xref: /freebsd/sys/net/iflib.c (revision 88640c0e8b6f503426cce9ea1337098c241d3801)
1 /*-
2  * Copyright (c) 2014-2018, Matthew Macy <mmacy@mattmacy.io>
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions are met:
7  *
8  *  1. Redistributions of source code must retain the above copyright notice,
9  *     this list of conditions and the following disclaimer.
10  *
11  *  2. Neither the name of Matthew Macy nor the names of its
12  *     contributors may be used to endorse or promote products derived from
13  *     this software without specific prior written permission.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
16  * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18  * ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
19  * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
20  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
21  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
22  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
23  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
24  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
25  * POSSIBILITY OF SUCH DAMAGE.
26  */
27 
28 #include <sys/cdefs.h>
29 __FBSDID("$FreeBSD$");
30 
31 #include "opt_inet.h"
32 #include "opt_inet6.h"
33 #include "opt_acpi.h"
34 #include "opt_sched.h"
35 
36 #include <sys/param.h>
37 #include <sys/types.h>
38 #include <sys/bus.h>
39 #include <sys/eventhandler.h>
40 #include <sys/jail.h>
41 #include <sys/kernel.h>
42 #include <sys/lock.h>
43 #include <sys/md5.h>
44 #include <sys/mutex.h>
45 #include <sys/module.h>
46 #include <sys/kobj.h>
47 #include <sys/rman.h>
48 #include <sys/proc.h>
49 #include <sys/sbuf.h>
50 #include <sys/smp.h>
51 #include <sys/socket.h>
52 #include <sys/sockio.h>
53 #include <sys/sysctl.h>
54 #include <sys/syslog.h>
55 #include <sys/taskqueue.h>
56 #include <sys/limits.h>
57 
58 #include <net/if.h>
59 #include <net/if_var.h>
60 #include <net/if_types.h>
61 #include <net/if_media.h>
62 #include <net/bpf.h>
63 #include <net/ethernet.h>
64 #include <net/mp_ring.h>
65 #include <net/vnet.h>
66 
67 #include <netinet/in.h>
68 #include <netinet/in_pcb.h>
69 #include <netinet/tcp_lro.h>
70 #include <netinet/in_systm.h>
71 #include <netinet/if_ether.h>
72 #include <netinet/ip.h>
73 #include <netinet/ip6.h>
74 #include <netinet/tcp.h>
75 #include <netinet/ip_var.h>
76 #include <netinet/netdump/netdump.h>
77 #include <netinet6/ip6_var.h>
78 
79 #include <machine/bus.h>
80 #include <machine/in_cksum.h>
81 
82 #include <vm/vm.h>
83 #include <vm/pmap.h>
84 
85 #include <dev/led/led.h>
86 #include <dev/pci/pcireg.h>
87 #include <dev/pci/pcivar.h>
88 #include <dev/pci/pci_private.h>
89 
90 #include <net/iflib.h>
91 #include <net/iflib_private.h>
92 
93 #include "ifdi_if.h"
94 
95 #ifdef PCI_IOV
96 #include <dev/pci/pci_iov.h>
97 #endif
98 
99 #include <sys/bitstring.h>
100 /*
101  * enable accounting of every mbuf as it comes in to and goes out of
102  * iflib's software descriptor references
103  */
104 #define MEMORY_LOGGING 0
105 /*
106  * Enable mbuf vectors for compressing long mbuf chains
107  */
108 
109 /*
110  * NB:
111  * - Prefetching in tx cleaning should perhaps be a tunable. The distance ahead
112  *   we prefetch needs to be determined by the time spent in m_free vis a vis
113  *   the cost of a prefetch. This will of course vary based on the workload:
114  *      - NFLX's m_free path is dominated by vm-based M_EXT manipulation which
115  *        is quite expensive, thus suggesting very little prefetch.
116  *      - small packet forwarding which is just returning a single mbuf to
117  *        UMA will typically be very fast vis a vis the cost of a memory
118  *        access.
119  */
120 
121 
122 /*
123  * File organization:
124  *  - private structures
125  *  - iflib private utility functions
126  *  - ifnet functions
127  *  - vlan registry and other exported functions
128  *  - iflib public core functions
129  *
130  *
131  */
132 MALLOC_DEFINE(M_IFLIB, "iflib", "ifnet library");
133 
134 struct iflib_txq;
135 typedef struct iflib_txq *iflib_txq_t;
136 struct iflib_rxq;
137 typedef struct iflib_rxq *iflib_rxq_t;
138 struct iflib_fl;
139 typedef struct iflib_fl *iflib_fl_t;
140 
141 struct iflib_ctx;
142 
143 static void iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid);
144 static void iflib_timer(void *arg);
145 
146 typedef struct iflib_filter_info {
147 	driver_filter_t *ifi_filter;
148 	void *ifi_filter_arg;
149 	struct grouptask *ifi_task;
150 	void *ifi_ctx;
151 } *iflib_filter_info_t;
152 
153 struct iflib_ctx {
154 	KOBJ_FIELDS;
155 	/*
156 	 * Pointer to hardware driver's softc
157 	 */
158 	void *ifc_softc;
159 	device_t ifc_dev;
160 	if_t ifc_ifp;
161 
162 	cpuset_t ifc_cpus;
163 	if_shared_ctx_t ifc_sctx;
164 	struct if_softc_ctx ifc_softc_ctx;
165 
166 	struct sx ifc_ctx_sx;
167 	struct mtx ifc_state_mtx;
168 
169 	iflib_txq_t ifc_txqs;
170 	iflib_rxq_t ifc_rxqs;
171 	uint32_t ifc_if_flags;
172 	uint32_t ifc_flags;
173 	uint32_t ifc_max_fl_buf_size;
174 
175 	int ifc_link_state;
176 	int ifc_link_irq;
177 	int ifc_watchdog_events;
178 	struct cdev *ifc_led_dev;
179 	struct resource *ifc_msix_mem;
180 
181 	struct if_irq ifc_legacy_irq;
182 	struct grouptask ifc_admin_task;
183 	struct grouptask ifc_vflr_task;
184 	struct iflib_filter_info ifc_filter_info;
185 	struct ifmedia	ifc_media;
186 
187 	struct sysctl_oid *ifc_sysctl_node;
188 	uint16_t ifc_sysctl_ntxqs;
189 	uint16_t ifc_sysctl_nrxqs;
190 	uint16_t ifc_sysctl_qs_eq_override;
191 	uint16_t ifc_sysctl_rx_budget;
192 	uint16_t ifc_sysctl_tx_abdicate;
193 
194 	qidx_t ifc_sysctl_ntxds[8];
195 	qidx_t ifc_sysctl_nrxds[8];
196 	struct if_txrx ifc_txrx;
197 #define isc_txd_encap  ifc_txrx.ift_txd_encap
198 #define isc_txd_flush  ifc_txrx.ift_txd_flush
199 #define isc_txd_credits_update  ifc_txrx.ift_txd_credits_update
200 #define isc_rxd_available ifc_txrx.ift_rxd_available
201 #define isc_rxd_pkt_get ifc_txrx.ift_rxd_pkt_get
202 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
203 #define isc_rxd_flush ifc_txrx.ift_rxd_flush
204 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
205 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
206 #define isc_legacy_intr ifc_txrx.ift_legacy_intr
207 	eventhandler_tag ifc_vlan_attach_event;
208 	eventhandler_tag ifc_vlan_detach_event;
209 	uint8_t ifc_mac[ETHER_ADDR_LEN];
210 	char ifc_mtx_name[16];
211 };
212 
213 
214 void *
215 iflib_get_softc(if_ctx_t ctx)
216 {
217 
218 	return (ctx->ifc_softc);
219 }
220 
221 device_t
222 iflib_get_dev(if_ctx_t ctx)
223 {
224 
225 	return (ctx->ifc_dev);
226 }
227 
228 if_t
229 iflib_get_ifp(if_ctx_t ctx)
230 {
231 
232 	return (ctx->ifc_ifp);
233 }
234 
235 struct ifmedia *
236 iflib_get_media(if_ctx_t ctx)
237 {
238 
239 	return (&ctx->ifc_media);
240 }
241 
242 uint32_t
243 iflib_get_flags(if_ctx_t ctx)
244 {
245 	return (ctx->ifc_flags);
246 }
247 
248 void
249 iflib_set_mac(if_ctx_t ctx, uint8_t mac[ETHER_ADDR_LEN])
250 {
251 
252 	bcopy(mac, ctx->ifc_mac, ETHER_ADDR_LEN);
253 }
254 
255 if_softc_ctx_t
256 iflib_get_softc_ctx(if_ctx_t ctx)
257 {
258 
259 	return (&ctx->ifc_softc_ctx);
260 }
261 
262 if_shared_ctx_t
263 iflib_get_sctx(if_ctx_t ctx)
264 {
265 
266 	return (ctx->ifc_sctx);
267 }
268 
269 #define IP_ALIGNED(m) ((((uintptr_t)(m)->m_data) & 0x3) == 0x2)
270 #define CACHE_PTR_INCREMENT (CACHE_LINE_SIZE/sizeof(void*))
271 #define CACHE_PTR_NEXT(ptr) ((void *)(((uintptr_t)(ptr)+CACHE_LINE_SIZE-1) & (CACHE_LINE_SIZE-1)))
272 
273 #define LINK_ACTIVE(ctx) ((ctx)->ifc_link_state == LINK_STATE_UP)
274 #define CTX_IS_VF(ctx) ((ctx)->ifc_sctx->isc_flags & IFLIB_IS_VF)
275 
276 typedef struct iflib_sw_rx_desc_array {
277 	bus_dmamap_t	*ifsd_map;         /* bus_dma maps for packet */
278 	struct mbuf	**ifsd_m;           /* pkthdr mbufs */
279 	caddr_t		*ifsd_cl;          /* direct cluster pointer for rx */
280 	bus_addr_t	*ifsd_ba;          /* bus addr of cluster for rx */
281 } iflib_rxsd_array_t;
282 
283 typedef struct iflib_sw_tx_desc_array {
284 	bus_dmamap_t    *ifsd_map;         /* bus_dma maps for packet */
285 	bus_dmamap_t	*ifsd_tso_map;     /* bus_dma maps for TSO packet */
286 	struct mbuf    **ifsd_m;           /* pkthdr mbufs */
287 } if_txsd_vec_t;
288 
289 
290 /* magic number that should be high enough for any hardware */
291 #define IFLIB_MAX_TX_SEGS		128
292 /* bnxt supports 64 with hardware LRO enabled */
293 #define IFLIB_MAX_RX_SEGS		64
294 #define IFLIB_RX_COPY_THRESH		128
295 #define IFLIB_MAX_RX_REFRESH		32
296 /* The minimum descriptors per second before we start coalescing */
297 #define IFLIB_MIN_DESC_SEC		16384
298 #define IFLIB_DEFAULT_TX_UPDATE_FREQ	16
299 #define IFLIB_QUEUE_IDLE		0
300 #define IFLIB_QUEUE_HUNG		1
301 #define IFLIB_QUEUE_WORKING		2
302 /* maximum number of txqs that can share an rx interrupt */
303 #define IFLIB_MAX_TX_SHARED_INTR	4
304 
305 /* this should really scale with ring size - this is a fairly arbitrary value */
306 #define TX_BATCH_SIZE			32
307 
308 #define IFLIB_RESTART_BUDGET		8
309 
310 
311 #define CSUM_OFFLOAD		(CSUM_IP_TSO|CSUM_IP6_TSO|CSUM_IP| \
312 				 CSUM_IP_UDP|CSUM_IP_TCP|CSUM_IP_SCTP| \
313 				 CSUM_IP6_UDP|CSUM_IP6_TCP|CSUM_IP6_SCTP)
314 struct iflib_txq {
315 	qidx_t		ift_in_use;
316 	qidx_t		ift_cidx;
317 	qidx_t		ift_cidx_processed;
318 	qidx_t		ift_pidx;
319 	uint8_t		ift_gen;
320 	uint8_t		ift_br_offset;
321 	uint16_t	ift_npending;
322 	uint16_t	ift_db_pending;
323 	uint16_t	ift_rs_pending;
324 	/* implicit pad */
325 	uint8_t		ift_txd_size[8];
326 	uint64_t	ift_processed;
327 	uint64_t	ift_cleaned;
328 	uint64_t	ift_cleaned_prev;
329 #if MEMORY_LOGGING
330 	uint64_t	ift_enqueued;
331 	uint64_t	ift_dequeued;
332 #endif
333 	uint64_t	ift_no_tx_dma_setup;
334 	uint64_t	ift_no_desc_avail;
335 	uint64_t	ift_mbuf_defrag_failed;
336 	uint64_t	ift_mbuf_defrag;
337 	uint64_t	ift_map_failed;
338 	uint64_t	ift_txd_encap_efbig;
339 	uint64_t	ift_pullups;
340 	uint64_t	ift_last_timer_tick;
341 
342 	struct mtx	ift_mtx;
343 	struct mtx	ift_db_mtx;
344 
345 	/* constant values */
346 	if_ctx_t	ift_ctx;
347 	struct ifmp_ring        *ift_br;
348 	struct grouptask	ift_task;
349 	qidx_t		ift_size;
350 	uint16_t	ift_id;
351 	struct callout	ift_timer;
352 
353 	if_txsd_vec_t	ift_sds;
354 	uint8_t		ift_qstatus;
355 	uint8_t		ift_closed;
356 	uint8_t		ift_update_freq;
357 	struct iflib_filter_info ift_filter_info;
358 	bus_dma_tag_t		ift_desc_tag;
359 	bus_dma_tag_t		ift_tso_desc_tag;
360 	iflib_dma_info_t	ift_ifdi;
361 #define MTX_NAME_LEN 16
362 	char                    ift_mtx_name[MTX_NAME_LEN];
363 	char                    ift_db_mtx_name[MTX_NAME_LEN];
364 	bus_dma_segment_t	ift_segs[IFLIB_MAX_TX_SEGS]  __aligned(CACHE_LINE_SIZE);
365 #ifdef IFLIB_DIAGNOSTICS
366 	uint64_t ift_cpu_exec_count[256];
367 #endif
368 } __aligned(CACHE_LINE_SIZE);
369 
370 struct iflib_fl {
371 	qidx_t		ifl_cidx;
372 	qidx_t		ifl_pidx;
373 	qidx_t		ifl_credits;
374 	uint8_t		ifl_gen;
375 	uint8_t		ifl_rxd_size;
376 #if MEMORY_LOGGING
377 	uint64_t	ifl_m_enqueued;
378 	uint64_t	ifl_m_dequeued;
379 	uint64_t	ifl_cl_enqueued;
380 	uint64_t	ifl_cl_dequeued;
381 #endif
382 	/* implicit pad */
383 
384 	bitstr_t 	*ifl_rx_bitmap;
385 	qidx_t		ifl_fragidx;
386 	/* constant */
387 	qidx_t		ifl_size;
388 	uint16_t	ifl_buf_size;
389 	uint16_t	ifl_cltype;
390 	uma_zone_t	ifl_zone;
391 	iflib_rxsd_array_t	ifl_sds;
392 	iflib_rxq_t	ifl_rxq;
393 	uint8_t		ifl_id;
394 	bus_dma_tag_t           ifl_desc_tag;
395 	iflib_dma_info_t	ifl_ifdi;
396 	uint64_t	ifl_bus_addrs[IFLIB_MAX_RX_REFRESH] __aligned(CACHE_LINE_SIZE);
397 	caddr_t		ifl_vm_addrs[IFLIB_MAX_RX_REFRESH];
398 	qidx_t	ifl_rxd_idxs[IFLIB_MAX_RX_REFRESH];
399 }  __aligned(CACHE_LINE_SIZE);
400 
401 static inline qidx_t
402 get_inuse(int size, qidx_t cidx, qidx_t pidx, uint8_t gen)
403 {
404 	qidx_t used;
405 
406 	if (pidx > cidx)
407 		used = pidx - cidx;
408 	else if (pidx < cidx)
409 		used = size - cidx + pidx;
410 	else if (gen == 0 && pidx == cidx)
411 		used = 0;
412 	else if (gen == 1 && pidx == cidx)
413 		used = size;
414 	else
415 		panic("bad state");
416 
417 	return (used);
418 }
419 
420 #define TXQ_AVAIL(txq) (txq->ift_size - get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx, txq->ift_gen))
421 
422 #define IDXDIFF(head, tail, wrap) \
423 	((head) >= (tail) ? (head) - (tail) : (wrap) - (tail) + (head))
424 
425 struct iflib_rxq {
426 	/* If there is a separate completion queue -
427 	 * these are the cq cidx and pidx. Otherwise
428 	 * these are unused.
429 	 */
430 	qidx_t		ifr_size;
431 	qidx_t		ifr_cq_cidx;
432 	qidx_t		ifr_cq_pidx;
433 	uint8_t		ifr_cq_gen;
434 	uint8_t		ifr_fl_offset;
435 
436 	if_ctx_t	ifr_ctx;
437 	iflib_fl_t	ifr_fl;
438 	uint64_t	ifr_rx_irq;
439 	uint16_t	ifr_id;
440 	uint8_t		ifr_lro_enabled;
441 	uint8_t		ifr_nfl;
442 	uint8_t		ifr_ntxqirq;
443 	uint8_t		ifr_txqid[IFLIB_MAX_TX_SHARED_INTR];
444 	struct lro_ctrl			ifr_lc;
445 	struct grouptask        ifr_task;
446 	struct iflib_filter_info ifr_filter_info;
447 	iflib_dma_info_t		ifr_ifdi;
448 
449 	/* dynamically allocate if any drivers need a value substantially larger than this */
450 	struct if_rxd_frag	ifr_frags[IFLIB_MAX_RX_SEGS] __aligned(CACHE_LINE_SIZE);
451 #ifdef IFLIB_DIAGNOSTICS
452 	uint64_t ifr_cpu_exec_count[256];
453 #endif
454 }  __aligned(CACHE_LINE_SIZE);
455 
456 typedef struct if_rxsd {
457 	caddr_t *ifsd_cl;
458 	struct mbuf **ifsd_m;
459 	iflib_fl_t ifsd_fl;
460 	qidx_t ifsd_cidx;
461 } *if_rxsd_t;
462 
463 /* multiple of word size */
464 #ifdef __LP64__
465 #define PKT_INFO_SIZE	6
466 #define RXD_INFO_SIZE	5
467 #define PKT_TYPE uint64_t
468 #else
469 #define PKT_INFO_SIZE	11
470 #define RXD_INFO_SIZE	8
471 #define PKT_TYPE uint32_t
472 #endif
473 #define PKT_LOOP_BOUND  ((PKT_INFO_SIZE/3)*3)
474 #define RXD_LOOP_BOUND  ((RXD_INFO_SIZE/4)*4)
475 
476 typedef struct if_pkt_info_pad {
477 	PKT_TYPE pkt_val[PKT_INFO_SIZE];
478 } *if_pkt_info_pad_t;
479 typedef struct if_rxd_info_pad {
480 	PKT_TYPE rxd_val[RXD_INFO_SIZE];
481 } *if_rxd_info_pad_t;
482 
483 CTASSERT(sizeof(struct if_pkt_info_pad) == sizeof(struct if_pkt_info));
484 CTASSERT(sizeof(struct if_rxd_info_pad) == sizeof(struct if_rxd_info));
485 
486 
487 static inline void
488 pkt_info_zero(if_pkt_info_t pi)
489 {
490 	if_pkt_info_pad_t pi_pad;
491 
492 	pi_pad = (if_pkt_info_pad_t)pi;
493 	pi_pad->pkt_val[0] = 0; pi_pad->pkt_val[1] = 0; pi_pad->pkt_val[2] = 0;
494 	pi_pad->pkt_val[3] = 0; pi_pad->pkt_val[4] = 0; pi_pad->pkt_val[5] = 0;
495 #ifndef __LP64__
496 	pi_pad->pkt_val[6] = 0; pi_pad->pkt_val[7] = 0; pi_pad->pkt_val[8] = 0;
497 	pi_pad->pkt_val[9] = 0; pi_pad->pkt_val[10] = 0;
498 #endif
499 }
500 
501 static device_method_t iflib_pseudo_methods[] = {
502 	DEVMETHOD(device_attach, noop_attach),
503 	DEVMETHOD(device_detach, iflib_pseudo_detach),
504 	DEVMETHOD_END
505 };
506 
507 driver_t iflib_pseudodriver = {
508 	"iflib_pseudo", iflib_pseudo_methods, sizeof(struct iflib_ctx),
509 };
510 
511 static inline void
512 rxd_info_zero(if_rxd_info_t ri)
513 {
514 	if_rxd_info_pad_t ri_pad;
515 	int i;
516 
517 	ri_pad = (if_rxd_info_pad_t)ri;
518 	for (i = 0; i < RXD_LOOP_BOUND; i += 4) {
519 		ri_pad->rxd_val[i] = 0;
520 		ri_pad->rxd_val[i+1] = 0;
521 		ri_pad->rxd_val[i+2] = 0;
522 		ri_pad->rxd_val[i+3] = 0;
523 	}
524 #ifdef __LP64__
525 	ri_pad->rxd_val[RXD_INFO_SIZE-1] = 0;
526 #endif
527 }
528 
529 /*
530  * Only allow a single packet to take up most 1/nth of the tx ring
531  */
532 #define MAX_SINGLE_PACKET_FRACTION 12
533 #define IF_BAD_DMA (bus_addr_t)-1
534 
535 #define CTX_ACTIVE(ctx) ((if_getdrvflags((ctx)->ifc_ifp) & IFF_DRV_RUNNING))
536 
537 #define CTX_LOCK_INIT(_sc)  sx_init(&(_sc)->ifc_ctx_sx, "iflib ctx lock")
538 #define CTX_LOCK(ctx) sx_xlock(&(ctx)->ifc_ctx_sx)
539 #define CTX_UNLOCK(ctx) sx_xunlock(&(ctx)->ifc_ctx_sx)
540 #define CTX_LOCK_DESTROY(ctx) sx_destroy(&(ctx)->ifc_ctx_sx)
541 
542 
543 #define STATE_LOCK_INIT(_sc, _name)  mtx_init(&(_sc)->ifc_state_mtx, _name, "iflib state lock", MTX_DEF)
544 #define STATE_LOCK(ctx) mtx_lock(&(ctx)->ifc_state_mtx)
545 #define STATE_UNLOCK(ctx) mtx_unlock(&(ctx)->ifc_state_mtx)
546 #define STATE_LOCK_DESTROY(ctx) mtx_destroy(&(ctx)->ifc_state_mtx)
547 
548 
549 
550 #define CALLOUT_LOCK(txq)	mtx_lock(&txq->ift_mtx)
551 #define CALLOUT_UNLOCK(txq) 	mtx_unlock(&txq->ift_mtx)
552 
553 void
554 iflib_set_detach(if_ctx_t ctx)
555 {
556 	STATE_LOCK(ctx);
557 	ctx->ifc_flags |= IFC_IN_DETACH;
558 	STATE_UNLOCK(ctx);
559 }
560 
561 /* Our boot-time initialization hook */
562 static int	iflib_module_event_handler(module_t, int, void *);
563 
564 static moduledata_t iflib_moduledata = {
565 	"iflib",
566 	iflib_module_event_handler,
567 	NULL
568 };
569 
570 DECLARE_MODULE(iflib, iflib_moduledata, SI_SUB_INIT_IF, SI_ORDER_ANY);
571 MODULE_VERSION(iflib, 1);
572 
573 MODULE_DEPEND(iflib, pci, 1, 1, 1);
574 MODULE_DEPEND(iflib, ether, 1, 1, 1);
575 
576 TASKQGROUP_DEFINE(if_io_tqg, mp_ncpus, 1);
577 TASKQGROUP_DEFINE(if_config_tqg, 1, 1);
578 
579 #ifndef IFLIB_DEBUG_COUNTERS
580 #ifdef INVARIANTS
581 #define IFLIB_DEBUG_COUNTERS 1
582 #else
583 #define IFLIB_DEBUG_COUNTERS 0
584 #endif /* !INVARIANTS */
585 #endif
586 
587 static SYSCTL_NODE(_net, OID_AUTO, iflib, CTLFLAG_RD, 0,
588                    "iflib driver parameters");
589 
590 /*
591  * XXX need to ensure that this can't accidentally cause the head to be moved backwards
592  */
593 static int iflib_min_tx_latency = 0;
594 SYSCTL_INT(_net_iflib, OID_AUTO, min_tx_latency, CTLFLAG_RW,
595 		   &iflib_min_tx_latency, 0, "minimize transmit latency at the possible expense of throughput");
596 static int iflib_no_tx_batch = 0;
597 SYSCTL_INT(_net_iflib, OID_AUTO, no_tx_batch, CTLFLAG_RW,
598 		   &iflib_no_tx_batch, 0, "minimize transmit latency at the possible expense of throughput");
599 
600 
601 #if IFLIB_DEBUG_COUNTERS
602 
603 static int iflib_tx_seen;
604 static int iflib_tx_sent;
605 static int iflib_tx_encap;
606 static int iflib_rx_allocs;
607 static int iflib_fl_refills;
608 static int iflib_fl_refills_large;
609 static int iflib_tx_frees;
610 
611 SYSCTL_INT(_net_iflib, OID_AUTO, tx_seen, CTLFLAG_RD,
612 		   &iflib_tx_seen, 0, "# tx mbufs seen");
613 SYSCTL_INT(_net_iflib, OID_AUTO, tx_sent, CTLFLAG_RD,
614 		   &iflib_tx_sent, 0, "# tx mbufs sent");
615 SYSCTL_INT(_net_iflib, OID_AUTO, tx_encap, CTLFLAG_RD,
616 		   &iflib_tx_encap, 0, "# tx mbufs encapped");
617 SYSCTL_INT(_net_iflib, OID_AUTO, tx_frees, CTLFLAG_RD,
618 		   &iflib_tx_frees, 0, "# tx frees");
619 SYSCTL_INT(_net_iflib, OID_AUTO, rx_allocs, CTLFLAG_RD,
620 		   &iflib_rx_allocs, 0, "# rx allocations");
621 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills, CTLFLAG_RD,
622 		   &iflib_fl_refills, 0, "# refills");
623 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills_large, CTLFLAG_RD,
624 		   &iflib_fl_refills_large, 0, "# large refills");
625 
626 
627 static int iflib_txq_drain_flushing;
628 static int iflib_txq_drain_oactive;
629 static int iflib_txq_drain_notready;
630 
631 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_flushing, CTLFLAG_RD,
632 		   &iflib_txq_drain_flushing, 0, "# drain flushes");
633 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_oactive, CTLFLAG_RD,
634 		   &iflib_txq_drain_oactive, 0, "# drain oactives");
635 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_notready, CTLFLAG_RD,
636 		   &iflib_txq_drain_notready, 0, "# drain notready");
637 
638 
639 static int iflib_encap_load_mbuf_fail;
640 static int iflib_encap_pad_mbuf_fail;
641 static int iflib_encap_txq_avail_fail;
642 static int iflib_encap_txd_encap_fail;
643 
644 SYSCTL_INT(_net_iflib, OID_AUTO, encap_load_mbuf_fail, CTLFLAG_RD,
645 		   &iflib_encap_load_mbuf_fail, 0, "# busdma load failures");
646 SYSCTL_INT(_net_iflib, OID_AUTO, encap_pad_mbuf_fail, CTLFLAG_RD,
647 		   &iflib_encap_pad_mbuf_fail, 0, "# runt frame pad failures");
648 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txq_avail_fail, CTLFLAG_RD,
649 		   &iflib_encap_txq_avail_fail, 0, "# txq avail failures");
650 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txd_encap_fail, CTLFLAG_RD,
651 		   &iflib_encap_txd_encap_fail, 0, "# driver encap failures");
652 
653 static int iflib_task_fn_rxs;
654 static int iflib_rx_intr_enables;
655 static int iflib_fast_intrs;
656 static int iflib_rx_unavail;
657 static int iflib_rx_ctx_inactive;
658 static int iflib_rx_if_input;
659 static int iflib_rx_mbuf_null;
660 static int iflib_rxd_flush;
661 
662 static int iflib_verbose_debug;
663 
664 SYSCTL_INT(_net_iflib, OID_AUTO, task_fn_rx, CTLFLAG_RD,
665 		   &iflib_task_fn_rxs, 0, "# task_fn_rx calls");
666 SYSCTL_INT(_net_iflib, OID_AUTO, rx_intr_enables, CTLFLAG_RD,
667 		   &iflib_rx_intr_enables, 0, "# rx intr enables");
668 SYSCTL_INT(_net_iflib, OID_AUTO, fast_intrs, CTLFLAG_RD,
669 		   &iflib_fast_intrs, 0, "# fast_intr calls");
670 SYSCTL_INT(_net_iflib, OID_AUTO, rx_unavail, CTLFLAG_RD,
671 		   &iflib_rx_unavail, 0, "# times rxeof called with no available data");
672 SYSCTL_INT(_net_iflib, OID_AUTO, rx_ctx_inactive, CTLFLAG_RD,
673 		   &iflib_rx_ctx_inactive, 0, "# times rxeof called with inactive context");
674 SYSCTL_INT(_net_iflib, OID_AUTO, rx_if_input, CTLFLAG_RD,
675 		   &iflib_rx_if_input, 0, "# times rxeof called if_input");
676 SYSCTL_INT(_net_iflib, OID_AUTO, rx_mbuf_null, CTLFLAG_RD,
677 		   &iflib_rx_mbuf_null, 0, "# times rxeof got null mbuf");
678 SYSCTL_INT(_net_iflib, OID_AUTO, rxd_flush, CTLFLAG_RD,
679 	         &iflib_rxd_flush, 0, "# times rxd_flush called");
680 SYSCTL_INT(_net_iflib, OID_AUTO, verbose_debug, CTLFLAG_RW,
681 		   &iflib_verbose_debug, 0, "enable verbose debugging");
682 
683 #define DBG_COUNTER_INC(name) atomic_add_int(&(iflib_ ## name), 1)
684 static void
685 iflib_debug_reset(void)
686 {
687 	iflib_tx_seen = iflib_tx_sent = iflib_tx_encap = iflib_rx_allocs =
688 		iflib_fl_refills = iflib_fl_refills_large = iflib_tx_frees =
689 		iflib_txq_drain_flushing = iflib_txq_drain_oactive =
690 		iflib_txq_drain_notready =
691 		iflib_encap_load_mbuf_fail = iflib_encap_pad_mbuf_fail =
692 		iflib_encap_txq_avail_fail = iflib_encap_txd_encap_fail =
693 		iflib_task_fn_rxs = iflib_rx_intr_enables = iflib_fast_intrs =
694 		iflib_rx_unavail =
695 		iflib_rx_ctx_inactive = iflib_rx_if_input =
696 		iflib_rx_mbuf_null = iflib_rxd_flush = 0;
697 }
698 
699 #else
700 #define DBG_COUNTER_INC(name)
701 static void iflib_debug_reset(void) {}
702 #endif
703 
704 #define IFLIB_DEBUG 0
705 
706 static void iflib_tx_structures_free(if_ctx_t ctx);
707 static void iflib_rx_structures_free(if_ctx_t ctx);
708 static int iflib_queues_alloc(if_ctx_t ctx);
709 static int iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq);
710 static int iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget);
711 static int iflib_qset_structures_setup(if_ctx_t ctx);
712 static int iflib_msix_init(if_ctx_t ctx);
713 static int iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filterarg, int *rid, const char *str);
714 static void iflib_txq_check_drain(iflib_txq_t txq, int budget);
715 static uint32_t iflib_txq_can_drain(struct ifmp_ring *);
716 #ifdef ALTQ
717 static void iflib_altq_if_start(if_t ifp);
718 static int iflib_altq_if_transmit(if_t ifp, struct mbuf *m);
719 #endif
720 static int iflib_register(if_ctx_t);
721 static void iflib_init_locked(if_ctx_t ctx);
722 static void iflib_add_device_sysctl_pre(if_ctx_t ctx);
723 static void iflib_add_device_sysctl_post(if_ctx_t ctx);
724 static void iflib_ifmp_purge(iflib_txq_t txq);
725 static void _iflib_pre_assert(if_softc_ctx_t scctx);
726 static void iflib_if_init_locked(if_ctx_t ctx);
727 static void iflib_free_intr_mem(if_ctx_t ctx);
728 #ifndef __NO_STRICT_ALIGNMENT
729 static struct mbuf * iflib_fixup_rx(struct mbuf *m);
730 #endif
731 
732 NETDUMP_DEFINE(iflib);
733 
734 #ifdef DEV_NETMAP
735 #include <sys/selinfo.h>
736 #include <net/netmap.h>
737 #include <dev/netmap/netmap_kern.h>
738 
739 MODULE_DEPEND(iflib, netmap, 1, 1, 1);
740 
741 static int netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, uint32_t nm_i, bool init);
742 
743 /*
744  * device-specific sysctl variables:
745  *
746  * iflib_crcstrip: 0: keep CRC in rx frames (default), 1: strip it.
