xref: /freebsd/sys/net/iflib.c (revision dd7fbcf1936a055e20850a9765ca98cfd2605684)
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 
201 	qidx_t ifc_sysctl_ntxds[8];
202 	qidx_t ifc_sysctl_nrxds[8];
203 	struct if_txrx ifc_txrx;
204 #define isc_txd_encap  ifc_txrx.ift_txd_encap
205 #define isc_txd_flush  ifc_txrx.ift_txd_flush
206 #define isc_txd_credits_update  ifc_txrx.ift_txd_credits_update
207 #define isc_rxd_available ifc_txrx.ift_rxd_available
208 #define isc_rxd_pkt_get ifc_txrx.ift_rxd_pkt_get
209 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
210 #define isc_rxd_flush ifc_txrx.ift_rxd_flush
211 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
212 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
213 #define isc_legacy_intr ifc_txrx.ift_legacy_intr
214 	eventhandler_tag ifc_vlan_attach_event;
215 	eventhandler_tag ifc_vlan_detach_event;
216 	uint8_t ifc_mac[ETHER_ADDR_LEN];
217 	char ifc_mtx_name[16];
218 };
219 
220 
221 void *
222 iflib_get_softc(if_ctx_t ctx)
223 {
224 
225 	return (ctx->ifc_softc);
226 }
227 
228 device_t
229 iflib_get_dev(if_ctx_t ctx)
230 {
231 
232 	return (ctx->ifc_dev);
233 }
234 
235 if_t
236 iflib_get_ifp(if_ctx_t ctx)
237 {
238 
239 	return (ctx->ifc_ifp);
240 }
241 
242 struct ifmedia *
243 iflib_get_media(if_ctx_t ctx)
244 {
245 
246 	return (&ctx->ifc_media);
247 }
248 
249 uint32_t
250 iflib_get_flags(if_ctx_t ctx)
251 {
252 	return (ctx->ifc_flags);
253 }
254 
255 void
256 iflib_set_detach(if_ctx_t ctx)
257 {
258 	ctx->ifc_in_detach = 1;
259 }
260 
261 void
262 iflib_set_mac(if_ctx_t ctx, uint8_t mac[ETHER_ADDR_LEN])
263 {
264 
265 	bcopy(mac, ctx->ifc_mac, ETHER_ADDR_LEN);
266 }
267 
268 if_softc_ctx_t
269 iflib_get_softc_ctx(if_ctx_t ctx)
270 {
271 
272 	return (&ctx->ifc_softc_ctx);
273 }
274 
275 if_shared_ctx_t
276 iflib_get_sctx(if_ctx_t ctx)
277 {
278 
279 	return (ctx->ifc_sctx);
280 }
281 
282 #define IP_ALIGNED(m) ((((uintptr_t)(m)->m_data) & 0x3) == 0x2)
283 #define CACHE_PTR_INCREMENT (CACHE_LINE_SIZE/sizeof(void*))
284 #define CACHE_PTR_NEXT(ptr) ((void *)(((uintptr_t)(ptr)+CACHE_LINE_SIZE-1) & (CACHE_LINE_SIZE-1)))
285 
286 #define LINK_ACTIVE(ctx) ((ctx)->ifc_link_state == LINK_STATE_UP)
287 #define CTX_IS_VF(ctx) ((ctx)->ifc_sctx->isc_flags & IFLIB_IS_VF)
288 
289 #define RX_SW_DESC_MAP_CREATED	(1 << 0)
290 #define TX_SW_DESC_MAP_CREATED	(1 << 1)
291 #define RX_SW_DESC_INUSE        (1 << 3)
292 #define TX_SW_DESC_MAPPED       (1 << 4)
293 
294 #define	M_TOOBIG		M_PROTO1
295 
296 typedef struct iflib_sw_rx_desc_array {
297 	bus_dmamap_t	*ifsd_map;         /* bus_dma maps for packet */
298 	struct mbuf	**ifsd_m;           /* pkthdr mbufs */
299 	caddr_t		*ifsd_cl;          /* direct cluster pointer for rx */
300 	uint8_t		*ifsd_flags;
301 } iflib_rxsd_array_t;
302 
303 typedef struct iflib_sw_tx_desc_array {
304 	bus_dmamap_t    *ifsd_map;         /* bus_dma maps for packet */
305 	struct mbuf    **ifsd_m;           /* pkthdr mbufs */
306 	uint8_t		*ifsd_flags;
307 } if_txsd_vec_t;
308 
309 
310 /* magic number that should be high enough for any hardware */
311 #define IFLIB_MAX_TX_SEGS		128
312 /* bnxt supports 64 with hardware LRO enabled */
313 #define IFLIB_MAX_RX_SEGS		64
314 #define IFLIB_RX_COPY_THRESH		128
315 #define IFLIB_MAX_RX_REFRESH		32
316 /* The minimum descriptors per second before we start coalescing */
317 #define IFLIB_MIN_DESC_SEC		16384
318 #define IFLIB_DEFAULT_TX_UPDATE_FREQ	16
319 #define IFLIB_QUEUE_IDLE		0
320 #define IFLIB_QUEUE_HUNG		1
321 #define IFLIB_QUEUE_WORKING		2
322 /* maximum number of txqs that can share an rx interrupt */
323 #define IFLIB_MAX_TX_SHARED_INTR	4
324 
325 /* this should really scale with ring size - this is a fairly arbitrary value */
326 #define TX_BATCH_SIZE			32
327 
328 #define IFLIB_RESTART_BUDGET		8
329 
330 
331 #define CSUM_OFFLOAD		(CSUM_IP_TSO|CSUM_IP6_TSO|CSUM_IP| \
332 				 CSUM_IP_UDP|CSUM_IP_TCP|CSUM_IP_SCTP| \
333 				 CSUM_IP6_UDP|CSUM_IP6_TCP|CSUM_IP6_SCTP)
334 struct iflib_txq {
335 	qidx_t		ift_in_use;
336 	qidx_t		ift_cidx;
337 	qidx_t		ift_cidx_processed;
338 	qidx_t		ift_pidx;
339 	uint8_t		ift_gen;
340 	uint8_t		ift_br_offset;
341 	uint16_t	ift_npending;
342 	uint16_t	ift_db_pending;
343 	uint16_t	ift_rs_pending;
344 	/* implicit pad */
345 	uint8_t		ift_txd_size[8];
346 	uint64_t	ift_processed;
347 	uint64_t	ift_cleaned;
348 	uint64_t	ift_cleaned_prev;
349 #if MEMORY_LOGGING
350 	uint64_t	ift_enqueued;
351 	uint64_t	ift_dequeued;
352 #endif
353 	uint64_t	ift_no_tx_dma_setup;
354 	uint64_t	ift_no_desc_avail;
355 	uint64_t	ift_mbuf_defrag_failed;
356 	uint64_t	ift_mbuf_defrag;
357 	uint64_t	ift_map_failed;
358 	uint64_t	ift_txd_encap_efbig;
359 	uint64_t	ift_pullups;
360 	uint64_t	ift_last_timer_tick;
361 
362 	struct mtx	ift_mtx;
363 	struct mtx	ift_db_mtx;
364 
365 	/* constant values */
366 	if_ctx_t	ift_ctx;
367 	struct ifmp_ring        *ift_br;
368 	struct grouptask	ift_task;
369 	qidx_t		ift_size;
370 	uint16_t	ift_id;
371 	struct callout	ift_timer;
372 
373 	if_txsd_vec_t	ift_sds;
374 	uint8_t		ift_qstatus;
375 	uint8_t		ift_closed;
376 	uint8_t		ift_update_freq;
377 	struct iflib_filter_info ift_filter_info;
378 	bus_dma_tag_t		ift_desc_tag;
379 	bus_dma_tag_t		ift_tso_desc_tag;
380 	iflib_dma_info_t	ift_ifdi;
381 #define MTX_NAME_LEN 16
382 	char                    ift_mtx_name[MTX_NAME_LEN];
383 	char                    ift_db_mtx_name[MTX_NAME_LEN];
384 	bus_dma_segment_t	ift_segs[IFLIB_MAX_TX_SEGS]  __aligned(CACHE_LINE_SIZE);
385 #ifdef IFLIB_DIAGNOSTICS
386 	uint64_t ift_cpu_exec_count[256];
387 #endif
388 } __aligned(CACHE_LINE_SIZE);
389 
390 struct iflib_fl {
391 	qidx_t		ifl_cidx;
392 	qidx_t		ifl_pidx;
393 	qidx_t		ifl_credits;
394 	uint8_t		ifl_gen;
395 	uint8_t		ifl_rxd_size;
396 #if MEMORY_LOGGING
397 	uint64_t	ifl_m_enqueued;
398 	uint64_t	ifl_m_dequeued;
399 	uint64_t	ifl_cl_enqueued;
400 	uint64_t	ifl_cl_dequeued;
401 #endif
402 	/* implicit pad */
403 
404 	bitstr_t 	*ifl_rx_bitmap;
405 	qidx_t		ifl_fragidx;
406 	/* constant */
407 	qidx_t		ifl_size;
408 	uint16_t	ifl_buf_size;
409 	uint16_t	ifl_cltype;
410 	uma_zone_t	ifl_zone;
411 	iflib_rxsd_array_t	ifl_sds;
412 	iflib_rxq_t	ifl_rxq;
413 	uint8_t		ifl_id;
414 	bus_dma_tag_t           ifl_desc_tag;
415 	iflib_dma_info_t	ifl_ifdi;
416 	uint64_t	ifl_bus_addrs[IFLIB_MAX_RX_REFRESH] __aligned(CACHE_LINE_SIZE);
417 	caddr_t		ifl_vm_addrs[IFLIB_MAX_RX_REFRESH];
418 	qidx_t	ifl_rxd_idxs[IFLIB_MAX_RX_REFRESH];
419 }  __aligned(CACHE_LINE_SIZE);
420 
421 static inline qidx_t
422 get_inuse(int size, qidx_t cidx, qidx_t pidx, uint8_t gen)
423 {
424 	qidx_t used;
425 
426 	if (pidx > cidx)
427 		used = pidx - cidx;
428 	else if (pidx < cidx)
429 		used = size - cidx + pidx;
430 	else if (gen == 0 && pidx == cidx)
431 		used = 0;
432 	else if (gen == 1 && pidx == cidx)
433 		used = size;
434 	else
435 		panic("bad state");
436 
437 	return (used);
438 }
439 
440 #define TXQ_AVAIL(txq) (txq->ift_size - get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx, txq->ift_gen))
441 
442 #define IDXDIFF(head, tail, wrap) \
443 	((head) >= (tail) ? (head) - (tail) : (wrap) - (tail) + (head))
444 
445 struct iflib_rxq {
446 	/* If there is a separate completion queue -
447 	 * these are the cq cidx and pidx. Otherwise
448 	 * these are unused.
449 	 */
450 	qidx_t		ifr_size;
451 	qidx_t		ifr_cq_cidx;
452 	qidx_t		ifr_cq_pidx;
453 	uint8_t		ifr_cq_gen;
454 	uint8_t		ifr_fl_offset;
455 
456 	if_ctx_t	ifr_ctx;
457 	iflib_fl_t	ifr_fl;
458 	uint64_t	ifr_rx_irq;
459 	uint16_t	ifr_id;
460 	uint8_t		ifr_lro_enabled;
461 	uint8_t		ifr_nfl;
462 	uint8_t		ifr_ntxqirq;
463 	uint8_t		ifr_txqid[IFLIB_MAX_TX_SHARED_INTR];
464 	struct lro_ctrl			ifr_lc;
465 	struct grouptask        ifr_task;
466 	struct iflib_filter_info ifr_filter_info;
467 	iflib_dma_info_t		ifr_ifdi;
468 
469 	/* dynamically allocate if any drivers need a value substantially larger than this */
470 	struct if_rxd_frag	ifr_frags[IFLIB_MAX_RX_SEGS] __aligned(CACHE_LINE_SIZE);
471 #ifdef IFLIB_DIAGNOSTICS
472 	uint64_t ifr_cpu_exec_count[256];
473 #endif
474 }  __aligned(CACHE_LINE_SIZE);
475 
476 typedef struct if_rxsd {
477 	caddr_t *ifsd_cl;
478 	struct mbuf **ifsd_m;
479 	iflib_fl_t ifsd_fl;
480 	qidx_t ifsd_cidx;
481 } *if_rxsd_t;
482 
483 /* multiple of word size */
484 #ifdef __LP64__
485 #define PKT_INFO_SIZE	6
486 #define RXD_INFO_SIZE	5
487 #define PKT_TYPE uint64_t
488 #else
489 #define PKT_INFO_SIZE	11
490 #define RXD_INFO_SIZE	8
491 #define PKT_TYPE uint32_t
492 #endif
493 #define PKT_LOOP_BOUND  ((PKT_INFO_SIZE/3)*3)
494 #define RXD_LOOP_BOUND  ((RXD_INFO_SIZE/4)*4)
495 
496 typedef struct if_pkt_info_pad {
497 	PKT_TYPE pkt_val[PKT_INFO_SIZE];
498 } *if_pkt_info_pad_t;
499 typedef struct if_rxd_info_pad {
500 	PKT_TYPE rxd_val[RXD_INFO_SIZE];
501 } *if_rxd_info_pad_t;
502 
503 CTASSERT(sizeof(struct if_pkt_info_pad) == sizeof(struct if_pkt_info));
504 CTASSERT(sizeof(struct if_rxd_info_pad) == sizeof(struct if_rxd_info));
505 
506 
507 static inline void
508 pkt_info_zero(if_pkt_info_t pi)
509 {
510 	if_pkt_info_pad_t pi_pad;
511 
512 	pi_pad = (if_pkt_info_pad_t)pi;
513 	pi_pad->pkt_val[0] = 0; pi_pad->pkt_val[1] = 0; pi_pad->pkt_val[2] = 0;
514 	pi_pad->pkt_val[3] = 0; pi_pad->pkt_val[4] = 0; pi_pad->pkt_val[5] = 0;
515 #ifndef __LP64__
516 	pi_pad->pkt_val[6] = 0; pi_pad->pkt_val[7] = 0; pi_pad->pkt_val[8] = 0;
517 	pi_pad->pkt_val[9] = 0; pi_pad->pkt_val[10] = 0;
518 #endif
519 }
520 
521 static device_method_t iflib_pseudo_methods[] = {
522 	DEVMETHOD(device_attach, noop_attach),
523 	DEVMETHOD(device_detach, iflib_pseudo_detach),
524 	DEVMETHOD_END
525 };
526 
527 driver_t iflib_pseudodriver = {
528 	"iflib_pseudo", iflib_pseudo_methods, sizeof(struct iflib_ctx),
529 };
530 
531 static inline void
532 rxd_info_zero(if_rxd_info_t ri)
533 {
534 	if_rxd_info_pad_t ri_pad;
535 	int i;
536 
537 	ri_pad = (if_rxd_info_pad_t)ri;
538 	for (i = 0; i < RXD_LOOP_BOUND; i += 4) {
539 		ri_pad->rxd_val[i] = 0;
540 		ri_pad->rxd_val[i+1] = 0;
541 		ri_pad->rxd_val[i+2] = 0;
542 		ri_pad->rxd_val[i+3] = 0;
543 	}
544 #ifdef __LP64__
545 	ri_pad->rxd_val[RXD_INFO_SIZE-1] = 0;
546 #endif
547 }
548 
549 /*
550  * Only allow a single packet to take up most 1/nth of the tx ring
551  */
552 #define MAX_SINGLE_PACKET_FRACTION 12
553 #define IF_BAD_DMA (bus_addr_t)-1
554 
555 #define CTX_ACTIVE(ctx) ((if_getdrvflags((ctx)->ifc_ifp) & IFF_DRV_RUNNING))
556 
557 #define CTX_LOCK_INIT(_sc)  sx_init(&(_sc)->ifc_ctx_sx, "iflib ctx lock")
558 #define CTX_LOCK(ctx) sx_xlock(&(ctx)->ifc_ctx_sx)
559 #define CTX_UNLOCK(ctx) sx_xunlock(&(ctx)->ifc_ctx_sx)
560 #define CTX_LOCK_DESTROY(ctx) sx_destroy(&(ctx)->ifc_ctx_sx)
561 
562 
563 #define STATE_LOCK_INIT(_sc, _name)  mtx_init(&(_sc)->ifc_state_mtx, _name, "iflib state lock", MTX_DEF)
564 #define STATE_LOCK(ctx) mtx_lock(&(ctx)->ifc_state_mtx)
565 #define STATE_UNLOCK(ctx) mtx_unlock(&(ctx)->ifc_state_mtx)
566 #define STATE_LOCK_DESTROY(ctx) mtx_destroy(&(ctx)->ifc_state_mtx)
567 
568 
569 
570 #define CALLOUT_LOCK(txq)	mtx_lock(&txq->ift_mtx)
571 #define CALLOUT_UNLOCK(txq) 	mtx_unlock(&txq->ift_mtx)
572 
573 
574 /* Our boot-time initialization hook */
575 static int	iflib_module_event_handler(module_t, int, void *);
576 
577 static moduledata_t iflib_moduledata = {
578 	"iflib",
579 	iflib_module_event_handler,
580 	NULL
581 };
582 
583 DECLARE_MODULE(iflib, iflib_moduledata, SI_SUB_INIT_IF, SI_ORDER_ANY);
584 MODULE_VERSION(iflib, 1);
585 
586 MODULE_DEPEND(iflib, pci, 1, 1, 1);
587 MODULE_DEPEND(iflib, ether, 1, 1, 1);
588 
589 TASKQGROUP_DEFINE(if_io_tqg, mp_ncpus, 1);
590 TASKQGROUP_DEFINE(if_config_tqg, 1, 1);
591 
592 #ifndef IFLIB_DEBUG_COUNTERS
593 #ifdef INVARIANTS
594 #define IFLIB_DEBUG_COUNTERS 1
595 #else
596 #define IFLIB_DEBUG_COUNTERS 0
597 #endif /* !INVARIANTS */
598 #endif
599 
600 static SYSCTL_NODE(_net, OID_AUTO, iflib, CTLFLAG_RD, 0,
601                    "iflib driver parameters");
602 
603 /*
604  * XXX need to ensure that this can't accidentally cause the head to be moved backwards
605  */
606 static int iflib_min_tx_latency = 0;
607 SYSCTL_INT(_net_iflib, OID_AUTO, min_tx_latency, CTLFLAG_RW,
608 		   &iflib_min_tx_latency, 0, "minimize transmit latency at the possible expense of throughput");
609 static int iflib_no_tx_batch = 0;
610 SYSCTL_INT(_net_iflib, OID_AUTO, no_tx_batch, CTLFLAG_RW,
611 		   &iflib_no_tx_batch, 0, "minimize transmit latency at the possible expense of throughput");
612 
613 
614 #if IFLIB_DEBUG_COUNTERS
615 
616 static int iflib_tx_seen;
617 static int iflib_tx_sent;
618 static int iflib_tx_encap;
619 static int iflib_rx_allocs;
620 static int iflib_fl_refills;
621 static int iflib_fl_refills_large;
622 static int iflib_tx_frees;
623 
624 SYSCTL_INT(_net_iflib, OID_AUTO, tx_seen, CTLFLAG_RD,
625 		   &iflib_tx_seen, 0, "# tx mbufs seen");
626 SYSCTL_INT(_net_iflib, OID_AUTO, tx_sent, CTLFLAG_RD,
627 		   &iflib_tx_sent, 0, "# tx mbufs sent");
628 SYSCTL_INT(_net_iflib, OID_AUTO, tx_encap, CTLFLAG_RD,
629 		   &iflib_tx_encap, 0, "# tx mbufs encapped");
630 SYSCTL_INT(_net_iflib, OID_AUTO, tx_frees, CTLFLAG_RD,
631 		   &iflib_tx_frees, 0, "# tx frees");
632 SYSCTL_INT(_net_iflib, OID_AUTO, rx_allocs, CTLFLAG_RD,
633 		   &iflib_rx_allocs, 0, "# rx allocations");
634 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills, CTLFLAG_RD,
635 		   &iflib_fl_refills, 0, "# refills");
636 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills_large, CTLFLAG_RD,
637 		   &iflib_fl_refills_large, 0, "# large refills");
638 
639 
640 static int iflib_txq_drain_flushing;
641 static int iflib_txq_drain_oactive;
642 static int iflib_txq_drain_notready;
643 static int iflib_txq_drain_encapfail;
644 
645 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_flushing, CTLFLAG_RD,
646 		   &iflib_txq_drain_flushing, 0, "# drain flushes");
647 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_oactive, CTLFLAG_RD,
648 		   &iflib_txq_drain_oactive, 0, "# drain oactives");
649 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_notready, CTLFLAG_RD,
650 		   &iflib_txq_drain_notready, 0, "# drain notready");
651 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_encapfail, CTLFLAG_RD,
652 		   &iflib_txq_drain_encapfail, 0, "# drain encap fails");
653 
654 
655 static int iflib_encap_load_mbuf_fail;
656 static int iflib_encap_pad_mbuf_fail;
657 static int iflib_encap_txq_avail_fail;
658 static int iflib_encap_txd_encap_fail;
659 
660 SYSCTL_INT(_net_iflib, OID_AUTO, encap_load_mbuf_fail, CTLFLAG_RD,
661 		   &iflib_encap_load_mbuf_fail, 0, "# busdma load failures");
662 SYSCTL_INT(_net_iflib, OID_AUTO, encap_pad_mbuf_fail, CTLFLAG_RD,
663 		   &iflib_encap_pad_mbuf_fail, 0, "# runt frame pad failures");
664 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txq_avail_fail, CTLFLAG_RD,
665 		   &iflib_encap_txq_avail_fail, 0, "# txq avail failures");
666 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txd_encap_fail, CTLFLAG_RD,
667 		   &iflib_encap_txd_encap_fail, 0, "# driver encap failures");
668 
669 static int iflib_task_fn_rxs;
670 static int iflib_rx_intr_enables;
671 static int iflib_fast_intrs;
672 static int iflib_intr_link;
673 static int iflib_intr_msix;
674 static int iflib_rx_unavail;
675 static int iflib_rx_ctx_inactive;
676 static int iflib_rx_zero_len;
677 static int iflib_rx_if_input;
678 static int iflib_rx_mbuf_null;
679 static int iflib_rxd_flush;
680 
681 static int iflib_verbose_debug;
682 
683 SYSCTL_INT(_net_iflib, OID_AUTO, intr_link, CTLFLAG_RD,
684 		   &iflib_intr_link, 0, "# intr link calls");
685 SYSCTL_INT(_net_iflib, OID_AUTO, intr_msix, CTLFLAG_RD,
686 		   &iflib_intr_msix, 0, "# intr msix calls");
687 SYSCTL_INT(_net_iflib, OID_AUTO, task_fn_rx, CTLFLAG_RD,
688 		   &iflib_task_fn_rxs, 0, "# task_fn_rx calls");
689 SYSCTL_INT(_net_iflib, OID_AUTO, rx_intr_enables, CTLFLAG_RD,
690 		   &iflib_rx_intr_enables, 0, "# rx intr enables");
691 SYSCTL_INT(_net_iflib, OID_AUTO, fast_intrs, CTLFLAG_RD,
692 		   &iflib_fast_intrs, 0, "# fast_intr calls");
693 SYSCTL_INT(_net_iflib, OID_AUTO, rx_unavail, CTLFLAG_RD,
694 		   &iflib_rx_unavail, 0, "# times rxeof called with no available data");
695 SYSCTL_INT(_net_iflib, OID_AUTO, rx_ctx_inactive, CTLFLAG_RD,
696 		   &iflib_rx_ctx_inactive, 0, "# times rxeof called with inactive context");
697 SYSCTL_INT(_net_iflib, OID_AUTO, rx_zero_len, CTLFLAG_RD,
698 		   &iflib_rx_zero_len, 0, "# times rxeof saw zero len mbuf");
699 SYSCTL_INT(_net_iflib, OID_AUTO, rx_if_input, CTLFLAG_RD,
700 		   &iflib_rx_if_input, 0, "# times rxeof called if_input");
701 SYSCTL_INT(_net_iflib, OID_AUTO, rx_mbuf_null, CTLFLAG_RD,
702 		   &iflib_rx_mbuf_null, 0, "# times rxeof got null mbuf");
703 SYSCTL_INT(_net_iflib, OID_AUTO, rxd_flush, CTLFLAG_RD,
704 	         &iflib_rxd_flush, 0, "# times rxd_flush called");
705 SYSCTL_INT(_net_iflib, OID_AUTO, verbose_debug, CTLFLAG_RW,
706 		   &iflib_verbose_debug, 0, "enable verbose debugging");
707 
708 #define DBG_COUNTER_INC(name) atomic_add_int(&(iflib_ ## name), 1)
709 static void
710 iflib_debug_reset(void)
711 {
712 	iflib_tx_seen = iflib_tx_sent = iflib_tx_encap = iflib_rx_allocs =
713 		iflib_fl_refills = iflib_fl_refills_large = iflib_tx_frees =
714 		iflib_txq_drain_flushing = iflib_txq_drain_oactive =
715 		iflib_txq_drain_notready = iflib_txq_drain_encapfail =
716 		iflib_encap_load_mbuf_fail = iflib_encap_pad_mbuf_fail =
717 		iflib_encap_txq_avail_fail = iflib_encap_txd_encap_fail =
718 		iflib_task_fn_rxs = iflib_rx_intr_enables = iflib_fast_intrs =
719 		iflib_intr_link = iflib_intr_msix = iflib_rx_unavail =
720 		iflib_rx_ctx_inactive = iflib_rx_zero_len = iflib_rx_if_input =
721 		iflib_rx_mbuf_null = iflib_rxd_flush = 0;
722 }
723 
724 #else
725 #define DBG_COUNTER_INC(name)
726 static void iflib_debug_reset(void) {}
727 #endif
728 
729 #define IFLIB_DEBUG 0
730 
731 static void iflib_tx_structures_free(if_ctx_t ctx);
732 static void iflib_rx_structures_free(if_ctx_t ctx);
733 static int iflib_queues_alloc(if_ctx_t ctx);
734 static int iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq);
735 static int iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget);
736 static int iflib_qset_structures_setup(if_ctx_t ctx);
737 static int iflib_msix_init(if_ctx_t ctx);
738 static int iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filterarg, int *rid, const char *str);
739 static void iflib_txq_check_drain(iflib_txq_t txq, int budget);
740 static uint32_t iflib_txq_can_drain(struct ifmp_ring *);
741 static int iflib_register(if_ctx_t);
742 static void iflib_init_locked(if_ctx_t ctx);
743 static void iflib_add_device_sysctl_pre(if_ctx_t ctx);
744 static void iflib_add_device_sysctl_post(if_ctx_t ctx);
745 static void iflib_ifmp_purge(iflib_txq_t txq);
746 static void _iflib_pre_assert(if_softc_ctx_t scctx);
747 static void iflib_if_init_locked(if_ctx_t ctx);
748 #ifndef __NO_STRICT_ALIGNMENT
749 static struct mbuf * iflib_fixup_rx(struct mbuf *m);
750 #endif
751 
752 NETDUMP_DEFINE(iflib);
753 
754 #ifdef DEV_NETMAP
755 #include <sys/selinfo.h>
756 #include <net/netmap.h>
757 #include <dev/netmap/netmap_kern.h>
758 
759 MODULE_DEPEND(iflib, netmap, 1, 1, 1);
760 
761 static int netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, uint32_t nm_i, bool init);
762 
763 /*
764  * device-specific sysctl variables:
765  *
766  * iflib_crcstrip: 0: keep CRC in rx frames (default), 1: strip it.
767  *	During regular operations the CRC is stripped, but on some
768  *	hardware reception of frames not multiple of 64 is slower,
769  *	so using crcstrip=0 helps in benchmarks.
770  *
771  * iflib_rx_miss, iflib_rx_miss_bufs:
772  *	count packets that might be missed due to lost interrupts.
773  */
774 SYSCTL_DECL(_dev_netmap);
775 /*
776  * The xl driver by default strips CRCs and we do not override it.
777  */
778 
779 int iflib_crcstrip = 1;
780 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_crcstrip,
781     CTLFLAG_RW, &iflib_crcstrip, 1, "strip CRC on rx frames");
782 
783 int iflib_rx_miss, iflib_rx_miss_bufs;
784 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss,
785     CTLFLAG_RW, &iflib_rx_miss, 0, "potentially missed rx intr");
786 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss_bufs,
787     CTLFLAG_RW, &iflib_rx_miss_bufs, 0, "potentially missed rx intr bufs");
788 
789 /*
790  * Register/unregister. We are already under netmap lock.
791  * Only called on the first register or the last unregister.
