx_t ctx)
273 {
274 
275 	return (ctx->ifc_sctx);
276 }
277 
278 #define IP_ALIGNED(m) ((((uintptr_t)(m)->m_data) & 0x3) == 0x2)
279 #define CACHE_PTR_INCREMENT (CACHE_LINE_SIZE/sizeof(void*))
280 #define CACHE_PTR_NEXT(ptr) ((void *)(((uintptr_t)(ptr)+CACHE_LINE_SIZE-1) & (CACHE_LINE_SIZE-1)))
281 
282 #define LINK_ACTIVE(ctx) ((ctx)->ifc_link_state == LINK_STATE_UP)
283 #define CTX_IS_VF(ctx) ((ctx)->ifc_sctx->isc_flags & IFLIB_IS_VF)
284 
285 #define RX_SW_DESC_MAP_CREATED	(1 << 0)
286 #define TX_SW_DESC_MAP_CREATED	(1 << 1)
287 #define RX_SW_DESC_INUSE        (1 << 3)
288 #define TX_SW_DESC_MAPPED       (1 << 4)
289 
290 #define	M_TOOBIG		M_PROTO1
291 
292 typedef struct iflib_sw_rx_desc_array {
293 	bus_dmamap_t	*ifsd_map;         /* bus_dma maps for packet */
294 	struct mbuf	**ifsd_m;           /* pkthdr mbufs */
295 	caddr_t		*ifsd_cl;          /* direct cluster pointer for rx */
296 	uint8_t		*ifsd_flags;
297 } iflib_rxsd_array_t;
298 
299 typedef struct iflib_sw_tx_desc_array {
300 	bus_dmamap_t    *ifsd_map;         /* bus_dma maps for packet */
301 	struct mbuf    **ifsd_m;           /* pkthdr mbufs */
302 	uint8_t		*ifsd_flags;
303 } if_txsd_vec_t;
304 
305 
306 /* magic number that should be high enough for any hardware */
307 #define IFLIB_MAX_TX_SEGS		128
308 /* bnxt supports 64 with hardware LRO enabled */
309 #define IFLIB_MAX_RX_SEGS		64
310 #define IFLIB_RX_COPY_THRESH		128
311 #define IFLIB_MAX_RX_REFRESH		32
312 /* The minimum descriptors per second before we start coalescing */
313 #define IFLIB_MIN_DESC_SEC		16384
314 #define IFLIB_DEFAULT_TX_UPDATE_FREQ	16
315 #define IFLIB_QUEUE_IDLE		0
316 #define IFLIB_QUEUE_HUNG		1
317 #define IFLIB_QUEUE_WORKING		2
318 /* maximum number of txqs that can share an rx interrupt */
319 #define IFLIB_MAX_TX_SHARED_INTR	4
320 
321 /* this should really scale with ring size - this is a fairly arbitrary value */
322 #define TX_BATCH_SIZE			32
323 
324 #define IFLIB_RESTART_BUDGET		8
325 
326 
327 #define CSUM_OFFLOAD		(CSUM_IP_TSO|CSUM_IP6_TSO|CSUM_IP| \
328 				 CSUM_IP_UDP|CSUM_IP_TCP|CSUM_IP_SCTP| \
329 				 CSUM_IP6_UDP|CSUM_IP6_TCP|CSUM_IP6_SCTP)
330 struct iflib_txq {
331 	qidx_t		ift_in_use;
332 	qidx_t		ift_cidx;
333 	qidx_t		ift_cidx_processed;
334 	qidx_t		ift_pidx;
335 	uint8_t		ift_gen;
336 	uint8_t		ift_br_offset;
337 	uint16_t	ift_npending;
338 	uint16_t	ift_db_pending;
339 	uint16_t	ift_rs_pending;
340 	/* implicit pad */
341 	uint8_t		ift_txd_size[8];
342 	uint64_t	ift_processed;
343 	uint64_t	ift_cleaned;
344 	uint64_t	ift_cleaned_prev;
345 #if MEMORY_LOGGING
346 	uint64_t	ift_enqueued;
347 	uint64_t	ift_dequeued;
348 #endif
349 	uint64_t	ift_no_tx_dma_setup;
350 	uint64_t	ift_no_desc_avail;
351 	uint64_t	ift_mbuf_defrag_failed;
352 	uint64_t	ift_mbuf_defrag;
353 	uint64_t	ift_map_failed;
354 	uint64_t	ift_txd_encap_efbig;
355 	uint64_t	ift_pullups;
356 	uint64_t	ift_last_timer_tick;
357 
358 	struct mtx	ift_mtx;
359 	struct mtx	ift_db_mtx;
360 
361 	/* constant values */
362 	if_ctx_t	ift_ctx;
363 	struct ifmp_ring        *ift_br;
364 	struct grouptask	ift_task;
365 	qidx_t		ift_size;
366 	uint16_t	ift_id;
367 	struct callout	ift_timer;
368 
369 	if_txsd_vec_t	ift_sds;
370 	uint8_t		ift_qstatus;
371 	uint8_t		ift_closed;
372 	uint8_t		ift_update_freq;
373 	struct iflib_filter_info ift_filter_info;
374 	bus_dma_tag_t		ift_desc_tag;
375 	bus_dma_tag_t		ift_tso_desc_tag;
376 	iflib_dma_info_t	ift_ifdi;
377 #define MTX_NAME_LEN 16
378 	char                    ift_mtx_name[MTX_NAME_LEN];
379 	char                    ift_db_mtx_name[MTX_NAME_LEN];
380 	bus_dma_segment_t	ift_segs[IFLIB_MAX_TX_SEGS]  __aligned(CACHE_LINE_SIZE);
381 #ifdef IFLIB_DIAGNOSTICS
382 	uint64_t ift_cpu_exec_count[256];
383 #endif
384 } __aligned(CACHE_LINE_SIZE);
385 
386 struct iflib_fl {
387 	qidx_t		ifl_cidx;
388 	qidx_t		ifl_pidx;
389 	qidx_t		ifl_credits;
390 	uint8_t		ifl_gen;
391 	uint8_t		ifl_rxd_size;
392 #if MEMORY_LOGGING
393 	uint64_t	ifl_m_enqueued;
394 	uint64_t	ifl_m_dequeued;
395 	uint64_t	ifl_cl_enqueued;
396 	uint64_t	ifl_cl_dequeued;
397 #endif
398 	/* implicit pad */
399 
400 	bitstr_t 	*ifl_rx_bitmap;
401 	qidx_t		ifl_fragidx;
402 	/* constant */
403 	qidx_t		ifl_size;
404 	uint16_t	ifl_buf_size;
405 	uint16_t	ifl_cltype;
406 	uma_zone_t	ifl_zone;
407 	iflib_rxsd_array_t	ifl_sds;
408 	iflib_rxq_t	ifl_rxq;
409 	uint8_t		ifl_id;
410 	bus_dma_tag_t           ifl_desc_tag;
411 	iflib_dma_info_t	ifl_ifdi;
412 	uint64_t	ifl_bus_addrs[IFLIB_MAX_RX_REFRESH] __aligned(CACHE_LINE_SIZE);
413 	caddr_t		ifl_vm_addrs[IFLIB_MAX_RX_REFRESH];
414 	qidx_t	ifl_rxd_idxs[IFLIB_MAX_RX_REFRESH];
415 }  __aligned(CACHE_LINE_SIZE);
416 
417 static inline qidx_t
418 get_inuse(int size, qidx_t cidx, qidx_t pidx, uint8_t gen)
419 {
420 	qidx_t used;
421 
422 	if (pidx > cidx)
423 		used = pidx - cidx;
424 	else if (pidx < cidx)
425 		used = size - cidx + pidx;
426 	else if (gen == 0 && pidx == cidx)
427 		used = 0;
428 	else if (gen == 1 && pidx == cidx)
429 		used = size;
430 	else
431 		panic("bad state");
432 
433 	return (used);
434 }
435 
436 #define TXQ_AVAIL(txq) (txq->ift_size - get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx, txq->ift_gen))
437 
438 #define IDXDIFF(head, tail, wrap) \
439 	((head) >= (tail) ? (head) - (tail) : (wrap) - (tail) + (head))
440 
441 struct iflib_rxq {
442 	/* If there is a separate completion queue -
443 	 * these are the cq cidx and pidx. Otherwise
444 	 * these are unused.
445 	 */
446 	qidx_t		ifr_size;
447 	qidx_t		ifr_cq_cidx;
448 	qidx_t		ifr_cq_pidx;
449 	uint8_t		ifr_cq_gen;
450 	uint8_t		ifr_fl_offset;
451 
452 	if_ctx_t	ifr_ctx;
453 	iflib_fl_t	ifr_fl;
454 	uint64_t	ifr_rx_irq;
455 	uint16_t	ifr_id;
456 	uint8_t		ifr_lro_enabled;
457 	uint8_t		ifr_nfl;
458 	uint8_t		ifr_ntxqirq;
459 	uint8_t		ifr_txqid[IFLIB_MAX_TX_SHARED_INTR];
460 	struct lro_ctrl			ifr_lc;
461 	struct grouptask        ifr_task;
462 	struct iflib_filter_info ifr_filter_info;
463 	iflib_dma_info_t		ifr_ifdi;
464 
465 	/* dynamically allocate if any drivers need a value substantially larger than this */
466 	struct if_rxd_frag	ifr_frags[IFLIB_MAX_RX_SEGS] __aligned(CACHE_LINE_SIZE);
467 #ifdef IFLIB_DIAGNOSTICS
468 	uint64_t ifr_cpu_exec_count[256];
469 #endif
470 }  __aligned(CACHE_LINE_SIZE);
471 
472 typedef struct if_rxsd {
473 	caddr_t *ifsd_cl;
474 	struct mbuf **ifsd_m;
475 	iflib_fl_t ifsd_fl;
476 	qidx_t ifsd_cidx;
477 } *if_rxsd_t;
478 
479 /* multiple of word size */
480 #ifdef __LP64__
481 #define PKT_INFO_SIZE	6
482 #define RXD_INFO_SIZE	5
483 #define PKT_TYPE uint64_t
484 #else
485 #define PKT_INFO_SIZE	11
486 #define RXD_INFO_SIZE	8
487 #define PKT_TYPE uint32_t
488 #endif
489 #define PKT_LOOP_BOUND  ((PKT_INFO_SIZE/3)*3)
490 #define RXD_LOOP_BOUND  ((RXD_INFO_SIZE/4)*4)
491 
492 typedef struct if_pkt_info_pad {
493 	PKT_TYPE pkt_val[PKT_INFO_SIZE];
494 } *if_pkt_info_pad_t;
495 typedef struct if_rxd_info_pad {
496 	PKT_TYPE rxd_val[RXD_INFO_SIZE];
497 } *if_rxd_info_pad_t;
498 
499 CTASSERT(sizeof(struct if_pkt_info_pad) == sizeof(struct if_pkt_info));
500 CTASSERT(sizeof(struct if_rxd_info_pad) == sizeof(struct if_rxd_info));
501 
502 
503 static inline void
504 pkt_info_zero(if_pkt_info_t pi)
505 {
506 	if_pkt_info_pad_t pi_pad;
507 
508 	pi_pad = (if_pkt_info_pad_t)pi;
509 	pi_pad->pkt_val[0] = 0; pi_pad->pkt_val[1] = 0; pi_pad->pkt_val[2] = 0;
510 	pi_pad->pkt_val[3] = 0; pi_pad->pkt_val[4] = 0; pi_pad->pkt_val[5] = 0;
511 #ifndef __LP64__
512 	pi_pad->pkt_val[6] = 0; pi_pad->pkt_val[7] = 0; pi_pad->pkt_val[8] = 0;
513 	pi_pad->pkt_val[9] = 0; pi_pad->pkt_val[10] = 0;
514 #endif
515 }
516 
517 static device_method_t iflib_pseudo_methods[] = {
518 	DEVMETHOD(device_attach, noop_attach),
519 	DEVMETHOD(device_detach, iflib_pseudo_detach),
520 	DEVMETHOD_END
521 };
522 
523 driver_t iflib_pseudodriver = {
524 	"iflib_pseudo", iflib_pseudo_methods, sizeof(struct iflib_ctx),
525 };
526 
527 static inline void
528 rxd_info_zero(if_rxd_info_t ri)
529 {
530 	if_rxd_info_pad_t ri_pad;
531 	int i;
532 
533 	ri_pad = (if_rxd_info_pad_t)ri;
534 	for (i = 0; i < RXD_LOOP_BOUND; i += 4) {
535 		ri_pad->rxd_val[i] = 0;
536 		ri_pad->rxd_val[i+1] = 0;
537 		ri_pad->rxd_val[i+2] = 0;
538 		ri_pad->rxd_val[i+3] = 0;
539 	}
540 #ifdef __LP64__
541 	ri_pad->rxd_val[RXD_INFO_SIZE-1] = 0;
542 #endif
543 }
544 
545 /*
546  * Only allow a single packet to take up most 1/nth of the tx ring
547  */
548 #define MAX_SINGLE_PACKET_FRACTION 12
549 #define IF_BAD_DMA (bus_addr_t)-1
550 
551 #define CTX_ACTIVE(ctx) ((if_getdrvflags((ctx)->ifc_ifp) & IFF_DRV_RUNNING))
552 
553 #define CTX_LOCK_INIT(_sc)  sx_init(&(_sc)->ifc_ctx_sx, "iflib ctx lock")
554 #define CTX_LOCK(ctx) sx_xlock(&(ctx)->ifc_ctx_sx)
555 #define CTX_UNLOCK(ctx) sx_xunlock(&(ctx)->ifc_ctx_sx)
556 #define CTX_LOCK_DESTROY(ctx) sx_destroy(&(ctx)->ifc_ctx_sx)
557 
558 
559 #define STATE_LOCK_INIT(_sc, _name)  mtx_init(&(_sc)->ifc_state_mtx, _name, "iflib state lock", MTX_DEF)
560 #define STATE_LOCK(ctx) mtx_lock(&(ctx)->ifc_state_mtx)
561 #define STATE_UNLOCK(ctx) mtx_unlock(&(ctx)->ifc_state_mtx)
562 #define STATE_LOCK_DESTROY(ctx) mtx_destroy(&(ctx)->ifc_state_mtx)
563 
564 
565 
566 #define CALLOUT_LOCK(txq)	mtx_lock(&txq->ift_mtx)
567 #define CALLOUT_UNLOCK(txq) 	mtx_unlock(&txq->ift_mtx)
568 
569 void
570 iflib_set_detach(if_ctx_t ctx)
571 {
572 	STATE_LOCK(ctx);
573 	ctx->ifc_flags |= IFC_IN_DETACH;
574 	STATE_UNLOCK(ctx);
575 }
576 
577 /* Our boot-time initialization hook */
578 static int	iflib_module_event_handler(module_t, int, void *);
579 
580 static moduledata_t iflib_moduledata = {
581 	"iflib",
582 	iflib_module_event_handler,
583 	NULL
584 };
585 
586 DECLARE_MODULE(iflib, iflib_moduledata, SI_SUB_INIT_IF, SI_ORDER_ANY);
587 MODULE_VERSION(iflib, 1);
588 
589 MODULE_DEPEND(iflib, pci, 1, 1, 1);
590 MODULE_DEPEND(iflib, ether, 1, 1, 1);
591 
592 TASKQGROUP_DEFINE(if_io_tqg, mp_ncpus, 1);
593 TASKQGROUP_DEFINE(if_config_tqg, 1, 1);
594 
595 #ifndef IFLIB_DEBUG_COUNTERS
596 #ifdef INVARIANTS
597 #define IFLIB_DEBUG_COUNTERS 1
598 #else
599 #define IFLIB_DEBUG_COUNTERS 0
600 #endif /* !INVARIANTS */
601 #endif
602 
603 static SYSCTL_NODE(_net, OID_AUTO, iflib, CTLFLAG_RD, 0,
604                    "iflib driver parameters");
605 
606 /*
607  * XXX need to ensure that this can't accidentally cause the head to be moved backwards
608  */
609 static int iflib_min_tx_latency = 0;
610 SYSCTL_INT(_net_iflib, OID_AUTO, min_tx_latency, CTLFLAG_RW,
611 		   &iflib_min_tx_latency, 0, "minimize transmit latency at the possible expense of throughput");
612 static int iflib_no_tx_batch = 0;
613 SYSCTL_INT(_net_iflib, OID_AUTO, no_tx_batch, CTLFLAG_RW,
614 		   &iflib_no_tx_batch, 0, "minimize transmit latency at the possible expense of throughput");
615 
616 
617 #if IFLIB_DEBUG_COUNTERS
618 
619 static int iflib_tx_seen;
620 static int iflib_tx_sent;
621 static int iflib_tx_encap;
622 static int iflib_rx_allocs;
623 static int iflib_fl_refills;
624 static int iflib_fl_refills_large;
625 static int iflib_tx_frees;
626 
627 SYSCTL_INT(_net_iflib, OID_AUTO, tx_seen, CTLFLAG_RD,
628 		   &iflib_tx_seen, 0, "# tx mbufs seen");
629 SYSCTL_INT(_net_iflib, OID_AUTO, tx_sent, CTLFLAG_RD,
630 		   &iflib_tx_sent, 0, "# tx mbufs sent");
631 SYSCTL_INT(_net_iflib, OID_AUTO, tx_encap, CTLFLAG_RD,
632 		   &iflib_tx_encap, 0, "# tx mbufs encapped");
633 SYSCTL_INT(_net_iflib, OID_AUTO, tx_frees, CTLFLAG_RD,
634 		   &iflib_tx_frees, 0, "# tx frees");
635 SYSCTL_INT(_net_iflib, OID_AUTO, rx_allocs, CTLFLAG_RD,
636 		   &iflib_rx_allocs, 0, "# rx allocations");
637 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills, CTLFLAG_RD,
638 		   &iflib_fl_refills, 0, "# refills");
639 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills_large, CTLFLAG_RD,
640 		   &iflib_fl_refills_large, 0, "# large refills");
641 
642 
643 static int iflib_txq_drain_flushing;
644 static int iflib_txq_drain_oactive;
645 static int iflib_txq_drain_notready;
646 
647 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_flushing, CTLFLAG_RD,
648 		   &iflib_txq_drain_flushing, 0, "# drain flushes");
649 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_oactive, CTLFLAG_RD,
650 		   &iflib_txq_drain_oactive, 0, "# drain oactives");
651 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_notready, CTLFLAG_RD,
652 		   &iflib_txq_drain_notready, 0, "# drain notready");
653 
654 
655 static int iflib_encap_load_mbuf_fail;
656 static int iflib_encap_pad_mbuf_fail;
657 static int iflib_encap_txq_avail_fail;
658 static int iflib_encap_txd_encap_fail;
659 
660 SYSCTL_INT(_net_iflib, OID_AUTO, encap_load_mbuf_fail, CTLFLAG_RD,
661 		   &iflib_encap_load_mbuf_fail, 0, "# busdma load failures");
662 SYSCTL_INT(_net_iflib, OID_AUTO, encap_pad_mbuf_fail, CTLFLAG_RD,
663 		   &iflib_encap_pad_mbuf_fail, 0, "# runt frame pad failures");
664 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txq_avail_fail, CTLFLAG_RD,
665 		   &iflib_encap_txq_avail_fail, 0, "# txq avail failures");
666 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txd_encap_fail, CTLFLAG_RD,
667 		   &iflib_encap_txd_encap_fail, 0, "# driver encap failures");
668 
669 static int iflib_task_fn_rxs;
670 static int iflib_rx_intr_enables;
671 static int iflib_fast_intrs;
672 static int iflib_rx_unavail;
673 static int iflib_rx_ctx_inactive;
674 static int iflib_rx_if_input;
675 static int iflib_rx_mbuf_null;
676 static int iflib_rxd_flush;
677 
678 static int iflib_verbose_debug;
679 
680 SYSCTL_INT(_net_iflib, OID_AUTO, task_fn_rx, CTLFLAG_RD,
681 		   &iflib_task_fn_rxs, 0, "# task_fn_rx calls");
682 SYSCTL_INT(_net_iflib, OID_AUTO, rx_intr_enables, CTLFLAG_RD,
683 		   &iflib_rx_intr_enables, 0, "# rx intr enables");
684 SYSCTL_INT(_net_iflib, OID_AUTO, fast_intrs, CTLFLAG_RD,
685 		   &iflib_fast_intrs, 0, "# fast_intr calls");
686 SYSCTL_INT(_net_iflib, OID_AUTO, rx_unavail, CTLFLAG_RD,
687 		   &iflib_rx_unavail, 0, "# times rxeof called with no available data");
688 SYSCTL_INT(_net_iflib, OID_AUTO, rx_ctx_inactive, CTLFLAG_RD,
689 		   &iflib_rx_ctx_inactive, 0, "# times rxeof called with inactive context");
690 SYSCTL_INT(_net_iflib, OID_AUTO, rx_if_input, CTLFLAG_RD,
691 		   &iflib_rx_if_input, 0, "# times rxeof called if_input");
692 SYSCTL_INT(_net_iflib, OID_AUTO, rx_mbuf_null, CTLFLAG_RD,
693 		   &iflib_rx_mbuf_null, 0, "# times rxeof got null mbuf");
694 SYSCTL_INT(_net_iflib, OID_AUTO, rxd_flush, CTLFLAG_RD,
695 	         &iflib_rxd_flush, 0, "# times rxd_flush called");
696 SYSCTL_INT(_net_iflib, OID_AUTO, verbose_debug, CTLFLAG_RW,
697 		   &iflib_verbose_debug, 0, "enable verbose debugging");
698 
699 #define DBG_COUNTER_INC(name) atomic_add_int(&(iflib_ ## name), 1)
700 static void
701 iflib_debug_reset(void)
702 {
703 	iflib_tx_seen = iflib_tx_sent = iflib_tx_encap = iflib_rx_allocs =
704 		iflib_fl_refills = iflib_fl_refills_large = iflib_tx_frees =
705 		iflib_txq_drain_flushing = iflib_txq_drain_oactive =
706 		iflib_txq_drain_notready =
707 		iflib_encap_load_mbuf_fail = iflib_encap_pad_mbuf_fail =
708 		iflib_encap_txq_avail_fail = iflib_encap_txd_encap_fail =
709 		iflib_task_fn_rxs = iflib_rx_intr_enables = iflib_fast_intrs =
710 		iflib_rx_unavail =
711 		iflib_rx_ctx_inactive = iflib_rx_if_input =
712 		iflib_rx_mbuf_null = iflib_rxd_flush = 0;
713 }
714 
715 #else
716 #define DBG_COUNTER_INC(name)
717 static void iflib_debug_reset(void) {}
718 #endif
719 
720 #define IFLIB_DEBUG 0
721 
722 static void iflib_tx_structures_free(if_ctx_t ctx);
723 static void iflib_rx_structures_free(if_ctx_t ctx);
724 static int iflib_queues_alloc(if_ctx_t ctx);
725 static int iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq);
726 static int iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget);
727 static int iflib_qset_structures_setup(if_ctx_t ctx);
728 static int iflib_msix_init(if_ctx_t ctx);
729 static int iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filterarg, int *rid, const char *str);
730 static void iflib_txq_check_drain(iflib_txq_t txq, int budget);
731 static uint32_t iflib_txq_can_drain(struct ifmp_ring *);
732 #ifdef ALTQ
733 static void iflib_altq_if_start(if_t ifp);
734 static int iflib_altq_if_transmit(if_t ifp, struct mbuf *m);
735 #endif
736 static int iflib_register(if_ctx_t);
737 static void iflib_init_locked(if_ctx_t ctx);
738 static void iflib_add_device_sysctl_pre(if_ctx_t ctx);
739 static void iflib_add_device_sysctl_post(if_ctx_t ctx);
740 static void iflib_ifmp_purge(iflib_txq_t txq);
741 static void _iflib_pre_assert(if_softc_ctx_t scctx);
742 static void iflib_if_init_locked(if_ctx_t ctx);
743 static void iflib_free_intr_mem(if_ctx_t ctx);
744 #ifndef __NO_STRICT_ALIGNMENT
745 static struct mbuf * iflib_fixup_rx(struct mbuf *m);
746 #endif
747 
748 NETDUMP_DEFINE(iflib);
749 
750 #ifdef DEV_NETMAP
751 #include <sys/selinfo.h>
752 #include <net/netmap.h>
753 #include <dev/netmap/netmap_kern.h>
754 
755 MODULE_DEPEND(iflib, netmap, 1, 1, 1);
756 
757 static int netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, uint32_t nm_i, bool init);
758 
759 /*
760  * device-specific sysctl variables:
761  *
762  * iflib_crcstrip: 0: keep CRC in rx frames (default), 1: strip it.
763  *	During regular operations the CRC is stripped, but on some
764  *	hardware reception of frames not multiple of 64 is slower,
765  *	so using crcstrip=0 helps in benchmarks.
766  *
767  * iflib_rx_miss, iflib_rx_miss_bufs:
768  *	count packets that might be missed due to lost interrupts.
769  */
770 SYSCTL_DECL(_dev_netmap);
771 /*
772  * The xl driver by default strips CRCs and we do not override it.
773  */
774 
775 int iflib_crcstrip = 1;
776 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_crcstrip,
777     CTLFLAG_RW, &iflib_crcstrip, 1, "strip CRC on rx frames");
778 
779 int iflib_rx_miss, iflib_rx_miss_bufs;
780 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss,
781     CTLFLAG_RW, &iflib_rx_miss, 0, "potentially missed rx intr");
782 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss_bufs,
783     CTLFLAG_RW, &iflib_rx_miss_bufs, 0, "potentially missed rx intr bufs");
784 
785 /*
786  * Register/unregister. We are already under netmap lock.
787  * Only called on the first register or the last unregister.
788  */
789 static int
790 iflib_netmap_register(struct netmap_adapter *na, int onoff)
791 {
792 	struct ifnet *ifp = na->ifp;
793 	if_ctx_t ctx = ifp->if_softc;
794 	int status;
795 
796 	CTX_LOCK(ctx);
797 	IFDI_INTR_DISABLE(ctx);
798 
799 	/* Tell the stack that the interface is no longer active */
800 	ifp->if_drv_flags &= ~(IFF_DRV_RUNNING | IFF_DRV_OACTIVE);
801 
802 	if (!CTX_IS_VF(ctx))
803 		IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip);
804 
805 	/* enable or disable flags and callbacks in na and ifp */
806 	if (onoff) {
807 		nm_set_native_flags(na);
808 	} else {
809 		nm_clear_native_flags(na);
810 	}
811 	iflib_stop(ctx);
812 	iflib_init_locked(ctx);
813 	IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip); // XXX why twice ?
814 	status = ifp->if_drv_flags & IFF_DRV_RUNNING ? 0 : 1;
815 	if (status)
816 		nm_clear_native_flags(na);
817 	CTX_UNLOCK(ctx);
818 	return (status);
819 }
820 
821 static int
822 netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, uint32_t nm_i, bool init)
823 {
824 	struct netmap_adapter *na = kring->na;
825 	u_int const lim = kring->nkr_num_slots - 1;
826 	u_int head = kring->rhead;
827 	struct netmap_ring *ring = kring->ring;
828 	bus_dmamap_t *map;
829 	struct if_rxd_update iru;
830 	if_ctx_t ctx = rxq->ifr_ctx;
831 	iflib_fl_t fl = &rxq->ifr_fl[0];
832 	uint32_t refill_pidx, nic_i;
833 #if IFLIB_DEBUG_COUNTERS
834 	int rf_count = 0;
835 #endif
836 
837 	if (nm_i == head && __predict_true(!init))
838 		return 0;
839 	iru_init(&iru, rxq, 0 /* flid */);
840 	map = fl->ifl_sds.ifsd_map;
841 	refill_pidx = netmap_idx_k2n(kring, nm_i);
842 	/*
843 	 * IMPORTANT: we must leave one free slot in the ring,
844 	 * so move head back by one unit
845 	 */
846 	head = nm_prev(head, lim);
847 	nic_i = UINT_MAX;
848 	DBG_COUNTER_INC(fl_refills);
849 	while (nm_i != head) {
850 #if IFLIB_DEBUG_COUNTERS
851 		if (++rf_count == 9)
852 			DBG_COUNTER_INC(fl_refills_large);
853 #endif
854 		for (int tmp_pidx = 0; tmp_pidx < IFLIB_MAX_RX_REFRESH && nm_i != head; tmp_pidx++) {
855 			struct netmap_slot *slot = &ring->slot[nm_i];
856 			void *addr = PNMB(na, slot, &fl->ifl_bus_addrs[tmp_pidx]);
857 			uint32_t nic_i_dma = refill_pidx;
858 			nic_i = netmap_idx_k2n(kring, nm_i);
859 
860 			MPASS(tmp_pidx < IFLIB_MAX_RX_REFRESH);
861 
862 			if (addr == NETMAP_BUF_BASE(na)) /* bad buf */
863 			        return netmap_ring_reinit(kring);
864 
865 			fl->ifl_vm_addrs[tmp_pidx] = addr;
866 			if (__predict_false(init) && map) {
867 				netmap_load_map(na, fl->ifl_ifdi->idi_tag, map[nic_i], addr);
868 			} else if (map && (slot->flags & NS_BUF_CHANGED)) {
869 				/* buffer has changed, reload map */
870 				netmap_reload_map(na, fl->ifl_ifdi->idi_tag, map[nic_i], addr);
871 			}
872 			slot->flags &= ~NS_BUF_CHANGED;
873 
874 			nm_i = nm_next(nm_i, lim);
875 			fl->ifl_rxd_idxs[tmp_pidx] = nic_i = nm_next(nic_i, lim);
876 			if (nm_i != head && tmp_pidx < IFLIB_MAX_RX_REFRESH-1)
877 				continue;
878 
879 			iru.iru_pidx = refill_pidx;
880 			iru.iru_count = tmp_pidx+1;
881 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
882 
883 			refill_pidx = nic_i;
884 			if (map == NULL)
885 				continue;
886 
887 			for (int n = 0; n < iru.iru_count; n++) {
888 				bus_dmamap_sync(fl->ifl_ifdi->idi_tag, map[nic_i_dma],
889 						BUS_DMASYNC_PREREAD);
890 				/* XXX - change this to not use the netmap func*/
891 				nic_i_dma = nm_next(nic_i_dma, lim);
892 			}
893 		}
894 	}
895 	kring->nr_hwcur = head;
896 
897 	if (map)
898 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
899 				BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
900 	if (__predict_true(nic_i != UINT_MAX)) {
901 		ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id, nic_i);
902 		DBG_COUNTER_INC(rxd_flush);
903 	}
904 	return (0);
905 }
906 
907 /*
908  * Reconcile kernel and user view of the transmit ring.
909  *
910  * All information is in the kring.
911  * Userspace wants to send packets up to the one before kring->rhead,
912  * kernel knows kring->nr_hwcur is the first unsent packet.
913  *
914  * Here we push packets out (as many as possible), and possibly
915  * reclaim buffers from previously completed transmission.
916  *
917  * The caller (netmap) guarantees that there is only one instance
918  * running at any time. Any interference with other driver
919  * methods should be handled by the individual drivers.
