xref: /freebsd/sys/net/iflib.c (revision 43d1e6ee299ad4e143d90d3ad374d1c24bd3306f)
1 /*-
2  * Copyright (c) 2014-2018, Matthew Macy <mmacy@mattmacy.io>
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions are met:
7  *
8  *  1. Redistributions of source code must retain the above copyright notice,
9  *     this list of conditions and the following disclaimer.
10  *
11  *  2. Neither the name of Matthew Macy nor the names of its
12  *     contributors may be used to endorse or promote products derived from
13  *     this software without specific prior written permission.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
16  * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18  * ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
19  * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
20  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
21  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
22  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
23  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
24  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
25  * POSSIBILITY OF SUCH DAMAGE.
26  */
27 
28 #include <sys/cdefs.h>
29 __FBSDID("$FreeBSD$");
30 
31 #include "opt_inet.h"
32 #include "opt_inet6.h"
33 #include "opt_acpi.h"
34 #include "opt_sched.h"
35 
36 #include <sys/param.h>
37 #include <sys/types.h>
38 #include <sys/bus.h>
39 #include <sys/eventhandler.h>
40 #include <sys/kernel.h>
41 #include <sys/lock.h>
42 #include <sys/mutex.h>
43 #include <sys/module.h>
44 #include <sys/kobj.h>
45 #include <sys/rman.h>
46 #include <sys/sbuf.h>
47 #include <sys/smp.h>
48 #include <sys/socket.h>
49 #include <sys/sockio.h>
50 #include <sys/sysctl.h>
51 #include <sys/syslog.h>
52 #include <sys/taskqueue.h>
53 #include <sys/limits.h>
54 
55 #include <net/if.h>
56 #include <net/if_var.h>
57 #include <net/if_types.h>
58 #include <net/if_media.h>
59 #include <net/bpf.h>
60 #include <net/ethernet.h>
61 #include <net/mp_ring.h>
62 #include <net/vnet.h>
63 
64 #include <netinet/in.h>
65 #include <netinet/in_pcb.h>
66 #include <netinet/tcp_lro.h>
67 #include <netinet/in_systm.h>
68 #include <netinet/if_ether.h>
69 #include <netinet/ip.h>
70 #include <netinet/ip6.h>
71 #include <netinet/tcp.h>
72 #include <netinet/ip_var.h>
73 #include <netinet/netdump/netdump.h>
74 #include <netinet6/ip6_var.h>
75 
76 #include <machine/bus.h>
77 #include <machine/in_cksum.h>
78 
79 #include <vm/vm.h>
80 #include <vm/pmap.h>
81 
82 #include <dev/led/led.h>
83 #include <dev/pci/pcireg.h>
84 #include <dev/pci/pcivar.h>
85 #include <dev/pci/pci_private.h>
86 
87 #include <net/iflib.h>
88 #include <net/iflib_private.h>
89 
90 #include "ifdi_if.h"
91 
92 #ifdef PCI_IOV
93 #include <dev/pci/pci_iov.h>
94 #endif
95 
96 #include <sys/bitstring.h>
97 /*
98  * enable accounting of every mbuf as it comes in to and goes out of
99  * iflib's software descriptor references
100  */
101 #define MEMORY_LOGGING 0
102 /*
103  * Enable mbuf vectors for compressing long mbuf chains
104  */
105 
106 /*
107  * NB:
108  * - Prefetching in tx cleaning should perhaps be a tunable. The distance ahead
109  *   we prefetch needs to be determined by the time spent in m_free vis a vis
110  *   the cost of a prefetch. This will of course vary based on the workload:
111  *      - NFLX's m_free path is dominated by vm-based M_EXT manipulation which
112  *        is quite expensive, thus suggesting very little prefetch.
113  *      - small packet forwarding which is just returning a single mbuf to
114  *        UMA will typically be very fast vis a vis the cost of a memory
115  *        access.
116  */
117 
118 
119 /*
120  * File organization:
121  *  - private structures
122  *  - iflib private utility functions
123  *  - ifnet functions
124  *  - vlan registry and other exported functions
125  *  - iflib public core functions
126  *
127  *
128  */
129 MALLOC_DEFINE(M_IFLIB, "iflib", "ifnet library");
130 
131 struct iflib_txq;
132 typedef struct iflib_txq *iflib_txq_t;
133 struct iflib_rxq;
134 typedef struct iflib_rxq *iflib_rxq_t;
135 struct iflib_fl;
136 typedef struct iflib_fl *iflib_fl_t;
137 
138 struct iflib_ctx;
139 
140 static void iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid);
141 static void iflib_timer(void *arg);
142 
143 typedef struct iflib_filter_info {
144 	driver_filter_t *ifi_filter;
145 	void *ifi_filter_arg;
146 	struct grouptask *ifi_task;
147 	void *ifi_ctx;
148 } *iflib_filter_info_t;
149 
150 struct iflib_ctx {
151 	KOBJ_FIELDS;
152 	/*
153 	 * Pointer to hardware driver's softc
154 	 */
155 	void *ifc_softc;
156 	device_t ifc_dev;
157 	if_t ifc_ifp;
158 
159 	cpuset_t ifc_cpus;
160 	if_shared_ctx_t ifc_sctx;
161 	struct if_softc_ctx ifc_softc_ctx;
162 
163 	struct sx ifc_ctx_sx;
164 	struct mtx ifc_state_mtx;
165 
166 	iflib_txq_t ifc_txqs;
167 	iflib_rxq_t ifc_rxqs;
168 	uint32_t ifc_if_flags;
169 	uint32_t ifc_flags;
170 	uint32_t ifc_max_fl_buf_size;
171 	uint32_t ifc_rx_mbuf_sz;
172 
173 	int ifc_link_state;
174 	int ifc_link_irq;
175 	int ifc_watchdog_events;
176 	struct cdev *ifc_led_dev;
177 	struct resource *ifc_msix_mem;
178 
179 	struct if_irq ifc_legacy_irq;
180 	struct grouptask ifc_admin_task;
181 	struct grouptask ifc_vflr_task;
182 	struct iflib_filter_info ifc_filter_info;
183 	struct ifmedia	ifc_media;
184 
185 	struct sysctl_oid *ifc_sysctl_node;
186 	uint16_t ifc_sysctl_ntxqs;
187 	uint16_t ifc_sysctl_nrxqs;
188 	uint16_t ifc_sysctl_qs_eq_override;
189 	uint16_t ifc_sysctl_rx_budget;
190 	uint16_t ifc_sysctl_tx_abdicate;
191 
192 	qidx_t ifc_sysctl_ntxds[8];
193 	qidx_t ifc_sysctl_nrxds[8];
194 	struct if_txrx ifc_txrx;
195 #define isc_txd_encap  ifc_txrx.ift_txd_encap
196 #define isc_txd_flush  ifc_txrx.ift_txd_flush
197 #define isc_txd_credits_update  ifc_txrx.ift_txd_credits_update
198 #define isc_rxd_available ifc_txrx.ift_rxd_available
199 #define isc_rxd_pkt_get ifc_txrx.ift_rxd_pkt_get
200 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
201 #define isc_rxd_flush ifc_txrx.ift_rxd_flush
202 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
203 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
204 #define isc_legacy_intr ifc_txrx.ift_legacy_intr
205 	eventhandler_tag ifc_vlan_attach_event;
206 	eventhandler_tag ifc_vlan_detach_event;
207 	struct ether_addr ifc_mac;
208 	char ifc_mtx_name[16];
209 };
210 
211 
212 void *
213 iflib_get_softc(if_ctx_t ctx)
214 {
215 
216 	return (ctx->ifc_softc);
217 }
218 
219 device_t
220 iflib_get_dev(if_ctx_t ctx)
221 {
222 
223 	return (ctx->ifc_dev);
224 }
225 
226 if_t
227 iflib_get_ifp(if_ctx_t ctx)
228 {
229 
230 	return (ctx->ifc_ifp);
231 }
232 
233 struct ifmedia *
234 iflib_get_media(if_ctx_t ctx)
235 {
236 
237 	return (&ctx->ifc_media);
238 }
239 
240 uint32_t
241 iflib_get_flags(if_ctx_t ctx)
242 {
243 	return (ctx->ifc_flags);
244 }
245 
246 void
247 iflib_set_mac(if_ctx_t ctx, uint8_t mac[ETHER_ADDR_LEN])
248 {
249 
250 	bcopy(mac, ctx->ifc_mac.octet, ETHER_ADDR_LEN);
251 }
252 
253 if_softc_ctx_t
254 iflib_get_softc_ctx(if_ctx_t ctx)
255 {
256 
257 	return (&ctx->ifc_softc_ctx);
258 }
259 
260 if_shared_ctx_t
261 iflib_get_sctx(if_ctx_t ctx)
262 {
263 
264 	return (ctx->ifc_sctx);
265 }
266 
267 #define IP_ALIGNED(m) ((((uintptr_t)(m)->m_data) & 0x3) == 0x2)
268 #define CACHE_PTR_INCREMENT (CACHE_LINE_SIZE/sizeof(void*))
269 #define CACHE_PTR_NEXT(ptr) ((void *)(((uintptr_t)(ptr)+CACHE_LINE_SIZE-1) & (CACHE_LINE_SIZE-1)))
270 
271 #define LINK_ACTIVE(ctx) ((ctx)->ifc_link_state == LINK_STATE_UP)
272 #define CTX_IS_VF(ctx) ((ctx)->ifc_sctx->isc_flags & IFLIB_IS_VF)
273 
274 typedef struct iflib_sw_rx_desc_array {
275 	bus_dmamap_t	*ifsd_map;         /* bus_dma maps for packet */
276 	struct mbuf	**ifsd_m;           /* pkthdr mbufs */
277 	caddr_t		*ifsd_cl;          /* direct cluster pointer for rx */
278 	bus_addr_t	*ifsd_ba;          /* bus addr of cluster for rx */
279 } iflib_rxsd_array_t;
280 
281 typedef struct iflib_sw_tx_desc_array {
282 	bus_dmamap_t    *ifsd_map;         /* bus_dma maps for packet */
283 	bus_dmamap_t	*ifsd_tso_map;     /* bus_dma maps for TSO packet */
284 	struct mbuf    **ifsd_m;           /* pkthdr mbufs */
285 } if_txsd_vec_t;
286 
287 
288 /* magic number that should be high enough for any hardware */
289 #define IFLIB_MAX_TX_SEGS		128
290 #define IFLIB_RX_COPY_THRESH		128
291 #define IFLIB_MAX_RX_REFRESH		32
292 /* The minimum descriptors per second before we start coalescing */
293 #define IFLIB_MIN_DESC_SEC		16384
294 #define IFLIB_DEFAULT_TX_UPDATE_FREQ	16
295 #define IFLIB_QUEUE_IDLE		0
296 #define IFLIB_QUEUE_HUNG		1
297 #define IFLIB_QUEUE_WORKING		2
298 /* maximum number of txqs that can share an rx interrupt */
299 #define IFLIB_MAX_TX_SHARED_INTR	4
300 
301 /* this should really scale with ring size - this is a fairly arbitrary value */
302 #define TX_BATCH_SIZE			32
303 
304 #define IFLIB_RESTART_BUDGET		8
305 
306 
307 #define CSUM_OFFLOAD		(CSUM_IP_TSO|CSUM_IP6_TSO|CSUM_IP| \
308 				 CSUM_IP_UDP|CSUM_IP_TCP|CSUM_IP_SCTP| \
309 				 CSUM_IP6_UDP|CSUM_IP6_TCP|CSUM_IP6_SCTP)
310 struct iflib_txq {
311 	qidx_t		ift_in_use;
312 	qidx_t		ift_cidx;
313 	qidx_t		ift_cidx_processed;
314 	qidx_t		ift_pidx;
315 	uint8_t		ift_gen;
316 	uint8_t		ift_br_offset;
317 	uint16_t	ift_npending;
318 	uint16_t	ift_db_pending;
319 	uint16_t	ift_rs_pending;
320 	/* implicit pad */
321 	uint8_t		ift_txd_size[8];
322 	uint64_t	ift_processed;
323 	uint64_t	ift_cleaned;
324 	uint64_t	ift_cleaned_prev;
325 #if MEMORY_LOGGING
326 	uint64_t	ift_enqueued;
327 	uint64_t	ift_dequeued;
328 #endif
329 	uint64_t	ift_no_tx_dma_setup;
330 	uint64_t	ift_no_desc_avail;
331 	uint64_t	ift_mbuf_defrag_failed;
332 	uint64_t	ift_mbuf_defrag;
333 	uint64_t	ift_map_failed;
334 	uint64_t	ift_txd_encap_efbig;
335 	uint64_t	ift_pullups;
336 	uint64_t	ift_last_timer_tick;
337 
338 	struct mtx	ift_mtx;
339 	struct mtx	ift_db_mtx;
340 
341 	/* constant values */
342 	if_ctx_t	ift_ctx;
343 	struct ifmp_ring        *ift_br;
344 	struct grouptask	ift_task;
345 	qidx_t		ift_size;
346 	uint16_t	ift_id;
347 	struct callout	ift_timer;
348 
349 	if_txsd_vec_t	ift_sds;
350 	uint8_t		ift_qstatus;
351 	uint8_t		ift_closed;
352 	uint8_t		ift_update_freq;
353 	struct iflib_filter_info ift_filter_info;
354 	bus_dma_tag_t	ift_buf_tag;
355 	bus_dma_tag_t	ift_tso_buf_tag;
356 	iflib_dma_info_t	ift_ifdi;
357 #define MTX_NAME_LEN 16
358 	char                    ift_mtx_name[MTX_NAME_LEN];
359 	char                    ift_db_mtx_name[MTX_NAME_LEN];
360 	bus_dma_segment_t	ift_segs[IFLIB_MAX_TX_SEGS]  __aligned(CACHE_LINE_SIZE);
361 #ifdef IFLIB_DIAGNOSTICS
362 	uint64_t ift_cpu_exec_count[256];
363 #endif
364 } __aligned(CACHE_LINE_SIZE);
365 
366 struct iflib_fl {
367 	qidx_t		ifl_cidx;
368 	qidx_t		ifl_pidx;
369 	qidx_t		ifl_credits;
370 	uint8_t		ifl_gen;
371 	uint8_t		ifl_rxd_size;
372 #if MEMORY_LOGGING
373 	uint64_t	ifl_m_enqueued;
374 	uint64_t	ifl_m_dequeued;
375 	uint64_t	ifl_cl_enqueued;
376 	uint64_t	ifl_cl_dequeued;
377 #endif
378 	/* implicit pad */
379 
380 	bitstr_t 	*ifl_rx_bitmap;
381 	qidx_t		ifl_fragidx;
382 	/* constant */
383 	qidx_t		ifl_size;
384 	uint16_t	ifl_buf_size;
385 	uint16_t	ifl_cltype;
386 	uma_zone_t	ifl_zone;
387 	iflib_rxsd_array_t	ifl_sds;
388 	iflib_rxq_t	ifl_rxq;
389 	uint8_t		ifl_id;
390 	bus_dma_tag_t	ifl_buf_tag;
391 	iflib_dma_info_t	ifl_ifdi;
392 	uint64_t	ifl_bus_addrs[IFLIB_MAX_RX_REFRESH] __aligned(CACHE_LINE_SIZE);
393 	caddr_t		ifl_vm_addrs[IFLIB_MAX_RX_REFRESH];
394 	qidx_t	ifl_rxd_idxs[IFLIB_MAX_RX_REFRESH];
395 }  __aligned(CACHE_LINE_SIZE);
396 
397 static inline qidx_t
398 get_inuse(int size, qidx_t cidx, qidx_t pidx, uint8_t gen)
399 {
400 	qidx_t used;
401 
402 	if (pidx > cidx)
403 		used = pidx - cidx;
404 	else if (pidx < cidx)
405 		used = size - cidx + pidx;
406 	else if (gen == 0 && pidx == cidx)
407 		used = 0;
408 	else if (gen == 1 && pidx == cidx)
409 		used = size;
410 	else
411 		panic("bad state");
412 
413 	return (used);
414 }
415 
416 #define TXQ_AVAIL(txq) (txq->ift_size - get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx, txq->ift_gen))
417 
418 #define IDXDIFF(head, tail, wrap) \
419 	((head) >= (tail) ? (head) - (tail) : (wrap) - (tail) + (head))
420 
421 struct iflib_rxq {
422 	/* If there is a separate completion queue -
423 	 * these are the cq cidx and pidx. Otherwise
424 	 * these are unused.
425 	 */
426 	qidx_t		ifr_size;
427 	qidx_t		ifr_cq_cidx;
428 	qidx_t		ifr_cq_pidx;
429 	uint8_t		ifr_cq_gen;
430 	uint8_t		ifr_fl_offset;
431 
432 	if_ctx_t	ifr_ctx;
433 	iflib_fl_t	ifr_fl;
434 	uint64_t	ifr_rx_irq;
435 	uint16_t	ifr_id;
436 	uint8_t		ifr_lro_enabled;
437 	uint8_t		ifr_nfl;
438 	uint8_t		ifr_ntxqirq;
439 	uint8_t		ifr_txqid[IFLIB_MAX_TX_SHARED_INTR];
440 	struct lro_ctrl			ifr_lc;
441 	struct grouptask        ifr_task;
442 	struct iflib_filter_info ifr_filter_info;
443 	iflib_dma_info_t		ifr_ifdi;
444 
445 	/* dynamically allocate if any drivers need a value substantially larger than this */
446 	struct if_rxd_frag	ifr_frags[IFLIB_MAX_RX_SEGS] __aligned(CACHE_LINE_SIZE);
447 #ifdef IFLIB_DIAGNOSTICS
448 	uint64_t ifr_cpu_exec_count[256];
449 #endif
450 }  __aligned(CACHE_LINE_SIZE);
451 
452 typedef struct if_rxsd {
453 	caddr_t *ifsd_cl;
454 	struct mbuf **ifsd_m;
455 	iflib_fl_t ifsd_fl;
456 	qidx_t ifsd_cidx;
457 } *if_rxsd_t;
458 
459 /* multiple of word size */
460 #ifdef __LP64__
461 #define PKT_INFO_SIZE	6
462 #define RXD_INFO_SIZE	5
463 #define PKT_TYPE uint64_t
464 #else
465 #define PKT_INFO_SIZE	11
466 #define RXD_INFO_SIZE	8
467 #define PKT_TYPE uint32_t
468 #endif
469 #define PKT_LOOP_BOUND  ((PKT_INFO_SIZE/3)*3)
470 #define RXD_LOOP_BOUND  ((RXD_INFO_SIZE/4)*4)
471 
472 typedef struct if_pkt_info_pad {
473 	PKT_TYPE pkt_val[PKT_INFO_SIZE];
474 } *if_pkt_info_pad_t;
475 typedef struct if_rxd_info_pad {
476 	PKT_TYPE rxd_val[RXD_INFO_SIZE];
477 } *if_rxd_info_pad_t;
478 
479 CTASSERT(sizeof(struct if_pkt_info_pad) == sizeof(struct if_pkt_info));
480 CTASSERT(sizeof(struct if_rxd_info_pad) == sizeof(struct if_rxd_info));
481 
482 
483 static inline void
484 pkt_info_zero(if_pkt_info_t pi)
485 {
486 	if_pkt_info_pad_t pi_pad;
487 
488 	pi_pad = (if_pkt_info_pad_t)pi;
489 	pi_pad->pkt_val[0] = 0; pi_pad->pkt_val[1] = 0; pi_pad->pkt_val[2] = 0;
490 	pi_pad->pkt_val[3] = 0; pi_pad->pkt_val[4] = 0; pi_pad->pkt_val[5] = 0;
491 #ifndef __LP64__
492 	pi_pad->pkt_val[6] = 0; pi_pad->pkt_val[7] = 0; pi_pad->pkt_val[8] = 0;
493 	pi_pad->pkt_val[9] = 0; pi_pad->pkt_val[10] = 0;
494 #endif
495 }
496 
497 static device_method_t iflib_pseudo_methods[] = {
498 	DEVMETHOD(device_attach, noop_attach),
499 	DEVMETHOD(device_detach, iflib_pseudo_detach),
500 	DEVMETHOD_END
501 };
502 
503 driver_t iflib_pseudodriver = {
504 	"iflib_pseudo", iflib_pseudo_methods, sizeof(struct iflib_ctx),
505 };
506 
507 static inline void
508 rxd_info_zero(if_rxd_info_t ri)
509 {
510 	if_rxd_info_pad_t ri_pad;
511 	int i;
512 
513 	ri_pad = (if_rxd_info_pad_t)ri;
514 	for (i = 0; i < RXD_LOOP_BOUND; i += 4) {
515 		ri_pad->rxd_val[i] = 0;
516 		ri_pad->rxd_val[i+1] = 0;
517 		ri_pad->rxd_val[i+2] = 0;
518 		ri_pad->rxd_val[i+3] = 0;
519 	}
520 #ifdef __LP64__
521 	ri_pad->rxd_val[RXD_INFO_SIZE-1] = 0;
522 #endif
523 }
524 
525 /*
526  * Only allow a single packet to take up most 1/nth of the tx ring
527  */
528 #define MAX_SINGLE_PACKET_FRACTION 12
529 #define IF_BAD_DMA (bus_addr_t)-1
530 
531 #define CTX_ACTIVE(ctx) ((if_getdrvflags((ctx)->ifc_ifp) & IFF_DRV_RUNNING))
532 
533 #define CTX_LOCK_INIT(_sc)  sx_init(&(_sc)->ifc_ctx_sx, "iflib ctx lock")
534 #define CTX_LOCK(ctx) sx_xlock(&(ctx)->ifc_ctx_sx)
535 #define CTX_UNLOCK(ctx) sx_xunlock(&(ctx)->ifc_ctx_sx)
536 #define CTX_LOCK_DESTROY(ctx) sx_destroy(&(ctx)->ifc_ctx_sx)
537 
538 
539 #define STATE_LOCK_INIT(_sc, _name)  mtx_init(&(_sc)->ifc_state_mtx, _name, "iflib state lock", MTX_DEF)
540 #define STATE_LOCK(ctx) mtx_lock(&(ctx)->ifc_state_mtx)
541 #define STATE_UNLOCK(ctx) mtx_unlock(&(ctx)->ifc_state_mtx)
542 #define STATE_LOCK_DESTROY(ctx) mtx_destroy(&(ctx)->ifc_state_mtx)
543 
544 
545 
546 #define CALLOUT_LOCK(txq)	mtx_lock(&txq->ift_mtx)
547 #define CALLOUT_UNLOCK(txq) 	mtx_unlock(&txq->ift_mtx)
548 
549 void
550 iflib_set_detach(if_ctx_t ctx)
551 {
552 	STATE_LOCK(ctx);
553 	ctx->ifc_flags |= IFC_IN_DETACH;
554 	STATE_UNLOCK(ctx);
555 }
556 
557 /* Our boot-time initialization hook */
558 static int	iflib_module_event_handler(module_t, int, void *);
559 
560 static moduledata_t iflib_moduledata = {
561 	"iflib",
562 	iflib_module_event_handler,
563 	NULL
564 };
565 
566 DECLARE_MODULE(iflib, iflib_moduledata, SI_SUB_INIT_IF, SI_ORDER_ANY);
567 MODULE_VERSION(iflib, 1);
568 
569 MODULE_DEPEND(iflib, pci, 1, 1, 1);
570 MODULE_DEPEND(iflib, ether, 1, 1, 1);
571 
572 TASKQGROUP_DEFINE(if_io_tqg, mp_ncpus, 1);
573 TASKQGROUP_DEFINE(if_config_tqg, 1, 1);
574 
575 #ifndef IFLIB_DEBUG_COUNTERS
576 #ifdef INVARIANTS
577 #define IFLIB_DEBUG_COUNTERS 1
578 #else
579 #define IFLIB_DEBUG_COUNTERS 0
580 #endif /* !INVARIANTS */
581 #endif
582 
583 static SYSCTL_NODE(_net, OID_AUTO, iflib, CTLFLAG_RD, 0,
584                    "iflib driver parameters");
585 
586 /*
587  * XXX need to ensure that this can't accidentally cause the head to be moved backwards
588  */
589 static int iflib_min_tx_latency = 0;
590 SYSCTL_INT(_net_iflib, OID_AUTO, min_tx_latency, CTLFLAG_RW,
591 		   &iflib_min_tx_latency, 0, "minimize transmit latency at the possible expense of throughput");
592 static int iflib_no_tx_batch = 0;
593 SYSCTL_INT(_net_iflib, OID_AUTO, no_tx_batch, CTLFLAG_RW,
594 		   &iflib_no_tx_batch, 0, "minimize transmit latency at the possible expense of throughput");
595 
596 
597 #if IFLIB_DEBUG_COUNTERS
598 
599 static int iflib_tx_seen;
600 static int iflib_tx_sent;
601 static int iflib_tx_encap;
602 static int iflib_rx_allocs;
603 static int iflib_fl_refills;
604 static int iflib_fl_refills_large;
605 static int iflib_tx_frees;
606 
607 SYSCTL_INT(_net_iflib, OID_AUTO, tx_seen, CTLFLAG_RD,
608 		   &iflib_tx_seen, 0, "# tx mbufs seen");
609 SYSCTL_INT(_net_iflib, OID_AUTO, tx_sent, CTLFLAG_RD,
610 		   &iflib_tx_sent, 0, "# tx mbufs sent");
611 SYSCTL_INT(_net_iflib, OID_AUTO, tx_encap, CTLFLAG_RD,
612 		   &iflib_tx_encap, 0, "# tx mbufs encapped");
613 SYSCTL_INT(_net_iflib, OID_AUTO, tx_frees, CTLFLAG_RD,
614 		   &iflib_tx_frees, 0, "# tx frees");
615 SYSCTL_INT(_net_iflib, OID_AUTO, rx_allocs, CTLFLAG_RD,
616 		   &iflib_rx_allocs, 0, "# rx allocations");
617 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills, CTLFLAG_RD,
618 		   &iflib_fl_refills, 0, "# refills");
619 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills_large, CTLFLAG_RD,
620 		   &iflib_fl_refills_large, 0, "# large refills");
621 
622 
623 static int iflib_txq_drain_flushing;
624 static int iflib_txq_drain_oactive;
625 static int iflib_txq_drain_notready;
626 
627 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_flushing, CTLFLAG_RD,
628 		   &iflib_txq_drain_flushing, 0, "# drain flushes");
629 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_oactive, CTLFLAG_RD,
630 		   &iflib_txq_drain_oactive, 0, "# drain oactives");
631 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_notready, CTLFLAG_RD,
632 		   &iflib_txq_drain_notready, 0, "# drain notready");
633 
634 
635 static int iflib_encap_load_mbuf_fail;
636 static int iflib_encap_pad_mbuf_fail;
637 static int iflib_encap_txq_avail_fail;
638 static int iflib_encap_txd_encap_fail;
639 
640 SYSCTL_INT(_net_iflib, OID_AUTO, encap_load_mbuf_fail, CTLFLAG_RD,
641 		   &iflib_encap_load_mbuf_fail, 0, "# busdma load failures");
642 SYSCTL_INT(_net_iflib, OID_AUTO, encap_pad_mbuf_fail, CTLFLAG_RD,
643 		   &iflib_encap_pad_mbuf_fail, 0, "# runt frame pad failures");
644 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txq_avail_fail, CTLFLAG_RD,
645 		   &iflib_encap_txq_avail_fail, 0, "# txq avail failures");
646 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txd_encap_fail, CTLFLAG_RD,
647 		   &iflib_encap_txd_encap_fail, 0, "# driver encap failures");
648 
649 static int iflib_task_fn_rxs;
650 static int iflib_rx_intr_enables;
651 static int iflib_fast_intrs;
652 static int iflib_rx_unavail;
653 static int iflib_rx_ctx_inactive;
654 static int iflib_rx_if_input;
655 static int iflib_rx_mbuf_null;
656 static int iflib_rxd_flush;
657 
658 static int iflib_verbose_debug;
659 
660 SYSCTL_INT(_net_iflib, OID_AUTO, task_fn_rx, CTLFLAG_RD,
661 		   &iflib_task_fn_rxs, 0, "# task_fn_rx calls");
662 SYSCTL_INT(_net_iflib, OID_AUTO, rx_intr_enables, CTLFLAG_RD,
663 		   &iflib_rx_intr_enables, 0, "# rx intr enables");
664 SYSCTL_INT(_net_iflib, OID_AUTO, fast_intrs, CTLFLAG_RD,
665 		   &iflib_fast_intrs, 0, "# fast_intr calls");
666 SYSCTL_INT(_net_iflib, OID_AUTO, rx_unavail, CTLFLAG_RD,
667 		   &iflib_rx_unavail, 0, "# times rxeof called with no available data");
668 SYSCTL_INT(_net_iflib, OID_AUTO, rx_ctx_inactive, CTLFLAG_RD,
669 		   &iflib_rx_ctx_inactive, 0, "# times rxeof called with inactive context");
670 SYSCTL_INT(_net_iflib, OID_AUTO, rx_if_input, CTLFLAG_RD,
671 		   &iflib_rx_if_input, 0, "# times rxeof called if_input");
672 SYSCTL_INT(_net_iflib, OID_AUTO, rx_mbuf_null, CTLFLAG_RD,
673 		   &iflib_rx_mbuf_null, 0, "# times rxeof got null mbuf");
674 SYSCTL_INT(_net_iflib, OID_AUTO, rxd_flush, CTLFLAG_RD,
675 	         &iflib_rxd_flush, 0, "# times rxd_flush called");
676 SYSCTL_INT(_net_iflib, OID_AUTO, verbose_debug, CTLFLAG_RW,
677 		   &iflib_verbose_debug, 0, "enable verbose debugging");
678 
679 #define DBG_COUNTER_INC(name) atomic_add_int(&(iflib_ ## name), 1)
680 static void
681 iflib_debug_reset(void)
682 {
683 	iflib_tx_seen = iflib_tx_sent = iflib_tx_encap = iflib_rx_allocs =
684 		iflib_fl_refills = iflib_fl_refills_large = iflib_tx_frees =
685 		iflib_txq_drain_flushing = iflib_txq_drain_oactive =
686 		iflib_txq_drain_notready =
687 		iflib_encap_load_mbuf_fail = iflib_encap_pad_mbuf_fail =
688 		iflib_encap_txq_avail_fail = iflib_encap_txd_encap_fail =
689 		iflib_task_fn_rxs = iflib_rx_intr_enables = iflib_fast_intrs =
690 		iflib_rx_unavail =
691 		iflib_rx_ctx_inactive = iflib_rx_if_input =
692 		iflib_rx_mbuf_null = iflib_rxd_flush = 0;
693 }
694 
695 #else
696 #define DBG_COUNTER_INC(name)
697 static void iflib_debug_reset(void) {}
698 #endif
699 
700 #define IFLIB_DEBUG 0
701 
702 static void iflib_tx_structures_free(if_ctx_t ctx);
703 static void iflib_rx_structures_free(if_ctx_t ctx);
704 static int iflib_queues_alloc(if_ctx_t ctx);
705 static int iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq);
706 static int iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget);
707 static int iflib_qset_structures_setup(if_ctx_t ctx);
708 static int iflib_msix_init(if_ctx_t ctx);
709 static int iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filterarg, int *rid, const char *str);
710 static void iflib_txq_check_drain(iflib_txq_t txq, int budget);
711 static uint32_t iflib_txq_can_drain(struct ifmp_ring *);
712 #ifdef ALTQ
713 static void iflib_altq_if_start(if_t ifp);
714 static int iflib_altq_if_transmit(if_t ifp, struct mbuf *m);
715 #endif
716 static int iflib_register(if_ctx_t);
717 static void iflib_init_locked(if_ctx_t ctx);
718 static void iflib_add_device_sysctl_pre(if_ctx_t ctx);
719 static void iflib_add_device_sysctl_post(if_ctx_t ctx);
720 static void iflib_ifmp_purge(iflib_txq_t txq);
721 static void _iflib_pre_assert(if_softc_ctx_t scctx);
722 static void iflib_if_init_locked(if_ctx_t ctx);
723 static void iflib_free_intr_mem(if_ctx_t ctx);
724 #ifndef __NO_STRICT_ALIGNMENT
725 static struct mbuf * iflib_fixup_rx(struct mbuf *m);
726 #endif
727 
728 NETDUMP_DEFINE(iflib);
729 
730 #ifdef DEV_NETMAP
731 #include <sys/selinfo.h>
732 #include <net/netmap.h>
733 #include <dev/netmap/netmap_kern.h>
734 
735 MODULE_DEPEND(iflib, netmap, 1, 1, 1);
736 
737 static int netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, uint32_t nm_i, bool init);
738 
739 /*
740  * device-specific sysctl variables:
741  *
742  * iflib_crcstrip: 0: keep CRC in rx frames (default), 1: strip it.