747  *	During regular operations the CRC is stripped, but on some
748  *	hardware reception of frames not multiple of 64 is slower,
749  *	so using crcstrip=0 helps in benchmarks.
750  *
751  * iflib_rx_miss, iflib_rx_miss_bufs:
752  *	count packets that might be missed due to lost interrupts.
753  */
754 SYSCTL_DECL(_dev_netmap);
755 /*
756  * The xl driver by default strips CRCs and we do not override it.
757  */
758 
759 int iflib_crcstrip = 1;
760 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_crcstrip,
761     CTLFLAG_RW, &iflib_crcstrip, 1, "strip CRC on rx frames");
762 
763 int iflib_rx_miss, iflib_rx_miss_bufs;
764 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss,
765     CTLFLAG_RW, &iflib_rx_miss, 0, "potentially missed rx intr");
766 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss_bufs,
767     CTLFLAG_RW, &iflib_rx_miss_bufs, 0, "potentially missed rx intr bufs");
768 
769 /*
770  * Register/unregister. We are already under netmap lock.
771  * Only called on the first register or the last unregister.
772  */
773 static int
774 iflib_netmap_register(struct netmap_adapter *na, int onoff)
775 {
776 	struct ifnet *ifp = na->ifp;
777 	if_ctx_t ctx = ifp->if_softc;
778 	int status;
779 
780 	CTX_LOCK(ctx);
781 	IFDI_INTR_DISABLE(ctx);
782 
783 	/* Tell the stack that the interface is no longer active */
784 	ifp->if_drv_flags &= ~(IFF_DRV_RUNNING | IFF_DRV_OACTIVE);
785 
786 	if (!CTX_IS_VF(ctx))
787 		IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip);
788 
789 	/* enable or disable flags and callbacks in na and ifp */
790 	if (onoff) {
791 		nm_set_native_flags(na);
792 	} else {
793 		nm_clear_native_flags(na);
794 	}
795 	iflib_stop(ctx);
796 	iflib_init_locked(ctx);
797 	IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip); // XXX why twice ?
798 	status = ifp->if_drv_flags & IFF_DRV_RUNNING ? 0 : 1;
799 	if (status)
800 		nm_clear_native_flags(na);
801 	CTX_UNLOCK(ctx);
802 	return (status);
803 }
804 
805 static int
806 netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, uint32_t nm_i, bool init)
807 {
808 	struct netmap_adapter *na = kring->na;
809 	u_int const lim = kring->nkr_num_slots - 1;
810 	u_int head = kring->rhead;
811 	struct netmap_ring *ring = kring->ring;
812 	bus_dmamap_t *map;
813 	struct if_rxd_update iru;
814 	if_ctx_t ctx = rxq->ifr_ctx;
815 	iflib_fl_t fl = &rxq->ifr_fl[0];
816 	uint32_t refill_pidx, nic_i;
817 #if IFLIB_DEBUG_COUNTERS
818 	int rf_count = 0;
819 #endif
820 
821 	if (nm_i == head && __predict_true(!init))
822 		return 0;
823 	iru_init(&iru, rxq, 0 /* flid */);
824 	map = fl->ifl_sds.ifsd_map;
825 	refill_pidx = netmap_idx_k2n(kring, nm_i);
826 	/*
827 	 * IMPORTANT: we must leave one free slot in the ring,
828 	 * so move head back by one unit
829 	 */
830 	head = nm_prev(head, lim);
831 	nic_i = UINT_MAX;
832 	DBG_COUNTER_INC(fl_refills);
833 	while (nm_i != head) {
834 #if IFLIB_DEBUG_COUNTERS
835 		if (++rf_count == 9)
836 			DBG_COUNTER_INC(fl_refills_large);
837 #endif
838 		for (int tmp_pidx = 0; tmp_pidx < IFLIB_MAX_RX_REFRESH && nm_i != head; tmp_pidx++) {
839 			struct netmap_slot *slot = &ring->slot[nm_i];
840 			void *addr = PNMB(na, slot, &fl->ifl_bus_addrs[tmp_pidx]);
841 			uint32_t nic_i_dma = refill_pidx;
842 			nic_i = netmap_idx_k2n(kring, nm_i);
843 
844 			MPASS(tmp_pidx < IFLIB_MAX_RX_REFRESH);
845 
846 			if (addr == NETMAP_BUF_BASE(na)) /* bad buf */
847 			        return netmap_ring_reinit(kring);
848 
849 			fl->ifl_vm_addrs[tmp_pidx] = addr;
850 			if (__predict_false(init) && map) {
851 				netmap_load_map(na, fl->ifl_ifdi->idi_tag, map[nic_i], addr);
852 			} else if (map && (slot->flags & NS_BUF_CHANGED)) {
853 				/* buffer has changed, reload map */
854 				netmap_reload_map(na, fl->ifl_ifdi->idi_tag, map[nic_i], addr);
855 			}
856 			slot->flags &= ~NS_BUF_CHANGED;
857 
858 			nm_i = nm_next(nm_i, lim);
859 			fl->ifl_rxd_idxs[tmp_pidx] = nic_i = nm_next(nic_i, lim);
860 			if (nm_i != head && tmp_pidx < IFLIB_MAX_RX_REFRESH-1)
861 				continue;
862 
863 			iru.iru_pidx = refill_pidx;
864 			iru.iru_count = tmp_pidx+1;
865 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
866 
867 			refill_pidx = nic_i;
868 			if (map == NULL)
869 				continue;
870 
871 			for (int n = 0; n < iru.iru_count; n++) {
872 				bus_dmamap_sync(fl->ifl_ifdi->idi_tag, map[nic_i_dma],
873 						BUS_DMASYNC_PREREAD);
874 				/* XXX - change this to not use the netmap func*/
875 				nic_i_dma = nm_next(nic_i_dma, lim);
876 			}
877 		}
878 	}
879 	kring->nr_hwcur = head;
880 
881 	if (map)
882 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
883 				BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
884 	if (__predict_true(nic_i != UINT_MAX)) {
885 		ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id, nic_i);
886 		DBG_COUNTER_INC(rxd_flush);
887 	}
888 	return (0);
889 }
890 
891 /*
892  * Reconcile kernel and user view of the transmit ring.
893  *
894  * All information is in the kring.
895  * Userspace wants to send packets up to the one before kring->rhead,
896  * kernel knows kring->nr_hwcur is the first unsent packet.
897  *
898  * Here we push packets out (as many as possible), and possibly
899  * reclaim buffers from previously completed transmission.
900  *
901  * The caller (netmap) guarantees that there is only one instance
902  * running at any time. Any interference with other driver
903  * methods should be handled by the individual drivers.
904  */
905 static int
906 iflib_netmap_txsync(struct netmap_kring *kring, int flags)
907 {
908 	struct netmap_adapter *na = kring->na;
909 	struct ifnet *ifp = na->ifp;
910 	struct netmap_ring *ring = kring->ring;
911 	u_int nm_i;	/* index into the netmap kring */
912 	u_int nic_i;	/* index into the NIC ring */
913 	u_int n;
914 	u_int const lim = kring->nkr_num_slots - 1;
915 	u_int const head = kring->rhead;
916 	struct if_pkt_info pi;
917 
918 	/*
919 	 * interrupts on every tx packet are expensive so request
920 	 * them every half ring, or where NS_REPORT is set
921 	 */
922 	u_int report_frequency = kring->nkr_num_slots >> 1;
923 	/* device-specific */
924 	if_ctx_t ctx = ifp->if_softc;
925 	iflib_txq_t txq = &ctx->ifc_txqs[kring->ring_id];
926 
927 	bus_dmamap_sync(txq->ift_desc_tag, txq->ift_ifdi->idi_map,
928 	    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
929 
930 
931 	/*
932 	 * First part: process new packets to send.
933 	 * nm_i is the current index in the netmap kring,
934 	 * nic_i is the corresponding index in the NIC ring.
935 	 *
936 	 * If we have packets to send (nm_i != head)
937 	 * iterate over the netmap ring, fetch length and update
938 	 * the corresponding slot in the NIC ring. Some drivers also
939 	 * need to update the buffer's physical address in the NIC slot
940 	 * even NS_BUF_CHANGED is not set (PNMB computes the addresses).
941 	 *
942 	 * The netmap_reload_map() calls is especially expensive,
943 	 * even when (as in this case) the tag is 0, so do only
944 	 * when the buffer has actually changed.
945 	 *
946 	 * If possible do not set the report/intr bit on all slots,
947 	 * but only a few times per ring or when NS_REPORT is set.
948 	 *
949 	 * Finally, on 10G and faster drivers, it might be useful
950 	 * to prefetch the next slot and txr entry.
951 	 */
952 
953 	nm_i = kring->nr_hwcur;
954 	if (nm_i != head) {	/* we have new packets to send */
955 		pkt_info_zero(&pi);
956 		pi.ipi_segs = txq->ift_segs;
957 		pi.ipi_qsidx = kring->ring_id;
958 		nic_i = netmap_idx_k2n(kring, nm_i);
959 
960 		__builtin_prefetch(&ring->slot[nm_i]);
961 		__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i]);
962 		if (txq->ift_sds.ifsd_map)
963 			__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i]);
964 
965 		for (n = 0; nm_i != head; n++) {
966 			struct netmap_slot *slot = &ring->slot[nm_i];
967 			u_int len = slot->len;
968 			uint64_t paddr;
969 			void *addr = PNMB(na, slot, &paddr);
970 			int flags = (slot->flags & NS_REPORT ||
971 				nic_i == 0 || nic_i == report_frequency) ?
972 				IPI_TX_INTR : 0;
973 
974 			/* device-specific */
975 			pi.ipi_len = len;
976 			pi.ipi_segs[0].ds_addr = paddr;
977 			pi.ipi_segs[0].ds_len = len;
978 			pi.ipi_nsegs = 1;
979 			pi.ipi_ndescs = 0;
980 			pi.ipi_pidx = nic_i;
981 			pi.ipi_flags = flags;
982 
983 			/* Fill the slot in the NIC ring. */
984 			ctx->isc_txd_encap(ctx->ifc_softc, &pi);
985 			DBG_COUNTER_INC(tx_encap);
986 
987 			/* prefetch for next round */
988 			__builtin_prefetch(&ring->slot[nm_i + 1]);
989 			__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i + 1]);
990 			if (txq->ift_sds.ifsd_map) {
991 				__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i + 1]);
992 
993 				NM_CHECK_ADDR_LEN(na, addr, len);
994 
995 				if (slot->flags & NS_BUF_CHANGED) {
996 					/* buffer has changed, reload map */
997 					netmap_reload_map(na, txq->ift_desc_tag, txq->ift_sds.ifsd_map[nic_i], addr);
998 				}
999 				/* make sure changes to the buffer are synced */
1000 				bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_sds.ifsd_map[nic_i],
1001 						BUS_DMASYNC_PREWRITE);
1002 			}
1003 			slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED);
1004 			nm_i = nm_next(nm_i, lim);
1005 			nic_i = nm_next(nic_i, lim);
1006 		}
1007 		kring->nr_hwcur = nm_i;
1008 
1009 		/* synchronize the NIC ring */
1010 		bus_dmamap_sync(txq->ift_desc_tag, txq->ift_ifdi->idi_map,
1011 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1012 
1013 		/* (re)start the tx unit up to slot nic_i (excluded) */
1014 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, nic_i);
1015 	}
1016 
1017 	/*
1018 	 * Second part: reclaim buffers for completed transmissions.
1019 	 *
1020 	 * If there are unclaimed buffers, attempt to reclaim them.
1021 	 * If none are reclaimed, and TX IRQs are not in use, do an initial
1022 	 * minimal delay, then trigger the tx handler which will spin in the
1023 	 * group task queue.
1024 	 */
1025 	if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1026 		if (iflib_tx_credits_update(ctx, txq)) {
1027 			/* some tx completed, increment avail */
1028 			nic_i = txq->ift_cidx_processed;
1029 			kring->nr_hwtail = nm_prev(netmap_idx_n2k(kring, nic_i), lim);
1030 		}
1031 	}
1032 	if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ))
1033 		if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1034 			callout_reset_on(&txq->ift_timer, hz < 2000 ? 1 : hz / 1000,
1035 			    iflib_timer, txq, txq->ift_timer.c_cpu);
1036 	}
1037 	return (0);
1038 }
1039 
1040 /*
1041  * Reconcile kernel and user view of the receive ring.
1042  * Same as for the txsync, this routine must be efficient.
1043  * The caller guarantees a single invocations, but races against
1044  * the rest of the driver should be handled here.
1045  *
1046  * On call, kring->rhead is the first packet that userspace wants
1047  * to keep, and kring->rcur is the wakeup point.
1048  * The kernel has previously reported packets up to kring->rtail.
1049  *
1050  * If (flags & NAF_FORCE_READ) also check for incoming packets irrespective
1051  * of whether or not we received an interrupt.
1052  */
1053 static int
1054 iflib_netmap_rxsync(struct netmap_kring *kring, int flags)
1055 {
1056 	struct netmap_adapter *na = kring->na;
1057 	struct netmap_ring *ring = kring->ring;
1058 	uint32_t nm_i;	/* index into the netmap ring */
1059 	uint32_t nic_i;	/* index into the NIC ring */
1060 	u_int i, n;
1061 	u_int const lim = kring->nkr_num_slots - 1;
1062 	u_int const head = kring->rhead;
1063 	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
1064 	struct if_rxd_info ri;
1065 
1066 	struct ifnet *ifp = na->ifp;
1067 	if_ctx_t ctx = ifp->if_softc;
1068 	iflib_rxq_t rxq = &ctx->ifc_rxqs[kring->ring_id];
1069 	iflib_fl_t fl = rxq->ifr_fl;
1070 	if (head > lim)
1071 		return netmap_ring_reinit(kring);
1072 
1073 	/* XXX check sync modes */
1074 	for (i = 0, fl = rxq->ifr_fl; i < rxq->ifr_nfl; i++, fl++) {
1075 		if (fl->ifl_sds.ifsd_map == NULL)
1076 			continue;
1077 		bus_dmamap_sync(rxq->ifr_fl[i].ifl_desc_tag, fl->ifl_ifdi->idi_map,
1078 				BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1079 	}
1080 	/*
1081 	 * First part: import newly received packets.
1082 	 *
1083 	 * nm_i is the index of the next free slot in the netmap ring,
1084 	 * nic_i is the index of the next received packet in the NIC ring,
1085 	 * and they may differ in case if_init() has been called while
1086 	 * in netmap mode. For the receive ring we have
1087 	 *
1088 	 *	nic_i = rxr->next_check;
1089 	 *	nm_i = kring->nr_hwtail (previous)
1090 	 * and
1091 	 *	nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1092 	 *
1093 	 * rxr->next_check is set to 0 on a ring reinit
1094 	 */
1095 	if (netmap_no_pendintr || force_update) {
1096 		int crclen = iflib_crcstrip ? 0 : 4;
1097 		int error, avail;
1098 
1099 		for (i = 0; i < rxq->ifr_nfl; i++) {
1100 			fl = &rxq->ifr_fl[i];
1101 			nic_i = fl->ifl_cidx;
1102 			nm_i = netmap_idx_n2k(kring, nic_i);
1103 			avail = iflib_rxd_avail(ctx, rxq, nic_i, USHRT_MAX);
1104 			for (n = 0; avail > 0; n++, avail--) {
1105 				rxd_info_zero(&ri);
1106 				ri.iri_frags = rxq->ifr_frags;
1107 				ri.iri_qsidx = kring->ring_id;
1108 				ri.iri_ifp = ctx->ifc_ifp;
1109 				ri.iri_cidx = nic_i;
1110 
1111 				error = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
1112 				ring->slot[nm_i].len = error ? 0 : ri.iri_len - crclen;
1113 				ring->slot[nm_i].flags = 0;
1114 				bus_dmamap_sync(fl->ifl_ifdi->idi_tag,
1115 				    fl->ifl_sds.ifsd_map[nic_i], BUS_DMASYNC_POSTREAD);
1116 				nm_i = nm_next(nm_i, lim);
1117 				nic_i = nm_next(nic_i, lim);
1118 			}
1119 			if (n) { /* update the state variables */
1120 				if (netmap_no_pendintr && !force_update) {
1121 					/* diagnostics */
1122 					iflib_rx_miss ++;
1123 					iflib_rx_miss_bufs += n;
1124 				}
1125 				fl->ifl_cidx = nic_i;
1126 				kring->nr_hwtail = nm_i;
1127 			}
1128 			kring->nr_kflags &= ~NKR_PENDINTR;
1129 		}
1130 	}
1131 	/*
1132 	 * Second part: skip past packets that userspace has released.
1133 	 * (kring->nr_hwcur to head excluded),
1134 	 * and make the buffers available for reception.
1135 	 * As usual nm_i is the index in the netmap ring,
1136 	 * nic_i is the index in the NIC ring, and
1137 	 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1138 	 */
1139 	/* XXX not sure how this will work with multiple free lists */
1140 	nm_i = kring->nr_hwcur;
1141 
1142 	return (netmap_fl_refill(rxq, kring, nm_i, false));
1143 }
1144 
1145 static void
1146 iflib_netmap_intr(struct netmap_adapter *na, int onoff)
1147 {
1148 	struct ifnet *ifp = na->ifp;
1149 	if_ctx_t ctx = ifp->if_softc;
1150 
1151 	CTX_LOCK(ctx);
1152 	if (onoff) {
1153 		IFDI_INTR_ENABLE(ctx);
1154 	} else {
1155 		IFDI_INTR_DISABLE(ctx);
1156 	}
1157 	CTX_UNLOCK(ctx);
1158 }
1159 
1160 
1161 static int
1162 iflib_netmap_attach(if_ctx_t ctx)
1163 {
1164 	struct netmap_adapter na;
1165 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1166 
1167 	bzero(&na, sizeof(na));
1168 
1169 	na.ifp = ctx->ifc_ifp;
1170 	na.na_flags = NAF_BDG_MAYSLEEP;
1171 	MPASS(ctx->ifc_softc_ctx.isc_ntxqsets);
1172 	MPASS(ctx->ifc_softc_ctx.isc_nrxqsets);
1173 
1174 	na.num_tx_desc = scctx->isc_ntxd[0];
1175 	na.num_rx_desc = scctx->isc_nrxd[0];
1176 	na.nm_txsync = iflib_netmap_txsync;
1177 	na.nm_rxsync = iflib_netmap_rxsync;
1178 	na.nm_register = iflib_netmap_register;
1179 	na.nm_intr = iflib_netmap_intr;
1180 	na.num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
1181 	na.num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
1182 	return (netmap_attach(&na));
1183 }
1184 
1185 static void
1186 iflib_netmap_txq_init(if_ctx_t ctx, iflib_txq_t txq)
1187 {
1188 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1189 	struct netmap_slot *slot;
1190 
1191 	slot = netmap_reset(na, NR_TX, txq->ift_id, 0);
1192 	if (slot == NULL)
1193 		return;
1194 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxd[0]; i++) {
1195 
1196 		/*
1197 		 * In netmap mode, set the map for the packet buffer.
1198 		 * NOTE: Some drivers (not this one) also need to set
1199 		 * the physical buffer address in the NIC ring.
1200 		 * netmap_idx_n2k() maps a nic index, i, into the corresponding
1201 		 * netmap slot index, si
1202 		 */
1203 		int si = netmap_idx_n2k(na->tx_rings[txq->ift_id], i);
1204 		netmap_load_map(na, txq->ift_desc_tag, txq->ift_sds.ifsd_map[i], NMB(na, slot + si));
1205 	}
1206 }
1207 
1208 static void
1209 iflib_netmap_rxq_init(if_ctx_t ctx, iflib_rxq_t rxq)
1210 {
1211 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1212 	struct netmap_kring *kring = na->rx_rings[rxq->ifr_id];
1213 	struct netmap_slot *slot;
1214 	uint32_t nm_i;
1215 
1216 	slot = netmap_reset(na, NR_RX, rxq->ifr_id, 0);
1217 	if (slot == NULL)
1218 		return;
1219 	nm_i = netmap_idx_n2k(kring, 0);
1220 	netmap_fl_refill(rxq, kring, nm_i, true);
1221 }
1222 
1223 static void
1224 iflib_netmap_timer_adjust(if_ctx_t ctx, uint16_t txqid, uint32_t *reset_on)
1225 {
1226 	struct netmap_kring *kring;
1227 
1228 	kring = NA(ctx->ifc_ifp)->tx_rings[txqid];
1229 
1230 	if (kring->nr_hwcur != nm_next(kring->nr_hwtail, kring->nkr_num_slots - 1)) {
1231 		if (ctx->isc_txd_credits_update(ctx->ifc_softc, txqid, false))
1232 			netmap_tx_irq(ctx->ifc_ifp, txqid);
1233 		if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ)) {
1234 			if (hz < 2000)
1235 				*reset_on = 1;
1236 			else
1237 				*reset_on = hz / 1000;
1238 		}
1239 	}
1240 }
1241 
1242 #define iflib_netmap_detach(ifp) netmap_detach(ifp)
1243 
1244 #else
1245 #define iflib_netmap_txq_init(ctx, txq)
1246 #define iflib_netmap_rxq_init(ctx, rxq)
1247 #define iflib_netmap_detach(ifp)
1248 
1249 #define iflib_netmap_attach(ctx) (0)
1250 #define netmap_rx_irq(ifp, qid, budget) (0)
1251 #define netmap_tx_irq(ifp, qid) do {} while (0)
1252 #define iflib_netmap_timer_adjust(ctx, txqid, reset_on)
1253 
1254 #endif
1255 
1256 #if defined(__i386__) || defined(__amd64__)
1257 static __inline void
1258 prefetch(void *x)
1259 {
1260 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1261 }
1262 static __inline void
1263 prefetch2cachelines(void *x)
1264 {
1265 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1266 #if (CACHE_LINE_SIZE < 128)
1267 	__asm volatile("prefetcht0 %0" :: "m" (*(((unsigned long *)x)+CACHE_LINE_SIZE/(sizeof(unsigned long)))));
1268 #endif
1269 }
1270 #else
1271 #define prefetch(x)
1272 #define prefetch2cachelines(x)
1273 #endif
1274 
1275 static void
1276 iflib_gen_mac(if_ctx_t ctx)
1277 {
1278 	struct thread *td;
1279 	MD5_CTX mdctx;
1280 	char uuid[HOSTUUIDLEN+1];
1281 	char buf[HOSTUUIDLEN+16];
1282 	uint8_t *mac;
1283 	unsigned char digest[16];
1284 
1285 	td = curthread;
1286 	mac = ctx->ifc_mac;
1287 	uuid[HOSTUUIDLEN] = 0;
1288 	bcopy(td->td_ucred->cr_prison->pr_hostuuid, uuid, HOSTUUIDLEN);
1289 	snprintf(buf, HOSTUUIDLEN+16, "%s-%s", uuid, device_get_nameunit(ctx->ifc_dev));
1290 	/*
1291 	 * Generate a pseudo-random, deterministic MAC
1292 	 * address based on the UUID and unit number.
1293 	 * The FreeBSD Foundation OUI of 58-9C-FC is used.
1294 	 */
1295 	MD5Init(&mdctx);
1296 	MD5Update(&mdctx, buf, strlen(buf));
1297 	MD5Final(digest, &mdctx);
1298 
1299 	mac[0] = 0x58;
1300 	mac[1] = 0x9C;
1301 	mac[2] = 0xFC;
1302 	mac[3] = digest[0];
1303 	mac[4] = digest[1];
1304 	mac[5] = digest[2];
1305 }
1306 
1307 static void
1308 iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid)
1309 {
1310 	iflib_fl_t fl;
1311 
1312 	fl = &rxq->ifr_fl[flid];
1313 	iru->iru_paddrs = fl->ifl_bus_addrs;
1314 	iru->iru_vaddrs = &fl->ifl_vm_addrs[0];
1315 	iru->iru_idxs = fl->ifl_rxd_idxs;
1316 	iru->iru_qsidx = rxq->ifr_id;
1317 	iru->iru_buf_size = fl->ifl_buf_size;
1318 	iru->iru_flidx = fl->ifl_id;
1319 }
1320 
1321 static void
1322 _iflib_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int err)
1323 {
1324 	if (err)
1325 		return;
1326 	*(bus_addr_t *) arg = segs[0].ds_addr;
1327 }
1328 
1329 int
1330 iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags)
1331 {
1332 	int err;
1333 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1334 	device_t dev = ctx->ifc_dev;
1335 
1336 	KASSERT(sctx->isc_q_align != 0, ("alignment value not initialized"));
1337 
1338 	err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1339 				sctx->isc_q_align, 0,	/* alignment, bounds */
1340 				BUS_SPACE_MAXADDR,	/* lowaddr */
1341 				BUS_SPACE_MAXADDR,	/* highaddr */
1342 				NULL, NULL,		/* filter, filterarg */
1343 				size,			/* maxsize */
1344 				1,			/* nsegments */
1345 				size,			/* maxsegsize */
1346 				BUS_DMA_ALLOCNOW,	/* flags */
1347 				NULL,			/* lockfunc */
1348 				NULL,			/* lockarg */
1349 				&dma->idi_tag);
1350 	if (err) {
1351 		device_printf(dev,
1352 		    "%s: bus_dma_tag_create failed: %d\n",
1353 		    __func__, err);
1354 		goto fail_0;
1355 	}
1356 
1357 	err = bus_dmamem_alloc(dma->idi_tag, (void**) &dma->idi_vaddr,
1358 	    BUS_DMA_NOWAIT | BUS_DMA_COHERENT | BUS_DMA_ZERO, &dma->idi_map);
1359 	if (err) {
1360 		device_printf(dev,
1361 		    "%s: bus_dmamem_alloc(%ju) failed: %d\n",
1362 		    __func__, (uintmax_t)size, err);
1363 		goto fail_1;
1364 	}
1365 
1366 	dma->idi_paddr = IF_BAD_DMA;
1367 	err = bus_dmamap_load(dma->idi_tag, dma->idi_map, dma->idi_vaddr,
1368 	    size, _iflib_dmamap_cb, &dma->idi_paddr, mapflags | BUS_DMA_NOWAIT);
1369 	if (err || dma->idi_paddr == IF_BAD_DMA) {
1370 		device_printf(dev,
1371 		    "%s: bus_dmamap_load failed: %d\n",
1372 		    __func__, err);
1373 		goto fail_2;
1374 	}
1375 
1376 	dma->idi_size = size;
1377 	return (0);
1378 
1379 fail_2:
1380 	bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1381 fail_1:
1382 	bus_dma_tag_destroy(dma->idi_tag);
1383 fail_0:
1384 	dma->idi_tag = NULL;
1385 
1386 	return (err);
1387 }
1388 
1389 int
1390 iflib_dma_alloc_multi(if_ctx_t ctx, int *sizes, iflib_dma_info_t *dmalist, int mapflags, int count)
1391 {
1392 	int i, err;
1393 	iflib_dma_info_t *dmaiter;
1394 
1395 	dmaiter = dmalist;
1396 	for (i = 0; i < count; i++, dmaiter++) {
1397 		if ((err = iflib_dma_alloc(ctx, sizes[i], *dmaiter, mapflags)) != 0)
1398 			break;
1399 	}
1400 	if (err)
1401 		iflib_dma_free_multi(dmalist, i);
1402 	return (err);
1403 }
1404 
1405 void
1406 iflib_dma_free(iflib_dma_info_t dma)
1407 {
1408 	if (dma->idi_tag == NULL)
1409 		return;
1410 	if (dma->idi_paddr != IF_BAD_DMA) {
1411 		bus_dmamap_sync(dma->idi_tag, dma->idi_map,
1412 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1413 		bus_dmamap_unload(dma->idi_tag, dma->idi_map);
1414 		dma->idi_paddr = IF_BAD_DMA;
1415 	}
1416 	if (dma->idi_vaddr != NULL) {
1417 		bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1418 		dma->idi_vaddr = NULL;
1419 	}
1420 	bus_dma_tag_destroy(dma->idi_tag);
1421 	dma->idi_tag = NULL;
1422 }
1423 
1424 void
1425 iflib_dma_free_multi(iflib_dma_info_t *dmalist, int count)
1426 {
1427 	int i;
1428 	iflib_dma_info_t *dmaiter = dmalist;
1429 
1430 	for (i = 0; i < count; i++, dmaiter++)
1431 		iflib_dma_free(*dmaiter);
1432 }
1433 
1434 #ifdef EARLY_AP_STARTUP
1435 static const int iflib_started = 1;
1436 #else
1437 /*
1438  * We used to abuse the smp_started flag to decide if the queues have been
1439  * fully initialized (by late taskqgroup_adjust() calls in a SYSINIT()).
1440  * That gave bad races, since the SYSINIT() runs strictly after smp_started
1441  * is set.  Run a SYSINIT() strictly after that to just set a usable
1442  * completion flag.
