xref: /freebsd/sys/net/iflib.c (revision c7432537b1a67e964ed1d7b35977049301dc9189)
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/kernel.h>
41 #include <sys/lock.h>
42 #include <sys/mutex.h>
43 #include <sys/module.h>
44 #include <sys/kobj.h>
45 #include <sys/rman.h>
46 #include <sys/sbuf.h>
47 #include <sys/smp.h>
48 #include <sys/socket.h>
49 #include <sys/sockio.h>
50 #include <sys/sysctl.h>
51 #include <sys/syslog.h>
52 #include <sys/taskqueue.h>
53 #include <sys/limits.h>
54 
55 #include <net/if.h>
56 #include <net/if_var.h>
57 #include <net/if_types.h>
58 #include <net/if_media.h>
59 #include <net/bpf.h>
60 #include <net/ethernet.h>
61 #include <net/mp_ring.h>
62 #include <net/pfil.h>
63 #include <net/vnet.h>
64 
65 #include <netinet/in.h>
66 #include <netinet/in_pcb.h>
67 #include <netinet/tcp_lro.h>
68 #include <netinet/in_systm.h>
69 #include <netinet/if_ether.h>
70 #include <netinet/ip.h>
71 #include <netinet/ip6.h>
72 #include <netinet/tcp.h>
73 #include <netinet/ip_var.h>
74 #include <netinet/netdump/netdump.h>
75 #include <netinet6/ip6_var.h>
76 
77 #include <machine/bus.h>
78 #include <machine/in_cksum.h>
79 
80 #include <vm/vm.h>
81 #include <vm/pmap.h>
82 
83 #include <dev/led/led.h>
84 #include <dev/pci/pcireg.h>
85 #include <dev/pci/pcivar.h>
86 #include <dev/pci/pci_private.h>
87 
88 #include <net/iflib.h>
89 #include <net/iflib_private.h>
90 
91 #include "ifdi_if.h"
92 
93 #ifdef PCI_IOV
94 #include <dev/pci/pci_iov.h>
95 #endif
96 
97 #include <sys/bitstring.h>
98 /*
99  * enable accounting of every mbuf as it comes in to and goes out of
100  * iflib's software descriptor references
101  */
102 #define MEMORY_LOGGING 0
103 /*
104  * Enable mbuf vectors for compressing long mbuf chains
105  */
106 
107 /*
108  * NB:
109  * - Prefetching in tx cleaning should perhaps be a tunable. The distance ahead
110  *   we prefetch needs to be determined by the time spent in m_free vis a vis
111  *   the cost of a prefetch. This will of course vary based on the workload:
112  *      - NFLX's m_free path is dominated by vm-based M_EXT manipulation which
113  *        is quite expensive, thus suggesting very little prefetch.
114  *      - small packet forwarding which is just returning a single mbuf to
115  *        UMA will typically be very fast vis a vis the cost of a memory
116  *        access.
117  */
118 
119 
120 /*
121  * File organization:
122  *  - private structures
123  *  - iflib private utility functions
124  *  - ifnet functions
125  *  - vlan registry and other exported functions
126  *  - iflib public core functions
127  *
128  *
129  */
130 MALLOC_DEFINE(M_IFLIB, "iflib", "ifnet library");
131 
132 struct iflib_txq;
133 typedef struct iflib_txq *iflib_txq_t;
134 struct iflib_rxq;
135 typedef struct iflib_rxq *iflib_rxq_t;
136 struct iflib_fl;
137 typedef struct iflib_fl *iflib_fl_t;
138 
139 struct iflib_ctx;
140 
141 static void iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid);
142 static void iflib_timer(void *arg);
143 
144 typedef struct iflib_filter_info {
145 	driver_filter_t *ifi_filter;
146 	void *ifi_filter_arg;
147 	struct grouptask *ifi_task;
148 	void *ifi_ctx;
149 } *iflib_filter_info_t;
150 
151 struct iflib_ctx {
152 	KOBJ_FIELDS;
153 	/*
154 	 * Pointer to hardware driver's softc
155 	 */
156 	void *ifc_softc;
157 	device_t ifc_dev;
158 	if_t ifc_ifp;
159 
160 	cpuset_t ifc_cpus;
161 	if_shared_ctx_t ifc_sctx;
162 	struct if_softc_ctx ifc_softc_ctx;
163 
164 	struct sx ifc_ctx_sx;
165 	struct mtx ifc_state_mtx;
166 
167 	iflib_txq_t ifc_txqs;
168 	iflib_rxq_t ifc_rxqs;
169 	uint32_t ifc_if_flags;
170 	uint32_t ifc_flags;
171 	uint32_t ifc_max_fl_buf_size;
172 	uint32_t ifc_rx_mbuf_sz;
173 
174 	int ifc_link_state;
175 	int ifc_link_irq;
176 	int ifc_watchdog_events;
177 	struct cdev *ifc_led_dev;
178 	struct resource *ifc_msix_mem;
179 
180 	struct if_irq ifc_legacy_irq;
181 	struct grouptask ifc_admin_task;
182 	struct grouptask ifc_vflr_task;
183 	struct iflib_filter_info ifc_filter_info;
184 	struct ifmedia	ifc_media;
185 
186 	struct sysctl_oid *ifc_sysctl_node;
187 	uint16_t ifc_sysctl_ntxqs;
188 	uint16_t ifc_sysctl_nrxqs;
189 	uint16_t ifc_sysctl_qs_eq_override;
190 	uint16_t ifc_sysctl_rx_budget;
191 	uint16_t ifc_sysctl_tx_abdicate;
192 
193 	qidx_t ifc_sysctl_ntxds[8];
194 	qidx_t ifc_sysctl_nrxds[8];
195 	struct if_txrx ifc_txrx;
196 #define isc_txd_encap  ifc_txrx.ift_txd_encap
197 #define isc_txd_flush  ifc_txrx.ift_txd_flush
198 #define isc_txd_credits_update  ifc_txrx.ift_txd_credits_update
199 #define isc_rxd_available ifc_txrx.ift_rxd_available
200 #define isc_rxd_pkt_get ifc_txrx.ift_rxd_pkt_get
201 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
202 #define isc_rxd_flush ifc_txrx.ift_rxd_flush
203 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
204 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
205 #define isc_legacy_intr ifc_txrx.ift_legacy_intr
206 	eventhandler_tag ifc_vlan_attach_event;
207 	eventhandler_tag ifc_vlan_detach_event;
208 	struct ether_addr ifc_mac;
209 	char ifc_mtx_name[16];
210 };
211 
212 
213 void *
214 iflib_get_softc(if_ctx_t ctx)
215 {
216 
217 	return (ctx->ifc_softc);
218 }
219 
220 device_t
221 iflib_get_dev(if_ctx_t ctx)
222 {
223 
224 	return (ctx->ifc_dev);
225 }
226 
227 if_t
228 iflib_get_ifp(if_ctx_t ctx)
229 {
230 
231 	return (ctx->ifc_ifp);
232 }
233 
234 struct ifmedia *
235 iflib_get_media(if_ctx_t ctx)
236 {
237 
238 	return (&ctx->ifc_media);
239 }
240 
241 uint32_t
242 iflib_get_flags(if_ctx_t ctx)
243 {
244 	return (ctx->ifc_flags);
245 }
246 
247 void
248 iflib_set_mac(if_ctx_t ctx, uint8_t mac[ETHER_ADDR_LEN])
249 {
250 
251 	bcopy(mac, ctx->ifc_mac.octet, ETHER_ADDR_LEN);
252 }
253 
254 if_softc_ctx_t
255 iflib_get_softc_ctx(if_ctx_t ctx)
256 {
257 
258 	return (&ctx->ifc_softc_ctx);
259 }
260 
261 if_shared_ctx_t
262 iflib_get_sctx(if_ctx_t ctx)
263 {
264 
265 	return (ctx->ifc_sctx);
266 }
267 
268 #define IP_ALIGNED(m) ((((uintptr_t)(m)->m_data) & 0x3) == 0x2)
269 #define CACHE_PTR_INCREMENT (CACHE_LINE_SIZE/sizeof(void*))
270 #define CACHE_PTR_NEXT(ptr) ((void *)(((uintptr_t)(ptr)+CACHE_LINE_SIZE-1) & (CACHE_LINE_SIZE-1)))
271 
272 #define LINK_ACTIVE(ctx) ((ctx)->ifc_link_state == LINK_STATE_UP)
273 #define CTX_IS_VF(ctx) ((ctx)->ifc_sctx->isc_flags & IFLIB_IS_VF)
274 
275 typedef struct iflib_sw_rx_desc_array {
276 	bus_dmamap_t	*ifsd_map;         /* bus_dma maps for packet */
277 	struct mbuf	**ifsd_m;           /* pkthdr mbufs */
278 	caddr_t		*ifsd_cl;          /* direct cluster pointer for rx */
279 	bus_addr_t	*ifsd_ba;          /* bus addr of cluster for rx */
280 } iflib_rxsd_array_t;
281 
282 typedef struct iflib_sw_tx_desc_array {
283 	bus_dmamap_t    *ifsd_map;         /* bus_dma maps for packet */
284 	bus_dmamap_t	*ifsd_tso_map;     /* bus_dma maps for TSO packet */
285 	struct mbuf    **ifsd_m;           /* pkthdr mbufs */
286 } if_txsd_vec_t;
287 
288 
289 /* magic number that should be high enough for any hardware */
290 #define IFLIB_MAX_TX_SEGS		128
291 #define IFLIB_RX_COPY_THRESH		128
292 #define IFLIB_MAX_RX_REFRESH		32
293 /* The minimum descriptors per second before we start coalescing */
294 #define IFLIB_MIN_DESC_SEC		16384
295 #define IFLIB_DEFAULT_TX_UPDATE_FREQ	16
296 #define IFLIB_QUEUE_IDLE		0
297 #define IFLIB_QUEUE_HUNG		1
298 #define IFLIB_QUEUE_WORKING		2
299 /* maximum number of txqs that can share an rx interrupt */
300 #define IFLIB_MAX_TX_SHARED_INTR	4
301 
302 /* this should really scale with ring size - this is a fairly arbitrary value */
303 #define TX_BATCH_SIZE			32
304 
305 #define IFLIB_RESTART_BUDGET		8
306 
307 
308 #define CSUM_OFFLOAD		(CSUM_IP_TSO|CSUM_IP6_TSO|CSUM_IP| \
309 				 CSUM_IP_UDP|CSUM_IP_TCP|CSUM_IP_SCTP| \
310 				 CSUM_IP6_UDP|CSUM_IP6_TCP|CSUM_IP6_SCTP)
311 struct iflib_txq {
312 	qidx_t		ift_in_use;
313 	qidx_t		ift_cidx;
314 	qidx_t		ift_cidx_processed;
315 	qidx_t		ift_pidx;
316 	uint8_t		ift_gen;
317 	uint8_t		ift_br_offset;
318 	uint16_t	ift_npending;
319 	uint16_t	ift_db_pending;
320 	uint16_t	ift_rs_pending;
321 	/* implicit pad */
322 	uint8_t		ift_txd_size[8];
323 	uint64_t	ift_processed;
324 	uint64_t	ift_cleaned;
325 	uint64_t	ift_cleaned_prev;
326 #if MEMORY_LOGGING
327 	uint64_t	ift_enqueued;
328 	uint64_t	ift_dequeued;
329 #endif
330 	uint64_t	ift_no_tx_dma_setup;
331 	uint64_t	ift_no_desc_avail;
332 	uint64_t	ift_mbuf_defrag_failed;
333 	uint64_t	ift_mbuf_defrag;
334 	uint64_t	ift_map_failed;
335 	uint64_t	ift_txd_encap_efbig;
336 	uint64_t	ift_pullups;
337 	uint64_t	ift_last_timer_tick;
338 
339 	struct mtx	ift_mtx;
340 	struct mtx	ift_db_mtx;
341 
342 	/* constant values */
343 	if_ctx_t	ift_ctx;
344 	struct ifmp_ring        *ift_br;
345 	struct grouptask	ift_task;
346 	qidx_t		ift_size;
347 	uint16_t	ift_id;
348 	struct callout	ift_timer;
349 
350 	if_txsd_vec_t	ift_sds;
351 	uint8_t		ift_qstatus;
352 	uint8_t		ift_closed;
353 	uint8_t		ift_update_freq;
354 	struct iflib_filter_info ift_filter_info;
355 	bus_dma_tag_t	ift_buf_tag;
356 	bus_dma_tag_t	ift_tso_buf_tag;
357 	iflib_dma_info_t	ift_ifdi;
358 #define MTX_NAME_LEN 16
359 	char                    ift_mtx_name[MTX_NAME_LEN];
360 	char                    ift_db_mtx_name[MTX_NAME_LEN];
361 	bus_dma_segment_t	ift_segs[IFLIB_MAX_TX_SEGS]  __aligned(CACHE_LINE_SIZE);
362 #ifdef IFLIB_DIAGNOSTICS
363 	uint64_t ift_cpu_exec_count[256];
364 #endif
365 } __aligned(CACHE_LINE_SIZE);
366 
367 struct iflib_fl {
368 	qidx_t		ifl_cidx;
369 	qidx_t		ifl_pidx;
370 	qidx_t		ifl_credits;
371 	uint8_t		ifl_gen;
372 	uint8_t		ifl_rxd_size;
373 #if MEMORY_LOGGING
374 	uint64_t	ifl_m_enqueued;
375 	uint64_t	ifl_m_dequeued;
376 	uint64_t	ifl_cl_enqueued;
377 	uint64_t	ifl_cl_dequeued;
378 #endif
379 	/* implicit pad */
380 
381 	bitstr_t 	*ifl_rx_bitmap;
382 	qidx_t		ifl_fragidx;
383 	/* constant */
384 	qidx_t		ifl_size;
385 	uint16_t	ifl_buf_size;
386 	uint16_t	ifl_cltype;
387 	uma_zone_t	ifl_zone;
388 	iflib_rxsd_array_t	ifl_sds;
389 	iflib_rxq_t	ifl_rxq;
390 	uint8_t		ifl_id;
391 	bus_dma_tag_t	ifl_buf_tag;
392 	iflib_dma_info_t	ifl_ifdi;
393 	uint64_t	ifl_bus_addrs[IFLIB_MAX_RX_REFRESH] __aligned(CACHE_LINE_SIZE);
394 	caddr_t		ifl_vm_addrs[IFLIB_MAX_RX_REFRESH];
395 	qidx_t	ifl_rxd_idxs[IFLIB_MAX_RX_REFRESH];
396 }  __aligned(CACHE_LINE_SIZE);
397 
398 static inline qidx_t
399 get_inuse(int size, qidx_t cidx, qidx_t pidx, uint8_t gen)
400 {
401 	qidx_t used;
402 
403 	if (pidx > cidx)
404 		used = pidx - cidx;
405 	else if (pidx < cidx)
406 		used = size - cidx + pidx;
407 	else if (gen == 0 && pidx == cidx)
408 		used = 0;
409 	else if (gen == 1 && pidx == cidx)
410 		used = size;
411 	else
412 		panic("bad state");
413 
414 	return (used);
415 }
416 
417 #define TXQ_AVAIL(txq) (txq->ift_size - get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx, txq->ift_gen))
418 
419 #define IDXDIFF(head, tail, wrap) \
420 	((head) >= (tail) ? (head) - (tail) : (wrap) - (tail) + (head))
421 
422 struct iflib_rxq {
423 	/* If there is a separate completion queue -
424 	 * these are the cq cidx and pidx. Otherwise
425 	 * these are unused.
426 	 */
427 	qidx_t		ifr_size;
428 	qidx_t		ifr_cq_cidx;
429 	qidx_t		ifr_cq_pidx;
430 	uint8_t		ifr_cq_gen;
431 	uint8_t		ifr_fl_offset;
432 
433 	if_ctx_t	ifr_ctx;
434 	iflib_fl_t	ifr_fl;
435 	uint64_t	ifr_rx_irq;
436 	struct pfil_head	*pfil;
437 	uint16_t	ifr_id;
438 	uint8_t		ifr_lro_enabled;
439 	uint8_t		ifr_nfl;
440 	uint8_t		ifr_ntxqirq;
441 	uint8_t		ifr_txqid[IFLIB_MAX_TX_SHARED_INTR];
442 	struct lro_ctrl			ifr_lc;
443 	struct grouptask        ifr_task;
444 	struct iflib_filter_info ifr_filter_info;
445 	iflib_dma_info_t		ifr_ifdi;
446 
447 	/* dynamically allocate if any drivers need a value substantially larger than this */
448 	struct if_rxd_frag	ifr_frags[IFLIB_MAX_RX_SEGS] __aligned(CACHE_LINE_SIZE);
449 #ifdef IFLIB_DIAGNOSTICS
450 	uint64_t ifr_cpu_exec_count[256];
451 #endif
452 }  __aligned(CACHE_LINE_SIZE);
453 
454 typedef struct if_rxsd {
455 	caddr_t *ifsd_cl;
456 	iflib_fl_t ifsd_fl;
457 	qidx_t ifsd_cidx;
458 } *if_rxsd_t;
459 
460 /* multiple of word size */
461 #ifdef __LP64__
462 #define PKT_INFO_SIZE	6
463 #define RXD_INFO_SIZE	5
464 #define PKT_TYPE uint64_t
465 #else
466 #define PKT_INFO_SIZE	11
467 #define RXD_INFO_SIZE	8
468 #define PKT_TYPE uint32_t
469 #endif
470 #define PKT_LOOP_BOUND  ((PKT_INFO_SIZE/3)*3)
471 #define RXD_LOOP_BOUND  ((RXD_INFO_SIZE/4)*4)
472 
473 typedef struct if_pkt_info_pad {
474 	PKT_TYPE pkt_val[PKT_INFO_SIZE];
475 } *if_pkt_info_pad_t;
476 typedef struct if_rxd_info_pad {
477 	PKT_TYPE rxd_val[RXD_INFO_SIZE];
478 } *if_rxd_info_pad_t;
479 
480 CTASSERT(sizeof(struct if_pkt_info_pad) == sizeof(struct if_pkt_info));
481 CTASSERT(sizeof(struct if_rxd_info_pad) == sizeof(struct if_rxd_info));
482 
483 
484 static inline void
485 pkt_info_zero(if_pkt_info_t pi)
486 {
487 	if_pkt_info_pad_t pi_pad;
488 
489 	pi_pad = (if_pkt_info_pad_t)pi;
490 	pi_pad->pkt_val[0] = 0; pi_pad->pkt_val[1] = 0; pi_pad->pkt_val[2] = 0;
491 	pi_pad->pkt_val[3] = 0; pi_pad->pkt_val[4] = 0; pi_pad->pkt_val[5] = 0;
492 #ifndef __LP64__
493 	pi_pad->pkt_val[6] = 0; pi_pad->pkt_val[7] = 0; pi_pad->pkt_val[8] = 0;
494 	pi_pad->pkt_val[9] = 0; pi_pad->pkt_val[10] = 0;
495 #endif
496 }
497 
498 static device_method_t iflib_pseudo_methods[] = {
499 	DEVMETHOD(device_attach, noop_attach),
500 	DEVMETHOD(device_detach, iflib_pseudo_detach),
501 	DEVMETHOD_END
502 };
503 
504 driver_t iflib_pseudodriver = {
505 	"iflib_pseudo", iflib_pseudo_methods, sizeof(struct iflib_ctx),
506 };
507 
508 static inline void
509 rxd_info_zero(if_rxd_info_t ri)
510 {
511 	if_rxd_info_pad_t ri_pad;
512 	int i;
513 
514 	ri_pad = (if_rxd_info_pad_t)ri;
515 	for (i = 0; i < RXD_LOOP_BOUND; i += 4) {
516 		ri_pad->rxd_val[i] = 0;
517 		ri_pad->rxd_val[i+1] = 0;
518 		ri_pad->rxd_val[i+2] = 0;
519 		ri_pad->rxd_val[i+3] = 0;
520 	}
521 #ifdef __LP64__
522 	ri_pad->rxd_val[RXD_INFO_SIZE-1] = 0;
523 #endif
524 }
525 
526 /*
527  * Only allow a single packet to take up most 1/nth of the tx ring
528  */
529 #define MAX_SINGLE_PACKET_FRACTION 12
530 #define IF_BAD_DMA (bus_addr_t)-1
531 
532 #define CTX_ACTIVE(ctx) ((if_getdrvflags((ctx)->ifc_ifp) & IFF_DRV_RUNNING))
533 
534 #define CTX_LOCK_INIT(_sc)  sx_init(&(_sc)->ifc_ctx_sx, "iflib ctx lock")
535 #define CTX_LOCK(ctx) sx_xlock(&(ctx)->ifc_ctx_sx)
536 #define CTX_UNLOCK(ctx) sx_xunlock(&(ctx)->ifc_ctx_sx)
537 #define CTX_LOCK_DESTROY(ctx) sx_destroy(&(ctx)->ifc_ctx_sx)
538 
539 
540 #define STATE_LOCK_INIT(_sc, _name)  mtx_init(&(_sc)->ifc_state_mtx, _name, "iflib state lock", MTX_DEF)
541 #define STATE_LOCK(ctx) mtx_lock(&(ctx)->ifc_state_mtx)
542 #define STATE_UNLOCK(ctx) mtx_unlock(&(ctx)->ifc_state_mtx)
543 #define STATE_LOCK_DESTROY(ctx) mtx_destroy(&(ctx)->ifc_state_mtx)
544 
545 
546 
547 #define CALLOUT_LOCK(txq)	mtx_lock(&txq->ift_mtx)
548 #define CALLOUT_UNLOCK(txq) 	mtx_unlock(&txq->ift_mtx)
549 
550 void
551 iflib_set_detach(if_ctx_t ctx)
552 {
553 	STATE_LOCK(ctx);
554 	ctx->ifc_flags |= IFC_IN_DETACH;
555 	STATE_UNLOCK(ctx);
556 }
557 
558 /* Our boot-time initialization hook */
559 static int	iflib_module_event_handler(module_t, int, void *);
560 
561 static moduledata_t iflib_moduledata = {
562 	"iflib",
563 	iflib_module_event_handler,
564 	NULL
565 };
566 
567 DECLARE_MODULE(iflib, iflib_moduledata, SI_SUB_INIT_IF, SI_ORDER_ANY);
568 MODULE_VERSION(iflib, 1);
569 
570 MODULE_DEPEND(iflib, pci, 1, 1, 1);
571 MODULE_DEPEND(iflib, ether, 1, 1, 1);
572 
573 TASKQGROUP_DEFINE(if_io_tqg, mp_ncpus, 1);
574 TASKQGROUP_DEFINE(if_config_tqg, 1, 1);
575 
576 #ifndef IFLIB_DEBUG_COUNTERS
577 #ifdef INVARIANTS
578 #define IFLIB_DEBUG_COUNTERS 1
579 #else
580 #define IFLIB_DEBUG_COUNTERS 0
581 #endif /* !INVARIANTS */
582 #endif
583 
584 static SYSCTL_NODE(_net, OID_AUTO, iflib, CTLFLAG_RD, 0,
585                    "iflib driver parameters");
586 
587 /*
588  * XXX need to ensure that this can't accidentally cause the head to be moved backwards
589  */
590 static int iflib_min_tx_latency = 0;
591 SYSCTL_INT(_net_iflib, OID_AUTO, min_tx_latency, CTLFLAG_RW,
592 		   &iflib_min_tx_latency, 0, "minimize transmit latency at the possible expense of throughput");
593 static int iflib_no_tx_batch = 0;
594 SYSCTL_INT(_net_iflib, OID_AUTO, no_tx_batch, CTLFLAG_RW,
595 		   &iflib_no_tx_batch, 0, "minimize transmit latency at the possible expense of throughput");
596 
597 
598 #if IFLIB_DEBUG_COUNTERS
599 
600 static int iflib_tx_seen;
601 static int iflib_tx_sent;
602 static int iflib_tx_encap;
603 static int iflib_rx_allocs;
604 static int iflib_fl_refills;
605 static int iflib_fl_refills_large;
606 static int iflib_tx_frees;
607 
608 SYSCTL_INT(_net_iflib, OID_AUTO, tx_seen, CTLFLAG_RD,
609 		   &iflib_tx_seen, 0, "# tx mbufs seen");
610 SYSCTL_INT(_net_iflib, OID_AUTO, tx_sent, CTLFLAG_RD,
611 		   &iflib_tx_sent, 0, "# tx mbufs sent");
612 SYSCTL_INT(_net_iflib, OID_AUTO, tx_encap, CTLFLAG_RD,
613 		   &iflib_tx_encap, 0, "# tx mbufs encapped");
614 SYSCTL_INT(_net_iflib, OID_AUTO, tx_frees, CTLFLAG_RD,
615 		   &iflib_tx_frees, 0, "# tx frees");
616 SYSCTL_INT(_net_iflib, OID_AUTO, rx_allocs, CTLFLAG_RD,
617 		   &iflib_rx_allocs, 0, "# rx allocations");
618 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills, CTLFLAG_RD,
619 		   &iflib_fl_refills, 0, "# refills");
620 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills_large, CTLFLAG_RD,
621 		   &iflib_fl_refills_large, 0, "# large refills");
622 
623 
624 static int iflib_txq_drain_flushing;
625 static int iflib_txq_drain_oactive;
626 static int iflib_txq_drain_notready;
627 
628 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_flushing, CTLFLAG_RD,
629 		   &iflib_txq_drain_flushing, 0, "# drain flushes");
630 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_oactive, CTLFLAG_RD,
631 		   &iflib_txq_drain_oactive, 0, "# drain oactives");
632 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_notready, CTLFLAG_RD,
633 		   &iflib_txq_drain_notready, 0, "# drain notready");
634 
635 
636 static int iflib_encap_load_mbuf_fail;
637 static int iflib_encap_pad_mbuf_fail;
638 static int iflib_encap_txq_avail_fail;
639 static int iflib_encap_txd_encap_fail;
640 
641 SYSCTL_INT(_net_iflib, OID_AUTO, encap_load_mbuf_fail, CTLFLAG_RD,
642 		   &iflib_encap_load_mbuf_fail, 0, "# busdma load failures");
643 SYSCTL_INT(_net_iflib, OID_AUTO, encap_pad_mbuf_fail, CTLFLAG_RD,
644 		   &iflib_encap_pad_mbuf_fail, 0, "# runt frame pad failures");
645 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txq_avail_fail, CTLFLAG_RD,
646 		   &iflib_encap_txq_avail_fail, 0, "# txq avail failures");
647 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txd_encap_fail, CTLFLAG_RD,
648 		   &iflib_encap_txd_encap_fail, 0, "# driver encap failures");
649 
650 static int iflib_task_fn_rxs;
651 static int iflib_rx_intr_enables;
652 static int iflib_fast_intrs;
653 static int iflib_rx_unavail;
654 static int iflib_rx_ctx_inactive;
655 static int iflib_rx_if_input;
656 static int iflib_rxd_flush;
657 
658 static int iflib_verbose_debug;
659 
660 SYSCTL_INT(_net_iflib, OID_AUTO, task_fn_rx, CTLFLAG_RD,
661 		   &iflib_task_fn_rxs, 0, "# task_fn_rx calls");
662 SYSCTL_INT(_net_iflib, OID_AUTO, rx_intr_enables, CTLFLAG_RD,
663 		   &iflib_rx_intr_enables, 0, "# rx intr enables");
664 SYSCTL_INT(_net_iflib, OID_AUTO, fast_intrs, CTLFLAG_RD,
665 		   &iflib_fast_intrs, 0, "# fast_intr calls");
666 SYSCTL_INT(_net_iflib, OID_AUTO, rx_unavail, CTLFLAG_RD,
667 		   &iflib_rx_unavail, 0, "# times rxeof called with no available data");
668 SYSCTL_INT(_net_iflib, OID_AUTO, rx_ctx_inactive, CTLFLAG_RD,
669 		   &iflib_rx_ctx_inactive, 0, "# times rxeof called with inactive context");
670 SYSCTL_INT(_net_iflib, OID_AUTO, rx_if_input, CTLFLAG_RD,
671 		   &iflib_rx_if_input, 0, "# times rxeof called if_input");
672 SYSCTL_INT(_net_iflib, OID_AUTO, rxd_flush, CTLFLAG_RD,
673 	         &iflib_rxd_flush, 0, "# times rxd_flush called");
674 SYSCTL_INT(_net_iflib, OID_AUTO, verbose_debug, CTLFLAG_RW,
675 		   &iflib_verbose_debug, 0, "enable verbose debugging");
676 
677 #define DBG_COUNTER_INC(name) atomic_add_int(&(iflib_ ## name), 1)
678 static void
679 iflib_debug_reset(void)
680 {
681 	iflib_tx_seen = iflib_tx_sent = iflib_tx_encap = iflib_rx_allocs =
682 		iflib_fl_refills = iflib_fl_refills_large = iflib_tx_frees =
683 		iflib_txq_drain_flushing = iflib_txq_drain_oactive =
684 		iflib_txq_drain_notready =
685 		iflib_encap_load_mbuf_fail = iflib_encap_pad_mbuf_fail =
686 		iflib_encap_txq_avail_fail = iflib_encap_txd_encap_fail =
687 		iflib_task_fn_rxs = iflib_rx_intr_enables = iflib_fast_intrs =
688 		iflib_rx_unavail =
689 		iflib_rx_ctx_inactive = iflib_rx_if_input =
690 		iflib_rxd_flush = 0;
691 }
692 
693 #else
694 #define DBG_COUNTER_INC(name)
695 static void iflib_debug_reset(void) {}
696 #endif
697 
698 #define IFLIB_DEBUG 0
699 
700 static void iflib_tx_structures_free(if_ctx_t ctx);
701 static void iflib_rx_structures_free(if_ctx_t ctx);
702 static int iflib_queues_alloc(if_ctx_t ctx);
703 static int iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq);
704 static int iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget);
705 static int iflib_qset_structures_setup(if_ctx_t ctx);
706 static int iflib_msix_init(if_ctx_t ctx);
707 static int iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filterarg, int *rid, const char *str);
708 static void iflib_txq_check_drain(iflib_txq_t txq, int budget);
709 static uint32_t iflib_txq_can_drain(struct ifmp_ring *);
710 #ifdef ALTQ
711 static void iflib_altq_if_start(if_t ifp);
712 static int iflib_altq_if_transmit(if_t ifp, struct mbuf *m);
713 #endif
714 static int iflib_register(if_ctx_t);
715 static void iflib_init_locked(if_ctx_t ctx);
716 static void iflib_add_device_sysctl_pre(if_ctx_t ctx);
717 static void iflib_add_device_sysctl_post(if_ctx_t ctx);
718 static void iflib_ifmp_purge(iflib_txq_t txq);
719 static void _iflib_pre_assert(if_softc_ctx_t scctx);
720 static void iflib_if_init_locked(if_ctx_t ctx);
721 static void iflib_free_intr_mem(if_ctx_t ctx);
722 #ifndef __NO_STRICT_ALIGNMENT
723 static struct mbuf * iflib_fixup_rx(struct mbuf *m);
724 #endif
725 
726 NETDUMP_DEFINE(iflib);
727 
728 #ifdef DEV_NETMAP
729 #include <sys/selinfo.h>
730 #include <net/netmap.h>
731 #include <dev/netmap/netmap_kern.h>
732 
733 MODULE_DEPEND(iflib, netmap, 1, 1, 1);
734 
735 static int netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, uint32_t nm_i, bool init);
736 
737 /*
738  * device-specific sysctl variables:
739  *
740  * iflib_crcstrip: 0: keep CRC in rx frames (default), 1: strip it.
741  *	During regular operations the CRC is stripped, but on some
742  *	hardware reception of frames not multiple of 64 is slower,
743  *	so using crcstrip=0 helps in benchmarks.
