xref: /freebsd/sys/net/iflib.c (revision 7877fdebeeb35fad1cbbafce22598b1bdf97c786)
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/debugnet.h>
63 #include <net/pfil.h>
64 #include <net/vnet.h>
65 
66 #include <netinet/in.h>
67 #include <netinet/in_pcb.h>
68 #include <netinet/tcp_lro.h>
69 #include <netinet/in_systm.h>
70 #include <netinet/if_ether.h>
71 #include <netinet/ip.h>
72 #include <netinet/ip6.h>
73 #include <netinet/tcp.h>
74 #include <netinet/ip_var.h>
75 #include <netinet6/ip6_var.h>
76 
77 #include <machine/bus.h>
78 #include <machine/in_cksum.h>
79 
80 #include <vm/vm.h>
81 #include <vm/pmap.h>
82 
83 #include <dev/led/led.h>
84 #include <dev/pci/pcireg.h>
85 #include <dev/pci/pcivar.h>
86 #include <dev/pci/pci_private.h>
87 
88 #include <net/iflib.h>
89 #include <net/iflib_private.h>
90 
91 #include "ifdi_if.h"
92 
93 #ifdef PCI_IOV
94 #include <dev/pci/pci_iov.h>
95 #endif
96 
97 #include <sys/bitstring.h>
98 /*
99  * enable accounting of every mbuf as it comes in to and goes out of
100  * iflib's software descriptor references
101  */
102 #define MEMORY_LOGGING 0
103 /*
104  * Enable mbuf vectors for compressing long mbuf chains
105  */
106 
107 /*
108  * NB:
109  * - Prefetching in tx cleaning should perhaps be a tunable. The distance ahead
110  *   we prefetch needs to be determined by the time spent in m_free vis a vis
111  *   the cost of a prefetch. This will of course vary based on the workload:
112  *      - NFLX's m_free path is dominated by vm-based M_EXT manipulation which
113  *        is quite expensive, thus suggesting very little prefetch.
114  *      - small packet forwarding which is just returning a single mbuf to
115  *        UMA will typically be very fast vis a vis the cost of a memory
116  *        access.
117  */
118 
119 /*
120  * 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 #define	IFLIB_RXEOF_MORE (1U << 0)
132 #define	IFLIB_RXEOF_EMPTY (2U << 0)
133 
134 struct iflib_txq;
135 typedef struct iflib_txq *iflib_txq_t;
136 struct iflib_rxq;
137 typedef struct iflib_rxq *iflib_rxq_t;
138 struct iflib_fl;
139 typedef struct iflib_fl *iflib_fl_t;
140 
141 struct iflib_ctx;
142 
143 static void iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid);
144 static void iflib_timer(void *arg);
145 static void iflib_tqg_detach(if_ctx_t ctx);
146 
147 typedef struct iflib_filter_info {
148 	driver_filter_t *ifi_filter;
149 	void *ifi_filter_arg;
150 	struct grouptask *ifi_task;
151 	void *ifi_ctx;
152 } *iflib_filter_info_t;
153 
154 struct iflib_ctx {
155 	KOBJ_FIELDS;
156 	/*
157 	 * Pointer to hardware driver's softc
158 	 */
159 	void *ifc_softc;
160 	device_t ifc_dev;
161 	if_t ifc_ifp;
162 
163 	cpuset_t ifc_cpus;
164 	if_shared_ctx_t ifc_sctx;
165 	struct if_softc_ctx ifc_softc_ctx;
166 
167 	struct sx ifc_ctx_sx;
168 	struct mtx ifc_state_mtx;
169 
170 	iflib_txq_t ifc_txqs;
171 	iflib_rxq_t ifc_rxqs;
172 	uint32_t ifc_if_flags;
173 	uint32_t ifc_flags;
174 	uint32_t ifc_max_fl_buf_size;
175 	uint32_t ifc_rx_mbuf_sz;
176 
177 	int ifc_link_state;
178 	int ifc_watchdog_events;
179 	struct cdev *ifc_led_dev;
180 	struct resource *ifc_msix_mem;
181 
182 	struct if_irq ifc_legacy_irq;
183 	struct grouptask ifc_admin_task;
184 	struct grouptask ifc_vflr_task;
185 	struct iflib_filter_info ifc_filter_info;
186 	struct ifmedia	ifc_media;
187 	struct ifmedia	*ifc_mediap;
188 
189 	struct sysctl_oid *ifc_sysctl_node;
190 	uint16_t ifc_sysctl_ntxqs;
191 	uint16_t ifc_sysctl_nrxqs;
192 	uint16_t ifc_sysctl_qs_eq_override;
193 	uint16_t ifc_sysctl_rx_budget;
194 	uint16_t ifc_sysctl_tx_abdicate;
195 	uint16_t ifc_sysctl_core_offset;
196 #define	CORE_OFFSET_UNSPECIFIED	0xffff
197 	uint8_t  ifc_sysctl_separate_txrx;
198 
199 	qidx_t ifc_sysctl_ntxds[8];
200 	qidx_t ifc_sysctl_nrxds[8];
201 	struct if_txrx ifc_txrx;
202 #define isc_txd_encap  ifc_txrx.ift_txd_encap
203 #define isc_txd_flush  ifc_txrx.ift_txd_flush
204 #define isc_txd_credits_update  ifc_txrx.ift_txd_credits_update
205 #define isc_rxd_available ifc_txrx.ift_rxd_available
206 #define isc_rxd_pkt_get ifc_txrx.ift_rxd_pkt_get
207 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
208 #define isc_rxd_flush ifc_txrx.ift_rxd_flush
209 #define isc_legacy_intr ifc_txrx.ift_legacy_intr
210 	eventhandler_tag ifc_vlan_attach_event;
211 	eventhandler_tag ifc_vlan_detach_event;
212 	struct ether_addr ifc_mac;
213 };
214 
215 void *
216 iflib_get_softc(if_ctx_t ctx)
217 {
218 
219 	return (ctx->ifc_softc);
220 }
221 
222 device_t
223 iflib_get_dev(if_ctx_t ctx)
224 {
225 
226 	return (ctx->ifc_dev);
227 }
228 
229 if_t
230 iflib_get_ifp(if_ctx_t ctx)
231 {
232 
233 	return (ctx->ifc_ifp);
234 }
235 
236 struct ifmedia *
237 iflib_get_media(if_ctx_t ctx)
238 {
239 
240 	return (ctx->ifc_mediap);
241 }
242 
243 uint32_t
244 iflib_get_flags(if_ctx_t ctx)
245 {
246 	return (ctx->ifc_flags);
247 }
248 
249 void
250 iflib_set_mac(if_ctx_t ctx, uint8_t mac[ETHER_ADDR_LEN])
251 {
252 
253 	bcopy(mac, ctx->ifc_mac.octet, ETHER_ADDR_LEN);
254 }
255 
256 if_softc_ctx_t
257 iflib_get_softc_ctx(if_ctx_t ctx)
258 {
259 
260 	return (&ctx->ifc_softc_ctx);
261 }
262 
263 if_shared_ctx_t
264 iflib_get_sctx(if_ctx_t ctx)
265 {
266 
267 	return (ctx->ifc_sctx);
268 }
269 
270 #define IP_ALIGNED(m) ((((uintptr_t)(m)->m_data) & 0x3) == 0x2)
271 #define CACHE_PTR_INCREMENT (CACHE_LINE_SIZE/sizeof(void*))
272 #define CACHE_PTR_NEXT(ptr) ((void *)(((uintptr_t)(ptr)+CACHE_LINE_SIZE-1) & (CACHE_LINE_SIZE-1)))
273 
274 #define LINK_ACTIVE(ctx) ((ctx)->ifc_link_state == LINK_STATE_UP)
275 #define CTX_IS_VF(ctx) ((ctx)->ifc_sctx->isc_flags & IFLIB_IS_VF)
276 
277 typedef struct iflib_sw_rx_desc_array {
278 	bus_dmamap_t	*ifsd_map;         /* bus_dma maps for packet */
279 	struct mbuf	**ifsd_m;           /* pkthdr mbufs */
280 	caddr_t		*ifsd_cl;          /* direct cluster pointer for rx */
281 	bus_addr_t	*ifsd_ba;          /* bus addr of cluster for rx */
282 } iflib_rxsd_array_t;
283 
284 typedef struct iflib_sw_tx_desc_array {
285 	bus_dmamap_t    *ifsd_map;         /* bus_dma maps for packet */
286 	bus_dmamap_t	*ifsd_tso_map;     /* bus_dma maps for TSO packet */
287 	struct mbuf    **ifsd_m;           /* pkthdr mbufs */
288 } if_txsd_vec_t;
289 
290 /* magic number that should be high enough for any hardware */
291 #define IFLIB_MAX_TX_SEGS		128
292 #define IFLIB_RX_COPY_THRESH		128
293 #define IFLIB_MAX_RX_REFRESH		32
294 /* The minimum descriptors per second before we start coalescing */
295 #define IFLIB_MIN_DESC_SEC		16384
296 #define IFLIB_DEFAULT_TX_UPDATE_FREQ	16
297 #define IFLIB_QUEUE_IDLE		0
298 #define IFLIB_QUEUE_HUNG		1
299 #define IFLIB_QUEUE_WORKING		2
300 /* maximum number of txqs that can share an rx interrupt */
301 #define IFLIB_MAX_TX_SHARED_INTR	4
302 
303 /* this should really scale with ring size - this is a fairly arbitrary value */
304 #define TX_BATCH_SIZE			32
305 
306 #define IFLIB_RESTART_BUDGET		8
307 
308 #define CSUM_OFFLOAD		(CSUM_IP_TSO|CSUM_IP6_TSO|CSUM_IP| \
309 				 CSUM_IP_UDP|CSUM_IP_TCP|CSUM_IP_SCTP| \
310 				 CSUM_IP6_UDP|CSUM_IP6_TCP|CSUM_IP6_SCTP)
311 
312 struct iflib_txq {
313 	qidx_t		ift_in_use;
314 	qidx_t		ift_cidx;
315 	qidx_t		ift_cidx_processed;
316 	qidx_t		ift_pidx;
317 	uint8_t		ift_gen;
318 	uint8_t		ift_br_offset;
319 	uint16_t	ift_npending;
320 	uint16_t	ift_db_pending;
321 	uint16_t	ift_rs_pending;
322 	/* implicit pad */
323 	uint8_t		ift_txd_size[8];
324 	uint64_t	ift_processed;
325 	uint64_t	ift_cleaned;
326 	uint64_t	ift_cleaned_prev;
327 #if MEMORY_LOGGING
328 	uint64_t	ift_enqueued;
329 	uint64_t	ift_dequeued;
330 #endif
331 	uint64_t	ift_no_tx_dma_setup;
332 	uint64_t	ift_no_desc_avail;
333 	uint64_t	ift_mbuf_defrag_failed;
334 	uint64_t	ift_mbuf_defrag;
335 	uint64_t	ift_map_failed;
336 	uint64_t	ift_txd_encap_efbig;
337 	uint64_t	ift_pullups;
338 	uint64_t	ift_last_timer_tick;
339 
340 	struct mtx	ift_mtx;
341 	struct mtx	ift_db_mtx;
342 
343 	/* constant values */
344 	if_ctx_t	ift_ctx;
345 	struct ifmp_ring        *ift_br;
346 	struct grouptask	ift_task;
347 	qidx_t		ift_size;
348 	uint16_t	ift_id;
349 	struct callout	ift_timer;
350 #ifdef DEV_NETMAP
351 	struct callout	ift_netmap_timer;
352 #endif /* DEV_NETMAP */
353 
354 	if_txsd_vec_t	ift_sds;
355 	uint8_t		ift_qstatus;
356 	uint8_t		ift_closed;
357 	uint8_t		ift_update_freq;
358 	struct iflib_filter_info ift_filter_info;
359 	bus_dma_tag_t	ift_buf_tag;
360 	bus_dma_tag_t	ift_tso_buf_tag;
361 	iflib_dma_info_t	ift_ifdi;
362 #define	MTX_NAME_LEN	32
363 	char                    ift_mtx_name[MTX_NAME_LEN];
364 	bus_dma_segment_t	ift_segs[IFLIB_MAX_TX_SEGS]  __aligned(CACHE_LINE_SIZE);
365 #ifdef IFLIB_DIAGNOSTICS
366 	uint64_t ift_cpu_exec_count[256];
367 #endif
368 } __aligned(CACHE_LINE_SIZE);
369 
370 struct iflib_fl {
371 	qidx_t		ifl_cidx;
372 	qidx_t		ifl_pidx;
373 	qidx_t		ifl_credits;
374 	uint8_t		ifl_gen;
375 	uint8_t		ifl_rxd_size;
376 #if MEMORY_LOGGING
377 	uint64_t	ifl_m_enqueued;
378 	uint64_t	ifl_m_dequeued;
379 	uint64_t	ifl_cl_enqueued;
380 	uint64_t	ifl_cl_dequeued;
381 #endif
382 	/* implicit pad */
383 	bitstr_t 	*ifl_rx_bitmap;
384 	qidx_t		ifl_fragidx;
385 	/* constant */
386 	qidx_t		ifl_size;
387 	uint16_t	ifl_buf_size;
388 	uint16_t	ifl_cltype;
389 	uma_zone_t	ifl_zone;
390 	iflib_rxsd_array_t	ifl_sds;
391 	iflib_rxq_t	ifl_rxq;
392 	uint8_t		ifl_id;
393 	bus_dma_tag_t	ifl_buf_tag;
394 	iflib_dma_info_t	ifl_ifdi;
395 	uint64_t	ifl_bus_addrs[IFLIB_MAX_RX_REFRESH] __aligned(CACHE_LINE_SIZE);
396 	qidx_t		ifl_rxd_idxs[IFLIB_MAX_RX_REFRESH];
397 }  __aligned(CACHE_LINE_SIZE);
398 
399 static inline qidx_t
400 get_inuse(int size, qidx_t cidx, qidx_t pidx, uint8_t gen)
401 {
402 	qidx_t used;
403 
404 	if (pidx > cidx)
405 		used = pidx - cidx;
406 	else if (pidx < cidx)
407 		used = size - cidx + pidx;
408 	else if (gen == 0 && pidx == cidx)
409 		used = 0;
410 	else if (gen == 1 && pidx == cidx)
411 		used = size;
412 	else
413 		panic("bad state");
414 
415 	return (used);
416 }
417 
418 #define TXQ_AVAIL(txq) (txq->ift_size - get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx, txq->ift_gen))
419 
420 #define IDXDIFF(head, tail, wrap) \
421 	((head) >= (tail) ? (head) - (tail) : (wrap) - (tail) + (head))
422 
423 struct iflib_rxq {
424 	if_ctx_t	ifr_ctx;
425 	iflib_fl_t	ifr_fl;
426 	uint64_t	ifr_rx_irq;
427 	struct pfil_head	*pfil;
428 	/*
429 	 * If there is a separate completion queue (IFLIB_HAS_RXCQ), this is
430 	 * the completion queue consumer index.  Otherwise it's unused.
431 	 */
432 	qidx_t		ifr_cq_cidx;
433 	uint16_t	ifr_id;
434 	uint8_t		ifr_nfl;
435 	uint8_t		ifr_ntxqirq;
436 	uint8_t		ifr_txqid[IFLIB_MAX_TX_SHARED_INTR];
437 	uint8_t		ifr_fl_offset;
438 	struct lro_ctrl			ifr_lc;
439 	struct grouptask        ifr_task;
440 	struct callout		ifr_watchdog;
441 	struct iflib_filter_info ifr_filter_info;
442 	iflib_dma_info_t		ifr_ifdi;
443 
444 	/* dynamically allocate if any drivers need a value substantially larger than this */
445 	struct if_rxd_frag	ifr_frags[IFLIB_MAX_RX_SEGS] __aligned(CACHE_LINE_SIZE);
446 #ifdef IFLIB_DIAGNOSTICS
447 	uint64_t ifr_cpu_exec_count[256];
448 #endif
449 }  __aligned(CACHE_LINE_SIZE);
450 
451 typedef struct if_rxsd {
452 	caddr_t *ifsd_cl;
453 	iflib_fl_t ifsd_fl;
454 } *if_rxsd_t;
455 
456 /* multiple of word size */
457 #ifdef __LP64__
458 #define PKT_INFO_SIZE	6
459 #define RXD_INFO_SIZE	5
460 #define PKT_TYPE uint64_t
461 #else
462 #define PKT_INFO_SIZE	11
463 #define RXD_INFO_SIZE	8
464 #define PKT_TYPE uint32_t
465 #endif
466 #define PKT_LOOP_BOUND  ((PKT_INFO_SIZE/3)*3)
467 #define RXD_LOOP_BOUND  ((RXD_INFO_SIZE/4)*4)
468 
469 typedef struct if_pkt_info_pad {
470 	PKT_TYPE pkt_val[PKT_INFO_SIZE];
471 } *if_pkt_info_pad_t;
472 typedef struct if_rxd_info_pad {
473 	PKT_TYPE rxd_val[RXD_INFO_SIZE];
474 } *if_rxd_info_pad_t;
475 
476 CTASSERT(sizeof(struct if_pkt_info_pad) == sizeof(struct if_pkt_info));
477 CTASSERT(sizeof(struct if_rxd_info_pad) == sizeof(struct if_rxd_info));
478 
479 static inline void
480 pkt_info_zero(if_pkt_info_t pi)
481 {
482 	if_pkt_info_pad_t pi_pad;
483 
484 	pi_pad = (if_pkt_info_pad_t)pi;
485 	pi_pad->pkt_val[0] = 0; pi_pad->pkt_val[1] = 0; pi_pad->pkt_val[2] = 0;
486 	pi_pad->pkt_val[3] = 0; pi_pad->pkt_val[4] = 0; pi_pad->pkt_val[5] = 0;
487 #ifndef __LP64__
488 	pi_pad->pkt_val[6] = 0; pi_pad->pkt_val[7] = 0; pi_pad->pkt_val[8] = 0;
489 	pi_pad->pkt_val[9] = 0; pi_pad->pkt_val[10] = 0;
490 #endif
491 }
492 
493 static device_method_t iflib_pseudo_methods[] = {
494 	DEVMETHOD(device_attach, noop_attach),
495 	DEVMETHOD(device_detach, iflib_pseudo_detach),
496 	DEVMETHOD_END
497 };
498 
499 driver_t iflib_pseudodriver = {
500 	"iflib_pseudo", iflib_pseudo_methods, sizeof(struct iflib_ctx),
501 };
502 
503 static inline void
504 rxd_info_zero(if_rxd_info_t ri)
505 {
506 	if_rxd_info_pad_t ri_pad;
507 	int i;
508 
509 	ri_pad = (if_rxd_info_pad_t)ri;
510 	for (i = 0; i < RXD_LOOP_BOUND; i += 4) {
511 		ri_pad->rxd_val[i] = 0;
512 		ri_pad->rxd_val[i+1] = 0;
513 		ri_pad->rxd_val[i+2] = 0;
514 		ri_pad->rxd_val[i+3] = 0;
515 	}
516 #ifdef __LP64__
517 	ri_pad->rxd_val[RXD_INFO_SIZE-1] = 0;
518 #endif
519 }
520 
521 /*
522  * Only allow a single packet to take up most 1/nth of the tx ring
523  */
524 #define MAX_SINGLE_PACKET_FRACTION 12
525 #define IF_BAD_DMA (bus_addr_t)-1
526 
527 #define CTX_ACTIVE(ctx) ((if_getdrvflags((ctx)->ifc_ifp) & IFF_DRV_RUNNING))
528 
529 #define CTX_LOCK_INIT(_sc)  sx_init(&(_sc)->ifc_ctx_sx, "iflib ctx lock")
530 #define CTX_LOCK(ctx) sx_xlock(&(ctx)->ifc_ctx_sx)
531 #define CTX_UNLOCK(ctx) sx_xunlock(&(ctx)->ifc_ctx_sx)
532 #define CTX_LOCK_DESTROY(ctx) sx_destroy(&(ctx)->ifc_ctx_sx)
533 
534 #define STATE_LOCK_INIT(_sc, _name)  mtx_init(&(_sc)->ifc_state_mtx, _name, "iflib state lock", MTX_DEF)
535 #define STATE_LOCK(ctx) mtx_lock(&(ctx)->ifc_state_mtx)
536 #define STATE_UNLOCK(ctx) mtx_unlock(&(ctx)->ifc_state_mtx)
537 #define STATE_LOCK_DESTROY(ctx) mtx_destroy(&(ctx)->ifc_state_mtx)
538 
539 #define CALLOUT_LOCK(txq)	mtx_lock(&txq->ift_mtx)
540 #define CALLOUT_UNLOCK(txq) 	mtx_unlock(&txq->ift_mtx)
541 
542 void
543 iflib_set_detach(if_ctx_t ctx)
544 {
545 	STATE_LOCK(ctx);
546 	ctx->ifc_flags |= IFC_IN_DETACH;
547 	STATE_UNLOCK(ctx);
548 }
549 
550 /* Our boot-time initialization hook */
551 static int	iflib_module_event_handler(module_t, int, void *);
552 
553 static moduledata_t iflib_moduledata = {
554 	"iflib",
555 	iflib_module_event_handler,
556 	NULL
557 };
558 
559 DECLARE_MODULE(iflib, iflib_moduledata, SI_SUB_INIT_IF, SI_ORDER_ANY);
560 MODULE_VERSION(iflib, 1);
561 
562 MODULE_DEPEND(iflib, pci, 1, 1, 1);
563 MODULE_DEPEND(iflib, ether, 1, 1, 1);
564 
565 TASKQGROUP_DEFINE(if_io_tqg, mp_ncpus, 1);
566 TASKQGROUP_DEFINE(if_config_tqg, 1, 1);
567 
568 #ifndef IFLIB_DEBUG_COUNTERS
569 #ifdef INVARIANTS
570 #define IFLIB_DEBUG_COUNTERS 1
571 #else
572 #define IFLIB_DEBUG_COUNTERS 0
573 #endif /* !INVARIANTS */
574 #endif
575 
576 static SYSCTL_NODE(_net, OID_AUTO, iflib, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
577     "iflib driver parameters");
578 
579 /*
580  * XXX need to ensure that this can't accidentally cause the head to be moved backwards
581  */
582 static int iflib_min_tx_latency = 0;
583 SYSCTL_INT(_net_iflib, OID_AUTO, min_tx_latency, CTLFLAG_RW,
584 		   &iflib_min_tx_latency, 0, "minimize transmit latency at the possible expense of throughput");
585 static int iflib_no_tx_batch = 0;
586 SYSCTL_INT(_net_iflib, OID_AUTO, no_tx_batch, CTLFLAG_RW,
587 		   &iflib_no_tx_batch, 0, "minimize transmit latency at the possible expense of throughput");
588 static int iflib_timer_default = 1000;
589 SYSCTL_INT(_net_iflib, OID_AUTO, timer_default, CTLFLAG_RW,
590 		   &iflib_timer_default, 0, "number of ticks between iflib_timer calls");
591 
592 
593 #if IFLIB_DEBUG_COUNTERS
594 
595 static int iflib_tx_seen;
596 static int iflib_tx_sent;
597 static int iflib_tx_encap;
598 static int iflib_rx_allocs;
599 static int iflib_fl_refills;
600 static int iflib_fl_refills_large;
601 static int iflib_tx_frees;
602 
603 SYSCTL_INT(_net_iflib, OID_AUTO, tx_seen, CTLFLAG_RD,
604 		   &iflib_tx_seen, 0, "# TX mbufs seen");
605 SYSCTL_INT(_net_iflib, OID_AUTO, tx_sent, CTLFLAG_RD,
606 		   &iflib_tx_sent, 0, "# TX mbufs sent");
607 SYSCTL_INT(_net_iflib, OID_AUTO, tx_encap, CTLFLAG_RD,
608 		   &iflib_tx_encap, 0, "# TX mbufs encapped");
609 SYSCTL_INT(_net_iflib, OID_AUTO, tx_frees, CTLFLAG_RD,
610 		   &iflib_tx_frees, 0, "# TX frees");
611 SYSCTL_INT(_net_iflib, OID_AUTO, rx_allocs, CTLFLAG_RD,
612 		   &iflib_rx_allocs, 0, "# RX allocations");
613 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills, CTLFLAG_RD,
614 		   &iflib_fl_refills, 0, "# refills");
615 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills_large, CTLFLAG_RD,
616 		   &iflib_fl_refills_large, 0, "# large refills");
617 
618 static int iflib_txq_drain_flushing;
619 static int iflib_txq_drain_oactive;
620 static int iflib_txq_drain_notready;
621 
622 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_flushing, CTLFLAG_RD,
623 		   &iflib_txq_drain_flushing, 0, "# drain flushes");
624 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_oactive, CTLFLAG_RD,
625 		   &iflib_txq_drain_oactive, 0, "# drain oactives");
626 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_notready, CTLFLAG_RD,
627 		   &iflib_txq_drain_notready, 0, "# drain notready");
628 
629 static int iflib_encap_load_mbuf_fail;
630 static int iflib_encap_pad_mbuf_fail;
631 static int iflib_encap_txq_avail_fail;
632 static int iflib_encap_txd_encap_fail;
633 
634 SYSCTL_INT(_net_iflib, OID_AUTO, encap_load_mbuf_fail, CTLFLAG_RD,
635 		   &iflib_encap_load_mbuf_fail, 0, "# busdma load failures");
636 SYSCTL_INT(_net_iflib, OID_AUTO, encap_pad_mbuf_fail, CTLFLAG_RD,
637 		   &iflib_encap_pad_mbuf_fail, 0, "# runt frame pad failures");
638 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txq_avail_fail, CTLFLAG_RD,
639 		   &iflib_encap_txq_avail_fail, 0, "# txq avail failures");
640 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txd_encap_fail, CTLFLAG_RD,
641 		   &iflib_encap_txd_encap_fail, 0, "# driver encap failures");
642 
643 static int iflib_task_fn_rxs;
644 static int iflib_rx_intr_enables;
645 static int iflib_fast_intrs;
646 static int iflib_rx_unavail;
647 static int iflib_rx_ctx_inactive;
648 static int iflib_rx_if_input;
649 static int iflib_rxd_flush;
650 
651 static int iflib_verbose_debug;
652 
653 SYSCTL_INT(_net_iflib, OID_AUTO, task_fn_rx, CTLFLAG_RD,
654 		   &iflib_task_fn_rxs, 0, "# task_fn_rx calls");
655 SYSCTL_INT(_net_iflib, OID_AUTO, rx_intr_enables, CTLFLAG_RD,
656 		   &iflib_rx_intr_enables, 0, "# RX intr enables");
657 SYSCTL_INT(_net_iflib, OID_AUTO, fast_intrs, CTLFLAG_RD,
658 		   &iflib_fast_intrs, 0, "# fast_intr calls");
659 SYSCTL_INT(_net_iflib, OID_AUTO, rx_unavail, CTLFLAG_RD,
660 		   &iflib_rx_unavail, 0, "# times rxeof called with no available data");
661 SYSCTL_INT(_net_iflib, OID_AUTO, rx_ctx_inactive, CTLFLAG_RD,
662 		   &iflib_rx_ctx_inactive, 0, "# times rxeof called with inactive context");
663 SYSCTL_INT(_net_iflib, OID_AUTO, rx_if_input, CTLFLAG_RD,
664 		   &iflib_rx_if_input, 0, "# times rxeof called if_input");
665 SYSCTL_INT(_net_iflib, OID_AUTO, rxd_flush, CTLFLAG_RD,
666 	         &iflib_rxd_flush, 0, "# times rxd_flush called");
667 SYSCTL_INT(_net_iflib, OID_AUTO, verbose_debug, CTLFLAG_RW,
668 		   &iflib_verbose_debug, 0, "enable verbose debugging");
669 
670 #define DBG_COUNTER_INC(name) atomic_add_int(&(iflib_ ## name), 1)
671 static void
672 iflib_debug_reset(void)
673 {
674 	iflib_tx_seen = iflib_tx_sent = iflib_tx_encap = iflib_rx_allocs =
675 		iflib_fl_refills = iflib_fl_refills_large = iflib_tx_frees =
676 		iflib_txq_drain_flushing = iflib_txq_drain_oactive =
677 		iflib_txq_drain_notready =
678 		iflib_encap_load_mbuf_fail = iflib_encap_pad_mbuf_fail =
679 		iflib_encap_txq_avail_fail = iflib_encap_txd_encap_fail =
680 		iflib_task_fn_rxs = iflib_rx_intr_enables = iflib_fast_intrs =
681 		iflib_rx_unavail =
682 		iflib_rx_ctx_inactive = iflib_rx_if_input =
683 		iflib_rxd_flush = 0;
684 }
685 
686 #else
687 #define DBG_COUNTER_INC(name)
688 static void iflib_debug_reset(void) {}
689 #endif
690 
691 #define IFLIB_DEBUG 0
692 
693 static void iflib_tx_structures_free(if_ctx_t ctx);
694 static void iflib_rx_structures_free(if_ctx_t ctx);
695 static int iflib_queues_alloc(if_ctx_t ctx);
696 static int iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq);
697 static int iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget);
698 static int iflib_qset_structures_setup(if_ctx_t ctx);
699 static int iflib_msix_init(if_ctx_t ctx);
700 static int iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filterarg, int *rid, const char *str);
701 static void iflib_txq_check_drain(iflib_txq_t txq, int budget);
702 static uint32_t iflib_txq_can_drain(struct ifmp_ring *);
703 #ifdef ALTQ
704 static void iflib_altq_if_start(if_t ifp);
705 static int iflib_altq_if_transmit(if_t ifp, struct mbuf *m);
706 #endif
707 static int iflib_register(if_ctx_t);
708 static void iflib_deregister(if_ctx_t);
709 static void iflib_unregister_vlan_handlers(if_ctx_t ctx);
710 static uint16_t iflib_get_mbuf_size_for(unsigned int size);
711 static void iflib_init_locked(if_ctx_t ctx);
712 static void iflib_add_device_sysctl_pre(if_ctx_t ctx);
713 static void iflib_add_device_sysctl_post(if_ctx_t ctx);
714 static void iflib_ifmp_purge(iflib_txq_t txq);
715 static void _iflib_pre_assert(if_softc_ctx_t scctx);
716 static void iflib_if_init_locked(if_ctx_t ctx);
717 static void iflib_free_intr_mem(if_ctx_t ctx);
718 #ifndef __NO_STRICT_ALIGNMENT
719 static struct mbuf * iflib_fixup_rx(struct mbuf *m);
720 #endif
721 
722 static SLIST_HEAD(cpu_offset_list, cpu_offset) cpu_offsets =
723     SLIST_HEAD_INITIALIZER(cpu_offsets);
724 struct cpu_offset {
725 	SLIST_ENTRY(cpu_offset) entries;
726 	cpuset_t	set;
727 	unsigned int	refcount;
728 	uint16_t	offset;
729 };
730 static struct mtx cpu_offset_mtx;
731 MTX_SYSINIT(iflib_cpu_offset, &cpu_offset_mtx, "iflib_cpu_offset lock",
732     MTX_DEF);
733 
734 DEBUGNET_DEFINE(iflib);
735 
736 static int
737 iflib_num_rx_descs(if_ctx_t ctx)
738 {
739 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
740 	if_shared_ctx_t sctx = ctx->ifc_sctx;
741 	uint16_t first_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
742 
743 	return scctx->isc_nrxd[first_rxq];
744 }
745 
746 static int
747 iflib_num_tx_descs(if_ctx_t ctx)
748 {
749 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
750 	if_shared_ctx_t sctx = ctx->ifc_sctx;
751 	uint16_t first_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
752 
753 	return scctx->isc_ntxd[first_txq];
754 }
755 
756 #ifdef DEV_NETMAP
757 #include <sys/selinfo.h>
758 #include <net/netmap.h>
759 #include <dev/netmap/netmap_kern.h>
760 
761 MODULE_DEPEND(iflib, netmap, 1, 1, 1);
762 
763 static int netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, bool init);
764 static void iflib_netmap_timer(void *arg);
765 
766 /*
767  * device-specific sysctl variables:
768  *
769  * iflib_crcstrip: 0: keep CRC in rx frames (default), 1: strip it.