792  */
793 static int
794 iflib_netmap_register(struct netmap_adapter *na, int onoff)
795 {
796 	struct ifnet *ifp = na->ifp;
797 	if_ctx_t ctx = ifp->if_softc;
798 	int status;
799 
800 	CTX_LOCK(ctx);
801 	IFDI_INTR_DISABLE(ctx);
802 
803 	/* Tell the stack that the interface is no longer active */
804 	ifp->if_drv_flags &= ~(IFF_DRV_RUNNING | IFF_DRV_OACTIVE);
805 
806 	if (!CTX_IS_VF(ctx))
807 		IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip);
808 
809 	/* enable or disable flags and callbacks in na and ifp */
810 	if (onoff) {
811 		nm_set_native_flags(na);
812 	} else {
813 		nm_clear_native_flags(na);
814 	}
815 	iflib_stop(ctx);
816 	iflib_init_locked(ctx);
817 	IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip); // XXX why twice ?
818 	status = ifp->if_drv_flags & IFF_DRV_RUNNING ? 0 : 1;
819 	if (status)
820 		nm_clear_native_flags(na);
821 	CTX_UNLOCK(ctx);
822 	return (status);
823 }
824 
825 static int
826 netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, uint32_t nm_i, bool init)
827 {
828 	struct netmap_adapter *na = kring->na;
829 	u_int const lim = kring->nkr_num_slots - 1;
830 	u_int head = kring->rhead;
831 	struct netmap_ring *ring = kring->ring;
832 	bus_dmamap_t *map;
833 	struct if_rxd_update iru;
834 	if_ctx_t ctx = rxq->ifr_ctx;
835 	iflib_fl_t fl = &rxq->ifr_fl[0];
836 	uint32_t refill_pidx, nic_i;
837 
838 	if (nm_i == head && __predict_true(!init))
839 		return 0;
840 	iru_init(&iru, rxq, 0 /* flid */);
841 	map = fl->ifl_sds.ifsd_map;
842 	refill_pidx = netmap_idx_k2n(kring, nm_i);
843 	/*
844 	 * IMPORTANT: we must leave one free slot in the ring,
845 	 * so move head back by one unit
846 	 */
847 	head = nm_prev(head, lim);
848 	nic_i = UINT_MAX;
849 	while (nm_i != head) {
850 		for (int tmp_pidx = 0; tmp_pidx < IFLIB_MAX_RX_REFRESH && nm_i != head; tmp_pidx++) {
851 			struct netmap_slot *slot = &ring->slot[nm_i];
852 			void *addr = PNMB(na, slot, &fl->ifl_bus_addrs[tmp_pidx]);
853 			uint32_t nic_i_dma = refill_pidx;
854 			nic_i = netmap_idx_k2n(kring, nm_i);
855 
856 			MPASS(tmp_pidx < IFLIB_MAX_RX_REFRESH);
857 
858 			if (addr == NETMAP_BUF_BASE(na)) /* bad buf */
859 			        return netmap_ring_reinit(kring);
860 
861 			fl->ifl_vm_addrs[tmp_pidx] = addr;
862 			if (__predict_false(init) && map) {
863 				netmap_load_map(na, fl->ifl_ifdi->idi_tag, map[nic_i], addr);
864 			} else if (map && (slot->flags & NS_BUF_CHANGED)) {
865 				/* buffer has changed, reload map */
866 				netmap_reload_map(na, fl->ifl_ifdi->idi_tag, map[nic_i], addr);
867 			}
868 			slot->flags &= ~NS_BUF_CHANGED;
869 
870 			nm_i = nm_next(nm_i, lim);
871 			fl->ifl_rxd_idxs[tmp_pidx] = nic_i = nm_next(nic_i, lim);
872 			if (nm_i != head && tmp_pidx < IFLIB_MAX_RX_REFRESH-1)
873 				continue;
874 
875 			iru.iru_pidx = refill_pidx;
876 			iru.iru_count = tmp_pidx+1;
877 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
878 
879 			refill_pidx = nic_i;
880 			if (map == NULL)
881 				continue;
882 
883 			for (int n = 0; n < iru.iru_count; n++) {
884 				bus_dmamap_sync(fl->ifl_ifdi->idi_tag, map[nic_i_dma],
885 						BUS_DMASYNC_PREREAD);
886 				/* XXX - change this to not use the netmap func*/
887 				nic_i_dma = nm_next(nic_i_dma, lim);
888 			}
889 		}
890 	}
891 	kring->nr_hwcur = head;
892 
893 	if (map)
894 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
895 				BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
896 	if (__predict_true(nic_i != UINT_MAX))
897 		ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id, nic_i);
898 	return (0);
899 }
900 
901 /*
902  * Reconcile kernel and user view of the transmit ring.
903  *
904  * All information is in the kring.
905  * Userspace wants to send packets up to the one before kring->rhead,
906  * kernel knows kring->nr_hwcur is the first unsent packet.
907  *
908  * Here we push packets out (as many as possible), and possibly
909  * reclaim buffers from previously completed transmission.
910  *
911  * The caller (netmap) guarantees that there is only one instance
912  * running at any time. Any interference with other driver
913  * methods should be handled by the individual drivers.
914  */
915 static int
916 iflib_netmap_txsync(struct netmap_kring *kring, int flags)
917 {
918 	struct netmap_adapter *na = kring->na;
919 	struct ifnet *ifp = na->ifp;
920 	struct netmap_ring *ring = kring->ring;
921 	u_int nm_i;	/* index into the netmap kring */
922 	u_int nic_i;	/* index into the NIC ring */
923 	u_int n;
924 	u_int const lim = kring->nkr_num_slots - 1;
925 	u_int const head = kring->rhead;
926 	struct if_pkt_info pi;
927 
928 	/*
929 	 * interrupts on every tx packet are expensive so request
930 	 * them every half ring, or where NS_REPORT is set
931 	 */
932 	u_int report_frequency = kring->nkr_num_slots >> 1;
933 	/* device-specific */
934 	if_ctx_t ctx = ifp->if_softc;
935 	iflib_txq_t txq = &ctx->ifc_txqs[kring->ring_id];
936 
937 	if (txq->ift_sds.ifsd_map)
938 		bus_dmamap_sync(txq->ift_desc_tag, txq->ift_ifdi->idi_map,
939 				BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
940 
941 
942 	/*
943 	 * First part: process new packets to send.
944 	 * nm_i is the current index in the netmap kring,
945 	 * nic_i is the corresponding index in the NIC ring.
946 	 *
947 	 * If we have packets to send (nm_i != head)
948 	 * iterate over the netmap ring, fetch length and update
949 	 * the corresponding slot in the NIC ring. Some drivers also
950 	 * need to update the buffer's physical address in the NIC slot
951 	 * even NS_BUF_CHANGED is not set (PNMB computes the addresses).
952 	 *
953 	 * The netmap_reload_map() calls is especially expensive,
954 	 * even when (as in this case) the tag is 0, so do only
955 	 * when the buffer has actually changed.
956 	 *
957 	 * If possible do not set the report/intr bit on all slots,
958 	 * but only a few times per ring or when NS_REPORT is set.
959 	 *
960 	 * Finally, on 10G and faster drivers, it might be useful
961 	 * to prefetch the next slot and txr entry.
962 	 */
963 
964 	nm_i = kring->nr_hwcur;
965 	if (nm_i != head) {	/* we have new packets to send */
966 		pkt_info_zero(&pi);
967 		pi.ipi_segs = txq->ift_segs;
968 		pi.ipi_qsidx = kring->ring_id;
969 		nic_i = netmap_idx_k2n(kring, nm_i);
970 
971 		__builtin_prefetch(&ring->slot[nm_i]);
972 		__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i]);
973 		if (txq->ift_sds.ifsd_map)
974 			__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i]);
975 
976 		for (n = 0; nm_i != head; n++) {
977 			struct netmap_slot *slot = &ring->slot[nm_i];
978 			u_int len = slot->len;
979 			uint64_t paddr;
980 			void *addr = PNMB(na, slot, &paddr);
981 			int flags = (slot->flags & NS_REPORT ||
982 				nic_i == 0 || nic_i == report_frequency) ?
983 				IPI_TX_INTR : 0;
984 
985 			/* device-specific */
986 			pi.ipi_len = len;
987 			pi.ipi_segs[0].ds_addr = paddr;
988 			pi.ipi_segs[0].ds_len = len;
989 			pi.ipi_nsegs = 1;
990 			pi.ipi_ndescs = 0;
991 			pi.ipi_pidx = nic_i;
992 			pi.ipi_flags = flags;
993 
994 			/* Fill the slot in the NIC ring. */
995 			ctx->isc_txd_encap(ctx->ifc_softc, &pi);
996 
997 			/* prefetch for next round */
998 			__builtin_prefetch(&ring->slot[nm_i + 1]);
999 			__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i + 1]);
1000 			if (txq->ift_sds.ifsd_map) {
1001 				__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i + 1]);
1002 
1003 				NM_CHECK_ADDR_LEN(na, addr, len);
1004 
1005 				if (slot->flags & NS_BUF_CHANGED) {
1006 					/* buffer has changed, reload map */
1007 					netmap_reload_map(na, txq->ift_desc_tag, txq->ift_sds.ifsd_map[nic_i], addr);
1008 				}
1009 				/* make sure changes to the buffer are synced */
1010 				bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_sds.ifsd_map[nic_i],
1011 						BUS_DMASYNC_PREWRITE);
1012 			}
1013 			slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED);
1014 			nm_i = nm_next(nm_i, lim);
1015 			nic_i = nm_next(nic_i, lim);
1016 		}
1017 		kring->nr_hwcur = nm_i;
1018 
1019 		/* synchronize the NIC ring */
1020 		if (txq->ift_sds.ifsd_map)
1021 			bus_dmamap_sync(txq->ift_desc_tag, txq->ift_ifdi->idi_map,
1022 						BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1023 
1024 		/* (re)start the tx unit up to slot nic_i (excluded) */
1025 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, nic_i);
1026 	}
1027 
1028 	/*
1029 	 * Second part: reclaim buffers for completed transmissions.
1030 	 *
1031 	 * If there are unclaimed buffers, attempt to reclaim them.
1032 	 * If none are reclaimed, and TX IRQs are not in use, do an initial
1033 	 * minimal delay, then trigger the tx handler which will spin in the
1034 	 * group task queue.
1035 	 */
1036 	if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1037 		if (iflib_tx_credits_update(ctx, txq)) {
1038 			/* some tx completed, increment avail */
1039 			nic_i = txq->ift_cidx_processed;
1040 			kring->nr_hwtail = nm_prev(netmap_idx_n2k(kring, nic_i), lim);
1041 		}
1042 	}
1043 	if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ))
1044 		if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1045 			callout_reset_on(&txq->ift_timer, hz < 2000 ? 1 : hz / 1000,
1046 			    iflib_timer, txq, txq->ift_timer.c_cpu);
1047 	}
1048 	return (0);
1049 }
1050 
1051 /*
1052  * Reconcile kernel and user view of the receive ring.
1053  * Same as for the txsync, this routine must be efficient.
1054  * The caller guarantees a single invocations, but races against
1055  * the rest of the driver should be handled here.
1056  *
1057  * On call, kring->rhead is the first packet that userspace wants
1058  * to keep, and kring->rcur is the wakeup point.
1059  * The kernel has previously reported packets up to kring->rtail.
1060  *
1061  * If (flags & NAF_FORCE_READ) also check for incoming packets irrespective
1062  * of whether or not we received an interrupt.
1063  */
1064 static int
1065 iflib_netmap_rxsync(struct netmap_kring *kring, int flags)
1066 {
1067 	struct netmap_adapter *na = kring->na;
1068 	struct netmap_ring *ring = kring->ring;
1069 	uint32_t nm_i;	/* index into the netmap ring */
1070 	uint32_t nic_i;	/* index into the NIC ring */
1071 	u_int i, n;
1072 	u_int const lim = kring->nkr_num_slots - 1;
1073 	u_int const head = kring->rhead;
1074 	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
1075 	struct if_rxd_info ri;
1076 
1077 	struct ifnet *ifp = na->ifp;
1078 	if_ctx_t ctx = ifp->if_softc;
1079 	iflib_rxq_t rxq = &ctx->ifc_rxqs[kring->ring_id];
1080 	iflib_fl_t fl = rxq->ifr_fl;
1081 	if (head > lim)
1082 		return netmap_ring_reinit(kring);
1083 
1084 	/* XXX check sync modes */
1085 	for (i = 0, fl = rxq->ifr_fl; i < rxq->ifr_nfl; i++, fl++) {
1086 		if (fl->ifl_sds.ifsd_map == NULL)
1087 			continue;
1088 		bus_dmamap_sync(rxq->ifr_fl[i].ifl_desc_tag, fl->ifl_ifdi->idi_map,
1089 				BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1090 	}
1091 	/*
1092 	 * First part: import newly received packets.
1093 	 *
1094 	 * nm_i is the index of the next free slot in the netmap ring,
1095 	 * nic_i is the index of the next received packet in the NIC ring,
1096 	 * and they may differ in case if_init() has been called while
1097 	 * in netmap mode. For the receive ring we have
1098 	 *
1099 	 *	nic_i = rxr->next_check;
1100 	 *	nm_i = kring->nr_hwtail (previous)
1101 	 * and
1102 	 *	nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1103 	 *
1104 	 * rxr->next_check is set to 0 on a ring reinit
1105 	 */
1106 	if (netmap_no_pendintr || force_update) {
1107 		int crclen = iflib_crcstrip ? 0 : 4;
1108 		int error, avail;
1109 
1110 		for (i = 0; i < rxq->ifr_nfl; i++) {
1111 			fl = &rxq->ifr_fl[i];
1112 			nic_i = fl->ifl_cidx;
1113 			nm_i = netmap_idx_n2k(kring, nic_i);
1114 			avail = iflib_rxd_avail(ctx, rxq, nic_i, USHRT_MAX);
1115 			for (n = 0; avail > 0; n++, avail--) {
1116 				rxd_info_zero(&ri);
1117 				ri.iri_frags = rxq->ifr_frags;
1118 				ri.iri_qsidx = kring->ring_id;
1119 				ri.iri_ifp = ctx->ifc_ifp;
1120 				ri.iri_cidx = nic_i;
1121 
1122 				error = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
1123 				ring->slot[nm_i].len = error ? 0 : ri.iri_len - crclen;
1124 				ring->slot[nm_i].flags = 0;
1125 				if (fl->ifl_sds.ifsd_map)
1126 					bus_dmamap_sync(fl->ifl_ifdi->idi_tag,
1127 							fl->ifl_sds.ifsd_map[nic_i], BUS_DMASYNC_POSTREAD);
1128 				nm_i = nm_next(nm_i, lim);
1129 				nic_i = nm_next(nic_i, lim);
1130 			}
1131 			if (n) { /* update the state variables */
1132 				if (netmap_no_pendintr && !force_update) {
1133 					/* diagnostics */
1134 					iflib_rx_miss ++;
1135 					iflib_rx_miss_bufs += n;
1136 				}
1137 				fl->ifl_cidx = nic_i;
1138 				kring->nr_hwtail = nm_i;
1139 			}
1140 			kring->nr_kflags &= ~NKR_PENDINTR;
1141 		}
1142 	}
1143 	/*
1144 	 * Second part: skip past packets that userspace has released.
1145 	 * (kring->nr_hwcur to head excluded),
1146 	 * and make the buffers available for reception.
1147 	 * As usual nm_i is the index in the netmap ring,
1148 	 * nic_i is the index in the NIC ring, and
1149 	 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1150 	 */
1151 	/* XXX not sure how this will work with multiple free lists */
1152 	nm_i = kring->nr_hwcur;
1153 
1154 	return (netmap_fl_refill(rxq, kring, nm_i, false));
1155 }
1156 
1157 static void
1158 iflib_netmap_intr(struct netmap_adapter *na, int onoff)
1159 {
1160 	struct ifnet *ifp = na->ifp;
1161 	if_ctx_t ctx = ifp->if_softc;
1162 
1163 	CTX_LOCK(ctx);
1164 	if (onoff) {
1165 		IFDI_INTR_ENABLE(ctx);
1166 	} else {
1167 		IFDI_INTR_DISABLE(ctx);
1168 	}
1169 	CTX_UNLOCK(ctx);
1170 }
1171 
1172 
1173 static int
1174 iflib_netmap_attach(if_ctx_t ctx)
1175 {
1176 	struct netmap_adapter na;
1177 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1178 
1179 	bzero(&na, sizeof(na));
1180 
1181 	na.ifp = ctx->ifc_ifp;
1182 	na.na_flags = NAF_BDG_MAYSLEEP;
1183 	MPASS(ctx->ifc_softc_ctx.isc_ntxqsets);
1184 	MPASS(ctx->ifc_softc_ctx.isc_nrxqsets);
1185 
1186 	na.num_tx_desc = scctx->isc_ntxd[0];
1187 	na.num_rx_desc = scctx->isc_nrxd[0];
1188 	na.nm_txsync = iflib_netmap_txsync;
1189 	na.nm_rxsync = iflib_netmap_rxsync;
1190 	na.nm_register = iflib_netmap_register;
1191 	na.nm_intr = iflib_netmap_intr;
1192 	na.num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
1193 	na.num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
1194 	return (netmap_attach(&na));
1195 }
1196 
1197 static void
1198 iflib_netmap_txq_init(if_ctx_t ctx, iflib_txq_t txq)
1199 {
1200 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1201 	struct netmap_slot *slot;
1202 
1203 	slot = netmap_reset(na, NR_TX, txq->ift_id, 0);
1204 	if (slot == NULL)
1205 		return;
1206 	if (txq->ift_sds.ifsd_map == NULL)
1207 		return;
1208 
1209 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxd[0]; i++) {
1210 
1211 		/*
1212 		 * In netmap mode, set the map for the packet buffer.
1213 		 * NOTE: Some drivers (not this one) also need to set
1214 		 * the physical buffer address in the NIC ring.
1215 		 * netmap_idx_n2k() maps a nic index, i, into the corresponding
1216 		 * netmap slot index, si
1217 		 */
1218 		int si = netmap_idx_n2k(na->tx_rings[txq->ift_id], i);
1219 		netmap_load_map(na, txq->ift_desc_tag, txq->ift_sds.ifsd_map[i], NMB(na, slot + si));
1220 	}
1221 }
1222 
1223 static void
1224 iflib_netmap_rxq_init(if_ctx_t ctx, iflib_rxq_t rxq)
1225 {
1226 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1227 	struct netmap_kring *kring = na->rx_rings[rxq->ifr_id];
1228 	struct netmap_slot *slot;
1229 	uint32_t nm_i;
1230 
1231 	slot = netmap_reset(na, NR_RX, rxq->ifr_id, 0);
1232 	if (slot == NULL)
1233 		return;
1234 	nm_i = netmap_idx_n2k(kring, 0);
1235 	netmap_fl_refill(rxq, kring, nm_i, true);
1236 }
1237 
1238 static void
1239 iflib_netmap_timer_adjust(if_ctx_t ctx, uint16_t txqid, uint32_t *reset_on)
1240 {
1241 	struct netmap_kring *kring;
1242 
1243 	kring = NA(ctx->ifc_ifp)->tx_rings[txqid];
1244 
1245 	if (kring->nr_hwcur != nm_next(kring->nr_hwtail, kring->nkr_num_slots - 1)) {
1246 		if (ctx->isc_txd_credits_update(ctx->ifc_softc, txqid, false))
1247 			netmap_tx_irq(ctx->ifc_ifp, txqid);
1248 		if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ)) {
1249 			if (hz < 2000)
1250 				*reset_on = 1;
1251 			else
1252 				*reset_on = hz / 1000;
1253 		}
1254 	}
1255 }
1256 
1257 #define iflib_netmap_detach(ifp) netmap_detach(ifp)
1258 
1259 #else
1260 #define iflib_netmap_txq_init(ctx, txq)
1261 #define iflib_netmap_rxq_init(ctx, rxq)
1262 #define iflib_netmap_detach(ifp)
1263 
1264 #define iflib_netmap_attach(ctx) (0)
1265 #define netmap_rx_irq(ifp, qid, budget) (0)
1266 #define netmap_tx_irq(ifp, qid) do {} while (0)
1267 #define iflib_netmap_timer_adjust(ctx, txqid, reset_on)
1268 
1269 #endif
1270 
1271 #if defined(__i386__) || defined(__amd64__)
1272 static __inline void
1273 prefetch(void *x)
1274 {
1275 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1276 }
1277 static __inline void
1278 prefetch2cachelines(void *x)
1279 {
1280 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1281 #if (CACHE_LINE_SIZE < 128)
1282 	__asm volatile("prefetcht0 %0" :: "m" (*(((unsigned long *)x)+CACHE_LINE_SIZE/(sizeof(unsigned long)))));
1283 #endif
1284 }
1285 #else
1286 #define prefetch(x)
1287 #define prefetch2cachelines(x)
1288 #endif
1289 
1290 static void
1291 iflib_gen_mac(if_ctx_t ctx)
1292 {
1293 	struct thread *td;
1294 	MD5_CTX mdctx;
1295 	char uuid[HOSTUUIDLEN+1];
1296 	char buf[HOSTUUIDLEN+16];
1297 	uint8_t *mac;
1298 	unsigned char digest[16];
1299 
1300 	td = curthread;
1301 	mac = ctx->ifc_mac;
1302 	uuid[HOSTUUIDLEN] = 0;
1303 	bcopy(td->td_ucred->cr_prison->pr_hostuuid, uuid, HOSTUUIDLEN);
1304 	snprintf(buf, HOSTUUIDLEN+16, "%s-%s", uuid, device_get_nameunit(ctx->ifc_dev));
1305 	/*
1306 	 * Generate a pseudo-random, deterministic MAC
1307 	 * address based on the UUID and unit number.
1308 	 * The FreeBSD Foundation OUI of 58-9C-FC is used.
1309 	 */
1310 	MD5Init(&mdctx);
1311 	MD5Update(&mdctx, buf, strlen(buf));
1312 	MD5Final(digest, &mdctx);
1313 
1314 	mac[0] = 0x58;
1315 	mac[1] = 0x9C;
1316 	mac[2] = 0xFC;
1317 	mac[3] = digest[0];
1318 	mac[4] = digest[1];
1319 	mac[5] = digest[2];
1320 }
1321 
1322 static void
1323 iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid)
1324 {
1325 	iflib_fl_t fl;
1326 
1327 	fl = &rxq->ifr_fl[flid];
1328 	iru->iru_paddrs = fl->ifl_bus_addrs;
1329 	iru->iru_vaddrs = &fl->ifl_vm_addrs[0];
1330 	iru->iru_idxs = fl->ifl_rxd_idxs;
1331 	iru->iru_qsidx = rxq->ifr_id;
1332 	iru->iru_buf_size = fl->ifl_buf_size;
1333 	iru->iru_flidx = fl->ifl_id;
1334 }
1335 
1336 static void
1337 _iflib_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int err)
1338 {
1339 	if (err)
1340 		return;
1341 	*(bus_addr_t *) arg = segs[0].ds_addr;
1342 }
1343 
1344 int
1345 iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags)
1346 {
1347 	int err;
1348 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1349 	device_t dev = ctx->ifc_dev;
1350 
1351 	KASSERT(sctx->isc_q_align != 0, ("alignment value not initialized"));
1352 
1353 	err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1354 				sctx->isc_q_align, 0,	/* alignment, bounds */
1355 				BUS_SPACE_MAXADDR,	/* lowaddr */
1356 				BUS_SPACE_MAXADDR,	/* highaddr */
1357 				NULL, NULL,		/* filter, filterarg */
1358 				size,			/* maxsize */
1359 				1,			/* nsegments */
1360 				size,			/* maxsegsize */
1361 				BUS_DMA_ALLOCNOW,	/* flags */
1362 				NULL,			/* lockfunc */
1363 				NULL,			/* lockarg */
1364 				&dma->idi_tag);
1365 	if (err) {
1366 		device_printf(dev,
1367 		    "%s: bus_dma_tag_create failed: %d\n",
1368 		    __func__, err);
1369 		goto fail_0;
1370 	}
1371 
1372 	err = bus_dmamem_alloc(dma->idi_tag, (void**) &dma->idi_vaddr,
1373 	    BUS_DMA_NOWAIT | BUS_DMA_COHERENT | BUS_DMA_ZERO, &dma->idi_map);
1374 	if (err) {
1375 		device_printf(dev,
1376 		    "%s: bus_dmamem_alloc(%ju) failed: %d\n",
1377 		    __func__, (uintmax_t)size, err);
1378 		goto fail_1;
1379 	}
1380 
1381 	dma->idi_paddr = IF_BAD_DMA;
1382 	err = bus_dmamap_load(dma->idi_tag, dma->idi_map, dma->idi_vaddr,
1383 	    size, _iflib_dmamap_cb, &dma->idi_paddr, mapflags | BUS_DMA_NOWAIT);
1384 	if (err || dma->idi_paddr == IF_BAD_DMA) {
1385 		device_printf(dev,
1386 		    "%s: bus_dmamap_load failed: %d\n",
1387 		    __func__, err);
1388 		goto fail_2;
1389 	}
1390 
1391 	dma->idi_size = size;
1392 	return (0);
1393 
1394 fail_2:
1395 	bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1396 fail_1:
1397 	bus_dma_tag_destroy(dma->idi_tag);
1398 fail_0:
1399 	dma->idi_tag = NULL;
1400 
1401 	return (err);
1402 }
1403 
1404 int
1405 iflib_dma_alloc_multi(if_ctx_t ctx, int *sizes, iflib_dma_info_t *dmalist, int mapflags, int count)
1406 {
1407 	int i, err;
1408 	iflib_dma_info_t *dmaiter;
1409 
1410 	dmaiter = dmalist;
1411 	for (i = 0; i < count; i++, dmaiter++) {
1412 		if ((err = iflib_dma_alloc(ctx, sizes[i], *dmaiter, mapflags)) != 0)
1413 			break;
1414 	}
1415 	if (err)
1416 		iflib_dma_free_multi(dmalist, i);
1417 	return (err);
1418 }
1419 
1420 void
1421 iflib_dma_free(iflib_dma_info_t dma)
1422 {
1423 	if (dma->idi_tag == NULL)
1424 		return;
1425 	if (dma->idi_paddr != IF_BAD_DMA) {
1426 		bus_dmamap_sync(dma->idi_tag, dma->idi_map,
1427 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1428 		bus_dmamap_unload(dma->idi_tag, dma->idi_map);
1429 		dma->idi_paddr = IF_BAD_DMA;
1430 	}
1431 	if (dma->idi_vaddr != NULL) {
1432 		bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1433 		dma->idi_vaddr = NULL;
1434 	}
1435 	bus_dma_tag_destroy(dma->idi_tag);
1436 	dma->idi_tag = NULL;
1437 }
1438 
1439 void
1440 iflib_dma_free_multi(iflib_dma_info_t *dmalist, int count)
1441 {
1442 	int i;
1443 	iflib_dma_info_t *dmaiter = dmalist;
1444 
1445 	for (i = 0; i < count; i++, dmaiter++)
1446 		iflib_dma_free(*dmaiter);
1447 }
1448 
1449 #ifdef EARLY_AP_STARTUP
1450 static const int iflib_started = 1;
1451 #else
1452 /*
1453  * We used to abuse the smp_started flag to decide if the queues have been
1454  * fully initialized (by late taskqgroup_adjust() calls in a SYSINIT()).
1455  * That gave bad races, since the SYSINIT() runs strictly after smp_started
1456  * is set.  Run a SYSINIT() strictly after that to just set a usable
1457  * completion flag.
1458  */
1459 
1460 static int iflib_started;
1461 
1462 static void
1463 iflib_record_started(void *arg)
1464 {
1465 	iflib_started = 1;
1466 }
1467 
1468 SYSINIT(iflib_record_started, SI_SUB_SMP + 1, SI_ORDER_FIRST,
1469 	iflib_record_started, NULL);
1470 #endif
1471 
1472 static int
1473 iflib_fast_intr(void *arg)
1474 {
1475 	iflib_filter_info_t info = arg;
1476 	struct grouptask *gtask = info->ifi_task;
1477 	if (!iflib_started)
1478 		return (FILTER_HANDLED);
1479 
1480 	DBG_COUNTER_INC(fast_intrs);
1481 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1482 		return (FILTER_HANDLED);
1483 
1484 	GROUPTASK_ENQUEUE(gtask);
1485 	return (FILTER_HANDLED);
1486 }
1487 
1488 static int
1489 iflib_fast_intr_rxtx(void *arg)
1490 {
1491 	iflib_filter_info_t info = arg;
1492 	struct grouptask *gtask = info->ifi_task;
1493 	iflib_rxq_t rxq = (iflib_rxq_t)info->ifi_ctx;
1494 	if_ctx_t ctx = NULL;;
1495 	int i, cidx;
1496 
1497 	if (!iflib_started)
1498 		return (FILTER_HANDLED);
1499 
1500 	DBG_COUNTER_INC(fast_intrs);
1501 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1502 		return (FILTER_HANDLED);
1503 
1504 	MPASS(rxq->ifr_ntxqirq);
1505 	for (i = 0; i < rxq->ifr_ntxqirq; i++) {
1506 		qidx_t txqid = rxq->ifr_txqid[i];
1507 
1508 		ctx = rxq->ifr_ctx;
1509 
1510 		if (!ctx->isc_txd_credits_update(ctx->ifc_softc, txqid, false)) {
1511 			IFDI_TX_QUEUE_INTR_ENABLE(ctx, txqid);
1512 			continue;
1513 		}
1514 		GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
1515 	}
1516 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_RXCQ)
1517 		cidx = rxq->ifr_cq_cidx;
1518 	else
1519 		cidx = rxq->ifr_fl[0].ifl_cidx;
1520 	if (iflib_rxd_avail(ctx, rxq, cidx, 1))
1521 		GROUPTASK_ENQUEUE(gtask);
1522 	else
1523 		IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
1524 	return (FILTER_HANDLED);
1525 }
1526 
1527 
1528 static int
1529 iflib_fast_intr_ctx(void *arg)
1530 {
1531 	iflib_filter_info_t info = arg;
1532 	struct grouptask *gtask = info->ifi_task;
1533 
1534 	if (!iflib_started)
1535 		return (FILTER_HANDLED);
1536 
1537 	DBG_COUNTER_INC(fast_intrs);
1538 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1539 		return (FILTER_HANDLED);
1540 
1541 	GROUPTASK_ENQUEUE(gtask);
1542 	return (FILTER_HANDLED);
1543 }
1544 
1545 static int
1546 _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
1547 		 driver_filter_t filter, driver_intr_t handler, void *arg,
1548 		 const char *name)
1549 {
1550 	int rc, flags;
1551 	struct resource *res;
1552 	void *tag = NULL;
1553 	device_t dev = ctx->ifc_dev;
1554 
1555 	flags = RF_ACTIVE;
1556 	if (ctx->ifc_flags & IFC_LEGACY)
1557 		flags |= RF_SHAREABLE;
1558 	MPASS(rid < 512);
1559 	irq->ii_rid = rid;
1560 	res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &irq->ii_rid, flags);
1561 	if (res == NULL) {
1562 		device_printf(dev,
1563 		    "failed to allocate IRQ for rid %d, name %s.\n", rid, name);
1564 		return (ENOMEM);
1565 	}
1566 	irq->ii_res = res;
1567 	KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL"));
1568 	rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET,
1569 						filter, handler, arg, &tag);
1570 	if (rc != 0) {
1571 		device_printf(dev,
1572 		    "failed to setup interrupt for rid %d, name %s: %d\n",
1573 					  rid, name ? name : "unknown", rc);
1574 		return (rc);
1575 	} else if (name)
1576 		bus_describe_intr(dev, res, tag, "%s", name);
1577 
1578 	irq->ii_tag = tag;
1579 	return (0);
1580 }
1581 
1582 
1583 /*********************************************************************
1584  *
1585  *  Allocate memory for tx_buffer structures. The tx_buffer stores all
1586  *  the information needed to transmit a packet on the wire. This is
1587  *  called only once at attach, setup is done every reset.
1588  *
1589  **********************************************************************/
1590 
1591 static int
1592 iflib_txsd_alloc(iflib_txq_t txq)
1593 {
1594 	if_ctx_t ctx = txq->ift_ctx;
1595 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1596 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1597 	device_t dev = ctx->ifc_dev;
1598 	bus_size_t tsomaxsize;
1599 	int err, nsegments, ntsosegments;
1600 
1601 	nsegments = scctx->isc_tx_nsegments;
1602 	ntsosegments = scctx->isc_tx_tso_segments_max;
1603 	tsomaxsize = scctx->isc_tx_tso_size_max;
1604 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_VLAN_MTU)
1605 		tsomaxsize += sizeof(struct ether_vlan_header);
1606 	MPASS(scctx->isc_ntxd[0] > 0);
1607 	MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0);
1608 	MPASS(nsegments > 0);
1609 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) {
1610 		MPASS(ntsosegments > 0);
1611 		MPASS(sctx->isc_tso_maxsize >= tsomaxsize);
1612 	}
1613 
1614 	/*
1615 	 * Setup DMA descriptor areas.