744  *
745  * iflib_rx_miss, iflib_rx_miss_bufs:
746  *	count packets that might be missed due to lost interrupts.
747  */
748 SYSCTL_DECL(_dev_netmap);
749 /*
750  * The xl driver by default strips CRCs and we do not override it.
751  */
752 
753 int iflib_crcstrip = 1;
754 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_crcstrip,
755     CTLFLAG_RW, &iflib_crcstrip, 1, "strip CRC on rx frames");
756 
757 int iflib_rx_miss, iflib_rx_miss_bufs;
758 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss,
759     CTLFLAG_RW, &iflib_rx_miss, 0, "potentially missed rx intr");
760 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss_bufs,
761     CTLFLAG_RW, &iflib_rx_miss_bufs, 0, "potentially missed rx intr bufs");
762 
763 /*
764  * Register/unregister. We are already under netmap lock.
765  * Only called on the first register or the last unregister.
766  */
767 static int
768 iflib_netmap_register(struct netmap_adapter *na, int onoff)
769 {
770 	struct ifnet *ifp = na->ifp;
771 	if_ctx_t ctx = ifp->if_softc;
772 	int status;
773 
774 	CTX_LOCK(ctx);
775 	IFDI_INTR_DISABLE(ctx);
776 
777 	/* Tell the stack that the interface is no longer active */
778 	ifp->if_drv_flags &= ~(IFF_DRV_RUNNING | IFF_DRV_OACTIVE);
779 
780 	if (!CTX_IS_VF(ctx))
781 		IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip);
782 
783 	/* enable or disable flags and callbacks in na and ifp */
784 	if (onoff) {
785 		nm_set_native_flags(na);
786 	} else {
787 		nm_clear_native_flags(na);
788 	}
789 	iflib_stop(ctx);
790 	iflib_init_locked(ctx);
791 	IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip); // XXX why twice ?
792 	status = ifp->if_drv_flags & IFF_DRV_RUNNING ? 0 : 1;
793 	if (status)
794 		nm_clear_native_flags(na);
795 	CTX_UNLOCK(ctx);
796 	return (status);
797 }
798 
799 static int
800 netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, uint32_t nm_i, bool init)
801 {
802 	struct netmap_adapter *na = kring->na;
803 	u_int const lim = kring->nkr_num_slots - 1;
804 	u_int head = kring->rhead;
805 	struct netmap_ring *ring = kring->ring;
806 	bus_dmamap_t *map;
807 	struct if_rxd_update iru;
808 	if_ctx_t ctx = rxq->ifr_ctx;
809 	iflib_fl_t fl = &rxq->ifr_fl[0];
810 	uint32_t refill_pidx, nic_i;
811 #if IFLIB_DEBUG_COUNTERS
812 	int rf_count = 0;
813 #endif
814 
815 	if (nm_i == head && __predict_true(!init))
816 		return 0;
817 	iru_init(&iru, rxq, 0 /* flid */);
818 	map = fl->ifl_sds.ifsd_map;
819 	refill_pidx = netmap_idx_k2n(kring, nm_i);
820 	/*
821 	 * IMPORTANT: we must leave one free slot in the ring,
822 	 * so move head back by one unit
823 	 */
824 	head = nm_prev(head, lim);
825 	nic_i = UINT_MAX;
826 	DBG_COUNTER_INC(fl_refills);
827 	while (nm_i != head) {
828 #if IFLIB_DEBUG_COUNTERS
829 		if (++rf_count == 9)
830 			DBG_COUNTER_INC(fl_refills_large);
831 #endif
832 		for (int tmp_pidx = 0; tmp_pidx < IFLIB_MAX_RX_REFRESH && nm_i != head; tmp_pidx++) {
833 			struct netmap_slot *slot = &ring->slot[nm_i];
834 			void *addr = PNMB(na, slot, &fl->ifl_bus_addrs[tmp_pidx]);
835 			uint32_t nic_i_dma = refill_pidx;
836 			nic_i = netmap_idx_k2n(kring, nm_i);
837 
838 			MPASS(tmp_pidx < IFLIB_MAX_RX_REFRESH);
839 
840 			if (addr == NETMAP_BUF_BASE(na)) /* bad buf */
841 			        return netmap_ring_reinit(kring);
842 
843 			fl->ifl_vm_addrs[tmp_pidx] = addr;
844 			if (__predict_false(init)) {
845 				netmap_load_map(na, fl->ifl_buf_tag,
846 				    map[nic_i], addr);
847 			} else if (slot->flags & NS_BUF_CHANGED) {
848 				/* buffer has changed, reload map */
849 				netmap_reload_map(na, fl->ifl_buf_tag,
850 				    map[nic_i], addr);
851 			}
852 			slot->flags &= ~NS_BUF_CHANGED;
853 
854 			nm_i = nm_next(nm_i, lim);
855 			fl->ifl_rxd_idxs[tmp_pidx] = nic_i = nm_next(nic_i, lim);
856 			if (nm_i != head && tmp_pidx < IFLIB_MAX_RX_REFRESH-1)
857 				continue;
858 
859 			iru.iru_pidx = refill_pidx;
860 			iru.iru_count = tmp_pidx+1;
861 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
862 			refill_pidx = nic_i;
863 			for (int n = 0; n < iru.iru_count; n++) {
864 				bus_dmamap_sync(fl->ifl_buf_tag, map[nic_i_dma],
865 						BUS_DMASYNC_PREREAD);
866 				/* XXX - change this to not use the netmap func*/
867 				nic_i_dma = nm_next(nic_i_dma, lim);
868 			}
869 		}
870 	}
871 	kring->nr_hwcur = head;
872 
873 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
874 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
875 	if (__predict_true(nic_i != UINT_MAX)) {
876 		ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id, nic_i);
877 		DBG_COUNTER_INC(rxd_flush);
878 	}
879 	return (0);
880 }
881 
882 /*
883  * Reconcile kernel and user view of the transmit ring.
884  *
885  * All information is in the kring.
886  * Userspace wants to send packets up to the one before kring->rhead,
887  * kernel knows kring->nr_hwcur is the first unsent packet.
888  *
889  * Here we push packets out (as many as possible), and possibly
890  * reclaim buffers from previously completed transmission.
891  *
892  * The caller (netmap) guarantees that there is only one instance
893  * running at any time. Any interference with other driver
894  * methods should be handled by the individual drivers.
895  */
896 static int
897 iflib_netmap_txsync(struct netmap_kring *kring, int flags)
898 {
899 	struct netmap_adapter *na = kring->na;
900 	struct ifnet *ifp = na->ifp;
901 	struct netmap_ring *ring = kring->ring;
902 	u_int nm_i;	/* index into the netmap kring */
903 	u_int nic_i;	/* index into the NIC ring */
904 	u_int n;
905 	u_int const lim = kring->nkr_num_slots - 1;
906 	u_int const head = kring->rhead;
907 	struct if_pkt_info pi;
908 
909 	/*
910 	 * interrupts on every tx packet are expensive so request
911 	 * them every half ring, or where NS_REPORT is set
912 	 */
913 	u_int report_frequency = kring->nkr_num_slots >> 1;
914 	/* device-specific */
915 	if_ctx_t ctx = ifp->if_softc;
916 	iflib_txq_t txq = &ctx->ifc_txqs[kring->ring_id];
917 
918 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
919 	    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
920 
921 	/*
922 	 * First part: process new packets to send.
923 	 * nm_i is the current index in the netmap kring,
924 	 * nic_i is the corresponding index in the NIC ring.
925 	 *
926 	 * If we have packets to send (nm_i != head)
927 	 * iterate over the netmap ring, fetch length and update
928 	 * the corresponding slot in the NIC ring. Some drivers also
929 	 * need to update the buffer's physical address in the NIC slot
930 	 * even NS_BUF_CHANGED is not set (PNMB computes the addresses).
931 	 *
932 	 * The netmap_reload_map() calls is especially expensive,
933 	 * even when (as in this case) the tag is 0, so do only
934 	 * when the buffer has actually changed.
935 	 *
936 	 * If possible do not set the report/intr bit on all slots,
937 	 * but only a few times per ring or when NS_REPORT is set.
938 	 *
939 	 * Finally, on 10G and faster drivers, it might be useful
940 	 * to prefetch the next slot and txr entry.
941 	 */
942 
943 	nm_i = kring->nr_hwcur;
944 	if (nm_i != head) {	/* we have new packets to send */
945 		pkt_info_zero(&pi);
946 		pi.ipi_segs = txq->ift_segs;
947 		pi.ipi_qsidx = kring->ring_id;
948 		nic_i = netmap_idx_k2n(kring, nm_i);
949 
950 		__builtin_prefetch(&ring->slot[nm_i]);
951 		__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i]);
952 		__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i]);
953 
954 		for (n = 0; nm_i != head; n++) {
955 			struct netmap_slot *slot = &ring->slot[nm_i];
956 			u_int len = slot->len;
957 			uint64_t paddr;
958 			void *addr = PNMB(na, slot, &paddr);
959 			int flags = (slot->flags & NS_REPORT ||
960 				nic_i == 0 || nic_i == report_frequency) ?
961 				IPI_TX_INTR : 0;
962 
963 			/* device-specific */
964 			pi.ipi_len = len;
965 			pi.ipi_segs[0].ds_addr = paddr;
966 			pi.ipi_segs[0].ds_len = len;
967 			pi.ipi_nsegs = 1;
968 			pi.ipi_ndescs = 0;
969 			pi.ipi_pidx = nic_i;
970 			pi.ipi_flags = flags;
971 
972 			/* Fill the slot in the NIC ring. */
973 			ctx->isc_txd_encap(ctx->ifc_softc, &pi);
974 			DBG_COUNTER_INC(tx_encap);
975 
976 			/* prefetch for next round */
977 			__builtin_prefetch(&ring->slot[nm_i + 1]);
978 			__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i + 1]);
979 			__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i + 1]);
980 
981 			NM_CHECK_ADDR_LEN(na, addr, len);
982 
983 			if (slot->flags & NS_BUF_CHANGED) {
984 				/* buffer has changed, reload map */
985 				netmap_reload_map(na, txq->ift_buf_tag,
986 				    txq->ift_sds.ifsd_map[nic_i], addr);
987 			}
988 			/* make sure changes to the buffer are synced */
989 			bus_dmamap_sync(txq->ift_buf_tag,
990 			    txq->ift_sds.ifsd_map[nic_i],
991 			    BUS_DMASYNC_PREWRITE);
992 
993 			slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED);
994 			nm_i = nm_next(nm_i, lim);
995 			nic_i = nm_next(nic_i, lim);
996 		}
997 		kring->nr_hwcur = nm_i;
998 
999 		/* synchronize the NIC ring */
1000 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1001 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1002 
1003 		/* (re)start the tx unit up to slot nic_i (excluded) */
1004 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, nic_i);
1005 	}
1006 
1007 	/*
1008 	 * Second part: reclaim buffers for completed transmissions.
1009 	 *
1010 	 * If there are unclaimed buffers, attempt to reclaim them.
1011 	 * If none are reclaimed, and TX IRQs are not in use, do an initial
1012 	 * minimal delay, then trigger the tx handler which will spin in the
1013 	 * group task queue.
1014 	 */
1015 	if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1016 		if (iflib_tx_credits_update(ctx, txq)) {
1017 			/* some tx completed, increment avail */
1018 			nic_i = txq->ift_cidx_processed;
1019 			kring->nr_hwtail = nm_prev(netmap_idx_n2k(kring, nic_i), lim);
1020 		}
1021 	}
1022 	if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ))
1023 		if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1024 			callout_reset_on(&txq->ift_timer, hz < 2000 ? 1 : hz / 1000,
1025 			    iflib_timer, txq, txq->ift_timer.c_cpu);
1026 	}
1027 	return (0);
1028 }
1029 
1030 /*
1031  * Reconcile kernel and user view of the receive ring.
1032  * Same as for the txsync, this routine must be efficient.
1033  * The caller guarantees a single invocations, but races against
1034  * the rest of the driver should be handled here.
1035  *
1036  * On call, kring->rhead is the first packet that userspace wants
1037  * to keep, and kring->rcur is the wakeup point.
1038  * The kernel has previously reported packets up to kring->rtail.
1039  *
1040  * If (flags & NAF_FORCE_READ) also check for incoming packets irrespective
1041  * of whether or not we received an interrupt.
1042  */
1043 static int
1044 iflib_netmap_rxsync(struct netmap_kring *kring, int flags)
1045 {
1046 	struct netmap_adapter *na = kring->na;
1047 	struct netmap_ring *ring = kring->ring;
1048 	iflib_fl_t fl;
1049 	uint32_t nm_i;	/* index into the netmap ring */
1050 	uint32_t nic_i;	/* index into the NIC ring */
1051 	u_int i, n;
1052 	u_int const lim = kring->nkr_num_slots - 1;
1053 	u_int const head = kring->rhead;
1054 	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
1055 	struct if_rxd_info ri;
1056 
1057 	struct ifnet *ifp = na->ifp;
1058 	if_ctx_t ctx = ifp->if_softc;
1059 	iflib_rxq_t rxq = &ctx->ifc_rxqs[kring->ring_id];
1060 	if (head > lim)
1061 		return netmap_ring_reinit(kring);
1062 
1063 	/*
1064 	 * XXX netmap_fl_refill() only ever (re)fills free list 0 so far.
1065 	 */
1066 
1067 	for (i = 0, fl = rxq->ifr_fl; i < rxq->ifr_nfl; i++, fl++) {
1068 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
1069 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1070 	}
1071 
1072 	/*
1073 	 * First part: import newly received packets.
1074 	 *
1075 	 * nm_i is the index of the next free slot in the netmap ring,
1076 	 * nic_i is the index of the next received packet in the NIC ring,
1077 	 * and they may differ in case if_init() has been called while
1078 	 * in netmap mode. For the receive ring we have
1079 	 *
1080 	 *	nic_i = rxr->next_check;
1081 	 *	nm_i = kring->nr_hwtail (previous)
1082 	 * and
1083 	 *	nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1084 	 *
1085 	 * rxr->next_check is set to 0 on a ring reinit
1086 	 */
1087 	if (netmap_no_pendintr || force_update) {
1088 		int crclen = iflib_crcstrip ? 0 : 4;
1089 		int error, avail;
1090 
1091 		for (i = 0; i < rxq->ifr_nfl; i++) {
1092 			fl = &rxq->ifr_fl[i];
1093 			nic_i = fl->ifl_cidx;
1094 			nm_i = netmap_idx_n2k(kring, nic_i);
1095 			avail = ctx->isc_rxd_available(ctx->ifc_softc,
1096 			    rxq->ifr_id, nic_i, USHRT_MAX);
1097 			for (n = 0; avail > 0; n++, avail--) {
1098 				rxd_info_zero(&ri);
1099 				ri.iri_frags = rxq->ifr_frags;
1100 				ri.iri_qsidx = kring->ring_id;
1101 				ri.iri_ifp = ctx->ifc_ifp;
1102 				ri.iri_cidx = nic_i;
1103 
1104 				error = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
1105 				ring->slot[nm_i].len = error ? 0 : ri.iri_len - crclen;
1106 				ring->slot[nm_i].flags = 0;
1107 				bus_dmamap_sync(fl->ifl_buf_tag,
1108 				    fl->ifl_sds.ifsd_map[nic_i], BUS_DMASYNC_POSTREAD);
1109 				nm_i = nm_next(nm_i, lim);
1110 				nic_i = nm_next(nic_i, lim);
1111 			}
1112 			if (n) { /* update the state variables */
1113 				if (netmap_no_pendintr && !force_update) {
1114 					/* diagnostics */
1115 					iflib_rx_miss ++;
1116 					iflib_rx_miss_bufs += n;
1117 				}
1118 				fl->ifl_cidx = nic_i;
1119 				kring->nr_hwtail = nm_i;
1120 			}
1121 			kring->nr_kflags &= ~NKR_PENDINTR;
1122 		}
1123 	}
1124 	/*
1125 	 * Second part: skip past packets that userspace has released.
1126 	 * (kring->nr_hwcur to head excluded),
1127 	 * and make the buffers available for reception.
1128 	 * As usual nm_i is the index in the netmap ring,
1129 	 * nic_i is the index in the NIC ring, and
1130 	 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1131 	 */
1132 	/* XXX not sure how this will work with multiple free lists */
1133 	nm_i = kring->nr_hwcur;
1134 
1135 	return (netmap_fl_refill(rxq, kring, nm_i, false));
1136 }
1137 
1138 static void
1139 iflib_netmap_intr(struct netmap_adapter *na, int onoff)
1140 {
1141 	struct ifnet *ifp = na->ifp;
1142 	if_ctx_t ctx = ifp->if_softc;
1143 
1144 	CTX_LOCK(ctx);
1145 	if (onoff) {
1146 		IFDI_INTR_ENABLE(ctx);
1147 	} else {
1148 		IFDI_INTR_DISABLE(ctx);
1149 	}
1150 	CTX_UNLOCK(ctx);
1151 }
1152 
1153 
1154 static int
1155 iflib_netmap_attach(if_ctx_t ctx)
1156 {
1157 	struct netmap_adapter na;
1158 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1159 
1160 	bzero(&na, sizeof(na));
1161 
1162 	na.ifp = ctx->ifc_ifp;
1163 	na.na_flags = NAF_BDG_MAYSLEEP;
1164 	MPASS(ctx->ifc_softc_ctx.isc_ntxqsets);
1165 	MPASS(ctx->ifc_softc_ctx.isc_nrxqsets);
1166 
1167 	na.num_tx_desc = scctx->isc_ntxd[0];
1168 	na.num_rx_desc = scctx->isc_nrxd[0];
1169 	na.nm_txsync = iflib_netmap_txsync;
1170 	na.nm_rxsync = iflib_netmap_rxsync;
1171 	na.nm_register = iflib_netmap_register;
1172 	na.nm_intr = iflib_netmap_intr;
1173 	na.num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
1174 	na.num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
1175 	return (netmap_attach(&na));
1176 }
1177 
1178 static void
1179 iflib_netmap_txq_init(if_ctx_t ctx, iflib_txq_t txq)
1180 {
1181 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1182 	struct netmap_slot *slot;
1183 
1184 	slot = netmap_reset(na, NR_TX, txq->ift_id, 0);
1185 	if (slot == NULL)
1186 		return;
1187 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxd[0]; i++) {
1188 
1189 		/*
1190 		 * In netmap mode, set the map for the packet buffer.
1191 		 * NOTE: Some drivers (not this one) also need to set
1192 		 * the physical buffer address in the NIC ring.
1193 		 * netmap_idx_n2k() maps a nic index, i, into the corresponding
1194 		 * netmap slot index, si
1195 		 */
1196 		int si = netmap_idx_n2k(na->tx_rings[txq->ift_id], i);
1197 		netmap_load_map(na, txq->ift_buf_tag, txq->ift_sds.ifsd_map[i],
1198 		    NMB(na, slot + si));
1199 	}
1200 }
1201 
1202 static void
1203 iflib_netmap_rxq_init(if_ctx_t ctx, iflib_rxq_t rxq)
1204 {
1205 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1206 	struct netmap_kring *kring = na->rx_rings[rxq->ifr_id];
1207 	struct netmap_slot *slot;
1208 	uint32_t nm_i;
1209 
1210 	slot = netmap_reset(na, NR_RX, rxq->ifr_id, 0);
1211 	if (slot == NULL)
1212 		return;
1213 	nm_i = netmap_idx_n2k(kring, 0);
1214 	netmap_fl_refill(rxq, kring, nm_i, true);
1215 }
1216 
1217 static void
1218 iflib_netmap_timer_adjust(if_ctx_t ctx, iflib_txq_t txq, uint32_t *reset_on)
1219 {
1220 	struct netmap_kring *kring;
1221 	uint16_t txqid;
1222 
1223 	txqid = txq->ift_id;
1224 	kring = NA(ctx->ifc_ifp)->tx_rings[txqid];
1225 
1226 	if (kring->nr_hwcur != nm_next(kring->nr_hwtail, kring->nkr_num_slots - 1)) {
1227 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1228 		    BUS_DMASYNC_POSTREAD);
1229 		if (ctx->isc_txd_credits_update(ctx->ifc_softc, txqid, false))
1230 			netmap_tx_irq(ctx->ifc_ifp, txqid);
1231 		if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ)) {
1232 			if (hz < 2000)
1233 				*reset_on = 1;
1234 			else
1235 				*reset_on = hz / 1000;
1236 		}
1237 	}
1238 }
1239 
1240 #define iflib_netmap_detach(ifp) netmap_detach(ifp)
1241 
1242 #else
1243 #define iflib_netmap_txq_init(ctx, txq)
1244 #define iflib_netmap_rxq_init(ctx, rxq)
1245 #define iflib_netmap_detach(ifp)
1246 
1247 #define iflib_netmap_attach(ctx) (0)
1248 #define netmap_rx_irq(ifp, qid, budget) (0)
1249 #define netmap_tx_irq(ifp, qid) do {} while (0)
1250 #define iflib_netmap_timer_adjust(ctx, txq, reset_on)
1251 
1252 #endif
1253 
1254 #if defined(__i386__) || defined(__amd64__)
1255 static __inline void
1256 prefetch(void *x)
1257 {
1258 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1259 }
1260 static __inline void
1261 prefetch2cachelines(void *x)
1262 {
1263 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1264 #if (CACHE_LINE_SIZE < 128)
1265 	__asm volatile("prefetcht0 %0" :: "m" (*(((unsigned long *)x)+CACHE_LINE_SIZE/(sizeof(unsigned long)))));
1266 #endif
1267 }
1268 #else
1269 #define prefetch(x)
1270 #define prefetch2cachelines(x)
1271 #endif
1272 
1273 static void
1274 iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid)
1275 {
1276 	iflib_fl_t fl;
1277 
1278 	fl = &rxq->ifr_fl[flid];
1279 	iru->iru_paddrs = fl->ifl_bus_addrs;
1280 	iru->iru_vaddrs = &fl->ifl_vm_addrs[0];
1281 	iru->iru_idxs = fl->ifl_rxd_idxs;
1282 	iru->iru_qsidx = rxq->ifr_id;
1283 	iru->iru_buf_size = fl->ifl_buf_size;
1284 	iru->iru_flidx = fl->ifl_id;
1285 }
1286 
1287 static void
1288 _iflib_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int err)
1289 {
1290 	if (err)
1291 		return;
1292 	*(bus_addr_t *) arg = segs[0].ds_addr;
1293 }
1294 
1295 int
1296 iflib_dma_alloc_align(if_ctx_t ctx, int size, int align, iflib_dma_info_t dma, int mapflags)
1297 {
1298 	int err;
1299 	device_t dev = ctx->ifc_dev;
1300 
1301 	err = bus_dma_tag_create(bus_get_dma_tag(dev),	/* parent */
1302 				align, 0,		/* alignment, bounds */
1303 				BUS_SPACE_MAXADDR,	/* lowaddr */
1304 				BUS_SPACE_MAXADDR,	/* highaddr */
1305 				NULL, NULL,		/* filter, filterarg */
1306 				size,			/* maxsize */
1307 				1,			/* nsegments */
1308 				size,			/* maxsegsize */
1309 				BUS_DMA_ALLOCNOW,	/* flags */
1310 				NULL,			/* lockfunc */
1311 				NULL,			/* lockarg */
1312 				&dma->idi_tag);
1313 	if (err) {
1314 		device_printf(dev,
1315 		    "%s: bus_dma_tag_create failed: %d\n",
1316 		    __func__, err);
1317 		goto fail_0;
1318 	}
1319 
1320 	err = bus_dmamem_alloc(dma->idi_tag, (void**) &dma->idi_vaddr,
1321 	    BUS_DMA_NOWAIT | BUS_DMA_COHERENT | BUS_DMA_ZERO, &dma->idi_map);
1322 	if (err) {
1323 		device_printf(dev,
1324 		    "%s: bus_dmamem_alloc(%ju) failed: %d\n",
1325 		    __func__, (uintmax_t)size, err);
1326 		goto fail_1;
1327 	}
1328 
1329 	dma->idi_paddr = IF_BAD_DMA;
1330 	err = bus_dmamap_load(dma->idi_tag, dma->idi_map, dma->idi_vaddr,
1331 	    size, _iflib_dmamap_cb, &dma->idi_paddr, mapflags | BUS_DMA_NOWAIT);
1332 	if (err || dma->idi_paddr == IF_BAD_DMA) {
1333 		device_printf(dev,
1334 		    "%s: bus_dmamap_load failed: %d\n",
1335 		    __func__, err);
1336 		goto fail_2;
1337 	}
1338 
1339 	dma->idi_size = size;
1340 	return (0);
1341 
1342 fail_2:
1343 	bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1344 fail_1:
1345 	bus_dma_tag_destroy(dma->idi_tag);
1346 fail_0:
1347 	dma->idi_tag = NULL;
1348 
1349 	return (err);
1350 }
1351 
1352 int
1353 iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags)
1354 {
1355 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1356 
1357 	KASSERT(sctx->isc_q_align != 0, ("alignment value not initialized"));
1358 
1359 	return (iflib_dma_alloc_align(ctx, size, sctx->isc_q_align, dma, mapflags));
1360 }
1361 
1362 int
1363 iflib_dma_alloc_multi(if_ctx_t ctx, int *sizes, iflib_dma_info_t *dmalist, int mapflags, int count)
1364 {
1365 	int i, err;
1366 	iflib_dma_info_t *dmaiter;
1367 
1368 	dmaiter = dmalist;
1369 	for (i = 0; i < count; i++, dmaiter++) {
1370 		if ((err = iflib_dma_alloc(ctx, sizes[i], *dmaiter, mapflags)) != 0)
1371 			break;
1372 	}
1373 	if (err)
1374 		iflib_dma_free_multi(dmalist, i);
1375 	return (err);
1376 }
1377 
1378 void
1379 iflib_dma_free(iflib_dma_info_t dma)
1380 {
1381 	if (dma->idi_tag == NULL)
1382 		return;
1383 	if (dma->idi_paddr != IF_BAD_DMA) {
1384 		bus_dmamap_sync(dma->idi_tag, dma->idi_map,
1385 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1386 		bus_dmamap_unload(dma->idi_tag, dma->idi_map);
1387 		dma->idi_paddr = IF_BAD_DMA;
1388 	}
1389 	if (dma->idi_vaddr != NULL) {
1390 		bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1391 		dma->idi_vaddr = NULL;
1392 	}
1393 	bus_dma_tag_destroy(dma->idi_tag);
1394 	dma->idi_tag = NULL;
1395 }
1396 
1397 void
1398 iflib_dma_free_multi(iflib_dma_info_t *dmalist, int count)
1399 {
1400 	int i;
1401 	iflib_dma_info_t *dmaiter = dmalist;
1402 
1403 	for (i = 0; i < count; i++, dmaiter++)
1404 		iflib_dma_free(*dmaiter);
1405 }
1406 
1407 #ifdef EARLY_AP_STARTUP
1408 static const int iflib_started = 1;
1409 #else
1410 /*
1411  * We used to abuse the smp_started flag to decide if the queues have been
1412  * fully initialized (by late taskqgroup_adjust() calls in a SYSINIT()).
1413  * That gave bad races, since the SYSINIT() runs strictly after smp_started
1414  * is set.  Run a SYSINIT() strictly after that to just set a usable
1415  * completion flag.
1416  */
1417 
1418 static int iflib_started;
1419 
1420 static void
1421 iflib_record_started(void *arg)
1422 {
1423 	iflib_started = 1;
1424 }
1425 
1426 SYSINIT(iflib_record_started, SI_SUB_SMP + 1, SI_ORDER_FIRST,
1427 	iflib_record_started, NULL);
1428 #endif
1429 
1430 static int
1431 iflib_fast_intr(void *arg)
1432 {
1433 	iflib_filter_info_t info = arg;
1434 	struct grouptask *gtask = info->ifi_task;
1435 	int result;
1436 
1437 	if (!iflib_started)
1438 		return (FILTER_STRAY);
1439 
1440 	DBG_COUNTER_INC(fast_intrs);
1441 	if (info->ifi_filter != NULL) {
1442 		result = info->ifi_filter(info->ifi_filter_arg);
1443 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1444 			return (result);
1445 	}
1446 
1447 	GROUPTASK_ENQUEUE(gtask);
1448 	return (FILTER_HANDLED);
1449 }
1450 
1451 static int
1452 iflib_fast_intr_rxtx(void *arg)
1453 {
1454 	iflib_filter_info_t info = arg;
1455 	struct grouptask *gtask = info->ifi_task;
1456 	if_ctx_t ctx;
1457 	iflib_rxq_t rxq = (iflib_rxq_t)info->ifi_ctx;
1458 	iflib_txq_t txq;
1459 	void *sc;
1460 	int i, cidx, result;
1461 	qidx_t txqid;
1462 
1463 	if (!iflib_started)
1464 		return (FILTER_STRAY);
1465 
1466 	DBG_COUNTER_INC(fast_intrs);
1467 	if (info->ifi_filter != NULL) {
1468 		result = info->ifi_filter(info->ifi_filter_arg);
1469 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1470 			return (result);
1471 	}
1472 
1473 	ctx = rxq->ifr_ctx;
1474 	sc = ctx->ifc_softc;
1475 	MPASS(rxq->ifr_ntxqirq);
1476 	for (i = 0; i < rxq->ifr_ntxqirq; i++) {
1477 		txqid = rxq->ifr_txqid[i];
1478 		txq = &ctx->ifc_txqs[txqid];
1479 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1480 		    BUS_DMASYNC_POSTREAD);
1481 		if (!ctx->isc_txd_credits_update(sc, txqid, false)) {
1482 			IFDI_TX_QUEUE_INTR_ENABLE(ctx, txqid);
1483 			continue;
1484 		}
1485 		GROUPTASK_ENQUEUE(&txq->ift_task);
1486 	}
1487 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_RXCQ)
1488 		cidx = rxq->ifr_cq_cidx;
1489 	else
1490 		cidx = rxq->ifr_fl[0].ifl_cidx;
1491 	if (iflib_rxd_avail(ctx, rxq, cidx, 1))
1492 		GROUPTASK_ENQUEUE(gtask);
1493 	else {
1494 		IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
1495 		DBG_COUNTER_INC(rx_intr_enables);
1496 	}
1497 	return (FILTER_HANDLED);
1498 }
1499 
1500 
1501 static int
1502 iflib_fast_intr_ctx(void *arg)
1503 {
1504 	iflib_filter_info_t info = arg;
1505 	struct grouptask *gtask = info->ifi_task;
1506 	int result;
1507 
1508 	if (!iflib_started)
1509 		return (FILTER_STRAY);
1510 
1511 	DBG_COUNTER_INC(fast_intrs);
1512 	if (info->ifi_filter != NULL) {
1513 		result = info->ifi_filter(info->ifi_filter_arg);
1514 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1515 			return (result);
1516 	}
1517 
1518 	GROUPTASK_ENQUEUE(gtask);
1519 	return (FILTER_HANDLED);
1520 }
1521 
1522 static int
1523 _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
1524 		 driver_filter_t filter, driver_intr_t handler, void *arg,
1525 		 const char *name)
1526 {
1527 	int rc, flags;
1528 	struct resource *res;
1529 	void *tag = NULL;
1530 	device_t dev = ctx->ifc_dev;
1531 
1532 	flags = RF_ACTIVE;
1533 	if (ctx->ifc_flags & IFC_LEGACY)
1534 		flags |= RF_SHAREABLE;
1535 	MPASS(rid < 512);
1536 	irq->ii_rid = rid;
1537 	res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &irq->ii_rid, flags);
1538 	if (res == NULL) {
1539 		device_printf(dev,
1540 		    "failed to allocate IRQ for rid %d, name %s.\n", rid, name);
1541 		return (ENOMEM);
1542 	}
1543 	irq->ii_res = res;
1544 	KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL"));
1545 	rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET,
1546 						filter, handler, arg, &tag);
1547 	if (rc != 0) {
1548 		device_printf(dev,
1549 		    "failed to setup interrupt for rid %d, name %s: %d\n",
1550 					  rid, name ? name : "unknown", rc);
1551 		return (rc);
1552 	} else if (name)
1553 		bus_describe_intr(dev, res, tag, "%s", name);
1554 
1555 	irq->ii_tag = tag;
1556 	return (0);
1557 }
1558 
1559 
1560 /*********************************************************************
1561  *
1562  *  Allocate DMA resources for TX buffers as well as memory for the TX
1563  *  mbuf map.  TX DMA maps (non-TSO/TSO) and TX mbuf map are kept in a
1564  *  iflib_sw_tx_desc_array structure, storing all the information that
1565  *  is needed to transmit a packet on the wire.  This is called only
1566  *  once at attach, setup is done every reset.