743  *	During regular operations the CRC is stripped, but on some
744  *	hardware reception of frames not multiple of 64 is slower,
745  *	so using crcstrip=0 helps in benchmarks.
746  *
747  * iflib_rx_miss, iflib_rx_miss_bufs:
748  *	count packets that might be missed due to lost interrupts.
749  */
750 SYSCTL_DECL(_dev_netmap);
751 /*
752  * The xl driver by default strips CRCs and we do not override it.
753  */
754 
755 int iflib_crcstrip = 1;
756 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_crcstrip,
757     CTLFLAG_RW, &iflib_crcstrip, 1, "strip CRC on rx frames");
758 
759 int iflib_rx_miss, iflib_rx_miss_bufs;
760 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss,
761     CTLFLAG_RW, &iflib_rx_miss, 0, "potentially missed rx intr");
762 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss_bufs,
763     CTLFLAG_RW, &iflib_rx_miss_bufs, 0, "potentially missed rx intr bufs");
764 
765 /*
766  * Register/unregister. We are already under netmap lock.
767  * Only called on the first register or the last unregister.
768  */
769 static int
770 iflib_netmap_register(struct netmap_adapter *na, int onoff)
771 {
772 	struct ifnet *ifp = na->ifp;
773 	if_ctx_t ctx = ifp->if_softc;
774 	int status;
775 
776 	CTX_LOCK(ctx);
777 	IFDI_INTR_DISABLE(ctx);
778 
779 	/* Tell the stack that the interface is no longer active */
780 	ifp->if_drv_flags &= ~(IFF_DRV_RUNNING | IFF_DRV_OACTIVE);
781 
782 	if (!CTX_IS_VF(ctx))
783 		IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip);
784 
785 	/* enable or disable flags and callbacks in na and ifp */
786 	if (onoff) {
787 		nm_set_native_flags(na);
788 	} else {
789 		nm_clear_native_flags(na);
790 	}
791 	iflib_stop(ctx);
792 	iflib_init_locked(ctx);
793 	IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip); // XXX why twice ?
794 	status = ifp->if_drv_flags & IFF_DRV_RUNNING ? 0 : 1;
795 	if (status)
796 		nm_clear_native_flags(na);
797 	CTX_UNLOCK(ctx);
798 	return (status);
799 }
800 
801 static int
802 netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, uint32_t nm_i, bool init)
803 {
804 	struct netmap_adapter *na = kring->na;
805 	u_int const lim = kring->nkr_num_slots - 1;
806 	u_int head = kring->rhead;
807 	struct netmap_ring *ring = kring->ring;
808 	bus_dmamap_t *map;
809 	struct if_rxd_update iru;
810 	if_ctx_t ctx = rxq->ifr_ctx;
811 	iflib_fl_t fl = &rxq->ifr_fl[0];
812 	uint32_t refill_pidx, nic_i;
813 #if IFLIB_DEBUG_COUNTERS
814 	int rf_count = 0;
815 #endif
816 
817 	if (nm_i == head && __predict_true(!init))
818 		return 0;
819 	iru_init(&iru, rxq, 0 /* flid */);
820 	map = fl->ifl_sds.ifsd_map;
821 	refill_pidx = netmap_idx_k2n(kring, nm_i);
822 	/*
823 	 * IMPORTANT: we must leave one free slot in the ring,
824 	 * so move head back by one unit
825 	 */
826 	head = nm_prev(head, lim);
827 	nic_i = UINT_MAX;
828 	DBG_COUNTER_INC(fl_refills);
829 	while (nm_i != head) {
830 #if IFLIB_DEBUG_COUNTERS
831 		if (++rf_count == 9)
832 			DBG_COUNTER_INC(fl_refills_large);
833 #endif
834 		for (int tmp_pidx = 0; tmp_pidx < IFLIB_MAX_RX_REFRESH && nm_i != head; tmp_pidx++) {
835 			struct netmap_slot *slot = &ring->slot[nm_i];
836 			void *addr = PNMB(na, slot, &fl->ifl_bus_addrs[tmp_pidx]);
837 			uint32_t nic_i_dma = refill_pidx;
838 			nic_i = netmap_idx_k2n(kring, nm_i);
839 
840 			MPASS(tmp_pidx < IFLIB_MAX_RX_REFRESH);
841 
842 			if (addr == NETMAP_BUF_BASE(na)) /* bad buf */
843 			        return netmap_ring_reinit(kring);
844 
845 			fl->ifl_vm_addrs[tmp_pidx] = addr;
846 			if (__predict_false(init)) {
847 				netmap_load_map(na, fl->ifl_buf_tag,
848 				    map[nic_i], addr);
849 			} else if (slot->flags & NS_BUF_CHANGED) {
850 				/* buffer has changed, reload map */
851 				netmap_reload_map(na, fl->ifl_buf_tag,
852 				    map[nic_i], addr);
853 			}
854 			slot->flags &= ~NS_BUF_CHANGED;
855 
856 			nm_i = nm_next(nm_i, lim);
857 			fl->ifl_rxd_idxs[tmp_pidx] = nic_i = nm_next(nic_i, lim);
858 			if (nm_i != head && tmp_pidx < IFLIB_MAX_RX_REFRESH-1)
859 				continue;
860 
861 			iru.iru_pidx = refill_pidx;
862 			iru.iru_count = tmp_pidx+1;
863 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
864 			refill_pidx = nic_i;
865 			for (int n = 0; n < iru.iru_count; n++) {
866 				bus_dmamap_sync(fl->ifl_buf_tag, map[nic_i_dma],
867 						BUS_DMASYNC_PREREAD);
868 				/* XXX - change this to not use the netmap func*/
869 				nic_i_dma = nm_next(nic_i_dma, lim);
870 			}
871 		}
872 	}
873 	kring->nr_hwcur = head;
874 
875 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
876 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
877 	if (__predict_true(nic_i != UINT_MAX)) {
878 		ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id, nic_i);
879 		DBG_COUNTER_INC(rxd_flush);
880 	}
881 	return (0);
882 }
883 
884 /*
885  * Reconcile kernel and user view of the transmit ring.
886  *
887  * All information is in the kring.
888  * Userspace wants to send packets up to the one before kring->rhead,
889  * kernel knows kring->nr_hwcur is the first unsent packet.
890  *
891  * Here we push packets out (as many as possible), and possibly
892  * reclaim buffers from previously completed transmission.
893  *
894  * The caller (netmap) guarantees that there is only one instance
895  * running at any time. Any interference with other driver
896  * methods should be handled by the individual drivers.
897  */
898 static int
899 iflib_netmap_txsync(struct netmap_kring *kring, int flags)
900 {
901 	struct netmap_adapter *na = kring->na;
902 	struct ifnet *ifp = na->ifp;
903 	struct netmap_ring *ring = kring->ring;
904 	u_int nm_i;	/* index into the netmap kring */
905 	u_int nic_i;	/* index into the NIC ring */
906 	u_int n;
907 	u_int const lim = kring->nkr_num_slots - 1;
908 	u_int const head = kring->rhead;
909 	struct if_pkt_info pi;
910 
911 	/*
912 	 * interrupts on every tx packet are expensive so request
913 	 * them every half ring, or where NS_REPORT is set
914 	 */
915 	u_int report_frequency = kring->nkr_num_slots >> 1;
916 	/* device-specific */
917 	if_ctx_t ctx = ifp->if_softc;
918 	iflib_txq_t txq = &ctx->ifc_txqs[kring->ring_id];
919 
920 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
921 	    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
922 
923 	/*
924 	 * First part: process new packets to send.
925 	 * nm_i is the current index in the netmap kring,
926 	 * nic_i is the corresponding index in the NIC ring.
927 	 *
928 	 * If we have packets to send (nm_i != head)
929 	 * iterate over the netmap ring, fetch length and update
930 	 * the corresponding slot in the NIC ring. Some drivers also
931 	 * need to update the buffer's physical address in the NIC slot
932 	 * even NS_BUF_CHANGED is not set (PNMB computes the addresses).
933 	 *
934 	 * The netmap_reload_map() calls is especially expensive,
935 	 * even when (as in this case) the tag is 0, so do only
936 	 * when the buffer has actually changed.
937 	 *
938 	 * If possible do not set the report/intr bit on all slots,
939 	 * but only a few times per ring or when NS_REPORT is set.
940 	 *
941 	 * Finally, on 10G and faster drivers, it might be useful
942 	 * to prefetch the next slot and txr entry.
943 	 */
944 
945 	nm_i = kring->nr_hwcur;
946 	if (nm_i != head) {	/* we have new packets to send */
947 		pkt_info_zero(&pi);
948 		pi.ipi_segs = txq->ift_segs;
949 		pi.ipi_qsidx = kring->ring_id;
950 		nic_i = netmap_idx_k2n(kring, nm_i);
951 
952 		__builtin_prefetch(&ring->slot[nm_i]);
953 		__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i]);
954 		__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i]);
955 
956 		for (n = 0; nm_i != head; n++) {
957 			struct netmap_slot *slot = &ring->slot[nm_i];
958 			u_int len = slot->len;
959 			uint64_t paddr;
960 			void *addr = PNMB(na, slot, &paddr);
961 			int flags = (slot->flags & NS_REPORT ||
962 				nic_i == 0 || nic_i == report_frequency) ?
963 				IPI_TX_INTR : 0;
964 
965 			/* device-specific */
966 			pi.ipi_len = len;
967 			pi.ipi_segs[0].ds_addr = paddr;
968 			pi.ipi_segs[0].ds_len = len;
969 			pi.ipi_nsegs = 1;
970 			pi.ipi_ndescs = 0;
971 			pi.ipi_pidx = nic_i;
972 			pi.ipi_flags = flags;
973 
974 			/* Fill the slot in the NIC ring. */
975 			ctx->isc_txd_encap(ctx->ifc_softc, &pi);
976 			DBG_COUNTER_INC(tx_encap);
977 
978 			/* prefetch for next round */
979 			__builtin_prefetch(&ring->slot[nm_i + 1]);
980 			__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i + 1]);
981 			__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i + 1]);
982 
983 			NM_CHECK_ADDR_LEN(na, addr, len);
984 
985 			if (slot->flags & NS_BUF_CHANGED) {
986 				/* buffer has changed, reload map */
987 				netmap_reload_map(na, txq->ift_buf_tag,
988 				    txq->ift_sds.ifsd_map[nic_i], addr);
989 			}
990 			/* make sure changes to the buffer are synced */
991 			bus_dmamap_sync(txq->ift_buf_tag,
992 			    txq->ift_sds.ifsd_map[nic_i],
993 			    BUS_DMASYNC_PREWRITE);
994 
995 			slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED);
996 			nm_i = nm_next(nm_i, lim);
997 			nic_i = nm_next(nic_i, lim);
998 		}
999 		kring->nr_hwcur = nm_i;
1000 
1001 		/* synchronize the NIC ring */
1002 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1003 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1004 
1005 		/* (re)start the tx unit up to slot nic_i (excluded) */
1006 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, nic_i);
1007 	}
1008 
1009 	/*
1010 	 * Second part: reclaim buffers for completed transmissions.
1011 	 *
1012 	 * If there are unclaimed buffers, attempt to reclaim them.
1013 	 * If none are reclaimed, and TX IRQs are not in use, do an initial
1014 	 * minimal delay, then trigger the tx handler which will spin in the
1015 	 * group task queue.
1016 	 */
1017 	if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1018 		if (iflib_tx_credits_update(ctx, txq)) {
1019 			/* some tx completed, increment avail */
1020 			nic_i = txq->ift_cidx_processed;
1021 			kring->nr_hwtail = nm_prev(netmap_idx_n2k(kring, nic_i), lim);
1022 		}
1023 	}
1024 	if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ))
1025 		if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1026 			callout_reset_on(&txq->ift_timer, hz < 2000 ? 1 : hz / 1000,
1027 			    iflib_timer, txq, txq->ift_timer.c_cpu);
1028 	}
1029 	return (0);
1030 }
1031 
1032 /*
1033  * Reconcile kernel and user view of the receive ring.
1034  * Same as for the txsync, this routine must be efficient.
1035  * The caller guarantees a single invocations, but races against
1036  * the rest of the driver should be handled here.
1037  *
1038  * On call, kring->rhead is the first packet that userspace wants
1039  * to keep, and kring->rcur is the wakeup point.
1040  * The kernel has previously reported packets up to kring->rtail.
1041  *
1042  * If (flags & NAF_FORCE_READ) also check for incoming packets irrespective
1043  * of whether or not we received an interrupt.
1044  */
1045 static int
1046 iflib_netmap_rxsync(struct netmap_kring *kring, int flags)
1047 {
1048 	struct netmap_adapter *na = kring->na;
1049 	struct netmap_ring *ring = kring->ring;
1050 	iflib_fl_t fl;
1051 	uint32_t nm_i;	/* index into the netmap ring */
1052 	uint32_t nic_i;	/* index into the NIC ring */
1053 	u_int i, n;
1054 	u_int const lim = kring->nkr_num_slots - 1;
1055 	u_int const head = kring->rhead;
1056 	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
1057 	struct if_rxd_info ri;
1058 
1059 	struct ifnet *ifp = na->ifp;
1060 	if_ctx_t ctx = ifp->if_softc;
1061 	iflib_rxq_t rxq = &ctx->ifc_rxqs[kring->ring_id];
1062 	if (head > lim)
1063 		return netmap_ring_reinit(kring);
1064 
1065 	/*
1066 	 * XXX netmap_fl_refill() only ever (re)fills free list 0 so far.
1067 	 */
1068 
1069 	for (i = 0, fl = rxq->ifr_fl; i < rxq->ifr_nfl; i++, fl++) {
1070 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
1071 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1072 	}
1073 
1074 	/*
1075 	 * First part: import newly received packets.
1076 	 *
1077 	 * nm_i is the index of the next free slot in the netmap ring,
1078 	 * nic_i is the index of the next received packet in the NIC ring,
1079 	 * and they may differ in case if_init() has been called while
1080 	 * in netmap mode. For the receive ring we have
1081 	 *
1082 	 *	nic_i = rxr->next_check;
1083 	 *	nm_i = kring->nr_hwtail (previous)
1084 	 * and
1085 	 *	nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1086 	 *
1087 	 * rxr->next_check is set to 0 on a ring reinit
1088 	 */
1089 	if (netmap_no_pendintr || force_update) {
1090 		int crclen = iflib_crcstrip ? 0 : 4;
1091 		int error, avail;
1092 
1093 		for (i = 0; i < rxq->ifr_nfl; i++) {
1094 			fl = &rxq->ifr_fl[i];
1095 			nic_i = fl->ifl_cidx;
1096 			nm_i = netmap_idx_n2k(kring, nic_i);
1097 			avail = ctx->isc_rxd_available(ctx->ifc_softc,
1098 			    rxq->ifr_id, nic_i, USHRT_MAX);
1099 			for (n = 0; avail > 0; n++, avail--) {
1100 				rxd_info_zero(&ri);
1101 				ri.iri_frags = rxq->ifr_frags;
1102 				ri.iri_qsidx = kring->ring_id;
1103 				ri.iri_ifp = ctx->ifc_ifp;
1104 				ri.iri_cidx = nic_i;
1105 
1106 				error = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
1107 				ring->slot[nm_i].len = error ? 0 : ri.iri_len - crclen;
1108 				ring->slot[nm_i].flags = 0;
1109 				bus_dmamap_sync(fl->ifl_buf_tag,
1110 				    fl->ifl_sds.ifsd_map[nic_i], BUS_DMASYNC_POSTREAD);
1111 				nm_i = nm_next(nm_i, lim);
1112 				nic_i = nm_next(nic_i, lim);
1113 			}
1114 			if (n) { /* update the state variables */
1115 				if (netmap_no_pendintr && !force_update) {
1116 					/* diagnostics */
1117 					iflib_rx_miss ++;
1118 					iflib_rx_miss_bufs += n;
1119 				}
1120 				fl->ifl_cidx = nic_i;
1121 				kring->nr_hwtail = nm_i;
1122 			}
1123 			kring->nr_kflags &= ~NKR_PENDINTR;
1124 		}
1125 	}
1126 	/*
1127 	 * Second part: skip past packets that userspace has released.
1128 	 * (kring->nr_hwcur to head excluded),
1129 	 * and make the buffers available for reception.
1130 	 * As usual nm_i is the index in the netmap ring,
1131 	 * nic_i is the index in the NIC ring, and
1132 	 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1133 	 */
1134 	/* XXX not sure how this will work with multiple free lists */
1135 	nm_i = kring->nr_hwcur;
1136 
1137 	return (netmap_fl_refill(rxq, kring, nm_i, false));
1138 }
1139 
1140 static void
1141 iflib_netmap_intr(struct netmap_adapter *na, int onoff)
1142 {
1143 	struct ifnet *ifp = na->ifp;
1144 	if_ctx_t ctx = ifp->if_softc;
1145 
1146 	CTX_LOCK(ctx);
1147 	if (onoff) {
1148 		IFDI_INTR_ENABLE(ctx);
1149 	} else {
1150 		IFDI_INTR_DISABLE(ctx);
1151 	}
1152 	CTX_UNLOCK(ctx);
1153 }
1154 
1155 
1156 static int
1157 iflib_netmap_attach(if_ctx_t ctx)
1158 {
1159 	struct netmap_adapter na;
1160 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1161 
1162 	bzero(&na, sizeof(na));
1163 
1164 	na.ifp = ctx->ifc_ifp;
1165 	na.na_flags = NAF_BDG_MAYSLEEP;
1166 	MPASS(ctx->ifc_softc_ctx.isc_ntxqsets);
1167 	MPASS(ctx->ifc_softc_ctx.isc_nrxqsets);
1168 
1169 	na.num_tx_desc = scctx->isc_ntxd[0];
1170 	na.num_rx_desc = scctx->isc_nrxd[0];
1171 	na.nm_txsync = iflib_netmap_txsync;
1172 	na.nm_rxsync = iflib_netmap_rxsync;
1173 	na.nm_register = iflib_netmap_register;
1174 	na.nm_intr = iflib_netmap_intr;
1175 	na.num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
1176 	na.num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
1177 	return (netmap_attach(&na));
1178 }
1179 
1180 static void
1181 iflib_netmap_txq_init(if_ctx_t ctx, iflib_txq_t txq)
1182 {
1183 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1184 	struct netmap_slot *slot;
1185 
1186 	slot = netmap_reset(na, NR_TX, txq->ift_id, 0);
1187 	if (slot == NULL)
1188 		return;
1189 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxd[0]; i++) {
1190 
1191 		/*
1192 		 * In netmap mode, set the map for the packet buffer.
1193 		 * NOTE: Some drivers (not this one) also need to set
1194 		 * the physical buffer address in the NIC ring.
1195 		 * netmap_idx_n2k() maps a nic index, i, into the corresponding
1196 		 * netmap slot index, si
1197 		 */
1198 		int si = netmap_idx_n2k(na->tx_rings[txq->ift_id], i);
1199 		netmap_load_map(na, txq->ift_buf_tag, txq->ift_sds.ifsd_map[i],
1200 		    NMB(na, slot + si));
1201 	}
1202 }
1203 
1204 static void
1205 iflib_netmap_rxq_init(if_ctx_t ctx, iflib_rxq_t rxq)
1206 {
1207 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1208 	struct netmap_kring *kring = na->rx_rings[rxq->ifr_id];
1209 	struct netmap_slot *slot;
1210 	uint32_t nm_i;
1211 
1212 	slot = netmap_reset(na, NR_RX, rxq->ifr_id, 0);
1213 	if (slot == NULL)
1214 		return;
1215 	nm_i = netmap_idx_n2k(kring, 0);
1216 	netmap_fl_refill(rxq, kring, nm_i, true);
1217 }
1218 
1219 static void
1220 iflib_netmap_timer_adjust(if_ctx_t ctx, iflib_txq_t txq, uint32_t *reset_on)
1221 {
1222 	struct netmap_kring *kring;
1223 	uint16_t txqid;
1224 
1225 	txqid = txq->ift_id;
1226 	kring = NA(ctx->ifc_ifp)->tx_rings[txqid];
1227 
1228 	if (kring->nr_hwcur != nm_next(kring->nr_hwtail, kring->nkr_num_slots - 1)) {
1229 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1230 		    BUS_DMASYNC_POSTREAD);
1231 		if (ctx->isc_txd_credits_update(ctx->ifc_softc, txqid, false))
1232 			netmap_tx_irq(ctx->ifc_ifp, txqid);
1233 		if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ)) {
1234 			if (hz < 2000)
1235 				*reset_on = 1;
1236 			else
1237 				*reset_on = hz / 1000;
1238 		}
1239 	}
1240 }
1241 
1242 #define iflib_netmap_detach(ifp) netmap_detach(ifp)
1243 
1244 #else
1245 #define iflib_netmap_txq_init(ctx, txq)
1246 #define iflib_netmap_rxq_init(ctx, rxq)
1247 #define iflib_netmap_detach(ifp)
1248 
1249 #define iflib_netmap_attach(ctx) (0)
1250 #define netmap_rx_irq(ifp, qid, budget) (0)
1251 #define netmap_tx_irq(ifp, qid) do {} while (0)
1252 #define iflib_netmap_timer_adjust(ctx, txq, reset_on)
1253 
1254 #endif
1255 
1256 #if defined(__i386__) || defined(__amd64__)
1257 static __inline void
1258 prefetch(void *x)
1259 {
1260 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1261 }
1262 static __inline void
1263 prefetch2cachelines(void *x)
1264 {
1265 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1266 #if (CACHE_LINE_SIZE < 128)
1267 	__asm volatile("prefetcht0 %0" :: "m" (*(((unsigned long *)x)+CACHE_LINE_SIZE/(sizeof(unsigned long)))));
1268 #endif
1269 }
1270 #else
1271 #define prefetch(x)
1272 #define prefetch2cachelines(x)
1273 #endif
1274 
1275 static void
1276 iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid)
1277 {
1278 	iflib_fl_t fl;
1279 
1280 	fl = &rxq->ifr_fl[flid];
1281 	iru->iru_paddrs = fl->ifl_bus_addrs;
1282 	iru->iru_vaddrs = &fl->ifl_vm_addrs[0];
1283 	iru->iru_idxs = fl->ifl_rxd_idxs;
1284 	iru->iru_qsidx = rxq->ifr_id;
1285 	iru->iru_buf_size = fl->ifl_buf_size;
1286 	iru->iru_flidx = fl->ifl_id;
1287 }
1288 
1289 static void
1290 _iflib_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int err)
1291 {
1292 	if (err)
1293 		return;
1294 	*(bus_addr_t *) arg = segs[0].ds_addr;
1295 }
1296 
1297 int
1298 iflib_dma_alloc_align(if_ctx_t ctx, int size, int align, iflib_dma_info_t dma, int mapflags)
1299 {
1300 	int err;
1301 	device_t dev = ctx->ifc_dev;
1302 
1303 	err = bus_dma_tag_create(bus_get_dma_tag(dev),	/* parent */
1304 				align, 0,		/* alignment, bounds */
1305 				BUS_SPACE_MAXADDR,	/* lowaddr */
1306 				BUS_SPACE_MAXADDR,	/* highaddr */
1307 				NULL, NULL,		/* filter, filterarg */
1308 				size,			/* maxsize */
1309 				1,			/* nsegments */
1310 				size,			/* maxsegsize */
1311 				BUS_DMA_ALLOCNOW,	/* flags */
1312 				NULL,			/* lockfunc */
1313 				NULL,			/* lockarg */
1314 				&dma->idi_tag);
1315 	if (err) {
1316 		device_printf(dev,
1317 		    "%s: bus_dma_tag_create failed: %d\n",
1318 		    __func__, err);
1319 		goto fail_0;
1320 	}
1321 
1322 	err = bus_dmamem_alloc(dma->idi_tag, (void**) &dma->idi_vaddr,
1323 	    BUS_DMA_NOWAIT | BUS_DMA_COHERENT | BUS_DMA_ZERO, &dma->idi_map);
1324 	if (err) {
1325 		device_printf(dev,
1326 		    "%s: bus_dmamem_alloc(%ju) failed: %d\n",
1327 		    __func__, (uintmax_t)size, err);
1328 		goto fail_1;
1329 	}
1330 
1331 	dma->idi_paddr = IF_BAD_DMA;
1332 	err = bus_dmamap_load(dma->idi_tag, dma->idi_map, dma->idi_vaddr,
1333 	    size, _iflib_dmamap_cb, &dma->idi_paddr, mapflags | BUS_DMA_NOWAIT);
1334 	if (err || dma->idi_paddr == IF_BAD_DMA) {
1335 		device_printf(dev,
1336 		    "%s: bus_dmamap_load failed: %d\n",
1337 		    __func__, err);
1338 		goto fail_2;
1339 	}
1340 
1341 	dma->idi_size = size;
1342 	return (0);
1343 
1344 fail_2:
1345 	bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1346 fail_1:
1347 	bus_dma_tag_destroy(dma->idi_tag);
1348 fail_0:
1349 	dma->idi_tag = NULL;
1350 
1351 	return (err);
1352 }
1353 
1354 int
1355 iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags)
1356 {
1357 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1358 
1359 	KASSERT(sctx->isc_q_align != 0, ("alignment value not initialized"));
1360 
1361 	return (iflib_dma_alloc_align(ctx, size, sctx->isc_q_align, dma, mapflags));
1362 }
1363 
1364 int
1365 iflib_dma_alloc_multi(if_ctx_t ctx, int *sizes, iflib_dma_info_t *dmalist, int mapflags, int count)
1366 {
1367 	int i, err;
1368 	iflib_dma_info_t *dmaiter;
1369 
1370 	dmaiter = dmalist;
1371 	for (i = 0; i < count; i++, dmaiter++) {
1372 		if ((err = iflib_dma_alloc(ctx, sizes[i], *dmaiter, mapflags)) != 0)
1373 			break;
1374 	}
1375 	if (err)
1376 		iflib_dma_free_multi(dmalist, i);
1377 	return (err);
1378 }
1379 
1380 void
1381 iflib_dma_free(iflib_dma_info_t dma)
1382 {
1383 	if (dma->idi_tag == NULL)
1384 		return;
1385 	if (dma->idi_paddr != IF_BAD_DMA) {
1386 		bus_dmamap_sync(dma->idi_tag, dma->idi_map,
1387 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1388 		bus_dmamap_unload(dma->idi_tag, dma->idi_map);
1389 		dma->idi_paddr = IF_BAD_DMA;
1390 	}
1391 	if (dma->idi_vaddr != NULL) {
1392 		bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1393 		dma->idi_vaddr = NULL;
1394 	}
1395 	bus_dma_tag_destroy(dma->idi_tag);
1396 	dma->idi_tag = NULL;
1397 }
1398 
1399 void
1400 iflib_dma_free_multi(iflib_dma_info_t *dmalist, int count)
1401 {
1402 	int i;
1403 	iflib_dma_info_t *dmaiter = dmalist;
1404 
1405 	for (i = 0; i < count; i++, dmaiter++)
1406 		iflib_dma_free(*dmaiter);
1407 }
1408 
1409 #ifdef EARLY_AP_STARTUP
1410 static const int iflib_started = 1;
1411 #else
1412 /*
1413  * We used to abuse the smp_started flag to decide if the queues have been
1414  * fully initialized (by late taskqgroup_adjust() calls in a SYSINIT()).
1415  * That gave bad races, since the SYSINIT() runs strictly after smp_started
1416  * is set.  Run a SYSINIT() strictly after that to just set a usable
1417  * completion flag.
1418  */
1419 
1420 static int iflib_started;
1421 
1422 static void
1423 iflib_record_started(void *arg)
1424 {
1425 	iflib_started = 1;
1426 }
1427 
1428 SYSINIT(iflib_record_started, SI_SUB_SMP + 1, SI_ORDER_FIRST,
1429 	iflib_record_started, NULL);
1430 #endif
1431 
1432 static int
1433 iflib_fast_intr(void *arg)
1434 {
1435 	iflib_filter_info_t info = arg;
1436 	struct grouptask *gtask = info->ifi_task;
1437 	int result;
1438 
1439 	if (!iflib_started)
1440 		return (FILTER_STRAY);
1441 
1442 	DBG_COUNTER_INC(fast_intrs);
1443 	if (info->ifi_filter != NULL) {
1444 		result = info->ifi_filter(info->ifi_filter_arg);
1445 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1446 			return (result);
1447 	}
1448 
1449 	GROUPTASK_ENQUEUE(gtask);
1450 	return (FILTER_HANDLED);
1451 }
1452 
1453 static int
1454 iflib_fast_intr_rxtx(void *arg)
1455 {
1456 	iflib_filter_info_t info = arg;
1457 	struct grouptask *gtask = info->ifi_task;
1458 	if_ctx_t ctx;
1459 	iflib_rxq_t rxq = (iflib_rxq_t)info->ifi_ctx;
1460 	iflib_txq_t txq;
1461 	void *sc;
1462 	int i, cidx, result;
1463 	qidx_t txqid;
1464 
1465 	if (!iflib_started)
1466 		return (FILTER_STRAY);
1467 
1468 	DBG_COUNTER_INC(fast_intrs);
1469 	if (info->ifi_filter != NULL) {
1470 		result = info->ifi_filter(info->ifi_filter_arg);
1471 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1472 			return (result);
1473 	}
1474 
1475 	ctx = rxq->ifr_ctx;
1476 	sc = ctx->ifc_softc;
1477 	MPASS(rxq->ifr_ntxqirq);
1478 	for (i = 0; i < rxq->ifr_ntxqirq; i++) {
1479 		txqid = rxq->ifr_txqid[i];
1480 		txq = &ctx->ifc_txqs[txqid];
1481 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1482 		    BUS_DMASYNC_POSTREAD);
1483 		if (!ctx->isc_txd_credits_update(sc, txqid, false)) {
1484 			IFDI_TX_QUEUE_INTR_ENABLE(ctx, txqid);
1485 			continue;
1486 		}
1487 		GROUPTASK_ENQUEUE(&txq->ift_task);
1488 	}
1489 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_RXCQ)
1490 		cidx = rxq->ifr_cq_cidx;
1491 	else
1492 		cidx = rxq->ifr_fl[0].ifl_cidx;
1493 	if (iflib_rxd_avail(ctx, rxq, cidx, 1))
1494 		GROUPTASK_ENQUEUE(gtask);
1495 	else {
1496 		IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
1497 		DBG_COUNTER_INC(rx_intr_enables);
1498 	}
1499 	return (FILTER_HANDLED);
1500 }
1501 
1502 
1503 static int
1504 iflib_fast_intr_ctx(void *arg)
1505 {
1506 	iflib_filter_info_t info = arg;
1507 	struct grouptask *gtask = info->ifi_task;
1508 	int result;
1509 
1510 	if (!iflib_started)
1511 		return (FILTER_STRAY);
1512 
1513 	DBG_COUNTER_INC(fast_intrs);
1514 	if (info->ifi_filter != NULL) {
1515 		result = info->ifi_filter(info->ifi_filter_arg);
1516 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1517 			return (result);
1518 	}
1519 
1520 	GROUPTASK_ENQUEUE(gtask);
1521 	return (FILTER_HANDLED);
1522 }
1523 
1524 static int
1525 _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
1526 		 driver_filter_t filter, driver_intr_t handler, void *arg,
1527 		 const char *name)
1528 {
1529 	int rc, flags;
1530 	struct resource *res;
1531 	void *tag = NULL;
1532 	device_t dev = ctx->ifc_dev;
1533 
1534 	flags = RF_ACTIVE;
1535 	if (ctx->ifc_flags & IFC_LEGACY)
1536 		flags |= RF_SHAREABLE;
1537 	MPASS(rid < 512);
1538 	irq->ii_rid = rid;
1539 	res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &irq->ii_rid, flags);
1540 	if (res == NULL) {
1541 		device_printf(dev,
1542 		    "failed to allocate IRQ for rid %d, name %s.\n", rid, name);
1543 		return (ENOMEM);
1544 	}
1545 	irq->ii_res = res;
1546 	KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL"));
1547 	rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET,
1548 						filter, handler, arg, &tag);
1549 	if (rc != 0) {
1550 		device_printf(dev,
1551 		    "failed to setup interrupt for rid %d, name %s: %d\n",
1552 					  rid, name ? name : "unknown", rc);
1553 		return (rc);
1554 	} else if (name)
1555 		bus_describe_intr(dev, res, tag, "%s", name);
1556 
1557 	irq->ii_tag = tag;
1558 	return (0);
1559 }
1560 
1561 
1562 /*********************************************************************
1563  *
1564  *  Allocate DMA resources for TX buffers as well as memory for the TX
1565  *  mbuf map.  TX DMA maps (non-TSO/TSO) and TX mbuf map are kept in a
1566  *  iflib_sw_tx_desc_array structure, storing all the information that
1567  *  is needed to transmit a packet on the wire.  This is called only
1568  *  once at attach, setup is done every reset.
