xref: /freebsd/sys/net/iflib.c (revision 8b8d90931d8089fc8329f9c67883b7ccd9266513)
1 /*-
2  * Copyright (c) 2014-2018, Matthew Macy <mmacy@mattmacy.io>
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions are met:
7  *
8  *  1. Redistributions of source code must retain the above copyright notice,
9  *     this list of conditions and the following disclaimer.
10  *
11  *  2. Neither the name of Matthew Macy nor the names of its
12  *     contributors may be used to endorse or promote products derived from
13  *     this software without specific prior written permission.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
16  * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18  * ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
19  * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
20  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
21  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
22  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
23  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
24  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
25  * POSSIBILITY OF SUCH DAMAGE.
26  */
27 
28 #include <sys/cdefs.h>
29 __FBSDID("$FreeBSD$");
30 
31 #include "opt_inet.h"
32 #include "opt_inet6.h"
33 #include "opt_acpi.h"
34 #include "opt_sched.h"
35 
36 #include <sys/param.h>
37 #include <sys/types.h>
38 #include <sys/bus.h>
39 #include <sys/eventhandler.h>
40 #include <sys/jail.h>
41 #include <sys/kernel.h>
42 #include <sys/lock.h>
43 #include <sys/md5.h>
44 #include <sys/mutex.h>
45 #include <sys/module.h>
46 #include <sys/kobj.h>
47 #include <sys/rman.h>
48 #include <sys/proc.h>
49 #include <sys/sbuf.h>
50 #include <sys/smp.h>
51 #include <sys/socket.h>
52 #include <sys/sockio.h>
53 #include <sys/sysctl.h>
54 #include <sys/syslog.h>
55 #include <sys/taskqueue.h>
56 #include <sys/limits.h>
57 
58 #include <net/if.h>
59 #include <net/if_var.h>
60 #include <net/if_types.h>
61 #include <net/if_media.h>
62 #include <net/bpf.h>
63 #include <net/ethernet.h>
64 #include <net/mp_ring.h>
65 #include <net/vnet.h>
66 
67 #include <netinet/in.h>
68 #include <netinet/in_pcb.h>
69 #include <netinet/tcp_lro.h>
70 #include <netinet/in_systm.h>
71 #include <netinet/if_ether.h>
72 #include <netinet/ip.h>
73 #include <netinet/ip6.h>
74 #include <netinet/tcp.h>
75 #include <netinet/ip_var.h>
76 #include <netinet/netdump/netdump.h>
77 #include <netinet6/ip6_var.h>
78 
79 #include <machine/bus.h>
80 #include <machine/in_cksum.h>
81 
82 #include <vm/vm.h>
83 #include <vm/pmap.h>
84 
85 #include <dev/led/led.h>
86 #include <dev/pci/pcireg.h>
87 #include <dev/pci/pcivar.h>
88 #include <dev/pci/pci_private.h>
89 
90 #include <net/iflib.h>
91 #include <net/iflib_private.h>
92 
93 #include "ifdi_if.h"
94 
95 #if defined(__i386__) || defined(__amd64__)
96 #include <sys/memdesc.h>
97 #include <machine/bus.h>
98 #include <machine/md_var.h>
99 #include <machine/specialreg.h>
100 #include <x86/include/busdma_impl.h>
101 #include <x86/iommu/busdma_dmar.h>
102 #endif
103 
104 #include <sys/bitstring.h>
105 /*
106  * enable accounting of every mbuf as it comes in to and goes out of
107  * iflib's software descriptor references
108  */
109 #define MEMORY_LOGGING 0
110 /*
111  * Enable mbuf vectors for compressing long mbuf chains
112  */
113 
114 /*
115  * NB:
116  * - Prefetching in tx cleaning should perhaps be a tunable. The distance ahead
117  *   we prefetch needs to be determined by the time spent in m_free vis a vis
118  *   the cost of a prefetch. This will of course vary based on the workload:
119  *      - NFLX's m_free path is dominated by vm-based M_EXT manipulation which
120  *        is quite expensive, thus suggesting very little prefetch.
121  *      - small packet forwarding which is just returning a single mbuf to
122  *        UMA will typically be very fast vis a vis the cost of a memory
123  *        access.
124  */
125 
126 
127 /*
128  * File organization:
129  *  - private structures
130  *  - iflib private utility functions
131  *  - ifnet functions
132  *  - vlan registry and other exported functions
133  *  - iflib public core functions
134  *
135  *
136  */
137 MALLOC_DEFINE(M_IFLIB, "iflib", "ifnet library");
138 
139 struct iflib_txq;
140 typedef struct iflib_txq *iflib_txq_t;
141 struct iflib_rxq;
142 typedef struct iflib_rxq *iflib_rxq_t;
143 struct iflib_fl;
144 typedef struct iflib_fl *iflib_fl_t;
145 
146 struct iflib_ctx;
147 
148 static void iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid);
149 static void iflib_timer(void *arg);
150 
151 typedef struct iflib_filter_info {
152 	driver_filter_t *ifi_filter;
153 	void *ifi_filter_arg;
154 	struct grouptask *ifi_task;
155 	void *ifi_ctx;
156 } *iflib_filter_info_t;
157 
158 struct iflib_ctx {
159 	KOBJ_FIELDS;
160    /*
161    * Pointer to hardware driver's softc
162    */
163 	void *ifc_softc;
164 	device_t ifc_dev;
165 	if_t ifc_ifp;
166 
167 	cpuset_t ifc_cpus;
168 	if_shared_ctx_t ifc_sctx;
169 	struct if_softc_ctx ifc_softc_ctx;
170 
171 	struct sx ifc_ctx_sx;
172 	struct mtx ifc_state_mtx;
173 
174 	uint16_t ifc_nhwtxqs;
175 
176 	iflib_txq_t ifc_txqs;
177 	iflib_rxq_t ifc_rxqs;
178 	uint32_t ifc_if_flags;
179 	uint32_t ifc_flags;
180 	uint32_t ifc_max_fl_buf_size;
181 	int ifc_in_detach;
182 
183 	int ifc_link_state;
184 	int ifc_link_irq;
185 	int ifc_watchdog_events;
186 	struct cdev *ifc_led_dev;
187 	struct resource *ifc_msix_mem;
188 
189 	struct if_irq ifc_legacy_irq;
190 	struct grouptask ifc_admin_task;
191 	struct grouptask ifc_vflr_task;
192 	struct iflib_filter_info ifc_filter_info;
193 	struct ifmedia	ifc_media;
194 
195 	struct sysctl_oid *ifc_sysctl_node;
196 	uint16_t ifc_sysctl_ntxqs;
197 	uint16_t ifc_sysctl_nrxqs;
198 	uint16_t ifc_sysctl_qs_eq_override;
199 	uint16_t ifc_sysctl_rx_budget;
200 	uint16_t ifc_sysctl_tx_abdicate;
201 
202 	qidx_t ifc_sysctl_ntxds[8];
203 	qidx_t ifc_sysctl_nrxds[8];
204 	struct if_txrx ifc_txrx;
205 #define isc_txd_encap  ifc_txrx.ift_txd_encap
206 #define isc_txd_flush  ifc_txrx.ift_txd_flush
207 #define isc_txd_credits_update  ifc_txrx.ift_txd_credits_update
208 #define isc_rxd_available ifc_txrx.ift_rxd_available
209 #define isc_rxd_pkt_get ifc_txrx.ift_rxd_pkt_get
210 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
211 #define isc_rxd_flush ifc_txrx.ift_rxd_flush
212 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
213 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
214 #define isc_legacy_intr ifc_txrx.ift_legacy_intr
215 	eventhandler_tag ifc_vlan_attach_event;
216 	eventhandler_tag ifc_vlan_detach_event;
217 	uint8_t ifc_mac[ETHER_ADDR_LEN];
218 	char ifc_mtx_name[16];
219 };
220 
221 
222 void *
223 iflib_get_softc(if_ctx_t ctx)
224 {
225 
226 	return (ctx->ifc_softc);
227 }
228 
229 device_t
230 iflib_get_dev(if_ctx_t ctx)
231 {
232 
233 	return (ctx->ifc_dev);
234 }
235 
236 if_t
237 iflib_get_ifp(if_ctx_t ctx)
238 {
239 
240 	return (ctx->ifc_ifp);
241 }
242 
243 struct ifmedia *
244 iflib_get_media(if_ctx_t ctx)
245 {
246 
247 	return (&ctx->ifc_media);
248 }
249 
250 uint32_t
251 iflib_get_flags(if_ctx_t ctx)
252 {
253 	return (ctx->ifc_flags);
254 }
255 
256 void
257 iflib_set_detach(if_ctx_t ctx)
258 {
259 	ctx->ifc_in_detach = 1;
260 }
261 
262 void
263 iflib_set_mac(if_ctx_t ctx, uint8_t mac[ETHER_ADDR_LEN])
264 {
265 
266 	bcopy(mac, ctx->ifc_mac, ETHER_ADDR_LEN);
267 }
268 
269 if_softc_ctx_t
270 iflib_get_softc_ctx(if_ctx_t ctx)
271 {
272 
273 	return (&ctx->ifc_softc_ctx);
274 }
275 
276 if_shared_ctx_t
277 iflib_get_sctx(if_ctx_t ctx)
278 {
279 
280 	return (ctx->ifc_sctx);
281 }
282 
283 #define IP_ALIGNED(m) ((((uintptr_t)(m)->m_data) & 0x3) == 0x2)
284 #define CACHE_PTR_INCREMENT (CACHE_LINE_SIZE/sizeof(void*))
285 #define CACHE_PTR_NEXT(ptr) ((void *)(((uintptr_t)(ptr)+CACHE_LINE_SIZE-1) & (CACHE_LINE_SIZE-1)))
286 
287 #define LINK_ACTIVE(ctx) ((ctx)->ifc_link_state == LINK_STATE_UP)
288 #define CTX_IS_VF(ctx) ((ctx)->ifc_sctx->isc_flags & IFLIB_IS_VF)
289 
290 #define RX_SW_DESC_MAP_CREATED	(1 << 0)
291 #define TX_SW_DESC_MAP_CREATED	(1 << 1)
292 #define RX_SW_DESC_INUSE        (1 << 3)
293 #define TX_SW_DESC_MAPPED       (1 << 4)
294 
295 #define	M_TOOBIG		M_PROTO1
296 
297 typedef struct iflib_sw_rx_desc_array {
298 	bus_dmamap_t	*ifsd_map;         /* bus_dma maps for packet */
299 	struct mbuf	**ifsd_m;           /* pkthdr mbufs */
300 	caddr_t		*ifsd_cl;          /* direct cluster pointer for rx */
301 	uint8_t		*ifsd_flags;
302 } iflib_rxsd_array_t;
303 
304 typedef struct iflib_sw_tx_desc_array {
305 	bus_dmamap_t    *ifsd_map;         /* bus_dma maps for packet */
306 	struct mbuf    **ifsd_m;           /* pkthdr mbufs */
307 	uint8_t		*ifsd_flags;
308 } if_txsd_vec_t;
309 
310 
311 /* magic number that should be high enough for any hardware */
312 #define IFLIB_MAX_TX_SEGS		128
313 /* bnxt supports 64 with hardware LRO enabled */
314 #define IFLIB_MAX_RX_SEGS		64
315 #define IFLIB_RX_COPY_THRESH		128
316 #define IFLIB_MAX_RX_REFRESH		32
317 /* The minimum descriptors per second before we start coalescing */
318 #define IFLIB_MIN_DESC_SEC		16384
319 #define IFLIB_DEFAULT_TX_UPDATE_FREQ	16
320 #define IFLIB_QUEUE_IDLE		0
321 #define IFLIB_QUEUE_HUNG		1
322 #define IFLIB_QUEUE_WORKING		2
323 /* maximum number of txqs that can share an rx interrupt */
324 #define IFLIB_MAX_TX_SHARED_INTR	4
325 
326 /* this should really scale with ring size - this is a fairly arbitrary value */
327 #define TX_BATCH_SIZE			32
328 
329 #define IFLIB_RESTART_BUDGET		8
330 
331 
332 #define CSUM_OFFLOAD		(CSUM_IP_TSO|CSUM_IP6_TSO|CSUM_IP| \
333 				 CSUM_IP_UDP|CSUM_IP_TCP|CSUM_IP_SCTP| \
334 				 CSUM_IP6_UDP|CSUM_IP6_TCP|CSUM_IP6_SCTP)
335 struct iflib_txq {
336 	qidx_t		ift_in_use;
337 	qidx_t		ift_cidx;
338 	qidx_t		ift_cidx_processed;
339 	qidx_t		ift_pidx;
340 	uint8_t		ift_gen;
341 	uint8_t		ift_br_offset;
342 	uint16_t	ift_npending;
343 	uint16_t	ift_db_pending;
344 	uint16_t	ift_rs_pending;
345 	/* implicit pad */
346 	uint8_t		ift_txd_size[8];
347 	uint64_t	ift_processed;
348 	uint64_t	ift_cleaned;
349 	uint64_t	ift_cleaned_prev;
350 #if MEMORY_LOGGING
351 	uint64_t	ift_enqueued;
352 	uint64_t	ift_dequeued;
353 #endif
354 	uint64_t	ift_no_tx_dma_setup;
355 	uint64_t	ift_no_desc_avail;
356 	uint64_t	ift_mbuf_defrag_failed;
357 	uint64_t	ift_mbuf_defrag;
358 	uint64_t	ift_map_failed;
359 	uint64_t	ift_txd_encap_efbig;
360 	uint64_t	ift_pullups;
361 	uint64_t	ift_last_timer_tick;
362 
363 	struct mtx	ift_mtx;
364 	struct mtx	ift_db_mtx;
365 
366 	/* constant values */
367 	if_ctx_t	ift_ctx;
368 	struct ifmp_ring        *ift_br;
369 	struct grouptask	ift_task;
370 	qidx_t		ift_size;
371 	uint16_t	ift_id;
372 	struct callout	ift_timer;
373 
374 	if_txsd_vec_t	ift_sds;
375 	uint8_t		ift_qstatus;
376 	uint8_t		ift_closed;
377 	uint8_t		ift_update_freq;
378 	struct iflib_filter_info ift_filter_info;
379 	bus_dma_tag_t		ift_desc_tag;
380 	bus_dma_tag_t		ift_tso_desc_tag;
381 	iflib_dma_info_t	ift_ifdi;
382 #define MTX_NAME_LEN 16
383 	char                    ift_mtx_name[MTX_NAME_LEN];
384 	char                    ift_db_mtx_name[MTX_NAME_LEN];
385 	bus_dma_segment_t	ift_segs[IFLIB_MAX_TX_SEGS]  __aligned(CACHE_LINE_SIZE);
386 #ifdef IFLIB_DIAGNOSTICS
387 	uint64_t ift_cpu_exec_count[256];
388 #endif
389 } __aligned(CACHE_LINE_SIZE);
390 
391 struct iflib_fl {
392 	qidx_t		ifl_cidx;
393 	qidx_t		ifl_pidx;
394 	qidx_t		ifl_credits;
395 	uint8_t		ifl_gen;
396 	uint8_t		ifl_rxd_size;
397 #if MEMORY_LOGGING
398 	uint64_t	ifl_m_enqueued;
399 	uint64_t	ifl_m_dequeued;
400 	uint64_t	ifl_cl_enqueued;
401 	uint64_t	ifl_cl_dequeued;
402 #endif
403 	/* implicit pad */
404 
405 	bitstr_t 	*ifl_rx_bitmap;
406 	qidx_t		ifl_fragidx;
407 	/* constant */
408 	qidx_t		ifl_size;
409 	uint16_t	ifl_buf_size;
410 	uint16_t	ifl_cltype;
411 	uma_zone_t	ifl_zone;
412 	iflib_rxsd_array_t	ifl_sds;
413 	iflib_rxq_t	ifl_rxq;
414 	uint8_t		ifl_id;
415 	bus_dma_tag_t           ifl_desc_tag;
416 	iflib_dma_info_t	ifl_ifdi;
417 	uint64_t	ifl_bus_addrs[IFLIB_MAX_RX_REFRESH] __aligned(CACHE_LINE_SIZE);
418 	caddr_t		ifl_vm_addrs[IFLIB_MAX_RX_REFRESH];
419 	qidx_t	ifl_rxd_idxs[IFLIB_MAX_RX_REFRESH];
420 }  __aligned(CACHE_LINE_SIZE);
421 
422 static inline qidx_t
423 get_inuse(int size, qidx_t cidx, qidx_t pidx, uint8_t gen)
424 {
425 	qidx_t used;
426 
427 	if (pidx > cidx)
428 		used = pidx - cidx;
429 	else if (pidx < cidx)
430 		used = size - cidx + pidx;
431 	else if (gen == 0 && pidx == cidx)
432 		used = 0;
433 	else if (gen == 1 && pidx == cidx)
434 		used = size;
435 	else
436 		panic("bad state");
437 
438 	return (used);
439 }
440 
441 #define TXQ_AVAIL(txq) (txq->ift_size - get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx, txq->ift_gen))
442 
443 #define IDXDIFF(head, tail, wrap) \
444 	((head) >= (tail) ? (head) - (tail) : (wrap) - (tail) + (head))
445 
446 struct iflib_rxq {
447 	/* If there is a separate completion queue -
448 	 * these are the cq cidx and pidx. Otherwise
449 	 * these are unused.
450 	 */
451 	qidx_t		ifr_size;
452 	qidx_t		ifr_cq_cidx;
453 	qidx_t		ifr_cq_pidx;
454 	uint8_t		ifr_cq_gen;
455 	uint8_t		ifr_fl_offset;
456 
457 	if_ctx_t	ifr_ctx;
458 	iflib_fl_t	ifr_fl;
459 	uint64_t	ifr_rx_irq;
460 	uint16_t	ifr_id;
461 	uint8_t		ifr_lro_enabled;
462 	uint8_t		ifr_nfl;
463 	uint8_t		ifr_ntxqirq;
464 	uint8_t		ifr_txqid[IFLIB_MAX_TX_SHARED_INTR];
465 	struct lro_ctrl			ifr_lc;
466 	struct grouptask        ifr_task;
467 	struct iflib_filter_info ifr_filter_info;
468 	iflib_dma_info_t		ifr_ifdi;
469 
470 	/* dynamically allocate if any drivers need a value substantially larger than this */
471 	struct if_rxd_frag	ifr_frags[IFLIB_MAX_RX_SEGS] __aligned(CACHE_LINE_SIZE);
472 #ifdef IFLIB_DIAGNOSTICS
473 	uint64_t ifr_cpu_exec_count[256];
474 #endif
475 }  __aligned(CACHE_LINE_SIZE);
476 
477 typedef struct if_rxsd {
478 	caddr_t *ifsd_cl;
479 	struct mbuf **ifsd_m;
480 	iflib_fl_t ifsd_fl;
481 	qidx_t ifsd_cidx;
482 } *if_rxsd_t;
483 
484 /* multiple of word size */
485 #ifdef __LP64__
486 #define PKT_INFO_SIZE	6
487 #define RXD_INFO_SIZE	5
488 #define PKT_TYPE uint64_t
489 #else
490 #define PKT_INFO_SIZE	11
491 #define RXD_INFO_SIZE	8
492 #define PKT_TYPE uint32_t
493 #endif
494 #define PKT_LOOP_BOUND  ((PKT_INFO_SIZE/3)*3)
495 #define RXD_LOOP_BOUND  ((RXD_INFO_SIZE/4)*4)
496 
497 typedef struct if_pkt_info_pad {
498 	PKT_TYPE pkt_val[PKT_INFO_SIZE];
499 } *if_pkt_info_pad_t;
500 typedef struct if_rxd_info_pad {
501 	PKT_TYPE rxd_val[RXD_INFO_SIZE];
502 } *if_rxd_info_pad_t;
503 
504 CTASSERT(sizeof(struct if_pkt_info_pad) == sizeof(struct if_pkt_info));
505 CTASSERT(sizeof(struct if_rxd_info_pad) == sizeof(struct if_rxd_info));
506 
507 
508 static inline void
509 pkt_info_zero(if_pkt_info_t pi)
510 {
511 	if_pkt_info_pad_t pi_pad;
512 
513 	pi_pad = (if_pkt_info_pad_t)pi;
514 	pi_pad->pkt_val[0] = 0; pi_pad->pkt_val[1] = 0; pi_pad->pkt_val[2] = 0;
515 	pi_pad->pkt_val[3] = 0; pi_pad->pkt_val[4] = 0; pi_pad->pkt_val[5] = 0;
516 #ifndef __LP64__
517 	pi_pad->pkt_val[6] = 0; pi_pad->pkt_val[7] = 0; pi_pad->pkt_val[8] = 0;
518 	pi_pad->pkt_val[9] = 0; pi_pad->pkt_val[10] = 0;
519 #endif
520 }
521 
522 static device_method_t iflib_pseudo_methods[] = {
523 	DEVMETHOD(device_attach, noop_attach),
524 	DEVMETHOD(device_detach, iflib_pseudo_detach),
525 	DEVMETHOD_END
526 };
527 
528 driver_t iflib_pseudodriver = {
529 	"iflib_pseudo", iflib_pseudo_methods, sizeof(struct iflib_ctx),
530 };
531 
532 static inline void
533 rxd_info_zero(if_rxd_info_t ri)
534 {
535 	if_rxd_info_pad_t ri_pad;
536 	int i;
537 
538 	ri_pad = (if_rxd_info_pad_t)ri;
539 	for (i = 0; i < RXD_LOOP_BOUND; i += 4) {
540 		ri_pad->rxd_val[i] = 0;
541 		ri_pad->rxd_val[i+1] = 0;
542 		ri_pad->rxd_val[i+2] = 0;
543 		ri_pad->rxd_val[i+3] = 0;
544 	}
545 #ifdef __LP64__
546 	ri_pad->rxd_val[RXD_INFO_SIZE-1] = 0;
547 #endif
548 }
549 
550 /*
551  * Only allow a single packet to take up most 1/nth of the tx ring
552  */
553 #define MAX_SINGLE_PACKET_FRACTION 12
554 #define IF_BAD_DMA (bus_addr_t)-1
555 
556 #define CTX_ACTIVE(ctx) ((if_getdrvflags((ctx)->ifc_ifp) & IFF_DRV_RUNNING))
557 
558 #define CTX_LOCK_INIT(_sc)  sx_init(&(_sc)->ifc_ctx_sx, "iflib ctx lock")
559 #define CTX_LOCK(ctx) sx_xlock(&(ctx)->ifc_ctx_sx)
560 #define CTX_UNLOCK(ctx) sx_xunlock(&(ctx)->ifc_ctx_sx)
561 #define CTX_LOCK_DESTROY(ctx) sx_destroy(&(ctx)->ifc_ctx_sx)
562 
563 
564 #define STATE_LOCK_INIT(_sc, _name)  mtx_init(&(_sc)->ifc_state_mtx, _name, "iflib state lock", MTX_DEF)
565 #define STATE_LOCK(ctx) mtx_lock(&(ctx)->ifc_state_mtx)
566 #define STATE_UNLOCK(ctx) mtx_unlock(&(ctx)->ifc_state_mtx)
567 #define STATE_LOCK_DESTROY(ctx) mtx_destroy(&(ctx)->ifc_state_mtx)
568 
569 
570 
571 #define CALLOUT_LOCK(txq)	mtx_lock(&txq->ift_mtx)
572 #define CALLOUT_UNLOCK(txq) 	mtx_unlock(&txq->ift_mtx)
573 
574 
575 /* Our boot-time initialization hook */
576 static int	iflib_module_event_handler(module_t, int, void *);
577 
578 static moduledata_t iflib_moduledata = {
579 	"iflib",
580 	iflib_module_event_handler,
581 	NULL
582 };
583 
584 DECLARE_MODULE(iflib, iflib_moduledata, SI_SUB_INIT_IF, SI_ORDER_ANY);
585 MODULE_VERSION(iflib, 1);
586 
587 MODULE_DEPEND(iflib, pci, 1, 1, 1);
588 MODULE_DEPEND(iflib, ether, 1, 1, 1);
589 
590 TASKQGROUP_DEFINE(if_io_tqg, mp_ncpus, 1);
591 TASKQGROUP_DEFINE(if_config_tqg, 1, 1);
592 
593 #ifndef IFLIB_DEBUG_COUNTERS
594 #ifdef INVARIANTS
595 #define IFLIB_DEBUG_COUNTERS 1
596 #else
597 #define IFLIB_DEBUG_COUNTERS 0
598 #endif /* !INVARIANTS */
599 #endif
600 
601 static SYSCTL_NODE(_net, OID_AUTO, iflib, CTLFLAG_RD, 0,
602                    "iflib driver parameters");
603 
604 /*
605  * XXX need to ensure that this can't accidentally cause the head to be moved backwards
606  */
607 static int iflib_min_tx_latency = 0;
608 SYSCTL_INT(_net_iflib, OID_AUTO, min_tx_latency, CTLFLAG_RW,
609 		   &iflib_min_tx_latency, 0, "minimize transmit latency at the possible expense of throughput");
610 static int iflib_no_tx_batch = 0;
611 SYSCTL_INT(_net_iflib, OID_AUTO, no_tx_batch, CTLFLAG_RW,
612 		   &iflib_no_tx_batch, 0, "minimize transmit latency at the possible expense of throughput");
613 
614 
615 #if IFLIB_DEBUG_COUNTERS
616 
617 static int iflib_tx_seen;
618 static int iflib_tx_sent;
619 static int iflib_tx_encap;
620 static int iflib_rx_allocs;
621 static int iflib_fl_refills;
622 static int iflib_fl_refills_large;
623 static int iflib_tx_frees;
624 
625 SYSCTL_INT(_net_iflib, OID_AUTO, tx_seen, CTLFLAG_RD,
626 		   &iflib_tx_seen, 0, "# tx mbufs seen");
627 SYSCTL_INT(_net_iflib, OID_AUTO, tx_sent, CTLFLAG_RD,
628 		   &iflib_tx_sent, 0, "# tx mbufs sent");
629 SYSCTL_INT(_net_iflib, OID_AUTO, tx_encap, CTLFLAG_RD,
630 		   &iflib_tx_encap, 0, "# tx mbufs encapped");
631 SYSCTL_INT(_net_iflib, OID_AUTO, tx_frees, CTLFLAG_RD,
632 		   &iflib_tx_frees, 0, "# tx frees");
633 SYSCTL_INT(_net_iflib, OID_AUTO, rx_allocs, CTLFLAG_RD,
634 		   &iflib_rx_allocs, 0, "# rx allocations");
635 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills, CTLFLAG_RD,
636 		   &iflib_fl_refills, 0, "# refills");
637 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills_large, CTLFLAG_RD,
638 		   &iflib_fl_refills_large, 0, "# large refills");
639 
640 
641 static int iflib_txq_drain_flushing;
642 static int iflib_txq_drain_oactive;
643 static int iflib_txq_drain_notready;
644 static int iflib_txq_drain_encapfail;
645 
646 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_flushing, CTLFLAG_RD,
647 		   &iflib_txq_drain_flushing, 0, "# drain flushes");
648 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_oactive, CTLFLAG_RD,
649 		   &iflib_txq_drain_oactive, 0, "# drain oactives");
650 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_notready, CTLFLAG_RD,
651 		   &iflib_txq_drain_notready, 0, "# drain notready");
652 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_encapfail, CTLFLAG_RD,
653 		   &iflib_txq_drain_encapfail, 0, "# drain encap fails");
654 
655 
656 static int iflib_encap_load_mbuf_fail;
657 static int iflib_encap_pad_mbuf_fail;
658 static int iflib_encap_txq_avail_fail;
659 static int iflib_encap_txd_encap_fail;
660 
661 SYSCTL_INT(_net_iflib, OID_AUTO, encap_load_mbuf_fail, CTLFLAG_RD,
662 		   &iflib_encap_load_mbuf_fail, 0, "# busdma load failures");
663 SYSCTL_INT(_net_iflib, OID_AUTO, encap_pad_mbuf_fail, CTLFLAG_RD,
664 		   &iflib_encap_pad_mbuf_fail, 0, "# runt frame pad failures");
665 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txq_avail_fail, CTLFLAG_RD,
666 		   &iflib_encap_txq_avail_fail, 0, "# txq avail failures");
667 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txd_encap_fail, CTLFLAG_RD,
668 		   &iflib_encap_txd_encap_fail, 0, "# driver encap failures");
669 
670 static int iflib_task_fn_rxs;
671 static int iflib_rx_intr_enables;
672 static int iflib_fast_intrs;
673 static int iflib_intr_link;
674 static int iflib_intr_msix;
675 static int iflib_rx_unavail;
676 static int iflib_rx_ctx_inactive;
677 static int iflib_rx_zero_len;
678 static int iflib_rx_if_input;
679 static int iflib_rx_mbuf_null;
680 static int iflib_rxd_flush;
681 
682 static int iflib_verbose_debug;
683 
684 SYSCTL_INT(_net_iflib, OID_AUTO, intr_link, CTLFLAG_RD,
685 		   &iflib_intr_link, 0, "# intr link calls");
686 SYSCTL_INT(_net_iflib, OID_AUTO, intr_msix, CTLFLAG_RD,
687 		   &iflib_intr_msix, 0, "# intr msix calls");
688 SYSCTL_INT(_net_iflib, OID_AUTO, task_fn_rx, CTLFLAG_RD,
689 		   &iflib_task_fn_rxs, 0, "# task_fn_rx calls");
690 SYSCTL_INT(_net_iflib, OID_AUTO, rx_intr_enables, CTLFLAG_RD,
691 		   &iflib_rx_intr_enables, 0, "# rx intr enables");
692 SYSCTL_INT(_net_iflib, OID_AUTO, fast_intrs, CTLFLAG_RD,
693 		   &iflib_fast_intrs, 0, "# fast_intr calls");
694 SYSCTL_INT(_net_iflib, OID_AUTO, rx_unavail, CTLFLAG_RD,
695 		   &iflib_rx_unavail, 0, "# times rxeof called with no available data");
696 SYSCTL_INT(_net_iflib, OID_AUTO, rx_ctx_inactive, CTLFLAG_RD,
697 		   &iflib_rx_ctx_inactive, 0, "# times rxeof called with inactive context");
698 SYSCTL_INT(_net_iflib, OID_AUTO, rx_zero_len, CTLFLAG_RD,
699 		   &iflib_rx_zero_len, 0, "# times rxeof saw zero len mbuf");
700 SYSCTL_INT(_net_iflib, OID_AUTO, rx_if_input, CTLFLAG_RD,
701 		   &iflib_rx_if_input, 0, "# times rxeof called if_input");
702 SYSCTL_INT(_net_iflib, OID_AUTO, rx_mbuf_null, CTLFLAG_RD,
703 		   &iflib_rx_mbuf_null, 0, "# times rxeof got null mbuf");
704 SYSCTL_INT(_net_iflib, OID_AUTO, rxd_flush, CTLFLAG_RD,
705 	         &iflib_rxd_flush, 0, "# times rxd_flush called");
706 SYSCTL_INT(_net_iflib, OID_AUTO, verbose_debug, CTLFLAG_RW,
707 		   &iflib_verbose_debug, 0, "enable verbose debugging");
708 
709 #define DBG_COUNTER_INC(name) atomic_add_int(&(iflib_ ## name), 1)
710 static void
711 iflib_debug_reset(void)
712 {
713 	iflib_tx_seen = iflib_tx_sent = iflib_tx_encap = iflib_rx_allocs =
714 		iflib_fl_refills = iflib_fl_refills_large = iflib_tx_frees =
715 		iflib_txq_drain_flushing = iflib_txq_drain_oactive =
716 		iflib_txq_drain_notready = iflib_txq_drain_encapfail =
717 		iflib_encap_load_mbuf_fail = iflib_encap_pad_mbuf_fail =
718 		iflib_encap_txq_avail_fail = iflib_encap_txd_encap_fail =
719 		iflib_task_fn_rxs = iflib_rx_intr_enables = iflib_fast_intrs =
720 		iflib_intr_link = iflib_intr_msix = iflib_rx_unavail =
721 		iflib_rx_ctx_inactive = iflib_rx_zero_len = iflib_rx_if_input =
722 		iflib_rx_mbuf_null = iflib_rxd_flush = 0;
723 }
724 
725 #else
726 #define DBG_COUNTER_INC(name)
727 static void iflib_debug_reset(void) {}
728 #endif
729 
730 #define IFLIB_DEBUG 0
731 
732 static void iflib_tx_structures_free(if_ctx_t ctx);
733 static void iflib_rx_structures_free(if_ctx_t ctx);
734 static int iflib_queues_alloc(if_ctx_t ctx);
735 static int iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq);
736 static int iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget);
737 static int iflib_qset_structures_setup(if_ctx_t ctx);
738 static int iflib_msix_init(if_ctx_t ctx);
739 static int iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filterarg, int *rid, const char *str);
740 static void iflib_txq_check_drain(iflib_txq_t txq, int budget);
741 static uint32_t iflib_txq_can_drain(struct ifmp_ring *);
742 static int iflib_register(if_ctx_t);
743 static void iflib_init_locked(if_ctx_t ctx);
744 static void iflib_add_device_sysctl_pre(if_ctx_t ctx);
745 static void iflib_add_device_sysctl_post(if_ctx_t ctx);
746 static void iflib_ifmp_purge(iflib_txq_t txq);
747 static void _iflib_pre_assert(if_softc_ctx_t scctx);
748 static void iflib_if_init_locked(if_ctx_t ctx);
749 #ifndef __NO_STRICT_ALIGNMENT
750 static struct mbuf * iflib_fixup_rx(struct mbuf *m);
751 #endif
752 
753 NETDUMP_DEFINE(iflib);
754 
755 #ifdef DEV_NETMAP
756 #include <sys/selinfo.h>
757 #include <net/netmap.h>
758 #include <dev/netmap/netmap_kern.h>
759 
760 MODULE_DEPEND(iflib, netmap, 1, 1, 1);
761 
762 static int netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, uint32_t nm_i, bool init);
763 
764 /*
765  * device-specific sysctl variables:
766  *
767  * iflib_crcstrip: 0: keep CRC in rx frames (default), 1: strip it.