1567  *
1568  **********************************************************************/
1569 static int
1570 iflib_txsd_alloc(iflib_txq_t txq)
1571 {
1572 	if_ctx_t ctx = txq->ift_ctx;
1573 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1574 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1575 	device_t dev = ctx->ifc_dev;
1576 	bus_size_t tsomaxsize;
1577 	int err, nsegments, ntsosegments;
1578 	bool tso;
1579 
1580 	nsegments = scctx->isc_tx_nsegments;
1581 	ntsosegments = scctx->isc_tx_tso_segments_max;
1582 	tsomaxsize = scctx->isc_tx_tso_size_max;
1583 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_VLAN_MTU)
1584 		tsomaxsize += sizeof(struct ether_vlan_header);
1585 	MPASS(scctx->isc_ntxd[0] > 0);
1586 	MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0);
1587 	MPASS(nsegments > 0);
1588 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) {
1589 		MPASS(ntsosegments > 0);
1590 		MPASS(sctx->isc_tso_maxsize >= tsomaxsize);
1591 	}
1592 
1593 	/*
1594 	 * Set up DMA tags for TX buffers.
1595 	 */
1596 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1597 			       1, 0,			/* alignment, bounds */
1598 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1599 			       BUS_SPACE_MAXADDR,	/* highaddr */
1600 			       NULL, NULL,		/* filter, filterarg */
1601 			       sctx->isc_tx_maxsize,		/* maxsize */
1602 			       nsegments,	/* nsegments */
1603 			       sctx->isc_tx_maxsegsize,	/* maxsegsize */
1604 			       0,			/* flags */
1605 			       NULL,			/* lockfunc */
1606 			       NULL,			/* lockfuncarg */
1607 			       &txq->ift_buf_tag))) {
1608 		device_printf(dev,"Unable to allocate TX DMA tag: %d\n", err);
1609 		device_printf(dev,"maxsize: %ju nsegments: %d maxsegsize: %ju\n",
1610 		    (uintmax_t)sctx->isc_tx_maxsize, nsegments, (uintmax_t)sctx->isc_tx_maxsegsize);
1611 		goto fail;
1612 	}
1613 	tso = (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) != 0;
1614 	if (tso && (err = bus_dma_tag_create(bus_get_dma_tag(dev),
1615 			       1, 0,			/* alignment, bounds */
1616 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1617 			       BUS_SPACE_MAXADDR,	/* highaddr */
1618 			       NULL, NULL,		/* filter, filterarg */
1619 			       tsomaxsize,		/* maxsize */
1620 			       ntsosegments,	/* nsegments */
1621 			       sctx->isc_tso_maxsegsize,/* maxsegsize */
1622 			       0,			/* flags */
1623 			       NULL,			/* lockfunc */
1624 			       NULL,			/* lockfuncarg */
1625 			       &txq->ift_tso_buf_tag))) {
1626 		device_printf(dev, "Unable to allocate TSO TX DMA tag: %d\n",
1627 		    err);
1628 		goto fail;
1629 	}
1630 
1631 	/* Allocate memory for the TX mbuf map. */
1632 	if (!(txq->ift_sds.ifsd_m =
1633 	    (struct mbuf **) malloc(sizeof(struct mbuf *) *
1634 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1635 		device_printf(dev, "Unable to allocate TX mbuf map memory\n");
1636 		err = ENOMEM;
1637 		goto fail;
1638 	}
1639 
1640 	/*
1641 	 * Create the DMA maps for TX buffers.
1642 	 */
1643 	if ((txq->ift_sds.ifsd_map = (bus_dmamap_t *)malloc(
1644 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1645 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1646 		device_printf(dev,
1647 		    "Unable to allocate TX buffer DMA map memory\n");
1648 		err = ENOMEM;
1649 		goto fail;
1650 	}
1651 	if (tso && (txq->ift_sds.ifsd_tso_map = (bus_dmamap_t *)malloc(
1652 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1653 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1654 		device_printf(dev,
1655 		    "Unable to allocate TSO TX buffer map memory\n");
1656 		err = ENOMEM;
1657 		goto fail;
1658 	}
1659 	for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) {
1660 		err = bus_dmamap_create(txq->ift_buf_tag, 0,
1661 		    &txq->ift_sds.ifsd_map[i]);
1662 		if (err != 0) {
1663 			device_printf(dev, "Unable to create TX DMA map\n");
1664 			goto fail;
1665 		}
1666 		if (!tso)
1667 			continue;
1668 		err = bus_dmamap_create(txq->ift_tso_buf_tag, 0,
1669 		    &txq->ift_sds.ifsd_tso_map[i]);
1670 		if (err != 0) {
1671 			device_printf(dev, "Unable to create TSO TX DMA map\n");
1672 			goto fail;
1673 		}
1674 	}
1675 	return (0);
1676 fail:
1677 	/* We free all, it handles case where we are in the middle */
1678 	iflib_tx_structures_free(ctx);
1679 	return (err);
1680 }
1681 
1682 static void
1683 iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i)
1684 {
1685 	bus_dmamap_t map;
1686 
1687 	map = NULL;
1688 	if (txq->ift_sds.ifsd_map != NULL)
1689 		map = txq->ift_sds.ifsd_map[i];
1690 	if (map != NULL) {
1691 		bus_dmamap_sync(txq->ift_buf_tag, map, BUS_DMASYNC_POSTWRITE);
1692 		bus_dmamap_unload(txq->ift_buf_tag, map);
1693 		bus_dmamap_destroy(txq->ift_buf_tag, map);
1694 		txq->ift_sds.ifsd_map[i] = NULL;
1695 	}
1696 
1697 	map = NULL;
1698 	if (txq->ift_sds.ifsd_tso_map != NULL)
1699 		map = txq->ift_sds.ifsd_tso_map[i];
1700 	if (map != NULL) {
1701 		bus_dmamap_sync(txq->ift_tso_buf_tag, map,
1702 		    BUS_DMASYNC_POSTWRITE);
1703 		bus_dmamap_unload(txq->ift_tso_buf_tag, map);
1704 		bus_dmamap_destroy(txq->ift_tso_buf_tag, map);
1705 		txq->ift_sds.ifsd_tso_map[i] = NULL;
1706 	}
1707 }
1708 
1709 static void
1710 iflib_txq_destroy(iflib_txq_t txq)
1711 {
1712 	if_ctx_t ctx = txq->ift_ctx;
1713 
1714 	for (int i = 0; i < txq->ift_size; i++)
1715 		iflib_txsd_destroy(ctx, txq, i);
1716 	if (txq->ift_sds.ifsd_map != NULL) {
1717 		free(txq->ift_sds.ifsd_map, M_IFLIB);
1718 		txq->ift_sds.ifsd_map = NULL;
1719 	}
1720 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1721 		free(txq->ift_sds.ifsd_tso_map, M_IFLIB);
1722 		txq->ift_sds.ifsd_tso_map = NULL;
1723 	}
1724 	if (txq->ift_sds.ifsd_m != NULL) {
1725 		free(txq->ift_sds.ifsd_m, M_IFLIB);
1726 		txq->ift_sds.ifsd_m = NULL;
1727 	}
1728 	if (txq->ift_buf_tag != NULL) {
1729 		bus_dma_tag_destroy(txq->ift_buf_tag);
1730 		txq->ift_buf_tag = NULL;
1731 	}
1732 	if (txq->ift_tso_buf_tag != NULL) {
1733 		bus_dma_tag_destroy(txq->ift_tso_buf_tag);
1734 		txq->ift_tso_buf_tag = NULL;
1735 	}
1736 }
1737 
1738 static void
1739 iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i)
1740 {
1741 	struct mbuf **mp;
1742 
1743 	mp = &txq->ift_sds.ifsd_m[i];
1744 	if (*mp == NULL)
1745 		return;
1746 
1747 	if (txq->ift_sds.ifsd_map != NULL) {
1748 		bus_dmamap_sync(txq->ift_buf_tag,
1749 		    txq->ift_sds.ifsd_map[i], BUS_DMASYNC_POSTWRITE);
1750 		bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[i]);
1751 	}
1752 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1753 		bus_dmamap_sync(txq->ift_tso_buf_tag,
1754 		    txq->ift_sds.ifsd_tso_map[i], BUS_DMASYNC_POSTWRITE);
1755 		bus_dmamap_unload(txq->ift_tso_buf_tag,
1756 		    txq->ift_sds.ifsd_tso_map[i]);
1757 	}
1758 	m_free(*mp);
1759 	DBG_COUNTER_INC(tx_frees);
1760 	*mp = NULL;
1761 }
1762 
1763 static int
1764 iflib_txq_setup(iflib_txq_t txq)
1765 {
1766 	if_ctx_t ctx = txq->ift_ctx;
1767 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1768 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1769 	iflib_dma_info_t di;
1770 	int i;
1771 
1772 	/* Set number of descriptors available */
1773 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
1774 	/* XXX make configurable */
1775 	txq->ift_update_freq = IFLIB_DEFAULT_TX_UPDATE_FREQ;
1776 
1777 	/* Reset indices */
1778 	txq->ift_cidx_processed = 0;
1779 	txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0;
1780 	txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset];
1781 
1782 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1783 		bzero((void *)di->idi_vaddr, di->idi_size);
1784 
1785 	IFDI_TXQ_SETUP(ctx, txq->ift_id);
1786 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1787 		bus_dmamap_sync(di->idi_tag, di->idi_map,
1788 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1789 	return (0);
1790 }
1791 
1792 /*********************************************************************
1793  *
1794  *  Allocate DMA resources for RX buffers as well as memory for the RX
1795  *  mbuf map, direct RX cluster pointer map and RX cluster bus address
1796  *  map.  RX DMA map, RX mbuf map, direct RX cluster pointer map and
1797  *  RX cluster map are kept in a iflib_sw_rx_desc_array structure.
1798  *  Since we use use one entry in iflib_sw_rx_desc_array per received
1799  *  packet, the maximum number of entries we'll need is equal to the
1800  *  number of hardware receive descriptors that we've allocated.
1801  *
1802  **********************************************************************/
1803 static int
1804 iflib_rxsd_alloc(iflib_rxq_t rxq)
1805 {
1806 	if_ctx_t ctx = rxq->ifr_ctx;
1807 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1808 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1809 	device_t dev = ctx->ifc_dev;
1810 	iflib_fl_t fl;
1811 	int			err;
1812 
1813 	MPASS(scctx->isc_nrxd[0] > 0);
1814 	MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0);
1815 
1816 	fl = rxq->ifr_fl;
1817 	for (int i = 0; i <  rxq->ifr_nfl; i++, fl++) {
1818 		fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */
1819 		/* Set up DMA tag for RX buffers. */
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_buf_tag);
1832 		if (err) {
1833 			device_printf(dev,
1834 			    "Unable to allocate RX DMA tag: %d\n", err);
1835 			goto fail;
1836 		}
1837 
1838 		/* Allocate memory for the RX mbuf map. */
1839 		if (!(fl->ifl_sds.ifsd_m =
1840 		      (struct mbuf **) malloc(sizeof(struct mbuf *) *
1841 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1842 			device_printf(dev,
1843 			    "Unable to allocate RX mbuf map memory\n");
1844 			err = ENOMEM;
1845 			goto fail;
1846 		}
1847 
1848 		/* Allocate memory for the direct RX cluster pointer map. */
1849 		if (!(fl->ifl_sds.ifsd_cl =
1850 		      (caddr_t *) malloc(sizeof(caddr_t) *
1851 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1852 			device_printf(dev,
1853 			    "Unable to allocate RX cluster map memory\n");
1854 			err = ENOMEM;
1855 			goto fail;
1856 		}
1857 
1858 		/* Allocate memory for the RX cluster bus address map. */
1859 		if (!(fl->ifl_sds.ifsd_ba =
1860 		      (bus_addr_t *) malloc(sizeof(bus_addr_t) *
1861 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1862 			device_printf(dev,
1863 			    "Unable to allocate RX bus address map memory\n");
1864 			err = ENOMEM;
1865 			goto fail;
1866 		}
1867 
1868 		/*
1869 		 * Create the DMA maps for RX buffers.
1870 		 */
1871 		if (!(fl->ifl_sds.ifsd_map =
1872 		      (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1873 			device_printf(dev,
1874 			    "Unable to allocate RX buffer DMA map memory\n");
1875 			err = ENOMEM;
1876 			goto fail;
1877 		}
1878 		for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++) {
1879 			err = bus_dmamap_create(fl->ifl_buf_tag, 0,
1880 			    &fl->ifl_sds.ifsd_map[i]);
1881 			if (err != 0) {
1882 				device_printf(dev, "Unable to create RX buffer DMA map\n");
1883 				goto fail;
1884 			}
1885 		}
1886 	}
1887 	return (0);
1888 
1889 fail:
1890 	iflib_rx_structures_free(ctx);
1891 	return (err);
1892 }
1893 
1894 
1895 /*
1896  * Internal service routines
1897  */
1898 
1899 struct rxq_refill_cb_arg {
1900 	int               error;
1901 	bus_dma_segment_t seg;
1902 	int               nseg;
1903 };
1904 
1905 static void
1906 _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
1907 {
1908 	struct rxq_refill_cb_arg *cb_arg = arg;
1909 
1910 	cb_arg->error = error;
1911 	cb_arg->seg = segs[0];
1912 	cb_arg->nseg = nseg;
1913 }
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 if_rxd_update iru;
1928 	struct rxq_refill_cb_arg cb_arg;
1929 	struct mbuf *m;
1930 	caddr_t cl, *sd_cl;
1931 	struct mbuf **sd_m;
1932 	bus_dmamap_t *sd_map;
1933 	bus_addr_t bus_addr, *sd_ba;
1934 	int err, frag_idx, i, idx, n, pidx;
1935 	qidx_t credits;
1936 
1937 	sd_m = fl->ifl_sds.ifsd_m;
1938 	sd_map = fl->ifl_sds.ifsd_map;
1939 	sd_cl = fl->ifl_sds.ifsd_cl;
1940 	sd_ba = fl->ifl_sds.ifsd_ba;
1941 	pidx = fl->ifl_pidx;
1942 	idx = pidx;
1943 	frag_idx = fl->ifl_fragidx;
1944 	credits = fl->ifl_credits;
1945 
1946 	i = 0;
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,
1970 		    &frag_idx);
1971 		if (frag_idx < 0)
1972 			bit_ffc(fl->ifl_rx_bitmap, fl->ifl_size, &frag_idx);
1973 		MPASS(frag_idx >= 0);
1974 		if ((cl = sd_cl[frag_idx]) == NULL) {
1975 			if ((cl = m_cljget(NULL, M_NOWAIT, fl->ifl_buf_size)) == NULL)
1976 				break;
1977 
1978 			cb_arg.error = 0;
1979 			MPASS(sd_map != NULL);
1980 			err = bus_dmamap_load(fl->ifl_buf_tag, sd_map[frag_idx],
1981 			    cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg,
1982 			    BUS_DMA_NOWAIT);
1983 			if (err != 0 || cb_arg.error) {
1984 				/*
1985 				 * !zone_pack ?
1986 				 */
1987 				if (fl->ifl_zone == zone_pack)
1988 					uma_zfree(fl->ifl_zone, cl);
1989 				break;
1990 			}
1991 
1992 			sd_ba[frag_idx] =  bus_addr = cb_arg.seg.ds_addr;
1993 			sd_cl[frag_idx] = cl;
1994 #if MEMORY_LOGGING
1995 			fl->ifl_cl_enqueued++;
1996 #endif
1997 		} else {
1998 			bus_addr = sd_ba[frag_idx];
1999 		}
2000 		bus_dmamap_sync(fl->ifl_buf_tag, sd_map[frag_idx],
2001 		    BUS_DMASYNC_PREREAD);
2002 
2003 		if (sd_m[frag_idx] == NULL) {
2004 			if ((m = m_gethdr(M_NOWAIT, MT_NOINIT)) == NULL) {
2005 				break;
2006 			}
2007 			sd_m[frag_idx] = m;
2008 		}
2009 		bit_set(fl->ifl_rx_bitmap, frag_idx);
2010 #if MEMORY_LOGGING
2011 		fl->ifl_m_enqueued++;
2012 #endif
2013 
2014 		DBG_COUNTER_INC(rx_allocs);
2015 		fl->ifl_rxd_idxs[i] = frag_idx;
2016 		fl->ifl_bus_addrs[i] = bus_addr;
2017 		fl->ifl_vm_addrs[i] = cl;
2018 		credits++;
2019 		i++;
2020 		MPASS(credits <= fl->ifl_size);
2021 		if (++idx == fl->ifl_size) {
2022 			fl->ifl_gen = 1;
2023 			idx = 0;
2024 		}
2025 		if (n == 0 || i == IFLIB_MAX_RX_REFRESH) {
2026 			iru.iru_pidx = pidx;
2027 			iru.iru_count = i;
2028 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2029 			i = 0;
2030 			pidx = idx;
2031 			fl->ifl_pidx = idx;
2032 			fl->ifl_credits = credits;
2033 		}
2034 	}
2035 
2036 	if (i) {
2037 		iru.iru_pidx = pidx;
2038 		iru.iru_count = i;
2039 		ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2040 		fl->ifl_pidx = idx;
2041 		fl->ifl_credits = credits;
2042 	}
2043 	DBG_COUNTER_INC(rxd_flush);
2044 	if (fl->ifl_pidx == 0)
2045 		pidx = fl->ifl_size - 1;
2046 	else
2047 		pidx = fl->ifl_pidx - 1;
2048 
2049 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2050 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2051 	ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id, fl->ifl_id, pidx);
2052 	fl->ifl_fragidx = frag_idx;
2053 }
2054 
2055 static __inline void
2056 __iflib_fl_refill_lt(if_ctx_t ctx, iflib_fl_t fl, int max)
2057 {
2058 	/* we avoid allowing pidx to catch up with cidx as it confuses ixl */
2059 	int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1;
2060 #ifdef INVARIANTS
2061 	int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1;
2062 #endif
2063 
2064 	MPASS(fl->ifl_credits <= fl->ifl_size);
2065 	MPASS(reclaimable == delta);
2066 
2067 	if (reclaimable > 0)
2068 		_iflib_fl_refill(ctx, fl, min(max, reclaimable));
2069 }
2070 
2071 uint8_t
2072 iflib_in_detach(if_ctx_t ctx)
2073 {
2074 	bool in_detach;
2075 	STATE_LOCK(ctx);
2076 	in_detach = !!(ctx->ifc_flags & IFC_IN_DETACH);
2077 	STATE_UNLOCK(ctx);
2078 	return (in_detach);
2079 }
2080 
2081 static void
2082 iflib_fl_bufs_free(iflib_fl_t fl)
2083 {
2084 	iflib_dma_info_t idi = fl->ifl_ifdi;
2085 	bus_dmamap_t sd_map;
2086 	uint32_t i;
2087 
2088 	for (i = 0; i < fl->ifl_size; i++) {
2089 		struct mbuf **sd_m = &fl->ifl_sds.ifsd_m[i];
2090 		caddr_t *sd_cl = &fl->ifl_sds.ifsd_cl[i];
2091 
2092 		if (*sd_cl != NULL) {
2093 			sd_map = fl->ifl_sds.ifsd_map[i];
2094 			bus_dmamap_sync(fl->ifl_buf_tag, sd_map,
2095 			    BUS_DMASYNC_POSTREAD);
2096 			bus_dmamap_unload(fl->ifl_buf_tag, sd_map);
2097 			if (*sd_cl != NULL)
2098 				uma_zfree(fl->ifl_zone, *sd_cl);
2099 			// XXX: Should this get moved out?
2100 			if (iflib_in_detach(fl->ifl_rxq->ifr_ctx))
2101 				bus_dmamap_destroy(fl->ifl_buf_tag, sd_map);
2102 			if (*sd_m != NULL) {
2103 				m_init(*sd_m, M_NOWAIT, MT_DATA, 0);
2104 				uma_zfree(zone_mbuf, *sd_m);
2105 			}
2106 		} else {
2107 			MPASS(*sd_cl == NULL);
2108 			MPASS(*sd_m == NULL);
2109 		}
2110 #if MEMORY_LOGGING
2111 		fl->ifl_m_dequeued++;
2112 		fl->ifl_cl_dequeued++;
2113 #endif
2114 		*sd_cl = NULL;
2115 		*sd_m = NULL;
2116 	}
2117 #ifdef INVARIANTS
2118 	for (i = 0; i < fl->ifl_size; i++) {
2119 		MPASS(fl->ifl_sds.ifsd_cl[i] == NULL);
2120 		MPASS(fl->ifl_sds.ifsd_m[i] == NULL);
2121 	}
2122 #endif
2123 	/*
2124 	 * Reset free list values
2125 	 */
2126 	fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = fl->ifl_fragidx = 0;
2127 	bzero(idi->idi_vaddr, idi->idi_size);
2128 }
2129 
2130 /*********************************************************************
2131  *
2132  *  Initialize a receive ring and its buffers.
2133  *
2134  **********************************************************************/
2135 static int
2136 iflib_fl_setup(iflib_fl_t fl)
2137 {
2138 	iflib_rxq_t rxq = fl->ifl_rxq;
2139 	if_ctx_t ctx = rxq->ifr_ctx;
2140 
2141 	bit_nclear(fl->ifl_rx_bitmap, 0, fl->ifl_size - 1);
2142 	/*
2143 	** Free current RX buffer structs and their mbufs
2144 	*/
2145 	iflib_fl_bufs_free(fl);
2146 	/* Now replenish the mbufs */
2147 	MPASS(fl->ifl_credits == 0);
2148 	fl->ifl_buf_size = ctx->ifc_rx_mbuf_sz;
2149 	if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size)
2150 		ctx->ifc_max_fl_buf_size = fl->ifl_buf_size;
2151 	fl->ifl_cltype = m_gettype(fl->ifl_buf_size);
2152 	fl->ifl_zone = m_getzone(fl->ifl_buf_size);
2153 
2154 
2155 	/* avoid pre-allocating zillions of clusters to an idle card
2156 	 * potentially speeding up attach
2157 	 */
2158 	_iflib_fl_refill(ctx, fl, min(128, fl->ifl_size));
2159 	MPASS(min(128, fl->ifl_size) == fl->ifl_credits);
2160 	if (min(128, fl->ifl_size) != fl->ifl_credits)
2161 		return (ENOBUFS);
2162 	/*
2163 	 * handle failure
2164 	 */
2165 	MPASS(rxq != NULL);
2166 	MPASS(fl->ifl_ifdi != NULL);
2167 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2168 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2169 	return (0);
2170 }
2171 
2172 /*********************************************************************
2173  *
2174  *  Free receive ring data structures
2175  *
2176  **********************************************************************/
2177 static void
2178 iflib_rx_sds_free(iflib_rxq_t rxq)
2179 {
2180 	iflib_fl_t fl;
2181 	int i, j;
2182 
2183 	if (rxq->ifr_fl != NULL) {
2184 		for (i = 0; i < rxq->ifr_nfl; i++) {
2185 			fl = &rxq->ifr_fl[i];
2186 			if (fl->ifl_buf_tag != NULL) {
2187 				if (fl->ifl_sds.ifsd_map != NULL) {
2188 					for (j = 0; j < fl->ifl_size; j++) {
2189 						if (fl->ifl_sds.ifsd_map[j] ==
2190 						    NULL)
2191 							continue;
2192 						bus_dmamap_sync(
2193 						    fl->ifl_buf_tag,
2194 						    fl->ifl_sds.ifsd_map[j],
2195 						    BUS_DMASYNC_POSTREAD);
2196 						bus_dmamap_unload(
2197 						    fl->ifl_buf_tag,
2198 						    fl->ifl_sds.ifsd_map[j]);
2199 					}
2200 				}
2201 				bus_dma_tag_destroy(fl->ifl_buf_tag);
2202 				fl->ifl_buf_tag = NULL;
2203 			}
2204 			free(fl->ifl_sds.ifsd_m, M_IFLIB);
2205 			free(fl->ifl_sds.ifsd_cl, M_IFLIB);
2206 			free(fl->ifl_sds.ifsd_ba, M_IFLIB);
2207 			free(fl->ifl_sds.ifsd_map, M_IFLIB);
2208 			fl->ifl_sds.ifsd_m = NULL;
2209 			fl->ifl_sds.ifsd_cl = NULL;
2210 			fl->ifl_sds.ifsd_ba = NULL;
2211 			fl->ifl_sds.ifsd_map = NULL;
2212 		}
2213 		free(rxq->ifr_fl, M_IFLIB);
2214 		rxq->ifr_fl = NULL;
2215 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
2216 	}
2217 }
2218 
2219 /*
2220  * MI independent logic
2221  *
2222  */
2223 static void
2224 iflib_timer(void *arg)
2225 {
2226 	iflib_txq_t txq = arg;
2227 	if_ctx_t ctx = txq->ift_ctx;
2228 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2229 	uint64_t this_tick = ticks;
2230 	uint32_t reset_on = hz / 2;
2231 
2232 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
2233 		return;
2234 	/*
2235 	** Check on the state of the TX queue(s), this
2236 	** can be done without the lock because its RO
2237 	** and the HUNG state will be static if set.
2238 	*/
2239 	if (this_tick - txq->ift_last_timer_tick >= hz / 2) {
2240 		txq->ift_last_timer_tick = this_tick;
2241 		IFDI_TIMER(ctx, txq->ift_id);
2242 		if ((txq->ift_qstatus == IFLIB_QUEUE_HUNG) &&
2243 		    ((txq->ift_cleaned_prev == txq->ift_cleaned) ||
2244 		     (sctx->isc_pause_frames == 0)))
2245 			goto hung;
2246 
2247 		if (ifmp_ring_is_stalled(txq->ift_br))
2248 			txq->ift_qstatus = IFLIB_QUEUE_HUNG;
2249 		txq->ift_cleaned_prev = txq->ift_cleaned;
2250 	}
2251 #ifdef DEV_NETMAP
2252 	if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP)
2253 		iflib_netmap_timer_adjust(ctx, txq, &reset_on);
2254 #endif
2255 	/* handle any laggards */
2256 	if (txq->ift_db_pending)
2257 		GROUPTASK_ENQUEUE(&txq->ift_task);
2258 
2259 	sctx->isc_pause_frames = 0;
2260 	if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)
2261 		callout_reset_on(&txq->ift_timer, reset_on, iflib_timer, txq, txq->ift_timer.c_cpu);
2262 	return;
2263  hung:
2264 	device_printf(ctx->ifc_dev,  "TX(%d) desc avail = %d, pidx = %d\n",
2265 				  txq->ift_id, TXQ_AVAIL(txq), txq->ift_pidx);
2266 	STATE_LOCK(ctx);
2267 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2268 	ctx->ifc_flags |= (IFC_DO_WATCHDOG|IFC_DO_RESET);
2269 	iflib_admin_intr_deferred(ctx);
2270 	STATE_UNLOCK(ctx);
2271 }
2272 
2273 static void
2274 iflib_calc_rx_mbuf_sz(if_ctx_t ctx)
2275 {
2276 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2277 
2278 	/*
2279 	 * XXX don't set the max_frame_size to larger
2280 	 * than the hardware can handle
2281 	 */
2282 	if (sctx->isc_max_frame_size <= MCLBYTES)
2283 		ctx->ifc_rx_mbuf_sz = MCLBYTES;
2284 	else
2285 		ctx->ifc_rx_mbuf_sz = MJUMPAGESIZE;
2286 }
2287 
2288 uint32_t
2289 iflib_get_rx_mbuf_sz(if_ctx_t ctx)
2290 {
2291 	return (ctx->ifc_rx_mbuf_sz);
2292 }
2293 
2294 static void
2295 iflib_init_locked(if_ctx_t ctx)
2296 {
2297 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2298 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2299 	if_t ifp = ctx->ifc_ifp;
2300 	iflib_fl_t fl;
2301 	iflib_txq_t txq;
2302 	iflib_rxq_t rxq;
2303 	int i, j, tx_ip_csum_flags, tx_ip6_csum_flags;
2304 
2305 
2306 	if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2307 	IFDI_INTR_DISABLE(ctx);
2308 
2309 	tx_ip_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP);
2310 	tx_ip6_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | CSUM_IP6_SCTP);
2311 	/* Set hardware offload abilities */
2312 	if_clearhwassist(ifp);
2313 	if (if_getcapenable(ifp) & IFCAP_TXCSUM)
2314 		if_sethwassistbits(ifp, tx_ip_csum_flags, 0);
2315 	if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6)
2316 		if_sethwassistbits(ifp,  tx_ip6_csum_flags, 0);
2317 	if (if_getcapenable(ifp) & IFCAP_TSO4)
2318 		if_sethwassistbits(ifp, CSUM_IP_TSO, 0);
2319 	if (if_getcapenable(ifp) & IFCAP_TSO6)
2320 		if_sethwassistbits(ifp, CSUM_IP6_TSO, 0);
2321 
2322 	for (i = 0, txq = ctx->ifc_txqs; i < sctx->isc_ntxqsets; i++, txq++) {
2323 		CALLOUT_LOCK(txq);
2324 		callout_stop(&txq->ift_timer);
2325 		CALLOUT_UNLOCK(txq);
2326 		iflib_netmap_txq_init(ctx, txq);
2327 	}
2328 
2329 	/*
2330 	 * Calculate a suitable Rx mbuf size prior to calling IFDI_INIT, so
2331 	 * that drivers can use the value when setting up the hardware receive
2332 	 * buffers.