1443  */
1444 
1445 static int iflib_started;
1446 
1447 static void
1448 iflib_record_started(void *arg)
1449 {
1450 	iflib_started = 1;
1451 }
1452 
1453 SYSINIT(iflib_record_started, SI_SUB_SMP + 1, SI_ORDER_FIRST,
1454 	iflib_record_started, NULL);
1455 #endif
1456 
1457 static int
1458 iflib_fast_intr(void *arg)
1459 {
1460 	iflib_filter_info_t info = arg;
1461 	struct grouptask *gtask = info->ifi_task;
1462 	if (!iflib_started)
1463 		return (FILTER_HANDLED);
1464 
1465 	DBG_COUNTER_INC(fast_intrs);
1466 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1467 		return (FILTER_HANDLED);
1468 
1469 	GROUPTASK_ENQUEUE(gtask);
1470 	return (FILTER_HANDLED);
1471 }
1472 
1473 static int
1474 iflib_fast_intr_rxtx(void *arg)
1475 {
1476 	iflib_filter_info_t info = arg;
1477 	struct grouptask *gtask = info->ifi_task;
1478 	iflib_rxq_t rxq = (iflib_rxq_t)info->ifi_ctx;
1479 	if_ctx_t ctx = NULL;;
1480 	int i, cidx;
1481 
1482 	if (!iflib_started)
1483 		return (FILTER_HANDLED);
1484 
1485 	DBG_COUNTER_INC(fast_intrs);
1486 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1487 		return (FILTER_HANDLED);
1488 
1489 	MPASS(rxq->ifr_ntxqirq);
1490 	for (i = 0; i < rxq->ifr_ntxqirq; i++) {
1491 		qidx_t txqid = rxq->ifr_txqid[i];
1492 
1493 		ctx = rxq->ifr_ctx;
1494 
1495 		bus_dmamap_sync(rxq->ifr_ifdi->idi_tag, rxq->ifr_ifdi->idi_map,
1496 		    BUS_DMASYNC_POSTREAD);
1497 		if (!ctx->isc_txd_credits_update(ctx->ifc_softc, txqid, false)) {
1498 			IFDI_TX_QUEUE_INTR_ENABLE(ctx, txqid);
1499 			continue;
1500 		}
1501 		GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
1502 	}
1503 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_RXCQ)
1504 		cidx = rxq->ifr_cq_cidx;
1505 	else
1506 		cidx = rxq->ifr_fl[0].ifl_cidx;
1507 	if (iflib_rxd_avail(ctx, rxq, cidx, 1))
1508 		GROUPTASK_ENQUEUE(gtask);
1509 	else {
1510 		IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
1511 		DBG_COUNTER_INC(rx_intr_enables);
1512 	}
1513 	return (FILTER_HANDLED);
1514 }
1515 
1516 
1517 static int
1518 iflib_fast_intr_ctx(void *arg)
1519 {
1520 	iflib_filter_info_t info = arg;
1521 	struct grouptask *gtask = info->ifi_task;
1522 
1523 	if (!iflib_started)
1524 		return (FILTER_HANDLED);
1525 
1526 	DBG_COUNTER_INC(fast_intrs);
1527 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1528 		return (FILTER_HANDLED);
1529 
1530 	GROUPTASK_ENQUEUE(gtask);
1531 	return (FILTER_HANDLED);
1532 }
1533 
1534 static int
1535 _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
1536 		 driver_filter_t filter, driver_intr_t handler, void *arg,
1537 		 const char *name)
1538 {
1539 	int rc, flags;
1540 	struct resource *res;
1541 	void *tag = NULL;
1542 	device_t dev = ctx->ifc_dev;
1543 
1544 	flags = RF_ACTIVE;
1545 	if (ctx->ifc_flags & IFC_LEGACY)
1546 		flags |= RF_SHAREABLE;
1547 	MPASS(rid < 512);
1548 	irq->ii_rid = rid;
1549 	res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &irq->ii_rid, flags);
1550 	if (res == NULL) {
1551 		device_printf(dev,
1552 		    "failed to allocate IRQ for rid %d, name %s.\n", rid, name);
1553 		return (ENOMEM);
1554 	}
1555 	irq->ii_res = res;
1556 	KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL"));
1557 	rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET,
1558 						filter, handler, arg, &tag);
1559 	if (rc != 0) {
1560 		device_printf(dev,
1561 		    "failed to setup interrupt for rid %d, name %s: %d\n",
1562 					  rid, name ? name : "unknown", rc);
1563 		return (rc);
1564 	} else if (name)
1565 		bus_describe_intr(dev, res, tag, "%s", name);
1566 
1567 	irq->ii_tag = tag;
1568 	return (0);
1569 }
1570 
1571 
1572 /*********************************************************************
1573  *
1574  *  Allocate memory for tx_buffer structures. The tx_buffer stores all
1575  *  the information needed to transmit a packet on the wire. This is
1576  *  called only once at attach, setup is done every reset.
1577  *
1578  **********************************************************************/
1579 
1580 static int
1581 iflib_txsd_alloc(iflib_txq_t txq)
1582 {
1583 	if_ctx_t ctx = txq->ift_ctx;
1584 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1585 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1586 	device_t dev = ctx->ifc_dev;
1587 	bus_size_t tsomaxsize;
1588 	int err, nsegments, ntsosegments;
1589 	bool tso;
1590 
1591 	nsegments = scctx->isc_tx_nsegments;
1592 	ntsosegments = scctx->isc_tx_tso_segments_max;
1593 	tsomaxsize = scctx->isc_tx_tso_size_max;
1594 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_VLAN_MTU)
1595 		tsomaxsize += sizeof(struct ether_vlan_header);
1596 	MPASS(scctx->isc_ntxd[0] > 0);
1597 	MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0);
1598 	MPASS(nsegments > 0);
1599 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) {
1600 		MPASS(ntsosegments > 0);
1601 		MPASS(sctx->isc_tso_maxsize >= tsomaxsize);
1602 	}
1603 
1604 	/*
1605 	 * Setup DMA descriptor areas.
1606 	 */
1607 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1608 			       1, 0,			/* alignment, bounds */
1609 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1610 			       BUS_SPACE_MAXADDR,	/* highaddr */
1611 			       NULL, NULL,		/* filter, filterarg */
1612 			       sctx->isc_tx_maxsize,		/* maxsize */
1613 			       nsegments,	/* nsegments */
1614 			       sctx->isc_tx_maxsegsize,	/* maxsegsize */
1615 			       0,			/* flags */
1616 			       NULL,			/* lockfunc */
1617 			       NULL,			/* lockfuncarg */
1618 			       &txq->ift_desc_tag))) {
1619 		device_printf(dev,"Unable to allocate TX DMA tag: %d\n", err);
1620 		device_printf(dev,"maxsize: %ju nsegments: %d maxsegsize: %ju\n",
1621 		    (uintmax_t)sctx->isc_tx_maxsize, nsegments, (uintmax_t)sctx->isc_tx_maxsegsize);
1622 		goto fail;
1623 	}
1624 	tso = (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) != 0;
1625 	if (tso && (err = bus_dma_tag_create(bus_get_dma_tag(dev),
1626 			       1, 0,			/* alignment, bounds */
1627 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1628 			       BUS_SPACE_MAXADDR,	/* highaddr */
1629 			       NULL, NULL,		/* filter, filterarg */
1630 			       tsomaxsize,		/* maxsize */
1631 			       ntsosegments,	/* nsegments */
1632 			       sctx->isc_tso_maxsegsize,/* maxsegsize */
1633 			       0,			/* flags */
1634 			       NULL,			/* lockfunc */
1635 			       NULL,			/* lockfuncarg */
1636 			       &txq->ift_tso_desc_tag))) {
1637 		device_printf(dev,"Unable to allocate TX TSO DMA tag: %d\n", err);
1638 		goto fail;
1639 	}
1640 	if (!(txq->ift_sds.ifsd_m =
1641 	    (struct mbuf **) malloc(sizeof(struct mbuf *) *
1642 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1643 		device_printf(dev, "Unable to allocate tx_buffer memory\n");
1644 		err = ENOMEM;
1645 		goto fail;
1646 	}
1647 
1648         /* Create the descriptor buffer dma maps */
1649 	if ((txq->ift_sds.ifsd_map = (bus_dmamap_t *)malloc(
1650 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1651 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1652 		device_printf(dev, "Unable to allocate tx_buffer map memory\n");
1653 		err = ENOMEM;
1654 		goto fail;
1655 	}
1656 
1657 	if (tso && (txq->ift_sds.ifsd_tso_map = (bus_dmamap_t *)malloc(
1658 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1659 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1660 		device_printf(dev, "Unable to allocate TSO tx_buffer "
1661 		    "map memory\n");
1662 		err = ENOMEM;
1663 		goto fail;
1664 	}
1665 
1666 	for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) {
1667 		err = bus_dmamap_create(txq->ift_desc_tag, 0,
1668 		    &txq->ift_sds.ifsd_map[i]);
1669 		if (err != 0) {
1670 			device_printf(dev, "Unable to create TX DMA map\n");
1671 			goto fail;
1672 		}
1673 		if (!tso)
1674 			continue;
1675 		err = bus_dmamap_create(txq->ift_tso_desc_tag, 0,
1676 		    &txq->ift_sds.ifsd_tso_map[i]);
1677 		if (err != 0) {
1678 			device_printf(dev, "Unable to create TSO TX DMA map\n");
1679 			goto fail;
1680 		}
1681 	}
1682 	return (0);
1683 fail:
1684 	/* We free all, it handles case where we are in the middle */
1685 	iflib_tx_structures_free(ctx);
1686 	return (err);
1687 }
1688 
1689 static void
1690 iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i)
1691 {
1692 	bus_dmamap_t map;
1693 
1694 	map = NULL;
1695 	if (txq->ift_sds.ifsd_map != NULL)
1696 		map = txq->ift_sds.ifsd_map[i];
1697 	if (map != NULL) {
1698 		bus_dmamap_sync(txq->ift_desc_tag, map, BUS_DMASYNC_POSTWRITE);
1699 		bus_dmamap_unload(txq->ift_desc_tag, map);
1700 		bus_dmamap_destroy(txq->ift_desc_tag, map);
1701 		txq->ift_sds.ifsd_map[i] = NULL;
1702 	}
1703 
1704 	map = NULL;
1705 	if (txq->ift_sds.ifsd_tso_map != NULL)
1706 		map = txq->ift_sds.ifsd_tso_map[i];
1707 	if (map != NULL) {
1708 		bus_dmamap_sync(txq->ift_tso_desc_tag, map,
1709 		    BUS_DMASYNC_POSTWRITE);
1710 		bus_dmamap_unload(txq->ift_tso_desc_tag, map);
1711 		bus_dmamap_destroy(txq->ift_tso_desc_tag, map);
1712 		txq->ift_sds.ifsd_tso_map[i] = NULL;
1713 	}
1714 }
1715 
1716 static void
1717 iflib_txq_destroy(iflib_txq_t txq)
1718 {
1719 	if_ctx_t ctx = txq->ift_ctx;
1720 
1721 	for (int i = 0; i < txq->ift_size; i++)
1722 		iflib_txsd_destroy(ctx, txq, i);
1723 	if (txq->ift_sds.ifsd_map != NULL) {
1724 		free(txq->ift_sds.ifsd_map, M_IFLIB);
1725 		txq->ift_sds.ifsd_map = NULL;
1726 	}
1727 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1728 		free(txq->ift_sds.ifsd_tso_map, M_IFLIB);
1729 		txq->ift_sds.ifsd_tso_map = NULL;
1730 	}
1731 	if (txq->ift_sds.ifsd_m != NULL) {
1732 		free(txq->ift_sds.ifsd_m, M_IFLIB);
1733 		txq->ift_sds.ifsd_m = NULL;
1734 	}
1735 	if (txq->ift_desc_tag != NULL) {
1736 		bus_dma_tag_destroy(txq->ift_desc_tag);
1737 		txq->ift_desc_tag = NULL;
1738 	}
1739 	if (txq->ift_tso_desc_tag != NULL) {
1740 		bus_dma_tag_destroy(txq->ift_tso_desc_tag);
1741 		txq->ift_tso_desc_tag = NULL;
1742 	}
1743 }
1744 
1745 static void
1746 iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i)
1747 {
1748 	struct mbuf **mp;
1749 
1750 	mp = &txq->ift_sds.ifsd_m[i];
1751 	if (*mp == NULL)
1752 		return;
1753 
1754 	if (txq->ift_sds.ifsd_map != NULL) {
1755 		bus_dmamap_sync(txq->ift_desc_tag,
1756 		    txq->ift_sds.ifsd_map[i], BUS_DMASYNC_POSTWRITE);
1757 		bus_dmamap_unload(txq->ift_desc_tag, txq->ift_sds.ifsd_map[i]);
1758 	}
1759 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1760 		bus_dmamap_sync(txq->ift_tso_desc_tag,
1761 		    txq->ift_sds.ifsd_tso_map[i], BUS_DMASYNC_POSTWRITE);
1762 		bus_dmamap_unload(txq->ift_tso_desc_tag,
1763 		    txq->ift_sds.ifsd_tso_map[i]);
1764 	}
1765 	m_free(*mp);
1766 	DBG_COUNTER_INC(tx_frees);
1767 	*mp = NULL;
1768 }
1769 
1770 static int
1771 iflib_txq_setup(iflib_txq_t txq)
1772 {
1773 	if_ctx_t ctx = txq->ift_ctx;
1774 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1775 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1776 	iflib_dma_info_t di;
1777 	int i;
1778 
1779 	/* Set number of descriptors available */
1780 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
1781 	/* XXX make configurable */
1782 	txq->ift_update_freq = IFLIB_DEFAULT_TX_UPDATE_FREQ;
1783 
1784 	/* Reset indices */
1785 	txq->ift_cidx_processed = 0;
1786 	txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0;
1787 	txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset];
1788 
1789 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1790 		bzero((void *)di->idi_vaddr, di->idi_size);
1791 
1792 	IFDI_TXQ_SETUP(ctx, txq->ift_id);
1793 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1794 		bus_dmamap_sync(di->idi_tag, di->idi_map,
1795 						BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1796 	return (0);
1797 }
1798 
1799 /*********************************************************************
1800  *
1801  *  Allocate memory for rx_buffer structures. Since we use one
1802  *  rx_buffer per received packet, the maximum number of rx_buffer's
1803  *  that we'll need is equal to the number of receive descriptors
1804  *  that we've allocated.
1805  *
1806  **********************************************************************/
1807 static int
1808 iflib_rxsd_alloc(iflib_rxq_t rxq)
1809 {
1810 	if_ctx_t ctx = rxq->ifr_ctx;
1811 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1812 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1813 	device_t dev = ctx->ifc_dev;
1814 	iflib_fl_t fl;
1815 	int			err;
1816 
1817 	MPASS(scctx->isc_nrxd[0] > 0);
1818 	MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0);
1819 
1820 	fl = rxq->ifr_fl;
1821 	for (int i = 0; i <  rxq->ifr_nfl; i++, fl++) {
1822 		fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */
1823 		err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1824 					 1, 0,			/* alignment, bounds */
1825 					 BUS_SPACE_MAXADDR,	/* lowaddr */
1826 					 BUS_SPACE_MAXADDR,	/* highaddr */
1827 					 NULL, NULL,		/* filter, filterarg */
1828 					 sctx->isc_rx_maxsize,	/* maxsize */
1829 					 sctx->isc_rx_nsegments,	/* nsegments */
1830 					 sctx->isc_rx_maxsegsize,	/* maxsegsize */
1831 					 0,			/* flags */
1832 					 NULL,			/* lockfunc */
1833 					 NULL,			/* lockarg */
1834 					 &fl->ifl_desc_tag);
1835 		if (err) {
1836 			device_printf(dev, "%s: bus_dma_tag_create failed %d\n",
1837 				__func__, err);
1838 			goto fail;
1839 		}
1840 		if (!(fl->ifl_sds.ifsd_m =
1841 		      (struct mbuf **) malloc(sizeof(struct mbuf *) *
1842 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1843 			device_printf(dev, "Unable to allocate tx_buffer memory\n");
1844 			err = ENOMEM;
1845 			goto fail;
1846 		}
1847 		if (!(fl->ifl_sds.ifsd_cl =
1848 		      (caddr_t *) malloc(sizeof(caddr_t) *
1849 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1850 			device_printf(dev, "Unable to allocate tx_buffer memory\n");
1851 			err = ENOMEM;
1852 			goto fail;
1853 		}
1854 
1855 		if (!(fl->ifl_sds.ifsd_ba =
1856 		      (bus_addr_t *) malloc(sizeof(bus_addr_t) *
1857 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1858 			device_printf(dev, "Unable to allocate rx bus addr memory\n");
1859 			err = ENOMEM;
1860 			goto fail;
1861 		}
1862 
1863 		/* Create the descriptor buffer dma maps */
1864 		if (!(fl->ifl_sds.ifsd_map =
1865 		      (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1866 			device_printf(dev, "Unable to allocate tx_buffer map memory\n");
1867 			err = ENOMEM;
1868 			goto fail;
1869 		}
1870 
1871 		for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++) {
1872 			err = bus_dmamap_create(fl->ifl_desc_tag, 0, &fl->ifl_sds.ifsd_map[i]);
1873 			if (err != 0) {
1874 				device_printf(dev, "Unable to create RX buffer DMA map\n");
1875 				goto fail;
1876 			}
1877 		}
1878 	}
1879 	return (0);
1880 
1881 fail:
1882 	iflib_rx_structures_free(ctx);
1883 	return (err);
1884 }
1885 
1886 
1887 /*
1888  * Internal service routines
1889  */
1890 
1891 struct rxq_refill_cb_arg {
1892 	int               error;
1893 	bus_dma_segment_t seg;
1894 	int               nseg;
1895 };
1896 
1897 static void
1898 _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
1899 {
1900 	struct rxq_refill_cb_arg *cb_arg = arg;
1901 
1902 	cb_arg->error = error;
1903 	cb_arg->seg = segs[0];
1904 	cb_arg->nseg = nseg;
1905 }
1906 
1907 /**
1908  *	rxq_refill - refill an rxq  free-buffer list
1909  *	@ctx: the iflib context
1910  *	@rxq: the free-list to refill
1911  *	@n: the number of new buffers to allocate
1912  *
1913  *	(Re)populate an rxq free-buffer list with up to @n new packet buffers.
1914  *	The caller must assure that @n does not exceed the queue's capacity.
1915  */
1916 static void
1917 _iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
1918 {
1919 	struct mbuf *m;
1920 	int idx, frag_idx = fl->ifl_fragidx;
1921         int pidx = fl->ifl_pidx;
1922 	caddr_t cl, *sd_cl;
1923 	struct mbuf **sd_m;
1924 	struct if_rxd_update iru;
1925 	struct rxq_refill_cb_arg cb_arg;
1926 	bus_dmamap_t *sd_map;
1927 	int n, i = 0;
1928 	bus_addr_t bus_addr, *sd_ba;
1929 	int err;
1930 	qidx_t credits;
1931 
1932 	sd_m = fl->ifl_sds.ifsd_m;
1933 	sd_map = fl->ifl_sds.ifsd_map;
1934 	sd_cl = fl->ifl_sds.ifsd_cl;
1935 	sd_ba = fl->ifl_sds.ifsd_ba;
1936 	idx = pidx;
1937 	credits = fl->ifl_credits;
1938 
1939 	n  = count;
1940 	MPASS(n > 0);
1941 	MPASS(credits + n <= fl->ifl_size);
1942 
1943 	if (pidx < fl->ifl_cidx)
1944 		MPASS(pidx + n <= fl->ifl_cidx);
1945 	if (pidx == fl->ifl_cidx && (credits < fl->ifl_size))
1946 		MPASS(fl->ifl_gen == 0);
1947 	if (pidx > fl->ifl_cidx)
1948 		MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx);
1949 
1950 	DBG_COUNTER_INC(fl_refills);
1951 	if (n > 8)
1952 		DBG_COUNTER_INC(fl_refills_large);
1953 	iru_init(&iru, fl->ifl_rxq, fl->ifl_id);
1954 	while (n--) {
1955 		/*
1956 		 * We allocate an uninitialized mbuf + cluster, mbuf is
1957 		 * initialized after rx.
1958 		 *
1959 		 * If the cluster is still set then we know a minimum sized packet was received
1960 		 */
1961 		bit_ffc_at(fl->ifl_rx_bitmap, frag_idx, fl->ifl_size,  &frag_idx);
1962 		if ((frag_idx < 0) || (frag_idx >= fl->ifl_size))
1963                 	bit_ffc(fl->ifl_rx_bitmap, fl->ifl_size, &frag_idx);
1964 		if ((cl = sd_cl[frag_idx]) == NULL) {
1965 			if ((cl = m_cljget(NULL, M_NOWAIT, fl->ifl_buf_size)) == NULL)
1966 				break;
1967 
1968 			cb_arg.error = 0;
1969 			MPASS(sd_map != NULL);
1970 			err = bus_dmamap_load(fl->ifl_desc_tag, sd_map[frag_idx],
1971 			    cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg,
1972 			    BUS_DMA_NOWAIT);
1973 			if (err != 0 || cb_arg.error) {
1974 				/*
1975 				 * !zone_pack ?
1976 				 */
1977 				if (fl->ifl_zone == zone_pack)
1978 					uma_zfree(fl->ifl_zone, cl);
1979 				break;
1980 			}
1981 
1982 			bus_dmamap_sync(fl->ifl_desc_tag, sd_map[frag_idx],
1983 			    BUS_DMASYNC_PREREAD);
1984 			sd_ba[frag_idx] =  bus_addr = cb_arg.seg.ds_addr;
1985 			sd_cl[frag_idx] = cl;
1986 #if MEMORY_LOGGING
1987 			fl->ifl_cl_enqueued++;
1988 #endif
1989 		} else {
1990 			bus_addr = sd_ba[frag_idx];
1991 		}
1992 
1993 		bit_set(fl->ifl_rx_bitmap, frag_idx);
1994 		MPASS(sd_m[frag_idx] == NULL);
1995 		if ((m = m_gethdr(M_NOWAIT, MT_NOINIT)) == NULL) {
1996 			break;
1997 		}
1998 		sd_m[frag_idx] = m;
1999 #if MEMORY_LOGGING
2000 		fl->ifl_m_enqueued++;
2001 #endif
2002 
2003 		DBG_COUNTER_INC(rx_allocs);
2004 		fl->ifl_rxd_idxs[i] = frag_idx;
2005 		fl->ifl_bus_addrs[i] = bus_addr;
2006 		fl->ifl_vm_addrs[i] = cl;
2007 		credits++;
2008 		i++;
2009 		MPASS(credits <= fl->ifl_size);
2010 		if (++idx == fl->ifl_size) {
2011 			fl->ifl_gen = 1;
2012 			idx = 0;
2013 		}
2014 		if (n == 0 || i == IFLIB_MAX_RX_REFRESH) {
2015 			iru.iru_pidx = pidx;
2016 			iru.iru_count = i;
2017 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2018 			i = 0;
2019 			pidx = idx;
2020 			fl->ifl_pidx = idx;
2021 			fl->ifl_credits = credits;
2022 		}
2023 
2024 	}
2025 
2026 	if (i) {
2027 		iru.iru_pidx = pidx;
2028 		iru.iru_count = i;
2029 		ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2030 		fl->ifl_pidx = idx;
2031 		fl->ifl_credits = credits;
2032 	}
2033 	DBG_COUNTER_INC(rxd_flush);
2034 	if (fl->ifl_pidx == 0)
2035 		pidx = fl->ifl_size - 1;
2036 	else
2037 		pidx = fl->ifl_pidx - 1;
2038 
2039 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2040 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2041 	ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id, fl->ifl_id, pidx);
2042 	fl->ifl_fragidx = frag_idx;
2043 }
2044 
2045 static __inline void
2046 __iflib_fl_refill_lt(if_ctx_t ctx, iflib_fl_t fl, int max)
2047 {
2048 	/* we avoid allowing pidx to catch up with cidx as it confuses ixl */
2049 	int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1;
2050 #ifdef INVARIANTS
2051 	int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1;
2052 #endif
2053 
2054 	MPASS(fl->ifl_credits <= fl->ifl_size);
2055 	MPASS(reclaimable == delta);
2056 
2057 	if (reclaimable > 0)
2058 		_iflib_fl_refill(ctx, fl, min(max, reclaimable));
2059 }
2060 
2061 uint8_t
2062 iflib_in_detach(if_ctx_t ctx)
2063 {
2064 	bool in_detach;
2065 	STATE_LOCK(ctx);
2066 	in_detach = !!(ctx->ifc_flags & IFC_IN_DETACH);
2067 	STATE_UNLOCK(ctx);
2068 	return (in_detach);
2069 }
2070 
2071 static void
2072 iflib_fl_bufs_free(iflib_fl_t fl)
2073 {
2074 	iflib_dma_info_t idi = fl->ifl_ifdi;
2075 	bus_dmamap_t sd_map;
2076 	uint32_t i;
2077 
2078 	for (i = 0; i < fl->ifl_size; i++) {
2079 		struct mbuf **sd_m = &fl->ifl_sds.ifsd_m[i];
2080 		caddr_t *sd_cl = &fl->ifl_sds.ifsd_cl[i];
2081 
2082 		if (*sd_cl != NULL) {
2083 			sd_map = fl->ifl_sds.ifsd_map[i];
2084 			bus_dmamap_sync(fl->ifl_desc_tag, sd_map,
2085 			    BUS_DMASYNC_POSTREAD);
2086 			bus_dmamap_unload(fl->ifl_desc_tag, sd_map);
2087 			if (*sd_cl != NULL)
2088 				uma_zfree(fl->ifl_zone, *sd_cl);
2089 			// XXX: Should this get moved out?
2090 			if (iflib_in_detach(fl->ifl_rxq->ifr_ctx))
2091 				bus_dmamap_destroy(fl->ifl_desc_tag, sd_map);
2092 			if (*sd_m != NULL) {
2093 				m_init(*sd_m, M_NOWAIT, MT_DATA, 0);
2094 				uma_zfree(zone_mbuf, *sd_m);
2095 			}
2096 		} else {
2097 			MPASS(*sd_cl == NULL);
2098 			MPASS(*sd_m == NULL);
2099 		}
2100 #if MEMORY_LOGGING
2101 		fl->ifl_m_dequeued++;
2102 		fl->ifl_cl_dequeued++;
2103 #endif
2104 		*sd_cl = NULL;
2105 		*sd_m = NULL;
2106 	}
2107 #ifdef INVARIANTS
2108 	for (i = 0; i < fl->ifl_size; i++) {
2109 		MPASS(fl->ifl_sds.ifsd_cl[i] == NULL);
2110 		MPASS(fl->ifl_sds.ifsd_m[i] == NULL);
2111 	}
2112 #endif
2113 	/*
2114 	 * Reset free list values
2115 	 */
2116 	fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = fl->ifl_fragidx = 0;
2117 	bzero(idi->idi_vaddr, idi->idi_size);
2118 }
2119 
2120 /*********************************************************************
2121  *
2122  *  Initialize a receive ring and its buffers.
2123  *
2124  **********************************************************************/
2125 static int
2126 iflib_fl_setup(iflib_fl_t fl)
2127 {
2128 	iflib_rxq_t rxq = fl->ifl_rxq;
2129 	if_ctx_t ctx = rxq->ifr_ctx;
2130 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2131 
2132 	bit_nclear(fl->ifl_rx_bitmap, 0, fl->ifl_size - 1);
2133 	/*
2134 	** Free current RX buffer structs and their mbufs
2135 	*/
2136 	iflib_fl_bufs_free(fl);
2137 	/* Now replenish the mbufs */
2138 	MPASS(fl->ifl_credits == 0);
2139 	/*
2140 	 * XXX don't set the max_frame_size to larger
2141 	 * than the hardware can handle
2142 	 */
2143 	if (sctx->isc_max_frame_size <= 2048)
2144 		fl->ifl_buf_size = MCLBYTES;
2145 #ifndef CONTIGMALLOC_WORKS
2146 	else
2147 		fl->ifl_buf_size = MJUMPAGESIZE;
2148 #else
2149 	else if (sctx->isc_max_frame_size <= 4096)
2150 		fl->ifl_buf_size = MJUMPAGESIZE;
2151 	else if (sctx->isc_max_frame_size <= 9216)
2152 		fl->ifl_buf_size = MJUM9BYTES;
2153 	else
2154 		fl->ifl_buf_size = MJUM16BYTES;
2155 #endif
2156 	if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size)
2157 		ctx->ifc_max_fl_buf_size = fl->ifl_buf_size;
2158 	fl->ifl_cltype = m_gettype(fl->ifl_buf_size);
2159 	fl->ifl_zone = m_getzone(fl->ifl_buf_size);
2160 
2161 
2162 	/* avoid pre-allocating zillions of clusters to an idle card
2163 	 * potentially speeding up attach
2164 	 */
2165 	_iflib_fl_refill(ctx, fl, min(128, fl->ifl_size));
2166 	MPASS(min(128, fl->ifl_size) == fl->ifl_credits);
2167 	if (min(128, fl->ifl_size) != fl->ifl_credits)
2168 		return (ENOBUFS);
2169 	/*
2170 	 * handle failure
2171 	 */
2172 	MPASS(rxq != NULL);
2173 	MPASS(fl->ifl_ifdi != NULL);
2174 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2175 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2176 	return (0);
2177 }
2178 
2179 /*********************************************************************
2180  *
2181  *  Free receive ring data structures
2182  *
2183  **********************************************************************/
2184 static void
2185 iflib_rx_sds_free(iflib_rxq_t rxq)
2186 {
2187 	iflib_fl_t fl;
2188 	int i, j;
2189 
2190 	if (rxq->ifr_fl != NULL) {
2191 		for (i = 0; i < rxq->ifr_nfl; i++) {
2192 			fl = &rxq->ifr_fl[i];
2193 			if (fl->ifl_desc_tag != NULL) {
2194 				if (fl->ifl_sds.ifsd_map != NULL) {
2195 					for (j = 0; j < fl->ifl_size; i++) {
2196 						if (fl->ifl_sds.ifsd_map[i] ==
2197 						    NULL)
2198 						continue;
2199 						bus_dmamap_sync(
2200 						    fl->ifl_desc_tag,
2201 						    fl->ifl_sds.ifsd_map[i],
2202 						    BUS_DMASYNC_POSTREAD);
2203 						bus_dmamap_unload(
2204 						    fl->ifl_desc_tag,
2205 						    fl->ifl_sds.ifsd_map[i]);
2206 					}
2207 				}
2208 				bus_dma_tag_destroy(fl->ifl_desc_tag);
2209 				fl->ifl_desc_tag = NULL;
2210 			}
2211 			free(fl->ifl_sds.ifsd_m, M_IFLIB);
2212 			free(fl->ifl_sds.ifsd_cl, M_IFLIB);
2213 			free(fl->ifl_sds.ifsd_ba, M_IFLIB);
2214 			free(fl->ifl_sds.ifsd_map, M_IFLIB);
2215 			fl->ifl_sds.ifsd_m = NULL;
2216 			fl->ifl_sds.ifsd_cl = NULL;
2217 			fl->ifl_sds.ifsd_ba = NULL;
2218 			fl->ifl_sds.ifsd_map = NULL;
2219 		}
2220 		free(rxq->ifr_fl, M_IFLIB);
2221 		rxq->ifr_fl = NULL;
2222 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
2223 	}
2224 }
2225 
2226 /*
2227  * MI independent logic
2228  *
2229  */
2230 static void
2231 iflib_timer(void *arg)
2232 {
2233 	iflib_txq_t txq = arg;
2234 	if_ctx_t ctx = txq->ift_ctx;
2235 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2236 	uint64_t this_tick = ticks;
2237 	uint32_t reset_on = hz / 2;
2238 
2239 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
2240 		return;
2241 	/*
2242 	** Check on the state of the TX queue(s), this
2243 	** can be done without the lock because its RO
2244 	** and the HUNG state will be static if set.