920  */
921 static int
922 iflib_netmap_txsync(struct netmap_kring *kring, int flags)
923 {
924 	struct netmap_adapter *na = kring->na;
925 	struct ifnet *ifp = na->ifp;
926 	struct netmap_ring *ring = kring->ring;
927 	u_int nm_i;	/* index into the netmap kring */
928 	u_int nic_i;	/* index into the NIC ring */
929 	u_int n;
930 	u_int const lim = kring->nkr_num_slots - 1;
931 	u_int const head = kring->rhead;
932 	struct if_pkt_info pi;
933 
934 	/*
935 	 * interrupts on every tx packet are expensive so request
936 	 * them every half ring, or where NS_REPORT is set
937 	 */
938 	u_int report_frequency = kring->nkr_num_slots >> 1;
939 	/* device-specific */
940 	if_ctx_t ctx = ifp->if_softc;
941 	iflib_txq_t txq = &ctx->ifc_txqs[kring->ring_id];
942 
943 	if (txq->ift_sds.ifsd_map)
944 		bus_dmamap_sync(txq->ift_desc_tag, txq->ift_ifdi->idi_map,
945 				BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
946 
947 
948 	/*
949 	 * First part: process new packets to send.
950 	 * nm_i is the current index in the netmap kring,
951 	 * nic_i is the corresponding index in the NIC ring.
952 	 *
953 	 * If we have packets to send (nm_i != head)
954 	 * iterate over the netmap ring, fetch length and update
955 	 * the corresponding slot in the NIC ring. Some drivers also
956 	 * need to update the buffer's physical address in the NIC slot
957 	 * even NS_BUF_CHANGED is not set (PNMB computes the addresses).
958 	 *
959 	 * The netmap_reload_map() calls is especially expensive,
960 	 * even when (as in this case) the tag is 0, so do only
961 	 * when the buffer has actually changed.
962 	 *
963 	 * If possible do not set the report/intr bit on all slots,
964 	 * but only a few times per ring or when NS_REPORT is set.
965 	 *
966 	 * Finally, on 10G and faster drivers, it might be useful
967 	 * to prefetch the next slot and txr entry.
968 	 */
969 
970 	nm_i = kring->nr_hwcur;
971 	if (nm_i != head) {	/* we have new packets to send */
972 		pkt_info_zero(&pi);
973 		pi.ipi_segs = txq->ift_segs;
974 		pi.ipi_qsidx = kring->ring_id;
975 		nic_i = netmap_idx_k2n(kring, nm_i);
976 
977 		__builtin_prefetch(&ring->slot[nm_i]);
978 		__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i]);
979 		if (txq->ift_sds.ifsd_map)
980 			__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i]);
981 
982 		for (n = 0; nm_i != head; n++) {
983 			struct netmap_slot *slot = &ring->slot[nm_i];
984 			u_int len = slot->len;
985 			uint64_t paddr;
986 			void *addr = PNMB(na, slot, &paddr);
987 			int flags = (slot->flags & NS_REPORT ||
988 				nic_i == 0 || nic_i == report_frequency) ?
989 				IPI_TX_INTR : 0;
990 
991 			/* device-specific */
992 			pi.ipi_len = len;
993 			pi.ipi_segs[0].ds_addr = paddr;
994 			pi.ipi_segs[0].ds_len = len;
995 			pi.ipi_nsegs = 1;
996 			pi.ipi_ndescs = 0;
997 			pi.ipi_pidx = nic_i;
998 			pi.ipi_flags = flags;
999 
1000 			/* Fill the slot in the NIC ring. */
1001 			ctx->isc_txd_encap(ctx->ifc_softc, &pi);
1002 			DBG_COUNTER_INC(tx_encap);
1003 
1004 			/* prefetch for next round */
1005 			__builtin_prefetch(&ring->slot[nm_i + 1]);
1006 			__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i + 1]);
1007 			if (txq->ift_sds.ifsd_map) {
1008 				__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i + 1]);
1009 
1010 				NM_CHECK_ADDR_LEN(na, addr, len);
1011 
1012 				if (slot->flags & NS_BUF_CHANGED) {
1013 					/* buffer has changed, reload map */
1014 					netmap_reload_map(na, txq->ift_desc_tag, txq->ift_sds.ifsd_map[nic_i], addr);
1015 				}
1016 				/* make sure changes to the buffer are synced */
1017 				bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_sds.ifsd_map[nic_i],
1018 						BUS_DMASYNC_PREWRITE);
1019 			}
1020 			slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED);
1021 			nm_i = nm_next(nm_i, lim);
1022 			nic_i = nm_next(nic_i, lim);
1023 		}
1024 		kring->nr_hwcur = nm_i;
1025 
1026 		/* synchronize the NIC ring */
1027 		if (txq->ift_sds.ifsd_map)
1028 			bus_dmamap_sync(txq->ift_desc_tag, txq->ift_ifdi->idi_map,
1029 						BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1030 
1031 		/* (re)start the tx unit up to slot nic_i (excluded) */
1032 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, nic_i);
1033 	}
1034 
1035 	/*
1036 	 * Second part: reclaim buffers for completed transmissions.
1037 	 *
1038 	 * If there are unclaimed buffers, attempt to reclaim them.
1039 	 * If none are reclaimed, and TX IRQs are not in use, do an initial
1040 	 * minimal delay, then trigger the tx handler which will spin in the
1041 	 * group task queue.
1042 	 */
1043 	if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1044 		if (iflib_tx_credits_update(ctx, txq)) {
1045 			/* some tx completed, increment avail */
1046 			nic_i = txq->ift_cidx_processed;
1047 			kring->nr_hwtail = nm_prev(netmap_idx_n2k(kring, nic_i), lim);
1048 		}
1049 	}
1050 	if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ))
1051 		if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1052 			callout_reset_on(&txq->ift_timer, hz < 2000 ? 1 : hz / 1000,
1053 			    iflib_timer, txq, txq->ift_timer.c_cpu);
1054 	}
1055 	return (0);
1056 }
1057 
1058 /*
1059  * Reconcile kernel and user view of the receive ring.
1060  * Same as for the txsync, this routine must be efficient.
1061  * The caller guarantees a single invocations, but races against
1062  * the rest of the driver should be handled here.
1063  *
1064  * On call, kring->rhead is the first packet that userspace wants
1065  * to keep, and kring->rcur is the wakeup point.
1066  * The kernel has previously reported packets up to kring->rtail.
1067  *
1068  * If (flags & NAF_FORCE_READ) also check for incoming packets irrespective
1069  * of whether or not we received an interrupt.
1070  */
1071 static int
1072 iflib_netmap_rxsync(struct netmap_kring *kring, int flags)
1073 {
1074 	struct netmap_adapter *na = kring->na;
1075 	struct netmap_ring *ring = kring->ring;
1076 	uint32_t nm_i;	/* index into the netmap ring */
1077 	uint32_t nic_i;	/* index into the NIC ring */
1078 	u_int i, n;
1079 	u_int const lim = kring->nkr_num_slots - 1;
1080 	u_int const head = kring->rhead;
1081 	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
1082 	struct if_rxd_info ri;
1083 
1084 	struct ifnet *ifp = na->ifp;
1085 	if_ctx_t ctx = ifp->if_softc;
1086 	iflib_rxq_t rxq = &ctx->ifc_rxqs[kring->ring_id];
1087 	iflib_fl_t fl = rxq->ifr_fl;
1088 	if (head > lim)
1089 		return netmap_ring_reinit(kring);
1090 
1091 	/* XXX check sync modes */
1092 	for (i = 0, fl = rxq->ifr_fl; i < rxq->ifr_nfl; i++, fl++) {
1093 		if (fl->ifl_sds.ifsd_map == NULL)
1094 			continue;
1095 		bus_dmamap_sync(rxq->ifr_fl[i].ifl_desc_tag, fl->ifl_ifdi->idi_map,
1096 				BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1097 	}
1098 	/*
1099 	 * First part: import newly received packets.
1100 	 *
1101 	 * nm_i is the index of the next free slot in the netmap ring,
1102 	 * nic_i is the index of the next received packet in the NIC ring,
1103 	 * and they may differ in case if_init() has been called while
1104 	 * in netmap mode. For the receive ring we have
1105 	 *
1106 	 *	nic_i = rxr->next_check;
1107 	 *	nm_i = kring->nr_hwtail (previous)
1108 	 * and
1109 	 *	nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1110 	 *
1111 	 * rxr->next_check is set to 0 on a ring reinit
1112 	 */
1113 	if (netmap_no_pendintr || force_update) {
1114 		int crclen = iflib_crcstrip ? 0 : 4;
1115 		int error, avail;
1116 
1117 		for (i = 0; i < rxq->ifr_nfl; i++) {
1118 			fl = &rxq->ifr_fl[i];
1119 			nic_i = fl->ifl_cidx;
1120 			nm_i = netmap_idx_n2k(kring, nic_i);
1121 			avail = iflib_rxd_avail(ctx, rxq, nic_i, USHRT_MAX);
1122 			for (n = 0; avail > 0; n++, avail--) {
1123 				rxd_info_zero(&ri);
1124 				ri.iri_frags = rxq->ifr_frags;
1125 				ri.iri_qsidx = kring->ring_id;
1126 				ri.iri_ifp = ctx->ifc_ifp;
1127 				ri.iri_cidx = nic_i;
1128 
1129 				error = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
1130 				ring->slot[nm_i].len = error ? 0 : ri.iri_len - crclen;
1131 				ring->slot[nm_i].flags = 0;
1132 				if (fl->ifl_sds.ifsd_map)
1133 					bus_dmamap_sync(fl->ifl_ifdi->idi_tag,
1134 							fl->ifl_sds.ifsd_map[nic_i], BUS_DMASYNC_POSTREAD);
1135 				nm_i = nm_next(nm_i, lim);
1136 				nic_i = nm_next(nic_i, lim);
1137 			}
1138 			if (n) { /* update the state variables */
1139 				if (netmap_no_pendintr && !force_update) {
1140 					/* diagnostics */
1141 					iflib_rx_miss ++;
1142 					iflib_rx_miss_bufs += n;
1143 				}
1144 				fl->ifl_cidx = nic_i;
1145 				kring->nr_hwtail = nm_i;
1146 			}
1147 			kring->nr_kflags &= ~NKR_PENDINTR;
1148 		}
1149 	}
1150 	/*
1151 	 * Second part: skip past packets that userspace has released.
1152 	 * (kring->nr_hwcur to head excluded),
1153 	 * and make the buffers available for reception.
1154 	 * As usual nm_i is the index in the netmap ring,
1155 	 * nic_i is the index in the NIC ring, and
1156 	 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1157 	 */
1158 	/* XXX not sure how this will work with multiple free lists */
1159 	nm_i = kring->nr_hwcur;
1160 
1161 	return (netmap_fl_refill(rxq, kring, nm_i, false));
1162 }
1163 
1164 static void
1165 iflib_netmap_intr(struct netmap_adapter *na, int onoff)
1166 {
1167 	struct ifnet *ifp = na->ifp;
1168 	if_ctx_t ctx = ifp->if_softc;
1169 
1170 	CTX_LOCK(ctx);
1171 	if (onoff) {
1172 		IFDI_INTR_ENABLE(ctx);
1173 	} else {
1174 		IFDI_INTR_DISABLE(ctx);
1175 	}
1176 	CTX_UNLOCK(ctx);
1177 }
1178 
1179 
1180 static int
1181 iflib_netmap_attach(if_ctx_t ctx)
1182 {
1183 	struct netmap_adapter na;
1184 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1185 
1186 	bzero(&na, sizeof(na));
1187 
1188 	na.ifp = ctx->ifc_ifp;
1189 	na.na_flags = NAF_BDG_MAYSLEEP;
1190 	MPASS(ctx->ifc_softc_ctx.isc_ntxqsets);
1191 	MPASS(ctx->ifc_softc_ctx.isc_nrxqsets);
1192 
1193 	na.num_tx_desc = scctx->isc_ntxd[0];
1194 	na.num_rx_desc = scctx->isc_nrxd[0];
1195 	na.nm_txsync = iflib_netmap_txsync;
1196 	na.nm_rxsync = iflib_netmap_rxsync;
1197 	na.nm_register = iflib_netmap_register;
1198 	na.nm_intr = iflib_netmap_intr;
1199 	na.num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
1200 	na.num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
1201 	return (netmap_attach(&na));
1202 }
1203 
1204 static void
1205 iflib_netmap_txq_init(if_ctx_t ctx, iflib_txq_t txq)
1206 {
1207 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1208 	struct netmap_slot *slot;
1209 
1210 	slot = netmap_reset(na, NR_TX, txq->ift_id, 0);
1211 	if (slot == NULL)
1212 		return;
1213 	if (txq->ift_sds.ifsd_map == NULL)
1214 		return;
1215 
1216 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxd[0]; i++) {
1217 
1218 		/*
1219 		 * In netmap mode, set the map for the packet buffer.
1220 		 * NOTE: Some drivers (not this one) also need to set
1221 		 * the physical buffer address in the NIC ring.
1222 		 * netmap_idx_n2k() maps a nic index, i, into the corresponding
1223 		 * netmap slot index, si
1224 		 */
1225 		int si = netmap_idx_n2k(na->tx_rings[txq->ift_id], i);
1226 		netmap_load_map(na, txq->ift_desc_tag, txq->ift_sds.ifsd_map[i], NMB(na, slot + si));
1227 	}
1228 }
1229 
1230 static void
1231 iflib_netmap_rxq_init(if_ctx_t ctx, iflib_rxq_t rxq)
1232 {
1233 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1234 	struct netmap_kring *kring = na->rx_rings[rxq->ifr_id];
1235 	struct netmap_slot *slot;
1236 	uint32_t nm_i;
1237 
1238 	slot = netmap_reset(na, NR_RX, rxq->ifr_id, 0);
1239 	if (slot == NULL)
1240 		return;
1241 	nm_i = netmap_idx_n2k(kring, 0);
1242 	netmap_fl_refill(rxq, kring, nm_i, true);
1243 }
1244 
1245 static void
1246 iflib_netmap_timer_adjust(if_ctx_t ctx, uint16_t txqid, uint32_t *reset_on)
1247 {
1248 	struct netmap_kring *kring;
1249 
1250 	kring = NA(ctx->ifc_ifp)->tx_rings[txqid];
1251 
1252 	if (kring->nr_hwcur != nm_next(kring->nr_hwtail, kring->nkr_num_slots - 1)) {
1253 		if (ctx->isc_txd_credits_update(ctx->ifc_softc, txqid, false))
1254 			netmap_tx_irq(ctx->ifc_ifp, txqid);
1255 		if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ)) {
1256 			if (hz < 2000)
1257 				*reset_on = 1;
1258 			else
1259 				*reset_on = hz / 1000;
1260 		}
1261 	}
1262 }
1263 
1264 #define iflib_netmap_detach(ifp) netmap_detach(ifp)
1265 
1266 #else
1267 #define iflib_netmap_txq_init(ctx, txq)
1268 #define iflib_netmap_rxq_init(ctx, rxq)
1269 #define iflib_netmap_detach(ifp)
1270 
1271 #define iflib_netmap_attach(ctx) (0)
1272 #define netmap_rx_irq(ifp, qid, budget) (0)
1273 #define netmap_tx_irq(ifp, qid) do {} while (0)
1274 #define iflib_netmap_timer_adjust(ctx, txqid, reset_on)
1275 
1276 #endif
1277 
1278 #if defined(__i386__) || defined(__amd64__)
1279 static __inline void
1280 prefetch(void *x)
1281 {
1282 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1283 }
1284 static __inline void
1285 prefetch2cachelines(void *x)
1286 {
1287 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1288 #if (CACHE_LINE_SIZE < 128)
1289 	__asm volatile("prefetcht0 %0" :: "m" (*(((unsigned long *)x)+CACHE_LINE_SIZE/(sizeof(unsigned long)))));
1290 #endif
1291 }
1292 #else
1293 #define prefetch(x)
1294 #define prefetch2cachelines(x)
1295 #endif
1296 
1297 static void
1298 iflib_gen_mac(if_ctx_t ctx)
1299 {
1300 	struct thread *td;
1301 	MD5_CTX mdctx;
1302 	char uuid[HOSTUUIDLEN+1];
1303 	char buf[HOSTUUIDLEN+16];
1304 	uint8_t *mac;
1305 	unsigned char digest[16];
1306 
1307 	td = curthread;
1308 	mac = ctx->ifc_mac;
1309 	uuid[HOSTUUIDLEN] = 0;
1310 	bcopy(td->td_ucred->cr_prison->pr_hostuuid, uuid, HOSTUUIDLEN);
1311 	snprintf(buf, HOSTUUIDLEN+16, "%s-%s", uuid, device_get_nameunit(ctx->ifc_dev));
1312 	/*
1313 	 * Generate a pseudo-random, deterministic MAC
1314 	 * address based on the UUID and unit number.
1315 	 * The FreeBSD Foundation OUI of 58-9C-FC is used.
1316 	 */
1317 	MD5Init(&mdctx);
1318 	MD5Update(&mdctx, buf, strlen(buf));
1319 	MD5Final(digest, &mdctx);
1320 
1321 	mac[0] = 0x58;
1322 	mac[1] = 0x9C;
1323 	mac[2] = 0xFC;
1324 	mac[3] = digest[0];
1325 	mac[4] = digest[1];
1326 	mac[5] = digest[2];
1327 }
1328 
1329 static void
1330 iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid)
1331 {
1332 	iflib_fl_t fl;
1333 
1334 	fl = &rxq->ifr_fl[flid];
1335 	iru->iru_paddrs = fl->ifl_bus_addrs;
1336 	iru->iru_vaddrs = &fl->ifl_vm_addrs[0];
1337 	iru->iru_idxs = fl->ifl_rxd_idxs;
1338 	iru->iru_qsidx = rxq->ifr_id;
1339 	iru->iru_buf_size = fl->ifl_buf_size;
1340 	iru->iru_flidx = fl->ifl_id;
1341 }
1342 
1343 static void
1344 _iflib_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int err)
1345 {
1346 	if (err)
1347 		return;
1348 	*(bus_addr_t *) arg = segs[0].ds_addr;
1349 }
1350 
1351 int
1352 iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags)
1353 {
1354 	int err;
1355 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1356 	device_t dev = ctx->ifc_dev;
1357 
1358 	KASSERT(sctx->isc_q_align != 0, ("alignment value not initialized"));
1359 
1360 	err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1361 				sctx->isc_q_align, 0,	/* alignment, bounds */
1362 				BUS_SPACE_MAXADDR,	/* lowaddr */
1363 				BUS_SPACE_MAXADDR,	/* highaddr */
1364 				NULL, NULL,		/* filter, filterarg */
1365 				size,			/* maxsize */
1366 				1,			/* nsegments */
1367 				size,			/* maxsegsize */
1368 				BUS_DMA_ALLOCNOW,	/* flags */
1369 				NULL,			/* lockfunc */
1370 				NULL,			/* lockarg */
1371 				&dma->idi_tag);
1372 	if (err) {
1373 		device_printf(dev,
1374 		    "%s: bus_dma_tag_create failed: %d\n",
1375 		    __func__, err);
1376 		goto fail_0;
1377 	}
1378 
1379 	err = bus_dmamem_alloc(dma->idi_tag, (void**) &dma->idi_vaddr,
1380 	    BUS_DMA_NOWAIT | BUS_DMA_COHERENT | BUS_DMA_ZERO, &dma->idi_map);
1381 	if (err) {
1382 		device_printf(dev,
1383 		    "%s: bus_dmamem_alloc(%ju) failed: %d\n",
1384 		    __func__, (uintmax_t)size, err);
1385 		goto fail_1;
1386 	}
1387 
1388 	dma->idi_paddr = IF_BAD_DMA;
1389 	err = bus_dmamap_load(dma->idi_tag, dma->idi_map, dma->idi_vaddr,
1390 	    size, _iflib_dmamap_cb, &dma->idi_paddr, mapflags | BUS_DMA_NOWAIT);
1391 	if (err || dma->idi_paddr == IF_BAD_DMA) {
1392 		device_printf(dev,
1393 		    "%s: bus_dmamap_load failed: %d\n",
1394 		    __func__, err);
1395 		goto fail_2;
1396 	}
1397 
1398 	dma->idi_size = size;
1399 	return (0);
1400 
1401 fail_2:
1402 	bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1403 fail_1:
1404 	bus_dma_tag_destroy(dma->idi_tag);
1405 fail_0:
1406 	dma->idi_tag = NULL;
1407 
1408 	return (err);
1409 }
1410 
1411 int
1412 iflib_dma_alloc_multi(if_ctx_t ctx, int *sizes, iflib_dma_info_t *dmalist, int mapflags, int count)
1413 {
1414 	int i, err;
1415 	iflib_dma_info_t *dmaiter;
1416 
1417 	dmaiter = dmalist;
1418 	for (i = 0; i < count; i++, dmaiter++) {
1419 		if ((err = iflib_dma_alloc(ctx, sizes[i], *dmaiter, mapflags)) != 0)
1420 			break;
1421 	}
1422 	if (err)
1423 		iflib_dma_free_multi(dmalist, i);
1424 	return (err);
1425 }
1426 
1427 void
1428 iflib_dma_free(iflib_dma_info_t dma)
1429 {
1430 	if (dma->idi_tag == NULL)
1431 		return;
1432 	if (dma->idi_paddr != IF_BAD_DMA) {
1433 		bus_dmamap_sync(dma->idi_tag, dma->idi_map,
1434 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1435 		bus_dmamap_unload(dma->idi_tag, dma->idi_map);
1436 		dma->idi_paddr = IF_BAD_DMA;
1437 	}
1438 	if (dma->idi_vaddr != NULL) {
1439 		bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1440 		dma->idi_vaddr = NULL;
1441 	}
1442 	bus_dma_tag_destroy(dma->idi_tag);
1443 	dma->idi_tag = NULL;
1444 }
1445 
1446 void
1447 iflib_dma_free_multi(iflib_dma_info_t *dmalist, int count)
1448 {
1449 	int i;
1450 	iflib_dma_info_t *dmaiter = dmalist;
1451 
1452 	for (i = 0; i < count; i++, dmaiter++)
1453 		iflib_dma_free(*dmaiter);
1454 }
1455 
1456 #ifdef EARLY_AP_STARTUP
1457 static const int iflib_started = 1;
1458 #else
1459 /*
1460  * We used to abuse the smp_started flag to decide if the queues have been
1461  * fully initialized (by late taskqgroup_adjust() calls in a SYSINIT()).
1462  * That gave bad races, since the SYSINIT() runs strictly after smp_started
1463  * is set.  Run a SYSINIT() strictly after that to just set a usable
1464  * completion flag.
1465  */
1466 
1467 static int iflib_started;
1468 
1469 static void
1470 iflib_record_started(void *arg)
1471 {
1472 	iflib_started = 1;
1473 }
1474 
1475 SYSINIT(iflib_record_started, SI_SUB_SMP + 1, SI_ORDER_FIRST,
1476 	iflib_record_started, NULL);
1477 #endif
1478 
1479 static int
1480 iflib_fast_intr(void *arg)
1481 {
1482 	iflib_filter_info_t info = arg;
1483 	struct grouptask *gtask = info->ifi_task;
1484 	if (!iflib_started)
1485 		return (FILTER_HANDLED);
1486 
1487 	DBG_COUNTER_INC(fast_intrs);
1488 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1489 		return (FILTER_HANDLED);
1490 
1491 	GROUPTASK_ENQUEUE(gtask);
1492 	return (FILTER_HANDLED);
1493 }
1494 
1495 static int
1496 iflib_fast_intr_rxtx(void *arg)
1497 {
1498 	iflib_filter_info_t info = arg;
1499 	struct grouptask *gtask = info->ifi_task;
1500 	iflib_rxq_t rxq = (iflib_rxq_t)info->ifi_ctx;
1501 	if_ctx_t ctx = NULL;;
1502 	int i, cidx;
1503 
1504 	if (!iflib_started)
1505 		return (FILTER_HANDLED);
1506 
1507 	DBG_COUNTER_INC(fast_intrs);
1508 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1509 		return (FILTER_HANDLED);
1510 
1511 	MPASS(rxq->ifr_ntxqirq);
1512 	for (i = 0; i < rxq->ifr_ntxqirq; i++) {
1513 		qidx_t txqid = rxq->ifr_txqid[i];
1514 
1515 		ctx = rxq->ifr_ctx;
1516 
1517 		if (!ctx->isc_txd_credits_update(ctx->ifc_softc, txqid, false)) {
1518 			IFDI_TX_QUEUE_INTR_ENABLE(ctx, txqid);
1519 			continue;
1520 		}
1521 		GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
1522 	}
1523 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_RXCQ)
1524 		cidx = rxq->ifr_cq_cidx;
1525 	else
1526 		cidx = rxq->ifr_fl[0].ifl_cidx;
1527 	if (iflib_rxd_avail(ctx, rxq, cidx, 1))
1528 		GROUPTASK_ENQUEUE(gtask);
1529 	else {
1530 		IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
1531 		DBG_COUNTER_INC(rx_intr_enables);
1532 	}
1533 	return (FILTER_HANDLED);
1534 }
1535 
1536 
1537 static int
1538 iflib_fast_intr_ctx(void *arg)
1539 {
1540 	iflib_filter_info_t info = arg;
1541 	struct grouptask *gtask = info->ifi_task;
1542 
1543 	if (!iflib_started)
1544 		return (FILTER_HANDLED);
1545 
1546 	DBG_COUNTER_INC(fast_intrs);
1547 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1548 		return (FILTER_HANDLED);
1549 
1550 	GROUPTASK_ENQUEUE(gtask);
1551 	return (FILTER_HANDLED);
1552 }
1553 
1554 static int
1555 _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
1556 		 driver_filter_t filter, driver_intr_t handler, void *arg,
1557 		 const char *name)
1558 {
1559 	int rc, flags;
1560 	struct resource *res;
1561 	void *tag = NULL;
1562 	device_t dev = ctx->ifc_dev;
1563 
1564 	flags = RF_ACTIVE;
1565 	if (ctx->ifc_flags & IFC_LEGACY)
1566 		flags |= RF_SHAREABLE;
1567 	MPASS(rid < 512);
1568 	irq->ii_rid = rid;
1569 	res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &irq->ii_rid, flags);
1570 	if (res == NULL) {
1571 		device_printf(dev,
1572 		    "failed to allocate IRQ for rid %d, name %s.\n", rid, name);
1573 		return (ENOMEM);
1574 	}
1575 	irq->ii_res = res;
1576 	KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL"));
1577 	rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET,
1578 						filter, handler, arg, &tag);
1579 	if (rc != 0) {
1580 		device_printf(dev,
1581 		    "failed to setup interrupt for rid %d, name %s: %d\n",
1582 					  rid, name ? name : "unknown", rc);
1583 		return (rc);
1584 	} else if (name)
1585 		bus_describe_intr(dev, res, tag, "%s", name);
1586 
1587 	irq->ii_tag = tag;
1588 	return (0);
1589 }
1590 
1591 
1592 /*********************************************************************
1593  *
1594  *  Allocate memory for tx_buffer structures. The tx_buffer stores all
1595  *  the information needed to transmit a packet on the wire. This is
1596  *  called only once at attach, setup is done every reset.
1597  *
1598  **********************************************************************/
1599 
1600 static int
1601 iflib_txsd_alloc(iflib_txq_t txq)
1602 {
1603 	if_ctx_t ctx = txq->ift_ctx;
1604 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1605 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1606 	device_t dev = ctx->ifc_dev;
1607 	bus_size_t tsomaxsize;
1608 	int err, nsegments, ntsosegments;
1609 
1610 	nsegments = scctx->isc_tx_nsegments;
1611 	ntsosegments = scctx->isc_tx_tso_segments_max;
1612 	tsomaxsize = scctx->isc_tx_tso_size_max;
1613 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_VLAN_MTU)
1614 		tsomaxsize += sizeof(struct ether_vlan_header);
1615 	MPASS(scctx->isc_ntxd[0] > 0);
1616 	MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0);
1617 	MPASS(nsegments > 0);
1618 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) {
1619 		MPASS(ntsosegments > 0);
1620 		MPASS(sctx->isc_tso_maxsize >= tsomaxsize);
1621 	}
1622 
1623 	/*
1624 	 * Setup DMA descriptor areas.