2333 	 */
2334 	iflib_calc_rx_mbuf_sz(ctx);
2335 
2336 #ifdef INVARIANTS
2337 	i = if_getdrvflags(ifp);
2338 #endif
2339 	IFDI_INIT(ctx);
2340 	MPASS(if_getdrvflags(ifp) == i);
2341 	for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
2342 		/* XXX this should really be done on a per-queue basis */
2343 		if (if_getcapenable(ifp) & IFCAP_NETMAP) {
2344 			MPASS(rxq->ifr_id == i);
2345 			iflib_netmap_rxq_init(ctx, rxq);
2346 			continue;
2347 		}
2348 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
2349 			if (iflib_fl_setup(fl)) {
2350 				device_printf(ctx->ifc_dev, "freelist setup failed - check cluster settings\n");
2351 				goto done;
2352 			}
2353 		}
2354 	}
2355 done:
2356 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
2357 	IFDI_INTR_ENABLE(ctx);
2358 	txq = ctx->ifc_txqs;
2359 	for (i = 0; i < sctx->isc_ntxqsets; i++, txq++)
2360 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq,
2361 			txq->ift_timer.c_cpu);
2362 }
2363 
2364 static int
2365 iflib_media_change(if_t ifp)
2366 {
2367 	if_ctx_t ctx = if_getsoftc(ifp);
2368 	int err;
2369 
2370 	CTX_LOCK(ctx);
2371 	if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0)
2372 		iflib_init_locked(ctx);
2373 	CTX_UNLOCK(ctx);
2374 	return (err);
2375 }
2376 
2377 static void
2378 iflib_media_status(if_t ifp, struct ifmediareq *ifmr)
2379 {
2380 	if_ctx_t ctx = if_getsoftc(ifp);
2381 
2382 	CTX_LOCK(ctx);
2383 	IFDI_UPDATE_ADMIN_STATUS(ctx);
2384 	IFDI_MEDIA_STATUS(ctx, ifmr);
2385 	CTX_UNLOCK(ctx);
2386 }
2387 
2388 void
2389 iflib_stop(if_ctx_t ctx)
2390 {
2391 	iflib_txq_t txq = ctx->ifc_txqs;
2392 	iflib_rxq_t rxq = ctx->ifc_rxqs;
2393 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2394 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2395 	iflib_dma_info_t di;
2396 	iflib_fl_t fl;
2397 	int i, j;
2398 
2399 	/* Tell the stack that the interface is no longer active */
2400 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2401 
2402 	IFDI_INTR_DISABLE(ctx);
2403 	DELAY(1000);
2404 	IFDI_STOP(ctx);
2405 	DELAY(1000);
2406 
2407 	iflib_debug_reset();
2408 	/* Wait for current tx queue users to exit to disarm watchdog timer. */
2409 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) {
2410 		/* make sure all transmitters have completed before proceeding XXX */
2411 
2412 		CALLOUT_LOCK(txq);
2413 		callout_stop(&txq->ift_timer);
2414 		CALLOUT_UNLOCK(txq);
2415 
2416 		/* clean any enqueued buffers */
2417 		iflib_ifmp_purge(txq);
2418 		/* Free any existing tx buffers. */
2419 		for (j = 0; j < txq->ift_size; j++) {
2420 			iflib_txsd_free(ctx, txq, j);
2421 		}
2422 		txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0;
2423 		txq->ift_in_use = txq->ift_gen = txq->ift_cidx = txq->ift_pidx = txq->ift_no_desc_avail = 0;
2424 		txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0;
2425 		txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0;
2426 		txq->ift_pullups = 0;
2427 		ifmp_ring_reset_stats(txq->ift_br);
2428 		for (j = 0, di = txq->ift_ifdi; j < sctx->isc_ntxqs; j++, di++)
2429 			bzero((void *)di->idi_vaddr, di->idi_size);
2430 	}
2431 	for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) {
2432 		/* make sure all transmitters have completed before proceeding XXX */
2433 
2434 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
2435 		for (j = 0, di = rxq->ifr_ifdi; j < sctx->isc_nrxqs; j++, di++)
2436 			bzero((void *)di->idi_vaddr, di->idi_size);
2437 		/* also resets the free lists pidx/cidx */
2438 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
2439 			iflib_fl_bufs_free(fl);
2440 	}
2441 }
2442 
2443 static inline caddr_t
2444 calc_next_rxd(iflib_fl_t fl, int cidx)
2445 {
2446 	qidx_t size;
2447 	int nrxd;
2448 	caddr_t start, end, cur, next;
2449 
2450 	nrxd = fl->ifl_size;
2451 	size = fl->ifl_rxd_size;
2452 	start = fl->ifl_ifdi->idi_vaddr;
2453 
2454 	if (__predict_false(size == 0))
2455 		return (start);
2456 	cur = start + size*cidx;
2457 	end = start + size*nrxd;
2458 	next = CACHE_PTR_NEXT(cur);
2459 	return (next < end ? next : start);
2460 }
2461 
2462 static inline void
2463 prefetch_pkts(iflib_fl_t fl, int cidx)
2464 {
2465 	int nextptr;
2466 	int nrxd = fl->ifl_size;
2467 	caddr_t next_rxd;
2468 
2469 
2470 	nextptr = (cidx + CACHE_PTR_INCREMENT) & (nrxd-1);
2471 	prefetch(&fl->ifl_sds.ifsd_m[nextptr]);
2472 	prefetch(&fl->ifl_sds.ifsd_cl[nextptr]);
2473 	next_rxd = calc_next_rxd(fl, cidx);
2474 	prefetch(next_rxd);
2475 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 1) & (nrxd-1)]);
2476 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 2) & (nrxd-1)]);
2477 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 3) & (nrxd-1)]);
2478 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 4) & (nrxd-1)]);
2479 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 1) & (nrxd-1)]);
2480 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 2) & (nrxd-1)]);
2481 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 3) & (nrxd-1)]);
2482 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 4) & (nrxd-1)]);
2483 }
2484 
2485 static struct mbuf *
2486 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, bool unload, if_rxsd_t sd,
2487     int *pf_rv, if_rxd_info_t ri)
2488 {
2489 	bus_dmamap_t map;
2490 	iflib_fl_t fl;
2491 	caddr_t payload;
2492 	struct mbuf *m;
2493 	int flid, cidx, len, next;
2494 
2495 	map = NULL;
2496 	flid = irf->irf_flid;
2497 	cidx = irf->irf_idx;
2498 	fl = &rxq->ifr_fl[flid];
2499 	sd->ifsd_fl = fl;
2500 	sd->ifsd_cidx = cidx;
2501 	m = fl->ifl_sds.ifsd_m[cidx];
2502 	sd->ifsd_cl = &fl->ifl_sds.ifsd_cl[cidx];
2503 	fl->ifl_credits--;
2504 #if MEMORY_LOGGING
2505 	fl->ifl_m_dequeued++;
2506 #endif
2507 	if (rxq->ifr_ctx->ifc_flags & IFC_PREFETCH)
2508 		prefetch_pkts(fl, cidx);
2509 	next = (cidx + CACHE_PTR_INCREMENT) & (fl->ifl_size-1);
2510 	prefetch(&fl->ifl_sds.ifsd_map[next]);
2511 	map = fl->ifl_sds.ifsd_map[cidx];
2512 	next = (cidx + CACHE_LINE_SIZE) & (fl->ifl_size-1);
2513 
2514 	/* not valid assert if bxe really does SGE from non-contiguous elements */
2515 	MPASS(fl->ifl_cidx == cidx);
2516 	bus_dmamap_sync(fl->ifl_buf_tag, map, BUS_DMASYNC_POSTREAD);
2517 
2518 	if (rxq->pfil != NULL && PFIL_HOOKED_IN(rxq->pfil) && pf_rv != NULL) {
2519 		payload  = *sd->ifsd_cl;
2520 		payload +=  ri->iri_pad;
2521 		len = ri->iri_len - ri->iri_pad;
2522 		*pf_rv = pfil_run_hooks(rxq->pfil, payload, ri->iri_ifp,
2523 		    len | PFIL_MEMPTR | PFIL_IN, NULL);
2524 		switch (*pf_rv) {
2525 		case PFIL_DROPPED:
2526 		case PFIL_CONSUMED:
2527 			/*
2528 			 * The filter ate it.  Everything is recycled.
2529 			 */
2530 			m = NULL;
2531 			unload = 0;
2532 			break;
2533 		case PFIL_REALLOCED:
2534 			/*
2535 			 * The filter copied it.  Everything is recycled.
2536 			 */
2537 			m = pfil_mem2mbuf(payload);
2538 			unload = 0;
2539 			break;
2540 		case PFIL_PASS:
2541 			/*
2542 			 * Filter said it was OK, so receive like
2543 			 * normal
2544 			 */
2545 			fl->ifl_sds.ifsd_m[cidx] = NULL;
2546 			break;
2547 		default:
2548 			MPASS(0);
2549 		}
2550 	} else {
2551 		fl->ifl_sds.ifsd_m[cidx] = NULL;
2552 		*pf_rv = PFIL_PASS;
2553 	}
2554 
2555 	if (unload)
2556 		bus_dmamap_unload(fl->ifl_buf_tag, map);
2557 	fl->ifl_cidx = (fl->ifl_cidx + 1) & (fl->ifl_size-1);
2558 	if (__predict_false(fl->ifl_cidx == 0))
2559 		fl->ifl_gen = 0;
2560 	bit_clear(fl->ifl_rx_bitmap, cidx);
2561 	return (m);
2562 }
2563 
2564 static struct mbuf *
2565 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri, if_rxsd_t sd, int *pf_rv)
2566 {
2567 	struct mbuf *m, *mh, *mt;
2568 	caddr_t cl;
2569 	int  *pf_rv_ptr, flags, i, padlen;
2570 	bool consumed;
2571 
2572 	i = 0;
2573 	mh = NULL;
2574 	consumed = false;
2575 	*pf_rv = PFIL_PASS;
2576 	pf_rv_ptr = pf_rv;
2577 	do {
2578 		m = rxd_frag_to_sd(rxq, &ri->iri_frags[i], !consumed, sd,
2579 		    pf_rv_ptr, ri);
2580 
2581 		MPASS(*sd->ifsd_cl != NULL);
2582 
2583 		/*
2584 		 * Exclude zero-length frags & frags from
2585 		 * packets the filter has consumed or dropped
2586 		 */
2587 		if (ri->iri_frags[i].irf_len == 0 || consumed ||
2588 		    *pf_rv == PFIL_CONSUMED || *pf_rv == PFIL_DROPPED) {
2589 			if (mh == NULL) {
2590 				/* everything saved here */
2591 				consumed = true;
2592 				pf_rv_ptr = NULL;
2593 				continue;
2594 			}
2595 			/* XXX we can save the cluster here, but not the mbuf */
2596 			m_init(m, M_NOWAIT, MT_DATA, 0);
2597 			m_free(m);
2598 			continue;
2599 		}
2600 		if (mh == NULL) {
2601 			flags = M_PKTHDR|M_EXT;
2602 			mh = mt = m;
2603 			padlen = ri->iri_pad;
2604 		} else {
2605 			flags = M_EXT;
2606 			mt->m_next = m;
2607 			mt = m;
2608 			/* assuming padding is only on the first fragment */
2609 			padlen = 0;
2610 		}
2611 		cl = *sd->ifsd_cl;
2612 		*sd->ifsd_cl = NULL;
2613 
2614 		/* Can these two be made one ? */
2615 		m_init(m, M_NOWAIT, MT_DATA, flags);
2616 		m_cljset(m, cl, sd->ifsd_fl->ifl_cltype);
2617 		/*
2618 		 * These must follow m_init and m_cljset
2619 		 */
2620 		m->m_data += padlen;
2621 		ri->iri_len -= padlen;
2622 		m->m_len = ri->iri_frags[i].irf_len;
2623 	} while (++i < ri->iri_nfrags);
2624 
2625 	return (mh);
2626 }
2627 
2628 /*
2629  * Process one software descriptor
2630  */
2631 static struct mbuf *
2632 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri)
2633 {
2634 	struct if_rxsd sd;
2635 	struct mbuf *m;
2636 	int pf_rv;
2637 
2638 	/* should I merge this back in now that the two paths are basically duplicated? */
2639 	if (ri->iri_nfrags == 1 &&
2640 	    ri->iri_frags[0].irf_len <= MIN(IFLIB_RX_COPY_THRESH, MHLEN)) {
2641 		m = rxd_frag_to_sd(rxq, &ri->iri_frags[0], false, &sd,
2642 		    &pf_rv, ri);
2643 		if (pf_rv != PFIL_PASS && pf_rv != PFIL_REALLOCED)
2644 			return (m);
2645 		if (pf_rv == PFIL_PASS) {
2646 			m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR);
2647 #ifndef __NO_STRICT_ALIGNMENT
2648 			if (!IP_ALIGNED(m))
2649 				m->m_data += 2;
2650 #endif
2651 			memcpy(m->m_data, *sd.ifsd_cl, ri->iri_len);
2652 			m->m_len = ri->iri_frags[0].irf_len;
2653 		}
2654 	} else {
2655 		m = assemble_segments(rxq, ri, &sd, &pf_rv);
2656 		if (pf_rv != PFIL_PASS && pf_rv != PFIL_REALLOCED)
2657 			return (m);
2658 	}
2659 	m->m_pkthdr.len = ri->iri_len;
2660 	m->m_pkthdr.rcvif = ri->iri_ifp;
2661 	m->m_flags |= ri->iri_flags;
2662 	m->m_pkthdr.ether_vtag = ri->iri_vtag;
2663 	m->m_pkthdr.flowid = ri->iri_flowid;
2664 	M_HASHTYPE_SET(m, ri->iri_rsstype);
2665 	m->m_pkthdr.csum_flags = ri->iri_csum_flags;
2666 	m->m_pkthdr.csum_data = ri->iri_csum_data;
2667 	return (m);
2668 }
2669 
2670 #if defined(INET6) || defined(INET)
2671 static void
2672 iflib_get_ip_forwarding(struct lro_ctrl *lc, bool *v4, bool *v6)
2673 {
2674 	CURVNET_SET(lc->ifp->if_vnet);
2675 #if defined(INET6)
2676 	*v6 = VNET(ip6_forwarding);
2677 #endif
2678 #if defined(INET)
2679 	*v4 = VNET(ipforwarding);
2680 #endif
2681 	CURVNET_RESTORE();
2682 }
2683 
2684 /*
2685  * Returns true if it's possible this packet could be LROed.
2686  * if it returns false, it is guaranteed that tcp_lro_rx()
2687  * would not return zero.
2688  */
2689 static bool
2690 iflib_check_lro_possible(struct mbuf *m, bool v4_forwarding, bool v6_forwarding)
2691 {
2692 	struct ether_header *eh;
2693 	uint16_t eh_type;
2694 
2695 	eh = mtod(m, struct ether_header *);
2696 	eh_type = ntohs(eh->ether_type);
2697 	switch (eh_type) {
2698 #if defined(INET6)
2699 		case ETHERTYPE_IPV6:
2700 			return !v6_forwarding;
2701 #endif
2702 #if defined (INET)
2703 		case ETHERTYPE_IP:
2704 			return !v4_forwarding;
2705 #endif
2706 	}
2707 
2708 	return false;
2709 }
2710 #else
2711 static void
2712 iflib_get_ip_forwarding(struct lro_ctrl *lc __unused, bool *v4 __unused, bool *v6 __unused)
2713 {
2714 }
2715 #endif
2716 
2717 static bool
2718 iflib_rxeof(iflib_rxq_t rxq, qidx_t budget)
2719 {
2720 	if_ctx_t ctx = rxq->ifr_ctx;
2721 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2722 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2723 	int avail, i;
2724 	qidx_t *cidxp;
2725 	struct if_rxd_info ri;
2726 	int err, budget_left, rx_bytes, rx_pkts;
2727 	iflib_fl_t fl;
2728 	struct ifnet *ifp;
2729 	int lro_enabled;
2730 	bool v4_forwarding, v6_forwarding, lro_possible;
2731 
2732 	/*
2733 	 * XXX early demux data packets so that if_input processing only handles
2734 	 * acks in interrupt context
2735 	 */
2736 	struct mbuf *m, *mh, *mt, *mf;
2737 
2738 	lro_possible = v4_forwarding = v6_forwarding = false;
2739 	ifp = ctx->ifc_ifp;
2740 	mh = mt = NULL;
2741 	MPASS(budget > 0);
2742 	rx_pkts	= rx_bytes = 0;
2743 	if (sctx->isc_flags & IFLIB_HAS_RXCQ)
2744 		cidxp = &rxq->ifr_cq_cidx;
2745 	else
2746 		cidxp = &rxq->ifr_fl[0].ifl_cidx;
2747 	if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) {
2748 		for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2749 			__iflib_fl_refill_lt(ctx, fl, budget + 8);
2750 		DBG_COUNTER_INC(rx_unavail);
2751 		return (false);
2752 	}
2753 
2754 	/* pfil needs the vnet to be set */
2755 	CURVNET_SET_QUIET(ifp->if_vnet);
2756 	for (budget_left = budget; budget_left > 0 && avail > 0;) {
2757 		if (__predict_false(!CTX_ACTIVE(ctx))) {
2758 			DBG_COUNTER_INC(rx_ctx_inactive);
2759 			break;
2760 		}
2761 		/*
2762 		 * Reset client set fields to their default values
2763 		 */
2764 		rxd_info_zero(&ri);
2765 		ri.iri_qsidx = rxq->ifr_id;
2766 		ri.iri_cidx = *cidxp;
2767 		ri.iri_ifp = ifp;
2768 		ri.iri_frags = rxq->ifr_frags;
2769 		err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
2770 
2771 		if (err)
2772 			goto err;
2773 		rx_pkts += 1;
2774 		rx_bytes += ri.iri_len;
2775 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
2776 			*cidxp = ri.iri_cidx;
2777 			/* Update our consumer index */
2778 			/* XXX NB: shurd - check if this is still safe */
2779 			while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0]) {
2780 				rxq->ifr_cq_cidx -= scctx->isc_nrxd[0];
2781 				rxq->ifr_cq_gen = 0;
2782 			}
2783 			/* was this only a completion queue message? */
2784 			if (__predict_false(ri.iri_nfrags == 0))
2785 				continue;
2786 		}
2787 		MPASS(ri.iri_nfrags != 0);
2788 		MPASS(ri.iri_len != 0);
2789 
2790 		/* will advance the cidx on the corresponding free lists */
2791 		m = iflib_rxd_pkt_get(rxq, &ri);
2792 		avail--;
2793 		budget_left--;
2794 		if (avail == 0 && budget_left)
2795 			avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left);
2796 
2797 		if (__predict_false(m == NULL))
2798 			continue;
2799 
2800 		/* imm_pkt: -- cxgb */
2801 		if (mh == NULL)
2802 			mh = mt = m;
2803 		else {
2804 			mt->m_nextpkt = m;
2805 			mt = m;
2806 		}
2807 	}
2808 	CURVNET_RESTORE();
2809 	/* make sure that we can refill faster than drain */
2810 	for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2811 		__iflib_fl_refill_lt(ctx, fl, budget + 8);
2812 
2813 	lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO);
2814 	if (lro_enabled)
2815 		iflib_get_ip_forwarding(&rxq->ifr_lc, &v4_forwarding, &v6_forwarding);
2816 	mt = mf = NULL;
2817 	while (mh != NULL) {
2818 		m = mh;
2819 		mh = mh->m_nextpkt;
2820 		m->m_nextpkt = NULL;
2821 #ifndef __NO_STRICT_ALIGNMENT
2822 		if (!IP_ALIGNED(m) && (m = iflib_fixup_rx(m)) == NULL)
2823 			continue;
2824 #endif
2825 		rx_bytes += m->m_pkthdr.len;
2826 		rx_pkts++;
2827 #if defined(INET6) || defined(INET)
2828 		if (lro_enabled) {
2829 			if (!lro_possible) {
2830 				lro_possible = iflib_check_lro_possible(m, v4_forwarding, v6_forwarding);
2831 				if (lro_possible && mf != NULL) {
2832 					ifp->if_input(ifp, mf);
2833 					DBG_COUNTER_INC(rx_if_input);
2834 					mt = mf = NULL;
2835 				}
2836 			}
2837 			if ((m->m_pkthdr.csum_flags & (CSUM_L4_CALC|CSUM_L4_VALID)) ==
2838 			    (CSUM_L4_CALC|CSUM_L4_VALID)) {
2839 				if (lro_possible && tcp_lro_rx(&rxq->ifr_lc, m, 0) == 0)
2840 					continue;
2841 			}
2842 		}
2843 #endif
2844 		if (lro_possible) {
2845 			ifp->if_input(ifp, m);
2846 			DBG_COUNTER_INC(rx_if_input);
2847 			continue;
2848 		}
2849 
2850 		if (mf == NULL)
2851 			mf = m;
2852 		if (mt != NULL)
2853 			mt->m_nextpkt = m;
2854 		mt = m;
2855 	}
2856 	if (mf != NULL) {
2857 		ifp->if_input(ifp, mf);
2858 		DBG_COUNTER_INC(rx_if_input);
2859 	}
2860 
2861 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
2862 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
2863 
2864 	/*
2865 	 * Flush any outstanding LRO work
2866 	 */
2867 #if defined(INET6) || defined(INET)
2868 	tcp_lro_flush_all(&rxq->ifr_lc);
2869 #endif
2870 	if (avail)
2871 		return true;
2872 	return (iflib_rxd_avail(ctx, rxq, *cidxp, 1));
2873 err:
2874 	STATE_LOCK(ctx);
2875 	ctx->ifc_flags |= IFC_DO_RESET;
2876 	iflib_admin_intr_deferred(ctx);
2877 	STATE_UNLOCK(ctx);
2878 	return (false);
2879 }
2880 
2881 #define TXD_NOTIFY_COUNT(txq) (((txq)->ift_size / (txq)->ift_update_freq)-1)
2882 static inline qidx_t
2883 txq_max_db_deferred(iflib_txq_t txq, qidx_t in_use)
2884 {
2885 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2886 	qidx_t minthresh = txq->ift_size / 8;
2887 	if (in_use > 4*minthresh)
2888 		return (notify_count);
2889 	if (in_use > 2*minthresh)
2890 		return (notify_count >> 1);
2891 	if (in_use > minthresh)
2892 		return (notify_count >> 3);
2893 	return (0);
2894 }
2895 
2896 static inline qidx_t
2897 txq_max_rs_deferred(iflib_txq_t txq)
2898 {
2899 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2900 	qidx_t minthresh = txq->ift_size / 8;
2901 	if (txq->ift_in_use > 4*minthresh)
2902 		return (notify_count);
2903 	if (txq->ift_in_use > 2*minthresh)
2904 		return (notify_count >> 1);
2905 	if (txq->ift_in_use > minthresh)
2906 		return (notify_count >> 2);
2907 	return (2);
2908 }
2909 
2910 #define M_CSUM_FLAGS(m) ((m)->m_pkthdr.csum_flags)
2911 #define M_HAS_VLANTAG(m) (m->m_flags & M_VLANTAG)
2912 
2913 #define TXQ_MAX_DB_DEFERRED(txq, in_use) txq_max_db_deferred((txq), (in_use))
2914 #define TXQ_MAX_RS_DEFERRED(txq) txq_max_rs_deferred(txq)
2915 #define TXQ_MAX_DB_CONSUMED(size) (size >> 4)
2916 
2917 /* forward compatibility for cxgb */
2918 #define FIRST_QSET(ctx) 0
2919 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets)
2920 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets)
2921 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx))
2922 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments))
2923 
2924 /* XXX we should be setting this to something other than zero */
2925 #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh)
2926 #define	MAX_TX_DESC(ctx) max((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max, \
2927     (ctx)->ifc_softc_ctx.isc_tx_nsegments)
2928 
2929 static inline bool
2930 iflib_txd_db_check(if_ctx_t ctx, iflib_txq_t txq, int ring, qidx_t in_use)
2931 {
2932 	qidx_t dbval, max;
2933 	bool rang;
2934 
2935 	rang = false;
2936 	max = TXQ_MAX_DB_DEFERRED(txq, in_use);
2937 	if (ring || txq->ift_db_pending >= max) {
2938 		dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx;
2939 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
2940 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2941 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval);
2942 		txq->ift_db_pending = txq->ift_npending = 0;
2943 		rang = true;
2944 	}
2945 	return (rang);
2946 }
2947 
2948 #ifdef PKT_DEBUG
2949 static void
2950 print_pkt(if_pkt_info_t pi)
2951 {
2952 	printf("pi len:  %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n",
2953 	       pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx);
2954 	printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n",
2955 	       pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag);
2956 	printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n",
2957 	       pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto);
2958 }
2959 #endif
2960 
2961 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO)
2962 #define IS_TX_OFFLOAD4(pi) ((pi)->ipi_csum_flags & (CSUM_IP_TCP | CSUM_IP_TSO))
2963 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO)
2964 #define IS_TX_OFFLOAD6(pi) ((pi)->ipi_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_TSO))
2965 
2966 static int
2967 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp)
2968 {
2969 	if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx;
2970 	struct ether_vlan_header *eh;
2971 	struct mbuf *m;
2972 
2973 	m = *mp;
2974 	if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) &&
2975 	    M_WRITABLE(m) == 0) {
2976 		if ((m = m_dup(m, M_NOWAIT)) == NULL) {
2977 			return (ENOMEM);
2978 		} else {
2979 			m_freem(*mp);
2980 			DBG_COUNTER_INC(tx_frees);
2981 			*mp = m;
2982 		}
2983 	}
2984 
2985 	/*
2986 	 * Determine where frame payload starts.
2987 	 * Jump over vlan headers if already present,
2988 	 * helpful for QinQ too.
2989 	 */
2990 	if (__predict_false(m->m_len < sizeof(*eh))) {
2991 		txq->ift_pullups++;
2992 		if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL))
2993 			return (ENOMEM);
2994 	}
2995 	eh = mtod(m, struct ether_vlan_header *);
2996 	if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
2997 		pi->ipi_etype = ntohs(eh->evl_proto);
2998 		pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
2999 	} else {
3000 		pi->ipi_etype = ntohs(eh->evl_encap_proto);
3001 		pi->ipi_ehdrlen = ETHER_HDR_LEN;
3002 	}
3003 
3004 	switch (pi->ipi_etype) {
3005 #ifdef INET
3006 	case ETHERTYPE_IP:
3007 	{
3008 		struct mbuf *n;
3009 		struct ip *ip = NULL;
3010 		struct tcphdr *th = NULL;
3011 		int minthlen;
3012 
3013 		minthlen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip) + sizeof(*th));
3014 		if (__predict_false(m->m_len < minthlen)) {
3015 			/*
3016 			 * if this code bloat is causing too much of a hit
3017 			 * move it to a separate function and mark it noinline
3018 			 */
3019 			if (m->m_len == pi->ipi_ehdrlen) {
3020 				n = m->m_next;
3021 				MPASS(n);
3022 				if (n->m_len >= sizeof(*ip))  {
3023 					ip = (struct ip *)n->m_data;
3024 					if (n->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3025 						th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3026 				} else {
3027 					txq->ift_pullups++;
3028 					if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
3029 						return (ENOMEM);
3030 					ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3031 				}
3032 			} else {
3033 				txq->ift_pullups++;
3034 				if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
3035 					return (ENOMEM);
3036 				ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3037 				if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3038 					th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3039 			}
3040 		} else {
3041 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3042 			if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3043 				th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3044 		}
3045 		pi->ipi_ip_hlen = ip->ip_hl << 2;
3046 		pi->ipi_ipproto = ip->ip_p;
3047 		pi->ipi_flags |= IPI_TX_IPV4;
3048 
3049 		/* TCP checksum offload may require TCP header length */
3050 		if (IS_TX_OFFLOAD4(pi)) {
3051 			if (__predict_true(pi->ipi_ipproto == IPPROTO_TCP)) {
3052 				if (__predict_false(th == NULL)) {
3053 					txq->ift_pullups++;
3054 					if (__predict_false((m = m_pullup(m, (ip->ip_hl << 2) + sizeof(*th))) == NULL))
3055 						return (ENOMEM);
3056 					th = (struct tcphdr *)((caddr_t)ip + pi->ipi_ip_hlen);
3057 				}
3058 				pi->ipi_tcp_hflags = th->th_flags;
3059 				pi->ipi_tcp_hlen = th->th_off << 2;
3060 				pi->ipi_tcp_seq = th->th_seq;
3061 			}
3062 			if (IS_TSO4(pi)) {
3063 				if (__predict_false(ip->ip_p != IPPROTO_TCP))
3064 					return (ENXIO);
3065 				/*
3066 				 * TSO always requires hardware checksum offload.
3067 				 */
3068 				pi->ipi_csum_flags |= (CSUM_IP_TCP | CSUM_IP);
3069 				th->th_sum = in_pseudo(ip->ip_src.s_addr,
3070 						       ip->ip_dst.s_addr, htons(IPPROTO_TCP));
3071 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3072 				if (sctx->isc_flags & IFLIB_TSO_INIT_IP) {
3073 					ip->ip_sum = 0;
3074 					ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz);
3075 				}
3076 			}
3077 		}
3078 		if ((sctx->isc_flags & IFLIB_NEED_ZERO_CSUM) && (pi->ipi_csum_flags & CSUM_IP))
3079                        ip->ip_sum = 0;
3080 
3081 		break;
3082 	}
3083 #endif
3084 #ifdef INET6
3085 	case ETHERTYPE_IPV6:
3086 	{
3087 		struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
3088 		struct tcphdr *th;
3089 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
3090 
3091 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
3092 			txq->ift_pullups++;
3093 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
3094 				return (ENOMEM);
3095 		}
3096 		th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
3097 
3098 		/* XXX-BZ this will go badly in case of ext hdrs. */
3099 		pi->ipi_ipproto = ip6->ip6_nxt;
3100 		pi->ipi_flags |= IPI_TX_IPV6;
3101 
3102 		/* TCP checksum offload may require TCP header length */
3103 		if (IS_TX_OFFLOAD6(pi)) {
3104 			if (pi->ipi_ipproto == IPPROTO_TCP) {
3105 				if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) {
3106 					txq->ift_pullups++;
3107 					if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL))
3108 						return (ENOMEM);
3109 				}
3110 				pi->ipi_tcp_hflags = th->th_flags;
3111 				pi->ipi_tcp_hlen = th->th_off << 2;
3112 				pi->ipi_tcp_seq = th->th_seq;
3113 			}
3114 			if (IS_TSO6(pi)) {
3115 				if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP))
3116 					return (ENXIO);
3117 				/*
3118 				 * TSO always requires hardware checksum offload.
3119 				 */
3120 				pi->ipi_csum_flags |= CSUM_IP6_TCP;
3121 				th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0);
3122 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3123 			}
3124 		}
3125 		break;
3126 	}
3127 #endif
3128 	default:
3129 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
3130 		pi->ipi_ip_hlen = 0;
3131 		break;
3132 	}
3133 	*mp = m;
3134 
3135 	return (0);
3136 }
3137 
3138 /*
3139  * If dodgy hardware rejects the scatter gather chain we've handed it
3140  * we'll need to remove the mbuf chain from ifsg_m[] before we can add the
3141  * m_defrag'd mbufs
3142  */
3143 static __noinline struct mbuf *
3144 iflib_remove_mbuf(iflib_txq_t txq)
3145 {
3146 	int ntxd, pidx;
3147 	struct mbuf *m, **ifsd_m;
3148 
3149 	ifsd_m = txq->ift_sds.ifsd_m;
3150 	ntxd = txq->ift_size;
3151 	pidx = txq->ift_pidx & (ntxd - 1);
3152 	ifsd_m = txq->ift_sds.ifsd_m;
3153 	m = ifsd_m[pidx];
3154 	ifsd_m[pidx] = NULL;
3155 	bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[pidx]);
3156 	if (txq->ift_sds.ifsd_tso_map != NULL)
3157 		bus_dmamap_unload(txq->ift_tso_buf_tag,
3158 		    txq->ift_sds.ifsd_tso_map[pidx]);
3159 #if MEMORY_LOGGING
3160 	txq->ift_dequeued++;
3161 #endif
3162 	return (m);
3163 }
3164 
3165 static inline caddr_t
3166 calc_next_txd(iflib_txq_t txq, int cidx, uint8_t qid)
3167 {
3168 	qidx_t size;
3169 	int ntxd;
3170 	caddr_t start, end, cur, next;
3171 
3172 	ntxd = txq->ift_size;
3173 	size = txq->ift_txd_size[qid];
3174 	start = txq->ift_ifdi[qid].idi_vaddr;
3175 
3176 	if (__predict_false(size == 0))
3177 		return (start);
3178 	cur = start + size*cidx;
3179 	end = start + size*ntxd;
3180 	next = CACHE_PTR_NEXT(cur);
3181 	return (next < end ? next : start);
3182 }
3183 
3184 /*
3185  * Pad an mbuf to ensure a minimum ethernet frame size.