2245 	*/
2246 	if (this_tick - txq->ift_last_timer_tick >= hz / 2) {
2247 		txq->ift_last_timer_tick = this_tick;
2248 		IFDI_TIMER(ctx, txq->ift_id);
2249 		if ((txq->ift_qstatus == IFLIB_QUEUE_HUNG) &&
2250 		    ((txq->ift_cleaned_prev == txq->ift_cleaned) ||
2251 		     (sctx->isc_pause_frames == 0)))
2252 			goto hung;
2253 
2254 		if (ifmp_ring_is_stalled(txq->ift_br))
2255 			txq->ift_qstatus = IFLIB_QUEUE_HUNG;
2256 		txq->ift_cleaned_prev = txq->ift_cleaned;
2257 	}
2258 #ifdef DEV_NETMAP
2259 	if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP)
2260 		iflib_netmap_timer_adjust(ctx, txq->ift_id, &reset_on);
2261 #endif
2262 	/* handle any laggards */
2263 	if (txq->ift_db_pending)
2264 		GROUPTASK_ENQUEUE(&txq->ift_task);
2265 
2266 	sctx->isc_pause_frames = 0;
2267 	if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)
2268 		callout_reset_on(&txq->ift_timer, reset_on, iflib_timer, txq, txq->ift_timer.c_cpu);
2269 	return;
2270  hung:
2271 	device_printf(ctx->ifc_dev,  "TX(%d) desc avail = %d, pidx = %d\n",
2272 				  txq->ift_id, TXQ_AVAIL(txq), txq->ift_pidx);
2273 	STATE_LOCK(ctx);
2274 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2275 	ctx->ifc_flags |= (IFC_DO_WATCHDOG|IFC_DO_RESET);
2276 	iflib_admin_intr_deferred(ctx);
2277 	STATE_UNLOCK(ctx);
2278 }
2279 
2280 static void
2281 iflib_init_locked(if_ctx_t ctx)
2282 {
2283 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2284 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2285 	if_t ifp = ctx->ifc_ifp;
2286 	iflib_fl_t fl;
2287 	iflib_txq_t txq;
2288 	iflib_rxq_t rxq;
2289 	int i, j, tx_ip_csum_flags, tx_ip6_csum_flags;
2290 
2291 
2292 	if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2293 	IFDI_INTR_DISABLE(ctx);
2294 
2295 	tx_ip_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP);
2296 	tx_ip6_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | CSUM_IP6_SCTP);
2297 	/* Set hardware offload abilities */
2298 	if_clearhwassist(ifp);
2299 	if (if_getcapenable(ifp) & IFCAP_TXCSUM)
2300 		if_sethwassistbits(ifp, tx_ip_csum_flags, 0);
2301 	if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6)
2302 		if_sethwassistbits(ifp,  tx_ip6_csum_flags, 0);
2303 	if (if_getcapenable(ifp) & IFCAP_TSO4)
2304 		if_sethwassistbits(ifp, CSUM_IP_TSO, 0);
2305 	if (if_getcapenable(ifp) & IFCAP_TSO6)
2306 		if_sethwassistbits(ifp, CSUM_IP6_TSO, 0);
2307 
2308 	for (i = 0, txq = ctx->ifc_txqs; i < sctx->isc_ntxqsets; i++, txq++) {
2309 		CALLOUT_LOCK(txq);
2310 		callout_stop(&txq->ift_timer);
2311 		CALLOUT_UNLOCK(txq);
2312 		iflib_netmap_txq_init(ctx, txq);
2313 	}
2314 #ifdef INVARIANTS
2315 	i = if_getdrvflags(ifp);
2316 #endif
2317 	IFDI_INIT(ctx);
2318 	MPASS(if_getdrvflags(ifp) == i);
2319 	for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
2320 		/* XXX this should really be done on a per-queue basis */
2321 		if (if_getcapenable(ifp) & IFCAP_NETMAP) {
2322 			MPASS(rxq->ifr_id == i);
2323 			iflib_netmap_rxq_init(ctx, rxq);
2324 			continue;
2325 		}
2326 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
2327 			if (iflib_fl_setup(fl)) {
2328 				device_printf(ctx->ifc_dev, "freelist setup failed - check cluster settings\n");
2329 				goto done;
2330 			}
2331 		}
2332 	}
2333 done:
2334 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
2335 	IFDI_INTR_ENABLE(ctx);
2336 	txq = ctx->ifc_txqs;
2337 	for (i = 0; i < sctx->isc_ntxqsets; i++, txq++)
2338 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq,
2339 			txq->ift_timer.c_cpu);
2340 }
2341 
2342 static int
2343 iflib_media_change(if_t ifp)
2344 {
2345 	if_ctx_t ctx = if_getsoftc(ifp);
2346 	int err;
2347 
2348 	CTX_LOCK(ctx);
2349 	if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0)
2350 		iflib_init_locked(ctx);
2351 	CTX_UNLOCK(ctx);
2352 	return (err);
2353 }
2354 
2355 static void
2356 iflib_media_status(if_t ifp, struct ifmediareq *ifmr)
2357 {
2358 	if_ctx_t ctx = if_getsoftc(ifp);
2359 
2360 	CTX_LOCK(ctx);
2361 	IFDI_UPDATE_ADMIN_STATUS(ctx);
2362 	IFDI_MEDIA_STATUS(ctx, ifmr);
2363 	CTX_UNLOCK(ctx);
2364 }
2365 
2366 void
2367 iflib_stop(if_ctx_t ctx)
2368 {
2369 	iflib_txq_t txq = ctx->ifc_txqs;
2370 	iflib_rxq_t rxq = ctx->ifc_rxqs;
2371 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2372 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2373 	iflib_dma_info_t di;
2374 	iflib_fl_t fl;
2375 	int i, j;
2376 
2377 	/* Tell the stack that the interface is no longer active */
2378 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2379 
2380 	IFDI_INTR_DISABLE(ctx);
2381 	DELAY(1000);
2382 	IFDI_STOP(ctx);
2383 	DELAY(1000);
2384 
2385 	iflib_debug_reset();
2386 	/* Wait for current tx queue users to exit to disarm watchdog timer. */
2387 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) {
2388 		/* make sure all transmitters have completed before proceeding XXX */
2389 
2390 		CALLOUT_LOCK(txq);
2391 		callout_stop(&txq->ift_timer);
2392 		CALLOUT_UNLOCK(txq);
2393 
2394 		/* clean any enqueued buffers */
2395 		iflib_ifmp_purge(txq);
2396 		/* Free any existing tx buffers. */
2397 		for (j = 0; j < txq->ift_size; j++) {
2398 			iflib_txsd_free(ctx, txq, j);
2399 		}
2400 		txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0;
2401 		txq->ift_in_use = txq->ift_gen = txq->ift_cidx = txq->ift_pidx = txq->ift_no_desc_avail = 0;
2402 		txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0;
2403 		txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0;
2404 		txq->ift_pullups = 0;
2405 		ifmp_ring_reset_stats(txq->ift_br);
2406 		for (j = 0, di = txq->ift_ifdi; j < sctx->isc_ntxqs; j++, di++)
2407 			bzero((void *)di->idi_vaddr, di->idi_size);
2408 	}
2409 	for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) {
2410 		/* make sure all transmitters have completed before proceeding XXX */
2411 
2412 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
2413 		for (j = 0, di = rxq->ifr_ifdi; j < sctx->isc_nrxqs; j++, di++)
2414 			bzero((void *)di->idi_vaddr, di->idi_size);
2415 		/* also resets the free lists pidx/cidx */
2416 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
2417 			iflib_fl_bufs_free(fl);
2418 	}
2419 }
2420 
2421 static inline caddr_t
2422 calc_next_rxd(iflib_fl_t fl, int cidx)
2423 {
2424 	qidx_t size;
2425 	int nrxd;
2426 	caddr_t start, end, cur, next;
2427 
2428 	nrxd = fl->ifl_size;
2429 	size = fl->ifl_rxd_size;
2430 	start = fl->ifl_ifdi->idi_vaddr;
2431 
2432 	if (__predict_false(size == 0))
2433 		return (start);
2434 	cur = start + size*cidx;
2435 	end = start + size*nrxd;
2436 	next = CACHE_PTR_NEXT(cur);
2437 	return (next < end ? next : start);
2438 }
2439 
2440 static inline void
2441 prefetch_pkts(iflib_fl_t fl, int cidx)
2442 {
2443 	int nextptr;
2444 	int nrxd = fl->ifl_size;
2445 	caddr_t next_rxd;
2446 
2447 
2448 	nextptr = (cidx + CACHE_PTR_INCREMENT) & (nrxd-1);
2449 	prefetch(&fl->ifl_sds.ifsd_m[nextptr]);
2450 	prefetch(&fl->ifl_sds.ifsd_cl[nextptr]);
2451 	next_rxd = calc_next_rxd(fl, cidx);
2452 	prefetch(next_rxd);
2453 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 1) & (nrxd-1)]);
2454 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 2) & (nrxd-1)]);
2455 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 3) & (nrxd-1)]);
2456 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 4) & (nrxd-1)]);
2457 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 1) & (nrxd-1)]);
2458 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 2) & (nrxd-1)]);
2459 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 3) & (nrxd-1)]);
2460 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 4) & (nrxd-1)]);
2461 }
2462 
2463 static void
2464 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, int unload, if_rxsd_t sd)
2465 {
2466 	int flid, cidx;
2467 	bus_dmamap_t map;
2468 	iflib_fl_t fl;
2469 	iflib_dma_info_t di;
2470 	int next;
2471 
2472 	map = NULL;
2473 	flid = irf->irf_flid;
2474 	cidx = irf->irf_idx;
2475 	fl = &rxq->ifr_fl[flid];
2476 	sd->ifsd_fl = fl;
2477 	sd->ifsd_cidx = cidx;
2478 	sd->ifsd_m = &fl->ifl_sds.ifsd_m[cidx];
2479 	sd->ifsd_cl = &fl->ifl_sds.ifsd_cl[cidx];
2480 	fl->ifl_credits--;
2481 #if MEMORY_LOGGING
2482 	fl->ifl_m_dequeued++;
2483 #endif
2484 	if (rxq->ifr_ctx->ifc_flags & IFC_PREFETCH)
2485 		prefetch_pkts(fl, cidx);
2486 	next = (cidx + CACHE_PTR_INCREMENT) & (fl->ifl_size-1);
2487 	prefetch(&fl->ifl_sds.ifsd_map[next]);
2488 	map = fl->ifl_sds.ifsd_map[cidx];
2489 	di = fl->ifl_ifdi;
2490 	next = (cidx + CACHE_LINE_SIZE) & (fl->ifl_size-1);
2491 
2492 	/* not valid assert if bxe really does SGE from non-contiguous elements */
2493 	MPASS(fl->ifl_cidx == cidx);
2494 	bus_dmamap_sync(fl->ifl_desc_tag, map, BUS_DMASYNC_POSTREAD);
2495 	if (unload)
2496 		bus_dmamap_unload(fl->ifl_desc_tag, map);
2497 	fl->ifl_cidx = (fl->ifl_cidx + 1) & (fl->ifl_size-1);
2498 	if (__predict_false(fl->ifl_cidx == 0))
2499 		fl->ifl_gen = 0;
2500 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2501 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2502 	bit_clear(fl->ifl_rx_bitmap, cidx);
2503 }
2504 
2505 static struct mbuf *
2506 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri, if_rxsd_t sd)
2507 {
2508 	int i, padlen , flags;
2509 	struct mbuf *m, *mh, *mt;
2510 	caddr_t cl;
2511 
2512 	i = 0;
2513 	mh = NULL;
2514 	do {
2515 		rxd_frag_to_sd(rxq, &ri->iri_frags[i], TRUE, sd);
2516 
2517 		MPASS(*sd->ifsd_cl != NULL);
2518 		MPASS(*sd->ifsd_m != NULL);
2519 
2520 		/* Don't include zero-length frags */
2521 		if (ri->iri_frags[i].irf_len == 0) {
2522 			/* XXX we can save the cluster here, but not the mbuf */
2523 			m_init(*sd->ifsd_m, M_NOWAIT, MT_DATA, 0);
2524 			m_free(*sd->ifsd_m);
2525 			*sd->ifsd_m = NULL;
2526 			continue;
2527 		}
2528 		m = *sd->ifsd_m;
2529 		*sd->ifsd_m = NULL;
2530 		if (mh == NULL) {
2531 			flags = M_PKTHDR|M_EXT;
2532 			mh = mt = m;
2533 			padlen = ri->iri_pad;
2534 		} else {
2535 			flags = M_EXT;
2536 			mt->m_next = m;
2537 			mt = m;
2538 			/* assuming padding is only on the first fragment */
2539 			padlen = 0;
2540 		}
2541 		cl = *sd->ifsd_cl;
2542 		*sd->ifsd_cl = NULL;
2543 
2544 		/* Can these two be made one ? */
2545 		m_init(m, M_NOWAIT, MT_DATA, flags);
2546 		m_cljset(m, cl, sd->ifsd_fl->ifl_cltype);
2547 		/*
2548 		 * These must follow m_init and m_cljset
2549 		 */
2550 		m->m_data += padlen;
2551 		ri->iri_len -= padlen;
2552 		m->m_len = ri->iri_frags[i].irf_len;
2553 	} while (++i < ri->iri_nfrags);
2554 
2555 	return (mh);
2556 }
2557 
2558 /*
2559  * Process one software descriptor
2560  */
2561 static struct mbuf *
2562 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri)
2563 {
2564 	struct if_rxsd sd;
2565 	struct mbuf *m;
2566 
2567 	/* should I merge this back in now that the two paths are basically duplicated? */
2568 	if (ri->iri_nfrags == 1 &&
2569 	    ri->iri_frags[0].irf_len <= MIN(IFLIB_RX_COPY_THRESH, MHLEN)) {
2570 		rxd_frag_to_sd(rxq, &ri->iri_frags[0], FALSE, &sd);
2571 		m = *sd.ifsd_m;
2572 		*sd.ifsd_m = NULL;
2573 		m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR);
2574 #ifndef __NO_STRICT_ALIGNMENT
2575 		if (!IP_ALIGNED(m))
2576 			m->m_data += 2;
2577 #endif
2578 		memcpy(m->m_data, *sd.ifsd_cl, ri->iri_len);
2579 		bus_dmamap_sync(rxq->ifr_fl->ifl_desc_tag,
2580 		    rxq->ifr_fl->ifl_sds.ifsd_map[ri->iri_frags[0].irf_idx],
2581 		    BUS_DMASYNC_PREREAD);
2582 		m->m_len = ri->iri_frags[0].irf_len;
2583        } else {
2584 		m = assemble_segments(rxq, ri, &sd);
2585 	}
2586 	m->m_pkthdr.len = ri->iri_len;
2587 	m->m_pkthdr.rcvif = ri->iri_ifp;
2588 	m->m_flags |= ri->iri_flags;
2589 	m->m_pkthdr.ether_vtag = ri->iri_vtag;
2590 	m->m_pkthdr.flowid = ri->iri_flowid;
2591 	M_HASHTYPE_SET(m, ri->iri_rsstype);
2592 	m->m_pkthdr.csum_flags = ri->iri_csum_flags;
2593 	m->m_pkthdr.csum_data = ri->iri_csum_data;
2594 	return (m);
2595 }
2596 
2597 #if defined(INET6) || defined(INET)
2598 static void
2599 iflib_get_ip_forwarding(struct lro_ctrl *lc, bool *v4, bool *v6)
2600 {
2601 	CURVNET_SET(lc->ifp->if_vnet);
2602 #if defined(INET6)
2603 	*v6 = VNET(ip6_forwarding);
2604 #endif
2605 #if defined(INET)
2606 	*v4 = VNET(ipforwarding);
2607 #endif
2608 	CURVNET_RESTORE();
2609 }
2610 
2611 /*
2612  * Returns true if it's possible this packet could be LROed.
2613  * if it returns false, it is guaranteed that tcp_lro_rx()
2614  * would not return zero.
2615  */
2616 static bool
2617 iflib_check_lro_possible(struct mbuf *m, bool v4_forwarding, bool v6_forwarding)
2618 {
2619 	struct ether_header *eh;
2620 	uint16_t eh_type;
2621 
2622 	eh = mtod(m, struct ether_header *);
2623 	eh_type = ntohs(eh->ether_type);
2624 	switch (eh_type) {
2625 #if defined(INET6)
2626 		case ETHERTYPE_IPV6:
2627 			return !v6_forwarding;
2628 #endif
2629 #if defined (INET)
2630 		case ETHERTYPE_IP:
2631 			return !v4_forwarding;
2632 #endif
2633 	}
2634 
2635 	return false;
2636 }
2637 #else
2638 static void
2639 iflib_get_ip_forwarding(struct lro_ctrl *lc __unused, bool *v4 __unused, bool *v6 __unused)
2640 {
2641 }
2642 #endif
2643 
2644 static bool
2645 iflib_rxeof(iflib_rxq_t rxq, qidx_t budget)
2646 {
2647 	if_ctx_t ctx = rxq->ifr_ctx;
2648 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2649 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2650 	iflib_dma_info_t di;
2651 	int avail, i;
2652 	qidx_t *cidxp;
2653 	struct if_rxd_info ri;
2654 	int err, budget_left, rx_bytes, rx_pkts;
2655 	iflib_fl_t fl;
2656 	struct ifnet *ifp;
2657 	int lro_enabled;
2658 	bool v4_forwarding, v6_forwarding, lro_possible;
2659 
2660 	/*
2661 	 * XXX early demux data packets so that if_input processing only handles
2662 	 * acks in interrupt context
2663 	 */
2664 	struct mbuf *m, *mh, *mt, *mf;
2665 
2666 	lro_possible = v4_forwarding = v6_forwarding = false;
2667 	ifp = ctx->ifc_ifp;
2668 	mh = mt = NULL;
2669 	MPASS(budget > 0);
2670 	rx_pkts	= rx_bytes = 0;
2671 	if (sctx->isc_flags & IFLIB_HAS_RXCQ)
2672 		cidxp = &rxq->ifr_cq_cidx;
2673 	else
2674 		cidxp = &rxq->ifr_fl[0].ifl_cidx;
2675 	if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) {
2676 		for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2677 			__iflib_fl_refill_lt(ctx, fl, budget + 8);
2678 		DBG_COUNTER_INC(rx_unavail);
2679 		return (false);
2680 	}
2681 
2682 	for (budget_left = budget; budget_left > 0 && avail > 0;) {
2683 		if (__predict_false(!CTX_ACTIVE(ctx))) {
2684 			DBG_COUNTER_INC(rx_ctx_inactive);
2685 			break;
2686 		}
2687 		/*
2688 		 * Reset client set fields to their default values
2689 		 */
2690 		rxd_info_zero(&ri);
2691 		ri.iri_qsidx = rxq->ifr_id;
2692 		ri.iri_cidx = *cidxp;
2693 		ri.iri_ifp = ifp;
2694 		ri.iri_frags = rxq->ifr_frags;
2695 		di = rxq->ifr_fl[rxq->ifr_frags[0].irf_flid].ifl_ifdi;
2696 		bus_dmamap_sync(di->idi_tag, di->idi_map,
2697 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
2698 		err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
2699 
2700 		if (err)
2701 			goto err;
2702 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
2703 			*cidxp = ri.iri_cidx;
2704 			/* Update our consumer index */
2705 			/* XXX NB: shurd - check if this is still safe */
2706 			while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0]) {
2707 				rxq->ifr_cq_cidx -= scctx->isc_nrxd[0];
2708 				rxq->ifr_cq_gen = 0;
2709 			}
2710 			/* was this only a completion queue message? */
2711 			if (__predict_false(ri.iri_nfrags == 0))
2712 				continue;
2713 		}
2714 		MPASS(ri.iri_nfrags != 0);
2715 		MPASS(ri.iri_len != 0);
2716 
2717 		/* will advance the cidx on the corresponding free lists */
2718 		m = iflib_rxd_pkt_get(rxq, &ri);
2719 		avail--;
2720 		budget_left--;
2721 		if (avail == 0 && budget_left)
2722 			avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left);
2723 
2724 		if (__predict_false(m == NULL)) {
2725 			DBG_COUNTER_INC(rx_mbuf_null);
2726 			continue;
2727 		}
2728 		/* imm_pkt: -- cxgb */
2729 		if (mh == NULL)
2730 			mh = mt = m;
2731 		else {
2732 			mt->m_nextpkt = m;
2733 			mt = m;
2734 		}
2735 	}
2736 	/* make sure that we can refill faster than drain */
2737 	for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2738 		__iflib_fl_refill_lt(ctx, fl, budget + 8);
2739 
2740 	lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO);
2741 	if (lro_enabled)
2742 		iflib_get_ip_forwarding(&rxq->ifr_lc, &v4_forwarding, &v6_forwarding);
2743 	mt = mf = NULL;
2744 	while (mh != NULL) {
2745 		m = mh;
2746 		mh = mh->m_nextpkt;
2747 		m->m_nextpkt = NULL;
2748 #ifndef __NO_STRICT_ALIGNMENT
2749 		if (!IP_ALIGNED(m) && (m = iflib_fixup_rx(m)) == NULL)
2750 			continue;
2751 #endif
2752 		rx_bytes += m->m_pkthdr.len;
2753 		rx_pkts++;
2754 #if defined(INET6) || defined(INET)
2755 		if (lro_enabled) {
2756 			if (!lro_possible) {
2757 				lro_possible = iflib_check_lro_possible(m, v4_forwarding, v6_forwarding);
2758 				if (lro_possible && mf != NULL) {
2759 					ifp->if_input(ifp, mf);
2760 					DBG_COUNTER_INC(rx_if_input);
2761 					mt = mf = NULL;
2762 				}
2763 			}
2764 			if ((m->m_pkthdr.csum_flags & (CSUM_L4_CALC|CSUM_L4_VALID)) ==
2765 			    (CSUM_L4_CALC|CSUM_L4_VALID)) {
2766 				if (lro_possible && tcp_lro_rx(&rxq->ifr_lc, m, 0) == 0)
2767 					continue;
2768 			}
2769 		}
2770 #endif
2771 		if (lro_possible) {
2772 			ifp->if_input(ifp, m);
2773 			DBG_COUNTER_INC(rx_if_input);
2774 			continue;
2775 		}
2776 
2777 		if (mf == NULL)
2778 			mf = m;
2779 		if (mt != NULL)
2780 			mt->m_nextpkt = m;
2781 		mt = m;
2782 	}
2783 	if (mf != NULL) {
2784 		ifp->if_input(ifp, mf);
2785 		DBG_COUNTER_INC(rx_if_input);
2786 	}
2787 
2788 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
2789 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
2790 
2791 	/*
2792 	 * Flush any outstanding LRO work
2793 	 */
2794 #if defined(INET6) || defined(INET)
2795 	tcp_lro_flush_all(&rxq->ifr_lc);
2796 #endif
2797 	if (avail)
2798 		return true;
2799 	return (iflib_rxd_avail(ctx, rxq, *cidxp, 1));
2800 err:
2801 	STATE_LOCK(ctx);
2802 	ctx->ifc_flags |= IFC_DO_RESET;
2803 	iflib_admin_intr_deferred(ctx);
2804 	STATE_UNLOCK(ctx);
2805 	return (false);
2806 }
2807 
2808 #define TXD_NOTIFY_COUNT(txq) (((txq)->ift_size / (txq)->ift_update_freq)-1)
2809 static inline qidx_t
2810 txq_max_db_deferred(iflib_txq_t txq, qidx_t in_use)
2811 {
2812 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2813 	qidx_t minthresh = txq->ift_size / 8;
2814 	if (in_use > 4*minthresh)
2815 		return (notify_count);
2816 	if (in_use > 2*minthresh)
2817 		return (notify_count >> 1);
2818 	if (in_use > minthresh)
2819 		return (notify_count >> 3);
2820 	return (0);
2821 }
2822 
2823 static inline qidx_t
2824 txq_max_rs_deferred(iflib_txq_t txq)
2825 {
2826 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2827 	qidx_t minthresh = txq->ift_size / 8;
2828 	if (txq->ift_in_use > 4*minthresh)
2829 		return (notify_count);
2830 	if (txq->ift_in_use > 2*minthresh)
2831 		return (notify_count >> 1);
2832 	if (txq->ift_in_use > minthresh)
2833 		return (notify_count >> 2);
2834 	return (2);
2835 }
2836 
2837 #define M_CSUM_FLAGS(m) ((m)->m_pkthdr.csum_flags)
2838 #define M_HAS_VLANTAG(m) (m->m_flags & M_VLANTAG)
2839 
2840 #define TXQ_MAX_DB_DEFERRED(txq, in_use) txq_max_db_deferred((txq), (in_use))
2841 #define TXQ_MAX_RS_DEFERRED(txq) txq_max_rs_deferred(txq)
2842 #define TXQ_MAX_DB_CONSUMED(size) (size >> 4)
2843 
2844 /* forward compatibility for cxgb */
2845 #define FIRST_QSET(ctx) 0
2846 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets)
2847 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets)
2848 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx))
2849 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments))
2850 
2851 /* XXX we should be setting this to something other than zero */
2852 #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh)
2853 #define	MAX_TX_DESC(ctx) max((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max, \
2854     (ctx)->ifc_softc_ctx.isc_tx_nsegments)
2855 
2856 static inline bool
2857 iflib_txd_db_check(if_ctx_t ctx, iflib_txq_t txq, int ring, qidx_t in_use)
2858 {
2859 	qidx_t dbval, max;
2860 	bool rang;
2861 
2862 	rang = false;
2863 	max = TXQ_MAX_DB_DEFERRED(txq, in_use);
2864 	if (ring || txq->ift_db_pending >= max) {
2865 		dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx;
2866 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval);
2867 		txq->ift_db_pending = txq->ift_npending = 0;
2868 		rang = true;
2869 	}
2870 	return (rang);
2871 }
2872 
2873 #ifdef PKT_DEBUG
2874 static void
2875 print_pkt(if_pkt_info_t pi)
2876 {
2877 	printf("pi len:  %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n",
2878 	       pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx);
2879 	printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n",
2880 	       pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag);
2881 	printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n",
2882 	       pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto);
2883 }
2884 #endif
2885 
2886 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO)
2887 #define IS_TX_OFFLOAD4(pi) ((pi)->ipi_csum_flags & (CSUM_IP_TCP | CSUM_IP_TSO))
2888 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO)
2889 #define IS_TX_OFFLOAD6(pi) ((pi)->ipi_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_TSO))
2890 
2891 static int
2892 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp)
2893 {
2894 	if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx;
2895 	struct ether_vlan_header *eh;
2896 	struct mbuf *m;
2897 
2898 	m = *mp;
2899 	if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) &&
2900 	    M_WRITABLE(m) == 0) {
2901 		if ((m = m_dup(m, M_NOWAIT)) == NULL) {
2902 			return (ENOMEM);
2903 		} else {
2904 			m_freem(*mp);
2905 			DBG_COUNTER_INC(tx_frees);
2906 			*mp = m;
2907 		}
2908 	}
2909 
2910 	/*
2911 	 * Determine where frame payload starts.
2912 	 * Jump over vlan headers if already present,
2913 	 * helpful for QinQ too.
2914 	 */
2915 	if (__predict_false(m->m_len < sizeof(*eh))) {
2916 		txq->ift_pullups++;
2917 		if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL))
2918 			return (ENOMEM);
2919 	}
2920 	eh = mtod(m, struct ether_vlan_header *);
2921 	if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
2922 		pi->ipi_etype = ntohs(eh->evl_proto);
2923 		pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
2924 	} else {
2925 		pi->ipi_etype = ntohs(eh->evl_encap_proto);
2926 		pi->ipi_ehdrlen = ETHER_HDR_LEN;
2927 	}
2928 
2929 	switch (pi->ipi_etype) {
2930 #ifdef INET
2931 	case ETHERTYPE_IP:
2932 	{
2933 		struct mbuf *n;
2934 		struct ip *ip = NULL;
2935 		struct tcphdr *th = NULL;
2936 		int minthlen;
2937 
2938 		minthlen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip) + sizeof(*th));
2939 		if (__predict_false(m->m_len < minthlen)) {
2940 			/*
2941 			 * if this code bloat is causing too much of a hit
2942 			 * move it to a separate function and mark it noinline
2943 			 */
2944 			if (m->m_len == pi->ipi_ehdrlen) {
2945 				n = m->m_next;
2946 				MPASS(n);
2947 				if (n->m_len >= sizeof(*ip))  {
2948 					ip = (struct ip *)n->m_data;
2949 					if (n->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2950 						th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2951 				} else {
2952 					txq->ift_pullups++;
2953 					if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
2954 						return (ENOMEM);
2955 					ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2956 				}
2957 			} else {
2958 				txq->ift_pullups++;
2959 				if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
2960 					return (ENOMEM);
2961 				ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2962 				if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2963 					th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2964 			}
2965 		} else {
2966 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2967 			if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2968 				th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2969 		}
2970 		pi->ipi_ip_hlen = ip->ip_hl << 2;
2971 		pi->ipi_ipproto = ip->ip_p;
2972 		pi->ipi_flags |= IPI_TX_IPV4;
2973 
2974 		/* TCP checksum offload may require TCP header length */
2975 		if (IS_TX_OFFLOAD4(pi)) {
2976 			if (__predict_true(pi->ipi_ipproto == IPPROTO_TCP)) {
2977 				if (__predict_false(th == NULL)) {
2978 					txq->ift_pullups++;
2979 					if (__predict_false((m = m_pullup(m, (ip->ip_hl << 2) + sizeof(*th))) == NULL))
2980 						return (ENOMEM);
2981 					th = (struct tcphdr *)((caddr_t)ip + pi->ipi_ip_hlen);
2982 				}
2983 				pi->ipi_tcp_hflags = th->th_flags;
2984 				pi->ipi_tcp_hlen = th->th_off << 2;
2985 				pi->ipi_tcp_seq = th->th_seq;
2986 			}
2987 			if (IS_TSO4(pi)) {
2988 				if (__predict_false(ip->ip_p != IPPROTO_TCP))
2989 					return (ENXIO);
2990 				/*
2991 				 * TSO always requires hardware checksum offload.
2992 				 */
2993 				pi->ipi_csum_flags |= (CSUM_IP_TCP | CSUM_IP);
2994 				th->th_sum = in_pseudo(ip->ip_src.s_addr,
2995 						       ip->ip_dst.s_addr, htons(IPPROTO_TCP));
2996 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
2997 				if (sctx->isc_flags & IFLIB_TSO_INIT_IP) {
2998 					ip->ip_sum = 0;
2999 					ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz);
3000 				}
3001 			}
3002 		}
3003 		if ((sctx->isc_flags & IFLIB_NEED_ZERO_CSUM) && (pi->ipi_csum_flags & CSUM_IP))
3004                        ip->ip_sum = 0;
3005 
3006 		break;
3007 	}
3008 #endif
3009 #ifdef INET6
3010 	case ETHERTYPE_IPV6:
3011 	{
3012 		struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
3013 		struct tcphdr *th;
3014 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
3015 
3016 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
3017 			txq->ift_pullups++;
3018 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
3019 				return (ENOMEM);
3020 		}
3021 		th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
3022 
3023 		/* XXX-BZ this will go badly in case of ext hdrs. */
3024 		pi->ipi_ipproto = ip6->ip6_nxt;
3025 		pi->ipi_flags |= IPI_TX_IPV6;
3026 
3027 		/* TCP checksum offload may require TCP header length */
3028 		if (IS_TX_OFFLOAD6(pi)) {
3029 			if (pi->ipi_ipproto == IPPROTO_TCP) {
3030 				if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) {
3031 					txq->ift_pullups++;
3032 					if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL))
3033 						return (ENOMEM);
3034 				}
3035 				pi->ipi_tcp_hflags = th->th_flags;
3036 				pi->ipi_tcp_hlen = th->th_off << 2;
3037 				pi->ipi_tcp_seq = th->th_seq;
3038 			}
3039 			if (IS_TSO6(pi)) {
3040 				if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP))
3041 					return (ENXIO);
3042 				/*
3043 				 * TSO always requires hardware checksum offload.
3044 				 */
3045 				pi->ipi_csum_flags |= CSUM_IP6_TCP;
3046 				th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0);
3047 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3048 			}
3049 		}
3050 		break;
3051 	}
3052 #endif
3053 	default:
3054 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
3055 		pi->ipi_ip_hlen = 0;
3056 		break;
3057 	}
3058 	*mp = m;
3059 
3060 	return (0);
3061 }
3062 
3063 /*
3064  * If dodgy hardware rejects the scatter gather chain we've handed it
3065  * we'll need to remove the mbuf chain from ifsg_m[] before we can add the
3066  * m_defrag'd mbufs
3067  */
3068 static __noinline struct mbuf *
3069 iflib_remove_mbuf(iflib_txq_t txq)
3070 {
3071 	int ntxd, pidx;
3072 	struct mbuf *m, **ifsd_m;
3073 
3074 	ifsd_m = txq->ift_sds.ifsd_m;
3075 	ntxd = txq->ift_size;
3076 	pidx = txq->ift_pidx & (ntxd - 1);
3077 	ifsd_m = txq->ift_sds.ifsd_m;
3078 	m = ifsd_m[pidx];
3079 	ifsd_m[pidx] = NULL;
3080 	bus_dmamap_unload(txq->ift_desc_tag, txq->ift_sds.ifsd_map[pidx]);
3081 	if (txq->ift_sds.ifsd_tso_map != NULL)
3082 		bus_dmamap_unload(txq->ift_tso_desc_tag,
3083 		    txq->ift_sds.ifsd_tso_map[pidx]);
3084 #if MEMORY_LOGGING
3085 	txq->ift_dequeued++;
3086 #endif
3087 	return (m);
3088 }
3089 
3090 static inline caddr_t
3091 calc_next_txd(iflib_txq_t txq, int cidx, uint8_t qid)
3092 {
3093 	qidx_t size;
3094 	int ntxd;
3095 	caddr_t start, end, cur, next;
3096 
3097 	ntxd = txq->ift_size;
3098 	size = txq->ift_txd_size[qid];
3099 	start = txq->ift_ifdi[qid].idi_vaddr;
3100 
3101 	if (__predict_false(size == 0))
3102 		return (start);
3103 	cur = start + size*cidx;
3104 	end = start + size*ntxd;
3105 	next = CACHE_PTR_NEXT(cur);
3106 	return (next < end ? next : start);
3107 }
3108 
3109 /*
3110  * Pad an mbuf to ensure a minimum ethernet frame size.