768  *	During regular operations the CRC is stripped, but on some
769  *	hardware reception of frames not multiple of 64 is slower,
770  *	so using crcstrip=0 helps in benchmarks.
771  *
772  * iflib_rx_miss, iflib_rx_miss_bufs:
773  *	count packets that might be missed due to lost interrupts.
774  */
775 SYSCTL_DECL(_dev_netmap);
776 /*
777  * The xl driver by default strips CRCs and we do not override it.
778  */
779 
780 int iflib_crcstrip = 1;
781 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_crcstrip,
782     CTLFLAG_RW, &iflib_crcstrip, 1, "strip CRC on rx frames");
783 
784 int iflib_rx_miss, iflib_rx_miss_bufs;
785 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss,
786     CTLFLAG_RW, &iflib_rx_miss, 0, "potentially missed rx intr");
787 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss_bufs,
788     CTLFLAG_RW, &iflib_rx_miss_bufs, 0, "potentially missed rx intr bufs");
789 
790 /*
791  * Register/unregister. We are already under netmap lock.
792  * Only called on the first register or the last unregister.
793  */
794 static int
795 iflib_netmap_register(struct netmap_adapter *na, int onoff)
796 {
797 	struct ifnet *ifp = na->ifp;
798 	if_ctx_t ctx = ifp->if_softc;
799 	int status;
800 
801 	CTX_LOCK(ctx);
802 	IFDI_INTR_DISABLE(ctx);
803 
804 	/* Tell the stack that the interface is no longer active */
805 	ifp->if_drv_flags &= ~(IFF_DRV_RUNNING | IFF_DRV_OACTIVE);
806 
807 	if (!CTX_IS_VF(ctx))
808 		IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip);
809 
810 	/* enable or disable flags and callbacks in na and ifp */
811 	if (onoff) {
812 		nm_set_native_flags(na);
813 	} else {
814 		nm_clear_native_flags(na);
815 	}
816 	iflib_stop(ctx);
817 	iflib_init_locked(ctx);
818 	IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip); // XXX why twice ?
819 	status = ifp->if_drv_flags & IFF_DRV_RUNNING ? 0 : 1;
820 	if (status)
821 		nm_clear_native_flags(na);
822 	CTX_UNLOCK(ctx);
823 	return (status);
824 }
825 
826 static int
827 netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, uint32_t nm_i, bool init)
828 {
829 	struct netmap_adapter *na = kring->na;
830 	u_int const lim = kring->nkr_num_slots - 1;
831 	u_int head = kring->rhead;
832 	struct netmap_ring *ring = kring->ring;
833 	bus_dmamap_t *map;
834 	struct if_rxd_update iru;
835 	if_ctx_t ctx = rxq->ifr_ctx;
836 	iflib_fl_t fl = &rxq->ifr_fl[0];
837 	uint32_t refill_pidx, nic_i;
838 
839 	if (nm_i == head && __predict_true(!init))
840 		return 0;
841 	iru_init(&iru, rxq, 0 /* flid */);
842 	map = fl->ifl_sds.ifsd_map;
843 	refill_pidx = netmap_idx_k2n(kring, nm_i);
844 	/*
845 	 * IMPORTANT: we must leave one free slot in the ring,
846 	 * so move head back by one unit
847 	 */
848 	head = nm_prev(head, lim);
849 	nic_i = UINT_MAX;
850 	while (nm_i != head) {
851 		for (int tmp_pidx = 0; tmp_pidx < IFLIB_MAX_RX_REFRESH && nm_i != head; tmp_pidx++) {
852 			struct netmap_slot *slot = &ring->slot[nm_i];
853 			void *addr = PNMB(na, slot, &fl->ifl_bus_addrs[tmp_pidx]);
854 			uint32_t nic_i_dma = refill_pidx;
855 			nic_i = netmap_idx_k2n(kring, nm_i);
856 
857 			MPASS(tmp_pidx < IFLIB_MAX_RX_REFRESH);
858 
859 			if (addr == NETMAP_BUF_BASE(na)) /* bad buf */
860 			        return netmap_ring_reinit(kring);
861 
862 			fl->ifl_vm_addrs[tmp_pidx] = addr;
863 			if (__predict_false(init) && map) {
864 				netmap_load_map(na, fl->ifl_ifdi->idi_tag, map[nic_i], addr);
865 			} else if (map && (slot->flags & NS_BUF_CHANGED)) {
866 				/* buffer has changed, reload map */
867 				netmap_reload_map(na, fl->ifl_ifdi->idi_tag, map[nic_i], addr);
868 			}
869 			slot->flags &= ~NS_BUF_CHANGED;
870 
871 			nm_i = nm_next(nm_i, lim);
872 			fl->ifl_rxd_idxs[tmp_pidx] = nic_i = nm_next(nic_i, lim);
873 			if (nm_i != head && tmp_pidx < IFLIB_MAX_RX_REFRESH-1)
874 				continue;
875 
876 			iru.iru_pidx = refill_pidx;
877 			iru.iru_count = tmp_pidx+1;
878 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
879 
880 			refill_pidx = nic_i;
881 			if (map == NULL)
882 				continue;
883 
884 			for (int n = 0; n < iru.iru_count; n++) {
885 				bus_dmamap_sync(fl->ifl_ifdi->idi_tag, map[nic_i_dma],
886 						BUS_DMASYNC_PREREAD);
887 				/* XXX - change this to not use the netmap func*/
888 				nic_i_dma = nm_next(nic_i_dma, lim);
889 			}
890 		}
891 	}
892 	kring->nr_hwcur = head;
893 
894 	if (map)
895 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
896 				BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
897 	if (__predict_true(nic_i != UINT_MAX))
898 		ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id, nic_i);
899 	return (0);
900 }
901 
902 /*
903  * Reconcile kernel and user view of the transmit ring.
904  *
905  * All information is in the kring.
906  * Userspace wants to send packets up to the one before kring->rhead,
907  * kernel knows kring->nr_hwcur is the first unsent packet.
908  *
909  * Here we push packets out (as many as possible), and possibly
910  * reclaim buffers from previously completed transmission.
911  *
912  * The caller (netmap) guarantees that there is only one instance
913  * running at any time. Any interference with other driver
914  * methods should be handled by the individual drivers.
915  */
916 static int
917 iflib_netmap_txsync(struct netmap_kring *kring, int flags)
918 {
919 	struct netmap_adapter *na = kring->na;
920 	struct ifnet *ifp = na->ifp;
921 	struct netmap_ring *ring = kring->ring;
922 	u_int nm_i;	/* index into the netmap kring */
923 	u_int nic_i;	/* index into the NIC ring */
924 	u_int n;
925 	u_int const lim = kring->nkr_num_slots - 1;
926 	u_int const head = kring->rhead;
927 	struct if_pkt_info pi;
928 
929 	/*
930 	 * interrupts on every tx packet are expensive so request
931 	 * them every half ring, or where NS_REPORT is set
932 	 */
933 	u_int report_frequency = kring->nkr_num_slots >> 1;
934 	/* device-specific */
935 	if_ctx_t ctx = ifp->if_softc;
936 	iflib_txq_t txq = &ctx->ifc_txqs[kring->ring_id];
937 
938 	if (txq->ift_sds.ifsd_map)
939 		bus_dmamap_sync(txq->ift_desc_tag, txq->ift_ifdi->idi_map,
940 				BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
941 
942 
943 	/*
944 	 * First part: process new packets to send.
945 	 * nm_i is the current index in the netmap kring,
946 	 * nic_i is the corresponding index in the NIC ring.
947 	 *
948 	 * If we have packets to send (nm_i != head)
949 	 * iterate over the netmap ring, fetch length and update
950 	 * the corresponding slot in the NIC ring. Some drivers also
951 	 * need to update the buffer's physical address in the NIC slot
952 	 * even NS_BUF_CHANGED is not set (PNMB computes the addresses).
953 	 *
954 	 * The netmap_reload_map() calls is especially expensive,
955 	 * even when (as in this case) the tag is 0, so do only
956 	 * when the buffer has actually changed.
957 	 *
958 	 * If possible do not set the report/intr bit on all slots,
959 	 * but only a few times per ring or when NS_REPORT is set.
960 	 *
961 	 * Finally, on 10G and faster drivers, it might be useful
962 	 * to prefetch the next slot and txr entry.
963 	 */
964 
965 	nm_i = kring->nr_hwcur;
966 	if (nm_i != head) {	/* we have new packets to send */
967 		pkt_info_zero(&pi);
968 		pi.ipi_segs = txq->ift_segs;
969 		pi.ipi_qsidx = kring->ring_id;
970 		nic_i = netmap_idx_k2n(kring, nm_i);
971 
972 		__builtin_prefetch(&ring->slot[nm_i]);
973 		__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i]);
974 		if (txq->ift_sds.ifsd_map)
975 			__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i]);
976 
977 		for (n = 0; nm_i != head; n++) {
978 			struct netmap_slot *slot = &ring->slot[nm_i];
979 			u_int len = slot->len;
980 			uint64_t paddr;
981 			void *addr = PNMB(na, slot, &paddr);
982 			int flags = (slot->flags & NS_REPORT ||
983 				nic_i == 0 || nic_i == report_frequency) ?
984 				IPI_TX_INTR : 0;
985 
986 			/* device-specific */
987 			pi.ipi_len = len;
988 			pi.ipi_segs[0].ds_addr = paddr;
989 			pi.ipi_segs[0].ds_len = len;
990 			pi.ipi_nsegs = 1;
991 			pi.ipi_ndescs = 0;
992 			pi.ipi_pidx = nic_i;
993 			pi.ipi_flags = flags;
994 
995 			/* Fill the slot in the NIC ring. */
996 			ctx->isc_txd_encap(ctx->ifc_softc, &pi);
997 
998 			/* prefetch for next round */
999 			__builtin_prefetch(&ring->slot[nm_i + 1]);
1000 			__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i + 1]);
1001 			if (txq->ift_sds.ifsd_map) {
1002 				__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i + 1]);
1003 
1004 				NM_CHECK_ADDR_LEN(na, addr, len);
1005 
1006 				if (slot->flags & NS_BUF_CHANGED) {
1007 					/* buffer has changed, reload map */
1008 					netmap_reload_map(na, txq->ift_desc_tag, txq->ift_sds.ifsd_map[nic_i], addr);
1009 				}
1010 				/* make sure changes to the buffer are synced */
1011 				bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_sds.ifsd_map[nic_i],
1012 						BUS_DMASYNC_PREWRITE);
1013 			}
1014 			slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED);
1015 			nm_i = nm_next(nm_i, lim);
1016 			nic_i = nm_next(nic_i, lim);
1017 		}
1018 		kring->nr_hwcur = nm_i;
1019 
1020 		/* synchronize the NIC ring */
1021 		if (txq->ift_sds.ifsd_map)
1022 			bus_dmamap_sync(txq->ift_desc_tag, txq->ift_ifdi->idi_map,
1023 						BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1024 
1025 		/* (re)start the tx unit up to slot nic_i (excluded) */
1026 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, nic_i);
1027 	}
1028 
1029 	/*
1030 	 * Second part: reclaim buffers for completed transmissions.
1031 	 *
1032 	 * If there are unclaimed buffers, attempt to reclaim them.
1033 	 * If none are reclaimed, and TX IRQs are not in use, do an initial
1034 	 * minimal delay, then trigger the tx handler which will spin in the
1035 	 * group task queue.
1036 	 */
1037 	if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1038 		if (iflib_tx_credits_update(ctx, txq)) {
1039 			/* some tx completed, increment avail */
1040 			nic_i = txq->ift_cidx_processed;
1041 			kring->nr_hwtail = nm_prev(netmap_idx_n2k(kring, nic_i), lim);
1042 		}
1043 	}
1044 	if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ))
1045 		if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1046 			callout_reset_on(&txq->ift_timer, hz < 2000 ? 1 : hz / 1000,
1047 			    iflib_timer, txq, txq->ift_timer.c_cpu);
1048 	}
1049 	return (0);
1050 }
1051 
1052 /*
1053  * Reconcile kernel and user view of the receive ring.
1054  * Same as for the txsync, this routine must be efficient.
1055  * The caller guarantees a single invocations, but races against
1056  * the rest of the driver should be handled here.
1057  *
1058  * On call, kring->rhead is the first packet that userspace wants
1059  * to keep, and kring->rcur is the wakeup point.
1060  * The kernel has previously reported packets up to kring->rtail.
1061  *
1062  * If (flags & NAF_FORCE_READ) also check for incoming packets irrespective
1063  * of whether or not we received an interrupt.
1064  */
1065 static int
1066 iflib_netmap_rxsync(struct netmap_kring *kring, int flags)
1067 {
1068 	struct netmap_adapter *na = kring->na;
1069 	struct netmap_ring *ring = kring->ring;
1070 	uint32_t nm_i;	/* index into the netmap ring */
1071 	uint32_t nic_i;	/* index into the NIC ring */
1072 	u_int i, n;
1073 	u_int const lim = kring->nkr_num_slots - 1;
1074 	u_int const head = kring->rhead;
1075 	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
1076 	struct if_rxd_info ri;
1077 
1078 	struct ifnet *ifp = na->ifp;
1079 	if_ctx_t ctx = ifp->if_softc;
1080 	iflib_rxq_t rxq = &ctx->ifc_rxqs[kring->ring_id];
1081 	iflib_fl_t fl = rxq->ifr_fl;
1082 	if (head > lim)
1083 		return netmap_ring_reinit(kring);
1084 
1085 	/* XXX check sync modes */
1086 	for (i = 0, fl = rxq->ifr_fl; i < rxq->ifr_nfl; i++, fl++) {
1087 		if (fl->ifl_sds.ifsd_map == NULL)
1088 			continue;
1089 		bus_dmamap_sync(rxq->ifr_fl[i].ifl_desc_tag, fl->ifl_ifdi->idi_map,
1090 				BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1091 	}
1092 	/*
1093 	 * First part: import newly received packets.
1094 	 *
1095 	 * nm_i is the index of the next free slot in the netmap ring,
1096 	 * nic_i is the index of the next received packet in the NIC ring,
1097 	 * and they may differ in case if_init() has been called while
1098 	 * in netmap mode. For the receive ring we have
1099 	 *
1100 	 *	nic_i = rxr->next_check;
1101 	 *	nm_i = kring->nr_hwtail (previous)
1102 	 * and
1103 	 *	nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1104 	 *
1105 	 * rxr->next_check is set to 0 on a ring reinit
1106 	 */
1107 	if (netmap_no_pendintr || force_update) {
1108 		int crclen = iflib_crcstrip ? 0 : 4;
1109 		int error, avail;
1110 
1111 		for (i = 0; i < rxq->ifr_nfl; i++) {
1112 			fl = &rxq->ifr_fl[i];
1113 			nic_i = fl->ifl_cidx;
1114 			nm_i = netmap_idx_n2k(kring, nic_i);
1115 			avail = iflib_rxd_avail(ctx, rxq, nic_i, USHRT_MAX);
1116 			for (n = 0; avail > 0; n++, avail--) {
1117 				rxd_info_zero(&ri);
1118 				ri.iri_frags = rxq->ifr_frags;
1119 				ri.iri_qsidx = kring->ring_id;
1120 				ri.iri_ifp = ctx->ifc_ifp;
1121 				ri.iri_cidx = nic_i;
1122 
1123 				error = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
1124 				ring->slot[nm_i].len = error ? 0 : ri.iri_len - crclen;
1125 				ring->slot[nm_i].flags = 0;
1126 				if (fl->ifl_sds.ifsd_map)
1127 					bus_dmamap_sync(fl->ifl_ifdi->idi_tag,
1128 							fl->ifl_sds.ifsd_map[nic_i], BUS_DMASYNC_POSTREAD);
1129 				nm_i = nm_next(nm_i, lim);
1130 				nic_i = nm_next(nic_i, lim);
1131 			}
1132 			if (n) { /* update the state variables */
1133 				if (netmap_no_pendintr && !force_update) {
1134 					/* diagnostics */
1135 					iflib_rx_miss ++;
1136 					iflib_rx_miss_bufs += n;
1137 				}
1138 				fl->ifl_cidx = nic_i;
1139 				kring->nr_hwtail = nm_i;
1140 			}
1141 			kring->nr_kflags &= ~NKR_PENDINTR;
1142 		}
1143 	}
1144 	/*
1145 	 * Second part: skip past packets that userspace has released.
1146 	 * (kring->nr_hwcur to head excluded),
1147 	 * and make the buffers available for reception.
1148 	 * As usual nm_i is the index in the netmap ring,
1149 	 * nic_i is the index in the NIC ring, and
1150 	 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1151 	 */
1152 	/* XXX not sure how this will work with multiple free lists */
1153 	nm_i = kring->nr_hwcur;
1154 
1155 	return (netmap_fl_refill(rxq, kring, nm_i, false));
1156 }
1157 
1158 static void
1159 iflib_netmap_intr(struct netmap_adapter *na, int onoff)
1160 {
1161 	struct ifnet *ifp = na->ifp;
1162 	if_ctx_t ctx = ifp->if_softc;
1163 
1164 	CTX_LOCK(ctx);
1165 	if (onoff) {
1166 		IFDI_INTR_ENABLE(ctx);
1167 	} else {
1168 		IFDI_INTR_DISABLE(ctx);
1169 	}
1170 	CTX_UNLOCK(ctx);
1171 }
1172 
1173 
1174 static int
1175 iflib_netmap_attach(if_ctx_t ctx)
1176 {
1177 	struct netmap_adapter na;
1178 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1179 
1180 	bzero(&na, sizeof(na));
1181 
1182 	na.ifp = ctx->ifc_ifp;
1183 	na.na_flags = NAF_BDG_MAYSLEEP;
1184 	MPASS(ctx->ifc_softc_ctx.isc_ntxqsets);
1185 	MPASS(ctx->ifc_softc_ctx.isc_nrxqsets);
1186 
1187 	na.num_tx_desc = scctx->isc_ntxd[0];
1188 	na.num_rx_desc = scctx->isc_nrxd[0];
1189 	na.nm_txsync = iflib_netmap_txsync;
1190 	na.nm_rxsync = iflib_netmap_rxsync;
1191 	na.nm_register = iflib_netmap_register;
1192 	na.nm_intr = iflib_netmap_intr;
1193 	na.num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
1194 	na.num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
1195 	return (netmap_attach(&na));
1196 }
1197 
1198 static void
1199 iflib_netmap_txq_init(if_ctx_t ctx, iflib_txq_t txq)
1200 {
1201 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1202 	struct netmap_slot *slot;
1203 
1204 	slot = netmap_reset(na, NR_TX, txq->ift_id, 0);
1205 	if (slot == NULL)
1206 		return;
1207 	if (txq->ift_sds.ifsd_map == NULL)
1208 		return;
1209 
1210 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxd[0]; i++) {
1211 
1212 		/*
1213 		 * In netmap mode, set the map for the packet buffer.
1214 		 * NOTE: Some drivers (not this one) also need to set
1215 		 * the physical buffer address in the NIC ring.
1216 		 * netmap_idx_n2k() maps a nic index, i, into the corresponding
1217 		 * netmap slot index, si
1218 		 */
1219 		int si = netmap_idx_n2k(na->tx_rings[txq->ift_id], i);
1220 		netmap_load_map(na, txq->ift_desc_tag, txq->ift_sds.ifsd_map[i], NMB(na, slot + si));
1221 	}
1222 }
1223 
1224 static void
1225 iflib_netmap_rxq_init(if_ctx_t ctx, iflib_rxq_t rxq)
1226 {
1227 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1228 	struct netmap_kring *kring = na->rx_rings[rxq->ifr_id];
1229 	struct netmap_slot *slot;
1230 	uint32_t nm_i;
1231 
1232 	slot = netmap_reset(na, NR_RX, rxq->ifr_id, 0);
1233 	if (slot == NULL)
1234 		return;
1235 	nm_i = netmap_idx_n2k(kring, 0);
1236 	netmap_fl_refill(rxq, kring, nm_i, true);
1237 }
1238 
1239 static void
1240 iflib_netmap_timer_adjust(if_ctx_t ctx, uint16_t txqid, uint32_t *reset_on)
1241 {
1242 	struct netmap_kring *kring;
1243 
1244 	kring = NA(ctx->ifc_ifp)->tx_rings[txqid];
1245 
1246 	if (kring->nr_hwcur != nm_next(kring->nr_hwtail, kring->nkr_num_slots - 1)) {
1247 		if (ctx->isc_txd_credits_update(ctx->ifc_softc, txqid, false))
1248 			netmap_tx_irq(ctx->ifc_ifp, txqid);
1249 		if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ)) {
1250 			if (hz < 2000)
1251 				*reset_on = 1;
1252 			else
1253 				*reset_on = hz / 1000;
1254 		}
1255 	}
1256 }
1257 
1258 #define iflib_netmap_detach(ifp) netmap_detach(ifp)
1259 
1260 #else
1261 #define iflib_netmap_txq_init(ctx, txq)
1262 #define iflib_netmap_rxq_init(ctx, rxq)
1263 #define iflib_netmap_detach(ifp)
1264 
1265 #define iflib_netmap_attach(ctx) (0)
1266 #define netmap_rx_irq(ifp, qid, budget) (0)
1267 #define netmap_tx_irq(ifp, qid) do {} while (0)
1268 #define iflib_netmap_timer_adjust(ctx, txqid, reset_on)
1269 
1270 #endif
1271 
1272 #if defined(__i386__) || defined(__amd64__)
1273 static __inline void
1274 prefetch(void *x)
1275 {
1276 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1277 }
1278 static __inline void
1279 prefetch2cachelines(void *x)
1280 {
1281 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1282 #if (CACHE_LINE_SIZE < 128)
1283 	__asm volatile("prefetcht0 %0" :: "m" (*(((unsigned long *)x)+CACHE_LINE_SIZE/(sizeof(unsigned long)))));
1284 #endif
1285 }
1286 #else
1287 #define prefetch(x)
1288 #define prefetch2cachelines(x)
1289 #endif
1290 
1291 static void
1292 iflib_gen_mac(if_ctx_t ctx)
1293 {
1294 	struct thread *td;
1295 	MD5_CTX mdctx;
1296 	char uuid[HOSTUUIDLEN+1];
1297 	char buf[HOSTUUIDLEN+16];
1298 	uint8_t *mac;
1299 	unsigned char digest[16];
1300 
1301 	td = curthread;
1302 	mac = ctx->ifc_mac;
1303 	uuid[HOSTUUIDLEN] = 0;
1304 	bcopy(td->td_ucred->cr_prison->pr_hostuuid, uuid, HOSTUUIDLEN);
1305 	snprintf(buf, HOSTUUIDLEN+16, "%s-%s", uuid, device_get_nameunit(ctx->ifc_dev));
1306 	/*
1307 	 * Generate a pseudo-random, deterministic MAC
1308 	 * address based on the UUID and unit number.
1309 	 * The FreeBSD Foundation OUI of 58-9C-FC is used.
1310 	 */
1311 	MD5Init(&mdctx);
1312 	MD5Update(&mdctx, buf, strlen(buf));
1313 	MD5Final(digest, &mdctx);
1314 
1315 	mac[0] = 0x58;
1316 	mac[1] = 0x9C;
1317 	mac[2] = 0xFC;
1318 	mac[3] = digest[0];
1319 	mac[4] = digest[1];
1320 	mac[5] = digest[2];
1321 }
1322 
1323 static void
1324 iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid)
1325 {
1326 	iflib_fl_t fl;
1327 
1328 	fl = &rxq->ifr_fl[flid];
1329 	iru->iru_paddrs = fl->ifl_bus_addrs;
1330 	iru->iru_vaddrs = &fl->ifl_vm_addrs[0];
1331 	iru->iru_idxs = fl->ifl_rxd_idxs;
1332 	iru->iru_qsidx = rxq->ifr_id;
1333 	iru->iru_buf_size = fl->ifl_buf_size;
1334 	iru->iru_flidx = fl->ifl_id;
1335 }
1336 
1337 static void
1338 _iflib_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int err)
1339 {
1340 	if (err)
1341 		return;
1342 	*(bus_addr_t *) arg = segs[0].ds_addr;
1343 }
1344 
1345 int
1346 iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags)
1347 {
1348 	int err;
1349 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1350 	device_t dev = ctx->ifc_dev;
1351 
1352 	KASSERT(sctx->isc_q_align != 0, ("alignment value not initialized"));
1353 
1354 	err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1355 				sctx->isc_q_align, 0,	/* alignment, bounds */
1356 				BUS_SPACE_MAXADDR,	/* lowaddr */
1357 				BUS_SPACE_MAXADDR,	/* highaddr */
1358 				NULL, NULL,		/* filter, filterarg */
1359 				size,			/* maxsize */
1360 				1,			/* nsegments */
1361 				size,			/* maxsegsize */
1362 				BUS_DMA_ALLOCNOW,	/* flags */
1363 				NULL,			/* lockfunc */
1364 				NULL,			/* lockarg */
1365 				&dma->idi_tag);
1366 	if (err) {
1367 		device_printf(dev,
1368 		    "%s: bus_dma_tag_create failed: %d\n",
1369 		    __func__, err);
1370 		goto fail_0;
1371 	}
1372 
1373 	err = bus_dmamem_alloc(dma->idi_tag, (void**) &dma->idi_vaddr,
1374 	    BUS_DMA_NOWAIT | BUS_DMA_COHERENT | BUS_DMA_ZERO, &dma->idi_map);
1375 	if (err) {
1376 		device_printf(dev,
1377 		    "%s: bus_dmamem_alloc(%ju) failed: %d\n",
1378 		    __func__, (uintmax_t)size, err);
1379 		goto fail_1;
1380 	}
1381 
1382 	dma->idi_paddr = IF_BAD_DMA;
1383 	err = bus_dmamap_load(dma->idi_tag, dma->idi_map, dma->idi_vaddr,
1384 	    size, _iflib_dmamap_cb, &dma->idi_paddr, mapflags | BUS_DMA_NOWAIT);
1385 	if (err || dma->idi_paddr == IF_BAD_DMA) {
1386 		device_printf(dev,
1387 		    "%s: bus_dmamap_load failed: %d\n",
1388 		    __func__, err);
1389 		goto fail_2;
1390 	}
1391 
1392 	dma->idi_size = size;
1393 	return (0);
1394 
1395 fail_2:
1396 	bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1397 fail_1:
1398 	bus_dma_tag_destroy(dma->idi_tag);
1399 fail_0:
1400 	dma->idi_tag = NULL;
1401 
1402 	return (err);
1403 }
1404 
1405 int
1406 iflib_dma_alloc_multi(if_ctx_t ctx, int *sizes, iflib_dma_info_t *dmalist, int mapflags, int count)
1407 {
1408 	int i, err;
1409 	iflib_dma_info_t *dmaiter;
1410 
1411 	dmaiter = dmalist;
1412 	for (i = 0; i < count; i++, dmaiter++) {
1413 		if ((err = iflib_dma_alloc(ctx, sizes[i], *dmaiter, mapflags)) != 0)
1414 			break;
1415 	}
1416 	if (err)
1417 		iflib_dma_free_multi(dmalist, i);
1418 	return (err);
1419 }
1420 
1421 void
1422 iflib_dma_free(iflib_dma_info_t dma)
1423 {
1424 	if (dma->idi_tag == NULL)
1425 		return;
1426 	if (dma->idi_paddr != IF_BAD_DMA) {
1427 		bus_dmamap_sync(dma->idi_tag, dma->idi_map,
1428 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1429 		bus_dmamap_unload(dma->idi_tag, dma->idi_map);
1430 		dma->idi_paddr = IF_BAD_DMA;
1431 	}
1432 	if (dma->idi_vaddr != NULL) {
1433 		bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1434 		dma->idi_vaddr = NULL;
1435 	}
1436 	bus_dma_tag_destroy(dma->idi_tag);
1437 	dma->idi_tag = NULL;
1438 }
1439 
1440 void
1441 iflib_dma_free_multi(iflib_dma_info_t *dmalist, int count)
1442 {
1443 	int i;
1444 	iflib_dma_info_t *dmaiter = dmalist;
1445 
1446 	for (i = 0; i < count; i++, dmaiter++)
1447 		iflib_dma_free(*dmaiter);
1448 }
1449 
1450 #ifdef EARLY_AP_STARTUP
1451 static const int iflib_started = 1;
1452 #else
1453 /*
1454  * We used to abuse the smp_started flag to decide if the queues have been
1455  * fully initialized (by late taskqgroup_adjust() calls in a SYSINIT()).
1456  * That gave bad races, since the SYSINIT() runs strictly after smp_started
1457  * is set.  Run a SYSINIT() strictly after that to just set a usable
1458  * completion flag.