770  *	During regular operations the CRC is stripped, but on some
771  *	hardware reception of frames not multiple of 64 is slower,
772  *	so using crcstrip=0 helps in benchmarks.
773  *
774  * iflib_rx_miss, iflib_rx_miss_bufs:
775  *	count packets that might be missed due to lost interrupts.
776  */
777 SYSCTL_DECL(_dev_netmap);
778 /*
779  * The xl driver by default strips CRCs and we do not override it.
780  */
781 
782 int iflib_crcstrip = 1;
783 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_crcstrip,
784     CTLFLAG_RW, &iflib_crcstrip, 1, "strip CRC on RX frames");
785 
786 int iflib_rx_miss, iflib_rx_miss_bufs;
787 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss,
788     CTLFLAG_RW, &iflib_rx_miss, 0, "potentially missed RX intr");
789 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss_bufs,
790     CTLFLAG_RW, &iflib_rx_miss_bufs, 0, "potentially missed RX intr bufs");
791 
792 /*
793  * Register/unregister. We are already under netmap lock.
794  * Only called on the first register or the last unregister.
795  */
796 static int
797 iflib_netmap_register(struct netmap_adapter *na, int onoff)
798 {
799 	if_t ifp = na->ifp;
800 	if_ctx_t ctx = ifp->if_softc;
801 	int status;
802 
803 	CTX_LOCK(ctx);
804 	IFDI_INTR_DISABLE(ctx);
805 
806 	/* Tell the stack that the interface is no longer active */
807 	ifp->if_drv_flags &= ~(IFF_DRV_RUNNING | IFF_DRV_OACTIVE);
808 
809 	if (!CTX_IS_VF(ctx))
810 		IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip);
811 
812 	iflib_stop(ctx);
813 
814 	/*
815 	 * Enable (or disable) netmap flags, and intercept (or restore)
816 	 * ifp->if_transmit. This is done once the device has been stopped
817 	 * to prevent race conditions.
818 	 */
819 	if (onoff) {
820 		nm_set_native_flags(na);
821 	} else {
822 		nm_clear_native_flags(na);
823 	}
824 
825 	iflib_init_locked(ctx);
826 	IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip); // XXX why twice ?
827 	status = ifp->if_drv_flags & IFF_DRV_RUNNING ? 0 : 1;
828 	if (status)
829 		nm_clear_native_flags(na);
830 	CTX_UNLOCK(ctx);
831 	return (status);
832 }
833 
834 static int
835 netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, bool init)
836 {
837 	struct netmap_adapter *na = kring->na;
838 	u_int const lim = kring->nkr_num_slots - 1;
839 	u_int nm_i = kring->nr_hwcur;
840 	struct netmap_ring *ring = kring->ring;
841 	bus_dmamap_t *map;
842 	struct if_rxd_update iru;
843 	if_ctx_t ctx = rxq->ifr_ctx;
844 	iflib_fl_t fl = &rxq->ifr_fl[0];
845 	u_int nic_i_first, nic_i;
846 	int i, n;
847 #if IFLIB_DEBUG_COUNTERS
848 	int rf_count = 0;
849 #endif
850 
851 	/*
852 	 * This function is used both at initialization and in rxsync.
853 	 * At initialization we need to prepare (with isc_rxd_refill())
854 	 * all the (N) netmap buffers in the ring, in such a way to keep
855 	 * fl->ifl_pidx and kring->nr_hwcur in sync (except for
856 	 * kring->nkr_hwofs); at rxsync time, both indexes point to the
857 	 * next buffer to be refilled.
858 	 * In any case we publish (with isc_rxd_flush()) up to
859 	 * (fl->ifl_pidx - 1) % N (included), to avoid the NIC tail/prod
860 	 * pointer to overrun the head/cons pointer, although this is
861 	 * not necessary for some NICs (e.g. vmx).
862 	 */
863 	if (__predict_false(init))
864 		n = kring->nkr_num_slots;
865 	else {
866 		n = kring->rhead - nm_i;
867 		if (n == 0)
868 			return (0); /* Nothing to do. */
869 		if (n < 0)
870 			n += kring->nkr_num_slots;
871 	}
872 
873 	/* Start to refill from nr_hwcur, publishing n buffers. */
874 	iru_init(&iru, rxq, 0 /* flid */);
875 	map = fl->ifl_sds.ifsd_map;
876 	nic_i = fl->ifl_pidx;
877 	MPASS(nic_i == netmap_idx_k2n(kring, nm_i));
878 	DBG_COUNTER_INC(fl_refills);
879 	while (n > 0) {
880 #if IFLIB_DEBUG_COUNTERS
881 		if (++rf_count == 9)
882 			DBG_COUNTER_INC(fl_refills_large);
883 #endif
884 		nic_i_first = nic_i;
885 		for (i = 0; n > 0 && i < IFLIB_MAX_RX_REFRESH; n--, i++) {
886 			struct netmap_slot *slot = &ring->slot[nm_i];
887 			void *addr = PNMB(na, slot, &fl->ifl_bus_addrs[i]);
888 
889 			MPASS(i < IFLIB_MAX_RX_REFRESH);
890 
891 			if (addr == NETMAP_BUF_BASE(na)) /* bad buf */
892 			        return netmap_ring_reinit(kring);
893 
894 			fl->ifl_rxd_idxs[i] = nic_i;
895 
896 			if (__predict_false(init)) {
897 				netmap_load_map(na, fl->ifl_buf_tag,
898 				    map[nic_i], addr);
899 			} else if (slot->flags & NS_BUF_CHANGED) {
900 				/* buffer has changed, reload map */
901 				netmap_reload_map(na, fl->ifl_buf_tag,
902 				    map[nic_i], addr);
903 			}
904 			bus_dmamap_sync(fl->ifl_buf_tag, map[nic_i],
905 			    BUS_DMASYNC_PREREAD);
906 			slot->flags &= ~NS_BUF_CHANGED;
907 
908 			nm_i = nm_next(nm_i, lim);
909 			nic_i = nm_next(nic_i, lim);
910 		}
911 
912 		iru.iru_pidx = nic_i_first;
913 		iru.iru_count = i;
914 		ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
915 	}
916 	fl->ifl_pidx = nic_i;
917 	MPASS(!init || nm_i == 0);
918 	MPASS(nm_i == kring->rhead);
919 	kring->nr_hwcur = nm_i;
920 
921 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
922 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
923 	ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id,
924 	    nm_prev(nic_i, lim));
925 	DBG_COUNTER_INC(rxd_flush);
926 
927 	return (0);
928 }
929 
930 #define NETMAP_TX_TIMER_US	90
931 
932 /*
933  * Reconcile kernel and user view of the transmit ring.
934  *
935  * All information is in the kring.
936  * Userspace wants to send packets up to the one before kring->rhead,
937  * kernel knows kring->nr_hwcur is the first unsent packet.
938  *
939  * Here we push packets out (as many as possible), and possibly
940  * reclaim buffers from previously completed transmission.
941  *
942  * The caller (netmap) guarantees that there is only one instance
943  * running at any time. Any interference with other driver
944  * methods should be handled by the individual drivers.
945  */
946 static int
947 iflib_netmap_txsync(struct netmap_kring *kring, int flags)
948 {
949 	struct netmap_adapter *na = kring->na;
950 	if_t ifp = na->ifp;
951 	struct netmap_ring *ring = kring->ring;
952 	u_int nm_i;	/* index into the netmap kring */
953 	u_int nic_i;	/* index into the NIC ring */
954 	u_int n;
955 	u_int const lim = kring->nkr_num_slots - 1;
956 	u_int const head = kring->rhead;
957 	struct if_pkt_info pi;
958 
959 	/*
960 	 * interrupts on every tx packet are expensive so request
961 	 * them every half ring, or where NS_REPORT is set
962 	 */
963 	u_int report_frequency = kring->nkr_num_slots >> 1;
964 	/* device-specific */
965 	if_ctx_t ctx = ifp->if_softc;
966 	iflib_txq_t txq = &ctx->ifc_txqs[kring->ring_id];
967 
968 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
969 	    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
970 
971 	/*
972 	 * First part: process new packets to send.
973 	 * nm_i is the current index in the netmap kring,
974 	 * nic_i is the corresponding index in the NIC ring.
975 	 *
976 	 * If we have packets to send (nm_i != head)
977 	 * iterate over the netmap ring, fetch length and update
978 	 * the corresponding slot in the NIC ring. Some drivers also
979 	 * need to update the buffer's physical address in the NIC slot
980 	 * even NS_BUF_CHANGED is not set (PNMB computes the addresses).
981 	 *
982 	 * The netmap_reload_map() calls is especially expensive,
983 	 * even when (as in this case) the tag is 0, so do only
984 	 * when the buffer has actually changed.
985 	 *
986 	 * If possible do not set the report/intr bit on all slots,
987 	 * but only a few times per ring or when NS_REPORT is set.
988 	 *
989 	 * Finally, on 10G and faster drivers, it might be useful
990 	 * to prefetch the next slot and txr entry.
991 	 */
992 
993 	nm_i = kring->nr_hwcur;
994 	if (nm_i != head) {	/* we have new packets to send */
995 		pkt_info_zero(&pi);
996 		pi.ipi_segs = txq->ift_segs;
997 		pi.ipi_qsidx = kring->ring_id;
998 		nic_i = netmap_idx_k2n(kring, nm_i);
999 
1000 		__builtin_prefetch(&ring->slot[nm_i]);
1001 		__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i]);
1002 		__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i]);
1003 
1004 		for (n = 0; nm_i != head; n++) {
1005 			struct netmap_slot *slot = &ring->slot[nm_i];
1006 			u_int len = slot->len;
1007 			uint64_t paddr;
1008 			void *addr = PNMB(na, slot, &paddr);
1009 			int flags = (slot->flags & NS_REPORT ||
1010 				nic_i == 0 || nic_i == report_frequency) ?
1011 				IPI_TX_INTR : 0;
1012 
1013 			/* device-specific */
1014 			pi.ipi_len = len;
1015 			pi.ipi_segs[0].ds_addr = paddr;
1016 			pi.ipi_segs[0].ds_len = len;
1017 			pi.ipi_nsegs = 1;
1018 			pi.ipi_ndescs = 0;
1019 			pi.ipi_pidx = nic_i;
1020 			pi.ipi_flags = flags;
1021 
1022 			/* Fill the slot in the NIC ring. */
1023 			ctx->isc_txd_encap(ctx->ifc_softc, &pi);
1024 			DBG_COUNTER_INC(tx_encap);
1025 
1026 			/* prefetch for next round */
1027 			__builtin_prefetch(&ring->slot[nm_i + 1]);
1028 			__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i + 1]);
1029 			__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i + 1]);
1030 
1031 			NM_CHECK_ADDR_LEN(na, addr, len);
1032 
1033 			if (slot->flags & NS_BUF_CHANGED) {
1034 				/* buffer has changed, reload map */
1035 				netmap_reload_map(na, txq->ift_buf_tag,
1036 				    txq->ift_sds.ifsd_map[nic_i], addr);
1037 			}
1038 			/* make sure changes to the buffer are synced */
1039 			bus_dmamap_sync(txq->ift_buf_tag,
1040 			    txq->ift_sds.ifsd_map[nic_i],
1041 			    BUS_DMASYNC_PREWRITE);
1042 
1043 			slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED);
1044 			nm_i = nm_next(nm_i, lim);
1045 			nic_i = nm_next(nic_i, lim);
1046 		}
1047 		kring->nr_hwcur = nm_i;
1048 
1049 		/* synchronize the NIC ring */
1050 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1051 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1052 
1053 		/* (re)start the tx unit up to slot nic_i (excluded) */
1054 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, nic_i);
1055 	}
1056 
1057 	/*
1058 	 * Second part: reclaim buffers for completed transmissions.
1059 	 *
1060 	 * If there are unclaimed buffers, attempt to reclaim them.
1061 	 * If we don't manage to reclaim them all, and TX IRQs are not in use,
1062 	 * trigger a per-tx-queue timer to try again later.
1063 	 */
1064 	if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1065 		if (iflib_tx_credits_update(ctx, txq)) {
1066 			/* some tx completed, increment avail */
1067 			nic_i = txq->ift_cidx_processed;
1068 			kring->nr_hwtail = nm_prev(netmap_idx_n2k(kring, nic_i), lim);
1069 		}
1070 	}
1071 
1072 	if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ))
1073 		if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1074 			callout_reset_sbt_on(&txq->ift_netmap_timer,
1075 			    NETMAP_TX_TIMER_US * SBT_1US, SBT_1US,
1076 			    iflib_netmap_timer, txq,
1077 			    txq->ift_netmap_timer.c_cpu, 0);
1078 		}
1079 	return (0);
1080 }
1081 
1082 /*
1083  * Reconcile kernel and user view of the receive ring.
1084  * Same as for the txsync, this routine must be efficient.
1085  * The caller guarantees a single invocations, but races against
1086  * the rest of the driver should be handled here.
1087  *
1088  * On call, kring->rhead is the first packet that userspace wants
1089  * to keep, and kring->rcur is the wakeup point.
1090  * The kernel has previously reported packets up to kring->rtail.
1091  *
1092  * If (flags & NAF_FORCE_READ) also check for incoming packets irrespective
1093  * of whether or not we received an interrupt.
1094  */
1095 static int
1096 iflib_netmap_rxsync(struct netmap_kring *kring, int flags)
1097 {
1098 	struct netmap_adapter *na = kring->na;
1099 	struct netmap_ring *ring = kring->ring;
1100 	if_t ifp = na->ifp;
1101 	uint32_t nm_i;	/* index into the netmap ring */
1102 	uint32_t nic_i;	/* index into the NIC ring */
1103 	u_int n;
1104 	u_int const lim = kring->nkr_num_slots - 1;
1105 	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
1106 
1107 	if_ctx_t ctx = ifp->if_softc;
1108 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1109 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1110 	iflib_rxq_t rxq = &ctx->ifc_rxqs[kring->ring_id];
1111 	iflib_fl_t fl = &rxq->ifr_fl[0];
1112 	struct if_rxd_info ri;
1113 	qidx_t *cidxp;
1114 
1115 	/*
1116 	 * netmap only uses free list 0, to avoid out of order consumption
1117 	 * of receive buffers
1118 	 */
1119 
1120 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
1121 	    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1122 
1123 	/*
1124 	 * First part: import newly received packets.
1125 	 *
1126 	 * nm_i is the index of the next free slot in the netmap ring,
1127 	 * nic_i is the index of the next received packet in the NIC ring
1128 	 * (or in the free list 0 if IFLIB_HAS_RXCQ is set), and they may
1129 	 * differ in case if_init() has been called while
1130 	 * in netmap mode. For the receive ring we have
1131 	 *
1132 	 *	nic_i = fl->ifl_cidx;
1133 	 *	nm_i = kring->nr_hwtail (previous)
1134 	 * and
1135 	 *	nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1136 	 *
1137 	 * fl->ifl_cidx is set to 0 on a ring reinit
1138 	 */
1139 	if (netmap_no_pendintr || force_update) {
1140 		uint32_t hwtail_lim = nm_prev(kring->nr_hwcur, lim);
1141 		bool have_rxcq = sctx->isc_flags & IFLIB_HAS_RXCQ;
1142 		int crclen = iflib_crcstrip ? 0 : 4;
1143 		int error, avail;
1144 
1145 		/*
1146 		 * For the free list consumer index, we use the same
1147 		 * logic as in iflib_rxeof().
1148 		 */
1149 		if (have_rxcq)
1150 			cidxp = &rxq->ifr_cq_cidx;
1151 		else
1152 			cidxp = &fl->ifl_cidx;
1153 		avail = ctx->isc_rxd_available(ctx->ifc_softc,
1154 		    rxq->ifr_id, *cidxp, USHRT_MAX);
1155 
1156 		nic_i = fl->ifl_cidx;
1157 		nm_i = netmap_idx_n2k(kring, nic_i);
1158 		MPASS(nm_i == kring->nr_hwtail);
1159 		for (n = 0; avail > 0 && nm_i != hwtail_lim; n++, avail--) {
1160 			rxd_info_zero(&ri);
1161 			ri.iri_frags = rxq->ifr_frags;
1162 			ri.iri_qsidx = kring->ring_id;
1163 			ri.iri_ifp = ctx->ifc_ifp;
1164 			ri.iri_cidx = *cidxp;
1165 
1166 			error = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
1167 			ring->slot[nm_i].len = error ? 0 : ri.iri_len - crclen;
1168 			ring->slot[nm_i].flags = 0;
1169 			if (have_rxcq) {
1170 				*cidxp = ri.iri_cidx;
1171 				while (*cidxp >= scctx->isc_nrxd[0])
1172 					*cidxp -= scctx->isc_nrxd[0];
1173 			}
1174 			bus_dmamap_sync(fl->ifl_buf_tag,
1175 			    fl->ifl_sds.ifsd_map[nic_i], BUS_DMASYNC_POSTREAD);
1176 			nm_i = nm_next(nm_i, lim);
1177 			fl->ifl_cidx = nic_i = nm_next(nic_i, lim);
1178 		}
1179 		if (n) { /* update the state variables */
1180 			if (netmap_no_pendintr && !force_update) {
1181 				/* diagnostics */
1182 				iflib_rx_miss ++;
1183 				iflib_rx_miss_bufs += n;
1184 			}
1185 			kring->nr_hwtail = nm_i;
1186 		}
1187 		kring->nr_kflags &= ~NKR_PENDINTR;
1188 	}
1189 	/*
1190 	 * Second part: skip past packets that userspace has released.
1191 	 * (kring->nr_hwcur to head excluded),
1192 	 * and make the buffers available for reception.
1193 	 * As usual nm_i is the index in the netmap ring,
1194 	 * nic_i is the index in the NIC ring, and
1195 	 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1196 	 */
1197 	netmap_fl_refill(rxq, kring, false);
1198 
1199 	return (0);
1200 }
1201 
1202 static void
1203 iflib_netmap_intr(struct netmap_adapter *na, int onoff)
1204 {
1205 	if_ctx_t ctx = na->ifp->if_softc;
1206 
1207 	CTX_LOCK(ctx);
1208 	if (onoff) {
1209 		IFDI_INTR_ENABLE(ctx);
1210 	} else {
1211 		IFDI_INTR_DISABLE(ctx);
1212 	}
1213 	CTX_UNLOCK(ctx);
1214 }
1215 
1216 static int
1217 iflib_netmap_attach(if_ctx_t ctx)
1218 {
1219 	struct netmap_adapter na;
1220 
1221 	bzero(&na, sizeof(na));
1222 
1223 	na.ifp = ctx->ifc_ifp;
1224 	na.na_flags = NAF_BDG_MAYSLEEP;
1225 	MPASS(ctx->ifc_softc_ctx.isc_ntxqsets);
1226 	MPASS(ctx->ifc_softc_ctx.isc_nrxqsets);
1227 
1228 	na.num_tx_desc = iflib_num_tx_descs(ctx);
1229 	na.num_rx_desc = iflib_num_rx_descs(ctx);
1230 	na.nm_txsync = iflib_netmap_txsync;
1231 	na.nm_rxsync = iflib_netmap_rxsync;
1232 	na.nm_register = iflib_netmap_register;
1233 	na.nm_intr = iflib_netmap_intr;
1234 	na.num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
1235 	na.num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
1236 	return (netmap_attach(&na));
1237 }
1238 
1239 static int
1240 iflib_netmap_txq_init(if_ctx_t ctx, iflib_txq_t txq)
1241 {
1242 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1243 	struct netmap_slot *slot;
1244 
1245 	slot = netmap_reset(na, NR_TX, txq->ift_id, 0);
1246 	if (slot == NULL)
1247 		return (0);
1248 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxd[0]; i++) {
1249 		/*
1250 		 * In netmap mode, set the map for the packet buffer.
1251 		 * NOTE: Some drivers (not this one) also need to set
1252 		 * the physical buffer address in the NIC ring.
1253 		 * netmap_idx_n2k() maps a nic index, i, into the corresponding
1254 		 * netmap slot index, si
1255 		 */
1256 		int si = netmap_idx_n2k(na->tx_rings[txq->ift_id], i);
1257 		netmap_load_map(na, txq->ift_buf_tag, txq->ift_sds.ifsd_map[i],
1258 		    NMB(na, slot + si));
1259 	}
1260 	return (1);
1261 }
1262 
1263 static int
1264 iflib_netmap_rxq_init(if_ctx_t ctx, iflib_rxq_t rxq)
1265 {
1266 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1267 	struct netmap_kring *kring;
1268 	struct netmap_slot *slot;
1269 
1270 	slot = netmap_reset(na, NR_RX, rxq->ifr_id, 0);
1271 	if (slot == NULL)
1272 		return (0);
1273 	kring = na->rx_rings[rxq->ifr_id];
1274 	netmap_fl_refill(rxq, kring, true);
1275 	return (1);
1276 }
1277 
1278 static void
1279 iflib_netmap_timer(void *arg)
1280 {
1281 	iflib_txq_t txq = arg;
1282 	if_ctx_t ctx = txq->ift_ctx;
1283 
1284 	/*
1285 	 * Wake up the netmap application, to give it a chance to
1286 	 * call txsync and reclaim more completed TX buffers.