1569  *
1570  **********************************************************************/
1571 static int
1572 iflib_txsd_alloc(iflib_txq_t txq)
1573 {
1574 	if_ctx_t ctx = txq->ift_ctx;
1575 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1576 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1577 	device_t dev = ctx->ifc_dev;
1578 	bus_size_t tsomaxsize;
1579 	int err, nsegments, ntsosegments;
1580 	bool tso;
1581 
1582 	nsegments = scctx->isc_tx_nsegments;
1583 	ntsosegments = scctx->isc_tx_tso_segments_max;
1584 	tsomaxsize = scctx->isc_tx_tso_size_max;
1585 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_VLAN_MTU)
1586 		tsomaxsize += sizeof(struct ether_vlan_header);
1587 	MPASS(scctx->isc_ntxd[0] > 0);
1588 	MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0);
1589 	MPASS(nsegments > 0);
1590 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) {
1591 		MPASS(ntsosegments > 0);
1592 		MPASS(sctx->isc_tso_maxsize >= tsomaxsize);
1593 	}
1594 
1595 	/*
1596 	 * Set up DMA tags for TX buffers.
1597 	 */
1598 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1599 			       1, 0,			/* alignment, bounds */
1600 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1601 			       BUS_SPACE_MAXADDR,	/* highaddr */
1602 			       NULL, NULL,		/* filter, filterarg */
1603 			       sctx->isc_tx_maxsize,		/* maxsize */
1604 			       nsegments,	/* nsegments */
1605 			       sctx->isc_tx_maxsegsize,	/* maxsegsize */
1606 			       0,			/* flags */
1607 			       NULL,			/* lockfunc */
1608 			       NULL,			/* lockfuncarg */
1609 			       &txq->ift_buf_tag))) {
1610 		device_printf(dev,"Unable to allocate TX DMA tag: %d\n", err);
1611 		device_printf(dev,"maxsize: %ju nsegments: %d maxsegsize: %ju\n",
1612 		    (uintmax_t)sctx->isc_tx_maxsize, nsegments, (uintmax_t)sctx->isc_tx_maxsegsize);
1613 		goto fail;
1614 	}
1615 	tso = (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) != 0;
1616 	if (tso && (err = bus_dma_tag_create(bus_get_dma_tag(dev),
1617 			       1, 0,			/* alignment, bounds */
1618 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1619 			       BUS_SPACE_MAXADDR,	/* highaddr */
1620 			       NULL, NULL,		/* filter, filterarg */
1621 			       tsomaxsize,		/* maxsize */
1622 			       ntsosegments,	/* nsegments */
1623 			       sctx->isc_tso_maxsegsize,/* maxsegsize */
1624 			       0,			/* flags */
1625 			       NULL,			/* lockfunc */
1626 			       NULL,			/* lockfuncarg */
1627 			       &txq->ift_tso_buf_tag))) {
1628 		device_printf(dev, "Unable to allocate TSO TX DMA tag: %d\n",
1629 		    err);
1630 		goto fail;
1631 	}
1632 
1633 	/* Allocate memory for the TX mbuf map. */
1634 	if (!(txq->ift_sds.ifsd_m =
1635 	    (struct mbuf **) malloc(sizeof(struct mbuf *) *
1636 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1637 		device_printf(dev, "Unable to allocate TX mbuf map memory\n");
1638 		err = ENOMEM;
1639 		goto fail;
1640 	}
1641 
1642 	/*
1643 	 * Create the DMA maps for TX buffers.
1644 	 */
1645 	if ((txq->ift_sds.ifsd_map = (bus_dmamap_t *)malloc(
1646 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1647 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1648 		device_printf(dev,
1649 		    "Unable to allocate TX buffer DMA map memory\n");
1650 		err = ENOMEM;
1651 		goto fail;
1652 	}
1653 	if (tso && (txq->ift_sds.ifsd_tso_map = (bus_dmamap_t *)malloc(
1654 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1655 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1656 		device_printf(dev,
1657 		    "Unable to allocate TSO TX buffer map memory\n");
1658 		err = ENOMEM;
1659 		goto fail;
1660 	}
1661 	for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) {
1662 		err = bus_dmamap_create(txq->ift_buf_tag, 0,
1663 		    &txq->ift_sds.ifsd_map[i]);
1664 		if (err != 0) {
1665 			device_printf(dev, "Unable to create TX DMA map\n");
1666 			goto fail;
1667 		}
1668 		if (!tso)
1669 			continue;
1670 		err = bus_dmamap_create(txq->ift_tso_buf_tag, 0,
1671 		    &txq->ift_sds.ifsd_tso_map[i]);
1672 		if (err != 0) {
1673 			device_printf(dev, "Unable to create TSO TX DMA map\n");
1674 			goto fail;
1675 		}
1676 	}
1677 	return (0);
1678 fail:
1679 	/* We free all, it handles case where we are in the middle */
1680 	iflib_tx_structures_free(ctx);
1681 	return (err);
1682 }
1683 
1684 static void
1685 iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i)
1686 {
1687 	bus_dmamap_t map;
1688 
1689 	map = NULL;
1690 	if (txq->ift_sds.ifsd_map != NULL)
1691 		map = txq->ift_sds.ifsd_map[i];
1692 	if (map != NULL) {
1693 		bus_dmamap_sync(txq->ift_buf_tag, map, BUS_DMASYNC_POSTWRITE);
1694 		bus_dmamap_unload(txq->ift_buf_tag, map);
1695 		bus_dmamap_destroy(txq->ift_buf_tag, map);
1696 		txq->ift_sds.ifsd_map[i] = NULL;
1697 	}
1698 
1699 	map = NULL;
1700 	if (txq->ift_sds.ifsd_tso_map != NULL)
1701 		map = txq->ift_sds.ifsd_tso_map[i];
1702 	if (map != NULL) {
1703 		bus_dmamap_sync(txq->ift_tso_buf_tag, map,
1704 		    BUS_DMASYNC_POSTWRITE);
1705 		bus_dmamap_unload(txq->ift_tso_buf_tag, map);
1706 		bus_dmamap_destroy(txq->ift_tso_buf_tag, map);
1707 		txq->ift_sds.ifsd_tso_map[i] = NULL;
1708 	}
1709 }
1710 
1711 static void
1712 iflib_txq_destroy(iflib_txq_t txq)
1713 {
1714 	if_ctx_t ctx = txq->ift_ctx;
1715 
1716 	for (int i = 0; i < txq->ift_size; i++)
1717 		iflib_txsd_destroy(ctx, txq, i);
1718 	if (txq->ift_sds.ifsd_map != NULL) {
1719 		free(txq->ift_sds.ifsd_map, M_IFLIB);
1720 		txq->ift_sds.ifsd_map = NULL;
1721 	}
1722 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1723 		free(txq->ift_sds.ifsd_tso_map, M_IFLIB);
1724 		txq->ift_sds.ifsd_tso_map = NULL;
1725 	}
1726 	if (txq->ift_sds.ifsd_m != NULL) {
1727 		free(txq->ift_sds.ifsd_m, M_IFLIB);
1728 		txq->ift_sds.ifsd_m = NULL;
1729 	}
1730 	if (txq->ift_buf_tag != NULL) {
1731 		bus_dma_tag_destroy(txq->ift_buf_tag);
1732 		txq->ift_buf_tag = NULL;
1733 	}
1734 	if (txq->ift_tso_buf_tag != NULL) {
1735 		bus_dma_tag_destroy(txq->ift_tso_buf_tag);
1736 		txq->ift_tso_buf_tag = NULL;
1737 	}
1738 }
1739 
1740 static void
1741 iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i)
1742 {
1743 	struct mbuf **mp;
1744 
1745 	mp = &txq->ift_sds.ifsd_m[i];
1746 	if (*mp == NULL)
1747 		return;
1748 
1749 	if (txq->ift_sds.ifsd_map != NULL) {
1750 		bus_dmamap_sync(txq->ift_buf_tag,
1751 		    txq->ift_sds.ifsd_map[i], BUS_DMASYNC_POSTWRITE);
1752 		bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[i]);
1753 	}
1754 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1755 		bus_dmamap_sync(txq->ift_tso_buf_tag,
1756 		    txq->ift_sds.ifsd_tso_map[i], BUS_DMASYNC_POSTWRITE);
1757 		bus_dmamap_unload(txq->ift_tso_buf_tag,
1758 		    txq->ift_sds.ifsd_tso_map[i]);
1759 	}
1760 	m_free(*mp);
1761 	DBG_COUNTER_INC(tx_frees);
1762 	*mp = NULL;
1763 }
1764 
1765 static int
1766 iflib_txq_setup(iflib_txq_t txq)
1767 {
1768 	if_ctx_t ctx = txq->ift_ctx;
1769 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1770 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1771 	iflib_dma_info_t di;
1772 	int i;
1773 
1774 	/* Set number of descriptors available */
1775 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
1776 	/* XXX make configurable */
1777 	txq->ift_update_freq = IFLIB_DEFAULT_TX_UPDATE_FREQ;
1778 
1779 	/* Reset indices */
1780 	txq->ift_cidx_processed = 0;
1781 	txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0;
1782 	txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset];
1783 
1784 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1785 		bzero((void *)di->idi_vaddr, di->idi_size);
1786 
1787 	IFDI_TXQ_SETUP(ctx, txq->ift_id);
1788 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1789 		bus_dmamap_sync(di->idi_tag, di->idi_map,
1790 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1791 	return (0);
1792 }
1793 
1794 /*********************************************************************
1795  *
1796  *  Allocate DMA resources for RX buffers as well as memory for the RX
1797  *  mbuf map, direct RX cluster pointer map and RX cluster bus address
1798  *  map.  RX DMA map, RX mbuf map, direct RX cluster pointer map and
1799  *  RX cluster map are kept in a iflib_sw_rx_desc_array structure.
1800  *  Since we use use one entry in iflib_sw_rx_desc_array per received
1801  *  packet, the maximum number of entries we'll need is equal to the
1802  *  number of hardware receive descriptors that we've allocated.
1803  *
1804  **********************************************************************/
1805 static int
1806 iflib_rxsd_alloc(iflib_rxq_t rxq)
1807 {
1808 	if_ctx_t ctx = rxq->ifr_ctx;
1809 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1810 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1811 	device_t dev = ctx->ifc_dev;
1812 	iflib_fl_t fl;
1813 	int			err;
1814 
1815 	MPASS(scctx->isc_nrxd[0] > 0);
1816 	MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0);
1817 
1818 	fl = rxq->ifr_fl;
1819 	for (int i = 0; i <  rxq->ifr_nfl; i++, fl++) {
1820 		fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */
1821 		/* Set up DMA tag for RX buffers. */
1822 		err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1823 					 1, 0,			/* alignment, bounds */
1824 					 BUS_SPACE_MAXADDR,	/* lowaddr */
1825 					 BUS_SPACE_MAXADDR,	/* highaddr */
1826 					 NULL, NULL,		/* filter, filterarg */
1827 					 sctx->isc_rx_maxsize,	/* maxsize */
1828 					 sctx->isc_rx_nsegments,	/* nsegments */
1829 					 sctx->isc_rx_maxsegsize,	/* maxsegsize */
1830 					 0,			/* flags */
1831 					 NULL,			/* lockfunc */
1832 					 NULL,			/* lockarg */
1833 					 &fl->ifl_buf_tag);
1834 		if (err) {
1835 			device_printf(dev,
1836 			    "Unable to allocate RX DMA tag: %d\n", err);
1837 			goto fail;
1838 		}
1839 
1840 		/* Allocate memory for the RX mbuf map. */
1841 		if (!(fl->ifl_sds.ifsd_m =
1842 		      (struct mbuf **) malloc(sizeof(struct mbuf *) *
1843 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1844 			device_printf(dev,
1845 			    "Unable to allocate RX mbuf map memory\n");
1846 			err = ENOMEM;
1847 			goto fail;
1848 		}
1849 
1850 		/* Allocate memory for the direct RX cluster pointer map. */
1851 		if (!(fl->ifl_sds.ifsd_cl =
1852 		      (caddr_t *) malloc(sizeof(caddr_t) *
1853 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1854 			device_printf(dev,
1855 			    "Unable to allocate RX cluster map memory\n");
1856 			err = ENOMEM;
1857 			goto fail;
1858 		}
1859 
1860 		/* Allocate memory for the RX cluster bus address map. */
1861 		if (!(fl->ifl_sds.ifsd_ba =
1862 		      (bus_addr_t *) malloc(sizeof(bus_addr_t) *
1863 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1864 			device_printf(dev,
1865 			    "Unable to allocate RX bus address map memory\n");
1866 			err = ENOMEM;
1867 			goto fail;
1868 		}
1869 
1870 		/*
1871 		 * Create the DMA maps for RX buffers.
1872 		 */
1873 		if (!(fl->ifl_sds.ifsd_map =
1874 		      (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1875 			device_printf(dev,
1876 			    "Unable to allocate RX buffer DMA map memory\n");
1877 			err = ENOMEM;
1878 			goto fail;
1879 		}
1880 		for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++) {
1881 			err = bus_dmamap_create(fl->ifl_buf_tag, 0,
1882 			    &fl->ifl_sds.ifsd_map[i]);
1883 			if (err != 0) {
1884 				device_printf(dev, "Unable to create RX buffer DMA map\n");
1885 				goto fail;
1886 			}
1887 		}
1888 	}
1889 	return (0);
1890 
1891 fail:
1892 	iflib_rx_structures_free(ctx);
1893 	return (err);
1894 }
1895 
1896 
1897 /*
1898  * Internal service routines
1899  */
1900 
1901 struct rxq_refill_cb_arg {
1902 	int               error;
1903 	bus_dma_segment_t seg;
1904 	int               nseg;
1905 };
1906 
1907 static void
1908 _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
1909 {
1910 	struct rxq_refill_cb_arg *cb_arg = arg;
1911 
1912 	cb_arg->error = error;
1913 	cb_arg->seg = segs[0];
1914 	cb_arg->nseg = nseg;
1915 }
1916 
1917 /**
1918  *	rxq_refill - refill an rxq  free-buffer list
1919  *	@ctx: the iflib context
1920  *	@rxq: the free-list to refill
1921  *	@n: the number of new buffers to allocate
1922  *
1923  *	(Re)populate an rxq free-buffer list with up to @n new packet buffers.
1924  *	The caller must assure that @n does not exceed the queue's capacity.
1925  */
1926 static void
1927 _iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
1928 {
1929 	struct if_rxd_update iru;
1930 	struct rxq_refill_cb_arg cb_arg;
1931 	struct mbuf *m;
1932 	caddr_t cl, *sd_cl;
1933 	struct mbuf **sd_m;
1934 	bus_dmamap_t *sd_map;
1935 	bus_addr_t bus_addr, *sd_ba;
1936 	int err, frag_idx, i, idx, n, pidx;
1937 	qidx_t credits;
1938 
1939 	sd_m = fl->ifl_sds.ifsd_m;
1940 	sd_map = fl->ifl_sds.ifsd_map;
1941 	sd_cl = fl->ifl_sds.ifsd_cl;
1942 	sd_ba = fl->ifl_sds.ifsd_ba;
1943 	pidx = fl->ifl_pidx;
1944 	idx = pidx;
1945 	frag_idx = fl->ifl_fragidx;
1946 	credits = fl->ifl_credits;
1947 
1948 	i = 0;
1949 	n = count;
1950 	MPASS(n > 0);
1951 	MPASS(credits + n <= fl->ifl_size);
1952 
1953 	if (pidx < fl->ifl_cidx)
1954 		MPASS(pidx + n <= fl->ifl_cidx);
1955 	if (pidx == fl->ifl_cidx && (credits < fl->ifl_size))
1956 		MPASS(fl->ifl_gen == 0);
1957 	if (pidx > fl->ifl_cidx)
1958 		MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx);
1959 
1960 	DBG_COUNTER_INC(fl_refills);
1961 	if (n > 8)
1962 		DBG_COUNTER_INC(fl_refills_large);
1963 	iru_init(&iru, fl->ifl_rxq, fl->ifl_id);
1964 	while (n--) {
1965 		/*
1966 		 * We allocate an uninitialized mbuf + cluster, mbuf is
1967 		 * initialized after rx.
1968 		 *
1969 		 * If the cluster is still set then we know a minimum sized packet was received
1970 		 */
1971 		bit_ffc_at(fl->ifl_rx_bitmap, frag_idx, fl->ifl_size,
1972 		    &frag_idx);
1973 		if (frag_idx < 0)
1974 			bit_ffc(fl->ifl_rx_bitmap, fl->ifl_size, &frag_idx);
1975 		MPASS(frag_idx >= 0);
1976 		if ((cl = sd_cl[frag_idx]) == NULL) {
1977 			if ((cl = m_cljget(NULL, M_NOWAIT, fl->ifl_buf_size)) == NULL)
1978 				break;
1979 
1980 			cb_arg.error = 0;
1981 			MPASS(sd_map != NULL);
1982 			err = bus_dmamap_load(fl->ifl_buf_tag, sd_map[frag_idx],
1983 			    cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg,
1984 			    BUS_DMA_NOWAIT);
1985 			if (err != 0 || cb_arg.error) {
1986 				/*
1987 				 * !zone_pack ?
1988 				 */
1989 				if (fl->ifl_zone == zone_pack)
1990 					uma_zfree(fl->ifl_zone, cl);
1991 				break;
1992 			}
1993 
1994 			sd_ba[frag_idx] =  bus_addr = cb_arg.seg.ds_addr;
1995 			sd_cl[frag_idx] = cl;
1996 #if MEMORY_LOGGING
1997 			fl->ifl_cl_enqueued++;
1998 #endif
1999 		} else {
2000 			bus_addr = sd_ba[frag_idx];
2001 		}
2002 		bus_dmamap_sync(fl->ifl_buf_tag, sd_map[frag_idx],
2003 		    BUS_DMASYNC_PREREAD);
2004 
2005 		MPASS(sd_m[frag_idx] == NULL);
2006 		if ((m = m_gethdr(M_NOWAIT, MT_NOINIT)) == NULL) {
2007 			break;
2008 		}
2009 		sd_m[frag_idx] = m;
2010 		bit_set(fl->ifl_rx_bitmap, frag_idx);
2011 #if MEMORY_LOGGING
2012 		fl->ifl_m_enqueued++;
2013 #endif
2014 
2015 		DBG_COUNTER_INC(rx_allocs);
2016 		fl->ifl_rxd_idxs[i] = frag_idx;
2017 		fl->ifl_bus_addrs[i] = bus_addr;
2018 		fl->ifl_vm_addrs[i] = cl;
2019 		credits++;
2020 		i++;
2021 		MPASS(credits <= fl->ifl_size);
2022 		if (++idx == fl->ifl_size) {
2023 			fl->ifl_gen = 1;
2024 			idx = 0;
2025 		}
2026 		if (n == 0 || i == IFLIB_MAX_RX_REFRESH) {
2027 			iru.iru_pidx = pidx;
2028 			iru.iru_count = i;
2029 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2030 			i = 0;
2031 			pidx = idx;
2032 			fl->ifl_pidx = idx;
2033 			fl->ifl_credits = credits;
2034 		}
2035 	}
2036 
2037 	if (i) {
2038 		iru.iru_pidx = pidx;
2039 		iru.iru_count = i;
2040 		ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2041 		fl->ifl_pidx = idx;
2042 		fl->ifl_credits = credits;
2043 	}
2044 	DBG_COUNTER_INC(rxd_flush);
2045 	if (fl->ifl_pidx == 0)
2046 		pidx = fl->ifl_size - 1;
2047 	else
2048 		pidx = fl->ifl_pidx - 1;
2049 
2050 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2051 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2052 	ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id, fl->ifl_id, pidx);
2053 	fl->ifl_fragidx = frag_idx;
2054 }
2055 
2056 static __inline void
2057 __iflib_fl_refill_lt(if_ctx_t ctx, iflib_fl_t fl, int max)
2058 {
2059 	/* we avoid allowing pidx to catch up with cidx as it confuses ixl */
2060 	int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1;
2061 #ifdef INVARIANTS
2062 	int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1;
2063 #endif
2064 
2065 	MPASS(fl->ifl_credits <= fl->ifl_size);
2066 	MPASS(reclaimable == delta);
2067 
2068 	if (reclaimable > 0)
2069 		_iflib_fl_refill(ctx, fl, min(max, reclaimable));
2070 }
2071 
2072 uint8_t
2073 iflib_in_detach(if_ctx_t ctx)
2074 {
2075 	bool in_detach;
2076 	STATE_LOCK(ctx);
2077 	in_detach = !!(ctx->ifc_flags & IFC_IN_DETACH);
2078 	STATE_UNLOCK(ctx);
2079 	return (in_detach);
2080 }
2081 
2082 static void
2083 iflib_fl_bufs_free(iflib_fl_t fl)
2084 {
2085 	iflib_dma_info_t idi = fl->ifl_ifdi;
2086 	bus_dmamap_t sd_map;
2087 	uint32_t i;
2088 
2089 	for (i = 0; i < fl->ifl_size; i++) {
2090 		struct mbuf **sd_m = &fl->ifl_sds.ifsd_m[i];
2091 		caddr_t *sd_cl = &fl->ifl_sds.ifsd_cl[i];
2092 
2093 		if (*sd_cl != NULL) {
2094 			sd_map = fl->ifl_sds.ifsd_map[i];
2095 			bus_dmamap_sync(fl->ifl_buf_tag, sd_map,
2096 			    BUS_DMASYNC_POSTREAD);
2097 			bus_dmamap_unload(fl->ifl_buf_tag, sd_map);
2098 			if (*sd_cl != NULL)
2099 				uma_zfree(fl->ifl_zone, *sd_cl);
2100 			// XXX: Should this get moved out?
2101 			if (iflib_in_detach(fl->ifl_rxq->ifr_ctx))
2102 				bus_dmamap_destroy(fl->ifl_buf_tag, sd_map);
2103 			if (*sd_m != NULL) {
2104 				m_init(*sd_m, M_NOWAIT, MT_DATA, 0);
2105 				uma_zfree(zone_mbuf, *sd_m);
2106 			}
2107 		} else {
2108 			MPASS(*sd_cl == NULL);
2109 			MPASS(*sd_m == NULL);
2110 		}
2111 #if MEMORY_LOGGING
2112 		fl->ifl_m_dequeued++;
2113 		fl->ifl_cl_dequeued++;
2114 #endif
2115 		*sd_cl = NULL;
2116 		*sd_m = NULL;
2117 	}
2118 #ifdef INVARIANTS
2119 	for (i = 0; i < fl->ifl_size; i++) {
2120 		MPASS(fl->ifl_sds.ifsd_cl[i] == NULL);
2121 		MPASS(fl->ifl_sds.ifsd_m[i] == NULL);
2122 	}
2123 #endif
2124 	/*
2125 	 * Reset free list values
2126 	 */
2127 	fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = fl->ifl_fragidx = 0;
2128 	bzero(idi->idi_vaddr, idi->idi_size);
2129 }
2130 
2131 /*********************************************************************
2132  *
2133  *  Initialize a receive ring and its buffers.
2134  *
2135  **********************************************************************/
2136 static int
2137 iflib_fl_setup(iflib_fl_t fl)
2138 {
2139 	iflib_rxq_t rxq = fl->ifl_rxq;
2140 	if_ctx_t ctx = rxq->ifr_ctx;
2141 
2142 	bit_nclear(fl->ifl_rx_bitmap, 0, fl->ifl_size - 1);
2143 	/*
2144 	** Free current RX buffer structs and their mbufs
2145 	*/
2146 	iflib_fl_bufs_free(fl);
2147 	/* Now replenish the mbufs */
2148 	MPASS(fl->ifl_credits == 0);
2149 	fl->ifl_buf_size = ctx->ifc_rx_mbuf_sz;
2150 	if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size)
2151 		ctx->ifc_max_fl_buf_size = fl->ifl_buf_size;
2152 	fl->ifl_cltype = m_gettype(fl->ifl_buf_size);
2153 	fl->ifl_zone = m_getzone(fl->ifl_buf_size);
2154 
2155 
2156 	/* avoid pre-allocating zillions of clusters to an idle card
2157 	 * potentially speeding up attach
2158 	 */
2159 	_iflib_fl_refill(ctx, fl, min(128, fl->ifl_size));
2160 	MPASS(min(128, fl->ifl_size) == fl->ifl_credits);
2161 	if (min(128, fl->ifl_size) != fl->ifl_credits)
2162 		return (ENOBUFS);
2163 	/*
2164 	 * handle failure
2165 	 */
2166 	MPASS(rxq != NULL);
2167 	MPASS(fl->ifl_ifdi != NULL);
2168 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2169 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2170 	return (0);
2171 }
2172 
2173 /*********************************************************************
2174  *
2175  *  Free receive ring data structures
2176  *
2177  **********************************************************************/
2178 static void
2179 iflib_rx_sds_free(iflib_rxq_t rxq)
2180 {
2181 	iflib_fl_t fl;
2182 	int i, j;
2183 
2184 	if (rxq->ifr_fl != NULL) {
2185 		for (i = 0; i < rxq->ifr_nfl; i++) {
2186 			fl = &rxq->ifr_fl[i];
2187 			if (fl->ifl_buf_tag != NULL) {
2188 				if (fl->ifl_sds.ifsd_map != NULL) {
2189 					for (j = 0; j < fl->ifl_size; j++) {
2190 						if (fl->ifl_sds.ifsd_map[j] ==
2191 						    NULL)
2192 							continue;
2193 						bus_dmamap_sync(
2194 						    fl->ifl_buf_tag,
2195 						    fl->ifl_sds.ifsd_map[j],
2196 						    BUS_DMASYNC_POSTREAD);
2197 						bus_dmamap_unload(
2198 						    fl->ifl_buf_tag,
2199 						    fl->ifl_sds.ifsd_map[j]);
2200 					}
2201 				}
2202 				bus_dma_tag_destroy(fl->ifl_buf_tag);
2203 				fl->ifl_buf_tag = NULL;
2204 			}
2205 			free(fl->ifl_sds.ifsd_m, M_IFLIB);
2206 			free(fl->ifl_sds.ifsd_cl, M_IFLIB);
2207 			free(fl->ifl_sds.ifsd_ba, M_IFLIB);
2208 			free(fl->ifl_sds.ifsd_map, M_IFLIB);
2209 			fl->ifl_sds.ifsd_m = NULL;
2210 			fl->ifl_sds.ifsd_cl = NULL;
2211 			fl->ifl_sds.ifsd_ba = NULL;
2212 			fl->ifl_sds.ifsd_map = NULL;
2213 		}
2214 		free(rxq->ifr_fl, M_IFLIB);
2215 		rxq->ifr_fl = NULL;
2216 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
2217 	}
2218 }
2219 
2220 /*
2221  * MI independent logic
2222  *
2223  */
2224 static void
2225 iflib_timer(void *arg)
2226 {
2227 	iflib_txq_t txq = arg;
2228 	if_ctx_t ctx = txq->ift_ctx;
2229 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2230 	uint64_t this_tick = ticks;
2231 	uint32_t reset_on = hz / 2;
2232 
2233 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
2234 		return;
2235 	/*
2236 	** Check on the state of the TX queue(s), this
2237 	** can be done without the lock because its RO
2238 	** and the HUNG state will be static if set.
2239 	*/
2240 	if (this_tick - txq->ift_last_timer_tick >= hz / 2) {
2241 		txq->ift_last_timer_tick = this_tick;
2242 		IFDI_TIMER(ctx, txq->ift_id);
2243 		if ((txq->ift_qstatus == IFLIB_QUEUE_HUNG) &&
2244 		    ((txq->ift_cleaned_prev == txq->ift_cleaned) ||
2245 		     (sctx->isc_pause_frames == 0)))
2246 			goto hung;
2247 
2248 		if (ifmp_ring_is_stalled(txq->ift_br))
2249 			txq->ift_qstatus = IFLIB_QUEUE_HUNG;
2250 		txq->ift_cleaned_prev = txq->ift_cleaned;
2251 	}
2252 #ifdef DEV_NETMAP
2253 	if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP)
2254 		iflib_netmap_timer_adjust(ctx, txq, &reset_on);
2255 #endif
2256 	/* handle any laggards */
2257 	if (txq->ift_db_pending)
2258 		GROUPTASK_ENQUEUE(&txq->ift_task);
2259 
2260 	sctx->isc_pause_frames = 0;
2261 	if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)
2262 		callout_reset_on(&txq->ift_timer, reset_on, iflib_timer, txq, txq->ift_timer.c_cpu);
2263 	return;
2264  hung:
2265 	device_printf(ctx->ifc_dev,  "TX(%d) desc avail = %d, pidx = %d\n",
2266 				  txq->ift_id, TXQ_AVAIL(txq), txq->ift_pidx);
2267 	STATE_LOCK(ctx);
2268 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2269 	ctx->ifc_flags |= (IFC_DO_WATCHDOG|IFC_DO_RESET);
2270 	iflib_admin_intr_deferred(ctx);
2271 	STATE_UNLOCK(ctx);
2272 }
2273 
2274 static void
2275 iflib_calc_rx_mbuf_sz(if_ctx_t ctx)
2276 {
2277 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2278 
2279 	/*
2280 	 * XXX don't set the max_frame_size to larger
2281 	 * than the hardware can handle
2282 	 */
2283 	if (sctx->isc_max_frame_size <= MCLBYTES)
2284 		ctx->ifc_rx_mbuf_sz = MCLBYTES;
2285 	else
2286 		ctx->ifc_rx_mbuf_sz = MJUMPAGESIZE;
2287 }
2288 
2289 uint32_t
2290 iflib_get_rx_mbuf_sz(if_ctx_t ctx)
2291 {
2292 	return (ctx->ifc_rx_mbuf_sz);
2293 }
2294 
2295 static void
2296 iflib_init_locked(if_ctx_t ctx)
2297 {
2298 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2299 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2300 	if_t ifp = ctx->ifc_ifp;
2301 	iflib_fl_t fl;
2302 	iflib_txq_t txq;
2303 	iflib_rxq_t rxq;
2304 	int i, j, tx_ip_csum_flags, tx_ip6_csum_flags;
2305 
2306 
2307 	if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2308 	IFDI_INTR_DISABLE(ctx);
2309 
2310 	tx_ip_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP);
2311 	tx_ip6_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | CSUM_IP6_SCTP);
2312 	/* Set hardware offload abilities */
2313 	if_clearhwassist(ifp);
2314 	if (if_getcapenable(ifp) & IFCAP_TXCSUM)
2315 		if_sethwassistbits(ifp, tx_ip_csum_flags, 0);
2316 	if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6)
2317 		if_sethwassistbits(ifp,  tx_ip6_csum_flags, 0);
2318 	if (if_getcapenable(ifp) & IFCAP_TSO4)
2319 		if_sethwassistbits(ifp, CSUM_IP_TSO, 0);
2320 	if (if_getcapenable(ifp) & IFCAP_TSO6)
2321 		if_sethwassistbits(ifp, CSUM_IP6_TSO, 0);
2322 
2323 	for (i = 0, txq = ctx->ifc_txqs; i < sctx->isc_ntxqsets; i++, txq++) {
2324 		CALLOUT_LOCK(txq);
2325 		callout_stop(&txq->ift_timer);
2326 		CALLOUT_UNLOCK(txq);
2327 		iflib_netmap_txq_init(ctx, txq);
2328 	}
2329 
2330 	/*
2331 	 * Calculate a suitable Rx mbuf size prior to calling IFDI_INIT, so
2332 	 * that drivers can use the value when setting up the hardware receive
2333 	 * buffers.