1625 	 */
1626 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1627 			       1, 0,			/* alignment, bounds */
1628 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1629 			       BUS_SPACE_MAXADDR,	/* highaddr */
1630 			       NULL, NULL,		/* filter, filterarg */
1631 			       sctx->isc_tx_maxsize,		/* maxsize */
1632 			       nsegments,	/* nsegments */
1633 			       sctx->isc_tx_maxsegsize,	/* maxsegsize */
1634 			       0,			/* flags */
1635 			       NULL,			/* lockfunc */
1636 			       NULL,			/* lockfuncarg */
1637 			       &txq->ift_desc_tag))) {
1638 		device_printf(dev,"Unable to allocate TX DMA tag: %d\n", err);
1639 		device_printf(dev,"maxsize: %ju nsegments: %d maxsegsize: %ju\n",
1640 		    (uintmax_t)sctx->isc_tx_maxsize, nsegments, (uintmax_t)sctx->isc_tx_maxsegsize);
1641 		goto fail;
1642 	}
1643 	if ((if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) &
1644 	    (err = bus_dma_tag_create(bus_get_dma_tag(dev),
1645 			       1, 0,			/* alignment, bounds */
1646 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1647 			       BUS_SPACE_MAXADDR,	/* highaddr */
1648 			       NULL, NULL,		/* filter, filterarg */
1649 			       tsomaxsize,		/* maxsize */
1650 			       ntsosegments,	/* nsegments */
1651 			       sctx->isc_tso_maxsegsize,/* maxsegsize */
1652 			       0,			/* flags */
1653 			       NULL,			/* lockfunc */
1654 			       NULL,			/* lockfuncarg */
1655 			       &txq->ift_tso_desc_tag))) {
1656 		device_printf(dev,"Unable to allocate TX TSO DMA tag: %d\n", err);
1657 
1658 		goto fail;
1659 	}
1660 	if (!(txq->ift_sds.ifsd_flags =
1661 	    (uint8_t *) malloc(sizeof(uint8_t) *
1662 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1663 		device_printf(dev, "Unable to allocate tx_buffer memory\n");
1664 		err = ENOMEM;
1665 		goto fail;
1666 	}
1667 	if (!(txq->ift_sds.ifsd_m =
1668 	    (struct mbuf **) malloc(sizeof(struct mbuf *) *
1669 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1670 		device_printf(dev, "Unable to allocate tx_buffer memory\n");
1671 		err = ENOMEM;
1672 		goto fail;
1673 	}
1674 
1675         /* Create the descriptor buffer dma maps */
1676 #if defined(ACPI_DMAR) || (! (defined(__i386__) || defined(__amd64__)))
1677 	if ((ctx->ifc_flags & IFC_DMAR) == 0)
1678 		return (0);
1679 
1680 	if (!(txq->ift_sds.ifsd_map =
1681 	    (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1682 		device_printf(dev, "Unable to allocate tx_buffer map memory\n");
1683 		err = ENOMEM;
1684 		goto fail;
1685 	}
1686 
1687 	for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) {
1688 		err = bus_dmamap_create(txq->ift_desc_tag, 0, &txq->ift_sds.ifsd_map[i]);
1689 		if (err != 0) {
1690 			device_printf(dev, "Unable to create TX DMA map\n");
1691 			goto fail;
1692 		}
1693 	}
1694 #endif
1695 	return (0);
1696 fail:
1697 	/* We free all, it handles case where we are in the middle */
1698 	iflib_tx_structures_free(ctx);
1699 	return (err);
1700 }
1701 
1702 static void
1703 iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i)
1704 {
1705 	bus_dmamap_t map;
1706 
1707 	map = NULL;
1708 	if (txq->ift_sds.ifsd_map != NULL)
1709 		map = txq->ift_sds.ifsd_map[i];
1710 	if (map != NULL) {
1711 		bus_dmamap_unload(txq->ift_desc_tag, map);
1712 		bus_dmamap_destroy(txq->ift_desc_tag, map);
1713 		txq->ift_sds.ifsd_map[i] = NULL;
1714 	}
1715 }
1716 
1717 static void
1718 iflib_txq_destroy(iflib_txq_t txq)
1719 {
1720 	if_ctx_t ctx = txq->ift_ctx;
1721 
1722 	for (int i = 0; i < txq->ift_size; i++)
1723 		iflib_txsd_destroy(ctx, txq, i);
1724 	if (txq->ift_sds.ifsd_map != NULL) {
1725 		free(txq->ift_sds.ifsd_map, M_IFLIB);
1726 		txq->ift_sds.ifsd_map = NULL;
1727 	}
1728 	if (txq->ift_sds.ifsd_m != NULL) {
1729 		free(txq->ift_sds.ifsd_m, M_IFLIB);
1730 		txq->ift_sds.ifsd_m = NULL;
1731 	}
1732 	if (txq->ift_sds.ifsd_flags != NULL) {
1733 		free(txq->ift_sds.ifsd_flags, M_IFLIB);
1734 		txq->ift_sds.ifsd_flags = NULL;
1735 	}
1736 	if (txq->ift_desc_tag != NULL) {
1737 		bus_dma_tag_destroy(txq->ift_desc_tag);
1738 		txq->ift_desc_tag = NULL;
1739 	}
1740 	if (txq->ift_tso_desc_tag != NULL) {
1741 		bus_dma_tag_destroy(txq->ift_tso_desc_tag);
1742 		txq->ift_tso_desc_tag = NULL;
1743 	}
1744 }
1745 
1746 static void
1747 iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i)
1748 {
1749 	struct mbuf **mp;
1750 
1751 	mp = &txq->ift_sds.ifsd_m[i];
1752 	if (*mp == NULL)
1753 		return;
1754 
1755 	if (txq->ift_sds.ifsd_map != NULL) {
1756 		bus_dmamap_sync(txq->ift_desc_tag,
1757 				txq->ift_sds.ifsd_map[i],
1758 				BUS_DMASYNC_POSTWRITE);
1759 		bus_dmamap_unload(txq->ift_desc_tag,
1760 				  txq->ift_sds.ifsd_map[i]);
1761 	}
1762 	m_free(*mp);
1763 	DBG_COUNTER_INC(tx_frees);
1764 	*mp = NULL;
1765 }
1766 
1767 static int
1768 iflib_txq_setup(iflib_txq_t txq)
1769 {
1770 	if_ctx_t ctx = txq->ift_ctx;
1771 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1772 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1773 	iflib_dma_info_t di;
1774 	int i;
1775 
1776 	/* Set number of descriptors available */
1777 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
1778 	/* XXX make configurable */
1779 	txq->ift_update_freq = IFLIB_DEFAULT_TX_UPDATE_FREQ;
1780 
1781 	/* Reset indices */
1782 	txq->ift_cidx_processed = 0;
1783 	txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0;
1784 	txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset];
1785 
1786 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1787 		bzero((void *)di->idi_vaddr, di->idi_size);
1788 
1789 	IFDI_TXQ_SETUP(ctx, txq->ift_id);
1790 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1791 		bus_dmamap_sync(di->idi_tag, di->idi_map,
1792 						BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1793 	return (0);
1794 }
1795 
1796 /*********************************************************************
1797  *
1798  *  Allocate memory for rx_buffer structures. Since we use one
1799  *  rx_buffer per received packet, the maximum number of rx_buffer's
1800  *  that we'll need is equal to the number of receive descriptors
1801  *  that we've allocated.
1802  *
1803  **********************************************************************/
1804 static int
1805 iflib_rxsd_alloc(iflib_rxq_t rxq)
1806 {
1807 	if_ctx_t ctx = rxq->ifr_ctx;
1808 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1809 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1810 	device_t dev = ctx->ifc_dev;
1811 	iflib_fl_t fl;
1812 	int			err;
1813 
1814 	MPASS(scctx->isc_nrxd[0] > 0);
1815 	MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0);
1816 
1817 	fl = rxq->ifr_fl;
1818 	for (int i = 0; i <  rxq->ifr_nfl; i++, fl++) {
1819 		fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */
1820 		err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1821 					 1, 0,			/* alignment, bounds */
1822 					 BUS_SPACE_MAXADDR,	/* lowaddr */
1823 					 BUS_SPACE_MAXADDR,	/* highaddr */
1824 					 NULL, NULL,		/* filter, filterarg */
1825 					 sctx->isc_rx_maxsize,	/* maxsize */
1826 					 sctx->isc_rx_nsegments,	/* nsegments */
1827 					 sctx->isc_rx_maxsegsize,	/* maxsegsize */
1828 					 0,			/* flags */
1829 					 NULL,			/* lockfunc */
1830 					 NULL,			/* lockarg */
1831 					 &fl->ifl_desc_tag);
1832 		if (err) {
1833 			device_printf(dev, "%s: bus_dma_tag_create failed %d\n",
1834 				__func__, err);
1835 			goto fail;
1836 		}
1837 		if (!(fl->ifl_sds.ifsd_flags =
1838 		      (uint8_t *) malloc(sizeof(uint8_t) *
1839 					 scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1840 			device_printf(dev, "Unable to allocate tx_buffer memory\n");
1841 			err = ENOMEM;
1842 			goto fail;
1843 		}
1844 		if (!(fl->ifl_sds.ifsd_m =
1845 		      (struct mbuf **) malloc(sizeof(struct mbuf *) *
1846 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1847 			device_printf(dev, "Unable to allocate tx_buffer memory\n");
1848 			err = ENOMEM;
1849 			goto fail;
1850 		}
1851 		if (!(fl->ifl_sds.ifsd_cl =
1852 		      (caddr_t *) malloc(sizeof(caddr_t) *
1853 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1854 			device_printf(dev, "Unable to allocate tx_buffer memory\n");
1855 			err = ENOMEM;
1856 			goto fail;
1857 		}
1858 
1859 		/* Create the descriptor buffer dma maps */
1860 #if defined(ACPI_DMAR) || (! (defined(__i386__) || defined(__amd64__)))
1861 		if ((ctx->ifc_flags & IFC_DMAR) == 0)
1862 			continue;
1863 
1864 		if (!(fl->ifl_sds.ifsd_map =
1865 		      (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1866 			device_printf(dev, "Unable to allocate tx_buffer map memory\n");
1867 			err = ENOMEM;
1868 			goto fail;
1869 		}
1870 
1871 		for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++) {
1872 			err = bus_dmamap_create(fl->ifl_desc_tag, 0, &fl->ifl_sds.ifsd_map[i]);
1873 			if (err != 0) {
1874 				device_printf(dev, "Unable to create RX buffer DMA map\n");
1875 				goto fail;
1876 			}
1877 		}
1878 #endif
1879 	}
1880 	return (0);
1881 
1882 fail:
1883 	iflib_rx_structures_free(ctx);
1884 	return (err);
1885 }
1886 
1887 
1888 /*
1889  * Internal service routines
1890  */
1891 
1892 struct rxq_refill_cb_arg {
1893 	int               error;
1894 	bus_dma_segment_t seg;
1895 	int               nseg;
1896 };
1897 
1898 static void
1899 _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
1900 {
1901 	struct rxq_refill_cb_arg *cb_arg = arg;
1902 
1903 	cb_arg->error = error;
1904 	cb_arg->seg = segs[0];
1905 	cb_arg->nseg = nseg;
1906 }
1907 
1908 
1909 #ifdef ACPI_DMAR
1910 #define IS_DMAR(ctx) (ctx->ifc_flags & IFC_DMAR)
1911 #else
1912 #define IS_DMAR(ctx) (0)
1913 #endif
1914 
1915 /**
1916  *	rxq_refill - refill an rxq  free-buffer list
1917  *	@ctx: the iflib context
1918  *	@rxq: the free-list to refill
1919  *	@n: the number of new buffers to allocate
1920  *
1921  *	(Re)populate an rxq free-buffer list with up to @n new packet buffers.
1922  *	The caller must assure that @n does not exceed the queue's capacity.
1923  */
1924 static void
1925 _iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
1926 {
1927 	struct mbuf *m;
1928 	int idx, frag_idx = fl->ifl_fragidx;
1929         int pidx = fl->ifl_pidx;
1930 	caddr_t cl, *sd_cl;
1931 	struct mbuf **sd_m;
1932 	uint8_t *sd_flags;
1933 	struct if_rxd_update iru;
1934 	bus_dmamap_t *sd_map;
1935 	int n, i = 0;
1936 	uint64_t bus_addr;
1937 	int err;
1938 	qidx_t credits;
1939 
1940 	sd_m = fl->ifl_sds.ifsd_m;
1941 	sd_map = fl->ifl_sds.ifsd_map;
1942 	sd_cl = fl->ifl_sds.ifsd_cl;
1943 	sd_flags = fl->ifl_sds.ifsd_flags;
1944 	idx = pidx;
1945 	credits = fl->ifl_credits;
1946 
1947 	n  = count;
1948 	MPASS(n > 0);
1949 	MPASS(credits + n <= fl->ifl_size);
1950 
1951 	if (pidx < fl->ifl_cidx)
1952 		MPASS(pidx + n <= fl->ifl_cidx);
1953 	if (pidx == fl->ifl_cidx && (credits < fl->ifl_size))
1954 		MPASS(fl->ifl_gen == 0);
1955 	if (pidx > fl->ifl_cidx)
1956 		MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx);
1957 
1958 	DBG_COUNTER_INC(fl_refills);
1959 	if (n > 8)
1960 		DBG_COUNTER_INC(fl_refills_large);
1961 	iru_init(&iru, fl->ifl_rxq, fl->ifl_id);
1962 	while (n--) {
1963 		/*
1964 		 * We allocate an uninitialized mbuf + cluster, mbuf is
1965 		 * initialized after rx.
1966 		 *
1967 		 * If the cluster is still set then we know a minimum sized packet was received
1968 		 */
1969 		bit_ffc_at(fl->ifl_rx_bitmap, frag_idx, fl->ifl_size,  &frag_idx);
1970 		if ((frag_idx < 0) || (frag_idx >= fl->ifl_size))
1971                 	bit_ffc(fl->ifl_rx_bitmap, fl->ifl_size, &frag_idx);
1972 		if ((cl = sd_cl[frag_idx]) == NULL) {
1973                        if ((cl = sd_cl[frag_idx] = m_cljget(NULL, M_NOWAIT, fl->ifl_buf_size)) == NULL)
1974 				break;
1975 #if MEMORY_LOGGING
1976 			fl->ifl_cl_enqueued++;
1977 #endif
1978 		}
1979 		if ((m = m_gethdr(M_NOWAIT, MT_NOINIT)) == NULL) {
1980 			break;
1981 		}
1982 #if MEMORY_LOGGING
1983 		fl->ifl_m_enqueued++;
1984 #endif
1985 
1986 		DBG_COUNTER_INC(rx_allocs);
1987 #if defined(__i386__) || defined(__amd64__)
1988 		if (!IS_DMAR(ctx)) {
1989 			bus_addr = pmap_kextract((vm_offset_t)cl);
1990 		} else
1991 #endif
1992 		{
1993 			struct rxq_refill_cb_arg cb_arg;
1994 
1995 			cb_arg.error = 0;
1996 			MPASS(sd_map != NULL);
1997 			MPASS(sd_map[frag_idx] != NULL);
1998 			err = bus_dmamap_load(fl->ifl_desc_tag, sd_map[frag_idx],
1999 		         cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg, 0);
2000 			bus_dmamap_sync(fl->ifl_desc_tag, sd_map[frag_idx],
2001 					BUS_DMASYNC_PREREAD);
2002 
2003 			if (err != 0 || cb_arg.error) {
2004 				/*
2005 				 * !zone_pack ?
2006 				 */
2007 				if (fl->ifl_zone == zone_pack)
2008 					uma_zfree(fl->ifl_zone, cl);
2009 				m_free(m);
2010 				n = 0;
2011 				goto done;
2012 			}
2013 			bus_addr = cb_arg.seg.ds_addr;
2014 		}
2015                 bit_set(fl->ifl_rx_bitmap, frag_idx);
2016 		sd_flags[frag_idx] |= RX_SW_DESC_INUSE;
2017 
2018 		MPASS(sd_m[frag_idx] == NULL);
2019 		sd_cl[frag_idx] = cl;
2020 		sd_m[frag_idx] = m;
2021 		fl->ifl_rxd_idxs[i] = frag_idx;
2022 		fl->ifl_bus_addrs[i] = bus_addr;
2023 		fl->ifl_vm_addrs[i] = cl;
2024 		credits++;
2025 		i++;
2026 		MPASS(credits <= fl->ifl_size);
2027 		if (++idx == fl->ifl_size) {
2028 			fl->ifl_gen = 1;
2029 			idx = 0;
2030 		}
2031 		if (n == 0 || i == IFLIB_MAX_RX_REFRESH) {
2032 			iru.iru_pidx = pidx;
2033 			iru.iru_count = i;
2034 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2035 			i = 0;
2036 			pidx = idx;
2037 			fl->ifl_pidx = idx;
2038 			fl->ifl_credits = credits;
2039 		}
2040 
2041 	}
2042 done:
2043 	if (i) {
2044 		iru.iru_pidx = pidx;
2045 		iru.iru_count = i;
2046 		ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2047 		fl->ifl_pidx = idx;
2048 		fl->ifl_credits = credits;
2049 	}
2050 	DBG_COUNTER_INC(rxd_flush);
2051 	if (fl->ifl_pidx == 0)
2052 		pidx = fl->ifl_size - 1;
2053 	else
2054 		pidx = fl->ifl_pidx - 1;
2055 
2056 	if (sd_map)
2057 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2058 				BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2059 	ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id, fl->ifl_id, pidx);
2060 	fl->ifl_fragidx = frag_idx;
2061 }
2062 
2063 static __inline void
2064 __iflib_fl_refill_lt(if_ctx_t ctx, iflib_fl_t fl, int max)
2065 {
2066 	/* we avoid allowing pidx to catch up with cidx as it confuses ixl */
2067 	int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1;
2068 #ifdef INVARIANTS
2069 	int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1;
2070 #endif
2071 
2072 	MPASS(fl->ifl_credits <= fl->ifl_size);
2073 	MPASS(reclaimable == delta);
2074 
2075 	if (reclaimable > 0)
2076 		_iflib_fl_refill(ctx, fl, min(max, reclaimable));
2077 }
2078 
2079 uint8_t
2080 iflib_in_detach(if_ctx_t ctx)
2081 {
2082 	bool in_detach;
2083 	STATE_LOCK(ctx);
2084 	in_detach = !!(ctx->ifc_flags & IFC_IN_DETACH);
2085 	STATE_UNLOCK(ctx);
2086 	return (in_detach);
2087 }
2088 
2089 static void
2090 iflib_fl_bufs_free(iflib_fl_t fl)
2091 {
2092 	iflib_dma_info_t idi = fl->ifl_ifdi;
2093 	uint32_t i;
2094 
2095 	for (i = 0; i < fl->ifl_size; i++) {
2096 		struct mbuf **sd_m = &fl->ifl_sds.ifsd_m[i];
2097 		uint8_t *sd_flags = &fl->ifl_sds.ifsd_flags[i];
2098 		caddr_t *sd_cl = &fl->ifl_sds.ifsd_cl[i];
2099 
2100 		if (*sd_flags & RX_SW_DESC_INUSE) {
2101 			if (fl->ifl_sds.ifsd_map != NULL) {
2102 				bus_dmamap_t sd_map = fl->ifl_sds.ifsd_map[i];
2103 				bus_dmamap_unload(fl->ifl_desc_tag, sd_map);
2104 				// XXX: Should this get moved out?
2105 				if (iflib_in_detach(fl->ifl_rxq->ifr_ctx))
2106 					bus_dmamap_destroy(fl->ifl_desc_tag, sd_map);
2107 			}
2108 			if (*sd_m != NULL) {
2109 				m_init(*sd_m, M_NOWAIT, MT_DATA, 0);
2110 				uma_zfree(zone_mbuf, *sd_m);
2111 			}
2112 			if (*sd_cl != NULL)
2113 				uma_zfree(fl->ifl_zone, *sd_cl);
2114 			*sd_flags = 0;
2115 		} else {
2116 			MPASS(*sd_cl == NULL);
2117 			MPASS(*sd_m == NULL);
2118 		}
2119 #if MEMORY_LOGGING
2120 		fl->ifl_m_dequeued++;
2121 		fl->ifl_cl_dequeued++;
2122 #endif
2123 		*sd_cl = NULL;
2124 		*sd_m = NULL;
2125 	}
2126 #ifdef INVARIANTS
2127 	for (i = 0; i < fl->ifl_size; i++) {
2128 		MPASS(fl->ifl_sds.ifsd_flags[i] == 0);
2129 		MPASS(fl->ifl_sds.ifsd_cl[i] == NULL);
2130 		MPASS(fl->ifl_sds.ifsd_m[i] == NULL);
2131 	}
2132 #endif
2133 	/*
2134 	 * Reset free list values
2135 	 */
2136 	fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = fl->ifl_fragidx = 0;
2137 	bzero(idi->idi_vaddr, idi->idi_size);
2138 }
2139 
2140 /*********************************************************************
2141  *
2142  *  Initialize a receive ring and its buffers.
2143  *
2144  **********************************************************************/
2145 static int
2146 iflib_fl_setup(iflib_fl_t fl)
2147 {
2148 	iflib_rxq_t rxq = fl->ifl_rxq;
2149 	if_ctx_t ctx = rxq->ifr_ctx;
2150 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2151 
2152 	bit_nclear(fl->ifl_rx_bitmap, 0, fl->ifl_size - 1);
2153 	/*
2154 	** Free current RX buffer structs and their mbufs
2155 	*/
2156 	iflib_fl_bufs_free(fl);
2157 	/* Now replenish the mbufs */
2158 	MPASS(fl->ifl_credits == 0);
2159 	/*
2160 	 * XXX don't set the max_frame_size to larger
2161 	 * than the hardware can handle
2162 	 */
2163 	if (sctx->isc_max_frame_size <= 2048)
2164 		fl->ifl_buf_size = MCLBYTES;
2165 #ifndef CONTIGMALLOC_WORKS
2166 	else
2167 		fl->ifl_buf_size = MJUMPAGESIZE;
2168 #else
2169 	else if (sctx->isc_max_frame_size <= 4096)
2170 		fl->ifl_buf_size = MJUMPAGESIZE;
2171 	else if (sctx->isc_max_frame_size <= 9216)
2172 		fl->ifl_buf_size = MJUM9BYTES;
2173 	else
2174 		fl->ifl_buf_size = MJUM16BYTES;
2175 #endif
2176 	if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size)
2177 		ctx->ifc_max_fl_buf_size = fl->ifl_buf_size;
2178 	fl->ifl_cltype = m_gettype(fl->ifl_buf_size);
2179 	fl->ifl_zone = m_getzone(fl->ifl_buf_size);
2180 
2181 
2182 	/* avoid pre-allocating zillions of clusters to an idle card
2183 	 * potentially speeding up attach
2184 	 */
2185 	_iflib_fl_refill(ctx, fl, min(128, fl->ifl_size));
2186 	MPASS(min(128, fl->ifl_size) == fl->ifl_credits);
2187 	if (min(128, fl->ifl_size) != fl->ifl_credits)
2188 		return (ENOBUFS);
2189 	/*
2190 	 * handle failure
2191 	 */
2192 	MPASS(rxq != NULL);
2193 	MPASS(fl->ifl_ifdi != NULL);
2194 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2195 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2196 	return (0);
2197 }
2198 
2199 /*********************************************************************
2200  *
2201  *  Free receive ring data structures
2202  *
2203  **********************************************************************/
2204 static void
2205 iflib_rx_sds_free(iflib_rxq_t rxq)
2206 {
2207 	iflib_fl_t fl;
2208 	int i;
2209 
2210 	if (rxq->ifr_fl != NULL) {
2211 		for (i = 0; i < rxq->ifr_nfl; i++) {
2212 			fl = &rxq->ifr_fl[i];
2213 			if (fl->ifl_desc_tag != NULL) {
2214 				bus_dma_tag_destroy(fl->ifl_desc_tag);
2215 				fl->ifl_desc_tag = NULL;
2216 			}
2217 			free(fl->ifl_sds.ifsd_m, M_IFLIB);
2218 			free(fl->ifl_sds.ifsd_cl, M_IFLIB);
2219 			/* XXX destroy maps first */
2220 			free(fl->ifl_sds.ifsd_map, M_IFLIB);
2221 			fl->ifl_sds.ifsd_m = NULL;
2222 			fl->ifl_sds.ifsd_cl = NULL;
2223 			fl->ifl_sds.ifsd_map = NULL;
2224 		}
2225 		free(rxq->ifr_fl, M_IFLIB);
2226 		rxq->ifr_fl = NULL;
2227 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
2228 	}
2229 }
2230 
2231 /*
2232  * MI independent logic
2233  *
2234  */
2235 static void
2236 iflib_timer(void *arg)
2237 {
2238 	iflib_txq_t txq = arg;
2239 	if_ctx_t ctx = txq->ift_ctx;
2240 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2241 	uint64_t this_tick = ticks;
2242 	uint32_t reset_on = hz / 2;
2243 
2244 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
2245 		return;
2246 	/*
2247 	** Check on the state of the TX queue(s), this
2248 	** can be done without the lock because its RO
2249 	** and the HUNG state will be static if set.
2250 	*/
2251 	if (this_tick - txq->ift_last_timer_tick >= hz / 2) {
2252 		txq->ift_last_timer_tick = this_tick;
2253 		IFDI_TIMER(ctx, txq->ift_id);
2254 		if ((txq->ift_qstatus == IFLIB_QUEUE_HUNG) &&
2255 		    ((txq->ift_cleaned_prev == txq->ift_cleaned) ||
2256 		     (sctx->isc_pause_frames == 0)))
2257 			goto hung;
2258 
2259 		if (ifmp_ring_is_stalled(txq->ift_br))
2260 			txq->ift_qstatus = IFLIB_QUEUE_HUNG;
2261 		txq->ift_cleaned_prev = txq->ift_cleaned;
2262 	}
2263 #ifdef DEV_NETMAP
2264 	if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP)
2265 		iflib_netmap_timer_adjust(ctx, txq->ift_id, &reset_on);
2266 #endif
2267 	/* handle any laggards */
2268 	if (txq->ift_db_pending)
2269 		GROUPTASK_ENQUEUE(&txq->ift_task);
2270 
2271 	sctx->isc_pause_frames = 0;
2272 	if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)
2273 		callout_reset_on(&txq->ift_timer, reset_on, iflib_timer, txq, txq->ift_timer.c_cpu);
2274 	return;
2275  hung:
2276 	device_printf(ctx->ifc_dev,  "TX(%d) desc avail = %d, pidx = %d\n",
2277 				  txq->ift_id, TXQ_AVAIL(txq), txq->ift_pidx);
2278 	STATE_LOCK(ctx);
2279 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2280 	ctx->ifc_flags |= (IFC_DO_WATCHDOG|IFC_DO_RESET);
2281 	iflib_admin_intr_deferred(ctx);
2282 	STATE_UNLOCK(ctx);
2283 }
2284 
2285 static void
2286 iflib_init_locked(if_ctx_t ctx)
2287 {
2288 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2289 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2290 	if_t ifp = ctx->ifc_ifp;
2291 	iflib_fl_t fl;
2292 	iflib_txq_t txq;
2293 	iflib_rxq_t rxq;
2294 	int i, j, tx_ip_csum_flags, tx_ip6_csum_flags;
2295 
2296 
2297 	if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2298 	IFDI_INTR_DISABLE(ctx);
2299 
2300 	tx_ip_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP);
2301 	tx_ip6_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | CSUM_IP6_SCTP);
2302 	/* Set hardware offload abilities */
2303 	if_clearhwassist(ifp);
2304 	if (if_getcapenable(ifp) & IFCAP_TXCSUM)
2305 		if_sethwassistbits(ifp, tx_ip_csum_flags, 0);
2306 	if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6)
2307 		if_sethwassistbits(ifp,  tx_ip6_csum_flags, 0);
2308 	if (if_getcapenable(ifp) & IFCAP_TSO4)
2309 		if_sethwassistbits(ifp, CSUM_IP_TSO, 0);
2310 	if (if_getcapenable(ifp) & IFCAP_TSO6)
2311 		if_sethwassistbits(ifp, CSUM_IP6_TSO, 0);
2312 
2313 	for (i = 0, txq = ctx->ifc_txqs; i < sctx->isc_ntxqsets; i++, txq++) {
2314 		CALLOUT_LOCK(txq);
2315 		callout_stop(&txq->ift_timer);
2316 		CALLOUT_UNLOCK(txq);
2317 		iflib_netmap_txq_init(ctx, txq);
2318 	}
2319 #ifdef INVARIANTS
2320 	i = if_getdrvflags(ifp);
2321 #endif
2322 	IFDI_INIT(ctx);
2323 	MPASS(if_getdrvflags(ifp) == i);
2324 	for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
2325 		/* XXX this should really be done on a per-queue basis */
2326 		if (if_getcapenable(ifp) & IFCAP_NETMAP) {
2327 			MPASS(rxq->ifr_id == i);
2328 			iflib_netmap_rxq_init(ctx, rxq);
2329 			continue;
2330 		}
2331 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
2332 			if (iflib_fl_setup(fl)) {
2333 				device_printf(ctx->ifc_dev, "freelist setup failed - check cluster settings\n");
2334 				goto done;
2335 			}
2336 		}
2337 	}
2338 done:
2339 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
2340 	IFDI_INTR_ENABLE(ctx);
2341 	txq = ctx->ifc_txqs;
2342 	for (i = 0; i < sctx->isc_ntxqsets; i++, txq++)
2343 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq,
2344 			txq->ift_timer.c_cpu);
2345 }
2346 
2347 static int
2348 iflib_media_change(if_t ifp)
2349 {
2350 	if_ctx_t ctx = if_getsoftc(ifp);
2351 	int err;
2352 
2353 	CTX_LOCK(ctx);
2354 	if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0)
2355 		iflib_init_locked(ctx);
2356 	CTX_UNLOCK(ctx);
2357 	return (err);
2358 }
2359 
2360 static void
2361 iflib_media_status(if_t ifp, struct ifmediareq *ifmr)
2362 {
2363 	if_ctx_t ctx = if_getsoftc(ifp);
2364 
2365 	CTX_LOCK(ctx);
2366 	IFDI_UPDATE_ADMIN_STATUS(ctx);
2367 	IFDI_MEDIA_STATUS(ctx, ifmr);
2368 	CTX_UNLOCK(ctx);
2369 }
2370 
2371 void
2372 iflib_stop(if_ctx_t ctx)
2373 {
2374 	iflib_txq_t txq = ctx->ifc_txqs;
2375 	iflib_rxq_t rxq = ctx->ifc_rxqs;
2376 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2377 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2378 	iflib_dma_info_t di;
2379 	iflib_fl_t fl;
2380 	int i, j;
2381 
2382 	/* Tell the stack that the interface is no longer active */
2383 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2384 
2385 	IFDI_INTR_DISABLE(ctx);
2386 	DELAY(1000);
2387 	IFDI_STOP(ctx);
2388 	DELAY(1000);
2389 
2390 	iflib_debug_reset();
2391 	/* Wait for current tx queue users to exit to disarm watchdog timer. */
2392 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) {
2393 		/* make sure all transmitters have completed before proceeding XXX */
2394 
2395 		CALLOUT_LOCK(txq);
2396 		callout_stop(&txq->ift_timer);
2397 		CALLOUT_UNLOCK(txq);
2398 
2399 		/* clean any enqueued buffers */
2400 		iflib_ifmp_purge(txq);
2401 		/* Free any existing tx buffers. */
2402 		for (j = 0; j < txq->ift_size; j++) {
2403 			iflib_txsd_free(ctx, txq, j);
2404 		}
2405 		txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0;
2406 		txq->ift_in_use = txq->ift_gen = txq->ift_cidx = txq->ift_pidx = txq->ift_no_desc_avail = 0;
2407 		txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0;
2408 		txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0;
2409 		txq->ift_pullups = 0;
2410 		ifmp_ring_reset_stats(txq->ift_br);
2411 		for (j = 0, di = txq->ift_ifdi; j < sctx->isc_ntxqs; j++, di++)
2412 			bzero((void *)di->idi_vaddr, di->idi_size);
2413 	}
2414 	for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) {
2415 		/* make sure all transmitters have completed before proceeding XXX */
2416 
2417 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
2418 		for (j = 0, di = rxq->ifr_ifdi; j < sctx->isc_nrxqs; j++, di++)
2419 			bzero((void *)di->idi_vaddr, di->idi_size);
2420 		/* also resets the free lists pidx/cidx */
2421 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
2422 			iflib_fl_bufs_free(fl);
2423 	}
2424 }
2425 
2426 static inline caddr_t
2427 calc_next_rxd(iflib_fl_t fl, int cidx)
2428 {
2429 	qidx_t size;
2430 	int nrxd;
2431 	caddr_t start, end, cur, next;
2432 
2433 	nrxd = fl->ifl_size;
2434 	size = fl->ifl_rxd_size;
2435 	start = fl->ifl_ifdi->idi_vaddr;
2436 
2437 	if (__predict_false(size == 0))
2438 		return (start);
2439 	cur = start + size*cidx;
2440 	end = start + size*nrxd;
2441 	next = CACHE_PTR_NEXT(cur);
2442 	return (next < end ? next : start);
2443 }
2444 
2445 static inline void
2446 prefetch_pkts(iflib_fl_t fl, int cidx)
2447 {
2448 	int nextptr;
2449 	int nrxd = fl->ifl_size;
2450 	caddr_t next_rxd;
2451 
2452 
2453 	nextptr = (cidx + CACHE_PTR_INCREMENT) & (nrxd-1);
2454 	prefetch(&fl->ifl_sds.ifsd_m[nextptr]);
2455 	prefetch(&fl->ifl_sds.ifsd_cl[nextptr]);
2456 	next_rxd = calc_next_rxd(fl, cidx);
2457 	prefetch(next_rxd);
2458 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 1) & (nrxd-1)]);
2459 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 2) & (nrxd-1)]);
2460 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 3) & (nrxd-1)]);
2461 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 4) & (nrxd-1)]);
2462 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 1) & (nrxd-1)]);
2463 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 2) & (nrxd-1)]);
2464 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 3) & (nrxd-1)]);
2465 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 4) & (nrxd-1)]);
2466 }
2467 
2468 static void
2469 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, int unload, if_rxsd_t sd)
2470 {
2471 	int flid, cidx;
2472 	bus_dmamap_t map;
2473 	iflib_fl_t fl;
2474 	iflib_dma_info_t di;
2475 	int next;
2476 
2477 	map = NULL;
2478 	flid = irf->irf_flid;
2479 	cidx = irf->irf_idx;
2480 	fl = &rxq->ifr_fl[flid];
2481 	sd->ifsd_fl = fl;
2482 	sd->ifsd_cidx = cidx;
2483 	sd->ifsd_m = &fl->ifl_sds.ifsd_m[cidx];
2484 	sd->ifsd_cl = &fl->ifl_sds.ifsd_cl[cidx];
2485 	fl->ifl_credits--;
2486 #if MEMORY_LOGGING
2487 	fl->ifl_m_dequeued++;
2488 #endif
2489 	if (rxq->ifr_ctx->ifc_flags & IFC_PREFETCH)
2490 		prefetch_pkts(fl, cidx);
2491 	if (fl->ifl_sds.ifsd_map != NULL) {
2492 		next = (cidx + CACHE_PTR_INCREMENT) & (fl->ifl_size-1);
2493 		prefetch(&fl->ifl_sds.ifsd_map[next]);
2494 		map = fl->ifl_sds.ifsd_map[cidx];
2495 		di = fl->ifl_ifdi;
2496 		next = (cidx + CACHE_LINE_SIZE) & (fl->ifl_size-1);
2497 		prefetch(&fl->ifl_sds.ifsd_flags[next]);
2498 		bus_dmamap_sync(di->idi_tag, di->idi_map,
2499 				BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
2500 
2501 	/* not valid assert if bxe really does SGE from non-contiguous elements */
2502 		MPASS(fl->ifl_cidx == cidx);
2503 		if (unload)
2504 			bus_dmamap_unload(fl->ifl_desc_tag, map);
2505 	}
2506 	fl->ifl_cidx = (fl->ifl_cidx + 1) & (fl->ifl_size-1);
2507 	if (__predict_false(fl->ifl_cidx == 0))
2508 		fl->ifl_gen = 0;
2509 	if (map != NULL)
2510 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2511 			BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2512         bit_clear(fl->ifl_rx_bitmap, cidx);
2513 }
2514 
2515 static struct mbuf *
2516 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri, if_rxsd_t sd)
2517 {
2518 	int i, padlen , flags;
2519 	struct mbuf *m, *mh, *mt;
2520 	caddr_t cl;
2521 
2522 	i = 0;
2523 	mh = NULL;
2524 	do {
2525 		rxd_frag_to_sd(rxq, &ri->iri_frags[i], TRUE, sd);
2526 
2527 		MPASS(*sd->ifsd_cl != NULL);
2528 		MPASS(*sd->ifsd_m != NULL);
2529 
2530 		/* Don't include zero-length frags */
2531 		if (ri->iri_frags[i].irf_len == 0) {
2532 			/* XXX we can save the cluster here, but not the mbuf */
2533 			m_init(*sd->ifsd_m, M_NOWAIT, MT_DATA, 0);
2534 			m_free(*sd->ifsd_m);
2535 			*sd->ifsd_m = NULL;
2536 			continue;
2537 		}
2538 		m = *sd->ifsd_m;
2539 		*sd->ifsd_m = NULL;
2540 		if (mh == NULL) {
2541 			flags = M_PKTHDR|M_EXT;
2542 			mh = mt = m;
2543 			padlen = ri->iri_pad;
2544 		} else {
2545 			flags = M_EXT;
2546 			mt->m_next = m;
2547 			mt = m;
2548 			/* assuming padding is only on the first fragment */
2549 			padlen = 0;
2550 		}
2551 		cl = *sd->ifsd_cl;
2552 		*sd->ifsd_cl = NULL;
2553 
2554 		/* Can these two be made one ? */
2555 		m_init(m, M_NOWAIT, MT_DATA, flags);
2556 		m_cljset(m, cl, sd->ifsd_fl->ifl_cltype);
2557 		/*
2558 		 * These must follow m_init and m_cljset
2559 		 */
2560 		m->m_data += padlen;
2561 		ri->iri_len -= padlen;
2562 		m->m_len = ri->iri_frags[i].irf_len;
2563 	} while (++i < ri->iri_nfrags);
2564 
2565 	return (mh);
2566 }
2567 
2568 /*
2569  * Process one software descriptor
2570  */
2571 static struct mbuf *
2572 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri)
2573 {
2574 	struct if_rxsd sd;
2575 	struct mbuf *m;
2576 
2577 	/* should I merge this back in now that the two paths are basically duplicated? */
2578 	if (ri->iri_nfrags == 1 &&
2579 	    ri->iri_frags[0].irf_len <= MIN(IFLIB_RX_COPY_THRESH, MHLEN)) {
2580 		rxd_frag_to_sd(rxq, &ri->iri_frags[0], FALSE, &sd);
2581 		m = *sd.ifsd_m;
2582 		*sd.ifsd_m = NULL;
2583 		m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR);
2584 #ifndef __NO_STRICT_ALIGNMENT
2585 		if (!IP_ALIGNED(m))
2586 			m->m_data += 2;
2587 #endif
2588 		memcpy(m->m_data, *sd.ifsd_cl, ri->iri_len);
2589 		m->m_len = ri->iri_frags[0].irf_len;
2590        } else {
2591 		m = assemble_segments(rxq, ri, &sd);
2592 	}
2593 	m->m_pkthdr.len = ri->iri_len;
2594 	m->m_pkthdr.rcvif = ri->iri_ifp;
2595 	m->m_flags |= ri->iri_flags;
2596 	m->m_pkthdr.ether_vtag = ri->iri_vtag;
2597 	m->m_pkthdr.flowid = ri->iri_flowid;
2598 	M_HASHTYPE_SET(m, ri->iri_rsstype);
2599 	m->m_pkthdr.csum_flags = ri->iri_csum_flags;
2600 	m->m_pkthdr.csum_data = ri->iri_csum_data;
2601 	return (m);
2602 }
2603 
2604 #if defined(INET6) || defined(INET)
2605 static void
2606 iflib_get_ip_forwarding(struct lro_ctrl *lc, bool *v4, bool *v6)
2607 {
2608 	CURVNET_SET(lc->ifp->if_vnet);
2609 #if defined(INET6)
2610 	*v6 = VNET(ip6_forwarding);
2611 #endif
2612 #if defined(INET)
2613 	*v4 = VNET(ipforwarding);
2614 #endif
2615 	CURVNET_RESTORE();
2616 }
2617 
2618 /*
2619  * Returns true if it's possible this packet could be LROed.