1287 	 */
1288 	netmap_tx_irq(ctx->ifc_ifp, txq->ift_id);
1289 }
1290 
1291 #define iflib_netmap_detach(ifp) netmap_detach(ifp)
1292 
1293 #else
1294 #define iflib_netmap_txq_init(ctx, txq) (0)
1295 #define iflib_netmap_rxq_init(ctx, rxq) (0)
1296 #define iflib_netmap_detach(ifp)
1297 
1298 #define iflib_netmap_attach(ctx) (0)
1299 #define netmap_rx_irq(ifp, qid, budget) (0)
1300 #endif
1301 
1302 #if defined(__i386__) || defined(__amd64__)
1303 static __inline void
1304 prefetch(void *x)
1305 {
1306 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1307 }
1308 static __inline void
1309 prefetch2cachelines(void *x)
1310 {
1311 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1312 #if (CACHE_LINE_SIZE < 128)
1313 	__asm volatile("prefetcht0 %0" :: "m" (*(((unsigned long *)x)+CACHE_LINE_SIZE/(sizeof(unsigned long)))));
1314 #endif
1315 }
1316 #else
1317 #define prefetch(x)
1318 #define prefetch2cachelines(x)
1319 #endif
1320 
1321 static void
1322 iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid)
1323 {
1324 	iflib_fl_t fl;
1325 
1326 	fl = &rxq->ifr_fl[flid];
1327 	iru->iru_paddrs = fl->ifl_bus_addrs;
1328 	iru->iru_idxs = fl->ifl_rxd_idxs;
1329 	iru->iru_qsidx = rxq->ifr_id;
1330 	iru->iru_buf_size = fl->ifl_buf_size;
1331 	iru->iru_flidx = fl->ifl_id;
1332 }
1333 
1334 static void
1335 _iflib_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int err)
1336 {
1337 	if (err)
1338 		return;
1339 	*(bus_addr_t *) arg = segs[0].ds_addr;
1340 }
1341 
1342 int
1343 iflib_dma_alloc_align(if_ctx_t ctx, int size, int align, iflib_dma_info_t dma, int mapflags)
1344 {
1345 	int err;
1346 	device_t dev = ctx->ifc_dev;
1347 
1348 	err = bus_dma_tag_create(bus_get_dma_tag(dev),	/* parent */
1349 				align, 0,		/* alignment, bounds */
1350 				BUS_SPACE_MAXADDR,	/* lowaddr */
1351 				BUS_SPACE_MAXADDR,	/* highaddr */
1352 				NULL, NULL,		/* filter, filterarg */
1353 				size,			/* maxsize */
1354 				1,			/* nsegments */
1355 				size,			/* maxsegsize */
1356 				BUS_DMA_ALLOCNOW,	/* flags */
1357 				NULL,			/* lockfunc */
1358 				NULL,			/* lockarg */
1359 				&dma->idi_tag);
1360 	if (err) {
1361 		device_printf(dev,
1362 		    "%s: bus_dma_tag_create failed: %d\n",
1363 		    __func__, err);
1364 		goto fail_0;
1365 	}
1366 
1367 	err = bus_dmamem_alloc(dma->idi_tag, (void**) &dma->idi_vaddr,
1368 	    BUS_DMA_NOWAIT | BUS_DMA_COHERENT | BUS_DMA_ZERO, &dma->idi_map);
1369 	if (err) {
1370 		device_printf(dev,
1371 		    "%s: bus_dmamem_alloc(%ju) failed: %d\n",
1372 		    __func__, (uintmax_t)size, err);
1373 		goto fail_1;
1374 	}
1375 
1376 	dma->idi_paddr = IF_BAD_DMA;
1377 	err = bus_dmamap_load(dma->idi_tag, dma->idi_map, dma->idi_vaddr,
1378 	    size, _iflib_dmamap_cb, &dma->idi_paddr, mapflags | BUS_DMA_NOWAIT);
1379 	if (err || dma->idi_paddr == IF_BAD_DMA) {
1380 		device_printf(dev,
1381 		    "%s: bus_dmamap_load failed: %d\n",
1382 		    __func__, err);
1383 		goto fail_2;
1384 	}
1385 
1386 	dma->idi_size = size;
1387 	return (0);
1388 
1389 fail_2:
1390 	bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1391 fail_1:
1392 	bus_dma_tag_destroy(dma->idi_tag);
1393 fail_0:
1394 	dma->idi_tag = NULL;
1395 
1396 	return (err);
1397 }
1398 
1399 int
1400 iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags)
1401 {
1402 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1403 
1404 	KASSERT(sctx->isc_q_align != 0, ("alignment value not initialized"));
1405 
1406 	return (iflib_dma_alloc_align(ctx, size, sctx->isc_q_align, dma, mapflags));
1407 }
1408 
1409 int
1410 iflib_dma_alloc_multi(if_ctx_t ctx, int *sizes, iflib_dma_info_t *dmalist, int mapflags, int count)
1411 {
1412 	int i, err;
1413 	iflib_dma_info_t *dmaiter;
1414 
1415 	dmaiter = dmalist;
1416 	for (i = 0; i < count; i++, dmaiter++) {
1417 		if ((err = iflib_dma_alloc(ctx, sizes[i], *dmaiter, mapflags)) != 0)
1418 			break;
1419 	}
1420 	if (err)
1421 		iflib_dma_free_multi(dmalist, i);
1422 	return (err);
1423 }
1424 
1425 void
1426 iflib_dma_free(iflib_dma_info_t dma)
1427 {
1428 	if (dma->idi_tag == NULL)
1429 		return;
1430 	if (dma->idi_paddr != IF_BAD_DMA) {
1431 		bus_dmamap_sync(dma->idi_tag, dma->idi_map,
1432 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1433 		bus_dmamap_unload(dma->idi_tag, dma->idi_map);
1434 		dma->idi_paddr = IF_BAD_DMA;
1435 	}
1436 	if (dma->idi_vaddr != NULL) {
1437 		bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1438 		dma->idi_vaddr = NULL;
1439 	}
1440 	bus_dma_tag_destroy(dma->idi_tag);
1441 	dma->idi_tag = NULL;
1442 }
1443 
1444 void
1445 iflib_dma_free_multi(iflib_dma_info_t *dmalist, int count)
1446 {
1447 	int i;
1448 	iflib_dma_info_t *dmaiter = dmalist;
1449 
1450 	for (i = 0; i < count; i++, dmaiter++)
1451 		iflib_dma_free(*dmaiter);
1452 }
1453 
1454 static int
1455 iflib_fast_intr(void *arg)
1456 {
1457 	iflib_filter_info_t info = arg;
1458 	struct grouptask *gtask = info->ifi_task;
1459 	int result;
1460 
1461 	DBG_COUNTER_INC(fast_intrs);
1462 	if (info->ifi_filter != NULL) {
1463 		result = info->ifi_filter(info->ifi_filter_arg);
1464 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1465 			return (result);
1466 	}
1467 
1468 	GROUPTASK_ENQUEUE(gtask);
1469 	return (FILTER_HANDLED);
1470 }
1471 
1472 static int
1473 iflib_fast_intr_rxtx(void *arg)
1474 {
1475 	iflib_filter_info_t info = arg;
1476 	struct grouptask *gtask = info->ifi_task;
1477 	if_ctx_t ctx;
1478 	iflib_rxq_t rxq = (iflib_rxq_t)info->ifi_ctx;
1479 	iflib_txq_t txq;
1480 	void *sc;
1481 	int i, cidx, result;
1482 	qidx_t txqid;
1483 	bool intr_enable, intr_legacy;
1484 
1485 	DBG_COUNTER_INC(fast_intrs);
1486 	if (info->ifi_filter != NULL) {
1487 		result = info->ifi_filter(info->ifi_filter_arg);
1488 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1489 			return (result);
1490 	}
1491 
1492 	ctx = rxq->ifr_ctx;
1493 	sc = ctx->ifc_softc;
1494 	intr_enable = false;
1495 	intr_legacy = !!(ctx->ifc_flags & IFC_LEGACY);
1496 	MPASS(rxq->ifr_ntxqirq);
1497 	for (i = 0; i < rxq->ifr_ntxqirq; i++) {
1498 		txqid = rxq->ifr_txqid[i];
1499 		txq = &ctx->ifc_txqs[txqid];
1500 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1501 		    BUS_DMASYNC_POSTREAD);
1502 		if (!ctx->isc_txd_credits_update(sc, txqid, false)) {
1503 			if (intr_legacy)
1504 				intr_enable = true;
1505 			else
1506 				IFDI_TX_QUEUE_INTR_ENABLE(ctx, txqid);
1507 			continue;
1508 		}
1509 		GROUPTASK_ENQUEUE(&txq->ift_task);
1510 	}
1511 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_RXCQ)
1512 		cidx = rxq->ifr_cq_cidx;
1513 	else
1514 		cidx = rxq->ifr_fl[0].ifl_cidx;
1515 	if (iflib_rxd_avail(ctx, rxq, cidx, 1))
1516 		GROUPTASK_ENQUEUE(gtask);
1517 	else {
1518 		if (intr_legacy)
1519 			intr_enable = true;
1520 		else
1521 			IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
1522 		DBG_COUNTER_INC(rx_intr_enables);
1523 	}
1524 	if (intr_enable)
1525 		IFDI_INTR_ENABLE(ctx);
1526 	return (FILTER_HANDLED);
1527 }
1528 
1529 static int
1530 iflib_fast_intr_ctx(void *arg)
1531 {
1532 	iflib_filter_info_t info = arg;
1533 	struct grouptask *gtask = info->ifi_task;
1534 	int result;
1535 
1536 	DBG_COUNTER_INC(fast_intrs);
1537 	if (info->ifi_filter != NULL) {
1538 		result = info->ifi_filter(info->ifi_filter_arg);
1539 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1540 			return (result);
1541 	}
1542 
1543 	GROUPTASK_ENQUEUE(gtask);
1544 	return (FILTER_HANDLED);
1545 }
1546 
1547 static int
1548 _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
1549 		 driver_filter_t filter, driver_intr_t handler, void *arg,
1550 		 const char *name)
1551 {
1552 	struct resource *res;
1553 	void *tag = NULL;
1554 	device_t dev = ctx->ifc_dev;
1555 	int flags, i, rc;
1556 
1557 	flags = RF_ACTIVE;
1558 	if (ctx->ifc_flags & IFC_LEGACY)
1559 		flags |= RF_SHAREABLE;
1560 	MPASS(rid < 512);
1561 	i = rid;
1562 	res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &i, flags);
1563 	if (res == NULL) {
1564 		device_printf(dev,
1565 		    "failed to allocate IRQ for rid %d, name %s.\n", rid, name);
1566 		return (ENOMEM);
1567 	}
1568 	irq->ii_res = res;
1569 	KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL"));
1570 	rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET,
1571 						filter, handler, arg, &tag);
1572 	if (rc != 0) {
1573 		device_printf(dev,
1574 		    "failed to setup interrupt for rid %d, name %s: %d\n",
1575 					  rid, name ? name : "unknown", rc);
1576 		return (rc);
1577 	} else if (name)
1578 		bus_describe_intr(dev, res, tag, "%s", name);
1579 
1580 	irq->ii_tag = tag;
1581 	return (0);
1582 }
1583 
1584 /*********************************************************************
1585  *
1586  *  Allocate DMA resources for TX buffers as well as memory for the TX
1587  *  mbuf map.  TX DMA maps (non-TSO/TSO) and TX mbuf map are kept in a
1588  *  iflib_sw_tx_desc_array structure, storing all the information that
1589  *  is needed to transmit a packet on the wire.  This is called only
1590  *  once at attach, setup is done every reset.
1591  *
1592  **********************************************************************/
1593 static int
1594 iflib_txsd_alloc(iflib_txq_t txq)
1595 {
1596 	if_ctx_t ctx = txq->ift_ctx;
1597 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1598 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1599 	device_t dev = ctx->ifc_dev;
1600 	bus_size_t tsomaxsize;
1601 	int err, nsegments, ntsosegments;
1602 	bool tso;
1603 
1604 	nsegments = scctx->isc_tx_nsegments;
1605 	ntsosegments = scctx->isc_tx_tso_segments_max;
1606 	tsomaxsize = scctx->isc_tx_tso_size_max;
1607 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_VLAN_MTU)
1608 		tsomaxsize += sizeof(struct ether_vlan_header);
1609 	MPASS(scctx->isc_ntxd[0] > 0);
1610 	MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0);
1611 	MPASS(nsegments > 0);
1612 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) {
1613 		MPASS(ntsosegments > 0);
1614 		MPASS(sctx->isc_tso_maxsize >= tsomaxsize);
1615 	}
1616 
1617 	/*
1618 	 * Set up DMA tags for TX buffers.
1619 	 */
1620 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1621 			       1, 0,			/* alignment, bounds */
1622 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1623 			       BUS_SPACE_MAXADDR,	/* highaddr */
1624 			       NULL, NULL,		/* filter, filterarg */
1625 			       sctx->isc_tx_maxsize,		/* maxsize */
1626 			       nsegments,	/* nsegments */
1627 			       sctx->isc_tx_maxsegsize,	/* maxsegsize */
1628 			       0,			/* flags */
1629 			       NULL,			/* lockfunc */
1630 			       NULL,			/* lockfuncarg */
1631 			       &txq->ift_buf_tag))) {
1632 		device_printf(dev,"Unable to allocate TX DMA tag: %d\n", err);
1633 		device_printf(dev,"maxsize: %ju nsegments: %d maxsegsize: %ju\n",
1634 		    (uintmax_t)sctx->isc_tx_maxsize, nsegments, (uintmax_t)sctx->isc_tx_maxsegsize);
1635 		goto fail;
1636 	}
1637 	tso = (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) != 0;
1638 	if (tso && (err = bus_dma_tag_create(bus_get_dma_tag(dev),
1639 			       1, 0,			/* alignment, bounds */
1640 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1641 			       BUS_SPACE_MAXADDR,	/* highaddr */
1642 			       NULL, NULL,		/* filter, filterarg */
1643 			       tsomaxsize,		/* maxsize */
1644 			       ntsosegments,	/* nsegments */
1645 			       sctx->isc_tso_maxsegsize,/* maxsegsize */
1646 			       0,			/* flags */
1647 			       NULL,			/* lockfunc */
1648 			       NULL,			/* lockfuncarg */
1649 			       &txq->ift_tso_buf_tag))) {
1650 		device_printf(dev, "Unable to allocate TSO TX DMA tag: %d\n",
1651 		    err);
1652 		goto fail;
1653 	}
1654 
1655 	/* Allocate memory for the TX mbuf map. */
1656 	if (!(txq->ift_sds.ifsd_m =
1657 	    (struct mbuf **) malloc(sizeof(struct mbuf *) *
1658 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1659 		device_printf(dev, "Unable to allocate TX mbuf map memory\n");
1660 		err = ENOMEM;
1661 		goto fail;
1662 	}
1663 
1664 	/*
1665 	 * Create the DMA maps for TX buffers.
1666 	 */
1667 	if ((txq->ift_sds.ifsd_map = (bus_dmamap_t *)malloc(
1668 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1669 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1670 		device_printf(dev,
1671 		    "Unable to allocate TX buffer DMA map memory\n");
1672 		err = ENOMEM;
1673 		goto fail;
1674 	}
1675 	if (tso && (txq->ift_sds.ifsd_tso_map = (bus_dmamap_t *)malloc(
1676 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1677 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1678 		device_printf(dev,
1679 		    "Unable to allocate TSO TX buffer map memory\n");
1680 		err = ENOMEM;
1681 		goto fail;
1682 	}
1683 	for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) {
1684 		err = bus_dmamap_create(txq->ift_buf_tag, 0,
1685 		    &txq->ift_sds.ifsd_map[i]);
1686 		if (err != 0) {
1687 			device_printf(dev, "Unable to create TX DMA map\n");
1688 			goto fail;
1689 		}
1690 		if (!tso)
1691 			continue;
1692 		err = bus_dmamap_create(txq->ift_tso_buf_tag, 0,
1693 		    &txq->ift_sds.ifsd_tso_map[i]);
1694 		if (err != 0) {
1695 			device_printf(dev, "Unable to create TSO TX DMA map\n");
1696 			goto fail;
1697 		}
1698 	}
1699 	return (0);
1700 fail:
1701 	/* We free all, it handles case where we are in the middle */
1702 	iflib_tx_structures_free(ctx);
1703 	return (err);
1704 }
1705 
1706 static void
1707 iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i)
1708 {
1709 	bus_dmamap_t map;
1710 
1711 	if (txq->ift_sds.ifsd_map != NULL) {
1712 		map = txq->ift_sds.ifsd_map[i];
1713 		bus_dmamap_sync(txq->ift_buf_tag, map, BUS_DMASYNC_POSTWRITE);
1714 		bus_dmamap_unload(txq->ift_buf_tag, map);
1715 		bus_dmamap_destroy(txq->ift_buf_tag, map);
1716 		txq->ift_sds.ifsd_map[i] = NULL;
1717 	}
1718 
1719 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1720 		map = txq->ift_sds.ifsd_tso_map[i];
1721 		bus_dmamap_sync(txq->ift_tso_buf_tag, map,
1722 		    BUS_DMASYNC_POSTWRITE);
1723 		bus_dmamap_unload(txq->ift_tso_buf_tag, map);
1724 		bus_dmamap_destroy(txq->ift_tso_buf_tag, map);
1725 		txq->ift_sds.ifsd_tso_map[i] = NULL;
1726 	}
1727 }
1728 
1729 static void
1730 iflib_txq_destroy(iflib_txq_t txq)
1731 {
1732 	if_ctx_t ctx = txq->ift_ctx;
1733 
1734 	for (int i = 0; i < txq->ift_size; i++)
1735 		iflib_txsd_destroy(ctx, txq, i);
1736 
1737 	if (txq->ift_br != NULL) {
1738 		ifmp_ring_free(txq->ift_br);
1739 		txq->ift_br = NULL;
1740 	}
1741 
1742 	mtx_destroy(&txq->ift_mtx);
1743 
1744 	if (txq->ift_sds.ifsd_map != NULL) {
1745 		free(txq->ift_sds.ifsd_map, M_IFLIB);
1746 		txq->ift_sds.ifsd_map = NULL;
1747 	}
1748 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1749 		free(txq->ift_sds.ifsd_tso_map, M_IFLIB);
1750 		txq->ift_sds.ifsd_tso_map = NULL;
1751 	}
1752 	if (txq->ift_sds.ifsd_m != NULL) {
1753 		free(txq->ift_sds.ifsd_m, M_IFLIB);
1754 		txq->ift_sds.ifsd_m = NULL;
1755 	}
1756 	if (txq->ift_buf_tag != NULL) {
1757 		bus_dma_tag_destroy(txq->ift_buf_tag);
1758 		txq->ift_buf_tag = NULL;
1759 	}
1760 	if (txq->ift_tso_buf_tag != NULL) {
1761 		bus_dma_tag_destroy(txq->ift_tso_buf_tag);
1762 		txq->ift_tso_buf_tag = NULL;
1763 	}
1764 	if (txq->ift_ifdi != NULL) {
1765 		free(txq->ift_ifdi, M_IFLIB);
1766 	}
1767 }
1768 
1769 static void
1770 iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i)
1771 {
1772 	struct mbuf **mp;
1773 
1774 	mp = &txq->ift_sds.ifsd_m[i];
1775 	if (*mp == NULL)
1776 		return;
1777 
1778 	if (txq->ift_sds.ifsd_map != NULL) {
1779 		bus_dmamap_sync(txq->ift_buf_tag,
1780 		    txq->ift_sds.ifsd_map[i], BUS_DMASYNC_POSTWRITE);
1781 		bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[i]);
1782 	}
1783 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1784 		bus_dmamap_sync(txq->ift_tso_buf_tag,
1785 		    txq->ift_sds.ifsd_tso_map[i], BUS_DMASYNC_POSTWRITE);
1786 		bus_dmamap_unload(txq->ift_tso_buf_tag,
1787 		    txq->ift_sds.ifsd_tso_map[i]);
1788 	}
1789 	m_freem(*mp);
1790 	DBG_COUNTER_INC(tx_frees);
1791 	*mp = NULL;
1792 }
1793 
1794 static int
1795 iflib_txq_setup(iflib_txq_t txq)
1796 {
1797 	if_ctx_t ctx = txq->ift_ctx;
1798 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1799 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1800 	iflib_dma_info_t di;
1801 	int i;
1802 
1803 	/* Set number of descriptors available */
1804 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
1805 	/* XXX make configurable */
1806 	txq->ift_update_freq = IFLIB_DEFAULT_TX_UPDATE_FREQ;
1807 
1808 	/* Reset indices */
1809 	txq->ift_cidx_processed = 0;
1810 	txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0;
1811 	txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset];
1812 
1813 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1814 		bzero((void *)di->idi_vaddr, di->idi_size);
1815 
1816 	IFDI_TXQ_SETUP(ctx, txq->ift_id);
1817 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1818 		bus_dmamap_sync(di->idi_tag, di->idi_map,
1819 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1820 	return (0);
1821 }
1822 
1823 /*********************************************************************
1824  *
1825  *  Allocate DMA resources for RX buffers as well as memory for the RX
1826  *  mbuf map, direct RX cluster pointer map and RX cluster bus address
1827  *  map.  RX DMA map, RX mbuf map, direct RX cluster pointer map and
1828  *  RX cluster map are kept in a iflib_sw_rx_desc_array structure.
1829  *  Since we use use one entry in iflib_sw_rx_desc_array per received
1830  *  packet, the maximum number of entries we'll need is equal to the
1831  *  number of hardware receive descriptors that we've allocated.
1832  *
1833  **********************************************************************/
1834 static int
1835 iflib_rxsd_alloc(iflib_rxq_t rxq)
1836 {
1837 	if_ctx_t ctx = rxq->ifr_ctx;
1838 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1839 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1840 	device_t dev = ctx->ifc_dev;
1841 	iflib_fl_t fl;
1842 	int			err;
1843 
1844 	MPASS(scctx->isc_nrxd[0] > 0);
1845 	MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0);
1846 
1847 	fl = rxq->ifr_fl;
1848 	for (int i = 0; i <  rxq->ifr_nfl; i++, fl++) {
1849 		fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */
1850 		/* Set up DMA tag for RX buffers. */
1851 		err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1852 					 1, 0,			/* alignment, bounds */
1853 					 BUS_SPACE_MAXADDR,	/* lowaddr */
1854 					 BUS_SPACE_MAXADDR,	/* highaddr */
1855 					 NULL, NULL,		/* filter, filterarg */
1856 					 sctx->isc_rx_maxsize,	/* maxsize */
1857 					 sctx->isc_rx_nsegments,	/* nsegments */
1858 					 sctx->isc_rx_maxsegsize,	/* maxsegsize */
1859 					 0,			/* flags */
1860 					 NULL,			/* lockfunc */
1861 					 NULL,			/* lockarg */
1862 					 &fl->ifl_buf_tag);
1863 		if (err) {
1864 			device_printf(dev,
1865 			    "Unable to allocate RX DMA tag: %d\n", err);
1866 			goto fail;
1867 		}
1868 
1869 		/* Allocate memory for the RX mbuf map. */
1870 		if (!(fl->ifl_sds.ifsd_m =
1871 		      (struct mbuf **) malloc(sizeof(struct mbuf *) *
1872 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1873 			device_printf(dev,
1874 			    "Unable to allocate RX mbuf map memory\n");
1875 			err = ENOMEM;
1876 			goto fail;
1877 		}
1878 
1879 		/* Allocate memory for the direct RX cluster pointer map. */
1880 		if (!(fl->ifl_sds.ifsd_cl =
1881 		      (caddr_t *) malloc(sizeof(caddr_t) *
1882 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1883 			device_printf(dev,
1884 			    "Unable to allocate RX cluster map memory\n");
1885 			err = ENOMEM;
1886 			goto fail;
1887 		}
1888 
1889 		/* Allocate memory for the RX cluster bus address map. */
1890 		if (!(fl->ifl_sds.ifsd_ba =
1891 		      (bus_addr_t *) malloc(sizeof(bus_addr_t) *
1892 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1893 			device_printf(dev,
1894 			    "Unable to allocate RX bus address map memory\n");
1895 			err = ENOMEM;
1896 			goto fail;
1897 		}
1898 
1899 		/*
1900 		 * Create the DMA maps for RX buffers.
1901 		 */
1902 		if (!(fl->ifl_sds.ifsd_map =
1903 		      (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1904 			device_printf(dev,
1905 			    "Unable to allocate RX buffer DMA map memory\n");
1906 			err = ENOMEM;
1907 			goto fail;
1908 		}
1909 		for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++) {
1910 			err = bus_dmamap_create(fl->ifl_buf_tag, 0,
1911 			    &fl->ifl_sds.ifsd_map[i]);
1912 			if (err != 0) {
1913 				device_printf(dev, "Unable to create RX buffer DMA map\n");
1914 				goto fail;
1915 			}
1916 		}
1917 	}
1918 	return (0);
1919 
1920 fail:
1921 	iflib_rx_structures_free(ctx);
1922 	return (err);
1923 }
1924 
1925 /*
1926  * Internal service routines
1927  */
1928 
1929 struct rxq_refill_cb_arg {
1930 	int               error;
1931 	bus_dma_segment_t seg;
1932 	int               nseg;
1933 };
1934 
1935 static void
1936 _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
1937 {
1938 	struct rxq_refill_cb_arg *cb_arg = arg;
1939 
1940 	cb_arg->error = error;
1941 	cb_arg->seg = segs[0];
1942 	cb_arg->nseg = nseg;
1943 }
1944 
1945 /**
1946  * iflib_fl_refill - refill an rxq free-buffer list
1947  * @ctx: the iflib context
1948  * @fl: the free list to refill
1949  * @count: the number of new buffers to allocate
1950  *
1951  * (Re)populate an rxq free-buffer list with up to @count new packet buffers.
1952  * The caller must assure that @count does not exceed the queue's capacity
1953  * minus one (since we always leave a descriptor unavailable).
1954  */
1955 static uint8_t
1956 iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
1957 {
1958 	struct if_rxd_update iru;
1959 	struct rxq_refill_cb_arg cb_arg;
1960 	struct mbuf *m;
1961 	caddr_t cl, *sd_cl;
1962 	struct mbuf **sd_m;
1963 	bus_dmamap_t *sd_map;
1964 	bus_addr_t bus_addr, *sd_ba;
1965 	int err, frag_idx, i, idx, n, pidx;
1966 	qidx_t credits;
1967 
1968 	MPASS(count <= fl->ifl_size - fl->ifl_credits - 1);
1969 
1970 	sd_m = fl->ifl_sds.ifsd_m;
1971 	sd_map = fl->ifl_sds.ifsd_map;
1972 	sd_cl = fl->ifl_sds.ifsd_cl;
1973 	sd_ba = fl->ifl_sds.ifsd_ba;
1974 	pidx = fl->ifl_pidx;
1975 	idx = pidx;
1976 	frag_idx = fl->ifl_fragidx;
1977 	credits = fl->ifl_credits;
1978 
1979 	i = 0;
1980 	n = count;
1981 	MPASS(n > 0);
1982 	MPASS(credits + n <= fl->ifl_size);
1983 
1984 	if (pidx < fl->ifl_cidx)
1985 		MPASS(pidx + n <= fl->ifl_cidx);
1986 	if (pidx == fl->ifl_cidx && (credits < fl->ifl_size))
1987 		MPASS(fl->ifl_gen == 0);
1988 	if (pidx > fl->ifl_cidx)
1989 		MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx);
1990 
1991 	DBG_COUNTER_INC(fl_refills);
1992 	if (n > 8)
1993 		DBG_COUNTER_INC(fl_refills_large);
1994 	iru_init(&iru, fl->ifl_rxq, fl->ifl_id);
1995 	while (n-- > 0) {
1996 		/*
1997 		 * We allocate an uninitialized mbuf + cluster, mbuf is
1998 		 * initialized after rx.