1616 	 */
1617 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1618 			       1, 0,			/* alignment, bounds */
1619 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1620 			       BUS_SPACE_MAXADDR,	/* highaddr */
1621 			       NULL, NULL,		/* filter, filterarg */
1622 			       sctx->isc_tx_maxsize,		/* maxsize */
1623 			       nsegments,	/* nsegments */
1624 			       sctx->isc_tx_maxsegsize,	/* maxsegsize */
1625 			       0,			/* flags */
1626 			       NULL,			/* lockfunc */
1627 			       NULL,			/* lockfuncarg */
1628 			       &txq->ift_desc_tag))) {
1629 		device_printf(dev,"Unable to allocate TX DMA tag: %d\n", err);
1630 		device_printf(dev,"maxsize: %ju nsegments: %d maxsegsize: %ju\n",
1631 		    (uintmax_t)sctx->isc_tx_maxsize, nsegments, (uintmax_t)sctx->isc_tx_maxsegsize);
1632 		goto fail;
1633 	}
1634 	if ((if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) &
1635 	    (err = bus_dma_tag_create(bus_get_dma_tag(dev),
1636 			       1, 0,			/* alignment, bounds */
1637 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1638 			       BUS_SPACE_MAXADDR,	/* highaddr */
1639 			       NULL, NULL,		/* filter, filterarg */
1640 			       tsomaxsize,		/* maxsize */
1641 			       ntsosegments,	/* nsegments */
1642 			       sctx->isc_tso_maxsegsize,/* maxsegsize */
1643 			       0,			/* flags */
1644 			       NULL,			/* lockfunc */
1645 			       NULL,			/* lockfuncarg */
1646 			       &txq->ift_tso_desc_tag))) {
1647 		device_printf(dev,"Unable to allocate TX TSO DMA tag: %d\n", err);
1648 
1649 		goto fail;
1650 	}
1651 	if (!(txq->ift_sds.ifsd_flags =
1652 	    (uint8_t *) malloc(sizeof(uint8_t) *
1653 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1654 		device_printf(dev, "Unable to allocate tx_buffer memory\n");
1655 		err = ENOMEM;
1656 		goto fail;
1657 	}
1658 	if (!(txq->ift_sds.ifsd_m =
1659 	    (struct mbuf **) malloc(sizeof(struct mbuf *) *
1660 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1661 		device_printf(dev, "Unable to allocate tx_buffer memory\n");
1662 		err = ENOMEM;
1663 		goto fail;
1664 	}
1665 
1666         /* Create the descriptor buffer dma maps */
1667 #if defined(ACPI_DMAR) || (! (defined(__i386__) || defined(__amd64__)))
1668 	if ((ctx->ifc_flags & IFC_DMAR) == 0)
1669 		return (0);
1670 
1671 	if (!(txq->ift_sds.ifsd_map =
1672 	    (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1673 		device_printf(dev, "Unable to allocate tx_buffer map memory\n");
1674 		err = ENOMEM;
1675 		goto fail;
1676 	}
1677 
1678 	for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) {
1679 		err = bus_dmamap_create(txq->ift_desc_tag, 0, &txq->ift_sds.ifsd_map[i]);
1680 		if (err != 0) {
1681 			device_printf(dev, "Unable to create TX DMA map\n");
1682 			goto fail;
1683 		}
1684 	}
1685 #endif
1686 	return (0);
1687 fail:
1688 	/* We free all, it handles case where we are in the middle */
1689 	iflib_tx_structures_free(ctx);
1690 	return (err);
1691 }
1692 
1693 static void
1694 iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i)
1695 {
1696 	bus_dmamap_t map;
1697 
1698 	map = NULL;
1699 	if (txq->ift_sds.ifsd_map != NULL)
1700 		map = txq->ift_sds.ifsd_map[i];
1701 	if (map != NULL) {
1702 		bus_dmamap_unload(txq->ift_desc_tag, map);
1703 		bus_dmamap_destroy(txq->ift_desc_tag, map);
1704 		txq->ift_sds.ifsd_map[i] = NULL;
1705 	}
1706 }
1707 
1708 static void
1709 iflib_txq_destroy(iflib_txq_t txq)
1710 {
1711 	if_ctx_t ctx = txq->ift_ctx;
1712 
1713 	for (int i = 0; i < txq->ift_size; i++)
1714 		iflib_txsd_destroy(ctx, txq, i);
1715 	if (txq->ift_sds.ifsd_map != NULL) {
1716 		free(txq->ift_sds.ifsd_map, M_IFLIB);
1717 		txq->ift_sds.ifsd_map = NULL;
1718 	}
1719 	if (txq->ift_sds.ifsd_m != NULL) {
1720 		free(txq->ift_sds.ifsd_m, M_IFLIB);
1721 		txq->ift_sds.ifsd_m = NULL;
1722 	}
1723 	if (txq->ift_sds.ifsd_flags != NULL) {
1724 		free(txq->ift_sds.ifsd_flags, M_IFLIB);
1725 		txq->ift_sds.ifsd_flags = NULL;
1726 	}
1727 	if (txq->ift_desc_tag != NULL) {
1728 		bus_dma_tag_destroy(txq->ift_desc_tag);
1729 		txq->ift_desc_tag = NULL;
1730 	}
1731 	if (txq->ift_tso_desc_tag != NULL) {
1732 		bus_dma_tag_destroy(txq->ift_tso_desc_tag);
1733 		txq->ift_tso_desc_tag = NULL;
1734 	}
1735 }
1736 
1737 static void
1738 iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i)
1739 {
1740 	struct mbuf **mp;
1741 
1742 	mp = &txq->ift_sds.ifsd_m[i];
1743 	if (*mp == NULL)
1744 		return;
1745 
1746 	if (txq->ift_sds.ifsd_map != NULL) {
1747 		bus_dmamap_sync(txq->ift_desc_tag,
1748 				txq->ift_sds.ifsd_map[i],
1749 				BUS_DMASYNC_POSTWRITE);
1750 		bus_dmamap_unload(txq->ift_desc_tag,
1751 				  txq->ift_sds.ifsd_map[i]);
1752 	}
1753 	m_free(*mp);
1754 	DBG_COUNTER_INC(tx_frees);
1755 	*mp = NULL;
1756 }
1757 
1758 static int
1759 iflib_txq_setup(iflib_txq_t txq)
1760 {
1761 	if_ctx_t ctx = txq->ift_ctx;
1762 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1763 	iflib_dma_info_t di;
1764 	int i;
1765 
1766 	/* Set number of descriptors available */
1767 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
1768 	/* XXX make configurable */
1769 	txq->ift_update_freq = IFLIB_DEFAULT_TX_UPDATE_FREQ;
1770 
1771 	/* Reset indices */
1772 	txq->ift_cidx_processed = 0;
1773 	txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0;
1774 	txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset];
1775 
1776 	for (i = 0, di = txq->ift_ifdi; i < ctx->ifc_nhwtxqs; i++, di++)
1777 		bzero((void *)di->idi_vaddr, di->idi_size);
1778 
1779 	IFDI_TXQ_SETUP(ctx, txq->ift_id);
1780 	for (i = 0, di = txq->ift_ifdi; i < ctx->ifc_nhwtxqs; i++, di++)
1781 		bus_dmamap_sync(di->idi_tag, di->idi_map,
1782 						BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1783 	return (0);
1784 }
1785 
1786 /*********************************************************************
1787  *
1788  *  Allocate memory for rx_buffer structures. Since we use one
1789  *  rx_buffer per received packet, the maximum number of rx_buffer's
1790  *  that we'll need is equal to the number of receive descriptors
1791  *  that we've allocated.
1792  *
1793  **********************************************************************/
1794 static int
1795 iflib_rxsd_alloc(iflib_rxq_t rxq)
1796 {
1797 	if_ctx_t ctx = rxq->ifr_ctx;
1798 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1799 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1800 	device_t dev = ctx->ifc_dev;
1801 	iflib_fl_t fl;
1802 	int			err;
1803 
1804 	MPASS(scctx->isc_nrxd[0] > 0);
1805 	MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0);
1806 
1807 	fl = rxq->ifr_fl;
1808 	for (int i = 0; i <  rxq->ifr_nfl; i++, fl++) {
1809 		fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */
1810 		err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1811 					 1, 0,			/* alignment, bounds */
1812 					 BUS_SPACE_MAXADDR,	/* lowaddr */
1813 					 BUS_SPACE_MAXADDR,	/* highaddr */
1814 					 NULL, NULL,		/* filter, filterarg */
1815 					 sctx->isc_rx_maxsize,	/* maxsize */
1816 					 sctx->isc_rx_nsegments,	/* nsegments */
1817 					 sctx->isc_rx_maxsegsize,	/* maxsegsize */
1818 					 0,			/* flags */
1819 					 NULL,			/* lockfunc */
1820 					 NULL,			/* lockarg */
1821 					 &fl->ifl_desc_tag);
1822 		if (err) {
1823 			device_printf(dev, "%s: bus_dma_tag_create failed %d\n",
1824 				__func__, err);
1825 			goto fail;
1826 		}
1827 		if (!(fl->ifl_sds.ifsd_flags =
1828 		      (uint8_t *) malloc(sizeof(uint8_t) *
1829 					 scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1830 			device_printf(dev, "Unable to allocate tx_buffer memory\n");
1831 			err = ENOMEM;
1832 			goto fail;
1833 		}
1834 		if (!(fl->ifl_sds.ifsd_m =
1835 		      (struct mbuf **) malloc(sizeof(struct mbuf *) *
1836 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1837 			device_printf(dev, "Unable to allocate tx_buffer memory\n");
1838 			err = ENOMEM;
1839 			goto fail;
1840 		}
1841 		if (!(fl->ifl_sds.ifsd_cl =
1842 		      (caddr_t *) malloc(sizeof(caddr_t) *
1843 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1844 			device_printf(dev, "Unable to allocate tx_buffer memory\n");
1845 			err = ENOMEM;
1846 			goto fail;
1847 		}
1848 
1849 		/* Create the descriptor buffer dma maps */
1850 #if defined(ACPI_DMAR) || (! (defined(__i386__) || defined(__amd64__)))
1851 		if ((ctx->ifc_flags & IFC_DMAR) == 0)
1852 			continue;
1853 
1854 		if (!(fl->ifl_sds.ifsd_map =
1855 		      (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1856 			device_printf(dev, "Unable to allocate tx_buffer map memory\n");
1857 			err = ENOMEM;
1858 			goto fail;
1859 		}
1860 
1861 		for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++) {
1862 			err = bus_dmamap_create(fl->ifl_desc_tag, 0, &fl->ifl_sds.ifsd_map[i]);
1863 			if (err != 0) {
1864 				device_printf(dev, "Unable to create RX buffer DMA map\n");
1865 				goto fail;
1866 			}
1867 		}
1868 #endif
1869 	}
1870 	return (0);
1871 
1872 fail:
1873 	iflib_rx_structures_free(ctx);
1874 	return (err);
1875 }
1876 
1877 
1878 /*
1879  * Internal service routines
1880  */
1881 
1882 struct rxq_refill_cb_arg {
1883 	int               error;
1884 	bus_dma_segment_t seg;
1885 	int               nseg;
1886 };
1887 
1888 static void
1889 _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
1890 {
1891 	struct rxq_refill_cb_arg *cb_arg = arg;
1892 
1893 	cb_arg->error = error;
1894 	cb_arg->seg = segs[0];
1895 	cb_arg->nseg = nseg;
1896 }
1897 
1898 
1899 #ifdef ACPI_DMAR
1900 #define IS_DMAR(ctx) (ctx->ifc_flags & IFC_DMAR)
1901 #else
1902 #define IS_DMAR(ctx) (0)
1903 #endif
1904 
1905 /**
1906  *	rxq_refill - refill an rxq  free-buffer list
1907  *	@ctx: the iflib context
1908  *	@rxq: the free-list to refill
1909  *	@n: the number of new buffers to allocate
1910  *
1911  *	(Re)populate an rxq free-buffer list with up to @n new packet buffers.
1912  *	The caller must assure that @n does not exceed the queue's capacity.
1913  */
1914 static void
1915 _iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
1916 {
1917 	struct mbuf *m;
1918 	int idx, frag_idx = fl->ifl_fragidx;
1919         int pidx = fl->ifl_pidx;
1920 	caddr_t cl, *sd_cl;
1921 	struct mbuf **sd_m;
1922 	uint8_t *sd_flags;
1923 	struct if_rxd_update iru;
1924 	bus_dmamap_t *sd_map;
1925 	int n, i = 0;
1926 	uint64_t bus_addr;
1927 	int err;
1928 	qidx_t credits;
1929 
1930 	sd_m = fl->ifl_sds.ifsd_m;
1931 	sd_map = fl->ifl_sds.ifsd_map;
1932 	sd_cl = fl->ifl_sds.ifsd_cl;
1933 	sd_flags = fl->ifl_sds.ifsd_flags;
1934 	idx = pidx;
1935 	credits = fl->ifl_credits;
1936 
1937 	n  = count;
1938 	MPASS(n > 0);
1939 	MPASS(credits + n <= fl->ifl_size);
1940 
1941 	if (pidx < fl->ifl_cidx)
1942 		MPASS(pidx + n <= fl->ifl_cidx);
1943 	if (pidx == fl->ifl_cidx && (credits < fl->ifl_size))
1944 		MPASS(fl->ifl_gen == 0);
1945 	if (pidx > fl->ifl_cidx)
1946 		MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx);
1947 
1948 	DBG_COUNTER_INC(fl_refills);
1949 	if (n > 8)
1950 		DBG_COUNTER_INC(fl_refills_large);
1951 	iru_init(&iru, fl->ifl_rxq, fl->ifl_id);
1952 	while (n--) {
1953 		/*
1954 		 * We allocate an uninitialized mbuf + cluster, mbuf is
1955 		 * initialized after rx.
1956 		 *
1957 		 * If the cluster is still set then we know a minimum sized packet was received
1958 		 */
1959 		bit_ffc_at(fl->ifl_rx_bitmap, frag_idx, fl->ifl_size,  &frag_idx);
1960 		if ((frag_idx < 0) || (frag_idx >= fl->ifl_size))
1961                 	bit_ffc(fl->ifl_rx_bitmap, fl->ifl_size, &frag_idx);
1962 		if ((cl = sd_cl[frag_idx]) == NULL) {
1963                        if ((cl = sd_cl[frag_idx] = m_cljget(NULL, M_NOWAIT, fl->ifl_buf_size)) == NULL)
1964 				break;
1965 #if MEMORY_LOGGING
1966 			fl->ifl_cl_enqueued++;
1967 #endif
1968 		}
1969 		if ((m = m_gethdr(M_NOWAIT, MT_NOINIT)) == NULL) {
1970 			break;
1971 		}
1972 #if MEMORY_LOGGING
1973 		fl->ifl_m_enqueued++;
1974 #endif
1975 
1976 		DBG_COUNTER_INC(rx_allocs);
1977 #if defined(__i386__) || defined(__amd64__)
1978 		if (!IS_DMAR(ctx)) {
1979 			bus_addr = pmap_kextract((vm_offset_t)cl);
1980 		} else
1981 #endif
1982 		{
1983 			struct rxq_refill_cb_arg cb_arg;
1984 
1985 			cb_arg.error = 0;
1986 			MPASS(sd_map != NULL);
1987 			MPASS(sd_map[frag_idx] != NULL);
1988 			err = bus_dmamap_load(fl->ifl_desc_tag, sd_map[frag_idx],
1989 		         cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg, 0);
1990 			bus_dmamap_sync(fl->ifl_desc_tag, sd_map[frag_idx],
1991 					BUS_DMASYNC_PREREAD);
1992 
1993 			if (err != 0 || cb_arg.error) {
1994 				/*
1995 				 * !zone_pack ?
1996 				 */
1997 				if (fl->ifl_zone == zone_pack)
1998 					uma_zfree(fl->ifl_zone, cl);
1999 				m_free(m);
2000 				n = 0;
2001 				goto done;
2002 			}
2003 			bus_addr = cb_arg.seg.ds_addr;
2004 		}
2005                 bit_set(fl->ifl_rx_bitmap, frag_idx);
2006 		sd_flags[frag_idx] |= RX_SW_DESC_INUSE;
2007 
2008 		MPASS(sd_m[frag_idx] == NULL);
2009 		sd_cl[frag_idx] = cl;
2010 		sd_m[frag_idx] = m;
2011 		fl->ifl_rxd_idxs[i] = frag_idx;
2012 		fl->ifl_bus_addrs[i] = bus_addr;
2013 		fl->ifl_vm_addrs[i] = cl;
2014 		credits++;
2015 		i++;
2016 		MPASS(credits <= fl->ifl_size);
2017 		if (++idx == fl->ifl_size) {
2018 			fl->ifl_gen = 1;
2019 			idx = 0;
2020 		}
2021 		if (n == 0 || i == IFLIB_MAX_RX_REFRESH) {
2022 			iru.iru_pidx = pidx;
2023 			iru.iru_count = i;
2024 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2025 			i = 0;
2026 			pidx = idx;
2027 			fl->ifl_pidx = idx;
2028 			fl->ifl_credits = credits;
2029 		}
2030 
2031 	}
2032 done:
2033 	if (i) {
2034 		iru.iru_pidx = pidx;
2035 		iru.iru_count = i;
2036 		ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2037 		fl->ifl_pidx = idx;
2038 		fl->ifl_credits = credits;
2039 	}
2040 	DBG_COUNTER_INC(rxd_flush);
2041 	if (fl->ifl_pidx == 0)
2042 		pidx = fl->ifl_size - 1;
2043 	else
2044 		pidx = fl->ifl_pidx - 1;
2045 
2046 	if (sd_map)
2047 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2048 				BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2049 	ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id, fl->ifl_id, pidx);
2050 	fl->ifl_fragidx = frag_idx;
2051 }
2052 
2053 static __inline void
2054 __iflib_fl_refill_lt(if_ctx_t ctx, iflib_fl_t fl, int max)
2055 {
2056 	/* we avoid allowing pidx to catch up with cidx as it confuses ixl */
2057 	int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1;
2058 #ifdef INVARIANTS
2059 	int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1;
2060 #endif
2061 
2062 	MPASS(fl->ifl_credits <= fl->ifl_size);
2063 	MPASS(reclaimable == delta);
2064 
2065 	if (reclaimable > 0)
2066 		_iflib_fl_refill(ctx, fl, min(max, reclaimable));
2067 }
2068 
2069 static void
2070 iflib_fl_bufs_free(iflib_fl_t fl)
2071 {
2072 	iflib_dma_info_t idi = fl->ifl_ifdi;
2073 	uint32_t i;
2074 
2075 	for (i = 0; i < fl->ifl_size; i++) {
2076 		struct mbuf **sd_m = &fl->ifl_sds.ifsd_m[i];
2077 		uint8_t *sd_flags = &fl->ifl_sds.ifsd_flags[i];
2078 		caddr_t *sd_cl = &fl->ifl_sds.ifsd_cl[i];
2079 
2080 		if (*sd_flags & RX_SW_DESC_INUSE) {
2081 			if (fl->ifl_sds.ifsd_map != NULL) {
2082 				bus_dmamap_t sd_map = fl->ifl_sds.ifsd_map[i];
2083 				bus_dmamap_unload(fl->ifl_desc_tag, sd_map);
2084 				if (fl->ifl_rxq->ifr_ctx->ifc_in_detach)
2085 					bus_dmamap_destroy(fl->ifl_desc_tag, sd_map);
2086 			}
2087 			if (*sd_m != NULL) {
2088 				m_init(*sd_m, M_NOWAIT, MT_DATA, 0);
2089 				uma_zfree(zone_mbuf, *sd_m);
2090 			}
2091 			if (*sd_cl != NULL)
2092 				uma_zfree(fl->ifl_zone, *sd_cl);
2093 			*sd_flags = 0;
2094 		} else {
2095 			MPASS(*sd_cl == NULL);
2096 			MPASS(*sd_m == NULL);
2097 		}
2098 #if MEMORY_LOGGING
2099 		fl->ifl_m_dequeued++;
2100 		fl->ifl_cl_dequeued++;
2101 #endif
2102 		*sd_cl = NULL;
2103 		*sd_m = NULL;
2104 	}
2105 #ifdef INVARIANTS
2106 	for (i = 0; i < fl->ifl_size; i++) {
2107 		MPASS(fl->ifl_sds.ifsd_flags[i] == 0);
2108 		MPASS(fl->ifl_sds.ifsd_cl[i] == NULL);
2109 		MPASS(fl->ifl_sds.ifsd_m[i] == NULL);
2110 	}
2111 #endif
2112 	/*
2113 	 * Reset free list values
2114 	 */
2115 	fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = fl->ifl_fragidx = 0;
2116 	bzero(idi->idi_vaddr, idi->idi_size);
2117 }
2118 
2119 /*********************************************************************
2120  *
2121  *  Initialize a receive ring and its buffers.
2122  *
2123  **********************************************************************/
2124 static int
2125 iflib_fl_setup(iflib_fl_t fl)
2126 {
2127 	iflib_rxq_t rxq = fl->ifl_rxq;
2128 	if_ctx_t ctx = rxq->ifr_ctx;
2129 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2130 
2131 	bit_nclear(fl->ifl_rx_bitmap, 0, fl->ifl_size - 1);
2132 	/*
2133 	** Free current RX buffer structs and their mbufs
2134 	*/
2135 	iflib_fl_bufs_free(fl);
2136 	/* Now replenish the mbufs */
2137 	MPASS(fl->ifl_credits == 0);
2138 	/*
2139 	 * XXX don't set the max_frame_size to larger
2140 	 * than the hardware can handle
2141 	 */
2142 	if (sctx->isc_max_frame_size <= 2048)
2143 		fl->ifl_buf_size = MCLBYTES;
2144 #ifndef CONTIGMALLOC_WORKS
2145 	else
2146 		fl->ifl_buf_size = MJUMPAGESIZE;
2147 #else
2148 	else if (sctx->isc_max_frame_size <= 4096)
2149 		fl->ifl_buf_size = MJUMPAGESIZE;
2150 	else if (sctx->isc_max_frame_size <= 9216)
2151 		fl->ifl_buf_size = MJUM9BYTES;
2152 	else
2153 		fl->ifl_buf_size = MJUM16BYTES;
2154 #endif
2155 	if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size)
2156 		ctx->ifc_max_fl_buf_size = fl->ifl_buf_size;
2157 	fl->ifl_cltype = m_gettype(fl->ifl_buf_size);
2158 	fl->ifl_zone = m_getzone(fl->ifl_buf_size);
2159 
2160 
2161 	/* avoid pre-allocating zillions of clusters to an idle card
2162 	 * potentially speeding up attach
2163 	 */
2164 	_iflib_fl_refill(ctx, fl, min(128, fl->ifl_size));
2165 	MPASS(min(128, fl->ifl_size) == fl->ifl_credits);
2166 	if (min(128, fl->ifl_size) != fl->ifl_credits)
2167 		return (ENOBUFS);
2168 	/*
2169 	 * handle failure
2170 	 */
2171 	MPASS(rxq != NULL);
2172 	MPASS(fl->ifl_ifdi != NULL);
2173 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2174 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2175 	return (0);
2176 }
2177 
2178 /*********************************************************************
2179  *
2180  *  Free receive ring data structures
2181  *
2182  **********************************************************************/
2183 static void
2184 iflib_rx_sds_free(iflib_rxq_t rxq)
2185 {
2186 	iflib_fl_t fl;
2187 	int i;
2188 
2189 	if (rxq->ifr_fl != NULL) {
2190 		for (i = 0; i < rxq->ifr_nfl; i++) {
2191 			fl = &rxq->ifr_fl[i];
2192 			if (fl->ifl_desc_tag != NULL) {
2193 				bus_dma_tag_destroy(fl->ifl_desc_tag);
2194 				fl->ifl_desc_tag = NULL;
2195 			}
2196 			free(fl->ifl_sds.ifsd_m, M_IFLIB);
2197 			free(fl->ifl_sds.ifsd_cl, M_IFLIB);
2198 			/* XXX destroy maps first */
2199 			free(fl->ifl_sds.ifsd_map, M_IFLIB);
2200 			fl->ifl_sds.ifsd_m = NULL;
2201 			fl->ifl_sds.ifsd_cl = NULL;
2202 			fl->ifl_sds.ifsd_map = NULL;
2203 		}
2204 		free(rxq->ifr_fl, M_IFLIB);
2205 		rxq->ifr_fl = NULL;
2206 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
2207 	}
2208 }
2209 
2210 /*
2211  * MI independent logic
2212  *
2213  */
2214 static void
2215 iflib_timer(void *arg)
2216 {
2217 	iflib_txq_t txq = arg;
2218 	if_ctx_t ctx = txq->ift_ctx;
2219 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2220 	uint64_t this_tick = ticks;
2221 	uint32_t reset_on = hz / 2;
2222 
2223 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
2224 		return;
2225 	/*
2226 	** Check on the state of the TX queue(s), this
2227 	** can be done without the lock because its RO
2228 	** and the HUNG state will be static if set.