1459  */
1460 
1461 static int iflib_started;
1462 
1463 static void
1464 iflib_record_started(void *arg)
1465 {
1466 	iflib_started = 1;
1467 }
1468 
1469 SYSINIT(iflib_record_started, SI_SUB_SMP + 1, SI_ORDER_FIRST,
1470 	iflib_record_started, NULL);
1471 #endif
1472 
1473 static int
1474 iflib_fast_intr(void *arg)
1475 {
1476 	iflib_filter_info_t info = arg;
1477 	struct grouptask *gtask = info->ifi_task;
1478 	if (!iflib_started)
1479 		return (FILTER_HANDLED);
1480 
1481 	DBG_COUNTER_INC(fast_intrs);
1482 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1483 		return (FILTER_HANDLED);
1484 
1485 	GROUPTASK_ENQUEUE(gtask);
1486 	return (FILTER_HANDLED);
1487 }
1488 
1489 static int
1490 iflib_fast_intr_rxtx(void *arg)
1491 {
1492 	iflib_filter_info_t info = arg;
1493 	struct grouptask *gtask = info->ifi_task;
1494 	iflib_rxq_t rxq = (iflib_rxq_t)info->ifi_ctx;
1495 	if_ctx_t ctx = NULL;;
1496 	int i, cidx;
1497 
1498 	if (!iflib_started)
1499 		return (FILTER_HANDLED);
1500 
1501 	DBG_COUNTER_INC(fast_intrs);
1502 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1503 		return (FILTER_HANDLED);
1504 
1505 	MPASS(rxq->ifr_ntxqirq);
1506 	for (i = 0; i < rxq->ifr_ntxqirq; i++) {
1507 		qidx_t txqid = rxq->ifr_txqid[i];
1508 
1509 		ctx = rxq->ifr_ctx;
1510 
1511 		if (!ctx->isc_txd_credits_update(ctx->ifc_softc, txqid, false)) {
1512 			IFDI_TX_QUEUE_INTR_ENABLE(ctx, txqid);
1513 			continue;
1514 		}
1515 		GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
1516 	}
1517 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_RXCQ)
1518 		cidx = rxq->ifr_cq_cidx;
1519 	else
1520 		cidx = rxq->ifr_fl[0].ifl_cidx;
1521 	if (iflib_rxd_avail(ctx, rxq, cidx, 1))
1522 		GROUPTASK_ENQUEUE(gtask);
1523 	else
1524 		IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
1525 	return (FILTER_HANDLED);
1526 }
1527 
1528 
1529 static int
1530 iflib_fast_intr_ctx(void *arg)
1531 {
1532 	iflib_filter_info_t info = arg;
1533 	struct grouptask *gtask = info->ifi_task;
1534 
1535 	if (!iflib_started)
1536 		return (FILTER_HANDLED);
1537 
1538 	DBG_COUNTER_INC(fast_intrs);
1539 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1540 		return (FILTER_HANDLED);
1541 
1542 	GROUPTASK_ENQUEUE(gtask);
1543 	return (FILTER_HANDLED);
1544 }
1545 
1546 static int
1547 _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
1548 		 driver_filter_t filter, driver_intr_t handler, void *arg,
1549 		 const char *name)
1550 {
1551 	int rc, flags;
1552 	struct resource *res;
1553 	void *tag = NULL;
1554 	device_t dev = ctx->ifc_dev;
1555 
1556 	flags = RF_ACTIVE;
1557 	if (ctx->ifc_flags & IFC_LEGACY)
1558 		flags |= RF_SHAREABLE;
1559 	MPASS(rid < 512);
1560 	irq->ii_rid = rid;
1561 	res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &irq->ii_rid, flags);
1562 	if (res == NULL) {
1563 		device_printf(dev,
1564 		    "failed to allocate IRQ for rid %d, name %s.\n", rid, name);
1565 		return (ENOMEM);
1566 	}
1567 	irq->ii_res = res;
1568 	KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL"));
1569 	rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET,
1570 						filter, handler, arg, &tag);
1571 	if (rc != 0) {
1572 		device_printf(dev,
1573 		    "failed to setup interrupt for rid %d, name %s: %d\n",
1574 					  rid, name ? name : "unknown", rc);
1575 		return (rc);
1576 	} else if (name)
1577 		bus_describe_intr(dev, res, tag, "%s", name);
1578 
1579 	irq->ii_tag = tag;
1580 	return (0);
1581 }
1582 
1583 
1584 /*********************************************************************
1585  *
1586  *  Allocate memory for tx_buffer structures. The tx_buffer stores all
1587  *  the information needed to transmit a packet on the wire. This is
1588  *  called only once at attach, setup is done every reset.
1589  *
1590  **********************************************************************/
1591 
1592 static int
1593 iflib_txsd_alloc(iflib_txq_t txq)
1594 {
1595 	if_ctx_t ctx = txq->ift_ctx;
1596 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1597 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1598 	device_t dev = ctx->ifc_dev;
1599 	bus_size_t tsomaxsize;
1600 	int err, nsegments, ntsosegments;
1601 
1602 	nsegments = scctx->isc_tx_nsegments;
1603 	ntsosegments = scctx->isc_tx_tso_segments_max;
1604 	tsomaxsize = scctx->isc_tx_tso_size_max;
1605 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_VLAN_MTU)
1606 		tsomaxsize += sizeof(struct ether_vlan_header);
1607 	MPASS(scctx->isc_ntxd[0] > 0);
1608 	MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0);
1609 	MPASS(nsegments > 0);
1610 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) {
1611 		MPASS(ntsosegments > 0);
1612 		MPASS(sctx->isc_tso_maxsize >= tsomaxsize);
1613 	}
1614 
1615 	/*
1616 	 * Setup DMA descriptor areas.
1617 	 */
1618 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1619 			       1, 0,			/* alignment, bounds */
1620 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1621 			       BUS_SPACE_MAXADDR,	/* highaddr */
1622 			       NULL, NULL,		/* filter, filterarg */
1623 			       sctx->isc_tx_maxsize,		/* maxsize */
1624 			       nsegments,	/* nsegments */
1625 			       sctx->isc_tx_maxsegsize,	/* maxsegsize */
1626 			       0,			/* flags */
1627 			       NULL,			/* lockfunc */
1628 			       NULL,			/* lockfuncarg */
1629 			       &txq->ift_desc_tag))) {
1630 		device_printf(dev,"Unable to allocate TX DMA tag: %d\n", err);
1631 		device_printf(dev,"maxsize: %ju nsegments: %d maxsegsize: %ju\n",
1632 		    (uintmax_t)sctx->isc_tx_maxsize, nsegments, (uintmax_t)sctx->isc_tx_maxsegsize);
1633 		goto fail;
1634 	}
1635 	if ((if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) &
1636 	    (err = bus_dma_tag_create(bus_get_dma_tag(dev),
1637 			       1, 0,			/* alignment, bounds */
1638 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1639 			       BUS_SPACE_MAXADDR,	/* highaddr */
1640 			       NULL, NULL,		/* filter, filterarg */
1641 			       tsomaxsize,		/* maxsize */
1642 			       ntsosegments,	/* nsegments */
1643 			       sctx->isc_tso_maxsegsize,/* maxsegsize */
1644 			       0,			/* flags */
1645 			       NULL,			/* lockfunc */
1646 			       NULL,			/* lockfuncarg */
1647 			       &txq->ift_tso_desc_tag))) {
1648 		device_printf(dev,"Unable to allocate TX TSO DMA tag: %d\n", err);
1649 
1650 		goto fail;
1651 	}
1652 	if (!(txq->ift_sds.ifsd_flags =
1653 	    (uint8_t *) malloc(sizeof(uint8_t) *
1654 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1655 		device_printf(dev, "Unable to allocate tx_buffer memory\n");
1656 		err = ENOMEM;
1657 		goto fail;
1658 	}
1659 	if (!(txq->ift_sds.ifsd_m =
1660 	    (struct mbuf **) malloc(sizeof(struct mbuf *) *
1661 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1662 		device_printf(dev, "Unable to allocate tx_buffer memory\n");
1663 		err = ENOMEM;
1664 		goto fail;
1665 	}
1666 
1667         /* Create the descriptor buffer dma maps */
1668 #if defined(ACPI_DMAR) || (! (defined(__i386__) || defined(__amd64__)))
1669 	if ((ctx->ifc_flags & IFC_DMAR) == 0)
1670 		return (0);
1671 
1672 	if (!(txq->ift_sds.ifsd_map =
1673 	    (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1674 		device_printf(dev, "Unable to allocate tx_buffer map memory\n");
1675 		err = ENOMEM;
1676 		goto fail;
1677 	}
1678 
1679 	for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) {
1680 		err = bus_dmamap_create(txq->ift_desc_tag, 0, &txq->ift_sds.ifsd_map[i]);
1681 		if (err != 0) {
1682 			device_printf(dev, "Unable to create TX DMA map\n");
1683 			goto fail;
1684 		}
1685 	}
1686 #endif
1687 	return (0);
1688 fail:
1689 	/* We free all, it handles case where we are in the middle */
1690 	iflib_tx_structures_free(ctx);
1691 	return (err);
1692 }
1693 
1694 static void
1695 iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i)
1696 {
1697 	bus_dmamap_t map;
1698 
1699 	map = NULL;
1700 	if (txq->ift_sds.ifsd_map != NULL)
1701 		map = txq->ift_sds.ifsd_map[i];
1702 	if (map != NULL) {
1703 		bus_dmamap_unload(txq->ift_desc_tag, map);
1704 		bus_dmamap_destroy(txq->ift_desc_tag, map);
1705 		txq->ift_sds.ifsd_map[i] = NULL;
1706 	}
1707 }
1708 
1709 static void
1710 iflib_txq_destroy(iflib_txq_t txq)
1711 {
1712 	if_ctx_t ctx = txq->ift_ctx;
1713 
1714 	for (int i = 0; i < txq->ift_size; i++)
1715 		iflib_txsd_destroy(ctx, txq, i);
1716 	if (txq->ift_sds.ifsd_map != NULL) {
1717 		free(txq->ift_sds.ifsd_map, M_IFLIB);
1718 		txq->ift_sds.ifsd_map = NULL;
1719 	}
1720 	if (txq->ift_sds.ifsd_m != NULL) {
1721 		free(txq->ift_sds.ifsd_m, M_IFLIB);
1722 		txq->ift_sds.ifsd_m = NULL;
1723 	}
1724 	if (txq->ift_sds.ifsd_flags != NULL) {
1725 		free(txq->ift_sds.ifsd_flags, M_IFLIB);
1726 		txq->ift_sds.ifsd_flags = NULL;
1727 	}
1728 	if (txq->ift_desc_tag != NULL) {
1729 		bus_dma_tag_destroy(txq->ift_desc_tag);
1730 		txq->ift_desc_tag = NULL;
1731 	}
1732 	if (txq->ift_tso_desc_tag != NULL) {
1733 		bus_dma_tag_destroy(txq->ift_tso_desc_tag);
1734 		txq->ift_tso_desc_tag = NULL;
1735 	}
1736 }
1737 
1738 static void
1739 iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i)
1740 {
1741 	struct mbuf **mp;
1742 
1743 	mp = &txq->ift_sds.ifsd_m[i];
1744 	if (*mp == NULL)
1745 		return;
1746 
1747 	if (txq->ift_sds.ifsd_map != NULL) {
1748 		bus_dmamap_sync(txq->ift_desc_tag,
1749 				txq->ift_sds.ifsd_map[i],
1750 				BUS_DMASYNC_POSTWRITE);
1751 		bus_dmamap_unload(txq->ift_desc_tag,
1752 				  txq->ift_sds.ifsd_map[i]);
1753 	}
1754 	m_free(*mp);
1755 	DBG_COUNTER_INC(tx_frees);
1756 	*mp = NULL;
1757 }
1758 
1759 static int
1760 iflib_txq_setup(iflib_txq_t txq)
1761 {
1762 	if_ctx_t ctx = txq->ift_ctx;
1763 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1764 	iflib_dma_info_t di;
1765 	int i;
1766 
1767 	/* Set number of descriptors available */
1768 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
1769 	/* XXX make configurable */
1770 	txq->ift_update_freq = IFLIB_DEFAULT_TX_UPDATE_FREQ;
1771 
1772 	/* Reset indices */
1773 	txq->ift_cidx_processed = 0;
1774 	txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0;
1775 	txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset];
1776 
1777 	for (i = 0, di = txq->ift_ifdi; i < ctx->ifc_nhwtxqs; i++, di++)
1778 		bzero((void *)di->idi_vaddr, di->idi_size);
1779 
1780 	IFDI_TXQ_SETUP(ctx, txq->ift_id);
1781 	for (i = 0, di = txq->ift_ifdi; i < ctx->ifc_nhwtxqs; i++, di++)
1782 		bus_dmamap_sync(di->idi_tag, di->idi_map,
1783 						BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1784 	return (0);
1785 }
1786 
1787 /*********************************************************************
1788  *
1789  *  Allocate memory for rx_buffer structures. Since we use one
1790  *  rx_buffer per received packet, the maximum number of rx_buffer's
1791  *  that we'll need is equal to the number of receive descriptors
1792  *  that we've allocated.
1793  *
1794  **********************************************************************/
1795 static int
1796 iflib_rxsd_alloc(iflib_rxq_t rxq)
1797 {
1798 	if_ctx_t ctx = rxq->ifr_ctx;
1799 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1800 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1801 	device_t dev = ctx->ifc_dev;
1802 	iflib_fl_t fl;
1803 	int			err;
1804 
1805 	MPASS(scctx->isc_nrxd[0] > 0);
1806 	MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0);
1807 
1808 	fl = rxq->ifr_fl;
1809 	for (int i = 0; i <  rxq->ifr_nfl; i++, fl++) {
1810 		fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */
1811 		err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1812 					 1, 0,			/* alignment, bounds */
1813 					 BUS_SPACE_MAXADDR,	/* lowaddr */
1814 					 BUS_SPACE_MAXADDR,	/* highaddr */
1815 					 NULL, NULL,		/* filter, filterarg */
1816 					 sctx->isc_rx_maxsize,	/* maxsize */
1817 					 sctx->isc_rx_nsegments,	/* nsegments */
1818 					 sctx->isc_rx_maxsegsize,	/* maxsegsize */
1819 					 0,			/* flags */
1820 					 NULL,			/* lockfunc */
1821 					 NULL,			/* lockarg */
1822 					 &fl->ifl_desc_tag);
1823 		if (err) {
1824 			device_printf(dev, "%s: bus_dma_tag_create failed %d\n",
1825 				__func__, err);
1826 			goto fail;
1827 		}
1828 		if (!(fl->ifl_sds.ifsd_flags =
1829 		      (uint8_t *) malloc(sizeof(uint8_t) *
1830 					 scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1831 			device_printf(dev, "Unable to allocate tx_buffer memory\n");
1832 			err = ENOMEM;
1833 			goto fail;
1834 		}
1835 		if (!(fl->ifl_sds.ifsd_m =
1836 		      (struct mbuf **) malloc(sizeof(struct mbuf *) *
1837 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1838 			device_printf(dev, "Unable to allocate tx_buffer memory\n");
1839 			err = ENOMEM;
1840 			goto fail;
1841 		}
1842 		if (!(fl->ifl_sds.ifsd_cl =
1843 		      (caddr_t *) malloc(sizeof(caddr_t) *
1844 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1845 			device_printf(dev, "Unable to allocate tx_buffer memory\n");
1846 			err = ENOMEM;
1847 			goto fail;
1848 		}
1849 
1850 		/* Create the descriptor buffer dma maps */
1851 #if defined(ACPI_DMAR) || (! (defined(__i386__) || defined(__amd64__)))
1852 		if ((ctx->ifc_flags & IFC_DMAR) == 0)
1853 			continue;
1854 
1855 		if (!(fl->ifl_sds.ifsd_map =
1856 		      (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1857 			device_printf(dev, "Unable to allocate tx_buffer map memory\n");
1858 			err = ENOMEM;
1859 			goto fail;
1860 		}
1861 
1862 		for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++) {
1863 			err = bus_dmamap_create(fl->ifl_desc_tag, 0, &fl->ifl_sds.ifsd_map[i]);
1864 			if (err != 0) {
1865 				device_printf(dev, "Unable to create RX buffer DMA map\n");
1866 				goto fail;
1867 			}
1868 		}
1869 #endif
1870 	}
1871 	return (0);
1872 
1873 fail:
1874 	iflib_rx_structures_free(ctx);
1875 	return (err);
1876 }
1877 
1878 
1879 /*
1880  * Internal service routines
1881  */
1882 
1883 struct rxq_refill_cb_arg {
1884 	int               error;
1885 	bus_dma_segment_t seg;
1886 	int               nseg;
1887 };
1888 
1889 static void
1890 _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
1891 {
1892 	struct rxq_refill_cb_arg *cb_arg = arg;
1893 
1894 	cb_arg->error = error;
1895 	cb_arg->seg = segs[0];
1896 	cb_arg->nseg = nseg;
1897 }
1898 
1899 
1900 #ifdef ACPI_DMAR
1901 #define IS_DMAR(ctx) (ctx->ifc_flags & IFC_DMAR)
1902 #else
1903 #define IS_DMAR(ctx) (0)
1904 #endif
1905 
1906 /**
1907  *	rxq_refill - refill an rxq  free-buffer list
1908  *	@ctx: the iflib context
1909  *	@rxq: the free-list to refill
1910  *	@n: the number of new buffers to allocate
1911  *
1912  *	(Re)populate an rxq free-buffer list with up to @n new packet buffers.
1913  *	The caller must assure that @n does not exceed the queue's capacity.
1914  */
1915 static void
1916 _iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
1917 {
1918 	struct mbuf *m;
1919 	int idx, frag_idx = fl->ifl_fragidx;
1920         int pidx = fl->ifl_pidx;
1921 	caddr_t cl, *sd_cl;
1922 	struct mbuf **sd_m;
1923 	uint8_t *sd_flags;
1924 	struct if_rxd_update iru;
1925 	bus_dmamap_t *sd_map;
1926 	int n, i = 0;
1927 	uint64_t bus_addr;
1928 	int err;
1929 	qidx_t credits;
1930 
1931 	sd_m = fl->ifl_sds.ifsd_m;
1932 	sd_map = fl->ifl_sds.ifsd_map;
1933 	sd_cl = fl->ifl_sds.ifsd_cl;
1934 	sd_flags = fl->ifl_sds.ifsd_flags;
1935 	idx = pidx;
1936 	credits = fl->ifl_credits;
1937 
1938 	n  = count;
1939 	MPASS(n > 0);
1940 	MPASS(credits + n <= fl->ifl_size);
1941 
1942 	if (pidx < fl->ifl_cidx)
1943 		MPASS(pidx + n <= fl->ifl_cidx);
1944 	if (pidx == fl->ifl_cidx && (credits < fl->ifl_size))
1945 		MPASS(fl->ifl_gen == 0);
1946 	if (pidx > fl->ifl_cidx)
1947 		MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx);
1948 
1949 	DBG_COUNTER_INC(fl_refills);
1950 	if (n > 8)
1951 		DBG_COUNTER_INC(fl_refills_large);
1952 	iru_init(&iru, fl->ifl_rxq, fl->ifl_id);
1953 	while (n--) {
1954 		/*
1955 		 * We allocate an uninitialized mbuf + cluster, mbuf is
1956 		 * initialized after rx.
1957 		 *
1958 		 * If the cluster is still set then we know a minimum sized packet was received
1959 		 */
1960 		bit_ffc_at(fl->ifl_rx_bitmap, frag_idx, fl->ifl_size,  &frag_idx);
1961 		if ((frag_idx < 0) || (frag_idx >= fl->ifl_size))
1962                 	bit_ffc(fl->ifl_rx_bitmap, fl->ifl_size, &frag_idx);
1963 		if ((cl = sd_cl[frag_idx]) == NULL) {
1964                        if ((cl = sd_cl[frag_idx] = m_cljget(NULL, M_NOWAIT, fl->ifl_buf_size)) == NULL)
1965 				break;
1966 #if MEMORY_LOGGING
1967 			fl->ifl_cl_enqueued++;
1968 #endif
1969 		}
1970 		if ((m = m_gethdr(M_NOWAIT, MT_NOINIT)) == NULL) {
1971 			break;
1972 		}
1973 #if MEMORY_LOGGING
1974 		fl->ifl_m_enqueued++;
1975 #endif
1976 
1977 		DBG_COUNTER_INC(rx_allocs);
1978 #if defined(__i386__) || defined(__amd64__)
1979 		if (!IS_DMAR(ctx)) {
1980 			bus_addr = pmap_kextract((vm_offset_t)cl);
1981 		} else
1982 #endif
1983 		{
1984 			struct rxq_refill_cb_arg cb_arg;
1985 
1986 			cb_arg.error = 0;
1987 			MPASS(sd_map != NULL);
1988 			MPASS(sd_map[frag_idx] != NULL);
1989 			err = bus_dmamap_load(fl->ifl_desc_tag, sd_map[frag_idx],
1990 		         cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg, 0);
1991 			bus_dmamap_sync(fl->ifl_desc_tag, sd_map[frag_idx],
1992 					BUS_DMASYNC_PREREAD);
1993 
1994 			if (err != 0 || cb_arg.error) {
1995 				/*
1996 				 * !zone_pack ?
1997 				 */
1998 				if (fl->ifl_zone == zone_pack)
1999 					uma_zfree(fl->ifl_zone, cl);
2000 				m_free(m);
2001 				n = 0;
2002 				goto done;
2003 			}
2004 			bus_addr = cb_arg.seg.ds_addr;
2005 		}
2006                 bit_set(fl->ifl_rx_bitmap, frag_idx);
2007 		sd_flags[frag_idx] |= RX_SW_DESC_INUSE;
2008 
2009 		MPASS(sd_m[frag_idx] == NULL);
2010 		sd_cl[frag_idx] = cl;
2011 		sd_m[frag_idx] = m;
2012 		fl->ifl_rxd_idxs[i] = frag_idx;
2013 		fl->ifl_bus_addrs[i] = bus_addr;
2014 		fl->ifl_vm_addrs[i] = cl;
2015 		credits++;
2016 		i++;
2017 		MPASS(credits <= fl->ifl_size);
2018 		if (++idx == fl->ifl_size) {
2019 			fl->ifl_gen = 1;
2020 			idx = 0;
2021 		}
2022 		if (n == 0 || i == IFLIB_MAX_RX_REFRESH) {
2023 			iru.iru_pidx = pidx;
2024 			iru.iru_count = i;
2025 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2026 			i = 0;
2027 			pidx = idx;
2028 			fl->ifl_pidx = idx;
2029 			fl->ifl_credits = credits;
2030 		}
2031 
2032 	}
2033 done:
2034 	if (i) {
2035 		iru.iru_pidx = pidx;
2036 		iru.iru_count = i;
2037 		ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2038 		fl->ifl_pidx = idx;
2039 		fl->ifl_credits = credits;
2040 	}
2041 	DBG_COUNTER_INC(rxd_flush);
2042 	if (fl->ifl_pidx == 0)
2043 		pidx = fl->ifl_size - 1;
2044 	else
2045 		pidx = fl->ifl_pidx - 1;
2046 
2047 	if (sd_map)
2048 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2049 				BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2050 	ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id, fl->ifl_id, pidx);
2051 	fl->ifl_fragidx = frag_idx;
2052 }
2053 
2054 static __inline void
2055 __iflib_fl_refill_lt(if_ctx_t ctx, iflib_fl_t fl, int max)
2056 {
2057 	/* we avoid allowing pidx to catch up with cidx as it confuses ixl */
2058 	int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1;
2059 #ifdef INVARIANTS
2060 	int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1;
2061 #endif
2062 
2063 	MPASS(fl->ifl_credits <= fl->ifl_size);
2064 	MPASS(reclaimable == delta);
2065 
2066 	if (reclaimable > 0)
2067 		_iflib_fl_refill(ctx, fl, min(max, reclaimable));
2068 }
2069 
2070 static void
2071 iflib_fl_bufs_free(iflib_fl_t fl)
2072 {
2073 	iflib_dma_info_t idi = fl->ifl_ifdi;
2074 	uint32_t i;
2075 
2076 	for (i = 0; i < fl->ifl_size; i++) {
2077 		struct mbuf **sd_m = &fl->ifl_sds.ifsd_m[i];
2078 		uint8_t *sd_flags = &fl->ifl_sds.ifsd_flags[i];
2079 		caddr_t *sd_cl = &fl->ifl_sds.ifsd_cl[i];
2080 
2081 		if (*sd_flags & RX_SW_DESC_INUSE) {
2082 			if (fl->ifl_sds.ifsd_map != NULL) {
2083 				bus_dmamap_t sd_map = fl->ifl_sds.ifsd_map[i];
2084 				bus_dmamap_unload(fl->ifl_desc_tag, sd_map);
2085 				if (fl->ifl_rxq->ifr_ctx->ifc_in_detach)
2086 					bus_dmamap_destroy(fl->ifl_desc_tag, sd_map);
2087 			}
2088 			if (*sd_m != NULL) {
2089 				m_init(*sd_m, M_NOWAIT, MT_DATA, 0);
2090 				uma_zfree(zone_mbuf, *sd_m);
2091 			}
2092 			if (*sd_cl != NULL)
2093 				uma_zfree(fl->ifl_zone, *sd_cl);
2094 			*sd_flags = 0;
2095 		} else {
2096 			MPASS(*sd_cl == NULL);
2097 			MPASS(*sd_m == NULL);
2098 		}
2099 #if MEMORY_LOGGING
2100 		fl->ifl_m_dequeued++;
2101 		fl->ifl_cl_dequeued++;
2102 #endif
2103 		*sd_cl = NULL;
2104 		*sd_m = NULL;
2105 	}
2106 #ifdef INVARIANTS
2107 	for (i = 0; i < fl->ifl_size; i++) {
2108 		MPASS(fl->ifl_sds.ifsd_flags[i] == 0);
2109 		MPASS(fl->ifl_sds.ifsd_cl[i] == NULL);
2110 		MPASS(fl->ifl_sds.ifsd_m[i] == NULL);
2111 	}
2112 #endif
2113 	/*
2114 	 * Reset free list values
2115 	 */
2116 	fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = fl->ifl_fragidx = 0;
2117 	bzero(idi->idi_vaddr, idi->idi_size);
2118 }
2119 
2120 /*********************************************************************
2121  *
2122  *  Initialize a receive ring and its buffers.
2123  *
2124  **********************************************************************/
2125 static int
2126 iflib_fl_setup(iflib_fl_t fl)
2127 {
2128 	iflib_rxq_t rxq = fl->ifl_rxq;
2129 	if_ctx_t ctx = rxq->ifr_ctx;
2130 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2131 
2132 	bit_nclear(fl->ifl_rx_bitmap, 0, fl->ifl_size - 1);
2133 	/*
2134 	** Free current RX buffer structs and their mbufs
2135 	*/
2136 	iflib_fl_bufs_free(fl);
2137 	/* Now replenish the mbufs */
2138 	MPASS(fl->ifl_credits == 0);
2139 	/*
2140 	 * XXX don't set the max_frame_size to larger
2141 	 * than the hardware can handle
2142 	 */
2143 	if (sctx->isc_max_frame_size <= 2048)
2144 		fl->ifl_buf_size = MCLBYTES;
2145 #ifndef CONTIGMALLOC_WORKS
2146 	else
2147 		fl->ifl_buf_size = MJUMPAGESIZE;
2148 #else
2149 	else if (sctx->isc_max_frame_size <= 4096)
2150 		fl->ifl_buf_size = MJUMPAGESIZE;
2151 	else if (sctx->isc_max_frame_size <= 9216)
2152 		fl->ifl_buf_size = MJUM9BYTES;
2153 	else
2154 		fl->ifl_buf_size = MJUM16BYTES;
2155 #endif
2156 	if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size)
2157 		ctx->ifc_max_fl_buf_size = fl->ifl_buf_size;
2158 	fl->ifl_cltype = m_gettype(fl->ifl_buf_size);
2159 	fl->ifl_zone = m_getzone(fl->ifl_buf_size);
2160 
2161 
2162 	/* avoid pre-allocating zillions of clusters to an idle card
2163 	 * potentially speeding up attach
2164 	 */
2165 	_iflib_fl_refill(ctx, fl, min(128, fl->ifl_size));
2166 	MPASS(min(128, fl->ifl_size) == fl->ifl_credits);
2167 	if (min(128, fl->ifl_size) != fl->ifl_credits)
2168 		return (ENOBUFS);
2169 	/*
2170 	 * handle failure
2171 	 */
2172 	MPASS(rxq != NULL);
2173 	MPASS(fl->ifl_ifdi != NULL);
2174 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2175 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2176 	return (0);
2177 }
2178 
2179 /*********************************************************************
2180  *
2181  *  Free receive ring data structures
2182  *
2183  **********************************************************************/
2184 static void
2185 iflib_rx_sds_free(iflib_rxq_t rxq)
2186 {
2187 	iflib_fl_t fl;
2188 	int i;
2189 
2190 	if (rxq->ifr_fl != NULL) {
2191 		for (i = 0; i < rxq->ifr_nfl; i++) {
2192 			fl = &rxq->ifr_fl[i];
2193 			if (fl->ifl_desc_tag != NULL) {
2194 				bus_dma_tag_destroy(fl->ifl_desc_tag);
2195 				fl->ifl_desc_tag = NULL;
2196 			}
2197 			free(fl->ifl_sds.ifsd_m, M_IFLIB);
2198 			free(fl->ifl_sds.ifsd_cl, M_IFLIB);
2199 			/* XXX destroy maps first */
2200 			free(fl->ifl_sds.ifsd_map, M_IFLIB);
2201 			fl->ifl_sds.ifsd_m = NULL;
2202 			fl->ifl_sds.ifsd_cl = NULL;
2203 			fl->ifl_sds.ifsd_map = NULL;
2204 		}
2205 		free(rxq->ifr_fl, M_IFLIB);
2206 		rxq->ifr_fl = NULL;
2207 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
2208 	}
2209 }
2210 
2211 /*
2212  * MI independent logic
2213  *
2214  */
2215 static void
2216 iflib_timer(void *arg)
2217 {
2218 	iflib_txq_t txq = arg;
2219 	if_ctx_t ctx = txq->ift_ctx;
2220 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2221 	uint64_t this_tick = ticks;
2222 	uint32_t reset_on = hz / 2;
2223 
2224 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
2225 		return;
2226 	/*
2227 	** Check on the state of the TX queue(s), this
2228 	** can be done without the lock because its RO
2229 	** and the HUNG state will be static if set.