1999 		 *
2000 		 * If the cluster is still set then we know a minimum sized
2001 		 * packet was received
2002 		 */
2003 		bit_ffc_at(fl->ifl_rx_bitmap, frag_idx, fl->ifl_size,
2004 		    &frag_idx);
2005 		if (frag_idx < 0)
2006 			bit_ffc(fl->ifl_rx_bitmap, fl->ifl_size, &frag_idx);
2007 		MPASS(frag_idx >= 0);
2008 		if ((cl = sd_cl[frag_idx]) == NULL) {
2009 			cl = uma_zalloc(fl->ifl_zone, M_NOWAIT);
2010 			if (__predict_false(cl == NULL))
2011 				break;
2012 
2013 			cb_arg.error = 0;
2014 			MPASS(sd_map != NULL);
2015 			err = bus_dmamap_load(fl->ifl_buf_tag, sd_map[frag_idx],
2016 			    cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg,
2017 			    BUS_DMA_NOWAIT);
2018 			if (__predict_false(err != 0 || cb_arg.error)) {
2019 				uma_zfree(fl->ifl_zone, cl);
2020 				break;
2021 			}
2022 
2023 			sd_ba[frag_idx] = bus_addr = cb_arg.seg.ds_addr;
2024 			sd_cl[frag_idx] = cl;
2025 #if MEMORY_LOGGING
2026 			fl->ifl_cl_enqueued++;
2027 #endif
2028 		} else {
2029 			bus_addr = sd_ba[frag_idx];
2030 		}
2031 		bus_dmamap_sync(fl->ifl_buf_tag, sd_map[frag_idx],
2032 		    BUS_DMASYNC_PREREAD);
2033 
2034 		if (sd_m[frag_idx] == NULL) {
2035 			m = m_gethdr(M_NOWAIT, MT_NOINIT);
2036 			if (__predict_false(m == NULL))
2037 				break;
2038 			sd_m[frag_idx] = m;
2039 		}
2040 		bit_set(fl->ifl_rx_bitmap, frag_idx);
2041 #if MEMORY_LOGGING
2042 		fl->ifl_m_enqueued++;
2043 #endif
2044 
2045 		DBG_COUNTER_INC(rx_allocs);
2046 		fl->ifl_rxd_idxs[i] = frag_idx;
2047 		fl->ifl_bus_addrs[i] = bus_addr;
2048 		credits++;
2049 		i++;
2050 		MPASS(credits <= fl->ifl_size);
2051 		if (++idx == fl->ifl_size) {
2052 #ifdef INVARIANTS
2053 			fl->ifl_gen = 1;
2054 #endif
2055 			idx = 0;
2056 		}
2057 		if (n == 0 || i == IFLIB_MAX_RX_REFRESH) {
2058 			iru.iru_pidx = pidx;
2059 			iru.iru_count = i;
2060 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2061 			fl->ifl_pidx = idx;
2062 			fl->ifl_credits = credits;
2063 			pidx = idx;
2064 			i = 0;
2065 		}
2066 	}
2067 
2068 	if (n < count - 1) {
2069 		if (i != 0) {
2070 			iru.iru_pidx = pidx;
2071 			iru.iru_count = i;
2072 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2073 			fl->ifl_pidx = idx;
2074 			fl->ifl_credits = credits;
2075 		}
2076 		DBG_COUNTER_INC(rxd_flush);
2077 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2078 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2079 		ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id,
2080 		    fl->ifl_id, fl->ifl_pidx);
2081 		if (__predict_true(bit_test(fl->ifl_rx_bitmap, frag_idx))) {
2082 			fl->ifl_fragidx = frag_idx + 1;
2083 			if (fl->ifl_fragidx == fl->ifl_size)
2084 				fl->ifl_fragidx = 0;
2085 		} else {
2086 			fl->ifl_fragidx = frag_idx;
2087 		}
2088 	}
2089 
2090 	return (n == -1 ? 0 : IFLIB_RXEOF_EMPTY);
2091 }
2092 
2093 static inline uint8_t
2094 iflib_fl_refill_all(if_ctx_t ctx, iflib_fl_t fl)
2095 {
2096 	/*
2097 	 * We leave an unused descriptor to avoid pidx to catch up with cidx.
2098 	 * This is important as it confuses most NICs. For instance,
2099 	 * Intel NICs have (per receive ring) RDH and RDT registers, where
2100 	 * RDH points to the next receive descriptor to be used by the NIC,
2101 	 * and RDT for the next receive descriptor to be published by the
2102 	 * driver to the NIC (RDT - 1 is thus the last valid one).
2103 	 * The condition RDH == RDT means no descriptors are available to
2104 	 * the NIC, and thus it would be ambiguous if it also meant that
2105 	 * all the descriptors are available to the NIC.
2106 	 */
2107 	int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1;
2108 #ifdef INVARIANTS
2109 	int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1;
2110 #endif
2111 
2112 	MPASS(fl->ifl_credits <= fl->ifl_size);
2113 	MPASS(reclaimable == delta);
2114 
2115 	if (reclaimable > 0)
2116 		return (iflib_fl_refill(ctx, fl, reclaimable));
2117 	return (0);
2118 }
2119 
2120 uint8_t
2121 iflib_in_detach(if_ctx_t ctx)
2122 {
2123 	bool in_detach;
2124 
2125 	STATE_LOCK(ctx);
2126 	in_detach = !!(ctx->ifc_flags & IFC_IN_DETACH);
2127 	STATE_UNLOCK(ctx);
2128 	return (in_detach);
2129 }
2130 
2131 static void
2132 iflib_fl_bufs_free(iflib_fl_t fl)
2133 {
2134 	iflib_dma_info_t idi = fl->ifl_ifdi;
2135 	bus_dmamap_t sd_map;
2136 	uint32_t i;
2137 
2138 	for (i = 0; i < fl->ifl_size; i++) {
2139 		struct mbuf **sd_m = &fl->ifl_sds.ifsd_m[i];
2140 		caddr_t *sd_cl = &fl->ifl_sds.ifsd_cl[i];
2141 
2142 		if (*sd_cl != NULL) {
2143 			sd_map = fl->ifl_sds.ifsd_map[i];
2144 			bus_dmamap_sync(fl->ifl_buf_tag, sd_map,
2145 			    BUS_DMASYNC_POSTREAD);
2146 			bus_dmamap_unload(fl->ifl_buf_tag, sd_map);
2147 			uma_zfree(fl->ifl_zone, *sd_cl);
2148 			*sd_cl = NULL;
2149 			if (*sd_m != NULL) {
2150 				m_init(*sd_m, M_NOWAIT, MT_DATA, 0);
2151 				uma_zfree(zone_mbuf, *sd_m);
2152 				*sd_m = NULL;
2153 			}
2154 		} else {
2155 			MPASS(*sd_m == NULL);
2156 		}
2157 #if MEMORY_LOGGING
2158 		fl->ifl_m_dequeued++;
2159 		fl->ifl_cl_dequeued++;
2160 #endif
2161 	}
2162 #ifdef INVARIANTS
2163 	for (i = 0; i < fl->ifl_size; i++) {
2164 		MPASS(fl->ifl_sds.ifsd_cl[i] == NULL);
2165 		MPASS(fl->ifl_sds.ifsd_m[i] == NULL);
2166 	}
2167 #endif
2168 	/*
2169 	 * Reset free list values
2170 	 */
2171 	fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = fl->ifl_fragidx = 0;
2172 	bzero(idi->idi_vaddr, idi->idi_size);
2173 }
2174 
2175 /*********************************************************************
2176  *
2177  *  Initialize a free list and its buffers.
2178  *
2179  **********************************************************************/
2180 static int
2181 iflib_fl_setup(iflib_fl_t fl)
2182 {
2183 	iflib_rxq_t rxq = fl->ifl_rxq;
2184 	if_ctx_t ctx = rxq->ifr_ctx;
2185 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2186 	int qidx;
2187 
2188 	bit_nclear(fl->ifl_rx_bitmap, 0, fl->ifl_size - 1);
2189 	/*
2190 	** Free current RX buffer structs and their mbufs
2191 	*/
2192 	iflib_fl_bufs_free(fl);
2193 	/* Now replenish the mbufs */
2194 	MPASS(fl->ifl_credits == 0);
2195 	qidx = rxq->ifr_fl_offset + fl->ifl_id;
2196 	if (scctx->isc_rxd_buf_size[qidx] != 0)
2197 		fl->ifl_buf_size = scctx->isc_rxd_buf_size[qidx];
2198 	else
2199 		fl->ifl_buf_size = ctx->ifc_rx_mbuf_sz;
2200 	/*
2201 	 * ifl_buf_size may be a driver-supplied value, so pull it up
2202 	 * to the selected mbuf size.
2203 	 */
2204 	fl->ifl_buf_size = iflib_get_mbuf_size_for(fl->ifl_buf_size);
2205 	if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size)
2206 		ctx->ifc_max_fl_buf_size = fl->ifl_buf_size;
2207 	fl->ifl_cltype = m_gettype(fl->ifl_buf_size);
2208 	fl->ifl_zone = m_getzone(fl->ifl_buf_size);
2209 
2210 	/*
2211 	 * Avoid pre-allocating zillions of clusters to an idle card
2212 	 * potentially speeding up attach. In any case make sure
2213 	 * to leave a descriptor unavailable. See the comment in
2214 	 * iflib_fl_refill_all().
2215 	 */
2216 	MPASS(fl->ifl_size > 0);
2217 	(void)iflib_fl_refill(ctx, fl, min(128, fl->ifl_size - 1));
2218 	if (min(128, fl->ifl_size - 1) != fl->ifl_credits)
2219 		return (ENOBUFS);
2220 	/*
2221 	 * handle failure
2222 	 */
2223 	MPASS(rxq != NULL);
2224 	MPASS(fl->ifl_ifdi != NULL);
2225 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2226 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2227 	return (0);
2228 }
2229 
2230 /*********************************************************************
2231  *
2232  *  Free receive ring data structures
2233  *
2234  **********************************************************************/
2235 static void
2236 iflib_rx_sds_free(iflib_rxq_t rxq)
2237 {
2238 	iflib_fl_t fl;
2239 	int i, j;
2240 
2241 	if (rxq->ifr_fl != NULL) {
2242 		for (i = 0; i < rxq->ifr_nfl; i++) {
2243 			fl = &rxq->ifr_fl[i];
2244 			if (fl->ifl_buf_tag != NULL) {
2245 				if (fl->ifl_sds.ifsd_map != NULL) {
2246 					for (j = 0; j < fl->ifl_size; j++) {
2247 						bus_dmamap_sync(
2248 						    fl->ifl_buf_tag,
2249 						    fl->ifl_sds.ifsd_map[j],
2250 						    BUS_DMASYNC_POSTREAD);
2251 						bus_dmamap_unload(
2252 						    fl->ifl_buf_tag,
2253 						    fl->ifl_sds.ifsd_map[j]);
2254 						bus_dmamap_destroy(
2255 						    fl->ifl_buf_tag,
2256 						    fl->ifl_sds.ifsd_map[j]);
2257 					}
2258 				}
2259 				bus_dma_tag_destroy(fl->ifl_buf_tag);
2260 				fl->ifl_buf_tag = NULL;
2261 			}
2262 			free(fl->ifl_sds.ifsd_m, M_IFLIB);
2263 			free(fl->ifl_sds.ifsd_cl, M_IFLIB);
2264 			free(fl->ifl_sds.ifsd_ba, M_IFLIB);
2265 			free(fl->ifl_sds.ifsd_map, M_IFLIB);
2266 			free(fl->ifl_rx_bitmap, M_IFLIB);
2267 			fl->ifl_sds.ifsd_m = NULL;
2268 			fl->ifl_sds.ifsd_cl = NULL;
2269 			fl->ifl_sds.ifsd_ba = NULL;
2270 			fl->ifl_sds.ifsd_map = NULL;
2271 			fl->ifl_rx_bitmap = NULL;
2272 		}
2273 		free(rxq->ifr_fl, M_IFLIB);
2274 		rxq->ifr_fl = NULL;
2275 		free(rxq->ifr_ifdi, M_IFLIB);
2276 		rxq->ifr_ifdi = NULL;
2277 		rxq->ifr_cq_cidx = 0;
2278 	}
2279 }
2280 
2281 /*
2282  * Timer routine
2283  */
2284 static void
2285 iflib_timer(void *arg)
2286 {
2287 	iflib_txq_t txq = arg;
2288 	if_ctx_t ctx = txq->ift_ctx;
2289 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2290 	uint64_t this_tick = ticks;
2291 
2292 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
2293 		return;
2294 
2295 	/*
2296 	** Check on the state of the TX queue(s), this
2297 	** can be done without the lock because its RO
2298 	** and the HUNG state will be static if set.
2299 	*/
2300 	if (this_tick - txq->ift_last_timer_tick >= iflib_timer_default) {
2301 		txq->ift_last_timer_tick = this_tick;
2302 		IFDI_TIMER(ctx, txq->ift_id);
2303 		if ((txq->ift_qstatus == IFLIB_QUEUE_HUNG) &&
2304 		    ((txq->ift_cleaned_prev == txq->ift_cleaned) ||
2305 		     (sctx->isc_pause_frames == 0)))
2306 			goto hung;
2307 
2308 		if (txq->ift_qstatus != IFLIB_QUEUE_IDLE &&
2309 		    ifmp_ring_is_stalled(txq->ift_br)) {
2310 			KASSERT(ctx->ifc_link_state == LINK_STATE_UP,
2311 			    ("queue can't be marked as hung if interface is down"));
2312 			txq->ift_qstatus = IFLIB_QUEUE_HUNG;
2313 		}
2314 		txq->ift_cleaned_prev = txq->ift_cleaned;
2315 	}
2316 	/* handle any laggards */
2317 	if (txq->ift_db_pending)
2318 		GROUPTASK_ENQUEUE(&txq->ift_task);
2319 
2320 	sctx->isc_pause_frames = 0;
2321 	if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)
2322 		callout_reset_on(&txq->ift_timer, iflib_timer_default, iflib_timer,
2323 		    txq, txq->ift_timer.c_cpu);
2324 	return;
2325 
2326  hung:
2327 	device_printf(ctx->ifc_dev,
2328 	    "Watchdog timeout (TX: %d desc avail: %d pidx: %d) -- resetting\n",
2329 	    txq->ift_id, TXQ_AVAIL(txq), txq->ift_pidx);
2330 	STATE_LOCK(ctx);
2331 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2332 	ctx->ifc_flags |= (IFC_DO_WATCHDOG|IFC_DO_RESET);
2333 	iflib_admin_intr_deferred(ctx);
2334 	STATE_UNLOCK(ctx);
2335 }
2336 
2337 static uint16_t
2338 iflib_get_mbuf_size_for(unsigned int size)
2339 {
2340 
2341 	if (size <= MCLBYTES)
2342 		return (MCLBYTES);
2343 	else
2344 		return (MJUMPAGESIZE);
2345 }
2346 
2347 static void
2348 iflib_calc_rx_mbuf_sz(if_ctx_t ctx)
2349 {
2350 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2351 
2352 	/*
2353 	 * XXX don't set the max_frame_size to larger
2354 	 * than the hardware can handle
2355 	 */
2356 	ctx->ifc_rx_mbuf_sz =
2357 	    iflib_get_mbuf_size_for(sctx->isc_max_frame_size);
2358 }
2359 
2360 uint32_t
2361 iflib_get_rx_mbuf_sz(if_ctx_t ctx)
2362 {
2363 
2364 	return (ctx->ifc_rx_mbuf_sz);
2365 }
2366 
2367 static void
2368 iflib_init_locked(if_ctx_t ctx)
2369 {
2370 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2371 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2372 	if_t ifp = ctx->ifc_ifp;
2373 	iflib_fl_t fl;
2374 	iflib_txq_t txq;
2375 	iflib_rxq_t rxq;
2376 	int i, j, tx_ip_csum_flags, tx_ip6_csum_flags;
2377 
2378 	if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2379 	IFDI_INTR_DISABLE(ctx);
2380 
2381 	tx_ip_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP);
2382 	tx_ip6_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | CSUM_IP6_SCTP);
2383 	/* Set hardware offload abilities */
2384 	if_clearhwassist(ifp);
2385 	if (if_getcapenable(ifp) & IFCAP_TXCSUM)
2386 		if_sethwassistbits(ifp, tx_ip_csum_flags, 0);
2387 	if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6)
2388 		if_sethwassistbits(ifp,  tx_ip6_csum_flags, 0);
2389 	if (if_getcapenable(ifp) & IFCAP_TSO4)
2390 		if_sethwassistbits(ifp, CSUM_IP_TSO, 0);
2391 	if (if_getcapenable(ifp) & IFCAP_TSO6)
2392 		if_sethwassistbits(ifp, CSUM_IP6_TSO, 0);
2393 
2394 	for (i = 0, txq = ctx->ifc_txqs; i < sctx->isc_ntxqsets; i++, txq++) {
2395 		CALLOUT_LOCK(txq);
2396 		callout_stop(&txq->ift_timer);
2397 #ifdef DEV_NETMAP
2398 		callout_stop(&txq->ift_netmap_timer);
2399 #endif /* DEV_NETMAP */
2400 		CALLOUT_UNLOCK(txq);
2401 		iflib_netmap_txq_init(ctx, txq);
2402 	}
2403 
2404 	/*
2405 	 * Calculate a suitable Rx mbuf size prior to calling IFDI_INIT, so
2406 	 * that drivers can use the value when setting up the hardware receive
2407 	 * buffers.
2408 	 */
2409 	iflib_calc_rx_mbuf_sz(ctx);
2410 
2411 #ifdef INVARIANTS
2412 	i = if_getdrvflags(ifp);
2413 #endif
2414 	IFDI_INIT(ctx);
2415 	MPASS(if_getdrvflags(ifp) == i);
2416 	for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
2417 		if (iflib_netmap_rxq_init(ctx, rxq) > 0) {
2418 			/* This rxq is in netmap mode. Skip normal init. */
2419 			continue;
2420 		}
2421 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
2422 			if (iflib_fl_setup(fl)) {
2423 				device_printf(ctx->ifc_dev,
2424 				    "setting up free list %d failed - "
2425 				    "check cluster settings\n", j);
2426 				goto done;
2427 			}
2428 		}
2429 	}
2430 done:
2431 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
2432 	IFDI_INTR_ENABLE(ctx);
2433 	txq = ctx->ifc_txqs;
2434 	for (i = 0; i < sctx->isc_ntxqsets; i++, txq++)
2435 		callout_reset_on(&txq->ift_timer, iflib_timer_default, iflib_timer, txq,
2436 			txq->ift_timer.c_cpu);
2437 }
2438 
2439 static int
2440 iflib_media_change(if_t ifp)
2441 {
2442 	if_ctx_t ctx = if_getsoftc(ifp);
2443 	int err;
2444 
2445 	CTX_LOCK(ctx);
2446 	if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0)
2447 		iflib_init_locked(ctx);
2448 	CTX_UNLOCK(ctx);
2449 	return (err);
2450 }
2451 
2452 static void
2453 iflib_media_status(if_t ifp, struct ifmediareq *ifmr)
2454 {
2455 	if_ctx_t ctx = if_getsoftc(ifp);
2456 
2457 	CTX_LOCK(ctx);
2458 	IFDI_UPDATE_ADMIN_STATUS(ctx);
2459 	IFDI_MEDIA_STATUS(ctx, ifmr);
2460 	CTX_UNLOCK(ctx);
2461 }
2462 
2463 void
2464 iflib_stop(if_ctx_t ctx)
2465 {
2466 	iflib_txq_t txq = ctx->ifc_txqs;
2467 	iflib_rxq_t rxq = ctx->ifc_rxqs;
2468 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2469 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2470 	iflib_dma_info_t di;
2471 	iflib_fl_t fl;
2472 	int i, j;
2473 
2474 	/* Tell the stack that the interface is no longer active */
2475 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2476 
2477 	IFDI_INTR_DISABLE(ctx);
2478 	DELAY(1000);
2479 	IFDI_STOP(ctx);
2480 	DELAY(1000);
2481 
2482 	iflib_debug_reset();
2483 	/* Wait for current tx queue users to exit to disarm watchdog timer. */
2484 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) {
2485 		/* make sure all transmitters have completed before proceeding XXX */
2486 
2487 		CALLOUT_LOCK(txq);
2488 		callout_stop(&txq->ift_timer);
2489 #ifdef DEV_NETMAP
2490 		callout_stop(&txq->ift_netmap_timer);
2491 #endif /* DEV_NETMAP */
2492 		CALLOUT_UNLOCK(txq);
2493 
2494 		/* clean any enqueued buffers */
2495 		iflib_ifmp_purge(txq);
2496 		/* Free any existing tx buffers. */
2497 		for (j = 0; j < txq->ift_size; j++) {
2498 			iflib_txsd_free(ctx, txq, j);
2499 		}
2500 		txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0;
2501 		txq->ift_in_use = txq->ift_gen = txq->ift_cidx = txq->ift_pidx = txq->ift_no_desc_avail = 0;
2502 		txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0;
2503 		txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0;
2504 		txq->ift_pullups = 0;
2505 		ifmp_ring_reset_stats(txq->ift_br);
2506 		for (j = 0, di = txq->ift_ifdi; j < sctx->isc_ntxqs; j++, di++)
2507 			bzero((void *)di->idi_vaddr, di->idi_size);
2508 	}
2509 	for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) {
2510 		/* make sure all transmitters have completed before proceeding XXX */
2511 
2512 		rxq->ifr_cq_cidx = 0;
2513 		for (j = 0, di = rxq->ifr_ifdi; j < sctx->isc_nrxqs; j++, di++)
2514 			bzero((void *)di->idi_vaddr, di->idi_size);
2515 		/* also resets the free lists pidx/cidx */
2516 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
2517 			iflib_fl_bufs_free(fl);
2518 	}
2519 }
2520 
2521 static inline caddr_t
2522 calc_next_rxd(iflib_fl_t fl, int cidx)
2523 {
2524 	qidx_t size;
2525 	int nrxd;
2526 	caddr_t start, end, cur, next;
2527 
2528 	nrxd = fl->ifl_size;
2529 	size = fl->ifl_rxd_size;
2530 	start = fl->ifl_ifdi->idi_vaddr;
2531 
2532 	if (__predict_false(size == 0))
2533 		return (start);
2534 	cur = start + size*cidx;
2535 	end = start + size*nrxd;
2536 	next = CACHE_PTR_NEXT(cur);
2537 	return (next < end ? next : start);
2538 }
2539 
2540 static inline void
2541 prefetch_pkts(iflib_fl_t fl, int cidx)
2542 {
2543 	int nextptr;
2544 	int nrxd = fl->ifl_size;
2545 	caddr_t next_rxd;
2546 
2547 	nextptr = (cidx + CACHE_PTR_INCREMENT) & (nrxd-1);
2548 	prefetch(&fl->ifl_sds.ifsd_m[nextptr]);
2549 	prefetch(&fl->ifl_sds.ifsd_cl[nextptr]);
2550 	next_rxd = calc_next_rxd(fl, cidx);
2551 	prefetch(next_rxd);
2552 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 1) & (nrxd-1)]);
2553 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 2) & (nrxd-1)]);
2554 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 3) & (nrxd-1)]);
2555 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 4) & (nrxd-1)]);
2556 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 1) & (nrxd-1)]);
2557 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 2) & (nrxd-1)]);
2558 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 3) & (nrxd-1)]);
2559 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 4) & (nrxd-1)]);
2560 }
2561 
2562 static struct mbuf *
2563 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, bool unload, if_rxsd_t sd,
2564     int *pf_rv, if_rxd_info_t ri)
2565 {
2566 	bus_dmamap_t map;
2567 	iflib_fl_t fl;
2568 	caddr_t payload;
2569 	struct mbuf *m;
2570 	int flid, cidx, len, next;
2571 
2572 	map = NULL;
2573 	flid = irf->irf_flid;
2574 	cidx = irf->irf_idx;
2575 	fl = &rxq->ifr_fl[flid];
2576 	sd->ifsd_fl = fl;
2577 	m = fl->ifl_sds.ifsd_m[cidx];
2578 	sd->ifsd_cl = &fl->ifl_sds.ifsd_cl[cidx];
2579 	fl->ifl_credits--;
2580 #if MEMORY_LOGGING
2581 	fl->ifl_m_dequeued++;
2582 #endif
2583 	if (rxq->ifr_ctx->ifc_flags & IFC_PREFETCH)
2584 		prefetch_pkts(fl, cidx);
2585 	next = (cidx + CACHE_PTR_INCREMENT) & (fl->ifl_size-1);
2586 	prefetch(&fl->ifl_sds.ifsd_map[next]);
2587 	map = fl->ifl_sds.ifsd_map[cidx];
2588 
2589 	bus_dmamap_sync(fl->ifl_buf_tag, map, BUS_DMASYNC_POSTREAD);
2590 
2591 	if (rxq->pfil != NULL && PFIL_HOOKED_IN(rxq->pfil) && pf_rv != NULL &&
2592 	    irf->irf_len != 0) {
2593 		payload  = *sd->ifsd_cl;
2594 		payload +=  ri->iri_pad;
2595 		len = ri->iri_len - ri->iri_pad;
2596 		*pf_rv = pfil_run_hooks(rxq->pfil, payload, ri->iri_ifp,
2597 		    len | PFIL_MEMPTR | PFIL_IN, NULL);
2598 		switch (*pf_rv) {
2599 		case PFIL_DROPPED:
2600 		case PFIL_CONSUMED:
2601 			/*
2602 			 * The filter ate it.  Everything is recycled.
2603 			 */
2604 			m = NULL;
2605 			unload = 0;
2606 			break;
2607 		case PFIL_REALLOCED:
2608 			/*
2609 			 * The filter copied it.  Everything is recycled.
2610 			 */
2611 			m = pfil_mem2mbuf(payload);
2612 			unload = 0;
2613 			break;
2614 		case PFIL_PASS:
2615 			/*
2616 			 * Filter said it was OK, so receive like
2617 			 * normal
2618 			 */
2619 			fl->ifl_sds.ifsd_m[cidx] = NULL;
2620 			break;
2621 		default:
2622 			MPASS(0);
2623 		}
2624 	} else {
2625 		fl->ifl_sds.ifsd_m[cidx] = NULL;
2626 		*pf_rv = PFIL_PASS;
2627 	}
2628 
2629 	if (unload && irf->irf_len != 0)
2630 		bus_dmamap_unload(fl->ifl_buf_tag, map);
2631 	fl->ifl_cidx = (fl->ifl_cidx + 1) & (fl->ifl_size-1);
2632 	if (__predict_false(fl->ifl_cidx == 0))
2633 		fl->ifl_gen = 0;
2634 	bit_clear(fl->ifl_rx_bitmap, cidx);
2635 	return (m);
2636 }
2637 
2638 static struct mbuf *
2639 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri, if_rxsd_t sd, int *pf_rv)
2640 {
2641 	struct mbuf *m, *mh, *mt;
2642 	caddr_t cl;
2643 	int  *pf_rv_ptr, flags, i, padlen;
2644 	bool consumed;
2645 
2646 	i = 0;
2647 	mh = NULL;
2648 	consumed = false;
2649 	*pf_rv = PFIL_PASS;
2650 	pf_rv_ptr = pf_rv;
2651 	do {
2652 		m = rxd_frag_to_sd(rxq, &ri->iri_frags[i], !consumed, sd,
2653 		    pf_rv_ptr, ri);
2654 
2655 		MPASS(*sd->ifsd_cl != NULL);
2656 
2657 		/*
2658 		 * Exclude zero-length frags & frags from
2659 		 * packets the filter has consumed or dropped
2660 		 */
2661 		if (ri->iri_frags[i].irf_len == 0 || consumed ||
2662 		    *pf_rv == PFIL_CONSUMED || *pf_rv == PFIL_DROPPED) {
2663 			if (mh == NULL) {
2664 				/* everything saved here */
2665 				consumed = true;
2666 				pf_rv_ptr = NULL;
2667 				continue;
2668 			}
2669 			/* XXX we can save the cluster here, but not the mbuf */
2670 			m_init(m, M_NOWAIT, MT_DATA, 0);
2671 			m_free(m);
2672 			continue;
2673 		}
2674 		if (mh == NULL) {
2675 			flags = M_PKTHDR|M_EXT;
2676 			mh = mt = m;
2677 			padlen = ri->iri_pad;
2678 		} else {
2679 			flags = M_EXT;
2680 			mt->m_next = m;
2681 			mt = m;
2682 			/* assuming padding is only on the first fragment */
2683 			padlen = 0;
2684 		}
2685 		cl = *sd->ifsd_cl;
2686 		*sd->ifsd_cl = NULL;
2687 
2688 		/* Can these two be made one ? */
2689 		m_init(m, M_NOWAIT, MT_DATA, flags);
2690 		m_cljset(m, cl, sd->ifsd_fl->ifl_cltype);
2691 		/*
2692 		 * These must follow m_init and m_cljset
2693 		 */
2694 		m->m_data += padlen;
2695 		ri->iri_len -= padlen;
2696 		m->m_len = ri->iri_frags[i].irf_len;
2697 	} while (++i < ri->iri_nfrags);
2698 
2699 	return (mh);
2700 }
2701 
2702 /*
2703  * Process one software descriptor
2704  */
2705 static struct mbuf *
2706 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri)
2707 {
2708 	struct if_rxsd sd;
2709 	struct mbuf *m;
2710 	int pf_rv;
2711 
2712 	/* should I merge this back in now that the two paths are basically duplicated? */
2713 	if (ri->iri_nfrags == 1 &&
2714 	    ri->iri_frags[0].irf_len != 0 &&
2715 	    ri->iri_frags[0].irf_len <= MIN(IFLIB_RX_COPY_THRESH, MHLEN)) {
2716 		m = rxd_frag_to_sd(rxq, &ri->iri_frags[0], false, &sd,
2717 		    &pf_rv, ri);
2718 		if (pf_rv != PFIL_PASS && pf_rv != PFIL_REALLOCED)
2719 			return (m);
2720 		if (pf_rv == PFIL_PASS) {
2721 			m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR);
2722 #ifndef __NO_STRICT_ALIGNMENT
2723 			if (!IP_ALIGNED(m))
2724 				m->m_data += 2;
2725 #endif
2726 			memcpy(m->m_data, *sd.ifsd_cl, ri->iri_len);
2727 			m->m_len = ri->iri_frags[0].irf_len;
2728 		}
2729 	} else {
2730 		m = assemble_segments(rxq, ri, &sd, &pf_rv);
2731 		if (m == NULL)
2732 			return (NULL);
2733 		if (pf_rv != PFIL_PASS && pf_rv != PFIL_REALLOCED)
2734 			return (m);
2735 	}
2736 	m->m_pkthdr.len = ri->iri_len;
2737 	m->m_pkthdr.rcvif = ri->iri_ifp;
2738 	m->m_flags |= ri->iri_flags;
2739 	m->m_pkthdr.ether_vtag = ri->iri_vtag;
2740 	m->m_pkthdr.flowid = ri->iri_flowid;
2741 	M_HASHTYPE_SET(m, ri->iri_rsstype);
2742 	m->m_pkthdr.csum_flags = ri->iri_csum_flags;
2743 	m->m_pkthdr.csum_data = ri->iri_csum_data;
2744 	return (m);
2745 }
2746 
2747 #if defined(INET6) || defined(INET)
2748 static void
2749 iflib_get_ip_forwarding(struct lro_ctrl *lc, bool *v4, bool *v6)
2750 {
2751 	CURVNET_SET(lc->ifp->if_vnet);
2752 #if defined(INET6)
2753 	*v6 = V_ip6_forwarding;
2754 #endif
2755 #if defined(INET)
2756 	*v4 = V_ipforwarding;
2757 #endif
2758 	CURVNET_RESTORE();
2759 }
2760 
2761 /*
2762  * Returns true if it's possible this packet could be LROed.