2229 	*/
2230 	if (this_tick - txq->ift_last_timer_tick >= hz / 2) {
2231 		txq->ift_last_timer_tick = this_tick;
2232 		IFDI_TIMER(ctx, txq->ift_id);
2233 		if ((txq->ift_qstatus == IFLIB_QUEUE_HUNG) &&
2234 		    ((txq->ift_cleaned_prev == txq->ift_cleaned) ||
2235 		     (sctx->isc_pause_frames == 0)))
2236 			goto hung;
2237 
2238 		if (ifmp_ring_is_stalled(txq->ift_br))
2239 			txq->ift_qstatus = IFLIB_QUEUE_HUNG;
2240 		txq->ift_cleaned_prev = txq->ift_cleaned;
2241 	}
2242 #ifdef DEV_NETMAP
2243 	if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP)
2244 		iflib_netmap_timer_adjust(ctx, txq->ift_id, &reset_on);
2245 #endif
2246 	/* handle any laggards */
2247 	if (txq->ift_db_pending)
2248 		GROUPTASK_ENQUEUE(&txq->ift_task);
2249 
2250 	sctx->isc_pause_frames = 0;
2251 	if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)
2252 		callout_reset_on(&txq->ift_timer, reset_on, iflib_timer, txq, txq->ift_timer.c_cpu);
2253 	return;
2254  hung:
2255 	device_printf(ctx->ifc_dev,  "TX(%d) desc avail = %d, pidx = %d\n",
2256 				  txq->ift_id, TXQ_AVAIL(txq), txq->ift_pidx);
2257 	STATE_LOCK(ctx);
2258 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2259 	ctx->ifc_flags |= (IFC_DO_WATCHDOG|IFC_DO_RESET);
2260 	iflib_admin_intr_deferred(ctx);
2261 	STATE_UNLOCK(ctx);
2262 }
2263 
2264 static void
2265 iflib_init_locked(if_ctx_t ctx)
2266 {
2267 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2268 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2269 	if_t ifp = ctx->ifc_ifp;
2270 	iflib_fl_t fl;
2271 	iflib_txq_t txq;
2272 	iflib_rxq_t rxq;
2273 	int i, j, tx_ip_csum_flags, tx_ip6_csum_flags;
2274 
2275 
2276 	if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2277 	IFDI_INTR_DISABLE(ctx);
2278 
2279 	tx_ip_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP);
2280 	tx_ip6_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | CSUM_IP6_SCTP);
2281 	/* Set hardware offload abilities */
2282 	if_clearhwassist(ifp);
2283 	if (if_getcapenable(ifp) & IFCAP_TXCSUM)
2284 		if_sethwassistbits(ifp, tx_ip_csum_flags, 0);
2285 	if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6)
2286 		if_sethwassistbits(ifp,  tx_ip6_csum_flags, 0);
2287 	if (if_getcapenable(ifp) & IFCAP_TSO4)
2288 		if_sethwassistbits(ifp, CSUM_IP_TSO, 0);
2289 	if (if_getcapenable(ifp) & IFCAP_TSO6)
2290 		if_sethwassistbits(ifp, CSUM_IP6_TSO, 0);
2291 
2292 	for (i = 0, txq = ctx->ifc_txqs; i < sctx->isc_ntxqsets; i++, txq++) {
2293 		CALLOUT_LOCK(txq);
2294 		callout_stop(&txq->ift_timer);
2295 		CALLOUT_UNLOCK(txq);
2296 		iflib_netmap_txq_init(ctx, txq);
2297 	}
2298 #ifdef INVARIANTS
2299 	i = if_getdrvflags(ifp);
2300 #endif
2301 	IFDI_INIT(ctx);
2302 	MPASS(if_getdrvflags(ifp) == i);
2303 	for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
2304 		/* XXX this should really be done on a per-queue basis */
2305 		if (if_getcapenable(ifp) & IFCAP_NETMAP) {
2306 			MPASS(rxq->ifr_id == i);
2307 			iflib_netmap_rxq_init(ctx, rxq);
2308 			continue;
2309 		}
2310 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
2311 			if (iflib_fl_setup(fl)) {
2312 				device_printf(ctx->ifc_dev, "freelist setup failed - check cluster settings\n");
2313 				goto done;
2314 			}
2315 		}
2316 	}
2317 done:
2318 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
2319 	IFDI_INTR_ENABLE(ctx);
2320 	txq = ctx->ifc_txqs;
2321 	for (i = 0; i < sctx->isc_ntxqsets; i++, txq++)
2322 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq,
2323 			txq->ift_timer.c_cpu);
2324 }
2325 
2326 static int
2327 iflib_media_change(if_t ifp)
2328 {
2329 	if_ctx_t ctx = if_getsoftc(ifp);
2330 	int err;
2331 
2332 	CTX_LOCK(ctx);
2333 	if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0)
2334 		iflib_init_locked(ctx);
2335 	CTX_UNLOCK(ctx);
2336 	return (err);
2337 }
2338 
2339 static void
2340 iflib_media_status(if_t ifp, struct ifmediareq *ifmr)
2341 {
2342 	if_ctx_t ctx = if_getsoftc(ifp);
2343 
2344 	CTX_LOCK(ctx);
2345 	IFDI_UPDATE_ADMIN_STATUS(ctx);
2346 	IFDI_MEDIA_STATUS(ctx, ifmr);
2347 	CTX_UNLOCK(ctx);
2348 }
2349 
2350 void
2351 iflib_stop(if_ctx_t ctx)
2352 {
2353 	iflib_txq_t txq = ctx->ifc_txqs;
2354 	iflib_rxq_t rxq = ctx->ifc_rxqs;
2355 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2356 	iflib_dma_info_t di;
2357 	iflib_fl_t fl;
2358 	int i, j;
2359 
2360 	/* Tell the stack that the interface is no longer active */
2361 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2362 
2363 	IFDI_INTR_DISABLE(ctx);
2364 	DELAY(1000);
2365 	IFDI_STOP(ctx);
2366 	DELAY(1000);
2367 
2368 	iflib_debug_reset();
2369 	/* Wait for current tx queue users to exit to disarm watchdog timer. */
2370 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) {
2371 		/* make sure all transmitters have completed before proceeding XXX */
2372 
2373 		CALLOUT_LOCK(txq);
2374 		callout_stop(&txq->ift_timer);
2375 		CALLOUT_UNLOCK(txq);
2376 
2377 		/* clean any enqueued buffers */
2378 		iflib_ifmp_purge(txq);
2379 		/* Free any existing tx buffers. */
2380 		for (j = 0; j < txq->ift_size; j++) {
2381 			iflib_txsd_free(ctx, txq, j);
2382 		}
2383 		txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0;
2384 		txq->ift_in_use = txq->ift_gen = txq->ift_cidx = txq->ift_pidx = txq->ift_no_desc_avail = 0;
2385 		txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0;
2386 		txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0;
2387 		txq->ift_pullups = 0;
2388 		ifmp_ring_reset_stats(txq->ift_br);
2389 		for (j = 0, di = txq->ift_ifdi; j < ctx->ifc_nhwtxqs; j++, di++)
2390 			bzero((void *)di->idi_vaddr, di->idi_size);
2391 	}
2392 	for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) {
2393 		/* make sure all transmitters have completed before proceeding XXX */
2394 
2395 		for (j = 0, di = rxq->ifr_ifdi; j < rxq->ifr_nfl; j++, di++)
2396 			bzero((void *)di->idi_vaddr, di->idi_size);
2397 		/* also resets the free lists pidx/cidx */
2398 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
2399 			iflib_fl_bufs_free(fl);
2400 	}
2401 }
2402 
2403 static inline caddr_t
2404 calc_next_rxd(iflib_fl_t fl, int cidx)
2405 {
2406 	qidx_t size;
2407 	int nrxd;
2408 	caddr_t start, end, cur, next;
2409 
2410 	nrxd = fl->ifl_size;
2411 	size = fl->ifl_rxd_size;
2412 	start = fl->ifl_ifdi->idi_vaddr;
2413 
2414 	if (__predict_false(size == 0))
2415 		return (start);
2416 	cur = start + size*cidx;
2417 	end = start + size*nrxd;
2418 	next = CACHE_PTR_NEXT(cur);
2419 	return (next < end ? next : start);
2420 }
2421 
2422 static inline void
2423 prefetch_pkts(iflib_fl_t fl, int cidx)
2424 {
2425 	int nextptr;
2426 	int nrxd = fl->ifl_size;
2427 	caddr_t next_rxd;
2428 
2429 
2430 	nextptr = (cidx + CACHE_PTR_INCREMENT) & (nrxd-1);
2431 	prefetch(&fl->ifl_sds.ifsd_m[nextptr]);
2432 	prefetch(&fl->ifl_sds.ifsd_cl[nextptr]);
2433 	next_rxd = calc_next_rxd(fl, cidx);
2434 	prefetch(next_rxd);
2435 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 1) & (nrxd-1)]);
2436 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 2) & (nrxd-1)]);
2437 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 3) & (nrxd-1)]);
2438 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 4) & (nrxd-1)]);
2439 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 1) & (nrxd-1)]);
2440 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 2) & (nrxd-1)]);
2441 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 3) & (nrxd-1)]);
2442 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 4) & (nrxd-1)]);
2443 }
2444 
2445 static void
2446 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, int unload, if_rxsd_t sd)
2447 {
2448 	int flid, cidx;
2449 	bus_dmamap_t map;
2450 	iflib_fl_t fl;
2451 	iflib_dma_info_t di;
2452 	int next;
2453 
2454 	map = NULL;
2455 	flid = irf->irf_flid;
2456 	cidx = irf->irf_idx;
2457 	fl = &rxq->ifr_fl[flid];
2458 	sd->ifsd_fl = fl;
2459 	sd->ifsd_cidx = cidx;
2460 	sd->ifsd_m = &fl->ifl_sds.ifsd_m[cidx];
2461 	sd->ifsd_cl = &fl->ifl_sds.ifsd_cl[cidx];
2462 	fl->ifl_credits--;
2463 #if MEMORY_LOGGING
2464 	fl->ifl_m_dequeued++;
2465 #endif
2466 	if (rxq->ifr_ctx->ifc_flags & IFC_PREFETCH)
2467 		prefetch_pkts(fl, cidx);
2468 	if (fl->ifl_sds.ifsd_map != NULL) {
2469 		next = (cidx + CACHE_PTR_INCREMENT) & (fl->ifl_size-1);
2470 		prefetch(&fl->ifl_sds.ifsd_map[next]);
2471 		map = fl->ifl_sds.ifsd_map[cidx];
2472 		di = fl->ifl_ifdi;
2473 		next = (cidx + CACHE_LINE_SIZE) & (fl->ifl_size-1);
2474 		prefetch(&fl->ifl_sds.ifsd_flags[next]);
2475 		bus_dmamap_sync(di->idi_tag, di->idi_map,
2476 				BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
2477 
2478 	/* not valid assert if bxe really does SGE from non-contiguous elements */
2479 		MPASS(fl->ifl_cidx == cidx);
2480 		if (unload)
2481 			bus_dmamap_unload(fl->ifl_desc_tag, map);
2482 	}
2483 	fl->ifl_cidx = (fl->ifl_cidx + 1) & (fl->ifl_size-1);
2484 	if (__predict_false(fl->ifl_cidx == 0))
2485 		fl->ifl_gen = 0;
2486 	if (map != NULL)
2487 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2488 			BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2489         bit_clear(fl->ifl_rx_bitmap, cidx);
2490 }
2491 
2492 static struct mbuf *
2493 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri, if_rxsd_t sd)
2494 {
2495 	int i, padlen , flags;
2496 	struct mbuf *m, *mh, *mt;
2497 	caddr_t cl;
2498 
2499 	i = 0;
2500 	mh = NULL;
2501 	do {
2502 		rxd_frag_to_sd(rxq, &ri->iri_frags[i], TRUE, sd);
2503 
2504 		MPASS(*sd->ifsd_cl != NULL);
2505 		MPASS(*sd->ifsd_m != NULL);
2506 
2507 		/* Don't include zero-length frags */
2508 		if (ri->iri_frags[i].irf_len == 0) {
2509 			/* XXX we can save the cluster here, but not the mbuf */
2510 			m_init(*sd->ifsd_m, M_NOWAIT, MT_DATA, 0);
2511 			m_free(*sd->ifsd_m);
2512 			*sd->ifsd_m = NULL;
2513 			continue;
2514 		}
2515 		m = *sd->ifsd_m;
2516 		*sd->ifsd_m = NULL;
2517 		if (mh == NULL) {
2518 			flags = M_PKTHDR|M_EXT;
2519 			mh = mt = m;
2520 			padlen = ri->iri_pad;
2521 		} else {
2522 			flags = M_EXT;
2523 			mt->m_next = m;
2524 			mt = m;
2525 			/* assuming padding is only on the first fragment */
2526 			padlen = 0;
2527 		}
2528 		cl = *sd->ifsd_cl;
2529 		*sd->ifsd_cl = NULL;
2530 
2531 		/* Can these two be made one ? */
2532 		m_init(m, M_NOWAIT, MT_DATA, flags);
2533 		m_cljset(m, cl, sd->ifsd_fl->ifl_cltype);
2534 		/*
2535 		 * These must follow m_init and m_cljset
2536 		 */
2537 		m->m_data += padlen;
2538 		ri->iri_len -= padlen;
2539 		m->m_len = ri->iri_frags[i].irf_len;
2540 	} while (++i < ri->iri_nfrags);
2541 
2542 	return (mh);
2543 }
2544 
2545 /*
2546  * Process one software descriptor
2547  */
2548 static struct mbuf *
2549 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri)
2550 {
2551 	struct if_rxsd sd;
2552 	struct mbuf *m;
2553 
2554 	/* should I merge this back in now that the two paths are basically duplicated? */
2555 	if (ri->iri_nfrags == 1 &&
2556 	    ri->iri_frags[0].irf_len <= MIN(IFLIB_RX_COPY_THRESH, MHLEN)) {
2557 		rxd_frag_to_sd(rxq, &ri->iri_frags[0], FALSE, &sd);
2558 		m = *sd.ifsd_m;
2559 		*sd.ifsd_m = NULL;
2560 		m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR);
2561 #ifndef __NO_STRICT_ALIGNMENT
2562 		if (!IP_ALIGNED(m))
2563 			m->m_data += 2;
2564 #endif
2565 		memcpy(m->m_data, *sd.ifsd_cl, ri->iri_len);
2566 		m->m_len = ri->iri_frags[0].irf_len;
2567        } else {
2568 		m = assemble_segments(rxq, ri, &sd);
2569 	}
2570 	m->m_pkthdr.len = ri->iri_len;
2571 	m->m_pkthdr.rcvif = ri->iri_ifp;
2572 	m->m_flags |= ri->iri_flags;
2573 	m->m_pkthdr.ether_vtag = ri->iri_vtag;
2574 	m->m_pkthdr.flowid = ri->iri_flowid;
2575 	M_HASHTYPE_SET(m, ri->iri_rsstype);
2576 	m->m_pkthdr.csum_flags = ri->iri_csum_flags;
2577 	m->m_pkthdr.csum_data = ri->iri_csum_data;
2578 	return (m);
2579 }
2580 
2581 #if defined(INET6) || defined(INET)
2582 static void
2583 iflib_get_ip_forwarding(struct lro_ctrl *lc, bool *v4, bool *v6)
2584 {
2585 	CURVNET_SET(lc->ifp->if_vnet);
2586 #if defined(INET6)
2587 	*v6 = VNET(ip6_forwarding);
2588 #endif
2589 #if defined(INET)
2590 	*v4 = VNET(ipforwarding);
2591 #endif
2592 	CURVNET_RESTORE();
2593 }
2594 
2595 /*
2596  * Returns true if it's possible this packet could be LROed.
2597  * if it returns false, it is guaranteed that tcp_lro_rx()
2598  * would not return zero.
2599  */
2600 static bool
2601 iflib_check_lro_possible(struct mbuf *m, bool v4_forwarding, bool v6_forwarding)
2602 {
2603 	struct ether_header *eh;
2604 	uint16_t eh_type;
2605 
2606 	eh = mtod(m, struct ether_header *);
2607 	eh_type = ntohs(eh->ether_type);
2608 	switch (eh_type) {
2609 #if defined(INET6)
2610 		case ETHERTYPE_IPV6:
2611 			return !v6_forwarding;
2612 #endif
2613 #if defined (INET)
2614 		case ETHERTYPE_IP:
2615 			return !v4_forwarding;
2616 #endif
2617 	}
2618 
2619 	return false;
2620 }
2621 #else
2622 static void
2623 iflib_get_ip_forwarding(struct lro_ctrl *lc __unused, bool *v4 __unused, bool *v6 __unused)
2624 {
2625 }
2626 #endif
2627 
2628 static bool
2629 iflib_rxeof(iflib_rxq_t rxq, qidx_t budget)
2630 {
2631 	if_ctx_t ctx = rxq->ifr_ctx;
2632 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2633 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2634 	int avail, i;
2635 	qidx_t *cidxp;
2636 	struct if_rxd_info ri;
2637 	int err, budget_left, rx_bytes, rx_pkts;
2638 	iflib_fl_t fl;
2639 	struct ifnet *ifp;
2640 	int lro_enabled;
2641 	bool v4_forwarding, v6_forwarding, lro_possible;
2642 
2643 	/*
2644 	 * XXX early demux data packets so that if_input processing only handles
2645 	 * acks in interrupt context
2646 	 */
2647 	struct mbuf *m, *mh, *mt, *mf;
2648 
2649 	lro_possible = v4_forwarding = v6_forwarding = false;
2650 	ifp = ctx->ifc_ifp;
2651 	mh = mt = NULL;
2652 	MPASS(budget > 0);
2653 	rx_pkts	= rx_bytes = 0;
2654 	if (sctx->isc_flags & IFLIB_HAS_RXCQ)
2655 		cidxp = &rxq->ifr_cq_cidx;
2656 	else
2657 		cidxp = &rxq->ifr_fl[0].ifl_cidx;
2658 	if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) {
2659 		for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2660 			__iflib_fl_refill_lt(ctx, fl, budget + 8);
2661 		DBG_COUNTER_INC(rx_unavail);
2662 		return (false);
2663 	}
2664 
2665 	for (budget_left = budget; (budget_left > 0) && (avail > 0); budget_left--, avail--) {
2666 		if (__predict_false(!CTX_ACTIVE(ctx))) {
2667 			DBG_COUNTER_INC(rx_ctx_inactive);
2668 			break;
2669 		}
2670 		/*
2671 		 * Reset client set fields to their default values
2672 		 */
2673 		rxd_info_zero(&ri);
2674 		ri.iri_qsidx = rxq->ifr_id;
2675 		ri.iri_cidx = *cidxp;
2676 		ri.iri_ifp = ifp;
2677 		ri.iri_frags = rxq->ifr_frags;
2678 		err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
2679 
2680 		if (err)
2681 			goto err;
2682 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
2683 			*cidxp = ri.iri_cidx;
2684 			/* Update our consumer index */
2685 			/* XXX NB: shurd - check if this is still safe */
2686 			while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0]) {
2687 				rxq->ifr_cq_cidx -= scctx->isc_nrxd[0];
2688 				rxq->ifr_cq_gen = 0;
2689 			}
2690 			/* was this only a completion queue message? */
2691 			if (__predict_false(ri.iri_nfrags == 0))
2692 				continue;
2693 		}
2694 		MPASS(ri.iri_nfrags != 0);
2695 		MPASS(ri.iri_len != 0);
2696 
2697 		/* will advance the cidx on the corresponding free lists */
2698 		m = iflib_rxd_pkt_get(rxq, &ri);
2699 		if (avail == 0 && budget_left)
2700 			avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left);
2701 
2702 		if (__predict_false(m == NULL)) {
2703 			DBG_COUNTER_INC(rx_mbuf_null);
2704 			continue;
2705 		}
2706 		/* imm_pkt: -- cxgb */
2707 		if (mh == NULL)
2708 			mh = mt = m;
2709 		else {
2710 			mt->m_nextpkt = m;
2711 			mt = m;
2712 		}
2713 	}
2714 	/* make sure that we can refill faster than drain */
2715 	for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2716 		__iflib_fl_refill_lt(ctx, fl, budget + 8);
2717 
2718 	lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO);
2719 	if (lro_enabled)
2720 		iflib_get_ip_forwarding(&rxq->ifr_lc, &v4_forwarding, &v6_forwarding);
2721 	mt = mf = NULL;
2722 	while (mh != NULL) {
2723 		m = mh;
2724 		mh = mh->m_nextpkt;
2725 		m->m_nextpkt = NULL;
2726 #ifndef __NO_STRICT_ALIGNMENT
2727 		if (!IP_ALIGNED(m) && (m = iflib_fixup_rx(m)) == NULL)
2728 			continue;
2729 #endif
2730 		rx_bytes += m->m_pkthdr.len;
2731 		rx_pkts++;
2732 #if defined(INET6) || defined(INET)
2733 		if (lro_enabled) {
2734 			if (!lro_possible) {
2735 				lro_possible = iflib_check_lro_possible(m, v4_forwarding, v6_forwarding);
2736 				if (lro_possible && mf != NULL) {
2737 					ifp->if_input(ifp, mf);
2738 					DBG_COUNTER_INC(rx_if_input);
2739 					mt = mf = NULL;
2740 				}
2741 			}
2742 			if ((m->m_pkthdr.csum_flags & (CSUM_L4_CALC|CSUM_L4_VALID)) ==
2743 			    (CSUM_L4_CALC|CSUM_L4_VALID)) {
2744 				if (lro_possible && tcp_lro_rx(&rxq->ifr_lc, m, 0) == 0)
2745 					continue;
2746 			}
2747 		}
2748 #endif
2749 		if (lro_possible) {
2750 			ifp->if_input(ifp, m);
2751 			DBG_COUNTER_INC(rx_if_input);
2752 			continue;
2753 		}
2754 
2755 		if (mf == NULL)
2756 			mf = m;
2757 		if (mt != NULL)
2758 			mt->m_nextpkt = m;
2759 		mt = m;
2760 	}
2761 	if (mf != NULL) {
2762 		ifp->if_input(ifp, mf);
2763 		DBG_COUNTER_INC(rx_if_input);
2764 	}
2765 
2766 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
2767 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
2768 
2769 	/*
2770 	 * Flush any outstanding LRO work
2771 	 */
2772 #if defined(INET6) || defined(INET)
2773 	tcp_lro_flush_all(&rxq->ifr_lc);
2774 #endif
2775 	if (avail)
2776 		return true;
2777 	return (iflib_rxd_avail(ctx, rxq, *cidxp, 1));
2778 err:
2779 	STATE_LOCK(ctx);
2780 	ctx->ifc_flags |= IFC_DO_RESET;
2781 	iflib_admin_intr_deferred(ctx);
2782 	STATE_UNLOCK(ctx);
2783 	return (false);
2784 }
2785 
2786 #define TXD_NOTIFY_COUNT(txq) (((txq)->ift_size / (txq)->ift_update_freq)-1)
2787 static inline qidx_t
2788 txq_max_db_deferred(iflib_txq_t txq, qidx_t in_use)
2789 {
2790 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2791 	qidx_t minthresh = txq->ift_size / 8;
2792 	if (in_use > 4*minthresh)
2793 		return (notify_count);
2794 	if (in_use > 2*minthresh)
2795 		return (notify_count >> 1);
2796 	if (in_use > minthresh)
2797 		return (notify_count >> 3);
2798 	return (0);
2799 }
2800 
2801 static inline qidx_t
2802 txq_max_rs_deferred(iflib_txq_t txq)
2803 {
2804 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2805 	qidx_t minthresh = txq->ift_size / 8;
2806 	if (txq->ift_in_use > 4*minthresh)
2807 		return (notify_count);
2808 	if (txq->ift_in_use > 2*minthresh)
2809 		return (notify_count >> 1);
2810 	if (txq->ift_in_use > minthresh)
2811 		return (notify_count >> 2);
2812 	return (2);
2813 }
2814 
2815 #define M_CSUM_FLAGS(m) ((m)->m_pkthdr.csum_flags)
2816 #define M_HAS_VLANTAG(m) (m->m_flags & M_VLANTAG)
2817 
2818 #define TXQ_MAX_DB_DEFERRED(txq, in_use) txq_max_db_deferred((txq), (in_use))
2819 #define TXQ_MAX_RS_DEFERRED(txq) txq_max_rs_deferred(txq)
2820 #define TXQ_MAX_DB_CONSUMED(size) (size >> 4)
2821 
2822 /* forward compatibility for cxgb */
2823 #define FIRST_QSET(ctx) 0
2824 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets)
2825 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets)
2826 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx))
2827 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments))
2828 
2829 /* XXX we should be setting this to something other than zero */
2830 #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh)
2831 #define MAX_TX_DESC(ctx) ((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max)
2832 
2833 static inline bool
2834 iflib_txd_db_check(if_ctx_t ctx, iflib_txq_t txq, int ring, qidx_t in_use)
2835 {
2836 	qidx_t dbval, max;
2837 	bool rang;
2838 
2839 	rang = false;
2840 	max = TXQ_MAX_DB_DEFERRED(txq, in_use);
2841 	if (ring || txq->ift_db_pending >= max) {
2842 		dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx;
2843 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval);
2844 		txq->ift_db_pending = txq->ift_npending = 0;
2845 		rang = true;
2846 	}
2847 	return (rang);
2848 }
2849 
2850 #ifdef PKT_DEBUG
2851 static void
2852 print_pkt(if_pkt_info_t pi)
2853 {
2854 	printf("pi len:  %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n",
2855 	       pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx);
2856 	printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n",
2857 	       pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag);
2858 	printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n",
2859 	       pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto);
2860 }
2861 #endif
2862 
2863 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO)
2864 #define IS_TX_OFFLOAD4(pi) ((pi)->ipi_csum_flags & (CSUM_IP_TCP | CSUM_IP_TSO))
2865 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO)
2866 #define IS_TX_OFFLOAD6(pi) ((pi)->ipi_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_TSO))
2867 
2868 static int
2869 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp)
2870 {
2871 	if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx;
2872 	struct ether_vlan_header *eh;
2873 	struct mbuf *m, *n;
2874 
2875 	n = m = *mp;
2876 	if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) &&
2877 	    M_WRITABLE(m) == 0) {
2878 		if ((m = m_dup(m, M_NOWAIT)) == NULL) {
2879 			return (ENOMEM);
2880 		} else {
2881 			m_freem(*mp);
2882 			n = *mp = m;
2883 		}
2884 	}
2885 
2886 	/*
2887 	 * Determine where frame payload starts.
2888 	 * Jump over vlan headers if already present,
2889 	 * helpful for QinQ too.
2890 	 */
2891 	if (__predict_false(m->m_len < sizeof(*eh))) {
2892 		txq->ift_pullups++;
2893 		if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL))
2894 			return (ENOMEM);
2895 	}
2896 	eh = mtod(m, struct ether_vlan_header *);
2897 	if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
2898 		pi->ipi_etype = ntohs(eh->evl_proto);
2899 		pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
2900 	} else {
2901 		pi->ipi_etype = ntohs(eh->evl_encap_proto);
2902 		pi->ipi_ehdrlen = ETHER_HDR_LEN;
2903 	}
2904 
2905 	switch (pi->ipi_etype) {
2906 #ifdef INET
2907 	case ETHERTYPE_IP:
2908 	{
2909 		struct ip *ip = NULL;
2910 		struct tcphdr *th = NULL;
2911 		int minthlen;
2912 
2913 		minthlen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip) + sizeof(*th));
2914 		if (__predict_false(m->m_len < minthlen)) {
2915 			/*
2916 			 * if this code bloat is causing too much of a hit
2917 			 * move it to a separate function and mark it noinline
2918 			 */
2919 			if (m->m_len == pi->ipi_ehdrlen) {
2920 				n = m->m_next;
2921 				MPASS(n);
2922 				if (n->m_len >= sizeof(*ip))  {
2923 					ip = (struct ip *)n->m_data;
2924 					if (n->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2925 						th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2926 				} else {
2927 					txq->ift_pullups++;
2928 					if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
2929 						return (ENOMEM);
2930 					ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2931 				}
2932 			} else {
2933 				txq->ift_pullups++;
2934 				if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
2935 					return (ENOMEM);
2936 				ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2937 				if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2938 					th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2939 			}
2940 		} else {
2941 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2942 			if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2943 				th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2944 		}
2945 		pi->ipi_ip_hlen = ip->ip_hl << 2;
2946 		pi->ipi_ipproto = ip->ip_p;
2947 		pi->ipi_flags |= IPI_TX_IPV4;
2948 
2949 		if ((sctx->isc_flags & IFLIB_NEED_ZERO_CSUM) && (pi->ipi_csum_flags & CSUM_IP))
2950                        ip->ip_sum = 0;
2951 
2952 		/* TCP checksum offload may require TCP header length */
2953 		if (IS_TX_OFFLOAD4(pi)) {
2954 			if (__predict_true(pi->ipi_ipproto == IPPROTO_TCP)) {
2955 				if (__predict_false(th == NULL)) {
2956 					txq->ift_pullups++;
2957 					if (__predict_false((m = m_pullup(m, (ip->ip_hl << 2) + sizeof(*th))) == NULL))
2958 						return (ENOMEM);
2959 					th = (struct tcphdr *)((caddr_t)ip + pi->ipi_ip_hlen);
2960 				}
2961 				pi->ipi_tcp_hflags = th->th_flags;
2962 				pi->ipi_tcp_hlen = th->th_off << 2;
2963 				pi->ipi_tcp_seq = th->th_seq;
2964 			}
2965 			if (IS_TSO4(pi)) {
2966 				if (__predict_false(ip->ip_p != IPPROTO_TCP))
2967 					return (ENXIO);
2968 				th->th_sum = in_pseudo(ip->ip_src.s_addr,
2969 						       ip->ip_dst.s_addr, htons(IPPROTO_TCP));
2970 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
2971 				if (sctx->isc_flags & IFLIB_TSO_INIT_IP) {
2972 					ip->ip_sum = 0;
2973 					ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz);
2974 				}
2975 			}
2976 		}
2977 		break;
2978 	}
2979 #endif
2980 #ifdef INET6
2981 	case ETHERTYPE_IPV6:
2982 	{
2983 		struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
2984 		struct tcphdr *th;
2985 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
2986 
2987 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
2988 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
2989 				return (ENOMEM);
2990 		}
2991 		th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
2992 
2993 		/* XXX-BZ this will go badly in case of ext hdrs. */
2994 		pi->ipi_ipproto = ip6->ip6_nxt;
2995 		pi->ipi_flags |= IPI_TX_IPV6;
2996 
2997 		/* TCP checksum offload may require TCP header length */
2998 		if (IS_TX_OFFLOAD6(pi)) {
2999 			if (pi->ipi_ipproto == IPPROTO_TCP) {
3000 				if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) {
3001 					txq->ift_pullups++;
3002 					if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL))
3003 						return (ENOMEM);
3004 				}
3005 				pi->ipi_tcp_hflags = th->th_flags;
3006 				pi->ipi_tcp_hlen = th->th_off << 2;
3007 				pi->ipi_tcp_seq = th->th_seq;
3008 			}
3009 			if (IS_TSO6(pi)) {
3010 				if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP))
3011 					return (ENXIO);
3012 				/*
3013 				 * The corresponding flag is set by the stack in the IPv4
3014 				 * TSO case, but not in IPv6 (at least in FreeBSD 10.2).
3015 				 * So, set it here because the rest of the flow requires it.
3016 				 */
3017 				pi->ipi_csum_flags |= CSUM_IP6_TCP;
3018 				th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0);
3019 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3020 			}
3021 		}
3022 		break;
3023 	}
3024 #endif
3025 	default:
3026 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
3027 		pi->ipi_ip_hlen = 0;
3028 		break;
3029 	}
3030 	*mp = m;
3031 
3032 	return (0);
3033 }
3034 
3035 static  __noinline  struct mbuf *
3036 collapse_pkthdr(struct mbuf *m0)
3037 {
3038 	struct mbuf *m, *m_next, *tmp;
3039 
3040 	m = m0;
3041 	m_next = m->m_next;
3042 	while (m_next != NULL && m_next->m_len == 0) {
3043 		m = m_next;
3044 		m->m_next = NULL;
3045 		m_free(m);
3046 		m_next = m_next->m_next;
3047 	}
3048 	m = m0;
3049 	m->m_next = m_next;
3050 	if ((m_next->m_flags & M_EXT) == 0) {
3051 		m = m_defrag(m, M_NOWAIT);
3052 	} else {
3053 		tmp = m_next->m_next;
3054 		memcpy(m_next, m, MPKTHSIZE);
3055 		m = m_next;
3056 		m->m_next = tmp;
3057 	}
3058 	return (m);
3059 }
3060 
3061 /*
3062  * If dodgy hardware rejects the scatter gather chain we've handed it
3063  * we'll need to remove the mbuf chain from ifsg_m[] before we can add the
3064  * m_defrag'd mbufs
3065  */
3066 static __noinline struct mbuf *
3067 iflib_remove_mbuf(iflib_txq_t txq)
3068 {
3069 	int ntxd, i, pidx;
3070 	struct mbuf *m, *mh, **ifsd_m;
3071 
3072 	pidx = txq->ift_pidx;
3073 	ifsd_m = txq->ift_sds.ifsd_m;
3074 	ntxd = txq->ift_size;
3075 	mh = m = ifsd_m[pidx];
3076 	ifsd_m[pidx] = NULL;
3077 #if MEMORY_LOGGING
3078 	txq->ift_dequeued++;
3079 #endif
3080 	i = 1;
3081 
3082 	while (m) {
3083 		ifsd_m[(pidx + i) & (ntxd -1)] = NULL;
3084 #if MEMORY_LOGGING
3085 		txq->ift_dequeued++;
3086 #endif
3087 		m = m->m_next;
3088 		i++;
3089 	}
3090 	return (mh);
3091 }
3092 
3093 static int
3094 iflib_busdma_load_mbuf_sg(iflib_txq_t txq, bus_dma_tag_t tag, bus_dmamap_t map,
3095 			  struct mbuf **m0, bus_dma_segment_t *segs, int *nsegs,
3096 			  int max_segs, int flags)
3097 {
3098 	if_ctx_t ctx;
3099 	if_shared_ctx_t		sctx;
3100 	if_softc_ctx_t		scctx;
3101 	int i, next, pidx, err, ntxd, count;
3102 	struct mbuf *m, *tmp, **ifsd_m;
3103 
3104 	m = *m0;
3105 
3106 	/*
3107 	 * Please don't ever do this
3108 	 */
3109 	if (__predict_false(m->m_len == 0))
3110 		*m0 = m = collapse_pkthdr(m);
3111 
3112 	ctx = txq->ift_ctx;
3113 	sctx = ctx->ifc_sctx;
3114 	scctx = &ctx->ifc_softc_ctx;
3115 	ifsd_m = txq->ift_sds.ifsd_m;
3116 	ntxd = txq->ift_size;
3117 	pidx = txq->ift_pidx;
3118 	if (map != NULL) {
3119 		uint8_t *ifsd_flags = txq->ift_sds.ifsd_flags;
3120 
3121 		err = bus_dmamap_load_mbuf_sg(tag, map,
3122 					      *m0, segs, nsegs, BUS_DMA_NOWAIT);
3123 		if (err)
3124 			return (err);
3125 		ifsd_flags[pidx] |= TX_SW_DESC_MAPPED;
3126 		count = 0;
3127 		m = *m0;
3128 		do {
3129 			if (__predict_false(m->m_len <= 0)) {
3130 				tmp = m;
3131 				m = m->m_next;
3132 				tmp->m_next = NULL;
3133 				m_free(tmp);
3134 				continue;
3135 			}
3136 			m = m->m_next;
3137 			count++;
3138 		} while (m != NULL);
3139 		if (count > *nsegs) {
3140 			ifsd_m[pidx] = *m0;
3141 			ifsd_m[pidx]->m_flags |= M_TOOBIG;
3142 			return (0);
3143 		}
3144 		m = *m0;
3145 		count = 0;
3146 		do {
3147 			next = (pidx + count) & (ntxd-1);
3148 			MPASS(ifsd_m[next] == NULL);
3149 			ifsd_m[next] = m;
3150 			count++;
3151 			tmp = m;
3152 			m = m->m_next;
3153 		} while (m != NULL);
3154 	} else {
3155 		int buflen, sgsize, maxsegsz, max_sgsize;
3156 		vm_offset_t vaddr;
3157 		vm_paddr_t curaddr;
3158 
3159 		count = i = 0;
3160 		m = *m0;
3161 		if (m->m_pkthdr.csum_flags & CSUM_TSO)
3162 			maxsegsz = scctx->isc_tx_tso_segsize_max;
3163 		else
3164 			maxsegsz = sctx->isc_tx_maxsegsize;
3165 
3166 		do {
3167 			if (__predict_false(m->m_len <= 0)) {
3168 				tmp = m;
3169 				m = m->m_next;
3170 				tmp->m_next = NULL;
3171 				m_free(tmp);
3172 				continue;
3173 			}
3174 			buflen = m->m_len;
3175 			vaddr = (vm_offset_t)m->m_data;
3176 			/*
3177 			 * see if we can't be smarter about physically
3178 			 * contiguous mappings
3179 			 */
3180 			next = (pidx + count) & (ntxd-1);
3181 			MPASS(ifsd_m[next] == NULL);
3182 #if MEMORY_LOGGING
3183 			txq->ift_enqueued++;
3184 #endif
3185 			ifsd_m[next] = m;
3186 			while (buflen > 0) {
3187 				if (i >= max_segs)
3188 					goto err;
3189 				max_sgsize = MIN(buflen, maxsegsz);
3190 				curaddr = pmap_kextract(vaddr);
3191 				sgsize = PAGE_SIZE - (curaddr & PAGE_MASK);
3192 				sgsize = MIN(sgsize, max_sgsize);
3193 				segs[i].ds_addr = curaddr;
3194 				segs[i].ds_len = sgsize;
3195 				vaddr += sgsize;
3196 				buflen -= sgsize;
3197 				i++;
3198 			}
3199 			count++;
3200 			tmp = m;
3201 			m = m->m_next;
3202 		} while (m != NULL);
3203 		*nsegs = i;
3204 	}
3205 	return (0);
3206 err:
3207 	*m0 = iflib_remove_mbuf(txq);
3208 	return (EFBIG);
3209 }
3210 
3211 static inline caddr_t
3212 calc_next_txd(iflib_txq_t txq, int cidx, uint8_t qid)
3213 {
3214 	qidx_t size;
3215 	int ntxd;
3216 	caddr_t start, end, cur, next;
3217 
3218 	ntxd = txq->ift_size;
3219 	size = txq->ift_txd_size[qid];
3220 	start = txq->ift_ifdi[qid].idi_vaddr;
3221 
3222 	if (__predict_false(size == 0))
3223 		return (start);
3224 	cur = start + size*cidx;
3225 	end = start + size*ntxd;
3226 	next = CACHE_PTR_NEXT(cur);
3227 	return (next < end ? next : start);
3228 }
3229 
3230 /*
3231  * Pad an mbuf to ensure a minimum ethernet frame size.