3111  * min_frame_size is the frame size (less CRC) to pad the mbuf to
3112  */
3113 static __noinline int
3114 iflib_ether_pad(device_t dev, struct mbuf **m_head, uint16_t min_frame_size)
3115 {
3116 	/*
3117 	 * 18 is enough bytes to pad an ARP packet to 46 bytes, and
3118 	 * and ARP message is the smallest common payload I can think of
3119 	 */
3120 	static char pad[18];	/* just zeros */
3121 	int n;
3122 	struct mbuf *new_head;
3123 
3124 	if (!M_WRITABLE(*m_head)) {
3125 		new_head = m_dup(*m_head, M_NOWAIT);
3126 		if (new_head == NULL) {
3127 			m_freem(*m_head);
3128 			device_printf(dev, "cannot pad short frame, m_dup() failed");
3129 			DBG_COUNTER_INC(encap_pad_mbuf_fail);
3130 			DBG_COUNTER_INC(tx_frees);
3131 			return ENOMEM;
3132 		}
3133 		m_freem(*m_head);
3134 		*m_head = new_head;
3135 	}
3136 
3137 	for (n = min_frame_size - (*m_head)->m_pkthdr.len;
3138 	     n > 0; n -= sizeof(pad))
3139 		if (!m_append(*m_head, min(n, sizeof(pad)), pad))
3140 			break;
3141 
3142 	if (n > 0) {
3143 		m_freem(*m_head);
3144 		device_printf(dev, "cannot pad short frame\n");
3145 		DBG_COUNTER_INC(encap_pad_mbuf_fail);
3146 		DBG_COUNTER_INC(tx_frees);
3147 		return (ENOBUFS);
3148 	}
3149 
3150 	return 0;
3151 }
3152 
3153 static int
3154 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp)
3155 {
3156 	if_ctx_t		ctx;
3157 	if_shared_ctx_t		sctx;
3158 	if_softc_ctx_t		scctx;
3159 	bus_dma_segment_t	*segs;
3160 	struct mbuf		*m_head, **ifsd_m;
3161 	void			*next_txd;
3162 	bus_dmamap_t		map;
3163 	struct if_pkt_info	pi;
3164 	int remap = 0;
3165 	int err, nsegs, ndesc, max_segs, pidx, cidx, next, ntxd;
3166 	bus_dma_tag_t desc_tag;
3167 
3168 	ctx = txq->ift_ctx;
3169 	sctx = ctx->ifc_sctx;
3170 	scctx = &ctx->ifc_softc_ctx;
3171 	segs = txq->ift_segs;
3172 	ntxd = txq->ift_size;
3173 	m_head = *m_headp;
3174 	map = NULL;
3175 
3176 	/*
3177 	 * If we're doing TSO the next descriptor to clean may be quite far ahead
3178 	 */
3179 	cidx = txq->ift_cidx;
3180 	pidx = txq->ift_pidx;
3181 	if (ctx->ifc_flags & IFC_PREFETCH) {
3182 		next = (cidx + CACHE_PTR_INCREMENT) & (ntxd-1);
3183 		if (!(ctx->ifc_flags & IFLIB_HAS_TXCQ)) {
3184 			next_txd = calc_next_txd(txq, cidx, 0);
3185 			prefetch(next_txd);
3186 		}
3187 
3188 		/* prefetch the next cache line of mbuf pointers and flags */
3189 		prefetch(&txq->ift_sds.ifsd_m[next]);
3190 		prefetch(&txq->ift_sds.ifsd_map[next]);
3191 		next = (cidx + CACHE_LINE_SIZE) & (ntxd-1);
3192 	}
3193 	map = txq->ift_sds.ifsd_map[pidx];
3194 	ifsd_m = txq->ift_sds.ifsd_m;
3195 
3196 	if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
3197 		desc_tag = txq->ift_tso_desc_tag;
3198 		max_segs = scctx->isc_tx_tso_segments_max;
3199 		map = txq->ift_sds.ifsd_tso_map[pidx];
3200 		MPASS(desc_tag != NULL);
3201 		MPASS(max_segs > 0);
3202 	} else {
3203 		desc_tag = txq->ift_desc_tag;
3204 		max_segs = scctx->isc_tx_nsegments;
3205 		map = txq->ift_sds.ifsd_map[pidx];
3206 	}
3207 	if ((sctx->isc_flags & IFLIB_NEED_ETHER_PAD) &&
3208 	    __predict_false(m_head->m_pkthdr.len < scctx->isc_min_frame_size)) {
3209 		err = iflib_ether_pad(ctx->ifc_dev, m_headp, scctx->isc_min_frame_size);
3210 		if (err) {
3211 			DBG_COUNTER_INC(encap_txd_encap_fail);
3212 			return err;
3213 		}
3214 	}
3215 	m_head = *m_headp;
3216 
3217 	pkt_info_zero(&pi);
3218 	pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG|M_BCAST|M_MCAST));
3219 	pi.ipi_pidx = pidx;
3220 	pi.ipi_qsidx = txq->ift_id;
3221 	pi.ipi_len = m_head->m_pkthdr.len;
3222 	pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags;
3223 	pi.ipi_vtag = (m_head->m_flags & M_VLANTAG) ? m_head->m_pkthdr.ether_vtag : 0;
3224 
3225 	/* deliberate bitwise OR to make one condition */
3226 	if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) {
3227 		if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0)) {
3228 			DBG_COUNTER_INC(encap_txd_encap_fail);
3229 			return (err);
3230 		}
3231 		m_head = *m_headp;
3232 	}
3233 
3234 retry:
3235 	err = bus_dmamap_load_mbuf_sg(desc_tag, map, m_head, segs, &nsegs,
3236 	    BUS_DMA_NOWAIT);
3237 defrag:
3238 	if (__predict_false(err)) {
3239 		switch (err) {
3240 		case EFBIG:
3241 			/* try collapse once and defrag once */
3242 			if (remap == 0) {
3243 				m_head = m_collapse(*m_headp, M_NOWAIT, max_segs);
3244 				/* try defrag if collapsing fails */
3245 				if (m_head == NULL)
3246 					remap++;
3247 			}
3248 			if (remap == 1) {
3249 				txq->ift_mbuf_defrag++;
3250 				m_head = m_defrag(*m_headp, M_NOWAIT);
3251 			}
3252 			remap++;
3253 			if (__predict_false(m_head == NULL))
3254 				goto defrag_failed;
3255 			*m_headp = m_head;
3256 			goto retry;
3257 			break;
3258 		case ENOMEM:
3259 			txq->ift_no_tx_dma_setup++;
3260 			break;
3261 		default:
3262 			txq->ift_no_tx_dma_setup++;
3263 			m_freem(*m_headp);
3264 			DBG_COUNTER_INC(tx_frees);
3265 			*m_headp = NULL;
3266 			break;
3267 		}
3268 		txq->ift_map_failed++;
3269 		DBG_COUNTER_INC(encap_load_mbuf_fail);
3270 		DBG_COUNTER_INC(encap_txd_encap_fail);
3271 		return (err);
3272 	}
3273 	ifsd_m[pidx] = m_head;
3274 	/*
3275 	 * XXX assumes a 1 to 1 relationship between segments and
3276 	 *        descriptors - this does not hold true on all drivers, e.g.
3277 	 *        cxgb
3278 	 */
3279 	if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) {
3280 		txq->ift_no_desc_avail++;
3281 		bus_dmamap_unload(desc_tag, map);
3282 		DBG_COUNTER_INC(encap_txq_avail_fail);
3283 		DBG_COUNTER_INC(encap_txd_encap_fail);
3284 		if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0)
3285 			GROUPTASK_ENQUEUE(&txq->ift_task);
3286 		return (ENOBUFS);
3287 	}
3288 	/*
3289 	 * On Intel cards we can greatly reduce the number of TX interrupts
3290 	 * we see by only setting report status on every Nth descriptor.
3291 	 * However, this also means that the driver will need to keep track
3292 	 * of the descriptors that RS was set on to check them for the DD bit.
3293 	 */
3294 	txq->ift_rs_pending += nsegs + 1;
3295 	if (txq->ift_rs_pending > TXQ_MAX_RS_DEFERRED(txq) ||
3296 	     iflib_no_tx_batch || (TXQ_AVAIL(txq) - nsegs) <= MAX_TX_DESC(ctx) + 2) {
3297 		pi.ipi_flags |= IPI_TX_INTR;
3298 		txq->ift_rs_pending = 0;
3299 	}
3300 
3301 	pi.ipi_segs = segs;
3302 	pi.ipi_nsegs = nsegs;
3303 
3304 	MPASS(pidx >= 0 && pidx < txq->ift_size);
3305 #ifdef PKT_DEBUG
3306 	print_pkt(&pi);
3307 #endif
3308 	bus_dmamap_sync(desc_tag, map, BUS_DMASYNC_PREWRITE);
3309 	if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) {
3310 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3311 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
3312 		DBG_COUNTER_INC(tx_encap);
3313 		MPASS(pi.ipi_new_pidx < txq->ift_size);
3314 
3315 		ndesc = pi.ipi_new_pidx - pi.ipi_pidx;
3316 		if (pi.ipi_new_pidx < pi.ipi_pidx) {
3317 			ndesc += txq->ift_size;
3318 			txq->ift_gen = 1;
3319 		}
3320 		/*
3321 		 * drivers can need as many as
3322 		 * two sentinels
3323 		 */
3324 		MPASS(ndesc <= pi.ipi_nsegs + 2);
3325 		MPASS(pi.ipi_new_pidx != pidx);
3326 		MPASS(ndesc > 0);
3327 		txq->ift_in_use += ndesc;
3328 
3329 		/*
3330 		 * We update the last software descriptor again here because there may
3331 		 * be a sentinel and/or there may be more mbufs than segments
3332 		 */
3333 		txq->ift_pidx = pi.ipi_new_pidx;
3334 		txq->ift_npending += pi.ipi_ndescs;
3335 	} else {
3336 		*m_headp = m_head = iflib_remove_mbuf(txq);
3337 		if (err == EFBIG) {
3338 			txq->ift_txd_encap_efbig++;
3339 			if (remap < 2) {
3340 				remap = 1;
3341 				goto defrag;
3342 			}
3343 		}
3344 		goto defrag_failed;
3345 	}
3346 	/*
3347 	 * err can't possibly be non-zero here, so we don't neet to test it
3348 	 * to see if we need to DBG_COUNTER_INC(encap_txd_encap_fail).
3349 	 */
3350 	return (err);
3351 
3352 defrag_failed:
3353 	txq->ift_mbuf_defrag_failed++;
3354 	txq->ift_map_failed++;
3355 	m_freem(*m_headp);
3356 	DBG_COUNTER_INC(tx_frees);
3357 	*m_headp = NULL;
3358 	DBG_COUNTER_INC(encap_txd_encap_fail);
3359 	return (ENOMEM);
3360 }
3361 
3362 static void
3363 iflib_tx_desc_free(iflib_txq_t txq, int n)
3364 {
3365 	uint32_t qsize, cidx, mask, gen;
3366 	struct mbuf *m, **ifsd_m;
3367 	bool do_prefetch;
3368 
3369 	cidx = txq->ift_cidx;
3370 	gen = txq->ift_gen;
3371 	qsize = txq->ift_size;
3372 	mask = qsize-1;
3373 	ifsd_m = txq->ift_sds.ifsd_m;
3374 	do_prefetch = (txq->ift_ctx->ifc_flags & IFC_PREFETCH);
3375 
3376 	while (n-- > 0) {
3377 		if (do_prefetch) {
3378 			prefetch(ifsd_m[(cidx + 3) & mask]);
3379 			prefetch(ifsd_m[(cidx + 4) & mask]);
3380 		}
3381 		if ((m = ifsd_m[cidx]) != NULL) {
3382 			prefetch(&ifsd_m[(cidx + CACHE_PTR_INCREMENT) & mask]);
3383 			if (m->m_pkthdr.csum_flags & CSUM_TSO) {
3384 				bus_dmamap_sync(txq->ift_tso_desc_tag,
3385 				    txq->ift_sds.ifsd_tso_map[cidx],
3386 				    BUS_DMASYNC_POSTWRITE);
3387 				bus_dmamap_unload(txq->ift_tso_desc_tag,
3388 				    txq->ift_sds.ifsd_tso_map[cidx]);
3389 			} else {
3390 				bus_dmamap_sync(txq->ift_desc_tag,
3391 				    txq->ift_sds.ifsd_map[cidx],
3392 				    BUS_DMASYNC_POSTWRITE);
3393 				bus_dmamap_unload(txq->ift_desc_tag,
3394 				    txq->ift_sds.ifsd_map[cidx]);
3395 			}
3396 			/* XXX we don't support any drivers that batch packets yet */
3397 			MPASS(m->m_nextpkt == NULL);
3398 			m_freem(m);
3399 			ifsd_m[cidx] = NULL;
3400 #if MEMORY_LOGGING
3401 			txq->ift_dequeued++;
3402 #endif
3403 			DBG_COUNTER_INC(tx_frees);
3404 		}
3405 		if (__predict_false(++cidx == qsize)) {
3406 			cidx = 0;
3407 			gen = 0;
3408 		}
3409 	}
3410 	txq->ift_cidx = cidx;
3411 	txq->ift_gen = gen;
3412 }
3413 
3414 static __inline int
3415 iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh)
3416 {
3417 	int reclaim;
3418 	if_ctx_t ctx = txq->ift_ctx;
3419 
3420 	KASSERT(thresh >= 0, ("invalid threshold to reclaim"));
3421 	MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size);
3422 
3423 	/*
3424 	 * Need a rate-limiting check so that this isn't called every time
3425 	 */
3426 	iflib_tx_credits_update(ctx, txq);
3427 	reclaim = DESC_RECLAIMABLE(txq);
3428 
3429 	if (reclaim <= thresh /* + MAX_TX_DESC(txq->ift_ctx) */) {
3430 #ifdef INVARIANTS
3431 		if (iflib_verbose_debug) {
3432 			printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __FUNCTION__,
3433 			       txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments,
3434 			       reclaim, thresh);
3435 
3436 		}
3437 #endif
3438 		return (0);
3439 	}
3440 	iflib_tx_desc_free(txq, reclaim);
3441 	txq->ift_cleaned += reclaim;
3442 	txq->ift_in_use -= reclaim;
3443 
3444 	return (reclaim);
3445 }
3446 
3447 static struct mbuf **
3448 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset, int remaining)
3449 {
3450 	int next, size;
3451 	struct mbuf **items;
3452 
3453 	size = r->size;
3454 	next = (cidx + CACHE_PTR_INCREMENT) & (size-1);
3455 	items = __DEVOLATILE(struct mbuf **, &r->items[0]);
3456 
3457 	prefetch(items[(cidx + offset) & (size-1)]);
3458 	if (remaining > 1) {
3459 		prefetch2cachelines(&items[next]);
3460 		prefetch2cachelines(items[(cidx + offset + 1) & (size-1)]);
3461 		prefetch2cachelines(items[(cidx + offset + 2) & (size-1)]);
3462 		prefetch2cachelines(items[(cidx + offset + 3) & (size-1)]);
3463 	}
3464 	return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (size-1)]));
3465 }
3466 
3467 static void
3468 iflib_txq_check_drain(iflib_txq_t txq, int budget)
3469 {
3470 
3471 	ifmp_ring_check_drainage(txq->ift_br, budget);
3472 }
3473 
3474 static uint32_t
3475 iflib_txq_can_drain(struct ifmp_ring *r)
3476 {
3477 	iflib_txq_t txq = r->cookie;
3478 	if_ctx_t ctx = txq->ift_ctx;
3479 
3480 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3481 	    BUS_DMASYNC_POSTREAD);
3482 	return ((TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2) ||
3483 		ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false));
3484 }
3485 
3486 static uint32_t
3487 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3488 {
3489 	iflib_txq_t txq = r->cookie;
3490 	if_ctx_t ctx = txq->ift_ctx;
3491 	struct ifnet *ifp = ctx->ifc_ifp;
3492 	struct mbuf **mp, *m;
3493 	int i, count, consumed, pkt_sent, bytes_sent, mcast_sent, avail;
3494 	int reclaimed, err, in_use_prev, desc_used;
3495 	bool do_prefetch, ring, rang;
3496 
3497 	if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) ||
3498 			    !LINK_ACTIVE(ctx))) {
3499 		DBG_COUNTER_INC(txq_drain_notready);
3500 		return (0);
3501 	}
3502 	reclaimed = iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
3503 	rang = iflib_txd_db_check(ctx, txq, reclaimed, txq->ift_in_use);
3504 	avail = IDXDIFF(pidx, cidx, r->size);
3505 	if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) {
3506 		DBG_COUNTER_INC(txq_drain_flushing);
3507 		for (i = 0; i < avail; i++) {
3508 			if (__predict_true(r->items[(cidx + i) & (r->size-1)] != (void *)txq))
3509 				m_free(r->items[(cidx + i) & (r->size-1)]);
3510 			r->items[(cidx + i) & (r->size-1)] = NULL;
3511 		}
3512 		return (avail);
3513 	}
3514 
3515 	if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3516 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3517 		CALLOUT_LOCK(txq);
3518 		callout_stop(&txq->ift_timer);
3519 		CALLOUT_UNLOCK(txq);
3520 		DBG_COUNTER_INC(txq_drain_oactive);
3521 		return (0);
3522 	}
3523 	if (reclaimed)
3524 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3525 	consumed = mcast_sent = bytes_sent = pkt_sent = 0;
3526 	count = MIN(avail, TX_BATCH_SIZE);
3527 #ifdef INVARIANTS
3528 	if (iflib_verbose_debug)
3529 		printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __FUNCTION__,
3530 		       avail, ctx->ifc_flags, TXQ_AVAIL(txq));
3531 #endif
3532 	do_prefetch = (ctx->ifc_flags & IFC_PREFETCH);
3533 	avail = TXQ_AVAIL(txq);
3534 	err = 0;
3535 	for (desc_used = i = 0; i < count && avail > MAX_TX_DESC(ctx) + 2; i++) {
3536 		int rem = do_prefetch ? count - i : 0;
3537 
3538 		mp = _ring_peek_one(r, cidx, i, rem);
3539 		MPASS(mp != NULL && *mp != NULL);
3540 		if (__predict_false(*mp == (struct mbuf *)txq)) {
3541 			consumed++;
3542 			reclaimed++;
3543 			continue;
3544 		}
3545 		in_use_prev = txq->ift_in_use;
3546 		err = iflib_encap(txq, mp);
3547 		if (__predict_false(err)) {
3548 			/* no room - bail out */
3549 			if (err == ENOBUFS)
3550 				break;
3551 			consumed++;
3552 			/* we can't send this packet - skip it */
3553 			continue;
3554 		}
3555 		consumed++;
3556 		pkt_sent++;
3557 		m = *mp;
3558 		DBG_COUNTER_INC(tx_sent);
3559 		bytes_sent += m->m_pkthdr.len;
3560 		mcast_sent += !!(m->m_flags & M_MCAST);
3561 		avail = TXQ_AVAIL(txq);
3562 
3563 		txq->ift_db_pending += (txq->ift_in_use - in_use_prev);
3564 		desc_used += (txq->ift_in_use - in_use_prev);
3565 		ETHER_BPF_MTAP(ifp, m);
3566 		if (__predict_false(!(ifp->if_drv_flags & IFF_DRV_RUNNING)))
3567 			break;
3568 		rang = iflib_txd_db_check(ctx, txq, false, in_use_prev);
3569 	}
3570 
3571 	/* deliberate use of bitwise or to avoid gratuitous short-circuit */
3572 	ring = rang ? false  : (iflib_min_tx_latency | err) || (TXQ_AVAIL(txq) < MAX_TX_DESC(ctx));
3573 	iflib_txd_db_check(ctx, txq, ring, txq->ift_in_use);
3574 	if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
3575 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
3576 	if (mcast_sent)
3577 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
3578 #ifdef INVARIANTS
3579 	if (iflib_verbose_debug)
3580 		printf("consumed=%d\n", consumed);
3581 #endif
3582 	return (consumed);
3583 }
3584 
3585 static uint32_t
3586 iflib_txq_drain_always(struct ifmp_ring *r)
3587 {
3588 	return (1);
3589 }
3590 
3591 static uint32_t
3592 iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3593 {
3594 	int i, avail;
3595 	struct mbuf **mp;
3596 	iflib_txq_t txq;
3597 
3598 	txq = r->cookie;
3599 
3600 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3601 	CALLOUT_LOCK(txq);
3602 	callout_stop(&txq->ift_timer);
3603 	CALLOUT_UNLOCK(txq);
3604 
3605 	avail = IDXDIFF(pidx, cidx, r->size);
3606 	for (i = 0; i < avail; i++) {
3607 		mp = _ring_peek_one(r, cidx, i, avail - i);
3608 		if (__predict_false(*mp == (struct mbuf *)txq))
3609 			continue;
3610 		m_freem(*mp);
3611 		DBG_COUNTER_INC(tx_frees);
3612 	}
3613 	MPASS(ifmp_ring_is_stalled(r) == 0);
3614 	return (avail);
3615 }
3616 
3617 static void
3618 iflib_ifmp_purge(iflib_txq_t txq)
3619 {
3620 	struct ifmp_ring *r;
3621 
3622 	r = txq->ift_br;
3623 	r->drain = iflib_txq_drain_free;
3624 	r->can_drain = iflib_txq_drain_always;
3625 
3626 	ifmp_ring_check_drainage(r, r->size);
3627 
3628 	r->drain = iflib_txq_drain;
3629 	r->can_drain = iflib_txq_can_drain;
3630 }
3631 
3632 static void
3633 _task_fn_tx(void *context)
3634 {
3635 	iflib_txq_t txq = context;
3636 	if_ctx_t ctx = txq->ift_ctx;
3637 	struct ifnet *ifp = ctx->ifc_ifp;
3638 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
3639 
3640 #ifdef IFLIB_DIAGNOSTICS
3641 	txq->ift_cpu_exec_count[curcpu]++;
3642 #endif
3643 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3644 		return;
3645 	if (if_getcapenable(ifp) & IFCAP_NETMAP) {
3646 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3647 		    BUS_DMASYNC_POSTREAD);
3648 		if (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false))
3649 			netmap_tx_irq(ifp, txq->ift_id);
3650 		IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3651 		return;
3652 	}
3653 #ifdef ALTQ
3654 	if (ALTQ_IS_ENABLED(&ifp->if_snd))
3655 		iflib_altq_if_start(ifp);
3656 #endif
3657 	if (txq->ift_db_pending)
3658 		ifmp_ring_enqueue(txq->ift_br, (void **)&txq, 1, TX_BATCH_SIZE, abdicate);
3659 	else if (!abdicate)
3660 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3661 	/*
3662 	 * When abdicating, we always need to check drainage, not just when we don't enqueue
3663 	 */
3664 	if (abdicate)
3665 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3666 	if (ctx->ifc_flags & IFC_LEGACY)
3667 		IFDI_INTR_ENABLE(ctx);
3668 	else {
3669 #ifdef INVARIANTS
3670 		int rc =
3671 #endif
3672 			IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3673 			KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver"));
3674 	}
3675 }
3676 
3677 static void
3678 _task_fn_rx(void *context)
3679 {
3680 	iflib_rxq_t rxq = context;
3681 	if_ctx_t ctx = rxq->ifr_ctx;
3682 	bool more;
3683 	uint16_t budget;
3684 
3685 #ifdef IFLIB_DIAGNOSTICS
3686 	rxq->ifr_cpu_exec_count[curcpu]++;
3687 #endif
3688 	DBG_COUNTER_INC(task_fn_rxs);
3689 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3690 		return;
3691 	more = true;
3692 #ifdef DEV_NETMAP
3693 	if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP) {
3694 		u_int work = 0;
3695 		if (netmap_rx_irq(ctx->ifc_ifp, rxq->ifr_id, &work)) {
3696 			more = false;
3697 		}
3698 	}
3699 #endif
3700 	budget = ctx->ifc_sysctl_rx_budget;
3701 	if (budget == 0)
3702 		budget = 16;	/* XXX */
3703 	if (more == false || (more = iflib_rxeof(rxq, budget)) == false) {
3704 		if (ctx->ifc_flags & IFC_LEGACY)
3705 			IFDI_INTR_ENABLE(ctx);
3706 		else {
3707 #ifdef INVARIANTS
3708 			int rc =
3709 #endif
3710 				IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
3711 			KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver"));
3712 			DBG_COUNTER_INC(rx_intr_enables);
3713 		}
3714 	}
3715 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3716 		return;
3717 	if (more)
3718 		GROUPTASK_ENQUEUE(&rxq->ifr_task);
3719 }
3720 
3721 static void
3722 _task_fn_admin(void *context)
3723 {
3724 	if_ctx_t ctx = context;
3725 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
3726 	iflib_txq_t txq;
3727 	int i;
3728 	bool oactive, running, do_reset, do_watchdog, in_detach;
3729 	uint32_t reset_on = hz / 2;
3730 
3731 	STATE_LOCK(ctx);
3732 	running = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING);
3733 	oactive = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE);
3734 	do_reset = (ctx->ifc_flags & IFC_DO_RESET);
3735 	do_watchdog = (ctx->ifc_flags & IFC_DO_WATCHDOG);
3736 	in_detach = (ctx->ifc_flags & IFC_IN_DETACH);
3737 	ctx->ifc_flags &= ~(IFC_DO_RESET|IFC_DO_WATCHDOG);
3738 	STATE_UNLOCK(ctx);
3739 
3740 	if ((!running && !oactive) && !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
3741 		return;
3742 	if (in_detach)
3743 		return;
3744 
3745 	CTX_LOCK(ctx);
3746 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3747 		CALLOUT_LOCK(txq);
3748 		callout_stop(&txq->ift_timer);
3749 		CALLOUT_UNLOCK(txq);
3750 	}
3751 	if (do_watchdog) {
3752 		ctx->ifc_watchdog_events++;
3753 		IFDI_WATCHDOG_RESET(ctx);
3754 	}
3755 	IFDI_UPDATE_ADMIN_STATUS(ctx);
3756 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3757 #ifdef DEV_NETMAP
3758 		reset_on = hz / 2;
3759 		if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP)
3760 			iflib_netmap_timer_adjust(ctx, txq->ift_id, &reset_on);
3761 #endif
3762 		callout_reset_on(&txq->ift_timer, reset_on, iflib_timer, txq, txq->ift_timer.c_cpu);
3763 	}
3764 	IFDI_LINK_INTR_ENABLE(ctx);
3765 	if (do_reset)
3766 		iflib_if_init_locked(ctx);
3767 	CTX_UNLOCK(ctx);
3768 
3769 	if (LINK_ACTIVE(ctx) == 0)
3770 		return;
3771 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3772 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
3773 }
3774 
3775 
3776 static void
3777 _task_fn_iov(void *context)
3778 {
3779 	if_ctx_t ctx = context;
3780 
3781 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING) &&
3782 	    !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
3783 		return;
3784 
3785 	CTX_LOCK(ctx);
3786 	IFDI_VFLR_HANDLE(ctx);
3787 	CTX_UNLOCK(ctx);
3788 }
3789 
3790 static int
3791 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)
3792 {
3793 	int err;
3794 	if_int_delay_info_t info;
3795 	if_ctx_t ctx;
3796 
3797 	info = (if_int_delay_info_t)arg1;
3798 	ctx = info->iidi_ctx;
3799 	info->iidi_req = req;
3800 	info->iidi_oidp = oidp;
3801 	CTX_LOCK(ctx);
3802 	err = IFDI_SYSCTL_INT_DELAY(ctx, info);
3803 	CTX_UNLOCK(ctx);
3804 	return (err);
3805 }
3806 
3807 /*********************************************************************
3808  *
3809  *  IFNET FUNCTIONS
3810  *
3811  **********************************************************************/
3812 
3813 static void
3814 iflib_if_init_locked(if_ctx_t ctx)
3815 {
3816 	iflib_stop(ctx);
3817 	iflib_init_locked(ctx);
3818 }
3819 
3820 
3821 static void
3822 iflib_if_init(void *arg)
3823 {
3824 	if_ctx_t ctx = arg;
3825 
3826 	CTX_LOCK(ctx);
3827 	iflib_if_init_locked(ctx);
3828 	CTX_UNLOCK(ctx);
3829 }
3830 
3831 static int
3832 iflib_if_transmit(if_t ifp, struct mbuf *m)
3833 {
3834 	if_ctx_t	ctx = if_getsoftc(ifp);
3835 
3836 	iflib_txq_t txq;
3837 	int err, qidx;
3838 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
3839 
3840 	if (__predict_false((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) {
3841 		DBG_COUNTER_INC(tx_frees);
3842 		m_freem(m);
3843 		return (ENOBUFS);
3844 	}
3845 
3846 	MPASS(m->m_nextpkt == NULL);
3847 	/* ALTQ-enabled interfaces always use queue 0. */
3848 	qidx = 0;
3849 	if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m) && !ALTQ_IS_ENABLED(&ifp->if_snd))
3850 		qidx = QIDX(ctx, m);
3851 	/*
3852 	 * XXX calculate buf_ring based on flowid (divvy up bits?)
3853 	 */
3854 	txq = &ctx->ifc_txqs[qidx];
3855 
3856 #ifdef DRIVER_BACKPRESSURE
3857 	if (txq->ift_closed) {
3858 		while (m != NULL) {
3859 			next = m->m_nextpkt;
3860 			m->m_nextpkt = NULL;
3861 			m_freem(m);
3862 			DBG_COUNTER_INC(tx_frees);
3863 			m = next;
3864 		}
3865 		return (ENOBUFS);
3866 	}
3867 #endif
3868 #ifdef notyet
3869 	qidx = count = 0;
3870 	mp = marr;
3871 	next = m;
3872 	do {
3873 		count++;
3874 		next = next->m_nextpkt;
3875 	} while (next != NULL);
3876 
3877 	if (count > nitems(marr))
3878 		if ((mp = malloc(count*sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) {
3879 			/* XXX check nextpkt */
3880 			m_freem(m);
3881 			/* XXX simplify for now */
3882 			DBG_COUNTER_INC(tx_frees);
3883 			return (ENOBUFS);
3884 		}
3885 	for (next = m, i = 0; next != NULL; i++) {
3886 		mp[i] = next;
3887 		next = next->m_nextpkt;
3888 		mp[i]->m_nextpkt = NULL;
3889 	}
3890 #endif
3891 	DBG_COUNTER_INC(tx_seen);
3892 	err = ifmp_ring_enqueue(txq->ift_br, (void **)&m, 1, TX_BATCH_SIZE, abdicate);
3893 
3894 	if (abdicate)
3895 		GROUPTASK_ENQUEUE(&txq->ift_task);
3896  	if (err) {
3897 		if (!abdicate)
3898 			GROUPTASK_ENQUEUE(&txq->ift_task);
3899 		/* support forthcoming later */
3900 #ifdef DRIVER_BACKPRESSURE
3901 		txq->ift_closed = TRUE;
3902 #endif
3903 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3904 		m_freem(m);
3905 		DBG_COUNTER_INC(tx_frees);
3906 	}
3907 
3908 	return (err);
3909 }
3910 
3911 #ifdef ALTQ
3912 /*
3913  * The overall approach to integrating iflib with ALTQ is to continue to use
3914  * the iflib mp_ring machinery between the ALTQ queue(s) and the hardware
3915  * ring.  Technically, when using ALTQ, queueing to an intermediate mp_ring
3916  * is redundant/unnecessary, but doing so minimizes the amount of
3917  * ALTQ-specific code required in iflib.  It is assumed that the overhead of
3918  * redundantly queueing to an intermediate mp_ring is swamped by the
3919  * performance limitations inherent in using ALTQ.