2763  * if it returns false, it is guaranteed that tcp_lro_rx()
2764  * would not return zero.
2765  */
2766 static bool
2767 iflib_check_lro_possible(struct mbuf *m, bool v4_forwarding, bool v6_forwarding)
2768 {
2769 	struct ether_header *eh;
2770 
2771 	eh = mtod(m, struct ether_header *);
2772 	switch (eh->ether_type) {
2773 #if defined(INET6)
2774 		case htons(ETHERTYPE_IPV6):
2775 			return (!v6_forwarding);
2776 #endif
2777 #if defined (INET)
2778 		case htons(ETHERTYPE_IP):
2779 			return (!v4_forwarding);
2780 #endif
2781 	}
2782 
2783 	return false;
2784 }
2785 #else
2786 static void
2787 iflib_get_ip_forwarding(struct lro_ctrl *lc __unused, bool *v4 __unused, bool *v6 __unused)
2788 {
2789 }
2790 #endif
2791 
2792 static void
2793 _task_fn_rx_watchdog(void *context)
2794 {
2795 	iflib_rxq_t rxq = context;
2796 
2797 	GROUPTASK_ENQUEUE(&rxq->ifr_task);
2798 }
2799 
2800 static uint8_t
2801 iflib_rxeof(iflib_rxq_t rxq, qidx_t budget)
2802 {
2803 	if_t ifp;
2804 	if_ctx_t ctx = rxq->ifr_ctx;
2805 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2806 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2807 	int avail, i;
2808 	qidx_t *cidxp;
2809 	struct if_rxd_info ri;
2810 	int err, budget_left, rx_bytes, rx_pkts;
2811 	iflib_fl_t fl;
2812 	int lro_enabled;
2813 	bool v4_forwarding, v6_forwarding, lro_possible;
2814 	uint8_t retval = 0;
2815 
2816 	/*
2817 	 * XXX early demux data packets so that if_input processing only handles
2818 	 * acks in interrupt context
2819 	 */
2820 	struct mbuf *m, *mh, *mt, *mf;
2821 
2822 	NET_EPOCH_ASSERT();
2823 
2824 	lro_possible = v4_forwarding = v6_forwarding = false;
2825 	ifp = ctx->ifc_ifp;
2826 	mh = mt = NULL;
2827 	MPASS(budget > 0);
2828 	rx_pkts	= rx_bytes = 0;
2829 	if (sctx->isc_flags & IFLIB_HAS_RXCQ)
2830 		cidxp = &rxq->ifr_cq_cidx;
2831 	else
2832 		cidxp = &rxq->ifr_fl[0].ifl_cidx;
2833 	if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) {
2834 		for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2835 			retval |= iflib_fl_refill_all(ctx, fl);
2836 		DBG_COUNTER_INC(rx_unavail);
2837 		return (retval);
2838 	}
2839 
2840 	/* pfil needs the vnet to be set */
2841 	CURVNET_SET_QUIET(ifp->if_vnet);
2842 	for (budget_left = budget; budget_left > 0 && avail > 0;) {
2843 		if (__predict_false(!CTX_ACTIVE(ctx))) {
2844 			DBG_COUNTER_INC(rx_ctx_inactive);
2845 			break;
2846 		}
2847 		/*
2848 		 * Reset client set fields to their default values
2849 		 */
2850 		rxd_info_zero(&ri);
2851 		ri.iri_qsidx = rxq->ifr_id;
2852 		ri.iri_cidx = *cidxp;
2853 		ri.iri_ifp = ifp;
2854 		ri.iri_frags = rxq->ifr_frags;
2855 		err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
2856 
2857 		if (err)
2858 			goto err;
2859 		rx_pkts += 1;
2860 		rx_bytes += ri.iri_len;
2861 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
2862 			*cidxp = ri.iri_cidx;
2863 			/* Update our consumer index */
2864 			/* XXX NB: shurd - check if this is still safe */
2865 			while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0])
2866 				rxq->ifr_cq_cidx -= scctx->isc_nrxd[0];
2867 			/* was this only a completion queue message? */
2868 			if (__predict_false(ri.iri_nfrags == 0))
2869 				continue;
2870 		}
2871 		MPASS(ri.iri_nfrags != 0);
2872 		MPASS(ri.iri_len != 0);
2873 
2874 		/* will advance the cidx on the corresponding free lists */
2875 		m = iflib_rxd_pkt_get(rxq, &ri);
2876 		avail--;
2877 		budget_left--;
2878 		if (avail == 0 && budget_left)
2879 			avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left);
2880 
2881 		if (__predict_false(m == NULL))
2882 			continue;
2883 
2884 		/* imm_pkt: -- cxgb */
2885 		if (mh == NULL)
2886 			mh = mt = m;
2887 		else {
2888 			mt->m_nextpkt = m;
2889 			mt = m;
2890 		}
2891 	}
2892 	CURVNET_RESTORE();
2893 	/* make sure that we can refill faster than drain */
2894 	for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2895 		retval |= iflib_fl_refill_all(ctx, fl);
2896 
2897 	lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO);
2898 	if (lro_enabled)
2899 		iflib_get_ip_forwarding(&rxq->ifr_lc, &v4_forwarding, &v6_forwarding);
2900 	mt = mf = NULL;
2901 	while (mh != NULL) {
2902 		m = mh;
2903 		mh = mh->m_nextpkt;
2904 		m->m_nextpkt = NULL;
2905 #ifndef __NO_STRICT_ALIGNMENT
2906 		if (!IP_ALIGNED(m) && (m = iflib_fixup_rx(m)) == NULL)
2907 			continue;
2908 #endif
2909 		rx_bytes += m->m_pkthdr.len;
2910 		rx_pkts++;
2911 #if defined(INET6) || defined(INET)
2912 		if (lro_enabled) {
2913 			if (!lro_possible) {
2914 				lro_possible = iflib_check_lro_possible(m, v4_forwarding, v6_forwarding);
2915 				if (lro_possible && mf != NULL) {
2916 					ifp->if_input(ifp, mf);
2917 					DBG_COUNTER_INC(rx_if_input);
2918 					mt = mf = NULL;
2919 				}
2920 			}
2921 			if ((m->m_pkthdr.csum_flags & (CSUM_L4_CALC|CSUM_L4_VALID)) ==
2922 			    (CSUM_L4_CALC|CSUM_L4_VALID)) {
2923 				if (lro_possible && tcp_lro_rx(&rxq->ifr_lc, m, 0) == 0)
2924 					continue;
2925 			}
2926 		}
2927 #endif
2928 		if (lro_possible) {
2929 			ifp->if_input(ifp, m);
2930 			DBG_COUNTER_INC(rx_if_input);
2931 			continue;
2932 		}
2933 
2934 		if (mf == NULL)
2935 			mf = m;
2936 		if (mt != NULL)
2937 			mt->m_nextpkt = m;
2938 		mt = m;
2939 	}
2940 	if (mf != NULL) {
2941 		ifp->if_input(ifp, mf);
2942 		DBG_COUNTER_INC(rx_if_input);
2943 	}
2944 
2945 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
2946 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
2947 
2948 	/*
2949 	 * Flush any outstanding LRO work
2950 	 */
2951 #if defined(INET6) || defined(INET)
2952 	tcp_lro_flush_all(&rxq->ifr_lc);
2953 #endif
2954 	if (avail != 0 || iflib_rxd_avail(ctx, rxq, *cidxp, 1) != 0)
2955 		retval |= IFLIB_RXEOF_MORE;
2956 	return (retval);
2957 err:
2958 	STATE_LOCK(ctx);
2959 	ctx->ifc_flags |= IFC_DO_RESET;
2960 	iflib_admin_intr_deferred(ctx);
2961 	STATE_UNLOCK(ctx);
2962 	return (0);
2963 }
2964 
2965 #define TXD_NOTIFY_COUNT(txq) (((txq)->ift_size / (txq)->ift_update_freq)-1)
2966 static inline qidx_t
2967 txq_max_db_deferred(iflib_txq_t txq, qidx_t in_use)
2968 {
2969 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2970 	qidx_t minthresh = txq->ift_size / 8;
2971 	if (in_use > 4*minthresh)
2972 		return (notify_count);
2973 	if (in_use > 2*minthresh)
2974 		return (notify_count >> 1);
2975 	if (in_use > minthresh)
2976 		return (notify_count >> 3);
2977 	return (0);
2978 }
2979 
2980 static inline qidx_t
2981 txq_max_rs_deferred(iflib_txq_t txq)
2982 {
2983 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2984 	qidx_t minthresh = txq->ift_size / 8;
2985 	if (txq->ift_in_use > 4*minthresh)
2986 		return (notify_count);
2987 	if (txq->ift_in_use > 2*minthresh)
2988 		return (notify_count >> 1);
2989 	if (txq->ift_in_use > minthresh)
2990 		return (notify_count >> 2);
2991 	return (2);
2992 }
2993 
2994 #define M_CSUM_FLAGS(m) ((m)->m_pkthdr.csum_flags)
2995 #define M_HAS_VLANTAG(m) (m->m_flags & M_VLANTAG)
2996 
2997 #define TXQ_MAX_DB_DEFERRED(txq, in_use) txq_max_db_deferred((txq), (in_use))
2998 #define TXQ_MAX_RS_DEFERRED(txq) txq_max_rs_deferred(txq)
2999 #define TXQ_MAX_DB_CONSUMED(size) (size >> 4)
3000 
3001 /* forward compatibility for cxgb */
3002 #define FIRST_QSET(ctx) 0
3003 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets)
3004 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets)
3005 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx))
3006 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments))
3007 
3008 /* XXX we should be setting this to something other than zero */
3009 #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh)
3010 #define	MAX_TX_DESC(ctx) MAX((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max, \
3011     (ctx)->ifc_softc_ctx.isc_tx_nsegments)
3012 
3013 static inline bool
3014 iflib_txd_db_check(iflib_txq_t txq, int ring)
3015 {
3016 	if_ctx_t ctx = txq->ift_ctx;
3017 	qidx_t dbval, max;
3018 
3019 	max = TXQ_MAX_DB_DEFERRED(txq, txq->ift_in_use);
3020 
3021 	/* force || threshold exceeded || at the edge of the ring */
3022 	if (ring || (txq->ift_db_pending >= max) || (TXQ_AVAIL(txq) <= MAX_TX_DESC(ctx) + 2)) {
3023 
3024 		/*
3025 		 * 'npending' is used if the card's doorbell is in terms of the number of descriptors
3026 		 * pending flush (BRCM). 'pidx' is used in cases where the card's doorbeel uses the
3027 		 * producer index explicitly (INTC).
3028 		 */
3029 		dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx;
3030 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3031 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
3032 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval);
3033 
3034 		/*
3035 		 * Absent bugs there are zero packets pending so reset pending counts to zero.
3036 		 */
3037 		txq->ift_db_pending = txq->ift_npending = 0;
3038 		return (true);
3039 	}
3040 	return (false);
3041 }
3042 
3043 #ifdef PKT_DEBUG
3044 static void
3045 print_pkt(if_pkt_info_t pi)
3046 {
3047 	printf("pi len:  %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n",
3048 	       pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx);
3049 	printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n",
3050 	       pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag);
3051 	printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n",
3052 	       pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto);
3053 }
3054 #endif
3055 
3056 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO)
3057 #define IS_TX_OFFLOAD4(pi) ((pi)->ipi_csum_flags & (CSUM_IP_TCP | CSUM_IP_TSO))
3058 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO)
3059 #define IS_TX_OFFLOAD6(pi) ((pi)->ipi_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_TSO))
3060 
3061 static int
3062 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp)
3063 {
3064 	if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx;
3065 	struct ether_vlan_header *eh;
3066 	struct mbuf *m;
3067 
3068 	m = *mp;
3069 	if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) &&
3070 	    M_WRITABLE(m) == 0) {
3071 		if ((m = m_dup(m, M_NOWAIT)) == NULL) {
3072 			return (ENOMEM);
3073 		} else {
3074 			m_freem(*mp);
3075 			DBG_COUNTER_INC(tx_frees);
3076 			*mp = m;
3077 		}
3078 	}
3079 
3080 	/*
3081 	 * Determine where frame payload starts.
3082 	 * Jump over vlan headers if already present,
3083 	 * helpful for QinQ too.
3084 	 */
3085 	if (__predict_false(m->m_len < sizeof(*eh))) {
3086 		txq->ift_pullups++;
3087 		if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL))
3088 			return (ENOMEM);
3089 	}
3090 	eh = mtod(m, struct ether_vlan_header *);
3091 	if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
3092 		pi->ipi_etype = ntohs(eh->evl_proto);
3093 		pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
3094 	} else {
3095 		pi->ipi_etype = ntohs(eh->evl_encap_proto);
3096 		pi->ipi_ehdrlen = ETHER_HDR_LEN;
3097 	}
3098 
3099 	switch (pi->ipi_etype) {
3100 #ifdef INET
3101 	case ETHERTYPE_IP:
3102 	{
3103 		struct mbuf *n;
3104 		struct ip *ip = NULL;
3105 		struct tcphdr *th = NULL;
3106 		int minthlen;
3107 
3108 		minthlen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip) + sizeof(*th));
3109 		if (__predict_false(m->m_len < minthlen)) {
3110 			/*
3111 			 * if this code bloat is causing too much of a hit
3112 			 * move it to a separate function and mark it noinline
3113 			 */
3114 			if (m->m_len == pi->ipi_ehdrlen) {
3115 				n = m->m_next;
3116 				MPASS(n);
3117 				if (n->m_len >= sizeof(*ip))  {
3118 					ip = (struct ip *)n->m_data;
3119 					if (n->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3120 						th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3121 				} else {
3122 					txq->ift_pullups++;
3123 					if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
3124 						return (ENOMEM);
3125 					ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3126 				}
3127 			} else {
3128 				txq->ift_pullups++;
3129 				if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
3130 					return (ENOMEM);
3131 				ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3132 				if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3133 					th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3134 			}
3135 		} else {
3136 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3137 			if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3138 				th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3139 		}
3140 		pi->ipi_ip_hlen = ip->ip_hl << 2;
3141 		pi->ipi_ipproto = ip->ip_p;
3142 		pi->ipi_flags |= IPI_TX_IPV4;
3143 
3144 		/* TCP checksum offload may require TCP header length */
3145 		if (IS_TX_OFFLOAD4(pi)) {
3146 			if (__predict_true(pi->ipi_ipproto == IPPROTO_TCP)) {
3147 				if (__predict_false(th == NULL)) {
3148 					txq->ift_pullups++;
3149 					if (__predict_false((m = m_pullup(m, (ip->ip_hl << 2) + sizeof(*th))) == NULL))
3150 						return (ENOMEM);
3151 					th = (struct tcphdr *)((caddr_t)ip + pi->ipi_ip_hlen);
3152 				}
3153 				pi->ipi_tcp_hflags = th->th_flags;
3154 				pi->ipi_tcp_hlen = th->th_off << 2;
3155 				pi->ipi_tcp_seq = th->th_seq;
3156 			}
3157 			if (IS_TSO4(pi)) {
3158 				if (__predict_false(ip->ip_p != IPPROTO_TCP))
3159 					return (ENXIO);
3160 				/*
3161 				 * TSO always requires hardware checksum offload.
3162 				 */
3163 				pi->ipi_csum_flags |= (CSUM_IP_TCP | CSUM_IP);
3164 				th->th_sum = in_pseudo(ip->ip_src.s_addr,
3165 						       ip->ip_dst.s_addr, htons(IPPROTO_TCP));
3166 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3167 				if (sctx->isc_flags & IFLIB_TSO_INIT_IP) {
3168 					ip->ip_sum = 0;
3169 					ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz);
3170 				}
3171 			}
3172 		}
3173 		if ((sctx->isc_flags & IFLIB_NEED_ZERO_CSUM) && (pi->ipi_csum_flags & CSUM_IP))
3174                        ip->ip_sum = 0;
3175 
3176 		break;
3177 	}
3178 #endif
3179 #ifdef INET6
3180 	case ETHERTYPE_IPV6:
3181 	{
3182 		struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
3183 		struct tcphdr *th;
3184 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
3185 
3186 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
3187 			txq->ift_pullups++;
3188 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
3189 				return (ENOMEM);
3190 		}
3191 		th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
3192 
3193 		/* XXX-BZ this will go badly in case of ext hdrs. */
3194 		pi->ipi_ipproto = ip6->ip6_nxt;
3195 		pi->ipi_flags |= IPI_TX_IPV6;
3196 
3197 		/* TCP checksum offload may require TCP header length */
3198 		if (IS_TX_OFFLOAD6(pi)) {
3199 			if (pi->ipi_ipproto == IPPROTO_TCP) {
3200 				if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) {
3201 					txq->ift_pullups++;
3202 					if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL))
3203 						return (ENOMEM);
3204 				}
3205 				pi->ipi_tcp_hflags = th->th_flags;
3206 				pi->ipi_tcp_hlen = th->th_off << 2;
3207 				pi->ipi_tcp_seq = th->th_seq;
3208 			}
3209 			if (IS_TSO6(pi)) {
3210 				if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP))
3211 					return (ENXIO);
3212 				/*
3213 				 * TSO always requires hardware checksum offload.
3214 				 */
3215 				pi->ipi_csum_flags |= CSUM_IP6_TCP;
3216 				th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0);
3217 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3218 			}
3219 		}
3220 		break;
3221 	}
3222 #endif
3223 	default:
3224 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
3225 		pi->ipi_ip_hlen = 0;
3226 		break;
3227 	}
3228 	*mp = m;
3229 
3230 	return (0);
3231 }
3232 
3233 /*
3234  * If dodgy hardware rejects the scatter gather chain we've handed it
3235  * we'll need to remove the mbuf chain from ifsg_m[] before we can add the
3236  * m_defrag'd mbufs
3237  */
3238 static __noinline struct mbuf *
3239 iflib_remove_mbuf(iflib_txq_t txq)
3240 {
3241 	int ntxd, pidx;
3242 	struct mbuf *m, **ifsd_m;
3243 
3244 	ifsd_m = txq->ift_sds.ifsd_m;
3245 	ntxd = txq->ift_size;
3246 	pidx = txq->ift_pidx & (ntxd - 1);
3247 	ifsd_m = txq->ift_sds.ifsd_m;
3248 	m = ifsd_m[pidx];
3249 	ifsd_m[pidx] = NULL;
3250 	bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[pidx]);
3251 	if (txq->ift_sds.ifsd_tso_map != NULL)
3252 		bus_dmamap_unload(txq->ift_tso_buf_tag,
3253 		    txq->ift_sds.ifsd_tso_map[pidx]);
3254 #if MEMORY_LOGGING
3255 	txq->ift_dequeued++;
3256 #endif
3257 	return (m);
3258 }
3259 
3260 static inline caddr_t
3261 calc_next_txd(iflib_txq_t txq, int cidx, uint8_t qid)
3262 {
3263 	qidx_t size;
3264 	int ntxd;
3265 	caddr_t start, end, cur, next;
3266 
3267 	ntxd = txq->ift_size;
3268 	size = txq->ift_txd_size[qid];
3269 	start = txq->ift_ifdi[qid].idi_vaddr;
3270 
3271 	if (__predict_false(size == 0))
3272 		return (start);
3273 	cur = start + size*cidx;
3274 	end = start + size*ntxd;
3275 	next = CACHE_PTR_NEXT(cur);
3276 	return (next < end ? next : start);
3277 }
3278 
3279 /*
3280  * Pad an mbuf to ensure a minimum ethernet frame size.
3281  * min_frame_size is the frame size (less CRC) to pad the mbuf to
3282  */
3283 static __noinline int
3284 iflib_ether_pad(device_t dev, struct mbuf **m_head, uint16_t min_frame_size)
3285 {
3286 	/*
3287 	 * 18 is enough bytes to pad an ARP packet to 46 bytes, and
3288 	 * and ARP message is the smallest common payload I can think of
3289 	 */
3290 	static char pad[18];	/* just zeros */
3291 	int n;
3292 	struct mbuf *new_head;
3293 
3294 	if (!M_WRITABLE(*m_head)) {
3295 		new_head = m_dup(*m_head, M_NOWAIT);
3296 		if (new_head == NULL) {
3297 			m_freem(*m_head);
3298 			device_printf(dev, "cannot pad short frame, m_dup() failed");
3299 			DBG_COUNTER_INC(encap_pad_mbuf_fail);
3300 			DBG_COUNTER_INC(tx_frees);
3301 			return ENOMEM;
3302 		}
3303 		m_freem(*m_head);
3304 		*m_head = new_head;
3305 	}
3306 
3307 	for (n = min_frame_size - (*m_head)->m_pkthdr.len;
3308 	     n > 0; n -= sizeof(pad))
3309 		if (!m_append(*m_head, min(n, sizeof(pad)), pad))
3310 			break;
3311 
3312 	if (n > 0) {
3313 		m_freem(*m_head);
3314 		device_printf(dev, "cannot pad short frame\n");
3315 		DBG_COUNTER_INC(encap_pad_mbuf_fail);
3316 		DBG_COUNTER_INC(tx_frees);
3317 		return (ENOBUFS);
3318 	}
3319 
3320 	return 0;
3321 }
3322 
3323 static int
3324 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp)
3325 {
3326 	if_ctx_t		ctx;
3327 	if_shared_ctx_t		sctx;
3328 	if_softc_ctx_t		scctx;
3329 	bus_dma_tag_t		buf_tag;
3330 	bus_dma_segment_t	*segs;
3331 	struct mbuf		*m_head, **ifsd_m;
3332 	void			*next_txd;
3333 	bus_dmamap_t		map;
3334 	struct if_pkt_info	pi;
3335 	int remap = 0;
3336 	int err, nsegs, ndesc, max_segs, pidx, cidx, next, ntxd;
3337 
3338 	ctx = txq->ift_ctx;
3339 	sctx = ctx->ifc_sctx;
3340 	scctx = &ctx->ifc_softc_ctx;
3341 	segs = txq->ift_segs;
3342 	ntxd = txq->ift_size;
3343 	m_head = *m_headp;
3344 	map = NULL;
3345 
3346 	/*
3347 	 * If we're doing TSO the next descriptor to clean may be quite far ahead
3348 	 */
3349 	cidx = txq->ift_cidx;
3350 	pidx = txq->ift_pidx;
3351 	if (ctx->ifc_flags & IFC_PREFETCH) {
3352 		next = (cidx + CACHE_PTR_INCREMENT) & (ntxd-1);
3353 		if (!(ctx->ifc_flags & IFLIB_HAS_TXCQ)) {
3354 			next_txd = calc_next_txd(txq, cidx, 0);
3355 			prefetch(next_txd);
3356 		}
3357 
3358 		/* prefetch the next cache line of mbuf pointers and flags */
3359 		prefetch(&txq->ift_sds.ifsd_m[next]);
3360 		prefetch(&txq->ift_sds.ifsd_map[next]);
3361 		next = (cidx + CACHE_LINE_SIZE) & (ntxd-1);
3362 	}
3363 	map = txq->ift_sds.ifsd_map[pidx];
3364 	ifsd_m = txq->ift_sds.ifsd_m;
3365 
3366 	if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
3367 		buf_tag = txq->ift_tso_buf_tag;
3368 		max_segs = scctx->isc_tx_tso_segments_max;
3369 		map = txq->ift_sds.ifsd_tso_map[pidx];
3370 		MPASS(buf_tag != NULL);
3371 		MPASS(max_segs > 0);
3372 	} else {
3373 		buf_tag = txq->ift_buf_tag;
3374 		max_segs = scctx->isc_tx_nsegments;
3375 		map = txq->ift_sds.ifsd_map[pidx];
3376 	}
3377 	if ((sctx->isc_flags & IFLIB_NEED_ETHER_PAD) &&
3378 	    __predict_false(m_head->m_pkthdr.len < scctx->isc_min_frame_size)) {
3379 		err = iflib_ether_pad(ctx->ifc_dev, m_headp, scctx->isc_min_frame_size);
3380 		if (err) {
3381 			DBG_COUNTER_INC(encap_txd_encap_fail);
3382 			return err;
3383 		}
3384 	}
3385 	m_head = *m_headp;
3386 
3387 	pkt_info_zero(&pi);
3388 	pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG|M_BCAST|M_MCAST));
3389 	pi.ipi_pidx = pidx;
3390 	pi.ipi_qsidx = txq->ift_id;
3391 	pi.ipi_len = m_head->m_pkthdr.len;
3392 	pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags;
3393 	pi.ipi_vtag = M_HAS_VLANTAG(m_head) ? m_head->m_pkthdr.ether_vtag : 0;
3394 
3395 	/* deliberate bitwise OR to make one condition */
3396 	if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) {
3397 		if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0)) {
3398 			DBG_COUNTER_INC(encap_txd_encap_fail);
3399 			return (err);
3400 		}
3401 		m_head = *m_headp;
3402 	}
3403 
3404 retry:
3405 	err = bus_dmamap_load_mbuf_sg(buf_tag, map, m_head, segs, &nsegs,
3406 	    BUS_DMA_NOWAIT);
3407 defrag:
3408 	if (__predict_false(err)) {
3409 		switch (err) {
3410 		case EFBIG:
3411 			/* try collapse once and defrag once */
3412 			if (remap == 0) {
3413 				m_head = m_collapse(*m_headp, M_NOWAIT, max_segs);
3414 				/* try defrag if collapsing fails */
3415 				if (m_head == NULL)
3416 					remap++;
3417 			}
3418 			if (remap == 1) {
3419 				txq->ift_mbuf_defrag++;
3420 				m_head = m_defrag(*m_headp, M_NOWAIT);
3421 			}
3422 			/*
3423 			 * remap should never be >1 unless bus_dmamap_load_mbuf_sg
3424 			 * failed to map an mbuf that was run through m_defrag
3425 			 */
3426 			MPASS(remap <= 1);
3427 			if (__predict_false(m_head == NULL || remap > 1))
3428 				goto defrag_failed;
3429 			remap++;
3430 			*m_headp = m_head;
3431 			goto retry;
3432 			break;
3433 		case ENOMEM:
3434 			txq->ift_no_tx_dma_setup++;
3435 			break;
3436 		default:
3437 			txq->ift_no_tx_dma_setup++;
3438 			m_freem(*m_headp);
3439 			DBG_COUNTER_INC(tx_frees);
3440 			*m_headp = NULL;
3441 			break;
3442 		}
3443 		txq->ift_map_failed++;
3444 		DBG_COUNTER_INC(encap_load_mbuf_fail);
3445 		DBG_COUNTER_INC(encap_txd_encap_fail);
3446 		return (err);
3447 	}
3448 	ifsd_m[pidx] = m_head;
3449 	/*
3450 	 * XXX assumes a 1 to 1 relationship between segments and
3451 	 *        descriptors - this does not hold true on all drivers, e.g.