2334 	 */
2335 	iflib_calc_rx_mbuf_sz(ctx);
2336 
2337 #ifdef INVARIANTS
2338 	i = if_getdrvflags(ifp);
2339 #endif
2340 	IFDI_INIT(ctx);
2341 	MPASS(if_getdrvflags(ifp) == i);
2342 	for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
2343 		/* XXX this should really be done on a per-queue basis */
2344 		if (if_getcapenable(ifp) & IFCAP_NETMAP) {
2345 			MPASS(rxq->ifr_id == i);
2346 			iflib_netmap_rxq_init(ctx, rxq);
2347 			continue;
2348 		}
2349 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
2350 			if (iflib_fl_setup(fl)) {
2351 				device_printf(ctx->ifc_dev, "freelist setup failed - check cluster settings\n");
2352 				goto done;
2353 			}
2354 		}
2355 	}
2356 done:
2357 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
2358 	IFDI_INTR_ENABLE(ctx);
2359 	txq = ctx->ifc_txqs;
2360 	for (i = 0; i < sctx->isc_ntxqsets; i++, txq++)
2361 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq,
2362 			txq->ift_timer.c_cpu);
2363 }
2364 
2365 static int
2366 iflib_media_change(if_t ifp)
2367 {
2368 	if_ctx_t ctx = if_getsoftc(ifp);
2369 	int err;
2370 
2371 	CTX_LOCK(ctx);
2372 	if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0)
2373 		iflib_init_locked(ctx);
2374 	CTX_UNLOCK(ctx);
2375 	return (err);
2376 }
2377 
2378 static void
2379 iflib_media_status(if_t ifp, struct ifmediareq *ifmr)
2380 {
2381 	if_ctx_t ctx = if_getsoftc(ifp);
2382 
2383 	CTX_LOCK(ctx);
2384 	IFDI_UPDATE_ADMIN_STATUS(ctx);
2385 	IFDI_MEDIA_STATUS(ctx, ifmr);
2386 	CTX_UNLOCK(ctx);
2387 }
2388 
2389 void
2390 iflib_stop(if_ctx_t ctx)
2391 {
2392 	iflib_txq_t txq = ctx->ifc_txqs;
2393 	iflib_rxq_t rxq = ctx->ifc_rxqs;
2394 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2395 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2396 	iflib_dma_info_t di;
2397 	iflib_fl_t fl;
2398 	int i, j;
2399 
2400 	/* Tell the stack that the interface is no longer active */
2401 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2402 
2403 	IFDI_INTR_DISABLE(ctx);
2404 	DELAY(1000);
2405 	IFDI_STOP(ctx);
2406 	DELAY(1000);
2407 
2408 	iflib_debug_reset();
2409 	/* Wait for current tx queue users to exit to disarm watchdog timer. */
2410 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) {
2411 		/* make sure all transmitters have completed before proceeding XXX */
2412 
2413 		CALLOUT_LOCK(txq);
2414 		callout_stop(&txq->ift_timer);
2415 		CALLOUT_UNLOCK(txq);
2416 
2417 		/* clean any enqueued buffers */
2418 		iflib_ifmp_purge(txq);
2419 		/* Free any existing tx buffers. */
2420 		for (j = 0; j < txq->ift_size; j++) {
2421 			iflib_txsd_free(ctx, txq, j);
2422 		}
2423 		txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0;
2424 		txq->ift_in_use = txq->ift_gen = txq->ift_cidx = txq->ift_pidx = txq->ift_no_desc_avail = 0;
2425 		txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0;
2426 		txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0;
2427 		txq->ift_pullups = 0;
2428 		ifmp_ring_reset_stats(txq->ift_br);
2429 		for (j = 0, di = txq->ift_ifdi; j < sctx->isc_ntxqs; j++, di++)
2430 			bzero((void *)di->idi_vaddr, di->idi_size);
2431 	}
2432 	for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) {
2433 		/* make sure all transmitters have completed before proceeding XXX */
2434 
2435 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
2436 		for (j = 0, di = rxq->ifr_ifdi; j < sctx->isc_nrxqs; j++, di++)
2437 			bzero((void *)di->idi_vaddr, di->idi_size);
2438 		/* also resets the free lists pidx/cidx */
2439 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
2440 			iflib_fl_bufs_free(fl);
2441 	}
2442 }
2443 
2444 static inline caddr_t
2445 calc_next_rxd(iflib_fl_t fl, int cidx)
2446 {
2447 	qidx_t size;
2448 	int nrxd;
2449 	caddr_t start, end, cur, next;
2450 
2451 	nrxd = fl->ifl_size;
2452 	size = fl->ifl_rxd_size;
2453 	start = fl->ifl_ifdi->idi_vaddr;
2454 
2455 	if (__predict_false(size == 0))
2456 		return (start);
2457 	cur = start + size*cidx;
2458 	end = start + size*nrxd;
2459 	next = CACHE_PTR_NEXT(cur);
2460 	return (next < end ? next : start);
2461 }
2462 
2463 static inline void
2464 prefetch_pkts(iflib_fl_t fl, int cidx)
2465 {
2466 	int nextptr;
2467 	int nrxd = fl->ifl_size;
2468 	caddr_t next_rxd;
2469 
2470 
2471 	nextptr = (cidx + CACHE_PTR_INCREMENT) & (nrxd-1);
2472 	prefetch(&fl->ifl_sds.ifsd_m[nextptr]);
2473 	prefetch(&fl->ifl_sds.ifsd_cl[nextptr]);
2474 	next_rxd = calc_next_rxd(fl, cidx);
2475 	prefetch(next_rxd);
2476 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 1) & (nrxd-1)]);
2477 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 2) & (nrxd-1)]);
2478 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 3) & (nrxd-1)]);
2479 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 4) & (nrxd-1)]);
2480 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 1) & (nrxd-1)]);
2481 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 2) & (nrxd-1)]);
2482 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 3) & (nrxd-1)]);
2483 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 4) & (nrxd-1)]);
2484 }
2485 
2486 static void
2487 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, int unload, if_rxsd_t sd)
2488 {
2489 	int flid, cidx;
2490 	bus_dmamap_t map;
2491 	iflib_fl_t fl;
2492 	int next;
2493 
2494 	map = NULL;
2495 	flid = irf->irf_flid;
2496 	cidx = irf->irf_idx;
2497 	fl = &rxq->ifr_fl[flid];
2498 	sd->ifsd_fl = fl;
2499 	sd->ifsd_cidx = cidx;
2500 	sd->ifsd_m = &fl->ifl_sds.ifsd_m[cidx];
2501 	sd->ifsd_cl = &fl->ifl_sds.ifsd_cl[cidx];
2502 	fl->ifl_credits--;
2503 #if MEMORY_LOGGING
2504 	fl->ifl_m_dequeued++;
2505 #endif
2506 	if (rxq->ifr_ctx->ifc_flags & IFC_PREFETCH)
2507 		prefetch_pkts(fl, cidx);
2508 	next = (cidx + CACHE_PTR_INCREMENT) & (fl->ifl_size-1);
2509 	prefetch(&fl->ifl_sds.ifsd_map[next]);
2510 	map = fl->ifl_sds.ifsd_map[cidx];
2511 	next = (cidx + CACHE_LINE_SIZE) & (fl->ifl_size-1);
2512 
2513 	/* not valid assert if bxe really does SGE from non-contiguous elements */
2514 	MPASS(fl->ifl_cidx == cidx);
2515 	bus_dmamap_sync(fl->ifl_buf_tag, map, BUS_DMASYNC_POSTREAD);
2516 	if (unload)
2517 		bus_dmamap_unload(fl->ifl_buf_tag, map);
2518 	fl->ifl_cidx = (fl->ifl_cidx + 1) & (fl->ifl_size-1);
2519 	if (__predict_false(fl->ifl_cidx == 0))
2520 		fl->ifl_gen = 0;
2521 	bit_clear(fl->ifl_rx_bitmap, cidx);
2522 }
2523 
2524 static struct mbuf *
2525 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri, if_rxsd_t sd)
2526 {
2527 	int i, padlen , flags;
2528 	struct mbuf *m, *mh, *mt;
2529 	caddr_t cl;
2530 
2531 	i = 0;
2532 	mh = NULL;
2533 	do {
2534 		rxd_frag_to_sd(rxq, &ri->iri_frags[i], TRUE, sd);
2535 
2536 		MPASS(*sd->ifsd_cl != NULL);
2537 		MPASS(*sd->ifsd_m != NULL);
2538 
2539 		/* Don't include zero-length frags */
2540 		if (ri->iri_frags[i].irf_len == 0) {
2541 			/* XXX we can save the cluster here, but not the mbuf */
2542 			m_init(*sd->ifsd_m, M_NOWAIT, MT_DATA, 0);
2543 			m_free(*sd->ifsd_m);
2544 			*sd->ifsd_m = NULL;
2545 			continue;
2546 		}
2547 		m = *sd->ifsd_m;
2548 		*sd->ifsd_m = NULL;
2549 		if (mh == NULL) {
2550 			flags = M_PKTHDR|M_EXT;
2551 			mh = mt = m;
2552 			padlen = ri->iri_pad;
2553 		} else {
2554 			flags = M_EXT;
2555 			mt->m_next = m;
2556 			mt = m;
2557 			/* assuming padding is only on the first fragment */
2558 			padlen = 0;
2559 		}
2560 		cl = *sd->ifsd_cl;
2561 		*sd->ifsd_cl = NULL;
2562 
2563 		/* Can these two be made one ? */
2564 		m_init(m, M_NOWAIT, MT_DATA, flags);
2565 		m_cljset(m, cl, sd->ifsd_fl->ifl_cltype);
2566 		/*
2567 		 * These must follow m_init and m_cljset
2568 		 */
2569 		m->m_data += padlen;
2570 		ri->iri_len -= padlen;
2571 		m->m_len = ri->iri_frags[i].irf_len;
2572 	} while (++i < ri->iri_nfrags);
2573 
2574 	return (mh);
2575 }
2576 
2577 /*
2578  * Process one software descriptor
2579  */
2580 static struct mbuf *
2581 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri)
2582 {
2583 	struct if_rxsd sd;
2584 	struct mbuf *m;
2585 
2586 	/* should I merge this back in now that the two paths are basically duplicated? */
2587 	if (ri->iri_nfrags == 1 &&
2588 	    ri->iri_frags[0].irf_len <= MIN(IFLIB_RX_COPY_THRESH, MHLEN)) {
2589 		rxd_frag_to_sd(rxq, &ri->iri_frags[0], FALSE, &sd);
2590 		m = *sd.ifsd_m;
2591 		*sd.ifsd_m = NULL;
2592 		m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR);
2593 #ifndef __NO_STRICT_ALIGNMENT
2594 		if (!IP_ALIGNED(m))
2595 			m->m_data += 2;
2596 #endif
2597 		memcpy(m->m_data, *sd.ifsd_cl, ri->iri_len);
2598 		m->m_len = ri->iri_frags[0].irf_len;
2599        } else {
2600 		m = assemble_segments(rxq, ri, &sd);
2601 	}
2602 	m->m_pkthdr.len = ri->iri_len;
2603 	m->m_pkthdr.rcvif = ri->iri_ifp;
2604 	m->m_flags |= ri->iri_flags;
2605 	m->m_pkthdr.ether_vtag = ri->iri_vtag;
2606 	m->m_pkthdr.flowid = ri->iri_flowid;
2607 	M_HASHTYPE_SET(m, ri->iri_rsstype);
2608 	m->m_pkthdr.csum_flags = ri->iri_csum_flags;
2609 	m->m_pkthdr.csum_data = ri->iri_csum_data;
2610 	return (m);
2611 }
2612 
2613 #if defined(INET6) || defined(INET)
2614 static void
2615 iflib_get_ip_forwarding(struct lro_ctrl *lc, bool *v4, bool *v6)
2616 {
2617 	CURVNET_SET(lc->ifp->if_vnet);
2618 #if defined(INET6)
2619 	*v6 = VNET(ip6_forwarding);
2620 #endif
2621 #if defined(INET)
2622 	*v4 = VNET(ipforwarding);
2623 #endif
2624 	CURVNET_RESTORE();
2625 }
2626 
2627 /*
2628  * Returns true if it's possible this packet could be LROed.
2629  * if it returns false, it is guaranteed that tcp_lro_rx()
2630  * would not return zero.
2631  */
2632 static bool
2633 iflib_check_lro_possible(struct mbuf *m, bool v4_forwarding, bool v6_forwarding)
2634 {
2635 	struct ether_header *eh;
2636 	uint16_t eh_type;
2637 
2638 	eh = mtod(m, struct ether_header *);
2639 	eh_type = ntohs(eh->ether_type);
2640 	switch (eh_type) {
2641 #if defined(INET6)
2642 		case ETHERTYPE_IPV6:
2643 			return !v6_forwarding;
2644 #endif
2645 #if defined (INET)
2646 		case ETHERTYPE_IP:
2647 			return !v4_forwarding;
2648 #endif
2649 	}
2650 
2651 	return false;
2652 }
2653 #else
2654 static void
2655 iflib_get_ip_forwarding(struct lro_ctrl *lc __unused, bool *v4 __unused, bool *v6 __unused)
2656 {
2657 }
2658 #endif
2659 
2660 static bool
2661 iflib_rxeof(iflib_rxq_t rxq, qidx_t budget)
2662 {
2663 	if_ctx_t ctx = rxq->ifr_ctx;
2664 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2665 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2666 	int avail, i;
2667 	qidx_t *cidxp;
2668 	struct if_rxd_info ri;
2669 	int err, budget_left, rx_bytes, rx_pkts;
2670 	iflib_fl_t fl;
2671 	struct ifnet *ifp;
2672 	int lro_enabled;
2673 	bool v4_forwarding, v6_forwarding, lro_possible;
2674 
2675 	/*
2676 	 * XXX early demux data packets so that if_input processing only handles
2677 	 * acks in interrupt context
2678 	 */
2679 	struct mbuf *m, *mh, *mt, *mf;
2680 
2681 	lro_possible = v4_forwarding = v6_forwarding = false;
2682 	ifp = ctx->ifc_ifp;
2683 	mh = mt = NULL;
2684 	MPASS(budget > 0);
2685 	rx_pkts	= rx_bytes = 0;
2686 	if (sctx->isc_flags & IFLIB_HAS_RXCQ)
2687 		cidxp = &rxq->ifr_cq_cidx;
2688 	else
2689 		cidxp = &rxq->ifr_fl[0].ifl_cidx;
2690 	if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) {
2691 		for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2692 			__iflib_fl_refill_lt(ctx, fl, budget + 8);
2693 		DBG_COUNTER_INC(rx_unavail);
2694 		return (false);
2695 	}
2696 
2697 	for (budget_left = budget; budget_left > 0 && avail > 0;) {
2698 		if (__predict_false(!CTX_ACTIVE(ctx))) {
2699 			DBG_COUNTER_INC(rx_ctx_inactive);
2700 			break;
2701 		}
2702 		/*
2703 		 * Reset client set fields to their default values
2704 		 */
2705 		rxd_info_zero(&ri);
2706 		ri.iri_qsidx = rxq->ifr_id;
2707 		ri.iri_cidx = *cidxp;
2708 		ri.iri_ifp = ifp;
2709 		ri.iri_frags = rxq->ifr_frags;
2710 		err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
2711 
2712 		if (err)
2713 			goto err;
2714 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
2715 			*cidxp = ri.iri_cidx;
2716 			/* Update our consumer index */
2717 			/* XXX NB: shurd - check if this is still safe */
2718 			while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0]) {
2719 				rxq->ifr_cq_cidx -= scctx->isc_nrxd[0];
2720 				rxq->ifr_cq_gen = 0;
2721 			}
2722 			/* was this only a completion queue message? */
2723 			if (__predict_false(ri.iri_nfrags == 0))
2724 				continue;
2725 		}
2726 		MPASS(ri.iri_nfrags != 0);
2727 		MPASS(ri.iri_len != 0);
2728 
2729 		/* will advance the cidx on the corresponding free lists */
2730 		m = iflib_rxd_pkt_get(rxq, &ri);
2731 		avail--;
2732 		budget_left--;
2733 		if (avail == 0 && budget_left)
2734 			avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left);
2735 
2736 		if (__predict_false(m == NULL)) {
2737 			DBG_COUNTER_INC(rx_mbuf_null);
2738 			continue;
2739 		}
2740 		/* imm_pkt: -- cxgb */
2741 		if (mh == NULL)
2742 			mh = mt = m;
2743 		else {
2744 			mt->m_nextpkt = m;
2745 			mt = m;
2746 		}
2747 	}
2748 	/* make sure that we can refill faster than drain */
2749 	for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2750 		__iflib_fl_refill_lt(ctx, fl, budget + 8);
2751 
2752 	lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO);
2753 	if (lro_enabled)
2754 		iflib_get_ip_forwarding(&rxq->ifr_lc, &v4_forwarding, &v6_forwarding);
2755 	mt = mf = NULL;
2756 	while (mh != NULL) {
2757 		m = mh;
2758 		mh = mh->m_nextpkt;
2759 		m->m_nextpkt = NULL;
2760 #ifndef __NO_STRICT_ALIGNMENT
2761 		if (!IP_ALIGNED(m) && (m = iflib_fixup_rx(m)) == NULL)
2762 			continue;
2763 #endif
2764 		rx_bytes += m->m_pkthdr.len;
2765 		rx_pkts++;
2766 #if defined(INET6) || defined(INET)
2767 		if (lro_enabled) {
2768 			if (!lro_possible) {
2769 				lro_possible = iflib_check_lro_possible(m, v4_forwarding, v6_forwarding);
2770 				if (lro_possible && mf != NULL) {
2771 					ifp->if_input(ifp, mf);
2772 					DBG_COUNTER_INC(rx_if_input);
2773 					mt = mf = NULL;
2774 				}
2775 			}
2776 			if ((m->m_pkthdr.csum_flags & (CSUM_L4_CALC|CSUM_L4_VALID)) ==
2777 			    (CSUM_L4_CALC|CSUM_L4_VALID)) {
2778 				if (lro_possible && tcp_lro_rx(&rxq->ifr_lc, m, 0) == 0)
2779 					continue;
2780 			}
2781 		}
2782 #endif
2783 		if (lro_possible) {
2784 			ifp->if_input(ifp, m);
2785 			DBG_COUNTER_INC(rx_if_input);
2786 			continue;
2787 		}
2788 
2789 		if (mf == NULL)
2790 			mf = m;
2791 		if (mt != NULL)
2792 			mt->m_nextpkt = m;
2793 		mt = m;
2794 	}
2795 	if (mf != NULL) {
2796 		ifp->if_input(ifp, mf);
2797 		DBG_COUNTER_INC(rx_if_input);
2798 	}
2799 
2800 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
2801 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
2802 
2803 	/*
2804 	 * Flush any outstanding LRO work
2805 	 */
2806 #if defined(INET6) || defined(INET)
2807 	tcp_lro_flush_all(&rxq->ifr_lc);
2808 #endif
2809 	if (avail)
2810 		return true;
2811 	return (iflib_rxd_avail(ctx, rxq, *cidxp, 1));
2812 err:
2813 	STATE_LOCK(ctx);
2814 	ctx->ifc_flags |= IFC_DO_RESET;
2815 	iflib_admin_intr_deferred(ctx);
2816 	STATE_UNLOCK(ctx);
2817 	return (false);
2818 }
2819 
2820 #define TXD_NOTIFY_COUNT(txq) (((txq)->ift_size / (txq)->ift_update_freq)-1)
2821 static inline qidx_t
2822 txq_max_db_deferred(iflib_txq_t txq, qidx_t in_use)
2823 {
2824 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2825 	qidx_t minthresh = txq->ift_size / 8;
2826 	if (in_use > 4*minthresh)
2827 		return (notify_count);
2828 	if (in_use > 2*minthresh)
2829 		return (notify_count >> 1);
2830 	if (in_use > minthresh)
2831 		return (notify_count >> 3);
2832 	return (0);
2833 }
2834 
2835 static inline qidx_t
2836 txq_max_rs_deferred(iflib_txq_t txq)
2837 {
2838 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2839 	qidx_t minthresh = txq->ift_size / 8;
2840 	if (txq->ift_in_use > 4*minthresh)
2841 		return (notify_count);
2842 	if (txq->ift_in_use > 2*minthresh)
2843 		return (notify_count >> 1);
2844 	if (txq->ift_in_use > minthresh)
2845 		return (notify_count >> 2);
2846 	return (2);
2847 }
2848 
2849 #define M_CSUM_FLAGS(m) ((m)->m_pkthdr.csum_flags)
2850 #define M_HAS_VLANTAG(m) (m->m_flags & M_VLANTAG)
2851 
2852 #define TXQ_MAX_DB_DEFERRED(txq, in_use) txq_max_db_deferred((txq), (in_use))
2853 #define TXQ_MAX_RS_DEFERRED(txq) txq_max_rs_deferred(txq)
2854 #define TXQ_MAX_DB_CONSUMED(size) (size >> 4)
2855 
2856 /* forward compatibility for cxgb */
2857 #define FIRST_QSET(ctx) 0
2858 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets)
2859 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets)
2860 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx))
2861 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments))
2862 
2863 /* XXX we should be setting this to something other than zero */
2864 #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh)
2865 #define	MAX_TX_DESC(ctx) max((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max, \
2866     (ctx)->ifc_softc_ctx.isc_tx_nsegments)
2867 
2868 static inline bool
2869 iflib_txd_db_check(if_ctx_t ctx, iflib_txq_t txq, int ring, qidx_t in_use)
2870 {
2871 	qidx_t dbval, max;
2872 	bool rang;
2873 
2874 	rang = false;
2875 	max = TXQ_MAX_DB_DEFERRED(txq, in_use);
2876 	if (ring || txq->ift_db_pending >= max) {
2877 		dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx;
2878 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
2879 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2880 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval);
2881 		txq->ift_db_pending = txq->ift_npending = 0;
2882 		rang = true;
2883 	}
2884 	return (rang);
2885 }
2886 
2887 #ifdef PKT_DEBUG
2888 static void
2889 print_pkt(if_pkt_info_t pi)
2890 {
2891 	printf("pi len:  %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n",
2892 	       pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx);
2893 	printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n",
2894 	       pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag);
2895 	printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n",
2896 	       pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto);
2897 }
2898 #endif
2899 
2900 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO)
2901 #define IS_TX_OFFLOAD4(pi) ((pi)->ipi_csum_flags & (CSUM_IP_TCP | CSUM_IP_TSO))
2902 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO)
2903 #define IS_TX_OFFLOAD6(pi) ((pi)->ipi_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_TSO))
2904 
2905 static int
2906 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp)
2907 {
2908 	if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx;
2909 	struct ether_vlan_header *eh;
2910 	struct mbuf *m;
2911 
2912 	m = *mp;
2913 	if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) &&
2914 	    M_WRITABLE(m) == 0) {
2915 		if ((m = m_dup(m, M_NOWAIT)) == NULL) {
2916 			return (ENOMEM);
2917 		} else {
2918 			m_freem(*mp);
2919 			DBG_COUNTER_INC(tx_frees);
2920 			*mp = m;
2921 		}
2922 	}
2923 
2924 	/*
2925 	 * Determine where frame payload starts.
2926 	 * Jump over vlan headers if already present,
2927 	 * helpful for QinQ too.
2928 	 */
2929 	if (__predict_false(m->m_len < sizeof(*eh))) {
2930 		txq->ift_pullups++;
2931 		if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL))
2932 			return (ENOMEM);
2933 	}
2934 	eh = mtod(m, struct ether_vlan_header *);
2935 	if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
2936 		pi->ipi_etype = ntohs(eh->evl_proto);
2937 		pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
2938 	} else {
2939 		pi->ipi_etype = ntohs(eh->evl_encap_proto);
2940 		pi->ipi_ehdrlen = ETHER_HDR_LEN;
2941 	}
2942 
2943 	switch (pi->ipi_etype) {
2944 #ifdef INET
2945 	case ETHERTYPE_IP:
2946 	{
2947 		struct mbuf *n;
2948 		struct ip *ip = NULL;
2949 		struct tcphdr *th = NULL;
2950 		int minthlen;
2951 
2952 		minthlen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip) + sizeof(*th));
2953 		if (__predict_false(m->m_len < minthlen)) {
2954 			/*
2955 			 * if this code bloat is causing too much of a hit
2956 			 * move it to a separate function and mark it noinline
2957 			 */
2958 			if (m->m_len == pi->ipi_ehdrlen) {
2959 				n = m->m_next;
2960 				MPASS(n);
2961 				if (n->m_len >= sizeof(*ip))  {
2962 					ip = (struct ip *)n->m_data;
2963 					if (n->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2964 						th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2965 				} else {
2966 					txq->ift_pullups++;
2967 					if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
2968 						return (ENOMEM);
2969 					ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2970 				}
2971 			} else {
2972 				txq->ift_pullups++;
2973 				if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
2974 					return (ENOMEM);
2975 				ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2976 				if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2977 					th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2978 			}
2979 		} else {
2980 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2981 			if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2982 				th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2983 		}
2984 		pi->ipi_ip_hlen = ip->ip_hl << 2;
2985 		pi->ipi_ipproto = ip->ip_p;
2986 		pi->ipi_flags |= IPI_TX_IPV4;
2987 
2988 		/* TCP checksum offload may require TCP header length */
2989 		if (IS_TX_OFFLOAD4(pi)) {
2990 			if (__predict_true(pi->ipi_ipproto == IPPROTO_TCP)) {
2991 				if (__predict_false(th == NULL)) {
2992 					txq->ift_pullups++;
2993 					if (__predict_false((m = m_pullup(m, (ip->ip_hl << 2) + sizeof(*th))) == NULL))
2994 						return (ENOMEM);
2995 					th = (struct tcphdr *)((caddr_t)ip + pi->ipi_ip_hlen);
2996 				}
2997 				pi->ipi_tcp_hflags = th->th_flags;
2998 				pi->ipi_tcp_hlen = th->th_off << 2;
2999 				pi->ipi_tcp_seq = th->th_seq;
3000 			}
3001 			if (IS_TSO4(pi)) {
3002 				if (__predict_false(ip->ip_p != IPPROTO_TCP))
3003 					return (ENXIO);
3004 				/*
3005 				 * TSO always requires hardware checksum offload.
3006 				 */
3007 				pi->ipi_csum_flags |= (CSUM_IP_TCP | CSUM_IP);
3008 				th->th_sum = in_pseudo(ip->ip_src.s_addr,
3009 						       ip->ip_dst.s_addr, htons(IPPROTO_TCP));
3010 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3011 				if (sctx->isc_flags & IFLIB_TSO_INIT_IP) {
3012 					ip->ip_sum = 0;
3013 					ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz);
3014 				}
3015 			}
3016 		}
3017 		if ((sctx->isc_flags & IFLIB_NEED_ZERO_CSUM) && (pi->ipi_csum_flags & CSUM_IP))
3018                        ip->ip_sum = 0;
3019 
3020 		break;
3021 	}
3022 #endif
3023 #ifdef INET6
3024 	case ETHERTYPE_IPV6:
3025 	{
3026 		struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
3027 		struct tcphdr *th;
3028 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
3029 
3030 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
3031 			txq->ift_pullups++;
3032 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
3033 				return (ENOMEM);
3034 		}
3035 		th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
3036 
3037 		/* XXX-BZ this will go badly in case of ext hdrs. */
3038 		pi->ipi_ipproto = ip6->ip6_nxt;
3039 		pi->ipi_flags |= IPI_TX_IPV6;
3040 
3041 		/* TCP checksum offload may require TCP header length */
3042 		if (IS_TX_OFFLOAD6(pi)) {
3043 			if (pi->ipi_ipproto == IPPROTO_TCP) {
3044 				if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) {
3045 					txq->ift_pullups++;
3046 					if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL))
3047 						return (ENOMEM);
3048 				}
3049 				pi->ipi_tcp_hflags = th->th_flags;
3050 				pi->ipi_tcp_hlen = th->th_off << 2;
3051 				pi->ipi_tcp_seq = th->th_seq;
3052 			}
3053 			if (IS_TSO6(pi)) {
3054 				if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP))
3055 					return (ENXIO);
3056 				/*
3057 				 * TSO always requires hardware checksum offload.
3058 				 */
3059 				pi->ipi_csum_flags |= CSUM_IP6_TCP;
3060 				th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0);
3061 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3062 			}
3063 		}
3064 		break;
3065 	}
3066 #endif
3067 	default:
3068 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
3069 		pi->ipi_ip_hlen = 0;
3070 		break;
3071 	}
3072 	*mp = m;
3073 
3074 	return (0);
3075 }
3076 
3077 /*
3078  * If dodgy hardware rejects the scatter gather chain we've handed it
3079  * we'll need to remove the mbuf chain from ifsg_m[] before we can add the
3080  * m_defrag'd mbufs
3081  */
3082 static __noinline struct mbuf *
3083 iflib_remove_mbuf(iflib_txq_t txq)
3084 {
3085 	int ntxd, pidx;
3086 	struct mbuf *m, **ifsd_m;
3087 
3088 	ifsd_m = txq->ift_sds.ifsd_m;
3089 	ntxd = txq->ift_size;
3090 	pidx = txq->ift_pidx & (ntxd - 1);
3091 	ifsd_m = txq->ift_sds.ifsd_m;
3092 	m = ifsd_m[pidx];
3093 	ifsd_m[pidx] = NULL;
3094 	bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[pidx]);
3095 	if (txq->ift_sds.ifsd_tso_map != NULL)
3096 		bus_dmamap_unload(txq->ift_tso_buf_tag,
3097 		    txq->ift_sds.ifsd_tso_map[pidx]);
3098 #if MEMORY_LOGGING
3099 	txq->ift_dequeued++;
3100 #endif
3101 	return (m);
3102 }
3103 
3104 static inline caddr_t
3105 calc_next_txd(iflib_txq_t txq, int cidx, uint8_t qid)
3106 {
3107 	qidx_t size;
3108 	int ntxd;
3109 	caddr_t start, end, cur, next;
3110 
3111 	ntxd = txq->ift_size;
3112 	size = txq->ift_txd_size[qid];
3113 	start = txq->ift_ifdi[qid].idi_vaddr;
3114 
3115 	if (__predict_false(size == 0))
3116 		return (start);
3117 	cur = start + size*cidx;
3118 	end = start + size*ntxd;
3119 	next = CACHE_PTR_NEXT(cur);
3120 	return (next < end ? next : start);
3121 }
3122 
3123 /*
3124  * Pad an mbuf to ensure a minimum ethernet frame size.