3232  * min_frame_size is the frame size (less CRC) to pad the mbuf to
3233  */
3234 static __noinline int
3235 iflib_ether_pad(device_t dev, struct mbuf **m_head, uint16_t min_frame_size)
3236 {
3237 	/*
3238 	 * 18 is enough bytes to pad an ARP packet to 46 bytes, and
3239 	 * and ARP message is the smallest common payload I can think of
3240 	 */
3241 	static char pad[18];	/* just zeros */
3242 	int n;
3243 	struct mbuf *new_head;
3244 
3245 	if (!M_WRITABLE(*m_head)) {
3246 		new_head = m_dup(*m_head, M_NOWAIT);
3247 		if (new_head == NULL) {
3248 			m_freem(*m_head);
3249 			device_printf(dev, "cannot pad short frame, m_dup() failed");
3250 			DBG_COUNTER_INC(encap_pad_mbuf_fail);
3251 			return ENOMEM;
3252 		}
3253 		m_freem(*m_head);
3254 		*m_head = new_head;
3255 	}
3256 
3257 	for (n = min_frame_size - (*m_head)->m_pkthdr.len;
3258 	     n > 0; n -= sizeof(pad))
3259 		if (!m_append(*m_head, min(n, sizeof(pad)), pad))
3260 			break;
3261 
3262 	if (n > 0) {
3263 		m_freem(*m_head);
3264 		device_printf(dev, "cannot pad short frame\n");
3265 		DBG_COUNTER_INC(encap_pad_mbuf_fail);
3266 		return (ENOBUFS);
3267 	}
3268 
3269 	return 0;
3270 }
3271 
3272 static int
3273 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp)
3274 {
3275 	if_ctx_t		ctx;
3276 	if_shared_ctx_t		sctx;
3277 	if_softc_ctx_t		scctx;
3278 	bus_dma_segment_t	*segs;
3279 	struct mbuf		*m_head;
3280 	void			*next_txd;
3281 	bus_dmamap_t		map;
3282 	struct if_pkt_info	pi;
3283 	int remap = 0;
3284 	int err, nsegs, ndesc, max_segs, pidx, cidx, next, ntxd;
3285 	bus_dma_tag_t desc_tag;
3286 
3287 	segs = txq->ift_segs;
3288 	ctx = txq->ift_ctx;
3289 	sctx = ctx->ifc_sctx;
3290 	scctx = &ctx->ifc_softc_ctx;
3291 	segs = txq->ift_segs;
3292 	ntxd = txq->ift_size;
3293 	m_head = *m_headp;
3294 	map = NULL;
3295 
3296 	/*
3297 	 * If we're doing TSO the next descriptor to clean may be quite far ahead
3298 	 */
3299 	cidx = txq->ift_cidx;
3300 	pidx = txq->ift_pidx;
3301 	if (ctx->ifc_flags & IFC_PREFETCH) {
3302 		next = (cidx + CACHE_PTR_INCREMENT) & (ntxd-1);
3303 		if (!(ctx->ifc_flags & IFLIB_HAS_TXCQ)) {
3304 			next_txd = calc_next_txd(txq, cidx, 0);
3305 			prefetch(next_txd);
3306 		}
3307 
3308 		/* prefetch the next cache line of mbuf pointers and flags */
3309 		prefetch(&txq->ift_sds.ifsd_m[next]);
3310 		if (txq->ift_sds.ifsd_map != NULL) {
3311 			prefetch(&txq->ift_sds.ifsd_map[next]);
3312 			next = (cidx + CACHE_LINE_SIZE) & (ntxd-1);
3313 			prefetch(&txq->ift_sds.ifsd_flags[next]);
3314 		}
3315 	} else if (txq->ift_sds.ifsd_map != NULL)
3316 		map = txq->ift_sds.ifsd_map[pidx];
3317 
3318 	if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
3319 		desc_tag = txq->ift_tso_desc_tag;
3320 		max_segs = scctx->isc_tx_tso_segments_max;
3321 		MPASS(desc_tag != NULL);
3322 		MPASS(max_segs > 0);
3323 	} else {
3324 		desc_tag = txq->ift_desc_tag;
3325 		max_segs = scctx->isc_tx_nsegments;
3326 	}
3327 	if ((sctx->isc_flags & IFLIB_NEED_ETHER_PAD) &&
3328 	    __predict_false(m_head->m_pkthdr.len < scctx->isc_min_frame_size)) {
3329 		err = iflib_ether_pad(ctx->ifc_dev, m_headp, scctx->isc_min_frame_size);
3330 		if (err)
3331 			return err;
3332 	}
3333 	m_head = *m_headp;
3334 
3335 	pkt_info_zero(&pi);
3336 	pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG|M_BCAST|M_MCAST));
3337 	pi.ipi_pidx = pidx;
3338 	pi.ipi_qsidx = txq->ift_id;
3339 	pi.ipi_len = m_head->m_pkthdr.len;
3340 	pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags;
3341 	pi.ipi_vtag = (m_head->m_flags & M_VLANTAG) ? m_head->m_pkthdr.ether_vtag : 0;
3342 
3343 	/* deliberate bitwise OR to make one condition */
3344 	if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) {
3345 		if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0))
3346 			return (err);
3347 		m_head = *m_headp;
3348 	}
3349 
3350 retry:
3351 	err = iflib_busdma_load_mbuf_sg(txq, desc_tag, map, m_headp, segs, &nsegs, max_segs, BUS_DMA_NOWAIT);
3352 defrag:
3353 	if (__predict_false(err)) {
3354 		switch (err) {
3355 		case EFBIG:
3356 			/* try collapse once and defrag once */
3357 			if (remap == 0) {
3358 				m_head = m_collapse(*m_headp, M_NOWAIT, max_segs);
3359 				/* try defrag if collapsing fails */
3360 				if (m_head == NULL)
3361 					remap++;
3362 			}
3363 			if (remap == 1)
3364 				m_head = m_defrag(*m_headp, M_NOWAIT);
3365 			remap++;
3366 			if (__predict_false(m_head == NULL))
3367 				goto defrag_failed;
3368 			txq->ift_mbuf_defrag++;
3369 			*m_headp = m_head;
3370 			goto retry;
3371 			break;
3372 		case ENOMEM:
3373 			txq->ift_no_tx_dma_setup++;
3374 			break;
3375 		default:
3376 			txq->ift_no_tx_dma_setup++;
3377 			m_freem(*m_headp);
3378 			DBG_COUNTER_INC(tx_frees);
3379 			*m_headp = NULL;
3380 			break;
3381 		}
3382 		txq->ift_map_failed++;
3383 		DBG_COUNTER_INC(encap_load_mbuf_fail);
3384 		return (err);
3385 	}
3386 
3387 	/*
3388 	 * XXX assumes a 1 to 1 relationship between segments and
3389 	 *        descriptors - this does not hold true on all drivers, e.g.
3390 	 *        cxgb
3391 	 */
3392 	if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) {
3393 		txq->ift_no_desc_avail++;
3394 		if (map != NULL)
3395 			bus_dmamap_unload(desc_tag, map);
3396 		DBG_COUNTER_INC(encap_txq_avail_fail);
3397 		if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0)
3398 			GROUPTASK_ENQUEUE(&txq->ift_task);
3399 		return (ENOBUFS);
3400 	}
3401 	/*
3402 	 * On Intel cards we can greatly reduce the number of TX interrupts
3403 	 * we see by only setting report status on every Nth descriptor.
3404 	 * However, this also means that the driver will need to keep track
3405 	 * of the descriptors that RS was set on to check them for the DD bit.
3406 	 */
3407 	txq->ift_rs_pending += nsegs + 1;
3408 	if (txq->ift_rs_pending > TXQ_MAX_RS_DEFERRED(txq) ||
3409 	     iflib_no_tx_batch || (TXQ_AVAIL(txq) - nsegs) <= MAX_TX_DESC(ctx) + 2) {
3410 		pi.ipi_flags |= IPI_TX_INTR;
3411 		txq->ift_rs_pending = 0;
3412 	}
3413 
3414 	pi.ipi_segs = segs;
3415 	pi.ipi_nsegs = nsegs;
3416 
3417 	MPASS(pidx >= 0 && pidx < txq->ift_size);
3418 #ifdef PKT_DEBUG
3419 	print_pkt(&pi);
3420 #endif
3421 	if (map != NULL)
3422 		bus_dmamap_sync(desc_tag, map, BUS_DMASYNC_PREWRITE);
3423 	if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) {
3424 		if (map != NULL)
3425 			bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3426 					BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
3427 		DBG_COUNTER_INC(tx_encap);
3428 		MPASS(pi.ipi_new_pidx < txq->ift_size);
3429 
3430 		ndesc = pi.ipi_new_pidx - pi.ipi_pidx;
3431 		if (pi.ipi_new_pidx < pi.ipi_pidx) {
3432 			ndesc += txq->ift_size;
3433 			txq->ift_gen = 1;
3434 		}
3435 		/*
3436 		 * drivers can need as many as
3437 		 * two sentinels
3438 		 */
3439 		MPASS(ndesc <= pi.ipi_nsegs + 2);
3440 		MPASS(pi.ipi_new_pidx != pidx);
3441 		MPASS(ndesc > 0);
3442 		txq->ift_in_use += ndesc;
3443 
3444 		/*
3445 		 * We update the last software descriptor again here because there may
3446 		 * be a sentinel and/or there may be more mbufs than segments
3447 		 */
3448 		txq->ift_pidx = pi.ipi_new_pidx;
3449 		txq->ift_npending += pi.ipi_ndescs;
3450 	} else {
3451 		*m_headp = m_head = iflib_remove_mbuf(txq);
3452 		if (err == EFBIG) {
3453 			txq->ift_txd_encap_efbig++;
3454 			if (remap < 2) {
3455 				remap = 1;
3456 				goto defrag;
3457 			}
3458 		}
3459 		DBG_COUNTER_INC(encap_txd_encap_fail);
3460 		goto defrag_failed;
3461 	}
3462 	return (err);
3463 
3464 defrag_failed:
3465 	txq->ift_mbuf_defrag_failed++;
3466 	txq->ift_map_failed++;
3467 	m_freem(*m_headp);
3468 	DBG_COUNTER_INC(tx_frees);
3469 	*m_headp = NULL;
3470 	return (ENOMEM);
3471 }
3472 
3473 static void
3474 iflib_tx_desc_free(iflib_txq_t txq, int n)
3475 {
3476 	int hasmap;
3477 	uint32_t qsize, cidx, mask, gen;
3478 	struct mbuf *m, **ifsd_m;
3479 	uint8_t *ifsd_flags;
3480 	bus_dmamap_t *ifsd_map;
3481 	bool do_prefetch;
3482 
3483 	cidx = txq->ift_cidx;
3484 	gen = txq->ift_gen;
3485 	qsize = txq->ift_size;
3486 	mask = qsize-1;
3487 	hasmap = txq->ift_sds.ifsd_map != NULL;
3488 	ifsd_flags = txq->ift_sds.ifsd_flags;
3489 	ifsd_m = txq->ift_sds.ifsd_m;
3490 	ifsd_map = txq->ift_sds.ifsd_map;
3491 	do_prefetch = (txq->ift_ctx->ifc_flags & IFC_PREFETCH);
3492 
3493 	while (n-- > 0) {
3494 		if (do_prefetch) {
3495 			prefetch(ifsd_m[(cidx + 3) & mask]);
3496 			prefetch(ifsd_m[(cidx + 4) & mask]);
3497 		}
3498 		if (ifsd_m[cidx] != NULL) {
3499 			prefetch(&ifsd_m[(cidx + CACHE_PTR_INCREMENT) & mask]);
3500 			prefetch(&ifsd_flags[(cidx + CACHE_PTR_INCREMENT) & mask]);
3501 			if (hasmap && (ifsd_flags[cidx] & TX_SW_DESC_MAPPED)) {
3502 				/*
3503 				 * does it matter if it's not the TSO tag? If so we'll
3504 				 * have to add the type to flags
3505 				 */
3506 				bus_dmamap_unload(txq->ift_desc_tag, ifsd_map[cidx]);
3507 				ifsd_flags[cidx] &= ~TX_SW_DESC_MAPPED;
3508 			}
3509 			if ((m = ifsd_m[cidx]) != NULL) {
3510 				/* XXX we don't support any drivers that batch packets yet */
3511 				MPASS(m->m_nextpkt == NULL);
3512 				/* if the number of clusters exceeds the number of segments
3513 				 * there won't be space on the ring to save a pointer to each
3514 				 * cluster so we simply free the list here
3515 				 */
3516 				if (m->m_flags & M_TOOBIG) {
3517 					m_freem(m);
3518 				} else {
3519 					m_free(m);
3520 				}
3521 				ifsd_m[cidx] = NULL;
3522 #if MEMORY_LOGGING
3523 				txq->ift_dequeued++;
3524 #endif
3525 				DBG_COUNTER_INC(tx_frees);
3526 			}
3527 		}
3528 		if (__predict_false(++cidx == qsize)) {
3529 			cidx = 0;
3530 			gen = 0;
3531 		}
3532 	}
3533 	txq->ift_cidx = cidx;
3534 	txq->ift_gen = gen;
3535 }
3536 
3537 static __inline int
3538 iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh)
3539 {
3540 	int reclaim;
3541 	if_ctx_t ctx = txq->ift_ctx;
3542 
3543 	KASSERT(thresh >= 0, ("invalid threshold to reclaim"));
3544 	MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size);
3545 
3546 	/*
3547 	 * Need a rate-limiting check so that this isn't called every time
3548 	 */
3549 	iflib_tx_credits_update(ctx, txq);
3550 	reclaim = DESC_RECLAIMABLE(txq);
3551 
3552 	if (reclaim <= thresh /* + MAX_TX_DESC(txq->ift_ctx) */) {
3553 #ifdef INVARIANTS
3554 		if (iflib_verbose_debug) {
3555 			printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __FUNCTION__,
3556 			       txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments,
3557 			       reclaim, thresh);
3558 
3559 		}
3560 #endif
3561 		return (0);
3562 	}
3563 	iflib_tx_desc_free(txq, reclaim);
3564 	txq->ift_cleaned += reclaim;
3565 	txq->ift_in_use -= reclaim;
3566 
3567 	return (reclaim);
3568 }
3569 
3570 static struct mbuf **
3571 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset, int remaining)
3572 {
3573 	int next, size;
3574 	struct mbuf **items;
3575 
3576 	size = r->size;
3577 	next = (cidx + CACHE_PTR_INCREMENT) & (size-1);
3578 	items = __DEVOLATILE(struct mbuf **, &r->items[0]);
3579 
3580 	prefetch(items[(cidx + offset) & (size-1)]);
3581 	if (remaining > 1) {
3582 		prefetch2cachelines(&items[next]);
3583 		prefetch2cachelines(items[(cidx + offset + 1) & (size-1)]);
3584 		prefetch2cachelines(items[(cidx + offset + 2) & (size-1)]);
3585 		prefetch2cachelines(items[(cidx + offset + 3) & (size-1)]);
3586 	}
3587 	return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (size-1)]));
3588 }
3589 
3590 static void
3591 iflib_txq_check_drain(iflib_txq_t txq, int budget)
3592 {
3593 
3594 	ifmp_ring_check_drainage(txq->ift_br, budget);
3595 }
3596 
3597 static uint32_t
3598 iflib_txq_can_drain(struct ifmp_ring *r)
3599 {
3600 	iflib_txq_t txq = r->cookie;
3601 	if_ctx_t ctx = txq->ift_ctx;
3602 
3603 	return ((TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2) ||
3604 		ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false));
3605 }
3606 
3607 static uint32_t
3608 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3609 {
3610 	iflib_txq_t txq = r->cookie;
3611 	if_ctx_t ctx = txq->ift_ctx;
3612 	struct ifnet *ifp = ctx->ifc_ifp;
3613 	struct mbuf **mp, *m;
3614 	int i, count, consumed, pkt_sent, bytes_sent, mcast_sent, avail;
3615 	int reclaimed, err, in_use_prev, desc_used;
3616 	bool do_prefetch, ring, rang;
3617 
3618 	if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) ||
3619 			    !LINK_ACTIVE(ctx))) {
3620 		DBG_COUNTER_INC(txq_drain_notready);
3621 		return (0);
3622 	}
3623 	reclaimed = iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
3624 	rang = iflib_txd_db_check(ctx, txq, reclaimed, txq->ift_in_use);
3625 	avail = IDXDIFF(pidx, cidx, r->size);
3626 	if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) {
3627 		DBG_COUNTER_INC(txq_drain_flushing);
3628 		for (i = 0; i < avail; i++) {
3629 			m_free(r->items[(cidx + i) & (r->size-1)]);
3630 			r->items[(cidx + i) & (r->size-1)] = NULL;
3631 		}
3632 		return (avail);
3633 	}
3634 
3635 	if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3636 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3637 		CALLOUT_LOCK(txq);
3638 		callout_stop(&txq->ift_timer);
3639 		CALLOUT_UNLOCK(txq);
3640 		DBG_COUNTER_INC(txq_drain_oactive);
3641 		return (0);
3642 	}
3643 	if (reclaimed)
3644 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3645 	consumed = mcast_sent = bytes_sent = pkt_sent = 0;
3646 	count = MIN(avail, TX_BATCH_SIZE);
3647 #ifdef INVARIANTS
3648 	if (iflib_verbose_debug)
3649 		printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __FUNCTION__,
3650 		       avail, ctx->ifc_flags, TXQ_AVAIL(txq));
3651 #endif
3652 	do_prefetch = (ctx->ifc_flags & IFC_PREFETCH);
3653 	avail = TXQ_AVAIL(txq);
3654 	err = 0;
3655 	for (desc_used = i = 0; i < count && avail > MAX_TX_DESC(ctx) + 2; i++) {
3656 		int rem = do_prefetch ? count - i : 0;
3657 
3658 		mp = _ring_peek_one(r, cidx, i, rem);
3659 		MPASS(mp != NULL && *mp != NULL);
3660 		if (__predict_false(*mp == (struct mbuf *)txq)) {
3661 			consumed++;
3662 			reclaimed++;
3663 			continue;
3664 		}
3665 		in_use_prev = txq->ift_in_use;
3666 		err = iflib_encap(txq, mp);
3667 		if (__predict_false(err)) {
3668 			DBG_COUNTER_INC(txq_drain_encapfail);
3669 			/* no room - bail out */
3670 			if (err == ENOBUFS)
3671 				break;
3672 			consumed++;
3673 			DBG_COUNTER_INC(txq_drain_encapfail);
3674 			/* we can't send this packet - skip it */
3675 			continue;
3676 		}
3677 		consumed++;
3678 		pkt_sent++;
3679 		m = *mp;
3680 		DBG_COUNTER_INC(tx_sent);
3681 		bytes_sent += m->m_pkthdr.len;
3682 		mcast_sent += !!(m->m_flags & M_MCAST);
3683 		avail = TXQ_AVAIL(txq);
3684 
3685 		txq->ift_db_pending += (txq->ift_in_use - in_use_prev);
3686 		desc_used += (txq->ift_in_use - in_use_prev);
3687 		ETHER_BPF_MTAP(ifp, m);
3688 		if (__predict_false(!(ifp->if_drv_flags & IFF_DRV_RUNNING)))
3689 			break;
3690 		rang = iflib_txd_db_check(ctx, txq, false, in_use_prev);
3691 	}
3692 
3693 	/* deliberate use of bitwise or to avoid gratuitous short-circuit */
3694 	ring = rang ? false  : (iflib_min_tx_latency | err) || (TXQ_AVAIL(txq) < MAX_TX_DESC(ctx));
3695 	iflib_txd_db_check(ctx, txq, ring, txq->ift_in_use);
3696 	if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
3697 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
3698 	if (mcast_sent)
3699 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
3700 #ifdef INVARIANTS
3701 	if (iflib_verbose_debug)
3702 		printf("consumed=%d\n", consumed);
3703 #endif
3704 	return (consumed);
3705 }
3706 
3707 static uint32_t
3708 iflib_txq_drain_always(struct ifmp_ring *r)
3709 {
3710 	return (1);
3711 }
3712 
3713 static uint32_t
3714 iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3715 {
3716 	int i, avail;
3717 	struct mbuf **mp;
3718 	iflib_txq_t txq;
3719 
3720 	txq = r->cookie;
3721 
3722 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3723 	CALLOUT_LOCK(txq);
3724 	callout_stop(&txq->ift_timer);
3725 	CALLOUT_UNLOCK(txq);
3726 
3727 	avail = IDXDIFF(pidx, cidx, r->size);
3728 	for (i = 0; i < avail; i++) {
3729 		mp = _ring_peek_one(r, cidx, i, avail - i);
3730 		if (__predict_false(*mp == (struct mbuf *)txq))
3731 			continue;
3732 		m_freem(*mp);
3733 	}
3734 	MPASS(ifmp_ring_is_stalled(r) == 0);
3735 	return (avail);
3736 }
3737 
3738 static void
3739 iflib_ifmp_purge(iflib_txq_t txq)
3740 {
3741 	struct ifmp_ring *r;
3742 
3743 	r = txq->ift_br;
3744 	r->drain = iflib_txq_drain_free;
3745 	r->can_drain = iflib_txq_drain_always;
3746 
3747 	ifmp_ring_check_drainage(r, r->size);
3748 
3749 	r->drain = iflib_txq_drain;
3750 	r->can_drain = iflib_txq_can_drain;
3751 }
3752 
3753 static void
3754 _task_fn_tx(void *context)
3755 {
3756 	iflib_txq_t txq = context;
3757 	if_ctx_t ctx = txq->ift_ctx;
3758 	struct ifnet *ifp = ctx->ifc_ifp;
3759 
3760 #ifdef IFLIB_DIAGNOSTICS
3761 	txq->ift_cpu_exec_count[curcpu]++;
3762 #endif
3763 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3764 		return;
3765 	if (if_getcapenable(ifp) & IFCAP_NETMAP) {
3766 		if (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false))
3767 			netmap_tx_irq(ifp, txq->ift_id);
3768 		IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3769 		return;
3770 	}
3771 	if (txq->ift_db_pending)
3772 		ifmp_ring_enqueue(txq->ift_br, (void **)&txq, 1, TX_BATCH_SIZE);
3773 	ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3774 	if (ctx->ifc_flags & IFC_LEGACY)
3775 		IFDI_INTR_ENABLE(ctx);
3776 	else {
3777 #ifdef INVARIANTS
3778 		int rc =
3779 #endif
3780 			IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3781 			KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver"));
3782 	}
3783 }
3784 
3785 static void
3786 _task_fn_rx(void *context)
3787 {
3788 	iflib_rxq_t rxq = context;
3789 	if_ctx_t ctx = rxq->ifr_ctx;
3790 	bool more;
3791 	uint16_t budget;
3792 
3793 #ifdef IFLIB_DIAGNOSTICS
3794 	rxq->ifr_cpu_exec_count[curcpu]++;
3795 #endif
3796 	DBG_COUNTER_INC(task_fn_rxs);
3797 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3798 		return;
3799 	more = true;
3800 #ifdef DEV_NETMAP
3801 	if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP) {
3802 		u_int work = 0;
3803 		if (netmap_rx_irq(ctx->ifc_ifp, rxq->ifr_id, &work)) {
3804 			more = false;
3805 		}
3806 	}
3807 #endif
3808 	budget = ctx->ifc_sysctl_rx_budget;
3809 	if (budget == 0)
3810 		budget = 16;	/* XXX */
3811 	if (more == false || (more = iflib_rxeof(rxq, budget)) == false) {
3812 		if (ctx->ifc_flags & IFC_LEGACY)
3813 			IFDI_INTR_ENABLE(ctx);
3814 		else {
3815 #ifdef INVARIANTS
3816 			int rc =
3817 #endif
3818 				IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
3819 			KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver"));
3820 			DBG_COUNTER_INC(rx_intr_enables);
3821 		}
3822 	}
3823 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3824 		return;
3825 	if (more)
3826 		GROUPTASK_ENQUEUE(&rxq->ifr_task);
3827 }
3828 
3829 static void
3830 _task_fn_admin(void *context)
3831 {
3832 	if_ctx_t ctx = context;
3833 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
3834 	iflib_txq_t txq;
3835 	int i;
3836 	bool oactive, running, do_reset, do_watchdog;
3837 	uint32_t reset_on = hz / 2;
3838 
3839 	STATE_LOCK(ctx);
3840 	running = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING);
3841 	oactive = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE);
3842 	do_reset = (ctx->ifc_flags & IFC_DO_RESET);
3843 	do_watchdog = (ctx->ifc_flags & IFC_DO_WATCHDOG);
3844 	ctx->ifc_flags &= ~(IFC_DO_RESET|IFC_DO_WATCHDOG);
3845 	STATE_UNLOCK(ctx);
3846 
3847 	if ((!running & !oactive) &&
3848 	    !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
3849 		return;
3850 
3851 	CTX_LOCK(ctx);
3852 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3853 		CALLOUT_LOCK(txq);
3854 		callout_stop(&txq->ift_timer);
3855 		CALLOUT_UNLOCK(txq);
3856 	}
3857 	if (do_watchdog) {
3858 		ctx->ifc_watchdog_events++;
3859 		IFDI_WATCHDOG_RESET(ctx);
3860 	}
3861 	IFDI_UPDATE_ADMIN_STATUS(ctx);
3862 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3863 #ifdef DEV_NETMAP
3864 		reset_on = hz / 2;
3865 		if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP)
3866 			iflib_netmap_timer_adjust(ctx, txq->ift_id, &reset_on);
3867 #endif
3868 		callout_reset_on(&txq->ift_timer, reset_on, iflib_timer, txq, txq->ift_timer.c_cpu);
3869 	}
3870 	IFDI_LINK_INTR_ENABLE(ctx);
3871 	if (do_reset)
3872 		iflib_if_init_locked(ctx);
3873 	CTX_UNLOCK(ctx);
3874 
3875 	if (LINK_ACTIVE(ctx) == 0)
3876 		return;
3877 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3878 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
3879 }
3880 
3881 
3882 static void
3883 _task_fn_iov(void *context)
3884 {
3885 	if_ctx_t ctx = context;
3886 
3887 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3888 		return;
3889 
3890 	CTX_LOCK(ctx);
3891 	IFDI_VFLR_HANDLE(ctx);
3892 	CTX_UNLOCK(ctx);
3893 }
3894 
3895 static int
3896 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)
3897 {
3898 	int err;
3899 	if_int_delay_info_t info;
3900 	if_ctx_t ctx;
3901 
3902 	info = (if_int_delay_info_t)arg1;
3903 	ctx = info->iidi_ctx;
3904 	info->iidi_req = req;
3905 	info->iidi_oidp = oidp;
3906 	CTX_LOCK(ctx);
3907 	err = IFDI_SYSCTL_INT_DELAY(ctx, info);
3908 	CTX_UNLOCK(ctx);
3909 	return (err);
3910 }
3911 
3912 /*********************************************************************
3913  *
3914  *  IFNET FUNCTIONS
3915  *
3916  **********************************************************************/
3917 
3918 static void
3919 iflib_if_init_locked(if_ctx_t ctx)
3920 {
3921 	iflib_stop(ctx);
3922 	iflib_init_locked(ctx);
3923 }
3924 
3925 
3926 static void
3927 iflib_if_init(void *arg)
3928 {
3929 	if_ctx_t ctx = arg;
3930 
3931 	CTX_LOCK(ctx);
3932 	iflib_if_init_locked(ctx);
3933 	CTX_UNLOCK(ctx);
3934 }
3935 
3936 static int
3937 iflib_if_transmit(if_t ifp, struct mbuf *m)
3938 {
3939 	if_ctx_t	ctx = if_getsoftc(ifp);
3940 
3941 	iflib_txq_t txq;
3942 	int err, qidx;
3943 
3944 	if (__predict_false((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) {
3945 		DBG_COUNTER_INC(tx_frees);
3946 		m_freem(m);
3947 		return (ENOBUFS);
3948 	}
3949 
3950 	MPASS(m->m_nextpkt == NULL);
3951 	qidx = 0;
3952 	if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m))
3953 		qidx = QIDX(ctx, m);
3954 	/*
3955 	 * XXX calculate buf_ring based on flowid (divvy up bits?)