2620  * if it returns false, it is guaranteed that tcp_lro_rx()
2621  * would not return zero.
2622  */
2623 static bool
2624 iflib_check_lro_possible(struct mbuf *m, bool v4_forwarding, bool v6_forwarding)
2625 {
2626 	struct ether_header *eh;
2627 	uint16_t eh_type;
2628 
2629 	eh = mtod(m, struct ether_header *);
2630 	eh_type = ntohs(eh->ether_type);
2631 	switch (eh_type) {
2632 #if defined(INET6)
2633 		case ETHERTYPE_IPV6:
2634 			return !v6_forwarding;
2635 #endif
2636 #if defined (INET)
2637 		case ETHERTYPE_IP:
2638 			return !v4_forwarding;
2639 #endif
2640 	}
2641 
2642 	return false;
2643 }
2644 #else
2645 static void
2646 iflib_get_ip_forwarding(struct lro_ctrl *lc __unused, bool *v4 __unused, bool *v6 __unused)
2647 {
2648 }
2649 #endif
2650 
2651 static bool
2652 iflib_rxeof(iflib_rxq_t rxq, qidx_t budget)
2653 {
2654 	if_ctx_t ctx = rxq->ifr_ctx;
2655 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2656 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2657 	int avail, i;
2658 	qidx_t *cidxp;
2659 	struct if_rxd_info ri;
2660 	int err, budget_left, rx_bytes, rx_pkts;
2661 	iflib_fl_t fl;
2662 	struct ifnet *ifp;
2663 	int lro_enabled;
2664 	bool v4_forwarding, v6_forwarding, lro_possible;
2665 
2666 	/*
2667 	 * XXX early demux data packets so that if_input processing only handles
2668 	 * acks in interrupt context
2669 	 */
2670 	struct mbuf *m, *mh, *mt, *mf;
2671 
2672 	lro_possible = v4_forwarding = v6_forwarding = false;
2673 	ifp = ctx->ifc_ifp;
2674 	mh = mt = NULL;
2675 	MPASS(budget > 0);
2676 	rx_pkts	= rx_bytes = 0;
2677 	if (sctx->isc_flags & IFLIB_HAS_RXCQ)
2678 		cidxp = &rxq->ifr_cq_cidx;
2679 	else
2680 		cidxp = &rxq->ifr_fl[0].ifl_cidx;
2681 	if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) {
2682 		for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2683 			__iflib_fl_refill_lt(ctx, fl, budget + 8);
2684 		DBG_COUNTER_INC(rx_unavail);
2685 		return (false);
2686 	}
2687 
2688 	for (budget_left = budget; budget_left > 0 && avail > 0;) {
2689 		if (__predict_false(!CTX_ACTIVE(ctx))) {
2690 			DBG_COUNTER_INC(rx_ctx_inactive);
2691 			break;
2692 		}
2693 		/*
2694 		 * Reset client set fields to their default values
2695 		 */
2696 		rxd_info_zero(&ri);
2697 		ri.iri_qsidx = rxq->ifr_id;
2698 		ri.iri_cidx = *cidxp;
2699 		ri.iri_ifp = ifp;
2700 		ri.iri_frags = rxq->ifr_frags;
2701 		err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
2702 
2703 		if (err)
2704 			goto err;
2705 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
2706 			*cidxp = ri.iri_cidx;
2707 			/* Update our consumer index */
2708 			/* XXX NB: shurd - check if this is still safe */
2709 			while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0]) {
2710 				rxq->ifr_cq_cidx -= scctx->isc_nrxd[0];
2711 				rxq->ifr_cq_gen = 0;
2712 			}
2713 			/* was this only a completion queue message? */
2714 			if (__predict_false(ri.iri_nfrags == 0))
2715 				continue;
2716 		}
2717 		MPASS(ri.iri_nfrags != 0);
2718 		MPASS(ri.iri_len != 0);
2719 
2720 		/* will advance the cidx on the corresponding free lists */
2721 		m = iflib_rxd_pkt_get(rxq, &ri);
2722 		avail--;
2723 		budget_left--;
2724 		if (avail == 0 && budget_left)
2725 			avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left);
2726 
2727 		if (__predict_false(m == NULL)) {
2728 			DBG_COUNTER_INC(rx_mbuf_null);
2729 			continue;
2730 		}
2731 		/* imm_pkt: -- cxgb */
2732 		if (mh == NULL)
2733 			mh = mt = m;
2734 		else {
2735 			mt->m_nextpkt = m;
2736 			mt = m;
2737 		}
2738 	}
2739 	/* make sure that we can refill faster than drain */
2740 	for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2741 		__iflib_fl_refill_lt(ctx, fl, budget + 8);
2742 
2743 	lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO);
2744 	if (lro_enabled)
2745 		iflib_get_ip_forwarding(&rxq->ifr_lc, &v4_forwarding, &v6_forwarding);
2746 	mt = mf = NULL;
2747 	while (mh != NULL) {
2748 		m = mh;
2749 		mh = mh->m_nextpkt;
2750 		m->m_nextpkt = NULL;
2751 #ifndef __NO_STRICT_ALIGNMENT
2752 		if (!IP_ALIGNED(m) && (m = iflib_fixup_rx(m)) == NULL)
2753 			continue;
2754 #endif
2755 		rx_bytes += m->m_pkthdr.len;
2756 		rx_pkts++;
2757 #if defined(INET6) || defined(INET)
2758 		if (lro_enabled) {
2759 			if (!lro_possible) {
2760 				lro_possible = iflib_check_lro_possible(m, v4_forwarding, v6_forwarding);
2761 				if (lro_possible && mf != NULL) {
2762 					ifp->if_input(ifp, mf);
2763 					DBG_COUNTER_INC(rx_if_input);
2764 					mt = mf = NULL;
2765 				}
2766 			}
2767 			if ((m->m_pkthdr.csum_flags & (CSUM_L4_CALC|CSUM_L4_VALID)) ==
2768 			    (CSUM_L4_CALC|CSUM_L4_VALID)) {
2769 				if (lro_possible && tcp_lro_rx(&rxq->ifr_lc, m, 0) == 0)
2770 					continue;
2771 			}
2772 		}
2773 #endif
2774 		if (lro_possible) {
2775 			ifp->if_input(ifp, m);
2776 			DBG_COUNTER_INC(rx_if_input);
2777 			continue;
2778 		}
2779 
2780 		if (mf == NULL)
2781 			mf = m;
2782 		if (mt != NULL)
2783 			mt->m_nextpkt = m;
2784 		mt = m;
2785 	}
2786 	if (mf != NULL) {
2787 		ifp->if_input(ifp, mf);
2788 		DBG_COUNTER_INC(rx_if_input);
2789 	}
2790 
2791 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
2792 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
2793 
2794 	/*
2795 	 * Flush any outstanding LRO work
2796 	 */
2797 #if defined(INET6) || defined(INET)
2798 	tcp_lro_flush_all(&rxq->ifr_lc);
2799 #endif
2800 	if (avail)
2801 		return true;
2802 	return (iflib_rxd_avail(ctx, rxq, *cidxp, 1));
2803 err:
2804 	STATE_LOCK(ctx);
2805 	ctx->ifc_flags |= IFC_DO_RESET;
2806 	iflib_admin_intr_deferred(ctx);
2807 	STATE_UNLOCK(ctx);
2808 	return (false);
2809 }
2810 
2811 #define TXD_NOTIFY_COUNT(txq) (((txq)->ift_size / (txq)->ift_update_freq)-1)
2812 static inline qidx_t
2813 txq_max_db_deferred(iflib_txq_t txq, qidx_t in_use)
2814 {
2815 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2816 	qidx_t minthresh = txq->ift_size / 8;
2817 	if (in_use > 4*minthresh)
2818 		return (notify_count);
2819 	if (in_use > 2*minthresh)
2820 		return (notify_count >> 1);
2821 	if (in_use > minthresh)
2822 		return (notify_count >> 3);
2823 	return (0);
2824 }
2825 
2826 static inline qidx_t
2827 txq_max_rs_deferred(iflib_txq_t txq)
2828 {
2829 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2830 	qidx_t minthresh = txq->ift_size / 8;
2831 	if (txq->ift_in_use > 4*minthresh)
2832 		return (notify_count);
2833 	if (txq->ift_in_use > 2*minthresh)
2834 		return (notify_count >> 1);
2835 	if (txq->ift_in_use > minthresh)
2836 		return (notify_count >> 2);
2837 	return (2);
2838 }
2839 
2840 #define M_CSUM_FLAGS(m) ((m)->m_pkthdr.csum_flags)
2841 #define M_HAS_VLANTAG(m) (m->m_flags & M_VLANTAG)
2842 
2843 #define TXQ_MAX_DB_DEFERRED(txq, in_use) txq_max_db_deferred((txq), (in_use))
2844 #define TXQ_MAX_RS_DEFERRED(txq) txq_max_rs_deferred(txq)
2845 #define TXQ_MAX_DB_CONSUMED(size) (size >> 4)
2846 
2847 /* forward compatibility for cxgb */
2848 #define FIRST_QSET(ctx) 0
2849 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets)
2850 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets)
2851 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx))
2852 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments))
2853 
2854 /* XXX we should be setting this to something other than zero */
2855 #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh)
2856 #define	MAX_TX_DESC(ctx) max((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max, \
2857     (ctx)->ifc_softc_ctx.isc_tx_nsegments)
2858 
2859 static inline bool
2860 iflib_txd_db_check(if_ctx_t ctx, iflib_txq_t txq, int ring, qidx_t in_use)
2861 {
2862 	qidx_t dbval, max;
2863 	bool rang;
2864 
2865 	rang = false;
2866 	max = TXQ_MAX_DB_DEFERRED(txq, in_use);
2867 	if (ring || txq->ift_db_pending >= max) {
2868 		dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx;
2869 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval);
2870 		txq->ift_db_pending = txq->ift_npending = 0;
2871 		rang = true;
2872 	}
2873 	return (rang);
2874 }
2875 
2876 #ifdef PKT_DEBUG
2877 static void
2878 print_pkt(if_pkt_info_t pi)
2879 {
2880 	printf("pi len:  %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n",
2881 	       pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx);
2882 	printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n",
2883 	       pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag);
2884 	printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n",
2885 	       pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto);
2886 }
2887 #endif
2888 
2889 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO)
2890 #define IS_TX_OFFLOAD4(pi) ((pi)->ipi_csum_flags & (CSUM_IP_TCP | CSUM_IP_TSO))
2891 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO)
2892 #define IS_TX_OFFLOAD6(pi) ((pi)->ipi_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_TSO))
2893 
2894 static int
2895 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp)
2896 {
2897 	if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx;
2898 	struct ether_vlan_header *eh;
2899 	struct mbuf *m;
2900 
2901 	m = *mp;
2902 	if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) &&
2903 	    M_WRITABLE(m) == 0) {
2904 		if ((m = m_dup(m, M_NOWAIT)) == NULL) {
2905 			return (ENOMEM);
2906 		} else {
2907 			m_freem(*mp);
2908 			DBG_COUNTER_INC(tx_frees);
2909 			*mp = m;
2910 		}
2911 	}
2912 
2913 	/*
2914 	 * Determine where frame payload starts.
2915 	 * Jump over vlan headers if already present,
2916 	 * helpful for QinQ too.
2917 	 */
2918 	if (__predict_false(m->m_len < sizeof(*eh))) {
2919 		txq->ift_pullups++;
2920 		if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL))
2921 			return (ENOMEM);
2922 	}
2923 	eh = mtod(m, struct ether_vlan_header *);
2924 	if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
2925 		pi->ipi_etype = ntohs(eh->evl_proto);
2926 		pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
2927 	} else {
2928 		pi->ipi_etype = ntohs(eh->evl_encap_proto);
2929 		pi->ipi_ehdrlen = ETHER_HDR_LEN;
2930 	}
2931 
2932 	switch (pi->ipi_etype) {
2933 #ifdef INET
2934 	case ETHERTYPE_IP:
2935 	{
2936 		struct mbuf *n;
2937 		struct ip *ip = NULL;
2938 		struct tcphdr *th = NULL;
2939 		int minthlen;
2940 
2941 		minthlen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip) + sizeof(*th));
2942 		if (__predict_false(m->m_len < minthlen)) {
2943 			/*
2944 			 * if this code bloat is causing too much of a hit
2945 			 * move it to a separate function and mark it noinline
2946 			 */
2947 			if (m->m_len == pi->ipi_ehdrlen) {
2948 				n = m->m_next;
2949 				MPASS(n);
2950 				if (n->m_len >= sizeof(*ip))  {
2951 					ip = (struct ip *)n->m_data;
2952 					if (n->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2953 						th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2954 				} else {
2955 					txq->ift_pullups++;
2956 					if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
2957 						return (ENOMEM);
2958 					ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2959 				}
2960 			} else {
2961 				txq->ift_pullups++;
2962 				if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
2963 					return (ENOMEM);
2964 				ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2965 				if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2966 					th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2967 			}
2968 		} else {
2969 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2970 			if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2971 				th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2972 		}
2973 		pi->ipi_ip_hlen = ip->ip_hl << 2;
2974 		pi->ipi_ipproto = ip->ip_p;
2975 		pi->ipi_flags |= IPI_TX_IPV4;
2976 
2977 		/* TCP checksum offload may require TCP header length */
2978 		if (IS_TX_OFFLOAD4(pi)) {
2979 			if (__predict_true(pi->ipi_ipproto == IPPROTO_TCP)) {
2980 				if (__predict_false(th == NULL)) {
2981 					txq->ift_pullups++;
2982 					if (__predict_false((m = m_pullup(m, (ip->ip_hl << 2) + sizeof(*th))) == NULL))
2983 						return (ENOMEM);
2984 					th = (struct tcphdr *)((caddr_t)ip + pi->ipi_ip_hlen);
2985 				}
2986 				pi->ipi_tcp_hflags = th->th_flags;
2987 				pi->ipi_tcp_hlen = th->th_off << 2;
2988 				pi->ipi_tcp_seq = th->th_seq;
2989 			}
2990 			if (IS_TSO4(pi)) {
2991 				if (__predict_false(ip->ip_p != IPPROTO_TCP))
2992 					return (ENXIO);
2993 				/*
2994 				 * TSO always requires hardware checksum offload.
2995 				 */
2996 				pi->ipi_csum_flags |= (CSUM_IP_TCP | CSUM_IP);
2997 				th->th_sum = in_pseudo(ip->ip_src.s_addr,
2998 						       ip->ip_dst.s_addr, htons(IPPROTO_TCP));
2999 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3000 				if (sctx->isc_flags & IFLIB_TSO_INIT_IP) {
3001 					ip->ip_sum = 0;
3002 					ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz);
3003 				}
3004 			}
3005 		}
3006 		if ((sctx->isc_flags & IFLIB_NEED_ZERO_CSUM) && (pi->ipi_csum_flags & CSUM_IP))
3007                        ip->ip_sum = 0;
3008 
3009 		break;
3010 	}
3011 #endif
3012 #ifdef INET6
3013 	case ETHERTYPE_IPV6:
3014 	{
3015 		struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
3016 		struct tcphdr *th;
3017 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
3018 
3019 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
3020 			txq->ift_pullups++;
3021 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
3022 				return (ENOMEM);
3023 		}
3024 		th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
3025 
3026 		/* XXX-BZ this will go badly in case of ext hdrs. */
3027 		pi->ipi_ipproto = ip6->ip6_nxt;
3028 		pi->ipi_flags |= IPI_TX_IPV6;
3029 
3030 		/* TCP checksum offload may require TCP header length */
3031 		if (IS_TX_OFFLOAD6(pi)) {
3032 			if (pi->ipi_ipproto == IPPROTO_TCP) {
3033 				if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) {
3034 					txq->ift_pullups++;
3035 					if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL))
3036 						return (ENOMEM);
3037 				}
3038 				pi->ipi_tcp_hflags = th->th_flags;
3039 				pi->ipi_tcp_hlen = th->th_off << 2;
3040 				pi->ipi_tcp_seq = th->th_seq;
3041 			}
3042 			if (IS_TSO6(pi)) {
3043 				if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP))
3044 					return (ENXIO);
3045 				/*
3046 				 * TSO always requires hardware checksum offload.
3047 				 */
3048 				pi->ipi_csum_flags |= CSUM_IP6_TCP;
3049 				th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0);
3050 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3051 			}
3052 		}
3053 		break;
3054 	}
3055 #endif
3056 	default:
3057 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
3058 		pi->ipi_ip_hlen = 0;
3059 		break;
3060 	}
3061 	*mp = m;
3062 
3063 	return (0);
3064 }
3065 
3066 /*
3067  * If dodgy hardware rejects the scatter gather chain we've handed it
3068  * we'll need to remove the mbuf chain from ifsg_m[] before we can add the
3069  * m_defrag'd mbufs
3070  */
3071 static __noinline struct mbuf *
3072 iflib_remove_mbuf(iflib_txq_t txq)
3073 {
3074 	int ntxd, i, pidx;
3075 	struct mbuf *m, *mh, **ifsd_m;
3076 
3077 	pidx = txq->ift_pidx;
3078 	ifsd_m = txq->ift_sds.ifsd_m;
3079 	ntxd = txq->ift_size;
3080 	mh = m = ifsd_m[pidx];
3081 	ifsd_m[pidx] = NULL;
3082 #if MEMORY_LOGGING
3083 	txq->ift_dequeued++;
3084 #endif
3085 	i = 1;
3086 
3087 	while (m) {
3088 		ifsd_m[(pidx + i) & (ntxd -1)] = NULL;
3089 #if MEMORY_LOGGING
3090 		txq->ift_dequeued++;
3091 #endif
3092 		m = m->m_next;
3093 		i++;
3094 	}
3095 	return (mh);
3096 }
3097 
3098 static int
3099 iflib_busdma_load_mbuf_sg(iflib_txq_t txq, bus_dma_tag_t tag, bus_dmamap_t map,
3100 			  struct mbuf **m0, bus_dma_segment_t *segs, int *nsegs,
3101 			  int max_segs, int flags)
3102 {
3103 	if_ctx_t ctx;
3104 	if_shared_ctx_t		sctx;
3105 	if_softc_ctx_t		scctx;
3106 	int i, next, pidx, err, ntxd, count;
3107 	struct mbuf *m, *tmp, **ifsd_m;
3108 
3109 	m = *m0;
3110 
3111 	/*
3112 	 * Please don't ever do this
3113 	 */
3114 	MPASS(__predict_true(m->m_len > 0));
3115 
3116 	ctx = txq->ift_ctx;
3117 	sctx = ctx->ifc_sctx;
3118 	scctx = &ctx->ifc_softc_ctx;
3119 	ifsd_m = txq->ift_sds.ifsd_m;
3120 	ntxd = txq->ift_size;
3121 	pidx = txq->ift_pidx;
3122 	if (map != NULL) {
3123 		uint8_t *ifsd_flags = txq->ift_sds.ifsd_flags;
3124 
3125 		err = bus_dmamap_load_mbuf_sg(tag, map,
3126 					      *m0, segs, nsegs, BUS_DMA_NOWAIT);
3127 		if (err)
3128 			return (err);
3129 		ifsd_flags[pidx] |= TX_SW_DESC_MAPPED;
3130 		count = 0;
3131 		m = *m0;
3132 		do {
3133 			if (__predict_false(m->m_len <= 0)) {
3134 				tmp = m;
3135 				m = m->m_next;
3136 				tmp->m_next = NULL;
3137 				m_free(tmp);
3138 				continue;
3139 			}
3140 			m = m->m_next;
3141 			count++;
3142 		} while (m != NULL);
3143 		if (count > *nsegs) {
3144 			ifsd_m[pidx] = *m0;
3145 			ifsd_m[pidx]->m_flags |= M_TOOBIG;
3146 			return (0);
3147 		}
3148 		m = *m0;
3149 		count = 0;
3150 		do {
3151 			next = (pidx + count) & (ntxd-1);
3152 			MPASS(ifsd_m[next] == NULL);
3153 			ifsd_m[next] = m;
3154 			count++;
3155 			tmp = m;
3156 			m = m->m_next;
3157 		} while (m != NULL);
3158 	} else {
3159 		int buflen, sgsize, maxsegsz, max_sgsize;
3160 		vm_offset_t vaddr;
3161 		vm_paddr_t curaddr;
3162 
3163 		count = i = 0;
3164 		m = *m0;
3165 		if (m->m_pkthdr.csum_flags & CSUM_TSO)
3166 			maxsegsz = scctx->isc_tx_tso_segsize_max;
3167 		else
3168 			maxsegsz = sctx->isc_tx_maxsegsize;
3169 
3170 		do {
3171 			if (__predict_false(m->m_len <= 0)) {
3172 				tmp = m;
3173 				m = m->m_next;
3174 				tmp->m_next = NULL;
3175 				m_free(tmp);
3176 				continue;
3177 			}
3178 			buflen = m->m_len;
3179 			vaddr = (vm_offset_t)m->m_data;
3180 			/*
3181 			 * see if we can't be smarter about physically
3182 			 * contiguous mappings
3183 			 */
3184 			next = (pidx + count) & (ntxd-1);
3185 			MPASS(ifsd_m[next] == NULL);
3186 #if MEMORY_LOGGING
3187 			txq->ift_enqueued++;
3188 #endif
3189 			ifsd_m[next] = m;
3190 			while (buflen > 0) {
3191 				if (i >= max_segs)
3192 					goto err;
3193 				max_sgsize = MIN(buflen, maxsegsz);
3194 				curaddr = pmap_kextract(vaddr);
3195 				sgsize = PAGE_SIZE - (curaddr & PAGE_MASK);
3196 				sgsize = MIN(sgsize, max_sgsize);
3197 				segs[i].ds_addr = curaddr;
3198 				segs[i].ds_len = sgsize;
3199 				vaddr += sgsize;
3200 				buflen -= sgsize;
3201 				i++;
3202 			}
3203 			count++;
3204 			tmp = m;
3205 			m = m->m_next;
3206 		} while (m != NULL);
3207 		*nsegs = i;
3208 	}
3209 	return (0);
3210 err:
3211 	*m0 = iflib_remove_mbuf(txq);
3212 	return (EFBIG);
3213 }
3214 
3215 static inline caddr_t
3216 calc_next_txd(iflib_txq_t txq, int cidx, uint8_t qid)
3217 {
3218 	qidx_t size;
3219 	int ntxd;
3220 	caddr_t start, end, cur, next;
3221 
3222 	ntxd = txq->ift_size;
3223 	size = txq->ift_txd_size[qid];
3224 	start = txq->ift_ifdi[qid].idi_vaddr;
3225 
3226 	if (__predict_false(size == 0))
3227 		return (start);
3228 	cur = start + size*cidx;
3229 	end = start + size*ntxd;
3230 	next = CACHE_PTR_NEXT(cur);
3231 	return (next < end ? next : start);
3232 }
3233 
3234 /*
3235  * Pad an mbuf to ensure a minimum ethernet frame size.