2230 	*/
2231 	if (this_tick - txq->ift_last_timer_tick >= hz / 2) {
2232 		txq->ift_last_timer_tick = this_tick;
2233 		IFDI_TIMER(ctx, txq->ift_id);
2234 		if ((txq->ift_qstatus == IFLIB_QUEUE_HUNG) &&
2235 		    ((txq->ift_cleaned_prev == txq->ift_cleaned) ||
2236 		     (sctx->isc_pause_frames == 0)))
2237 			goto hung;
2238 
2239 		if (ifmp_ring_is_stalled(txq->ift_br))
2240 			txq->ift_qstatus = IFLIB_QUEUE_HUNG;
2241 		txq->ift_cleaned_prev = txq->ift_cleaned;
2242 	}
2243 #ifdef DEV_NETMAP
2244 	if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP)
2245 		iflib_netmap_timer_adjust(ctx, txq->ift_id, &reset_on);
2246 #endif
2247 	/* handle any laggards */
2248 	if (txq->ift_db_pending)
2249 		GROUPTASK_ENQUEUE(&txq->ift_task);
2250 
2251 	sctx->isc_pause_frames = 0;
2252 	if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)
2253 		callout_reset_on(&txq->ift_timer, reset_on, iflib_timer, txq, txq->ift_timer.c_cpu);
2254 	return;
2255  hung:
2256 	device_printf(ctx->ifc_dev,  "TX(%d) desc avail = %d, pidx = %d\n",
2257 				  txq->ift_id, TXQ_AVAIL(txq), txq->ift_pidx);
2258 	STATE_LOCK(ctx);
2259 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2260 	ctx->ifc_flags |= (IFC_DO_WATCHDOG|IFC_DO_RESET);
2261 	iflib_admin_intr_deferred(ctx);
2262 	STATE_UNLOCK(ctx);
2263 }
2264 
2265 static void
2266 iflib_init_locked(if_ctx_t ctx)
2267 {
2268 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2269 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2270 	if_t ifp = ctx->ifc_ifp;
2271 	iflib_fl_t fl;
2272 	iflib_txq_t txq;
2273 	iflib_rxq_t rxq;
2274 	int i, j, tx_ip_csum_flags, tx_ip6_csum_flags;
2275 
2276 
2277 	if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2278 	IFDI_INTR_DISABLE(ctx);
2279 
2280 	tx_ip_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP);
2281 	tx_ip6_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | CSUM_IP6_SCTP);
2282 	/* Set hardware offload abilities */
2283 	if_clearhwassist(ifp);
2284 	if (if_getcapenable(ifp) & IFCAP_TXCSUM)
2285 		if_sethwassistbits(ifp, tx_ip_csum_flags, 0);
2286 	if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6)
2287 		if_sethwassistbits(ifp,  tx_ip6_csum_flags, 0);
2288 	if (if_getcapenable(ifp) & IFCAP_TSO4)
2289 		if_sethwassistbits(ifp, CSUM_IP_TSO, 0);
2290 	if (if_getcapenable(ifp) & IFCAP_TSO6)
2291 		if_sethwassistbits(ifp, CSUM_IP6_TSO, 0);
2292 
2293 	for (i = 0, txq = ctx->ifc_txqs; i < sctx->isc_ntxqsets; i++, txq++) {
2294 		CALLOUT_LOCK(txq);
2295 		callout_stop(&txq->ift_timer);
2296 		CALLOUT_UNLOCK(txq);
2297 		iflib_netmap_txq_init(ctx, txq);
2298 	}
2299 #ifdef INVARIANTS
2300 	i = if_getdrvflags(ifp);
2301 #endif
2302 	IFDI_INIT(ctx);
2303 	MPASS(if_getdrvflags(ifp) == i);
2304 	for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
2305 		/* XXX this should really be done on a per-queue basis */
2306 		if (if_getcapenable(ifp) & IFCAP_NETMAP) {
2307 			MPASS(rxq->ifr_id == i);
2308 			iflib_netmap_rxq_init(ctx, rxq);
2309 			continue;
2310 		}
2311 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
2312 			if (iflib_fl_setup(fl)) {
2313 				device_printf(ctx->ifc_dev, "freelist setup failed - check cluster settings\n");
2314 				goto done;
2315 			}
2316 		}
2317 	}
2318 done:
2319 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
2320 	IFDI_INTR_ENABLE(ctx);
2321 	txq = ctx->ifc_txqs;
2322 	for (i = 0; i < sctx->isc_ntxqsets; i++, txq++)
2323 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq,
2324 			txq->ift_timer.c_cpu);
2325 }
2326 
2327 static int
2328 iflib_media_change(if_t ifp)
2329 {
2330 	if_ctx_t ctx = if_getsoftc(ifp);
2331 	int err;
2332 
2333 	CTX_LOCK(ctx);
2334 	if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0)
2335 		iflib_init_locked(ctx);
2336 	CTX_UNLOCK(ctx);
2337 	return (err);
2338 }
2339 
2340 static void
2341 iflib_media_status(if_t ifp, struct ifmediareq *ifmr)
2342 {
2343 	if_ctx_t ctx = if_getsoftc(ifp);
2344 
2345 	CTX_LOCK(ctx);
2346 	IFDI_UPDATE_ADMIN_STATUS(ctx);
2347 	IFDI_MEDIA_STATUS(ctx, ifmr);
2348 	CTX_UNLOCK(ctx);
2349 }
2350 
2351 void
2352 iflib_stop(if_ctx_t ctx)
2353 {
2354 	iflib_txq_t txq = ctx->ifc_txqs;
2355 	iflib_rxq_t rxq = ctx->ifc_rxqs;
2356 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2357 	iflib_dma_info_t di;
2358 	iflib_fl_t fl;
2359 	int i, j;
2360 
2361 	/* Tell the stack that the interface is no longer active */
2362 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2363 
2364 	IFDI_INTR_DISABLE(ctx);
2365 	DELAY(1000);
2366 	IFDI_STOP(ctx);
2367 	DELAY(1000);
2368 
2369 	iflib_debug_reset();
2370 	/* Wait for current tx queue users to exit to disarm watchdog timer. */
2371 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) {
2372 		/* make sure all transmitters have completed before proceeding XXX */
2373 
2374 		CALLOUT_LOCK(txq);
2375 		callout_stop(&txq->ift_timer);
2376 		CALLOUT_UNLOCK(txq);
2377 
2378 		/* clean any enqueued buffers */
2379 		iflib_ifmp_purge(txq);
2380 		/* Free any existing tx buffers. */
2381 		for (j = 0; j < txq->ift_size; j++) {
2382 			iflib_txsd_free(ctx, txq, j);
2383 		}
2384 		txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0;
2385 		txq->ift_in_use = txq->ift_gen = txq->ift_cidx = txq->ift_pidx = txq->ift_no_desc_avail = 0;
2386 		txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0;
2387 		txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0;
2388 		txq->ift_pullups = 0;
2389 		ifmp_ring_reset_stats(txq->ift_br);
2390 		for (j = 0, di = txq->ift_ifdi; j < ctx->ifc_nhwtxqs; j++, di++)
2391 			bzero((void *)di->idi_vaddr, di->idi_size);
2392 	}
2393 	for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) {
2394 		/* make sure all transmitters have completed before proceeding XXX */
2395 
2396 		for (j = 0, di = rxq->ifr_ifdi; j < rxq->ifr_nfl; j++, di++)
2397 			bzero((void *)di->idi_vaddr, di->idi_size);
2398 		/* also resets the free lists pidx/cidx */
2399 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
2400 			iflib_fl_bufs_free(fl);
2401 	}
2402 }
2403 
2404 static inline caddr_t
2405 calc_next_rxd(iflib_fl_t fl, int cidx)
2406 {
2407 	qidx_t size;
2408 	int nrxd;
2409 	caddr_t start, end, cur, next;
2410 
2411 	nrxd = fl->ifl_size;
2412 	size = fl->ifl_rxd_size;
2413 	start = fl->ifl_ifdi->idi_vaddr;
2414 
2415 	if (__predict_false(size == 0))
2416 		return (start);
2417 	cur = start + size*cidx;
2418 	end = start + size*nrxd;
2419 	next = CACHE_PTR_NEXT(cur);
2420 	return (next < end ? next : start);
2421 }
2422 
2423 static inline void
2424 prefetch_pkts(iflib_fl_t fl, int cidx)
2425 {
2426 	int nextptr;
2427 	int nrxd = fl->ifl_size;
2428 	caddr_t next_rxd;
2429 
2430 
2431 	nextptr = (cidx + CACHE_PTR_INCREMENT) & (nrxd-1);
2432 	prefetch(&fl->ifl_sds.ifsd_m[nextptr]);
2433 	prefetch(&fl->ifl_sds.ifsd_cl[nextptr]);
2434 	next_rxd = calc_next_rxd(fl, cidx);
2435 	prefetch(next_rxd);
2436 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 1) & (nrxd-1)]);
2437 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 2) & (nrxd-1)]);
2438 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 3) & (nrxd-1)]);
2439 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 4) & (nrxd-1)]);
2440 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 1) & (nrxd-1)]);
2441 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 2) & (nrxd-1)]);
2442 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 3) & (nrxd-1)]);
2443 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 4) & (nrxd-1)]);
2444 }
2445 
2446 static void
2447 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, int unload, if_rxsd_t sd)
2448 {
2449 	int flid, cidx;
2450 	bus_dmamap_t map;
2451 	iflib_fl_t fl;
2452 	iflib_dma_info_t di;
2453 	int next;
2454 
2455 	map = NULL;
2456 	flid = irf->irf_flid;
2457 	cidx = irf->irf_idx;
2458 	fl = &rxq->ifr_fl[flid];
2459 	sd->ifsd_fl = fl;
2460 	sd->ifsd_cidx = cidx;
2461 	sd->ifsd_m = &fl->ifl_sds.ifsd_m[cidx];
2462 	sd->ifsd_cl = &fl->ifl_sds.ifsd_cl[cidx];
2463 	fl->ifl_credits--;
2464 #if MEMORY_LOGGING
2465 	fl->ifl_m_dequeued++;
2466 #endif
2467 	if (rxq->ifr_ctx->ifc_flags & IFC_PREFETCH)
2468 		prefetch_pkts(fl, cidx);
2469 	if (fl->ifl_sds.ifsd_map != NULL) {
2470 		next = (cidx + CACHE_PTR_INCREMENT) & (fl->ifl_size-1);
2471 		prefetch(&fl->ifl_sds.ifsd_map[next]);
2472 		map = fl->ifl_sds.ifsd_map[cidx];
2473 		di = fl->ifl_ifdi;
2474 		next = (cidx + CACHE_LINE_SIZE) & (fl->ifl_size-1);
2475 		prefetch(&fl->ifl_sds.ifsd_flags[next]);
2476 		bus_dmamap_sync(di->idi_tag, di->idi_map,
2477 				BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
2478 
2479 	/* not valid assert if bxe really does SGE from non-contiguous elements */
2480 		MPASS(fl->ifl_cidx == cidx);
2481 		if (unload)
2482 			bus_dmamap_unload(fl->ifl_desc_tag, map);
2483 	}
2484 	fl->ifl_cidx = (fl->ifl_cidx + 1) & (fl->ifl_size-1);
2485 	if (__predict_false(fl->ifl_cidx == 0))
2486 		fl->ifl_gen = 0;
2487 	if (map != NULL)
2488 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2489 			BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2490         bit_clear(fl->ifl_rx_bitmap, cidx);
2491 }
2492 
2493 static struct mbuf *
2494 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri, if_rxsd_t sd)
2495 {
2496 	int i, padlen , flags;
2497 	struct mbuf *m, *mh, *mt;
2498 	caddr_t cl;
2499 
2500 	i = 0;
2501 	mh = NULL;
2502 	do {
2503 		rxd_frag_to_sd(rxq, &ri->iri_frags[i], TRUE, sd);
2504 
2505 		MPASS(*sd->ifsd_cl != NULL);
2506 		MPASS(*sd->ifsd_m != NULL);
2507 
2508 		/* Don't include zero-length frags */
2509 		if (ri->iri_frags[i].irf_len == 0) {
2510 			/* XXX we can save the cluster here, but not the mbuf */
2511 			m_init(*sd->ifsd_m, M_NOWAIT, MT_DATA, 0);
2512 			m_free(*sd->ifsd_m);
2513 			*sd->ifsd_m = NULL;
2514 			continue;
2515 		}
2516 		m = *sd->ifsd_m;
2517 		*sd->ifsd_m = NULL;
2518 		if (mh == NULL) {
2519 			flags = M_PKTHDR|M_EXT;
2520 			mh = mt = m;
2521 			padlen = ri->iri_pad;
2522 		} else {
2523 			flags = M_EXT;
2524 			mt->m_next = m;
2525 			mt = m;
2526 			/* assuming padding is only on the first fragment */
2527 			padlen = 0;
2528 		}
2529 		cl = *sd->ifsd_cl;
2530 		*sd->ifsd_cl = NULL;
2531 
2532 		/* Can these two be made one ? */
2533 		m_init(m, M_NOWAIT, MT_DATA, flags);
2534 		m_cljset(m, cl, sd->ifsd_fl->ifl_cltype);
2535 		/*
2536 		 * These must follow m_init and m_cljset
2537 		 */
2538 		m->m_data += padlen;
2539 		ri->iri_len -= padlen;
2540 		m->m_len = ri->iri_frags[i].irf_len;
2541 	} while (++i < ri->iri_nfrags);
2542 
2543 	return (mh);
2544 }
2545 
2546 /*
2547  * Process one software descriptor
2548  */
2549 static struct mbuf *
2550 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri)
2551 {
2552 	struct if_rxsd sd;
2553 	struct mbuf *m;
2554 
2555 	/* should I merge this back in now that the two paths are basically duplicated? */
2556 	if (ri->iri_nfrags == 1 &&
2557 	    ri->iri_frags[0].irf_len <= MIN(IFLIB_RX_COPY_THRESH, MHLEN)) {
2558 		rxd_frag_to_sd(rxq, &ri->iri_frags[0], FALSE, &sd);
2559 		m = *sd.ifsd_m;
2560 		*sd.ifsd_m = NULL;
2561 		m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR);
2562 #ifndef __NO_STRICT_ALIGNMENT
2563 		if (!IP_ALIGNED(m))
2564 			m->m_data += 2;
2565 #endif
2566 		memcpy(m->m_data, *sd.ifsd_cl, ri->iri_len);
2567 		m->m_len = ri->iri_frags[0].irf_len;
2568        } else {
2569 		m = assemble_segments(rxq, ri, &sd);
2570 	}
2571 	m->m_pkthdr.len = ri->iri_len;
2572 	m->m_pkthdr.rcvif = ri->iri_ifp;
2573 	m->m_flags |= ri->iri_flags;
2574 	m->m_pkthdr.ether_vtag = ri->iri_vtag;
2575 	m->m_pkthdr.flowid = ri->iri_flowid;
2576 	M_HASHTYPE_SET(m, ri->iri_rsstype);
2577 	m->m_pkthdr.csum_flags = ri->iri_csum_flags;
2578 	m->m_pkthdr.csum_data = ri->iri_csum_data;
2579 	return (m);
2580 }
2581 
2582 #if defined(INET6) || defined(INET)
2583 static void
2584 iflib_get_ip_forwarding(struct lro_ctrl *lc, bool *v4, bool *v6)
2585 {
2586 	CURVNET_SET(lc->ifp->if_vnet);
2587 #if defined(INET6)
2588 	*v6 = VNET(ip6_forwarding);
2589 #endif
2590 #if defined(INET)
2591 	*v4 = VNET(ipforwarding);
2592 #endif
2593 	CURVNET_RESTORE();
2594 }
2595 
2596 /*
2597  * Returns true if it's possible this packet could be LROed.
2598  * if it returns false, it is guaranteed that tcp_lro_rx()
2599  * would not return zero.
2600  */
2601 static bool
2602 iflib_check_lro_possible(struct mbuf *m, bool v4_forwarding, bool v6_forwarding)
2603 {
2604 	struct ether_header *eh;
2605 	uint16_t eh_type;
2606 
2607 	eh = mtod(m, struct ether_header *);
2608 	eh_type = ntohs(eh->ether_type);
2609 	switch (eh_type) {
2610 #if defined(INET6)
2611 		case ETHERTYPE_IPV6:
2612 			return !v6_forwarding;
2613 #endif
2614 #if defined (INET)
2615 		case ETHERTYPE_IP:
2616 			return !v4_forwarding;
2617 #endif
2618 	}
2619 
2620 	return false;
2621 }
2622 #else
2623 static void
2624 iflib_get_ip_forwarding(struct lro_ctrl *lc __unused, bool *v4 __unused, bool *v6 __unused)
2625 {
2626 }
2627 #endif
2628 
2629 static bool
2630 iflib_rxeof(iflib_rxq_t rxq, qidx_t budget)
2631 {
2632 	if_ctx_t ctx = rxq->ifr_ctx;
2633 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2634 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2635 	int avail, i;
2636 	qidx_t *cidxp;
2637 	struct if_rxd_info ri;
2638 	int err, budget_left, rx_bytes, rx_pkts;
2639 	iflib_fl_t fl;
2640 	struct ifnet *ifp;
2641 	int lro_enabled;
2642 	bool v4_forwarding, v6_forwarding, lro_possible;
2643 
2644 	/*
2645 	 * XXX early demux data packets so that if_input processing only handles
2646 	 * acks in interrupt context
2647 	 */
2648 	struct mbuf *m, *mh, *mt, *mf;
2649 
2650 	lro_possible = v4_forwarding = v6_forwarding = false;
2651 	ifp = ctx->ifc_ifp;
2652 	mh = mt = NULL;
2653 	MPASS(budget > 0);
2654 	rx_pkts	= rx_bytes = 0;
2655 	if (sctx->isc_flags & IFLIB_HAS_RXCQ)
2656 		cidxp = &rxq->ifr_cq_cidx;
2657 	else
2658 		cidxp = &rxq->ifr_fl[0].ifl_cidx;
2659 	if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) {
2660 		for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2661 			__iflib_fl_refill_lt(ctx, fl, budget + 8);
2662 		DBG_COUNTER_INC(rx_unavail);
2663 		return (false);
2664 	}
2665 
2666 	for (budget_left = budget; budget_left > 0 && avail > 0;) {
2667 		if (__predict_false(!CTX_ACTIVE(ctx))) {
2668 			DBG_COUNTER_INC(rx_ctx_inactive);
2669 			break;
2670 		}
2671 		/*
2672 		 * Reset client set fields to their default values
2673 		 */
2674 		rxd_info_zero(&ri);
2675 		ri.iri_qsidx = rxq->ifr_id;
2676 		ri.iri_cidx = *cidxp;
2677 		ri.iri_ifp = ifp;
2678 		ri.iri_frags = rxq->ifr_frags;
2679 		err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
2680 
2681 		if (err)
2682 			goto err;
2683 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
2684 			*cidxp = ri.iri_cidx;
2685 			/* Update our consumer index */
2686 			/* XXX NB: shurd - check if this is still safe */
2687 			while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0]) {
2688 				rxq->ifr_cq_cidx -= scctx->isc_nrxd[0];
2689 				rxq->ifr_cq_gen = 0;
2690 			}
2691 			/* was this only a completion queue message? */
2692 			if (__predict_false(ri.iri_nfrags == 0))
2693 				continue;
2694 		}
2695 		MPASS(ri.iri_nfrags != 0);
2696 		MPASS(ri.iri_len != 0);
2697 
2698 		/* will advance the cidx on the corresponding free lists */
2699 		m = iflib_rxd_pkt_get(rxq, &ri);
2700 		avail--;
2701 		budget_left--;
2702 		if (avail == 0 && budget_left)
2703 			avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left);
2704 
2705 		if (__predict_false(m == NULL)) {
2706 			DBG_COUNTER_INC(rx_mbuf_null);
2707 			continue;
2708 		}
2709 		/* imm_pkt: -- cxgb */
2710 		if (mh == NULL)
2711 			mh = mt = m;
2712 		else {
2713 			mt->m_nextpkt = m;
2714 			mt = m;
2715 		}
2716 	}
2717 	/* make sure that we can refill faster than drain */
2718 	for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2719 		__iflib_fl_refill_lt(ctx, fl, budget + 8);
2720 
2721 	lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO);
2722 	if (lro_enabled)
2723 		iflib_get_ip_forwarding(&rxq->ifr_lc, &v4_forwarding, &v6_forwarding);
2724 	mt = mf = NULL;
2725 	while (mh != NULL) {
2726 		m = mh;
2727 		mh = mh->m_nextpkt;
2728 		m->m_nextpkt = NULL;
2729 #ifndef __NO_STRICT_ALIGNMENT
2730 		if (!IP_ALIGNED(m) && (m = iflib_fixup_rx(m)) == NULL)
2731 			continue;
2732 #endif
2733 		rx_bytes += m->m_pkthdr.len;
2734 		rx_pkts++;
2735 #if defined(INET6) || defined(INET)
2736 		if (lro_enabled) {
2737 			if (!lro_possible) {
2738 				lro_possible = iflib_check_lro_possible(m, v4_forwarding, v6_forwarding);
2739 				if (lro_possible && mf != NULL) {
2740 					ifp->if_input(ifp, mf);
2741 					DBG_COUNTER_INC(rx_if_input);
2742 					mt = mf = NULL;
2743 				}
2744 			}
2745 			if ((m->m_pkthdr.csum_flags & (CSUM_L4_CALC|CSUM_L4_VALID)) ==
2746 			    (CSUM_L4_CALC|CSUM_L4_VALID)) {
2747 				if (lro_possible && tcp_lro_rx(&rxq->ifr_lc, m, 0) == 0)
2748 					continue;
2749 			}
2750 		}
2751 #endif
2752 		if (lro_possible) {
2753 			ifp->if_input(ifp, m);
2754 			DBG_COUNTER_INC(rx_if_input);
2755 			continue;
2756 		}
2757 
2758 		if (mf == NULL)
2759 			mf = m;
2760 		if (mt != NULL)
2761 			mt->m_nextpkt = m;
2762 		mt = m;
2763 	}
2764 	if (mf != NULL) {
2765 		ifp->if_input(ifp, mf);
2766 		DBG_COUNTER_INC(rx_if_input);
2767 	}
2768 
2769 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
2770 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
2771 
2772 	/*
2773 	 * Flush any outstanding LRO work
2774 	 */
2775 #if defined(INET6) || defined(INET)
2776 	tcp_lro_flush_all(&rxq->ifr_lc);
2777 #endif
2778 	if (avail)
2779 		return true;
2780 	return (iflib_rxd_avail(ctx, rxq, *cidxp, 1));
2781 err:
2782 	STATE_LOCK(ctx);
2783 	ctx->ifc_flags |= IFC_DO_RESET;
2784 	iflib_admin_intr_deferred(ctx);
2785 	STATE_UNLOCK(ctx);
2786 	return (false);
2787 }
2788 
2789 #define TXD_NOTIFY_COUNT(txq) (((txq)->ift_size / (txq)->ift_update_freq)-1)
2790 static inline qidx_t
2791 txq_max_db_deferred(iflib_txq_t txq, qidx_t in_use)
2792 {
2793 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2794 	qidx_t minthresh = txq->ift_size / 8;
2795 	if (in_use > 4*minthresh)
2796 		return (notify_count);
2797 	if (in_use > 2*minthresh)
2798 		return (notify_count >> 1);
2799 	if (in_use > minthresh)
2800 		return (notify_count >> 3);
2801 	return (0);
2802 }
2803 
2804 static inline qidx_t
2805 txq_max_rs_deferred(iflib_txq_t txq)
2806 {
2807 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2808 	qidx_t minthresh = txq->ift_size / 8;
2809 	if (txq->ift_in_use > 4*minthresh)
2810 		return (notify_count);
2811 	if (txq->ift_in_use > 2*minthresh)
2812 		return (notify_count >> 1);
2813 	if (txq->ift_in_use > minthresh)
2814 		return (notify_count >> 2);
2815 	return (2);
2816 }
2817 
2818 #define M_CSUM_FLAGS(m) ((m)->m_pkthdr.csum_flags)
2819 #define M_HAS_VLANTAG(m) (m->m_flags & M_VLANTAG)
2820 
2821 #define TXQ_MAX_DB_DEFERRED(txq, in_use) txq_max_db_deferred((txq), (in_use))
2822 #define TXQ_MAX_RS_DEFERRED(txq) txq_max_rs_deferred(txq)
2823 #define TXQ_MAX_DB_CONSUMED(size) (size >> 4)
2824 
2825 /* forward compatibility for cxgb */
2826 #define FIRST_QSET(ctx) 0
2827 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets)
2828 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets)
2829 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx))
2830 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments))
2831 
2832 /* XXX we should be setting this to something other than zero */
2833 #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh)
2834 #define MAX_TX_DESC(ctx) ((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max)
2835 
2836 static inline bool
2837 iflib_txd_db_check(if_ctx_t ctx, iflib_txq_t txq, int ring, qidx_t in_use)
2838 {
2839 	qidx_t dbval, max;
2840 	bool rang;
2841 
2842 	rang = false;
2843 	max = TXQ_MAX_DB_DEFERRED(txq, in_use);
2844 	if (ring || txq->ift_db_pending >= max) {
2845 		dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx;
2846 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval);
2847 		txq->ift_db_pending = txq->ift_npending = 0;
2848 		rang = true;
2849 	}
2850 	return (rang);
2851 }
2852 
2853 #ifdef PKT_DEBUG
2854 static void
2855 print_pkt(if_pkt_info_t pi)
2856 {
2857 	printf("pi len:  %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n",
2858 	       pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx);
2859 	printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n",
2860 	       pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag);
2861 	printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n",
2862 	       pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto);
2863 }
2864 #endif
2865 
2866 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO)
2867 #define IS_TX_OFFLOAD4(pi) ((pi)->ipi_csum_flags & (CSUM_IP_TCP | CSUM_IP_TSO))
2868 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO)
2869 #define IS_TX_OFFLOAD6(pi) ((pi)->ipi_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_TSO))
2870 
2871 static int
2872 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp)
2873 {
2874 	if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx;
2875 	struct ether_vlan_header *eh;
2876 	struct mbuf *m, *n;
2877 
2878 	m = *mp;
2879 	if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) &&
2880 	    M_WRITABLE(m) == 0) {
2881 		if ((m = m_dup(m, M_NOWAIT)) == NULL) {
2882 			return (ENOMEM);
2883 		} else {
2884 			m_freem(*mp);
2885 			*mp = m;
2886 		}
2887 	}
2888 
2889 	/*
2890 	 * Determine where frame payload starts.
2891 	 * Jump over vlan headers if already present,
2892 	 * helpful for QinQ too.
2893 	 */
2894 	if (__predict_false(m->m_len < sizeof(*eh))) {
2895 		txq->ift_pullups++;
2896 		if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL))
2897 			return (ENOMEM);
2898 	}
2899 	eh = mtod(m, struct ether_vlan_header *);
2900 	if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
2901 		pi->ipi_etype = ntohs(eh->evl_proto);
2902 		pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
2903 	} else {
2904 		pi->ipi_etype = ntohs(eh->evl_encap_proto);
2905 		pi->ipi_ehdrlen = ETHER_HDR_LEN;
2906 	}
2907 
2908 	switch (pi->ipi_etype) {
2909 #ifdef INET
2910 	case ETHERTYPE_IP:
2911 	{
2912 		struct ip *ip = NULL;
2913 		struct tcphdr *th = NULL;
2914 		int minthlen;
2915 
2916 		minthlen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip) + sizeof(*th));
2917 		if (__predict_false(m->m_len < minthlen)) {
2918 			/*
2919 			 * if this code bloat is causing too much of a hit
2920 			 * move it to a separate function and mark it noinline
2921 			 */
2922 			if (m->m_len == pi->ipi_ehdrlen) {
2923 				n = m->m_next;
2924 				MPASS(n);
2925 				if (n->m_len >= sizeof(*ip))  {
2926 					ip = (struct ip *)n->m_data;
2927 					if (n->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2928 						th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2929 				} else {
2930 					txq->ift_pullups++;
2931 					if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
2932 						return (ENOMEM);
2933 					ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2934 				}
2935 			} else {
2936 				txq->ift_pullups++;
2937 				if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
2938 					return (ENOMEM);
2939 				ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2940 				if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2941 					th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2942 			}
2943 		} else {
2944 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2945 			if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2946 				th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2947 		}
2948 		pi->ipi_ip_hlen = ip->ip_hl << 2;
2949 		pi->ipi_ipproto = ip->ip_p;
2950 		pi->ipi_flags |= IPI_TX_IPV4;
2951 
2952 		if ((sctx->isc_flags & IFLIB_NEED_ZERO_CSUM) && (pi->ipi_csum_flags & CSUM_IP))
2953                        ip->ip_sum = 0;
2954 
2955 		/* TCP checksum offload may require TCP header length */
2956 		if (IS_TX_OFFLOAD4(pi)) {
2957 			if (__predict_true(pi->ipi_ipproto == IPPROTO_TCP)) {
2958 				if (__predict_false(th == NULL)) {
2959 					txq->ift_pullups++;
2960 					if (__predict_false((m = m_pullup(m, (ip->ip_hl << 2) + sizeof(*th))) == NULL))
2961 						return (ENOMEM);
2962 					th = (struct tcphdr *)((caddr_t)ip + pi->ipi_ip_hlen);
2963 				}
2964 				pi->ipi_tcp_hflags = th->th_flags;
2965 				pi->ipi_tcp_hlen = th->th_off << 2;
2966 				pi->ipi_tcp_seq = th->th_seq;
2967 			}
2968 			if (IS_TSO4(pi)) {
2969 				if (__predict_false(ip->ip_p != IPPROTO_TCP))
2970 					return (ENXIO);
2971 				th->th_sum = in_pseudo(ip->ip_src.s_addr,
2972 						       ip->ip_dst.s_addr, htons(IPPROTO_TCP));
2973 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
2974 				if (sctx->isc_flags & IFLIB_TSO_INIT_IP) {
2975 					ip->ip_sum = 0;
2976 					ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz);
2977 				}
2978 			}
2979 		}
2980 		break;
2981 	}
2982 #endif
2983 #ifdef INET6
2984 	case ETHERTYPE_IPV6:
2985 	{
2986 		struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
2987 		struct tcphdr *th;
2988 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
2989 
2990 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
2991 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
2992 				return (ENOMEM);
2993 		}
2994 		th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
2995 
2996 		/* XXX-BZ this will go badly in case of ext hdrs. */
2997 		pi->ipi_ipproto = ip6->ip6_nxt;
2998 		pi->ipi_flags |= IPI_TX_IPV6;
2999 
3000 		/* TCP checksum offload may require TCP header length */
3001 		if (IS_TX_OFFLOAD6(pi)) {
3002 			if (pi->ipi_ipproto == IPPROTO_TCP) {
3003 				if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) {
3004 					txq->ift_pullups++;
3005 					if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL))
3006 						return (ENOMEM);
3007 				}
3008 				pi->ipi_tcp_hflags = th->th_flags;
3009 				pi->ipi_tcp_hlen = th->th_off << 2;
3010 				pi->ipi_tcp_seq = th->th_seq;
3011 			}
3012 			if (IS_TSO6(pi)) {
3013 				if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP))
3014 					return (ENXIO);
3015 				/*
3016 				 * The corresponding flag is set by the stack in the IPv4
3017 				 * TSO case, but not in IPv6 (at least in FreeBSD 10.2).
3018 				 * So, set it here because the rest of the flow requires it.