3956 	 */
3957 	txq = &ctx->ifc_txqs[qidx];
3958 
3959 #ifdef DRIVER_BACKPRESSURE
3960 	if (txq->ift_closed) {
3961 		while (m != NULL) {
3962 			next = m->m_nextpkt;
3963 			m->m_nextpkt = NULL;
3964 			m_freem(m);
3965 			m = next;
3966 		}
3967 		return (ENOBUFS);
3968 	}
3969 #endif
3970 #ifdef notyet
3971 	qidx = count = 0;
3972 	mp = marr;
3973 	next = m;
3974 	do {
3975 		count++;
3976 		next = next->m_nextpkt;
3977 	} while (next != NULL);
3978 
3979 	if (count > nitems(marr))
3980 		if ((mp = malloc(count*sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) {
3981 			/* XXX check nextpkt */
3982 			m_freem(m);
3983 			/* XXX simplify for now */
3984 			DBG_COUNTER_INC(tx_frees);
3985 			return (ENOBUFS);
3986 		}
3987 	for (next = m, i = 0; next != NULL; i++) {
3988 		mp[i] = next;
3989 		next = next->m_nextpkt;
3990 		mp[i]->m_nextpkt = NULL;
3991 	}
3992 #endif
3993 	DBG_COUNTER_INC(tx_seen);
3994 	err = ifmp_ring_enqueue(txq->ift_br, (void **)&m, 1, TX_BATCH_SIZE);
3995 
3996 	GROUPTASK_ENQUEUE(&txq->ift_task);
3997 	if (err) {
3998 		/* support forthcoming later */
3999 #ifdef DRIVER_BACKPRESSURE
4000 		txq->ift_closed = TRUE;
4001 #endif
4002 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
4003 		m_freem(m);
4004 	}
4005 
4006 	return (err);
4007 }
4008 
4009 static void
4010 iflib_if_qflush(if_t ifp)
4011 {
4012 	if_ctx_t ctx = if_getsoftc(ifp);
4013 	iflib_txq_t txq = ctx->ifc_txqs;
4014 	int i;
4015 
4016 	STATE_LOCK(ctx);
4017 	ctx->ifc_flags |= IFC_QFLUSH;
4018 	STATE_UNLOCK(ctx);
4019 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
4020 		while (!(ifmp_ring_is_idle(txq->ift_br) || ifmp_ring_is_stalled(txq->ift_br)))
4021 			iflib_txq_check_drain(txq, 0);
4022 	STATE_LOCK(ctx);
4023 	ctx->ifc_flags &= ~IFC_QFLUSH;
4024 	STATE_UNLOCK(ctx);
4025 
4026 	if_qflush(ifp);
4027 }
4028 
4029 
4030 #define IFCAP_FLAGS (IFCAP_TXCSUM_IPV6 | IFCAP_RXCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \
4031 		     IFCAP_TSO4 | IFCAP_TSO6 | IFCAP_VLAN_HWTAGGING | IFCAP_HWSTATS | \
4032 		     IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | IFCAP_VLAN_HWTSO)
4033 
4034 static int
4035 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data)
4036 {
4037 	if_ctx_t ctx = if_getsoftc(ifp);
4038 	struct ifreq	*ifr = (struct ifreq *)data;
4039 #if defined(INET) || defined(INET6)
4040 	struct ifaddr	*ifa = (struct ifaddr *)data;
4041 #endif
4042 	bool		avoid_reset = FALSE;
4043 	int		err = 0, reinit = 0, bits;
4044 
4045 	switch (command) {
4046 	case SIOCSIFADDR:
4047 #ifdef INET
4048 		if (ifa->ifa_addr->sa_family == AF_INET)
4049 			avoid_reset = TRUE;
4050 #endif
4051 #ifdef INET6
4052 		if (ifa->ifa_addr->sa_family == AF_INET6)
4053 			avoid_reset = TRUE;
4054 #endif
4055 		/*
4056 		** Calling init results in link renegotiation,
4057 		** so we avoid doing it when possible.
4058 		*/
4059 		if (avoid_reset) {
4060 			if_setflagbits(ifp, IFF_UP,0);
4061 			if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
4062 				reinit = 1;
4063 #ifdef INET
4064 			if (!(if_getflags(ifp) & IFF_NOARP))
4065 				arp_ifinit(ifp, ifa);
4066 #endif
4067 		} else
4068 			err = ether_ioctl(ifp, command, data);
4069 		break;
4070 	case SIOCSIFMTU:
4071 		CTX_LOCK(ctx);
4072 		if (ifr->ifr_mtu == if_getmtu(ifp)) {
4073 			CTX_UNLOCK(ctx);
4074 			break;
4075 		}
4076 		bits = if_getdrvflags(ifp);
4077 		/* stop the driver and free any clusters before proceeding */
4078 		iflib_stop(ctx);
4079 
4080 		if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) {
4081 			STATE_LOCK(ctx);
4082 			if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size)
4083 				ctx->ifc_flags |= IFC_MULTISEG;
4084 			else
4085 				ctx->ifc_flags &= ~IFC_MULTISEG;
4086 			STATE_UNLOCK(ctx);
4087 			err = if_setmtu(ifp, ifr->ifr_mtu);
4088 		}
4089 		iflib_init_locked(ctx);
4090 		STATE_LOCK(ctx);
4091 		if_setdrvflags(ifp, bits);
4092 		STATE_UNLOCK(ctx);
4093 		CTX_UNLOCK(ctx);
4094 		break;
4095 	case SIOCSIFFLAGS:
4096 		CTX_LOCK(ctx);
4097 		if (if_getflags(ifp) & IFF_UP) {
4098 			if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4099 				if ((if_getflags(ifp) ^ ctx->ifc_if_flags) &
4100 				    (IFF_PROMISC | IFF_ALLMULTI)) {
4101 					err = IFDI_PROMISC_SET(ctx, if_getflags(ifp));
4102 				}
4103 			} else
4104 				reinit = 1;
4105 		} else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4106 			iflib_stop(ctx);
4107 		}
4108 		ctx->ifc_if_flags = if_getflags(ifp);
4109 		CTX_UNLOCK(ctx);
4110 		break;
4111 	case SIOCADDMULTI:
4112 	case SIOCDELMULTI:
4113 		if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4114 			CTX_LOCK(ctx);
4115 			IFDI_INTR_DISABLE(ctx);
4116 			IFDI_MULTI_SET(ctx);
4117 			IFDI_INTR_ENABLE(ctx);
4118 			CTX_UNLOCK(ctx);
4119 		}
4120 		break;
4121 	case SIOCSIFMEDIA:
4122 		CTX_LOCK(ctx);
4123 		IFDI_MEDIA_SET(ctx);
4124 		CTX_UNLOCK(ctx);
4125 		/* falls thru */
4126 	case SIOCGIFMEDIA:
4127 	case SIOCGIFXMEDIA:
4128 		err = ifmedia_ioctl(ifp, ifr, &ctx->ifc_media, command);
4129 		break;
4130 	case SIOCGI2C:
4131 	{
4132 		struct ifi2creq i2c;
4133 
4134 		err = copyin(ifr_data_get_ptr(ifr), &i2c, sizeof(i2c));
4135 		if (err != 0)
4136 			break;
4137 		if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) {
4138 			err = EINVAL;
4139 			break;
4140 		}
4141 		if (i2c.len > sizeof(i2c.data)) {
4142 			err = EINVAL;
4143 			break;
4144 		}
4145 
4146 		if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0)
4147 			err = copyout(&i2c, ifr_data_get_ptr(ifr),
4148 			    sizeof(i2c));
4149 		break;
4150 	}
4151 	case SIOCSIFCAP:
4152 	{
4153 		int mask, setmask;
4154 
4155 		mask = ifr->ifr_reqcap ^ if_getcapenable(ifp);
4156 		setmask = 0;
4157 #ifdef TCP_OFFLOAD
4158 		setmask |= mask & (IFCAP_TOE4|IFCAP_TOE6);
4159 #endif
4160 		setmask |= (mask & IFCAP_FLAGS);
4161 
4162 		if (setmask & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6))
4163 			setmask |= (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6);
4164 		if ((mask & IFCAP_WOL) &&
4165 		    (if_getcapabilities(ifp) & IFCAP_WOL) != 0)
4166 			setmask |= (mask & (IFCAP_WOL_MCAST|IFCAP_WOL_MAGIC));
4167 		if_vlancap(ifp);
4168 		/*
4169 		 * want to ensure that traffic has stopped before we change any of the flags
4170 		 */
4171 		if (setmask) {
4172 			CTX_LOCK(ctx);
4173 			bits = if_getdrvflags(ifp);
4174 			if (bits & IFF_DRV_RUNNING)
4175 				iflib_stop(ctx);
4176 			STATE_LOCK(ctx);
4177 			if_togglecapenable(ifp, setmask);
4178 			STATE_UNLOCK(ctx);
4179 			if (bits & IFF_DRV_RUNNING)
4180 				iflib_init_locked(ctx);
4181 			STATE_LOCK(ctx);
4182 			if_setdrvflags(ifp, bits);
4183 			STATE_UNLOCK(ctx);
4184 			CTX_UNLOCK(ctx);
4185 		}
4186 		break;
4187 	}
4188 	case SIOCGPRIVATE_0:
4189 	case SIOCSDRVSPEC:
4190 	case SIOCGDRVSPEC:
4191 		CTX_LOCK(ctx);
4192 		err = IFDI_PRIV_IOCTL(ctx, command, data);
4193 		CTX_UNLOCK(ctx);
4194 		break;
4195 	default:
4196 		err = ether_ioctl(ifp, command, data);
4197 		break;
4198 	}
4199 	if (reinit)
4200 		iflib_if_init(ctx);
4201 	return (err);
4202 }
4203 
4204 static uint64_t
4205 iflib_if_get_counter(if_t ifp, ift_counter cnt)
4206 {
4207 	if_ctx_t ctx = if_getsoftc(ifp);
4208 
4209 	return (IFDI_GET_COUNTER(ctx, cnt));
4210 }
4211 
4212 /*********************************************************************
4213  *
4214  *  OTHER FUNCTIONS EXPORTED TO THE STACK
4215  *
4216  **********************************************************************/
4217 
4218 static void
4219 iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag)
4220 {
4221 	if_ctx_t ctx = if_getsoftc(ifp);
4222 
4223 	if ((void *)ctx != arg)
4224 		return;
4225 
4226 	if ((vtag == 0) || (vtag > 4095))
4227 		return;
4228 
4229 	CTX_LOCK(ctx);
4230 	IFDI_VLAN_REGISTER(ctx, vtag);
4231 	/* Re-init to load the changes */
4232 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
4233 		iflib_if_init_locked(ctx);
4234 	CTX_UNLOCK(ctx);
4235 }
4236 
4237 static void
4238 iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag)
4239 {
4240 	if_ctx_t ctx = if_getsoftc(ifp);
4241 
4242 	if ((void *)ctx != arg)
4243 		return;
4244 
4245 	if ((vtag == 0) || (vtag > 4095))
4246 		return;
4247 
4248 	CTX_LOCK(ctx);
4249 	IFDI_VLAN_UNREGISTER(ctx, vtag);
4250 	/* Re-init to load the changes */
4251 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
4252 		iflib_if_init_locked(ctx);
4253 	CTX_UNLOCK(ctx);
4254 }
4255 
4256 static void
4257 iflib_led_func(void *arg, int onoff)
4258 {
4259 	if_ctx_t ctx = arg;
4260 
4261 	CTX_LOCK(ctx);
4262 	IFDI_LED_FUNC(ctx, onoff);
4263 	CTX_UNLOCK(ctx);
4264 }
4265 
4266 /*********************************************************************
4267  *
4268  *  BUS FUNCTION DEFINITIONS
4269  *
4270  **********************************************************************/
4271 
4272 int
4273 iflib_device_probe(device_t dev)
4274 {
4275 	pci_vendor_info_t *ent;
4276 
4277 	uint16_t	pci_vendor_id, pci_device_id;
4278 	uint16_t	pci_subvendor_id, pci_subdevice_id;
4279 	uint16_t	pci_rev_id;
4280 	if_shared_ctx_t sctx;
4281 
4282 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4283 		return (ENOTSUP);
4284 
4285 	pci_vendor_id = pci_get_vendor(dev);
4286 	pci_device_id = pci_get_device(dev);
4287 	pci_subvendor_id = pci_get_subvendor(dev);
4288 	pci_subdevice_id = pci_get_subdevice(dev);
4289 	pci_rev_id = pci_get_revid(dev);
4290 	if (sctx->isc_parse_devinfo != NULL)
4291 		sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id);
4292 
4293 	ent = sctx->isc_vendor_info;
4294 	while (ent->pvi_vendor_id != 0) {
4295 		if (pci_vendor_id != ent->pvi_vendor_id) {
4296 			ent++;
4297 			continue;
4298 		}
4299 		if ((pci_device_id == ent->pvi_device_id) &&
4300 		    ((pci_subvendor_id == ent->pvi_subvendor_id) ||
4301 		     (ent->pvi_subvendor_id == 0)) &&
4302 		    ((pci_subdevice_id == ent->pvi_subdevice_id) ||
4303 		     (ent->pvi_subdevice_id == 0)) &&
4304 		    ((pci_rev_id == ent->pvi_rev_id) ||
4305 		     (ent->pvi_rev_id == 0))) {
4306 
4307 			device_set_desc_copy(dev, ent->pvi_name);
4308 			/* this needs to be changed to zero if the bus probing code
4309 			 * ever stops re-probing on best match because the sctx
4310 			 * may have its values over written by register calls
4311 			 * in subsequent probes
4312 			 */
4313 			return (BUS_PROBE_DEFAULT);
4314 		}
4315 		ent++;
4316 	}
4317 	return (ENXIO);
4318 }
4319 
4320 static void
4321 iflib_reset_qvalues(if_ctx_t ctx)
4322 {
4323 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4324 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4325 	device_t dev = ctx->ifc_dev;
4326 	int i;
4327 
4328 	scctx->isc_txrx_budget_bytes_max = IFLIB_MAX_TX_BYTES;
4329 	scctx->isc_tx_qdepth = IFLIB_DEFAULT_TX_QDEPTH;
4330 	/*
4331 	 * XXX sanity check that ntxd & nrxd are a power of 2
4332 	 */
4333 	if (ctx->ifc_sysctl_ntxqs != 0)
4334 		scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs;
4335 	if (ctx->ifc_sysctl_nrxqs != 0)
4336 		scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs;
4337 
4338 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4339 		if (ctx->ifc_sysctl_ntxds[i] != 0)
4340 			scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i];
4341 		else
4342 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4343 	}
4344 
4345 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4346 		if (ctx->ifc_sysctl_nrxds[i] != 0)
4347 			scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i];
4348 		else
4349 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4350 	}
4351 
4352 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4353 		if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) {
4354 			device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n",
4355 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]);
4356 			scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i];
4357 		}
4358 		if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) {
4359 			device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n",
4360 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]);
4361 			scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i];
4362 		}
4363 	}
4364 
4365 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4366 		if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) {
4367 			device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n",
4368 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]);
4369 			scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i];
4370 		}
4371 		if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) {
4372 			device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n",
4373 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]);
4374 			scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i];
4375 		}
4376 	}
4377 }
4378 
4379 int
4380 iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp)
4381 {
4382 	int err, rid, msix;
4383 	if_ctx_t ctx;
4384 	if_t ifp;
4385 	if_softc_ctx_t scctx;
4386 	int i;
4387 	uint16_t main_txq;
4388 	uint16_t main_rxq;
4389 
4390 
4391 	ctx = malloc(sizeof(* ctx), M_IFLIB, M_WAITOK|M_ZERO);
4392 
4393 	if (sc == NULL) {
4394 		sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4395 		device_set_softc(dev, ctx);
4396 		ctx->ifc_flags |= IFC_SC_ALLOCATED;
4397 	}
4398 
4399 	ctx->ifc_sctx = sctx;
4400 	ctx->ifc_dev = dev;
4401 	ctx->ifc_softc = sc;
4402 
4403 	if ((err = iflib_register(ctx)) != 0) {
4404 		if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4405 			free(sc, M_IFLIB);
4406 		free(ctx, M_IFLIB);
4407 		device_printf(dev, "iflib_register failed %d\n", err);
4408 		return (err);
4409 	}
4410 	iflib_add_device_sysctl_pre(ctx);
4411 
4412 	scctx = &ctx->ifc_softc_ctx;
4413 	ifp = ctx->ifc_ifp;
4414 
4415 	iflib_reset_qvalues(ctx);
4416 	CTX_LOCK(ctx);
4417 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4418 		CTX_UNLOCK(ctx);
4419 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4420 		return (err);
4421 	}
4422 	_iflib_pre_assert(scctx);
4423 	ctx->ifc_txrx = *scctx->isc_txrx;
4424 
4425 #ifdef INVARIANTS
4426 	MPASS(scctx->isc_capabilities);
4427 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
4428 		MPASS(scctx->isc_tx_csum_flags);
4429 #endif
4430 
4431 	if_setcapabilities(ifp, scctx->isc_capabilities | IFCAP_HWSTATS);
4432 	if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS);
4433 
4434 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4435 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4436 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4437 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4438 
4439 #ifdef ACPI_DMAR
4440 	if (dmar_get_dma_tag(device_get_parent(dev), dev) != NULL)
4441 		ctx->ifc_flags |= IFC_DMAR;
4442 #elif !(defined(__i386__) || defined(__amd64__))
4443 	/* set unconditionally for !x86 */
4444 	ctx->ifc_flags |= IFC_DMAR;
4445 #endif
4446 
4447 	main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4448 	main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4449 
4450 	/* XXX change for per-queue sizes */
4451 	device_printf(dev, "using %d tx descriptors and %d rx descriptors\n",
4452 		      scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4453 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4454 		if (!powerof2(scctx->isc_nrxd[i])) {
4455 			/* round down instead? */
4456 			device_printf(dev, "# rx descriptors must be a power of 2\n");
4457 			err = EINVAL;
4458 			goto fail;
4459 		}
4460 	}
4461 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4462 		if (!powerof2(scctx->isc_ntxd[i])) {
4463 			device_printf(dev,
4464 			    "# tx descriptors must be a power of 2");
4465 			err = EINVAL;
4466 			goto fail;
4467 		}
4468 	}
4469 
4470 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4471 	    MAX_SINGLE_PACKET_FRACTION)
4472 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4473 		    MAX_SINGLE_PACKET_FRACTION);
4474 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4475 	    MAX_SINGLE_PACKET_FRACTION)
4476 		scctx->isc_tx_tso_segments_max = max(1,
4477 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4478 
4479 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4480 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
4481 		/*
4482 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
4483 		 * but some MACs do.
4484 		 */
4485 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
4486 		    IP_MAXPACKET));
4487 		/*
4488 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
4489 		 * into account.  In the worst case, each of these calls will
4490 		 * add another mbuf and, thus, the requirement for another DMA
4491 		 * segment.  So for best performance, it doesn't make sense to
4492 		 * advertize a maximum of TSO segments that typically will
4493 		 * require defragmentation in iflib_encap().
4494 		 */
4495 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
4496 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
4497 	}
4498 	if (scctx->isc_rss_table_size == 0)
4499 		scctx->isc_rss_table_size = 64;
4500 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4501 
4502 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4503 	/* XXX format name */
4504 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx, -1, "admin");
4505 
4506 	/* Set up cpu set.  If it fails, use the set of all CPUs. */
4507 	if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) != 0) {
4508 		device_printf(dev, "Unable to fetch CPU list\n");
4509 		CPU_COPY(&all_cpus, &ctx->ifc_cpus);
4510 	}
4511 	MPASS(CPU_COUNT(&ctx->ifc_cpus) > 0);
4512 
4513 	/*
4514 	** Now setup MSI or MSI/X, should
4515 	** return us the number of supported
4516 	** vectors. (Will be 1 for MSI)
4517 	*/
4518 	if (sctx->isc_flags & IFLIB_SKIP_MSIX) {
4519 		msix = scctx->isc_vectors;
4520 	} else if (scctx->isc_msix_bar != 0)
4521 	       /*
4522 		* The simple fact that isc_msix_bar is not 0 does not mean we
4523 		* we have a good value there that is known to work.
4524 		*/
4525 		msix = iflib_msix_init(ctx);
4526 	else {
4527 		scctx->isc_vectors = 1;
4528 		scctx->isc_ntxqsets = 1;
4529 		scctx->isc_nrxqsets = 1;
4530 		scctx->isc_intr = IFLIB_INTR_LEGACY;
4531 		msix = 0;
4532 	}
4533 	/* Get memory for the station queues */
4534 	if ((err = iflib_queues_alloc(ctx))) {
4535 		device_printf(dev, "Unable to allocate queue memory\n");
4536 		goto fail;
4537 	}
4538 
4539 	if ((err = iflib_qset_structures_setup(ctx)))
4540 		goto fail_queues;
4541 
4542 	/*
4543 	 * Group taskqueues aren't properly set up until SMP is started,
4544 	 * so we disable interrupts until we can handle them post
4545 	 * SI_SUB_SMP.
4546 	 *
4547 	 * XXX: disabling interrupts doesn't actually work, at least for
4548 	 * the non-MSI case.  When they occur before SI_SUB_SMP completes,
4549 	 * we do null handling and depend on this not causing too large an
4550 	 * interrupt storm.
4551 	 */
4552 	IFDI_INTR_DISABLE(ctx);
4553 	if (msix > 1 && (err = IFDI_MSIX_INTR_ASSIGN(ctx, msix)) != 0) {
4554 		device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n", err);
4555 		goto fail_intr_free;
4556 	}
4557 	if (msix <= 1) {
4558 		rid = 0;
4559 		if (scctx->isc_intr == IFLIB_INTR_MSI) {
4560 			MPASS(msix == 1);
4561 			rid = 1;
4562 		}
4563 		if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) {
4564 			device_printf(dev, "iflib_legacy_setup failed %d\n", err);
4565 			goto fail_intr_free;
4566 		}
4567 	}
4568 
4569 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac);
4570 
4571 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4572 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4573 		goto fail_detach;
4574 	}
4575 
4576 	/*
4577 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4578 	 * This must appear after the call to ether_ifattach() because
4579 	 * ether_ifattach() sets if_hdrlen to the default value.
4580 	 */
4581 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4582 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
4583 
4584 	if ((err = iflib_netmap_attach(ctx))) {
4585 		device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err);
4586 		goto fail_detach;
4587 	}
4588 	*ctxp = ctx;
4589 
4590 	NETDUMP_SET(ctx->ifc_ifp, iflib);
4591 
4592 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4593 	iflib_add_device_sysctl_post(ctx);
4594 	ctx->ifc_flags |= IFC_INIT_DONE;
4595 	CTX_UNLOCK(ctx);
4596 	return (0);
4597 fail_detach:
4598 	ether_ifdetach(ctx->ifc_ifp);
4599 fail_intr_free:
4600 	if (scctx->isc_intr == IFLIB_INTR_MSIX || scctx->isc_intr == IFLIB_INTR_MSI)
4601 		pci_release_msi(ctx->ifc_dev);
4602 fail_queues:
4603 	iflib_tx_structures_free(ctx);
4604 	iflib_rx_structures_free(ctx);
4605 fail:
4606 	IFDI_DETACH(ctx);
4607 	CTX_UNLOCK(ctx);
4608 	return (err);
4609 }
4610 
4611 int
4612 iflib_pseudo_register(device_t dev, if_shared_ctx_t sctx, if_ctx_t *ctxp,
4613 					  struct iflib_cloneattach_ctx *clctx)
4614 {
4615 	int err;
4616 	if_ctx_t ctx;
4617 	if_t ifp;
4618 	if_softc_ctx_t scctx;
4619 	int i;
4620 	void *sc;
4621 	uint16_t main_txq;
4622 	uint16_t main_rxq;
4623 
4624 	ctx = malloc(sizeof(*ctx), M_IFLIB, M_WAITOK|M_ZERO);
4625 	sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4626 	ctx->ifc_flags |= IFC_SC_ALLOCATED;
4627 	if (sctx->isc_flags & (IFLIB_PSEUDO|IFLIB_VIRTUAL))
4628 		ctx->ifc_flags |= IFC_PSEUDO;
4629 
4630 	ctx->ifc_sctx = sctx;
4631 	ctx->ifc_softc = sc;
4632 	ctx->ifc_dev = dev;
4633 
4634 	if ((err = iflib_register(ctx)) != 0) {
4635 		device_printf(dev, "%s: iflib_register failed %d\n", __func__, err);
4636 		free(sc, M_IFLIB);
4637 		free(ctx, M_IFLIB);
4638 		return (err);
4639 	}
4640 	iflib_add_device_sysctl_pre(ctx);
4641 
4642 	scctx = &ctx->ifc_softc_ctx;
4643 	ifp = ctx->ifc_ifp;
4644 
4645 	/*
4646 	 * XXX sanity check that ntxd & nrxd are a power of 2
4647 	 */
4648 	iflib_reset_qvalues(ctx);
4649 
4650 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4651 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4652 		return (err);
4653 	}
4654 	if (sctx->isc_flags & IFLIB_GEN_MAC)
4655 		iflib_gen_mac(ctx);
4656 	if ((err = IFDI_CLONEATTACH(ctx, clctx->cc_ifc, clctx->cc_name,
4657 								clctx->cc_params)) != 0) {
4658 		device_printf(dev, "IFDI_CLONEATTACH failed %d\n", err);
4659 		return (err);
4660 	}
4661 	ifmedia_add(&ctx->ifc_media, IFM_ETHER | IFM_1000_T | IFM_FDX, 0, NULL);
4662 	ifmedia_add(&ctx->ifc_media, IFM_ETHER | IFM_AUTO, 0, NULL);
4663 	ifmedia_set(&ctx->ifc_media, IFM_ETHER | IFM_AUTO);
4664 
4665 #ifdef INVARIANTS
4666 	MPASS(scctx->isc_capabilities);
4667 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
4668 		MPASS(scctx->isc_tx_csum_flags);
4669 #endif
4670 
4671 	if_setcapabilities(ifp, scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_LINKSTATE);
4672 	if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_LINKSTATE);
4673 
4674 	ifp->if_flags |= IFF_NOGROUP;
4675 	if (sctx->isc_flags & IFLIB_PSEUDO) {
4676 		ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac);
4677 
4678 		if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4679 			device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4680 			goto fail_detach;
4681 		}
4682 		*ctxp = ctx;
4683 
4684 		/*
4685 		 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4686 		 * This must appear after the call to ether_ifattach() because
4687 		 * ether_ifattach() sets if_hdrlen to the default value.
4688 		 */
4689 		if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4690 			if_setifheaderlen(ifp,
4691 			    sizeof(struct ether_vlan_header));
4692 
4693 		if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4694 		iflib_add_device_sysctl_post(ctx);
4695 		ctx->ifc_flags |= IFC_INIT_DONE;
4696 		return (0);
4697 	}
4698 	_iflib_pre_assert(scctx);
4699 	ctx->ifc_txrx = *scctx->isc_txrx;
4700 
4701 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4702 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4703 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4704 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4705 
4706 	main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4707 	main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4708 
4709 	/* XXX change for per-queue sizes */
4710 	device_printf(dev, "using %d tx descriptors and %d rx descriptors\n",
4711 		      scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4712 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4713 		if (!powerof2(scctx->isc_nrxd[i])) {
4714 			/* round down instead? */
4715 			device_printf(dev, "# rx descriptors must be a power of 2\n");
4716 			err = EINVAL;
4717 			goto fail;
4718 		}
4719 	}
4720 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4721 		if (!powerof2(scctx->isc_ntxd[i])) {
4722 			device_printf(dev,
4723 			    "# tx descriptors must be a power of 2");
4724 			err = EINVAL;
4725 			goto fail;
4726 		}
4727 	}
4728 
4729 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4730 	    MAX_SINGLE_PACKET_FRACTION)
4731 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4732 		    MAX_SINGLE_PACKET_FRACTION);
4733 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4734 	    MAX_SINGLE_PACKET_FRACTION)
4735 		scctx->isc_tx_tso_segments_max = max(1,
4736 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4737 
4738 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4739 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
4740 		/*
4741 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
4742 		 * but some MACs do.
4743 		 */
4744 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
4745 		    IP_MAXPACKET));
4746 		/*
4747 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
4748 		 * into account.  In the worst case, each of these calls will
4749 		 * add another mbuf and, thus, the requirement for another DMA
4750 		 * segment.  So for best performance, it doesn't make sense to
4751 		 * advertize a maximum of TSO segments that typically will
4752 		 * require defragmentation in iflib_encap().
4753 		 */
4754 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
4755 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
4756 	}
4757 	if (scctx->isc_rss_table_size == 0)
4758 		scctx->isc_rss_table_size = 64;
4759 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4760 
4761 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4762 	/* XXX format name */
4763 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx, -1, "admin");
4764 
4765 	/* XXX --- can support > 1 -- but keep it simple for now */
4766 	scctx->isc_intr = IFLIB_INTR_LEGACY;
4767 
4768 	/* Get memory for the station queues */
4769 	if ((err = iflib_queues_alloc(ctx))) {
4770 		device_printf(dev, "Unable to allocate queue memory\n");
4771 		goto fail;
4772 	}
4773 
4774 	if ((err = iflib_qset_structures_setup(ctx))) {
4775 		device_printf(dev, "qset structure setup failed %d\n", err);
4776 		goto fail_queues;
4777 	}
4778 
4779 	/*
4780 	 * XXX What if anything do we want to do about interrupts?
4781 	 */
4782 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac);
4783 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4784 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4785 		goto fail_detach;
4786 	}
4787 
4788 	/*
4789 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4790 	 * This must appear after the call to ether_ifattach() because
4791 	 * ether_ifattach() sets if_hdrlen to the default value.