3452 	 *        cxgb
3453 	 */
3454 	if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) {
3455 		txq->ift_no_desc_avail++;
3456 		bus_dmamap_unload(buf_tag, map);
3457 		DBG_COUNTER_INC(encap_txq_avail_fail);
3458 		DBG_COUNTER_INC(encap_txd_encap_fail);
3459 		if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0)
3460 			GROUPTASK_ENQUEUE(&txq->ift_task);
3461 		return (ENOBUFS);
3462 	}
3463 	/*
3464 	 * On Intel cards we can greatly reduce the number of TX interrupts
3465 	 * we see by only setting report status on every Nth descriptor.
3466 	 * However, this also means that the driver will need to keep track
3467 	 * of the descriptors that RS was set on to check them for the DD bit.
3468 	 */
3469 	txq->ift_rs_pending += nsegs + 1;
3470 	if (txq->ift_rs_pending > TXQ_MAX_RS_DEFERRED(txq) ||
3471 	     iflib_no_tx_batch || (TXQ_AVAIL(txq) - nsegs) <= MAX_TX_DESC(ctx) + 2) {
3472 		pi.ipi_flags |= IPI_TX_INTR;
3473 		txq->ift_rs_pending = 0;
3474 	}
3475 
3476 	pi.ipi_segs = segs;
3477 	pi.ipi_nsegs = nsegs;
3478 
3479 	MPASS(pidx >= 0 && pidx < txq->ift_size);
3480 #ifdef PKT_DEBUG
3481 	print_pkt(&pi);
3482 #endif
3483 	if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) {
3484 		bus_dmamap_sync(buf_tag, map, BUS_DMASYNC_PREWRITE);
3485 		DBG_COUNTER_INC(tx_encap);
3486 		MPASS(pi.ipi_new_pidx < txq->ift_size);
3487 
3488 		ndesc = pi.ipi_new_pidx - pi.ipi_pidx;
3489 		if (pi.ipi_new_pidx < pi.ipi_pidx) {
3490 			ndesc += txq->ift_size;
3491 			txq->ift_gen = 1;
3492 		}
3493 		/*
3494 		 * drivers can need as many as
3495 		 * two sentinels
3496 		 */
3497 		MPASS(ndesc <= pi.ipi_nsegs + 2);
3498 		MPASS(pi.ipi_new_pidx != pidx);
3499 		MPASS(ndesc > 0);
3500 		txq->ift_in_use += ndesc;
3501 		txq->ift_db_pending += ndesc;
3502 
3503 		/*
3504 		 * We update the last software descriptor again here because there may
3505 		 * be a sentinel and/or there may be more mbufs than segments
3506 		 */
3507 		txq->ift_pidx = pi.ipi_new_pidx;
3508 		txq->ift_npending += pi.ipi_ndescs;
3509 	} else {
3510 		*m_headp = m_head = iflib_remove_mbuf(txq);
3511 		if (err == EFBIG) {
3512 			txq->ift_txd_encap_efbig++;
3513 			if (remap < 2) {
3514 				remap = 1;
3515 				goto defrag;
3516 			}
3517 		}
3518 		goto defrag_failed;
3519 	}
3520 	/*
3521 	 * err can't possibly be non-zero here, so we don't neet to test it
3522 	 * to see if we need to DBG_COUNTER_INC(encap_txd_encap_fail).
3523 	 */
3524 	return (err);
3525 
3526 defrag_failed:
3527 	txq->ift_mbuf_defrag_failed++;
3528 	txq->ift_map_failed++;
3529 	m_freem(*m_headp);
3530 	DBG_COUNTER_INC(tx_frees);
3531 	*m_headp = NULL;
3532 	DBG_COUNTER_INC(encap_txd_encap_fail);
3533 	return (ENOMEM);
3534 }
3535 
3536 static void
3537 iflib_tx_desc_free(iflib_txq_t txq, int n)
3538 {
3539 	uint32_t qsize, cidx, mask, gen;
3540 	struct mbuf *m, **ifsd_m;
3541 	bool do_prefetch;
3542 
3543 	cidx = txq->ift_cidx;
3544 	gen = txq->ift_gen;
3545 	qsize = txq->ift_size;
3546 	mask = qsize-1;
3547 	ifsd_m = txq->ift_sds.ifsd_m;
3548 	do_prefetch = (txq->ift_ctx->ifc_flags & IFC_PREFETCH);
3549 
3550 	while (n-- > 0) {
3551 		if (do_prefetch) {
3552 			prefetch(ifsd_m[(cidx + 3) & mask]);
3553 			prefetch(ifsd_m[(cidx + 4) & mask]);
3554 		}
3555 		if ((m = ifsd_m[cidx]) != NULL) {
3556 			prefetch(&ifsd_m[(cidx + CACHE_PTR_INCREMENT) & mask]);
3557 			if (m->m_pkthdr.csum_flags & CSUM_TSO) {
3558 				bus_dmamap_sync(txq->ift_tso_buf_tag,
3559 				    txq->ift_sds.ifsd_tso_map[cidx],
3560 				    BUS_DMASYNC_POSTWRITE);
3561 				bus_dmamap_unload(txq->ift_tso_buf_tag,
3562 				    txq->ift_sds.ifsd_tso_map[cidx]);
3563 			} else {
3564 				bus_dmamap_sync(txq->ift_buf_tag,
3565 				    txq->ift_sds.ifsd_map[cidx],
3566 				    BUS_DMASYNC_POSTWRITE);
3567 				bus_dmamap_unload(txq->ift_buf_tag,
3568 				    txq->ift_sds.ifsd_map[cidx]);
3569 			}
3570 			/* XXX we don't support any drivers that batch packets yet */
3571 			MPASS(m->m_nextpkt == NULL);
3572 			m_freem(m);
3573 			ifsd_m[cidx] = NULL;
3574 #if MEMORY_LOGGING
3575 			txq->ift_dequeued++;
3576 #endif
3577 			DBG_COUNTER_INC(tx_frees);
3578 		}
3579 		if (__predict_false(++cidx == qsize)) {
3580 			cidx = 0;
3581 			gen = 0;
3582 		}
3583 	}
3584 	txq->ift_cidx = cidx;
3585 	txq->ift_gen = gen;
3586 }
3587 
3588 static __inline int
3589 iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh)
3590 {
3591 	int reclaim;
3592 	if_ctx_t ctx = txq->ift_ctx;
3593 
3594 	KASSERT(thresh >= 0, ("invalid threshold to reclaim"));
3595 	MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size);
3596 
3597 	/*
3598 	 * Need a rate-limiting check so that this isn't called every time
3599 	 */
3600 	iflib_tx_credits_update(ctx, txq);
3601 	reclaim = DESC_RECLAIMABLE(txq);
3602 
3603 	if (reclaim <= thresh /* + MAX_TX_DESC(txq->ift_ctx) */) {
3604 #ifdef INVARIANTS
3605 		if (iflib_verbose_debug) {
3606 			printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __FUNCTION__,
3607 			       txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments,
3608 			       reclaim, thresh);
3609 		}
3610 #endif
3611 		return (0);
3612 	}
3613 	iflib_tx_desc_free(txq, reclaim);
3614 	txq->ift_cleaned += reclaim;
3615 	txq->ift_in_use -= reclaim;
3616 
3617 	return (reclaim);
3618 }
3619 
3620 static struct mbuf **
3621 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset, int remaining)
3622 {
3623 	int next, size;
3624 	struct mbuf **items;
3625 
3626 	size = r->size;
3627 	next = (cidx + CACHE_PTR_INCREMENT) & (size-1);
3628 	items = __DEVOLATILE(struct mbuf **, &r->items[0]);
3629 
3630 	prefetch(items[(cidx + offset) & (size-1)]);
3631 	if (remaining > 1) {
3632 		prefetch2cachelines(&items[next]);
3633 		prefetch2cachelines(items[(cidx + offset + 1) & (size-1)]);
3634 		prefetch2cachelines(items[(cidx + offset + 2) & (size-1)]);
3635 		prefetch2cachelines(items[(cidx + offset + 3) & (size-1)]);
3636 	}
3637 	return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (size-1)]));
3638 }
3639 
3640 static void
3641 iflib_txq_check_drain(iflib_txq_t txq, int budget)
3642 {
3643 
3644 	ifmp_ring_check_drainage(txq->ift_br, budget);
3645 }
3646 
3647 static uint32_t
3648 iflib_txq_can_drain(struct ifmp_ring *r)
3649 {
3650 	iflib_txq_t txq = r->cookie;
3651 	if_ctx_t ctx = txq->ift_ctx;
3652 
3653 	if (TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2)
3654 		return (1);
3655 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3656 	    BUS_DMASYNC_POSTREAD);
3657 	return (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id,
3658 	    false));
3659 }
3660 
3661 static uint32_t
3662 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3663 {
3664 	iflib_txq_t txq = r->cookie;
3665 	if_ctx_t ctx = txq->ift_ctx;
3666 	if_t ifp = ctx->ifc_ifp;
3667 	struct mbuf *m, **mp;
3668 	int avail, bytes_sent, skipped, count, err, i;
3669 	int mcast_sent, pkt_sent, reclaimed;
3670 	bool do_prefetch, rang, ring;
3671 
3672 	if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) ||
3673 			    !LINK_ACTIVE(ctx))) {
3674 		DBG_COUNTER_INC(txq_drain_notready);
3675 		return (0);
3676 	}
3677 	reclaimed = iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
3678 	rang = iflib_txd_db_check(txq, reclaimed && txq->ift_db_pending);
3679 	avail = IDXDIFF(pidx, cidx, r->size);
3680 
3681 	if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) {
3682 		/*
3683 		 * The driver is unloading so we need to free all pending packets.
3684 		 */
3685 		DBG_COUNTER_INC(txq_drain_flushing);
3686 		for (i = 0; i < avail; i++) {
3687 			if (__predict_true(r->items[(cidx + i) & (r->size-1)] != (void *)txq))
3688 				m_freem(r->items[(cidx + i) & (r->size-1)]);
3689 			r->items[(cidx + i) & (r->size-1)] = NULL;
3690 		}
3691 		return (avail);
3692 	}
3693 
3694 	if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3695 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3696 		CALLOUT_LOCK(txq);
3697 		callout_stop(&txq->ift_timer);
3698 		CALLOUT_UNLOCK(txq);
3699 		DBG_COUNTER_INC(txq_drain_oactive);
3700 		return (0);
3701 	}
3702 
3703 	/*
3704 	 * If we've reclaimed any packets this queue cannot be hung.
3705 	 */
3706 	if (reclaimed)
3707 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3708 	skipped = mcast_sent = bytes_sent = pkt_sent = 0;
3709 	count = MIN(avail, TX_BATCH_SIZE);
3710 #ifdef INVARIANTS
3711 	if (iflib_verbose_debug)
3712 		printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __FUNCTION__,
3713 		       avail, ctx->ifc_flags, TXQ_AVAIL(txq));
3714 #endif
3715 	do_prefetch = (ctx->ifc_flags & IFC_PREFETCH);
3716 	err = 0;
3717 	for (i = 0; i < count && TXQ_AVAIL(txq) >= MAX_TX_DESC(ctx) + 2; i++) {
3718 		int rem = do_prefetch ? count - i : 0;
3719 
3720 		mp = _ring_peek_one(r, cidx, i, rem);
3721 		MPASS(mp != NULL && *mp != NULL);
3722 
3723 		/*
3724 		 * Completion interrupts will use the address of the txq
3725 		 * as a sentinel to enqueue _something_ in order to acquire
3726 		 * the lock on the mp_ring (there's no direct lock call).
3727 		 * We obviously whave to check for these sentinel cases
3728 		 * and skip them.
3729 		 */
3730 		if (__predict_false(*mp == (struct mbuf *)txq)) {
3731 			skipped++;
3732 			continue;
3733 		}
3734 		err = iflib_encap(txq, mp);
3735 		if (__predict_false(err)) {
3736 			/* no room - bail out */
3737 			if (err == ENOBUFS)
3738 				break;
3739 			skipped++;
3740 			/* we can't send this packet - skip it */
3741 			continue;
3742 		}
3743 		pkt_sent++;
3744 		m = *mp;
3745 		DBG_COUNTER_INC(tx_sent);
3746 		bytes_sent += m->m_pkthdr.len;
3747 		mcast_sent += !!(m->m_flags & M_MCAST);
3748 
3749 		if (__predict_false(!(ifp->if_drv_flags & IFF_DRV_RUNNING)))
3750 			break;
3751 		ETHER_BPF_MTAP(ifp, m);
3752 		rang = iflib_txd_db_check(txq, false);
3753 	}
3754 
3755 	/* deliberate use of bitwise or to avoid gratuitous short-circuit */
3756 	ring = rang ? false  : (iflib_min_tx_latency | err);
3757 	iflib_txd_db_check(txq, ring);
3758 	if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
3759 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
3760 	if (mcast_sent)
3761 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
3762 #ifdef INVARIANTS
3763 	if (iflib_verbose_debug)
3764 		printf("consumed=%d\n", skipped + pkt_sent);
3765 #endif
3766 	return (skipped + pkt_sent);
3767 }
3768 
3769 static uint32_t
3770 iflib_txq_drain_always(struct ifmp_ring *r)
3771 {
3772 	return (1);
3773 }
3774 
3775 static uint32_t
3776 iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3777 {
3778 	int i, avail;
3779 	struct mbuf **mp;
3780 	iflib_txq_t txq;
3781 
3782 	txq = r->cookie;
3783 
3784 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3785 	CALLOUT_LOCK(txq);
3786 	callout_stop(&txq->ift_timer);
3787 	CALLOUT_UNLOCK(txq);
3788 
3789 	avail = IDXDIFF(pidx, cidx, r->size);
3790 	for (i = 0; i < avail; i++) {
3791 		mp = _ring_peek_one(r, cidx, i, avail - i);
3792 		if (__predict_false(*mp == (struct mbuf *)txq))
3793 			continue;
3794 		m_freem(*mp);
3795 		DBG_COUNTER_INC(tx_frees);
3796 	}
3797 	MPASS(ifmp_ring_is_stalled(r) == 0);
3798 	return (avail);
3799 }
3800 
3801 static void
3802 iflib_ifmp_purge(iflib_txq_t txq)
3803 {
3804 	struct ifmp_ring *r;
3805 
3806 	r = txq->ift_br;
3807 	r->drain = iflib_txq_drain_free;
3808 	r->can_drain = iflib_txq_drain_always;
3809 
3810 	ifmp_ring_check_drainage(r, r->size);
3811 
3812 	r->drain = iflib_txq_drain;
3813 	r->can_drain = iflib_txq_can_drain;
3814 }
3815 
3816 static void
3817 _task_fn_tx(void *context)
3818 {
3819 	iflib_txq_t txq = context;
3820 	if_ctx_t ctx = txq->ift_ctx;
3821 	if_t ifp = ctx->ifc_ifp;
3822 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
3823 
3824 #ifdef IFLIB_DIAGNOSTICS
3825 	txq->ift_cpu_exec_count[curcpu]++;
3826 #endif
3827 	if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
3828 		return;
3829 #ifdef DEV_NETMAP
3830 	if ((if_getcapenable(ifp) & IFCAP_NETMAP) &&
3831 	    netmap_tx_irq(ifp, txq->ift_id))
3832 		goto skip_ifmp;
3833 #endif
3834 #ifdef ALTQ
3835 	if (ALTQ_IS_ENABLED(&ifp->if_snd))
3836 		iflib_altq_if_start(ifp);
3837 #endif
3838 	if (txq->ift_db_pending)
3839 		ifmp_ring_enqueue(txq->ift_br, (void **)&txq, 1, TX_BATCH_SIZE, abdicate);
3840 	else if (!abdicate)
3841 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3842 	/*
3843 	 * When abdicating, we always need to check drainage, not just when we don't enqueue
3844 	 */
3845 	if (abdicate)
3846 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3847 #ifdef DEV_NETMAP
3848 skip_ifmp:
3849 #endif
3850 	if (ctx->ifc_flags & IFC_LEGACY)
3851 		IFDI_INTR_ENABLE(ctx);
3852 	else
3853 		IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3854 }
3855 
3856 static void
3857 _task_fn_rx(void *context)
3858 {
3859 	iflib_rxq_t rxq = context;
3860 	if_ctx_t ctx = rxq->ifr_ctx;
3861 	uint8_t more;
3862 	uint16_t budget;
3863 #ifdef DEV_NETMAP
3864 	u_int work = 0;
3865 	int nmirq;
3866 #endif
3867 
3868 #ifdef IFLIB_DIAGNOSTICS
3869 	rxq->ifr_cpu_exec_count[curcpu]++;
3870 #endif
3871 	DBG_COUNTER_INC(task_fn_rxs);
3872 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3873 		return;
3874 #ifdef DEV_NETMAP
3875 	nmirq = netmap_rx_irq(ctx->ifc_ifp, rxq->ifr_id, &work);
3876 	if (nmirq != NM_IRQ_PASS) {
3877 		more = (nmirq == NM_IRQ_RESCHED) ? IFLIB_RXEOF_MORE : 0;
3878 		goto skip_rxeof;
3879 	}
3880 #endif
3881 	budget = ctx->ifc_sysctl_rx_budget;
3882 	if (budget == 0)
3883 		budget = 16;	/* XXX */
3884 	more = iflib_rxeof(rxq, budget);
3885 #ifdef DEV_NETMAP
3886 skip_rxeof:
3887 #endif
3888 	if ((more & IFLIB_RXEOF_MORE) == 0) {
3889 		if (ctx->ifc_flags & IFC_LEGACY)
3890 			IFDI_INTR_ENABLE(ctx);
3891 		else
3892 			IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
3893 		DBG_COUNTER_INC(rx_intr_enables);
3894 	}
3895 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3896 		return;
3897 
3898 	if (more & IFLIB_RXEOF_MORE)
3899 		GROUPTASK_ENQUEUE(&rxq->ifr_task);
3900 	else if (more & IFLIB_RXEOF_EMPTY)
3901 		callout_reset_curcpu(&rxq->ifr_watchdog, 1, &_task_fn_rx_watchdog, rxq);
3902 }
3903 
3904 static void
3905 _task_fn_admin(void *context)
3906 {
3907 	if_ctx_t ctx = context;
3908 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
3909 	iflib_txq_t txq;
3910 	int i;
3911 	bool oactive, running, do_reset, do_watchdog, in_detach;
3912 
3913 	STATE_LOCK(ctx);
3914 	running = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING);
3915 	oactive = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE);
3916 	do_reset = (ctx->ifc_flags & IFC_DO_RESET);
3917 	do_watchdog = (ctx->ifc_flags & IFC_DO_WATCHDOG);
3918 	in_detach = (ctx->ifc_flags & IFC_IN_DETACH);
3919 	ctx->ifc_flags &= ~(IFC_DO_RESET|IFC_DO_WATCHDOG);
3920 	STATE_UNLOCK(ctx);
3921 
3922 	if ((!running && !oactive) && !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
3923 		return;
3924 	if (in_detach)
3925 		return;
3926 
3927 	CTX_LOCK(ctx);
3928 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3929 		CALLOUT_LOCK(txq);
3930 		callout_stop(&txq->ift_timer);
3931 		CALLOUT_UNLOCK(txq);
3932 	}
3933 	if (do_watchdog) {
3934 		ctx->ifc_watchdog_events++;
3935 		IFDI_WATCHDOG_RESET(ctx);
3936 	}
3937 	IFDI_UPDATE_ADMIN_STATUS(ctx);
3938 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3939 		callout_reset_on(&txq->ift_timer, iflib_timer_default, iflib_timer, txq,
3940 		    txq->ift_timer.c_cpu);
3941 	}
3942 	IFDI_LINK_INTR_ENABLE(ctx);
3943 	if (do_reset)
3944 		iflib_if_init_locked(ctx);
3945 	CTX_UNLOCK(ctx);
3946 
3947 	if (LINK_ACTIVE(ctx) == 0)
3948 		return;
3949 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3950 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
3951 }
3952 
3953 static void
3954 _task_fn_iov(void *context)
3955 {
3956 	if_ctx_t ctx = context;
3957 
3958 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING) &&
3959 	    !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
3960 		return;
3961 
3962 	CTX_LOCK(ctx);
3963 	IFDI_VFLR_HANDLE(ctx);
3964 	CTX_UNLOCK(ctx);
3965 }
3966 
3967 static int
3968 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)
3969 {
3970 	int err;
3971 	if_int_delay_info_t info;
3972 	if_ctx_t ctx;
3973 
3974 	info = (if_int_delay_info_t)arg1;
3975 	ctx = info->iidi_ctx;
3976 	info->iidi_req = req;
3977 	info->iidi_oidp = oidp;
3978 	CTX_LOCK(ctx);
3979 	err = IFDI_SYSCTL_INT_DELAY(ctx, info);
3980 	CTX_UNLOCK(ctx);
3981 	return (err);
3982 }
3983 
3984 /*********************************************************************
3985  *
3986  *  IFNET FUNCTIONS
3987  *
3988  **********************************************************************/
3989 
3990 static void
3991 iflib_if_init_locked(if_ctx_t ctx)
3992 {
3993 	iflib_stop(ctx);
3994 	iflib_init_locked(ctx);
3995 }
3996 
3997 static void
3998 iflib_if_init(void *arg)
3999 {
4000 	if_ctx_t ctx = arg;
4001 
4002 	CTX_LOCK(ctx);
4003 	iflib_if_init_locked(ctx);
4004 	CTX_UNLOCK(ctx);
4005 }
4006 
4007 static int
4008 iflib_if_transmit(if_t ifp, struct mbuf *m)
4009 {
4010 	if_ctx_t	ctx = if_getsoftc(ifp);
4011 
4012 	iflib_txq_t txq;
4013 	int err, qidx;
4014 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
4015 
4016 	if (__predict_false((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) {
4017 		DBG_COUNTER_INC(tx_frees);
4018 		m_freem(m);
4019 		return (ENETDOWN);
4020 	}
4021 
4022 	MPASS(m->m_nextpkt == NULL);
4023 	/* ALTQ-enabled interfaces always use queue 0. */
4024 	qidx = 0;
4025 	if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m) && !ALTQ_IS_ENABLED(&ifp->if_snd))
4026 		qidx = QIDX(ctx, m);
4027 	/*
4028 	 * XXX calculate buf_ring based on flowid (divvy up bits?)
4029 	 */
4030 	txq = &ctx->ifc_txqs[qidx];
4031 
4032 #ifdef DRIVER_BACKPRESSURE
4033 	if (txq->ift_closed) {
4034 		while (m != NULL) {
4035 			next = m->m_nextpkt;
4036 			m->m_nextpkt = NULL;
4037 			m_freem(m);
4038 			DBG_COUNTER_INC(tx_frees);
4039 			m = next;
4040 		}
4041 		return (ENOBUFS);
4042 	}
4043 #endif
4044 #ifdef notyet
4045 	qidx = count = 0;
4046 	mp = marr;
4047 	next = m;
4048 	do {
4049 		count++;
4050 		next = next->m_nextpkt;
4051 	} while (next != NULL);
4052 
4053 	if (count > nitems(marr))
4054 		if ((mp = malloc(count*sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) {
4055 			/* XXX check nextpkt */
4056 			m_freem(m);
4057 			/* XXX simplify for now */
4058 			DBG_COUNTER_INC(tx_frees);
4059 			return (ENOBUFS);
4060 		}
4061 	for (next = m, i = 0; next != NULL; i++) {
4062 		mp[i] = next;
4063 		next = next->m_nextpkt;
4064 		mp[i]->m_nextpkt = NULL;
4065 	}
4066 #endif
4067 	DBG_COUNTER_INC(tx_seen);
4068 	err = ifmp_ring_enqueue(txq->ift_br, (void **)&m, 1, TX_BATCH_SIZE, abdicate);
4069 
4070 	if (abdicate)
4071 		GROUPTASK_ENQUEUE(&txq->ift_task);
4072  	if (err) {
4073 		if (!abdicate)
4074 			GROUPTASK_ENQUEUE(&txq->ift_task);
4075 		/* support forthcoming later */
4076 #ifdef DRIVER_BACKPRESSURE
4077 		txq->ift_closed = TRUE;
4078 #endif
4079 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
4080 		m_freem(m);
4081 		DBG_COUNTER_INC(tx_frees);
4082 	}
4083 
4084 	return (err);
4085 }
4086 
4087 #ifdef ALTQ
4088 /*
4089  * The overall approach to integrating iflib with ALTQ is to continue to use
4090  * the iflib mp_ring machinery between the ALTQ queue(s) and the hardware
4091  * ring.  Technically, when using ALTQ, queueing to an intermediate mp_ring
4092  * is redundant/unnecessary, but doing so minimizes the amount of
4093  * ALTQ-specific code required in iflib.  It is assumed that the overhead of
4094  * redundantly queueing to an intermediate mp_ring is swamped by the
4095  * performance limitations inherent in using ALTQ.
4096  *
4097  * When ALTQ support is compiled in, all iflib drivers will use a transmit
4098  * routine, iflib_altq_if_transmit(), that checks if ALTQ is enabled for the
4099  * given interface.  If ALTQ is enabled for an interface, then all
4100  * transmitted packets for that interface will be submitted to the ALTQ
4101  * subsystem via IFQ_ENQUEUE().  We don't use the legacy if_transmit()
4102  * implementation because it uses IFQ_HANDOFF(), which will duplicatively
4103  * update stats that the iflib machinery handles, and which is sensitve to
4104  * the disused IFF_DRV_OACTIVE flag.  Additionally, iflib_altq_if_start()
4105  * will be installed as the start routine for use by ALTQ facilities that
4106  * need to trigger queue drains on a scheduled basis.