3019 				 */
3020 				pi->ipi_csum_flags |= CSUM_IP6_TCP;
3021 				th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0);
3022 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3023 			}
3024 		}
3025 		break;
3026 	}
3027 #endif
3028 	default:
3029 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
3030 		pi->ipi_ip_hlen = 0;
3031 		break;
3032 	}
3033 	*mp = m;
3034 
3035 	return (0);
3036 }
3037 
3038 static  __noinline  struct mbuf *
3039 collapse_pkthdr(struct mbuf *m0)
3040 {
3041 	struct mbuf *m, *m_next, *tmp;
3042 
3043 	m = m0;
3044 	m_next = m->m_next;
3045 	while (m_next != NULL && m_next->m_len == 0) {
3046 		m = m_next;
3047 		m->m_next = NULL;
3048 		m_free(m);
3049 		m_next = m_next->m_next;
3050 	}
3051 	m = m0;
3052 	m->m_next = m_next;
3053 	if (m_next == NULL)
3054 		return (m);
3055 	if ((m_next->m_flags & M_EXT) == 0) {
3056 		m = m_defrag(m, M_NOWAIT);
3057 	} else {
3058 		tmp = m_next->m_next;
3059 		memcpy(m_next, m, MPKTHSIZE);
3060 		m = m_next;
3061 		m->m_next = tmp;
3062 	}
3063 	return (m);
3064 }
3065 
3066 /*
3067  * If dodgy hardware rejects the scatter gather chain we've handed it
3068  * we'll need to remove the mbuf chain from ifsg_m[] before we can add the
3069  * m_defrag'd mbufs
3070  */
3071 static __noinline struct mbuf *
3072 iflib_remove_mbuf(iflib_txq_t txq)
3073 {
3074 	int ntxd, i, pidx;
3075 	struct mbuf *m, *mh, **ifsd_m;
3076 
3077 	pidx = txq->ift_pidx;
3078 	ifsd_m = txq->ift_sds.ifsd_m;
3079 	ntxd = txq->ift_size;
3080 	mh = m = ifsd_m[pidx];
3081 	ifsd_m[pidx] = NULL;
3082 #if MEMORY_LOGGING
3083 	txq->ift_dequeued++;
3084 #endif
3085 	i = 1;
3086 
3087 	while (m) {
3088 		ifsd_m[(pidx + i) & (ntxd -1)] = NULL;
3089 #if MEMORY_LOGGING
3090 		txq->ift_dequeued++;
3091 #endif
3092 		m = m->m_next;
3093 		i++;
3094 	}
3095 	return (mh);
3096 }
3097 
3098 static int
3099 iflib_busdma_load_mbuf_sg(iflib_txq_t txq, bus_dma_tag_t tag, bus_dmamap_t map,
3100 			  struct mbuf **m0, bus_dma_segment_t *segs, int *nsegs,
3101 			  int max_segs, int flags)
3102 {
3103 	if_ctx_t ctx;
3104 	if_shared_ctx_t		sctx;
3105 	if_softc_ctx_t		scctx;
3106 	int i, next, pidx, err, ntxd, count;
3107 	struct mbuf *m, *tmp, **ifsd_m;
3108 
3109 	m = *m0;
3110 
3111 	/*
3112 	 * Please don't ever do this
3113 	 */
3114 	if (__predict_false(m->m_len == 0))
3115 		*m0 = collapse_pkthdr(m);
3116 
3117 	ctx = txq->ift_ctx;
3118 	sctx = ctx->ifc_sctx;
3119 	scctx = &ctx->ifc_softc_ctx;
3120 	ifsd_m = txq->ift_sds.ifsd_m;
3121 	ntxd = txq->ift_size;
3122 	pidx = txq->ift_pidx;
3123 	if (map != NULL) {
3124 		uint8_t *ifsd_flags = txq->ift_sds.ifsd_flags;
3125 
3126 		err = bus_dmamap_load_mbuf_sg(tag, map,
3127 					      *m0, segs, nsegs, BUS_DMA_NOWAIT);
3128 		if (err)
3129 			return (err);
3130 		ifsd_flags[pidx] |= TX_SW_DESC_MAPPED;
3131 		count = 0;
3132 		m = *m0;
3133 		do {
3134 			if (__predict_false(m->m_len <= 0)) {
3135 				tmp = m;
3136 				m = m->m_next;
3137 				tmp->m_next = NULL;
3138 				m_free(tmp);
3139 				continue;
3140 			}
3141 			m = m->m_next;
3142 			count++;
3143 		} while (m != NULL);
3144 		if (count > *nsegs) {
3145 			ifsd_m[pidx] = *m0;
3146 			ifsd_m[pidx]->m_flags |= M_TOOBIG;
3147 			return (0);
3148 		}
3149 		m = *m0;
3150 		count = 0;
3151 		do {
3152 			next = (pidx + count) & (ntxd-1);
3153 			MPASS(ifsd_m[next] == NULL);
3154 			ifsd_m[next] = m;
3155 			count++;
3156 			tmp = m;
3157 			m = m->m_next;
3158 		} while (m != NULL);
3159 	} else {
3160 		int buflen, sgsize, maxsegsz, max_sgsize;
3161 		vm_offset_t vaddr;
3162 		vm_paddr_t curaddr;
3163 
3164 		count = i = 0;
3165 		m = *m0;
3166 		if (m->m_pkthdr.csum_flags & CSUM_TSO)
3167 			maxsegsz = scctx->isc_tx_tso_segsize_max;
3168 		else
3169 			maxsegsz = sctx->isc_tx_maxsegsize;
3170 
3171 		do {
3172 			if (__predict_false(m->m_len <= 0)) {
3173 				tmp = m;
3174 				m = m->m_next;
3175 				tmp->m_next = NULL;
3176 				m_free(tmp);
3177 				continue;
3178 			}
3179 			buflen = m->m_len;
3180 			vaddr = (vm_offset_t)m->m_data;
3181 			/*
3182 			 * see if we can't be smarter about physically
3183 			 * contiguous mappings
3184 			 */
3185 			next = (pidx + count) & (ntxd-1);
3186 			MPASS(ifsd_m[next] == NULL);
3187 #if MEMORY_LOGGING
3188 			txq->ift_enqueued++;
3189 #endif
3190 			ifsd_m[next] = m;
3191 			while (buflen > 0) {
3192 				if (i >= max_segs)
3193 					goto err;
3194 				max_sgsize = MIN(buflen, maxsegsz);
3195 				curaddr = pmap_kextract(vaddr);
3196 				sgsize = PAGE_SIZE - (curaddr & PAGE_MASK);
3197 				sgsize = MIN(sgsize, max_sgsize);
3198 				segs[i].ds_addr = curaddr;
3199 				segs[i].ds_len = sgsize;
3200 				vaddr += sgsize;
3201 				buflen -= sgsize;
3202 				i++;
3203 			}
3204 			count++;
3205 			tmp = m;
3206 			m = m->m_next;
3207 		} while (m != NULL);
3208 		*nsegs = i;
3209 	}
3210 	return (0);
3211 err:
3212 	*m0 = iflib_remove_mbuf(txq);
3213 	return (EFBIG);
3214 }
3215 
3216 static inline caddr_t
3217 calc_next_txd(iflib_txq_t txq, int cidx, uint8_t qid)
3218 {
3219 	qidx_t size;
3220 	int ntxd;
3221 	caddr_t start, end, cur, next;
3222 
3223 	ntxd = txq->ift_size;
3224 	size = txq->ift_txd_size[qid];
3225 	start = txq->ift_ifdi[qid].idi_vaddr;
3226 
3227 	if (__predict_false(size == 0))
3228 		return (start);
3229 	cur = start + size*cidx;
3230 	end = start + size*ntxd;
3231 	next = CACHE_PTR_NEXT(cur);
3232 	return (next < end ? next : start);
3233 }
3234 
3235 /*
3236  * Pad an mbuf to ensure a minimum ethernet frame size.
3237  * min_frame_size is the frame size (less CRC) to pad the mbuf to
3238  */
3239 static __noinline int
3240 iflib_ether_pad(device_t dev, struct mbuf **m_head, uint16_t min_frame_size)
3241 {
3242 	/*
3243 	 * 18 is enough bytes to pad an ARP packet to 46 bytes, and
3244 	 * and ARP message is the smallest common payload I can think of
3245 	 */
3246 	static char pad[18];	/* just zeros */
3247 	int n;
3248 	struct mbuf *new_head;
3249 
3250 	if (!M_WRITABLE(*m_head)) {
3251 		new_head = m_dup(*m_head, M_NOWAIT);
3252 		if (new_head == NULL) {
3253 			m_freem(*m_head);
3254 			device_printf(dev, "cannot pad short frame, m_dup() failed");
3255 			DBG_COUNTER_INC(encap_pad_mbuf_fail);
3256 			return ENOMEM;
3257 		}
3258 		m_freem(*m_head);
3259 		*m_head = new_head;
3260 	}
3261 
3262 	for (n = min_frame_size - (*m_head)->m_pkthdr.len;
3263 	     n > 0; n -= sizeof(pad))
3264 		if (!m_append(*m_head, min(n, sizeof(pad)), pad))
3265 			break;
3266 
3267 	if (n > 0) {
3268 		m_freem(*m_head);
3269 		device_printf(dev, "cannot pad short frame\n");
3270 		DBG_COUNTER_INC(encap_pad_mbuf_fail);
3271 		return (ENOBUFS);
3272 	}
3273 
3274 	return 0;
3275 }
3276 
3277 static int
3278 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp)
3279 {
3280 	if_ctx_t		ctx;
3281 	if_shared_ctx_t		sctx;
3282 	if_softc_ctx_t		scctx;
3283 	bus_dma_segment_t	*segs;
3284 	struct mbuf		*m_head;
3285 	void			*next_txd;
3286 	bus_dmamap_t		map;
3287 	struct if_pkt_info	pi;
3288 	int remap = 0;
3289 	int err, nsegs, ndesc, max_segs, pidx, cidx, next, ntxd;
3290 	bus_dma_tag_t desc_tag;
3291 
3292 	ctx = txq->ift_ctx;
3293 	sctx = ctx->ifc_sctx;
3294 	scctx = &ctx->ifc_softc_ctx;
3295 	segs = txq->ift_segs;
3296 	ntxd = txq->ift_size;
3297 	m_head = *m_headp;
3298 	map = NULL;
3299 
3300 	/*
3301 	 * If we're doing TSO the next descriptor to clean may be quite far ahead
3302 	 */
3303 	cidx = txq->ift_cidx;
3304 	pidx = txq->ift_pidx;
3305 	if (ctx->ifc_flags & IFC_PREFETCH) {
3306 		next = (cidx + CACHE_PTR_INCREMENT) & (ntxd-1);
3307 		if (!(ctx->ifc_flags & IFLIB_HAS_TXCQ)) {
3308 			next_txd = calc_next_txd(txq, cidx, 0);
3309 			prefetch(next_txd);
3310 		}
3311 
3312 		/* prefetch the next cache line of mbuf pointers and flags */
3313 		prefetch(&txq->ift_sds.ifsd_m[next]);
3314 		if (txq->ift_sds.ifsd_map != NULL) {
3315 			prefetch(&txq->ift_sds.ifsd_map[next]);
3316 			next = (cidx + CACHE_LINE_SIZE) & (ntxd-1);
3317 			prefetch(&txq->ift_sds.ifsd_flags[next]);
3318 		}
3319 	} else if (txq->ift_sds.ifsd_map != NULL)
3320 		map = txq->ift_sds.ifsd_map[pidx];
3321 
3322 	if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
3323 		desc_tag = txq->ift_tso_desc_tag;
3324 		max_segs = scctx->isc_tx_tso_segments_max;
3325 		MPASS(desc_tag != NULL);
3326 		MPASS(max_segs > 0);
3327 	} else {
3328 		desc_tag = txq->ift_desc_tag;
3329 		max_segs = scctx->isc_tx_nsegments;
3330 	}
3331 	if ((sctx->isc_flags & IFLIB_NEED_ETHER_PAD) &&
3332 	    __predict_false(m_head->m_pkthdr.len < scctx->isc_min_frame_size)) {
3333 		err = iflib_ether_pad(ctx->ifc_dev, m_headp, scctx->isc_min_frame_size);
3334 		if (err)
3335 			return err;
3336 	}
3337 	m_head = *m_headp;
3338 
3339 	pkt_info_zero(&pi);
3340 	pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG|M_BCAST|M_MCAST));
3341 	pi.ipi_pidx = pidx;
3342 	pi.ipi_qsidx = txq->ift_id;
3343 	pi.ipi_len = m_head->m_pkthdr.len;
3344 	pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags;
3345 	pi.ipi_vtag = (m_head->m_flags & M_VLANTAG) ? m_head->m_pkthdr.ether_vtag : 0;
3346 
3347 	/* deliberate bitwise OR to make one condition */
3348 	if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) {
3349 		if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0))
3350 			return (err);
3351 		m_head = *m_headp;
3352 	}
3353 
3354 retry:
3355 	err = iflib_busdma_load_mbuf_sg(txq, desc_tag, map, m_headp, segs, &nsegs, max_segs, BUS_DMA_NOWAIT);
3356 defrag:
3357 	if (__predict_false(err)) {
3358 		switch (err) {
3359 		case EFBIG:
3360 			/* try collapse once and defrag once */
3361 			if (remap == 0) {
3362 				m_head = m_collapse(*m_headp, M_NOWAIT, max_segs);
3363 				/* try defrag if collapsing fails */
3364 				if (m_head == NULL)
3365 					remap++;
3366 			}
3367 			if (remap == 1)
3368 				m_head = m_defrag(*m_headp, M_NOWAIT);
3369 			remap++;
3370 			if (__predict_false(m_head == NULL))
3371 				goto defrag_failed;
3372 			txq->ift_mbuf_defrag++;
3373 			*m_headp = m_head;
3374 			goto retry;
3375 			break;
3376 		case ENOMEM:
3377 			txq->ift_no_tx_dma_setup++;
3378 			break;
3379 		default:
3380 			txq->ift_no_tx_dma_setup++;
3381 			m_freem(*m_headp);
3382 			DBG_COUNTER_INC(tx_frees);
3383 			*m_headp = NULL;
3384 			break;
3385 		}
3386 		txq->ift_map_failed++;
3387 		DBG_COUNTER_INC(encap_load_mbuf_fail);
3388 		return (err);
3389 	}
3390 
3391 	/*
3392 	 * XXX assumes a 1 to 1 relationship between segments and
3393 	 *        descriptors - this does not hold true on all drivers, e.g.
3394 	 *        cxgb
3395 	 */
3396 	if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) {
3397 		txq->ift_no_desc_avail++;
3398 		if (map != NULL)
3399 			bus_dmamap_unload(desc_tag, map);
3400 		DBG_COUNTER_INC(encap_txq_avail_fail);
3401 		if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0)
3402 			GROUPTASK_ENQUEUE(&txq->ift_task);
3403 		return (ENOBUFS);
3404 	}
3405 	/*
3406 	 * On Intel cards we can greatly reduce the number of TX interrupts
3407 	 * we see by only setting report status on every Nth descriptor.
3408 	 * However, this also means that the driver will need to keep track
3409 	 * of the descriptors that RS was set on to check them for the DD bit.
3410 	 */
3411 	txq->ift_rs_pending += nsegs + 1;
3412 	if (txq->ift_rs_pending > TXQ_MAX_RS_DEFERRED(txq) ||
3413 	     iflib_no_tx_batch || (TXQ_AVAIL(txq) - nsegs) <= MAX_TX_DESC(ctx) + 2) {
3414 		pi.ipi_flags |= IPI_TX_INTR;
3415 		txq->ift_rs_pending = 0;
3416 	}
3417 
3418 	pi.ipi_segs = segs;
3419 	pi.ipi_nsegs = nsegs;
3420 
3421 	MPASS(pidx >= 0 && pidx < txq->ift_size);
3422 #ifdef PKT_DEBUG
3423 	print_pkt(&pi);
3424 #endif
3425 	if (map != NULL)
3426 		bus_dmamap_sync(desc_tag, map, BUS_DMASYNC_PREWRITE);
3427 	if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) {
3428 		if (map != NULL)
3429 			bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3430 					BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
3431 		DBG_COUNTER_INC(tx_encap);
3432 		MPASS(pi.ipi_new_pidx < txq->ift_size);
3433 
3434 		ndesc = pi.ipi_new_pidx - pi.ipi_pidx;
3435 		if (pi.ipi_new_pidx < pi.ipi_pidx) {
3436 			ndesc += txq->ift_size;
3437 			txq->ift_gen = 1;
3438 		}
3439 		/*
3440 		 * drivers can need as many as
3441 		 * two sentinels
3442 		 */
3443 		MPASS(ndesc <= pi.ipi_nsegs + 2);
3444 		MPASS(pi.ipi_new_pidx != pidx);
3445 		MPASS(ndesc > 0);
3446 		txq->ift_in_use += ndesc;
3447 
3448 		/*
3449 		 * We update the last software descriptor again here because there may
3450 		 * be a sentinel and/or there may be more mbufs than segments
3451 		 */
3452 		txq->ift_pidx = pi.ipi_new_pidx;
3453 		txq->ift_npending += pi.ipi_ndescs;
3454 	} else {
3455 		*m_headp = m_head = iflib_remove_mbuf(txq);
3456 		if (err == EFBIG) {
3457 			txq->ift_txd_encap_efbig++;
3458 			if (remap < 2) {
3459 				remap = 1;
3460 				goto defrag;
3461 			}
3462 		}
3463 		DBG_COUNTER_INC(encap_txd_encap_fail);
3464 		goto defrag_failed;
3465 	}
3466 	return (err);
3467 
3468 defrag_failed:
3469 	txq->ift_mbuf_defrag_failed++;
3470 	txq->ift_map_failed++;
3471 	m_freem(*m_headp);
3472 	DBG_COUNTER_INC(tx_frees);
3473 	*m_headp = NULL;
3474 	return (ENOMEM);
3475 }
3476 
3477 static void
3478 iflib_tx_desc_free(iflib_txq_t txq, int n)
3479 {
3480 	int hasmap;
3481 	uint32_t qsize, cidx, mask, gen;
3482 	struct mbuf *m, **ifsd_m;
3483 	uint8_t *ifsd_flags;
3484 	bus_dmamap_t *ifsd_map;
3485 	bool do_prefetch;
3486 
3487 	cidx = txq->ift_cidx;
3488 	gen = txq->ift_gen;
3489 	qsize = txq->ift_size;
3490 	mask = qsize-1;
3491 	hasmap = txq->ift_sds.ifsd_map != NULL;
3492 	ifsd_flags = txq->ift_sds.ifsd_flags;
3493 	ifsd_m = txq->ift_sds.ifsd_m;
3494 	ifsd_map = txq->ift_sds.ifsd_map;
3495 	do_prefetch = (txq->ift_ctx->ifc_flags & IFC_PREFETCH);
3496 
3497 	while (n-- > 0) {
3498 		if (do_prefetch) {
3499 			prefetch(ifsd_m[(cidx + 3) & mask]);
3500 			prefetch(ifsd_m[(cidx + 4) & mask]);
3501 		}
3502 		if (ifsd_m[cidx] != NULL) {
3503 			prefetch(&ifsd_m[(cidx + CACHE_PTR_INCREMENT) & mask]);
3504 			prefetch(&ifsd_flags[(cidx + CACHE_PTR_INCREMENT) & mask]);
3505 			if (hasmap && (ifsd_flags[cidx] & TX_SW_DESC_MAPPED)) {
3506 				/*
3507 				 * does it matter if it's not the TSO tag? If so we'll
3508 				 * have to add the type to flags
3509 				 */
3510 				bus_dmamap_unload(txq->ift_desc_tag, ifsd_map[cidx]);
3511 				ifsd_flags[cidx] &= ~TX_SW_DESC_MAPPED;
3512 			}
3513 			if ((m = ifsd_m[cidx]) != NULL) {
3514 				/* XXX we don't support any drivers that batch packets yet */
3515 				MPASS(m->m_nextpkt == NULL);
3516 				/* if the number of clusters exceeds the number of segments
3517 				 * there won't be space on the ring to save a pointer to each
3518 				 * cluster so we simply free the list here
3519 				 */
3520 				if (m->m_flags & M_TOOBIG) {
3521 					m_freem(m);
3522 				} else {
3523 					m_free(m);
3524 				}
3525 				ifsd_m[cidx] = NULL;
3526 #if MEMORY_LOGGING
3527 				txq->ift_dequeued++;
3528 #endif
3529 				DBG_COUNTER_INC(tx_frees);
3530 			}
3531 		}
3532 		if (__predict_false(++cidx == qsize)) {
3533 			cidx = 0;
3534 			gen = 0;
3535 		}
3536 	}
3537 	txq->ift_cidx = cidx;
3538 	txq->ift_gen = gen;
3539 }
3540 
3541 static __inline int
3542 iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh)
3543 {
3544 	int reclaim;
3545 	if_ctx_t ctx = txq->ift_ctx;
3546 
3547 	KASSERT(thresh >= 0, ("invalid threshold to reclaim"));
3548 	MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size);
3549 
3550 	/*
3551 	 * Need a rate-limiting check so that this isn't called every time
3552 	 */
3553 	iflib_tx_credits_update(ctx, txq);
3554 	reclaim = DESC_RECLAIMABLE(txq);
3555 
3556 	if (reclaim <= thresh /* + MAX_TX_DESC(txq->ift_ctx) */) {
3557 #ifdef INVARIANTS
3558 		if (iflib_verbose_debug) {
3559 			printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __FUNCTION__,
3560 			       txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments,
3561 			       reclaim, thresh);
3562 
3563 		}
3564 #endif
3565 		return (0);
3566 	}
3567 	iflib_tx_desc_free(txq, reclaim);
3568 	txq->ift_cleaned += reclaim;
3569 	txq->ift_in_use -= reclaim;
3570 
3571 	return (reclaim);
3572 }
3573 
3574 static struct mbuf **
3575 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset, int remaining)
3576 {
3577 	int next, size;
3578 	struct mbuf **items;
3579 
3580 	size = r->size;
3581 	next = (cidx + CACHE_PTR_INCREMENT) & (size-1);
3582 	items = __DEVOLATILE(struct mbuf **, &r->items[0]);
3583 
3584 	prefetch(items[(cidx + offset) & (size-1)]);
3585 	if (remaining > 1) {
3586 		prefetch2cachelines(&items[next]);
3587 		prefetch2cachelines(items[(cidx + offset + 1) & (size-1)]);
3588 		prefetch2cachelines(items[(cidx + offset + 2) & (size-1)]);
3589 		prefetch2cachelines(items[(cidx + offset + 3) & (size-1)]);
3590 	}
3591 	return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (size-1)]));
3592 }
3593 
3594 static void
3595 iflib_txq_check_drain(iflib_txq_t txq, int budget)
3596 {
3597 
3598 	ifmp_ring_check_drainage(txq->ift_br, budget);
3599 }
3600 
3601 static uint32_t
3602 iflib_txq_can_drain(struct ifmp_ring *r)
3603 {
3604 	iflib_txq_t txq = r->cookie;
3605 	if_ctx_t ctx = txq->ift_ctx;
3606 
3607 	return ((TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2) ||
3608 		ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false));
3609 }
3610 
3611 static uint32_t
3612 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3613 {
3614 	iflib_txq_t txq = r->cookie;
3615 	if_ctx_t ctx = txq->ift_ctx;
3616 	struct ifnet *ifp = ctx->ifc_ifp;
3617 	struct mbuf **mp, *m;
3618 	int i, count, consumed, pkt_sent, bytes_sent, mcast_sent, avail;
3619 	int reclaimed, err, in_use_prev, desc_used;
3620 	bool do_prefetch, ring, rang;
3621 
3622 	if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) ||
3623 			    !LINK_ACTIVE(ctx))) {
3624 		DBG_COUNTER_INC(txq_drain_notready);
3625 		return (0);
3626 	}
3627 	reclaimed = iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
3628 	rang = iflib_txd_db_check(ctx, txq, reclaimed, txq->ift_in_use);
3629 	avail = IDXDIFF(pidx, cidx, r->size);
3630 	if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) {
3631 		DBG_COUNTER_INC(txq_drain_flushing);
3632 		for (i = 0; i < avail; i++) {
3633 			m_free(r->items[(cidx + i) & (r->size-1)]);
3634 			r->items[(cidx + i) & (r->size-1)] = NULL;
3635 		}
3636 		return (avail);
3637 	}
3638 
3639 	if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3640 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3641 		CALLOUT_LOCK(txq);
3642 		callout_stop(&txq->ift_timer);
3643 		CALLOUT_UNLOCK(txq);
3644 		DBG_COUNTER_INC(txq_drain_oactive);
3645 		return (0);
3646 	}
3647 	if (reclaimed)
3648 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3649 	consumed = mcast_sent = bytes_sent = pkt_sent = 0;
3650 	count = MIN(avail, TX_BATCH_SIZE);
3651 #ifdef INVARIANTS
3652 	if (iflib_verbose_debug)
3653 		printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __FUNCTION__,
3654 		       avail, ctx->ifc_flags, TXQ_AVAIL(txq));
3655 #endif
3656 	do_prefetch = (ctx->ifc_flags & IFC_PREFETCH);
3657 	avail = TXQ_AVAIL(txq);
3658 	err = 0;
3659 	for (desc_used = i = 0; i < count && avail > MAX_TX_DESC(ctx) + 2; i++) {
3660 		int rem = do_prefetch ? count - i : 0;
3661 
3662 		mp = _ring_peek_one(r, cidx, i, rem);
3663 		MPASS(mp != NULL && *mp != NULL);
3664 		if (__predict_false(*mp == (struct mbuf *)txq)) {
3665 			consumed++;
3666 			reclaimed++;
3667 			continue;
3668 		}
3669 		in_use_prev = txq->ift_in_use;
3670 		err = iflib_encap(txq, mp);
3671 		if (__predict_false(err)) {
3672 			DBG_COUNTER_INC(txq_drain_encapfail);
3673 			/* no room - bail out */
3674 			if (err == ENOBUFS)
3675 				break;
3676 			consumed++;
3677 			DBG_COUNTER_INC(txq_drain_encapfail);
3678 			/* we can't send this packet - skip it */
3679 			continue;
3680 		}
3681 		consumed++;
3682 		pkt_sent++;
3683 		m = *mp;
3684 		DBG_COUNTER_INC(tx_sent);
3685 		bytes_sent += m->m_pkthdr.len;
3686 		mcast_sent += !!(m->m_flags & M_MCAST);
3687 		avail = TXQ_AVAIL(txq);
3688 
3689 		txq->ift_db_pending += (txq->ift_in_use - in_use_prev);
3690 		desc_used += (txq->ift_in_use - in_use_prev);
3691 		ETHER_BPF_MTAP(ifp, m);
3692 		if (__predict_false(!(ifp->if_drv_flags & IFF_DRV_RUNNING)))
3693 			break;
3694 		rang = iflib_txd_db_check(ctx, txq, false, in_use_prev);
3695 	}
3696 
3697 	/* deliberate use of bitwise or to avoid gratuitous short-circuit */
3698 	ring = rang ? false  : (iflib_min_tx_latency | err) || (TXQ_AVAIL(txq) < MAX_TX_DESC(ctx));
3699 	iflib_txd_db_check(ctx, txq, ring, txq->ift_in_use);
3700 	if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
3701 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
3702 	if (mcast_sent)
3703 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
3704 #ifdef INVARIANTS
3705 	if (iflib_verbose_debug)
3706 		printf("consumed=%d\n", consumed);
3707 #endif
3708 	return (consumed);
3709 }
3710 
3711 static uint32_t
3712 iflib_txq_drain_always(struct ifmp_ring *r)
3713 {
3714 	return (1);
3715 }
3716 
3717 static uint32_t
3718 iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3719 {
3720 	int i, avail;
3721 	struct mbuf **mp;
3722 	iflib_txq_t txq;
3723 
3724 	txq = r->cookie;
3725 
3726 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3727 	CALLOUT_LOCK(txq);
3728 	callout_stop(&txq->ift_timer);
3729 	CALLOUT_UNLOCK(txq);
3730 
3731 	avail = IDXDIFF(pidx, cidx, r->size);
3732 	for (i = 0; i < avail; i++) {
3733 		mp = _ring_peek_one(r, cidx, i, avail - i);
3734 		if (__predict_false(*mp == (struct mbuf *)txq))
3735 			continue;
3736 		m_freem(*mp);
3737 	}
3738 	MPASS(ifmp_ring_is_stalled(r) == 0);
3739 	return (avail);
3740 }
3741 
3742 static void
3743 iflib_ifmp_purge(iflib_txq_t txq)
3744 {
3745 	struct ifmp_ring *r;
3746 
3747 	r = txq->ift_br;
3748 	r->drain = iflib_txq_drain_free;
3749 	r->can_drain = iflib_txq_drain_always;
3750 
3751 	ifmp_ring_check_drainage(r, r->size);
3752 
3753 	r->drain = iflib_txq_drain;
3754 	r->can_drain = iflib_txq_can_drain;
3755 }
3756 
3757 static void
3758 _task_fn_tx(void *context)
3759 {
3760 	iflib_txq_t txq = context;
3761 	if_ctx_t ctx = txq->ift_ctx;
3762 	struct ifnet *ifp = ctx->ifc_ifp;
3763 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
3764 
3765 #ifdef IFLIB_DIAGNOSTICS
3766 	txq->ift_cpu_exec_count[curcpu]++;
3767 #endif
3768 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3769 		return;
3770 	if (if_getcapenable(ifp) & IFCAP_NETMAP) {
3771 		if (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false))
3772 			netmap_tx_irq(ifp, txq->ift_id);
3773 		IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3774 		return;
3775 	}
3776 	if (txq->ift_db_pending)
3777 		ifmp_ring_enqueue(txq->ift_br, (void **)&txq, 1, TX_BATCH_SIZE, abdicate);
3778 	else if (!abdicate)
3779 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3780 	/*
3781 	 * When abdicating, we always need to check drainage, not just when we don't enqueue
3782 	 */
3783 	if (abdicate)
3784 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3785 	ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3786 	if (ctx->ifc_flags & IFC_LEGACY)
3787 		IFDI_INTR_ENABLE(ctx);
3788 	else {
3789 #ifdef INVARIANTS
3790 		int rc =
3791 #endif
3792 			IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3793 			KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver"));
3794 	}
3795 }
3796 
3797 static void
3798 _task_fn_rx(void *context)
3799 {
3800 	iflib_rxq_t rxq = context;
3801 	if_ctx_t ctx = rxq->ifr_ctx;
3802 	bool more;
3803 	uint16_t budget;
3804 
3805 #ifdef IFLIB_DIAGNOSTICS
3806 	rxq->ifr_cpu_exec_count[curcpu]++;
3807 #endif
3808 	DBG_COUNTER_INC(task_fn_rxs);
3809 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3810 		return;
3811 	more = true;
3812 #ifdef DEV_NETMAP
3813 	if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP) {
3814 		u_int work = 0;
3815 		if (netmap_rx_irq(ctx->ifc_ifp, rxq->ifr_id, &work)) {
3816 			more = false;
3817 		}
3818 	}
3819 #endif
3820 	budget = ctx->ifc_sysctl_rx_budget;
3821 	if (budget == 0)
3822 		budget = 16;	/* XXX */
3823 	if (more == false || (more = iflib_rxeof(rxq, budget)) == false) {
3824 		if (ctx->ifc_flags & IFC_LEGACY)
3825 			IFDI_INTR_ENABLE(ctx);
3826 		else {
3827 #ifdef INVARIANTS
3828 			int rc =
3829 #endif
3830 				IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
3831 			KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver"));
3832 			DBG_COUNTER_INC(rx_intr_enables);
3833 		}
3834 	}
3835 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3836 		return;
3837 	if (more)
3838 		GROUPTASK_ENQUEUE(&rxq->ifr_task);
3839 }
3840 
3841 static void
3842 _task_fn_admin(void *context)
3843 {
3844 	if_ctx_t ctx = context;
3845 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
3846 	iflib_txq_t txq;
3847 	int i;
3848 	bool oactive, running, do_reset, do_watchdog;
3849 	uint32_t reset_on = hz / 2;
3850 
3851 	STATE_LOCK(ctx);
3852 	running = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING);
3853 	oactive = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE);
3854 	do_reset = (ctx->ifc_flags & IFC_DO_RESET);
3855 	do_watchdog = (ctx->ifc_flags & IFC_DO_WATCHDOG);
3856 	ctx->ifc_flags &= ~(IFC_DO_RESET|IFC_DO_WATCHDOG);
3857 	STATE_UNLOCK(ctx);
3858 
3859 	if ((!running & !oactive) &&
3860 	    !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
3861 		return;
3862 
3863 	CTX_LOCK(ctx);
3864 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3865 		CALLOUT_LOCK(txq);
3866 		callout_stop(&txq->ift_timer);
3867 		CALLOUT_UNLOCK(txq);
3868 	}
3869 	if (do_watchdog) {
3870 		ctx->ifc_watchdog_events++;
3871 		IFDI_WATCHDOG_RESET(ctx);
3872 	}
3873 	IFDI_UPDATE_ADMIN_STATUS(ctx);
3874 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3875 #ifdef DEV_NETMAP
3876 		reset_on = hz / 2;
3877 		if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP)
3878 			iflib_netmap_timer_adjust(ctx, txq->ift_id, &reset_on);
3879 #endif
3880 		callout_reset_on(&txq->ift_timer, reset_on, iflib_timer, txq, txq->ift_timer.c_cpu);
3881 	}
3882 	IFDI_LINK_INTR_ENABLE(ctx);
3883 	if (do_reset)
3884 		iflib_if_init_locked(ctx);
3885 	CTX_UNLOCK(ctx);
3886 
3887 	if (LINK_ACTIVE(ctx) == 0)
3888 		return;
3889 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3890 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
3891 }
3892 
3893 
3894 static void
3895 _task_fn_iov(void *context)
3896 {
3897 	if_ctx_t ctx = context;
3898 
3899 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3900 		return;
3901 
3902 	CTX_LOCK(ctx);
3903 	IFDI_VFLR_HANDLE(ctx);
3904 	CTX_UNLOCK(ctx);
3905 }
3906 
3907 static int
3908 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)
3909 {
3910 	int err;
3911 	if_int_delay_info_t info;
3912 	if_ctx_t ctx;
3913 
3914 	info = (if_int_delay_info_t)arg1;
3915 	ctx = info->iidi_ctx;
3916 	info->iidi_req = req;
3917 	info->iidi_oidp = oidp;
3918 	CTX_LOCK(ctx);
3919 	err = IFDI_SYSCTL_INT_DELAY(ctx, info);
3920 	CTX_UNLOCK(ctx);
3921 	return (err);
3922 }
3923 
3924 /*********************************************************************
3925  *
3926  *  IFNET FUNCTIONS
3927  *
3928  **********************************************************************/
3929 
3930 static void
3931 iflib_if_init_locked(if_ctx_t ctx)
3932 {
3933 	iflib_stop(ctx);
3934 	iflib_init_locked(ctx);
3935 }
3936 
3937 
3938 static void
3939 iflib_if_init(void *arg)
3940 {
3941 	if_ctx_t ctx = arg;
3942 
3943 	CTX_LOCK(ctx);
3944 	iflib_if_init_locked(ctx);
3945 	CTX_UNLOCK(ctx);
3946 }
3947 
3948 static int
3949 iflib_if_transmit(if_t ifp, struct mbuf *m)
3950 {
3951 	if_ctx_t	ctx = if_getsoftc(ifp);
3952 
3953 	iflib_txq_t txq;
3954 	int err, qidx;
3955 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
3956 
3957 	if (__predict_false((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) {
3958 		DBG_COUNTER_INC(tx_frees);
3959 		m_freem(m);
3960 		return (ENOBUFS);
3961 	}
3962 
3963 	MPASS(m->m_nextpkt == NULL);
3964 	qidx = 0;
3965 	if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m))
3966 		qidx = QIDX(ctx, m);
3967 	/*
3968 	 * XXX calculate buf_ring based on flowid (divvy up bits?)