3186  * min_frame_size is the frame size (less CRC) to pad the mbuf to
3187  */
3188 static __noinline int
3189 iflib_ether_pad(device_t dev, struct mbuf **m_head, uint16_t min_frame_size)
3190 {
3191 	/*
3192 	 * 18 is enough bytes to pad an ARP packet to 46 bytes, and
3193 	 * and ARP message is the smallest common payload I can think of
3194 	 */
3195 	static char pad[18];	/* just zeros */
3196 	int n;
3197 	struct mbuf *new_head;
3198 
3199 	if (!M_WRITABLE(*m_head)) {
3200 		new_head = m_dup(*m_head, M_NOWAIT);
3201 		if (new_head == NULL) {
3202 			m_freem(*m_head);
3203 			device_printf(dev, "cannot pad short frame, m_dup() failed");
3204 			DBG_COUNTER_INC(encap_pad_mbuf_fail);
3205 			DBG_COUNTER_INC(tx_frees);
3206 			return ENOMEM;
3207 		}
3208 		m_freem(*m_head);
3209 		*m_head = new_head;
3210 	}
3211 
3212 	for (n = min_frame_size - (*m_head)->m_pkthdr.len;
3213 	     n > 0; n -= sizeof(pad))
3214 		if (!m_append(*m_head, min(n, sizeof(pad)), pad))
3215 			break;
3216 
3217 	if (n > 0) {
3218 		m_freem(*m_head);
3219 		device_printf(dev, "cannot pad short frame\n");
3220 		DBG_COUNTER_INC(encap_pad_mbuf_fail);
3221 		DBG_COUNTER_INC(tx_frees);
3222 		return (ENOBUFS);
3223 	}
3224 
3225 	return 0;
3226 }
3227 
3228 static int
3229 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp)
3230 {
3231 	if_ctx_t		ctx;
3232 	if_shared_ctx_t		sctx;
3233 	if_softc_ctx_t		scctx;
3234 	bus_dma_tag_t		buf_tag;
3235 	bus_dma_segment_t	*segs;
3236 	struct mbuf		*m_head, **ifsd_m;
3237 	void			*next_txd;
3238 	bus_dmamap_t		map;
3239 	struct if_pkt_info	pi;
3240 	int remap = 0;
3241 	int err, nsegs, ndesc, max_segs, pidx, cidx, next, ntxd;
3242 
3243 	ctx = txq->ift_ctx;
3244 	sctx = ctx->ifc_sctx;
3245 	scctx = &ctx->ifc_softc_ctx;
3246 	segs = txq->ift_segs;
3247 	ntxd = txq->ift_size;
3248 	m_head = *m_headp;
3249 	map = NULL;
3250 
3251 	/*
3252 	 * If we're doing TSO the next descriptor to clean may be quite far ahead
3253 	 */
3254 	cidx = txq->ift_cidx;
3255 	pidx = txq->ift_pidx;
3256 	if (ctx->ifc_flags & IFC_PREFETCH) {
3257 		next = (cidx + CACHE_PTR_INCREMENT) & (ntxd-1);
3258 		if (!(ctx->ifc_flags & IFLIB_HAS_TXCQ)) {
3259 			next_txd = calc_next_txd(txq, cidx, 0);
3260 			prefetch(next_txd);
3261 		}
3262 
3263 		/* prefetch the next cache line of mbuf pointers and flags */
3264 		prefetch(&txq->ift_sds.ifsd_m[next]);
3265 		prefetch(&txq->ift_sds.ifsd_map[next]);
3266 		next = (cidx + CACHE_LINE_SIZE) & (ntxd-1);
3267 	}
3268 	map = txq->ift_sds.ifsd_map[pidx];
3269 	ifsd_m = txq->ift_sds.ifsd_m;
3270 
3271 	if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
3272 		buf_tag = txq->ift_tso_buf_tag;
3273 		max_segs = scctx->isc_tx_tso_segments_max;
3274 		map = txq->ift_sds.ifsd_tso_map[pidx];
3275 		MPASS(buf_tag != NULL);
3276 		MPASS(max_segs > 0);
3277 	} else {
3278 		buf_tag = txq->ift_buf_tag;
3279 		max_segs = scctx->isc_tx_nsegments;
3280 		map = txq->ift_sds.ifsd_map[pidx];
3281 	}
3282 	if ((sctx->isc_flags & IFLIB_NEED_ETHER_PAD) &&
3283 	    __predict_false(m_head->m_pkthdr.len < scctx->isc_min_frame_size)) {
3284 		err = iflib_ether_pad(ctx->ifc_dev, m_headp, scctx->isc_min_frame_size);
3285 		if (err) {
3286 			DBG_COUNTER_INC(encap_txd_encap_fail);
3287 			return err;
3288 		}
3289 	}
3290 	m_head = *m_headp;
3291 
3292 	pkt_info_zero(&pi);
3293 	pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG|M_BCAST|M_MCAST));
3294 	pi.ipi_pidx = pidx;
3295 	pi.ipi_qsidx = txq->ift_id;
3296 	pi.ipi_len = m_head->m_pkthdr.len;
3297 	pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags;
3298 	pi.ipi_vtag = (m_head->m_flags & M_VLANTAG) ? m_head->m_pkthdr.ether_vtag : 0;
3299 
3300 	/* deliberate bitwise OR to make one condition */
3301 	if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) {
3302 		if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0)) {
3303 			DBG_COUNTER_INC(encap_txd_encap_fail);
3304 			return (err);
3305 		}
3306 		m_head = *m_headp;
3307 	}
3308 
3309 retry:
3310 	err = bus_dmamap_load_mbuf_sg(buf_tag, map, m_head, segs, &nsegs,
3311 	    BUS_DMA_NOWAIT);
3312 defrag:
3313 	if (__predict_false(err)) {
3314 		switch (err) {
3315 		case EFBIG:
3316 			/* try collapse once and defrag once */
3317 			if (remap == 0) {
3318 				m_head = m_collapse(*m_headp, M_NOWAIT, max_segs);
3319 				/* try defrag if collapsing fails */
3320 				if (m_head == NULL)
3321 					remap++;
3322 			}
3323 			if (remap == 1) {
3324 				txq->ift_mbuf_defrag++;
3325 				m_head = m_defrag(*m_headp, M_NOWAIT);
3326 			}
3327 			/*
3328 			 * remap should never be >1 unless bus_dmamap_load_mbuf_sg
3329 			 * failed to map an mbuf that was run through m_defrag
3330 			 */
3331 			MPASS(remap <= 1);
3332 			if (__predict_false(m_head == NULL || remap > 1))
3333 				goto defrag_failed;
3334 			remap++;
3335 			*m_headp = m_head;
3336 			goto retry;
3337 			break;
3338 		case ENOMEM:
3339 			txq->ift_no_tx_dma_setup++;
3340 			break;
3341 		default:
3342 			txq->ift_no_tx_dma_setup++;
3343 			m_freem(*m_headp);
3344 			DBG_COUNTER_INC(tx_frees);
3345 			*m_headp = NULL;
3346 			break;
3347 		}
3348 		txq->ift_map_failed++;
3349 		DBG_COUNTER_INC(encap_load_mbuf_fail);
3350 		DBG_COUNTER_INC(encap_txd_encap_fail);
3351 		return (err);
3352 	}
3353 	ifsd_m[pidx] = m_head;
3354 	/*
3355 	 * XXX assumes a 1 to 1 relationship between segments and
3356 	 *        descriptors - this does not hold true on all drivers, e.g.
3357 	 *        cxgb
3358 	 */
3359 	if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) {
3360 		txq->ift_no_desc_avail++;
3361 		bus_dmamap_unload(buf_tag, map);
3362 		DBG_COUNTER_INC(encap_txq_avail_fail);
3363 		DBG_COUNTER_INC(encap_txd_encap_fail);
3364 		if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0)
3365 			GROUPTASK_ENQUEUE(&txq->ift_task);
3366 		return (ENOBUFS);
3367 	}
3368 	/*
3369 	 * On Intel cards we can greatly reduce the number of TX interrupts
3370 	 * we see by only setting report status on every Nth descriptor.
3371 	 * However, this also means that the driver will need to keep track
3372 	 * of the descriptors that RS was set on to check them for the DD bit.
3373 	 */
3374 	txq->ift_rs_pending += nsegs + 1;
3375 	if (txq->ift_rs_pending > TXQ_MAX_RS_DEFERRED(txq) ||
3376 	     iflib_no_tx_batch || (TXQ_AVAIL(txq) - nsegs) <= MAX_TX_DESC(ctx) + 2) {
3377 		pi.ipi_flags |= IPI_TX_INTR;
3378 		txq->ift_rs_pending = 0;
3379 	}
3380 
3381 	pi.ipi_segs = segs;
3382 	pi.ipi_nsegs = nsegs;
3383 
3384 	MPASS(pidx >= 0 && pidx < txq->ift_size);
3385 #ifdef PKT_DEBUG
3386 	print_pkt(&pi);
3387 #endif
3388 	if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) {
3389 		bus_dmamap_sync(buf_tag, map, BUS_DMASYNC_PREWRITE);
3390 		DBG_COUNTER_INC(tx_encap);
3391 		MPASS(pi.ipi_new_pidx < txq->ift_size);
3392 
3393 		ndesc = pi.ipi_new_pidx - pi.ipi_pidx;
3394 		if (pi.ipi_new_pidx < pi.ipi_pidx) {
3395 			ndesc += txq->ift_size;
3396 			txq->ift_gen = 1;
3397 		}
3398 		/*
3399 		 * drivers can need as many as
3400 		 * two sentinels
3401 		 */
3402 		MPASS(ndesc <= pi.ipi_nsegs + 2);
3403 		MPASS(pi.ipi_new_pidx != pidx);
3404 		MPASS(ndesc > 0);
3405 		txq->ift_in_use += ndesc;
3406 
3407 		/*
3408 		 * We update the last software descriptor again here because there may
3409 		 * be a sentinel and/or there may be more mbufs than segments
3410 		 */
3411 		txq->ift_pidx = pi.ipi_new_pidx;
3412 		txq->ift_npending += pi.ipi_ndescs;
3413 	} else {
3414 		*m_headp = m_head = iflib_remove_mbuf(txq);
3415 		if (err == EFBIG) {
3416 			txq->ift_txd_encap_efbig++;
3417 			if (remap < 2) {
3418 				remap = 1;
3419 				goto defrag;
3420 			}
3421 		}
3422 		goto defrag_failed;
3423 	}
3424 	/*
3425 	 * err can't possibly be non-zero here, so we don't neet to test it
3426 	 * to see if we need to DBG_COUNTER_INC(encap_txd_encap_fail).
3427 	 */
3428 	return (err);
3429 
3430 defrag_failed:
3431 	txq->ift_mbuf_defrag_failed++;
3432 	txq->ift_map_failed++;
3433 	m_freem(*m_headp);
3434 	DBG_COUNTER_INC(tx_frees);
3435 	*m_headp = NULL;
3436 	DBG_COUNTER_INC(encap_txd_encap_fail);
3437 	return (ENOMEM);
3438 }
3439 
3440 static void
3441 iflib_tx_desc_free(iflib_txq_t txq, int n)
3442 {
3443 	uint32_t qsize, cidx, mask, gen;
3444 	struct mbuf *m, **ifsd_m;
3445 	bool do_prefetch;
3446 
3447 	cidx = txq->ift_cidx;
3448 	gen = txq->ift_gen;
3449 	qsize = txq->ift_size;
3450 	mask = qsize-1;
3451 	ifsd_m = txq->ift_sds.ifsd_m;
3452 	do_prefetch = (txq->ift_ctx->ifc_flags & IFC_PREFETCH);
3453 
3454 	while (n-- > 0) {
3455 		if (do_prefetch) {
3456 			prefetch(ifsd_m[(cidx + 3) & mask]);
3457 			prefetch(ifsd_m[(cidx + 4) & mask]);
3458 		}
3459 		if ((m = ifsd_m[cidx]) != NULL) {
3460 			prefetch(&ifsd_m[(cidx + CACHE_PTR_INCREMENT) & mask]);
3461 			if (m->m_pkthdr.csum_flags & CSUM_TSO) {
3462 				bus_dmamap_sync(txq->ift_tso_buf_tag,
3463 				    txq->ift_sds.ifsd_tso_map[cidx],
3464 				    BUS_DMASYNC_POSTWRITE);
3465 				bus_dmamap_unload(txq->ift_tso_buf_tag,
3466 				    txq->ift_sds.ifsd_tso_map[cidx]);
3467 			} else {
3468 				bus_dmamap_sync(txq->ift_buf_tag,
3469 				    txq->ift_sds.ifsd_map[cidx],
3470 				    BUS_DMASYNC_POSTWRITE);
3471 				bus_dmamap_unload(txq->ift_buf_tag,
3472 				    txq->ift_sds.ifsd_map[cidx]);
3473 			}
3474 			/* XXX we don't support any drivers that batch packets yet */
3475 			MPASS(m->m_nextpkt == NULL);
3476 			m_freem(m);
3477 			ifsd_m[cidx] = NULL;
3478 #if MEMORY_LOGGING
3479 			txq->ift_dequeued++;
3480 #endif
3481 			DBG_COUNTER_INC(tx_frees);
3482 		}
3483 		if (__predict_false(++cidx == qsize)) {
3484 			cidx = 0;
3485 			gen = 0;
3486 		}
3487 	}
3488 	txq->ift_cidx = cidx;
3489 	txq->ift_gen = gen;
3490 }
3491 
3492 static __inline int
3493 iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh)
3494 {
3495 	int reclaim;
3496 	if_ctx_t ctx = txq->ift_ctx;
3497 
3498 	KASSERT(thresh >= 0, ("invalid threshold to reclaim"));
3499 	MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size);
3500 
3501 	/*
3502 	 * Need a rate-limiting check so that this isn't called every time
3503 	 */
3504 	iflib_tx_credits_update(ctx, txq);
3505 	reclaim = DESC_RECLAIMABLE(txq);
3506 
3507 	if (reclaim <= thresh /* + MAX_TX_DESC(txq->ift_ctx) */) {
3508 #ifdef INVARIANTS
3509 		if (iflib_verbose_debug) {
3510 			printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __FUNCTION__,
3511 			       txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments,
3512 			       reclaim, thresh);
3513 
3514 		}
3515 #endif
3516 		return (0);
3517 	}
3518 	iflib_tx_desc_free(txq, reclaim);
3519 	txq->ift_cleaned += reclaim;
3520 	txq->ift_in_use -= reclaim;
3521 
3522 	return (reclaim);
3523 }
3524 
3525 static struct mbuf **
3526 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset, int remaining)
3527 {
3528 	int next, size;
3529 	struct mbuf **items;
3530 
3531 	size = r->size;
3532 	next = (cidx + CACHE_PTR_INCREMENT) & (size-1);
3533 	items = __DEVOLATILE(struct mbuf **, &r->items[0]);
3534 
3535 	prefetch(items[(cidx + offset) & (size-1)]);
3536 	if (remaining > 1) {
3537 		prefetch2cachelines(&items[next]);
3538 		prefetch2cachelines(items[(cidx + offset + 1) & (size-1)]);
3539 		prefetch2cachelines(items[(cidx + offset + 2) & (size-1)]);
3540 		prefetch2cachelines(items[(cidx + offset + 3) & (size-1)]);
3541 	}
3542 	return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (size-1)]));
3543 }
3544 
3545 static void
3546 iflib_txq_check_drain(iflib_txq_t txq, int budget)
3547 {
3548 
3549 	ifmp_ring_check_drainage(txq->ift_br, budget);
3550 }
3551 
3552 static uint32_t
3553 iflib_txq_can_drain(struct ifmp_ring *r)
3554 {
3555 	iflib_txq_t txq = r->cookie;
3556 	if_ctx_t ctx = txq->ift_ctx;
3557 
3558 	if (TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2)
3559 		return (1);
3560 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3561 	    BUS_DMASYNC_POSTREAD);
3562 	return (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id,
3563 	    false));
3564 }
3565 
3566 static uint32_t
3567 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3568 {
3569 	iflib_txq_t txq = r->cookie;
3570 	if_ctx_t ctx = txq->ift_ctx;
3571 	struct ifnet *ifp = ctx->ifc_ifp;
3572 	struct mbuf **mp, *m;
3573 	int i, count, consumed, pkt_sent, bytes_sent, mcast_sent, avail;
3574 	int reclaimed, err, in_use_prev, desc_used;
3575 	bool do_prefetch, ring, rang;
3576 
3577 	if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) ||
3578 			    !LINK_ACTIVE(ctx))) {
3579 		DBG_COUNTER_INC(txq_drain_notready);
3580 		return (0);
3581 	}
3582 	reclaimed = iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
3583 	rang = iflib_txd_db_check(ctx, txq, reclaimed, txq->ift_in_use);
3584 	avail = IDXDIFF(pidx, cidx, r->size);
3585 	if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) {
3586 		DBG_COUNTER_INC(txq_drain_flushing);
3587 		for (i = 0; i < avail; i++) {
3588 			if (__predict_true(r->items[(cidx + i) & (r->size-1)] != (void *)txq))
3589 				m_free(r->items[(cidx + i) & (r->size-1)]);
3590 			r->items[(cidx + i) & (r->size-1)] = NULL;
3591 		}
3592 		return (avail);
3593 	}
3594 
3595 	if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3596 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3597 		CALLOUT_LOCK(txq);
3598 		callout_stop(&txq->ift_timer);
3599 		CALLOUT_UNLOCK(txq);
3600 		DBG_COUNTER_INC(txq_drain_oactive);
3601 		return (0);
3602 	}
3603 	if (reclaimed)
3604 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3605 	consumed = mcast_sent = bytes_sent = pkt_sent = 0;
3606 	count = MIN(avail, TX_BATCH_SIZE);
3607 #ifdef INVARIANTS
3608 	if (iflib_verbose_debug)
3609 		printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __FUNCTION__,
3610 		       avail, ctx->ifc_flags, TXQ_AVAIL(txq));
3611 #endif
3612 	do_prefetch = (ctx->ifc_flags & IFC_PREFETCH);
3613 	avail = TXQ_AVAIL(txq);
3614 	err = 0;
3615 	for (desc_used = i = 0; i < count && avail > MAX_TX_DESC(ctx) + 2; i++) {
3616 		int rem = do_prefetch ? count - i : 0;
3617 
3618 		mp = _ring_peek_one(r, cidx, i, rem);
3619 		MPASS(mp != NULL && *mp != NULL);
3620 		if (__predict_false(*mp == (struct mbuf *)txq)) {
3621 			consumed++;
3622 			reclaimed++;
3623 			continue;
3624 		}
3625 		in_use_prev = txq->ift_in_use;
3626 		err = iflib_encap(txq, mp);
3627 		if (__predict_false(err)) {
3628 			/* no room - bail out */
3629 			if (err == ENOBUFS)
3630 				break;
3631 			consumed++;
3632 			/* we can't send this packet - skip it */
3633 			continue;
3634 		}
3635 		consumed++;
3636 		pkt_sent++;
3637 		m = *mp;
3638 		DBG_COUNTER_INC(tx_sent);
3639 		bytes_sent += m->m_pkthdr.len;
3640 		mcast_sent += !!(m->m_flags & M_MCAST);
3641 		avail = TXQ_AVAIL(txq);
3642 
3643 		txq->ift_db_pending += (txq->ift_in_use - in_use_prev);
3644 		desc_used += (txq->ift_in_use - in_use_prev);
3645 		ETHER_BPF_MTAP(ifp, m);
3646 		if (__predict_false(!(ifp->if_drv_flags & IFF_DRV_RUNNING)))
3647 			break;
3648 		rang = iflib_txd_db_check(ctx, txq, false, in_use_prev);
3649 	}
3650 
3651 	/* deliberate use of bitwise or to avoid gratuitous short-circuit */
3652 	ring = rang ? false  : (iflib_min_tx_latency | err) || (TXQ_AVAIL(txq) < MAX_TX_DESC(ctx));
3653 	iflib_txd_db_check(ctx, txq, ring, txq->ift_in_use);
3654 	if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
3655 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
3656 	if (mcast_sent)
3657 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
3658 #ifdef INVARIANTS
3659 	if (iflib_verbose_debug)
3660 		printf("consumed=%d\n", consumed);
3661 #endif
3662 	return (consumed);
3663 }
3664 
3665 static uint32_t
3666 iflib_txq_drain_always(struct ifmp_ring *r)
3667 {
3668 	return (1);
3669 }
3670 
3671 static uint32_t
3672 iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3673 {
3674 	int i, avail;
3675 	struct mbuf **mp;
3676 	iflib_txq_t txq;
3677 
3678 	txq = r->cookie;
3679 
3680 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3681 	CALLOUT_LOCK(txq);
3682 	callout_stop(&txq->ift_timer);
3683 	CALLOUT_UNLOCK(txq);
3684 
3685 	avail = IDXDIFF(pidx, cidx, r->size);
3686 	for (i = 0; i < avail; i++) {
3687 		mp = _ring_peek_one(r, cidx, i, avail - i);
3688 		if (__predict_false(*mp == (struct mbuf *)txq))
3689 			continue;
3690 		m_freem(*mp);
3691 		DBG_COUNTER_INC(tx_frees);
3692 	}
3693 	MPASS(ifmp_ring_is_stalled(r) == 0);
3694 	return (avail);
3695 }
3696 
3697 static void
3698 iflib_ifmp_purge(iflib_txq_t txq)
3699 {
3700 	struct ifmp_ring *r;
3701 
3702 	r = txq->ift_br;
3703 	r->drain = iflib_txq_drain_free;
3704 	r->can_drain = iflib_txq_drain_always;
3705 
3706 	ifmp_ring_check_drainage(r, r->size);
3707 
3708 	r->drain = iflib_txq_drain;
3709 	r->can_drain = iflib_txq_can_drain;
3710 }
3711 
3712 static void
3713 _task_fn_tx(void *context)
3714 {
3715 	iflib_txq_t txq = context;
3716 	if_ctx_t ctx = txq->ift_ctx;
3717 #if defined(ALTQ) || defined(DEV_NETMAP)
3718 	if_t ifp = ctx->ifc_ifp;
3719 #endif
3720 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
3721 
3722 #ifdef IFLIB_DIAGNOSTICS
3723 	txq->ift_cpu_exec_count[curcpu]++;
3724 #endif
3725 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3726 		return;
3727 #ifdef DEV_NETMAP
3728 	if (if_getcapenable(ifp) & IFCAP_NETMAP) {
3729 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3730 		    BUS_DMASYNC_POSTREAD);
3731 		if (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false))
3732 			netmap_tx_irq(ifp, txq->ift_id);
3733 		IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3734 		return;
3735 	}
3736 #endif
3737 #ifdef ALTQ
3738 	if (ALTQ_IS_ENABLED(&ifp->if_snd))
3739 		iflib_altq_if_start(ifp);
3740 #endif
3741 	if (txq->ift_db_pending)
3742 		ifmp_ring_enqueue(txq->ift_br, (void **)&txq, 1, TX_BATCH_SIZE, abdicate);
3743 	else if (!abdicate)
3744 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3745 	/*
3746 	 * When abdicating, we always need to check drainage, not just when we don't enqueue
3747 	 */
3748 	if (abdicate)
3749 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3750 	if (ctx->ifc_flags & IFC_LEGACY)
3751 		IFDI_INTR_ENABLE(ctx);
3752 	else {
3753 #ifdef INVARIANTS
3754 		int rc =
3755 #endif
3756 			IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3757 			KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver"));
3758 	}
3759 }
3760 
3761 static void
3762 _task_fn_rx(void *context)
3763 {
3764 	iflib_rxq_t rxq = context;
3765 	if_ctx_t ctx = rxq->ifr_ctx;
3766 	bool more;
3767 	uint16_t budget;
3768 
3769 #ifdef IFLIB_DIAGNOSTICS
3770 	rxq->ifr_cpu_exec_count[curcpu]++;
3771 #endif
3772 	DBG_COUNTER_INC(task_fn_rxs);
3773 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3774 		return;
3775 	more = true;
3776 #ifdef DEV_NETMAP
3777 	if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP) {
3778 		u_int work = 0;
3779 		if (netmap_rx_irq(ctx->ifc_ifp, rxq->ifr_id, &work)) {
3780 			more = false;
3781 		}
3782 	}
3783 #endif
3784 	budget = ctx->ifc_sysctl_rx_budget;
3785 	if (budget == 0)
3786 		budget = 16;	/* XXX */
3787 	if (more == false || (more = iflib_rxeof(rxq, budget)) == false) {
3788 		if (ctx->ifc_flags & IFC_LEGACY)
3789 			IFDI_INTR_ENABLE(ctx);
3790 		else {
3791 #ifdef INVARIANTS
3792 			int rc =
3793 #endif
3794 				IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
3795 			KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver"));
3796 			DBG_COUNTER_INC(rx_intr_enables);
3797 		}
3798 	}
3799 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3800 		return;
3801 	if (more)
3802 		GROUPTASK_ENQUEUE(&rxq->ifr_task);
3803 }
3804 
3805 static void
3806 _task_fn_admin(void *context)
3807 {
3808 	if_ctx_t ctx = context;
3809 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
3810 	iflib_txq_t txq;
3811 	int i;
3812 	bool oactive, running, do_reset, do_watchdog, in_detach;
3813 	uint32_t reset_on = hz / 2;
3814 
3815 	STATE_LOCK(ctx);
3816 	running = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING);
3817 	oactive = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE);
3818 	do_reset = (ctx->ifc_flags & IFC_DO_RESET);
3819 	do_watchdog = (ctx->ifc_flags & IFC_DO_WATCHDOG);
3820 	in_detach = (ctx->ifc_flags & IFC_IN_DETACH);
3821 	ctx->ifc_flags &= ~(IFC_DO_RESET|IFC_DO_WATCHDOG);
3822 	STATE_UNLOCK(ctx);
3823 
3824 	if ((!running && !oactive) && !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
3825 		return;
3826 	if (in_detach)
3827 		return;
3828 
3829 	CTX_LOCK(ctx);
3830 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3831 		CALLOUT_LOCK(txq);
3832 		callout_stop(&txq->ift_timer);
3833 		CALLOUT_UNLOCK(txq);
3834 	}
3835 	if (do_watchdog) {
3836 		ctx->ifc_watchdog_events++;
3837 		IFDI_WATCHDOG_RESET(ctx);
3838 	}
3839 	IFDI_UPDATE_ADMIN_STATUS(ctx);
3840 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3841 #ifdef DEV_NETMAP
3842 		reset_on = hz / 2;
3843 		if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP)
3844 			iflib_netmap_timer_adjust(ctx, txq, &reset_on);
3845 #endif
3846 		callout_reset_on(&txq->ift_timer, reset_on, iflib_timer, txq, txq->ift_timer.c_cpu);
3847 	}
3848 	IFDI_LINK_INTR_ENABLE(ctx);
3849 	if (do_reset)
3850 		iflib_if_init_locked(ctx);
3851 	CTX_UNLOCK(ctx);
3852 
3853 	if (LINK_ACTIVE(ctx) == 0)
3854 		return;
3855 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3856 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
3857 }
3858 
3859 
3860 static void
3861 _task_fn_iov(void *context)
3862 {
3863 	if_ctx_t ctx = context;
3864 
3865 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING) &&
3866 	    !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
3867 		return;
3868 
3869 	CTX_LOCK(ctx);
3870 	IFDI_VFLR_HANDLE(ctx);
3871 	CTX_UNLOCK(ctx);
3872 }
3873 
3874 static int
3875 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)
3876 {
3877 	int err;
3878 	if_int_delay_info_t info;
3879 	if_ctx_t ctx;
3880 
3881 	info = (if_int_delay_info_t)arg1;
3882 	ctx = info->iidi_ctx;
3883 	info->iidi_req = req;
3884 	info->iidi_oidp = oidp;
3885 	CTX_LOCK(ctx);
3886 	err = IFDI_SYSCTL_INT_DELAY(ctx, info);
3887 	CTX_UNLOCK(ctx);
3888 	return (err);
3889 }
3890 
3891 /*********************************************************************
3892  *
3893  *  IFNET FUNCTIONS
3894  *
3895  **********************************************************************/
3896 
3897 static void
3898 iflib_if_init_locked(if_ctx_t ctx)
3899 {
3900 	iflib_stop(ctx);
3901 	iflib_init_locked(ctx);
3902 }
3903 
3904 
3905 static void
3906 iflib_if_init(void *arg)
3907 {
3908 	if_ctx_t ctx = arg;
3909 
3910 	CTX_LOCK(ctx);
3911 	iflib_if_init_locked(ctx);
3912 	CTX_UNLOCK(ctx);
3913 }
3914 
3915 static int
3916 iflib_if_transmit(if_t ifp, struct mbuf *m)
3917 {
3918 	if_ctx_t	ctx = if_getsoftc(ifp);
3919 
3920 	iflib_txq_t txq;
3921 	int err, qidx;
3922 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
3923 
3924 	if (__predict_false((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) {
3925 		DBG_COUNTER_INC(tx_frees);
3926 		m_freem(m);
3927 		return (ENETDOWN);
3928 	}
3929 
3930 	MPASS(m->m_nextpkt == NULL);
3931 	/* ALTQ-enabled interfaces always use queue 0. */
3932 	qidx = 0;
3933 	if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m) && !ALTQ_IS_ENABLED(&ifp->if_snd))
3934 		qidx = QIDX(ctx, m);
3935 	/*
3936 	 * XXX calculate buf_ring based on flowid (divvy up bits?)
3937 	 */
3938 	txq = &ctx->ifc_txqs[qidx];
3939 
3940 #ifdef DRIVER_BACKPRESSURE
3941 	if (txq->ift_closed) {
3942 		while (m != NULL) {
3943 			next = m->m_nextpkt;
3944 			m->m_nextpkt = NULL;
3945 			m_freem(m);
3946 			DBG_COUNTER_INC(tx_frees);
3947 			m = next;
3948 		}
3949 		return (ENOBUFS);
3950 	}
3951 #endif
3952 #ifdef notyet
3953 	qidx = count = 0;
3954 	mp = marr;
3955 	next = m;
3956 	do {
3957 		count++;
3958 		next = next->m_nextpkt;
3959 	} while (next != NULL);
3960 
3961 	if (count > nitems(marr))
3962 		if ((mp = malloc(count*sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) {
3963 			/* XXX check nextpkt */
3964 			m_freem(m);
3965 			/* XXX simplify for now */
3966 			DBG_COUNTER_INC(tx_frees);
3967 			return (ENOBUFS);
3968 		}
3969 	for (next = m, i = 0; next != NULL; i++) {
3970 		mp[i] = next;
3971 		next = next->m_nextpkt;
3972 		mp[i]->m_nextpkt = NULL;
3973 	}
3974 #endif
3975 	DBG_COUNTER_INC(tx_seen);
3976 	err = ifmp_ring_enqueue(txq->ift_br, (void **)&m, 1, TX_BATCH_SIZE, abdicate);
3977 
3978 	if (abdicate)
3979 		GROUPTASK_ENQUEUE(&txq->ift_task);
3980  	if (err) {
3981 		if (!abdicate)
3982 			GROUPTASK_ENQUEUE(&txq->ift_task);
3983 		/* support forthcoming later */
3984 #ifdef DRIVER_BACKPRESSURE
3985 		txq->ift_closed = TRUE;
3986 #endif
3987 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3988 		m_freem(m);
3989 		DBG_COUNTER_INC(tx_frees);
3990 	}
3991 
3992 	return (err);
3993 }
3994 
3995 #ifdef ALTQ
3996 /*
3997  * The overall approach to integrating iflib with ALTQ is to continue to use
3998  * the iflib mp_ring machinery between the ALTQ queue(s) and the hardware
3999  * ring.  Technically, when using ALTQ, queueing to an intermediate mp_ring
4000  * is redundant/unnecessary, but doing so minimizes the amount of
4001  * ALTQ-specific code required in iflib.  It is assumed that the overhead of
4002  * redundantly queueing to an intermediate mp_ring is swamped by the
4003  * performance limitations inherent in using ALTQ.