3920  *
3921  * When ALTQ support is compiled in, all iflib drivers will use a transmit
3922  * routine, iflib_altq_if_transmit(), that checks if ALTQ is enabled for the
3923  * given interface.  If ALTQ is enabled for an interface, then all
3924  * transmitted packets for that interface will be submitted to the ALTQ
3925  * subsystem via IFQ_ENQUEUE().  We don't use the legacy if_transmit()
3926  * implementation because it uses IFQ_HANDOFF(), which will duplicatively
3927  * update stats that the iflib machinery handles, and which is sensitve to
3928  * the disused IFF_DRV_OACTIVE flag.  Additionally, iflib_altq_if_start()
3929  * will be installed as the start routine for use by ALTQ facilities that
3930  * need to trigger queue drains on a scheduled basis.
3931  *
3932  */
3933 static void
3934 iflib_altq_if_start(if_t ifp)
3935 {
3936 	struct ifaltq *ifq = &ifp->if_snd;
3937 	struct mbuf *m;
3938 
3939 	IFQ_LOCK(ifq);
3940 	IFQ_DEQUEUE_NOLOCK(ifq, m);
3941 	while (m != NULL) {
3942 		iflib_if_transmit(ifp, m);
3943 		IFQ_DEQUEUE_NOLOCK(ifq, m);
3944 	}
3945 	IFQ_UNLOCK(ifq);
3946 }
3947 
3948 static int
3949 iflib_altq_if_transmit(if_t ifp, struct mbuf *m)
3950 {
3951 	int err;
3952 
3953 	if (ALTQ_IS_ENABLED(&ifp->if_snd)) {
3954 		IFQ_ENQUEUE(&ifp->if_snd, m, err);
3955 		if (err == 0)
3956 			iflib_altq_if_start(ifp);
3957 	} else
3958 		err = iflib_if_transmit(ifp, m);
3959 
3960 	return (err);
3961 }
3962 #endif /* ALTQ */
3963 
3964 static void
3965 iflib_if_qflush(if_t ifp)
3966 {
3967 	if_ctx_t ctx = if_getsoftc(ifp);
3968 	iflib_txq_t txq = ctx->ifc_txqs;
3969 	int i;
3970 
3971 	STATE_LOCK(ctx);
3972 	ctx->ifc_flags |= IFC_QFLUSH;
3973 	STATE_UNLOCK(ctx);
3974 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
3975 		while (!(ifmp_ring_is_idle(txq->ift_br) || ifmp_ring_is_stalled(txq->ift_br)))
3976 			iflib_txq_check_drain(txq, 0);
3977 	STATE_LOCK(ctx);
3978 	ctx->ifc_flags &= ~IFC_QFLUSH;
3979 	STATE_UNLOCK(ctx);
3980 
3981 	/*
3982 	 * When ALTQ is enabled, this will also take care of purging the
3983 	 * ALTQ queue(s).
3984 	 */
3985 	if_qflush(ifp);
3986 }
3987 
3988 
3989 #define IFCAP_FLAGS (IFCAP_HWCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \
3990 		     IFCAP_TSO | IFCAP_VLAN_HWTAGGING | IFCAP_HWSTATS | \
3991 		     IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | \
3992 		     IFCAP_VLAN_HWTSO | IFCAP_VLAN_HWCSUM)
3993 
3994 static int
3995 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data)
3996 {
3997 	if_ctx_t ctx = if_getsoftc(ifp);
3998 	struct ifreq	*ifr = (struct ifreq *)data;
3999 #if defined(INET) || defined(INET6)
4000 	struct ifaddr	*ifa = (struct ifaddr *)data;
4001 #endif
4002 	bool		avoid_reset = FALSE;
4003 	int		err = 0, reinit = 0, bits;
4004 
4005 	switch (command) {
4006 	case SIOCSIFADDR:
4007 #ifdef INET
4008 		if (ifa->ifa_addr->sa_family == AF_INET)
4009 			avoid_reset = TRUE;
4010 #endif
4011 #ifdef INET6
4012 		if (ifa->ifa_addr->sa_family == AF_INET6)
4013 			avoid_reset = TRUE;
4014 #endif
4015 		/*
4016 		** Calling init results in link renegotiation,
4017 		** so we avoid doing it when possible.
4018 		*/
4019 		if (avoid_reset) {
4020 			if_setflagbits(ifp, IFF_UP,0);
4021 			if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
4022 				reinit = 1;
4023 #ifdef INET
4024 			if (!(if_getflags(ifp) & IFF_NOARP))
4025 				arp_ifinit(ifp, ifa);
4026 #endif
4027 		} else
4028 			err = ether_ioctl(ifp, command, data);
4029 		break;
4030 	case SIOCSIFMTU:
4031 		CTX_LOCK(ctx);
4032 		if (ifr->ifr_mtu == if_getmtu(ifp)) {
4033 			CTX_UNLOCK(ctx);
4034 			break;
4035 		}
4036 		bits = if_getdrvflags(ifp);
4037 		/* stop the driver and free any clusters before proceeding */
4038 		iflib_stop(ctx);
4039 
4040 		if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) {
4041 			STATE_LOCK(ctx);
4042 			if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size)
4043 				ctx->ifc_flags |= IFC_MULTISEG;
4044 			else
4045 				ctx->ifc_flags &= ~IFC_MULTISEG;
4046 			STATE_UNLOCK(ctx);
4047 			err = if_setmtu(ifp, ifr->ifr_mtu);
4048 		}
4049 		iflib_init_locked(ctx);
4050 		STATE_LOCK(ctx);
4051 		if_setdrvflags(ifp, bits);
4052 		STATE_UNLOCK(ctx);
4053 		CTX_UNLOCK(ctx);
4054 		break;
4055 	case SIOCSIFFLAGS:
4056 		CTX_LOCK(ctx);
4057 		if (if_getflags(ifp) & IFF_UP) {
4058 			if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4059 				if ((if_getflags(ifp) ^ ctx->ifc_if_flags) &
4060 				    (IFF_PROMISC | IFF_ALLMULTI)) {
4061 					err = IFDI_PROMISC_SET(ctx, if_getflags(ifp));
4062 				}
4063 			} else
4064 				reinit = 1;
4065 		} else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4066 			iflib_stop(ctx);
4067 		}
4068 		ctx->ifc_if_flags = if_getflags(ifp);
4069 		CTX_UNLOCK(ctx);
4070 		break;
4071 	case SIOCADDMULTI:
4072 	case SIOCDELMULTI:
4073 		if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4074 			CTX_LOCK(ctx);
4075 			IFDI_INTR_DISABLE(ctx);
4076 			IFDI_MULTI_SET(ctx);
4077 			IFDI_INTR_ENABLE(ctx);
4078 			CTX_UNLOCK(ctx);
4079 		}
4080 		break;
4081 	case SIOCSIFMEDIA:
4082 		CTX_LOCK(ctx);
4083 		IFDI_MEDIA_SET(ctx);
4084 		CTX_UNLOCK(ctx);
4085 		/* falls thru */
4086 	case SIOCGIFMEDIA:
4087 	case SIOCGIFXMEDIA:
4088 		err = ifmedia_ioctl(ifp, ifr, &ctx->ifc_media, command);
4089 		break;
4090 	case SIOCGI2C:
4091 	{
4092 		struct ifi2creq i2c;
4093 
4094 		err = copyin(ifr_data_get_ptr(ifr), &i2c, sizeof(i2c));
4095 		if (err != 0)
4096 			break;
4097 		if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) {
4098 			err = EINVAL;
4099 			break;
4100 		}
4101 		if (i2c.len > sizeof(i2c.data)) {
4102 			err = EINVAL;
4103 			break;
4104 		}
4105 
4106 		if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0)
4107 			err = copyout(&i2c, ifr_data_get_ptr(ifr),
4108 			    sizeof(i2c));
4109 		break;
4110 	}
4111 	case SIOCSIFCAP:
4112 	{
4113 		int mask, setmask, oldmask;
4114 
4115 		oldmask = if_getcapenable(ifp);
4116 		mask = ifr->ifr_reqcap ^ oldmask;
4117 		mask &= ctx->ifc_softc_ctx.isc_capabilities;
4118 		setmask = 0;
4119 #ifdef TCP_OFFLOAD
4120 		setmask |= mask & (IFCAP_TOE4|IFCAP_TOE6);
4121 #endif
4122 		setmask |= (mask & IFCAP_FLAGS);
4123 		setmask |= (mask & IFCAP_WOL);
4124 
4125 		/*
4126 		 * If any RX csum has changed, change all the ones that
4127 		 * are supported by the driver.
4128 		 */
4129 		if (setmask & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6)) {
4130 			setmask |= ctx->ifc_softc_ctx.isc_capabilities &
4131 			    (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6);
4132 		}
4133 
4134 		/*
4135 		 * want to ensure that traffic has stopped before we change any of the flags
4136 		 */
4137 		if (setmask) {
4138 			CTX_LOCK(ctx);
4139 			bits = if_getdrvflags(ifp);
4140 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4141 				iflib_stop(ctx);
4142 			STATE_LOCK(ctx);
4143 			if_togglecapenable(ifp, setmask);
4144 			STATE_UNLOCK(ctx);
4145 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4146 				iflib_init_locked(ctx);
4147 			STATE_LOCK(ctx);
4148 			if_setdrvflags(ifp, bits);
4149 			STATE_UNLOCK(ctx);
4150 			CTX_UNLOCK(ctx);
4151 		}
4152 		if_vlancap(ifp);
4153 		break;
4154 	}
4155 	case SIOCGPRIVATE_0:
4156 	case SIOCSDRVSPEC:
4157 	case SIOCGDRVSPEC:
4158 		CTX_LOCK(ctx);
4159 		err = IFDI_PRIV_IOCTL(ctx, command, data);
4160 		CTX_UNLOCK(ctx);
4161 		break;
4162 	default:
4163 		err = ether_ioctl(ifp, command, data);
4164 		break;
4165 	}
4166 	if (reinit)
4167 		iflib_if_init(ctx);
4168 	return (err);
4169 }
4170 
4171 static uint64_t
4172 iflib_if_get_counter(if_t ifp, ift_counter cnt)
4173 {
4174 	if_ctx_t ctx = if_getsoftc(ifp);
4175 
4176 	return (IFDI_GET_COUNTER(ctx, cnt));
4177 }
4178 
4179 /*********************************************************************
4180  *
4181  *  OTHER FUNCTIONS EXPORTED TO THE STACK
4182  *
4183  **********************************************************************/
4184 
4185 static void
4186 iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag)
4187 {
4188 	if_ctx_t ctx = if_getsoftc(ifp);
4189 
4190 	if ((void *)ctx != arg)
4191 		return;
4192 
4193 	if ((vtag == 0) || (vtag > 4095))
4194 		return;
4195 
4196 	CTX_LOCK(ctx);
4197 	IFDI_VLAN_REGISTER(ctx, vtag);
4198 	/* Re-init to load the changes */
4199 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
4200 		iflib_if_init_locked(ctx);
4201 	CTX_UNLOCK(ctx);
4202 }
4203 
4204 static void
4205 iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag)
4206 {
4207 	if_ctx_t ctx = if_getsoftc(ifp);
4208 
4209 	if ((void *)ctx != arg)
4210 		return;
4211 
4212 	if ((vtag == 0) || (vtag > 4095))
4213 		return;
4214 
4215 	CTX_LOCK(ctx);
4216 	IFDI_VLAN_UNREGISTER(ctx, vtag);
4217 	/* Re-init to load the changes */
4218 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
4219 		iflib_if_init_locked(ctx);
4220 	CTX_UNLOCK(ctx);
4221 }
4222 
4223 static void
4224 iflib_led_func(void *arg, int onoff)
4225 {
4226 	if_ctx_t ctx = arg;
4227 
4228 	CTX_LOCK(ctx);
4229 	IFDI_LED_FUNC(ctx, onoff);
4230 	CTX_UNLOCK(ctx);
4231 }
4232 
4233 /*********************************************************************
4234  *
4235  *  BUS FUNCTION DEFINITIONS
4236  *
4237  **********************************************************************/
4238 
4239 int
4240 iflib_device_probe(device_t dev)
4241 {
4242 	pci_vendor_info_t *ent;
4243 
4244 	uint16_t	pci_vendor_id, pci_device_id;
4245 	uint16_t	pci_subvendor_id, pci_subdevice_id;
4246 	uint16_t	pci_rev_id;
4247 	if_shared_ctx_t sctx;
4248 
4249 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4250 		return (ENOTSUP);
4251 
4252 	pci_vendor_id = pci_get_vendor(dev);
4253 	pci_device_id = pci_get_device(dev);
4254 	pci_subvendor_id = pci_get_subvendor(dev);
4255 	pci_subdevice_id = pci_get_subdevice(dev);
4256 	pci_rev_id = pci_get_revid(dev);
4257 	if (sctx->isc_parse_devinfo != NULL)
4258 		sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id);
4259 
4260 	ent = sctx->isc_vendor_info;
4261 	while (ent->pvi_vendor_id != 0) {
4262 		if (pci_vendor_id != ent->pvi_vendor_id) {
4263 			ent++;
4264 			continue;
4265 		}
4266 		if ((pci_device_id == ent->pvi_device_id) &&
4267 		    ((pci_subvendor_id == ent->pvi_subvendor_id) ||
4268 		     (ent->pvi_subvendor_id == 0)) &&
4269 		    ((pci_subdevice_id == ent->pvi_subdevice_id) ||
4270 		     (ent->pvi_subdevice_id == 0)) &&
4271 		    ((pci_rev_id == ent->pvi_rev_id) ||
4272 		     (ent->pvi_rev_id == 0))) {
4273 
4274 			device_set_desc_copy(dev, ent->pvi_name);
4275 			/* this needs to be changed to zero if the bus probing code
4276 			 * ever stops re-probing on best match because the sctx
4277 			 * may have its values over written by register calls
4278 			 * in subsequent probes
4279 			 */
4280 			return (BUS_PROBE_DEFAULT);
4281 		}
4282 		ent++;
4283 	}
4284 	return (ENXIO);
4285 }
4286 
4287 static void
4288 iflib_reset_qvalues(if_ctx_t ctx)
4289 {
4290 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4291 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4292 	device_t dev = ctx->ifc_dev;
4293 	int i;
4294 
4295 	scctx->isc_txrx_budget_bytes_max = IFLIB_MAX_TX_BYTES;
4296 	scctx->isc_tx_qdepth = IFLIB_DEFAULT_TX_QDEPTH;
4297 	/*
4298 	 * XXX sanity check that ntxd & nrxd are a power of 2
4299 	 */
4300 	if (ctx->ifc_sysctl_ntxqs != 0)
4301 		scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs;
4302 	if (ctx->ifc_sysctl_nrxqs != 0)
4303 		scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs;
4304 
4305 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4306 		if (ctx->ifc_sysctl_ntxds[i] != 0)
4307 			scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i];
4308 		else
4309 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4310 	}
4311 
4312 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4313 		if (ctx->ifc_sysctl_nrxds[i] != 0)
4314 			scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i];
4315 		else
4316 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4317 	}
4318 
4319 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4320 		if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) {
4321 			device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n",
4322 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]);
4323 			scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i];
4324 		}
4325 		if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) {
4326 			device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n",
4327 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]);
4328 			scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i];
4329 		}
4330 	}
4331 
4332 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4333 		if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) {
4334 			device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n",
4335 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]);
4336 			scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i];
4337 		}
4338 		if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) {
4339 			device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n",
4340 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]);
4341 			scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i];
4342 		}
4343 	}
4344 }
4345 
4346 int
4347 iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp)
4348 {
4349 	int err, rid, msix;
4350 	if_ctx_t ctx;
4351 	if_t ifp;
4352 	if_softc_ctx_t scctx;
4353 	int i;
4354 	uint16_t main_txq;
4355 	uint16_t main_rxq;
4356 
4357 
4358 	ctx = malloc(sizeof(* ctx), M_IFLIB, M_WAITOK|M_ZERO);
4359 
4360 	if (sc == NULL) {
4361 		sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4362 		device_set_softc(dev, ctx);
4363 		ctx->ifc_flags |= IFC_SC_ALLOCATED;
4364 	}
4365 
4366 	ctx->ifc_sctx = sctx;
4367 	ctx->ifc_dev = dev;
4368 	ctx->ifc_softc = sc;
4369 
4370 	if ((err = iflib_register(ctx)) != 0) {
4371 		if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4372 			free(sc, M_IFLIB);
4373 		free(ctx, M_IFLIB);
4374 		device_printf(dev, "iflib_register failed %d\n", err);
4375 		return (err);
4376 	}
4377 	iflib_add_device_sysctl_pre(ctx);
4378 
4379 	scctx = &ctx->ifc_softc_ctx;
4380 	ifp = ctx->ifc_ifp;
4381 
4382 	iflib_reset_qvalues(ctx);
4383 	CTX_LOCK(ctx);
4384 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4385 		CTX_UNLOCK(ctx);
4386 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4387 		return (err);
4388 	}
4389 	_iflib_pre_assert(scctx);
4390 	ctx->ifc_txrx = *scctx->isc_txrx;
4391 
4392 #ifdef INVARIANTS
4393 	MPASS(scctx->isc_capabilities);
4394 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
4395 		MPASS(scctx->isc_tx_csum_flags);
4396 #endif
4397 
4398 	if_setcapabilities(ifp, scctx->isc_capabilities | IFCAP_HWSTATS);
4399 	if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS);
4400 
4401 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4402 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4403 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4404 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4405 
4406 	main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4407 	main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4408 
4409 	/* XXX change for per-queue sizes */
4410 	device_printf(dev, "using %d tx descriptors and %d rx descriptors\n",
4411 		      scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4412 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4413 		if (!powerof2(scctx->isc_nrxd[i])) {
4414 			/* round down instead? */
4415 			device_printf(dev, "# rx descriptors must be a power of 2\n");
4416 			err = EINVAL;
4417 			goto fail;
4418 		}
4419 	}
4420 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4421 		if (!powerof2(scctx->isc_ntxd[i])) {
4422 			device_printf(dev,
4423 			    "# tx descriptors must be a power of 2");
4424 			err = EINVAL;
4425 			goto fail;
4426 		}
4427 	}
4428 
4429 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4430 	    MAX_SINGLE_PACKET_FRACTION)
4431 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4432 		    MAX_SINGLE_PACKET_FRACTION);
4433 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4434 	    MAX_SINGLE_PACKET_FRACTION)
4435 		scctx->isc_tx_tso_segments_max = max(1,
4436 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4437 
4438 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4439 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
4440 		/*
4441 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
4442 		 * but some MACs do.
4443 		 */
4444 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
4445 		    IP_MAXPACKET));
4446 		/*
4447 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
4448 		 * into account.  In the worst case, each of these calls will
4449 		 * add another mbuf and, thus, the requirement for another DMA
4450 		 * segment.  So for best performance, it doesn't make sense to
4451 		 * advertize a maximum of TSO segments that typically will
4452 		 * require defragmentation in iflib_encap().
4453 		 */
4454 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
4455 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
4456 	}
4457 	if (scctx->isc_rss_table_size == 0)
4458 		scctx->isc_rss_table_size = 64;
4459 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4460 
4461 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4462 	/* XXX format name */
4463 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx, -1, "admin");
4464 
4465 	/* Set up cpu set.  If it fails, use the set of all CPUs. */
4466 	if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) != 0) {
4467 		device_printf(dev, "Unable to fetch CPU list\n");
4468 		CPU_COPY(&all_cpus, &ctx->ifc_cpus);
4469 	}
4470 	MPASS(CPU_COUNT(&ctx->ifc_cpus) > 0);
4471 
4472 	/*
4473 	** Now setup MSI or MSI/X, should
4474 	** return us the number of supported
4475 	** vectors. (Will be 1 for MSI)
4476 	*/
4477 	if (sctx->isc_flags & IFLIB_SKIP_MSIX) {
4478 		msix = scctx->isc_vectors;
4479 	} else if (scctx->isc_msix_bar != 0)
4480 	       /*
4481 		* The simple fact that isc_msix_bar is not 0 does not mean we
4482 		* we have a good value there that is known to work.
4483 		*/
4484 		msix = iflib_msix_init(ctx);
4485 	else {
4486 		scctx->isc_vectors = 1;
4487 		scctx->isc_ntxqsets = 1;
4488 		scctx->isc_nrxqsets = 1;
4489 		scctx->isc_intr = IFLIB_INTR_LEGACY;
4490 		msix = 0;
4491 	}
4492 	/* Get memory for the station queues */
4493 	if ((err = iflib_queues_alloc(ctx))) {
4494 		device_printf(dev, "Unable to allocate queue memory\n");
4495 		goto fail;
4496 	}
4497 
4498 	if ((err = iflib_qset_structures_setup(ctx)))
4499 		goto fail_queues;
4500 
4501 	/*
4502 	 * Group taskqueues aren't properly set up until SMP is started,
4503 	 * so we disable interrupts until we can handle them post
4504 	 * SI_SUB_SMP.
4505 	 *
4506 	 * XXX: disabling interrupts doesn't actually work, at least for
4507 	 * the non-MSI case.  When they occur before SI_SUB_SMP completes,
4508 	 * we do null handling and depend on this not causing too large an
4509 	 * interrupt storm.
4510 	 */
4511 	IFDI_INTR_DISABLE(ctx);
4512 	if (msix > 1 && (err = IFDI_MSIX_INTR_ASSIGN(ctx, msix)) != 0) {
4513 		device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n", err);
4514 		goto fail_intr_free;
4515 	}
4516 	if (msix <= 1) {
4517 		rid = 0;
4518 		if (scctx->isc_intr == IFLIB_INTR_MSI) {
4519 			MPASS(msix == 1);
4520 			rid = 1;
4521 		}
4522 		if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) {
4523 			device_printf(dev, "iflib_legacy_setup failed %d\n", err);
4524 			goto fail_intr_free;
4525 		}
4526 	}
4527 
4528 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac);
4529 
4530 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4531 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4532 		goto fail_detach;
4533 	}
4534 
4535 	/*
4536 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4537 	 * This must appear after the call to ether_ifattach() because
4538 	 * ether_ifattach() sets if_hdrlen to the default value.
4539 	 */
4540 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4541 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
4542 
4543 	if ((err = iflib_netmap_attach(ctx))) {
4544 		device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err);
4545 		goto fail_detach;
4546 	}
4547 	*ctxp = ctx;
4548 
4549 	NETDUMP_SET(ctx->ifc_ifp, iflib);
4550 
4551 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4552 	iflib_add_device_sysctl_post(ctx);
4553 	ctx->ifc_flags |= IFC_INIT_DONE;
4554 	CTX_UNLOCK(ctx);
4555 	return (0);
4556 
4557 fail_detach:
4558 	ether_ifdetach(ctx->ifc_ifp);
4559 fail_intr_free:
4560 fail_queues:
4561 	iflib_tx_structures_free(ctx);
4562 	iflib_rx_structures_free(ctx);
4563 fail:
4564 	iflib_free_intr_mem(ctx);
4565 	IFDI_DETACH(ctx);
4566 	CTX_UNLOCK(ctx);
4567 
4568 	return (err);
4569 }
4570 
4571 int
4572 iflib_pseudo_register(device_t dev, if_shared_ctx_t sctx, if_ctx_t *ctxp,
4573 					  struct iflib_cloneattach_ctx *clctx)
4574 {
4575 	int err;
4576 	if_ctx_t ctx;
4577 	if_t ifp;
4578 	if_softc_ctx_t scctx;
4579 	int i;
4580 	void *sc;
4581 	uint16_t main_txq;
4582 	uint16_t main_rxq;
4583 
4584 	ctx = malloc(sizeof(*ctx), M_IFLIB, M_WAITOK|M_ZERO);
4585 	sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4586 	ctx->ifc_flags |= IFC_SC_ALLOCATED;
4587 	if (sctx->isc_flags & (IFLIB_PSEUDO|IFLIB_VIRTUAL))
4588 		ctx->ifc_flags |= IFC_PSEUDO;
4589 
4590 	ctx->ifc_sctx = sctx;
4591 	ctx->ifc_softc = sc;
4592 	ctx->ifc_dev = dev;
4593 
4594 	if ((err = iflib_register(ctx)) != 0) {
4595 		device_printf(dev, "%s: iflib_register failed %d\n", __func__, err);
4596 		free(sc, M_IFLIB);
4597 		free(ctx, M_IFLIB);
4598 		return (err);
4599 	}
4600 	iflib_add_device_sysctl_pre(ctx);
4601 
4602 	scctx = &ctx->ifc_softc_ctx;
4603 	ifp = ctx->ifc_ifp;
4604 
4605 	/*
4606 	 * XXX sanity check that ntxd & nrxd are a power of 2
4607 	 */
4608 	iflib_reset_qvalues(ctx);
4609 
4610 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4611 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4612 		return (err);
4613 	}
4614 	if (sctx->isc_flags & IFLIB_GEN_MAC)
4615 		iflib_gen_mac(ctx);
4616 	if ((err = IFDI_CLONEATTACH(ctx, clctx->cc_ifc, clctx->cc_name,
4617 								clctx->cc_params)) != 0) {
4618 		device_printf(dev, "IFDI_CLONEATTACH failed %d\n", err);
4619 		return (err);
4620 	}
4621 	ifmedia_add(&ctx->ifc_media, IFM_ETHER | IFM_1000_T | IFM_FDX, 0, NULL);
4622 	ifmedia_add(&ctx->ifc_media, IFM_ETHER | IFM_AUTO, 0, NULL);
4623 	ifmedia_set(&ctx->ifc_media, IFM_ETHER | IFM_AUTO);
4624 
4625 #ifdef INVARIANTS
4626 	MPASS(scctx->isc_capabilities);
4627 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
4628 		MPASS(scctx->isc_tx_csum_flags);
4629 #endif
4630 
4631 	if_setcapabilities(ifp, scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_LINKSTATE);
4632 	if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_LINKSTATE);
4633 
4634 	ifp->if_flags |= IFF_NOGROUP;
4635 	if (sctx->isc_flags & IFLIB_PSEUDO) {
4636 		ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac);
4637 
4638 		if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4639 			device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4640 			goto fail_detach;
4641 		}
4642 		*ctxp = ctx;
4643 
4644 		/*
4645 		 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4646 		 * This must appear after the call to ether_ifattach() because
4647 		 * ether_ifattach() sets if_hdrlen to the default value.
4648 		 */
4649 		if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4650 			if_setifheaderlen(ifp,
4651 			    sizeof(struct ether_vlan_header));
4652 
4653 		if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4654 		iflib_add_device_sysctl_post(ctx);
4655 		ctx->ifc_flags |= IFC_INIT_DONE;
4656 		return (0);
4657 	}
4658 	_iflib_pre_assert(scctx);
4659 	ctx->ifc_txrx = *scctx->isc_txrx;
4660 
4661 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4662 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4663 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4664 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4665 
4666 	main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4667 	main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4668 
4669 	/* XXX change for per-queue sizes */
4670 	device_printf(dev, "using %d tx descriptors and %d rx descriptors\n",
4671 		      scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4672 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4673 		if (!powerof2(scctx->isc_nrxd[i])) {
4674 			/* round down instead? */
4675 			device_printf(dev, "# rx descriptors must be a power of 2\n");
4676 			err = EINVAL;
4677 			goto fail;
4678 		}
4679 	}
4680 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4681 		if (!powerof2(scctx->isc_ntxd[i])) {
4682 			device_printf(dev,
4683 			    "# tx descriptors must be a power of 2");
4684 			err = EINVAL;
4685 			goto fail;
4686 		}
4687 	}
4688 
4689 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4690 	    MAX_SINGLE_PACKET_FRACTION)
4691 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4692 		    MAX_SINGLE_PACKET_FRACTION);
4693 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4694 	    MAX_SINGLE_PACKET_FRACTION)
4695 		scctx->isc_tx_tso_segments_max = max(1,
4696 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4697 
4698 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4699 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
4700 		/*
4701 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
4702 		 * but some MACs do.
4703 		 */
4704 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
4705 		    IP_MAXPACKET));
4706 		/*
4707 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
4708 		 * into account.  In the worst case, each of these calls will
4709 		 * add another mbuf and, thus, the requirement for another DMA
4710 		 * segment.  So for best performance, it doesn't make sense to
4711 		 * advertize a maximum of TSO segments that typically will
4712 		 * require defragmentation in iflib_encap().
4713 		 */
4714 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
4715 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
4716 	}
4717 	if (scctx->isc_rss_table_size == 0)
4718 		scctx->isc_rss_table_size = 64;
4719 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4720 
4721 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4722 	/* XXX format name */
4723 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx, -1, "admin");
4724 
4725 	/* XXX --- can support > 1 -- but keep it simple for now */
4726 	scctx->isc_intr = IFLIB_INTR_LEGACY;
4727 
4728 	/* Get memory for the station queues */
4729 	if ((err = iflib_queues_alloc(ctx))) {
4730 		device_printf(dev, "Unable to allocate queue memory\n");
4731 		goto fail;
4732 	}
4733 
4734 	if ((err = iflib_qset_structures_setup(ctx))) {
4735 		device_printf(dev, "qset structure setup failed %d\n", err);
4736 		goto fail_queues;
4737 	}
4738 
4739 	/*
4740 	 * XXX What if anything do we want to do about interrupts?
4741 	 */
4742 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac);
4743 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4744 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4745 		goto fail_detach;
4746 	}
4747 
4748 	/*
4749 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4750 	 * This must appear after the call to ether_ifattach() because
4751 	 * ether_ifattach() sets if_hdrlen to the default value.