3236  * min_frame_size is the frame size (less CRC) to pad the mbuf to
3237  */
3238 static __noinline int
3239 iflib_ether_pad(device_t dev, struct mbuf **m_head, uint16_t min_frame_size)
3240 {
3241 	/*
3242 	 * 18 is enough bytes to pad an ARP packet to 46 bytes, and
3243 	 * and ARP message is the smallest common payload I can think of
3244 	 */
3245 	static char pad[18];	/* just zeros */
3246 	int n;
3247 	struct mbuf *new_head;
3248 
3249 	if (!M_WRITABLE(*m_head)) {
3250 		new_head = m_dup(*m_head, M_NOWAIT);
3251 		if (new_head == NULL) {
3252 			m_freem(*m_head);
3253 			device_printf(dev, "cannot pad short frame, m_dup() failed");
3254 			DBG_COUNTER_INC(encap_pad_mbuf_fail);
3255 			DBG_COUNTER_INC(tx_frees);
3256 			return ENOMEM;
3257 		}
3258 		m_freem(*m_head);
3259 		*m_head = new_head;
3260 	}
3261 
3262 	for (n = min_frame_size - (*m_head)->m_pkthdr.len;
3263 	     n > 0; n -= sizeof(pad))
3264 		if (!m_append(*m_head, min(n, sizeof(pad)), pad))
3265 			break;
3266 
3267 	if (n > 0) {
3268 		m_freem(*m_head);
3269 		device_printf(dev, "cannot pad short frame\n");
3270 		DBG_COUNTER_INC(encap_pad_mbuf_fail);
3271 		DBG_COUNTER_INC(tx_frees);
3272 		return (ENOBUFS);
3273 	}
3274 
3275 	return 0;
3276 }
3277 
3278 static int
3279 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp)
3280 {
3281 	if_ctx_t		ctx;
3282 	if_shared_ctx_t		sctx;
3283 	if_softc_ctx_t		scctx;
3284 	bus_dma_segment_t	*segs;
3285 	struct mbuf		*m_head;
3286 	void			*next_txd;
3287 	bus_dmamap_t		map;
3288 	struct if_pkt_info	pi;
3289 	int remap = 0;
3290 	int err, nsegs, ndesc, max_segs, pidx, cidx, next, ntxd;
3291 	bus_dma_tag_t desc_tag;
3292 
3293 	ctx = txq->ift_ctx;
3294 	sctx = ctx->ifc_sctx;
3295 	scctx = &ctx->ifc_softc_ctx;
3296 	segs = txq->ift_segs;
3297 	ntxd = txq->ift_size;
3298 	m_head = *m_headp;
3299 	map = NULL;
3300 
3301 	/*
3302 	 * If we're doing TSO the next descriptor to clean may be quite far ahead
3303 	 */
3304 	cidx = txq->ift_cidx;
3305 	pidx = txq->ift_pidx;
3306 	if (ctx->ifc_flags & IFC_PREFETCH) {
3307 		next = (cidx + CACHE_PTR_INCREMENT) & (ntxd-1);
3308 		if (!(ctx->ifc_flags & IFLIB_HAS_TXCQ)) {
3309 			next_txd = calc_next_txd(txq, cidx, 0);
3310 			prefetch(next_txd);
3311 		}
3312 
3313 		/* prefetch the next cache line of mbuf pointers and flags */
3314 		prefetch(&txq->ift_sds.ifsd_m[next]);
3315 		if (txq->ift_sds.ifsd_map != NULL) {
3316 			prefetch(&txq->ift_sds.ifsd_map[next]);
3317 			next = (cidx + CACHE_LINE_SIZE) & (ntxd-1);
3318 			prefetch(&txq->ift_sds.ifsd_flags[next]);
3319 		}
3320 	} else if (txq->ift_sds.ifsd_map != NULL)
3321 		map = txq->ift_sds.ifsd_map[pidx];
3322 
3323 	if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
3324 		desc_tag = txq->ift_tso_desc_tag;
3325 		max_segs = scctx->isc_tx_tso_segments_max;
3326 		MPASS(desc_tag != NULL);
3327 		MPASS(max_segs > 0);
3328 	} else {
3329 		desc_tag = txq->ift_desc_tag;
3330 		max_segs = scctx->isc_tx_nsegments;
3331 	}
3332 	if ((sctx->isc_flags & IFLIB_NEED_ETHER_PAD) &&
3333 	    __predict_false(m_head->m_pkthdr.len < scctx->isc_min_frame_size)) {
3334 		err = iflib_ether_pad(ctx->ifc_dev, m_headp, scctx->isc_min_frame_size);
3335 		if (err) {
3336 			DBG_COUNTER_INC(encap_txd_encap_fail);
3337 			return err;
3338 		}
3339 	}
3340 	m_head = *m_headp;
3341 
3342 	pkt_info_zero(&pi);
3343 	pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG|M_BCAST|M_MCAST));
3344 	pi.ipi_pidx = pidx;
3345 	pi.ipi_qsidx = txq->ift_id;
3346 	pi.ipi_len = m_head->m_pkthdr.len;
3347 	pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags;
3348 	pi.ipi_vtag = (m_head->m_flags & M_VLANTAG) ? m_head->m_pkthdr.ether_vtag : 0;
3349 
3350 	/* deliberate bitwise OR to make one condition */
3351 	if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) {
3352 		if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0)) {
3353 			DBG_COUNTER_INC(encap_txd_encap_fail);
3354 			return (err);
3355 		}
3356 		m_head = *m_headp;
3357 	}
3358 
3359 retry:
3360 	err = iflib_busdma_load_mbuf_sg(txq, desc_tag, map, m_headp, segs, &nsegs, max_segs, BUS_DMA_NOWAIT);
3361 defrag:
3362 	if (__predict_false(err)) {
3363 		switch (err) {
3364 		case EFBIG:
3365 			/* try collapse once and defrag once */
3366 			if (remap == 0) {
3367 				m_head = m_collapse(*m_headp, M_NOWAIT, max_segs);
3368 				/* try defrag if collapsing fails */
3369 				if (m_head == NULL)
3370 					remap++;
3371 			}
3372 			if (remap == 1) {
3373 				txq->ift_mbuf_defrag++;
3374 				m_head = m_defrag(*m_headp, M_NOWAIT);
3375 			}
3376 			remap++;
3377 			if (__predict_false(m_head == NULL))
3378 				goto defrag_failed;
3379 			*m_headp = m_head;
3380 			goto retry;
3381 			break;
3382 		case ENOMEM:
3383 			txq->ift_no_tx_dma_setup++;
3384 			break;
3385 		default:
3386 			txq->ift_no_tx_dma_setup++;
3387 			m_freem(*m_headp);
3388 			DBG_COUNTER_INC(tx_frees);
3389 			*m_headp = NULL;
3390 			break;
3391 		}
3392 		txq->ift_map_failed++;
3393 		DBG_COUNTER_INC(encap_load_mbuf_fail);
3394 		DBG_COUNTER_INC(encap_txd_encap_fail);
3395 		return (err);
3396 	}
3397 
3398 	/*
3399 	 * XXX assumes a 1 to 1 relationship between segments and
3400 	 *        descriptors - this does not hold true on all drivers, e.g.
3401 	 *        cxgb
3402 	 */
3403 	if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) {
3404 		txq->ift_no_desc_avail++;
3405 		if (map != NULL)
3406 			bus_dmamap_unload(desc_tag, map);
3407 		DBG_COUNTER_INC(encap_txq_avail_fail);
3408 		DBG_COUNTER_INC(encap_txd_encap_fail);
3409 		if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0)
3410 			GROUPTASK_ENQUEUE(&txq->ift_task);
3411 		return (ENOBUFS);
3412 	}
3413 	/*
3414 	 * On Intel cards we can greatly reduce the number of TX interrupts
3415 	 * we see by only setting report status on every Nth descriptor.
3416 	 * However, this also means that the driver will need to keep track
3417 	 * of the descriptors that RS was set on to check them for the DD bit.
3418 	 */
3419 	txq->ift_rs_pending += nsegs + 1;
3420 	if (txq->ift_rs_pending > TXQ_MAX_RS_DEFERRED(txq) ||
3421 	     iflib_no_tx_batch || (TXQ_AVAIL(txq) - nsegs) <= MAX_TX_DESC(ctx) + 2) {
3422 		pi.ipi_flags |= IPI_TX_INTR;
3423 		txq->ift_rs_pending = 0;
3424 	}
3425 
3426 	pi.ipi_segs = segs;
3427 	pi.ipi_nsegs = nsegs;
3428 
3429 	MPASS(pidx >= 0 && pidx < txq->ift_size);
3430 #ifdef PKT_DEBUG
3431 	print_pkt(&pi);
3432 #endif
3433 	if (map != NULL)
3434 		bus_dmamap_sync(desc_tag, map, BUS_DMASYNC_PREWRITE);
3435 	if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) {
3436 		if (map != NULL)
3437 			bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3438 					BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
3439 		DBG_COUNTER_INC(tx_encap);
3440 		MPASS(pi.ipi_new_pidx < txq->ift_size);
3441 
3442 		ndesc = pi.ipi_new_pidx - pi.ipi_pidx;
3443 		if (pi.ipi_new_pidx < pi.ipi_pidx) {
3444 			ndesc += txq->ift_size;
3445 			txq->ift_gen = 1;
3446 		}
3447 		/*
3448 		 * drivers can need as many as
3449 		 * two sentinels
3450 		 */
3451 		MPASS(ndesc <= pi.ipi_nsegs + 2);
3452 		MPASS(pi.ipi_new_pidx != pidx);
3453 		MPASS(ndesc > 0);
3454 		txq->ift_in_use += ndesc;
3455 
3456 		/*
3457 		 * We update the last software descriptor again here because there may
3458 		 * be a sentinel and/or there may be more mbufs than segments
3459 		 */
3460 		txq->ift_pidx = pi.ipi_new_pidx;
3461 		txq->ift_npending += pi.ipi_ndescs;
3462 	} else {
3463 		*m_headp = m_head = iflib_remove_mbuf(txq);
3464 		if (err == EFBIG) {
3465 			txq->ift_txd_encap_efbig++;
3466 			if (remap < 2) {
3467 				remap = 1;
3468 				goto defrag;
3469 			}
3470 		}
3471 		goto defrag_failed;
3472 	}
3473 	/*
3474 	 * err can't possibly be non-zero here, so we don't neet to test it
3475 	 * to see if we need to DBG_COUNTER_INC(encap_txd_encap_fail).
3476 	 */
3477 	return (err);
3478 
3479 defrag_failed:
3480 	txq->ift_mbuf_defrag_failed++;
3481 	txq->ift_map_failed++;
3482 	m_freem(*m_headp);
3483 	DBG_COUNTER_INC(tx_frees);
3484 	*m_headp = NULL;
3485 	DBG_COUNTER_INC(encap_txd_encap_fail);
3486 	return (ENOMEM);
3487 }
3488 
3489 static void
3490 iflib_tx_desc_free(iflib_txq_t txq, int n)
3491 {
3492 	int hasmap;
3493 	uint32_t qsize, cidx, mask, gen;
3494 	struct mbuf *m, **ifsd_m;
3495 	uint8_t *ifsd_flags;
3496 	bus_dmamap_t *ifsd_map;
3497 	bool do_prefetch;
3498 
3499 	cidx = txq->ift_cidx;
3500 	gen = txq->ift_gen;
3501 	qsize = txq->ift_size;
3502 	mask = qsize-1;
3503 	hasmap = txq->ift_sds.ifsd_map != NULL;
3504 	ifsd_flags = txq->ift_sds.ifsd_flags;
3505 	ifsd_m = txq->ift_sds.ifsd_m;
3506 	ifsd_map = txq->ift_sds.ifsd_map;
3507 	do_prefetch = (txq->ift_ctx->ifc_flags & IFC_PREFETCH);
3508 
3509 	while (n-- > 0) {
3510 		if (do_prefetch) {
3511 			prefetch(ifsd_m[(cidx + 3) & mask]);
3512 			prefetch(ifsd_m[(cidx + 4) & mask]);
3513 		}
3514 		if (ifsd_m[cidx] != NULL) {
3515 			prefetch(&ifsd_m[(cidx + CACHE_PTR_INCREMENT) & mask]);
3516 			prefetch(&ifsd_flags[(cidx + CACHE_PTR_INCREMENT) & mask]);
3517 			if (hasmap && (ifsd_flags[cidx] & TX_SW_DESC_MAPPED)) {
3518 				/*
3519 				 * does it matter if it's not the TSO tag? If so we'll
3520 				 * have to add the type to flags
3521 				 */
3522 				bus_dmamap_unload(txq->ift_desc_tag, ifsd_map[cidx]);
3523 				ifsd_flags[cidx] &= ~TX_SW_DESC_MAPPED;
3524 			}
3525 			if ((m = ifsd_m[cidx]) != NULL) {
3526 				/* XXX we don't support any drivers that batch packets yet */
3527 				MPASS(m->m_nextpkt == NULL);
3528 				/* if the number of clusters exceeds the number of segments
3529 				 * there won't be space on the ring to save a pointer to each
3530 				 * cluster so we simply free the list here
3531 				 */
3532 				if (m->m_flags & M_TOOBIG) {
3533 					m_freem(m);
3534 				} else {
3535 					m_free(m);
3536 				}
3537 				ifsd_m[cidx] = NULL;
3538 #if MEMORY_LOGGING
3539 				txq->ift_dequeued++;
3540 #endif
3541 				DBG_COUNTER_INC(tx_frees);
3542 			}
3543 		}
3544 		if (__predict_false(++cidx == qsize)) {
3545 			cidx = 0;
3546 			gen = 0;
3547 		}
3548 	}
3549 	txq->ift_cidx = cidx;
3550 	txq->ift_gen = gen;
3551 }
3552 
3553 static __inline int
3554 iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh)
3555 {
3556 	int reclaim;
3557 	if_ctx_t ctx = txq->ift_ctx;
3558 
3559 	KASSERT(thresh >= 0, ("invalid threshold to reclaim"));
3560 	MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size);
3561 
3562 	/*
3563 	 * Need a rate-limiting check so that this isn't called every time
3564 	 */
3565 	iflib_tx_credits_update(ctx, txq);
3566 	reclaim = DESC_RECLAIMABLE(txq);
3567 
3568 	if (reclaim <= thresh /* + MAX_TX_DESC(txq->ift_ctx) */) {
3569 #ifdef INVARIANTS
3570 		if (iflib_verbose_debug) {
3571 			printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __FUNCTION__,
3572 			       txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments,
3573 			       reclaim, thresh);
3574 
3575 		}
3576 #endif
3577 		return (0);
3578 	}
3579 	iflib_tx_desc_free(txq, reclaim);
3580 	txq->ift_cleaned += reclaim;
3581 	txq->ift_in_use -= reclaim;
3582 
3583 	return (reclaim);
3584 }
3585 
3586 static struct mbuf **
3587 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset, int remaining)
3588 {
3589 	int next, size;
3590 	struct mbuf **items;
3591 
3592 	size = r->size;
3593 	next = (cidx + CACHE_PTR_INCREMENT) & (size-1);
3594 	items = __DEVOLATILE(struct mbuf **, &r->items[0]);
3595 
3596 	prefetch(items[(cidx + offset) & (size-1)]);
3597 	if (remaining > 1) {
3598 		prefetch2cachelines(&items[next]);
3599 		prefetch2cachelines(items[(cidx + offset + 1) & (size-1)]);
3600 		prefetch2cachelines(items[(cidx + offset + 2) & (size-1)]);
3601 		prefetch2cachelines(items[(cidx + offset + 3) & (size-1)]);
3602 	}
3603 	return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (size-1)]));
3604 }
3605 
3606 static void
3607 iflib_txq_check_drain(iflib_txq_t txq, int budget)
3608 {
3609 
3610 	ifmp_ring_check_drainage(txq->ift_br, budget);
3611 }
3612 
3613 static uint32_t
3614 iflib_txq_can_drain(struct ifmp_ring *r)
3615 {
3616 	iflib_txq_t txq = r->cookie;
3617 	if_ctx_t ctx = txq->ift_ctx;
3618 
3619 	return ((TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2) ||
3620 		ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false));
3621 }
3622 
3623 static uint32_t
3624 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3625 {
3626 	iflib_txq_t txq = r->cookie;
3627 	if_ctx_t ctx = txq->ift_ctx;
3628 	struct ifnet *ifp = ctx->ifc_ifp;
3629 	struct mbuf **mp, *m;
3630 	int i, count, consumed, pkt_sent, bytes_sent, mcast_sent, avail;
3631 	int reclaimed, err, in_use_prev, desc_used;
3632 	bool do_prefetch, ring, rang;
3633 
3634 	if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) ||
3635 			    !LINK_ACTIVE(ctx))) {
3636 		DBG_COUNTER_INC(txq_drain_notready);
3637 		return (0);
3638 	}
3639 	reclaimed = iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
3640 	rang = iflib_txd_db_check(ctx, txq, reclaimed, txq->ift_in_use);
3641 	avail = IDXDIFF(pidx, cidx, r->size);
3642 	if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) {
3643 		DBG_COUNTER_INC(txq_drain_flushing);
3644 		for (i = 0; i < avail; i++) {
3645 			if (__predict_true(r->items[(cidx + i) & (r->size-1)] != (void *)txq))
3646 				m_free(r->items[(cidx + i) & (r->size-1)]);
3647 			r->items[(cidx + i) & (r->size-1)] = NULL;
3648 		}
3649 		return (avail);
3650 	}
3651 
3652 	if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3653 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3654 		CALLOUT_LOCK(txq);
3655 		callout_stop(&txq->ift_timer);
3656 		CALLOUT_UNLOCK(txq);
3657 		DBG_COUNTER_INC(txq_drain_oactive);
3658 		return (0);
3659 	}
3660 	if (reclaimed)
3661 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3662 	consumed = mcast_sent = bytes_sent = pkt_sent = 0;
3663 	count = MIN(avail, TX_BATCH_SIZE);
3664 #ifdef INVARIANTS
3665 	if (iflib_verbose_debug)
3666 		printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __FUNCTION__,
3667 		       avail, ctx->ifc_flags, TXQ_AVAIL(txq));
3668 #endif
3669 	do_prefetch = (ctx->ifc_flags & IFC_PREFETCH);
3670 	avail = TXQ_AVAIL(txq);
3671 	err = 0;
3672 	for (desc_used = i = 0; i < count && avail > MAX_TX_DESC(ctx) + 2; i++) {
3673 		int rem = do_prefetch ? count - i : 0;
3674 
3675 		mp = _ring_peek_one(r, cidx, i, rem);
3676 		MPASS(mp != NULL && *mp != NULL);
3677 		if (__predict_false(*mp == (struct mbuf *)txq)) {
3678 			consumed++;
3679 			reclaimed++;
3680 			continue;
3681 		}
3682 		in_use_prev = txq->ift_in_use;
3683 		err = iflib_encap(txq, mp);
3684 		if (__predict_false(err)) {
3685 			/* no room - bail out */
3686 			if (err == ENOBUFS)
3687 				break;
3688 			consumed++;
3689 			/* we can't send this packet - skip it */
3690 			continue;
3691 		}
3692 		consumed++;
3693 		pkt_sent++;
3694 		m = *mp;
3695 		DBG_COUNTER_INC(tx_sent);
3696 		bytes_sent += m->m_pkthdr.len;
3697 		mcast_sent += !!(m->m_flags & M_MCAST);
3698 		avail = TXQ_AVAIL(txq);
3699 
3700 		txq->ift_db_pending += (txq->ift_in_use - in_use_prev);
3701 		desc_used += (txq->ift_in_use - in_use_prev);
3702 		ETHER_BPF_MTAP(ifp, m);
3703 		if (__predict_false(!(ifp->if_drv_flags & IFF_DRV_RUNNING)))
3704 			break;
3705 		rang = iflib_txd_db_check(ctx, txq, false, in_use_prev);
3706 	}
3707 
3708 	/* deliberate use of bitwise or to avoid gratuitous short-circuit */
3709 	ring = rang ? false  : (iflib_min_tx_latency | err) || (TXQ_AVAIL(txq) < MAX_TX_DESC(ctx));
3710 	iflib_txd_db_check(ctx, txq, ring, txq->ift_in_use);
3711 	if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
3712 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
3713 	if (mcast_sent)
3714 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
3715 #ifdef INVARIANTS
3716 	if (iflib_verbose_debug)
3717 		printf("consumed=%d\n", consumed);
3718 #endif
3719 	return (consumed);
3720 }
3721 
3722 static uint32_t
3723 iflib_txq_drain_always(struct ifmp_ring *r)
3724 {
3725 	return (1);
3726 }
3727 
3728 static uint32_t
3729 iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3730 {
3731 	int i, avail;
3732 	struct mbuf **mp;
3733 	iflib_txq_t txq;
3734 
3735 	txq = r->cookie;
3736 
3737 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3738 	CALLOUT_LOCK(txq);
3739 	callout_stop(&txq->ift_timer);
3740 	CALLOUT_UNLOCK(txq);
3741 
3742 	avail = IDXDIFF(pidx, cidx, r->size);
3743 	for (i = 0; i < avail; i++) {
3744 		mp = _ring_peek_one(r, cidx, i, avail - i);
3745 		if (__predict_false(*mp == (struct mbuf *)txq))
3746 			continue;
3747 		m_freem(*mp);
3748 		DBG_COUNTER_INC(tx_frees);
3749 	}
3750 	MPASS(ifmp_ring_is_stalled(r) == 0);
3751 	return (avail);
3752 }
3753 
3754 static void
3755 iflib_ifmp_purge(iflib_txq_t txq)
3756 {
3757 	struct ifmp_ring *r;
3758 
3759 	r = txq->ift_br;
3760 	r->drain = iflib_txq_drain_free;
3761 	r->can_drain = iflib_txq_drain_always;
3762 
3763 	ifmp_ring_check_drainage(r, r->size);
3764 
3765 	r->drain = iflib_txq_drain;
3766 	r->can_drain = iflib_txq_can_drain;
3767 }
3768 
3769 static void
3770 _task_fn_tx(void *context)
3771 {
3772 	iflib_txq_t txq = context;
3773 	if_ctx_t ctx = txq->ift_ctx;
3774 	struct ifnet *ifp = ctx->ifc_ifp;
3775 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
3776 
3777 #ifdef IFLIB_DIAGNOSTICS
3778 	txq->ift_cpu_exec_count[curcpu]++;
3779 #endif
3780 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3781 		return;
3782 	if (if_getcapenable(ifp) & IFCAP_NETMAP) {
3783 		if (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false))
3784 			netmap_tx_irq(ifp, txq->ift_id);
3785 		IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3786 		return;
3787 	}
3788 #ifdef ALTQ
3789 	if (ALTQ_IS_ENABLED(&ifp->if_snd))
3790 		iflib_altq_if_start(ifp);
3791 #endif
3792 	if (txq->ift_db_pending)
3793 		ifmp_ring_enqueue(txq->ift_br, (void **)&txq, 1, TX_BATCH_SIZE, abdicate);
3794 	else if (!abdicate)
3795 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3796 	/*
3797 	 * When abdicating, we always need to check drainage, not just when we don't enqueue
3798 	 */
3799 	if (abdicate)
3800 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3801 	ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3802 	if (ctx->ifc_flags & IFC_LEGACY)
3803 		IFDI_INTR_ENABLE(ctx);
3804 	else {
3805 #ifdef INVARIANTS
3806 		int rc =
3807 #endif
3808 			IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3809 			KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver"));
3810 	}
3811 }
3812 
3813 static void
3814 _task_fn_rx(void *context)
3815 {
3816 	iflib_rxq_t rxq = context;
3817 	if_ctx_t ctx = rxq->ifr_ctx;
3818 	bool more;
3819 	uint16_t budget;
3820 
3821 #ifdef IFLIB_DIAGNOSTICS
3822 	rxq->ifr_cpu_exec_count[curcpu]++;
3823 #endif
3824 	DBG_COUNTER_INC(task_fn_rxs);
3825 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3826 		return;
3827 	more = true;
3828 #ifdef DEV_NETMAP
3829 	if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP) {
3830 		u_int work = 0;
3831 		if (netmap_rx_irq(ctx->ifc_ifp, rxq->ifr_id, &work)) {
3832 			more = false;
3833 		}
3834 	}
3835 #endif
3836 	budget = ctx->ifc_sysctl_rx_budget;
3837 	if (budget == 0)
3838 		budget = 16;	/* XXX */
3839 	if (more == false || (more = iflib_rxeof(rxq, budget)) == false) {
3840 		if (ctx->ifc_flags & IFC_LEGACY)
3841 			IFDI_INTR_ENABLE(ctx);
3842 		else {
3843 #ifdef INVARIANTS
3844 			int rc =
3845 #endif
3846 				IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
3847 			KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver"));
3848 			DBG_COUNTER_INC(rx_intr_enables);
3849 		}
3850 	}
3851 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3852 		return;
3853 	if (more)
3854 		GROUPTASK_ENQUEUE(&rxq->ifr_task);
3855 }
3856 
3857 static void
3858 _task_fn_admin(void *context)
3859 {
3860 	if_ctx_t ctx = context;
3861 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
3862 	iflib_txq_t txq;
3863 	int i;
3864 	bool oactive, running, do_reset, do_watchdog, in_detach;
3865 	uint32_t reset_on = hz / 2;
3866 
3867 	STATE_LOCK(ctx);
3868 	running = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING);
3869 	oactive = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE);
3870 	do_reset = (ctx->ifc_flags & IFC_DO_RESET);
3871 	do_watchdog = (ctx->ifc_flags & IFC_DO_WATCHDOG);
3872 	in_detach = (ctx->ifc_flags & IFC_IN_DETACH);
3873 	ctx->ifc_flags &= ~(IFC_DO_RESET|IFC_DO_WATCHDOG);
3874 	STATE_UNLOCK(ctx);
3875 
3876 	if ((!running && !oactive) && !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
3877 		return;
3878 	if (in_detach)
3879 		return;
3880 
3881 	CTX_LOCK(ctx);
3882 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3883 		CALLOUT_LOCK(txq);
3884 		callout_stop(&txq->ift_timer);
3885 		CALLOUT_UNLOCK(txq);
3886 	}
3887 	if (do_watchdog) {
3888 		ctx->ifc_watchdog_events++;
3889 		IFDI_WATCHDOG_RESET(ctx);
3890 	}
3891 	IFDI_UPDATE_ADMIN_STATUS(ctx);
3892 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3893 #ifdef DEV_NETMAP
3894 		reset_on = hz / 2;
3895 		if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP)
3896 			iflib_netmap_timer_adjust(ctx, txq->ift_id, &reset_on);
3897 #endif
3898 		callout_reset_on(&txq->ift_timer, reset_on, iflib_timer, txq, txq->ift_timer.c_cpu);
3899 	}
3900 	IFDI_LINK_INTR_ENABLE(ctx);
3901 	if (do_reset)
3902 		iflib_if_init_locked(ctx);
3903 	CTX_UNLOCK(ctx);
3904 
3905 	if (LINK_ACTIVE(ctx) == 0)
3906 		return;
3907 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3908 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
3909 }
3910 
3911 
3912 static void
3913 _task_fn_iov(void *context)
3914 {
3915 	if_ctx_t ctx = context;
3916 
3917 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING) &&
3918 	    !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
3919 		return;
3920 
3921 	CTX_LOCK(ctx);
3922 	IFDI_VFLR_HANDLE(ctx);
3923 	CTX_UNLOCK(ctx);
3924 }
3925 
3926 static int
3927 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)
3928 {
3929 	int err;
3930 	if_int_delay_info_t info;
3931 	if_ctx_t ctx;
3932 
3933 	info = (if_int_delay_info_t)arg1;
3934 	ctx = info->iidi_ctx;
3935 	info->iidi_req = req;
3936 	info->iidi_oidp = oidp;
3937 	CTX_LOCK(ctx);
3938 	err = IFDI_SYSCTL_INT_DELAY(ctx, info);
3939 	CTX_UNLOCK(ctx);
3940 	return (err);
3941 }
3942 
3943 /*********************************************************************
3944  *
3945  *  IFNET FUNCTIONS
3946  *
3947  **********************************************************************/
3948 
3949 static void
3950 iflib_if_init_locked(if_ctx_t ctx)
3951 {
3952 	iflib_stop(ctx);
3953 	iflib_init_locked(ctx);
3954 }
3955 
3956 
3957 static void
3958 iflib_if_init(void *arg)
3959 {
3960 	if_ctx_t ctx = arg;
3961 
3962 	CTX_LOCK(ctx);
3963 	iflib_if_init_locked(ctx);
3964 	CTX_UNLOCK(ctx);
3965 }
3966 
3967 static int
3968 iflib_if_transmit(if_t ifp, struct mbuf *m)
3969 {
3970 	if_ctx_t	ctx = if_getsoftc(ifp);
3971 
3972 	iflib_txq_t txq;
3973 	int err, qidx;
3974 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
3975 
3976 	if (__predict_false((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) {
3977 		DBG_COUNTER_INC(tx_frees);
3978 		m_freem(m);
3979 		return (ENOBUFS);
3980 	}
3981 
3982 	MPASS(m->m_nextpkt == NULL);
3983 	/* ALTQ-enabled interfaces always use queue 0. */
3984 	qidx = 0;
3985 	if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m) && !ALTQ_IS_ENABLED(&ifp->if_snd))
3986 		qidx = QIDX(ctx, m);
3987 	/*
3988 	 * XXX calculate buf_ring based on flowid (divvy up bits?)
3989 	 */
3990 	txq = &ctx->ifc_txqs[qidx];
3991 
3992 #ifdef DRIVER_BACKPRESSURE
3993 	if (txq->ift_closed) {
3994 		while (m != NULL) {
3995 			next = m->m_nextpkt;
3996 			m->m_nextpkt = NULL;
3997 			m_freem(m);
3998 			DBG_COUNTER_INC(tx_frees);
3999 			m = next;
4000 		}
4001 		return (ENOBUFS);
4002 	}
4003 #endif
4004 #ifdef notyet
4005 	qidx = count = 0;
4006 	mp = marr;
4007 	next = m;
4008 	do {
4009 		count++;
4010 		next = next->m_nextpkt;
4011 	} while (next != NULL);
4012 
4013 	if (count > nitems(marr))
4014 		if ((mp = malloc(count*sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) {
4015 			/* XXX check nextpkt */
4016 			m_freem(m);
4017 			/* XXX simplify for now */
4018 			DBG_COUNTER_INC(tx_frees);
4019 			return (ENOBUFS);
4020 		}
4021 	for (next = m, i = 0; next != NULL; i++) {
4022 		mp[i] = next;
4023 		next = next->m_nextpkt;
4024 		mp[i]->m_nextpkt = NULL;
4025 	}
4026 #endif
4027 	DBG_COUNTER_INC(tx_seen);
4028 	err = ifmp_ring_enqueue(txq->ift_br, (void **)&m, 1, TX_BATCH_SIZE, abdicate);
4029 
4030 	if (abdicate)
4031 		GROUPTASK_ENQUEUE(&txq->ift_task);
4032  	if (err) {
4033 		if (!abdicate)
4034 			GROUPTASK_ENQUEUE(&txq->ift_task);
4035 		/* support forthcoming later */
4036 #ifdef DRIVER_BACKPRESSURE
4037 		txq->ift_closed = TRUE;
4038 #endif
4039 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
4040 		m_freem(m);
4041 		DBG_COUNTER_INC(tx_frees);
4042 	}
4043 
4044 	return (err);
4045 }
4046 
4047 #ifdef ALTQ
4048 /*
4049  * The overall approach to integrating iflib with ALTQ is to continue to use
4050  * the iflib mp_ring machinery between the ALTQ queue(s) and the hardware
4051  * ring.  Technically, when using ALTQ, queueing to an intermediate mp_ring
4052  * is redundant/unnecessary, but doing so minimizes the amount of
4053  * ALTQ-specific code required in iflib.  It is assumed that the overhead of
4054  * redundantly queueing to an intermediate mp_ring is swamped by the
4055  * performance limitations inherent in using ALTQ.
4056  *
4057  * When ALTQ support is compiled in, all iflib drivers will use a transmit
4058  * routine, iflib_altq_if_transmit(), that checks if ALTQ is enabled for the
4059  * given interface.  If ALTQ is enabled for an interface, then all
4060  * transmitted packets for that interface will be submitted to the ALTQ
4061  * subsystem via IFQ_ENQUEUE().  We don't use the legacy if_transmit()
4062  * implementation because it uses IFQ_HANDOFF(), which will duplicatively
4063  * update stats that the iflib machinery handles, and which is sensitve to
4064  * the disused IFF_DRV_OACTIVE flag.  Additionally, iflib_altq_if_start()
4065  * will be installed as the start routine for use by ALTQ facilities that
4066  * need to trigger queue drains on a scheduled basis.