4004  *
4005  * When ALTQ support is compiled in, all iflib drivers will use a transmit
4006  * routine, iflib_altq_if_transmit(), that checks if ALTQ is enabled for the
4007  * given interface.  If ALTQ is enabled for an interface, then all
4008  * transmitted packets for that interface will be submitted to the ALTQ
4009  * subsystem via IFQ_ENQUEUE().  We don't use the legacy if_transmit()
4010  * implementation because it uses IFQ_HANDOFF(), which will duplicatively
4011  * update stats that the iflib machinery handles, and which is sensitve to
4012  * the disused IFF_DRV_OACTIVE flag.  Additionally, iflib_altq_if_start()
4013  * will be installed as the start routine for use by ALTQ facilities that
4014  * need to trigger queue drains on a scheduled basis.
4015  *
4016  */
4017 static void
4018 iflib_altq_if_start(if_t ifp)
4019 {
4020 	struct ifaltq *ifq = &ifp->if_snd;
4021 	struct mbuf *m;
4022 
4023 	IFQ_LOCK(ifq);
4024 	IFQ_DEQUEUE_NOLOCK(ifq, m);
4025 	while (m != NULL) {
4026 		iflib_if_transmit(ifp, m);
4027 		IFQ_DEQUEUE_NOLOCK(ifq, m);
4028 	}
4029 	IFQ_UNLOCK(ifq);
4030 }
4031 
4032 static int
4033 iflib_altq_if_transmit(if_t ifp, struct mbuf *m)
4034 {
4035 	int err;
4036 
4037 	if (ALTQ_IS_ENABLED(&ifp->if_snd)) {
4038 		IFQ_ENQUEUE(&ifp->if_snd, m, err);
4039 		if (err == 0)
4040 			iflib_altq_if_start(ifp);
4041 	} else
4042 		err = iflib_if_transmit(ifp, m);
4043 
4044 	return (err);
4045 }
4046 #endif /* ALTQ */
4047 
4048 static void
4049 iflib_if_qflush(if_t ifp)
4050 {
4051 	if_ctx_t ctx = if_getsoftc(ifp);
4052 	iflib_txq_t txq = ctx->ifc_txqs;
4053 	int i;
4054 
4055 	STATE_LOCK(ctx);
4056 	ctx->ifc_flags |= IFC_QFLUSH;
4057 	STATE_UNLOCK(ctx);
4058 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
4059 		while (!(ifmp_ring_is_idle(txq->ift_br) || ifmp_ring_is_stalled(txq->ift_br)))
4060 			iflib_txq_check_drain(txq, 0);
4061 	STATE_LOCK(ctx);
4062 	ctx->ifc_flags &= ~IFC_QFLUSH;
4063 	STATE_UNLOCK(ctx);
4064 
4065 	/*
4066 	 * When ALTQ is enabled, this will also take care of purging the
4067 	 * ALTQ queue(s).
4068 	 */
4069 	if_qflush(ifp);
4070 }
4071 
4072 
4073 #define IFCAP_FLAGS (IFCAP_HWCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \
4074 		     IFCAP_TSO | IFCAP_VLAN_HWTAGGING | IFCAP_HWSTATS | \
4075 		     IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | \
4076 		     IFCAP_VLAN_HWTSO | IFCAP_VLAN_HWCSUM)
4077 
4078 static int
4079 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data)
4080 {
4081 	if_ctx_t ctx = if_getsoftc(ifp);
4082 	struct ifreq	*ifr = (struct ifreq *)data;
4083 #if defined(INET) || defined(INET6)
4084 	struct ifaddr	*ifa = (struct ifaddr *)data;
4085 #endif
4086 	bool		avoid_reset = FALSE;
4087 	int		err = 0, reinit = 0, bits;
4088 
4089 	switch (command) {
4090 	case SIOCSIFADDR:
4091 #ifdef INET
4092 		if (ifa->ifa_addr->sa_family == AF_INET)
4093 			avoid_reset = TRUE;
4094 #endif
4095 #ifdef INET6
4096 		if (ifa->ifa_addr->sa_family == AF_INET6)
4097 			avoid_reset = TRUE;
4098 #endif
4099 		/*
4100 		** Calling init results in link renegotiation,
4101 		** so we avoid doing it when possible.
4102 		*/
4103 		if (avoid_reset) {
4104 			if_setflagbits(ifp, IFF_UP,0);
4105 			if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
4106 				reinit = 1;
4107 #ifdef INET
4108 			if (!(if_getflags(ifp) & IFF_NOARP))
4109 				arp_ifinit(ifp, ifa);
4110 #endif
4111 		} else
4112 			err = ether_ioctl(ifp, command, data);
4113 		break;
4114 	case SIOCSIFMTU:
4115 		CTX_LOCK(ctx);
4116 		if (ifr->ifr_mtu == if_getmtu(ifp)) {
4117 			CTX_UNLOCK(ctx);
4118 			break;
4119 		}
4120 		bits = if_getdrvflags(ifp);
4121 		/* stop the driver and free any clusters before proceeding */
4122 		iflib_stop(ctx);
4123 
4124 		if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) {
4125 			STATE_LOCK(ctx);
4126 			if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size)
4127 				ctx->ifc_flags |= IFC_MULTISEG;
4128 			else
4129 				ctx->ifc_flags &= ~IFC_MULTISEG;
4130 			STATE_UNLOCK(ctx);
4131 			err = if_setmtu(ifp, ifr->ifr_mtu);
4132 		}
4133 		iflib_init_locked(ctx);
4134 		STATE_LOCK(ctx);
4135 		if_setdrvflags(ifp, bits);
4136 		STATE_UNLOCK(ctx);
4137 		CTX_UNLOCK(ctx);
4138 		break;
4139 	case SIOCSIFFLAGS:
4140 		CTX_LOCK(ctx);
4141 		if (if_getflags(ifp) & IFF_UP) {
4142 			if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4143 				if ((if_getflags(ifp) ^ ctx->ifc_if_flags) &
4144 				    (IFF_PROMISC | IFF_ALLMULTI)) {
4145 					err = IFDI_PROMISC_SET(ctx, if_getflags(ifp));
4146 				}
4147 			} else
4148 				reinit = 1;
4149 		} else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4150 			iflib_stop(ctx);
4151 		}
4152 		ctx->ifc_if_flags = if_getflags(ifp);
4153 		CTX_UNLOCK(ctx);
4154 		break;
4155 	case SIOCADDMULTI:
4156 	case SIOCDELMULTI:
4157 		if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4158 			CTX_LOCK(ctx);
4159 			IFDI_INTR_DISABLE(ctx);
4160 			IFDI_MULTI_SET(ctx);
4161 			IFDI_INTR_ENABLE(ctx);
4162 			CTX_UNLOCK(ctx);
4163 		}
4164 		break;
4165 	case SIOCSIFMEDIA:
4166 		CTX_LOCK(ctx);
4167 		IFDI_MEDIA_SET(ctx);
4168 		CTX_UNLOCK(ctx);
4169 		/* falls thru */
4170 	case SIOCGIFMEDIA:
4171 	case SIOCGIFXMEDIA:
4172 		err = ifmedia_ioctl(ifp, ifr, &ctx->ifc_media, command);
4173 		break;
4174 	case SIOCGI2C:
4175 	{
4176 		struct ifi2creq i2c;
4177 
4178 		err = copyin(ifr_data_get_ptr(ifr), &i2c, sizeof(i2c));
4179 		if (err != 0)
4180 			break;
4181 		if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) {
4182 			err = EINVAL;
4183 			break;
4184 		}
4185 		if (i2c.len > sizeof(i2c.data)) {
4186 			err = EINVAL;
4187 			break;
4188 		}
4189 
4190 		if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0)
4191 			err = copyout(&i2c, ifr_data_get_ptr(ifr),
4192 			    sizeof(i2c));
4193 		break;
4194 	}
4195 	case SIOCSIFCAP:
4196 	{
4197 		int mask, setmask, oldmask;
4198 
4199 		oldmask = if_getcapenable(ifp);
4200 		mask = ifr->ifr_reqcap ^ oldmask;
4201 		mask &= ctx->ifc_softc_ctx.isc_capabilities;
4202 		setmask = 0;
4203 #ifdef TCP_OFFLOAD
4204 		setmask |= mask & (IFCAP_TOE4|IFCAP_TOE6);
4205 #endif
4206 		setmask |= (mask & IFCAP_FLAGS);
4207 		setmask |= (mask & IFCAP_WOL);
4208 
4209 		/*
4210 		 * If any RX csum has changed, change all the ones that
4211 		 * are supported by the driver.
4212 		 */
4213 		if (setmask & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6)) {
4214 			setmask |= ctx->ifc_softc_ctx.isc_capabilities &
4215 			    (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6);
4216 		}
4217 
4218 		/*
4219 		 * want to ensure that traffic has stopped before we change any of the flags
4220 		 */
4221 		if (setmask) {
4222 			CTX_LOCK(ctx);
4223 			bits = if_getdrvflags(ifp);
4224 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4225 				iflib_stop(ctx);
4226 			STATE_LOCK(ctx);
4227 			if_togglecapenable(ifp, setmask);
4228 			STATE_UNLOCK(ctx);
4229 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4230 				iflib_init_locked(ctx);
4231 			STATE_LOCK(ctx);
4232 			if_setdrvflags(ifp, bits);
4233 			STATE_UNLOCK(ctx);
4234 			CTX_UNLOCK(ctx);
4235 		}
4236 		if_vlancap(ifp);
4237 		break;
4238 	}
4239 	case SIOCGPRIVATE_0:
4240 	case SIOCSDRVSPEC:
4241 	case SIOCGDRVSPEC:
4242 		CTX_LOCK(ctx);
4243 		err = IFDI_PRIV_IOCTL(ctx, command, data);
4244 		CTX_UNLOCK(ctx);
4245 		break;
4246 	default:
4247 		err = ether_ioctl(ifp, command, data);
4248 		break;
4249 	}
4250 	if (reinit)
4251 		iflib_if_init(ctx);
4252 	return (err);
4253 }
4254 
4255 static uint64_t
4256 iflib_if_get_counter(if_t ifp, ift_counter cnt)
4257 {
4258 	if_ctx_t ctx = if_getsoftc(ifp);
4259 
4260 	return (IFDI_GET_COUNTER(ctx, cnt));
4261 }
4262 
4263 /*********************************************************************
4264  *
4265  *  OTHER FUNCTIONS EXPORTED TO THE STACK
4266  *
4267  **********************************************************************/
4268 
4269 static void
4270 iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag)
4271 {
4272 	if_ctx_t ctx = if_getsoftc(ifp);
4273 
4274 	if ((void *)ctx != arg)
4275 		return;
4276 
4277 	if ((vtag == 0) || (vtag > 4095))
4278 		return;
4279 
4280 	CTX_LOCK(ctx);
4281 	IFDI_VLAN_REGISTER(ctx, vtag);
4282 	/* Re-init to load the changes */
4283 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
4284 		iflib_if_init_locked(ctx);
4285 	CTX_UNLOCK(ctx);
4286 }
4287 
4288 static void
4289 iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag)
4290 {
4291 	if_ctx_t ctx = if_getsoftc(ifp);
4292 
4293 	if ((void *)ctx != arg)
4294 		return;
4295 
4296 	if ((vtag == 0) || (vtag > 4095))
4297 		return;
4298 
4299 	CTX_LOCK(ctx);
4300 	IFDI_VLAN_UNREGISTER(ctx, vtag);
4301 	/* Re-init to load the changes */
4302 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
4303 		iflib_if_init_locked(ctx);
4304 	CTX_UNLOCK(ctx);
4305 }
4306 
4307 static void
4308 iflib_led_func(void *arg, int onoff)
4309 {
4310 	if_ctx_t ctx = arg;
4311 
4312 	CTX_LOCK(ctx);
4313 	IFDI_LED_FUNC(ctx, onoff);
4314 	CTX_UNLOCK(ctx);
4315 }
4316 
4317 /*********************************************************************
4318  *
4319  *  BUS FUNCTION DEFINITIONS
4320  *
4321  **********************************************************************/
4322 
4323 int
4324 iflib_device_probe(device_t dev)
4325 {
4326 	pci_vendor_info_t *ent;
4327 
4328 	uint16_t	pci_vendor_id, pci_device_id;
4329 	uint16_t	pci_subvendor_id, pci_subdevice_id;
4330 	uint16_t	pci_rev_id;
4331 	if_shared_ctx_t sctx;
4332 
4333 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4334 		return (ENOTSUP);
4335 
4336 	pci_vendor_id = pci_get_vendor(dev);
4337 	pci_device_id = pci_get_device(dev);
4338 	pci_subvendor_id = pci_get_subvendor(dev);
4339 	pci_subdevice_id = pci_get_subdevice(dev);
4340 	pci_rev_id = pci_get_revid(dev);
4341 	if (sctx->isc_parse_devinfo != NULL)
4342 		sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id);
4343 
4344 	ent = sctx->isc_vendor_info;
4345 	while (ent->pvi_vendor_id != 0) {
4346 		if (pci_vendor_id != ent->pvi_vendor_id) {
4347 			ent++;
4348 			continue;
4349 		}
4350 		if ((pci_device_id == ent->pvi_device_id) &&
4351 		    ((pci_subvendor_id == ent->pvi_subvendor_id) ||
4352 		     (ent->pvi_subvendor_id == 0)) &&
4353 		    ((pci_subdevice_id == ent->pvi_subdevice_id) ||
4354 		     (ent->pvi_subdevice_id == 0)) &&
4355 		    ((pci_rev_id == ent->pvi_rev_id) ||
4356 		     (ent->pvi_rev_id == 0))) {
4357 
4358 			device_set_desc_copy(dev, ent->pvi_name);
4359 			/* this needs to be changed to zero if the bus probing code
4360 			 * ever stops re-probing on best match because the sctx
4361 			 * may have its values over written by register calls
4362 			 * in subsequent probes
4363 			 */
4364 			return (BUS_PROBE_DEFAULT);
4365 		}
4366 		ent++;
4367 	}
4368 	return (ENXIO);
4369 }
4370 
4371 static void
4372 iflib_reset_qvalues(if_ctx_t ctx)
4373 {
4374 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4375 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4376 	device_t dev = ctx->ifc_dev;
4377 	int i;
4378 
4379 	scctx->isc_txrx_budget_bytes_max = IFLIB_MAX_TX_BYTES;
4380 	scctx->isc_tx_qdepth = IFLIB_DEFAULT_TX_QDEPTH;
4381 	/*
4382 	 * XXX sanity check that ntxd & nrxd are a power of 2
4383 	 */
4384 	if (ctx->ifc_sysctl_ntxqs != 0)
4385 		scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs;
4386 	if (ctx->ifc_sysctl_nrxqs != 0)
4387 		scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs;
4388 
4389 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4390 		if (ctx->ifc_sysctl_ntxds[i] != 0)
4391 			scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i];
4392 		else
4393 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4394 	}
4395 
4396 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4397 		if (ctx->ifc_sysctl_nrxds[i] != 0)
4398 			scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i];
4399 		else
4400 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4401 	}
4402 
4403 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4404 		if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) {
4405 			device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n",
4406 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]);
4407 			scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i];
4408 		}
4409 		if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) {
4410 			device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n",
4411 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]);
4412 			scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i];
4413 		}
4414 	}
4415 
4416 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4417 		if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) {
4418 			device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n",
4419 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]);
4420 			scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i];
4421 		}
4422 		if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) {
4423 			device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n",
4424 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]);
4425 			scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i];
4426 		}
4427 	}
4428 }
4429 
4430 static void
4431 iflib_add_pfil(if_ctx_t ctx)
4432 {
4433 	struct pfil_head *pfil;
4434 	struct pfil_head_args pa;
4435 	iflib_rxq_t rxq;
4436 	int i;
4437 
4438 	pa.pa_version = PFIL_VERSION;
4439 	pa.pa_flags = PFIL_IN;
4440 	pa.pa_type = PFIL_TYPE_ETHERNET;
4441 	pa.pa_headname = ctx->ifc_ifp->if_xname;
4442 	pfil = pfil_head_register(&pa);
4443 
4444 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4445 		rxq->pfil = pfil;
4446 	}
4447 }
4448 
4449 static void
4450 iflib_rem_pfil(if_ctx_t ctx)
4451 {
4452 	struct pfil_head *pfil;
4453 	iflib_rxq_t rxq;
4454 	int i;
4455 
4456 	rxq = ctx->ifc_rxqs;
4457 	pfil = rxq->pfil;
4458 	for (i = 0; i < NRXQSETS(ctx); i++, rxq++) {
4459 		rxq->pfil = NULL;
4460 	}
4461 	pfil_head_unregister(pfil);
4462 }
4463 
4464 int
4465 iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp)
4466 {
4467 	int err, rid, msix;
4468 	if_ctx_t ctx;
4469 	if_t ifp;
4470 	if_softc_ctx_t scctx;
4471 	int i;
4472 	uint16_t main_txq;
4473 	uint16_t main_rxq;
4474 
4475 
4476 	ctx = malloc(sizeof(* ctx), M_IFLIB, M_WAITOK|M_ZERO);
4477 
4478 	if (sc == NULL) {
4479 		sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4480 		device_set_softc(dev, ctx);
4481 		ctx->ifc_flags |= IFC_SC_ALLOCATED;
4482 	}
4483 
4484 	ctx->ifc_sctx = sctx;
4485 	ctx->ifc_dev = dev;
4486 	ctx->ifc_softc = sc;
4487 
4488 	if ((err = iflib_register(ctx)) != 0) {
4489 		device_printf(dev, "iflib_register failed %d\n", err);
4490 		goto fail_ctx_free;
4491 	}
4492 	iflib_add_device_sysctl_pre(ctx);
4493 
4494 	scctx = &ctx->ifc_softc_ctx;
4495 	ifp = ctx->ifc_ifp;
4496 
4497 	iflib_reset_qvalues(ctx);
4498 	CTX_LOCK(ctx);
4499 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4500 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4501 		goto fail_unlock;
4502 	}
4503 	_iflib_pre_assert(scctx);
4504 	ctx->ifc_txrx = *scctx->isc_txrx;
4505 
4506 #ifdef INVARIANTS
4507 	MPASS(scctx->isc_capabilities);
4508 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
4509 		MPASS(scctx->isc_tx_csum_flags);
4510 #endif
4511 
4512 	if_setcapabilities(ifp, scctx->isc_capabilities | IFCAP_HWSTATS);
4513 	if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS);
4514 
4515 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4516 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4517 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4518 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4519 
4520 	main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4521 	main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4522 
4523 	/* XXX change for per-queue sizes */
4524 	device_printf(dev, "Using %d tx descriptors and %d rx descriptors\n",
4525 	    scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4526 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4527 		if (!powerof2(scctx->isc_nrxd[i])) {
4528 			/* round down instead? */
4529 			device_printf(dev, "# rx descriptors must be a power of 2\n");
4530 			err = EINVAL;
4531 			goto fail_iflib_detach;
4532 		}
4533 	}
4534 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4535 		if (!powerof2(scctx->isc_ntxd[i])) {
4536 			device_printf(dev,
4537 			    "# tx descriptors must be a power of 2");
4538 			err = EINVAL;
4539 			goto fail_iflib_detach;
4540 		}
4541 	}
4542 
4543 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4544 	    MAX_SINGLE_PACKET_FRACTION)
4545 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4546 		    MAX_SINGLE_PACKET_FRACTION);
4547 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4548 	    MAX_SINGLE_PACKET_FRACTION)
4549 		scctx->isc_tx_tso_segments_max = max(1,
4550 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4551 
4552 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4553 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
4554 		/*
4555 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
4556 		 * but some MACs do.
4557 		 */
4558 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
4559 		    IP_MAXPACKET));
4560 		/*
4561 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
4562 		 * into account.  In the worst case, each of these calls will
4563 		 * add another mbuf and, thus, the requirement for another DMA
4564 		 * segment.  So for best performance, it doesn't make sense to
4565 		 * advertize a maximum of TSO segments that typically will
4566 		 * require defragmentation in iflib_encap().
4567 		 */
4568 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
4569 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
4570 	}
4571 	if (scctx->isc_rss_table_size == 0)
4572 		scctx->isc_rss_table_size = 64;
4573 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4574 
4575 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4576 	/* XXX format name */
4577 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx,
4578 	    NULL, NULL, "admin");
4579 
4580 	/* Set up cpu set.  If it fails, use the set of all CPUs. */
4581 	if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) != 0) {
4582 		device_printf(dev, "Unable to fetch CPU list\n");
4583 		CPU_COPY(&all_cpus, &ctx->ifc_cpus);
4584 	}
4585 	MPASS(CPU_COUNT(&ctx->ifc_cpus) > 0);
4586 
4587 	/*
4588 	** Now set up MSI or MSI-X, should return us the number of supported
4589 	** vectors (will be 1 for a legacy interrupt and MSI).
4590 	*/
4591 	if (sctx->isc_flags & IFLIB_SKIP_MSIX) {
4592 		msix = scctx->isc_vectors;
4593 	} else if (scctx->isc_msix_bar != 0)
4594 	       /*
4595 		* The simple fact that isc_msix_bar is not 0 does not mean we
4596 		* we have a good value there that is known to work.
4597 		*/
4598 		msix = iflib_msix_init(ctx);
4599 	else {
4600 		scctx->isc_vectors = 1;
4601 		scctx->isc_ntxqsets = 1;
4602 		scctx->isc_nrxqsets = 1;
4603 		scctx->isc_intr = IFLIB_INTR_LEGACY;
4604 		msix = 0;
4605 	}
4606 	/* Get memory for the station queues */
4607 	if ((err = iflib_queues_alloc(ctx))) {
4608 		device_printf(dev, "Unable to allocate queue memory\n");
4609 		goto fail_intr_free;
4610 	}
4611 
4612 	if ((err = iflib_qset_structures_setup(ctx)))
4613 		goto fail_queues;
4614 
4615 	/*
4616 	 * Group taskqueues aren't properly set up until SMP is started,
4617 	 * so we disable interrupts until we can handle them post
4618 	 * SI_SUB_SMP.
4619 	 *
4620 	 * XXX: disabling interrupts doesn't actually work, at least for
4621 	 * the non-MSI case.  When they occur before SI_SUB_SMP completes,
4622 	 * we do null handling and depend on this not causing too large an
4623 	 * interrupt storm.
4624 	 */
4625 	IFDI_INTR_DISABLE(ctx);
4626 	if (msix > 1 && (err = IFDI_MSIX_INTR_ASSIGN(ctx, msix)) != 0) {
4627 		device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n", err);
4628 		goto fail_queues;
4629 	}
4630 	if (msix <= 1) {
4631 		rid = 0;
4632 		if (scctx->isc_intr == IFLIB_INTR_MSI) {
4633 			MPASS(msix == 1);
4634 			rid = 1;
4635 		}
4636 		if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) {
4637 			device_printf(dev, "iflib_legacy_setup failed %d\n", err);
4638 			goto fail_queues;
4639 		}
4640 	}
4641 
4642 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
4643 
4644 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4645 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4646 		goto fail_detach;
4647 	}
4648 
4649 	/*
4650 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4651 	 * This must appear after the call to ether_ifattach() because
4652 	 * ether_ifattach() sets if_hdrlen to the default value.
4653 	 */
4654 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4655 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
4656 
4657 	if ((err = iflib_netmap_attach(ctx))) {
4658 		device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err);
4659 		goto fail_detach;
4660 	}
4661 	*ctxp = ctx;
4662 
4663 	NETDUMP_SET(ctx->ifc_ifp, iflib);
4664 
4665 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4666 	iflib_add_device_sysctl_post(ctx);
4667 	iflib_add_pfil(ctx);
4668 	ctx->ifc_flags |= IFC_INIT_DONE;
4669 	CTX_UNLOCK(ctx);
4670 	return (0);
4671 
4672 fail_detach:
4673 	ether_ifdetach(ctx->ifc_ifp);
4674 fail_intr_free:
4675 	iflib_free_intr_mem(ctx);
4676 fail_queues:
4677 	iflib_tx_structures_free(ctx);
4678 	iflib_rx_structures_free(ctx);
4679 fail_iflib_detach:
4680 	IFDI_DETACH(ctx);
4681 fail_unlock:
4682 	CTX_UNLOCK(ctx);
4683 fail_ctx_free:
4684         if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4685                 free(ctx->ifc_softc, M_IFLIB);
4686         free(ctx, M_IFLIB);
4687 	return (err);
4688 }
4689 
4690 int
4691 iflib_pseudo_register(device_t dev, if_shared_ctx_t sctx, if_ctx_t *ctxp,
4692 					  struct iflib_cloneattach_ctx *clctx)
4693 {
4694 	int err;
4695 	if_ctx_t ctx;
4696 	if_t ifp;
4697 	if_softc_ctx_t scctx;
4698 	int i;
4699 	void *sc;
4700 	uint16_t main_txq;
4701 	uint16_t main_rxq;
4702 
4703 	ctx = malloc(sizeof(*ctx), M_IFLIB, M_WAITOK|M_ZERO);
4704 	sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4705 	ctx->ifc_flags |= IFC_SC_ALLOCATED;
4706 	if (sctx->isc_flags & (IFLIB_PSEUDO|IFLIB_VIRTUAL))
4707 		ctx->ifc_flags |= IFC_PSEUDO;
4708 
4709 	ctx->ifc_sctx = sctx;
4710 	ctx->ifc_softc = sc;
4711 	ctx->ifc_dev = dev;
4712 
4713 	if ((err = iflib_register(ctx)) != 0) {
4714 		device_printf(dev, "%s: iflib_register failed %d\n", __func__, err);
4715 		goto fail_ctx_free;
4716 	}
4717 	iflib_add_device_sysctl_pre(ctx);
4718 
4719 	scctx = &ctx->ifc_softc_ctx;
4720 	ifp = ctx->ifc_ifp;
4721 
4722 	/*
4723 	 * XXX sanity check that ntxd & nrxd are a power of 2
4724 	 */
4725 	iflib_reset_qvalues(ctx);
4726 	CTX_LOCK(ctx);
4727 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4728 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4729 		goto fail_unlock;
4730 	}
4731 	if (sctx->isc_flags & IFLIB_GEN_MAC)
4732 		ether_gen_addr(ifp, &ctx->ifc_mac);
4733 	if ((err = IFDI_CLONEATTACH(ctx, clctx->cc_ifc, clctx->cc_name,
4734 								clctx->cc_params)) != 0) {
4735 		device_printf(dev, "IFDI_CLONEATTACH failed %d\n", err);
4736 		goto fail_ctx_free;
4737 	}
4738 	ifmedia_add(&ctx->ifc_media, IFM_ETHER | IFM_1000_T | IFM_FDX, 0, NULL);
4739 	ifmedia_add(&ctx->ifc_media, IFM_ETHER | IFM_AUTO, 0, NULL);
4740 	ifmedia_set(&ctx->ifc_media, IFM_ETHER | IFM_AUTO);
4741 
4742 #ifdef INVARIANTS
4743 	MPASS(scctx->isc_capabilities);
4744 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
4745 		MPASS(scctx->isc_tx_csum_flags);
4746 #endif
4747 
4748 	if_setcapabilities(ifp, scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_LINKSTATE);
4749 	if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_LINKSTATE);
4750 
4751 	ifp->if_flags |= IFF_NOGROUP;
4752 	if (sctx->isc_flags & IFLIB_PSEUDO) {
4753 		ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
4754 
4755 		if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4756 			device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4757 			goto fail_detach;
4758 		}
4759 		*ctxp = ctx;
4760 
4761 		/*
4762 		 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4763 		 * This must appear after the call to ether_ifattach() because
4764 		 * ether_ifattach() sets if_hdrlen to the default value.
4765 		 */
4766 		if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4767 			if_setifheaderlen(ifp,
4768 			    sizeof(struct ether_vlan_header));
4769 
4770 		if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4771 		iflib_add_device_sysctl_post(ctx);
4772 		ctx->ifc_flags |= IFC_INIT_DONE;
4773 		return (0);
4774 	}
4775 	_iflib_pre_assert(scctx);
4776 	ctx->ifc_txrx = *scctx->isc_txrx;
4777 
4778 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4779 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4780 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4781 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4782 
4783 	main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4784 	main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4785 
4786 	/* XXX change for per-queue sizes */
4787 	device_printf(dev, "Using %d tx descriptors and %d rx descriptors\n",
4788 	    scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4789 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4790 		if (!powerof2(scctx->isc_nrxd[i])) {
4791 			/* round down instead? */
4792 			device_printf(dev, "# rx descriptors must be a power of 2\n");
4793 			err = EINVAL;
4794 			goto fail_iflib_detach;
4795 		}
4796 	}
4797 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4798 		if (!powerof2(scctx->isc_ntxd[i])) {
4799 			device_printf(dev,
4800 			    "# tx descriptors must be a power of 2");
4801 			err = EINVAL;
4802 			goto fail_iflib_detach;
4803 		}
4804 	}
4805 
4806 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4807 	    MAX_SINGLE_PACKET_FRACTION)
4808 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4809 		    MAX_SINGLE_PACKET_FRACTION);
4810 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4811 	    MAX_SINGLE_PACKET_FRACTION)
4812 		scctx->isc_tx_tso_segments_max = max(1,
4813 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4814 
4815 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4816 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
4817 		/*
4818 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
4819 		 * but some MACs do.
4820 		 */
4821 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
4822 		    IP_MAXPACKET));
4823 		/*
4824 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
4825 		 * into account.  In the worst case, each of these calls will
4826 		 * add another mbuf and, thus, the requirement for another DMA
4827 		 * segment.  So for best performance, it doesn't make sense to
4828 		 * advertize a maximum of TSO segments that typically will
4829 		 * require defragmentation in iflib_encap().
4830 		 */
4831 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
4832 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
4833 	}
4834 	if (scctx->isc_rss_table_size == 0)
4835 		scctx->isc_rss_table_size = 64;
4836 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4837 
4838 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4839 	/* XXX format name */
4840 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx,
4841 	    NULL, NULL, "admin");
4842 
4843 	/* XXX --- can support > 1 -- but keep it simple for now */
4844 	scctx->isc_intr = IFLIB_INTR_LEGACY;
4845 
4846 	/* Get memory for the station queues */
4847 	if ((err = iflib_queues_alloc(ctx))) {
4848 		device_printf(dev, "Unable to allocate queue memory\n");
4849 		goto fail_iflib_detach;
4850 	}
4851 
4852 	if ((err = iflib_qset_structures_setup(ctx))) {
4853 		device_printf(dev, "qset structure setup failed %d\n", err);
4854 		goto fail_queues;
4855 	}
4856 
4857 	/*
4858 	 * XXX What if anything do we want to do about interrupts?
4859 	 */
4860 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
4861 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4862 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4863 		goto fail_detach;
4864 	}
4865 
4866 	/*
4867 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4868 	 * This must appear after the call to ether_ifattach() because
4869 	 * ether_ifattach() sets if_hdrlen to the default value.