3125  * min_frame_size is the frame size (less CRC) to pad the mbuf to
3126  */
3127 static __noinline int
3128 iflib_ether_pad(device_t dev, struct mbuf **m_head, uint16_t min_frame_size)
3129 {
3130 	/*
3131 	 * 18 is enough bytes to pad an ARP packet to 46 bytes, and
3132 	 * and ARP message is the smallest common payload I can think of
3133 	 */
3134 	static char pad[18];	/* just zeros */
3135 	int n;
3136 	struct mbuf *new_head;
3137 
3138 	if (!M_WRITABLE(*m_head)) {
3139 		new_head = m_dup(*m_head, M_NOWAIT);
3140 		if (new_head == NULL) {
3141 			m_freem(*m_head);
3142 			device_printf(dev, "cannot pad short frame, m_dup() failed");
3143 			DBG_COUNTER_INC(encap_pad_mbuf_fail);
3144 			DBG_COUNTER_INC(tx_frees);
3145 			return ENOMEM;
3146 		}
3147 		m_freem(*m_head);
3148 		*m_head = new_head;
3149 	}
3150 
3151 	for (n = min_frame_size - (*m_head)->m_pkthdr.len;
3152 	     n > 0; n -= sizeof(pad))
3153 		if (!m_append(*m_head, min(n, sizeof(pad)), pad))
3154 			break;
3155 
3156 	if (n > 0) {
3157 		m_freem(*m_head);
3158 		device_printf(dev, "cannot pad short frame\n");
3159 		DBG_COUNTER_INC(encap_pad_mbuf_fail);
3160 		DBG_COUNTER_INC(tx_frees);
3161 		return (ENOBUFS);
3162 	}
3163 
3164 	return 0;
3165 }
3166 
3167 static int
3168 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp)
3169 {
3170 	if_ctx_t		ctx;
3171 	if_shared_ctx_t		sctx;
3172 	if_softc_ctx_t		scctx;
3173 	bus_dma_tag_t		buf_tag;
3174 	bus_dma_segment_t	*segs;
3175 	struct mbuf		*m_head, **ifsd_m;
3176 	void			*next_txd;
3177 	bus_dmamap_t		map;
3178 	struct if_pkt_info	pi;
3179 	int remap = 0;
3180 	int err, nsegs, ndesc, max_segs, pidx, cidx, next, ntxd;
3181 
3182 	ctx = txq->ift_ctx;
3183 	sctx = ctx->ifc_sctx;
3184 	scctx = &ctx->ifc_softc_ctx;
3185 	segs = txq->ift_segs;
3186 	ntxd = txq->ift_size;
3187 	m_head = *m_headp;
3188 	map = NULL;
3189 
3190 	/*
3191 	 * If we're doing TSO the next descriptor to clean may be quite far ahead
3192 	 */
3193 	cidx = txq->ift_cidx;
3194 	pidx = txq->ift_pidx;
3195 	if (ctx->ifc_flags & IFC_PREFETCH) {
3196 		next = (cidx + CACHE_PTR_INCREMENT) & (ntxd-1);
3197 		if (!(ctx->ifc_flags & IFLIB_HAS_TXCQ)) {
3198 			next_txd = calc_next_txd(txq, cidx, 0);
3199 			prefetch(next_txd);
3200 		}
3201 
3202 		/* prefetch the next cache line of mbuf pointers and flags */
3203 		prefetch(&txq->ift_sds.ifsd_m[next]);
3204 		prefetch(&txq->ift_sds.ifsd_map[next]);
3205 		next = (cidx + CACHE_LINE_SIZE) & (ntxd-1);
3206 	}
3207 	map = txq->ift_sds.ifsd_map[pidx];
3208 	ifsd_m = txq->ift_sds.ifsd_m;
3209 
3210 	if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
3211 		buf_tag = txq->ift_tso_buf_tag;
3212 		max_segs = scctx->isc_tx_tso_segments_max;
3213 		map = txq->ift_sds.ifsd_tso_map[pidx];
3214 		MPASS(buf_tag != NULL);
3215 		MPASS(max_segs > 0);
3216 	} else {
3217 		buf_tag = txq->ift_buf_tag;
3218 		max_segs = scctx->isc_tx_nsegments;
3219 		map = txq->ift_sds.ifsd_map[pidx];
3220 	}
3221 	if ((sctx->isc_flags & IFLIB_NEED_ETHER_PAD) &&
3222 	    __predict_false(m_head->m_pkthdr.len < scctx->isc_min_frame_size)) {
3223 		err = iflib_ether_pad(ctx->ifc_dev, m_headp, scctx->isc_min_frame_size);
3224 		if (err) {
3225 			DBG_COUNTER_INC(encap_txd_encap_fail);
3226 			return err;
3227 		}
3228 	}
3229 	m_head = *m_headp;
3230 
3231 	pkt_info_zero(&pi);
3232 	pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG|M_BCAST|M_MCAST));
3233 	pi.ipi_pidx = pidx;
3234 	pi.ipi_qsidx = txq->ift_id;
3235 	pi.ipi_len = m_head->m_pkthdr.len;
3236 	pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags;
3237 	pi.ipi_vtag = (m_head->m_flags & M_VLANTAG) ? m_head->m_pkthdr.ether_vtag : 0;
3238 
3239 	/* deliberate bitwise OR to make one condition */
3240 	if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) {
3241 		if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0)) {
3242 			DBG_COUNTER_INC(encap_txd_encap_fail);
3243 			return (err);
3244 		}
3245 		m_head = *m_headp;
3246 	}
3247 
3248 retry:
3249 	err = bus_dmamap_load_mbuf_sg(buf_tag, map, m_head, segs, &nsegs,
3250 	    BUS_DMA_NOWAIT);
3251 defrag:
3252 	if (__predict_false(err)) {
3253 		switch (err) {
3254 		case EFBIG:
3255 			/* try collapse once and defrag once */
3256 			if (remap == 0) {
3257 				m_head = m_collapse(*m_headp, M_NOWAIT, max_segs);
3258 				/* try defrag if collapsing fails */
3259 				if (m_head == NULL)
3260 					remap++;
3261 			}
3262 			if (remap == 1) {
3263 				txq->ift_mbuf_defrag++;
3264 				m_head = m_defrag(*m_headp, M_NOWAIT);
3265 			}
3266 			/*
3267 			 * remap should never be >1 unless bus_dmamap_load_mbuf_sg
3268 			 * failed to map an mbuf that was run through m_defrag
3269 			 */
3270 			MPASS(remap <= 1);
3271 			if (__predict_false(m_head == NULL || remap > 1))
3272 				goto defrag_failed;
3273 			remap++;
3274 			*m_headp = m_head;
3275 			goto retry;
3276 			break;
3277 		case ENOMEM:
3278 			txq->ift_no_tx_dma_setup++;
3279 			break;
3280 		default:
3281 			txq->ift_no_tx_dma_setup++;
3282 			m_freem(*m_headp);
3283 			DBG_COUNTER_INC(tx_frees);
3284 			*m_headp = NULL;
3285 			break;
3286 		}
3287 		txq->ift_map_failed++;
3288 		DBG_COUNTER_INC(encap_load_mbuf_fail);
3289 		DBG_COUNTER_INC(encap_txd_encap_fail);
3290 		return (err);
3291 	}
3292 	ifsd_m[pidx] = m_head;
3293 	/*
3294 	 * XXX assumes a 1 to 1 relationship between segments and
3295 	 *        descriptors - this does not hold true on all drivers, e.g.
3296 	 *        cxgb
3297 	 */
3298 	if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) {
3299 		txq->ift_no_desc_avail++;
3300 		bus_dmamap_unload(buf_tag, map);
3301 		DBG_COUNTER_INC(encap_txq_avail_fail);
3302 		DBG_COUNTER_INC(encap_txd_encap_fail);
3303 		if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0)
3304 			GROUPTASK_ENQUEUE(&txq->ift_task);
3305 		return (ENOBUFS);
3306 	}
3307 	/*
3308 	 * On Intel cards we can greatly reduce the number of TX interrupts
3309 	 * we see by only setting report status on every Nth descriptor.
3310 	 * However, this also means that the driver will need to keep track
3311 	 * of the descriptors that RS was set on to check them for the DD bit.
3312 	 */
3313 	txq->ift_rs_pending += nsegs + 1;
3314 	if (txq->ift_rs_pending > TXQ_MAX_RS_DEFERRED(txq) ||
3315 	     iflib_no_tx_batch || (TXQ_AVAIL(txq) - nsegs) <= MAX_TX_DESC(ctx) + 2) {
3316 		pi.ipi_flags |= IPI_TX_INTR;
3317 		txq->ift_rs_pending = 0;
3318 	}
3319 
3320 	pi.ipi_segs = segs;
3321 	pi.ipi_nsegs = nsegs;
3322 
3323 	MPASS(pidx >= 0 && pidx < txq->ift_size);
3324 #ifdef PKT_DEBUG
3325 	print_pkt(&pi);
3326 #endif
3327 	if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) {
3328 		bus_dmamap_sync(buf_tag, map, BUS_DMASYNC_PREWRITE);
3329 		DBG_COUNTER_INC(tx_encap);
3330 		MPASS(pi.ipi_new_pidx < txq->ift_size);
3331 
3332 		ndesc = pi.ipi_new_pidx - pi.ipi_pidx;
3333 		if (pi.ipi_new_pidx < pi.ipi_pidx) {
3334 			ndesc += txq->ift_size;
3335 			txq->ift_gen = 1;
3336 		}
3337 		/*
3338 		 * drivers can need as many as
3339 		 * two sentinels
3340 		 */
3341 		MPASS(ndesc <= pi.ipi_nsegs + 2);
3342 		MPASS(pi.ipi_new_pidx != pidx);
3343 		MPASS(ndesc > 0);
3344 		txq->ift_in_use += ndesc;
3345 
3346 		/*
3347 		 * We update the last software descriptor again here because there may
3348 		 * be a sentinel and/or there may be more mbufs than segments
3349 		 */
3350 		txq->ift_pidx = pi.ipi_new_pidx;
3351 		txq->ift_npending += pi.ipi_ndescs;
3352 	} else {
3353 		*m_headp = m_head = iflib_remove_mbuf(txq);
3354 		if (err == EFBIG) {
3355 			txq->ift_txd_encap_efbig++;
3356 			if (remap < 2) {
3357 				remap = 1;
3358 				goto defrag;
3359 			}
3360 		}
3361 		goto defrag_failed;
3362 	}
3363 	/*
3364 	 * err can't possibly be non-zero here, so we don't neet to test it
3365 	 * to see if we need to DBG_COUNTER_INC(encap_txd_encap_fail).
3366 	 */
3367 	return (err);
3368 
3369 defrag_failed:
3370 	txq->ift_mbuf_defrag_failed++;
3371 	txq->ift_map_failed++;
3372 	m_freem(*m_headp);
3373 	DBG_COUNTER_INC(tx_frees);
3374 	*m_headp = NULL;
3375 	DBG_COUNTER_INC(encap_txd_encap_fail);
3376 	return (ENOMEM);
3377 }
3378 
3379 static void
3380 iflib_tx_desc_free(iflib_txq_t txq, int n)
3381 {
3382 	uint32_t qsize, cidx, mask, gen;
3383 	struct mbuf *m, **ifsd_m;
3384 	bool do_prefetch;
3385 
3386 	cidx = txq->ift_cidx;
3387 	gen = txq->ift_gen;
3388 	qsize = txq->ift_size;
3389 	mask = qsize-1;
3390 	ifsd_m = txq->ift_sds.ifsd_m;
3391 	do_prefetch = (txq->ift_ctx->ifc_flags & IFC_PREFETCH);
3392 
3393 	while (n-- > 0) {
3394 		if (do_prefetch) {
3395 			prefetch(ifsd_m[(cidx + 3) & mask]);
3396 			prefetch(ifsd_m[(cidx + 4) & mask]);
3397 		}
3398 		if ((m = ifsd_m[cidx]) != NULL) {
3399 			prefetch(&ifsd_m[(cidx + CACHE_PTR_INCREMENT) & mask]);
3400 			if (m->m_pkthdr.csum_flags & CSUM_TSO) {
3401 				bus_dmamap_sync(txq->ift_tso_buf_tag,
3402 				    txq->ift_sds.ifsd_tso_map[cidx],
3403 				    BUS_DMASYNC_POSTWRITE);
3404 				bus_dmamap_unload(txq->ift_tso_buf_tag,
3405 				    txq->ift_sds.ifsd_tso_map[cidx]);
3406 			} else {
3407 				bus_dmamap_sync(txq->ift_buf_tag,
3408 				    txq->ift_sds.ifsd_map[cidx],
3409 				    BUS_DMASYNC_POSTWRITE);
3410 				bus_dmamap_unload(txq->ift_buf_tag,
3411 				    txq->ift_sds.ifsd_map[cidx]);
3412 			}
3413 			/* XXX we don't support any drivers that batch packets yet */
3414 			MPASS(m->m_nextpkt == NULL);
3415 			m_freem(m);
3416 			ifsd_m[cidx] = NULL;
3417 #if MEMORY_LOGGING
3418 			txq->ift_dequeued++;
3419 #endif
3420 			DBG_COUNTER_INC(tx_frees);
3421 		}
3422 		if (__predict_false(++cidx == qsize)) {
3423 			cidx = 0;
3424 			gen = 0;
3425 		}
3426 	}
3427 	txq->ift_cidx = cidx;
3428 	txq->ift_gen = gen;
3429 }
3430 
3431 static __inline int
3432 iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh)
3433 {
3434 	int reclaim;
3435 	if_ctx_t ctx = txq->ift_ctx;
3436 
3437 	KASSERT(thresh >= 0, ("invalid threshold to reclaim"));
3438 	MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size);
3439 
3440 	/*
3441 	 * Need a rate-limiting check so that this isn't called every time
3442 	 */
3443 	iflib_tx_credits_update(ctx, txq);
3444 	reclaim = DESC_RECLAIMABLE(txq);
3445 
3446 	if (reclaim <= thresh /* + MAX_TX_DESC(txq->ift_ctx) */) {
3447 #ifdef INVARIANTS
3448 		if (iflib_verbose_debug) {
3449 			printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __FUNCTION__,
3450 			       txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments,
3451 			       reclaim, thresh);
3452 
3453 		}
3454 #endif
3455 		return (0);
3456 	}
3457 	iflib_tx_desc_free(txq, reclaim);
3458 	txq->ift_cleaned += reclaim;
3459 	txq->ift_in_use -= reclaim;
3460 
3461 	return (reclaim);
3462 }
3463 
3464 static struct mbuf **
3465 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset, int remaining)
3466 {
3467 	int next, size;
3468 	struct mbuf **items;
3469 
3470 	size = r->size;
3471 	next = (cidx + CACHE_PTR_INCREMENT) & (size-1);
3472 	items = __DEVOLATILE(struct mbuf **, &r->items[0]);
3473 
3474 	prefetch(items[(cidx + offset) & (size-1)]);
3475 	if (remaining > 1) {
3476 		prefetch2cachelines(&items[next]);
3477 		prefetch2cachelines(items[(cidx + offset + 1) & (size-1)]);
3478 		prefetch2cachelines(items[(cidx + offset + 2) & (size-1)]);
3479 		prefetch2cachelines(items[(cidx + offset + 3) & (size-1)]);
3480 	}
3481 	return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (size-1)]));
3482 }
3483 
3484 static void
3485 iflib_txq_check_drain(iflib_txq_t txq, int budget)
3486 {
3487 
3488 	ifmp_ring_check_drainage(txq->ift_br, budget);
3489 }
3490 
3491 static uint32_t
3492 iflib_txq_can_drain(struct ifmp_ring *r)
3493 {
3494 	iflib_txq_t txq = r->cookie;
3495 	if_ctx_t ctx = txq->ift_ctx;
3496 
3497 	if (TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2)
3498 		return (1);
3499 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3500 	    BUS_DMASYNC_POSTREAD);
3501 	return (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id,
3502 	    false));
3503 }
3504 
3505 static uint32_t
3506 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3507 {
3508 	iflib_txq_t txq = r->cookie;
3509 	if_ctx_t ctx = txq->ift_ctx;
3510 	struct ifnet *ifp = ctx->ifc_ifp;
3511 	struct mbuf **mp, *m;
3512 	int i, count, consumed, pkt_sent, bytes_sent, mcast_sent, avail;
3513 	int reclaimed, err, in_use_prev, desc_used;
3514 	bool do_prefetch, ring, rang;
3515 
3516 	if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) ||
3517 			    !LINK_ACTIVE(ctx))) {
3518 		DBG_COUNTER_INC(txq_drain_notready);
3519 		return (0);
3520 	}
3521 	reclaimed = iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
3522 	rang = iflib_txd_db_check(ctx, txq, reclaimed, txq->ift_in_use);
3523 	avail = IDXDIFF(pidx, cidx, r->size);
3524 	if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) {
3525 		DBG_COUNTER_INC(txq_drain_flushing);
3526 		for (i = 0; i < avail; i++) {
3527 			if (__predict_true(r->items[(cidx + i) & (r->size-1)] != (void *)txq))
3528 				m_free(r->items[(cidx + i) & (r->size-1)]);
3529 			r->items[(cidx + i) & (r->size-1)] = NULL;
3530 		}
3531 		return (avail);
3532 	}
3533 
3534 	if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3535 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3536 		CALLOUT_LOCK(txq);
3537 		callout_stop(&txq->ift_timer);
3538 		CALLOUT_UNLOCK(txq);
3539 		DBG_COUNTER_INC(txq_drain_oactive);
3540 		return (0);
3541 	}
3542 	if (reclaimed)
3543 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3544 	consumed = mcast_sent = bytes_sent = pkt_sent = 0;
3545 	count = MIN(avail, TX_BATCH_SIZE);
3546 #ifdef INVARIANTS
3547 	if (iflib_verbose_debug)
3548 		printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __FUNCTION__,
3549 		       avail, ctx->ifc_flags, TXQ_AVAIL(txq));
3550 #endif
3551 	do_prefetch = (ctx->ifc_flags & IFC_PREFETCH);
3552 	avail = TXQ_AVAIL(txq);
3553 	err = 0;
3554 	for (desc_used = i = 0; i < count && avail > MAX_TX_DESC(ctx) + 2; i++) {
3555 		int rem = do_prefetch ? count - i : 0;
3556 
3557 		mp = _ring_peek_one(r, cidx, i, rem);
3558 		MPASS(mp != NULL && *mp != NULL);
3559 		if (__predict_false(*mp == (struct mbuf *)txq)) {
3560 			consumed++;
3561 			reclaimed++;
3562 			continue;
3563 		}
3564 		in_use_prev = txq->ift_in_use;
3565 		err = iflib_encap(txq, mp);
3566 		if (__predict_false(err)) {
3567 			/* no room - bail out */
3568 			if (err == ENOBUFS)
3569 				break;
3570 			consumed++;
3571 			/* we can't send this packet - skip it */
3572 			continue;
3573 		}
3574 		consumed++;
3575 		pkt_sent++;
3576 		m = *mp;
3577 		DBG_COUNTER_INC(tx_sent);
3578 		bytes_sent += m->m_pkthdr.len;
3579 		mcast_sent += !!(m->m_flags & M_MCAST);
3580 		avail = TXQ_AVAIL(txq);
3581 
3582 		txq->ift_db_pending += (txq->ift_in_use - in_use_prev);
3583 		desc_used += (txq->ift_in_use - in_use_prev);
3584 		ETHER_BPF_MTAP(ifp, m);
3585 		if (__predict_false(!(ifp->if_drv_flags & IFF_DRV_RUNNING)))
3586 			break;
3587 		rang = iflib_txd_db_check(ctx, txq, false, in_use_prev);
3588 	}
3589 
3590 	/* deliberate use of bitwise or to avoid gratuitous short-circuit */
3591 	ring = rang ? false  : (iflib_min_tx_latency | err) || (TXQ_AVAIL(txq) < MAX_TX_DESC(ctx));
3592 	iflib_txd_db_check(ctx, txq, ring, txq->ift_in_use);
3593 	if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
3594 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
3595 	if (mcast_sent)
3596 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
3597 #ifdef INVARIANTS
3598 	if (iflib_verbose_debug)
3599 		printf("consumed=%d\n", consumed);
3600 #endif
3601 	return (consumed);
3602 }
3603 
3604 static uint32_t
3605 iflib_txq_drain_always(struct ifmp_ring *r)
3606 {
3607 	return (1);
3608 }
3609 
3610 static uint32_t
3611 iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3612 {
3613 	int i, avail;
3614 	struct mbuf **mp;
3615 	iflib_txq_t txq;
3616 
3617 	txq = r->cookie;
3618 
3619 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3620 	CALLOUT_LOCK(txq);
3621 	callout_stop(&txq->ift_timer);
3622 	CALLOUT_UNLOCK(txq);
3623 
3624 	avail = IDXDIFF(pidx, cidx, r->size);
3625 	for (i = 0; i < avail; i++) {
3626 		mp = _ring_peek_one(r, cidx, i, avail - i);
3627 		if (__predict_false(*mp == (struct mbuf *)txq))
3628 			continue;
3629 		m_freem(*mp);
3630 		DBG_COUNTER_INC(tx_frees);
3631 	}
3632 	MPASS(ifmp_ring_is_stalled(r) == 0);
3633 	return (avail);
3634 }
3635 
3636 static void
3637 iflib_ifmp_purge(iflib_txq_t txq)
3638 {
3639 	struct ifmp_ring *r;
3640 
3641 	r = txq->ift_br;
3642 	r->drain = iflib_txq_drain_free;
3643 	r->can_drain = iflib_txq_drain_always;
3644 
3645 	ifmp_ring_check_drainage(r, r->size);
3646 
3647 	r->drain = iflib_txq_drain;
3648 	r->can_drain = iflib_txq_can_drain;
3649 }
3650 
3651 static void
3652 _task_fn_tx(void *context)
3653 {
3654 	iflib_txq_t txq = context;
3655 	if_ctx_t ctx = txq->ift_ctx;
3656 #if defined(ALTQ) || defined(DEV_NETMAP)
3657 	if_t ifp = ctx->ifc_ifp;
3658 #endif
3659 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
3660 
3661 #ifdef IFLIB_DIAGNOSTICS
3662 	txq->ift_cpu_exec_count[curcpu]++;
3663 #endif
3664 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3665 		return;
3666 #ifdef DEV_NETMAP
3667 	if (if_getcapenable(ifp) & IFCAP_NETMAP) {
3668 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3669 		    BUS_DMASYNC_POSTREAD);
3670 		if (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false))
3671 			netmap_tx_irq(ifp, txq->ift_id);
3672 		IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3673 		return;
3674 	}
3675 #endif
3676 #ifdef ALTQ
3677 	if (ALTQ_IS_ENABLED(&ifp->if_snd))
3678 		iflib_altq_if_start(ifp);
3679 #endif
3680 	if (txq->ift_db_pending)
3681 		ifmp_ring_enqueue(txq->ift_br, (void **)&txq, 1, TX_BATCH_SIZE, abdicate);
3682 	else if (!abdicate)
3683 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3684 	/*
3685 	 * When abdicating, we always need to check drainage, not just when we don't enqueue
3686 	 */
3687 	if (abdicate)
3688 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3689 	if (ctx->ifc_flags & IFC_LEGACY)
3690 		IFDI_INTR_ENABLE(ctx);
3691 	else {
3692 #ifdef INVARIANTS
3693 		int rc =
3694 #endif
3695 			IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3696 			KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver"));
3697 	}
3698 }
3699 
3700 static void
3701 _task_fn_rx(void *context)
3702 {
3703 	iflib_rxq_t rxq = context;
3704 	if_ctx_t ctx = rxq->ifr_ctx;
3705 	bool more;
3706 	uint16_t budget;
3707 
3708 #ifdef IFLIB_DIAGNOSTICS
3709 	rxq->ifr_cpu_exec_count[curcpu]++;
3710 #endif
3711 	DBG_COUNTER_INC(task_fn_rxs);
3712 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3713 		return;
3714 	more = true;
3715 #ifdef DEV_NETMAP
3716 	if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP) {
3717 		u_int work = 0;
3718 		if (netmap_rx_irq(ctx->ifc_ifp, rxq->ifr_id, &work)) {
3719 			more = false;
3720 		}
3721 	}
3722 #endif
3723 	budget = ctx->ifc_sysctl_rx_budget;
3724 	if (budget == 0)
3725 		budget = 16;	/* XXX */
3726 	if (more == false || (more = iflib_rxeof(rxq, budget)) == false) {
3727 		if (ctx->ifc_flags & IFC_LEGACY)
3728 			IFDI_INTR_ENABLE(ctx);
3729 		else {
3730 #ifdef INVARIANTS
3731 			int rc =
3732 #endif
3733 				IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
3734 			KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver"));
3735 			DBG_COUNTER_INC(rx_intr_enables);
3736 		}
3737 	}
3738 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3739 		return;
3740 	if (more)
3741 		GROUPTASK_ENQUEUE(&rxq->ifr_task);
3742 }
3743 
3744 static void
3745 _task_fn_admin(void *context)
3746 {
3747 	if_ctx_t ctx = context;
3748 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
3749 	iflib_txq_t txq;
3750 	int i;
3751 	bool oactive, running, do_reset, do_watchdog, in_detach;
3752 	uint32_t reset_on = hz / 2;
3753 
3754 	STATE_LOCK(ctx);
3755 	running = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING);
3756 	oactive = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE);
3757 	do_reset = (ctx->ifc_flags & IFC_DO_RESET);
3758 	do_watchdog = (ctx->ifc_flags & IFC_DO_WATCHDOG);
3759 	in_detach = (ctx->ifc_flags & IFC_IN_DETACH);
3760 	ctx->ifc_flags &= ~(IFC_DO_RESET|IFC_DO_WATCHDOG);
3761 	STATE_UNLOCK(ctx);
3762 
3763 	if ((!running && !oactive) && !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
3764 		return;
3765 	if (in_detach)
3766 		return;
3767 
3768 	CTX_LOCK(ctx);
3769 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3770 		CALLOUT_LOCK(txq);
3771 		callout_stop(&txq->ift_timer);
3772 		CALLOUT_UNLOCK(txq);
3773 	}
3774 	if (do_watchdog) {
3775 		ctx->ifc_watchdog_events++;
3776 		IFDI_WATCHDOG_RESET(ctx);
3777 	}
3778 	IFDI_UPDATE_ADMIN_STATUS(ctx);
3779 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3780 #ifdef DEV_NETMAP
3781 		reset_on = hz / 2;
3782 		if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP)
3783 			iflib_netmap_timer_adjust(ctx, txq, &reset_on);
3784 #endif
3785 		callout_reset_on(&txq->ift_timer, reset_on, iflib_timer, txq, txq->ift_timer.c_cpu);
3786 	}
3787 	IFDI_LINK_INTR_ENABLE(ctx);
3788 	if (do_reset)
3789 		iflib_if_init_locked(ctx);
3790 	CTX_UNLOCK(ctx);
3791 
3792 	if (LINK_ACTIVE(ctx) == 0)
3793 		return;
3794 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3795 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
3796 }
3797 
3798 
3799 static void
3800 _task_fn_iov(void *context)
3801 {
3802 	if_ctx_t ctx = context;
3803 
3804 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING) &&
3805 	    !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
3806 		return;
3807 
3808 	CTX_LOCK(ctx);
3809 	IFDI_VFLR_HANDLE(ctx);
3810 	CTX_UNLOCK(ctx);
3811 }
3812 
3813 static int
3814 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)
3815 {
3816 	int err;
3817 	if_int_delay_info_t info;
3818 	if_ctx_t ctx;
3819 
3820 	info = (if_int_delay_info_t)arg1;
3821 	ctx = info->iidi_ctx;
3822 	info->iidi_req = req;
3823 	info->iidi_oidp = oidp;
3824 	CTX_LOCK(ctx);
3825 	err = IFDI_SYSCTL_INT_DELAY(ctx, info);
3826 	CTX_UNLOCK(ctx);
3827 	return (err);
3828 }
3829 
3830 /*********************************************************************
3831  *
3832  *  IFNET FUNCTIONS
3833  *
3834  **********************************************************************/
3835 
3836 static void
3837 iflib_if_init_locked(if_ctx_t ctx)
3838 {
3839 	iflib_stop(ctx);
3840 	iflib_init_locked(ctx);
3841 }
3842 
3843 
3844 static void
3845 iflib_if_init(void *arg)
3846 {
3847 	if_ctx_t ctx = arg;
3848 
3849 	CTX_LOCK(ctx);
3850 	iflib_if_init_locked(ctx);
3851 	CTX_UNLOCK(ctx);
3852 }
3853 
3854 static int
3855 iflib_if_transmit(if_t ifp, struct mbuf *m)
3856 {
3857 	if_ctx_t	ctx = if_getsoftc(ifp);
3858 
3859 	iflib_txq_t txq;
3860 	int err, qidx;
3861 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
3862 
3863 	if (__predict_false((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) {
3864 		DBG_COUNTER_INC(tx_frees);
3865 		m_freem(m);
3866 		return (ENETDOWN);
3867 	}
3868 
3869 	MPASS(m->m_nextpkt == NULL);
3870 	/* ALTQ-enabled interfaces always use queue 0. */
3871 	qidx = 0;
3872 	if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m) && !ALTQ_IS_ENABLED(&ifp->if_snd))
3873 		qidx = QIDX(ctx, m);
3874 	/*
3875 	 * XXX calculate buf_ring based on flowid (divvy up bits?)
3876 	 */
3877 	txq = &ctx->ifc_txqs[qidx];
3878 
3879 #ifdef DRIVER_BACKPRESSURE
3880 	if (txq->ift_closed) {
3881 		while (m != NULL) {
3882 			next = m->m_nextpkt;
3883 			m->m_nextpkt = NULL;
3884 			m_freem(m);
3885 			DBG_COUNTER_INC(tx_frees);
3886 			m = next;
3887 		}
3888 		return (ENOBUFS);
3889 	}
3890 #endif
3891 #ifdef notyet
3892 	qidx = count = 0;
3893 	mp = marr;
3894 	next = m;
3895 	do {
3896 		count++;
3897 		next = next->m_nextpkt;
3898 	} while (next != NULL);
3899 
3900 	if (count > nitems(marr))
3901 		if ((mp = malloc(count*sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) {
3902 			/* XXX check nextpkt */
3903 			m_freem(m);
3904 			/* XXX simplify for now */
3905 			DBG_COUNTER_INC(tx_frees);
3906 			return (ENOBUFS);
3907 		}
3908 	for (next = m, i = 0; next != NULL; i++) {
3909 		mp[i] = next;
3910 		next = next->m_nextpkt;
3911 		mp[i]->m_nextpkt = NULL;
3912 	}
3913 #endif
3914 	DBG_COUNTER_INC(tx_seen);
3915 	err = ifmp_ring_enqueue(txq->ift_br, (void **)&m, 1, TX_BATCH_SIZE, abdicate);
3916 
3917 	if (abdicate)
3918 		GROUPTASK_ENQUEUE(&txq->ift_task);
3919  	if (err) {
3920 		if (!abdicate)
3921 			GROUPTASK_ENQUEUE(&txq->ift_task);
3922 		/* support forthcoming later */
3923 #ifdef DRIVER_BACKPRESSURE
3924 		txq->ift_closed = TRUE;
3925 #endif
3926 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3927 		m_freem(m);
3928 		DBG_COUNTER_INC(tx_frees);
3929 	}
3930 
3931 	return (err);
3932 }
3933 
3934 #ifdef ALTQ
3935 /*
3936  * The overall approach to integrating iflib with ALTQ is to continue to use
3937  * the iflib mp_ring machinery between the ALTQ queue(s) and the hardware
3938  * ring.  Technically, when using ALTQ, queueing to an intermediate mp_ring
3939  * is redundant/unnecessary, but doing so minimizes the amount of
3940  * ALTQ-specific code required in iflib.  It is assumed that the overhead of
3941  * redundantly queueing to an intermediate mp_ring is swamped by the
3942  * performance limitations inherent in using ALTQ.