4752 	 */
4753 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4754 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
4755 
4756 	/* XXX handle more than one queue */
4757 	for (i = 0; i < scctx->isc_nrxqsets; i++)
4758 		IFDI_RX_CLSET(ctx, 0, i, ctx->ifc_rxqs[i].ifr_fl[0].ifl_sds.ifsd_cl);
4759 
4760 	*ctxp = ctx;
4761 
4762 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4763 	iflib_add_device_sysctl_post(ctx);
4764 	ctx->ifc_flags |= IFC_INIT_DONE;
4765 	return (0);
4766 fail_detach:
4767 	ether_ifdetach(ctx->ifc_ifp);
4768 fail_queues:
4769 	iflib_tx_structures_free(ctx);
4770 	iflib_rx_structures_free(ctx);
4771 fail:
4772 	IFDI_DETACH(ctx);
4773 	return (err);
4774 }
4775 
4776 int
4777 iflib_pseudo_deregister(if_ctx_t ctx)
4778 {
4779 	if_t ifp = ctx->ifc_ifp;
4780 	iflib_txq_t txq;
4781 	iflib_rxq_t rxq;
4782 	int i, j;
4783 	struct taskqgroup *tqg;
4784 	iflib_fl_t fl;
4785 
4786 	/* Unregister VLAN events */
4787 	if (ctx->ifc_vlan_attach_event != NULL)
4788 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
4789 	if (ctx->ifc_vlan_detach_event != NULL)
4790 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
4791 
4792 	ether_ifdetach(ifp);
4793 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
4794 	CTX_LOCK_DESTROY(ctx);
4795 	/* XXX drain any dependent tasks */
4796 	tqg = qgroup_if_io_tqg;
4797 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
4798 		callout_drain(&txq->ift_timer);
4799 		if (txq->ift_task.gt_uniq != NULL)
4800 			taskqgroup_detach(tqg, &txq->ift_task);
4801 	}
4802 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4803 		if (rxq->ifr_task.gt_uniq != NULL)
4804 			taskqgroup_detach(tqg, &rxq->ifr_task);
4805 
4806 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
4807 			free(fl->ifl_rx_bitmap, M_IFLIB);
4808 	}
4809 	tqg = qgroup_if_config_tqg;
4810 	if (ctx->ifc_admin_task.gt_uniq != NULL)
4811 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
4812 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
4813 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
4814 
4815 	if_free(ifp);
4816 
4817 	iflib_tx_structures_free(ctx);
4818 	iflib_rx_structures_free(ctx);
4819 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4820 		free(ctx->ifc_softc, M_IFLIB);
4821 	free(ctx, M_IFLIB);
4822 	return (0);
4823 }
4824 
4825 int
4826 iflib_device_attach(device_t dev)
4827 {
4828 	if_ctx_t ctx;
4829 	if_shared_ctx_t sctx;
4830 
4831 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4832 		return (ENOTSUP);
4833 
4834 	pci_enable_busmaster(dev);
4835 
4836 	return (iflib_device_register(dev, NULL, sctx, &ctx));
4837 }
4838 
4839 int
4840 iflib_device_deregister(if_ctx_t ctx)
4841 {
4842 	if_t ifp = ctx->ifc_ifp;
4843 	iflib_txq_t txq;
4844 	iflib_rxq_t rxq;
4845 	device_t dev = ctx->ifc_dev;
4846 	int i, j;
4847 	struct taskqgroup *tqg;
4848 	iflib_fl_t fl;
4849 
4850 	/* Make sure VLANS are not using driver */
4851 	if (if_vlantrunkinuse(ifp)) {
4852 		device_printf(dev, "Vlan in use, detach first\n");
4853 		return (EBUSY);
4854 	}
4855 #ifdef PCI_IOV
4856 	if (!CTX_IS_VF(ctx) && pci_iov_detach(dev) != 0) {
4857 		device_printf(dev, "SR-IOV in use; detach first.\n");
4858 		return (EBUSY);
4859 	}
4860 #endif
4861 
4862 	STATE_LOCK(ctx);
4863 	ctx->ifc_flags |= IFC_IN_DETACH;
4864 	STATE_UNLOCK(ctx);
4865 
4866 	CTX_LOCK(ctx);
4867 	iflib_stop(ctx);
4868 	CTX_UNLOCK(ctx);
4869 
4870 	/* Unregister VLAN events */
4871 	if (ctx->ifc_vlan_attach_event != NULL)
4872 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
4873 	if (ctx->ifc_vlan_detach_event != NULL)
4874 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
4875 
4876 	iflib_netmap_detach(ifp);
4877 	ether_ifdetach(ifp);
4878 	if (ctx->ifc_led_dev != NULL)
4879 		led_destroy(ctx->ifc_led_dev);
4880 	/* XXX drain any dependent tasks */
4881 	tqg = qgroup_if_io_tqg;
4882 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
4883 		callout_drain(&txq->ift_timer);
4884 		if (txq->ift_task.gt_uniq != NULL)
4885 			taskqgroup_detach(tqg, &txq->ift_task);
4886 	}
4887 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4888 		if (rxq->ifr_task.gt_uniq != NULL)
4889 			taskqgroup_detach(tqg, &rxq->ifr_task);
4890 
4891 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
4892 			free(fl->ifl_rx_bitmap, M_IFLIB);
4893 
4894 	}
4895 	tqg = qgroup_if_config_tqg;
4896 	if (ctx->ifc_admin_task.gt_uniq != NULL)
4897 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
4898 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
4899 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
4900 	CTX_LOCK(ctx);
4901 	IFDI_DETACH(ctx);
4902 	CTX_UNLOCK(ctx);
4903 
4904 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
4905 	CTX_LOCK_DESTROY(ctx);
4906 	device_set_softc(ctx->ifc_dev, NULL);
4907 	iflib_free_intr_mem(ctx);
4908 
4909 	bus_generic_detach(dev);
4910 	if_free(ifp);
4911 
4912 	iflib_tx_structures_free(ctx);
4913 	iflib_rx_structures_free(ctx);
4914 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4915 		free(ctx->ifc_softc, M_IFLIB);
4916 	STATE_LOCK_DESTROY(ctx);
4917 	free(ctx, M_IFLIB);
4918 	return (0);
4919 }
4920 
4921 static void
4922 iflib_free_intr_mem(if_ctx_t ctx)
4923 {
4924 
4925 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) {
4926 		pci_release_msi(ctx->ifc_dev);
4927 	}
4928 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) {
4929 		iflib_irq_free(ctx, &ctx->ifc_legacy_irq);
4930 	}
4931 	if (ctx->ifc_msix_mem != NULL) {
4932 		bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY,
4933 			ctx->ifc_softc_ctx.isc_msix_bar, ctx->ifc_msix_mem);
4934 		ctx->ifc_msix_mem = NULL;
4935 	}
4936 }
4937 
4938 int
4939 iflib_device_detach(device_t dev)
4940 {
4941 	if_ctx_t ctx = device_get_softc(dev);
4942 
4943 	return (iflib_device_deregister(ctx));
4944 }
4945 
4946 int
4947 iflib_device_suspend(device_t dev)
4948 {
4949 	if_ctx_t ctx = device_get_softc(dev);
4950 
4951 	CTX_LOCK(ctx);
4952 	IFDI_SUSPEND(ctx);
4953 	CTX_UNLOCK(ctx);
4954 
4955 	return bus_generic_suspend(dev);
4956 }
4957 int
4958 iflib_device_shutdown(device_t dev)
4959 {
4960 	if_ctx_t ctx = device_get_softc(dev);
4961 
4962 	CTX_LOCK(ctx);
4963 	IFDI_SHUTDOWN(ctx);
4964 	CTX_UNLOCK(ctx);
4965 
4966 	return bus_generic_suspend(dev);
4967 }
4968 
4969 
4970 int
4971 iflib_device_resume(device_t dev)
4972 {
4973 	if_ctx_t ctx = device_get_softc(dev);
4974 	iflib_txq_t txq = ctx->ifc_txqs;
4975 
4976 	CTX_LOCK(ctx);
4977 	IFDI_RESUME(ctx);
4978 	iflib_if_init_locked(ctx);
4979 	CTX_UNLOCK(ctx);
4980 	for (int i = 0; i < NTXQSETS(ctx); i++, txq++)
4981 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
4982 
4983 	return (bus_generic_resume(dev));
4984 }
4985 
4986 int
4987 iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params)
4988 {
4989 	int error;
4990 	if_ctx_t ctx = device_get_softc(dev);
4991 
4992 	CTX_LOCK(ctx);
4993 	error = IFDI_IOV_INIT(ctx, num_vfs, params);
4994 	CTX_UNLOCK(ctx);
4995 
4996 	return (error);
4997 }
4998 
4999 void
5000 iflib_device_iov_uninit(device_t dev)
5001 {
5002 	if_ctx_t ctx = device_get_softc(dev);
5003 
5004 	CTX_LOCK(ctx);
5005 	IFDI_IOV_UNINIT(ctx);
5006 	CTX_UNLOCK(ctx);
5007 }
5008 
5009 int
5010 iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params)
5011 {
5012 	int error;
5013 	if_ctx_t ctx = device_get_softc(dev);
5014 
5015 	CTX_LOCK(ctx);
5016 	error = IFDI_IOV_VF_ADD(ctx, vfnum, params);
5017 	CTX_UNLOCK(ctx);
5018 
5019 	return (error);
5020 }
5021 
5022 /*********************************************************************
5023  *
5024  *  MODULE FUNCTION DEFINITIONS
5025  *
5026  **********************************************************************/
5027 
5028 /*
5029  * - Start a fast taskqueue thread for each core
5030  * - Start a taskqueue for control operations
5031  */
5032 static int
5033 iflib_module_init(void)
5034 {
5035 	return (0);
5036 }
5037 
5038 static int
5039 iflib_module_event_handler(module_t mod, int what, void *arg)
5040 {
5041 	int err;
5042 
5043 	switch (what) {
5044 	case MOD_LOAD:
5045 		if ((err = iflib_module_init()) != 0)
5046 			return (err);
5047 		break;
5048 	case MOD_UNLOAD:
5049 		return (EBUSY);
5050 	default:
5051 		return (EOPNOTSUPP);
5052 	}
5053 
5054 	return (0);
5055 }
5056 
5057 /*********************************************************************
5058  *
5059  *  PUBLIC FUNCTION DEFINITIONS
5060  *     ordered as in iflib.h
5061  *
5062  **********************************************************************/
5063 
5064 
5065 static void
5066 _iflib_assert(if_shared_ctx_t sctx)
5067 {
5068 	MPASS(sctx->isc_tx_maxsize);
5069 	MPASS(sctx->isc_tx_maxsegsize);
5070 
5071 	MPASS(sctx->isc_rx_maxsize);
5072 	MPASS(sctx->isc_rx_nsegments);
5073 	MPASS(sctx->isc_rx_maxsegsize);
5074 
5075 	MPASS(sctx->isc_nrxd_min[0]);
5076 	MPASS(sctx->isc_nrxd_max[0]);
5077 	MPASS(sctx->isc_nrxd_default[0]);
5078 	MPASS(sctx->isc_ntxd_min[0]);
5079 	MPASS(sctx->isc_ntxd_max[0]);
5080 	MPASS(sctx->isc_ntxd_default[0]);
5081 }
5082 
5083 static void
5084 _iflib_pre_assert(if_softc_ctx_t scctx)
5085 {
5086 
5087 	MPASS(scctx->isc_txrx->ift_txd_encap);
5088 	MPASS(scctx->isc_txrx->ift_txd_flush);
5089 	MPASS(scctx->isc_txrx->ift_txd_credits_update);
5090 	MPASS(scctx->isc_txrx->ift_rxd_available);
5091 	MPASS(scctx->isc_txrx->ift_rxd_pkt_get);
5092 	MPASS(scctx->isc_txrx->ift_rxd_refill);
5093 	MPASS(scctx->isc_txrx->ift_rxd_flush);
5094 }
5095 
5096 static int
5097 iflib_register(if_ctx_t ctx)
5098 {
5099 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5100 	driver_t *driver = sctx->isc_driver;
5101 	device_t dev = ctx->ifc_dev;
5102 	if_t ifp;
5103 
5104 	_iflib_assert(sctx);
5105 
5106 	CTX_LOCK_INIT(ctx);
5107 	STATE_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev));
5108 	ifp = ctx->ifc_ifp = if_alloc(IFT_ETHER);
5109 	if (ifp == NULL) {
5110 		device_printf(dev, "can not allocate ifnet structure\n");
5111 		return (ENOMEM);
5112 	}
5113 
5114 	/*
5115 	 * Initialize our context's device specific methods
5116 	 */
5117 	kobj_init((kobj_t) ctx, (kobj_class_t) driver);
5118 	kobj_class_compile((kobj_class_t) driver);
5119 	driver->refs++;
5120 
5121 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
5122 	if_setsoftc(ifp, ctx);
5123 	if_setdev(ifp, dev);
5124 	if_setinitfn(ifp, iflib_if_init);
5125 	if_setioctlfn(ifp, iflib_if_ioctl);
5126 #ifdef ALTQ
5127 	if_setstartfn(ifp, iflib_altq_if_start);
5128 	if_settransmitfn(ifp, iflib_altq_if_transmit);
5129 	if_setsendqready(ifp);
5130 #else
5131 	if_settransmitfn(ifp, iflib_if_transmit);
5132 #endif
5133 	if_setqflushfn(ifp, iflib_if_qflush);
5134 	if_setflags(ifp, IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST);
5135 
5136 	ctx->ifc_vlan_attach_event =
5137 		EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx,
5138 							  EVENTHANDLER_PRI_FIRST);
5139 	ctx->ifc_vlan_detach_event =
5140 		EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx,
5141 							  EVENTHANDLER_PRI_FIRST);
5142 
5143 	ifmedia_init(&ctx->ifc_media, IFM_IMASK,
5144 					 iflib_media_change, iflib_media_status);
5145 
5146 	return (0);
5147 }
5148 
5149 
5150 static int
5151 iflib_queues_alloc(if_ctx_t ctx)
5152 {
5153 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5154 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5155 	device_t dev = ctx->ifc_dev;
5156 	int nrxqsets = scctx->isc_nrxqsets;
5157 	int ntxqsets = scctx->isc_ntxqsets;
5158 	iflib_txq_t txq;
5159 	iflib_rxq_t rxq;
5160 	iflib_fl_t fl = NULL;
5161 	int i, j, cpu, err, txconf, rxconf;
5162 	iflib_dma_info_t ifdip;
5163 	uint32_t *rxqsizes = scctx->isc_rxqsizes;
5164 	uint32_t *txqsizes = scctx->isc_txqsizes;
5165 	uint8_t nrxqs = sctx->isc_nrxqs;
5166 	uint8_t ntxqs = sctx->isc_ntxqs;
5167 	int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1;
5168 	caddr_t *vaddrs;
5169 	uint64_t *paddrs;
5170 
5171 	KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1"));
5172 	KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1"));
5173 
5174 	/* Allocate the TX ring struct memory */
5175 	if (!(ctx->ifc_txqs =
5176 	    (iflib_txq_t) malloc(sizeof(struct iflib_txq) *
5177 	    ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5178 		device_printf(dev, "Unable to allocate TX ring memory\n");
5179 		err = ENOMEM;
5180 		goto fail;
5181 	}
5182 
5183 	/* Now allocate the RX */
5184 	if (!(ctx->ifc_rxqs =
5185 	    (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) *
5186 	    nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5187 		device_printf(dev, "Unable to allocate RX ring memory\n");
5188 		err = ENOMEM;
5189 		goto rx_fail;
5190 	}
5191 
5192 	txq = ctx->ifc_txqs;
5193 	rxq = ctx->ifc_rxqs;
5194 
5195 	/*
5196 	 * XXX handle allocation failure
5197 	 */
5198 	for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) {
5199 		/* Set up some basics */
5200 
5201 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs, M_IFLIB, M_WAITOK|M_ZERO)) == NULL) {
5202 			device_printf(dev, "failed to allocate iflib_dma_info\n");
5203 			err = ENOMEM;
5204 			goto err_tx_desc;
5205 		}
5206 		txq->ift_ifdi = ifdip;
5207 		for (j = 0; j < ntxqs; j++, ifdip++) {
5208 			if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, BUS_DMA_NOWAIT)) {
5209 				device_printf(dev, "Unable to allocate Descriptor memory\n");
5210 				err = ENOMEM;
5211 				goto err_tx_desc;
5212 			}
5213 			txq->ift_txd_size[j] = scctx->isc_txd_size[j];
5214 			bzero((void *)ifdip->idi_vaddr, txqsizes[j]);
5215 		}
5216 		txq->ift_ctx = ctx;
5217 		txq->ift_id = i;
5218 		if (sctx->isc_flags & IFLIB_HAS_TXCQ) {
5219 			txq->ift_br_offset = 1;
5220 		} else {
5221 			txq->ift_br_offset = 0;
5222 		}
5223 		/* XXX fix this */
5224 		txq->ift_timer.c_cpu = cpu;
5225 
5226 		if (iflib_txsd_alloc(txq)) {
5227 			device_printf(dev, "Critical Failure setting up TX buffers\n");
5228 			err = ENOMEM;
5229 			goto err_tx_desc;
5230 		}
5231 
5232 		/* Initialize the TX lock */
5233 		snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:tx(%d):callout",
5234 		    device_get_nameunit(dev), txq->ift_id);
5235 		mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF);
5236 		callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0);
5237 
5238 		snprintf(txq->ift_db_mtx_name, MTX_NAME_LEN, "%s:tx(%d):db",
5239 			 device_get_nameunit(dev), txq->ift_id);
5240 
5241 		err = ifmp_ring_alloc(&txq->ift_br, 2048, txq, iflib_txq_drain,
5242 				      iflib_txq_can_drain, M_IFLIB, M_WAITOK);
5243 		if (err) {
5244 			/* XXX free any allocated rings */
5245 			device_printf(dev, "Unable to allocate buf_ring\n");
5246 			goto err_tx_desc;
5247 		}
5248 	}
5249 
5250 	for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) {
5251 		/* Set up some basics */
5252 
5253 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs, M_IFLIB, M_WAITOK|M_ZERO)) == NULL) {
5254 			device_printf(dev, "failed to allocate iflib_dma_info\n");
5255 			err = ENOMEM;
5256 			goto err_tx_desc;
5257 		}
5258 
5259 		rxq->ifr_ifdi = ifdip;
5260 		/* XXX this needs to be changed if #rx queues != #tx queues */
5261 		rxq->ifr_ntxqirq = 1;
5262 		rxq->ifr_txqid[0] = i;
5263 		for (j = 0; j < nrxqs; j++, ifdip++) {
5264 			if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, BUS_DMA_NOWAIT)) {
5265 				device_printf(dev, "Unable to allocate Descriptor memory\n");
5266 				err = ENOMEM;
5267 				goto err_tx_desc;
5268 			}
5269 			bzero((void *)ifdip->idi_vaddr, rxqsizes[j]);
5270 		}
5271 		rxq->ifr_ctx = ctx;
5272 		rxq->ifr_id = i;
5273 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
5274 			rxq->ifr_fl_offset = 1;
5275 		} else {
5276 			rxq->ifr_fl_offset = 0;
5277 		}
5278 		rxq->ifr_nfl = nfree_lists;
5279 		if (!(fl =
5280 			  (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) {
5281 			device_printf(dev, "Unable to allocate free list memory\n");
5282 			err = ENOMEM;
5283 			goto err_tx_desc;
5284 		}
5285 		rxq->ifr_fl = fl;
5286 		for (j = 0; j < nfree_lists; j++) {
5287 			fl[j].ifl_rxq = rxq;
5288 			fl[j].ifl_id = j;
5289 			fl[j].ifl_ifdi = &rxq->ifr_ifdi[j + rxq->ifr_fl_offset];
5290 			fl[j].ifl_rxd_size = scctx->isc_rxd_size[j];
5291 		}
5292 		/* Allocate receive buffers for the ring */
5293 		if (iflib_rxsd_alloc(rxq)) {
5294 			device_printf(dev,
5295 			    "Critical Failure setting up receive buffers\n");
5296 			err = ENOMEM;
5297 			goto err_rx_desc;
5298 		}
5299 
5300 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5301 			fl->ifl_rx_bitmap = bit_alloc(fl->ifl_size, M_IFLIB, M_WAITOK|M_ZERO);
5302 	}
5303 
5304 	/* TXQs */
5305 	vaddrs = malloc(sizeof(caddr_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5306 	paddrs = malloc(sizeof(uint64_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5307 	for (i = 0; i < ntxqsets; i++) {
5308 		iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi;
5309 
5310 		for (j = 0; j < ntxqs; j++, di++) {
5311 			vaddrs[i*ntxqs + j] = di->idi_vaddr;
5312 			paddrs[i*ntxqs + j] = di->idi_paddr;
5313 		}
5314 	}
5315 	if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) {
5316 		device_printf(ctx->ifc_dev, "device queue allocation failed\n");
5317 		iflib_tx_structures_free(ctx);
5318 		free(vaddrs, M_IFLIB);
5319 		free(paddrs, M_IFLIB);
5320 		goto err_rx_desc;
5321 	}
5322 	free(vaddrs, M_IFLIB);
5323 	free(paddrs, M_IFLIB);
5324 
5325 	/* RXQs */
5326 	vaddrs = malloc(sizeof(caddr_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5327 	paddrs = malloc(sizeof(uint64_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5328 	for (i = 0; i < nrxqsets; i++) {
5329 		iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi;
5330 
5331 		for (j = 0; j < nrxqs; j++, di++) {
5332 			vaddrs[i*nrxqs + j] = di->idi_vaddr;
5333 			paddrs[i*nrxqs + j] = di->idi_paddr;
5334 		}
5335 	}
5336 	if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) {
5337 		device_printf(ctx->ifc_dev, "device queue allocation failed\n");
5338 		iflib_tx_structures_free(ctx);
5339 		free(vaddrs, M_IFLIB);
5340 		free(paddrs, M_IFLIB);
5341 		goto err_rx_desc;
5342 	}
5343 	free(vaddrs, M_IFLIB);
5344 	free(paddrs, M_IFLIB);
5345 
5346 	return (0);
5347 
5348 /* XXX handle allocation failure changes */
5349 err_rx_desc:
5350 err_tx_desc:
5351 rx_fail:
5352 	if (ctx->ifc_rxqs != NULL)
5353 		free(ctx->ifc_rxqs, M_IFLIB);
5354 	ctx->ifc_rxqs = NULL;
5355 	if (ctx->ifc_txqs != NULL)
5356 		free(ctx->ifc_txqs, M_IFLIB);
5357 	ctx->ifc_txqs = NULL;
5358 fail:
5359 	return (err);
5360 }
5361 
5362 static int
5363 iflib_tx_structures_setup(if_ctx_t ctx)
5364 {
5365 	iflib_txq_t txq = ctx->ifc_txqs;
5366 	int i;
5367 
5368 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
5369 		iflib_txq_setup(txq);
5370 
5371 	return (0);
5372 }
5373 
5374 static void
5375 iflib_tx_structures_free(if_ctx_t ctx)
5376 {
5377 	iflib_txq_t txq = ctx->ifc_txqs;
5378 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5379 	int i, j;
5380 
5381 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
5382 		iflib_txq_destroy(txq);
5383 		for (j = 0; j < sctx->isc_ntxqs; j++)
5384 			iflib_dma_free(&txq->ift_ifdi[j]);
5385 	}
5386 	free(ctx->ifc_txqs, M_IFLIB);
5387 	ctx->ifc_txqs = NULL;
5388 	IFDI_QUEUES_FREE(ctx);
5389 }
5390 
5391 /*********************************************************************
5392  *
5393  *  Initialize all receive rings.
5394  *
5395  **********************************************************************/
5396 static int
5397 iflib_rx_structures_setup(if_ctx_t ctx)
5398 {
5399 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5400 	int q;
5401 #if defined(INET6) || defined(INET)
5402 	int i, err;
5403 #endif
5404 
5405 	for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) {
5406 #if defined(INET6) || defined(INET)
5407 		tcp_lro_free(&rxq->ifr_lc);
5408 		if ((err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp,
5409 		    TCP_LRO_ENTRIES, min(1024,
5410 		    ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset]))) != 0) {
5411 			device_printf(ctx->ifc_dev, "LRO Initialization failed!\n");
5412 			goto fail;
5413 		}
5414 		rxq->ifr_lro_enabled = TRUE;
5415 #endif
5416 		IFDI_RXQ_SETUP(ctx, rxq->ifr_id);
5417 	}
5418 	return (0);
5419 #if defined(INET6) || defined(INET)
5420 fail:
5421 	/*
5422 	 * Free RX software descriptors allocated so far, we will only handle
5423 	 * the rings that completed, the failing case will have
5424 	 * cleaned up for itself. 'q' failed, so its the terminus.
5425 	 */
5426 	rxq = ctx->ifc_rxqs;
5427 	for (i = 0; i < q; ++i, rxq++) {
5428 		iflib_rx_sds_free(rxq);
5429 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
5430 	}
5431 	return (err);
5432 #endif
5433 }
5434 
5435 /*********************************************************************
5436  *
5437  *  Free all receive rings.
5438  *
5439  **********************************************************************/
5440 static void
5441 iflib_rx_structures_free(if_ctx_t ctx)
5442 {
5443 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5444 
5445 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) {
5446 		iflib_rx_sds_free(rxq);
5447 	}
5448 	free(ctx->ifc_rxqs, M_IFLIB);
5449 	ctx->ifc_rxqs = NULL;
5450 }
5451 
5452 static int
5453 iflib_qset_structures_setup(if_ctx_t ctx)
5454 {
5455 	int err;
5456 
5457 	/*
5458 	 * It is expected that the caller takes care of freeing queues if this
5459 	 * fails.
5460 	 */
5461 	if ((err = iflib_tx_structures_setup(ctx)) != 0) {
5462 		device_printf(ctx->ifc_dev, "iflib_tx_structures_setup failed: %d\n", err);
5463 		return (err);
5464 	}
5465 
5466 	if ((err = iflib_rx_structures_setup(ctx)) != 0)
5467 		device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err);
5468 
5469 	return (err);
5470 }
5471 
5472 int
5473 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
5474 		driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, const char *name)
5475 {
5476 
5477 	return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name));
5478 }
5479 
5480 #ifdef SMP
5481 static int
5482 find_nth(if_ctx_t ctx, int qid)
5483 {
5484 	cpuset_t cpus;
5485 	int i, cpuid, eqid, count;
5486 
5487 	CPU_COPY(&ctx->ifc_cpus, &cpus);
5488 	count = CPU_COUNT(&cpus);
5489 	eqid = qid % count;
5490 	/* clear up to the qid'th bit */
5491 	for (i = 0; i < eqid; i++) {
5492 		cpuid = CPU_FFS(&cpus);
5493 		MPASS(cpuid != 0);
5494 		CPU_CLR(cpuid-1, &cpus);
5495 	}
5496 	cpuid = CPU_FFS(&cpus);
5497 	MPASS(cpuid != 0);
5498 	return (cpuid-1);
5499 }
5500 
5501 #ifdef SCHED_ULE
5502 extern struct cpu_group *cpu_top;              /* CPU topology */
5503 
5504 static int
5505 find_child_with_core(int cpu, struct cpu_group *grp)
5506 {
5507 	int i;
5508 
5509 	if (grp->cg_children == 0)
5510 		return -1;
5511 
5512 	MPASS(grp->cg_child);
5513 	for (i = 0; i < grp->cg_children; i++) {
5514 		if (CPU_ISSET(cpu, &grp->cg_child[i].cg_mask))
5515 			return i;
5516 	}
5517 
5518 	return -1;
5519 }
5520 
5521 /*
5522  * Find the nth "close" core to the specified core
5523  * "close" is defined as the deepest level that shares
5524  * at least an L2 cache.  With threads, this will be
5525  * threads on the same core.  If the sahred cache is L3
5526  * or higher, simply returns the same core.