4107  *
4108  */
4109 static void
4110 iflib_altq_if_start(if_t ifp)
4111 {
4112 	struct ifaltq *ifq = &ifp->if_snd;
4113 	struct mbuf *m;
4114 
4115 	IFQ_LOCK(ifq);
4116 	IFQ_DEQUEUE_NOLOCK(ifq, m);
4117 	while (m != NULL) {
4118 		iflib_if_transmit(ifp, m);
4119 		IFQ_DEQUEUE_NOLOCK(ifq, m);
4120 	}
4121 	IFQ_UNLOCK(ifq);
4122 }
4123 
4124 static int
4125 iflib_altq_if_transmit(if_t ifp, struct mbuf *m)
4126 {
4127 	int err;
4128 
4129 	if (ALTQ_IS_ENABLED(&ifp->if_snd)) {
4130 		IFQ_ENQUEUE(&ifp->if_snd, m, err);
4131 		if (err == 0)
4132 			iflib_altq_if_start(ifp);
4133 	} else
4134 		err = iflib_if_transmit(ifp, m);
4135 
4136 	return (err);
4137 }
4138 #endif /* ALTQ */
4139 
4140 static void
4141 iflib_if_qflush(if_t ifp)
4142 {
4143 	if_ctx_t ctx = if_getsoftc(ifp);
4144 	iflib_txq_t txq = ctx->ifc_txqs;
4145 	int i;
4146 
4147 	STATE_LOCK(ctx);
4148 	ctx->ifc_flags |= IFC_QFLUSH;
4149 	STATE_UNLOCK(ctx);
4150 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
4151 		while (!(ifmp_ring_is_idle(txq->ift_br) || ifmp_ring_is_stalled(txq->ift_br)))
4152 			iflib_txq_check_drain(txq, 0);
4153 	STATE_LOCK(ctx);
4154 	ctx->ifc_flags &= ~IFC_QFLUSH;
4155 	STATE_UNLOCK(ctx);
4156 
4157 	/*
4158 	 * When ALTQ is enabled, this will also take care of purging the
4159 	 * ALTQ queue(s).
4160 	 */
4161 	if_qflush(ifp);
4162 }
4163 
4164 #define IFCAP_FLAGS (IFCAP_HWCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \
4165 		     IFCAP_TSO | IFCAP_VLAN_HWTAGGING | IFCAP_HWSTATS | \
4166 		     IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | \
4167 		     IFCAP_VLAN_HWTSO | IFCAP_VLAN_HWCSUM | IFCAP_NOMAP)
4168 
4169 static int
4170 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data)
4171 {
4172 	if_ctx_t ctx = if_getsoftc(ifp);
4173 	struct ifreq	*ifr = (struct ifreq *)data;
4174 #if defined(INET) || defined(INET6)
4175 	struct ifaddr	*ifa = (struct ifaddr *)data;
4176 #endif
4177 	bool		avoid_reset = false;
4178 	int		err = 0, reinit = 0, bits;
4179 
4180 	switch (command) {
4181 	case SIOCSIFADDR:
4182 #ifdef INET
4183 		if (ifa->ifa_addr->sa_family == AF_INET)
4184 			avoid_reset = true;
4185 #endif
4186 #ifdef INET6
4187 		if (ifa->ifa_addr->sa_family == AF_INET6)
4188 			avoid_reset = true;
4189 #endif
4190 		/*
4191 		** Calling init results in link renegotiation,
4192 		** so we avoid doing it when possible.
4193 		*/
4194 		if (avoid_reset) {
4195 			if_setflagbits(ifp, IFF_UP,0);
4196 			if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
4197 				reinit = 1;
4198 #ifdef INET
4199 			if (!(if_getflags(ifp) & IFF_NOARP))
4200 				arp_ifinit(ifp, ifa);
4201 #endif
4202 		} else
4203 			err = ether_ioctl(ifp, command, data);
4204 		break;
4205 	case SIOCSIFMTU:
4206 		CTX_LOCK(ctx);
4207 		if (ifr->ifr_mtu == if_getmtu(ifp)) {
4208 			CTX_UNLOCK(ctx);
4209 			break;
4210 		}
4211 		bits = if_getdrvflags(ifp);
4212 		/* stop the driver and free any clusters before proceeding */
4213 		iflib_stop(ctx);
4214 
4215 		if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) {
4216 			STATE_LOCK(ctx);
4217 			if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size)
4218 				ctx->ifc_flags |= IFC_MULTISEG;
4219 			else
4220 				ctx->ifc_flags &= ~IFC_MULTISEG;
4221 			STATE_UNLOCK(ctx);
4222 			err = if_setmtu(ifp, ifr->ifr_mtu);
4223 		}
4224 		iflib_init_locked(ctx);
4225 		STATE_LOCK(ctx);
4226 		if_setdrvflags(ifp, bits);
4227 		STATE_UNLOCK(ctx);
4228 		CTX_UNLOCK(ctx);
4229 		break;
4230 	case SIOCSIFFLAGS:
4231 		CTX_LOCK(ctx);
4232 		if (if_getflags(ifp) & IFF_UP) {
4233 			if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4234 				if ((if_getflags(ifp) ^ ctx->ifc_if_flags) &
4235 				    (IFF_PROMISC | IFF_ALLMULTI)) {
4236 					CTX_UNLOCK(ctx);
4237 					err = IFDI_PROMISC_SET(ctx, if_getflags(ifp));
4238 					CTX_LOCK(ctx);
4239 				}
4240 			} else
4241 				reinit = 1;
4242 		} else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4243 			iflib_stop(ctx);
4244 		}
4245 		ctx->ifc_if_flags = if_getflags(ifp);
4246 		CTX_UNLOCK(ctx);
4247 		break;
4248 	case SIOCADDMULTI:
4249 	case SIOCDELMULTI:
4250 		if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4251 			CTX_LOCK(ctx);
4252 			IFDI_INTR_DISABLE(ctx);
4253 			IFDI_MULTI_SET(ctx);
4254 			IFDI_INTR_ENABLE(ctx);
4255 			CTX_UNLOCK(ctx);
4256 		}
4257 		break;
4258 	case SIOCSIFMEDIA:
4259 		CTX_LOCK(ctx);
4260 		IFDI_MEDIA_SET(ctx);
4261 		CTX_UNLOCK(ctx);
4262 		/* FALLTHROUGH */
4263 	case SIOCGIFMEDIA:
4264 	case SIOCGIFXMEDIA:
4265 		err = ifmedia_ioctl(ifp, ifr, ctx->ifc_mediap, command);
4266 		break;
4267 	case SIOCGI2C:
4268 	{
4269 		struct ifi2creq i2c;
4270 
4271 		err = copyin(ifr_data_get_ptr(ifr), &i2c, sizeof(i2c));
4272 		if (err != 0)
4273 			break;
4274 		if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) {
4275 			err = EINVAL;
4276 			break;
4277 		}
4278 		if (i2c.len > sizeof(i2c.data)) {
4279 			err = EINVAL;
4280 			break;
4281 		}
4282 
4283 		if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0)
4284 			err = copyout(&i2c, ifr_data_get_ptr(ifr),
4285 			    sizeof(i2c));
4286 		break;
4287 	}
4288 	case SIOCSIFCAP:
4289 	{
4290 		int mask, setmask, oldmask;
4291 
4292 		oldmask = if_getcapenable(ifp);
4293 		mask = ifr->ifr_reqcap ^ oldmask;
4294 		mask &= ctx->ifc_softc_ctx.isc_capabilities | IFCAP_NOMAP;
4295 		setmask = 0;
4296 #ifdef TCP_OFFLOAD
4297 		setmask |= mask & (IFCAP_TOE4|IFCAP_TOE6);
4298 #endif
4299 		setmask |= (mask & IFCAP_FLAGS);
4300 		setmask |= (mask & IFCAP_WOL);
4301 
4302 		/*
4303 		 * If any RX csum has changed, change all the ones that
4304 		 * are supported by the driver.
4305 		 */
4306 		if (setmask & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6)) {
4307 			setmask |= ctx->ifc_softc_ctx.isc_capabilities &
4308 			    (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6);
4309 		}
4310 
4311 		/*
4312 		 * want to ensure that traffic has stopped before we change any of the flags
4313 		 */
4314 		if (setmask) {
4315 			CTX_LOCK(ctx);
4316 			bits = if_getdrvflags(ifp);
4317 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4318 				iflib_stop(ctx);
4319 			STATE_LOCK(ctx);
4320 			if_togglecapenable(ifp, setmask);
4321 			STATE_UNLOCK(ctx);
4322 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4323 				iflib_init_locked(ctx);
4324 			STATE_LOCK(ctx);
4325 			if_setdrvflags(ifp, bits);
4326 			STATE_UNLOCK(ctx);
4327 			CTX_UNLOCK(ctx);
4328 		}
4329 		if_vlancap(ifp);
4330 		break;
4331 	}
4332 	case SIOCGPRIVATE_0:
4333 	case SIOCSDRVSPEC:
4334 	case SIOCGDRVSPEC:
4335 		CTX_LOCK(ctx);
4336 		err = IFDI_PRIV_IOCTL(ctx, command, data);
4337 		CTX_UNLOCK(ctx);
4338 		break;
4339 	default:
4340 		err = ether_ioctl(ifp, command, data);
4341 		break;
4342 	}
4343 	if (reinit)
4344 		iflib_if_init(ctx);
4345 	return (err);
4346 }
4347 
4348 static uint64_t
4349 iflib_if_get_counter(if_t ifp, ift_counter cnt)
4350 {
4351 	if_ctx_t ctx = if_getsoftc(ifp);
4352 
4353 	return (IFDI_GET_COUNTER(ctx, cnt));
4354 }
4355 
4356 /*********************************************************************
4357  *
4358  *  OTHER FUNCTIONS EXPORTED TO THE STACK
4359  *
4360  **********************************************************************/
4361 
4362 static void
4363 iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag)
4364 {
4365 	if_ctx_t ctx = if_getsoftc(ifp);
4366 
4367 	if ((void *)ctx != arg)
4368 		return;
4369 
4370 	if ((vtag == 0) || (vtag > 4095))
4371 		return;
4372 
4373 	if (iflib_in_detach(ctx))
4374 		return;
4375 
4376 	CTX_LOCK(ctx);
4377 	/* Driver may need all untagged packets to be flushed */
4378 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4379 		iflib_stop(ctx);
4380 	IFDI_VLAN_REGISTER(ctx, vtag);
4381 	/* Re-init to load the changes, if required */
4382 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4383 		iflib_init_locked(ctx);
4384 	CTX_UNLOCK(ctx);
4385 }
4386 
4387 static void
4388 iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag)
4389 {
4390 	if_ctx_t ctx = if_getsoftc(ifp);
4391 
4392 	if ((void *)ctx != arg)
4393 		return;
4394 
4395 	if ((vtag == 0) || (vtag > 4095))
4396 		return;
4397 
4398 	CTX_LOCK(ctx);
4399 	/* Driver may need all tagged packets to be flushed */
4400 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4401 		iflib_stop(ctx);
4402 	IFDI_VLAN_UNREGISTER(ctx, vtag);
4403 	/* Re-init to load the changes, if required */
4404 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4405 		iflib_init_locked(ctx);
4406 	CTX_UNLOCK(ctx);
4407 }
4408 
4409 static void
4410 iflib_led_func(void *arg, int onoff)
4411 {
4412 	if_ctx_t ctx = arg;
4413 
4414 	CTX_LOCK(ctx);
4415 	IFDI_LED_FUNC(ctx, onoff);
4416 	CTX_UNLOCK(ctx);
4417 }
4418 
4419 /*********************************************************************
4420  *
4421  *  BUS FUNCTION DEFINITIONS
4422  *
4423  **********************************************************************/
4424 
4425 int
4426 iflib_device_probe(device_t dev)
4427 {
4428 	const pci_vendor_info_t *ent;
4429 	if_shared_ctx_t sctx;
4430 	uint16_t pci_device_id, pci_rev_id, pci_subdevice_id, pci_subvendor_id;
4431 	uint16_t pci_vendor_id;
4432 
4433 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4434 		return (ENOTSUP);
4435 
4436 	pci_vendor_id = pci_get_vendor(dev);
4437 	pci_device_id = pci_get_device(dev);
4438 	pci_subvendor_id = pci_get_subvendor(dev);
4439 	pci_subdevice_id = pci_get_subdevice(dev);
4440 	pci_rev_id = pci_get_revid(dev);
4441 	if (sctx->isc_parse_devinfo != NULL)
4442 		sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id);
4443 
4444 	ent = sctx->isc_vendor_info;
4445 	while (ent->pvi_vendor_id != 0) {
4446 		if (pci_vendor_id != ent->pvi_vendor_id) {
4447 			ent++;
4448 			continue;
4449 		}
4450 		if ((pci_device_id == ent->pvi_device_id) &&
4451 		    ((pci_subvendor_id == ent->pvi_subvendor_id) ||
4452 		     (ent->pvi_subvendor_id == 0)) &&
4453 		    ((pci_subdevice_id == ent->pvi_subdevice_id) ||
4454 		     (ent->pvi_subdevice_id == 0)) &&
4455 		    ((pci_rev_id == ent->pvi_rev_id) ||
4456 		     (ent->pvi_rev_id == 0))) {
4457 			device_set_desc_copy(dev, ent->pvi_name);
4458 			/* this needs to be changed to zero if the bus probing code
4459 			 * ever stops re-probing on best match because the sctx
4460 			 * may have its values over written by register calls
4461 			 * in subsequent probes
4462 			 */
4463 			return (BUS_PROBE_DEFAULT);
4464 		}
4465 		ent++;
4466 	}
4467 	return (ENXIO);
4468 }
4469 
4470 int
4471 iflib_device_probe_vendor(device_t dev)
4472 {
4473 	int probe;
4474 
4475 	probe = iflib_device_probe(dev);
4476 	if (probe == BUS_PROBE_DEFAULT)
4477 		return (BUS_PROBE_VENDOR);
4478 	else
4479 		return (probe);
4480 }
4481 
4482 static void
4483 iflib_reset_qvalues(if_ctx_t ctx)
4484 {
4485 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4486 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4487 	device_t dev = ctx->ifc_dev;
4488 	int i;
4489 
4490 	if (ctx->ifc_sysctl_ntxqs != 0)
4491 		scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs;
4492 	if (ctx->ifc_sysctl_nrxqs != 0)
4493 		scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs;
4494 
4495 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4496 		if (ctx->ifc_sysctl_ntxds[i] != 0)
4497 			scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i];
4498 		else
4499 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4500 	}
4501 
4502 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4503 		if (ctx->ifc_sysctl_nrxds[i] != 0)
4504 			scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i];
4505 		else
4506 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4507 	}
4508 
4509 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4510 		if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) {
4511 			device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n",
4512 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]);
4513 			scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i];
4514 		}
4515 		if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) {
4516 			device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n",
4517 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]);
4518 			scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i];
4519 		}
4520 		if (!powerof2(scctx->isc_nrxd[i])) {
4521 			device_printf(dev, "nrxd%d: %d is not a power of 2 - using default value of %d\n",
4522 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_default[i]);
4523 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4524 		}
4525 	}
4526 
4527 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4528 		if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) {
4529 			device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n",
4530 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]);
4531 			scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i];
4532 		}
4533 		if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) {
4534 			device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n",
4535 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]);
4536 			scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i];
4537 		}
4538 		if (!powerof2(scctx->isc_ntxd[i])) {
4539 			device_printf(dev, "ntxd%d: %d is not a power of 2 - using default value of %d\n",
4540 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_default[i]);
4541 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4542 		}
4543 	}
4544 }
4545 
4546 static void
4547 iflib_add_pfil(if_ctx_t ctx)
4548 {
4549 	struct pfil_head *pfil;
4550 	struct pfil_head_args pa;
4551 	iflib_rxq_t rxq;
4552 	int i;
4553 
4554 	pa.pa_version = PFIL_VERSION;
4555 	pa.pa_flags = PFIL_IN;
4556 	pa.pa_type = PFIL_TYPE_ETHERNET;
4557 	pa.pa_headname = ctx->ifc_ifp->if_xname;
4558 	pfil = pfil_head_register(&pa);
4559 
4560 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4561 		rxq->pfil = pfil;
4562 	}
4563 }
4564 
4565 static void
4566 iflib_rem_pfil(if_ctx_t ctx)
4567 {
4568 	struct pfil_head *pfil;
4569 	iflib_rxq_t rxq;
4570 	int i;
4571 
4572 	rxq = ctx->ifc_rxqs;
4573 	pfil = rxq->pfil;
4574 	for (i = 0; i < NRXQSETS(ctx); i++, rxq++) {
4575 		rxq->pfil = NULL;
4576 	}
4577 	pfil_head_unregister(pfil);
4578 }
4579 
4580 static uint16_t
4581 get_ctx_core_offset(if_ctx_t ctx)
4582 {
4583 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4584 	struct cpu_offset *op;
4585 	uint16_t qc;
4586 	uint16_t ret = ctx->ifc_sysctl_core_offset;
4587 
4588 	if (ret != CORE_OFFSET_UNSPECIFIED)
4589 		return (ret);
4590 
4591 	if (ctx->ifc_sysctl_separate_txrx)
4592 		qc = scctx->isc_ntxqsets + scctx->isc_nrxqsets;
4593 	else
4594 		qc = max(scctx->isc_ntxqsets, scctx->isc_nrxqsets);
4595 
4596 	mtx_lock(&cpu_offset_mtx);
4597 	SLIST_FOREACH(op, &cpu_offsets, entries) {
4598 		if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) {
4599 			ret = op->offset;
4600 			op->offset += qc;
4601 			MPASS(op->refcount < UINT_MAX);
4602 			op->refcount++;
4603 			break;
4604 		}
4605 	}
4606 	if (ret == CORE_OFFSET_UNSPECIFIED) {
4607 		ret = 0;
4608 		op = malloc(sizeof(struct cpu_offset), M_IFLIB,
4609 		    M_NOWAIT | M_ZERO);
4610 		if (op == NULL) {
4611 			device_printf(ctx->ifc_dev,
4612 			    "allocation for cpu offset failed.\n");
4613 		} else {
4614 			op->offset = qc;
4615 			op->refcount = 1;
4616 			CPU_COPY(&ctx->ifc_cpus, &op->set);
4617 			SLIST_INSERT_HEAD(&cpu_offsets, op, entries);
4618 		}
4619 	}
4620 	mtx_unlock(&cpu_offset_mtx);
4621 
4622 	return (ret);
4623 }
4624 
4625 static void
4626 unref_ctx_core_offset(if_ctx_t ctx)
4627 {
4628 	struct cpu_offset *op, *top;
4629 
4630 	mtx_lock(&cpu_offset_mtx);
4631 	SLIST_FOREACH_SAFE(op, &cpu_offsets, entries, top) {
4632 		if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) {
4633 			MPASS(op->refcount > 0);
4634 			op->refcount--;
4635 			if (op->refcount == 0) {
4636 				SLIST_REMOVE(&cpu_offsets, op, cpu_offset, entries);
4637 				free(op, M_IFLIB);
4638 			}
4639 			break;
4640 		}
4641 	}
4642 	mtx_unlock(&cpu_offset_mtx);
4643 }
4644 
4645 int
4646 iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp)
4647 {
4648 	if_ctx_t ctx;
4649 	if_t ifp;
4650 	if_softc_ctx_t scctx;
4651 	kobjop_desc_t kobj_desc;
4652 	kobj_method_t *kobj_method;
4653 	int err, msix, rid;
4654 	int num_txd, num_rxd;
4655 
4656 	ctx = malloc(sizeof(* ctx), M_IFLIB, M_WAITOK|M_ZERO);
4657 
4658 	if (sc == NULL) {
4659 		sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4660 		device_set_softc(dev, ctx);
4661 		ctx->ifc_flags |= IFC_SC_ALLOCATED;
4662 	}
4663 
4664 	ctx->ifc_sctx = sctx;
4665 	ctx->ifc_dev = dev;
4666 	ctx->ifc_softc = sc;
4667 
4668 	if ((err = iflib_register(ctx)) != 0) {
4669 		device_printf(dev, "iflib_register failed %d\n", err);
4670 		goto fail_ctx_free;
4671 	}
4672 	iflib_add_device_sysctl_pre(ctx);
4673 
4674 	scctx = &ctx->ifc_softc_ctx;
4675 	ifp = ctx->ifc_ifp;
4676 
4677 	iflib_reset_qvalues(ctx);
4678 	CTX_LOCK(ctx);
4679 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4680 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4681 		goto fail_unlock;
4682 	}
4683 	_iflib_pre_assert(scctx);
4684 	ctx->ifc_txrx = *scctx->isc_txrx;
4685 
4686 	if (sctx->isc_flags & IFLIB_DRIVER_MEDIA)
4687 		ctx->ifc_mediap = scctx->isc_media;
4688 
4689 #ifdef INVARIANTS
4690 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
4691 		MPASS(scctx->isc_tx_csum_flags);
4692 #endif
4693 
4694 	if_setcapabilities(ifp,
4695 	    scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_NOMAP);
4696 	if_setcapenable(ifp,
4697 	    scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_NOMAP);
4698 
4699 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4700 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4701 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4702 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4703 
4704 	num_txd = iflib_num_tx_descs(ctx);
4705 	num_rxd = iflib_num_rx_descs(ctx);
4706 
4707 	/* XXX change for per-queue sizes */
4708 	device_printf(dev, "Using %d TX descriptors and %d RX descriptors\n",
4709 	    num_txd, num_rxd);
4710 
4711 	if (scctx->isc_tx_nsegments > num_txd / MAX_SINGLE_PACKET_FRACTION)
4712 		scctx->isc_tx_nsegments = max(1, num_txd /
4713 		    MAX_SINGLE_PACKET_FRACTION);
4714 	if (scctx->isc_tx_tso_segments_max > num_txd /
4715 	    MAX_SINGLE_PACKET_FRACTION)
4716 		scctx->isc_tx_tso_segments_max = max(1,
4717 		    num_txd / 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 	/* Set up cpu set.  If it fails, use the set of all CPUs. */
4748 	if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) != 0) {
4749 		device_printf(dev, "Unable to fetch CPU list\n");
4750 		CPU_COPY(&all_cpus, &ctx->ifc_cpus);
4751 	}
4752 	MPASS(CPU_COUNT(&ctx->ifc_cpus) > 0);
4753 
4754 	/*
4755 	** Now set up MSI or MSI-X, should return us the number of supported
4756 	** vectors (will be 1 for a legacy interrupt and MSI).
4757 	*/
4758 	if (sctx->isc_flags & IFLIB_SKIP_MSIX) {
4759 		msix = scctx->isc_vectors;
4760 	} else if (scctx->isc_msix_bar != 0)
4761 	       /*
4762 		* The simple fact that isc_msix_bar is not 0 does not mean we
4763 		* we have a good value there that is known to work.
4764 		*/
4765 		msix = iflib_msix_init(ctx);
4766 	else {
4767 		scctx->isc_vectors = 1;
4768 		scctx->isc_ntxqsets = 1;
4769 		scctx->isc_nrxqsets = 1;
4770 		scctx->isc_intr = IFLIB_INTR_LEGACY;
4771 		msix = 0;
4772 	}
4773 	/* Get memory for the station queues */
4774 	if ((err = iflib_queues_alloc(ctx))) {
4775 		device_printf(dev, "Unable to allocate queue memory\n");
4776 		goto fail_intr_free;
4777 	}
4778 
4779 	if ((err = iflib_qset_structures_setup(ctx)))
4780 		goto fail_queues;
4781 
4782 	/*
4783 	 * Now that we know how many queues there are, get the core offset.
4784 	 */
4785 	ctx->ifc_sysctl_core_offset = get_ctx_core_offset(ctx);
4786 
4787 	if (msix > 1) {
4788 		/*
4789 		 * When using MSI-X, ensure that ifdi_{r,t}x_queue_intr_enable
4790 		 * aren't the default NULL implementation.
4791 		 */
4792 		kobj_desc = &ifdi_rx_queue_intr_enable_desc;
4793 		kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL,
4794 		    kobj_desc);
4795 		if (kobj_method == &kobj_desc->deflt) {
4796 			device_printf(dev,
4797 			    "MSI-X requires ifdi_rx_queue_intr_enable method");
4798 			err = EOPNOTSUPP;
4799 			goto fail_queues;
4800 		}
4801 		kobj_desc = &ifdi_tx_queue_intr_enable_desc;
4802 		kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL,
4803 		    kobj_desc);
4804 		if (kobj_method == &kobj_desc->deflt) {
4805 			device_printf(dev,
4806 			    "MSI-X requires ifdi_tx_queue_intr_enable method");
4807 			err = EOPNOTSUPP;
4808 			goto fail_queues;
4809 		}
4810 
4811 		/*
4812 		 * Assign the MSI-X vectors.
4813 		 * Note that the default NULL ifdi_msix_intr_assign method will
4814 		 * fail here, too.
4815 		 */
4816 		err = IFDI_MSIX_INTR_ASSIGN(ctx, msix);
4817 		if (err != 0) {
4818 			device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n",
4819 			    err);
4820 			goto fail_queues;
4821 		}
4822 	} else if (scctx->isc_intr != IFLIB_INTR_MSIX) {
4823 		rid = 0;
4824 		if (scctx->isc_intr == IFLIB_INTR_MSI) {
4825 			MPASS(msix == 1);
4826 			rid = 1;
4827 		}
4828 		if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) {
4829 			device_printf(dev, "iflib_legacy_setup failed %d\n", err);
4830 			goto fail_queues;
4831 		}
4832 	} else {
4833 		device_printf(dev,
4834 		    "Cannot use iflib with only 1 MSI-X interrupt!\n");
4835 		err = ENODEV;
4836 		goto fail_intr_free;
4837 	}
4838 
4839 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
4840 
4841 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4842 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4843 		goto fail_detach;
4844 	}
4845 
4846 	/*
4847 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4848 	 * This must appear after the call to ether_ifattach() because
4849 	 * ether_ifattach() sets if_hdrlen to the default value.
4850 	 */
4851 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4852 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
4853 
4854 	if ((err = iflib_netmap_attach(ctx))) {
4855 		device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err);
4856 		goto fail_detach;
4857 	}
4858 	*ctxp = ctx;
4859 
4860 	DEBUGNET_SET(ctx->ifc_ifp, iflib);
4861 
4862 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4863 	iflib_add_device_sysctl_post(ctx);
4864 	iflib_add_pfil(ctx);
4865 	ctx->ifc_flags |= IFC_INIT_DONE;
4866 	CTX_UNLOCK(ctx);
4867 
4868 	return (0);
4869 
4870 fail_detach:
4871 	ether_ifdetach(ctx->ifc_ifp);
4872 fail_intr_free:
4873 	iflib_free_intr_mem(ctx);
4874 fail_queues:
4875 	iflib_tx_structures_free(ctx);
4876 	iflib_rx_structures_free(ctx);
4877 	iflib_tqg_detach(ctx);
4878 	IFDI_DETACH(ctx);
4879 fail_unlock:
4880 	CTX_UNLOCK(ctx);
4881 	iflib_deregister(ctx);
4882 fail_ctx_free:
4883 	device_set_softc(ctx->ifc_dev, NULL);
4884         if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4885                 free(ctx->ifc_softc, M_IFLIB);
4886         free(ctx, M_IFLIB);
4887 	return (err);
4888 }
4889 
4890 int
4891 iflib_pseudo_register(device_t dev, if_shared_ctx_t sctx, if_ctx_t *ctxp,
4892 					  struct iflib_cloneattach_ctx *clctx)
4893 {
4894 	int num_txd, num_rxd;
4895 	int err;
4896 	if_ctx_t ctx;
4897 	if_t ifp;
4898 	if_softc_ctx_t scctx;
4899 	int i;
4900 	void *sc;
4901 
4902 	ctx = malloc(sizeof(*ctx), M_IFLIB, M_WAITOK|M_ZERO);
4903 	sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4904 	ctx->ifc_flags |= IFC_SC_ALLOCATED;
4905 	if (sctx->isc_flags & (IFLIB_PSEUDO|IFLIB_VIRTUAL))
4906 		ctx->ifc_flags |= IFC_PSEUDO;
4907 
4908 	ctx->ifc_sctx = sctx;
4909 	ctx->ifc_softc = sc;
4910 	ctx->ifc_dev = dev;
4911 
4912 	if ((err = iflib_register(ctx)) != 0) {
4913 		device_printf(dev, "%s: iflib_register failed %d\n", __func__, err);
4914 		goto fail_ctx_free;
4915 	}
4916 	iflib_add_device_sysctl_pre(ctx);
4917 
4918 	scctx = &ctx->ifc_softc_ctx;
4919 	ifp = ctx->ifc_ifp;
4920 
4921 	iflib_reset_qvalues(ctx);
4922 	CTX_LOCK(ctx);
4923 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4924 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4925 		goto fail_unlock;
4926 	}
4927 	if (sctx->isc_flags & IFLIB_GEN_MAC)
4928 		ether_gen_addr(ifp, &ctx->ifc_mac);
4929 	if ((err = IFDI_CLONEATTACH(ctx, clctx->cc_ifc, clctx->cc_name,
4930 								clctx->cc_params)) != 0) {
4931 		device_printf(dev, "IFDI_CLONEATTACH failed %d\n", err);
4932 		goto fail_unlock;
4933 	}
4934 #ifdef INVARIANTS
4935 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
4936 		MPASS(scctx->isc_tx_csum_flags);
4937 #endif
4938 
4939 	if_setcapabilities(ifp, scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_LINKSTATE);
4940 	if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_LINKSTATE);
4941 
4942 	ifp->if_flags |= IFF_NOGROUP;
4943 	if (sctx->isc_flags & IFLIB_PSEUDO) {
4944 		ifmedia_add(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO, 0, NULL);
4945 		ifmedia_set(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO);
4946 		if (sctx->isc_flags & IFLIB_PSEUDO_ETHER) {
4947 			ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
4948 		} else {
4949 			if_attach(ctx->ifc_ifp);
4950 			bpfattach(ctx->ifc_ifp, DLT_NULL, sizeof(u_int32_t));
4951 		}
4952 
4953 		if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4954 			device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4955 			goto fail_detach;
4956 		}
4957 		*ctxp = ctx;
4958 
4959 		/*
4960 		 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4961 		 * This must appear after the call to ether_ifattach() because
4962 		 * ether_ifattach() sets if_hdrlen to the default value.