3969 	 */
3970 	txq = &ctx->ifc_txqs[qidx];
3971 
3972 #ifdef DRIVER_BACKPRESSURE
3973 	if (txq->ift_closed) {
3974 		while (m != NULL) {
3975 			next = m->m_nextpkt;
3976 			m->m_nextpkt = NULL;
3977 			m_freem(m);
3978 			m = next;
3979 		}
3980 		return (ENOBUFS);
3981 	}
3982 #endif
3983 #ifdef notyet
3984 	qidx = count = 0;
3985 	mp = marr;
3986 	next = m;
3987 	do {
3988 		count++;
3989 		next = next->m_nextpkt;
3990 	} while (next != NULL);
3991 
3992 	if (count > nitems(marr))
3993 		if ((mp = malloc(count*sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) {
3994 			/* XXX check nextpkt */
3995 			m_freem(m);
3996 			/* XXX simplify for now */
3997 			DBG_COUNTER_INC(tx_frees);
3998 			return (ENOBUFS);
3999 		}
4000 	for (next = m, i = 0; next != NULL; i++) {
4001 		mp[i] = next;
4002 		next = next->m_nextpkt;
4003 		mp[i]->m_nextpkt = NULL;
4004 	}
4005 #endif
4006 	DBG_COUNTER_INC(tx_seen);
4007 	err = ifmp_ring_enqueue(txq->ift_br, (void **)&m, 1, TX_BATCH_SIZE, abdicate);
4008 
4009 	if (abdicate)
4010 		GROUPTASK_ENQUEUE(&txq->ift_task);
4011  	if (err) {
4012 		if (!abdicate)
4013 			GROUPTASK_ENQUEUE(&txq->ift_task);
4014 		/* support forthcoming later */
4015 #ifdef DRIVER_BACKPRESSURE
4016 		txq->ift_closed = TRUE;
4017 #endif
4018 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
4019 		m_freem(m);
4020 	}
4021 
4022 	return (err);
4023 }
4024 
4025 static void
4026 iflib_if_qflush(if_t ifp)
4027 {
4028 	if_ctx_t ctx = if_getsoftc(ifp);
4029 	iflib_txq_t txq = ctx->ifc_txqs;
4030 	int i;
4031 
4032 	STATE_LOCK(ctx);
4033 	ctx->ifc_flags |= IFC_QFLUSH;
4034 	STATE_UNLOCK(ctx);
4035 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
4036 		while (!(ifmp_ring_is_idle(txq->ift_br) || ifmp_ring_is_stalled(txq->ift_br)))
4037 			iflib_txq_check_drain(txq, 0);
4038 	STATE_LOCK(ctx);
4039 	ctx->ifc_flags &= ~IFC_QFLUSH;
4040 	STATE_UNLOCK(ctx);
4041 
4042 	if_qflush(ifp);
4043 }
4044 
4045 
4046 #define IFCAP_FLAGS (IFCAP_TXCSUM_IPV6 | IFCAP_RXCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \
4047 		     IFCAP_TSO4 | IFCAP_TSO6 | IFCAP_VLAN_HWTAGGING | IFCAP_HWSTATS | \
4048 		     IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | IFCAP_VLAN_HWTSO)
4049 
4050 static int
4051 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data)
4052 {
4053 	if_ctx_t ctx = if_getsoftc(ifp);
4054 	struct ifreq	*ifr = (struct ifreq *)data;
4055 #if defined(INET) || defined(INET6)
4056 	struct ifaddr	*ifa = (struct ifaddr *)data;
4057 #endif
4058 	bool		avoid_reset = FALSE;
4059 	int		err = 0, reinit = 0, bits;
4060 
4061 	switch (command) {
4062 	case SIOCSIFADDR:
4063 #ifdef INET
4064 		if (ifa->ifa_addr->sa_family == AF_INET)
4065 			avoid_reset = TRUE;
4066 #endif
4067 #ifdef INET6
4068 		if (ifa->ifa_addr->sa_family == AF_INET6)
4069 			avoid_reset = TRUE;
4070 #endif
4071 		/*
4072 		** Calling init results in link renegotiation,
4073 		** so we avoid doing it when possible.
4074 		*/
4075 		if (avoid_reset) {
4076 			if_setflagbits(ifp, IFF_UP,0);
4077 			if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
4078 				reinit = 1;
4079 #ifdef INET
4080 			if (!(if_getflags(ifp) & IFF_NOARP))
4081 				arp_ifinit(ifp, ifa);
4082 #endif
4083 		} else
4084 			err = ether_ioctl(ifp, command, data);
4085 		break;
4086 	case SIOCSIFMTU:
4087 		CTX_LOCK(ctx);
4088 		if (ifr->ifr_mtu == if_getmtu(ifp)) {
4089 			CTX_UNLOCK(ctx);
4090 			break;
4091 		}
4092 		bits = if_getdrvflags(ifp);
4093 		/* stop the driver and free any clusters before proceeding */
4094 		iflib_stop(ctx);
4095 
4096 		if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) {
4097 			STATE_LOCK(ctx);
4098 			if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size)
4099 				ctx->ifc_flags |= IFC_MULTISEG;
4100 			else
4101 				ctx->ifc_flags &= ~IFC_MULTISEG;
4102 			STATE_UNLOCK(ctx);
4103 			err = if_setmtu(ifp, ifr->ifr_mtu);
4104 		}
4105 		iflib_init_locked(ctx);
4106 		STATE_LOCK(ctx);
4107 		if_setdrvflags(ifp, bits);
4108 		STATE_UNLOCK(ctx);
4109 		CTX_UNLOCK(ctx);
4110 		break;
4111 	case SIOCSIFFLAGS:
4112 		CTX_LOCK(ctx);
4113 		if (if_getflags(ifp) & IFF_UP) {
4114 			if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4115 				if ((if_getflags(ifp) ^ ctx->ifc_if_flags) &
4116 				    (IFF_PROMISC | IFF_ALLMULTI)) {
4117 					err = IFDI_PROMISC_SET(ctx, if_getflags(ifp));
4118 				}
4119 			} else
4120 				reinit = 1;
4121 		} else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4122 			iflib_stop(ctx);
4123 		}
4124 		ctx->ifc_if_flags = if_getflags(ifp);
4125 		CTX_UNLOCK(ctx);
4126 		break;
4127 	case SIOCADDMULTI:
4128 	case SIOCDELMULTI:
4129 		if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4130 			CTX_LOCK(ctx);
4131 			IFDI_INTR_DISABLE(ctx);
4132 			IFDI_MULTI_SET(ctx);
4133 			IFDI_INTR_ENABLE(ctx);
4134 			CTX_UNLOCK(ctx);
4135 		}
4136 		break;
4137 	case SIOCSIFMEDIA:
4138 		CTX_LOCK(ctx);
4139 		IFDI_MEDIA_SET(ctx);
4140 		CTX_UNLOCK(ctx);
4141 		/* falls thru */
4142 	case SIOCGIFMEDIA:
4143 	case SIOCGIFXMEDIA:
4144 		err = ifmedia_ioctl(ifp, ifr, &ctx->ifc_media, command);
4145 		break;
4146 	case SIOCGI2C:
4147 	{
4148 		struct ifi2creq i2c;
4149 
4150 		err = copyin(ifr_data_get_ptr(ifr), &i2c, sizeof(i2c));
4151 		if (err != 0)
4152 			break;
4153 		if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) {
4154 			err = EINVAL;
4155 			break;
4156 		}
4157 		if (i2c.len > sizeof(i2c.data)) {
4158 			err = EINVAL;
4159 			break;
4160 		}
4161 
4162 		if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0)
4163 			err = copyout(&i2c, ifr_data_get_ptr(ifr),
4164 			    sizeof(i2c));
4165 		break;
4166 	}
4167 	case SIOCSIFCAP:
4168 	{
4169 		int mask, setmask;
4170 
4171 		mask = ifr->ifr_reqcap ^ if_getcapenable(ifp);
4172 		setmask = 0;
4173 #ifdef TCP_OFFLOAD
4174 		setmask |= mask & (IFCAP_TOE4|IFCAP_TOE6);
4175 #endif
4176 		setmask |= (mask & IFCAP_FLAGS);
4177 
4178 		if (setmask & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6))
4179 			setmask |= (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6);
4180 		if ((mask & IFCAP_WOL) &&
4181 		    (if_getcapabilities(ifp) & IFCAP_WOL) != 0)
4182 			setmask |= (mask & (IFCAP_WOL_MCAST|IFCAP_WOL_MAGIC));
4183 		if_vlancap(ifp);
4184 		/*
4185 		 * want to ensure that traffic has stopped before we change any of the flags
4186 		 */
4187 		if (setmask) {
4188 			CTX_LOCK(ctx);
4189 			bits = if_getdrvflags(ifp);
4190 			if (bits & IFF_DRV_RUNNING)
4191 				iflib_stop(ctx);
4192 			STATE_LOCK(ctx);
4193 			if_togglecapenable(ifp, setmask);
4194 			STATE_UNLOCK(ctx);
4195 			if (bits & IFF_DRV_RUNNING)
4196 				iflib_init_locked(ctx);
4197 			STATE_LOCK(ctx);
4198 			if_setdrvflags(ifp, bits);
4199 			STATE_UNLOCK(ctx);
4200 			CTX_UNLOCK(ctx);
4201 		}
4202 		break;
4203 	}
4204 	case SIOCGPRIVATE_0:
4205 	case SIOCSDRVSPEC:
4206 	case SIOCGDRVSPEC:
4207 		CTX_LOCK(ctx);
4208 		err = IFDI_PRIV_IOCTL(ctx, command, data);
4209 		CTX_UNLOCK(ctx);
4210 		break;
4211 	default:
4212 		err = ether_ioctl(ifp, command, data);
4213 		break;
4214 	}
4215 	if (reinit)
4216 		iflib_if_init(ctx);
4217 	return (err);
4218 }
4219 
4220 static uint64_t
4221 iflib_if_get_counter(if_t ifp, ift_counter cnt)
4222 {
4223 	if_ctx_t ctx = if_getsoftc(ifp);
4224 
4225 	return (IFDI_GET_COUNTER(ctx, cnt));
4226 }
4227 
4228 /*********************************************************************
4229  *
4230  *  OTHER FUNCTIONS EXPORTED TO THE STACK
4231  *
4232  **********************************************************************/
4233 
4234 static void
4235 iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag)
4236 {
4237 	if_ctx_t ctx = if_getsoftc(ifp);
4238 
4239 	if ((void *)ctx != arg)
4240 		return;
4241 
4242 	if ((vtag == 0) || (vtag > 4095))
4243 		return;
4244 
4245 	CTX_LOCK(ctx);
4246 	IFDI_VLAN_REGISTER(ctx, vtag);
4247 	/* Re-init to load the changes */
4248 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
4249 		iflib_if_init_locked(ctx);
4250 	CTX_UNLOCK(ctx);
4251 }
4252 
4253 static void
4254 iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag)
4255 {
4256 	if_ctx_t ctx = if_getsoftc(ifp);
4257 
4258 	if ((void *)ctx != arg)
4259 		return;
4260 
4261 	if ((vtag == 0) || (vtag > 4095))
4262 		return;
4263 
4264 	CTX_LOCK(ctx);
4265 	IFDI_VLAN_UNREGISTER(ctx, vtag);
4266 	/* Re-init to load the changes */
4267 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
4268 		iflib_if_init_locked(ctx);
4269 	CTX_UNLOCK(ctx);
4270 }
4271 
4272 static void
4273 iflib_led_func(void *arg, int onoff)
4274 {
4275 	if_ctx_t ctx = arg;
4276 
4277 	CTX_LOCK(ctx);
4278 	IFDI_LED_FUNC(ctx, onoff);
4279 	CTX_UNLOCK(ctx);
4280 }
4281 
4282 /*********************************************************************
4283  *
4284  *  BUS FUNCTION DEFINITIONS
4285  *
4286  **********************************************************************/
4287 
4288 int
4289 iflib_device_probe(device_t dev)
4290 {
4291 	pci_vendor_info_t *ent;
4292 
4293 	uint16_t	pci_vendor_id, pci_device_id;
4294 	uint16_t	pci_subvendor_id, pci_subdevice_id;
4295 	uint16_t	pci_rev_id;
4296 	if_shared_ctx_t sctx;
4297 
4298 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4299 		return (ENOTSUP);
4300 
4301 	pci_vendor_id = pci_get_vendor(dev);
4302 	pci_device_id = pci_get_device(dev);
4303 	pci_subvendor_id = pci_get_subvendor(dev);
4304 	pci_subdevice_id = pci_get_subdevice(dev);
4305 	pci_rev_id = pci_get_revid(dev);
4306 	if (sctx->isc_parse_devinfo != NULL)
4307 		sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id);
4308 
4309 	ent = sctx->isc_vendor_info;
4310 	while (ent->pvi_vendor_id != 0) {
4311 		if (pci_vendor_id != ent->pvi_vendor_id) {
4312 			ent++;
4313 			continue;
4314 		}
4315 		if ((pci_device_id == ent->pvi_device_id) &&
4316 		    ((pci_subvendor_id == ent->pvi_subvendor_id) ||
4317 		     (ent->pvi_subvendor_id == 0)) &&
4318 		    ((pci_subdevice_id == ent->pvi_subdevice_id) ||
4319 		     (ent->pvi_subdevice_id == 0)) &&
4320 		    ((pci_rev_id == ent->pvi_rev_id) ||
4321 		     (ent->pvi_rev_id == 0))) {
4322 
4323 			device_set_desc_copy(dev, ent->pvi_name);
4324 			/* this needs to be changed to zero if the bus probing code
4325 			 * ever stops re-probing on best match because the sctx
4326 			 * may have its values over written by register calls
4327 			 * in subsequent probes
4328 			 */
4329 			return (BUS_PROBE_DEFAULT);
4330 		}
4331 		ent++;
4332 	}
4333 	return (ENXIO);
4334 }
4335 
4336 static void
4337 iflib_reset_qvalues(if_ctx_t ctx)
4338 {
4339 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4340 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4341 	device_t dev = ctx->ifc_dev;
4342 	int i;
4343 
4344 	scctx->isc_txrx_budget_bytes_max = IFLIB_MAX_TX_BYTES;
4345 	scctx->isc_tx_qdepth = IFLIB_DEFAULT_TX_QDEPTH;
4346 	/*
4347 	 * XXX sanity check that ntxd & nrxd are a power of 2
4348 	 */
4349 	if (ctx->ifc_sysctl_ntxqs != 0)
4350 		scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs;
4351 	if (ctx->ifc_sysctl_nrxqs != 0)
4352 		scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs;
4353 
4354 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4355 		if (ctx->ifc_sysctl_ntxds[i] != 0)
4356 			scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i];
4357 		else
4358 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4359 	}
4360 
4361 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4362 		if (ctx->ifc_sysctl_nrxds[i] != 0)
4363 			scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i];
4364 		else
4365 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4366 	}
4367 
4368 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4369 		if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) {
4370 			device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n",
4371 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]);
4372 			scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i];
4373 		}
4374 		if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) {
4375 			device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n",
4376 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]);
4377 			scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i];
4378 		}
4379 	}
4380 
4381 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4382 		if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) {
4383 			device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n",
4384 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]);
4385 			scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i];
4386 		}
4387 		if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) {
4388 			device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n",
4389 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]);
4390 			scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i];
4391 		}
4392 	}
4393 }
4394 
4395 int
4396 iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp)
4397 {
4398 	int err, rid, msix;
4399 	if_ctx_t ctx;
4400 	if_t ifp;
4401 	if_softc_ctx_t scctx;
4402 	int i;
4403 	uint16_t main_txq;
4404 	uint16_t main_rxq;
4405 
4406 
4407 	ctx = malloc(sizeof(* ctx), M_IFLIB, M_WAITOK|M_ZERO);
4408 
4409 	if (sc == NULL) {
4410 		sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4411 		device_set_softc(dev, ctx);
4412 		ctx->ifc_flags |= IFC_SC_ALLOCATED;
4413 	}
4414 
4415 	ctx->ifc_sctx = sctx;
4416 	ctx->ifc_dev = dev;
4417 	ctx->ifc_softc = sc;
4418 
4419 	if ((err = iflib_register(ctx)) != 0) {
4420 		if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4421 			free(sc, M_IFLIB);
4422 		free(ctx, M_IFLIB);
4423 		device_printf(dev, "iflib_register failed %d\n", err);
4424 		return (err);
4425 	}
4426 	iflib_add_device_sysctl_pre(ctx);
4427 
4428 	scctx = &ctx->ifc_softc_ctx;
4429 	ifp = ctx->ifc_ifp;
4430 
4431 	iflib_reset_qvalues(ctx);
4432 	CTX_LOCK(ctx);
4433 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4434 		CTX_UNLOCK(ctx);
4435 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4436 		return (err);
4437 	}
4438 	_iflib_pre_assert(scctx);
4439 	ctx->ifc_txrx = *scctx->isc_txrx;
4440 
4441 #ifdef INVARIANTS
4442 	MPASS(scctx->isc_capabilities);
4443 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
4444 		MPASS(scctx->isc_tx_csum_flags);
4445 #endif
4446 
4447 	if_setcapabilities(ifp, scctx->isc_capabilities | IFCAP_HWSTATS);
4448 	if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS);
4449 
4450 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4451 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4452 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4453 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4454 
4455 #ifdef ACPI_DMAR
4456 	if (dmar_get_dma_tag(device_get_parent(dev), dev) != NULL)
4457 		ctx->ifc_flags |= IFC_DMAR;
4458 #elif !(defined(__i386__) || defined(__amd64__))
4459 	/* set unconditionally for !x86 */
4460 	ctx->ifc_flags |= IFC_DMAR;
4461 #endif
4462 
4463 	main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4464 	main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4465 
4466 	/* XXX change for per-queue sizes */
4467 	device_printf(dev, "using %d tx descriptors and %d rx descriptors\n",
4468 		      scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4469 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4470 		if (!powerof2(scctx->isc_nrxd[i])) {
4471 			/* round down instead? */
4472 			device_printf(dev, "# rx descriptors must be a power of 2\n");
4473 			err = EINVAL;
4474 			goto fail;
4475 		}
4476 	}
4477 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4478 		if (!powerof2(scctx->isc_ntxd[i])) {
4479 			device_printf(dev,
4480 			    "# tx descriptors must be a power of 2");
4481 			err = EINVAL;
4482 			goto fail;
4483 		}
4484 	}
4485 
4486 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4487 	    MAX_SINGLE_PACKET_FRACTION)
4488 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4489 		    MAX_SINGLE_PACKET_FRACTION);
4490 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4491 	    MAX_SINGLE_PACKET_FRACTION)
4492 		scctx->isc_tx_tso_segments_max = max(1,
4493 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4494 
4495 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4496 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
4497 		/*
4498 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
4499 		 * but some MACs do.
4500 		 */
4501 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
4502 		    IP_MAXPACKET));
4503 		/*
4504 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
4505 		 * into account.  In the worst case, each of these calls will
4506 		 * add another mbuf and, thus, the requirement for another DMA
4507 		 * segment.  So for best performance, it doesn't make sense to
4508 		 * advertize a maximum of TSO segments that typically will
4509 		 * require defragmentation in iflib_encap().
4510 		 */
4511 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
4512 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
4513 	}
4514 	if (scctx->isc_rss_table_size == 0)
4515 		scctx->isc_rss_table_size = 64;
4516 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4517 
4518 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4519 	/* XXX format name */
4520 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx, -1, "admin");
4521 
4522 	/* Set up cpu set.  If it fails, use the set of all CPUs. */
4523 	if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) != 0) {
4524 		device_printf(dev, "Unable to fetch CPU list\n");
4525 		CPU_COPY(&all_cpus, &ctx->ifc_cpus);
4526 	}
4527 	MPASS(CPU_COUNT(&ctx->ifc_cpus) > 0);
4528 
4529 	/*
4530 	** Now setup MSI or MSI/X, should
4531 	** return us the number of supported
4532 	** vectors. (Will be 1 for MSI)
4533 	*/
4534 	if (sctx->isc_flags & IFLIB_SKIP_MSIX) {
4535 		msix = scctx->isc_vectors;
4536 	} else if (scctx->isc_msix_bar != 0)
4537 	       /*
4538 		* The simple fact that isc_msix_bar is not 0 does not mean we
4539 		* we have a good value there that is known to work.
4540 		*/
4541 		msix = iflib_msix_init(ctx);
4542 	else {
4543 		scctx->isc_vectors = 1;
4544 		scctx->isc_ntxqsets = 1;
4545 		scctx->isc_nrxqsets = 1;
4546 		scctx->isc_intr = IFLIB_INTR_LEGACY;
4547 		msix = 0;
4548 	}
4549 	/* Get memory for the station queues */
4550 	if ((err = iflib_queues_alloc(ctx))) {
4551 		device_printf(dev, "Unable to allocate queue memory\n");
4552 		goto fail;
4553 	}
4554 
4555 	if ((err = iflib_qset_structures_setup(ctx)))
4556 		goto fail_queues;
4557 
4558 	/*
4559 	 * Group taskqueues aren't properly set up until SMP is started,
4560 	 * so we disable interrupts until we can handle them post
4561 	 * SI_SUB_SMP.
4562 	 *
4563 	 * XXX: disabling interrupts doesn't actually work, at least for
4564 	 * the non-MSI case.  When they occur before SI_SUB_SMP completes,
4565 	 * we do null handling and depend on this not causing too large an
4566 	 * interrupt storm.
4567 	 */
4568 	IFDI_INTR_DISABLE(ctx);
4569 	if (msix > 1 && (err = IFDI_MSIX_INTR_ASSIGN(ctx, msix)) != 0) {
4570 		device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n", err);
4571 		goto fail_intr_free;
4572 	}
4573 	if (msix <= 1) {
4574 		rid = 0;
4575 		if (scctx->isc_intr == IFLIB_INTR_MSI) {
4576 			MPASS(msix == 1);
4577 			rid = 1;
4578 		}
4579 		if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) {
4580 			device_printf(dev, "iflib_legacy_setup failed %d\n", err);
4581 			goto fail_intr_free;
4582 		}
4583 	}
4584 
4585 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac);
4586 
4587 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4588 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4589 		goto fail_detach;
4590 	}
4591 
4592 	/*
4593 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4594 	 * This must appear after the call to ether_ifattach() because
4595 	 * ether_ifattach() sets if_hdrlen to the default value.
4596 	 */
4597 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4598 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
4599 
4600 	if ((err = iflib_netmap_attach(ctx))) {
4601 		device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err);
4602 		goto fail_detach;
4603 	}
4604 	*ctxp = ctx;
4605 
4606 	NETDUMP_SET(ctx->ifc_ifp, iflib);
4607 
4608 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4609 	iflib_add_device_sysctl_post(ctx);
4610 	ctx->ifc_flags |= IFC_INIT_DONE;
4611 	CTX_UNLOCK(ctx);
4612 	return (0);
4613 fail_detach:
4614 	ether_ifdetach(ctx->ifc_ifp);
4615 fail_intr_free:
4616 	if (scctx->isc_intr == IFLIB_INTR_MSIX || scctx->isc_intr == IFLIB_INTR_MSI)
4617 		pci_release_msi(ctx->ifc_dev);
4618 fail_queues:
4619 	iflib_tx_structures_free(ctx);
4620 	iflib_rx_structures_free(ctx);
4621 fail:
4622 	IFDI_DETACH(ctx);
4623 	CTX_UNLOCK(ctx);
4624 	return (err);
4625 }
4626 
4627 int
4628 iflib_pseudo_register(device_t dev, if_shared_ctx_t sctx, if_ctx_t *ctxp,
4629 					  struct iflib_cloneattach_ctx *clctx)
4630 {
4631 	int err;
4632 	if_ctx_t ctx;
4633 	if_t ifp;
4634 	if_softc_ctx_t scctx;
4635 	int i;
4636 	void *sc;
4637 	uint16_t main_txq;
4638 	uint16_t main_rxq;
4639 
4640 	ctx = malloc(sizeof(*ctx), M_IFLIB, M_WAITOK|M_ZERO);
4641 	sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4642 	ctx->ifc_flags |= IFC_SC_ALLOCATED;
4643 	if (sctx->isc_flags & (IFLIB_PSEUDO|IFLIB_VIRTUAL))
4644 		ctx->ifc_flags |= IFC_PSEUDO;
4645 
4646 	ctx->ifc_sctx = sctx;
4647 	ctx->ifc_softc = sc;
4648 	ctx->ifc_dev = dev;
4649 
4650 	if ((err = iflib_register(ctx)) != 0) {
4651 		device_printf(dev, "%s: iflib_register failed %d\n", __func__, err);
4652 		free(sc, M_IFLIB);
4653 		free(ctx, M_IFLIB);
4654 		return (err);
4655 	}
4656 	iflib_add_device_sysctl_pre(ctx);
4657 
4658 	scctx = &ctx->ifc_softc_ctx;
4659 	ifp = ctx->ifc_ifp;
4660 
4661 	/*
4662 	 * XXX sanity check that ntxd & nrxd are a power of 2
4663 	 */
4664 	iflib_reset_qvalues(ctx);
4665 
4666 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4667 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4668 		return (err);
4669 	}
4670 	if (sctx->isc_flags & IFLIB_GEN_MAC)
4671 		iflib_gen_mac(ctx);
4672 	if ((err = IFDI_CLONEATTACH(ctx, clctx->cc_ifc, clctx->cc_name,
4673 								clctx->cc_params)) != 0) {
4674 		device_printf(dev, "IFDI_CLONEATTACH failed %d\n", err);
4675 		return (err);
4676 	}
4677 	ifmedia_add(&ctx->ifc_media, IFM_ETHER | IFM_1000_T | IFM_FDX, 0, NULL);
4678 	ifmedia_add(&ctx->ifc_media, IFM_ETHER | IFM_AUTO, 0, NULL);
4679 	ifmedia_set(&ctx->ifc_media, IFM_ETHER | IFM_AUTO);
4680 
4681 #ifdef INVARIANTS
4682 	MPASS(scctx->isc_capabilities);
4683 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
4684 		MPASS(scctx->isc_tx_csum_flags);
4685 #endif
4686 
4687 	if_setcapabilities(ifp, scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_LINKSTATE);
4688 	if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_LINKSTATE);
4689 
4690 	ifp->if_flags |= IFF_NOGROUP;
4691 	if (sctx->isc_flags & IFLIB_PSEUDO) {
4692 		ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac);
4693 
4694 		if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4695 			device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4696 			goto fail_detach;
4697 		}
4698 		*ctxp = ctx;
4699 
4700 		/*
4701 		 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4702 		 * This must appear after the call to ether_ifattach() because
4703 		 * ether_ifattach() sets if_hdrlen to the default value.
4704 		 */
4705 		if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4706 			if_setifheaderlen(ifp,
4707 			    sizeof(struct ether_vlan_header));
4708 
4709 		if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4710 		iflib_add_device_sysctl_post(ctx);
4711 		ctx->ifc_flags |= IFC_INIT_DONE;
4712 		return (0);
4713 	}
4714 	_iflib_pre_assert(scctx);
4715 	ctx->ifc_txrx = *scctx->isc_txrx;
4716 
4717 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4718 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4719 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4720 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4721 
4722 	main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4723 	main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4724 
4725 	/* XXX change for per-queue sizes */
4726 	device_printf(dev, "using %d tx descriptors and %d rx descriptors\n",
4727 		      scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4728 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4729 		if (!powerof2(scctx->isc_nrxd[i])) {
4730 			/* round down instead? */
4731 			device_printf(dev, "# rx descriptors must be a power of 2\n");
4732 			err = EINVAL;
4733 			goto fail;
4734 		}
4735 	}
4736 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4737 		if (!powerof2(scctx->isc_ntxd[i])) {
4738 			device_printf(dev,
4739 			    "# tx descriptors must be a power of 2");
4740 			err = EINVAL;
4741 			goto fail;
4742 		}
4743 	}
4744 
4745 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4746 	    MAX_SINGLE_PACKET_FRACTION)
4747 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4748 		    MAX_SINGLE_PACKET_FRACTION);
4749 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4750 	    MAX_SINGLE_PACKET_FRACTION)
4751 		scctx->isc_tx_tso_segments_max = max(1,
4752 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4753 
4754 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4755 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
4756 		/*
4757 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
4758 		 * but some MACs do.
4759 		 */
4760 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
4761 		    IP_MAXPACKET));
4762 		/*
4763 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
4764 		 * into account.  In the worst case, each of these calls will
4765 		 * add another mbuf and, thus, the requirement for another DMA
4766 		 * segment.  So for best performance, it doesn't make sense to
4767 		 * advertize a maximum of TSO segments that typically will
4768 		 * require defragmentation in iflib_encap().
4769 		 */
4770 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
4771 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
4772 	}
4773 	if (scctx->isc_rss_table_size == 0)
4774 		scctx->isc_rss_table_size = 64;
4775 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4776 
4777 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4778 	/* XXX format name */
4779 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx, -1, "admin");
4780 
4781 	/* XXX --- can support > 1 -- but keep it simple for now */
4782 	scctx->isc_intr = IFLIB_INTR_LEGACY;
4783 
4784 	/* Get memory for the station queues */
4785 	if ((err = iflib_queues_alloc(ctx))) {
4786 		device_printf(dev, "Unable to allocate queue memory\n");
4787 		goto fail;
4788 	}
4789 
4790 	if ((err = iflib_qset_structures_setup(ctx))) {
4791 		device_printf(dev, "qset structure setup failed %d\n", err);
4792 		goto fail_queues;
4793 	}
4794 
4795 	/*
4796 	 * XXX What if anything do we want to do about interrupts?