5527  */
5528 static int
5529 find_close_core(int cpu, int core_offset)
5530 {
5531 	struct cpu_group *grp;
5532 	int i;
5533 	int fcpu;
5534 	cpuset_t cs;
5535 
5536 	grp = cpu_top;
5537 	if (grp == NULL)
5538 		return cpu;
5539 	i = 0;
5540 	while ((i = find_child_with_core(cpu, grp)) != -1) {
5541 		/* If the child only has one cpu, don't descend */
5542 		if (grp->cg_child[i].cg_count <= 1)
5543 			break;
5544 		grp = &grp->cg_child[i];
5545 	}
5546 
5547 	/* If they don't share at least an L2 cache, use the same CPU */
5548 	if (grp->cg_level > CG_SHARE_L2 || grp->cg_level == CG_SHARE_NONE)
5549 		return cpu;
5550 
5551 	/* Now pick one */
5552 	CPU_COPY(&grp->cg_mask, &cs);
5553 
5554 	/* Add the selected CPU offset to core offset. */
5555 	for (i = 0; (fcpu = CPU_FFS(&cs)) != 0; i++) {
5556 		if (fcpu - 1 == cpu)
5557 			break;
5558 		CPU_CLR(fcpu - 1, &cs);
5559 	}
5560 	MPASS(fcpu);
5561 
5562 	core_offset += i;
5563 
5564 	CPU_COPY(&grp->cg_mask, &cs);
5565 	for (i = core_offset % grp->cg_count; i > 0; i--) {
5566 		MPASS(CPU_FFS(&cs));
5567 		CPU_CLR(CPU_FFS(&cs) - 1, &cs);
5568 	}
5569 	MPASS(CPU_FFS(&cs));
5570 	return CPU_FFS(&cs) - 1;
5571 }
5572 #else
5573 static int
5574 find_close_core(int cpu, int core_offset __unused)
5575 {
5576 	return cpu;
5577 }
5578 #endif
5579 
5580 static int
5581 get_core_offset(if_ctx_t ctx, iflib_intr_type_t type, int qid)
5582 {
5583 	switch (type) {
5584 	case IFLIB_INTR_TX:
5585 		/* TX queues get cores which share at least an L2 cache with the corresponding RX queue */
5586 		/* XXX handle multiple RX threads per core and more than two core per L2 group */
5587 		return qid / CPU_COUNT(&ctx->ifc_cpus) + 1;
5588 	case IFLIB_INTR_RX:
5589 	case IFLIB_INTR_RXTX:
5590 		/* RX queues get the specified core */
5591 		return qid / CPU_COUNT(&ctx->ifc_cpus);
5592 	default:
5593 		return -1;
5594 	}
5595 }
5596 #else
5597 #define get_core_offset(ctx, type, qid)	CPU_FIRST()
5598 #define find_close_core(cpuid, tid)	CPU_FIRST()
5599 #define find_nth(ctx, gid)		CPU_FIRST()
5600 #endif
5601 
5602 /* Just to avoid copy/paste */
5603 static inline int
5604 iflib_irq_set_affinity(if_ctx_t ctx, int irq, iflib_intr_type_t type, int qid,
5605     struct grouptask *gtask, struct taskqgroup *tqg, void *uniq, const char *name)
5606 {
5607 	int cpuid;
5608 	int err, tid;
5609 
5610 	cpuid = find_nth(ctx, qid);
5611 	tid = get_core_offset(ctx, type, qid);
5612 	MPASS(tid >= 0);
5613 	cpuid = find_close_core(cpuid, tid);
5614 	err = taskqgroup_attach_cpu(tqg, gtask, uniq, cpuid, irq, name);
5615 	if (err) {
5616 		device_printf(ctx->ifc_dev, "taskqgroup_attach_cpu failed %d\n", err);
5617 		return (err);
5618 	}
5619 #ifdef notyet
5620 	if (cpuid > ctx->ifc_cpuid_highest)
5621 		ctx->ifc_cpuid_highest = cpuid;
5622 #endif
5623 	return 0;
5624 }
5625 
5626 int
5627 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid,
5628 			iflib_intr_type_t type, driver_filter_t *filter,
5629 			void *filter_arg, int qid, const char *name)
5630 {
5631 	struct grouptask *gtask;
5632 	struct taskqgroup *tqg;
5633 	iflib_filter_info_t info;
5634 	gtask_fn_t *fn;
5635 	int tqrid, err;
5636 	driver_filter_t *intr_fast;
5637 	void *q;
5638 
5639 	info = &ctx->ifc_filter_info;
5640 	tqrid = rid;
5641 
5642 	switch (type) {
5643 	/* XXX merge tx/rx for netmap? */
5644 	case IFLIB_INTR_TX:
5645 		q = &ctx->ifc_txqs[qid];
5646 		info = &ctx->ifc_txqs[qid].ift_filter_info;
5647 		gtask = &ctx->ifc_txqs[qid].ift_task;
5648 		tqg = qgroup_if_io_tqg;
5649 		fn = _task_fn_tx;
5650 		intr_fast = iflib_fast_intr;
5651 		GROUPTASK_INIT(gtask, 0, fn, q);
5652 		ctx->ifc_flags |= IFC_NETMAP_TX_IRQ;
5653 		break;
5654 	case IFLIB_INTR_RX:
5655 		q = &ctx->ifc_rxqs[qid];
5656 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
5657 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5658 		tqg = qgroup_if_io_tqg;
5659 		fn = _task_fn_rx;
5660 		intr_fast = iflib_fast_intr;
5661 		GROUPTASK_INIT(gtask, 0, fn, q);
5662 		break;
5663 	case IFLIB_INTR_RXTX:
5664 		q = &ctx->ifc_rxqs[qid];
5665 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
5666 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5667 		tqg = qgroup_if_io_tqg;
5668 		fn = _task_fn_rx;
5669 		intr_fast = iflib_fast_intr_rxtx;
5670 		GROUPTASK_INIT(gtask, 0, fn, q);
5671 		break;
5672 	case IFLIB_INTR_ADMIN:
5673 		q = ctx;
5674 		tqrid = -1;
5675 		info = &ctx->ifc_filter_info;
5676 		gtask = &ctx->ifc_admin_task;
5677 		tqg = qgroup_if_config_tqg;
5678 		fn = _task_fn_admin;
5679 		intr_fast = iflib_fast_intr_ctx;
5680 		break;
5681 	default:
5682 		panic("unknown net intr type");
5683 	}
5684 
5685 	info->ifi_filter = filter;
5686 	info->ifi_filter_arg = filter_arg;
5687 	info->ifi_task = gtask;
5688 	info->ifi_ctx = q;
5689 
5690 	err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info,  name);
5691 	if (err != 0) {
5692 		device_printf(ctx->ifc_dev, "_iflib_irq_alloc failed %d\n", err);
5693 		return (err);
5694 	}
5695 	if (type == IFLIB_INTR_ADMIN)
5696 		return (0);
5697 
5698 	if (tqrid != -1) {
5699 		err = iflib_irq_set_affinity(ctx, rman_get_start(irq->ii_res), type, qid, gtask, tqg, q, name);
5700 		if (err)
5701 			return (err);
5702 	} else {
5703 		taskqgroup_attach(tqg, gtask, q, rman_get_start(irq->ii_res), name);
5704 	}
5705 
5706 	return (0);
5707 }
5708 
5709 void
5710 iflib_softirq_alloc_generic(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type, void *arg, int qid, const char *name)
5711 {
5712 	struct grouptask *gtask;
5713 	struct taskqgroup *tqg;
5714 	gtask_fn_t *fn;
5715 	void *q;
5716 	int irq_num = -1;
5717 	int err;
5718 
5719 	switch (type) {
5720 	case IFLIB_INTR_TX:
5721 		q = &ctx->ifc_txqs[qid];
5722 		gtask = &ctx->ifc_txqs[qid].ift_task;
5723 		tqg = qgroup_if_io_tqg;
5724 		fn = _task_fn_tx;
5725 		if (irq != NULL)
5726 			irq_num = rman_get_start(irq->ii_res);
5727 		break;
5728 	case IFLIB_INTR_RX:
5729 		q = &ctx->ifc_rxqs[qid];
5730 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5731 		tqg = qgroup_if_io_tqg;
5732 		fn = _task_fn_rx;
5733 		if (irq != NULL)
5734 			irq_num = rman_get_start(irq->ii_res);
5735 		break;
5736 	case IFLIB_INTR_IOV:
5737 		q = ctx;
5738 		gtask = &ctx->ifc_vflr_task;
5739 		tqg = qgroup_if_config_tqg;
5740 		fn = _task_fn_iov;
5741 		break;
5742 	default:
5743 		panic("unknown net intr type");
5744 	}
5745 	GROUPTASK_INIT(gtask, 0, fn, q);
5746 	if (irq_num != -1) {
5747 		err = iflib_irq_set_affinity(ctx, irq_num, type, qid, gtask, tqg, q, name);
5748 		if (err)
5749 			taskqgroup_attach(tqg, gtask, q, irq_num, name);
5750 	}
5751 	else {
5752 		taskqgroup_attach(tqg, gtask, q, irq_num, name);
5753 	}
5754 }
5755 
5756 void
5757 iflib_irq_free(if_ctx_t ctx, if_irq_t irq)
5758 {
5759 	if (irq->ii_tag)
5760 		bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag);
5761 
5762 	if (irq->ii_res)
5763 		bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ, irq->ii_rid, irq->ii_res);
5764 }
5765 
5766 static int
5767 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, const char *name)
5768 {
5769 	iflib_txq_t txq = ctx->ifc_txqs;
5770 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5771 	if_irq_t irq = &ctx->ifc_legacy_irq;
5772 	iflib_filter_info_t info;
5773 	struct grouptask *gtask;
5774 	struct taskqgroup *tqg;
5775 	gtask_fn_t *fn;
5776 	int tqrid;
5777 	void *q;
5778 	int err;
5779 
5780 	q = &ctx->ifc_rxqs[0];
5781 	info = &rxq[0].ifr_filter_info;
5782 	gtask = &rxq[0].ifr_task;
5783 	tqg = qgroup_if_io_tqg;
5784 	tqrid = irq->ii_rid = *rid;
5785 	fn = _task_fn_rx;
5786 
5787 	ctx->ifc_flags |= IFC_LEGACY;
5788 	info->ifi_filter = filter;
5789 	info->ifi_filter_arg = filter_arg;
5790 	info->ifi_task = gtask;
5791 	info->ifi_ctx = ctx;
5792 
5793 	/* We allocate a single interrupt resource */
5794 	if ((err = _iflib_irq_alloc(ctx, irq, tqrid, iflib_fast_intr_ctx, NULL, info, name)) != 0)
5795 		return (err);
5796 	GROUPTASK_INIT(gtask, 0, fn, q);
5797 	taskqgroup_attach(tqg, gtask, q, rman_get_start(irq->ii_res), name);
5798 
5799 	GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq);
5800 	taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, rman_get_start(irq->ii_res), "tx");
5801 	return (0);
5802 }
5803 
5804 void
5805 iflib_led_create(if_ctx_t ctx)
5806 {
5807 
5808 	ctx->ifc_led_dev = led_create(iflib_led_func, ctx,
5809 	    device_get_nameunit(ctx->ifc_dev));
5810 }
5811 
5812 void
5813 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid)
5814 {
5815 
5816 	GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
5817 }
5818 
5819 void
5820 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid)
5821 {
5822 
5823 	GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task);
5824 }
5825 
5826 void
5827 iflib_admin_intr_deferred(if_ctx_t ctx)
5828 {
5829 #ifdef INVARIANTS
5830 	struct grouptask *gtask;
5831 
5832 	gtask = &ctx->ifc_admin_task;
5833 	MPASS(gtask != NULL && gtask->gt_taskqueue != NULL);
5834 #endif
5835 
5836 	GROUPTASK_ENQUEUE(&ctx->ifc_admin_task);
5837 }
5838 
5839 void
5840 iflib_iov_intr_deferred(if_ctx_t ctx)
5841 {
5842 
5843 	GROUPTASK_ENQUEUE(&ctx->ifc_vflr_task);
5844 }
5845 
5846 void
5847 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, char *name)
5848 {
5849 
5850 	taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, -1, name);
5851 }
5852 
5853 void
5854 iflib_config_gtask_init(void *ctx, struct grouptask *gtask, gtask_fn_t *fn,
5855 	const char *name)
5856 {
5857 
5858 	GROUPTASK_INIT(gtask, 0, fn, ctx);
5859 	taskqgroup_attach(qgroup_if_config_tqg, gtask, gtask, -1, name);
5860 }
5861 
5862 void
5863 iflib_config_gtask_deinit(struct grouptask *gtask)
5864 {
5865 
5866 	taskqgroup_detach(qgroup_if_config_tqg, gtask);
5867 }
5868 
5869 void
5870 iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate)
5871 {
5872 	if_t ifp = ctx->ifc_ifp;
5873 	iflib_txq_t txq = ctx->ifc_txqs;
5874 
5875 	if_setbaudrate(ifp, baudrate);
5876 	if (baudrate >= IF_Gbps(10)) {
5877 		STATE_LOCK(ctx);
5878 		ctx->ifc_flags |= IFC_PREFETCH;
5879 		STATE_UNLOCK(ctx);
5880 	}
5881 	/* If link down, disable watchdog */
5882 	if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) {
5883 		for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++)
5884 			txq->ift_qstatus = IFLIB_QUEUE_IDLE;
5885 	}
5886 	ctx->ifc_link_state = link_state;
5887 	if_link_state_change(ifp, link_state);
5888 }
5889 
5890 static int
5891 iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq)
5892 {
5893 	int credits;
5894 #ifdef INVARIANTS
5895 	int credits_pre = txq->ift_cidx_processed;
5896 #endif
5897 
5898 	if (ctx->isc_txd_credits_update == NULL)
5899 		return (0);
5900 
5901 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
5902 	    BUS_DMASYNC_POSTREAD);
5903 	if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, true)) == 0)
5904 		return (0);
5905 
5906 	txq->ift_processed += credits;
5907 	txq->ift_cidx_processed += credits;
5908 
5909 	MPASS(credits_pre + credits == txq->ift_cidx_processed);
5910 	if (txq->ift_cidx_processed >= txq->ift_size)
5911 		txq->ift_cidx_processed -= txq->ift_size;
5912 	return (credits);
5913 }
5914 
5915 static int
5916 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget)
5917 {
5918 
5919 	return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx,
5920 	    budget));
5921 }
5922 
5923 void
5924 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name,
5925 	const char *description, if_int_delay_info_t info,
5926 	int offset, int value)
5927 {
5928 	info->iidi_ctx = ctx;
5929 	info->iidi_offset = offset;
5930 	info->iidi_value = value;
5931 	SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev),
5932 	    SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)),
5933 	    OID_AUTO, name, CTLTYPE_INT|CTLFLAG_RW,
5934 	    info, 0, iflib_sysctl_int_delay, "I", description);
5935 }
5936 
5937 struct sx *
5938 iflib_ctx_lock_get(if_ctx_t ctx)
5939 {
5940 
5941 	return (&ctx->ifc_ctx_sx);
5942 }
5943 
5944 static int
5945 iflib_msix_init(if_ctx_t ctx)
5946 {
5947 	device_t dev = ctx->ifc_dev;
5948 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5949 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5950 	int vectors, queues, rx_queues, tx_queues, queuemsgs, msgs;
5951 	int iflib_num_tx_queues, iflib_num_rx_queues;
5952 	int err, admincnt, bar;
5953 
5954 	iflib_num_tx_queues = ctx->ifc_sysctl_ntxqs;
5955 	iflib_num_rx_queues = ctx->ifc_sysctl_nrxqs;
5956 
5957 	device_printf(dev, "msix_init qsets capped at %d\n", imax(scctx->isc_ntxqsets, scctx->isc_nrxqsets));
5958 
5959 	bar = ctx->ifc_softc_ctx.isc_msix_bar;
5960 	admincnt = sctx->isc_admin_intrcnt;
5961 	/* Override by tuneable */
5962 	if (scctx->isc_disable_msix)
5963 		goto msi;
5964 
5965 	/*
5966 	 * bar == -1 => "trust me I know what I'm doing"
5967 	 * Some drivers are for hardware that is so shoddily
5968 	 * documented that no one knows which bars are which
5969 	 * so the developer has to map all bars. This hack
5970 	 * allows shoddy garbage to use msix in this framework.
5971 	 */
5972 	if (bar != -1) {
5973 		ctx->ifc_msix_mem = bus_alloc_resource_any(dev,
5974 	            SYS_RES_MEMORY, &bar, RF_ACTIVE);
5975 		if (ctx->ifc_msix_mem == NULL) {
5976 			/* May not be enabled */
5977 			device_printf(dev, "Unable to map MSIX table \n");
5978 			goto msi;
5979 		}
5980 	}
5981 	/* First try MSI/X */
5982 	if ((msgs = pci_msix_count(dev)) == 0) { /* system has msix disabled */
5983 		device_printf(dev, "System has MSIX disabled \n");
5984 		bus_release_resource(dev, SYS_RES_MEMORY,
5985 		    bar, ctx->ifc_msix_mem);
5986 		ctx->ifc_msix_mem = NULL;
5987 		goto msi;
5988 	}
5989 #if IFLIB_DEBUG
5990 	/* use only 1 qset in debug mode */
5991 	queuemsgs = min(msgs - admincnt, 1);
5992 #else
5993 	queuemsgs = msgs - admincnt;
5994 #endif
5995 #ifdef RSS
5996 	queues = imin(queuemsgs, rss_getnumbuckets());
5997 #else
5998 	queues = queuemsgs;
5999 #endif
6000 	queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues);
6001 	device_printf(dev, "pxm cpus: %d queue msgs: %d admincnt: %d\n",
6002 				  CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt);
6003 #ifdef  RSS
6004 	/* If we're doing RSS, clamp at the number of RSS buckets */
6005 	if (queues > rss_getnumbuckets())
6006 		queues = rss_getnumbuckets();
6007 #endif
6008 	if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt)
6009 		rx_queues = iflib_num_rx_queues;
6010 	else
6011 		rx_queues = queues;
6012 
6013 	if (rx_queues > scctx->isc_nrxqsets)
6014 		rx_queues = scctx->isc_nrxqsets;
6015 
6016 	/*
6017 	 * We want this to be all logical CPUs by default
6018 	 */
6019 	if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues)
6020 		tx_queues = iflib_num_tx_queues;
6021 	else
6022 		tx_queues = mp_ncpus;
6023 
6024 	if (tx_queues > scctx->isc_ntxqsets)
6025 		tx_queues = scctx->isc_ntxqsets;
6026 
6027 	if (ctx->ifc_sysctl_qs_eq_override == 0) {
6028 #ifdef INVARIANTS
6029 		if (tx_queues != rx_queues)
6030 			device_printf(dev,
6031 			    "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n",
6032 			    min(rx_queues, tx_queues), min(rx_queues, tx_queues));
6033 #endif
6034 		tx_queues = min(rx_queues, tx_queues);
6035 		rx_queues = min(rx_queues, tx_queues);
6036 	}
6037 
6038 	device_printf(dev, "using %d rx queues %d tx queues \n", rx_queues, tx_queues);
6039 
6040 	vectors = rx_queues + admincnt;
6041 	if ((err = pci_alloc_msix(dev, &vectors)) == 0) {
6042 		device_printf(dev, "Using MSIX interrupts with %d vectors\n", vectors);
6043 		scctx->isc_vectors = vectors;
6044 		scctx->isc_nrxqsets = rx_queues;
6045 		scctx->isc_ntxqsets = tx_queues;
6046 		scctx->isc_intr = IFLIB_INTR_MSIX;
6047 
6048 		return (vectors);
6049 	} else {
6050 		device_printf(dev,
6051 		    "failed to allocate %d msix vectors, err: %d - using MSI\n", vectors, err);
6052 		bus_release_resource(dev, SYS_RES_MEMORY, bar,
6053 		    ctx->ifc_msix_mem);
6054 		ctx->ifc_msix_mem = NULL;
6055 	}
6056 msi:
6057 	vectors = pci_msi_count(dev);
6058 	scctx->isc_nrxqsets = 1;
6059 	scctx->isc_ntxqsets = 1;
6060 	scctx->isc_vectors = vectors;
6061 	if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) {
6062 		device_printf(dev,"Using an MSI interrupt\n");
6063 		scctx->isc_intr = IFLIB_INTR_MSI;
6064 	} else {
6065 		scctx->isc_vectors = 1;
6066 		device_printf(dev,"Using a Legacy interrupt\n");
6067 		scctx->isc_intr = IFLIB_INTR_LEGACY;
6068 	}
6069 
6070 	return (vectors);
6071 }
6072 
6073 static const char *ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" };
6074 
6075 static int
6076 mp_ring_state_handler(SYSCTL_HANDLER_ARGS)
6077 {
6078 	int rc;
6079 	uint16_t *state = ((uint16_t *)oidp->oid_arg1);
6080 	struct sbuf *sb;
6081 	const char *ring_state = "UNKNOWN";
6082 
6083 	/* XXX needed ? */
6084 	rc = sysctl_wire_old_buffer(req, 0);
6085 	MPASS(rc == 0);
6086 	if (rc != 0)
6087 		return (rc);
6088 	sb = sbuf_new_for_sysctl(NULL, NULL, 80, req);
6089 	MPASS(sb != NULL);
6090 	if (sb == NULL)
6091 		return (ENOMEM);
6092 	if (state[3] <= 3)
6093 		ring_state = ring_states[state[3]];
6094 
6095 	sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s",
6096 		    state[0], state[1], state[2], ring_state);
6097 	rc = sbuf_finish(sb);
6098 	sbuf_delete(sb);
6099         return(rc);
6100 }
6101 
6102 enum iflib_ndesc_handler {
6103 	IFLIB_NTXD_HANDLER,
6104 	IFLIB_NRXD_HANDLER,
6105 };
6106 
6107 static int
6108 mp_ndesc_handler(SYSCTL_HANDLER_ARGS)
6109 {
6110 	if_ctx_t ctx = (void *)arg1;
6111 	enum iflib_ndesc_handler type = arg2;
6112 	char buf[256] = {0};
6113 	qidx_t *ndesc;
6114 	char *p, *next;
6115 	int nqs, rc, i;
6116 
6117 	MPASS(type == IFLIB_NTXD_HANDLER || type == IFLIB_NRXD_HANDLER);
6118 
6119 	nqs = 8;
6120 	switch(type) {
6121 	case IFLIB_NTXD_HANDLER:
6122 		ndesc = ctx->ifc_sysctl_ntxds;
6123 		if (ctx->ifc_sctx)
6124 			nqs = ctx->ifc_sctx->isc_ntxqs;
6125 		break;
6126 	case IFLIB_NRXD_HANDLER:
6127 		ndesc = ctx->ifc_sysctl_nrxds;
6128 		if (ctx->ifc_sctx)
6129 			nqs = ctx->ifc_sctx->isc_nrxqs;
6130 		break;
6131 	default:
6132 			panic("unhandled type");
6133 	}
6134 	if (nqs == 0)
6135 		nqs = 8;
6136 
6137 	for (i=0; i<8; i++) {
6138 		if (i >= nqs)
6139 			break;
6140 		if (i)
6141 			strcat(buf, ",");
6142 		sprintf(strchr(buf, 0), "%d", ndesc[i]);
6143 	}
6144 
6145 	rc = sysctl_handle_string(oidp, buf, sizeof(buf), req);
6146 	if (rc || req->newptr == NULL)
6147 		return rc;
6148 
6149 	for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p;
6150 	    i++, p = strsep(&next, " ,")) {
6151 		ndesc[i] = strtoul(p, NULL, 10);
6152 	}
6153 
6154 	return(rc);
6155 }
6156 
6157 #define NAME_BUFLEN 32
6158 static void
6159 iflib_add_device_sysctl_pre(if_ctx_t ctx)
6160 {
6161         device_t dev = iflib_get_dev(ctx);
6162 	struct sysctl_oid_list *child, *oid_list;
6163 	struct sysctl_ctx_list *ctx_list;
6164 	struct sysctl_oid *node;
6165 
6166 	ctx_list = device_get_sysctl_ctx(dev);
6167 	child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev));
6168 	ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, "iflib",
6169 						      CTLFLAG_RD, NULL, "IFLIB fields");
6170 	oid_list = SYSCTL_CHILDREN(node);
6171 
6172 	SYSCTL_ADD_STRING(ctx_list, oid_list, OID_AUTO, "driver_version",
6173 		       CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version, 0,
6174 		       "driver version");
6175 
6176 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs",
6177 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0,
6178 			"# of txqs to use, 0 => use default #");
6179 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs",
6180 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0,
6181 			"# of rxqs to use, 0 => use default #");
6182 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_qs_enable",
6183 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0,
6184                        "permit #txq != #rxq");
6185 	SYSCTL_ADD_INT(ctx_list, oid_list, OID_AUTO, "disable_msix",
6186                       CTLFLAG_RWTUN, &ctx->ifc_softc_ctx.isc_disable_msix, 0,
6187                       "disable MSIX (default 0)");
6188 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "rx_budget",
6189 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_rx_budget, 0,
6190                        "set the rx budget");
6191 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "tx_abdicate",
6192 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_tx_abdicate, 0,
6193 		       "cause tx to abdicate instead of running to completion");
6194 
6195 	/* XXX change for per-queue sizes */
6196 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_ntxds",
6197 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NTXD_HANDLER,
6198                        mp_ndesc_handler, "A",
6199                        "list of # of tx descriptors to use, 0 = use default #");
6200 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_nrxds",
6201 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NRXD_HANDLER,
6202                        mp_ndesc_handler, "A",
6203                        "list of # of rx descriptors to use, 0 = use default #");
6204 }
6205 
6206 static void
6207 iflib_add_device_sysctl_post(if_ctx_t ctx)
6208 {
6209 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6210 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6211         device_t dev = iflib_get_dev(ctx);
6212 	struct sysctl_oid_list *child;
6213 	struct sysctl_ctx_list *ctx_list;
6214 	iflib_fl_t fl;
6215 	iflib_txq_t txq;
6216 	iflib_rxq_t rxq;
6217 	int i, j;
6218 	char namebuf[NAME_BUFLEN];
6219 	char *qfmt;
6220 	struct sysctl_oid *queue_node, *fl_node, *node;
6221 	struct sysctl_oid_list *queue_list, *fl_list;
6222 	ctx_list = device_get_sysctl_ctx(dev);
6223 
6224 	node = ctx->ifc_sysctl_node;
6225 	child = SYSCTL_CHILDREN(node);
6226 
6227 	if (scctx->isc_ntxqsets > 100)
6228 		qfmt = "txq%03d";
6229 	else if (scctx->isc_ntxqsets > 10)
6230 		qfmt = "txq%02d";
6231 	else
6232 		qfmt = "txq%d";
6233 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
6234 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6235 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6236 					     CTLFLAG_RD, NULL, "Queue Name");
6237 		queue_list = SYSCTL_CHILDREN(queue_node);
6238 #if MEMORY_LOGGING
6239 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued",
6240 				CTLFLAG_RD,
6241 				&txq->ift_dequeued, "total mbufs freed");
6242 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued",
6243 				CTLFLAG_RD,
6244 				&txq->ift_enqueued, "total mbufs enqueued");
6245 #endif
6246 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag",
6247 				   CTLFLAG_RD,
6248 				   &txq->ift_mbuf_defrag, "# of times m_defrag was called");
6249 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "m_pullups",
6250 				   CTLFLAG_RD,
6251 				   &txq->ift_pullups, "# of times m_pullup was called");
6252 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag_failed",
6253 				   CTLFLAG_RD,
6254 				   &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed");
6255 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_desc_avail",
6256 				   CTLFLAG_RD,
6257 				   &txq->ift_no_desc_avail, "# of times no descriptors were available");
6258 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "tx_map_failed",
6259 				   CTLFLAG_RD,
6260 				   &txq->ift_map_failed, "# of times dma map failed");
6261 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txd_encap_efbig",
6262 				   CTLFLAG_RD,
6263 				   &txq->ift_txd_encap_efbig, "# of times txd_encap returned EFBIG");
6264 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_tx_dma_setup",
6265 				   CTLFLAG_RD,
6266 				   &txq->ift_no_tx_dma_setup, "# of times map failed for other than EFBIG");
6267 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx",
6268 				   CTLFLAG_RD,
6269 				   &txq->ift_pidx, 1, "Producer Index");
6270 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx",
6271 				   CTLFLAG_RD,
6272 				   &txq->ift_cidx, 1, "Consumer Index");
6273 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx_processed",
6274 				   CTLFLAG_RD,
6275 				   &txq->ift_cidx_processed, 1, "Consumer Index seen by credit update");
6276 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use",
6277 				   CTLFLAG_RD,
6278 				   &txq->ift_in_use, 1, "descriptors in use");
6279 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_processed",
6280 				   CTLFLAG_RD,
6281 				   &txq->ift_processed, "descriptors procesed for clean");
6282 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned",
6283 				   CTLFLAG_RD,
6284 				   &txq->ift_cleaned, "total cleaned");
6285 		SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state",
6286 				CTLTYPE_STRING | CTLFLAG_RD, __DEVOLATILE(uint64_t *, &txq->ift_br->state),
6287 				0, mp_ring_state_handler, "A", "soft ring state");
6288 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_enqueues",
6289 				       CTLFLAG_RD, &txq->ift_br->enqueues,
6290 				       "# of enqueues to the mp_ring for this queue");
6291 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_drops",
6292 				       CTLFLAG_RD, &txq->ift_br->drops,
6293 				       "# of drops in the mp_ring for this queue");
6294 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_starts",
6295 				       CTLFLAG_RD, &txq->ift_br->starts,
6296 				       "# of normal consumer starts in the mp_ring for this queue");
6297 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_stalls",
6298 				       CTLFLAG_RD, &txq->ift_br->stalls,
6299 					       "# of consumer stalls in the mp_ring for this queue");
6300 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_restarts",
6301 			       CTLFLAG_RD, &txq->ift_br->restarts,
6302 				       "# of consumer restarts in the mp_ring for this queue");
6303 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_abdications",
6304 				       CTLFLAG_RD, &txq->ift_br->abdications,
6305 				       "# of consumer abdications in the mp_ring for this queue");
6306 	}
6307 
6308 	if (scctx->isc_nrxqsets > 100)
6309 		qfmt = "rxq%03d";
6310 	else if (scctx->isc_nrxqsets > 10)
6311 		qfmt = "rxq%02d";
6312 	else
6313 		qfmt = "rxq%d";
6314 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
6315 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6316 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6317 					     CTLFLAG_RD, NULL, "Queue Name");
6318 		queue_list = SYSCTL_CHILDREN(queue_node);
6319 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
6320 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_pidx",
6321 				       CTLFLAG_RD,
6322 				       &rxq->ifr_cq_pidx, 1, "Producer Index");
6323 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_cidx",
6324 				       CTLFLAG_RD,
6325 				       &rxq->ifr_cq_cidx, 1, "Consumer Index");
6326 		}
6327 
6328 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
6329 			snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j);
6330 			fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list, OID_AUTO, namebuf,
6331 						     CTLFLAG_RD, NULL, "freelist Name");
6332 			fl_list = SYSCTL_CHILDREN(fl_node);
6333 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx",
6334 				       CTLFLAG_RD,
6335 				       &fl->ifl_pidx, 1, "Producer Index");
6336 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx",
6337 				       CTLFLAG_RD,
6338 				       &fl->ifl_cidx, 1, "Consumer Index");
6339 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits",
6340 				       CTLFLAG_RD,
6341 				       &fl->ifl_credits, 1, "credits available");
6342 #if MEMORY_LOGGING
6343 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_enqueued",
6344 					CTLFLAG_RD,
6345 					&fl->ifl_m_enqueued, "mbufs allocated");
6346 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_dequeued",
6347 					CTLFLAG_RD,
6348 					&fl->ifl_m_dequeued, "mbufs freed");
6349 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_enqueued",
6350 					CTLFLAG_RD,
6351 					&fl->ifl_cl_enqueued, "clusters allocated");
6352 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_dequeued",
6353 					CTLFLAG_RD,
6354 					&fl->ifl_cl_dequeued, "clusters freed");
6355 #endif
6356 
6357 		}
6358 	}
6359 
6360 }
6361 
6362 void
6363 iflib_request_reset(if_ctx_t ctx)
6364 {
6365 
6366 	STATE_LOCK(ctx);
6367 	ctx->ifc_flags |= IFC_DO_RESET;
6368 	STATE_UNLOCK(ctx);
6369 }
6370 
6371 #ifndef __NO_STRICT_ALIGNMENT
6372 static struct mbuf *
6373 iflib_fixup_rx(struct mbuf *m)
6374 {
6375 	struct mbuf *n;
6376 
6377 	if (m->m_len <= (MCLBYTES - ETHER_HDR_LEN)) {
6378 		bcopy(m->m_data, m->m_data + ETHER_HDR_LEN, m->m_len);
6379 		m->m_data += ETHER_HDR_LEN;
6380 		n = m;
6381 	} else {
6382 		MGETHDR(n, M_NOWAIT, MT_DATA);
6383 		if (n == NULL) {
6384 			m_freem(m);
6385 			return (NULL);
6386 		}
6387 		bcopy(m->m_data, n->m_data, ETHER_HDR_LEN);
6388 		m->m_data += ETHER_HDR_LEN;
6389 		m->m_len -= ETHER_HDR_LEN;
6390 		n->m_len = ETHER_HDR_LEN;
6391 		M_MOVE_PKTHDR(n, m);
6392 		n->m_next = m;
6393 	}
6394 	return (n);
6395 }
6396 #endif
6397 
6398 #ifdef NETDUMP
6399 static void
6400 iflib_netdump_init(struct ifnet *ifp, int *nrxr, int *ncl, int *clsize)
6401 {
6402 	if_ctx_t ctx;
6403 
6404 	ctx = if_getsoftc(ifp);
6405 	CTX_LOCK(ctx);
6406 	*nrxr = NRXQSETS(ctx);
6407 	*ncl = ctx->ifc_rxqs[0].ifr_fl->ifl_size;
6408 	*clsize = ctx->ifc_rxqs[0].ifr_fl->ifl_buf_size;
6409 	CTX_UNLOCK(ctx);
6410 }
6411 
6412 static void
6413 iflib_netdump_event(struct ifnet *ifp, enum netdump_ev event)
6414 {
6415 	if_ctx_t ctx;
6416 	if_softc_ctx_t scctx;
6417 	iflib_fl_t fl;
6418 	iflib_rxq_t rxq;
6419 	int i, j;
6420 
6421 	ctx = if_getsoftc(ifp);
6422 	scctx = &ctx->ifc_softc_ctx;
6423 
6424 	switch (event) {
6425 	case NETDUMP_START:
6426 		for (i = 0; i < scctx->isc_nrxqsets; i++) {
6427 			rxq = &ctx->ifc_rxqs[i];
6428 			for (j = 0; j < rxq->ifr_nfl; j++) {
6429 				fl = rxq->ifr_fl;
6430 				fl->ifl_zone = m_getzone(fl->ifl_buf_size);
6431 			}
6432 		}
6433 		iflib_no_tx_batch = 1;
6434 		break;
6435 	default:
6436 		break;
6437 	}
6438 }
6439 
6440 static int
6441 iflib_netdump_transmit(struct ifnet *ifp, struct mbuf *m)
6442 {
6443 	if_ctx_t ctx;
6444 	iflib_txq_t txq;
6445 	int error;
6446 
6447 	ctx = if_getsoftc(ifp);
6448 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
6449 	    IFF_DRV_RUNNING)
6450 		return (EBUSY);
6451 
6452 	txq = &ctx->ifc_txqs[0];
6453 	error = iflib_encap(txq, &m);
6454 	if (error == 0)
6455 		(void)iflib_txd_db_check(ctx, txq, true, txq->ift_in_use);
6456 	return (error);
6457 }
6458 
6459 static int
6460 iflib_netdump_poll(struct ifnet *ifp, int count)
6461 {
6462 	if_ctx_t ctx;
6463 	if_softc_ctx_t scctx;
6464 	iflib_txq_t txq;
6465 	int i;
6466 
6467 	ctx = if_getsoftc(ifp);
6468 	scctx = &ctx->ifc_softc_ctx;
6469 
6470 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
6471 	    IFF_DRV_RUNNING)
6472 		return (EBUSY);
6473 
6474 	txq = &ctx->ifc_txqs[0];
6475 	(void)iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
6476 
6477 	for (i = 0; i < scctx->isc_nrxqsets; i++)
6478 		(void)iflib_rxeof(&ctx->ifc_rxqs[i], 16 /* XXX */);
6479 	return (0);
6480 }
6481 #endif /* NETDUMP */
6482