4792 	 */
4793 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4794 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
4795 
4796 	/* XXX handle more than one queue */
4797 	for (i = 0; i < scctx->isc_nrxqsets; i++)
4798 		IFDI_RX_CLSET(ctx, 0, i, ctx->ifc_rxqs[i].ifr_fl[0].ifl_sds.ifsd_cl);
4799 
4800 	*ctxp = ctx;
4801 
4802 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4803 	iflib_add_device_sysctl_post(ctx);
4804 	ctx->ifc_flags |= IFC_INIT_DONE;
4805 	return (0);
4806 fail_detach:
4807 	ether_ifdetach(ctx->ifc_ifp);
4808 fail_queues:
4809 	iflib_tx_structures_free(ctx);
4810 	iflib_rx_structures_free(ctx);
4811 fail:
4812 	IFDI_DETACH(ctx);
4813 	return (err);
4814 }
4815 
4816 int
4817 iflib_pseudo_deregister(if_ctx_t ctx)
4818 {
4819 	if_t ifp = ctx->ifc_ifp;
4820 	iflib_txq_t txq;
4821 	iflib_rxq_t rxq;
4822 	int i, j;
4823 	struct taskqgroup *tqg;
4824 	iflib_fl_t fl;
4825 
4826 	/* Unregister VLAN events */
4827 	if (ctx->ifc_vlan_attach_event != NULL)
4828 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
4829 	if (ctx->ifc_vlan_detach_event != NULL)
4830 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
4831 
4832 	ether_ifdetach(ifp);
4833 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
4834 	CTX_LOCK_DESTROY(ctx);
4835 	/* XXX drain any dependent tasks */
4836 	tqg = qgroup_if_io_tqg;
4837 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
4838 		callout_drain(&txq->ift_timer);
4839 		if (txq->ift_task.gt_uniq != NULL)
4840 			taskqgroup_detach(tqg, &txq->ift_task);
4841 	}
4842 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4843 		if (rxq->ifr_task.gt_uniq != NULL)
4844 			taskqgroup_detach(tqg, &rxq->ifr_task);
4845 
4846 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
4847 			free(fl->ifl_rx_bitmap, M_IFLIB);
4848 	}
4849 	tqg = qgroup_if_config_tqg;
4850 	if (ctx->ifc_admin_task.gt_uniq != NULL)
4851 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
4852 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
4853 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
4854 
4855 	if_free(ifp);
4856 
4857 	iflib_tx_structures_free(ctx);
4858 	iflib_rx_structures_free(ctx);
4859 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4860 		free(ctx->ifc_softc, M_IFLIB);
4861 	free(ctx, M_IFLIB);
4862 	return (0);
4863 }
4864 
4865 int
4866 iflib_device_attach(device_t dev)
4867 {
4868 	if_ctx_t ctx;
4869 	if_shared_ctx_t sctx;
4870 
4871 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4872 		return (ENOTSUP);
4873 
4874 	pci_enable_busmaster(dev);
4875 
4876 	return (iflib_device_register(dev, NULL, sctx, &ctx));
4877 }
4878 
4879 int
4880 iflib_device_deregister(if_ctx_t ctx)
4881 {
4882 	if_t ifp = ctx->ifc_ifp;
4883 	iflib_txq_t txq;
4884 	iflib_rxq_t rxq;
4885 	device_t dev = ctx->ifc_dev;
4886 	int i, j;
4887 	struct taskqgroup *tqg;
4888 	iflib_fl_t fl;
4889 
4890 	/* Make sure VLANS are not using driver */
4891 	if (if_vlantrunkinuse(ifp)) {
4892 		device_printf(dev,"Vlan in use, detach first\n");
4893 		return (EBUSY);
4894 	}
4895 
4896 	CTX_LOCK(ctx);
4897 	ctx->ifc_in_detach = 1;
4898 	iflib_stop(ctx);
4899 	CTX_UNLOCK(ctx);
4900 
4901 	/* Unregister VLAN events */
4902 	if (ctx->ifc_vlan_attach_event != NULL)
4903 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
4904 	if (ctx->ifc_vlan_detach_event != NULL)
4905 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
4906 
4907 	iflib_netmap_detach(ifp);
4908 	ether_ifdetach(ifp);
4909 	if (ctx->ifc_led_dev != NULL)
4910 		led_destroy(ctx->ifc_led_dev);
4911 	/* XXX drain any dependent tasks */
4912 	tqg = qgroup_if_io_tqg;
4913 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
4914 		callout_drain(&txq->ift_timer);
4915 		if (txq->ift_task.gt_uniq != NULL)
4916 			taskqgroup_detach(tqg, &txq->ift_task);
4917 	}
4918 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4919 		if (rxq->ifr_task.gt_uniq != NULL)
4920 			taskqgroup_detach(tqg, &rxq->ifr_task);
4921 
4922 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
4923 			free(fl->ifl_rx_bitmap, M_IFLIB);
4924 
4925 	}
4926 	tqg = qgroup_if_config_tqg;
4927 	if (ctx->ifc_admin_task.gt_uniq != NULL)
4928 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
4929 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
4930 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
4931 	CTX_LOCK(ctx);
4932 	IFDI_DETACH(ctx);
4933 	CTX_UNLOCK(ctx);
4934 
4935 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
4936 	CTX_LOCK_DESTROY(ctx);
4937 	device_set_softc(ctx->ifc_dev, NULL);
4938 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) {
4939 		pci_release_msi(dev);
4940 	}
4941 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) {
4942 		iflib_irq_free(ctx, &ctx->ifc_legacy_irq);
4943 	}
4944 	if (ctx->ifc_msix_mem != NULL) {
4945 		bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY,
4946 			ctx->ifc_softc_ctx.isc_msix_bar, ctx->ifc_msix_mem);
4947 		ctx->ifc_msix_mem = NULL;
4948 	}
4949 
4950 	bus_generic_detach(dev);
4951 	if_free(ifp);
4952 
4953 	iflib_tx_structures_free(ctx);
4954 	iflib_rx_structures_free(ctx);
4955 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4956 		free(ctx->ifc_softc, M_IFLIB);
4957 	free(ctx, M_IFLIB);
4958 	return (0);
4959 }
4960 
4961 
4962 int
4963 iflib_device_detach(device_t dev)
4964 {
4965 	if_ctx_t ctx = device_get_softc(dev);
4966 
4967 	return (iflib_device_deregister(ctx));
4968 }
4969 
4970 int
4971 iflib_device_suspend(device_t dev)
4972 {
4973 	if_ctx_t ctx = device_get_softc(dev);
4974 
4975 	CTX_LOCK(ctx);
4976 	IFDI_SUSPEND(ctx);
4977 	CTX_UNLOCK(ctx);
4978 
4979 	return bus_generic_suspend(dev);
4980 }
4981 int
4982 iflib_device_shutdown(device_t dev)
4983 {
4984 	if_ctx_t ctx = device_get_softc(dev);
4985 
4986 	CTX_LOCK(ctx);
4987 	IFDI_SHUTDOWN(ctx);
4988 	CTX_UNLOCK(ctx);
4989 
4990 	return bus_generic_suspend(dev);
4991 }
4992 
4993 
4994 int
4995 iflib_device_resume(device_t dev)
4996 {
4997 	if_ctx_t ctx = device_get_softc(dev);
4998 	iflib_txq_t txq = ctx->ifc_txqs;
4999 
5000 	CTX_LOCK(ctx);
5001 	IFDI_RESUME(ctx);
5002 	iflib_init_locked(ctx);
5003 	CTX_UNLOCK(ctx);
5004 	for (int i = 0; i < NTXQSETS(ctx); i++, txq++)
5005 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
5006 
5007 	return (bus_generic_resume(dev));
5008 }
5009 
5010 int
5011 iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params)
5012 {
5013 	int error;
5014 	if_ctx_t ctx = device_get_softc(dev);
5015 
5016 	CTX_LOCK(ctx);
5017 	error = IFDI_IOV_INIT(ctx, num_vfs, params);
5018 	CTX_UNLOCK(ctx);
5019 
5020 	return (error);
5021 }
5022 
5023 void
5024 iflib_device_iov_uninit(device_t dev)
5025 {
5026 	if_ctx_t ctx = device_get_softc(dev);
5027 
5028 	CTX_LOCK(ctx);
5029 	IFDI_IOV_UNINIT(ctx);
5030 	CTX_UNLOCK(ctx);
5031 }
5032 
5033 int
5034 iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params)
5035 {
5036 	int error;
5037 	if_ctx_t ctx = device_get_softc(dev);
5038 
5039 	CTX_LOCK(ctx);
5040 	error = IFDI_IOV_VF_ADD(ctx, vfnum, params);
5041 	CTX_UNLOCK(ctx);
5042 
5043 	return (error);
5044 }
5045 
5046 /*********************************************************************
5047  *
5048  *  MODULE FUNCTION DEFINITIONS
5049  *
5050  **********************************************************************/
5051 
5052 /*
5053  * - Start a fast taskqueue thread for each core
5054  * - Start a taskqueue for control operations
5055  */
5056 static int
5057 iflib_module_init(void)
5058 {
5059 	return (0);
5060 }
5061 
5062 static int
5063 iflib_module_event_handler(module_t mod, int what, void *arg)
5064 {
5065 	int err;
5066 
5067 	switch (what) {
5068 	case MOD_LOAD:
5069 		if ((err = iflib_module_init()) != 0)
5070 			return (err);
5071 		break;
5072 	case MOD_UNLOAD:
5073 		return (EBUSY);
5074 	default:
5075 		return (EOPNOTSUPP);
5076 	}
5077 
5078 	return (0);
5079 }
5080 
5081 /*********************************************************************
5082  *
5083  *  PUBLIC FUNCTION DEFINITIONS
5084  *     ordered as in iflib.h
5085  *
5086  **********************************************************************/
5087 
5088 
5089 static void
5090 _iflib_assert(if_shared_ctx_t sctx)
5091 {
5092 	MPASS(sctx->isc_tx_maxsize);
5093 	MPASS(sctx->isc_tx_maxsegsize);
5094 
5095 	MPASS(sctx->isc_rx_maxsize);
5096 	MPASS(sctx->isc_rx_nsegments);
5097 	MPASS(sctx->isc_rx_maxsegsize);
5098 
5099 	MPASS(sctx->isc_nrxd_min[0]);
5100 	MPASS(sctx->isc_nrxd_max[0]);
5101 	MPASS(sctx->isc_nrxd_default[0]);
5102 	MPASS(sctx->isc_ntxd_min[0]);
5103 	MPASS(sctx->isc_ntxd_max[0]);
5104 	MPASS(sctx->isc_ntxd_default[0]);
5105 }
5106 
5107 static void
5108 _iflib_pre_assert(if_softc_ctx_t scctx)
5109 {
5110 
5111 	MPASS(scctx->isc_txrx->ift_txd_encap);
5112 	MPASS(scctx->isc_txrx->ift_txd_flush);
5113 	MPASS(scctx->isc_txrx->ift_txd_credits_update);
5114 	MPASS(scctx->isc_txrx->ift_rxd_available);
5115 	MPASS(scctx->isc_txrx->ift_rxd_pkt_get);
5116 	MPASS(scctx->isc_txrx->ift_rxd_refill);
5117 	MPASS(scctx->isc_txrx->ift_rxd_flush);
5118 }
5119 
5120 static int
5121 iflib_register(if_ctx_t ctx)
5122 {
5123 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5124 	driver_t *driver = sctx->isc_driver;
5125 	device_t dev = ctx->ifc_dev;
5126 	if_t ifp;
5127 
5128 	_iflib_assert(sctx);
5129 
5130 	CTX_LOCK_INIT(ctx);
5131 	STATE_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev));
5132 	ifp = ctx->ifc_ifp = if_gethandle(IFT_ETHER);
5133 	if (ifp == NULL) {
5134 		device_printf(dev, "can not allocate ifnet structure\n");
5135 		return (ENOMEM);
5136 	}
5137 
5138 	/*
5139 	 * Initialize our context's device specific methods
5140 	 */
5141 	kobj_init((kobj_t) ctx, (kobj_class_t) driver);
5142 	kobj_class_compile((kobj_class_t) driver);
5143 	driver->refs++;
5144 
5145 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
5146 	if_setsoftc(ifp, ctx);
5147 	if_setdev(ifp, dev);
5148 	if_setinitfn(ifp, iflib_if_init);
5149 	if_setioctlfn(ifp, iflib_if_ioctl);
5150 	if_settransmitfn(ifp, iflib_if_transmit);
5151 	if_setqflushfn(ifp, iflib_if_qflush);
5152 	if_setflags(ifp, IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST);
5153 
5154 	ctx->ifc_vlan_attach_event =
5155 		EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx,
5156 							  EVENTHANDLER_PRI_FIRST);
5157 	ctx->ifc_vlan_detach_event =
5158 		EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx,
5159 							  EVENTHANDLER_PRI_FIRST);
5160 
5161 	ifmedia_init(&ctx->ifc_media, IFM_IMASK,
5162 					 iflib_media_change, iflib_media_status);
5163 
5164 	return (0);
5165 }
5166 
5167 
5168 static int
5169 iflib_queues_alloc(if_ctx_t ctx)
5170 {
5171 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5172 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5173 	device_t dev = ctx->ifc_dev;
5174 	int nrxqsets = scctx->isc_nrxqsets;
5175 	int ntxqsets = scctx->isc_ntxqsets;
5176 	iflib_txq_t txq;
5177 	iflib_rxq_t rxq;
5178 	iflib_fl_t fl = NULL;
5179 	int i, j, cpu, err, txconf, rxconf;
5180 	iflib_dma_info_t ifdip;
5181 	uint32_t *rxqsizes = scctx->isc_rxqsizes;
5182 	uint32_t *txqsizes = scctx->isc_txqsizes;
5183 	uint8_t nrxqs = sctx->isc_nrxqs;
5184 	uint8_t ntxqs = sctx->isc_ntxqs;
5185 	int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1;
5186 	caddr_t *vaddrs;
5187 	uint64_t *paddrs;
5188 
5189 	KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1"));
5190 	KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1"));
5191 
5192 	/* Allocate the TX ring struct memory */
5193 	if (!(ctx->ifc_txqs =
5194 	    (iflib_txq_t) malloc(sizeof(struct iflib_txq) *
5195 	    ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5196 		device_printf(dev, "Unable to allocate TX ring memory\n");
5197 		err = ENOMEM;
5198 		goto fail;
5199 	}
5200 
5201 	/* Now allocate the RX */
5202 	if (!(ctx->ifc_rxqs =
5203 	    (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) *
5204 	    nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5205 		device_printf(dev, "Unable to allocate RX ring memory\n");
5206 		err = ENOMEM;
5207 		goto rx_fail;
5208 	}
5209 
5210 	txq = ctx->ifc_txqs;
5211 	rxq = ctx->ifc_rxqs;
5212 
5213 	/*
5214 	 * XXX handle allocation failure
5215 	 */
5216 	for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) {
5217 		/* Set up some basics */
5218 
5219 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs, M_IFLIB, M_WAITOK|M_ZERO)) == NULL) {
5220 			device_printf(dev, "failed to allocate iflib_dma_info\n");
5221 			err = ENOMEM;
5222 			goto err_tx_desc;
5223 		}
5224 		txq->ift_ifdi = ifdip;
5225 		for (j = 0; j < ntxqs; j++, ifdip++) {
5226 			if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, BUS_DMA_NOWAIT)) {
5227 				device_printf(dev, "Unable to allocate Descriptor memory\n");
5228 				err = ENOMEM;
5229 				goto err_tx_desc;
5230 			}
5231 			txq->ift_txd_size[j] = scctx->isc_txd_size[j];
5232 			bzero((void *)ifdip->idi_vaddr, txqsizes[j]);
5233 		}
5234 		txq->ift_ctx = ctx;
5235 		txq->ift_id = i;
5236 		if (sctx->isc_flags & IFLIB_HAS_TXCQ) {
5237 			txq->ift_br_offset = 1;
5238 		} else {
5239 			txq->ift_br_offset = 0;
5240 		}
5241 		/* XXX fix this */
5242 		txq->ift_timer.c_cpu = cpu;
5243 
5244 		if (iflib_txsd_alloc(txq)) {
5245 			device_printf(dev, "Critical Failure setting up TX buffers\n");
5246 			err = ENOMEM;
5247 			goto err_tx_desc;
5248 		}
5249 
5250 		/* Initialize the TX lock */
5251 		snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:tx(%d):callout",
5252 		    device_get_nameunit(dev), txq->ift_id);
5253 		mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF);
5254 		callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0);
5255 
5256 		snprintf(txq->ift_db_mtx_name, MTX_NAME_LEN, "%s:tx(%d):db",
5257 			 device_get_nameunit(dev), txq->ift_id);
5258 
5259 		err = ifmp_ring_alloc(&txq->ift_br, 2048, txq, iflib_txq_drain,
5260 				      iflib_txq_can_drain, M_IFLIB, M_WAITOK);
5261 		if (err) {
5262 			/* XXX free any allocated rings */
5263 			device_printf(dev, "Unable to allocate buf_ring\n");
5264 			goto err_tx_desc;
5265 		}
5266 	}
5267 
5268 	for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) {
5269 		/* Set up some basics */
5270 
5271 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs, M_IFLIB, M_WAITOK|M_ZERO)) == NULL) {
5272 			device_printf(dev, "failed to allocate iflib_dma_info\n");
5273 			err = ENOMEM;
5274 			goto err_tx_desc;
5275 		}
5276 
5277 		rxq->ifr_ifdi = ifdip;
5278 		/* XXX this needs to be changed if #rx queues != #tx queues */
5279 		rxq->ifr_ntxqirq = 1;
5280 		rxq->ifr_txqid[0] = i;
5281 		for (j = 0; j < nrxqs; j++, ifdip++) {
5282 			if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, BUS_DMA_NOWAIT)) {
5283 				device_printf(dev, "Unable to allocate Descriptor memory\n");
5284 				err = ENOMEM;
5285 				goto err_tx_desc;
5286 			}
5287 			bzero((void *)ifdip->idi_vaddr, rxqsizes[j]);
5288 		}
5289 		rxq->ifr_ctx = ctx;
5290 		rxq->ifr_id = i;
5291 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
5292 			rxq->ifr_fl_offset = 1;
5293 		} else {
5294 			rxq->ifr_fl_offset = 0;
5295 		}
5296 		rxq->ifr_nfl = nfree_lists;
5297 		if (!(fl =
5298 			  (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) {
5299 			device_printf(dev, "Unable to allocate free list memory\n");
5300 			err = ENOMEM;
5301 			goto err_tx_desc;
5302 		}
5303 		rxq->ifr_fl = fl;
5304 		for (j = 0; j < nfree_lists; j++) {
5305 			fl[j].ifl_rxq = rxq;
5306 			fl[j].ifl_id = j;
5307 			fl[j].ifl_ifdi = &rxq->ifr_ifdi[j + rxq->ifr_fl_offset];
5308 			fl[j].ifl_rxd_size = scctx->isc_rxd_size[j];
5309 		}
5310         /* Allocate receive buffers for the ring*/
5311 		if (iflib_rxsd_alloc(rxq)) {
5312 			device_printf(dev,
5313 			    "Critical Failure setting up receive buffers\n");
5314 			err = ENOMEM;
5315 			goto err_rx_desc;
5316 		}
5317 
5318 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5319 			fl->ifl_rx_bitmap = bit_alloc(fl->ifl_size, M_IFLIB, M_WAITOK|M_ZERO);
5320 	}
5321 
5322 	/* TXQs */
5323 	vaddrs = malloc(sizeof(caddr_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5324 	paddrs = malloc(sizeof(uint64_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5325 	for (i = 0; i < ntxqsets; i++) {
5326 		iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi;
5327 
5328 		for (j = 0; j < ntxqs; j++, di++) {
5329 			vaddrs[i*ntxqs + j] = di->idi_vaddr;
5330 			paddrs[i*ntxqs + j] = di->idi_paddr;
5331 		}
5332 	}
5333 	if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) {
5334 		device_printf(ctx->ifc_dev, "device queue allocation failed\n");
5335 		iflib_tx_structures_free(ctx);
5336 		free(vaddrs, M_IFLIB);
5337 		free(paddrs, M_IFLIB);
5338 		goto err_rx_desc;
5339 	}
5340 	free(vaddrs, M_IFLIB);
5341 	free(paddrs, M_IFLIB);
5342 
5343 	/* RXQs */
5344 	vaddrs = malloc(sizeof(caddr_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5345 	paddrs = malloc(sizeof(uint64_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5346 	for (i = 0; i < nrxqsets; i++) {
5347 		iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi;
5348 
5349 		for (j = 0; j < nrxqs; j++, di++) {
5350 			vaddrs[i*nrxqs + j] = di->idi_vaddr;
5351 			paddrs[i*nrxqs + j] = di->idi_paddr;
5352 		}
5353 	}
5354 	if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) {
5355 		device_printf(ctx->ifc_dev, "device queue allocation failed\n");
5356 		iflib_tx_structures_free(ctx);
5357 		free(vaddrs, M_IFLIB);
5358 		free(paddrs, M_IFLIB);
5359 		goto err_rx_desc;
5360 	}
5361 	free(vaddrs, M_IFLIB);
5362 	free(paddrs, M_IFLIB);
5363 
5364 	return (0);
5365 
5366 /* XXX handle allocation failure changes */
5367 err_rx_desc:
5368 err_tx_desc:
5369 rx_fail:
5370 	if (ctx->ifc_rxqs != NULL)
5371 		free(ctx->ifc_rxqs, M_IFLIB);
5372 	ctx->ifc_rxqs = NULL;
5373 	if (ctx->ifc_txqs != NULL)
5374 		free(ctx->ifc_txqs, M_IFLIB);
5375 	ctx->ifc_txqs = NULL;
5376 fail:
5377 	return (err);
5378 }
5379 
5380 static int
5381 iflib_tx_structures_setup(if_ctx_t ctx)
5382 {
5383 	iflib_txq_t txq = ctx->ifc_txqs;
5384 	int i;
5385 
5386 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
5387 		iflib_txq_setup(txq);
5388 
5389 	return (0);
5390 }
5391 
5392 static void
5393 iflib_tx_structures_free(if_ctx_t ctx)
5394 {
5395 	iflib_txq_t txq = ctx->ifc_txqs;
5396 	int i, j;
5397 
5398 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
5399 		iflib_txq_destroy(txq);
5400 		for (j = 0; j < ctx->ifc_nhwtxqs; j++)
5401 			iflib_dma_free(&txq->ift_ifdi[j]);
5402 	}
5403 	free(ctx->ifc_txqs, M_IFLIB);
5404 	ctx->ifc_txqs = NULL;
5405 	IFDI_QUEUES_FREE(ctx);
5406 }
5407 
5408 /*********************************************************************
5409  *
5410  *  Initialize all receive rings.
5411  *
5412  **********************************************************************/
5413 static int
5414 iflib_rx_structures_setup(if_ctx_t ctx)
5415 {
5416 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5417 	int q;
5418 #if defined(INET6) || defined(INET)
5419 	int i, err;
5420 #endif
5421 
5422 	for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) {
5423 #if defined(INET6) || defined(INET)
5424 		tcp_lro_free(&rxq->ifr_lc);
5425 		if ((err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp,
5426 		    TCP_LRO_ENTRIES, min(1024,
5427 		    ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset]))) != 0) {
5428 			device_printf(ctx->ifc_dev, "LRO Initialization failed!\n");
5429 			goto fail;
5430 		}
5431 		rxq->ifr_lro_enabled = TRUE;
5432 #endif
5433 		IFDI_RXQ_SETUP(ctx, rxq->ifr_id);
5434 	}
5435 	return (0);
5436 #if defined(INET6) || defined(INET)
5437 fail:
5438 	/*
5439 	 * Free RX software descriptors allocated so far, we will only handle
5440 	 * the rings that completed, the failing case will have
5441 	 * cleaned up for itself. 'q' failed, so its the terminus.
5442 	 */
5443 	rxq = ctx->ifc_rxqs;
5444 	for (i = 0; i < q; ++i, rxq++) {
5445 		iflib_rx_sds_free(rxq);
5446 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
5447 	}
5448 	return (err);
5449 #endif
5450 }
5451 
5452 /*********************************************************************
5453  *
5454  *  Free all receive rings.
5455  *
5456  **********************************************************************/
5457 static void
5458 iflib_rx_structures_free(if_ctx_t ctx)
5459 {
5460 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5461 
5462 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) {
5463 		iflib_rx_sds_free(rxq);
5464 	}
5465 }
5466 
5467 static int
5468 iflib_qset_structures_setup(if_ctx_t ctx)
5469 {
5470 	int err;
5471 
5472 	/*
5473 	 * It is expected that the caller takes care of freeing queues if this
5474 	 * fails.
5475 	 */
5476 	if ((err = iflib_tx_structures_setup(ctx)) != 0) {
5477 		device_printf(ctx->ifc_dev, "iflib_tx_structures_setup failed: %d\n", err);
5478 		return (err);
5479 	}
5480 
5481 	if ((err = iflib_rx_structures_setup(ctx)) != 0)
5482 		device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err);
5483 
5484 	return (err);
5485 }
5486 
5487 int
5488 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
5489 		driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, const char *name)
5490 {
5491 
5492 	return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name));
5493 }
5494 
5495 #ifdef SMP
5496 static int
5497 find_nth(if_ctx_t ctx, int qid)
5498 {
5499 	cpuset_t cpus;
5500 	int i, cpuid, eqid, count;
5501 
5502 	CPU_COPY(&ctx->ifc_cpus, &cpus);
5503 	count = CPU_COUNT(&cpus);
5504 	eqid = qid % count;
5505 	/* clear up to the qid'th bit */
5506 	for (i = 0; i < eqid; i++) {
5507 		cpuid = CPU_FFS(&cpus);
5508 		MPASS(cpuid != 0);
5509 		CPU_CLR(cpuid-1, &cpus);
5510 	}
5511 	cpuid = CPU_FFS(&cpus);
5512 	MPASS(cpuid != 0);
5513 	return (cpuid-1);
5514 }
5515 
5516 #ifdef SCHED_ULE
5517 extern struct cpu_group *cpu_top;              /* CPU topology */
5518 
5519 static int
5520 find_child_with_core(int cpu, struct cpu_group *grp)
5521 {
5522 	int i;
5523 
5524 	if (grp->cg_children == 0)
5525 		return -1;
5526 
5527 	MPASS(grp->cg_child);
5528 	for (i = 0; i < grp->cg_children; i++) {
5529 		if (CPU_ISSET(cpu, &grp->cg_child[i].cg_mask))
5530 			return i;
5531 	}
5532 
5533 	return -1;
5534 }
5535 
5536 /*
5537  * Find the nth "close" core to the specified core
5538  * "close" is defined as the deepest level that shares
5539  * at least an L2 cache.  With threads, this will be
5540  * threads on the same core.  If the sahred cache is L3
5541  * or higher, simply returns the same core.