4870 	 */
4871 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4872 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
4873 
4874 	/* XXX handle more than one queue */
4875 	for (i = 0; i < scctx->isc_nrxqsets; i++)
4876 		IFDI_RX_CLSET(ctx, 0, i, ctx->ifc_rxqs[i].ifr_fl[0].ifl_sds.ifsd_cl);
4877 
4878 	*ctxp = ctx;
4879 
4880 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4881 	iflib_add_device_sysctl_post(ctx);
4882 	ctx->ifc_flags |= IFC_INIT_DONE;
4883 	CTX_UNLOCK(ctx);
4884 	return (0);
4885 fail_detach:
4886 	ether_ifdetach(ctx->ifc_ifp);
4887 fail_queues:
4888 	iflib_tx_structures_free(ctx);
4889 	iflib_rx_structures_free(ctx);
4890 fail_iflib_detach:
4891 	IFDI_DETACH(ctx);
4892 fail_unlock:
4893 	CTX_UNLOCK(ctx);
4894 fail_ctx_free:
4895 	free(ctx->ifc_softc, M_IFLIB);
4896 	free(ctx, M_IFLIB);
4897 	return (err);
4898 }
4899 
4900 int
4901 iflib_pseudo_deregister(if_ctx_t ctx)
4902 {
4903 	if_t ifp = ctx->ifc_ifp;
4904 	iflib_txq_t txq;
4905 	iflib_rxq_t rxq;
4906 	int i, j;
4907 	struct taskqgroup *tqg;
4908 	iflib_fl_t fl;
4909 
4910 	/* Unregister VLAN events */
4911 	if (ctx->ifc_vlan_attach_event != NULL)
4912 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
4913 	if (ctx->ifc_vlan_detach_event != NULL)
4914 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
4915 
4916 	ether_ifdetach(ifp);
4917 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
4918 	CTX_LOCK_DESTROY(ctx);
4919 	/* XXX drain any dependent tasks */
4920 	tqg = qgroup_if_io_tqg;
4921 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
4922 		callout_drain(&txq->ift_timer);
4923 		if (txq->ift_task.gt_uniq != NULL)
4924 			taskqgroup_detach(tqg, &txq->ift_task);
4925 	}
4926 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4927 		if (rxq->ifr_task.gt_uniq != NULL)
4928 			taskqgroup_detach(tqg, &rxq->ifr_task);
4929 
4930 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
4931 			free(fl->ifl_rx_bitmap, M_IFLIB);
4932 	}
4933 	tqg = qgroup_if_config_tqg;
4934 	if (ctx->ifc_admin_task.gt_uniq != NULL)
4935 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
4936 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
4937 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
4938 
4939 	if_free(ifp);
4940 
4941 	iflib_tx_structures_free(ctx);
4942 	iflib_rx_structures_free(ctx);
4943 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4944 		free(ctx->ifc_softc, M_IFLIB);
4945 	free(ctx, M_IFLIB);
4946 	return (0);
4947 }
4948 
4949 int
4950 iflib_device_attach(device_t dev)
4951 {
4952 	if_ctx_t ctx;
4953 	if_shared_ctx_t sctx;
4954 
4955 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4956 		return (ENOTSUP);
4957 
4958 	pci_enable_busmaster(dev);
4959 
4960 	return (iflib_device_register(dev, NULL, sctx, &ctx));
4961 }
4962 
4963 int
4964 iflib_device_deregister(if_ctx_t ctx)
4965 {
4966 	if_t ifp = ctx->ifc_ifp;
4967 	iflib_txq_t txq;
4968 	iflib_rxq_t rxq;
4969 	device_t dev = ctx->ifc_dev;
4970 	int i, j;
4971 	struct taskqgroup *tqg;
4972 	iflib_fl_t fl;
4973 
4974 	/* Make sure VLANS are not using driver */
4975 	if (if_vlantrunkinuse(ifp)) {
4976 		device_printf(dev, "Vlan in use, detach first\n");
4977 		return (EBUSY);
4978 	}
4979 #ifdef PCI_IOV
4980 	if (!CTX_IS_VF(ctx) && pci_iov_detach(dev) != 0) {
4981 		device_printf(dev, "SR-IOV in use; detach first.\n");
4982 		return (EBUSY);
4983 	}
4984 #endif
4985 
4986 	STATE_LOCK(ctx);
4987 	ctx->ifc_flags |= IFC_IN_DETACH;
4988 	STATE_UNLOCK(ctx);
4989 
4990 	CTX_LOCK(ctx);
4991 	iflib_stop(ctx);
4992 	CTX_UNLOCK(ctx);
4993 
4994 	/* Unregister VLAN events */
4995 	if (ctx->ifc_vlan_attach_event != NULL)
4996 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
4997 	if (ctx->ifc_vlan_detach_event != NULL)
4998 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
4999 
5000 	iflib_netmap_detach(ifp);
5001 	ether_ifdetach(ifp);
5002 	iflib_rem_pfil(ctx);
5003 	if (ctx->ifc_led_dev != NULL)
5004 		led_destroy(ctx->ifc_led_dev);
5005 	/* XXX drain any dependent tasks */
5006 	tqg = qgroup_if_io_tqg;
5007 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
5008 		callout_drain(&txq->ift_timer);
5009 		if (txq->ift_task.gt_uniq != NULL)
5010 			taskqgroup_detach(tqg, &txq->ift_task);
5011 	}
5012 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
5013 		if (rxq->ifr_task.gt_uniq != NULL)
5014 			taskqgroup_detach(tqg, &rxq->ifr_task);
5015 
5016 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5017 			free(fl->ifl_rx_bitmap, M_IFLIB);
5018 	}
5019 	tqg = qgroup_if_config_tqg;
5020 	if (ctx->ifc_admin_task.gt_uniq != NULL)
5021 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
5022 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
5023 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
5024 	CTX_LOCK(ctx);
5025 	IFDI_DETACH(ctx);
5026 	CTX_UNLOCK(ctx);
5027 
5028 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
5029 	CTX_LOCK_DESTROY(ctx);
5030 	device_set_softc(ctx->ifc_dev, NULL);
5031 	iflib_free_intr_mem(ctx);
5032 
5033 	bus_generic_detach(dev);
5034 	if_free(ifp);
5035 
5036 	iflib_tx_structures_free(ctx);
5037 	iflib_rx_structures_free(ctx);
5038 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5039 		free(ctx->ifc_softc, M_IFLIB);
5040 	STATE_LOCK_DESTROY(ctx);
5041 	free(ctx, M_IFLIB);
5042 	return (0);
5043 }
5044 
5045 static void
5046 iflib_free_intr_mem(if_ctx_t ctx)
5047 {
5048 
5049 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) {
5050 		iflib_irq_free(ctx, &ctx->ifc_legacy_irq);
5051 	}
5052 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) {
5053 		pci_release_msi(ctx->ifc_dev);
5054 	}
5055 	if (ctx->ifc_msix_mem != NULL) {
5056 		bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY,
5057 		    rman_get_rid(ctx->ifc_msix_mem), ctx->ifc_msix_mem);
5058 		ctx->ifc_msix_mem = NULL;
5059 	}
5060 }
5061 
5062 int
5063 iflib_device_detach(device_t dev)
5064 {
5065 	if_ctx_t ctx = device_get_softc(dev);
5066 
5067 	return (iflib_device_deregister(ctx));
5068 }
5069 
5070 int
5071 iflib_device_suspend(device_t dev)
5072 {
5073 	if_ctx_t ctx = device_get_softc(dev);
5074 
5075 	CTX_LOCK(ctx);
5076 	IFDI_SUSPEND(ctx);
5077 	CTX_UNLOCK(ctx);
5078 
5079 	return bus_generic_suspend(dev);
5080 }
5081 int
5082 iflib_device_shutdown(device_t dev)
5083 {
5084 	if_ctx_t ctx = device_get_softc(dev);
5085 
5086 	CTX_LOCK(ctx);
5087 	IFDI_SHUTDOWN(ctx);
5088 	CTX_UNLOCK(ctx);
5089 
5090 	return bus_generic_suspend(dev);
5091 }
5092 
5093 
5094 int
5095 iflib_device_resume(device_t dev)
5096 {
5097 	if_ctx_t ctx = device_get_softc(dev);
5098 	iflib_txq_t txq = ctx->ifc_txqs;
5099 
5100 	CTX_LOCK(ctx);
5101 	IFDI_RESUME(ctx);
5102 	iflib_if_init_locked(ctx);
5103 	CTX_UNLOCK(ctx);
5104 	for (int i = 0; i < NTXQSETS(ctx); i++, txq++)
5105 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
5106 
5107 	return (bus_generic_resume(dev));
5108 }
5109 
5110 int
5111 iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params)
5112 {
5113 	int error;
5114 	if_ctx_t ctx = device_get_softc(dev);
5115 
5116 	CTX_LOCK(ctx);
5117 	error = IFDI_IOV_INIT(ctx, num_vfs, params);
5118 	CTX_UNLOCK(ctx);
5119 
5120 	return (error);
5121 }
5122 
5123 void
5124 iflib_device_iov_uninit(device_t dev)
5125 {
5126 	if_ctx_t ctx = device_get_softc(dev);
5127 
5128 	CTX_LOCK(ctx);
5129 	IFDI_IOV_UNINIT(ctx);
5130 	CTX_UNLOCK(ctx);
5131 }
5132 
5133 int
5134 iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params)
5135 {
5136 	int error;
5137 	if_ctx_t ctx = device_get_softc(dev);
5138 
5139 	CTX_LOCK(ctx);
5140 	error = IFDI_IOV_VF_ADD(ctx, vfnum, params);
5141 	CTX_UNLOCK(ctx);
5142 
5143 	return (error);
5144 }
5145 
5146 /*********************************************************************
5147  *
5148  *  MODULE FUNCTION DEFINITIONS
5149  *
5150  **********************************************************************/
5151 
5152 /*
5153  * - Start a fast taskqueue thread for each core
5154  * - Start a taskqueue for control operations
5155  */
5156 static int
5157 iflib_module_init(void)
5158 {
5159 	return (0);
5160 }
5161 
5162 static int
5163 iflib_module_event_handler(module_t mod, int what, void *arg)
5164 {
5165 	int err;
5166 
5167 	switch (what) {
5168 	case MOD_LOAD:
5169 		if ((err = iflib_module_init()) != 0)
5170 			return (err);
5171 		break;
5172 	case MOD_UNLOAD:
5173 		return (EBUSY);
5174 	default:
5175 		return (EOPNOTSUPP);
5176 	}
5177 
5178 	return (0);
5179 }
5180 
5181 /*********************************************************************
5182  *
5183  *  PUBLIC FUNCTION DEFINITIONS
5184  *     ordered as in iflib.h
5185  *
5186  **********************************************************************/
5187 
5188 
5189 static void
5190 _iflib_assert(if_shared_ctx_t sctx)
5191 {
5192 	MPASS(sctx->isc_tx_maxsize);
5193 	MPASS(sctx->isc_tx_maxsegsize);
5194 
5195 	MPASS(sctx->isc_rx_maxsize);
5196 	MPASS(sctx->isc_rx_nsegments);
5197 	MPASS(sctx->isc_rx_maxsegsize);
5198 
5199 	MPASS(sctx->isc_nrxd_min[0]);
5200 	MPASS(sctx->isc_nrxd_max[0]);
5201 	MPASS(sctx->isc_nrxd_default[0]);
5202 	MPASS(sctx->isc_ntxd_min[0]);
5203 	MPASS(sctx->isc_ntxd_max[0]);
5204 	MPASS(sctx->isc_ntxd_default[0]);
5205 }
5206 
5207 static void
5208 _iflib_pre_assert(if_softc_ctx_t scctx)
5209 {
5210 
5211 	MPASS(scctx->isc_txrx->ift_txd_encap);
5212 	MPASS(scctx->isc_txrx->ift_txd_flush);
5213 	MPASS(scctx->isc_txrx->ift_txd_credits_update);
5214 	MPASS(scctx->isc_txrx->ift_rxd_available);
5215 	MPASS(scctx->isc_txrx->ift_rxd_pkt_get);
5216 	MPASS(scctx->isc_txrx->ift_rxd_refill);
5217 	MPASS(scctx->isc_txrx->ift_rxd_flush);
5218 }
5219 
5220 static int
5221 iflib_register(if_ctx_t ctx)
5222 {
5223 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5224 	driver_t *driver = sctx->isc_driver;
5225 	device_t dev = ctx->ifc_dev;
5226 	if_t ifp;
5227 
5228 	_iflib_assert(sctx);
5229 
5230 	CTX_LOCK_INIT(ctx);
5231 	STATE_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev));
5232 	ifp = ctx->ifc_ifp = if_alloc(IFT_ETHER);
5233 	if (ifp == NULL) {
5234 		device_printf(dev, "can not allocate ifnet structure\n");
5235 		return (ENOMEM);
5236 	}
5237 
5238 	/*
5239 	 * Initialize our context's device specific methods
5240 	 */
5241 	kobj_init((kobj_t) ctx, (kobj_class_t) driver);
5242 	kobj_class_compile((kobj_class_t) driver);
5243 	driver->refs++;
5244 
5245 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
5246 	if_setsoftc(ifp, ctx);
5247 	if_setdev(ifp, dev);
5248 	if_setinitfn(ifp, iflib_if_init);
5249 	if_setioctlfn(ifp, iflib_if_ioctl);
5250 #ifdef ALTQ
5251 	if_setstartfn(ifp, iflib_altq_if_start);
5252 	if_settransmitfn(ifp, iflib_altq_if_transmit);
5253 	if_setsendqready(ifp);
5254 #else
5255 	if_settransmitfn(ifp, iflib_if_transmit);
5256 #endif
5257 	if_setqflushfn(ifp, iflib_if_qflush);
5258 	if_setflags(ifp, IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST);
5259 
5260 	ctx->ifc_vlan_attach_event =
5261 		EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx,
5262 							  EVENTHANDLER_PRI_FIRST);
5263 	ctx->ifc_vlan_detach_event =
5264 		EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx,
5265 							  EVENTHANDLER_PRI_FIRST);
5266 
5267 	ifmedia_init(&ctx->ifc_media, IFM_IMASK,
5268 					 iflib_media_change, iflib_media_status);
5269 
5270 	return (0);
5271 }
5272 
5273 
5274 static int
5275 iflib_queues_alloc(if_ctx_t ctx)
5276 {
5277 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5278 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5279 	device_t dev = ctx->ifc_dev;
5280 	int nrxqsets = scctx->isc_nrxqsets;
5281 	int ntxqsets = scctx->isc_ntxqsets;
5282 	iflib_txq_t txq;
5283 	iflib_rxq_t rxq;
5284 	iflib_fl_t fl = NULL;
5285 	int i, j, cpu, err, txconf, rxconf;
5286 	iflib_dma_info_t ifdip;
5287 	uint32_t *rxqsizes = scctx->isc_rxqsizes;
5288 	uint32_t *txqsizes = scctx->isc_txqsizes;
5289 	uint8_t nrxqs = sctx->isc_nrxqs;
5290 	uint8_t ntxqs = sctx->isc_ntxqs;
5291 	int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1;
5292 	caddr_t *vaddrs;
5293 	uint64_t *paddrs;
5294 
5295 	KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1"));
5296 	KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1"));
5297 
5298 	/* Allocate the TX ring struct memory */
5299 	if (!(ctx->ifc_txqs =
5300 	    (iflib_txq_t) malloc(sizeof(struct iflib_txq) *
5301 	    ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5302 		device_printf(dev, "Unable to allocate TX ring memory\n");
5303 		err = ENOMEM;
5304 		goto fail;
5305 	}
5306 
5307 	/* Now allocate the RX */
5308 	if (!(ctx->ifc_rxqs =
5309 	    (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) *
5310 	    nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5311 		device_printf(dev, "Unable to allocate RX ring memory\n");
5312 		err = ENOMEM;
5313 		goto rx_fail;
5314 	}
5315 
5316 	txq = ctx->ifc_txqs;
5317 	rxq = ctx->ifc_rxqs;
5318 
5319 	/*
5320 	 * XXX handle allocation failure
5321 	 */
5322 	for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) {
5323 		/* Set up some basics */
5324 
5325 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs,
5326 		    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
5327 			device_printf(dev,
5328 			    "Unable to allocate TX DMA info memory\n");
5329 			err = ENOMEM;
5330 			goto err_tx_desc;
5331 		}
5332 		txq->ift_ifdi = ifdip;
5333 		for (j = 0; j < ntxqs; j++, ifdip++) {
5334 			if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, 0)) {
5335 				device_printf(dev,
5336 				    "Unable to allocate TX descriptors\n");
5337 				err = ENOMEM;
5338 				goto err_tx_desc;
5339 			}
5340 			txq->ift_txd_size[j] = scctx->isc_txd_size[j];
5341 			bzero((void *)ifdip->idi_vaddr, txqsizes[j]);
5342 		}
5343 		txq->ift_ctx = ctx;
5344 		txq->ift_id = i;
5345 		if (sctx->isc_flags & IFLIB_HAS_TXCQ) {
5346 			txq->ift_br_offset = 1;
5347 		} else {
5348 			txq->ift_br_offset = 0;
5349 		}
5350 		/* XXX fix this */
5351 		txq->ift_timer.c_cpu = cpu;
5352 
5353 		if (iflib_txsd_alloc(txq)) {
5354 			device_printf(dev, "Critical Failure setting up TX buffers\n");
5355 			err = ENOMEM;
5356 			goto err_tx_desc;
5357 		}
5358 
5359 		/* Initialize the TX lock */
5360 		snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:tx(%d):callout",
5361 		    device_get_nameunit(dev), txq->ift_id);
5362 		mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF);
5363 		callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0);
5364 
5365 		snprintf(txq->ift_db_mtx_name, MTX_NAME_LEN, "%s:tx(%d):db",
5366 			 device_get_nameunit(dev), txq->ift_id);
5367 
5368 		err = ifmp_ring_alloc(&txq->ift_br, 2048, txq, iflib_txq_drain,
5369 				      iflib_txq_can_drain, M_IFLIB, M_WAITOK);
5370 		if (err) {
5371 			/* XXX free any allocated rings */
5372 			device_printf(dev, "Unable to allocate buf_ring\n");
5373 			goto err_tx_desc;
5374 		}
5375 	}
5376 
5377 	for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) {
5378 		/* Set up some basics */
5379 
5380 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs,
5381 		   M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
5382 			device_printf(dev,
5383 			    "Unable to allocate RX DMA info memory\n");
5384 			err = ENOMEM;
5385 			goto err_tx_desc;
5386 		}
5387 
5388 		rxq->ifr_ifdi = ifdip;
5389 		/* XXX this needs to be changed if #rx queues != #tx queues */
5390 		rxq->ifr_ntxqirq = 1;
5391 		rxq->ifr_txqid[0] = i;
5392 		for (j = 0; j < nrxqs; j++, ifdip++) {
5393 			if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, 0)) {
5394 				device_printf(dev,
5395 				    "Unable to allocate RX descriptors\n");
5396 				err = ENOMEM;
5397 				goto err_tx_desc;
5398 			}
5399 			bzero((void *)ifdip->idi_vaddr, rxqsizes[j]);
5400 		}
5401 		rxq->ifr_ctx = ctx;
5402 		rxq->ifr_id = i;
5403 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
5404 			rxq->ifr_fl_offset = 1;
5405 		} else {
5406 			rxq->ifr_fl_offset = 0;
5407 		}
5408 		rxq->ifr_nfl = nfree_lists;
5409 		if (!(fl =
5410 			  (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) {
5411 			device_printf(dev, "Unable to allocate free list memory\n");
5412 			err = ENOMEM;
5413 			goto err_tx_desc;
5414 		}
5415 		rxq->ifr_fl = fl;
5416 		for (j = 0; j < nfree_lists; j++) {
5417 			fl[j].ifl_rxq = rxq;
5418 			fl[j].ifl_id = j;
5419 			fl[j].ifl_ifdi = &rxq->ifr_ifdi[j + rxq->ifr_fl_offset];
5420 			fl[j].ifl_rxd_size = scctx->isc_rxd_size[j];
5421 		}
5422 		/* Allocate receive buffers for the ring */
5423 		if (iflib_rxsd_alloc(rxq)) {
5424 			device_printf(dev,
5425 			    "Critical Failure setting up receive buffers\n");
5426 			err = ENOMEM;
5427 			goto err_rx_desc;
5428 		}
5429 
5430 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5431 			fl->ifl_rx_bitmap = bit_alloc(fl->ifl_size, M_IFLIB,
5432 			    M_WAITOK);
5433 	}
5434 
5435 	/* TXQs */
5436 	vaddrs = malloc(sizeof(caddr_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5437 	paddrs = malloc(sizeof(uint64_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5438 	for (i = 0; i < ntxqsets; i++) {
5439 		iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi;
5440 
5441 		for (j = 0; j < ntxqs; j++, di++) {
5442 			vaddrs[i*ntxqs + j] = di->idi_vaddr;
5443 			paddrs[i*ntxqs + j] = di->idi_paddr;
5444 		}
5445 	}
5446 	if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) {
5447 		device_printf(ctx->ifc_dev,
5448 		    "Unable to allocate device TX queue\n");
5449 		iflib_tx_structures_free(ctx);
5450 		free(vaddrs, M_IFLIB);
5451 		free(paddrs, M_IFLIB);
5452 		goto err_rx_desc;
5453 	}
5454 	free(vaddrs, M_IFLIB);
5455 	free(paddrs, M_IFLIB);
5456 
5457 	/* RXQs */
5458 	vaddrs = malloc(sizeof(caddr_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5459 	paddrs = malloc(sizeof(uint64_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5460 	for (i = 0; i < nrxqsets; i++) {
5461 		iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi;
5462 
5463 		for (j = 0; j < nrxqs; j++, di++) {
5464 			vaddrs[i*nrxqs + j] = di->idi_vaddr;
5465 			paddrs[i*nrxqs + j] = di->idi_paddr;
5466 		}
5467 	}
5468 	if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) {
5469 		device_printf(ctx->ifc_dev,
5470 		    "Unable to allocate device RX queue\n");
5471 		iflib_tx_structures_free(ctx);
5472 		free(vaddrs, M_IFLIB);
5473 		free(paddrs, M_IFLIB);
5474 		goto err_rx_desc;
5475 	}
5476 	free(vaddrs, M_IFLIB);
5477 	free(paddrs, M_IFLIB);
5478 
5479 	return (0);
5480 
5481 /* XXX handle allocation failure changes */
5482 err_rx_desc:
5483 err_tx_desc:
5484 rx_fail:
5485 	if (ctx->ifc_rxqs != NULL)
5486 		free(ctx->ifc_rxqs, M_IFLIB);
5487 	ctx->ifc_rxqs = NULL;
5488 	if (ctx->ifc_txqs != NULL)
5489 		free(ctx->ifc_txqs, M_IFLIB);
5490 	ctx->ifc_txqs = NULL;
5491 fail:
5492 	return (err);
5493 }
5494 
5495 static int
5496 iflib_tx_structures_setup(if_ctx_t ctx)
5497 {
5498 	iflib_txq_t txq = ctx->ifc_txqs;
5499 	int i;
5500 
5501 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
5502 		iflib_txq_setup(txq);
5503 
5504 	return (0);
5505 }
5506 
5507 static void
5508 iflib_tx_structures_free(if_ctx_t ctx)
5509 {
5510 	iflib_txq_t txq = ctx->ifc_txqs;
5511 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5512 	int i, j;
5513 
5514 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
5515 		iflib_txq_destroy(txq);
5516 		for (j = 0; j < sctx->isc_ntxqs; j++)
5517 			iflib_dma_free(&txq->ift_ifdi[j]);
5518 	}
5519 	free(ctx->ifc_txqs, M_IFLIB);
5520 	ctx->ifc_txqs = NULL;
5521 	IFDI_QUEUES_FREE(ctx);
5522 }
5523 
5524 /*********************************************************************
5525  *
5526  *  Initialize all receive rings.
5527  *
5528  **********************************************************************/
5529 static int
5530 iflib_rx_structures_setup(if_ctx_t ctx)
5531 {
5532 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5533 	int q;
5534 #if defined(INET6) || defined(INET)
5535 	int i, err;
5536 #endif
5537 
5538 	for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) {
5539 #if defined(INET6) || defined(INET)
5540 		tcp_lro_free(&rxq->ifr_lc);
5541 		if ((err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp,
5542 		    TCP_LRO_ENTRIES, min(1024,
5543 		    ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset]))) != 0) {
5544 			device_printf(ctx->ifc_dev, "LRO Initialization failed!\n");
5545 			goto fail;
5546 		}
5547 		rxq->ifr_lro_enabled = TRUE;
5548 #endif
5549 		IFDI_RXQ_SETUP(ctx, rxq->ifr_id);
5550 	}
5551 	return (0);
5552 #if defined(INET6) || defined(INET)
5553 fail:
5554 	/*
5555 	 * Free RX software descriptors allocated so far, we will only handle
5556 	 * the rings that completed, the failing case will have
5557 	 * cleaned up for itself. 'q' failed, so its the terminus.
5558 	 */
5559 	rxq = ctx->ifc_rxqs;
5560 	for (i = 0; i < q; ++i, rxq++) {
5561 		iflib_rx_sds_free(rxq);
5562 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
5563 	}
5564 	return (err);
5565 #endif
5566 }
5567 
5568 /*********************************************************************
5569  *
5570  *  Free all receive rings.
5571  *
5572  **********************************************************************/
5573 static void
5574 iflib_rx_structures_free(if_ctx_t ctx)
5575 {
5576 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5577 
5578 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) {
5579 		iflib_rx_sds_free(rxq);
5580 	}
5581 	free(ctx->ifc_rxqs, M_IFLIB);
5582 	ctx->ifc_rxqs = NULL;
5583 }
5584 
5585 static int
5586 iflib_qset_structures_setup(if_ctx_t ctx)
5587 {
5588 	int err;
5589 
5590 	/*
5591 	 * It is expected that the caller takes care of freeing queues if this
5592 	 * fails.
5593 	 */
5594 	if ((err = iflib_tx_structures_setup(ctx)) != 0) {
5595 		device_printf(ctx->ifc_dev, "iflib_tx_structures_setup failed: %d\n", err);
5596 		return (err);
5597 	}
5598 
5599 	if ((err = iflib_rx_structures_setup(ctx)) != 0)
5600 		device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err);
5601 
5602 	return (err);
5603 }
5604 
5605 int
5606 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
5607 		driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, const char *name)
5608 {
5609 
5610 	return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name));
5611 }
5612 
5613 #ifdef SMP
5614 static int
5615 find_nth(if_ctx_t ctx, int qid)
5616 {
5617 	cpuset_t cpus;
5618 	int i, cpuid, eqid, count;
5619 
5620 	CPU_COPY(&ctx->ifc_cpus, &cpus);
5621 	count = CPU_COUNT(&cpus);
5622 	eqid = qid % count;
5623 	/* clear up to the qid'th bit */
5624 	for (i = 0; i < eqid; i++) {
5625 		cpuid = CPU_FFS(&cpus);
5626 		MPASS(cpuid != 0);
5627 		CPU_CLR(cpuid-1, &cpus);
5628 	}
5629 	cpuid = CPU_FFS(&cpus);
5630 	MPASS(cpuid != 0);
5631 	return (cpuid-1);
5632 }
5633 
5634 #ifdef SCHED_ULE
5635 extern struct cpu_group *cpu_top;              /* CPU topology */
5636 
5637 static int
5638 find_child_with_core(int cpu, struct cpu_group *grp)
5639 {
5640 	int i;
5641 
5642 	if (grp->cg_children == 0)
5643 		return -1;
5644 
5645 	MPASS(grp->cg_child);
5646 	for (i = 0; i < grp->cg_children; i++) {
5647 		if (CPU_ISSET(cpu, &grp->cg_child[i].cg_mask))
5648 			return i;
5649 	}
5650 
5651 	return -1;
5652 }
5653 
5654 /*
5655  * Find the nth "close" core to the specified core
5656  * "close" is defined as the deepest level that shares
5657  * at least an L2 cache.  With threads, this will be
5658  * threads on the same core.  If the sahred cache is L3
5659  * or higher, simply returns the same core.