4797 	 */
4798 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac);
4799 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4800 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4801 		goto fail_detach;
4802 	}
4803 
4804 	/*
4805 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4806 	 * This must appear after the call to ether_ifattach() because
4807 	 * ether_ifattach() sets if_hdrlen to the default value.
4808 	 */
4809 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4810 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
4811 
4812 	/* XXX handle more than one queue */
4813 	for (i = 0; i < scctx->isc_nrxqsets; i++)
4814 		IFDI_RX_CLSET(ctx, 0, i, ctx->ifc_rxqs[i].ifr_fl[0].ifl_sds.ifsd_cl);
4815 
4816 	*ctxp = ctx;
4817 
4818 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4819 	iflib_add_device_sysctl_post(ctx);
4820 	ctx->ifc_flags |= IFC_INIT_DONE;
4821 	return (0);
4822 fail_detach:
4823 	ether_ifdetach(ctx->ifc_ifp);
4824 fail_queues:
4825 	iflib_tx_structures_free(ctx);
4826 	iflib_rx_structures_free(ctx);
4827 fail:
4828 	IFDI_DETACH(ctx);
4829 	return (err);
4830 }
4831 
4832 int
4833 iflib_pseudo_deregister(if_ctx_t ctx)
4834 {
4835 	if_t ifp = ctx->ifc_ifp;
4836 	iflib_txq_t txq;
4837 	iflib_rxq_t rxq;
4838 	int i, j;
4839 	struct taskqgroup *tqg;
4840 	iflib_fl_t fl;
4841 
4842 	/* Unregister VLAN events */
4843 	if (ctx->ifc_vlan_attach_event != NULL)
4844 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
4845 	if (ctx->ifc_vlan_detach_event != NULL)
4846 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
4847 
4848 	ether_ifdetach(ifp);
4849 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
4850 	CTX_LOCK_DESTROY(ctx);
4851 	/* XXX drain any dependent tasks */
4852 	tqg = qgroup_if_io_tqg;
4853 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
4854 		callout_drain(&txq->ift_timer);
4855 		if (txq->ift_task.gt_uniq != NULL)
4856 			taskqgroup_detach(tqg, &txq->ift_task);
4857 	}
4858 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4859 		if (rxq->ifr_task.gt_uniq != NULL)
4860 			taskqgroup_detach(tqg, &rxq->ifr_task);
4861 
4862 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
4863 			free(fl->ifl_rx_bitmap, M_IFLIB);
4864 	}
4865 	tqg = qgroup_if_config_tqg;
4866 	if (ctx->ifc_admin_task.gt_uniq != NULL)
4867 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
4868 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
4869 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
4870 
4871 	if_free(ifp);
4872 
4873 	iflib_tx_structures_free(ctx);
4874 	iflib_rx_structures_free(ctx);
4875 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4876 		free(ctx->ifc_softc, M_IFLIB);
4877 	free(ctx, M_IFLIB);
4878 	return (0);
4879 }
4880 
4881 int
4882 iflib_device_attach(device_t dev)
4883 {
4884 	if_ctx_t ctx;
4885 	if_shared_ctx_t sctx;
4886 
4887 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4888 		return (ENOTSUP);
4889 
4890 	pci_enable_busmaster(dev);
4891 
4892 	return (iflib_device_register(dev, NULL, sctx, &ctx));
4893 }
4894 
4895 int
4896 iflib_device_deregister(if_ctx_t ctx)
4897 {
4898 	if_t ifp = ctx->ifc_ifp;
4899 	iflib_txq_t txq;
4900 	iflib_rxq_t rxq;
4901 	device_t dev = ctx->ifc_dev;
4902 	int i, j;
4903 	struct taskqgroup *tqg;
4904 	iflib_fl_t fl;
4905 
4906 	/* Make sure VLANS are not using driver */
4907 	if (if_vlantrunkinuse(ifp)) {
4908 		device_printf(dev,"Vlan in use, detach first\n");
4909 		return (EBUSY);
4910 	}
4911 
4912 	CTX_LOCK(ctx);
4913 	ctx->ifc_in_detach = 1;
4914 	iflib_stop(ctx);
4915 	CTX_UNLOCK(ctx);
4916 
4917 	/* Unregister VLAN events */
4918 	if (ctx->ifc_vlan_attach_event != NULL)
4919 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
4920 	if (ctx->ifc_vlan_detach_event != NULL)
4921 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
4922 
4923 	iflib_netmap_detach(ifp);
4924 	ether_ifdetach(ifp);
4925 	if (ctx->ifc_led_dev != NULL)
4926 		led_destroy(ctx->ifc_led_dev);
4927 	/* XXX drain any dependent tasks */
4928 	tqg = qgroup_if_io_tqg;
4929 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
4930 		callout_drain(&txq->ift_timer);
4931 		if (txq->ift_task.gt_uniq != NULL)
4932 			taskqgroup_detach(tqg, &txq->ift_task);
4933 	}
4934 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4935 		if (rxq->ifr_task.gt_uniq != NULL)
4936 			taskqgroup_detach(tqg, &rxq->ifr_task);
4937 
4938 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
4939 			free(fl->ifl_rx_bitmap, M_IFLIB);
4940 
4941 	}
4942 	tqg = qgroup_if_config_tqg;
4943 	if (ctx->ifc_admin_task.gt_uniq != NULL)
4944 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
4945 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
4946 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
4947 	CTX_LOCK(ctx);
4948 	IFDI_DETACH(ctx);
4949 	CTX_UNLOCK(ctx);
4950 
4951 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
4952 	CTX_LOCK_DESTROY(ctx);
4953 	device_set_softc(ctx->ifc_dev, NULL);
4954 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) {
4955 		pci_release_msi(dev);
4956 	}
4957 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) {
4958 		iflib_irq_free(ctx, &ctx->ifc_legacy_irq);
4959 	}
4960 	if (ctx->ifc_msix_mem != NULL) {
4961 		bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY,
4962 			ctx->ifc_softc_ctx.isc_msix_bar, ctx->ifc_msix_mem);
4963 		ctx->ifc_msix_mem = NULL;
4964 	}
4965 
4966 	bus_generic_detach(dev);
4967 	if_free(ifp);
4968 
4969 	iflib_tx_structures_free(ctx);
4970 	iflib_rx_structures_free(ctx);
4971 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4972 		free(ctx->ifc_softc, M_IFLIB);
4973 	free(ctx, M_IFLIB);
4974 	return (0);
4975 }
4976 
4977 
4978 int
4979 iflib_device_detach(device_t dev)
4980 {
4981 	if_ctx_t ctx = device_get_softc(dev);
4982 
4983 	return (iflib_device_deregister(ctx));
4984 }
4985 
4986 int
4987 iflib_device_suspend(device_t dev)
4988 {
4989 	if_ctx_t ctx = device_get_softc(dev);
4990 
4991 	CTX_LOCK(ctx);
4992 	IFDI_SUSPEND(ctx);
4993 	CTX_UNLOCK(ctx);
4994 
4995 	return bus_generic_suspend(dev);
4996 }
4997 int
4998 iflib_device_shutdown(device_t dev)
4999 {
5000 	if_ctx_t ctx = device_get_softc(dev);
5001 
5002 	CTX_LOCK(ctx);
5003 	IFDI_SHUTDOWN(ctx);
5004 	CTX_UNLOCK(ctx);
5005 
5006 	return bus_generic_suspend(dev);
5007 }
5008 
5009 
5010 int
5011 iflib_device_resume(device_t dev)
5012 {
5013 	if_ctx_t ctx = device_get_softc(dev);
5014 	iflib_txq_t txq = ctx->ifc_txqs;
5015 
5016 	CTX_LOCK(ctx);
5017 	IFDI_RESUME(ctx);
5018 	iflib_init_locked(ctx);
5019 	CTX_UNLOCK(ctx);
5020 	for (int i = 0; i < NTXQSETS(ctx); i++, txq++)
5021 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
5022 
5023 	return (bus_generic_resume(dev));
5024 }
5025 
5026 int
5027 iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params)
5028 {
5029 	int error;
5030 	if_ctx_t ctx = device_get_softc(dev);
5031 
5032 	CTX_LOCK(ctx);
5033 	error = IFDI_IOV_INIT(ctx, num_vfs, params);
5034 	CTX_UNLOCK(ctx);
5035 
5036 	return (error);
5037 }
5038 
5039 void
5040 iflib_device_iov_uninit(device_t dev)
5041 {
5042 	if_ctx_t ctx = device_get_softc(dev);
5043 
5044 	CTX_LOCK(ctx);
5045 	IFDI_IOV_UNINIT(ctx);
5046 	CTX_UNLOCK(ctx);
5047 }
5048 
5049 int
5050 iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params)
5051 {
5052 	int error;
5053 	if_ctx_t ctx = device_get_softc(dev);
5054 
5055 	CTX_LOCK(ctx);
5056 	error = IFDI_IOV_VF_ADD(ctx, vfnum, params);
5057 	CTX_UNLOCK(ctx);
5058 
5059 	return (error);
5060 }
5061 
5062 /*********************************************************************
5063  *
5064  *  MODULE FUNCTION DEFINITIONS
5065  *
5066  **********************************************************************/
5067 
5068 /*
5069  * - Start a fast taskqueue thread for each core
5070  * - Start a taskqueue for control operations
5071  */
5072 static int
5073 iflib_module_init(void)
5074 {
5075 	return (0);
5076 }
5077 
5078 static int
5079 iflib_module_event_handler(module_t mod, int what, void *arg)
5080 {
5081 	int err;
5082 
5083 	switch (what) {
5084 	case MOD_LOAD:
5085 		if ((err = iflib_module_init()) != 0)
5086 			return (err);
5087 		break;
5088 	case MOD_UNLOAD:
5089 		return (EBUSY);
5090 	default:
5091 		return (EOPNOTSUPP);
5092 	}
5093 
5094 	return (0);
5095 }
5096 
5097 /*********************************************************************
5098  *
5099  *  PUBLIC FUNCTION DEFINITIONS
5100  *     ordered as in iflib.h
5101  *
5102  **********************************************************************/
5103 
5104 
5105 static void
5106 _iflib_assert(if_shared_ctx_t sctx)
5107 {
5108 	MPASS(sctx->isc_tx_maxsize);
5109 	MPASS(sctx->isc_tx_maxsegsize);
5110 
5111 	MPASS(sctx->isc_rx_maxsize);
5112 	MPASS(sctx->isc_rx_nsegments);
5113 	MPASS(sctx->isc_rx_maxsegsize);
5114 
5115 	MPASS(sctx->isc_nrxd_min[0]);
5116 	MPASS(sctx->isc_nrxd_max[0]);
5117 	MPASS(sctx->isc_nrxd_default[0]);
5118 	MPASS(sctx->isc_ntxd_min[0]);
5119 	MPASS(sctx->isc_ntxd_max[0]);
5120 	MPASS(sctx->isc_ntxd_default[0]);
5121 }
5122 
5123 static void
5124 _iflib_pre_assert(if_softc_ctx_t scctx)
5125 {
5126 
5127 	MPASS(scctx->isc_txrx->ift_txd_encap);
5128 	MPASS(scctx->isc_txrx->ift_txd_flush);
5129 	MPASS(scctx->isc_txrx->ift_txd_credits_update);
5130 	MPASS(scctx->isc_txrx->ift_rxd_available);
5131 	MPASS(scctx->isc_txrx->ift_rxd_pkt_get);
5132 	MPASS(scctx->isc_txrx->ift_rxd_refill);
5133 	MPASS(scctx->isc_txrx->ift_rxd_flush);
5134 }
5135 
5136 static int
5137 iflib_register(if_ctx_t ctx)
5138 {
5139 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5140 	driver_t *driver = sctx->isc_driver;
5141 	device_t dev = ctx->ifc_dev;
5142 	if_t ifp;
5143 
5144 	_iflib_assert(sctx);
5145 
5146 	CTX_LOCK_INIT(ctx);
5147 	STATE_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev));
5148 	ifp = ctx->ifc_ifp = if_gethandle(IFT_ETHER);
5149 	if (ifp == NULL) {
5150 		device_printf(dev, "can not allocate ifnet structure\n");
5151 		return (ENOMEM);
5152 	}
5153 
5154 	/*
5155 	 * Initialize our context's device specific methods
5156 	 */
5157 	kobj_init((kobj_t) ctx, (kobj_class_t) driver);
5158 	kobj_class_compile((kobj_class_t) driver);
5159 	driver->refs++;
5160 
5161 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
5162 	if_setsoftc(ifp, ctx);
5163 	if_setdev(ifp, dev);
5164 	if_setinitfn(ifp, iflib_if_init);
5165 	if_setioctlfn(ifp, iflib_if_ioctl);
5166 	if_settransmitfn(ifp, iflib_if_transmit);
5167 	if_setqflushfn(ifp, iflib_if_qflush);
5168 	if_setflags(ifp, IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST);
5169 
5170 	ctx->ifc_vlan_attach_event =
5171 		EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx,
5172 							  EVENTHANDLER_PRI_FIRST);
5173 	ctx->ifc_vlan_detach_event =
5174 		EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx,
5175 							  EVENTHANDLER_PRI_FIRST);
5176 
5177 	ifmedia_init(&ctx->ifc_media, IFM_IMASK,
5178 					 iflib_media_change, iflib_media_status);
5179 
5180 	return (0);
5181 }
5182 
5183 
5184 static int
5185 iflib_queues_alloc(if_ctx_t ctx)
5186 {
5187 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5188 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5189 	device_t dev = ctx->ifc_dev;
5190 	int nrxqsets = scctx->isc_nrxqsets;
5191 	int ntxqsets = scctx->isc_ntxqsets;
5192 	iflib_txq_t txq;
5193 	iflib_rxq_t rxq;
5194 	iflib_fl_t fl = NULL;
5195 	int i, j, cpu, err, txconf, rxconf;
5196 	iflib_dma_info_t ifdip;
5197 	uint32_t *rxqsizes = scctx->isc_rxqsizes;
5198 	uint32_t *txqsizes = scctx->isc_txqsizes;
5199 	uint8_t nrxqs = sctx->isc_nrxqs;
5200 	uint8_t ntxqs = sctx->isc_ntxqs;
5201 	int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1;
5202 	caddr_t *vaddrs;
5203 	uint64_t *paddrs;
5204 
5205 	KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1"));
5206 	KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1"));
5207 
5208 	/* Allocate the TX ring struct memory */
5209 	if (!(ctx->ifc_txqs =
5210 	    (iflib_txq_t) malloc(sizeof(struct iflib_txq) *
5211 	    ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5212 		device_printf(dev, "Unable to allocate TX ring memory\n");
5213 		err = ENOMEM;
5214 		goto fail;
5215 	}
5216 
5217 	/* Now allocate the RX */
5218 	if (!(ctx->ifc_rxqs =
5219 	    (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) *
5220 	    nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5221 		device_printf(dev, "Unable to allocate RX ring memory\n");
5222 		err = ENOMEM;
5223 		goto rx_fail;
5224 	}
5225 
5226 	txq = ctx->ifc_txqs;
5227 	rxq = ctx->ifc_rxqs;
5228 
5229 	/*
5230 	 * XXX handle allocation failure
5231 	 */
5232 	for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) {
5233 		/* Set up some basics */
5234 
5235 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs, M_IFLIB, M_WAITOK|M_ZERO)) == NULL) {
5236 			device_printf(dev, "failed to allocate iflib_dma_info\n");
5237 			err = ENOMEM;
5238 			goto err_tx_desc;
5239 		}
5240 		txq->ift_ifdi = ifdip;
5241 		for (j = 0; j < ntxqs; j++, ifdip++) {
5242 			if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, BUS_DMA_NOWAIT)) {
5243 				device_printf(dev, "Unable to allocate Descriptor memory\n");
5244 				err = ENOMEM;
5245 				goto err_tx_desc;
5246 			}
5247 			txq->ift_txd_size[j] = scctx->isc_txd_size[j];
5248 			bzero((void *)ifdip->idi_vaddr, txqsizes[j]);
5249 		}
5250 		txq->ift_ctx = ctx;
5251 		txq->ift_id = i;
5252 		if (sctx->isc_flags & IFLIB_HAS_TXCQ) {
5253 			txq->ift_br_offset = 1;
5254 		} else {
5255 			txq->ift_br_offset = 0;
5256 		}
5257 		/* XXX fix this */
5258 		txq->ift_timer.c_cpu = cpu;
5259 
5260 		if (iflib_txsd_alloc(txq)) {
5261 			device_printf(dev, "Critical Failure setting up TX buffers\n");
5262 			err = ENOMEM;
5263 			goto err_tx_desc;
5264 		}
5265 
5266 		/* Initialize the TX lock */
5267 		snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:tx(%d):callout",
5268 		    device_get_nameunit(dev), txq->ift_id);
5269 		mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF);
5270 		callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0);
5271 
5272 		snprintf(txq->ift_db_mtx_name, MTX_NAME_LEN, "%s:tx(%d):db",
5273 			 device_get_nameunit(dev), txq->ift_id);
5274 
5275 		err = ifmp_ring_alloc(&txq->ift_br, 2048, txq, iflib_txq_drain,
5276 				      iflib_txq_can_drain, M_IFLIB, M_WAITOK);
5277 		if (err) {
5278 			/* XXX free any allocated rings */
5279 			device_printf(dev, "Unable to allocate buf_ring\n");
5280 			goto err_tx_desc;
5281 		}
5282 	}
5283 
5284 	for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) {
5285 		/* Set up some basics */
5286 
5287 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs, M_IFLIB, M_WAITOK|M_ZERO)) == NULL) {
5288 			device_printf(dev, "failed to allocate iflib_dma_info\n");
5289 			err = ENOMEM;
5290 			goto err_tx_desc;
5291 		}
5292 
5293 		rxq->ifr_ifdi = ifdip;
5294 		/* XXX this needs to be changed if #rx queues != #tx queues */
5295 		rxq->ifr_ntxqirq = 1;
5296 		rxq->ifr_txqid[0] = i;
5297 		for (j = 0; j < nrxqs; j++, ifdip++) {
5298 			if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, BUS_DMA_NOWAIT)) {
5299 				device_printf(dev, "Unable to allocate Descriptor memory\n");
5300 				err = ENOMEM;
5301 				goto err_tx_desc;
5302 			}
5303 			bzero((void *)ifdip->idi_vaddr, rxqsizes[j]);
5304 		}
5305 		rxq->ifr_ctx = ctx;
5306 		rxq->ifr_id = i;
5307 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
5308 			rxq->ifr_fl_offset = 1;
5309 		} else {
5310 			rxq->ifr_fl_offset = 0;
5311 		}
5312 		rxq->ifr_nfl = nfree_lists;
5313 		if (!(fl =
5314 			  (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) {
5315 			device_printf(dev, "Unable to allocate free list memory\n");
5316 			err = ENOMEM;
5317 			goto err_tx_desc;
5318 		}
5319 		rxq->ifr_fl = fl;
5320 		for (j = 0; j < nfree_lists; j++) {
5321 			fl[j].ifl_rxq = rxq;
5322 			fl[j].ifl_id = j;
5323 			fl[j].ifl_ifdi = &rxq->ifr_ifdi[j + rxq->ifr_fl_offset];
5324 			fl[j].ifl_rxd_size = scctx->isc_rxd_size[j];
5325 		}
5326         /* Allocate receive buffers for the ring*/
5327 		if (iflib_rxsd_alloc(rxq)) {
5328 			device_printf(dev,
5329 			    "Critical Failure setting up receive buffers\n");
5330 			err = ENOMEM;
5331 			goto err_rx_desc;
5332 		}
5333 
5334 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5335 			fl->ifl_rx_bitmap = bit_alloc(fl->ifl_size, M_IFLIB, M_WAITOK|M_ZERO);
5336 	}
5337 
5338 	/* TXQs */
5339 	vaddrs = malloc(sizeof(caddr_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5340 	paddrs = malloc(sizeof(uint64_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5341 	for (i = 0; i < ntxqsets; i++) {
5342 		iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi;
5343 
5344 		for (j = 0; j < ntxqs; j++, di++) {
5345 			vaddrs[i*ntxqs + j] = di->idi_vaddr;
5346 			paddrs[i*ntxqs + j] = di->idi_paddr;
5347 		}
5348 	}
5349 	if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) {
5350 		device_printf(ctx->ifc_dev, "device queue allocation failed\n");
5351 		iflib_tx_structures_free(ctx);
5352 		free(vaddrs, M_IFLIB);
5353 		free(paddrs, M_IFLIB);
5354 		goto err_rx_desc;
5355 	}
5356 	free(vaddrs, M_IFLIB);
5357 	free(paddrs, M_IFLIB);
5358 
5359 	/* RXQs */
5360 	vaddrs = malloc(sizeof(caddr_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5361 	paddrs = malloc(sizeof(uint64_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5362 	for (i = 0; i < nrxqsets; i++) {
5363 		iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi;
5364 
5365 		for (j = 0; j < nrxqs; j++, di++) {
5366 			vaddrs[i*nrxqs + j] = di->idi_vaddr;
5367 			paddrs[i*nrxqs + j] = di->idi_paddr;
5368 		}
5369 	}
5370 	if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) {
5371 		device_printf(ctx->ifc_dev, "device queue allocation failed\n");
5372 		iflib_tx_structures_free(ctx);
5373 		free(vaddrs, M_IFLIB);
5374 		free(paddrs, M_IFLIB);
5375 		goto err_rx_desc;
5376 	}
5377 	free(vaddrs, M_IFLIB);
5378 	free(paddrs, M_IFLIB);
5379 
5380 	return (0);
5381 
5382 /* XXX handle allocation failure changes */
5383 err_rx_desc:
5384 err_tx_desc:
5385 rx_fail:
5386 	if (ctx->ifc_rxqs != NULL)
5387 		free(ctx->ifc_rxqs, M_IFLIB);
5388 	ctx->ifc_rxqs = NULL;
5389 	if (ctx->ifc_txqs != NULL)
5390 		free(ctx->ifc_txqs, M_IFLIB);
5391 	ctx->ifc_txqs = NULL;
5392 fail:
5393 	return (err);
5394 }
5395 
5396 static int
5397 iflib_tx_structures_setup(if_ctx_t ctx)
5398 {
5399 	iflib_txq_t txq = ctx->ifc_txqs;
5400 	int i;
5401 
5402 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
5403 		iflib_txq_setup(txq);
5404 
5405 	return (0);
5406 }
5407 
5408 static void
5409 iflib_tx_structures_free(if_ctx_t ctx)
5410 {
5411 	iflib_txq_t txq = ctx->ifc_txqs;
5412 	int i, j;
5413 
5414 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
5415 		iflib_txq_destroy(txq);
5416 		for (j = 0; j < ctx->ifc_nhwtxqs; j++)
5417 			iflib_dma_free(&txq->ift_ifdi[j]);
5418 	}
5419 	free(ctx->ifc_txqs, M_IFLIB);
5420 	ctx->ifc_txqs = NULL;
5421 	IFDI_QUEUES_FREE(ctx);
5422 }
5423 
5424 /*********************************************************************
5425  *
5426  *  Initialize all receive rings.
5427  *
5428  **********************************************************************/
5429 static int
5430 iflib_rx_structures_setup(if_ctx_t ctx)
5431 {
5432 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5433 	int q;
5434 #if defined(INET6) || defined(INET)
5435 	int i, err;
5436 #endif
5437 
5438 	for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) {
5439 #if defined(INET6) || defined(INET)
5440 		tcp_lro_free(&rxq->ifr_lc);
5441 		if ((err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp,
5442 		    TCP_LRO_ENTRIES, min(1024,
5443 		    ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset]))) != 0) {
5444 			device_printf(ctx->ifc_dev, "LRO Initialization failed!\n");
5445 			goto fail;
5446 		}
5447 		rxq->ifr_lro_enabled = TRUE;
5448 #endif
5449 		IFDI_RXQ_SETUP(ctx, rxq->ifr_id);
5450 	}
5451 	return (0);
5452 #if defined(INET6) || defined(INET)
5453 fail:
5454 	/*
5455 	 * Free RX software descriptors allocated so far, we will only handle
5456 	 * the rings that completed, the failing case will have
5457 	 * cleaned up for itself. 'q' failed, so its the terminus.
5458 	 */
5459 	rxq = ctx->ifc_rxqs;
5460 	for (i = 0; i < q; ++i, rxq++) {
5461 		iflib_rx_sds_free(rxq);
5462 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
5463 	}
5464 	return (err);
5465 #endif
5466 }
5467 
5468 /*********************************************************************
5469  *
5470  *  Free all receive rings.
5471  *
5472  **********************************************************************/
5473 static void
5474 iflib_rx_structures_free(if_ctx_t ctx)
5475 {
5476 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5477 
5478 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) {
5479 		iflib_rx_sds_free(rxq);
5480 	}
5481 }
5482 
5483 static int
5484 iflib_qset_structures_setup(if_ctx_t ctx)
5485 {
5486 	int err;
5487 
5488 	/*
5489 	 * It is expected that the caller takes care of freeing queues if this
5490 	 * fails.
5491 	 */
5492 	if ((err = iflib_tx_structures_setup(ctx)) != 0) {
5493 		device_printf(ctx->ifc_dev, "iflib_tx_structures_setup failed: %d\n", err);
5494 		return (err);
5495 	}
5496 
5497 	if ((err = iflib_rx_structures_setup(ctx)) != 0)
5498 		device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err);
5499 
5500 	return (err);
5501 }
5502 
5503 int
5504 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
5505 		driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, const char *name)
5506 {
5507 
5508 	return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name));
5509 }
5510 
5511 #ifdef SMP
5512 static int
5513 find_nth(if_ctx_t ctx, int qid)
5514 {
5515 	cpuset_t cpus;
5516 	int i, cpuid, eqid, count;
5517 
5518 	CPU_COPY(&ctx->ifc_cpus, &cpus);
5519 	count = CPU_COUNT(&cpus);
5520 	eqid = qid % count;
5521 	/* clear up to the qid'th bit */
5522 	for (i = 0; i < eqid; i++) {
5523 		cpuid = CPU_FFS(&cpus);
5524 		MPASS(cpuid != 0);
5525 		CPU_CLR(cpuid-1, &cpus);
5526 	}
5527 	cpuid = CPU_FFS(&cpus);
5528 	MPASS(cpuid != 0);
5529 	return (cpuid-1);
5530 }
5531 
5532 #ifdef SCHED_ULE
5533 extern struct cpu_group *cpu_top;              /* CPU topology */
5534 
5535 static int
5536 find_child_with_core(int cpu, struct cpu_group *grp)
5537 {
5538 	int i;
5539 
5540 	if (grp->cg_children == 0)
5541 		return -1;
5542 
5543 	MPASS(grp->cg_child);
5544 	for (i = 0; i < grp->cg_children; i++) {
5545 		if (CPU_ISSET(cpu, &grp->cg_child[i].cg_mask))
5546 			return i;
5547 	}
5548 
5549 	return -1;
5550 }
5551 
5552 /*
5553  * Find the nth "close" core to the specified core
5554  * "close" is defined as the deepest level that shares
5555  * at least an L2 cache.  With threads, this will be
5556  * threads on the same core.  If the sahred cache is L3
5557  * or higher, simply returns the same core.