5542  */
5543 static int
5544 find_close_core(int cpu, int core_offset)
5545 {
5546 	struct cpu_group *grp;
5547 	int i;
5548 	int fcpu;
5549 	cpuset_t cs;
5550 
5551 	grp = cpu_top;
5552 	if (grp == NULL)
5553 		return cpu;
5554 	i = 0;
5555 	while ((i = find_child_with_core(cpu, grp)) != -1) {
5556 		/* If the child only has one cpu, don't descend */
5557 		if (grp->cg_child[i].cg_count <= 1)
5558 			break;
5559 		grp = &grp->cg_child[i];
5560 	}
5561 
5562 	/* If they don't share at least an L2 cache, use the same CPU */
5563 	if (grp->cg_level > CG_SHARE_L2 || grp->cg_level == CG_SHARE_NONE)
5564 		return cpu;
5565 
5566 	/* Now pick one */
5567 	CPU_COPY(&grp->cg_mask, &cs);
5568 
5569 	/* Add the selected CPU offset to core offset. */
5570 	for (i = 0; (fcpu = CPU_FFS(&cs)) != 0; i++) {
5571 		if (fcpu - 1 == cpu)
5572 			break;
5573 		CPU_CLR(fcpu - 1, &cs);
5574 	}
5575 	MPASS(fcpu);
5576 
5577 	core_offset += i;
5578 
5579 	CPU_COPY(&grp->cg_mask, &cs);
5580 	for (i = core_offset % grp->cg_count; i > 0; i--) {
5581 		MPASS(CPU_FFS(&cs));
5582 		CPU_CLR(CPU_FFS(&cs) - 1, &cs);
5583 	}
5584 	MPASS(CPU_FFS(&cs));
5585 	return CPU_FFS(&cs) - 1;
5586 }
5587 #else
5588 static int
5589 find_close_core(int cpu, int core_offset __unused)
5590 {
5591 	return cpu;
5592 }
5593 #endif
5594 
5595 static int
5596 get_core_offset(if_ctx_t ctx, iflib_intr_type_t type, int qid)
5597 {
5598 	switch (type) {
5599 	case IFLIB_INTR_TX:
5600 		/* TX queues get cores which share at least an L2 cache with the corresponding RX queue */
5601 		/* XXX handle multiple RX threads per core and more than two core per L2 group */
5602 		return qid / CPU_COUNT(&ctx->ifc_cpus) + 1;
5603 	case IFLIB_INTR_RX:
5604 	case IFLIB_INTR_RXTX:
5605 		/* RX queues get the specified core */
5606 		return qid / CPU_COUNT(&ctx->ifc_cpus);
5607 	default:
5608 		return -1;
5609 	}
5610 }
5611 #else
5612 #define get_core_offset(ctx, type, qid)	CPU_FIRST()
5613 #define find_close_core(cpuid, tid)	CPU_FIRST()
5614 #define find_nth(ctx, gid)		CPU_FIRST()
5615 #endif
5616 
5617 /* Just to avoid copy/paste */
5618 static inline int
5619 iflib_irq_set_affinity(if_ctx_t ctx, int irq, iflib_intr_type_t type, int qid,
5620     struct grouptask *gtask, struct taskqgroup *tqg, void *uniq, const char *name)
5621 {
5622 	int cpuid;
5623 	int err, tid;
5624 
5625 	cpuid = find_nth(ctx, qid);
5626 	tid = get_core_offset(ctx, type, qid);
5627 	MPASS(tid >= 0);
5628 	cpuid = find_close_core(cpuid, tid);
5629 	err = taskqgroup_attach_cpu(tqg, gtask, uniq, cpuid, irq, name);
5630 	if (err) {
5631 		device_printf(ctx->ifc_dev, "taskqgroup_attach_cpu failed %d\n", err);
5632 		return (err);
5633 	}
5634 #ifdef notyet
5635 	if (cpuid > ctx->ifc_cpuid_highest)
5636 		ctx->ifc_cpuid_highest = cpuid;
5637 #endif
5638 	return 0;
5639 }
5640 
5641 int
5642 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid,
5643 			iflib_intr_type_t type, driver_filter_t *filter,
5644 			void *filter_arg, int qid, const char *name)
5645 {
5646 	struct grouptask *gtask;
5647 	struct taskqgroup *tqg;
5648 	iflib_filter_info_t info;
5649 	gtask_fn_t *fn;
5650 	int tqrid, err;
5651 	driver_filter_t *intr_fast;
5652 	void *q;
5653 
5654 	info = &ctx->ifc_filter_info;
5655 	tqrid = rid;
5656 
5657 	switch (type) {
5658 	/* XXX merge tx/rx for netmap? */
5659 	case IFLIB_INTR_TX:
5660 		q = &ctx->ifc_txqs[qid];
5661 		info = &ctx->ifc_txqs[qid].ift_filter_info;
5662 		gtask = &ctx->ifc_txqs[qid].ift_task;
5663 		tqg = qgroup_if_io_tqg;
5664 		fn = _task_fn_tx;
5665 		intr_fast = iflib_fast_intr;
5666 		GROUPTASK_INIT(gtask, 0, fn, q);
5667 		ctx->ifc_flags |= IFC_NETMAP_TX_IRQ;
5668 		break;
5669 	case IFLIB_INTR_RX:
5670 		q = &ctx->ifc_rxqs[qid];
5671 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
5672 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5673 		tqg = qgroup_if_io_tqg;
5674 		fn = _task_fn_rx;
5675 		intr_fast = iflib_fast_intr;
5676 		GROUPTASK_INIT(gtask, 0, fn, q);
5677 		break;
5678 	case IFLIB_INTR_RXTX:
5679 		q = &ctx->ifc_rxqs[qid];
5680 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
5681 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5682 		tqg = qgroup_if_io_tqg;
5683 		fn = _task_fn_rx;
5684 		intr_fast = iflib_fast_intr_rxtx;
5685 		GROUPTASK_INIT(gtask, 0, fn, q);
5686 		break;
5687 	case IFLIB_INTR_ADMIN:
5688 		q = ctx;
5689 		tqrid = -1;
5690 		info = &ctx->ifc_filter_info;
5691 		gtask = &ctx->ifc_admin_task;
5692 		tqg = qgroup_if_config_tqg;
5693 		fn = _task_fn_admin;
5694 		intr_fast = iflib_fast_intr_ctx;
5695 		break;
5696 	default:
5697 		panic("unknown net intr type");
5698 	}
5699 
5700 	info->ifi_filter = filter;
5701 	info->ifi_filter_arg = filter_arg;
5702 	info->ifi_task = gtask;
5703 	info->ifi_ctx = q;
5704 
5705 	err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info,  name);
5706 	if (err != 0) {
5707 		device_printf(ctx->ifc_dev, "_iflib_irq_alloc failed %d\n", err);
5708 		return (err);
5709 	}
5710 	if (type == IFLIB_INTR_ADMIN)
5711 		return (0);
5712 
5713 	if (tqrid != -1) {
5714 		err = iflib_irq_set_affinity(ctx, rman_get_start(irq->ii_res), type, qid, gtask, tqg, q, name);
5715 		if (err)
5716 			return (err);
5717 	} else {
5718 		taskqgroup_attach(tqg, gtask, q, rman_get_start(irq->ii_res), name);
5719 	}
5720 
5721 	return (0);
5722 }
5723 
5724 void
5725 iflib_softirq_alloc_generic(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type,  void *arg, int qid, const char *name)
5726 {
5727 	struct grouptask *gtask;
5728 	struct taskqgroup *tqg;
5729 	gtask_fn_t *fn;
5730 	void *q;
5731 	int irq_num = -1;
5732 	int err;
5733 
5734 	switch (type) {
5735 	case IFLIB_INTR_TX:
5736 		q = &ctx->ifc_txqs[qid];
5737 		gtask = &ctx->ifc_txqs[qid].ift_task;
5738 		tqg = qgroup_if_io_tqg;
5739 		fn = _task_fn_tx;
5740 		if (irq != NULL)
5741 			irq_num = rman_get_start(irq->ii_res);
5742 		break;
5743 	case IFLIB_INTR_RX:
5744 		q = &ctx->ifc_rxqs[qid];
5745 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5746 		tqg = qgroup_if_io_tqg;
5747 		fn = _task_fn_rx;
5748 		if (irq != NULL)
5749 			irq_num = rman_get_start(irq->ii_res);
5750 		break;
5751 	case IFLIB_INTR_IOV:
5752 		q = ctx;
5753 		gtask = &ctx->ifc_vflr_task;
5754 		tqg = qgroup_if_config_tqg;
5755 		fn = _task_fn_iov;
5756 		break;
5757 	default:
5758 		panic("unknown net intr type");
5759 	}
5760 	GROUPTASK_INIT(gtask, 0, fn, q);
5761 	if (irq_num != -1) {
5762 		err = iflib_irq_set_affinity(ctx, irq_num, type, qid, gtask, tqg, q, name);
5763 		if (err)
5764 			taskqgroup_attach(tqg, gtask, q, irq_num, name);
5765 	}
5766 	else {
5767 		taskqgroup_attach(tqg, gtask, q, irq_num, name);
5768 	}
5769 }
5770 
5771 void
5772 iflib_irq_free(if_ctx_t ctx, if_irq_t irq)
5773 {
5774 	if (irq->ii_tag)
5775 		bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag);
5776 
5777 	if (irq->ii_res)
5778 		bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ, irq->ii_rid, irq->ii_res);
5779 }
5780 
5781 static int
5782 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, const char *name)
5783 {
5784 	iflib_txq_t txq = ctx->ifc_txqs;
5785 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5786 	if_irq_t irq = &ctx->ifc_legacy_irq;
5787 	iflib_filter_info_t info;
5788 	struct grouptask *gtask;
5789 	struct taskqgroup *tqg;
5790 	gtask_fn_t *fn;
5791 	int tqrid;
5792 	void *q;
5793 	int err;
5794 
5795 	q = &ctx->ifc_rxqs[0];
5796 	info = &rxq[0].ifr_filter_info;
5797 	gtask = &rxq[0].ifr_task;
5798 	tqg = qgroup_if_io_tqg;
5799 	tqrid = irq->ii_rid = *rid;
5800 	fn = _task_fn_rx;
5801 
5802 	ctx->ifc_flags |= IFC_LEGACY;
5803 	info->ifi_filter = filter;
5804 	info->ifi_filter_arg = filter_arg;
5805 	info->ifi_task = gtask;
5806 	info->ifi_ctx = ctx;
5807 
5808 	/* We allocate a single interrupt resource */
5809 	if ((err = _iflib_irq_alloc(ctx, irq, tqrid, iflib_fast_intr_ctx, NULL, info, name)) != 0)
5810 		return (err);
5811 	GROUPTASK_INIT(gtask, 0, fn, q);
5812 	taskqgroup_attach(tqg, gtask, q, rman_get_start(irq->ii_res), name);
5813 
5814 	GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq);
5815 	taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, rman_get_start(irq->ii_res), "tx");
5816 	return (0);
5817 }
5818 
5819 void
5820 iflib_led_create(if_ctx_t ctx)
5821 {
5822 
5823 	ctx->ifc_led_dev = led_create(iflib_led_func, ctx,
5824 	    device_get_nameunit(ctx->ifc_dev));
5825 }
5826 
5827 void
5828 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid)
5829 {
5830 
5831 	GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
5832 }
5833 
5834 void
5835 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid)
5836 {
5837 
5838 	GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task);
5839 }
5840 
5841 void
5842 iflib_admin_intr_deferred(if_ctx_t ctx)
5843 {
5844 #ifdef INVARIANTS
5845 	struct grouptask *gtask;
5846 
5847 	gtask = &ctx->ifc_admin_task;
5848 	MPASS(gtask != NULL && gtask->gt_taskqueue != NULL);
5849 #endif
5850 
5851 	GROUPTASK_ENQUEUE(&ctx->ifc_admin_task);
5852 }
5853 
5854 void
5855 iflib_iov_intr_deferred(if_ctx_t ctx)
5856 {
5857 
5858 	GROUPTASK_ENQUEUE(&ctx->ifc_vflr_task);
5859 }
5860 
5861 void
5862 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, char *name)
5863 {
5864 
5865 	taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, -1, name);
5866 }
5867 
5868 void
5869 iflib_config_gtask_init(void *ctx, struct grouptask *gtask, gtask_fn_t *fn,
5870 	const char *name)
5871 {
5872 
5873 	GROUPTASK_INIT(gtask, 0, fn, ctx);
5874 	taskqgroup_attach(qgroup_if_config_tqg, gtask, gtask, -1, name);
5875 }
5876 
5877 void
5878 iflib_config_gtask_deinit(struct grouptask *gtask)
5879 {
5880 
5881 	taskqgroup_detach(qgroup_if_config_tqg, gtask);
5882 }
5883 
5884 void
5885 iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate)
5886 {
5887 	if_t ifp = ctx->ifc_ifp;
5888 	iflib_txq_t txq = ctx->ifc_txqs;
5889 
5890 	if_setbaudrate(ifp, baudrate);
5891 	if (baudrate >= IF_Gbps(10)) {
5892 		STATE_LOCK(ctx);
5893 		ctx->ifc_flags |= IFC_PREFETCH;
5894 		STATE_UNLOCK(ctx);
5895 	}
5896 	/* If link down, disable watchdog */
5897 	if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) {
5898 		for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++)
5899 			txq->ift_qstatus = IFLIB_QUEUE_IDLE;
5900 	}
5901 	ctx->ifc_link_state = link_state;
5902 	if_link_state_change(ifp, link_state);
5903 }
5904 
5905 static int
5906 iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq)
5907 {
5908 	int credits;
5909 #ifdef INVARIANTS
5910 	int credits_pre = txq->ift_cidx_processed;
5911 #endif
5912 
5913 	if (ctx->isc_txd_credits_update == NULL)
5914 		return (0);
5915 
5916 	if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, true)) == 0)
5917 		return (0);
5918 
5919 	txq->ift_processed += credits;
5920 	txq->ift_cidx_processed += credits;
5921 
5922 	MPASS(credits_pre + credits == txq->ift_cidx_processed);
5923 	if (txq->ift_cidx_processed >= txq->ift_size)
5924 		txq->ift_cidx_processed -= txq->ift_size;
5925 	return (credits);
5926 }
5927 
5928 static int
5929 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget)
5930 {
5931 
5932 	return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx,
5933 	    budget));
5934 }
5935 
5936 void
5937 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name,
5938 	const char *description, if_int_delay_info_t info,
5939 	int offset, int value)
5940 {
5941 	info->iidi_ctx = ctx;
5942 	info->iidi_offset = offset;
5943 	info->iidi_value = value;
5944 	SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev),
5945 	    SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)),
5946 	    OID_AUTO, name, CTLTYPE_INT|CTLFLAG_RW,
5947 	    info, 0, iflib_sysctl_int_delay, "I", description);
5948 }
5949 
5950 struct sx *
5951 iflib_ctx_lock_get(if_ctx_t ctx)
5952 {
5953 
5954 	return (&ctx->ifc_ctx_sx);
5955 }
5956 
5957 static int
5958 iflib_msix_init(if_ctx_t ctx)
5959 {
5960 	device_t dev = ctx->ifc_dev;
5961 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5962 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5963 	int vectors, queues, rx_queues, tx_queues, queuemsgs, msgs;
5964 	int iflib_num_tx_queues, iflib_num_rx_queues;
5965 	int err, admincnt, bar;
5966 
5967 	iflib_num_tx_queues = ctx->ifc_sysctl_ntxqs;
5968 	iflib_num_rx_queues = ctx->ifc_sysctl_nrxqs;
5969 
5970 	device_printf(dev, "msix_init qsets capped at %d\n", imax(scctx->isc_ntxqsets, scctx->isc_nrxqsets));
5971 
5972 	bar = ctx->ifc_softc_ctx.isc_msix_bar;
5973 	admincnt = sctx->isc_admin_intrcnt;
5974 	/* Override by tuneable */
5975 	if (scctx->isc_disable_msix)
5976 		goto msi;
5977 
5978 	/*
5979 	 * bar == -1 => "trust me I know what I'm doing"
5980 	 * Some drivers are for hardware that is so shoddily
5981 	 * documented that no one knows which bars are which
5982 	 * so the developer has to map all bars. This hack
5983 	 * allows shoddy garbage to use msix in this framework.
5984 	 */
5985 	if (bar != -1) {
5986 		ctx->ifc_msix_mem = bus_alloc_resource_any(dev,
5987 	            SYS_RES_MEMORY, &bar, RF_ACTIVE);
5988 		if (ctx->ifc_msix_mem == NULL) {
5989 			/* May not be enabled */
5990 			device_printf(dev, "Unable to map MSIX table \n");
5991 			goto msi;
5992 		}
5993 	}
5994 	/* First try MSI/X */
5995 	if ((msgs = pci_msix_count(dev)) == 0) { /* system has msix disabled */
5996 		device_printf(dev, "System has MSIX disabled \n");
5997 		bus_release_resource(dev, SYS_RES_MEMORY,
5998 		    bar, ctx->ifc_msix_mem);
5999 		ctx->ifc_msix_mem = NULL;
6000 		goto msi;
6001 	}
6002 #if IFLIB_DEBUG
6003 	/* use only 1 qset in debug mode */
6004 	queuemsgs = min(msgs - admincnt, 1);
6005 #else
6006 	queuemsgs = msgs - admincnt;
6007 #endif
6008 #ifdef RSS
6009 	queues = imin(queuemsgs, rss_getnumbuckets());
6010 #else
6011 	queues = queuemsgs;
6012 #endif
6013 	queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues);
6014 	device_printf(dev, "pxm cpus: %d queue msgs: %d admincnt: %d\n",
6015 				  CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt);
6016 #ifdef  RSS
6017 	/* If we're doing RSS, clamp at the number of RSS buckets */
6018 	if (queues > rss_getnumbuckets())
6019 		queues = rss_getnumbuckets();
6020 #endif
6021 	if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt)
6022 		rx_queues = iflib_num_rx_queues;
6023 	else
6024 		rx_queues = queues;
6025 
6026 	if (rx_queues > scctx->isc_nrxqsets)
6027 		rx_queues = scctx->isc_nrxqsets;
6028 
6029 	/*
6030 	 * We want this to be all logical CPUs by default
6031 	 */
6032 	if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues)
6033 		tx_queues = iflib_num_tx_queues;
6034 	else
6035 		tx_queues = mp_ncpus;
6036 
6037 	if (tx_queues > scctx->isc_ntxqsets)
6038 		tx_queues = scctx->isc_ntxqsets;
6039 
6040 	if (ctx->ifc_sysctl_qs_eq_override == 0) {
6041 #ifdef INVARIANTS
6042 		if (tx_queues != rx_queues)
6043 			device_printf(dev, "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n",
6044 				      min(rx_queues, tx_queues), min(rx_queues, tx_queues));
6045 #endif
6046 		tx_queues = min(rx_queues, tx_queues);
6047 		rx_queues = min(rx_queues, tx_queues);
6048 	}
6049 
6050 	device_printf(dev, "using %d rx queues %d tx queues \n", rx_queues, tx_queues);
6051 
6052 	vectors = rx_queues + admincnt;
6053 	if ((err = pci_alloc_msix(dev, &vectors)) == 0) {
6054 		device_printf(dev,
6055 					  "Using MSIX interrupts with %d vectors\n", vectors);
6056 		scctx->isc_vectors = vectors;
6057 		scctx->isc_nrxqsets = rx_queues;
6058 		scctx->isc_ntxqsets = tx_queues;
6059 		scctx->isc_intr = IFLIB_INTR_MSIX;
6060 
6061 		return (vectors);
6062 	} else {
6063 		device_printf(dev, "failed to allocate %d msix vectors, err: %d - using MSI\n", vectors, err);
6064 		bus_release_resource(dev, SYS_RES_MEMORY, bar,
6065 		    ctx->ifc_msix_mem);
6066 		ctx->ifc_msix_mem = NULL;
6067 	}
6068 msi:
6069 	vectors = pci_msi_count(dev);
6070 	scctx->isc_nrxqsets = 1;
6071 	scctx->isc_ntxqsets = 1;
6072 	scctx->isc_vectors = vectors;
6073 	if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) {
6074 		device_printf(dev,"Using an MSI interrupt\n");
6075 		scctx->isc_intr = IFLIB_INTR_MSI;
6076 	} else {
6077 		scctx->isc_vectors = 1;
6078 		device_printf(dev,"Using a Legacy interrupt\n");
6079 		scctx->isc_intr = IFLIB_INTR_LEGACY;
6080 	}
6081 
6082 	return (vectors);
6083 }
6084 
6085 static const char *ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" };
6086 
6087 static int
6088 mp_ring_state_handler(SYSCTL_HANDLER_ARGS)
6089 {
6090 	int rc;
6091 	uint16_t *state = ((uint16_t *)oidp->oid_arg1);
6092 	struct sbuf *sb;
6093 	const char *ring_state = "UNKNOWN";
6094 
6095 	/* XXX needed ? */
6096 	rc = sysctl_wire_old_buffer(req, 0);
6097 	MPASS(rc == 0);
6098 	if (rc != 0)
6099 		return (rc);
6100 	sb = sbuf_new_for_sysctl(NULL, NULL, 80, req);
6101 	MPASS(sb != NULL);
6102 	if (sb == NULL)
6103 		return (ENOMEM);
6104 	if (state[3] <= 3)
6105 		ring_state = ring_states[state[3]];
6106 
6107 	sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s",
6108 		    state[0], state[1], state[2], ring_state);
6109 	rc = sbuf_finish(sb);
6110 	sbuf_delete(sb);
6111         return(rc);
6112 }
6113 
6114 enum iflib_ndesc_handler {
6115 	IFLIB_NTXD_HANDLER,
6116 	IFLIB_NRXD_HANDLER,
6117 };
6118 
6119 static int
6120 mp_ndesc_handler(SYSCTL_HANDLER_ARGS)
6121 {
6122 	if_ctx_t ctx = (void *)arg1;
6123 	enum iflib_ndesc_handler type = arg2;
6124 	char buf[256] = {0};
6125 	qidx_t *ndesc;
6126 	char *p, *next;
6127 	int nqs, rc, i;
6128 
6129 	MPASS(type == IFLIB_NTXD_HANDLER || type == IFLIB_NRXD_HANDLER);
6130 
6131 	nqs = 8;
6132 	switch(type) {
6133 	case IFLIB_NTXD_HANDLER:
6134 		ndesc = ctx->ifc_sysctl_ntxds;
6135 		if (ctx->ifc_sctx)
6136 			nqs = ctx->ifc_sctx->isc_ntxqs;
6137 		break;
6138 	case IFLIB_NRXD_HANDLER:
6139 		ndesc = ctx->ifc_sysctl_nrxds;
6140 		if (ctx->ifc_sctx)
6141 			nqs = ctx->ifc_sctx->isc_nrxqs;
6142 		break;
6143 	default:
6144 			panic("unhandled type");
6145 	}
6146 	if (nqs == 0)
6147 		nqs = 8;
6148 
6149 	for (i=0; i<8; i++) {
6150 		if (i >= nqs)
6151 			break;
6152 		if (i)
6153 			strcat(buf, ",");
6154 		sprintf(strchr(buf, 0), "%d", ndesc[i]);
6155 	}
6156 
6157 	rc = sysctl_handle_string(oidp, buf, sizeof(buf), req);
6158 	if (rc || req->newptr == NULL)
6159 		return rc;
6160 
6161 	for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p;
6162 	    i++, p = strsep(&next, " ,")) {
6163 		ndesc[i] = strtoul(p, NULL, 10);
6164 	}
6165 
6166 	return(rc);
6167 }
6168 
6169 #define NAME_BUFLEN 32
6170 static void
6171 iflib_add_device_sysctl_pre(if_ctx_t ctx)
6172 {
6173         device_t dev = iflib_get_dev(ctx);
6174 	struct sysctl_oid_list *child, *oid_list;
6175 	struct sysctl_ctx_list *ctx_list;
6176 	struct sysctl_oid *node;
6177 
6178 	ctx_list = device_get_sysctl_ctx(dev);
6179 	child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev));
6180 	ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, "iflib",
6181 						      CTLFLAG_RD, NULL, "IFLIB fields");
6182 	oid_list = SYSCTL_CHILDREN(node);
6183 
6184 	SYSCTL_ADD_STRING(ctx_list, oid_list, OID_AUTO, "driver_version",
6185 		       CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version, 0,
6186 		       "driver version");
6187 
6188 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs",
6189 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0,
6190 			"# of txqs to use, 0 => use default #");
6191 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs",
6192 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0,
6193 			"# of rxqs to use, 0 => use default #");
6194 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_qs_enable",
6195 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0,
6196                        "permit #txq != #rxq");
6197 	SYSCTL_ADD_INT(ctx_list, oid_list, OID_AUTO, "disable_msix",
6198                       CTLFLAG_RWTUN, &ctx->ifc_softc_ctx.isc_disable_msix, 0,
6199                       "disable MSIX (default 0)");
6200 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "rx_budget",
6201 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_rx_budget, 0,
6202                        "set the rx budget");
6203 
6204 	/* XXX change for per-queue sizes */
6205 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_ntxds",
6206 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NTXD_HANDLER,
6207                        mp_ndesc_handler, "A",
6208                        "list of # of tx descriptors to use, 0 = use default #");
6209 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_nrxds",
6210 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NRXD_HANDLER,
6211                        mp_ndesc_handler, "A",
6212                        "list of # of rx descriptors to use, 0 = use default #");
6213 }
6214 
6215 static void
6216 iflib_add_device_sysctl_post(if_ctx_t ctx)
6217 {
6218 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6219 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6220         device_t dev = iflib_get_dev(ctx);
6221 	struct sysctl_oid_list *child;
6222 	struct sysctl_ctx_list *ctx_list;
6223 	iflib_fl_t fl;
6224 	iflib_txq_t txq;
6225 	iflib_rxq_t rxq;
6226 	int i, j;
6227 	char namebuf[NAME_BUFLEN];
6228 	char *qfmt;
6229 	struct sysctl_oid *queue_node, *fl_node, *node;
6230 	struct sysctl_oid_list *queue_list, *fl_list;
6231 	ctx_list = device_get_sysctl_ctx(dev);
6232 
6233 	node = ctx->ifc_sysctl_node;
6234 	child = SYSCTL_CHILDREN(node);
6235 
6236 	if (scctx->isc_ntxqsets > 100)
6237 		qfmt = "txq%03d";
6238 	else if (scctx->isc_ntxqsets > 10)
6239 		qfmt = "txq%02d";
6240 	else
6241 		qfmt = "txq%d";
6242 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
6243 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6244 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6245 					     CTLFLAG_RD, NULL, "Queue Name");
6246 		queue_list = SYSCTL_CHILDREN(queue_node);
6247 #if MEMORY_LOGGING
6248 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued",
6249 				CTLFLAG_RD,
6250 				&txq->ift_dequeued, "total mbufs freed");
6251 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued",
6252 				CTLFLAG_RD,
6253 				&txq->ift_enqueued, "total mbufs enqueued");
6254 #endif
6255 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag",
6256 				   CTLFLAG_RD,
6257 				   &txq->ift_mbuf_defrag, "# of times m_defrag was called");
6258 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "m_pullups",
6259 				   CTLFLAG_RD,
6260 				   &txq->ift_pullups, "# of times m_pullup was called");
6261 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag_failed",
6262 				   CTLFLAG_RD,
6263 				   &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed");
6264 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_desc_avail",
6265 				   CTLFLAG_RD,
6266 				   &txq->ift_no_desc_avail, "# of times no descriptors were available");
6267 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "tx_map_failed",
6268 				   CTLFLAG_RD,
6269 				   &txq->ift_map_failed, "# of times dma map failed");
6270 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txd_encap_efbig",
6271 				   CTLFLAG_RD,
6272 				   &txq->ift_txd_encap_efbig, "# of times txd_encap returned EFBIG");
6273 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_tx_dma_setup",
6274 				   CTLFLAG_RD,
6275 				   &txq->ift_no_tx_dma_setup, "# of times map failed for other than EFBIG");
6276 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx",
6277 				   CTLFLAG_RD,
6278 				   &txq->ift_pidx, 1, "Producer Index");
6279 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx",
6280 				   CTLFLAG_RD,
6281 				   &txq->ift_cidx, 1, "Consumer Index");
6282 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx_processed",
6283 				   CTLFLAG_RD,
6284 				   &txq->ift_cidx_processed, 1, "Consumer Index seen by credit update");
6285 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use",
6286 				   CTLFLAG_RD,
6287 				   &txq->ift_in_use, 1, "descriptors in use");
6288 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_processed",
6289 				   CTLFLAG_RD,
6290 				   &txq->ift_processed, "descriptors procesed for clean");
6291 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned",
6292 				   CTLFLAG_RD,
6293 				   &txq->ift_cleaned, "total cleaned");
6294 		SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state",
6295 				CTLTYPE_STRING | CTLFLAG_RD, __DEVOLATILE(uint64_t *, &txq->ift_br->state),
6296 				0, mp_ring_state_handler, "A", "soft ring state");
6297 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_enqueues",
6298 				       CTLFLAG_RD, &txq->ift_br->enqueues,
6299 				       "# of enqueues to the mp_ring for this queue");
6300 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_drops",
6301 				       CTLFLAG_RD, &txq->ift_br->drops,
6302 				       "# of drops in the mp_ring for this queue");
6303 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_starts",
6304 				       CTLFLAG_RD, &txq->ift_br->starts,
6305 				       "# of normal consumer starts in the mp_ring for this queue");
6306 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_stalls",
6307 				       CTLFLAG_RD, &txq->ift_br->stalls,
6308 					       "# of consumer stalls in the mp_ring for this queue");
6309 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_restarts",
6310 			       CTLFLAG_RD, &txq->ift_br->restarts,
6311 				       "# of consumer restarts in the mp_ring for this queue");
6312 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_abdications",
6313 				       CTLFLAG_RD, &txq->ift_br->abdications,
6314 				       "# of consumer abdications in the mp_ring for this queue");
6315 	}
6316 
6317 	if (scctx->isc_nrxqsets > 100)
6318 		qfmt = "rxq%03d";
6319 	else if (scctx->isc_nrxqsets > 10)
6320 		qfmt = "rxq%02d";
6321 	else
6322 		qfmt = "rxq%d";
6323 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
6324 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6325 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6326 					     CTLFLAG_RD, NULL, "Queue Name");
6327 		queue_list = SYSCTL_CHILDREN(queue_node);
6328 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
6329 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_pidx",
6330 				       CTLFLAG_RD,
6331 				       &rxq->ifr_cq_pidx, 1, "Producer Index");
6332 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_cidx",
6333 				       CTLFLAG_RD,
6334 				       &rxq->ifr_cq_cidx, 1, "Consumer Index");
6335 		}
6336 
6337 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
6338 			snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j);
6339 			fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list, OID_AUTO, namebuf,
6340 						     CTLFLAG_RD, NULL, "freelist Name");
6341 			fl_list = SYSCTL_CHILDREN(fl_node);
6342 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx",
6343 				       CTLFLAG_RD,
6344 				       &fl->ifl_pidx, 1, "Producer Index");
6345 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx",
6346 				       CTLFLAG_RD,
6347 				       &fl->ifl_cidx, 1, "Consumer Index");
6348 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits",
6349 				       CTLFLAG_RD,
6350 				       &fl->ifl_credits, 1, "credits available");
6351 #if MEMORY_LOGGING
6352 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_enqueued",
6353 					CTLFLAG_RD,
6354 					&fl->ifl_m_enqueued, "mbufs allocated");
6355 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_dequeued",
6356 					CTLFLAG_RD,
6357 					&fl->ifl_m_dequeued, "mbufs freed");
6358 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_enqueued",
6359 					CTLFLAG_RD,
6360 					&fl->ifl_cl_enqueued, "clusters allocated");
6361 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_dequeued",
6362 					CTLFLAG_RD,
6363 					&fl->ifl_cl_dequeued, "clusters freed");
6364 #endif
6365 
6366 		}
6367 	}
6368 
6369 }
6370 
6371 #ifndef __NO_STRICT_ALIGNMENT
6372 static struct mbuf *
6373 iflib_fixup_rx(struct mbuf *m)
6374 {
6375 	struct mbuf *n;
6376 
6377 	if (m->m_len <= (MCLBYTES - ETHER_HDR_LEN)) {
6378 		bcopy(m->m_data, m->m_data + ETHER_HDR_LEN, m->m_len);
6379 		m->m_data += ETHER_HDR_LEN;
6380 		n = m;
6381 	} else {
6382 		MGETHDR(n, M_NOWAIT, MT_DATA);
6383 		if (n == NULL) {
6384 			m_freem(m);
6385 			return (NULL);
6386 		}
6387 		bcopy(m->m_data, n->m_data, ETHER_HDR_LEN);
6388 		m->m_data += ETHER_HDR_LEN;
6389 		m->m_len -= ETHER_HDR_LEN;
6390 		n->m_len = ETHER_HDR_LEN;
6391 		M_MOVE_PKTHDR(n, m);
6392 		n->m_next = m;
6393 	}
6394 	return (n);
6395 }
6396 #endif
6397 
6398 #ifdef NETDUMP
6399 static void
6400 iflib_netdump_init(struct ifnet *ifp, int *nrxr, int *ncl, int *clsize)
6401 {
6402 	if_ctx_t ctx;
6403 
6404 	ctx = if_getsoftc(ifp);
6405 	CTX_LOCK(ctx);
6406 	*nrxr = NRXQSETS(ctx);
6407 	*ncl = ctx->ifc_rxqs[0].ifr_fl->ifl_size;
6408 	*clsize = ctx->ifc_rxqs[0].ifr_fl->ifl_buf_size;
6409 	CTX_UNLOCK(ctx);
6410 }
6411 
6412 static void
6413 iflib_netdump_event(struct ifnet *ifp, enum netdump_ev event)
6414 {
6415 	if_ctx_t ctx;
6416 	if_softc_ctx_t scctx;
6417 	iflib_fl_t fl;
6418 	iflib_rxq_t rxq;
6419 	int i, j;
6420 
6421 	ctx = if_getsoftc(ifp);
6422 	scctx = &ctx->ifc_softc_ctx;
6423 
6424 	switch (event) {
6425 	case NETDUMP_START:
6426 		for (i = 0; i < scctx->isc_nrxqsets; i++) {
6427 			rxq = &ctx->ifc_rxqs[i];
6428 			for (j = 0; j < rxq->ifr_nfl; j++) {
6429 				fl = rxq->ifr_fl;
6430 				fl->ifl_zone = m_getzone(fl->ifl_buf_size);
6431 			}
6432 		}
6433 		iflib_no_tx_batch = 1;
6434 		break;
6435 	default:
6436 		break;
6437 	}
6438 }
6439 
6440 static int
6441 iflib_netdump_transmit(struct ifnet *ifp, struct mbuf *m)
6442 {
6443 	if_ctx_t ctx;
6444 	iflib_txq_t txq;
6445 	int error;
6446 
6447 	ctx = if_getsoftc(ifp);
6448 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
6449 	    IFF_DRV_RUNNING)
6450 		return (EBUSY);
6451 
6452 	txq = &ctx->ifc_txqs[0];
6453 	error = iflib_encap(txq, &m);
6454 	if (error == 0)
6455 		(void)iflib_txd_db_check(ctx, txq, true, txq->ift_in_use);
6456 	return (error);
6457 }
6458 
6459 static int
6460 iflib_netdump_poll(struct ifnet *ifp, int count)
6461 {
6462 	if_ctx_t ctx;
6463 	if_softc_ctx_t scctx;
6464 	iflib_txq_t txq;
6465 	int i;
6466 
6467 	ctx = if_getsoftc(ifp);
6468 	scctx = &ctx->ifc_softc_ctx;
6469 
6470 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
6471 	    IFF_DRV_RUNNING)
6472 		return (EBUSY);
6473 
6474 	txq = &ctx->ifc_txqs[0];
6475 	(void)iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
6476 
6477 	for (i = 0; i < scctx->isc_nrxqsets; i++)
6478 		(void)iflib_rxeof(&ctx->ifc_rxqs[i], 16 /* XXX */);
6479 	return (0);
6480 }
6481 #endif /* NETDUMP */
6482