3943  *
3944  * When ALTQ support is compiled in, all iflib drivers will use a transmit
3945  * routine, iflib_altq_if_transmit(), that checks if ALTQ is enabled for the
3946  * given interface.  If ALTQ is enabled for an interface, then all
3947  * transmitted packets for that interface will be submitted to the ALTQ
3948  * subsystem via IFQ_ENQUEUE().  We don't use the legacy if_transmit()
3949  * implementation because it uses IFQ_HANDOFF(), which will duplicatively
3950  * update stats that the iflib machinery handles, and which is sensitve to
3951  * the disused IFF_DRV_OACTIVE flag.  Additionally, iflib_altq_if_start()
3952  * will be installed as the start routine for use by ALTQ facilities that
3953  * need to trigger queue drains on a scheduled basis.
3954  *
3955  */
3956 static void
3957 iflib_altq_if_start(if_t ifp)
3958 {
3959 	struct ifaltq *ifq = &ifp->if_snd;
3960 	struct mbuf *m;
3961 
3962 	IFQ_LOCK(ifq);
3963 	IFQ_DEQUEUE_NOLOCK(ifq, m);
3964 	while (m != NULL) {
3965 		iflib_if_transmit(ifp, m);
3966 		IFQ_DEQUEUE_NOLOCK(ifq, m);
3967 	}
3968 	IFQ_UNLOCK(ifq);
3969 }
3970 
3971 static int
3972 iflib_altq_if_transmit(if_t ifp, struct mbuf *m)
3973 {
3974 	int err;
3975 
3976 	if (ALTQ_IS_ENABLED(&ifp->if_snd)) {
3977 		IFQ_ENQUEUE(&ifp->if_snd, m, err);
3978 		if (err == 0)
3979 			iflib_altq_if_start(ifp);
3980 	} else
3981 		err = iflib_if_transmit(ifp, m);
3982 
3983 	return (err);
3984 }
3985 #endif /* ALTQ */
3986 
3987 static void
3988 iflib_if_qflush(if_t ifp)
3989 {
3990 	if_ctx_t ctx = if_getsoftc(ifp);
3991 	iflib_txq_t txq = ctx->ifc_txqs;
3992 	int i;
3993 
3994 	STATE_LOCK(ctx);
3995 	ctx->ifc_flags |= IFC_QFLUSH;
3996 	STATE_UNLOCK(ctx);
3997 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
3998 		while (!(ifmp_ring_is_idle(txq->ift_br) || ifmp_ring_is_stalled(txq->ift_br)))
3999 			iflib_txq_check_drain(txq, 0);
4000 	STATE_LOCK(ctx);
4001 	ctx->ifc_flags &= ~IFC_QFLUSH;
4002 	STATE_UNLOCK(ctx);
4003 
4004 	/*
4005 	 * When ALTQ is enabled, this will also take care of purging the
4006 	 * ALTQ queue(s).
4007 	 */
4008 	if_qflush(ifp);
4009 }
4010 
4011 
4012 #define IFCAP_FLAGS (IFCAP_HWCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \
4013 		     IFCAP_TSO | IFCAP_VLAN_HWTAGGING | IFCAP_HWSTATS | \
4014 		     IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | \
4015 		     IFCAP_VLAN_HWTSO | IFCAP_VLAN_HWCSUM)
4016 
4017 static int
4018 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data)
4019 {
4020 	if_ctx_t ctx = if_getsoftc(ifp);
4021 	struct ifreq	*ifr = (struct ifreq *)data;
4022 #if defined(INET) || defined(INET6)
4023 	struct ifaddr	*ifa = (struct ifaddr *)data;
4024 #endif
4025 	bool		avoid_reset = FALSE;
4026 	int		err = 0, reinit = 0, bits;
4027 
4028 	switch (command) {
4029 	case SIOCSIFADDR:
4030 #ifdef INET
4031 		if (ifa->ifa_addr->sa_family == AF_INET)
4032 			avoid_reset = TRUE;
4033 #endif
4034 #ifdef INET6
4035 		if (ifa->ifa_addr->sa_family == AF_INET6)
4036 			avoid_reset = TRUE;
4037 #endif
4038 		/*
4039 		** Calling init results in link renegotiation,
4040 		** so we avoid doing it when possible.
4041 		*/
4042 		if (avoid_reset) {
4043 			if_setflagbits(ifp, IFF_UP,0);
4044 			if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
4045 				reinit = 1;
4046 #ifdef INET
4047 			if (!(if_getflags(ifp) & IFF_NOARP))
4048 				arp_ifinit(ifp, ifa);
4049 #endif
4050 		} else
4051 			err = ether_ioctl(ifp, command, data);
4052 		break;
4053 	case SIOCSIFMTU:
4054 		CTX_LOCK(ctx);
4055 		if (ifr->ifr_mtu == if_getmtu(ifp)) {
4056 			CTX_UNLOCK(ctx);
4057 			break;
4058 		}
4059 		bits = if_getdrvflags(ifp);
4060 		/* stop the driver and free any clusters before proceeding */
4061 		iflib_stop(ctx);
4062 
4063 		if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) {
4064 			STATE_LOCK(ctx);
4065 			if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size)
4066 				ctx->ifc_flags |= IFC_MULTISEG;
4067 			else
4068 				ctx->ifc_flags &= ~IFC_MULTISEG;
4069 			STATE_UNLOCK(ctx);
4070 			err = if_setmtu(ifp, ifr->ifr_mtu);
4071 		}
4072 		iflib_init_locked(ctx);
4073 		STATE_LOCK(ctx);
4074 		if_setdrvflags(ifp, bits);
4075 		STATE_UNLOCK(ctx);
4076 		CTX_UNLOCK(ctx);
4077 		break;
4078 	case SIOCSIFFLAGS:
4079 		CTX_LOCK(ctx);
4080 		if (if_getflags(ifp) & IFF_UP) {
4081 			if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4082 				if ((if_getflags(ifp) ^ ctx->ifc_if_flags) &
4083 				    (IFF_PROMISC | IFF_ALLMULTI)) {
4084 					err = IFDI_PROMISC_SET(ctx, if_getflags(ifp));
4085 				}
4086 			} else
4087 				reinit = 1;
4088 		} else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4089 			iflib_stop(ctx);
4090 		}
4091 		ctx->ifc_if_flags = if_getflags(ifp);
4092 		CTX_UNLOCK(ctx);
4093 		break;
4094 	case SIOCADDMULTI:
4095 	case SIOCDELMULTI:
4096 		if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4097 			CTX_LOCK(ctx);
4098 			IFDI_INTR_DISABLE(ctx);
4099 			IFDI_MULTI_SET(ctx);
4100 			IFDI_INTR_ENABLE(ctx);
4101 			CTX_UNLOCK(ctx);
4102 		}
4103 		break;
4104 	case SIOCSIFMEDIA:
4105 		CTX_LOCK(ctx);
4106 		IFDI_MEDIA_SET(ctx);
4107 		CTX_UNLOCK(ctx);
4108 		/* falls thru */
4109 	case SIOCGIFMEDIA:
4110 	case SIOCGIFXMEDIA:
4111 		err = ifmedia_ioctl(ifp, ifr, &ctx->ifc_media, command);
4112 		break;
4113 	case SIOCGI2C:
4114 	{
4115 		struct ifi2creq i2c;
4116 
4117 		err = copyin(ifr_data_get_ptr(ifr), &i2c, sizeof(i2c));
4118 		if (err != 0)
4119 			break;
4120 		if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) {
4121 			err = EINVAL;
4122 			break;
4123 		}
4124 		if (i2c.len > sizeof(i2c.data)) {
4125 			err = EINVAL;
4126 			break;
4127 		}
4128 
4129 		if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0)
4130 			err = copyout(&i2c, ifr_data_get_ptr(ifr),
4131 			    sizeof(i2c));
4132 		break;
4133 	}
4134 	case SIOCSIFCAP:
4135 	{
4136 		int mask, setmask, oldmask;
4137 
4138 		oldmask = if_getcapenable(ifp);
4139 		mask = ifr->ifr_reqcap ^ oldmask;
4140 		mask &= ctx->ifc_softc_ctx.isc_capabilities;
4141 		setmask = 0;
4142 #ifdef TCP_OFFLOAD
4143 		setmask |= mask & (IFCAP_TOE4|IFCAP_TOE6);
4144 #endif
4145 		setmask |= (mask & IFCAP_FLAGS);
4146 		setmask |= (mask & IFCAP_WOL);
4147 
4148 		/*
4149 		 * If any RX csum has changed, change all the ones that
4150 		 * are supported by the driver.
4151 		 */
4152 		if (setmask & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6)) {
4153 			setmask |= ctx->ifc_softc_ctx.isc_capabilities &
4154 			    (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6);
4155 		}
4156 
4157 		/*
4158 		 * want to ensure that traffic has stopped before we change any of the flags
4159 		 */
4160 		if (setmask) {
4161 			CTX_LOCK(ctx);
4162 			bits = if_getdrvflags(ifp);
4163 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4164 				iflib_stop(ctx);
4165 			STATE_LOCK(ctx);
4166 			if_togglecapenable(ifp, setmask);
4167 			STATE_UNLOCK(ctx);
4168 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4169 				iflib_init_locked(ctx);
4170 			STATE_LOCK(ctx);
4171 			if_setdrvflags(ifp, bits);
4172 			STATE_UNLOCK(ctx);
4173 			CTX_UNLOCK(ctx);
4174 		}
4175 		if_vlancap(ifp);
4176 		break;
4177 	}
4178 	case SIOCGPRIVATE_0:
4179 	case SIOCSDRVSPEC:
4180 	case SIOCGDRVSPEC:
4181 		CTX_LOCK(ctx);
4182 		err = IFDI_PRIV_IOCTL(ctx, command, data);
4183 		CTX_UNLOCK(ctx);
4184 		break;
4185 	default:
4186 		err = ether_ioctl(ifp, command, data);
4187 		break;
4188 	}
4189 	if (reinit)
4190 		iflib_if_init(ctx);
4191 	return (err);
4192 }
4193 
4194 static uint64_t
4195 iflib_if_get_counter(if_t ifp, ift_counter cnt)
4196 {
4197 	if_ctx_t ctx = if_getsoftc(ifp);
4198 
4199 	return (IFDI_GET_COUNTER(ctx, cnt));
4200 }
4201 
4202 /*********************************************************************
4203  *
4204  *  OTHER FUNCTIONS EXPORTED TO THE STACK
4205  *
4206  **********************************************************************/
4207 
4208 static void
4209 iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag)
4210 {
4211 	if_ctx_t ctx = if_getsoftc(ifp);
4212 
4213 	if ((void *)ctx != arg)
4214 		return;
4215 
4216 	if ((vtag == 0) || (vtag > 4095))
4217 		return;
4218 
4219 	CTX_LOCK(ctx);
4220 	IFDI_VLAN_REGISTER(ctx, vtag);
4221 	/* Re-init to load the changes */
4222 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
4223 		iflib_if_init_locked(ctx);
4224 	CTX_UNLOCK(ctx);
4225 }
4226 
4227 static void
4228 iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag)
4229 {
4230 	if_ctx_t ctx = if_getsoftc(ifp);
4231 
4232 	if ((void *)ctx != arg)
4233 		return;
4234 
4235 	if ((vtag == 0) || (vtag > 4095))
4236 		return;
4237 
4238 	CTX_LOCK(ctx);
4239 	IFDI_VLAN_UNREGISTER(ctx, vtag);
4240 	/* Re-init to load the changes */
4241 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
4242 		iflib_if_init_locked(ctx);
4243 	CTX_UNLOCK(ctx);
4244 }
4245 
4246 static void
4247 iflib_led_func(void *arg, int onoff)
4248 {
4249 	if_ctx_t ctx = arg;
4250 
4251 	CTX_LOCK(ctx);
4252 	IFDI_LED_FUNC(ctx, onoff);
4253 	CTX_UNLOCK(ctx);
4254 }
4255 
4256 /*********************************************************************
4257  *
4258  *  BUS FUNCTION DEFINITIONS
4259  *
4260  **********************************************************************/
4261 
4262 int
4263 iflib_device_probe(device_t dev)
4264 {
4265 	pci_vendor_info_t *ent;
4266 
4267 	uint16_t	pci_vendor_id, pci_device_id;
4268 	uint16_t	pci_subvendor_id, pci_subdevice_id;
4269 	uint16_t	pci_rev_id;
4270 	if_shared_ctx_t sctx;
4271 
4272 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4273 		return (ENOTSUP);
4274 
4275 	pci_vendor_id = pci_get_vendor(dev);
4276 	pci_device_id = pci_get_device(dev);
4277 	pci_subvendor_id = pci_get_subvendor(dev);
4278 	pci_subdevice_id = pci_get_subdevice(dev);
4279 	pci_rev_id = pci_get_revid(dev);
4280 	if (sctx->isc_parse_devinfo != NULL)
4281 		sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id);
4282 
4283 	ent = sctx->isc_vendor_info;
4284 	while (ent->pvi_vendor_id != 0) {
4285 		if (pci_vendor_id != ent->pvi_vendor_id) {
4286 			ent++;
4287 			continue;
4288 		}
4289 		if ((pci_device_id == ent->pvi_device_id) &&
4290 		    ((pci_subvendor_id == ent->pvi_subvendor_id) ||
4291 		     (ent->pvi_subvendor_id == 0)) &&
4292 		    ((pci_subdevice_id == ent->pvi_subdevice_id) ||
4293 		     (ent->pvi_subdevice_id == 0)) &&
4294 		    ((pci_rev_id == ent->pvi_rev_id) ||
4295 		     (ent->pvi_rev_id == 0))) {
4296 
4297 			device_set_desc_copy(dev, ent->pvi_name);
4298 			/* this needs to be changed to zero if the bus probing code
4299 			 * ever stops re-probing on best match because the sctx
4300 			 * may have its values over written by register calls
4301 			 * in subsequent probes
4302 			 */
4303 			return (BUS_PROBE_DEFAULT);
4304 		}
4305 		ent++;
4306 	}
4307 	return (ENXIO);
4308 }
4309 
4310 static void
4311 iflib_reset_qvalues(if_ctx_t ctx)
4312 {
4313 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4314 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4315 	device_t dev = ctx->ifc_dev;
4316 	int i;
4317 
4318 	scctx->isc_txrx_budget_bytes_max = IFLIB_MAX_TX_BYTES;
4319 	scctx->isc_tx_qdepth = IFLIB_DEFAULT_TX_QDEPTH;
4320 	/*
4321 	 * XXX sanity check that ntxd & nrxd are a power of 2
4322 	 */
4323 	if (ctx->ifc_sysctl_ntxqs != 0)
4324 		scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs;
4325 	if (ctx->ifc_sysctl_nrxqs != 0)
4326 		scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs;
4327 
4328 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4329 		if (ctx->ifc_sysctl_ntxds[i] != 0)
4330 			scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i];
4331 		else
4332 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4333 	}
4334 
4335 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4336 		if (ctx->ifc_sysctl_nrxds[i] != 0)
4337 			scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i];
4338 		else
4339 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4340 	}
4341 
4342 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4343 		if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) {
4344 			device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n",
4345 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]);
4346 			scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i];
4347 		}
4348 		if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) {
4349 			device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n",
4350 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]);
4351 			scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i];
4352 		}
4353 	}
4354 
4355 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4356 		if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) {
4357 			device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n",
4358 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]);
4359 			scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i];
4360 		}
4361 		if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) {
4362 			device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n",
4363 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]);
4364 			scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i];
4365 		}
4366 	}
4367 }
4368 
4369 int
4370 iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp)
4371 {
4372 	int err, rid, msix;
4373 	if_ctx_t ctx;
4374 	if_t ifp;
4375 	if_softc_ctx_t scctx;
4376 	int i;
4377 	uint16_t main_txq;
4378 	uint16_t main_rxq;
4379 
4380 
4381 	ctx = malloc(sizeof(* ctx), M_IFLIB, M_WAITOK|M_ZERO);
4382 
4383 	if (sc == NULL) {
4384 		sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4385 		device_set_softc(dev, ctx);
4386 		ctx->ifc_flags |= IFC_SC_ALLOCATED;
4387 	}
4388 
4389 	ctx->ifc_sctx = sctx;
4390 	ctx->ifc_dev = dev;
4391 	ctx->ifc_softc = sc;
4392 
4393 	if ((err = iflib_register(ctx)) != 0) {
4394 		device_printf(dev, "iflib_register failed %d\n", err);
4395 		goto fail_ctx_free;
4396 	}
4397 	iflib_add_device_sysctl_pre(ctx);
4398 
4399 	scctx = &ctx->ifc_softc_ctx;
4400 	ifp = ctx->ifc_ifp;
4401 
4402 	iflib_reset_qvalues(ctx);
4403 	CTX_LOCK(ctx);
4404 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4405 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4406 		goto fail_unlock;
4407 	}
4408 	_iflib_pre_assert(scctx);
4409 	ctx->ifc_txrx = *scctx->isc_txrx;
4410 
4411 #ifdef INVARIANTS
4412 	MPASS(scctx->isc_capabilities);
4413 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
4414 		MPASS(scctx->isc_tx_csum_flags);
4415 #endif
4416 
4417 	if_setcapabilities(ifp, scctx->isc_capabilities | IFCAP_HWSTATS);
4418 	if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS);
4419 
4420 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4421 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4422 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4423 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4424 
4425 	main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4426 	main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4427 
4428 	/* XXX change for per-queue sizes */
4429 	device_printf(dev, "Using %d tx descriptors and %d rx descriptors\n",
4430 	    scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4431 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4432 		if (!powerof2(scctx->isc_nrxd[i])) {
4433 			/* round down instead? */
4434 			device_printf(dev, "# rx descriptors must be a power of 2\n");
4435 			err = EINVAL;
4436 			goto fail_iflib_detach;
4437 		}
4438 	}
4439 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4440 		if (!powerof2(scctx->isc_ntxd[i])) {
4441 			device_printf(dev,
4442 			    "# tx descriptors must be a power of 2");
4443 			err = EINVAL;
4444 			goto fail_iflib_detach;
4445 		}
4446 	}
4447 
4448 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4449 	    MAX_SINGLE_PACKET_FRACTION)
4450 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4451 		    MAX_SINGLE_PACKET_FRACTION);
4452 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4453 	    MAX_SINGLE_PACKET_FRACTION)
4454 		scctx->isc_tx_tso_segments_max = max(1,
4455 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4456 
4457 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4458 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
4459 		/*
4460 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
4461 		 * but some MACs do.
4462 		 */
4463 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
4464 		    IP_MAXPACKET));
4465 		/*
4466 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
4467 		 * into account.  In the worst case, each of these calls will
4468 		 * add another mbuf and, thus, the requirement for another DMA
4469 		 * segment.  So for best performance, it doesn't make sense to
4470 		 * advertize a maximum of TSO segments that typically will
4471 		 * require defragmentation in iflib_encap().
4472 		 */
4473 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
4474 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
4475 	}
4476 	if (scctx->isc_rss_table_size == 0)
4477 		scctx->isc_rss_table_size = 64;
4478 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4479 
4480 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4481 	/* XXX format name */
4482 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx,
4483 	    NULL, NULL, "admin");
4484 
4485 	/* Set up cpu set.  If it fails, use the set of all CPUs. */
4486 	if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) != 0) {
4487 		device_printf(dev, "Unable to fetch CPU list\n");
4488 		CPU_COPY(&all_cpus, &ctx->ifc_cpus);
4489 	}
4490 	MPASS(CPU_COUNT(&ctx->ifc_cpus) > 0);
4491 
4492 	/*
4493 	** Now set up MSI or MSI-X, should return us the number of supported
4494 	** vectors (will be 1 for a legacy interrupt and MSI).
4495 	*/
4496 	if (sctx->isc_flags & IFLIB_SKIP_MSIX) {
4497 		msix = scctx->isc_vectors;
4498 	} else if (scctx->isc_msix_bar != 0)
4499 	       /*
4500 		* The simple fact that isc_msix_bar is not 0 does not mean we
4501 		* we have a good value there that is known to work.
4502 		*/
4503 		msix = iflib_msix_init(ctx);
4504 	else {
4505 		scctx->isc_vectors = 1;
4506 		scctx->isc_ntxqsets = 1;
4507 		scctx->isc_nrxqsets = 1;
4508 		scctx->isc_intr = IFLIB_INTR_LEGACY;
4509 		msix = 0;
4510 	}
4511 	/* Get memory for the station queues */
4512 	if ((err = iflib_queues_alloc(ctx))) {
4513 		device_printf(dev, "Unable to allocate queue memory\n");
4514 		goto fail_intr_free;
4515 	}
4516 
4517 	if ((err = iflib_qset_structures_setup(ctx)))
4518 		goto fail_queues;
4519 
4520 	/*
4521 	 * Group taskqueues aren't properly set up until SMP is started,
4522 	 * so we disable interrupts until we can handle them post
4523 	 * SI_SUB_SMP.
4524 	 *
4525 	 * XXX: disabling interrupts doesn't actually work, at least for
4526 	 * the non-MSI case.  When they occur before SI_SUB_SMP completes,
4527 	 * we do null handling and depend on this not causing too large an
4528 	 * interrupt storm.
4529 	 */
4530 	IFDI_INTR_DISABLE(ctx);
4531 	if (msix > 1 && (err = IFDI_MSIX_INTR_ASSIGN(ctx, msix)) != 0) {
4532 		device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n", err);
4533 		goto fail_queues;
4534 	}
4535 	if (msix <= 1) {
4536 		rid = 0;
4537 		if (scctx->isc_intr == IFLIB_INTR_MSI) {
4538 			MPASS(msix == 1);
4539 			rid = 1;
4540 		}
4541 		if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) {
4542 			device_printf(dev, "iflib_legacy_setup failed %d\n", err);
4543 			goto fail_queues;
4544 		}
4545 	}
4546 
4547 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
4548 
4549 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4550 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4551 		goto fail_detach;
4552 	}
4553 
4554 	/*
4555 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4556 	 * This must appear after the call to ether_ifattach() because
4557 	 * ether_ifattach() sets if_hdrlen to the default value.
4558 	 */
4559 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4560 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
4561 
4562 	if ((err = iflib_netmap_attach(ctx))) {
4563 		device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err);
4564 		goto fail_detach;
4565 	}
4566 	*ctxp = ctx;
4567 
4568 	NETDUMP_SET(ctx->ifc_ifp, iflib);
4569 
4570 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4571 	iflib_add_device_sysctl_post(ctx);
4572 	ctx->ifc_flags |= IFC_INIT_DONE;
4573 	CTX_UNLOCK(ctx);
4574 	return (0);
4575 
4576 fail_detach:
4577 	ether_ifdetach(ctx->ifc_ifp);
4578 fail_intr_free:
4579 	iflib_free_intr_mem(ctx);
4580 fail_queues:
4581 	iflib_tx_structures_free(ctx);
4582 	iflib_rx_structures_free(ctx);
4583 fail_iflib_detach:
4584 	IFDI_DETACH(ctx);
4585 fail_unlock:
4586 	CTX_UNLOCK(ctx);
4587 fail_ctx_free:
4588         if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4589                 free(ctx->ifc_softc, M_IFLIB);
4590         free(ctx, M_IFLIB);
4591 	return (err);
4592 }
4593 
4594 int
4595 iflib_pseudo_register(device_t dev, if_shared_ctx_t sctx, if_ctx_t *ctxp,
4596 					  struct iflib_cloneattach_ctx *clctx)
4597 {
4598 	int err;
4599 	if_ctx_t ctx;
4600 	if_t ifp;
4601 	if_softc_ctx_t scctx;
4602 	int i;
4603 	void *sc;
4604 	uint16_t main_txq;
4605 	uint16_t main_rxq;
4606 
4607 	ctx = malloc(sizeof(*ctx), M_IFLIB, M_WAITOK|M_ZERO);
4608 	sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4609 	ctx->ifc_flags |= IFC_SC_ALLOCATED;
4610 	if (sctx->isc_flags & (IFLIB_PSEUDO|IFLIB_VIRTUAL))
4611 		ctx->ifc_flags |= IFC_PSEUDO;
4612 
4613 	ctx->ifc_sctx = sctx;
4614 	ctx->ifc_softc = sc;
4615 	ctx->ifc_dev = dev;
4616 
4617 	if ((err = iflib_register(ctx)) != 0) {
4618 		device_printf(dev, "%s: iflib_register failed %d\n", __func__, err);
4619 		goto fail_ctx_free;
4620 	}
4621 	iflib_add_device_sysctl_pre(ctx);
4622 
4623 	scctx = &ctx->ifc_softc_ctx;
4624 	ifp = ctx->ifc_ifp;
4625 
4626 	/*
4627 	 * XXX sanity check that ntxd & nrxd are a power of 2
4628 	 */
4629 	iflib_reset_qvalues(ctx);
4630 	CTX_LOCK(ctx);
4631 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4632 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4633 		goto fail_unlock;
4634 	}
4635 	if (sctx->isc_flags & IFLIB_GEN_MAC)
4636 		ether_gen_addr(ifp, &ctx->ifc_mac);
4637 	if ((err = IFDI_CLONEATTACH(ctx, clctx->cc_ifc, clctx->cc_name,
4638 								clctx->cc_params)) != 0) {
4639 		device_printf(dev, "IFDI_CLONEATTACH failed %d\n", err);
4640 		goto fail_ctx_free;
4641 	}
4642 	ifmedia_add(&ctx->ifc_media, IFM_ETHER | IFM_1000_T | IFM_FDX, 0, NULL);
4643 	ifmedia_add(&ctx->ifc_media, IFM_ETHER | IFM_AUTO, 0, NULL);
4644 	ifmedia_set(&ctx->ifc_media, IFM_ETHER | IFM_AUTO);
4645 
4646 #ifdef INVARIANTS
4647 	MPASS(scctx->isc_capabilities);
4648 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
4649 		MPASS(scctx->isc_tx_csum_flags);
4650 #endif
4651 
4652 	if_setcapabilities(ifp, scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_LINKSTATE);
4653 	if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_LINKSTATE);
4654 
4655 	ifp->if_flags |= IFF_NOGROUP;
4656 	if (sctx->isc_flags & IFLIB_PSEUDO) {
4657 		ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
4658 
4659 		if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4660 			device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4661 			goto fail_detach;
4662 		}
4663 		*ctxp = ctx;
4664 
4665 		/*
4666 		 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4667 		 * This must appear after the call to ether_ifattach() because
4668 		 * ether_ifattach() sets if_hdrlen to the default value.
4669 		 */
4670 		if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4671 			if_setifheaderlen(ifp,
4672 			    sizeof(struct ether_vlan_header));
4673 
4674 		if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4675 		iflib_add_device_sysctl_post(ctx);
4676 		ctx->ifc_flags |= IFC_INIT_DONE;
4677 		return (0);
4678 	}
4679 	_iflib_pre_assert(scctx);
4680 	ctx->ifc_txrx = *scctx->isc_txrx;
4681 
4682 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4683 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4684 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4685 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4686 
4687 	main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4688 	main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4689 
4690 	/* XXX change for per-queue sizes */
4691 	device_printf(dev, "Using %d tx descriptors and %d rx descriptors\n",
4692 	    scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4693 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4694 		if (!powerof2(scctx->isc_nrxd[i])) {
4695 			/* round down instead? */
4696 			device_printf(dev, "# rx descriptors must be a power of 2\n");
4697 			err = EINVAL;
4698 			goto fail_iflib_detach;
4699 		}
4700 	}
4701 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4702 		if (!powerof2(scctx->isc_ntxd[i])) {
4703 			device_printf(dev,
4704 			    "# tx descriptors must be a power of 2");
4705 			err = EINVAL;
4706 			goto fail_iflib_detach;
4707 		}
4708 	}
4709 
4710 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4711 	    MAX_SINGLE_PACKET_FRACTION)
4712 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4713 		    MAX_SINGLE_PACKET_FRACTION);
4714 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4715 	    MAX_SINGLE_PACKET_FRACTION)
4716 		scctx->isc_tx_tso_segments_max = max(1,
4717 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4718 
4719 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4720 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
4721 		/*
4722 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
4723 		 * but some MACs do.
4724 		 */
4725 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
4726 		    IP_MAXPACKET));
4727 		/*
4728 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
4729 		 * into account.  In the worst case, each of these calls will
4730 		 * add another mbuf and, thus, the requirement for another DMA
4731 		 * segment.  So for best performance, it doesn't make sense to
4732 		 * advertize a maximum of TSO segments that typically will
4733 		 * require defragmentation in iflib_encap().
4734 		 */
4735 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
4736 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
4737 	}
4738 	if (scctx->isc_rss_table_size == 0)
4739 		scctx->isc_rss_table_size = 64;
4740 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4741 
4742 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4743 	/* XXX format name */
4744 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx,
4745 	    NULL, NULL, "admin");
4746 
4747 	/* XXX --- can support > 1 -- but keep it simple for now */
4748 	scctx->isc_intr = IFLIB_INTR_LEGACY;
4749 
4750 	/* Get memory for the station queues */
4751 	if ((err = iflib_queues_alloc(ctx))) {
4752 		device_printf(dev, "Unable to allocate queue memory\n");
4753 		goto fail_iflib_detach;
4754 	}
4755 
4756 	if ((err = iflib_qset_structures_setup(ctx))) {
4757 		device_printf(dev, "qset structure setup failed %d\n", err);
4758 		goto fail_queues;
4759 	}
4760 
4761 	/*
4762 	 * XXX What if anything do we want to do about interrupts?
4763 	 */
4764 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
4765 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4766 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4767 		goto fail_detach;
4768 	}
4769 
4770 	/*
4771 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4772 	 * This must appear after the call to ether_ifattach() because
4773 	 * ether_ifattach() sets if_hdrlen to the default value.