4067  *
4068  */
4069 static void
4070 iflib_altq_if_start(if_t ifp)
4071 {
4072 	struct ifaltq *ifq = &ifp->if_snd;
4073 	struct mbuf *m;
4074 
4075 	IFQ_LOCK(ifq);
4076 	IFQ_DEQUEUE_NOLOCK(ifq, m);
4077 	while (m != NULL) {
4078 		iflib_if_transmit(ifp, m);
4079 		IFQ_DEQUEUE_NOLOCK(ifq, m);
4080 	}
4081 	IFQ_UNLOCK(ifq);
4082 }
4083 
4084 static int
4085 iflib_altq_if_transmit(if_t ifp, struct mbuf *m)
4086 {
4087 	int err;
4088 
4089 	if (ALTQ_IS_ENABLED(&ifp->if_snd)) {
4090 		IFQ_ENQUEUE(&ifp->if_snd, m, err);
4091 		if (err == 0)
4092 			iflib_altq_if_start(ifp);
4093 	} else
4094 		err = iflib_if_transmit(ifp, m);
4095 
4096 	return (err);
4097 }
4098 #endif /* ALTQ */
4099 
4100 static void
4101 iflib_if_qflush(if_t ifp)
4102 {
4103 	if_ctx_t ctx = if_getsoftc(ifp);
4104 	iflib_txq_t txq = ctx->ifc_txqs;
4105 	int i;
4106 
4107 	STATE_LOCK(ctx);
4108 	ctx->ifc_flags |= IFC_QFLUSH;
4109 	STATE_UNLOCK(ctx);
4110 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
4111 		while (!(ifmp_ring_is_idle(txq->ift_br) || ifmp_ring_is_stalled(txq->ift_br)))
4112 			iflib_txq_check_drain(txq, 0);
4113 	STATE_LOCK(ctx);
4114 	ctx->ifc_flags &= ~IFC_QFLUSH;
4115 	STATE_UNLOCK(ctx);
4116 
4117 	/*
4118 	 * When ALTQ is enabled, this will also take care of purging the
4119 	 * ALTQ queue(s).
4120 	 */
4121 	if_qflush(ifp);
4122 }
4123 
4124 
4125 #define IFCAP_FLAGS (IFCAP_HWCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \
4126 		     IFCAP_TSO | IFCAP_VLAN_HWTAGGING | IFCAP_HWSTATS | \
4127 		     IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | \
4128 		     IFCAP_VLAN_HWTSO | IFCAP_VLAN_HWCSUM)
4129 
4130 static int
4131 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data)
4132 {
4133 	if_ctx_t ctx = if_getsoftc(ifp);
4134 	struct ifreq	*ifr = (struct ifreq *)data;
4135 #if defined(INET) || defined(INET6)
4136 	struct ifaddr	*ifa = (struct ifaddr *)data;
4137 #endif
4138 	bool		avoid_reset = FALSE;
4139 	int		err = 0, reinit = 0, bits;
4140 
4141 	switch (command) {
4142 	case SIOCSIFADDR:
4143 #ifdef INET
4144 		if (ifa->ifa_addr->sa_family == AF_INET)
4145 			avoid_reset = TRUE;
4146 #endif
4147 #ifdef INET6
4148 		if (ifa->ifa_addr->sa_family == AF_INET6)
4149 			avoid_reset = TRUE;
4150 #endif
4151 		/*
4152 		** Calling init results in link renegotiation,
4153 		** so we avoid doing it when possible.
4154 		*/
4155 		if (avoid_reset) {
4156 			if_setflagbits(ifp, IFF_UP,0);
4157 			if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
4158 				reinit = 1;
4159 #ifdef INET
4160 			if (!(if_getflags(ifp) & IFF_NOARP))
4161 				arp_ifinit(ifp, ifa);
4162 #endif
4163 		} else
4164 			err = ether_ioctl(ifp, command, data);
4165 		break;
4166 	case SIOCSIFMTU:
4167 		CTX_LOCK(ctx);
4168 		if (ifr->ifr_mtu == if_getmtu(ifp)) {
4169 			CTX_UNLOCK(ctx);
4170 			break;
4171 		}
4172 		bits = if_getdrvflags(ifp);
4173 		/* stop the driver and free any clusters before proceeding */
4174 		iflib_stop(ctx);
4175 
4176 		if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) {
4177 			STATE_LOCK(ctx);
4178 			if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size)
4179 				ctx->ifc_flags |= IFC_MULTISEG;
4180 			else
4181 				ctx->ifc_flags &= ~IFC_MULTISEG;
4182 			STATE_UNLOCK(ctx);
4183 			err = if_setmtu(ifp, ifr->ifr_mtu);
4184 		}
4185 		iflib_init_locked(ctx);
4186 		STATE_LOCK(ctx);
4187 		if_setdrvflags(ifp, bits);
4188 		STATE_UNLOCK(ctx);
4189 		CTX_UNLOCK(ctx);
4190 		break;
4191 	case SIOCSIFFLAGS:
4192 		CTX_LOCK(ctx);
4193 		if (if_getflags(ifp) & IFF_UP) {
4194 			if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4195 				if ((if_getflags(ifp) ^ ctx->ifc_if_flags) &
4196 				    (IFF_PROMISC | IFF_ALLMULTI)) {
4197 					err = IFDI_PROMISC_SET(ctx, if_getflags(ifp));
4198 				}
4199 			} else
4200 				reinit = 1;
4201 		} else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4202 			iflib_stop(ctx);
4203 		}
4204 		ctx->ifc_if_flags = if_getflags(ifp);
4205 		CTX_UNLOCK(ctx);
4206 		break;
4207 	case SIOCADDMULTI:
4208 	case SIOCDELMULTI:
4209 		if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4210 			CTX_LOCK(ctx);
4211 			IFDI_INTR_DISABLE(ctx);
4212 			IFDI_MULTI_SET(ctx);
4213 			IFDI_INTR_ENABLE(ctx);
4214 			CTX_UNLOCK(ctx);
4215 		}
4216 		break;
4217 	case SIOCSIFMEDIA:
4218 		CTX_LOCK(ctx);
4219 		IFDI_MEDIA_SET(ctx);
4220 		CTX_UNLOCK(ctx);
4221 		/* falls thru */
4222 	case SIOCGIFMEDIA:
4223 	case SIOCGIFXMEDIA:
4224 		err = ifmedia_ioctl(ifp, ifr, &ctx->ifc_media, command);
4225 		break;
4226 	case SIOCGI2C:
4227 	{
4228 		struct ifi2creq i2c;
4229 
4230 		err = copyin(ifr_data_get_ptr(ifr), &i2c, sizeof(i2c));
4231 		if (err != 0)
4232 			break;
4233 		if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) {
4234 			err = EINVAL;
4235 			break;
4236 		}
4237 		if (i2c.len > sizeof(i2c.data)) {
4238 			err = EINVAL;
4239 			break;
4240 		}
4241 
4242 		if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0)
4243 			err = copyout(&i2c, ifr_data_get_ptr(ifr),
4244 			    sizeof(i2c));
4245 		break;
4246 	}
4247 	case SIOCSIFCAP:
4248 	{
4249 		int mask, setmask, oldmask;
4250 
4251 		oldmask = if_getcapenable(ifp);
4252 		mask = ifr->ifr_reqcap ^ oldmask;
4253 		mask &= ctx->ifc_softc_ctx.isc_capabilities;
4254 		setmask = 0;
4255 #ifdef TCP_OFFLOAD
4256 		setmask |= mask & (IFCAP_TOE4|IFCAP_TOE6);
4257 #endif
4258 		setmask |= (mask & IFCAP_FLAGS);
4259 		setmask |= (mask & IFCAP_WOL);
4260 
4261 		/*
4262 		 * If any RX csum has changed, change all the ones that
4263 		 * are supported by the driver.
4264 		 */
4265 		if (setmask & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6)) {
4266 			setmask |= ctx->ifc_softc_ctx.isc_capabilities &
4267 			    (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6);
4268 		}
4269 
4270 		/*
4271 		 * want to ensure that traffic has stopped before we change any of the flags
4272 		 */
4273 		if (setmask) {
4274 			CTX_LOCK(ctx);
4275 			bits = if_getdrvflags(ifp);
4276 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4277 				iflib_stop(ctx);
4278 			STATE_LOCK(ctx);
4279 			if_togglecapenable(ifp, setmask);
4280 			STATE_UNLOCK(ctx);
4281 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4282 				iflib_init_locked(ctx);
4283 			STATE_LOCK(ctx);
4284 			if_setdrvflags(ifp, bits);
4285 			STATE_UNLOCK(ctx);
4286 			CTX_UNLOCK(ctx);
4287 		}
4288 		if_vlancap(ifp);
4289 		break;
4290 	}
4291 	case SIOCGPRIVATE_0:
4292 	case SIOCSDRVSPEC:
4293 	case SIOCGDRVSPEC:
4294 		CTX_LOCK(ctx);
4295 		err = IFDI_PRIV_IOCTL(ctx, command, data);
4296 		CTX_UNLOCK(ctx);
4297 		break;
4298 	default:
4299 		err = ether_ioctl(ifp, command, data);
4300 		break;
4301 	}
4302 	if (reinit)
4303 		iflib_if_init(ctx);
4304 	return (err);
4305 }
4306 
4307 static uint64_t
4308 iflib_if_get_counter(if_t ifp, ift_counter cnt)
4309 {
4310 	if_ctx_t ctx = if_getsoftc(ifp);
4311 
4312 	return (IFDI_GET_COUNTER(ctx, cnt));
4313 }
4314 
4315 /*********************************************************************
4316  *
4317  *  OTHER FUNCTIONS EXPORTED TO THE STACK
4318  *
4319  **********************************************************************/
4320 
4321 static void
4322 iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag)
4323 {
4324 	if_ctx_t ctx = if_getsoftc(ifp);
4325 
4326 	if ((void *)ctx != arg)
4327 		return;
4328 
4329 	if ((vtag == 0) || (vtag > 4095))
4330 		return;
4331 
4332 	CTX_LOCK(ctx);
4333 	IFDI_VLAN_REGISTER(ctx, vtag);
4334 	/* Re-init to load the changes */
4335 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
4336 		iflib_if_init_locked(ctx);
4337 	CTX_UNLOCK(ctx);
4338 }
4339 
4340 static void
4341 iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag)
4342 {
4343 	if_ctx_t ctx = if_getsoftc(ifp);
4344 
4345 	if ((void *)ctx != arg)
4346 		return;
4347 
4348 	if ((vtag == 0) || (vtag > 4095))
4349 		return;
4350 
4351 	CTX_LOCK(ctx);
4352 	IFDI_VLAN_UNREGISTER(ctx, vtag);
4353 	/* Re-init to load the changes */
4354 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
4355 		iflib_if_init_locked(ctx);
4356 	CTX_UNLOCK(ctx);
4357 }
4358 
4359 static void
4360 iflib_led_func(void *arg, int onoff)
4361 {
4362 	if_ctx_t ctx = arg;
4363 
4364 	CTX_LOCK(ctx);
4365 	IFDI_LED_FUNC(ctx, onoff);
4366 	CTX_UNLOCK(ctx);
4367 }
4368 
4369 /*********************************************************************
4370  *
4371  *  BUS FUNCTION DEFINITIONS
4372  *
4373  **********************************************************************/
4374 
4375 int
4376 iflib_device_probe(device_t dev)
4377 {
4378 	pci_vendor_info_t *ent;
4379 
4380 	uint16_t	pci_vendor_id, pci_device_id;
4381 	uint16_t	pci_subvendor_id, pci_subdevice_id;
4382 	uint16_t	pci_rev_id;
4383 	if_shared_ctx_t sctx;
4384 
4385 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4386 		return (ENOTSUP);
4387 
4388 	pci_vendor_id = pci_get_vendor(dev);
4389 	pci_device_id = pci_get_device(dev);
4390 	pci_subvendor_id = pci_get_subvendor(dev);
4391 	pci_subdevice_id = pci_get_subdevice(dev);
4392 	pci_rev_id = pci_get_revid(dev);
4393 	if (sctx->isc_parse_devinfo != NULL)
4394 		sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id);
4395 
4396 	ent = sctx->isc_vendor_info;
4397 	while (ent->pvi_vendor_id != 0) {
4398 		if (pci_vendor_id != ent->pvi_vendor_id) {
4399 			ent++;
4400 			continue;
4401 		}
4402 		if ((pci_device_id == ent->pvi_device_id) &&
4403 		    ((pci_subvendor_id == ent->pvi_subvendor_id) ||
4404 		     (ent->pvi_subvendor_id == 0)) &&
4405 		    ((pci_subdevice_id == ent->pvi_subdevice_id) ||
4406 		     (ent->pvi_subdevice_id == 0)) &&
4407 		    ((pci_rev_id == ent->pvi_rev_id) ||
4408 		     (ent->pvi_rev_id == 0))) {
4409 
4410 			device_set_desc_copy(dev, ent->pvi_name);
4411 			/* this needs to be changed to zero if the bus probing code
4412 			 * ever stops re-probing on best match because the sctx
4413 			 * may have its values over written by register calls
4414 			 * in subsequent probes
4415 			 */
4416 			return (BUS_PROBE_DEFAULT);
4417 		}
4418 		ent++;
4419 	}
4420 	return (ENXIO);
4421 }
4422 
4423 static void
4424 iflib_reset_qvalues(if_ctx_t ctx)
4425 {
4426 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4427 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4428 	device_t dev = ctx->ifc_dev;
4429 	int i;
4430 
4431 	scctx->isc_txrx_budget_bytes_max = IFLIB_MAX_TX_BYTES;
4432 	scctx->isc_tx_qdepth = IFLIB_DEFAULT_TX_QDEPTH;
4433 	/*
4434 	 * XXX sanity check that ntxd & nrxd are a power of 2
4435 	 */
4436 	if (ctx->ifc_sysctl_ntxqs != 0)
4437 		scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs;
4438 	if (ctx->ifc_sysctl_nrxqs != 0)
4439 		scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs;
4440 
4441 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4442 		if (ctx->ifc_sysctl_ntxds[i] != 0)
4443 			scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i];
4444 		else
4445 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4446 	}
4447 
4448 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4449 		if (ctx->ifc_sysctl_nrxds[i] != 0)
4450 			scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i];
4451 		else
4452 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4453 	}
4454 
4455 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4456 		if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) {
4457 			device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n",
4458 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]);
4459 			scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i];
4460 		}
4461 		if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) {
4462 			device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n",
4463 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]);
4464 			scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i];
4465 		}
4466 	}
4467 
4468 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4469 		if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) {
4470 			device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n",
4471 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]);
4472 			scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i];
4473 		}
4474 		if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) {
4475 			device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n",
4476 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]);
4477 			scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i];
4478 		}
4479 	}
4480 }
4481 
4482 int
4483 iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp)
4484 {
4485 	int err, rid, msix;
4486 	if_ctx_t ctx;
4487 	if_t ifp;
4488 	if_softc_ctx_t scctx;
4489 	int i;
4490 	uint16_t main_txq;
4491 	uint16_t main_rxq;
4492 
4493 
4494 	ctx = malloc(sizeof(* ctx), M_IFLIB, M_WAITOK|M_ZERO);
4495 
4496 	if (sc == NULL) {
4497 		sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4498 		device_set_softc(dev, ctx);
4499 		ctx->ifc_flags |= IFC_SC_ALLOCATED;
4500 	}
4501 
4502 	ctx->ifc_sctx = sctx;
4503 	ctx->ifc_dev = dev;
4504 	ctx->ifc_softc = sc;
4505 
4506 	if ((err = iflib_register(ctx)) != 0) {
4507 		if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4508 			free(sc, M_IFLIB);
4509 		free(ctx, M_IFLIB);
4510 		device_printf(dev, "iflib_register failed %d\n", err);
4511 		return (err);
4512 	}
4513 	iflib_add_device_sysctl_pre(ctx);
4514 
4515 	scctx = &ctx->ifc_softc_ctx;
4516 	ifp = ctx->ifc_ifp;
4517 
4518 	iflib_reset_qvalues(ctx);
4519 	CTX_LOCK(ctx);
4520 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4521 		CTX_UNLOCK(ctx);
4522 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4523 		return (err);
4524 	}
4525 	_iflib_pre_assert(scctx);
4526 	ctx->ifc_txrx = *scctx->isc_txrx;
4527 
4528 #ifdef INVARIANTS
4529 	MPASS(scctx->isc_capabilities);
4530 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
4531 		MPASS(scctx->isc_tx_csum_flags);
4532 #endif
4533 
4534 	if_setcapabilities(ifp, scctx->isc_capabilities | IFCAP_HWSTATS);
4535 	if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS);
4536 
4537 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4538 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4539 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4540 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4541 
4542 #ifdef ACPI_DMAR
4543 	if (dmar_get_dma_tag(device_get_parent(dev), dev) != NULL)
4544 		ctx->ifc_flags |= IFC_DMAR;
4545 #elif !(defined(__i386__) || defined(__amd64__))
4546 	/* set unconditionally for !x86 */
4547 	ctx->ifc_flags |= IFC_DMAR;
4548 #endif
4549 
4550 	main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4551 	main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4552 
4553 	/* XXX change for per-queue sizes */
4554 	device_printf(dev, "using %d tx descriptors and %d rx descriptors\n",
4555 		      scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4556 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4557 		if (!powerof2(scctx->isc_nrxd[i])) {
4558 			/* round down instead? */
4559 			device_printf(dev, "# rx descriptors must be a power of 2\n");
4560 			err = EINVAL;
4561 			goto fail;
4562 		}
4563 	}
4564 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4565 		if (!powerof2(scctx->isc_ntxd[i])) {
4566 			device_printf(dev,
4567 			    "# tx descriptors must be a power of 2");
4568 			err = EINVAL;
4569 			goto fail;
4570 		}
4571 	}
4572 
4573 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4574 	    MAX_SINGLE_PACKET_FRACTION)
4575 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4576 		    MAX_SINGLE_PACKET_FRACTION);
4577 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4578 	    MAX_SINGLE_PACKET_FRACTION)
4579 		scctx->isc_tx_tso_segments_max = max(1,
4580 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4581 
4582 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4583 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
4584 		/*
4585 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
4586 		 * but some MACs do.
4587 		 */
4588 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
4589 		    IP_MAXPACKET));
4590 		/*
4591 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
4592 		 * into account.  In the worst case, each of these calls will
4593 		 * add another mbuf and, thus, the requirement for another DMA
4594 		 * segment.  So for best performance, it doesn't make sense to
4595 		 * advertize a maximum of TSO segments that typically will
4596 		 * require defragmentation in iflib_encap().
4597 		 */
4598 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
4599 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
4600 	}
4601 	if (scctx->isc_rss_table_size == 0)
4602 		scctx->isc_rss_table_size = 64;
4603 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4604 
4605 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4606 	/* XXX format name */
4607 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx, -1, "admin");
4608 
4609 	/* Set up cpu set.  If it fails, use the set of all CPUs. */
4610 	if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) != 0) {
4611 		device_printf(dev, "Unable to fetch CPU list\n");
4612 		CPU_COPY(&all_cpus, &ctx->ifc_cpus);
4613 	}
4614 	MPASS(CPU_COUNT(&ctx->ifc_cpus) > 0);
4615 
4616 	/*
4617 	** Now setup MSI or MSI/X, should
4618 	** return us the number of supported
4619 	** vectors. (Will be 1 for MSI)
4620 	*/
4621 	if (sctx->isc_flags & IFLIB_SKIP_MSIX) {
4622 		msix = scctx->isc_vectors;
4623 	} else if (scctx->isc_msix_bar != 0)
4624 	       /*
4625 		* The simple fact that isc_msix_bar is not 0 does not mean we
4626 		* we have a good value there that is known to work.
4627 		*/
4628 		msix = iflib_msix_init(ctx);
4629 	else {
4630 		scctx->isc_vectors = 1;
4631 		scctx->isc_ntxqsets = 1;
4632 		scctx->isc_nrxqsets = 1;
4633 		scctx->isc_intr = IFLIB_INTR_LEGACY;
4634 		msix = 0;
4635 	}
4636 	/* Get memory for the station queues */
4637 	if ((err = iflib_queues_alloc(ctx))) {
4638 		device_printf(dev, "Unable to allocate queue memory\n");
4639 		goto fail;
4640 	}
4641 
4642 	if ((err = iflib_qset_structures_setup(ctx)))
4643 		goto fail_queues;
4644 
4645 	/*
4646 	 * Group taskqueues aren't properly set up until SMP is started,
4647 	 * so we disable interrupts until we can handle them post
4648 	 * SI_SUB_SMP.
4649 	 *
4650 	 * XXX: disabling interrupts doesn't actually work, at least for
4651 	 * the non-MSI case.  When they occur before SI_SUB_SMP completes,
4652 	 * we do null handling and depend on this not causing too large an
4653 	 * interrupt storm.
4654 	 */
4655 	IFDI_INTR_DISABLE(ctx);
4656 	if (msix > 1 && (err = IFDI_MSIX_INTR_ASSIGN(ctx, msix)) != 0) {
4657 		device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n", err);
4658 		goto fail_intr_free;
4659 	}
4660 	if (msix <= 1) {
4661 		rid = 0;
4662 		if (scctx->isc_intr == IFLIB_INTR_MSI) {
4663 			MPASS(msix == 1);
4664 			rid = 1;
4665 		}
4666 		if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) {
4667 			device_printf(dev, "iflib_legacy_setup failed %d\n", err);
4668 			goto fail_intr_free;
4669 		}
4670 	}
4671 
4672 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac);
4673 
4674 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4675 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4676 		goto fail_detach;
4677 	}
4678 
4679 	/*
4680 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4681 	 * This must appear after the call to ether_ifattach() because
4682 	 * ether_ifattach() sets if_hdrlen to the default value.
4683 	 */
4684 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4685 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
4686 
4687 	if ((err = iflib_netmap_attach(ctx))) {
4688 		device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err);
4689 		goto fail_detach;
4690 	}
4691 	*ctxp = ctx;
4692 
4693 	NETDUMP_SET(ctx->ifc_ifp, iflib);
4694 
4695 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4696 	iflib_add_device_sysctl_post(ctx);
4697 	ctx->ifc_flags |= IFC_INIT_DONE;
4698 	CTX_UNLOCK(ctx);
4699 	return (0);
4700 
4701 fail_detach:
4702 	ether_ifdetach(ctx->ifc_ifp);
4703 fail_intr_free:
4704 fail_queues:
4705 	iflib_tx_structures_free(ctx);
4706 	iflib_rx_structures_free(ctx);
4707 fail:
4708 	iflib_free_intr_mem(ctx);
4709 	IFDI_DETACH(ctx);
4710 	CTX_UNLOCK(ctx);
4711 
4712 	return (err);
4713 }
4714 
4715 int
4716 iflib_pseudo_register(device_t dev, if_shared_ctx_t sctx, if_ctx_t *ctxp,
4717 					  struct iflib_cloneattach_ctx *clctx)
4718 {
4719 	int err;
4720 	if_ctx_t ctx;
4721 	if_t ifp;
4722 	if_softc_ctx_t scctx;
4723 	int i;
4724 	void *sc;
4725 	uint16_t main_txq;
4726 	uint16_t main_rxq;
4727 
4728 	ctx = malloc(sizeof(*ctx), M_IFLIB, M_WAITOK|M_ZERO);
4729 	sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4730 	ctx->ifc_flags |= IFC_SC_ALLOCATED;
4731 	if (sctx->isc_flags & (IFLIB_PSEUDO|IFLIB_VIRTUAL))
4732 		ctx->ifc_flags |= IFC_PSEUDO;
4733 
4734 	ctx->ifc_sctx = sctx;
4735 	ctx->ifc_softc = sc;
4736 	ctx->ifc_dev = dev;
4737 
4738 	if ((err = iflib_register(ctx)) != 0) {
4739 		device_printf(dev, "%s: iflib_register failed %d\n", __func__, err);
4740 		free(sc, M_IFLIB);
4741 		free(ctx, M_IFLIB);
4742 		return (err);
4743 	}
4744 	iflib_add_device_sysctl_pre(ctx);
4745 
4746 	scctx = &ctx->ifc_softc_ctx;
4747 	ifp = ctx->ifc_ifp;
4748 
4749 	/*
4750 	 * XXX sanity check that ntxd & nrxd are a power of 2
4751 	 */
4752 	iflib_reset_qvalues(ctx);
4753 
4754 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4755 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4756 		return (err);
4757 	}
4758 	if (sctx->isc_flags & IFLIB_GEN_MAC)
4759 		iflib_gen_mac(ctx);
4760 	if ((err = IFDI_CLONEATTACH(ctx, clctx->cc_ifc, clctx->cc_name,
4761 								clctx->cc_params)) != 0) {
4762 		device_printf(dev, "IFDI_CLONEATTACH failed %d\n", err);
4763 		return (err);
4764 	}
4765 	ifmedia_add(&ctx->ifc_media, IFM_ETHER | IFM_1000_T | IFM_FDX, 0, NULL);
4766 	ifmedia_add(&ctx->ifc_media, IFM_ETHER | IFM_AUTO, 0, NULL);
4767 	ifmedia_set(&ctx->ifc_media, IFM_ETHER | IFM_AUTO);
4768 
4769 #ifdef INVARIANTS
4770 	MPASS(scctx->isc_capabilities);
4771 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
4772 		MPASS(scctx->isc_tx_csum_flags);
4773 #endif
4774 
4775 	if_setcapabilities(ifp, scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_LINKSTATE);
4776 	if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_LINKSTATE);
4777 
4778 	ifp->if_flags |= IFF_NOGROUP;
4779 	if (sctx->isc_flags & IFLIB_PSEUDO) {
4780 		ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac);
4781 
4782 		if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4783 			device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4784 			goto fail_detach;
4785 		}
4786 		*ctxp = ctx;
4787 
4788 		/*
4789 		 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4790 		 * This must appear after the call to ether_ifattach() because
4791 		 * ether_ifattach() sets if_hdrlen to the default value.
4792 		 */
4793 		if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4794 			if_setifheaderlen(ifp,
4795 			    sizeof(struct ether_vlan_header));
4796 
4797 		if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4798 		iflib_add_device_sysctl_post(ctx);
4799 		ctx->ifc_flags |= IFC_INIT_DONE;
4800 		return (0);
4801 	}
4802 	_iflib_pre_assert(scctx);
4803 	ctx->ifc_txrx = *scctx->isc_txrx;
4804 
4805 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4806 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4807 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4808 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4809 
4810 	main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4811 	main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4812 
4813 	/* XXX change for per-queue sizes */
4814 	device_printf(dev, "using %d tx descriptors and %d rx descriptors\n",
4815 		      scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4816 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4817 		if (!powerof2(scctx->isc_nrxd[i])) {
4818 			/* round down instead? */
4819 			device_printf(dev, "# rx descriptors must be a power of 2\n");
4820 			err = EINVAL;
4821 			goto fail;
4822 		}
4823 	}
4824 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4825 		if (!powerof2(scctx->isc_ntxd[i])) {
4826 			device_printf(dev,
4827 			    "# tx descriptors must be a power of 2");
4828 			err = EINVAL;
4829 			goto fail;
4830 		}
4831 	}
4832 
4833 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4834 	    MAX_SINGLE_PACKET_FRACTION)
4835 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4836 		    MAX_SINGLE_PACKET_FRACTION);
4837 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4838 	    MAX_SINGLE_PACKET_FRACTION)
4839 		scctx->isc_tx_tso_segments_max = max(1,
4840 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4841 
4842 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4843 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
4844 		/*
4845 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
4846 		 * but some MACs do.
4847 		 */
4848 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
4849 		    IP_MAXPACKET));
4850 		/*
4851 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
4852 		 * into account.  In the worst case, each of these calls will
4853 		 * add another mbuf and, thus, the requirement for another DMA
4854 		 * segment.  So for best performance, it doesn't make sense to
4855 		 * advertize a maximum of TSO segments that typically will
4856 		 * require defragmentation in iflib_encap().
4857 		 */
4858 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
4859 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
4860 	}
4861 	if (scctx->isc_rss_table_size == 0)
4862 		scctx->isc_rss_table_size = 64;
4863 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4864 
4865 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4866 	/* XXX format name */
4867 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx, -1, "admin");
4868 
4869 	/* XXX --- can support > 1 -- but keep it simple for now */
4870 	scctx->isc_intr = IFLIB_INTR_LEGACY;
4871 
4872 	/* Get memory for the station queues */
4873 	if ((err = iflib_queues_alloc(ctx))) {
4874 		device_printf(dev, "Unable to allocate queue memory\n");
4875 		goto fail;
4876 	}
4877 
4878 	if ((err = iflib_qset_structures_setup(ctx))) {
4879 		device_printf(dev, "qset structure setup failed %d\n", err);
4880 		goto fail_queues;
4881 	}
4882 
4883 	/*
4884 	 * XXX What if anything do we want to do about interrupts?
4885 	 */
4886 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac);
4887 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4888 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4889 		goto fail_detach;
4890 	}
4891 
4892 	/*
4893 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4894 	 * This must appear after the call to ether_ifattach() because
4895 	 * ether_ifattach() sets if_hdrlen to the default value.