5660  */
5661 static int
5662 find_close_core(int cpu, int core_offset)
5663 {
5664 	struct cpu_group *grp;
5665 	int i;
5666 	int fcpu;
5667 	cpuset_t cs;
5668 
5669 	grp = cpu_top;
5670 	if (grp == NULL)
5671 		return cpu;
5672 	i = 0;
5673 	while ((i = find_child_with_core(cpu, grp)) != -1) {
5674 		/* If the child only has one cpu, don't descend */
5675 		if (grp->cg_child[i].cg_count <= 1)
5676 			break;
5677 		grp = &grp->cg_child[i];
5678 	}
5679 
5680 	/* If they don't share at least an L2 cache, use the same CPU */
5681 	if (grp->cg_level > CG_SHARE_L2 || grp->cg_level == CG_SHARE_NONE)
5682 		return cpu;
5683 
5684 	/* Now pick one */
5685 	CPU_COPY(&grp->cg_mask, &cs);
5686 
5687 	/* Add the selected CPU offset to core offset. */
5688 	for (i = 0; (fcpu = CPU_FFS(&cs)) != 0; i++) {
5689 		if (fcpu - 1 == cpu)
5690 			break;
5691 		CPU_CLR(fcpu - 1, &cs);
5692 	}
5693 	MPASS(fcpu);
5694 
5695 	core_offset += i;
5696 
5697 	CPU_COPY(&grp->cg_mask, &cs);
5698 	for (i = core_offset % grp->cg_count; i > 0; i--) {
5699 		MPASS(CPU_FFS(&cs));
5700 		CPU_CLR(CPU_FFS(&cs) - 1, &cs);
5701 	}
5702 	MPASS(CPU_FFS(&cs));
5703 	return CPU_FFS(&cs) - 1;
5704 }
5705 #else
5706 static int
5707 find_close_core(int cpu, int core_offset __unused)
5708 {
5709 	return cpu;
5710 }
5711 #endif
5712 
5713 static int
5714 get_core_offset(if_ctx_t ctx, iflib_intr_type_t type, int qid)
5715 {
5716 	switch (type) {
5717 	case IFLIB_INTR_TX:
5718 		/* TX queues get cores which share at least an L2 cache with the corresponding RX queue */
5719 		/* XXX handle multiple RX threads per core and more than two core per L2 group */
5720 		return qid / CPU_COUNT(&ctx->ifc_cpus) + 1;
5721 	case IFLIB_INTR_RX:
5722 	case IFLIB_INTR_RXTX:
5723 		/* RX queues get the specified core */
5724 		return qid / CPU_COUNT(&ctx->ifc_cpus);
5725 	default:
5726 		return -1;
5727 	}
5728 }
5729 #else
5730 #define get_core_offset(ctx, type, qid)	CPU_FIRST()
5731 #define find_close_core(cpuid, tid)	CPU_FIRST()
5732 #define find_nth(ctx, gid)		CPU_FIRST()
5733 #endif
5734 
5735 /* Just to avoid copy/paste */
5736 static inline int
5737 iflib_irq_set_affinity(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type,
5738     int qid, struct grouptask *gtask, struct taskqgroup *tqg, void *uniq,
5739     const char *name)
5740 {
5741 	device_t dev;
5742 	int err, cpuid, tid;
5743 
5744 	dev = ctx->ifc_dev;
5745 	cpuid = find_nth(ctx, qid);
5746 	tid = get_core_offset(ctx, type, qid);
5747 	MPASS(tid >= 0);
5748 	cpuid = find_close_core(cpuid, tid);
5749 	err = taskqgroup_attach_cpu(tqg, gtask, uniq, cpuid, dev, irq->ii_res,
5750 	    name);
5751 	if (err) {
5752 		device_printf(dev, "taskqgroup_attach_cpu failed %d\n", err);
5753 		return (err);
5754 	}
5755 #ifdef notyet
5756 	if (cpuid > ctx->ifc_cpuid_highest)
5757 		ctx->ifc_cpuid_highest = cpuid;
5758 #endif
5759 	return 0;
5760 }
5761 
5762 int
5763 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid,
5764 			iflib_intr_type_t type, driver_filter_t *filter,
5765 			void *filter_arg, int qid, const char *name)
5766 {
5767 	device_t dev;
5768 	struct grouptask *gtask;
5769 	struct taskqgroup *tqg;
5770 	iflib_filter_info_t info;
5771 	gtask_fn_t *fn;
5772 	int tqrid, err;
5773 	driver_filter_t *intr_fast;
5774 	void *q;
5775 
5776 	info = &ctx->ifc_filter_info;
5777 	tqrid = rid;
5778 
5779 	switch (type) {
5780 	/* XXX merge tx/rx for netmap? */
5781 	case IFLIB_INTR_TX:
5782 		q = &ctx->ifc_txqs[qid];
5783 		info = &ctx->ifc_txqs[qid].ift_filter_info;
5784 		gtask = &ctx->ifc_txqs[qid].ift_task;
5785 		tqg = qgroup_if_io_tqg;
5786 		fn = _task_fn_tx;
5787 		intr_fast = iflib_fast_intr;
5788 		GROUPTASK_INIT(gtask, 0, fn, q);
5789 		ctx->ifc_flags |= IFC_NETMAP_TX_IRQ;
5790 		break;
5791 	case IFLIB_INTR_RX:
5792 		q = &ctx->ifc_rxqs[qid];
5793 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
5794 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5795 		tqg = qgroup_if_io_tqg;
5796 		fn = _task_fn_rx;
5797 		intr_fast = iflib_fast_intr;
5798 		GROUPTASK_INIT(gtask, 0, fn, q);
5799 		break;
5800 	case IFLIB_INTR_RXTX:
5801 		q = &ctx->ifc_rxqs[qid];
5802 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
5803 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5804 		tqg = qgroup_if_io_tqg;
5805 		fn = _task_fn_rx;
5806 		intr_fast = iflib_fast_intr_rxtx;
5807 		GROUPTASK_INIT(gtask, 0, fn, q);
5808 		break;
5809 	case IFLIB_INTR_ADMIN:
5810 		q = ctx;
5811 		tqrid = -1;
5812 		info = &ctx->ifc_filter_info;
5813 		gtask = &ctx->ifc_admin_task;
5814 		tqg = qgroup_if_config_tqg;
5815 		fn = _task_fn_admin;
5816 		intr_fast = iflib_fast_intr_ctx;
5817 		break;
5818 	default:
5819 		panic("unknown net intr type");
5820 	}
5821 
5822 	info->ifi_filter = filter;
5823 	info->ifi_filter_arg = filter_arg;
5824 	info->ifi_task = gtask;
5825 	info->ifi_ctx = q;
5826 
5827 	dev = ctx->ifc_dev;
5828 	err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info,  name);
5829 	if (err != 0) {
5830 		device_printf(dev, "_iflib_irq_alloc failed %d\n", err);
5831 		return (err);
5832 	}
5833 	if (type == IFLIB_INTR_ADMIN)
5834 		return (0);
5835 
5836 	if (tqrid != -1) {
5837 		err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg,
5838 		    q, name);
5839 		if (err)
5840 			return (err);
5841 	} else {
5842 		taskqgroup_attach(tqg, gtask, q, dev, irq->ii_res, name);
5843 	}
5844 
5845 	return (0);
5846 }
5847 
5848 void
5849 iflib_softirq_alloc_generic(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type, void *arg, int qid, const char *name)
5850 {
5851 	struct grouptask *gtask;
5852 	struct taskqgroup *tqg;
5853 	gtask_fn_t *fn;
5854 	void *q;
5855 	int err;
5856 
5857 	switch (type) {
5858 	case IFLIB_INTR_TX:
5859 		q = &ctx->ifc_txqs[qid];
5860 		gtask = &ctx->ifc_txqs[qid].ift_task;
5861 		tqg = qgroup_if_io_tqg;
5862 		fn = _task_fn_tx;
5863 		break;
5864 	case IFLIB_INTR_RX:
5865 		q = &ctx->ifc_rxqs[qid];
5866 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5867 		tqg = qgroup_if_io_tqg;
5868 		fn = _task_fn_rx;
5869 		break;
5870 	case IFLIB_INTR_IOV:
5871 		q = ctx;
5872 		gtask = &ctx->ifc_vflr_task;
5873 		tqg = qgroup_if_config_tqg;
5874 		fn = _task_fn_iov;
5875 		break;
5876 	default:
5877 		panic("unknown net intr type");
5878 	}
5879 	GROUPTASK_INIT(gtask, 0, fn, q);
5880 	if (irq != NULL) {
5881 		err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg,
5882 		    q, name);
5883 		if (err)
5884 			taskqgroup_attach(tqg, gtask, q, ctx->ifc_dev,
5885 			    irq->ii_res, name);
5886 	} else {
5887 		taskqgroup_attach(tqg, gtask, q, NULL, NULL, name);
5888 	}
5889 }
5890 
5891 void
5892 iflib_irq_free(if_ctx_t ctx, if_irq_t irq)
5893 {
5894 
5895 	if (irq->ii_tag)
5896 		bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag);
5897 
5898 	if (irq->ii_res)
5899 		bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ,
5900 		    rman_get_rid(irq->ii_res), irq->ii_res);
5901 }
5902 
5903 static int
5904 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, const char *name)
5905 {
5906 	iflib_txq_t txq = ctx->ifc_txqs;
5907 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5908 	if_irq_t irq = &ctx->ifc_legacy_irq;
5909 	iflib_filter_info_t info;
5910 	device_t dev;
5911 	struct grouptask *gtask;
5912 	struct resource *res;
5913 	struct taskqgroup *tqg;
5914 	gtask_fn_t *fn;
5915 	int tqrid;
5916 	void *q;
5917 	int err;
5918 
5919 	q = &ctx->ifc_rxqs[0];
5920 	info = &rxq[0].ifr_filter_info;
5921 	gtask = &rxq[0].ifr_task;
5922 	tqg = qgroup_if_io_tqg;
5923 	tqrid = irq->ii_rid = *rid;
5924 	fn = _task_fn_rx;
5925 
5926 	ctx->ifc_flags |= IFC_LEGACY;
5927 	info->ifi_filter = filter;
5928 	info->ifi_filter_arg = filter_arg;
5929 	info->ifi_task = gtask;
5930 	info->ifi_ctx = ctx;
5931 
5932 	dev = ctx->ifc_dev;
5933 	/* We allocate a single interrupt resource */
5934 	if ((err = _iflib_irq_alloc(ctx, irq, tqrid, iflib_fast_intr_ctx, NULL, info, name)) != 0)
5935 		return (err);
5936 	GROUPTASK_INIT(gtask, 0, fn, q);
5937 	res = irq->ii_res;
5938 	taskqgroup_attach(tqg, gtask, q, dev, res, name);
5939 
5940 	GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq);
5941 	taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, dev, res,
5942 	    "tx");
5943 	return (0);
5944 }
5945 
5946 void
5947 iflib_led_create(if_ctx_t ctx)
5948 {
5949 
5950 	ctx->ifc_led_dev = led_create(iflib_led_func, ctx,
5951 	    device_get_nameunit(ctx->ifc_dev));
5952 }
5953 
5954 void
5955 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid)
5956 {
5957 
5958 	GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
5959 }
5960 
5961 void
5962 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid)
5963 {
5964 
5965 	GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task);
5966 }
5967 
5968 void
5969 iflib_admin_intr_deferred(if_ctx_t ctx)
5970 {
5971 #ifdef INVARIANTS
5972 	struct grouptask *gtask;
5973 
5974 	gtask = &ctx->ifc_admin_task;
5975 	MPASS(gtask != NULL && gtask->gt_taskqueue != NULL);
5976 #endif
5977 
5978 	GROUPTASK_ENQUEUE(&ctx->ifc_admin_task);
5979 }
5980 
5981 void
5982 iflib_iov_intr_deferred(if_ctx_t ctx)
5983 {
5984 
5985 	GROUPTASK_ENQUEUE(&ctx->ifc_vflr_task);
5986 }
5987 
5988 void
5989 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, char *name)
5990 {
5991 
5992 	taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, NULL, NULL,
5993 	    name);
5994 }
5995 
5996 void
5997 iflib_config_gtask_init(void *ctx, struct grouptask *gtask, gtask_fn_t *fn,
5998 	const char *name)
5999 {
6000 
6001 	GROUPTASK_INIT(gtask, 0, fn, ctx);
6002 	taskqgroup_attach(qgroup_if_config_tqg, gtask, gtask, NULL, NULL,
6003 	    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 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
6046 	    BUS_DMASYNC_POSTREAD);
6047 	if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, true)) == 0)
6048 		return (0);
6049 
6050 	txq->ift_processed += credits;
6051 	txq->ift_cidx_processed += credits;
6052 
6053 	MPASS(credits_pre + credits == txq->ift_cidx_processed);
6054 	if (txq->ift_cidx_processed >= txq->ift_size)
6055 		txq->ift_cidx_processed -= txq->ift_size;
6056 	return (credits);
6057 }
6058 
6059 static int
6060 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget)
6061 {
6062 	iflib_fl_t fl;
6063 	u_int i;
6064 
6065 	for (i = 0, fl = &rxq->ifr_fl[0]; i < rxq->ifr_nfl; i++, fl++)
6066 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
6067 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
6068 	return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx,
6069 	    budget));
6070 }
6071 
6072 void
6073 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name,
6074 	const char *description, if_int_delay_info_t info,
6075 	int offset, int value)
6076 {
6077 	info->iidi_ctx = ctx;
6078 	info->iidi_offset = offset;
6079 	info->iidi_value = value;
6080 	SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev),
6081 	    SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)),
6082 	    OID_AUTO, name, CTLTYPE_INT|CTLFLAG_RW,
6083 	    info, 0, iflib_sysctl_int_delay, "I", description);
6084 }
6085 
6086 struct sx *
6087 iflib_ctx_lock_get(if_ctx_t ctx)
6088 {
6089 
6090 	return (&ctx->ifc_ctx_sx);
6091 }
6092 
6093 static int
6094 iflib_msix_init(if_ctx_t ctx)
6095 {
6096 	device_t dev = ctx->ifc_dev;
6097 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6098 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6099 	int vectors, queues, rx_queues, tx_queues, queuemsgs, msgs;
6100 	int iflib_num_tx_queues, iflib_num_rx_queues;
6101 	int err, admincnt, bar;
6102 
6103 	iflib_num_tx_queues = ctx->ifc_sysctl_ntxqs;
6104 	iflib_num_rx_queues = ctx->ifc_sysctl_nrxqs;
6105 
6106 	if (bootverbose)
6107 		device_printf(dev, "msix_init qsets capped at %d\n",
6108 		    imax(scctx->isc_ntxqsets, scctx->isc_nrxqsets));
6109 
6110 	bar = ctx->ifc_softc_ctx.isc_msix_bar;
6111 	admincnt = sctx->isc_admin_intrcnt;
6112 	/* Override by tuneable */
6113 	if (scctx->isc_disable_msix)
6114 		goto msi;
6115 
6116 	/* First try MSI-X */
6117 	if ((msgs = pci_msix_count(dev)) == 0) {
6118 		if (bootverbose)
6119 			device_printf(dev, "MSI-X not supported or disabled\n");
6120 		goto msi;
6121 	}
6122 	/*
6123 	 * bar == -1 => "trust me I know what I'm doing"
6124 	 * Some drivers are for hardware that is so shoddily
6125 	 * documented that no one knows which bars are which
6126 	 * so the developer has to map all bars. This hack
6127 	 * allows shoddy garbage to use MSI-X in this framework.
6128 	 */
6129 	if (bar != -1) {
6130 		ctx->ifc_msix_mem = bus_alloc_resource_any(dev,
6131 	            SYS_RES_MEMORY, &bar, RF_ACTIVE);
6132 		if (ctx->ifc_msix_mem == NULL) {
6133 			device_printf(dev, "Unable to map MSI-X table\n");
6134 			goto msi;
6135 		}
6136 	}
6137 #if IFLIB_DEBUG
6138 	/* use only 1 qset in debug mode */
6139 	queuemsgs = min(msgs - admincnt, 1);
6140 #else
6141 	queuemsgs = msgs - admincnt;
6142 #endif
6143 #ifdef RSS
6144 	queues = imin(queuemsgs, rss_getnumbuckets());
6145 #else
6146 	queues = queuemsgs;
6147 #endif
6148 	queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues);
6149 	if (bootverbose)
6150 		device_printf(dev,
6151 		    "intr CPUs: %d queue msgs: %d admincnt: %d\n",
6152 		    CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt);
6153 #ifdef  RSS
6154 	/* If we're doing RSS, clamp at the number of RSS buckets */
6155 	if (queues > rss_getnumbuckets())
6156 		queues = rss_getnumbuckets();
6157 #endif
6158 	if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt)
6159 		rx_queues = iflib_num_rx_queues;
6160 	else
6161 		rx_queues = queues;
6162 
6163 	if (rx_queues > scctx->isc_nrxqsets)
6164 		rx_queues = scctx->isc_nrxqsets;
6165 
6166 	/*
6167 	 * We want this to be all logical CPUs by default
6168 	 */
6169 	if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues)
6170 		tx_queues = iflib_num_tx_queues;
6171 	else
6172 		tx_queues = mp_ncpus;
6173 
6174 	if (tx_queues > scctx->isc_ntxqsets)
6175 		tx_queues = scctx->isc_ntxqsets;
6176 
6177 	if (ctx->ifc_sysctl_qs_eq_override == 0) {
6178 #ifdef INVARIANTS
6179 		if (tx_queues != rx_queues)
6180 			device_printf(dev,
6181 			    "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n",
6182 			    min(rx_queues, tx_queues), min(rx_queues, tx_queues));
6183 #endif
6184 		tx_queues = min(rx_queues, tx_queues);
6185 		rx_queues = min(rx_queues, tx_queues);
6186 	}
6187 
6188 	device_printf(dev, "Using %d rx queues %d tx queues\n",
6189 	    rx_queues, tx_queues);
6190 
6191 	vectors = rx_queues + admincnt;
6192 	if ((err = pci_alloc_msix(dev, &vectors)) == 0) {
6193 		device_printf(dev, "Using MSI-X interrupts with %d vectors\n",
6194 		    vectors);
6195 		scctx->isc_vectors = vectors;
6196 		scctx->isc_nrxqsets = rx_queues;
6197 		scctx->isc_ntxqsets = tx_queues;
6198 		scctx->isc_intr = IFLIB_INTR_MSIX;
6199 
6200 		return (vectors);
6201 	} else {
6202 		device_printf(dev,
6203 		    "failed to allocate %d MSI-X vectors, err: %d - using MSI\n",
6204 		    vectors, err);
6205 		bus_release_resource(dev, SYS_RES_MEMORY, bar,
6206 		    ctx->ifc_msix_mem);
6207 		ctx->ifc_msix_mem = NULL;
6208 	}
6209 msi:
6210 	vectors = pci_msi_count(dev);
6211 	scctx->isc_nrxqsets = 1;
6212 	scctx->isc_ntxqsets = 1;
6213 	scctx->isc_vectors = vectors;
6214 	if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) {
6215 		device_printf(dev,"Using an MSI interrupt\n");
6216 		scctx->isc_intr = IFLIB_INTR_MSI;
6217 	} else {
6218 		scctx->isc_vectors = 1;
6219 		device_printf(dev,"Using a Legacy interrupt\n");
6220 		scctx->isc_intr = IFLIB_INTR_LEGACY;
6221 	}
6222 
6223 	return (vectors);
6224 }
6225 
6226 static const char *ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" };
6227 
6228 static int
6229 mp_ring_state_handler(SYSCTL_HANDLER_ARGS)
6230 {
6231 	int rc;
6232 	uint16_t *state = ((uint16_t *)oidp->oid_arg1);
6233 	struct sbuf *sb;
6234 	const char *ring_state = "UNKNOWN";
6235 
6236 	/* XXX needed ? */
6237 	rc = sysctl_wire_old_buffer(req, 0);
6238 	MPASS(rc == 0);
6239 	if (rc != 0)
6240 		return (rc);
6241 	sb = sbuf_new_for_sysctl(NULL, NULL, 80, req);
6242 	MPASS(sb != NULL);
6243 	if (sb == NULL)
6244 		return (ENOMEM);
6245 	if (state[3] <= 3)
6246 		ring_state = ring_states[state[3]];
6247 
6248 	sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s",
6249 		    state[0], state[1], state[2], ring_state);
6250 	rc = sbuf_finish(sb);
6251 	sbuf_delete(sb);
6252         return(rc);
6253 }
6254 
6255 enum iflib_ndesc_handler {
6256 	IFLIB_NTXD_HANDLER,
6257 	IFLIB_NRXD_HANDLER,
6258 };
6259 
6260 static int
6261 mp_ndesc_handler(SYSCTL_HANDLER_ARGS)
6262 {
6263 	if_ctx_t ctx = (void *)arg1;
6264 	enum iflib_ndesc_handler type = arg2;
6265 	char buf[256] = {0};
6266 	qidx_t *ndesc;
6267 	char *p, *next;
6268 	int nqs, rc, i;
6269 
6270 	MPASS(type == IFLIB_NTXD_HANDLER || type == IFLIB_NRXD_HANDLER);
6271 
6272 	nqs = 8;
6273 	switch(type) {
6274 	case IFLIB_NTXD_HANDLER:
6275 		ndesc = ctx->ifc_sysctl_ntxds;
6276 		if (ctx->ifc_sctx)
6277 			nqs = ctx->ifc_sctx->isc_ntxqs;
6278 		break;
6279 	case IFLIB_NRXD_HANDLER:
6280 		ndesc = ctx->ifc_sysctl_nrxds;
6281 		if (ctx->ifc_sctx)
6282 			nqs = ctx->ifc_sctx->isc_nrxqs;
6283 		break;
6284 	default:
6285 			panic("unhandled type");
6286 	}
6287 	if (nqs == 0)
6288 		nqs = 8;
6289 
6290 	for (i=0; i<8; i++) {
6291 		if (i >= nqs)
6292 			break;
6293 		if (i)
6294 			strcat(buf, ",");
6295 		sprintf(strchr(buf, 0), "%d", ndesc[i]);
6296 	}
6297 
6298 	rc = sysctl_handle_string(oidp, buf, sizeof(buf), req);
6299 	if (rc || req->newptr == NULL)
6300 		return rc;
6301 
6302 	for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p;
6303 	    i++, p = strsep(&next, " ,")) {
6304 		ndesc[i] = strtoul(p, NULL, 10);
6305 	}
6306 
6307 	return(rc);
6308 }
6309 
6310 #define NAME_BUFLEN 32
6311 static void
6312 iflib_add_device_sysctl_pre(if_ctx_t ctx)
6313 {
6314         device_t dev = iflib_get_dev(ctx);
6315 	struct sysctl_oid_list *child, *oid_list;
6316 	struct sysctl_ctx_list *ctx_list;
6317 	struct sysctl_oid *node;
6318 
6319 	ctx_list = device_get_sysctl_ctx(dev);
6320 	child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev));
6321 	ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, "iflib",
6322 						      CTLFLAG_RD, NULL, "IFLIB fields");
6323 	oid_list = SYSCTL_CHILDREN(node);
6324 
6325 	SYSCTL_ADD_CONST_STRING(ctx_list, oid_list, OID_AUTO, "driver_version",
6326 		       CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version,
6327 		       "driver version");
6328 
6329 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs",
6330 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0,
6331 			"# of txqs to use, 0 => use default #");
6332 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs",
6333 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0,
6334 			"# of rxqs to use, 0 => use default #");
6335 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_qs_enable",
6336 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0,
6337                        "permit #txq != #rxq");
6338 	SYSCTL_ADD_INT(ctx_list, oid_list, OID_AUTO, "disable_msix",
6339                       CTLFLAG_RWTUN, &ctx->ifc_softc_ctx.isc_disable_msix, 0,
6340                       "disable MSI-X (default 0)");
6341 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "rx_budget",
6342 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_rx_budget, 0,
6343                        "set the rx budget");
6344 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "tx_abdicate",
6345 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_tx_abdicate, 0,
6346 		       "cause tx to abdicate instead of running to completion");
6347 
6348 	/* XXX change for per-queue sizes */
6349 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_ntxds",
6350 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NTXD_HANDLER,
6351                        mp_ndesc_handler, "A",
6352                        "list of # of tx descriptors to use, 0 = use default #");
6353 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_nrxds",
6354 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NRXD_HANDLER,
6355                        mp_ndesc_handler, "A",
6356                        "list of # of rx descriptors to use, 0 = use default #");
6357 }
6358 
6359 static void
6360 iflib_add_device_sysctl_post(if_ctx_t ctx)
6361 {
6362 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6363 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6364         device_t dev = iflib_get_dev(ctx);
6365 	struct sysctl_oid_list *child;
6366 	struct sysctl_ctx_list *ctx_list;
6367 	iflib_fl_t fl;
6368 	iflib_txq_t txq;
6369 	iflib_rxq_t rxq;
6370 	int i, j;
6371 	char namebuf[NAME_BUFLEN];
6372 	char *qfmt;
6373 	struct sysctl_oid *queue_node, *fl_node, *node;
6374 	struct sysctl_oid_list *queue_list, *fl_list;
6375 	ctx_list = device_get_sysctl_ctx(dev);
6376 
6377 	node = ctx->ifc_sysctl_node;
6378 	child = SYSCTL_CHILDREN(node);
6379 
6380 	if (scctx->isc_ntxqsets > 100)
6381 		qfmt = "txq%03d";
6382 	else if (scctx->isc_ntxqsets > 10)
6383 		qfmt = "txq%02d";
6384 	else
6385 		qfmt = "txq%d";
6386 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
6387 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6388 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6389 					     CTLFLAG_RD, NULL, "Queue Name");
6390 		queue_list = SYSCTL_CHILDREN(queue_node);
6391 #if MEMORY_LOGGING
6392 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued",
6393 				CTLFLAG_RD,
6394 				&txq->ift_dequeued, "total mbufs freed");
6395 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued",
6396 				CTLFLAG_RD,
6397 				&txq->ift_enqueued, "total mbufs enqueued");
6398 #endif
6399 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag",
6400 				   CTLFLAG_RD,
6401 				   &txq->ift_mbuf_defrag, "# of times m_defrag was called");
6402 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "m_pullups",
6403 				   CTLFLAG_RD,
6404 				   &txq->ift_pullups, "# of times m_pullup was called");
6405 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag_failed",
6406 				   CTLFLAG_RD,
6407 				   &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed");
6408 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_desc_avail",
6409 				   CTLFLAG_RD,
6410 				   &txq->ift_no_desc_avail, "# of times no descriptors were available");
6411 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "tx_map_failed",
6412 				   CTLFLAG_RD,
6413 				   &txq->ift_map_failed, "# of times dma map failed");
6414 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txd_encap_efbig",
6415 				   CTLFLAG_RD,
6416 				   &txq->ift_txd_encap_efbig, "# of times txd_encap returned EFBIG");
6417 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_tx_dma_setup",
6418 				   CTLFLAG_RD,
6419 				   &txq->ift_no_tx_dma_setup, "# of times map failed for other than EFBIG");
6420 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx",
6421 				   CTLFLAG_RD,
6422 				   &txq->ift_pidx, 1, "Producer Index");
6423 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx",
6424 				   CTLFLAG_RD,
6425 				   &txq->ift_cidx, 1, "Consumer Index");
6426 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx_processed",
6427 				   CTLFLAG_RD,
6428 				   &txq->ift_cidx_processed, 1, "Consumer Index seen by credit update");
6429 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use",
6430 				   CTLFLAG_RD,
6431 				   &txq->ift_in_use, 1, "descriptors in use");
6432 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_processed",
6433 				   CTLFLAG_RD,
6434 				   &txq->ift_processed, "descriptors procesed for clean");
6435 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned",
6436 				   CTLFLAG_RD,
6437 				   &txq->ift_cleaned, "total cleaned");
6438 		SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state",
6439 				CTLTYPE_STRING | CTLFLAG_RD, __DEVOLATILE(uint64_t *, &txq->ift_br->state),
6440 				0, mp_ring_state_handler, "A", "soft ring state");
6441 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_enqueues",
6442 				       CTLFLAG_RD, &txq->ift_br->enqueues,
6443 				       "# of enqueues to the mp_ring for this queue");
6444 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_drops",
6445 				       CTLFLAG_RD, &txq->ift_br->drops,
6446 				       "# of drops in the mp_ring for this queue");
6447 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_starts",
6448 				       CTLFLAG_RD, &txq->ift_br->starts,
6449 				       "# of normal consumer starts in the mp_ring for this queue");
6450 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_stalls",
6451 				       CTLFLAG_RD, &txq->ift_br->stalls,
6452 					       "# of consumer stalls in the mp_ring for this queue");
6453 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_restarts",
6454 			       CTLFLAG_RD, &txq->ift_br->restarts,
6455 				       "# of consumer restarts in the mp_ring for this queue");
6456 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_abdications",
6457 				       CTLFLAG_RD, &txq->ift_br->abdications,
6458 				       "# of consumer abdications in the mp_ring for this queue");
6459 	}
6460 
6461 	if (scctx->isc_nrxqsets > 100)
6462 		qfmt = "rxq%03d";
6463 	else if (scctx->isc_nrxqsets > 10)
6464 		qfmt = "rxq%02d";
6465 	else
6466 		qfmt = "rxq%d";
6467 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
6468 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6469 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6470 					     CTLFLAG_RD, NULL, "Queue Name");
6471 		queue_list = SYSCTL_CHILDREN(queue_node);
6472 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
6473 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_pidx",
6474 				       CTLFLAG_RD,
6475 				       &rxq->ifr_cq_pidx, 1, "Producer Index");
6476 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_cidx",
6477 				       CTLFLAG_RD,
6478 				       &rxq->ifr_cq_cidx, 1, "Consumer Index");
6479 		}
6480 
6481 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
6482 			snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j);
6483 			fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list, OID_AUTO, namebuf,
6484 						     CTLFLAG_RD, NULL, "freelist Name");
6485 			fl_list = SYSCTL_CHILDREN(fl_node);
6486 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx",
6487 				       CTLFLAG_RD,
6488 				       &fl->ifl_pidx, 1, "Producer Index");
6489 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx",
6490 				       CTLFLAG_RD,
6491 				       &fl->ifl_cidx, 1, "Consumer Index");
6492 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits",
6493 				       CTLFLAG_RD,
6494 				       &fl->ifl_credits, 1, "credits available");
6495 #if MEMORY_LOGGING
6496 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_enqueued",
6497 					CTLFLAG_RD,
6498 					&fl->ifl_m_enqueued, "mbufs allocated");
6499 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_dequeued",
6500 					CTLFLAG_RD,
6501 					&fl->ifl_m_dequeued, "mbufs freed");
6502 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_enqueued",
6503 					CTLFLAG_RD,
6504 					&fl->ifl_cl_enqueued, "clusters allocated");
6505 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_dequeued",
6506 					CTLFLAG_RD,
6507 					&fl->ifl_cl_dequeued, "clusters freed");
6508 #endif
6509 
6510 		}
6511 	}
6512 
6513 }
6514 
6515 void
6516 iflib_request_reset(if_ctx_t ctx)
6517 {
6518 
6519 	STATE_LOCK(ctx);
6520 	ctx->ifc_flags |= IFC_DO_RESET;
6521 	STATE_UNLOCK(ctx);
6522 }
6523 
6524 #ifndef __NO_STRICT_ALIGNMENT
6525 static struct mbuf *
6526 iflib_fixup_rx(struct mbuf *m)
6527 {
6528 	struct mbuf *n;
6529 
6530 	if (m->m_len <= (MCLBYTES - ETHER_HDR_LEN)) {
6531 		bcopy(m->m_data, m->m_data + ETHER_HDR_LEN, m->m_len);
6532 		m->m_data += ETHER_HDR_LEN;
6533 		n = m;
6534 	} else {
6535 		MGETHDR(n, M_NOWAIT, MT_DATA);
6536 		if (n == NULL) {
6537 			m_freem(m);
6538 			return (NULL);
6539 		}
6540 		bcopy(m->m_data, n->m_data, ETHER_HDR_LEN);
6541 		m->m_data += ETHER_HDR_LEN;
6542 		m->m_len -= ETHER_HDR_LEN;
6543 		n->m_len = ETHER_HDR_LEN;
6544 		M_MOVE_PKTHDR(n, m);
6545 		n->m_next = m;
6546 	}
6547 	return (n);
6548 }
6549 #endif
6550 
6551 #ifdef NETDUMP
6552 static void
6553 iflib_netdump_init(struct ifnet *ifp, int *nrxr, int *ncl, int *clsize)
6554 {
6555 	if_ctx_t ctx;
6556 
6557 	ctx = if_getsoftc(ifp);
6558 	CTX_LOCK(ctx);
6559 	*nrxr = NRXQSETS(ctx);
6560 	*ncl = ctx->ifc_rxqs[0].ifr_fl->ifl_size;
6561 	*clsize = ctx->ifc_rxqs[0].ifr_fl->ifl_buf_size;
6562 	CTX_UNLOCK(ctx);
6563 }
6564 
6565 static void
6566 iflib_netdump_event(struct ifnet *ifp, enum netdump_ev event)
6567 {
6568 	if_ctx_t ctx;
6569 	if_softc_ctx_t scctx;
6570 	iflib_fl_t fl;
6571 	iflib_rxq_t rxq;
6572 	int i, j;
6573 
6574 	ctx = if_getsoftc(ifp);
6575 	scctx = &ctx->ifc_softc_ctx;
6576 
6577 	switch (event) {
6578 	case NETDUMP_START:
6579 		for (i = 0; i < scctx->isc_nrxqsets; i++) {
6580 			rxq = &ctx->ifc_rxqs[i];
6581 			for (j = 0; j < rxq->ifr_nfl; j++) {
6582 				fl = rxq->ifr_fl;
6583 				fl->ifl_zone = m_getzone(fl->ifl_buf_size);
6584 			}
6585 		}
6586 		iflib_no_tx_batch = 1;
6587 		break;
6588 	default:
6589 		break;
6590 	}
6591 }
6592 
6593 static int
6594 iflib_netdump_transmit(struct ifnet *ifp, struct mbuf *m)
6595 {
6596 	if_ctx_t ctx;
6597 	iflib_txq_t txq;
6598 	int error;
6599 
6600 	ctx = if_getsoftc(ifp);
6601 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
6602 	    IFF_DRV_RUNNING)
6603 		return (EBUSY);
6604 
6605 	txq = &ctx->ifc_txqs[0];
6606 	error = iflib_encap(txq, &m);
6607 	if (error == 0)
6608 		(void)iflib_txd_db_check(ctx, txq, true, txq->ift_in_use);
6609 	return (error);
6610 }
6611 
6612 static int
6613 iflib_netdump_poll(struct ifnet *ifp, int count)
6614 {
6615 	if_ctx_t ctx;
6616 	if_softc_ctx_t scctx;
6617 	iflib_txq_t txq;
6618 	int i;
6619 
6620 	ctx = if_getsoftc(ifp);
6621 	scctx = &ctx->ifc_softc_ctx;
6622 
6623 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
6624 	    IFF_DRV_RUNNING)
6625 		return (EBUSY);
6626 
6627 	txq = &ctx->ifc_txqs[0];
6628 	(void)iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
6629 
6630 	for (i = 0; i < scctx->isc_nrxqsets; i++)
6631 		(void)iflib_rxeof(&ctx->ifc_rxqs[i], 16 /* XXX */);
6632 	return (0);
6633 }
6634 #endif /* NETDUMP */
6635