4774 	 */
4775 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4776 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
4777 
4778 	/* XXX handle more than one queue */
4779 	for (i = 0; i < scctx->isc_nrxqsets; i++)
4780 		IFDI_RX_CLSET(ctx, 0, i, ctx->ifc_rxqs[i].ifr_fl[0].ifl_sds.ifsd_cl);
4781 
4782 	*ctxp = ctx;
4783 
4784 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4785 	iflib_add_device_sysctl_post(ctx);
4786 	ctx->ifc_flags |= IFC_INIT_DONE;
4787 	CTX_UNLOCK(ctx);
4788 	return (0);
4789 fail_detach:
4790 	ether_ifdetach(ctx->ifc_ifp);
4791 fail_queues:
4792 	iflib_tx_structures_free(ctx);
4793 	iflib_rx_structures_free(ctx);
4794 fail_iflib_detach:
4795 	IFDI_DETACH(ctx);
4796 fail_unlock:
4797 	CTX_UNLOCK(ctx);
4798 fail_ctx_free:
4799 	free(ctx->ifc_softc, M_IFLIB);
4800 	free(ctx, M_IFLIB);
4801 	return (err);
4802 }
4803 
4804 int
4805 iflib_pseudo_deregister(if_ctx_t ctx)
4806 {
4807 	if_t ifp = ctx->ifc_ifp;
4808 	iflib_txq_t txq;
4809 	iflib_rxq_t rxq;
4810 	int i, j;
4811 	struct taskqgroup *tqg;
4812 	iflib_fl_t fl;
4813 
4814 	/* Unregister VLAN events */
4815 	if (ctx->ifc_vlan_attach_event != NULL)
4816 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
4817 	if (ctx->ifc_vlan_detach_event != NULL)
4818 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
4819 
4820 	ether_ifdetach(ifp);
4821 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
4822 	CTX_LOCK_DESTROY(ctx);
4823 	/* XXX drain any dependent tasks */
4824 	tqg = qgroup_if_io_tqg;
4825 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
4826 		callout_drain(&txq->ift_timer);
4827 		if (txq->ift_task.gt_uniq != NULL)
4828 			taskqgroup_detach(tqg, &txq->ift_task);
4829 	}
4830 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4831 		if (rxq->ifr_task.gt_uniq != NULL)
4832 			taskqgroup_detach(tqg, &rxq->ifr_task);
4833 
4834 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
4835 			free(fl->ifl_rx_bitmap, M_IFLIB);
4836 	}
4837 	tqg = qgroup_if_config_tqg;
4838 	if (ctx->ifc_admin_task.gt_uniq != NULL)
4839 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
4840 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
4841 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
4842 
4843 	if_free(ifp);
4844 
4845 	iflib_tx_structures_free(ctx);
4846 	iflib_rx_structures_free(ctx);
4847 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4848 		free(ctx->ifc_softc, M_IFLIB);
4849 	free(ctx, M_IFLIB);
4850 	return (0);
4851 }
4852 
4853 int
4854 iflib_device_attach(device_t dev)
4855 {
4856 	if_ctx_t ctx;
4857 	if_shared_ctx_t sctx;
4858 
4859 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4860 		return (ENOTSUP);
4861 
4862 	pci_enable_busmaster(dev);
4863 
4864 	return (iflib_device_register(dev, NULL, sctx, &ctx));
4865 }
4866 
4867 int
4868 iflib_device_deregister(if_ctx_t ctx)
4869 {
4870 	if_t ifp = ctx->ifc_ifp;
4871 	iflib_txq_t txq;
4872 	iflib_rxq_t rxq;
4873 	device_t dev = ctx->ifc_dev;
4874 	int i, j;
4875 	struct taskqgroup *tqg;
4876 	iflib_fl_t fl;
4877 
4878 	/* Make sure VLANS are not using driver */
4879 	if (if_vlantrunkinuse(ifp)) {
4880 		device_printf(dev, "Vlan in use, detach first\n");
4881 		return (EBUSY);
4882 	}
4883 #ifdef PCI_IOV
4884 	if (!CTX_IS_VF(ctx) && pci_iov_detach(dev) != 0) {
4885 		device_printf(dev, "SR-IOV in use; detach first.\n");
4886 		return (EBUSY);
4887 	}
4888 #endif
4889 
4890 	STATE_LOCK(ctx);
4891 	ctx->ifc_flags |= IFC_IN_DETACH;
4892 	STATE_UNLOCK(ctx);
4893 
4894 	CTX_LOCK(ctx);
4895 	iflib_stop(ctx);
4896 	CTX_UNLOCK(ctx);
4897 
4898 	/* Unregister VLAN events */
4899 	if (ctx->ifc_vlan_attach_event != NULL)
4900 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
4901 	if (ctx->ifc_vlan_detach_event != NULL)
4902 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
4903 
4904 	iflib_netmap_detach(ifp);
4905 	ether_ifdetach(ifp);
4906 	if (ctx->ifc_led_dev != NULL)
4907 		led_destroy(ctx->ifc_led_dev);
4908 	/* XXX drain any dependent tasks */
4909 	tqg = qgroup_if_io_tqg;
4910 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
4911 		callout_drain(&txq->ift_timer);
4912 		if (txq->ift_task.gt_uniq != NULL)
4913 			taskqgroup_detach(tqg, &txq->ift_task);
4914 	}
4915 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4916 		if (rxq->ifr_task.gt_uniq != NULL)
4917 			taskqgroup_detach(tqg, &rxq->ifr_task);
4918 
4919 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
4920 			free(fl->ifl_rx_bitmap, M_IFLIB);
4921 	}
4922 	tqg = qgroup_if_config_tqg;
4923 	if (ctx->ifc_admin_task.gt_uniq != NULL)
4924 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
4925 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
4926 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
4927 	CTX_LOCK(ctx);
4928 	IFDI_DETACH(ctx);
4929 	CTX_UNLOCK(ctx);
4930 
4931 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
4932 	CTX_LOCK_DESTROY(ctx);
4933 	device_set_softc(ctx->ifc_dev, NULL);
4934 	iflib_free_intr_mem(ctx);
4935 
4936 	bus_generic_detach(dev);
4937 	if_free(ifp);
4938 
4939 	iflib_tx_structures_free(ctx);
4940 	iflib_rx_structures_free(ctx);
4941 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4942 		free(ctx->ifc_softc, M_IFLIB);
4943 	STATE_LOCK_DESTROY(ctx);
4944 	free(ctx, M_IFLIB);
4945 	return (0);
4946 }
4947 
4948 static void
4949 iflib_free_intr_mem(if_ctx_t ctx)
4950 {
4951 
4952 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) {
4953 		iflib_irq_free(ctx, &ctx->ifc_legacy_irq);
4954 	}
4955 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) {
4956 		pci_release_msi(ctx->ifc_dev);
4957 	}
4958 	if (ctx->ifc_msix_mem != NULL) {
4959 		bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY,
4960 		    rman_get_rid(ctx->ifc_msix_mem), ctx->ifc_msix_mem);
4961 		ctx->ifc_msix_mem = NULL;
4962 	}
4963 }
4964 
4965 int
4966 iflib_device_detach(device_t dev)
4967 {
4968 	if_ctx_t ctx = device_get_softc(dev);
4969 
4970 	return (iflib_device_deregister(ctx));
4971 }
4972 
4973 int
4974 iflib_device_suspend(device_t dev)
4975 {
4976 	if_ctx_t ctx = device_get_softc(dev);
4977 
4978 	CTX_LOCK(ctx);
4979 	IFDI_SUSPEND(ctx);
4980 	CTX_UNLOCK(ctx);
4981 
4982 	return bus_generic_suspend(dev);
4983 }
4984 int
4985 iflib_device_shutdown(device_t dev)
4986 {
4987 	if_ctx_t ctx = device_get_softc(dev);
4988 
4989 	CTX_LOCK(ctx);
4990 	IFDI_SHUTDOWN(ctx);
4991 	CTX_UNLOCK(ctx);
4992 
4993 	return bus_generic_suspend(dev);
4994 }
4995 
4996 
4997 int
4998 iflib_device_resume(device_t dev)
4999 {
5000 	if_ctx_t ctx = device_get_softc(dev);
5001 	iflib_txq_t txq = ctx->ifc_txqs;
5002 
5003 	CTX_LOCK(ctx);
5004 	IFDI_RESUME(ctx);
5005 	iflib_if_init_locked(ctx);
5006 	CTX_UNLOCK(ctx);
5007 	for (int i = 0; i < NTXQSETS(ctx); i++, txq++)
5008 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
5009 
5010 	return (bus_generic_resume(dev));
5011 }
5012 
5013 int
5014 iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params)
5015 {
5016 	int error;
5017 	if_ctx_t ctx = device_get_softc(dev);
5018 
5019 	CTX_LOCK(ctx);
5020 	error = IFDI_IOV_INIT(ctx, num_vfs, params);
5021 	CTX_UNLOCK(ctx);
5022 
5023 	return (error);
5024 }
5025 
5026 void
5027 iflib_device_iov_uninit(device_t dev)
5028 {
5029 	if_ctx_t ctx = device_get_softc(dev);
5030 
5031 	CTX_LOCK(ctx);
5032 	IFDI_IOV_UNINIT(ctx);
5033 	CTX_UNLOCK(ctx);
5034 }
5035 
5036 int
5037 iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params)
5038 {
5039 	int error;
5040 	if_ctx_t ctx = device_get_softc(dev);
5041 
5042 	CTX_LOCK(ctx);
5043 	error = IFDI_IOV_VF_ADD(ctx, vfnum, params);
5044 	CTX_UNLOCK(ctx);
5045 
5046 	return (error);
5047 }
5048 
5049 /*********************************************************************
5050  *
5051  *  MODULE FUNCTION DEFINITIONS
5052  *
5053  **********************************************************************/
5054 
5055 /*
5056  * - Start a fast taskqueue thread for each core
5057  * - Start a taskqueue for control operations
5058  */
5059 static int
5060 iflib_module_init(void)
5061 {
5062 	return (0);
5063 }
5064 
5065 static int
5066 iflib_module_event_handler(module_t mod, int what, void *arg)
5067 {
5068 	int err;
5069 
5070 	switch (what) {
5071 	case MOD_LOAD:
5072 		if ((err = iflib_module_init()) != 0)
5073 			return (err);
5074 		break;
5075 	case MOD_UNLOAD:
5076 		return (EBUSY);
5077 	default:
5078 		return (EOPNOTSUPP);
5079 	}
5080 
5081 	return (0);
5082 }
5083 
5084 /*********************************************************************
5085  *
5086  *  PUBLIC FUNCTION DEFINITIONS
5087  *     ordered as in iflib.h
5088  *
5089  **********************************************************************/
5090 
5091 
5092 static void
5093 _iflib_assert(if_shared_ctx_t sctx)
5094 {
5095 	MPASS(sctx->isc_tx_maxsize);
5096 	MPASS(sctx->isc_tx_maxsegsize);
5097 
5098 	MPASS(sctx->isc_rx_maxsize);
5099 	MPASS(sctx->isc_rx_nsegments);
5100 	MPASS(sctx->isc_rx_maxsegsize);
5101 
5102 	MPASS(sctx->isc_nrxd_min[0]);
5103 	MPASS(sctx->isc_nrxd_max[0]);
5104 	MPASS(sctx->isc_nrxd_default[0]);
5105 	MPASS(sctx->isc_ntxd_min[0]);
5106 	MPASS(sctx->isc_ntxd_max[0]);
5107 	MPASS(sctx->isc_ntxd_default[0]);
5108 }
5109 
5110 static void
5111 _iflib_pre_assert(if_softc_ctx_t scctx)
5112 {
5113 
5114 	MPASS(scctx->isc_txrx->ift_txd_encap);
5115 	MPASS(scctx->isc_txrx->ift_txd_flush);
5116 	MPASS(scctx->isc_txrx->ift_txd_credits_update);
5117 	MPASS(scctx->isc_txrx->ift_rxd_available);
5118 	MPASS(scctx->isc_txrx->ift_rxd_pkt_get);
5119 	MPASS(scctx->isc_txrx->ift_rxd_refill);
5120 	MPASS(scctx->isc_txrx->ift_rxd_flush);
5121 }
5122 
5123 static int
5124 iflib_register(if_ctx_t ctx)
5125 {
5126 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5127 	driver_t *driver = sctx->isc_driver;
5128 	device_t dev = ctx->ifc_dev;
5129 	if_t ifp;
5130 
5131 	_iflib_assert(sctx);
5132 
5133 	CTX_LOCK_INIT(ctx);
5134 	STATE_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev));
5135 	ifp = ctx->ifc_ifp = if_alloc(IFT_ETHER);
5136 	if (ifp == NULL) {
5137 		device_printf(dev, "can not allocate ifnet structure\n");
5138 		return (ENOMEM);
5139 	}
5140 
5141 	/*
5142 	 * Initialize our context's device specific methods
5143 	 */
5144 	kobj_init((kobj_t) ctx, (kobj_class_t) driver);
5145 	kobj_class_compile((kobj_class_t) driver);
5146 	driver->refs++;
5147 
5148 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
5149 	if_setsoftc(ifp, ctx);
5150 	if_setdev(ifp, dev);
5151 	if_setinitfn(ifp, iflib_if_init);
5152 	if_setioctlfn(ifp, iflib_if_ioctl);
5153 #ifdef ALTQ
5154 	if_setstartfn(ifp, iflib_altq_if_start);
5155 	if_settransmitfn(ifp, iflib_altq_if_transmit);
5156 	if_setsendqready(ifp);
5157 #else
5158 	if_settransmitfn(ifp, iflib_if_transmit);
5159 #endif
5160 	if_setqflushfn(ifp, iflib_if_qflush);
5161 	if_setflags(ifp, IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST);
5162 
5163 	ctx->ifc_vlan_attach_event =
5164 		EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx,
5165 							  EVENTHANDLER_PRI_FIRST);
5166 	ctx->ifc_vlan_detach_event =
5167 		EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx,
5168 							  EVENTHANDLER_PRI_FIRST);
5169 
5170 	ifmedia_init(&ctx->ifc_media, IFM_IMASK,
5171 					 iflib_media_change, iflib_media_status);
5172 
5173 	return (0);
5174 }
5175 
5176 
5177 static int
5178 iflib_queues_alloc(if_ctx_t ctx)
5179 {
5180 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5181 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5182 	device_t dev = ctx->ifc_dev;
5183 	int nrxqsets = scctx->isc_nrxqsets;
5184 	int ntxqsets = scctx->isc_ntxqsets;
5185 	iflib_txq_t txq;
5186 	iflib_rxq_t rxq;
5187 	iflib_fl_t fl = NULL;
5188 	int i, j, cpu, err, txconf, rxconf;
5189 	iflib_dma_info_t ifdip;
5190 	uint32_t *rxqsizes = scctx->isc_rxqsizes;
5191 	uint32_t *txqsizes = scctx->isc_txqsizes;
5192 	uint8_t nrxqs = sctx->isc_nrxqs;
5193 	uint8_t ntxqs = sctx->isc_ntxqs;
5194 	int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1;
5195 	caddr_t *vaddrs;
5196 	uint64_t *paddrs;
5197 
5198 	KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1"));
5199 	KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1"));
5200 
5201 	/* Allocate the TX ring struct memory */
5202 	if (!(ctx->ifc_txqs =
5203 	    (iflib_txq_t) malloc(sizeof(struct iflib_txq) *
5204 	    ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5205 		device_printf(dev, "Unable to allocate TX ring memory\n");
5206 		err = ENOMEM;
5207 		goto fail;
5208 	}
5209 
5210 	/* Now allocate the RX */
5211 	if (!(ctx->ifc_rxqs =
5212 	    (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) *
5213 	    nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5214 		device_printf(dev, "Unable to allocate RX ring memory\n");
5215 		err = ENOMEM;
5216 		goto rx_fail;
5217 	}
5218 
5219 	txq = ctx->ifc_txqs;
5220 	rxq = ctx->ifc_rxqs;
5221 
5222 	/*
5223 	 * XXX handle allocation failure
5224 	 */
5225 	for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) {
5226 		/* Set up some basics */
5227 
5228 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs,
5229 		    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
5230 			device_printf(dev,
5231 			    "Unable to allocate TX DMA info memory\n");
5232 			err = ENOMEM;
5233 			goto err_tx_desc;
5234 		}
5235 		txq->ift_ifdi = ifdip;
5236 		for (j = 0; j < ntxqs; j++, ifdip++) {
5237 			if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, 0)) {
5238 				device_printf(dev,
5239 				    "Unable to allocate TX descriptors\n");
5240 				err = ENOMEM;
5241 				goto err_tx_desc;
5242 			}
5243 			txq->ift_txd_size[j] = scctx->isc_txd_size[j];
5244 			bzero((void *)ifdip->idi_vaddr, txqsizes[j]);
5245 		}
5246 		txq->ift_ctx = ctx;
5247 		txq->ift_id = i;
5248 		if (sctx->isc_flags & IFLIB_HAS_TXCQ) {
5249 			txq->ift_br_offset = 1;
5250 		} else {
5251 			txq->ift_br_offset = 0;
5252 		}
5253 		/* XXX fix this */
5254 		txq->ift_timer.c_cpu = cpu;
5255 
5256 		if (iflib_txsd_alloc(txq)) {
5257 			device_printf(dev, "Critical Failure setting up TX buffers\n");
5258 			err = ENOMEM;
5259 			goto err_tx_desc;
5260 		}
5261 
5262 		/* Initialize the TX lock */
5263 		snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:tx(%d):callout",
5264 		    device_get_nameunit(dev), txq->ift_id);
5265 		mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF);
5266 		callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0);
5267 
5268 		snprintf(txq->ift_db_mtx_name, MTX_NAME_LEN, "%s:tx(%d):db",
5269 			 device_get_nameunit(dev), txq->ift_id);
5270 
5271 		err = ifmp_ring_alloc(&txq->ift_br, 2048, txq, iflib_txq_drain,
5272 				      iflib_txq_can_drain, M_IFLIB, M_WAITOK);
5273 		if (err) {
5274 			/* XXX free any allocated rings */
5275 			device_printf(dev, "Unable to allocate buf_ring\n");
5276 			goto err_tx_desc;
5277 		}
5278 	}
5279 
5280 	for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) {
5281 		/* Set up some basics */
5282 
5283 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs,
5284 		   M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
5285 			device_printf(dev,
5286 			    "Unable to allocate RX DMA info memory\n");
5287 			err = ENOMEM;
5288 			goto err_tx_desc;
5289 		}
5290 
5291 		rxq->ifr_ifdi = ifdip;
5292 		/* XXX this needs to be changed if #rx queues != #tx queues */
5293 		rxq->ifr_ntxqirq = 1;
5294 		rxq->ifr_txqid[0] = i;
5295 		for (j = 0; j < nrxqs; j++, ifdip++) {
5296 			if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, 0)) {
5297 				device_printf(dev,
5298 				    "Unable to allocate RX descriptors\n");
5299 				err = ENOMEM;
5300 				goto err_tx_desc;
5301 			}
5302 			bzero((void *)ifdip->idi_vaddr, rxqsizes[j]);
5303 		}
5304 		rxq->ifr_ctx = ctx;
5305 		rxq->ifr_id = i;
5306 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
5307 			rxq->ifr_fl_offset = 1;
5308 		} else {
5309 			rxq->ifr_fl_offset = 0;
5310 		}
5311 		rxq->ifr_nfl = nfree_lists;
5312 		if (!(fl =
5313 			  (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) {
5314 			device_printf(dev, "Unable to allocate free list memory\n");
5315 			err = ENOMEM;
5316 			goto err_tx_desc;
5317 		}
5318 		rxq->ifr_fl = fl;
5319 		for (j = 0; j < nfree_lists; j++) {
5320 			fl[j].ifl_rxq = rxq;
5321 			fl[j].ifl_id = j;
5322 			fl[j].ifl_ifdi = &rxq->ifr_ifdi[j + rxq->ifr_fl_offset];
5323 			fl[j].ifl_rxd_size = scctx->isc_rxd_size[j];
5324 		}
5325 		/* Allocate receive buffers for the ring */
5326 		if (iflib_rxsd_alloc(rxq)) {
5327 			device_printf(dev,
5328 			    "Critical Failure setting up receive buffers\n");
5329 			err = ENOMEM;
5330 			goto err_rx_desc;
5331 		}
5332 
5333 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5334 			fl->ifl_rx_bitmap = bit_alloc(fl->ifl_size, M_IFLIB,
5335 			    M_WAITOK);
5336 	}
5337 
5338 	/* TXQs */
5339 	vaddrs = malloc(sizeof(caddr_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5340 	paddrs = malloc(sizeof(uint64_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5341 	for (i = 0; i < ntxqsets; i++) {
5342 		iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi;
5343 
5344 		for (j = 0; j < ntxqs; j++, di++) {
5345 			vaddrs[i*ntxqs + j] = di->idi_vaddr;
5346 			paddrs[i*ntxqs + j] = di->idi_paddr;
5347 		}
5348 	}
5349 	if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) {
5350 		device_printf(ctx->ifc_dev,
5351 		    "Unable to allocate device TX queue\n");
5352 		iflib_tx_structures_free(ctx);
5353 		free(vaddrs, M_IFLIB);
5354 		free(paddrs, M_IFLIB);
5355 		goto err_rx_desc;
5356 	}
5357 	free(vaddrs, M_IFLIB);
5358 	free(paddrs, M_IFLIB);
5359 
5360 	/* RXQs */
5361 	vaddrs = malloc(sizeof(caddr_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5362 	paddrs = malloc(sizeof(uint64_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5363 	for (i = 0; i < nrxqsets; i++) {
5364 		iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi;
5365 
5366 		for (j = 0; j < nrxqs; j++, di++) {
5367 			vaddrs[i*nrxqs + j] = di->idi_vaddr;
5368 			paddrs[i*nrxqs + j] = di->idi_paddr;
5369 		}
5370 	}
5371 	if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) {
5372 		device_printf(ctx->ifc_dev,
5373 		    "Unable to allocate device RX queue\n");
5374 		iflib_tx_structures_free(ctx);
5375 		free(vaddrs, M_IFLIB);
5376 		free(paddrs, M_IFLIB);
5377 		goto err_rx_desc;
5378 	}
5379 	free(vaddrs, M_IFLIB);
5380 	free(paddrs, M_IFLIB);
5381 
5382 	return (0);
5383 
5384 /* XXX handle allocation failure changes */
5385 err_rx_desc:
5386 err_tx_desc:
5387 rx_fail:
5388 	if (ctx->ifc_rxqs != NULL)
5389 		free(ctx->ifc_rxqs, M_IFLIB);
5390 	ctx->ifc_rxqs = NULL;
5391 	if (ctx->ifc_txqs != NULL)
5392 		free(ctx->ifc_txqs, M_IFLIB);
5393 	ctx->ifc_txqs = NULL;
5394 fail:
5395 	return (err);
5396 }
5397 
5398 static int
5399 iflib_tx_structures_setup(if_ctx_t ctx)
5400 {
5401 	iflib_txq_t txq = ctx->ifc_txqs;
5402 	int i;
5403 
5404 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
5405 		iflib_txq_setup(txq);
5406 
5407 	return (0);
5408 }
5409 
5410 static void
5411 iflib_tx_structures_free(if_ctx_t ctx)
5412 {
5413 	iflib_txq_t txq = ctx->ifc_txqs;
5414 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5415 	int i, j;
5416 
5417 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
5418 		iflib_txq_destroy(txq);
5419 		for (j = 0; j < sctx->isc_ntxqs; j++)
5420 			iflib_dma_free(&txq->ift_ifdi[j]);
5421 	}
5422 	free(ctx->ifc_txqs, M_IFLIB);
5423 	ctx->ifc_txqs = NULL;
5424 	IFDI_QUEUES_FREE(ctx);
5425 }
5426 
5427 /*********************************************************************
5428  *
5429  *  Initialize all receive rings.
5430  *
5431  **********************************************************************/
5432 static int
5433 iflib_rx_structures_setup(if_ctx_t ctx)
5434 {
5435 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5436 	int q;
5437 #if defined(INET6) || defined(INET)
5438 	int i, err;
5439 #endif
5440 
5441 	for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) {
5442 #if defined(INET6) || defined(INET)
5443 		tcp_lro_free(&rxq->ifr_lc);
5444 		if ((err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp,
5445 		    TCP_LRO_ENTRIES, min(1024,
5446 		    ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset]))) != 0) {
5447 			device_printf(ctx->ifc_dev, "LRO Initialization failed!\n");
5448 			goto fail;
5449 		}
5450 		rxq->ifr_lro_enabled = TRUE;
5451 #endif
5452 		IFDI_RXQ_SETUP(ctx, rxq->ifr_id);
5453 	}
5454 	return (0);
5455 #if defined(INET6) || defined(INET)
5456 fail:
5457 	/*
5458 	 * Free RX software descriptors allocated so far, we will only handle
5459 	 * the rings that completed, the failing case will have
5460 	 * cleaned up for itself. 'q' failed, so its the terminus.
5461 	 */
5462 	rxq = ctx->ifc_rxqs;
5463 	for (i = 0; i < q; ++i, rxq++) {
5464 		iflib_rx_sds_free(rxq);
5465 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
5466 	}
5467 	return (err);
5468 #endif
5469 }
5470 
5471 /*********************************************************************
5472  *
5473  *  Free all receive rings.
5474  *
5475  **********************************************************************/
5476 static void
5477 iflib_rx_structures_free(if_ctx_t ctx)
5478 {
5479 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5480 
5481 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) {
5482 		iflib_rx_sds_free(rxq);
5483 	}
5484 	free(ctx->ifc_rxqs, M_IFLIB);
5485 	ctx->ifc_rxqs = NULL;
5486 }
5487 
5488 static int
5489 iflib_qset_structures_setup(if_ctx_t ctx)
5490 {
5491 	int err;
5492 
5493 	/*
5494 	 * It is expected that the caller takes care of freeing queues if this
5495 	 * fails.
5496 	 */
5497 	if ((err = iflib_tx_structures_setup(ctx)) != 0) {
5498 		device_printf(ctx->ifc_dev, "iflib_tx_structures_setup failed: %d\n", err);
5499 		return (err);
5500 	}
5501 
5502 	if ((err = iflib_rx_structures_setup(ctx)) != 0)
5503 		device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err);
5504 
5505 	return (err);
5506 }
5507 
5508 int
5509 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
5510 		driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, const char *name)
5511 {
5512 
5513 	return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name));
5514 }
5515 
5516 #ifdef SMP
5517 static int
5518 find_nth(if_ctx_t ctx, int qid)
5519 {
5520 	cpuset_t cpus;
5521 	int i, cpuid, eqid, count;
5522 
5523 	CPU_COPY(&ctx->ifc_cpus, &cpus);
5524 	count = CPU_COUNT(&cpus);
5525 	eqid = qid % count;
5526 	/* clear up to the qid'th bit */
5527 	for (i = 0; i < eqid; i++) {
5528 		cpuid = CPU_FFS(&cpus);
5529 		MPASS(cpuid != 0);
5530 		CPU_CLR(cpuid-1, &cpus);
5531 	}
5532 	cpuid = CPU_FFS(&cpus);
5533 	MPASS(cpuid != 0);
5534 	return (cpuid-1);
5535 }
5536 
5537 #ifdef SCHED_ULE
5538 extern struct cpu_group *cpu_top;              /* CPU topology */
5539 
5540 static int
5541 find_child_with_core(int cpu, struct cpu_group *grp)
5542 {
5543 	int i;
5544 
5545 	if (grp->cg_children == 0)
5546 		return -1;
5547 
5548 	MPASS(grp->cg_child);
5549 	for (i = 0; i < grp->cg_children; i++) {
5550 		if (CPU_ISSET(cpu, &grp->cg_child[i].cg_mask))
5551 			return i;
5552 	}
5553 
5554 	return -1;
5555 }
5556 
5557 /*
5558  * Find the nth "close" core to the specified core
5559  * "close" is defined as the deepest level that shares
5560  * at least an L2 cache.  With threads, this will be
5561  * threads on the same core.  If the sahred cache is L3
5562  * or higher, simply returns the same core.