4896 	 */
4897 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4898 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
4899 
4900 	/* XXX handle more than one queue */
4901 	for (i = 0; i < scctx->isc_nrxqsets; i++)
4902 		IFDI_RX_CLSET(ctx, 0, i, ctx->ifc_rxqs[i].ifr_fl[0].ifl_sds.ifsd_cl);
4903 
4904 	*ctxp = ctx;
4905 
4906 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4907 	iflib_add_device_sysctl_post(ctx);
4908 	ctx->ifc_flags |= IFC_INIT_DONE;
4909 	return (0);
4910 fail_detach:
4911 	ether_ifdetach(ctx->ifc_ifp);
4912 fail_queues:
4913 	iflib_tx_structures_free(ctx);
4914 	iflib_rx_structures_free(ctx);
4915 fail:
4916 	IFDI_DETACH(ctx);
4917 	return (err);
4918 }
4919 
4920 int
4921 iflib_pseudo_deregister(if_ctx_t ctx)
4922 {
4923 	if_t ifp = ctx->ifc_ifp;
4924 	iflib_txq_t txq;
4925 	iflib_rxq_t rxq;
4926 	int i, j;
4927 	struct taskqgroup *tqg;
4928 	iflib_fl_t fl;
4929 
4930 	/* Unregister VLAN events */
4931 	if (ctx->ifc_vlan_attach_event != NULL)
4932 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
4933 	if (ctx->ifc_vlan_detach_event != NULL)
4934 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
4935 
4936 	ether_ifdetach(ifp);
4937 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
4938 	CTX_LOCK_DESTROY(ctx);
4939 	/* XXX drain any dependent tasks */
4940 	tqg = qgroup_if_io_tqg;
4941 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
4942 		callout_drain(&txq->ift_timer);
4943 		if (txq->ift_task.gt_uniq != NULL)
4944 			taskqgroup_detach(tqg, &txq->ift_task);
4945 	}
4946 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4947 		if (rxq->ifr_task.gt_uniq != NULL)
4948 			taskqgroup_detach(tqg, &rxq->ifr_task);
4949 
4950 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
4951 			free(fl->ifl_rx_bitmap, M_IFLIB);
4952 	}
4953 	tqg = qgroup_if_config_tqg;
4954 	if (ctx->ifc_admin_task.gt_uniq != NULL)
4955 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
4956 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
4957 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
4958 
4959 	if_free(ifp);
4960 
4961 	iflib_tx_structures_free(ctx);
4962 	iflib_rx_structures_free(ctx);
4963 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4964 		free(ctx->ifc_softc, M_IFLIB);
4965 	free(ctx, M_IFLIB);
4966 	return (0);
4967 }
4968 
4969 int
4970 iflib_device_attach(device_t dev)
4971 {
4972 	if_ctx_t ctx;
4973 	if_shared_ctx_t sctx;
4974 
4975 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4976 		return (ENOTSUP);
4977 
4978 	pci_enable_busmaster(dev);
4979 
4980 	return (iflib_device_register(dev, NULL, sctx, &ctx));
4981 }
4982 
4983 int
4984 iflib_device_deregister(if_ctx_t ctx)
4985 {
4986 	if_t ifp = ctx->ifc_ifp;
4987 	iflib_txq_t txq;
4988 	iflib_rxq_t rxq;
4989 	device_t dev = ctx->ifc_dev;
4990 	int i, j;
4991 	struct taskqgroup *tqg;
4992 	iflib_fl_t fl;
4993 
4994 	/* Make sure VLANS are not using driver */
4995 	if (if_vlantrunkinuse(ifp)) {
4996 		device_printf(dev, "Vlan in use, detach first\n");
4997 		return (EBUSY);
4998 	}
4999 #ifdef PCI_IOV
5000 	if (!CTX_IS_VF(ctx) && pci_iov_detach(dev) != 0) {
5001 		device_printf(dev, "SR-IOV in use; detach first.\n");
5002 		return (EBUSY);
5003 	}
5004 #endif
5005 
5006 	STATE_LOCK(ctx);
5007 	ctx->ifc_flags |= IFC_IN_DETACH;
5008 	STATE_UNLOCK(ctx);
5009 
5010 	CTX_LOCK(ctx);
5011 	iflib_stop(ctx);
5012 	CTX_UNLOCK(ctx);
5013 
5014 	/* Unregister VLAN events */
5015 	if (ctx->ifc_vlan_attach_event != NULL)
5016 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
5017 	if (ctx->ifc_vlan_detach_event != NULL)
5018 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
5019 
5020 	iflib_netmap_detach(ifp);
5021 	ether_ifdetach(ifp);
5022 	if (ctx->ifc_led_dev != NULL)
5023 		led_destroy(ctx->ifc_led_dev);
5024 	/* XXX drain any dependent tasks */
5025 	tqg = qgroup_if_io_tqg;
5026 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
5027 		callout_drain(&txq->ift_timer);
5028 		if (txq->ift_task.gt_uniq != NULL)
5029 			taskqgroup_detach(tqg, &txq->ift_task);
5030 	}
5031 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
5032 		if (rxq->ifr_task.gt_uniq != NULL)
5033 			taskqgroup_detach(tqg, &rxq->ifr_task);
5034 
5035 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5036 			free(fl->ifl_rx_bitmap, M_IFLIB);
5037 
5038 	}
5039 	tqg = qgroup_if_config_tqg;
5040 	if (ctx->ifc_admin_task.gt_uniq != NULL)
5041 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
5042 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
5043 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
5044 	CTX_LOCK(ctx);
5045 	IFDI_DETACH(ctx);
5046 	CTX_UNLOCK(ctx);
5047 
5048 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
5049 	CTX_LOCK_DESTROY(ctx);
5050 	device_set_softc(ctx->ifc_dev, NULL);
5051 	iflib_free_intr_mem(ctx);
5052 
5053 	bus_generic_detach(dev);
5054 	if_free(ifp);
5055 
5056 	iflib_tx_structures_free(ctx);
5057 	iflib_rx_structures_free(ctx);
5058 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5059 		free(ctx->ifc_softc, M_IFLIB);
5060 	STATE_LOCK_DESTROY(ctx);
5061 	free(ctx, M_IFLIB);
5062 	return (0);
5063 }
5064 
5065 static void
5066 iflib_free_intr_mem(if_ctx_t ctx)
5067 {
5068 
5069 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) {
5070 		pci_release_msi(ctx->ifc_dev);
5071 	}
5072 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) {
5073 		iflib_irq_free(ctx, &ctx->ifc_legacy_irq);
5074 	}
5075 	if (ctx->ifc_msix_mem != NULL) {
5076 		bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY,
5077 			ctx->ifc_softc_ctx.isc_msix_bar, ctx->ifc_msix_mem);
5078 		ctx->ifc_msix_mem = NULL;
5079 	}
5080 }
5081 
5082 int
5083 iflib_device_detach(device_t dev)
5084 {
5085 	if_ctx_t ctx = device_get_softc(dev);
5086 
5087 	return (iflib_device_deregister(ctx));
5088 }
5089 
5090 int
5091 iflib_device_suspend(device_t dev)
5092 {
5093 	if_ctx_t ctx = device_get_softc(dev);
5094 
5095 	CTX_LOCK(ctx);
5096 	IFDI_SUSPEND(ctx);
5097 	CTX_UNLOCK(ctx);
5098 
5099 	return bus_generic_suspend(dev);
5100 }
5101 int
5102 iflib_device_shutdown(device_t dev)
5103 {
5104 	if_ctx_t ctx = device_get_softc(dev);
5105 
5106 	CTX_LOCK(ctx);
5107 	IFDI_SHUTDOWN(ctx);
5108 	CTX_UNLOCK(ctx);
5109 
5110 	return bus_generic_suspend(dev);
5111 }
5112 
5113 
5114 int
5115 iflib_device_resume(device_t dev)
5116 {
5117 	if_ctx_t ctx = device_get_softc(dev);
5118 	iflib_txq_t txq = ctx->ifc_txqs;
5119 
5120 	CTX_LOCK(ctx);
5121 	IFDI_RESUME(ctx);
5122 	iflib_init_locked(ctx);
5123 	CTX_UNLOCK(ctx);
5124 	for (int i = 0; i < NTXQSETS(ctx); i++, txq++)
5125 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
5126 
5127 	return (bus_generic_resume(dev));
5128 }
5129 
5130 int
5131 iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params)
5132 {
5133 	int error;
5134 	if_ctx_t ctx = device_get_softc(dev);
5135 
5136 	CTX_LOCK(ctx);
5137 	error = IFDI_IOV_INIT(ctx, num_vfs, params);
5138 	CTX_UNLOCK(ctx);
5139 
5140 	return (error);
5141 }
5142 
5143 void
5144 iflib_device_iov_uninit(device_t dev)
5145 {
5146 	if_ctx_t ctx = device_get_softc(dev);
5147 
5148 	CTX_LOCK(ctx);
5149 	IFDI_IOV_UNINIT(ctx);
5150 	CTX_UNLOCK(ctx);
5151 }
5152 
5153 int
5154 iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params)
5155 {
5156 	int error;
5157 	if_ctx_t ctx = device_get_softc(dev);
5158 
5159 	CTX_LOCK(ctx);
5160 	error = IFDI_IOV_VF_ADD(ctx, vfnum, params);
5161 	CTX_UNLOCK(ctx);
5162 
5163 	return (error);
5164 }
5165 
5166 /*********************************************************************
5167  *
5168  *  MODULE FUNCTION DEFINITIONS
5169  *
5170  **********************************************************************/
5171 
5172 /*
5173  * - Start a fast taskqueue thread for each core
5174  * - Start a taskqueue for control operations
5175  */
5176 static int
5177 iflib_module_init(void)
5178 {
5179 	return (0);
5180 }
5181 
5182 static int
5183 iflib_module_event_handler(module_t mod, int what, void *arg)
5184 {
5185 	int err;
5186 
5187 	switch (what) {
5188 	case MOD_LOAD:
5189 		if ((err = iflib_module_init()) != 0)
5190 			return (err);
5191 		break;
5192 	case MOD_UNLOAD:
5193 		return (EBUSY);
5194 	default:
5195 		return (EOPNOTSUPP);
5196 	}
5197 
5198 	return (0);
5199 }
5200 
5201 /*********************************************************************
5202  *
5203  *  PUBLIC FUNCTION DEFINITIONS
5204  *     ordered as in iflib.h
5205  *
5206  **********************************************************************/
5207 
5208 
5209 static void
5210 _iflib_assert(if_shared_ctx_t sctx)
5211 {
5212 	MPASS(sctx->isc_tx_maxsize);
5213 	MPASS(sctx->isc_tx_maxsegsize);
5214 
5215 	MPASS(sctx->isc_rx_maxsize);
5216 	MPASS(sctx->isc_rx_nsegments);
5217 	MPASS(sctx->isc_rx_maxsegsize);
5218 
5219 	MPASS(sctx->isc_nrxd_min[0]);
5220 	MPASS(sctx->isc_nrxd_max[0]);
5221 	MPASS(sctx->isc_nrxd_default[0]);
5222 	MPASS(sctx->isc_ntxd_min[0]);
5223 	MPASS(sctx->isc_ntxd_max[0]);
5224 	MPASS(sctx->isc_ntxd_default[0]);
5225 }
5226 
5227 static void
5228 _iflib_pre_assert(if_softc_ctx_t scctx)
5229 {
5230 
5231 	MPASS(scctx->isc_txrx->ift_txd_encap);
5232 	MPASS(scctx->isc_txrx->ift_txd_flush);
5233 	MPASS(scctx->isc_txrx->ift_txd_credits_update);
5234 	MPASS(scctx->isc_txrx->ift_rxd_available);
5235 	MPASS(scctx->isc_txrx->ift_rxd_pkt_get);
5236 	MPASS(scctx->isc_txrx->ift_rxd_refill);
5237 	MPASS(scctx->isc_txrx->ift_rxd_flush);
5238 }
5239 
5240 static int
5241 iflib_register(if_ctx_t ctx)
5242 {
5243 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5244 	driver_t *driver = sctx->isc_driver;
5245 	device_t dev = ctx->ifc_dev;
5246 	if_t ifp;
5247 
5248 	_iflib_assert(sctx);
5249 
5250 	CTX_LOCK_INIT(ctx);
5251 	STATE_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev));
5252 	ifp = ctx->ifc_ifp = if_alloc(IFT_ETHER);
5253 	if (ifp == NULL) {
5254 		device_printf(dev, "can not allocate ifnet structure\n");
5255 		return (ENOMEM);
5256 	}
5257 
5258 	/*
5259 	 * Initialize our context's device specific methods
5260 	 */
5261 	kobj_init((kobj_t) ctx, (kobj_class_t) driver);
5262 	kobj_class_compile((kobj_class_t) driver);
5263 	driver->refs++;
5264 
5265 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
5266 	if_setsoftc(ifp, ctx);
5267 	if_setdev(ifp, dev);
5268 	if_setinitfn(ifp, iflib_if_init);
5269 	if_setioctlfn(ifp, iflib_if_ioctl);
5270 #ifdef ALTQ
5271 	if_setstartfn(ifp, iflib_altq_if_start);
5272 	if_settransmitfn(ifp, iflib_altq_if_transmit);
5273 	if_setsendqready(ifp);
5274 #else
5275 	if_settransmitfn(ifp, iflib_if_transmit);
5276 #endif
5277 	if_setqflushfn(ifp, iflib_if_qflush);
5278 	if_setflags(ifp, IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST);
5279 
5280 	ctx->ifc_vlan_attach_event =
5281 		EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx,
5282 							  EVENTHANDLER_PRI_FIRST);
5283 	ctx->ifc_vlan_detach_event =
5284 		EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx,
5285 							  EVENTHANDLER_PRI_FIRST);
5286 
5287 	ifmedia_init(&ctx->ifc_media, IFM_IMASK,
5288 					 iflib_media_change, iflib_media_status);
5289 
5290 	return (0);
5291 }
5292 
5293 
5294 static int
5295 iflib_queues_alloc(if_ctx_t ctx)
5296 {
5297 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5298 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5299 	device_t dev = ctx->ifc_dev;
5300 	int nrxqsets = scctx->isc_nrxqsets;
5301 	int ntxqsets = scctx->isc_ntxqsets;
5302 	iflib_txq_t txq;
5303 	iflib_rxq_t rxq;
5304 	iflib_fl_t fl = NULL;
5305 	int i, j, cpu, err, txconf, rxconf;
5306 	iflib_dma_info_t ifdip;
5307 	uint32_t *rxqsizes = scctx->isc_rxqsizes;
5308 	uint32_t *txqsizes = scctx->isc_txqsizes;
5309 	uint8_t nrxqs = sctx->isc_nrxqs;
5310 	uint8_t ntxqs = sctx->isc_ntxqs;
5311 	int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1;
5312 	caddr_t *vaddrs;
5313 	uint64_t *paddrs;
5314 
5315 	KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1"));
5316 	KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1"));
5317 
5318 	/* Allocate the TX ring struct memory */
5319 	if (!(ctx->ifc_txqs =
5320 	    (iflib_txq_t) malloc(sizeof(struct iflib_txq) *
5321 	    ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5322 		device_printf(dev, "Unable to allocate TX ring memory\n");
5323 		err = ENOMEM;
5324 		goto fail;
5325 	}
5326 
5327 	/* Now allocate the RX */
5328 	if (!(ctx->ifc_rxqs =
5329 	    (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) *
5330 	    nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5331 		device_printf(dev, "Unable to allocate RX ring memory\n");
5332 		err = ENOMEM;
5333 		goto rx_fail;
5334 	}
5335 
5336 	txq = ctx->ifc_txqs;
5337 	rxq = ctx->ifc_rxqs;
5338 
5339 	/*
5340 	 * XXX handle allocation failure
5341 	 */
5342 	for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) {
5343 		/* Set up some basics */
5344 
5345 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs, M_IFLIB, M_WAITOK|M_ZERO)) == NULL) {
5346 			device_printf(dev, "failed to allocate iflib_dma_info\n");
5347 			err = ENOMEM;
5348 			goto err_tx_desc;
5349 		}
5350 		txq->ift_ifdi = ifdip;
5351 		for (j = 0; j < ntxqs; j++, ifdip++) {
5352 			if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, BUS_DMA_NOWAIT)) {
5353 				device_printf(dev, "Unable to allocate Descriptor memory\n");
5354 				err = ENOMEM;
5355 				goto err_tx_desc;
5356 			}
5357 			txq->ift_txd_size[j] = scctx->isc_txd_size[j];
5358 			bzero((void *)ifdip->idi_vaddr, txqsizes[j]);
5359 		}
5360 		txq->ift_ctx = ctx;
5361 		txq->ift_id = i;
5362 		if (sctx->isc_flags & IFLIB_HAS_TXCQ) {
5363 			txq->ift_br_offset = 1;
5364 		} else {
5365 			txq->ift_br_offset = 0;
5366 		}
5367 		/* XXX fix this */
5368 		txq->ift_timer.c_cpu = cpu;
5369 
5370 		if (iflib_txsd_alloc(txq)) {
5371 			device_printf(dev, "Critical Failure setting up TX buffers\n");
5372 			err = ENOMEM;
5373 			goto err_tx_desc;
5374 		}
5375 
5376 		/* Initialize the TX lock */
5377 		snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:tx(%d):callout",
5378 		    device_get_nameunit(dev), txq->ift_id);
5379 		mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF);
5380 		callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0);
5381 
5382 		snprintf(txq->ift_db_mtx_name, MTX_NAME_LEN, "%s:tx(%d):db",
5383 			 device_get_nameunit(dev), txq->ift_id);
5384 
5385 		err = ifmp_ring_alloc(&txq->ift_br, 2048, txq, iflib_txq_drain,
5386 				      iflib_txq_can_drain, M_IFLIB, M_WAITOK);
5387 		if (err) {
5388 			/* XXX free any allocated rings */
5389 			device_printf(dev, "Unable to allocate buf_ring\n");
5390 			goto err_tx_desc;
5391 		}
5392 	}
5393 
5394 	for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) {
5395 		/* Set up some basics */
5396 
5397 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs, M_IFLIB, M_WAITOK|M_ZERO)) == NULL) {
5398 			device_printf(dev, "failed to allocate iflib_dma_info\n");
5399 			err = ENOMEM;
5400 			goto err_tx_desc;
5401 		}
5402 
5403 		rxq->ifr_ifdi = ifdip;
5404 		/* XXX this needs to be changed if #rx queues != #tx queues */
5405 		rxq->ifr_ntxqirq = 1;
5406 		rxq->ifr_txqid[0] = i;
5407 		for (j = 0; j < nrxqs; j++, ifdip++) {
5408 			if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, BUS_DMA_NOWAIT)) {
5409 				device_printf(dev, "Unable to allocate Descriptor memory\n");
5410 				err = ENOMEM;
5411 				goto err_tx_desc;
5412 			}
5413 			bzero((void *)ifdip->idi_vaddr, rxqsizes[j]);
5414 		}
5415 		rxq->ifr_ctx = ctx;
5416 		rxq->ifr_id = i;
5417 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
5418 			rxq->ifr_fl_offset = 1;
5419 		} else {
5420 			rxq->ifr_fl_offset = 0;
5421 		}
5422 		rxq->ifr_nfl = nfree_lists;
5423 		if (!(fl =
5424 			  (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) {
5425 			device_printf(dev, "Unable to allocate free list memory\n");
5426 			err = ENOMEM;
5427 			goto err_tx_desc;
5428 		}
5429 		rxq->ifr_fl = fl;
5430 		for (j = 0; j < nfree_lists; j++) {
5431 			fl[j].ifl_rxq = rxq;
5432 			fl[j].ifl_id = j;
5433 			fl[j].ifl_ifdi = &rxq->ifr_ifdi[j + rxq->ifr_fl_offset];
5434 			fl[j].ifl_rxd_size = scctx->isc_rxd_size[j];
5435 		}
5436 		/* Allocate receive buffers for the ring */
5437 		if (iflib_rxsd_alloc(rxq)) {
5438 			device_printf(dev,
5439 			    "Critical Failure setting up receive buffers\n");
5440 			err = ENOMEM;
5441 			goto err_rx_desc;
5442 		}
5443 
5444 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5445 			fl->ifl_rx_bitmap = bit_alloc(fl->ifl_size, M_IFLIB, M_WAITOK|M_ZERO);
5446 	}
5447 
5448 	/* TXQs */
5449 	vaddrs = malloc(sizeof(caddr_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5450 	paddrs = malloc(sizeof(uint64_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5451 	for (i = 0; i < ntxqsets; i++) {
5452 		iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi;
5453 
5454 		for (j = 0; j < ntxqs; j++, di++) {
5455 			vaddrs[i*ntxqs + j] = di->idi_vaddr;
5456 			paddrs[i*ntxqs + j] = di->idi_paddr;
5457 		}
5458 	}
5459 	if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) {
5460 		device_printf(ctx->ifc_dev, "device queue allocation failed\n");
5461 		iflib_tx_structures_free(ctx);
5462 		free(vaddrs, M_IFLIB);
5463 		free(paddrs, M_IFLIB);
5464 		goto err_rx_desc;
5465 	}
5466 	free(vaddrs, M_IFLIB);
5467 	free(paddrs, M_IFLIB);
5468 
5469 	/* RXQs */
5470 	vaddrs = malloc(sizeof(caddr_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5471 	paddrs = malloc(sizeof(uint64_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5472 	for (i = 0; i < nrxqsets; i++) {
5473 		iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi;
5474 
5475 		for (j = 0; j < nrxqs; j++, di++) {
5476 			vaddrs[i*nrxqs + j] = di->idi_vaddr;
5477 			paddrs[i*nrxqs + j] = di->idi_paddr;
5478 		}
5479 	}
5480 	if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) {
5481 		device_printf(ctx->ifc_dev, "device queue allocation failed\n");
5482 		iflib_tx_structures_free(ctx);
5483 		free(vaddrs, M_IFLIB);
5484 		free(paddrs, M_IFLIB);
5485 		goto err_rx_desc;
5486 	}
5487 	free(vaddrs, M_IFLIB);
5488 	free(paddrs, M_IFLIB);
5489 
5490 	return (0);
5491 
5492 /* XXX handle allocation failure changes */
5493 err_rx_desc:
5494 err_tx_desc:
5495 rx_fail:
5496 	if (ctx->ifc_rxqs != NULL)
5497 		free(ctx->ifc_rxqs, M_IFLIB);
5498 	ctx->ifc_rxqs = NULL;
5499 	if (ctx->ifc_txqs != NULL)
5500 		free(ctx->ifc_txqs, M_IFLIB);
5501 	ctx->ifc_txqs = NULL;
5502 fail:
5503 	return (err);
5504 }
5505 
5506 static int
5507 iflib_tx_structures_setup(if_ctx_t ctx)
5508 {
5509 	iflib_txq_t txq = ctx->ifc_txqs;
5510 	int i;
5511 
5512 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
5513 		iflib_txq_setup(txq);
5514 
5515 	return (0);
5516 }
5517 
5518 static void
5519 iflib_tx_structures_free(if_ctx_t ctx)
5520 {
5521 	iflib_txq_t txq = ctx->ifc_txqs;
5522 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5523 	int i, j;
5524 
5525 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
5526 		iflib_txq_destroy(txq);
5527 		for (j = 0; j < sctx->isc_ntxqs; j++)
5528 			iflib_dma_free(&txq->ift_ifdi[j]);
5529 	}
5530 	free(ctx->ifc_txqs, M_IFLIB);
5531 	ctx->ifc_txqs = NULL;
5532 	IFDI_QUEUES_FREE(ctx);
5533 }
5534 
5535 /*********************************************************************
5536  *
5537  *  Initialize all receive rings.
5538  *
5539  **********************************************************************/
5540 static int
5541 iflib_rx_structures_setup(if_ctx_t ctx)
5542 {
5543 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5544 	int q;
5545 #if defined(INET6) || defined(INET)
5546 	int i, err;
5547 #endif
5548 
5549 	for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) {
5550 #if defined(INET6) || defined(INET)
5551 		tcp_lro_free(&rxq->ifr_lc);
5552 		if ((err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp,
5553 		    TCP_LRO_ENTRIES, min(1024,
5554 		    ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset]))) != 0) {
5555 			device_printf(ctx->ifc_dev, "LRO Initialization failed!\n");
5556 			goto fail;
5557 		}
5558 		rxq->ifr_lro_enabled = TRUE;
5559 #endif
5560 		IFDI_RXQ_SETUP(ctx, rxq->ifr_id);
5561 	}
5562 	return (0);
5563 #if defined(INET6) || defined(INET)
5564 fail:
5565 	/*
5566 	 * Free RX software descriptors allocated so far, we will only handle
5567 	 * the rings that completed, the failing case will have
5568 	 * cleaned up for itself. 'q' failed, so its the terminus.
5569 	 */
5570 	rxq = ctx->ifc_rxqs;
5571 	for (i = 0; i < q; ++i, rxq++) {
5572 		iflib_rx_sds_free(rxq);
5573 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
5574 	}
5575 	return (err);
5576 #endif
5577 }
5578 
5579 /*********************************************************************
5580  *
5581  *  Free all receive rings.
5582  *
5583  **********************************************************************/
5584 static void
5585 iflib_rx_structures_free(if_ctx_t ctx)
5586 {
5587 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5588 
5589 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) {
5590 		iflib_rx_sds_free(rxq);
5591 	}
5592 	free(ctx->ifc_rxqs, M_IFLIB);
5593 	ctx->ifc_rxqs = NULL;
5594 }
5595 
5596 static int
5597 iflib_qset_structures_setup(if_ctx_t ctx)
5598 {
5599 	int err;
5600 
5601 	/*
5602 	 * It is expected that the caller takes care of freeing queues if this
5603 	 * fails.
5604 	 */
5605 	if ((err = iflib_tx_structures_setup(ctx)) != 0) {
5606 		device_printf(ctx->ifc_dev, "iflib_tx_structures_setup failed: %d\n", err);
5607 		return (err);
5608 	}
5609 
5610 	if ((err = iflib_rx_structures_setup(ctx)) != 0)
5611 		device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err);
5612 
5613 	return (err);
5614 }
5615 
5616 int
5617 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
5618 		driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, const char *name)
5619 {
5620 
5621 	return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name));
5622 }
5623 
5624 #ifdef SMP
5625 static int
5626 find_nth(if_ctx_t ctx, int qid)
5627 {
5628 	cpuset_t cpus;
5629 	int i, cpuid, eqid, count;
5630 
5631 	CPU_COPY(&ctx->ifc_cpus, &cpus);
5632 	count = CPU_COUNT(&cpus);
5633 	eqid = qid % count;
5634 	/* clear up to the qid'th bit */
5635 	for (i = 0; i < eqid; i++) {
5636 		cpuid = CPU_FFS(&cpus);
5637 		MPASS(cpuid != 0);
5638 		CPU_CLR(cpuid-1, &cpus);
5639 	}
5640 	cpuid = CPU_FFS(&cpus);
5641 	MPASS(cpuid != 0);
5642 	return (cpuid-1);
5643 }
5644 
5645 #ifdef SCHED_ULE
5646 extern struct cpu_group *cpu_top;              /* CPU topology */
5647 
5648 static int
5649 find_child_with_core(int cpu, struct cpu_group *grp)
5650 {
5651 	int i;
5652 
5653 	if (grp->cg_children == 0)
5654 		return -1;
5655 
5656 	MPASS(grp->cg_child);
5657 	for (i = 0; i < grp->cg_children; i++) {
5658 		if (CPU_ISSET(cpu, &grp->cg_child[i].cg_mask))
5659 			return i;
5660 	}
5661 
5662 	return -1;
5663 }
5664 
5665 /*
5666  * Find the nth "close" core to the specified core
5667  * "close" is defined as the deepest level that shares
5668  * at least an L2 cache.  With threads, this will be
5669  * threads on the same core.  If the sahred cache is L3
5670  * or higher, simply returns the same core.