5558  */
5559 static int
5560 find_close_core(int cpu, int core_offset)
5561 {
5562 	struct cpu_group *grp;
5563 	int i;
5564 	int fcpu;
5565 	cpuset_t cs;
5566 
5567 	grp = cpu_top;
5568 	if (grp == NULL)
5569 		return cpu;
5570 	i = 0;
5571 	while ((i = find_child_with_core(cpu, grp)) != -1) {
5572 		/* If the child only has one cpu, don't descend */
5573 		if (grp->cg_child[i].cg_count <= 1)
5574 			break;
5575 		grp = &grp->cg_child[i];
5576 	}
5577 
5578 	/* If they don't share at least an L2 cache, use the same CPU */
5579 	if (grp->cg_level > CG_SHARE_L2 || grp->cg_level == CG_SHARE_NONE)
5580 		return cpu;
5581 
5582 	/* Now pick one */
5583 	CPU_COPY(&grp->cg_mask, &cs);
5584 
5585 	/* Add the selected CPU offset to core offset. */
5586 	for (i = 0; (fcpu = CPU_FFS(&cs)) != 0; i++) {
5587 		if (fcpu - 1 == cpu)
5588 			break;
5589 		CPU_CLR(fcpu - 1, &cs);
5590 	}
5591 	MPASS(fcpu);
5592 
5593 	core_offset += i;
5594 
5595 	CPU_COPY(&grp->cg_mask, &cs);
5596 	for (i = core_offset % grp->cg_count; i > 0; i--) {
5597 		MPASS(CPU_FFS(&cs));
5598 		CPU_CLR(CPU_FFS(&cs) - 1, &cs);
5599 	}
5600 	MPASS(CPU_FFS(&cs));
5601 	return CPU_FFS(&cs) - 1;
5602 }
5603 #else
5604 static int
5605 find_close_core(int cpu, int core_offset __unused)
5606 {
5607 	return cpu;
5608 }
5609 #endif
5610 
5611 static int
5612 get_core_offset(if_ctx_t ctx, iflib_intr_type_t type, int qid)
5613 {
5614 	switch (type) {
5615 	case IFLIB_INTR_TX:
5616 		/* TX queues get cores which share at least an L2 cache with the corresponding RX queue */
5617 		/* XXX handle multiple RX threads per core and more than two core per L2 group */
5618 		return qid / CPU_COUNT(&ctx->ifc_cpus) + 1;
5619 	case IFLIB_INTR_RX:
5620 	case IFLIB_INTR_RXTX:
5621 		/* RX queues get the specified core */
5622 		return qid / CPU_COUNT(&ctx->ifc_cpus);
5623 	default:
5624 		return -1;
5625 	}
5626 }
5627 #else
5628 #define get_core_offset(ctx, type, qid)	CPU_FIRST()
5629 #define find_close_core(cpuid, tid)	CPU_FIRST()
5630 #define find_nth(ctx, gid)		CPU_FIRST()
5631 #endif
5632 
5633 /* Just to avoid copy/paste */
5634 static inline int
5635 iflib_irq_set_affinity(if_ctx_t ctx, int irq, iflib_intr_type_t type, int qid,
5636     struct grouptask *gtask, struct taskqgroup *tqg, void *uniq, const char *name)
5637 {
5638 	int cpuid;
5639 	int err, tid;
5640 
5641 	cpuid = find_nth(ctx, qid);
5642 	tid = get_core_offset(ctx, type, qid);
5643 	MPASS(tid >= 0);
5644 	cpuid = find_close_core(cpuid, tid);
5645 	err = taskqgroup_attach_cpu(tqg, gtask, uniq, cpuid, irq, name);
5646 	if (err) {
5647 		device_printf(ctx->ifc_dev, "taskqgroup_attach_cpu failed %d\n", err);
5648 		return (err);
5649 	}
5650 #ifdef notyet
5651 	if (cpuid > ctx->ifc_cpuid_highest)
5652 		ctx->ifc_cpuid_highest = cpuid;
5653 #endif
5654 	return 0;
5655 }
5656 
5657 int
5658 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid,
5659 			iflib_intr_type_t type, driver_filter_t *filter,
5660 			void *filter_arg, int qid, const char *name)
5661 {
5662 	struct grouptask *gtask;
5663 	struct taskqgroup *tqg;
5664 	iflib_filter_info_t info;
5665 	gtask_fn_t *fn;
5666 	int tqrid, err;
5667 	driver_filter_t *intr_fast;
5668 	void *q;
5669 
5670 	info = &ctx->ifc_filter_info;
5671 	tqrid = rid;
5672 
5673 	switch (type) {
5674 	/* XXX merge tx/rx for netmap? */
5675 	case IFLIB_INTR_TX:
5676 		q = &ctx->ifc_txqs[qid];
5677 		info = &ctx->ifc_txqs[qid].ift_filter_info;
5678 		gtask = &ctx->ifc_txqs[qid].ift_task;
5679 		tqg = qgroup_if_io_tqg;
5680 		fn = _task_fn_tx;
5681 		intr_fast = iflib_fast_intr;
5682 		GROUPTASK_INIT(gtask, 0, fn, q);
5683 		ctx->ifc_flags |= IFC_NETMAP_TX_IRQ;
5684 		break;
5685 	case IFLIB_INTR_RX:
5686 		q = &ctx->ifc_rxqs[qid];
5687 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
5688 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5689 		tqg = qgroup_if_io_tqg;
5690 		fn = _task_fn_rx;
5691 		intr_fast = iflib_fast_intr;
5692 		GROUPTASK_INIT(gtask, 0, fn, q);
5693 		break;
5694 	case IFLIB_INTR_RXTX:
5695 		q = &ctx->ifc_rxqs[qid];
5696 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
5697 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5698 		tqg = qgroup_if_io_tqg;
5699 		fn = _task_fn_rx;
5700 		intr_fast = iflib_fast_intr_rxtx;
5701 		GROUPTASK_INIT(gtask, 0, fn, q);
5702 		break;
5703 	case IFLIB_INTR_ADMIN:
5704 		q = ctx;
5705 		tqrid = -1;
5706 		info = &ctx->ifc_filter_info;
5707 		gtask = &ctx->ifc_admin_task;
5708 		tqg = qgroup_if_config_tqg;
5709 		fn = _task_fn_admin;
5710 		intr_fast = iflib_fast_intr_ctx;
5711 		break;
5712 	default:
5713 		panic("unknown net intr type");
5714 	}
5715 
5716 	info->ifi_filter = filter;
5717 	info->ifi_filter_arg = filter_arg;
5718 	info->ifi_task = gtask;
5719 	info->ifi_ctx = q;
5720 
5721 	err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info,  name);
5722 	if (err != 0) {
5723 		device_printf(ctx->ifc_dev, "_iflib_irq_alloc failed %d\n", err);
5724 		return (err);
5725 	}
5726 	if (type == IFLIB_INTR_ADMIN)
5727 		return (0);
5728 
5729 	if (tqrid != -1) {
5730 		err = iflib_irq_set_affinity(ctx, rman_get_start(irq->ii_res), type, qid, gtask, tqg, q, name);
5731 		if (err)
5732 			return (err);
5733 	} else {
5734 		taskqgroup_attach(tqg, gtask, q, rman_get_start(irq->ii_res), name);
5735 	}
5736 
5737 	return (0);
5738 }
5739 
5740 void
5741 iflib_softirq_alloc_generic(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type,  void *arg, int qid, const char *name)
5742 {
5743 	struct grouptask *gtask;
5744 	struct taskqgroup *tqg;
5745 	gtask_fn_t *fn;
5746 	void *q;
5747 	int irq_num = -1;
5748 	int err;
5749 
5750 	switch (type) {
5751 	case IFLIB_INTR_TX:
5752 		q = &ctx->ifc_txqs[qid];
5753 		gtask = &ctx->ifc_txqs[qid].ift_task;
5754 		tqg = qgroup_if_io_tqg;
5755 		fn = _task_fn_tx;
5756 		if (irq != NULL)
5757 			irq_num = rman_get_start(irq->ii_res);
5758 		break;
5759 	case IFLIB_INTR_RX:
5760 		q = &ctx->ifc_rxqs[qid];
5761 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5762 		tqg = qgroup_if_io_tqg;
5763 		fn = _task_fn_rx;
5764 		if (irq != NULL)
5765 			irq_num = rman_get_start(irq->ii_res);
5766 		break;
5767 	case IFLIB_INTR_IOV:
5768 		q = ctx;
5769 		gtask = &ctx->ifc_vflr_task;
5770 		tqg = qgroup_if_config_tqg;
5771 		fn = _task_fn_iov;
5772 		break;
5773 	default:
5774 		panic("unknown net intr type");
5775 	}
5776 	GROUPTASK_INIT(gtask, 0, fn, q);
5777 	if (irq_num != -1) {
5778 		err = iflib_irq_set_affinity(ctx, irq_num, type, qid, gtask, tqg, q, name);
5779 		if (err)
5780 			taskqgroup_attach(tqg, gtask, q, irq_num, name);
5781 	}
5782 	else {
5783 		taskqgroup_attach(tqg, gtask, q, irq_num, name);
5784 	}
5785 }
5786 
5787 void
5788 iflib_irq_free(if_ctx_t ctx, if_irq_t irq)
5789 {
5790 	if (irq->ii_tag)
5791 		bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag);
5792 
5793 	if (irq->ii_res)
5794 		bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ, irq->ii_rid, irq->ii_res);
5795 }
5796 
5797 static int
5798 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, const char *name)
5799 {
5800 	iflib_txq_t txq = ctx->ifc_txqs;
5801 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5802 	if_irq_t irq = &ctx->ifc_legacy_irq;
5803 	iflib_filter_info_t info;
5804 	struct grouptask *gtask;
5805 	struct taskqgroup *tqg;
5806 	gtask_fn_t *fn;
5807 	int tqrid;
5808 	void *q;
5809 	int err;
5810 
5811 	q = &ctx->ifc_rxqs[0];
5812 	info = &rxq[0].ifr_filter_info;
5813 	gtask = &rxq[0].ifr_task;
5814 	tqg = qgroup_if_io_tqg;
5815 	tqrid = irq->ii_rid = *rid;
5816 	fn = _task_fn_rx;
5817 
5818 	ctx->ifc_flags |= IFC_LEGACY;
5819 	info->ifi_filter = filter;
5820 	info->ifi_filter_arg = filter_arg;
5821 	info->ifi_task = gtask;
5822 	info->ifi_ctx = ctx;
5823 
5824 	/* We allocate a single interrupt resource */
5825 	if ((err = _iflib_irq_alloc(ctx, irq, tqrid, iflib_fast_intr_ctx, NULL, info, name)) != 0)
5826 		return (err);
5827 	GROUPTASK_INIT(gtask, 0, fn, q);
5828 	taskqgroup_attach(tqg, gtask, q, rman_get_start(irq->ii_res), name);
5829 
5830 	GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq);
5831 	taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, rman_get_start(irq->ii_res), "tx");
5832 	return (0);
5833 }
5834 
5835 void
5836 iflib_led_create(if_ctx_t ctx)
5837 {
5838 
5839 	ctx->ifc_led_dev = led_create(iflib_led_func, ctx,
5840 	    device_get_nameunit(ctx->ifc_dev));
5841 }
5842 
5843 void
5844 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid)
5845 {
5846 
5847 	GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
5848 }
5849 
5850 void
5851 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid)
5852 {
5853 
5854 	GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task);
5855 }
5856 
5857 void
5858 iflib_admin_intr_deferred(if_ctx_t ctx)
5859 {
5860 #ifdef INVARIANTS
5861 	struct grouptask *gtask;
5862 
5863 	gtask = &ctx->ifc_admin_task;
5864 	MPASS(gtask != NULL && gtask->gt_taskqueue != NULL);
5865 #endif
5866 
5867 	GROUPTASK_ENQUEUE(&ctx->ifc_admin_task);
5868 }
5869 
5870 void
5871 iflib_iov_intr_deferred(if_ctx_t ctx)
5872 {
5873 
5874 	GROUPTASK_ENQUEUE(&ctx->ifc_vflr_task);
5875 }
5876 
5877 void
5878 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, char *name)
5879 {
5880 
5881 	taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, -1, name);
5882 }
5883 
5884 void
5885 iflib_config_gtask_init(void *ctx, struct grouptask *gtask, gtask_fn_t *fn,
5886 	const char *name)
5887 {
5888 
5889 	GROUPTASK_INIT(gtask, 0, fn, ctx);
5890 	taskqgroup_attach(qgroup_if_config_tqg, gtask, gtask, -1, name);
5891 }
5892 
5893 void
5894 iflib_config_gtask_deinit(struct grouptask *gtask)
5895 {
5896 
5897 	taskqgroup_detach(qgroup_if_config_tqg, gtask);
5898 }
5899 
5900 void
5901 iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate)
5902 {
5903 	if_t ifp = ctx->ifc_ifp;
5904 	iflib_txq_t txq = ctx->ifc_txqs;
5905 
5906 	if_setbaudrate(ifp, baudrate);
5907 	if (baudrate >= IF_Gbps(10)) {
5908 		STATE_LOCK(ctx);
5909 		ctx->ifc_flags |= IFC_PREFETCH;
5910 		STATE_UNLOCK(ctx);
5911 	}
5912 	/* If link down, disable watchdog */
5913 	if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) {
5914 		for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++)
5915 			txq->ift_qstatus = IFLIB_QUEUE_IDLE;
5916 	}
5917 	ctx->ifc_link_state = link_state;
5918 	if_link_state_change(ifp, link_state);
5919 }
5920 
5921 static int
5922 iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq)
5923 {
5924 	int credits;
5925 #ifdef INVARIANTS
5926 	int credits_pre = txq->ift_cidx_processed;
5927 #endif
5928 
5929 	if (ctx->isc_txd_credits_update == NULL)
5930 		return (0);
5931 
5932 	if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, true)) == 0)
5933 		return (0);
5934 
5935 	txq->ift_processed += credits;
5936 	txq->ift_cidx_processed += credits;
5937 
5938 	MPASS(credits_pre + credits == txq->ift_cidx_processed);
5939 	if (txq->ift_cidx_processed >= txq->ift_size)
5940 		txq->ift_cidx_processed -= txq->ift_size;
5941 	return (credits);
5942 }
5943 
5944 static int
5945 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget)
5946 {
5947 
5948 	return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx,
5949 	    budget));
5950 }
5951 
5952 void
5953 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name,
5954 	const char *description, if_int_delay_info_t info,
5955 	int offset, int value)
5956 {
5957 	info->iidi_ctx = ctx;
5958 	info->iidi_offset = offset;
5959 	info->iidi_value = value;
5960 	SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev),
5961 	    SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)),
5962 	    OID_AUTO, name, CTLTYPE_INT|CTLFLAG_RW,
5963 	    info, 0, iflib_sysctl_int_delay, "I", description);
5964 }
5965 
5966 struct sx *
5967 iflib_ctx_lock_get(if_ctx_t ctx)
5968 {
5969 
5970 	return (&ctx->ifc_ctx_sx);
5971 }
5972 
5973 static int
5974 iflib_msix_init(if_ctx_t ctx)
5975 {
5976 	device_t dev = ctx->ifc_dev;
5977 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5978 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5979 	int vectors, queues, rx_queues, tx_queues, queuemsgs, msgs;
5980 	int iflib_num_tx_queues, iflib_num_rx_queues;
5981 	int err, admincnt, bar;
5982 
5983 	iflib_num_tx_queues = ctx->ifc_sysctl_ntxqs;
5984 	iflib_num_rx_queues = ctx->ifc_sysctl_nrxqs;
5985 
5986 	device_printf(dev, "msix_init qsets capped at %d\n", imax(scctx->isc_ntxqsets, scctx->isc_nrxqsets));
5987 
5988 	bar = ctx->ifc_softc_ctx.isc_msix_bar;
5989 	admincnt = sctx->isc_admin_intrcnt;
5990 	/* Override by tuneable */
5991 	if (scctx->isc_disable_msix)
5992 		goto msi;
5993 
5994 	/*
5995 	 * bar == -1 => "trust me I know what I'm doing"
5996 	 * Some drivers are for hardware that is so shoddily
5997 	 * documented that no one knows which bars are which
5998 	 * so the developer has to map all bars. This hack
5999 	 * allows shoddy garbage to use msix in this framework.
6000 	 */
6001 	if (bar != -1) {
6002 		ctx->ifc_msix_mem = bus_alloc_resource_any(dev,
6003 	            SYS_RES_MEMORY, &bar, RF_ACTIVE);
6004 		if (ctx->ifc_msix_mem == NULL) {
6005 			/* May not be enabled */
6006 			device_printf(dev, "Unable to map MSIX table \n");
6007 			goto msi;
6008 		}
6009 	}
6010 	/* First try MSI/X */
6011 	if ((msgs = pci_msix_count(dev)) == 0) { /* system has msix disabled */
6012 		device_printf(dev, "System has MSIX disabled \n");
6013 		bus_release_resource(dev, SYS_RES_MEMORY,
6014 		    bar, ctx->ifc_msix_mem);
6015 		ctx->ifc_msix_mem = NULL;
6016 		goto msi;
6017 	}
6018 #if IFLIB_DEBUG
6019 	/* use only 1 qset in debug mode */
6020 	queuemsgs = min(msgs - admincnt, 1);
6021 #else
6022 	queuemsgs = msgs - admincnt;
6023 #endif
6024 #ifdef RSS
6025 	queues = imin(queuemsgs, rss_getnumbuckets());
6026 #else
6027 	queues = queuemsgs;
6028 #endif
6029 	queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues);
6030 	device_printf(dev, "pxm cpus: %d queue msgs: %d admincnt: %d\n",
6031 				  CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt);
6032 #ifdef  RSS
6033 	/* If we're doing RSS, clamp at the number of RSS buckets */
6034 	if (queues > rss_getnumbuckets())
6035 		queues = rss_getnumbuckets();
6036 #endif
6037 	if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt)
6038 		rx_queues = iflib_num_rx_queues;
6039 	else
6040 		rx_queues = queues;
6041 
6042 	if (rx_queues > scctx->isc_nrxqsets)
6043 		rx_queues = scctx->isc_nrxqsets;
6044 
6045 	/*
6046 	 * We want this to be all logical CPUs by default
6047 	 */
6048 	if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues)
6049 		tx_queues = iflib_num_tx_queues;
6050 	else
6051 		tx_queues = mp_ncpus;
6052 
6053 	if (tx_queues > scctx->isc_ntxqsets)
6054 		tx_queues = scctx->isc_ntxqsets;
6055 
6056 	if (ctx->ifc_sysctl_qs_eq_override == 0) {
6057 #ifdef INVARIANTS
6058 		if (tx_queues != rx_queues)
6059 			device_printf(dev, "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n",
6060 				      min(rx_queues, tx_queues), min(rx_queues, tx_queues));
6061 #endif
6062 		tx_queues = min(rx_queues, tx_queues);
6063 		rx_queues = min(rx_queues, tx_queues);
6064 	}
6065 
6066 	device_printf(dev, "using %d rx queues %d tx queues \n", rx_queues, tx_queues);
6067 
6068 	vectors = rx_queues + admincnt;
6069 	if ((err = pci_alloc_msix(dev, &vectors)) == 0) {
6070 		device_printf(dev,
6071 					  "Using MSIX interrupts with %d vectors\n", vectors);
6072 		scctx->isc_vectors = vectors;
6073 		scctx->isc_nrxqsets = rx_queues;
6074 		scctx->isc_ntxqsets = tx_queues;
6075 		scctx->isc_intr = IFLIB_INTR_MSIX;
6076 
6077 		return (vectors);
6078 	} else {
6079 		device_printf(dev, "failed to allocate %d msix vectors, err: %d - using MSI\n", vectors, err);
6080 		bus_release_resource(dev, SYS_RES_MEMORY, bar,
6081 		    ctx->ifc_msix_mem);
6082 		ctx->ifc_msix_mem = NULL;
6083 	}
6084 msi:
6085 	vectors = pci_msi_count(dev);
6086 	scctx->isc_nrxqsets = 1;
6087 	scctx->isc_ntxqsets = 1;
6088 	scctx->isc_vectors = vectors;
6089 	if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) {
6090 		device_printf(dev,"Using an MSI interrupt\n");
6091 		scctx->isc_intr = IFLIB_INTR_MSI;
6092 	} else {
6093 		scctx->isc_vectors = 1;
6094 		device_printf(dev,"Using a Legacy interrupt\n");
6095 		scctx->isc_intr = IFLIB_INTR_LEGACY;
6096 	}
6097 
6098 	return (vectors);
6099 }
6100 
6101 static const char *ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" };
6102 
6103 static int
6104 mp_ring_state_handler(SYSCTL_HANDLER_ARGS)
6105 {
6106 	int rc;
6107 	uint16_t *state = ((uint16_t *)oidp->oid_arg1);
6108 	struct sbuf *sb;
6109 	const char *ring_state = "UNKNOWN";
6110 
6111 	/* XXX needed ? */
6112 	rc = sysctl_wire_old_buffer(req, 0);
6113 	MPASS(rc == 0);
6114 	if (rc != 0)
6115 		return (rc);
6116 	sb = sbuf_new_for_sysctl(NULL, NULL, 80, req);
6117 	MPASS(sb != NULL);
6118 	if (sb == NULL)
6119 		return (ENOMEM);
6120 	if (state[3] <= 3)
6121 		ring_state = ring_states[state[3]];
6122 
6123 	sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s",
6124 		    state[0], state[1], state[2], ring_state);
6125 	rc = sbuf_finish(sb);
6126 	sbuf_delete(sb);
6127         return(rc);
6128 }
6129 
6130 enum iflib_ndesc_handler {
6131 	IFLIB_NTXD_HANDLER,
6132 	IFLIB_NRXD_HANDLER,
6133 };
6134 
6135 static int
6136 mp_ndesc_handler(SYSCTL_HANDLER_ARGS)
6137 {
6138 	if_ctx_t ctx = (void *)arg1;
6139 	enum iflib_ndesc_handler type = arg2;
6140 	char buf[256] = {0};
6141 	qidx_t *ndesc;
6142 	char *p, *next;
6143 	int nqs, rc, i;
6144 
6145 	MPASS(type == IFLIB_NTXD_HANDLER || type == IFLIB_NRXD_HANDLER);
6146 
6147 	nqs = 8;
6148 	switch(type) {
6149 	case IFLIB_NTXD_HANDLER:
6150 		ndesc = ctx->ifc_sysctl_ntxds;
6151 		if (ctx->ifc_sctx)
6152 			nqs = ctx->ifc_sctx->isc_ntxqs;
6153 		break;
6154 	case IFLIB_NRXD_HANDLER:
6155 		ndesc = ctx->ifc_sysctl_nrxds;
6156 		if (ctx->ifc_sctx)
6157 			nqs = ctx->ifc_sctx->isc_nrxqs;
6158 		break;
6159 	default:
6160 			panic("unhandled type");
6161 	}
6162 	if (nqs == 0)
6163 		nqs = 8;
6164 
6165 	for (i=0; i<8; i++) {
6166 		if (i >= nqs)
6167 			break;
6168 		if (i)
6169 			strcat(buf, ",");
6170 		sprintf(strchr(buf, 0), "%d", ndesc[i]);
6171 	}
6172 
6173 	rc = sysctl_handle_string(oidp, buf, sizeof(buf), req);
6174 	if (rc || req->newptr == NULL)
6175 		return rc;
6176 
6177 	for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p;
6178 	    i++, p = strsep(&next, " ,")) {
6179 		ndesc[i] = strtoul(p, NULL, 10);
6180 	}
6181 
6182 	return(rc);
6183 }
6184 
6185 #define NAME_BUFLEN 32
6186 static void
6187 iflib_add_device_sysctl_pre(if_ctx_t ctx)
6188 {
6189         device_t dev = iflib_get_dev(ctx);
6190 	struct sysctl_oid_list *child, *oid_list;
6191 	struct sysctl_ctx_list *ctx_list;
6192 	struct sysctl_oid *node;
6193 
6194 	ctx_list = device_get_sysctl_ctx(dev);
6195 	child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev));
6196 	ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, "iflib",
6197 						      CTLFLAG_RD, NULL, "IFLIB fields");
6198 	oid_list = SYSCTL_CHILDREN(node);
6199 
6200 	SYSCTL_ADD_STRING(ctx_list, oid_list, OID_AUTO, "driver_version",
6201 		       CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version, 0,
6202 		       "driver version");
6203 
6204 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs",
6205 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0,
6206 			"# of txqs to use, 0 => use default #");
6207 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs",
6208 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0,
6209 			"# of rxqs to use, 0 => use default #");
6210 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_qs_enable",
6211 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0,
6212                        "permit #txq != #rxq");
6213 	SYSCTL_ADD_INT(ctx_list, oid_list, OID_AUTO, "disable_msix",
6214                       CTLFLAG_RWTUN, &ctx->ifc_softc_ctx.isc_disable_msix, 0,
6215                       "disable MSIX (default 0)");
6216 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "rx_budget",
6217 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_rx_budget, 0,
6218                        "set the rx budget");
6219 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "tx_abdicate",
6220 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_tx_abdicate, 0,
6221 		       "cause tx to abdicate instead of running to completion");
6222 
6223 	/* XXX change for per-queue sizes */
6224 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_ntxds",
6225 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NTXD_HANDLER,
6226                        mp_ndesc_handler, "A",
6227                        "list of # of tx descriptors to use, 0 = use default #");
6228 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_nrxds",
6229 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NRXD_HANDLER,
6230                        mp_ndesc_handler, "A",
6231                        "list of # of rx descriptors to use, 0 = use default #");
6232 }
6233 
6234 static void
6235 iflib_add_device_sysctl_post(if_ctx_t ctx)
6236 {
6237 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6238 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6239         device_t dev = iflib_get_dev(ctx);
6240 	struct sysctl_oid_list *child;
6241 	struct sysctl_ctx_list *ctx_list;
6242 	iflib_fl_t fl;
6243 	iflib_txq_t txq;
6244 	iflib_rxq_t rxq;
6245 	int i, j;
6246 	char namebuf[NAME_BUFLEN];
6247 	char *qfmt;
6248 	struct sysctl_oid *queue_node, *fl_node, *node;
6249 	struct sysctl_oid_list *queue_list, *fl_list;
6250 	ctx_list = device_get_sysctl_ctx(dev);
6251 
6252 	node = ctx->ifc_sysctl_node;
6253 	child = SYSCTL_CHILDREN(node);
6254 
6255 	if (scctx->isc_ntxqsets > 100)
6256 		qfmt = "txq%03d";
6257 	else if (scctx->isc_ntxqsets > 10)
6258 		qfmt = "txq%02d";
6259 	else
6260 		qfmt = "txq%d";
6261 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
6262 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6263 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6264 					     CTLFLAG_RD, NULL, "Queue Name");
6265 		queue_list = SYSCTL_CHILDREN(queue_node);
6266 #if MEMORY_LOGGING
6267 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued",
6268 				CTLFLAG_RD,
6269 				&txq->ift_dequeued, "total mbufs freed");
6270 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued",
6271 				CTLFLAG_RD,
6272 				&txq->ift_enqueued, "total mbufs enqueued");
6273 #endif
6274 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag",
6275 				   CTLFLAG_RD,
6276 				   &txq->ift_mbuf_defrag, "# of times m_defrag was called");
6277 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "m_pullups",
6278 				   CTLFLAG_RD,
6279 				   &txq->ift_pullups, "# of times m_pullup was called");
6280 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag_failed",
6281 				   CTLFLAG_RD,
6282 				   &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed");
6283 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_desc_avail",
6284 				   CTLFLAG_RD,
6285 				   &txq->ift_no_desc_avail, "# of times no descriptors were available");
6286 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "tx_map_failed",
6287 				   CTLFLAG_RD,
6288 				   &txq->ift_map_failed, "# of times dma map failed");
6289 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txd_encap_efbig",
6290 				   CTLFLAG_RD,
6291 				   &txq->ift_txd_encap_efbig, "# of times txd_encap returned EFBIG");
6292 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_tx_dma_setup",
6293 				   CTLFLAG_RD,
6294 				   &txq->ift_no_tx_dma_setup, "# of times map failed for other than EFBIG");
6295 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx",
6296 				   CTLFLAG_RD,
6297 				   &txq->ift_pidx, 1, "Producer Index");
6298 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx",
6299 				   CTLFLAG_RD,
6300 				   &txq->ift_cidx, 1, "Consumer Index");
6301 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx_processed",
6302 				   CTLFLAG_RD,
6303 				   &txq->ift_cidx_processed, 1, "Consumer Index seen by credit update");
6304 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use",
6305 				   CTLFLAG_RD,
6306 				   &txq->ift_in_use, 1, "descriptors in use");
6307 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_processed",
6308 				   CTLFLAG_RD,
6309 				   &txq->ift_processed, "descriptors procesed for clean");
6310 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned",
6311 				   CTLFLAG_RD,
6312 				   &txq->ift_cleaned, "total cleaned");
6313 		SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state",
6314 				CTLTYPE_STRING | CTLFLAG_RD, __DEVOLATILE(uint64_t *, &txq->ift_br->state),
6315 				0, mp_ring_state_handler, "A", "soft ring state");
6316 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_enqueues",
6317 				       CTLFLAG_RD, &txq->ift_br->enqueues,
6318 				       "# of enqueues to the mp_ring for this queue");
6319 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_drops",
6320 				       CTLFLAG_RD, &txq->ift_br->drops,
6321 				       "# of drops in the mp_ring for this queue");
6322 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_starts",
6323 				       CTLFLAG_RD, &txq->ift_br->starts,
6324 				       "# of normal consumer starts in the mp_ring for this queue");
6325 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_stalls",
6326 				       CTLFLAG_RD, &txq->ift_br->stalls,
6327 					       "# of consumer stalls in the mp_ring for this queue");
6328 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_restarts",
6329 			       CTLFLAG_RD, &txq->ift_br->restarts,
6330 				       "# of consumer restarts in the mp_ring for this queue");
6331 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_abdications",
6332 				       CTLFLAG_RD, &txq->ift_br->abdications,
6333 				       "# of consumer abdications in the mp_ring for this queue");
6334 	}
6335 
6336 	if (scctx->isc_nrxqsets > 100)
6337 		qfmt = "rxq%03d";
6338 	else if (scctx->isc_nrxqsets > 10)
6339 		qfmt = "rxq%02d";
6340 	else
6341 		qfmt = "rxq%d";
6342 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
6343 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6344 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6345 					     CTLFLAG_RD, NULL, "Queue Name");
6346 		queue_list = SYSCTL_CHILDREN(queue_node);
6347 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
6348 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_pidx",
6349 				       CTLFLAG_RD,
6350 				       &rxq->ifr_cq_pidx, 1, "Producer Index");
6351 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_cidx",
6352 				       CTLFLAG_RD,
6353 				       &rxq->ifr_cq_cidx, 1, "Consumer Index");
6354 		}
6355 
6356 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
6357 			snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j);
6358 			fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list, OID_AUTO, namebuf,
6359 						     CTLFLAG_RD, NULL, "freelist Name");
6360 			fl_list = SYSCTL_CHILDREN(fl_node);
6361 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx",
6362 				       CTLFLAG_RD,
6363 				       &fl->ifl_pidx, 1, "Producer Index");
6364 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx",
6365 				       CTLFLAG_RD,
6366 				       &fl->ifl_cidx, 1, "Consumer Index");
6367 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits",
6368 				       CTLFLAG_RD,
6369 				       &fl->ifl_credits, 1, "credits available");
6370 #if MEMORY_LOGGING
6371 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_enqueued",
6372 					CTLFLAG_RD,
6373 					&fl->ifl_m_enqueued, "mbufs allocated");
6374 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_dequeued",
6375 					CTLFLAG_RD,
6376 					&fl->ifl_m_dequeued, "mbufs freed");
6377 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_enqueued",
6378 					CTLFLAG_RD,
6379 					&fl->ifl_cl_enqueued, "clusters allocated");
6380 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_dequeued",
6381 					CTLFLAG_RD,
6382 					&fl->ifl_cl_dequeued, "clusters freed");
6383 #endif
6384 
6385 		}
6386 	}
6387 
6388 }
6389 
6390 #ifndef __NO_STRICT_ALIGNMENT
6391 static struct mbuf *
6392 iflib_fixup_rx(struct mbuf *m)
6393 {
6394 	struct mbuf *n;
6395 
6396 	if (m->m_len <= (MCLBYTES - ETHER_HDR_LEN)) {
6397 		bcopy(m->m_data, m->m_data + ETHER_HDR_LEN, m->m_len);
6398 		m->m_data += ETHER_HDR_LEN;
6399 		n = m;
6400 	} else {
6401 		MGETHDR(n, M_NOWAIT, MT_DATA);
6402 		if (n == NULL) {
6403 			m_freem(m);
6404 			return (NULL);
6405 		}
6406 		bcopy(m->m_data, n->m_data, ETHER_HDR_LEN);
6407 		m->m_data += ETHER_HDR_LEN;
6408 		m->m_len -= ETHER_HDR_LEN;
6409 		n->m_len = ETHER_HDR_LEN;
6410 		M_MOVE_PKTHDR(n, m);
6411 		n->m_next = m;
6412 	}
6413 	return (n);
6414 }
6415 #endif
6416 
6417 #ifdef NETDUMP
6418 static void
6419 iflib_netdump_init(struct ifnet *ifp, int *nrxr, int *ncl, int *clsize)
6420 {
6421 	if_ctx_t ctx;
6422 
6423 	ctx = if_getsoftc(ifp);
6424 	CTX_LOCK(ctx);
6425 	*nrxr = NRXQSETS(ctx);
6426 	*ncl = ctx->ifc_rxqs[0].ifr_fl->ifl_size;
6427 	*clsize = ctx->ifc_rxqs[0].ifr_fl->ifl_buf_size;
6428 	CTX_UNLOCK(ctx);
6429 }
6430 
6431 static void
6432 iflib_netdump_event(struct ifnet *ifp, enum netdump_ev event)
6433 {
6434 	if_ctx_t ctx;
6435 	if_softc_ctx_t scctx;
6436 	iflib_fl_t fl;
6437 	iflib_rxq_t rxq;
6438 	int i, j;
6439 
6440 	ctx = if_getsoftc(ifp);
6441 	scctx = &ctx->ifc_softc_ctx;
6442 
6443 	switch (event) {
6444 	case NETDUMP_START:
6445 		for (i = 0; i < scctx->isc_nrxqsets; i++) {
6446 			rxq = &ctx->ifc_rxqs[i];
6447 			for (j = 0; j < rxq->ifr_nfl; j++) {
6448 				fl = rxq->ifr_fl;
6449 				fl->ifl_zone = m_getzone(fl->ifl_buf_size);
6450 			}
6451 		}
6452 		iflib_no_tx_batch = 1;
6453 		break;
6454 	default:
6455 		break;
6456 	}
6457 }
6458 
6459 static int
6460 iflib_netdump_transmit(struct ifnet *ifp, struct mbuf *m)
6461 {
6462 	if_ctx_t ctx;
6463 	iflib_txq_t txq;
6464 	int error;
6465 
6466 	ctx = if_getsoftc(ifp);
6467 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
6468 	    IFF_DRV_RUNNING)
6469 		return (EBUSY);
6470 
6471 	txq = &ctx->ifc_txqs[0];
6472 	error = iflib_encap(txq, &m);
6473 	if (error == 0)
6474 		(void)iflib_txd_db_check(ctx, txq, true, txq->ift_in_use);
6475 	return (error);
6476 }
6477 
6478 static int
6479 iflib_netdump_poll(struct ifnet *ifp, int count)
6480 {
6481 	if_ctx_t ctx;
6482 	if_softc_ctx_t scctx;
6483 	iflib_txq_t txq;
6484 	int i;
6485 
6486 	ctx = if_getsoftc(ifp);
6487 	scctx = &ctx->ifc_softc_ctx;
6488 
6489 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
6490 	    IFF_DRV_RUNNING)
6491 		return (EBUSY);
6492 
6493 	txq = &ctx->ifc_txqs[0];
6494 	(void)iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
6495 
6496 	for (i = 0; i < scctx->isc_nrxqsets; i++)
6497 		(void)iflib_rxeof(&ctx->ifc_rxqs[i], 16 /* XXX */);
6498 	return (0);
6499 }
6500 #endif /* NETDUMP */
6501