5563  */
5564 static int
5565 find_close_core(int cpu, int core_offset)
5566 {
5567 	struct cpu_group *grp;
5568 	int i;
5569 	int fcpu;
5570 	cpuset_t cs;
5571 
5572 	grp = cpu_top;
5573 	if (grp == NULL)
5574 		return cpu;
5575 	i = 0;
5576 	while ((i = find_child_with_core(cpu, grp)) != -1) {
5577 		/* If the child only has one cpu, don't descend */
5578 		if (grp->cg_child[i].cg_count <= 1)
5579 			break;
5580 		grp = &grp->cg_child[i];
5581 	}
5582 
5583 	/* If they don't share at least an L2 cache, use the same CPU */
5584 	if (grp->cg_level > CG_SHARE_L2 || grp->cg_level == CG_SHARE_NONE)
5585 		return cpu;
5586 
5587 	/* Now pick one */
5588 	CPU_COPY(&grp->cg_mask, &cs);
5589 
5590 	/* Add the selected CPU offset to core offset. */
5591 	for (i = 0; (fcpu = CPU_FFS(&cs)) != 0; i++) {
5592 		if (fcpu - 1 == cpu)
5593 			break;
5594 		CPU_CLR(fcpu - 1, &cs);
5595 	}
5596 	MPASS(fcpu);
5597 
5598 	core_offset += i;
5599 
5600 	CPU_COPY(&grp->cg_mask, &cs);
5601 	for (i = core_offset % grp->cg_count; i > 0; i--) {
5602 		MPASS(CPU_FFS(&cs));
5603 		CPU_CLR(CPU_FFS(&cs) - 1, &cs);
5604 	}
5605 	MPASS(CPU_FFS(&cs));
5606 	return CPU_FFS(&cs) - 1;
5607 }
5608 #else
5609 static int
5610 find_close_core(int cpu, int core_offset __unused)
5611 {
5612 	return cpu;
5613 }
5614 #endif
5615 
5616 static int
5617 get_core_offset(if_ctx_t ctx, iflib_intr_type_t type, int qid)
5618 {
5619 	switch (type) {
5620 	case IFLIB_INTR_TX:
5621 		/* TX queues get cores which share at least an L2 cache with the corresponding RX queue */
5622 		/* XXX handle multiple RX threads per core and more than two core per L2 group */
5623 		return qid / CPU_COUNT(&ctx->ifc_cpus) + 1;
5624 	case IFLIB_INTR_RX:
5625 	case IFLIB_INTR_RXTX:
5626 		/* RX queues get the specified core */
5627 		return qid / CPU_COUNT(&ctx->ifc_cpus);
5628 	default:
5629 		return -1;
5630 	}
5631 }
5632 #else
5633 #define get_core_offset(ctx, type, qid)	CPU_FIRST()
5634 #define find_close_core(cpuid, tid)	CPU_FIRST()
5635 #define find_nth(ctx, gid)		CPU_FIRST()
5636 #endif
5637 
5638 /* Just to avoid copy/paste */
5639 static inline int
5640 iflib_irq_set_affinity(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type,
5641     int qid, struct grouptask *gtask, struct taskqgroup *tqg, void *uniq,
5642     const char *name)
5643 {
5644 	device_t dev;
5645 	int err, cpuid, tid;
5646 
5647 	dev = ctx->ifc_dev;
5648 	cpuid = find_nth(ctx, qid);
5649 	tid = get_core_offset(ctx, type, qid);
5650 	MPASS(tid >= 0);
5651 	cpuid = find_close_core(cpuid, tid);
5652 	err = taskqgroup_attach_cpu(tqg, gtask, uniq, cpuid, dev, irq->ii_res,
5653 	    name);
5654 	if (err) {
5655 		device_printf(dev, "taskqgroup_attach_cpu failed %d\n", err);
5656 		return (err);
5657 	}
5658 #ifdef notyet
5659 	if (cpuid > ctx->ifc_cpuid_highest)
5660 		ctx->ifc_cpuid_highest = cpuid;
5661 #endif
5662 	return 0;
5663 }
5664 
5665 int
5666 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid,
5667 			iflib_intr_type_t type, driver_filter_t *filter,
5668 			void *filter_arg, int qid, const char *name)
5669 {
5670 	device_t dev;
5671 	struct grouptask *gtask;
5672 	struct taskqgroup *tqg;
5673 	iflib_filter_info_t info;
5674 	gtask_fn_t *fn;
5675 	int tqrid, err;
5676 	driver_filter_t *intr_fast;
5677 	void *q;
5678 
5679 	info = &ctx->ifc_filter_info;
5680 	tqrid = rid;
5681 
5682 	switch (type) {
5683 	/* XXX merge tx/rx for netmap? */
5684 	case IFLIB_INTR_TX:
5685 		q = &ctx->ifc_txqs[qid];
5686 		info = &ctx->ifc_txqs[qid].ift_filter_info;
5687 		gtask = &ctx->ifc_txqs[qid].ift_task;
5688 		tqg = qgroup_if_io_tqg;
5689 		fn = _task_fn_tx;
5690 		intr_fast = iflib_fast_intr;
5691 		GROUPTASK_INIT(gtask, 0, fn, q);
5692 		ctx->ifc_flags |= IFC_NETMAP_TX_IRQ;
5693 		break;
5694 	case IFLIB_INTR_RX:
5695 		q = &ctx->ifc_rxqs[qid];
5696 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
5697 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5698 		tqg = qgroup_if_io_tqg;
5699 		fn = _task_fn_rx;
5700 		intr_fast = iflib_fast_intr;
5701 		GROUPTASK_INIT(gtask, 0, fn, q);
5702 		break;
5703 	case IFLIB_INTR_RXTX:
5704 		q = &ctx->ifc_rxqs[qid];
5705 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
5706 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5707 		tqg = qgroup_if_io_tqg;
5708 		fn = _task_fn_rx;
5709 		intr_fast = iflib_fast_intr_rxtx;
5710 		GROUPTASK_INIT(gtask, 0, fn, q);
5711 		break;
5712 	case IFLIB_INTR_ADMIN:
5713 		q = ctx;
5714 		tqrid = -1;
5715 		info = &ctx->ifc_filter_info;
5716 		gtask = &ctx->ifc_admin_task;
5717 		tqg = qgroup_if_config_tqg;
5718 		fn = _task_fn_admin;
5719 		intr_fast = iflib_fast_intr_ctx;
5720 		break;
5721 	default:
5722 		panic("unknown net intr type");
5723 	}
5724 
5725 	info->ifi_filter = filter;
5726 	info->ifi_filter_arg = filter_arg;
5727 	info->ifi_task = gtask;
5728 	info->ifi_ctx = q;
5729 
5730 	dev = ctx->ifc_dev;
5731 	err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info,  name);
5732 	if (err != 0) {
5733 		device_printf(dev, "_iflib_irq_alloc failed %d\n", err);
5734 		return (err);
5735 	}
5736 	if (type == IFLIB_INTR_ADMIN)
5737 		return (0);
5738 
5739 	if (tqrid != -1) {
5740 		err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg,
5741 		    q, name);
5742 		if (err)
5743 			return (err);
5744 	} else {
5745 		taskqgroup_attach(tqg, gtask, q, dev, irq->ii_res, name);
5746 	}
5747 
5748 	return (0);
5749 }
5750 
5751 void
5752 iflib_softirq_alloc_generic(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type, void *arg, int qid, const char *name)
5753 {
5754 	struct grouptask *gtask;
5755 	struct taskqgroup *tqg;
5756 	gtask_fn_t *fn;
5757 	void *q;
5758 	int err;
5759 
5760 	switch (type) {
5761 	case IFLIB_INTR_TX:
5762 		q = &ctx->ifc_txqs[qid];
5763 		gtask = &ctx->ifc_txqs[qid].ift_task;
5764 		tqg = qgroup_if_io_tqg;
5765 		fn = _task_fn_tx;
5766 		break;
5767 	case IFLIB_INTR_RX:
5768 		q = &ctx->ifc_rxqs[qid];
5769 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5770 		tqg = qgroup_if_io_tqg;
5771 		fn = _task_fn_rx;
5772 		break;
5773 	case IFLIB_INTR_IOV:
5774 		q = ctx;
5775 		gtask = &ctx->ifc_vflr_task;
5776 		tqg = qgroup_if_config_tqg;
5777 		fn = _task_fn_iov;
5778 		break;
5779 	default:
5780 		panic("unknown net intr type");
5781 	}
5782 	GROUPTASK_INIT(gtask, 0, fn, q);
5783 	if (irq != NULL) {
5784 		err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg,
5785 		    q, name);
5786 		if (err)
5787 			taskqgroup_attach(tqg, gtask, q, ctx->ifc_dev,
5788 			    irq->ii_res, name);
5789 	} else {
5790 		taskqgroup_attach(tqg, gtask, q, NULL, NULL, name);
5791 	}
5792 }
5793 
5794 void
5795 iflib_irq_free(if_ctx_t ctx, if_irq_t irq)
5796 {
5797 
5798 	if (irq->ii_tag)
5799 		bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag);
5800 
5801 	if (irq->ii_res)
5802 		bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ,
5803 		    rman_get_rid(irq->ii_res), irq->ii_res);
5804 }
5805 
5806 static int
5807 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, const char *name)
5808 {
5809 	iflib_txq_t txq = ctx->ifc_txqs;
5810 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5811 	if_irq_t irq = &ctx->ifc_legacy_irq;
5812 	iflib_filter_info_t info;
5813 	device_t dev;
5814 	struct grouptask *gtask;
5815 	struct resource *res;
5816 	struct taskqgroup *tqg;
5817 	gtask_fn_t *fn;
5818 	int tqrid;
5819 	void *q;
5820 	int err;
5821 
5822 	q = &ctx->ifc_rxqs[0];
5823 	info = &rxq[0].ifr_filter_info;
5824 	gtask = &rxq[0].ifr_task;
5825 	tqg = qgroup_if_io_tqg;
5826 	tqrid = irq->ii_rid = *rid;
5827 	fn = _task_fn_rx;
5828 
5829 	ctx->ifc_flags |= IFC_LEGACY;
5830 	info->ifi_filter = filter;
5831 	info->ifi_filter_arg = filter_arg;
5832 	info->ifi_task = gtask;
5833 	info->ifi_ctx = ctx;
5834 
5835 	dev = ctx->ifc_dev;
5836 	/* We allocate a single interrupt resource */
5837 	if ((err = _iflib_irq_alloc(ctx, irq, tqrid, iflib_fast_intr_ctx, NULL, info, name)) != 0)
5838 		return (err);
5839 	GROUPTASK_INIT(gtask, 0, fn, q);
5840 	res = irq->ii_res;
5841 	taskqgroup_attach(tqg, gtask, q, dev, res, name);
5842 
5843 	GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq);
5844 	taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, dev, res,
5845 	    "tx");
5846 	return (0);
5847 }
5848 
5849 void
5850 iflib_led_create(if_ctx_t ctx)
5851 {
5852 
5853 	ctx->ifc_led_dev = led_create(iflib_led_func, ctx,
5854 	    device_get_nameunit(ctx->ifc_dev));
5855 }
5856 
5857 void
5858 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid)
5859 {
5860 
5861 	GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
5862 }
5863 
5864 void
5865 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid)
5866 {
5867 
5868 	GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task);
5869 }
5870 
5871 void
5872 iflib_admin_intr_deferred(if_ctx_t ctx)
5873 {
5874 #ifdef INVARIANTS
5875 	struct grouptask *gtask;
5876 
5877 	gtask = &ctx->ifc_admin_task;
5878 	MPASS(gtask != NULL && gtask->gt_taskqueue != NULL);
5879 #endif
5880 
5881 	GROUPTASK_ENQUEUE(&ctx->ifc_admin_task);
5882 }
5883 
5884 void
5885 iflib_iov_intr_deferred(if_ctx_t ctx)
5886 {
5887 
5888 	GROUPTASK_ENQUEUE(&ctx->ifc_vflr_task);
5889 }
5890 
5891 void
5892 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, char *name)
5893 {
5894 
5895 	taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, NULL, NULL,
5896 	    name);
5897 }
5898 
5899 void
5900 iflib_config_gtask_init(void *ctx, struct grouptask *gtask, gtask_fn_t *fn,
5901 	const char *name)
5902 {
5903 
5904 	GROUPTASK_INIT(gtask, 0, fn, ctx);
5905 	taskqgroup_attach(qgroup_if_config_tqg, gtask, gtask, NULL, NULL,
5906 	    name);
5907 }
5908 
5909 void
5910 iflib_config_gtask_deinit(struct grouptask *gtask)
5911 {
5912 
5913 	taskqgroup_detach(qgroup_if_config_tqg, gtask);
5914 }
5915 
5916 void
5917 iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate)
5918 {
5919 	if_t ifp = ctx->ifc_ifp;
5920 	iflib_txq_t txq = ctx->ifc_txqs;
5921 
5922 	if_setbaudrate(ifp, baudrate);
5923 	if (baudrate >= IF_Gbps(10)) {
5924 		STATE_LOCK(ctx);
5925 		ctx->ifc_flags |= IFC_PREFETCH;
5926 		STATE_UNLOCK(ctx);
5927 	}
5928 	/* If link down, disable watchdog */
5929 	if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) {
5930 		for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++)
5931 			txq->ift_qstatus = IFLIB_QUEUE_IDLE;
5932 	}
5933 	ctx->ifc_link_state = link_state;
5934 	if_link_state_change(ifp, link_state);
5935 }
5936 
5937 static int
5938 iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq)
5939 {
5940 	int credits;
5941 #ifdef INVARIANTS
5942 	int credits_pre = txq->ift_cidx_processed;
5943 #endif
5944 
5945 	if (ctx->isc_txd_credits_update == NULL)
5946 		return (0);
5947 
5948 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
5949 	    BUS_DMASYNC_POSTREAD);
5950 	if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, true)) == 0)
5951 		return (0);
5952 
5953 	txq->ift_processed += credits;
5954 	txq->ift_cidx_processed += credits;
5955 
5956 	MPASS(credits_pre + credits == txq->ift_cidx_processed);
5957 	if (txq->ift_cidx_processed >= txq->ift_size)
5958 		txq->ift_cidx_processed -= txq->ift_size;
5959 	return (credits);
5960 }
5961 
5962 static int
5963 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget)
5964 {
5965 	iflib_fl_t fl;
5966 	u_int i;
5967 
5968 	for (i = 0, fl = &rxq->ifr_fl[0]; i < rxq->ifr_nfl; i++, fl++)
5969 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
5970 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
5971 	return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx,
5972 	    budget));
5973 }
5974 
5975 void
5976 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name,
5977 	const char *description, if_int_delay_info_t info,
5978 	int offset, int value)
5979 {
5980 	info->iidi_ctx = ctx;
5981 	info->iidi_offset = offset;
5982 	info->iidi_value = value;
5983 	SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev),
5984 	    SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)),
5985 	    OID_AUTO, name, CTLTYPE_INT|CTLFLAG_RW,
5986 	    info, 0, iflib_sysctl_int_delay, "I", description);
5987 }
5988 
5989 struct sx *
5990 iflib_ctx_lock_get(if_ctx_t ctx)
5991 {
5992 
5993 	return (&ctx->ifc_ctx_sx);
5994 }
5995 
5996 static int
5997 iflib_msix_init(if_ctx_t ctx)
5998 {
5999 	device_t dev = ctx->ifc_dev;
6000 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6001 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6002 	int vectors, queues, rx_queues, tx_queues, queuemsgs, msgs;
6003 	int iflib_num_tx_queues, iflib_num_rx_queues;
6004 	int err, admincnt, bar;
6005 
6006 	iflib_num_tx_queues = ctx->ifc_sysctl_ntxqs;
6007 	iflib_num_rx_queues = ctx->ifc_sysctl_nrxqs;
6008 
6009 	if (bootverbose)
6010 		device_printf(dev, "msix_init qsets capped at %d\n",
6011 		    imax(scctx->isc_ntxqsets, scctx->isc_nrxqsets));
6012 
6013 	bar = ctx->ifc_softc_ctx.isc_msix_bar;
6014 	admincnt = sctx->isc_admin_intrcnt;
6015 	/* Override by tuneable */
6016 	if (scctx->isc_disable_msix)
6017 		goto msi;
6018 
6019 	/* First try MSI-X */
6020 	if ((msgs = pci_msix_count(dev)) == 0) {
6021 		if (bootverbose)
6022 			device_printf(dev, "MSI-X not supported or disabled\n");
6023 		goto msi;
6024 	}
6025 	/*
6026 	 * bar == -1 => "trust me I know what I'm doing"
6027 	 * Some drivers are for hardware that is so shoddily
6028 	 * documented that no one knows which bars are which
6029 	 * so the developer has to map all bars. This hack
6030 	 * allows shoddy garbage to use MSI-X in this framework.
6031 	 */
6032 	if (bar != -1) {
6033 		ctx->ifc_msix_mem = bus_alloc_resource_any(dev,
6034 	            SYS_RES_MEMORY, &bar, RF_ACTIVE);
6035 		if (ctx->ifc_msix_mem == NULL) {
6036 			device_printf(dev, "Unable to map MSI-X table\n");
6037 			goto msi;
6038 		}
6039 	}
6040 #if IFLIB_DEBUG
6041 	/* use only 1 qset in debug mode */
6042 	queuemsgs = min(msgs - admincnt, 1);
6043 #else
6044 	queuemsgs = msgs - admincnt;
6045 #endif
6046 #ifdef RSS
6047 	queues = imin(queuemsgs, rss_getnumbuckets());
6048 #else
6049 	queues = queuemsgs;
6050 #endif
6051 	queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues);
6052 	if (bootverbose)
6053 		device_printf(dev,
6054 		    "intr CPUs: %d queue msgs: %d admincnt: %d\n",
6055 		    CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt);
6056 #ifdef  RSS
6057 	/* If we're doing RSS, clamp at the number of RSS buckets */
6058 	if (queues > rss_getnumbuckets())
6059 		queues = rss_getnumbuckets();
6060 #endif
6061 	if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt)
6062 		rx_queues = iflib_num_rx_queues;
6063 	else
6064 		rx_queues = queues;
6065 
6066 	if (rx_queues > scctx->isc_nrxqsets)
6067 		rx_queues = scctx->isc_nrxqsets;
6068 
6069 	/*
6070 	 * We want this to be all logical CPUs by default
6071 	 */
6072 	if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues)
6073 		tx_queues = iflib_num_tx_queues;
6074 	else
6075 		tx_queues = mp_ncpus;
6076 
6077 	if (tx_queues > scctx->isc_ntxqsets)
6078 		tx_queues = scctx->isc_ntxqsets;
6079 
6080 	if (ctx->ifc_sysctl_qs_eq_override == 0) {
6081 #ifdef INVARIANTS
6082 		if (tx_queues != rx_queues)
6083 			device_printf(dev,
6084 			    "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n",
6085 			    min(rx_queues, tx_queues), min(rx_queues, tx_queues));
6086 #endif
6087 		tx_queues = min(rx_queues, tx_queues);
6088 		rx_queues = min(rx_queues, tx_queues);
6089 	}
6090 
6091 	device_printf(dev, "Using %d rx queues %d tx queues\n",
6092 	    rx_queues, tx_queues);
6093 
6094 	vectors = rx_queues + admincnt;
6095 	if ((err = pci_alloc_msix(dev, &vectors)) == 0) {
6096 		device_printf(dev, "Using MSI-X interrupts with %d vectors\n",
6097 		    vectors);
6098 		scctx->isc_vectors = vectors;
6099 		scctx->isc_nrxqsets = rx_queues;
6100 		scctx->isc_ntxqsets = tx_queues;
6101 		scctx->isc_intr = IFLIB_INTR_MSIX;
6102 
6103 		return (vectors);
6104 	} else {
6105 		device_printf(dev,
6106 		    "failed to allocate %d MSI-X vectors, err: %d - using MSI\n",
6107 		    vectors, err);
6108 		bus_release_resource(dev, SYS_RES_MEMORY, bar,
6109 		    ctx->ifc_msix_mem);
6110 		ctx->ifc_msix_mem = NULL;
6111 	}
6112 msi:
6113 	vectors = pci_msi_count(dev);
6114 	scctx->isc_nrxqsets = 1;
6115 	scctx->isc_ntxqsets = 1;
6116 	scctx->isc_vectors = vectors;
6117 	if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) {
6118 		device_printf(dev,"Using an MSI interrupt\n");
6119 		scctx->isc_intr = IFLIB_INTR_MSI;
6120 	} else {
6121 		scctx->isc_vectors = 1;
6122 		device_printf(dev,"Using a Legacy interrupt\n");
6123 		scctx->isc_intr = IFLIB_INTR_LEGACY;
6124 	}
6125 
6126 	return (vectors);
6127 }
6128 
6129 static const char *ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" };
6130 
6131 static int
6132 mp_ring_state_handler(SYSCTL_HANDLER_ARGS)
6133 {
6134 	int rc;
6135 	uint16_t *state = ((uint16_t *)oidp->oid_arg1);
6136 	struct sbuf *sb;
6137 	const char *ring_state = "UNKNOWN";
6138 
6139 	/* XXX needed ? */
6140 	rc = sysctl_wire_old_buffer(req, 0);
6141 	MPASS(rc == 0);
6142 	if (rc != 0)
6143 		return (rc);
6144 	sb = sbuf_new_for_sysctl(NULL, NULL, 80, req);
6145 	MPASS(sb != NULL);
6146 	if (sb == NULL)
6147 		return (ENOMEM);
6148 	if (state[3] <= 3)
6149 		ring_state = ring_states[state[3]];
6150 
6151 	sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s",
6152 		    state[0], state[1], state[2], ring_state);
6153 	rc = sbuf_finish(sb);
6154 	sbuf_delete(sb);
6155         return(rc);
6156 }
6157 
6158 enum iflib_ndesc_handler {
6159 	IFLIB_NTXD_HANDLER,
6160 	IFLIB_NRXD_HANDLER,
6161 };
6162 
6163 static int
6164 mp_ndesc_handler(SYSCTL_HANDLER_ARGS)
6165 {
6166 	if_ctx_t ctx = (void *)arg1;
6167 	enum iflib_ndesc_handler type = arg2;
6168 	char buf[256] = {0};
6169 	qidx_t *ndesc;
6170 	char *p, *next;
6171 	int nqs, rc, i;
6172 
6173 	MPASS(type == IFLIB_NTXD_HANDLER || type == IFLIB_NRXD_HANDLER);
6174 
6175 	nqs = 8;
6176 	switch(type) {
6177 	case IFLIB_NTXD_HANDLER:
6178 		ndesc = ctx->ifc_sysctl_ntxds;
6179 		if (ctx->ifc_sctx)
6180 			nqs = ctx->ifc_sctx->isc_ntxqs;
6181 		break;
6182 	case IFLIB_NRXD_HANDLER:
6183 		ndesc = ctx->ifc_sysctl_nrxds;
6184 		if (ctx->ifc_sctx)
6185 			nqs = ctx->ifc_sctx->isc_nrxqs;
6186 		break;
6187 	default:
6188 			panic("unhandled type");
6189 	}
6190 	if (nqs == 0)
6191 		nqs = 8;
6192 
6193 	for (i=0; i<8; i++) {
6194 		if (i >= nqs)
6195 			break;
6196 		if (i)
6197 			strcat(buf, ",");
6198 		sprintf(strchr(buf, 0), "%d", ndesc[i]);
6199 	}
6200 
6201 	rc = sysctl_handle_string(oidp, buf, sizeof(buf), req);
6202 	if (rc || req->newptr == NULL)
6203 		return rc;
6204 
6205 	for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p;
6206 	    i++, p = strsep(&next, " ,")) {
6207 		ndesc[i] = strtoul(p, NULL, 10);
6208 	}
6209 
6210 	return(rc);
6211 }
6212 
6213 #define NAME_BUFLEN 32
6214 static void
6215 iflib_add_device_sysctl_pre(if_ctx_t ctx)
6216 {
6217         device_t dev = iflib_get_dev(ctx);
6218 	struct sysctl_oid_list *child, *oid_list;
6219 	struct sysctl_ctx_list *ctx_list;
6220 	struct sysctl_oid *node;
6221 
6222 	ctx_list = device_get_sysctl_ctx(dev);
6223 	child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev));
6224 	ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, "iflib",
6225 						      CTLFLAG_RD, NULL, "IFLIB fields");
6226 	oid_list = SYSCTL_CHILDREN(node);
6227 
6228 	SYSCTL_ADD_CONST_STRING(ctx_list, oid_list, OID_AUTO, "driver_version",
6229 		       CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version,
6230 		       "driver version");
6231 
6232 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs",
6233 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0,
6234 			"# of txqs to use, 0 => use default #");
6235 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs",
6236 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0,
6237 			"# of rxqs to use, 0 => use default #");
6238 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_qs_enable",
6239 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0,
6240                        "permit #txq != #rxq");
6241 	SYSCTL_ADD_INT(ctx_list, oid_list, OID_AUTO, "disable_msix",
6242                       CTLFLAG_RWTUN, &ctx->ifc_softc_ctx.isc_disable_msix, 0,
6243                       "disable MSI-X (default 0)");
6244 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "rx_budget",
6245 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_rx_budget, 0,
6246                        "set the rx budget");
6247 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "tx_abdicate",
6248 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_tx_abdicate, 0,
6249 		       "cause tx to abdicate instead of running to completion");
6250 
6251 	/* XXX change for per-queue sizes */
6252 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_ntxds",
6253 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NTXD_HANDLER,
6254                        mp_ndesc_handler, "A",
6255                        "list of # of tx descriptors to use, 0 = use default #");
6256 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_nrxds",
6257 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NRXD_HANDLER,
6258                        mp_ndesc_handler, "A",
6259                        "list of # of rx descriptors to use, 0 = use default #");
6260 }
6261 
6262 static void
6263 iflib_add_device_sysctl_post(if_ctx_t ctx)
6264 {
6265 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6266 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6267         device_t dev = iflib_get_dev(ctx);
6268 	struct sysctl_oid_list *child;
6269 	struct sysctl_ctx_list *ctx_list;
6270 	iflib_fl_t fl;
6271 	iflib_txq_t txq;
6272 	iflib_rxq_t rxq;
6273 	int i, j;
6274 	char namebuf[NAME_BUFLEN];
6275 	char *qfmt;
6276 	struct sysctl_oid *queue_node, *fl_node, *node;
6277 	struct sysctl_oid_list *queue_list, *fl_list;
6278 	ctx_list = device_get_sysctl_ctx(dev);
6279 
6280 	node = ctx->ifc_sysctl_node;
6281 	child = SYSCTL_CHILDREN(node);
6282 
6283 	if (scctx->isc_ntxqsets > 100)
6284 		qfmt = "txq%03d";
6285 	else if (scctx->isc_ntxqsets > 10)
6286 		qfmt = "txq%02d";
6287 	else
6288 		qfmt = "txq%d";
6289 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
6290 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6291 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6292 					     CTLFLAG_RD, NULL, "Queue Name");
6293 		queue_list = SYSCTL_CHILDREN(queue_node);
6294 #if MEMORY_LOGGING
6295 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued",
6296 				CTLFLAG_RD,
6297 				&txq->ift_dequeued, "total mbufs freed");
6298 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued",
6299 				CTLFLAG_RD,
6300 				&txq->ift_enqueued, "total mbufs enqueued");
6301 #endif
6302 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag",
6303 				   CTLFLAG_RD,
6304 				   &txq->ift_mbuf_defrag, "# of times m_defrag was called");
6305 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "m_pullups",
6306 				   CTLFLAG_RD,
6307 				   &txq->ift_pullups, "# of times m_pullup was called");
6308 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag_failed",
6309 				   CTLFLAG_RD,
6310 				   &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed");
6311 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_desc_avail",
6312 				   CTLFLAG_RD,
6313 				   &txq->ift_no_desc_avail, "# of times no descriptors were available");
6314 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "tx_map_failed",
6315 				   CTLFLAG_RD,
6316 				   &txq->ift_map_failed, "# of times dma map failed");
6317 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txd_encap_efbig",
6318 				   CTLFLAG_RD,
6319 				   &txq->ift_txd_encap_efbig, "# of times txd_encap returned EFBIG");
6320 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_tx_dma_setup",
6321 				   CTLFLAG_RD,
6322 				   &txq->ift_no_tx_dma_setup, "# of times map failed for other than EFBIG");
6323 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx",
6324 				   CTLFLAG_RD,
6325 				   &txq->ift_pidx, 1, "Producer Index");
6326 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx",
6327 				   CTLFLAG_RD,
6328 				   &txq->ift_cidx, 1, "Consumer Index");
6329 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx_processed",
6330 				   CTLFLAG_RD,
6331 				   &txq->ift_cidx_processed, 1, "Consumer Index seen by credit update");
6332 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use",
6333 				   CTLFLAG_RD,
6334 				   &txq->ift_in_use, 1, "descriptors in use");
6335 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_processed",
6336 				   CTLFLAG_RD,
6337 				   &txq->ift_processed, "descriptors procesed for clean");
6338 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned",
6339 				   CTLFLAG_RD,
6340 				   &txq->ift_cleaned, "total cleaned");
6341 		SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state",
6342 				CTLTYPE_STRING | CTLFLAG_RD, __DEVOLATILE(uint64_t *, &txq->ift_br->state),
6343 				0, mp_ring_state_handler, "A", "soft ring state");
6344 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_enqueues",
6345 				       CTLFLAG_RD, &txq->ift_br->enqueues,
6346 				       "# of enqueues to the mp_ring for this queue");
6347 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_drops",
6348 				       CTLFLAG_RD, &txq->ift_br->drops,
6349 				       "# of drops in the mp_ring for this queue");
6350 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_starts",
6351 				       CTLFLAG_RD, &txq->ift_br->starts,
6352 				       "# of normal consumer starts in the mp_ring for this queue");
6353 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_stalls",
6354 				       CTLFLAG_RD, &txq->ift_br->stalls,
6355 					       "# of consumer stalls in the mp_ring for this queue");
6356 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_restarts",
6357 			       CTLFLAG_RD, &txq->ift_br->restarts,
6358 				       "# of consumer restarts in the mp_ring for this queue");
6359 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_abdications",
6360 				       CTLFLAG_RD, &txq->ift_br->abdications,
6361 				       "# of consumer abdications in the mp_ring for this queue");
6362 	}
6363 
6364 	if (scctx->isc_nrxqsets > 100)
6365 		qfmt = "rxq%03d";
6366 	else if (scctx->isc_nrxqsets > 10)
6367 		qfmt = "rxq%02d";
6368 	else
6369 		qfmt = "rxq%d";
6370 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
6371 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6372 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6373 					     CTLFLAG_RD, NULL, "Queue Name");
6374 		queue_list = SYSCTL_CHILDREN(queue_node);
6375 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
6376 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_pidx",
6377 				       CTLFLAG_RD,
6378 				       &rxq->ifr_cq_pidx, 1, "Producer Index");
6379 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_cidx",
6380 				       CTLFLAG_RD,
6381 				       &rxq->ifr_cq_cidx, 1, "Consumer Index");
6382 		}
6383 
6384 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
6385 			snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j);
6386 			fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list, OID_AUTO, namebuf,
6387 						     CTLFLAG_RD, NULL, "freelist Name");
6388 			fl_list = SYSCTL_CHILDREN(fl_node);
6389 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx",
6390 				       CTLFLAG_RD,
6391 				       &fl->ifl_pidx, 1, "Producer Index");
6392 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx",
6393 				       CTLFLAG_RD,
6394 				       &fl->ifl_cidx, 1, "Consumer Index");
6395 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits",
6396 				       CTLFLAG_RD,
6397 				       &fl->ifl_credits, 1, "credits available");
6398 #if MEMORY_LOGGING
6399 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_enqueued",
6400 					CTLFLAG_RD,
6401 					&fl->ifl_m_enqueued, "mbufs allocated");
6402 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_dequeued",
6403 					CTLFLAG_RD,
6404 					&fl->ifl_m_dequeued, "mbufs freed");
6405 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_enqueued",
6406 					CTLFLAG_RD,
6407 					&fl->ifl_cl_enqueued, "clusters allocated");
6408 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_dequeued",
6409 					CTLFLAG_RD,
6410 					&fl->ifl_cl_dequeued, "clusters freed");
6411 #endif
6412 
6413 		}
6414 	}
6415 
6416 }
6417 
6418 void
6419 iflib_request_reset(if_ctx_t ctx)
6420 {
6421 
6422 	STATE_LOCK(ctx);
6423 	ctx->ifc_flags |= IFC_DO_RESET;
6424 	STATE_UNLOCK(ctx);
6425 }
6426 
6427 #ifndef __NO_STRICT_ALIGNMENT
6428 static struct mbuf *
6429 iflib_fixup_rx(struct mbuf *m)
6430 {
6431 	struct mbuf *n;
6432 
6433 	if (m->m_len <= (MCLBYTES - ETHER_HDR_LEN)) {
6434 		bcopy(m->m_data, m->m_data + ETHER_HDR_LEN, m->m_len);
6435 		m->m_data += ETHER_HDR_LEN;
6436 		n = m;
6437 	} else {
6438 		MGETHDR(n, M_NOWAIT, MT_DATA);
6439 		if (n == NULL) {
6440 			m_freem(m);
6441 			return (NULL);
6442 		}
6443 		bcopy(m->m_data, n->m_data, ETHER_HDR_LEN);
6444 		m->m_data += ETHER_HDR_LEN;
6445 		m->m_len -= ETHER_HDR_LEN;
6446 		n->m_len = ETHER_HDR_LEN;
6447 		M_MOVE_PKTHDR(n, m);
6448 		n->m_next = m;
6449 	}
6450 	return (n);
6451 }
6452 #endif
6453 
6454 #ifdef NETDUMP
6455 static void
6456 iflib_netdump_init(struct ifnet *ifp, int *nrxr, int *ncl, int *clsize)
6457 {
6458 	if_ctx_t ctx;
6459 
6460 	ctx = if_getsoftc(ifp);
6461 	CTX_LOCK(ctx);
6462 	*nrxr = NRXQSETS(ctx);
6463 	*ncl = ctx->ifc_rxqs[0].ifr_fl->ifl_size;
6464 	*clsize = ctx->ifc_rxqs[0].ifr_fl->ifl_buf_size;
6465 	CTX_UNLOCK(ctx);
6466 }
6467 
6468 static void
6469 iflib_netdump_event(struct ifnet *ifp, enum netdump_ev event)
6470 {
6471 	if_ctx_t ctx;
6472 	if_softc_ctx_t scctx;
6473 	iflib_fl_t fl;
6474 	iflib_rxq_t rxq;
6475 	int i, j;
6476 
6477 	ctx = if_getsoftc(ifp);
6478 	scctx = &ctx->ifc_softc_ctx;
6479 
6480 	switch (event) {
6481 	case NETDUMP_START:
6482 		for (i = 0; i < scctx->isc_nrxqsets; i++) {
6483 			rxq = &ctx->ifc_rxqs[i];
6484 			for (j = 0; j < rxq->ifr_nfl; j++) {
6485 				fl = rxq->ifr_fl;
6486 				fl->ifl_zone = m_getzone(fl->ifl_buf_size);
6487 			}
6488 		}
6489 		iflib_no_tx_batch = 1;
6490 		break;
6491 	default:
6492 		break;
6493 	}
6494 }
6495 
6496 static int
6497 iflib_netdump_transmit(struct ifnet *ifp, struct mbuf *m)
6498 {
6499 	if_ctx_t ctx;
6500 	iflib_txq_t txq;
6501 	int error;
6502 
6503 	ctx = if_getsoftc(ifp);
6504 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
6505 	    IFF_DRV_RUNNING)
6506 		return (EBUSY);
6507 
6508 	txq = &ctx->ifc_txqs[0];
6509 	error = iflib_encap(txq, &m);
6510 	if (error == 0)
6511 		(void)iflib_txd_db_check(ctx, txq, true, txq->ift_in_use);
6512 	return (error);
6513 }
6514 
6515 static int
6516 iflib_netdump_poll(struct ifnet *ifp, int count)
6517 {
6518 	if_ctx_t ctx;
6519 	if_softc_ctx_t scctx;
6520 	iflib_txq_t txq;
6521 	int i;
6522 
6523 	ctx = if_getsoftc(ifp);
6524 	scctx = &ctx->ifc_softc_ctx;
6525 
6526 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
6527 	    IFF_DRV_RUNNING)
6528 		return (EBUSY);
6529 
6530 	txq = &ctx->ifc_txqs[0];
6531 	(void)iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
6532 
6533 	for (i = 0; i < scctx->isc_nrxqsets; i++)
6534 		(void)iflib_rxeof(&ctx->ifc_rxqs[i], 16 /* XXX */);
6535 	return (0);
6536 }
6537 #endif /* NETDUMP */
6538