5671  */
5672 static int
5673 find_close_core(int cpu, int core_offset)
5674 {
5675 	struct cpu_group *grp;
5676 	int i;
5677 	int fcpu;
5678 	cpuset_t cs;
5679 
5680 	grp = cpu_top;
5681 	if (grp == NULL)
5682 		return cpu;
5683 	i = 0;
5684 	while ((i = find_child_with_core(cpu, grp)) != -1) {
5685 		/* If the child only has one cpu, don't descend */
5686 		if (grp->cg_child[i].cg_count <= 1)
5687 			break;
5688 		grp = &grp->cg_child[i];
5689 	}
5690 
5691 	/* If they don't share at least an L2 cache, use the same CPU */
5692 	if (grp->cg_level > CG_SHARE_L2 || grp->cg_level == CG_SHARE_NONE)
5693 		return cpu;
5694 
5695 	/* Now pick one */
5696 	CPU_COPY(&grp->cg_mask, &cs);
5697 
5698 	/* Add the selected CPU offset to core offset. */
5699 	for (i = 0; (fcpu = CPU_FFS(&cs)) != 0; i++) {
5700 		if (fcpu - 1 == cpu)
5701 			break;
5702 		CPU_CLR(fcpu - 1, &cs);
5703 	}
5704 	MPASS(fcpu);
5705 
5706 	core_offset += i;
5707 
5708 	CPU_COPY(&grp->cg_mask, &cs);
5709 	for (i = core_offset % grp->cg_count; i > 0; i--) {
5710 		MPASS(CPU_FFS(&cs));
5711 		CPU_CLR(CPU_FFS(&cs) - 1, &cs);
5712 	}
5713 	MPASS(CPU_FFS(&cs));
5714 	return CPU_FFS(&cs) - 1;
5715 }
5716 #else
5717 static int
5718 find_close_core(int cpu, int core_offset __unused)
5719 {
5720 	return cpu;
5721 }
5722 #endif
5723 
5724 static int
5725 get_core_offset(if_ctx_t ctx, iflib_intr_type_t type, int qid)
5726 {
5727 	switch (type) {
5728 	case IFLIB_INTR_TX:
5729 		/* TX queues get cores which share at least an L2 cache with the corresponding RX queue */
5730 		/* XXX handle multiple RX threads per core and more than two core per L2 group */
5731 		return qid / CPU_COUNT(&ctx->ifc_cpus) + 1;
5732 	case IFLIB_INTR_RX:
5733 	case IFLIB_INTR_RXTX:
5734 		/* RX queues get the specified core */
5735 		return qid / CPU_COUNT(&ctx->ifc_cpus);
5736 	default:
5737 		return -1;
5738 	}
5739 }
5740 #else
5741 #define get_core_offset(ctx, type, qid)	CPU_FIRST()
5742 #define find_close_core(cpuid, tid)	CPU_FIRST()
5743 #define find_nth(ctx, gid)		CPU_FIRST()
5744 #endif
5745 
5746 /* Just to avoid copy/paste */
5747 static inline int
5748 iflib_irq_set_affinity(if_ctx_t ctx, int irq, iflib_intr_type_t type, int qid,
5749     struct grouptask *gtask, struct taskqgroup *tqg, void *uniq, const char *name)
5750 {
5751 	int cpuid;
5752 	int err, tid;
5753 
5754 	cpuid = find_nth(ctx, qid);
5755 	tid = get_core_offset(ctx, type, qid);
5756 	MPASS(tid >= 0);
5757 	cpuid = find_close_core(cpuid, tid);
5758 	err = taskqgroup_attach_cpu(tqg, gtask, uniq, cpuid, irq, name);
5759 	if (err) {
5760 		device_printf(ctx->ifc_dev, "taskqgroup_attach_cpu failed %d\n", err);
5761 		return (err);
5762 	}
5763 #ifdef notyet
5764 	if (cpuid > ctx->ifc_cpuid_highest)
5765 		ctx->ifc_cpuid_highest = cpuid;
5766 #endif
5767 	return 0;
5768 }
5769 
5770 int
5771 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid,
5772 			iflib_intr_type_t type, driver_filter_t *filter,
5773 			void *filter_arg, int qid, const char *name)
5774 {
5775 	struct grouptask *gtask;
5776 	struct taskqgroup *tqg;
5777 	iflib_filter_info_t info;
5778 	gtask_fn_t *fn;
5779 	int tqrid, err;
5780 	driver_filter_t *intr_fast;
5781 	void *q;
5782 
5783 	info = &ctx->ifc_filter_info;
5784 	tqrid = rid;
5785 
5786 	switch (type) {
5787 	/* XXX merge tx/rx for netmap? */
5788 	case IFLIB_INTR_TX:
5789 		q = &ctx->ifc_txqs[qid];
5790 		info = &ctx->ifc_txqs[qid].ift_filter_info;
5791 		gtask = &ctx->ifc_txqs[qid].ift_task;
5792 		tqg = qgroup_if_io_tqg;
5793 		fn = _task_fn_tx;
5794 		intr_fast = iflib_fast_intr;
5795 		GROUPTASK_INIT(gtask, 0, fn, q);
5796 		ctx->ifc_flags |= IFC_NETMAP_TX_IRQ;
5797 		break;
5798 	case IFLIB_INTR_RX:
5799 		q = &ctx->ifc_rxqs[qid];
5800 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
5801 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5802 		tqg = qgroup_if_io_tqg;
5803 		fn = _task_fn_rx;
5804 		intr_fast = iflib_fast_intr;
5805 		GROUPTASK_INIT(gtask, 0, fn, q);
5806 		break;
5807 	case IFLIB_INTR_RXTX:
5808 		q = &ctx->ifc_rxqs[qid];
5809 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
5810 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5811 		tqg = qgroup_if_io_tqg;
5812 		fn = _task_fn_rx;
5813 		intr_fast = iflib_fast_intr_rxtx;
5814 		GROUPTASK_INIT(gtask, 0, fn, q);
5815 		break;
5816 	case IFLIB_INTR_ADMIN:
5817 		q = ctx;
5818 		tqrid = -1;
5819 		info = &ctx->ifc_filter_info;
5820 		gtask = &ctx->ifc_admin_task;
5821 		tqg = qgroup_if_config_tqg;
5822 		fn = _task_fn_admin;
5823 		intr_fast = iflib_fast_intr_ctx;
5824 		break;
5825 	default:
5826 		panic("unknown net intr type");
5827 	}
5828 
5829 	info->ifi_filter = filter;
5830 	info->ifi_filter_arg = filter_arg;
5831 	info->ifi_task = gtask;
5832 	info->ifi_ctx = q;
5833 
5834 	err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info,  name);
5835 	if (err != 0) {
5836 		device_printf(ctx->ifc_dev, "_iflib_irq_alloc failed %d\n", err);
5837 		return (err);
5838 	}
5839 	if (type == IFLIB_INTR_ADMIN)
5840 		return (0);
5841 
5842 	if (tqrid != -1) {
5843 		err = iflib_irq_set_affinity(ctx, rman_get_start(irq->ii_res), type, qid, gtask, tqg, q, name);
5844 		if (err)
5845 			return (err);
5846 	} else {
5847 		taskqgroup_attach(tqg, gtask, q, rman_get_start(irq->ii_res), name);
5848 	}
5849 
5850 	return (0);
5851 }
5852 
5853 void
5854 iflib_softirq_alloc_generic(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type, void *arg, int qid, const char *name)
5855 {
5856 	struct grouptask *gtask;
5857 	struct taskqgroup *tqg;
5858 	gtask_fn_t *fn;
5859 	void *q;
5860 	int irq_num = -1;
5861 	int err;
5862 
5863 	switch (type) {
5864 	case IFLIB_INTR_TX:
5865 		q = &ctx->ifc_txqs[qid];
5866 		gtask = &ctx->ifc_txqs[qid].ift_task;
5867 		tqg = qgroup_if_io_tqg;
5868 		fn = _task_fn_tx;
5869 		if (irq != NULL)
5870 			irq_num = rman_get_start(irq->ii_res);
5871 		break;
5872 	case IFLIB_INTR_RX:
5873 		q = &ctx->ifc_rxqs[qid];
5874 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5875 		tqg = qgroup_if_io_tqg;
5876 		fn = _task_fn_rx;
5877 		if (irq != NULL)
5878 			irq_num = rman_get_start(irq->ii_res);
5879 		break;
5880 	case IFLIB_INTR_IOV:
5881 		q = ctx;
5882 		gtask = &ctx->ifc_vflr_task;
5883 		tqg = qgroup_if_config_tqg;
5884 		fn = _task_fn_iov;
5885 		break;
5886 	default:
5887 		panic("unknown net intr type");
5888 	}
5889 	GROUPTASK_INIT(gtask, 0, fn, q);
5890 	if (irq_num != -1) {
5891 		err = iflib_irq_set_affinity(ctx, irq_num, type, qid, gtask, tqg, q, name);
5892 		if (err)
5893 			taskqgroup_attach(tqg, gtask, q, irq_num, name);
5894 	}
5895 	else {
5896 		taskqgroup_attach(tqg, gtask, q, irq_num, name);
5897 	}
5898 }
5899 
5900 void
5901 iflib_irq_free(if_ctx_t ctx, if_irq_t irq)
5902 {
5903 	if (irq->ii_tag)
5904 		bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag);
5905 
5906 	if (irq->ii_res)
5907 		bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ, irq->ii_rid, irq->ii_res);
5908 }
5909 
5910 static int
5911 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, const char *name)
5912 {
5913 	iflib_txq_t txq = ctx->ifc_txqs;
5914 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5915 	if_irq_t irq = &ctx->ifc_legacy_irq;
5916 	iflib_filter_info_t info;
5917 	struct grouptask *gtask;
5918 	struct taskqgroup *tqg;
5919 	gtask_fn_t *fn;
5920 	int tqrid;
5921 	void *q;
5922 	int err;
5923 
5924 	q = &ctx->ifc_rxqs[0];
5925 	info = &rxq[0].ifr_filter_info;
5926 	gtask = &rxq[0].ifr_task;
5927 	tqg = qgroup_if_io_tqg;
5928 	tqrid = irq->ii_rid = *rid;
5929 	fn = _task_fn_rx;
5930 
5931 	ctx->ifc_flags |= IFC_LEGACY;
5932 	info->ifi_filter = filter;
5933 	info->ifi_filter_arg = filter_arg;
5934 	info->ifi_task = gtask;
5935 	info->ifi_ctx = ctx;
5936 
5937 	/* We allocate a single interrupt resource */
5938 	if ((err = _iflib_irq_alloc(ctx, irq, tqrid, iflib_fast_intr_ctx, NULL, info, name)) != 0)
5939 		return (err);
5940 	GROUPTASK_INIT(gtask, 0, fn, q);
5941 	taskqgroup_attach(tqg, gtask, q, rman_get_start(irq->ii_res), name);
5942 
5943 	GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq);
5944 	taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, rman_get_start(irq->ii_res), "tx");
5945 	return (0);
5946 }
5947 
5948 void
5949 iflib_led_create(if_ctx_t ctx)
5950 {
5951 
5952 	ctx->ifc_led_dev = led_create(iflib_led_func, ctx,
5953 	    device_get_nameunit(ctx->ifc_dev));
5954 }
5955 
5956 void
5957 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid)
5958 {
5959 
5960 	GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
5961 }
5962 
5963 void
5964 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid)
5965 {
5966 
5967 	GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task);
5968 }
5969 
5970 void
5971 iflib_admin_intr_deferred(if_ctx_t ctx)
5972 {
5973 #ifdef INVARIANTS
5974 	struct grouptask *gtask;
5975 
5976 	gtask = &ctx->ifc_admin_task;
5977 	MPASS(gtask != NULL && gtask->gt_taskqueue != NULL);
5978 #endif
5979 
5980 	GROUPTASK_ENQUEUE(&ctx->ifc_admin_task);
5981 }
5982 
5983 void
5984 iflib_iov_intr_deferred(if_ctx_t ctx)
5985 {
5986 
5987 	GROUPTASK_ENQUEUE(&ctx->ifc_vflr_task);
5988 }
5989 
5990 void
5991 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, char *name)
5992 {
5993 
5994 	taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, -1, name);
5995 }
5996 
5997 void
5998 iflib_config_gtask_init(void *ctx, struct grouptask *gtask, gtask_fn_t *fn,
5999 	const char *name)
6000 {
6001 
6002 	GROUPTASK_INIT(gtask, 0, fn, ctx);
6003 	taskqgroup_attach(qgroup_if_config_tqg, gtask, gtask, -1, name);
6004 }
6005 
6006 void
6007 iflib_config_gtask_deinit(struct grouptask *gtask)
6008 {
6009 
6010 	taskqgroup_detach(qgroup_if_config_tqg, gtask);
6011 }
6012 
6013 void
6014 iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate)
6015 {
6016 	if_t ifp = ctx->ifc_ifp;
6017 	iflib_txq_t txq = ctx->ifc_txqs;
6018 
6019 	if_setbaudrate(ifp, baudrate);
6020 	if (baudrate >= IF_Gbps(10)) {
6021 		STATE_LOCK(ctx);
6022 		ctx->ifc_flags |= IFC_PREFETCH;
6023 		STATE_UNLOCK(ctx);
6024 	}
6025 	/* If link down, disable watchdog */
6026 	if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) {
6027 		for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++)
6028 			txq->ift_qstatus = IFLIB_QUEUE_IDLE;
6029 	}
6030 	ctx->ifc_link_state = link_state;
6031 	if_link_state_change(ifp, link_state);
6032 }
6033 
6034 static int
6035 iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq)
6036 {
6037 	int credits;
6038 #ifdef INVARIANTS
6039 	int credits_pre = txq->ift_cidx_processed;
6040 #endif
6041 
6042 	if (ctx->isc_txd_credits_update == NULL)
6043 		return (0);
6044 
6045 	if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, true)) == 0)
6046 		return (0);
6047 
6048 	txq->ift_processed += credits;
6049 	txq->ift_cidx_processed += credits;
6050 
6051 	MPASS(credits_pre + credits == txq->ift_cidx_processed);
6052 	if (txq->ift_cidx_processed >= txq->ift_size)
6053 		txq->ift_cidx_processed -= txq->ift_size;
6054 	return (credits);
6055 }
6056 
6057 static int
6058 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget)
6059 {
6060 
6061 	return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx,
6062 	    budget));
6063 }
6064 
6065 void
6066 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name,
6067 	const char *description, if_int_delay_info_t info,
6068 	int offset, int value)
6069 {
6070 	info->iidi_ctx = ctx;
6071 	info->iidi_offset = offset;
6072 	info->iidi_value = value;
6073 	SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev),
6074 	    SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)),
6075 	    OID_AUTO, name, CTLTYPE_INT|CTLFLAG_RW,
6076 	    info, 0, iflib_sysctl_int_delay, "I", description);
6077 }
6078 
6079 struct sx *
6080 iflib_ctx_lock_get(if_ctx_t ctx)
6081 {
6082 
6083 	return (&ctx->ifc_ctx_sx);
6084 }
6085 
6086 static int
6087 iflib_msix_init(if_ctx_t ctx)
6088 {
6089 	device_t dev = ctx->ifc_dev;
6090 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6091 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6092 	int vectors, queues, rx_queues, tx_queues, queuemsgs, msgs;
6093 	int iflib_num_tx_queues, iflib_num_rx_queues;
6094 	int err, admincnt, bar;
6095 
6096 	iflib_num_tx_queues = ctx->ifc_sysctl_ntxqs;
6097 	iflib_num_rx_queues = ctx->ifc_sysctl_nrxqs;
6098 
6099 	device_printf(dev, "msix_init qsets capped at %d\n", imax(scctx->isc_ntxqsets, scctx->isc_nrxqsets));
6100 
6101 	bar = ctx->ifc_softc_ctx.isc_msix_bar;
6102 	admincnt = sctx->isc_admin_intrcnt;
6103 	/* Override by tuneable */
6104 	if (scctx->isc_disable_msix)
6105 		goto msi;
6106 
6107 	/*
6108 	 * bar == -1 => "trust me I know what I'm doing"
6109 	 * Some drivers are for hardware that is so shoddily
6110 	 * documented that no one knows which bars are which
6111 	 * so the developer has to map all bars. This hack
6112 	 * allows shoddy garbage to use msix in this framework.
6113 	 */
6114 	if (bar != -1) {
6115 		ctx->ifc_msix_mem = bus_alloc_resource_any(dev,
6116 	            SYS_RES_MEMORY, &bar, RF_ACTIVE);
6117 		if (ctx->ifc_msix_mem == NULL) {
6118 			/* May not be enabled */
6119 			device_printf(dev, "Unable to map MSIX table \n");
6120 			goto msi;
6121 		}
6122 	}
6123 	/* First try MSI/X */
6124 	if ((msgs = pci_msix_count(dev)) == 0) { /* system has msix disabled */
6125 		device_printf(dev, "System has MSIX disabled \n");
6126 		bus_release_resource(dev, SYS_RES_MEMORY,
6127 		    bar, ctx->ifc_msix_mem);
6128 		ctx->ifc_msix_mem = NULL;
6129 		goto msi;
6130 	}
6131 #if IFLIB_DEBUG
6132 	/* use only 1 qset in debug mode */
6133 	queuemsgs = min(msgs - admincnt, 1);
6134 #else
6135 	queuemsgs = msgs - admincnt;
6136 #endif
6137 #ifdef RSS
6138 	queues = imin(queuemsgs, rss_getnumbuckets());
6139 #else
6140 	queues = queuemsgs;
6141 #endif
6142 	queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues);
6143 	device_printf(dev, "pxm cpus: %d queue msgs: %d admincnt: %d\n",
6144 				  CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt);
6145 #ifdef  RSS
6146 	/* If we're doing RSS, clamp at the number of RSS buckets */
6147 	if (queues > rss_getnumbuckets())
6148 		queues = rss_getnumbuckets();
6149 #endif
6150 	if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt)
6151 		rx_queues = iflib_num_rx_queues;
6152 	else
6153 		rx_queues = queues;
6154 
6155 	if (rx_queues > scctx->isc_nrxqsets)
6156 		rx_queues = scctx->isc_nrxqsets;
6157 
6158 	/*
6159 	 * We want this to be all logical CPUs by default
6160 	 */
6161 	if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues)
6162 		tx_queues = iflib_num_tx_queues;
6163 	else
6164 		tx_queues = mp_ncpus;
6165 
6166 	if (tx_queues > scctx->isc_ntxqsets)
6167 		tx_queues = scctx->isc_ntxqsets;
6168 
6169 	if (ctx->ifc_sysctl_qs_eq_override == 0) {
6170 #ifdef INVARIANTS
6171 		if (tx_queues != rx_queues)
6172 			device_printf(dev,
6173 			    "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n",
6174 			    min(rx_queues, tx_queues), min(rx_queues, tx_queues));
6175 #endif
6176 		tx_queues = min(rx_queues, tx_queues);
6177 		rx_queues = min(rx_queues, tx_queues);
6178 	}
6179 
6180 	device_printf(dev, "using %d rx queues %d tx queues \n", rx_queues, tx_queues);
6181 
6182 	vectors = rx_queues + admincnt;
6183 	if ((err = pci_alloc_msix(dev, &vectors)) == 0) {
6184 		device_printf(dev, "Using MSIX interrupts with %d vectors\n", vectors);
6185 		scctx->isc_vectors = vectors;
6186 		scctx->isc_nrxqsets = rx_queues;
6187 		scctx->isc_ntxqsets = tx_queues;
6188 		scctx->isc_intr = IFLIB_INTR_MSIX;
6189 
6190 		return (vectors);
6191 	} else {
6192 		device_printf(dev,
6193 		    "failed to allocate %d msix vectors, err: %d - using MSI\n", vectors, err);
6194 		bus_release_resource(dev, SYS_RES_MEMORY, bar,
6195 		    ctx->ifc_msix_mem);
6196 		ctx->ifc_msix_mem = NULL;
6197 	}
6198 msi:
6199 	vectors = pci_msi_count(dev);
6200 	scctx->isc_nrxqsets = 1;
6201 	scctx->isc_ntxqsets = 1;
6202 	scctx->isc_vectors = vectors;
6203 	if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) {
6204 		device_printf(dev,"Using an MSI interrupt\n");
6205 		scctx->isc_intr = IFLIB_INTR_MSI;
6206 	} else {
6207 		scctx->isc_vectors = 1;
6208 		device_printf(dev,"Using a Legacy interrupt\n");
6209 		scctx->isc_intr = IFLIB_INTR_LEGACY;
6210 	}
6211 
6212 	return (vectors);
6213 }
6214 
6215 static const char *ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" };
6216 
6217 static int
6218 mp_ring_state_handler(SYSCTL_HANDLER_ARGS)
6219 {
6220 	int rc;
6221 	uint16_t *state = ((uint16_t *)oidp->oid_arg1);
6222 	struct sbuf *sb;
6223 	const char *ring_state = "UNKNOWN";
6224 
6225 	/* XXX needed ? */
6226 	rc = sysctl_wire_old_buffer(req, 0);
6227 	MPASS(rc == 0);
6228 	if (rc != 0)
6229 		return (rc);
6230 	sb = sbuf_new_for_sysctl(NULL, NULL, 80, req);
6231 	MPASS(sb != NULL);
6232 	if (sb == NULL)
6233 		return (ENOMEM);
6234 	if (state[3] <= 3)
6235 		ring_state = ring_states[state[3]];
6236 
6237 	sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s",
6238 		    state[0], state[1], state[2], ring_state);
6239 	rc = sbuf_finish(sb);
6240 	sbuf_delete(sb);
6241         return(rc);
6242 }
6243 
6244 enum iflib_ndesc_handler {
6245 	IFLIB_NTXD_HANDLER,
6246 	IFLIB_NRXD_HANDLER,
6247 };
6248 
6249 static int
6250 mp_ndesc_handler(SYSCTL_HANDLER_ARGS)
6251 {
6252 	if_ctx_t ctx = (void *)arg1;
6253 	enum iflib_ndesc_handler type = arg2;
6254 	char buf[256] = {0};
6255 	qidx_t *ndesc;
6256 	char *p, *next;
6257 	int nqs, rc, i;
6258 
6259 	MPASS(type == IFLIB_NTXD_HANDLER || type == IFLIB_NRXD_HANDLER);
6260 
6261 	nqs = 8;
6262 	switch(type) {
6263 	case IFLIB_NTXD_HANDLER:
6264 		ndesc = ctx->ifc_sysctl_ntxds;
6265 		if (ctx->ifc_sctx)
6266 			nqs = ctx->ifc_sctx->isc_ntxqs;
6267 		break;
6268 	case IFLIB_NRXD_HANDLER:
6269 		ndesc = ctx->ifc_sysctl_nrxds;
6270 		if (ctx->ifc_sctx)
6271 			nqs = ctx->ifc_sctx->isc_nrxqs;
6272 		break;
6273 	default:
6274 			panic("unhandled type");
6275 	}
6276 	if (nqs == 0)
6277 		nqs = 8;
6278 
6279 	for (i=0; i<8; i++) {
6280 		if (i >= nqs)
6281 			break;
6282 		if (i)
6283 			strcat(buf, ",");
6284 		sprintf(strchr(buf, 0), "%d", ndesc[i]);
6285 	}
6286 
6287 	rc = sysctl_handle_string(oidp, buf, sizeof(buf), req);
6288 	if (rc || req->newptr == NULL)
6289 		return rc;
6290 
6291 	for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p;
6292 	    i++, p = strsep(&next, " ,")) {
6293 		ndesc[i] = strtoul(p, NULL, 10);
6294 	}
6295 
6296 	return(rc);
6297 }
6298 
6299 #define NAME_BUFLEN 32
6300 static void
6301 iflib_add_device_sysctl_pre(if_ctx_t ctx)
6302 {
6303         device_t dev = iflib_get_dev(ctx);
6304 	struct sysctl_oid_list *child, *oid_list;
6305 	struct sysctl_ctx_list *ctx_list;
6306 	struct sysctl_oid *node;
6307 
6308 	ctx_list = device_get_sysctl_ctx(dev);
6309 	child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev));
6310 	ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, "iflib",
6311 						      CTLFLAG_RD, NULL, "IFLIB fields");
6312 	oid_list = SYSCTL_CHILDREN(node);
6313 
6314 	SYSCTL_ADD_STRING(ctx_list, oid_list, OID_AUTO, "driver_version",
6315 		       CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version, 0,
6316 		       "driver version");
6317 
6318 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs",
6319 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0,
6320 			"# of txqs to use, 0 => use default #");
6321 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs",
6322 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0,
6323 			"# of rxqs to use, 0 => use default #");
6324 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_qs_enable",
6325 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0,
6326                        "permit #txq != #rxq");
6327 	SYSCTL_ADD_INT(ctx_list, oid_list, OID_AUTO, "disable_msix",
6328                       CTLFLAG_RWTUN, &ctx->ifc_softc_ctx.isc_disable_msix, 0,
6329                       "disable MSIX (default 0)");
6330 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "rx_budget",
6331 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_rx_budget, 0,
6332                        "set the rx budget");
6333 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "tx_abdicate",
6334 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_tx_abdicate, 0,
6335 		       "cause tx to abdicate instead of running to completion");
6336 
6337 	/* XXX change for per-queue sizes */
6338 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_ntxds",
6339 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NTXD_HANDLER,
6340                        mp_ndesc_handler, "A",
6341                        "list of # of tx descriptors to use, 0 = use default #");
6342 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_nrxds",
6343 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NRXD_HANDLER,
6344                        mp_ndesc_handler, "A",
6345                        "list of # of rx descriptors to use, 0 = use default #");
6346 }
6347 
6348 static void
6349 iflib_add_device_sysctl_post(if_ctx_t ctx)
6350 {
6351 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6352 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6353         device_t dev = iflib_get_dev(ctx);
6354 	struct sysctl_oid_list *child;
6355 	struct sysctl_ctx_list *ctx_list;
6356 	iflib_fl_t fl;
6357 	iflib_txq_t txq;
6358 	iflib_rxq_t rxq;
6359 	int i, j;
6360 	char namebuf[NAME_BUFLEN];
6361 	char *qfmt;
6362 	struct sysctl_oid *queue_node, *fl_node, *node;
6363 	struct sysctl_oid_list *queue_list, *fl_list;
6364 	ctx_list = device_get_sysctl_ctx(dev);
6365 
6366 	node = ctx->ifc_sysctl_node;
6367 	child = SYSCTL_CHILDREN(node);
6368 
6369 	if (scctx->isc_ntxqsets > 100)
6370 		qfmt = "txq%03d";
6371 	else if (scctx->isc_ntxqsets > 10)
6372 		qfmt = "txq%02d";
6373 	else
6374 		qfmt = "txq%d";
6375 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
6376 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6377 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6378 					     CTLFLAG_RD, NULL, "Queue Name");
6379 		queue_list = SYSCTL_CHILDREN(queue_node);
6380 #if MEMORY_LOGGING
6381 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued",
6382 				CTLFLAG_RD,
6383 				&txq->ift_dequeued, "total mbufs freed");
6384 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued",
6385 				CTLFLAG_RD,
6386 				&txq->ift_enqueued, "total mbufs enqueued");
6387 #endif
6388 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag",
6389 				   CTLFLAG_RD,
6390 				   &txq->ift_mbuf_defrag, "# of times m_defrag was called");
6391 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "m_pullups",
6392 				   CTLFLAG_RD,
6393 				   &txq->ift_pullups, "# of times m_pullup was called");
6394 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag_failed",
6395 				   CTLFLAG_RD,
6396 				   &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed");
6397 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_desc_avail",
6398 				   CTLFLAG_RD,
6399 				   &txq->ift_no_desc_avail, "# of times no descriptors were available");
6400 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "tx_map_failed",
6401 				   CTLFLAG_RD,
6402 				   &txq->ift_map_failed, "# of times dma map failed");
6403 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txd_encap_efbig",
6404 				   CTLFLAG_RD,
6405 				   &txq->ift_txd_encap_efbig, "# of times txd_encap returned EFBIG");
6406 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_tx_dma_setup",
6407 				   CTLFLAG_RD,
6408 				   &txq->ift_no_tx_dma_setup, "# of times map failed for other than EFBIG");
6409 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx",
6410 				   CTLFLAG_RD,
6411 				   &txq->ift_pidx, 1, "Producer Index");
6412 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx",
6413 				   CTLFLAG_RD,
6414 				   &txq->ift_cidx, 1, "Consumer Index");
6415 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx_processed",
6416 				   CTLFLAG_RD,
6417 				   &txq->ift_cidx_processed, 1, "Consumer Index seen by credit update");
6418 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use",
6419 				   CTLFLAG_RD,
6420 				   &txq->ift_in_use, 1, "descriptors in use");
6421 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_processed",
6422 				   CTLFLAG_RD,
6423 				   &txq->ift_processed, "descriptors procesed for clean");
6424 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned",
6425 				   CTLFLAG_RD,
6426 				   &txq->ift_cleaned, "total cleaned");
6427 		SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state",
6428 				CTLTYPE_STRING | CTLFLAG_RD, __DEVOLATILE(uint64_t *, &txq->ift_br->state),
6429 				0, mp_ring_state_handler, "A", "soft ring state");
6430 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_enqueues",
6431 				       CTLFLAG_RD, &txq->ift_br->enqueues,
6432 				       "# of enqueues to the mp_ring for this queue");
6433 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_drops",
6434 				       CTLFLAG_RD, &txq->ift_br->drops,
6435 				       "# of drops in the mp_ring for this queue");
6436 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_starts",
6437 				       CTLFLAG_RD, &txq->ift_br->starts,
6438 				       "# of normal consumer starts in the mp_ring for this queue");
6439 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_stalls",
6440 				       CTLFLAG_RD, &txq->ift_br->stalls,
6441 					       "# of consumer stalls in the mp_ring for this queue");
6442 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_restarts",
6443 			       CTLFLAG_RD, &txq->ift_br->restarts,
6444 				       "# of consumer restarts in the mp_ring for this queue");
6445 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_abdications",
6446 				       CTLFLAG_RD, &txq->ift_br->abdications,
6447 				       "# of consumer abdications in the mp_ring for this queue");
6448 	}
6449 
6450 	if (scctx->isc_nrxqsets > 100)
6451 		qfmt = "rxq%03d";
6452 	else if (scctx->isc_nrxqsets > 10)
6453 		qfmt = "rxq%02d";
6454 	else
6455 		qfmt = "rxq%d";
6456 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
6457 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6458 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6459 					     CTLFLAG_RD, NULL, "Queue Name");
6460 		queue_list = SYSCTL_CHILDREN(queue_node);
6461 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
6462 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_pidx",
6463 				       CTLFLAG_RD,
6464 				       &rxq->ifr_cq_pidx, 1, "Producer Index");
6465 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_cidx",
6466 				       CTLFLAG_RD,
6467 				       &rxq->ifr_cq_cidx, 1, "Consumer Index");
6468 		}
6469 
6470 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
6471 			snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j);
6472 			fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list, OID_AUTO, namebuf,
6473 						     CTLFLAG_RD, NULL, "freelist Name");
6474 			fl_list = SYSCTL_CHILDREN(fl_node);
6475 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx",
6476 				       CTLFLAG_RD,
6477 				       &fl->ifl_pidx, 1, "Producer Index");
6478 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx",
6479 				       CTLFLAG_RD,
6480 				       &fl->ifl_cidx, 1, "Consumer Index");
6481 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits",
6482 				       CTLFLAG_RD,
6483 				       &fl->ifl_credits, 1, "credits available");
6484 #if MEMORY_LOGGING
6485 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_enqueued",
6486 					CTLFLAG_RD,
6487 					&fl->ifl_m_enqueued, "mbufs allocated");
6488 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_dequeued",
6489 					CTLFLAG_RD,
6490 					&fl->ifl_m_dequeued, "mbufs freed");
6491 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_enqueued",
6492 					CTLFLAG_RD,
6493 					&fl->ifl_cl_enqueued, "clusters allocated");
6494 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_dequeued",
6495 					CTLFLAG_RD,
6496 					&fl->ifl_cl_dequeued, "clusters freed");
6497 #endif
6498 
6499 		}
6500 	}
6501 
6502 }
6503 
6504 void
6505 iflib_request_reset(if_ctx_t ctx)
6506 {
6507 
6508 	STATE_LOCK(ctx);
6509 	ctx->ifc_flags |= IFC_DO_RESET;
6510 	STATE_UNLOCK(ctx);
6511 }
6512 
6513 #ifndef __NO_STRICT_ALIGNMENT
6514 static struct mbuf *
6515 iflib_fixup_rx(struct mbuf *m)
6516 {
6517 	struct mbuf *n;
6518 
6519 	if (m->m_len <= (MCLBYTES - ETHER_HDR_LEN)) {
6520 		bcopy(m->m_data, m->m_data + ETHER_HDR_LEN, m->m_len);
6521 		m->m_data += ETHER_HDR_LEN;
6522 		n = m;
6523 	} else {
6524 		MGETHDR(n, M_NOWAIT, MT_DATA);
6525 		if (n == NULL) {
6526 			m_freem(m);
6527 			return (NULL);
6528 		}
6529 		bcopy(m->m_data, n->m_data, ETHER_HDR_LEN);
6530 		m->m_data += ETHER_HDR_LEN;
6531 		m->m_len -= ETHER_HDR_LEN;
6532 		n->m_len = ETHER_HDR_LEN;
6533 		M_MOVE_PKTHDR(n, m);
6534 		n->m_next = m;
6535 	}
6536 	return (n);
6537 }
6538 #endif
6539 
6540 #ifdef NETDUMP
6541 static void
6542 iflib_netdump_init(struct ifnet *ifp, int *nrxr, int *ncl, int *clsize)
6543 {
6544 	if_ctx_t ctx;
6545 
6546 	ctx = if_getsoftc(ifp);
6547 	CTX_LOCK(ctx);
6548 	*nrxr = NRXQSETS(ctx);
6549 	*ncl = ctx->ifc_rxqs[0].ifr_fl->ifl_size;
6550 	*clsize = ctx->ifc_rxqs[0].ifr_fl->ifl_buf_size;
6551 	CTX_UNLOCK(ctx);
6552 }
6553 
6554 static void
6555 iflib_netdump_event(struct ifnet *ifp, enum netdump_ev event)
6556 {
6557 	if_ctx_t ctx;
6558 	if_softc_ctx_t scctx;
6559 	iflib_fl_t fl;
6560 	iflib_rxq_t rxq;
6561 	int i, j;
6562 
6563 	ctx = if_getsoftc(ifp);
6564 	scctx = &ctx->ifc_softc_ctx;
6565 
6566 	switch (event) {
6567 	case NETDUMP_START:
6568 		for (i = 0; i < scctx->isc_nrxqsets; i++) {
6569 			rxq = &ctx->ifc_rxqs[i];
6570 			for (j = 0; j < rxq->ifr_nfl; j++) {
6571 				fl = rxq->ifr_fl;
6572 				fl->ifl_zone = m_getzone(fl->ifl_buf_size);
6573 			}
6574 		}
6575 		iflib_no_tx_batch = 1;
6576 		break;
6577 	default:
6578 		break;
6579 	}
6580 }
6581 
6582 static int
6583 iflib_netdump_transmit(struct ifnet *ifp, struct mbuf *m)
6584 {
6585 	if_ctx_t ctx;
6586 	iflib_txq_t txq;
6587 	int error;
6588 
6589 	ctx = if_getsoftc(ifp);
6590 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
6591 	    IFF_DRV_RUNNING)
6592 		return (EBUSY);
6593 
6594 	txq = &ctx->ifc_txqs[0];
6595 	error = iflib_encap(txq, &m);
6596 	if (error == 0)
6597 		(void)iflib_txd_db_check(ctx, txq, true, txq->ift_in_use);
6598 	return (error);
6599 }
6600 
6601 static int
6602 iflib_netdump_poll(struct ifnet *ifp, int count)
6603 {
6604 	if_ctx_t ctx;
6605 	if_softc_ctx_t scctx;
6606 	iflib_txq_t txq;
6607 	int i;
6608 
6609 	ctx = if_getsoftc(ifp);
6610 	scctx = &ctx->ifc_softc_ctx;
6611 
6612 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
6613 	    IFF_DRV_RUNNING)
6614 		return (EBUSY);
6615 
6616 	txq = &ctx->ifc_txqs[0];
6617 	(void)iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
6618 
6619 	for (i = 0; i < scctx->isc_nrxqsets; i++)
6620 		(void)iflib_rxeof(&ctx->ifc_rxqs[i], 16 /* XXX */);
6621 	return (0);
6622 }
6623 #endif /* NETDUMP */
6624