4963 		 */
4964 		if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4965 			if_setifheaderlen(ifp,
4966 			    sizeof(struct ether_vlan_header));
4967 
4968 		if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4969 		iflib_add_device_sysctl_post(ctx);
4970 		ctx->ifc_flags |= IFC_INIT_DONE;
4971 		CTX_UNLOCK(ctx);
4972 		return (0);
4973 	}
4974 	ifmedia_add(ctx->ifc_mediap, IFM_ETHER | IFM_1000_T | IFM_FDX, 0, NULL);
4975 	ifmedia_add(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO, 0, NULL);
4976 	ifmedia_set(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO);
4977 
4978 	_iflib_pre_assert(scctx);
4979 	ctx->ifc_txrx = *scctx->isc_txrx;
4980 
4981 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4982 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4983 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4984 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4985 
4986 	num_txd = iflib_num_tx_descs(ctx);
4987 	num_rxd = iflib_num_rx_descs(ctx);
4988 
4989 	/* XXX change for per-queue sizes */
4990 	device_printf(dev, "Using %d TX descriptors and %d RX descriptors\n",
4991 	    num_txd, num_rxd);
4992 
4993 	if (scctx->isc_tx_nsegments > num_txd / MAX_SINGLE_PACKET_FRACTION)
4994 		scctx->isc_tx_nsegments = max(1, num_txd /
4995 		    MAX_SINGLE_PACKET_FRACTION);
4996 	if (scctx->isc_tx_tso_segments_max > num_txd /
4997 	    MAX_SINGLE_PACKET_FRACTION)
4998 		scctx->isc_tx_tso_segments_max = max(1,
4999 		    num_txd / MAX_SINGLE_PACKET_FRACTION);
5000 
5001 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
5002 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
5003 		/*
5004 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
5005 		 * but some MACs do.
5006 		 */
5007 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
5008 		    IP_MAXPACKET));
5009 		/*
5010 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
5011 		 * into account.  In the worst case, each of these calls will
5012 		 * add another mbuf and, thus, the requirement for another DMA
5013 		 * segment.  So for best performance, it doesn't make sense to
5014 		 * advertize a maximum of TSO segments that typically will
5015 		 * require defragmentation in iflib_encap().
5016 		 */
5017 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
5018 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
5019 	}
5020 	if (scctx->isc_rss_table_size == 0)
5021 		scctx->isc_rss_table_size = 64;
5022 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
5023 
5024 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
5025 	/* XXX format name */
5026 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx,
5027 	    NULL, NULL, "admin");
5028 
5029 	/* XXX --- can support > 1 -- but keep it simple for now */
5030 	scctx->isc_intr = IFLIB_INTR_LEGACY;
5031 
5032 	/* Get memory for the station queues */
5033 	if ((err = iflib_queues_alloc(ctx))) {
5034 		device_printf(dev, "Unable to allocate queue memory\n");
5035 		goto fail_iflib_detach;
5036 	}
5037 
5038 	if ((err = iflib_qset_structures_setup(ctx))) {
5039 		device_printf(dev, "qset structure setup failed %d\n", err);
5040 		goto fail_queues;
5041 	}
5042 
5043 	/*
5044 	 * XXX What if anything do we want to do about interrupts?
5045 	 */
5046 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
5047 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
5048 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
5049 		goto fail_detach;
5050 	}
5051 
5052 	/*
5053 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
5054 	 * This must appear after the call to ether_ifattach() because
5055 	 * ether_ifattach() sets if_hdrlen to the default value.
5056 	 */
5057 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
5058 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
5059 
5060 	/* XXX handle more than one queue */
5061 	for (i = 0; i < scctx->isc_nrxqsets; i++)
5062 		IFDI_RX_CLSET(ctx, 0, i, ctx->ifc_rxqs[i].ifr_fl[0].ifl_sds.ifsd_cl);
5063 
5064 	*ctxp = ctx;
5065 
5066 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
5067 	iflib_add_device_sysctl_post(ctx);
5068 	ctx->ifc_flags |= IFC_INIT_DONE;
5069 	CTX_UNLOCK(ctx);
5070 
5071 	return (0);
5072 fail_detach:
5073 	ether_ifdetach(ctx->ifc_ifp);
5074 fail_queues:
5075 	iflib_tx_structures_free(ctx);
5076 	iflib_rx_structures_free(ctx);
5077 	iflib_tqg_detach(ctx);
5078 fail_iflib_detach:
5079 	IFDI_DETACH(ctx);
5080 fail_unlock:
5081 	CTX_UNLOCK(ctx);
5082 	iflib_deregister(ctx);
5083 fail_ctx_free:
5084 	free(ctx->ifc_softc, M_IFLIB);
5085 	free(ctx, M_IFLIB);
5086 	return (err);
5087 }
5088 
5089 int
5090 iflib_pseudo_deregister(if_ctx_t ctx)
5091 {
5092 	if_t ifp = ctx->ifc_ifp;
5093 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5094 
5095 	/* Unregister VLAN event handlers early */
5096 	iflib_unregister_vlan_handlers(ctx);
5097 
5098 	if ((sctx->isc_flags & IFLIB_PSEUDO)  &&
5099 		(sctx->isc_flags & IFLIB_PSEUDO_ETHER) == 0) {
5100 		bpfdetach(ifp);
5101 		if_detach(ifp);
5102 	} else {
5103 		ether_ifdetach(ifp);
5104 	}
5105 
5106 	iflib_tqg_detach(ctx);
5107 	iflib_tx_structures_free(ctx);
5108 	iflib_rx_structures_free(ctx);
5109 
5110 	iflib_deregister(ctx);
5111 
5112 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5113 		free(ctx->ifc_softc, M_IFLIB);
5114 	free(ctx, M_IFLIB);
5115 	return (0);
5116 }
5117 
5118 int
5119 iflib_device_attach(device_t dev)
5120 {
5121 	if_ctx_t ctx;
5122 	if_shared_ctx_t sctx;
5123 
5124 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
5125 		return (ENOTSUP);
5126 
5127 	pci_enable_busmaster(dev);
5128 
5129 	return (iflib_device_register(dev, NULL, sctx, &ctx));
5130 }
5131 
5132 int
5133 iflib_device_deregister(if_ctx_t ctx)
5134 {
5135 	if_t ifp = ctx->ifc_ifp;
5136 	device_t dev = ctx->ifc_dev;
5137 
5138 	/* Make sure VLANS are not using driver */
5139 	if (if_vlantrunkinuse(ifp)) {
5140 		device_printf(dev, "Vlan in use, detach first\n");
5141 		return (EBUSY);
5142 	}
5143 #ifdef PCI_IOV
5144 	if (!CTX_IS_VF(ctx) && pci_iov_detach(dev) != 0) {
5145 		device_printf(dev, "SR-IOV in use; detach first.\n");
5146 		return (EBUSY);
5147 	}
5148 #endif
5149 
5150 	STATE_LOCK(ctx);
5151 	ctx->ifc_flags |= IFC_IN_DETACH;
5152 	STATE_UNLOCK(ctx);
5153 
5154 	/* Unregister VLAN handlers before calling iflib_stop() */
5155 	iflib_unregister_vlan_handlers(ctx);
5156 
5157 	iflib_netmap_detach(ifp);
5158 	ether_ifdetach(ifp);
5159 
5160 	CTX_LOCK(ctx);
5161 	iflib_stop(ctx);
5162 	CTX_UNLOCK(ctx);
5163 
5164 	iflib_rem_pfil(ctx);
5165 	if (ctx->ifc_led_dev != NULL)
5166 		led_destroy(ctx->ifc_led_dev);
5167 
5168 	iflib_tqg_detach(ctx);
5169 	CTX_LOCK(ctx);
5170 	IFDI_DETACH(ctx);
5171 	CTX_UNLOCK(ctx);
5172 
5173 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
5174 	iflib_free_intr_mem(ctx);
5175 
5176 	bus_generic_detach(dev);
5177 
5178 	iflib_tx_structures_free(ctx);
5179 	iflib_rx_structures_free(ctx);
5180 
5181 	iflib_deregister(ctx);
5182 
5183 	device_set_softc(ctx->ifc_dev, NULL);
5184 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5185 		free(ctx->ifc_softc, M_IFLIB);
5186 	unref_ctx_core_offset(ctx);
5187 	free(ctx, M_IFLIB);
5188 	return (0);
5189 }
5190 
5191 static void
5192 iflib_tqg_detach(if_ctx_t ctx)
5193 {
5194 	iflib_txq_t txq;
5195 	iflib_rxq_t rxq;
5196 	int i;
5197 	struct taskqgroup *tqg;
5198 
5199 	/* XXX drain any dependent tasks */
5200 	tqg = qgroup_if_io_tqg;
5201 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
5202 		callout_drain(&txq->ift_timer);
5203 #ifdef DEV_NETMAP
5204 		callout_drain(&txq->ift_netmap_timer);
5205 #endif /* DEV_NETMAP */
5206 		if (txq->ift_task.gt_uniq != NULL)
5207 			taskqgroup_detach(tqg, &txq->ift_task);
5208 	}
5209 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
5210 		if (rxq->ifr_task.gt_uniq != NULL)
5211 			taskqgroup_detach(tqg, &rxq->ifr_task);
5212 	}
5213 	tqg = qgroup_if_config_tqg;
5214 	if (ctx->ifc_admin_task.gt_uniq != NULL)
5215 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
5216 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
5217 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
5218 }
5219 
5220 static void
5221 iflib_free_intr_mem(if_ctx_t ctx)
5222 {
5223 
5224 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) {
5225 		iflib_irq_free(ctx, &ctx->ifc_legacy_irq);
5226 	}
5227 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) {
5228 		pci_release_msi(ctx->ifc_dev);
5229 	}
5230 	if (ctx->ifc_msix_mem != NULL) {
5231 		bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY,
5232 		    rman_get_rid(ctx->ifc_msix_mem), ctx->ifc_msix_mem);
5233 		ctx->ifc_msix_mem = NULL;
5234 	}
5235 }
5236 
5237 int
5238 iflib_device_detach(device_t dev)
5239 {
5240 	if_ctx_t ctx = device_get_softc(dev);
5241 
5242 	return (iflib_device_deregister(ctx));
5243 }
5244 
5245 int
5246 iflib_device_suspend(device_t dev)
5247 {
5248 	if_ctx_t ctx = device_get_softc(dev);
5249 
5250 	CTX_LOCK(ctx);
5251 	IFDI_SUSPEND(ctx);
5252 	CTX_UNLOCK(ctx);
5253 
5254 	return bus_generic_suspend(dev);
5255 }
5256 int
5257 iflib_device_shutdown(device_t dev)
5258 {
5259 	if_ctx_t ctx = device_get_softc(dev);
5260 
5261 	CTX_LOCK(ctx);
5262 	IFDI_SHUTDOWN(ctx);
5263 	CTX_UNLOCK(ctx);
5264 
5265 	return bus_generic_suspend(dev);
5266 }
5267 
5268 int
5269 iflib_device_resume(device_t dev)
5270 {
5271 	if_ctx_t ctx = device_get_softc(dev);
5272 	iflib_txq_t txq = ctx->ifc_txqs;
5273 
5274 	CTX_LOCK(ctx);
5275 	IFDI_RESUME(ctx);
5276 	iflib_if_init_locked(ctx);
5277 	CTX_UNLOCK(ctx);
5278 	for (int i = 0; i < NTXQSETS(ctx); i++, txq++)
5279 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
5280 
5281 	return (bus_generic_resume(dev));
5282 }
5283 
5284 int
5285 iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params)
5286 {
5287 	int error;
5288 	if_ctx_t ctx = device_get_softc(dev);
5289 
5290 	CTX_LOCK(ctx);
5291 	error = IFDI_IOV_INIT(ctx, num_vfs, params);
5292 	CTX_UNLOCK(ctx);
5293 
5294 	return (error);
5295 }
5296 
5297 void
5298 iflib_device_iov_uninit(device_t dev)
5299 {
5300 	if_ctx_t ctx = device_get_softc(dev);
5301 
5302 	CTX_LOCK(ctx);
5303 	IFDI_IOV_UNINIT(ctx);
5304 	CTX_UNLOCK(ctx);
5305 }
5306 
5307 int
5308 iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params)
5309 {
5310 	int error;
5311 	if_ctx_t ctx = device_get_softc(dev);
5312 
5313 	CTX_LOCK(ctx);
5314 	error = IFDI_IOV_VF_ADD(ctx, vfnum, params);
5315 	CTX_UNLOCK(ctx);
5316 
5317 	return (error);
5318 }
5319 
5320 /*********************************************************************
5321  *
5322  *  MODULE FUNCTION DEFINITIONS
5323  *
5324  **********************************************************************/
5325 
5326 /*
5327  * - Start a fast taskqueue thread for each core
5328  * - Start a taskqueue for control operations
5329  */
5330 static int
5331 iflib_module_init(void)
5332 {
5333 	iflib_timer_default = hz / 2;
5334 	return (0);
5335 }
5336 
5337 static int
5338 iflib_module_event_handler(module_t mod, int what, void *arg)
5339 {
5340 	int err;
5341 
5342 	switch (what) {
5343 	case MOD_LOAD:
5344 		if ((err = iflib_module_init()) != 0)
5345 			return (err);
5346 		break;
5347 	case MOD_UNLOAD:
5348 		return (EBUSY);
5349 	default:
5350 		return (EOPNOTSUPP);
5351 	}
5352 
5353 	return (0);
5354 }
5355 
5356 /*********************************************************************
5357  *
5358  *  PUBLIC FUNCTION DEFINITIONS
5359  *     ordered as in iflib.h
5360  *
5361  **********************************************************************/
5362 
5363 static void
5364 _iflib_assert(if_shared_ctx_t sctx)
5365 {
5366 	int i;
5367 
5368 	MPASS(sctx->isc_tx_maxsize);
5369 	MPASS(sctx->isc_tx_maxsegsize);
5370 
5371 	MPASS(sctx->isc_rx_maxsize);
5372 	MPASS(sctx->isc_rx_nsegments);
5373 	MPASS(sctx->isc_rx_maxsegsize);
5374 
5375 	MPASS(sctx->isc_nrxqs >= 1 && sctx->isc_nrxqs <= 8);
5376 	for (i = 0; i < sctx->isc_nrxqs; i++) {
5377 		MPASS(sctx->isc_nrxd_min[i]);
5378 		MPASS(powerof2(sctx->isc_nrxd_min[i]));
5379 		MPASS(sctx->isc_nrxd_max[i]);
5380 		MPASS(powerof2(sctx->isc_nrxd_max[i]));
5381 		MPASS(sctx->isc_nrxd_default[i]);
5382 		MPASS(powerof2(sctx->isc_nrxd_default[i]));
5383 	}
5384 
5385 	MPASS(sctx->isc_ntxqs >= 1 && sctx->isc_ntxqs <= 8);
5386 	for (i = 0; i < sctx->isc_ntxqs; i++) {
5387 		MPASS(sctx->isc_ntxd_min[i]);
5388 		MPASS(powerof2(sctx->isc_ntxd_min[i]));
5389 		MPASS(sctx->isc_ntxd_max[i]);
5390 		MPASS(powerof2(sctx->isc_ntxd_max[i]));
5391 		MPASS(sctx->isc_ntxd_default[i]);
5392 		MPASS(powerof2(sctx->isc_ntxd_default[i]));
5393 	}
5394 }
5395 
5396 static void
5397 _iflib_pre_assert(if_softc_ctx_t scctx)
5398 {
5399 
5400 	MPASS(scctx->isc_txrx->ift_txd_encap);
5401 	MPASS(scctx->isc_txrx->ift_txd_flush);
5402 	MPASS(scctx->isc_txrx->ift_txd_credits_update);
5403 	MPASS(scctx->isc_txrx->ift_rxd_available);
5404 	MPASS(scctx->isc_txrx->ift_rxd_pkt_get);
5405 	MPASS(scctx->isc_txrx->ift_rxd_refill);
5406 	MPASS(scctx->isc_txrx->ift_rxd_flush);
5407 }
5408 
5409 static int
5410 iflib_register(if_ctx_t ctx)
5411 {
5412 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5413 	driver_t *driver = sctx->isc_driver;
5414 	device_t dev = ctx->ifc_dev;
5415 	if_t ifp;
5416 	u_char type;
5417 	int iflags;
5418 
5419 	if ((sctx->isc_flags & IFLIB_PSEUDO) == 0)
5420 		_iflib_assert(sctx);
5421 
5422 	CTX_LOCK_INIT(ctx);
5423 	STATE_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev));
5424 	if (sctx->isc_flags & IFLIB_PSEUDO) {
5425 		if (sctx->isc_flags & IFLIB_PSEUDO_ETHER)
5426 			type = IFT_ETHER;
5427 		else
5428 			type = IFT_PPP;
5429 	} else
5430 		type = IFT_ETHER;
5431 	ifp = ctx->ifc_ifp = if_alloc(type);
5432 	if (ifp == NULL) {
5433 		device_printf(dev, "can not allocate ifnet structure\n");
5434 		return (ENOMEM);
5435 	}
5436 
5437 	/*
5438 	 * Initialize our context's device specific methods
5439 	 */
5440 	kobj_init((kobj_t) ctx, (kobj_class_t) driver);
5441 	kobj_class_compile((kobj_class_t) driver);
5442 
5443 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
5444 	if_setsoftc(ifp, ctx);
5445 	if_setdev(ifp, dev);
5446 	if_setinitfn(ifp, iflib_if_init);
5447 	if_setioctlfn(ifp, iflib_if_ioctl);
5448 #ifdef ALTQ
5449 	if_setstartfn(ifp, iflib_altq_if_start);
5450 	if_settransmitfn(ifp, iflib_altq_if_transmit);
5451 	if_setsendqready(ifp);
5452 #else
5453 	if_settransmitfn(ifp, iflib_if_transmit);
5454 #endif
5455 	if_setqflushfn(ifp, iflib_if_qflush);
5456 	iflags = IFF_MULTICAST | IFF_KNOWSEPOCH;
5457 
5458 	if ((sctx->isc_flags & IFLIB_PSEUDO) &&
5459 		(sctx->isc_flags & IFLIB_PSEUDO_ETHER) == 0)
5460 		iflags |= IFF_POINTOPOINT;
5461 	else
5462 		iflags |= IFF_BROADCAST | IFF_SIMPLEX;
5463 	if_setflags(ifp, iflags);
5464 	ctx->ifc_vlan_attach_event =
5465 		EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx,
5466 							  EVENTHANDLER_PRI_FIRST);
5467 	ctx->ifc_vlan_detach_event =
5468 		EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx,
5469 							  EVENTHANDLER_PRI_FIRST);
5470 
5471 	if ((sctx->isc_flags & IFLIB_DRIVER_MEDIA) == 0) {
5472 		ctx->ifc_mediap = &ctx->ifc_media;
5473 		ifmedia_init(ctx->ifc_mediap, IFM_IMASK,
5474 		    iflib_media_change, iflib_media_status);
5475 	}
5476 	return (0);
5477 }
5478 
5479 static void
5480 iflib_unregister_vlan_handlers(if_ctx_t ctx)
5481 {
5482 	/* Unregister VLAN events */
5483 	if (ctx->ifc_vlan_attach_event != NULL) {
5484 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
5485 		ctx->ifc_vlan_attach_event = NULL;
5486 	}
5487 	if (ctx->ifc_vlan_detach_event != NULL) {
5488 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
5489 		ctx->ifc_vlan_detach_event = NULL;
5490 	}
5491 
5492 }
5493 
5494 static void
5495 iflib_deregister(if_ctx_t ctx)
5496 {
5497 	if_t ifp = ctx->ifc_ifp;
5498 
5499 	/* Remove all media */
5500 	ifmedia_removeall(&ctx->ifc_media);
5501 
5502 	/* Ensure that VLAN event handlers are unregistered */
5503 	iflib_unregister_vlan_handlers(ctx);
5504 
5505 	/* Release kobject reference */
5506 	kobj_delete((kobj_t) ctx, NULL);
5507 
5508 	/* Free the ifnet structure */
5509 	if_free(ifp);
5510 
5511 	STATE_LOCK_DESTROY(ctx);
5512 
5513 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
5514 	CTX_LOCK_DESTROY(ctx);
5515 }
5516 
5517 static int
5518 iflib_queues_alloc(if_ctx_t ctx)
5519 {
5520 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5521 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5522 	device_t dev = ctx->ifc_dev;
5523 	int nrxqsets = scctx->isc_nrxqsets;
5524 	int ntxqsets = scctx->isc_ntxqsets;
5525 	iflib_txq_t txq;
5526 	iflib_rxq_t rxq;
5527 	iflib_fl_t fl = NULL;
5528 	int i, j, cpu, err, txconf, rxconf;
5529 	iflib_dma_info_t ifdip;
5530 	uint32_t *rxqsizes = scctx->isc_rxqsizes;
5531 	uint32_t *txqsizes = scctx->isc_txqsizes;
5532 	uint8_t nrxqs = sctx->isc_nrxqs;
5533 	uint8_t ntxqs = sctx->isc_ntxqs;
5534 	int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1;
5535 	caddr_t *vaddrs;
5536 	uint64_t *paddrs;
5537 
5538 	KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1"));
5539 	KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1"));
5540 
5541 	/* Allocate the TX ring struct memory */
5542 	if (!(ctx->ifc_txqs =
5543 	    (iflib_txq_t) malloc(sizeof(struct iflib_txq) *
5544 	    ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5545 		device_printf(dev, "Unable to allocate TX ring memory\n");
5546 		err = ENOMEM;
5547 		goto fail;
5548 	}
5549 
5550 	/* Now allocate the RX */
5551 	if (!(ctx->ifc_rxqs =
5552 	    (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) *
5553 	    nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5554 		device_printf(dev, "Unable to allocate RX ring memory\n");
5555 		err = ENOMEM;
5556 		goto rx_fail;
5557 	}
5558 
5559 	txq = ctx->ifc_txqs;
5560 	rxq = ctx->ifc_rxqs;
5561 
5562 	/*
5563 	 * XXX handle allocation failure
5564 	 */
5565 	for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) {
5566 		/* Set up some basics */
5567 
5568 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs,
5569 		    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
5570 			device_printf(dev,
5571 			    "Unable to allocate TX DMA info memory\n");
5572 			err = ENOMEM;
5573 			goto err_tx_desc;
5574 		}
5575 		txq->ift_ifdi = ifdip;
5576 		for (j = 0; j < ntxqs; j++, ifdip++) {
5577 			if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, 0)) {
5578 				device_printf(dev,
5579 				    "Unable to allocate TX descriptors\n");
5580 				err = ENOMEM;
5581 				goto err_tx_desc;
5582 			}
5583 			txq->ift_txd_size[j] = scctx->isc_txd_size[j];
5584 			bzero((void *)ifdip->idi_vaddr, txqsizes[j]);
5585 		}
5586 		txq->ift_ctx = ctx;
5587 		txq->ift_id = i;
5588 		if (sctx->isc_flags & IFLIB_HAS_TXCQ) {
5589 			txq->ift_br_offset = 1;
5590 		} else {
5591 			txq->ift_br_offset = 0;
5592 		}
5593 
5594 		if (iflib_txsd_alloc(txq)) {
5595 			device_printf(dev, "Critical Failure setting up TX buffers\n");
5596 			err = ENOMEM;
5597 			goto err_tx_desc;
5598 		}
5599 
5600 		/* Initialize the TX lock */
5601 		snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:TX(%d):callout",
5602 		    device_get_nameunit(dev), txq->ift_id);
5603 		mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF);
5604 		callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0);
5605 		txq->ift_timer.c_cpu = cpu;
5606 #ifdef DEV_NETMAP
5607 		callout_init_mtx(&txq->ift_netmap_timer, &txq->ift_mtx, 0);
5608 		txq->ift_netmap_timer.c_cpu = cpu;
5609 #endif /* DEV_NETMAP */
5610 
5611 		err = ifmp_ring_alloc(&txq->ift_br, 2048, txq, iflib_txq_drain,
5612 				      iflib_txq_can_drain, M_IFLIB, M_WAITOK);
5613 		if (err) {
5614 			/* XXX free any allocated rings */
5615 			device_printf(dev, "Unable to allocate buf_ring\n");
5616 			goto err_tx_desc;
5617 		}
5618 	}
5619 
5620 	for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) {
5621 		/* Set up some basics */
5622 		callout_init(&rxq->ifr_watchdog, 1);
5623 
5624 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs,
5625 		   M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
5626 			device_printf(dev,
5627 			    "Unable to allocate RX DMA info memory\n");
5628 			err = ENOMEM;
5629 			goto err_tx_desc;
5630 		}
5631 
5632 		rxq->ifr_ifdi = ifdip;
5633 		/* XXX this needs to be changed if #rx queues != #tx queues */
5634 		rxq->ifr_ntxqirq = 1;
5635 		rxq->ifr_txqid[0] = i;
5636 		for (j = 0; j < nrxqs; j++, ifdip++) {
5637 			if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, 0)) {
5638 				device_printf(dev,
5639 				    "Unable to allocate RX descriptors\n");
5640 				err = ENOMEM;
5641 				goto err_tx_desc;
5642 			}
5643 			bzero((void *)ifdip->idi_vaddr, rxqsizes[j]);
5644 		}
5645 		rxq->ifr_ctx = ctx;
5646 		rxq->ifr_id = i;
5647 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
5648 			rxq->ifr_fl_offset = 1;
5649 		} else {
5650 			rxq->ifr_fl_offset = 0;
5651 		}
5652 		rxq->ifr_nfl = nfree_lists;
5653 		if (!(fl =
5654 			  (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) {
5655 			device_printf(dev, "Unable to allocate free list memory\n");
5656 			err = ENOMEM;
5657 			goto err_tx_desc;
5658 		}
5659 		rxq->ifr_fl = fl;
5660 		for (j = 0; j < nfree_lists; j++) {
5661 			fl[j].ifl_rxq = rxq;
5662 			fl[j].ifl_id = j;
5663 			fl[j].ifl_ifdi = &rxq->ifr_ifdi[j + rxq->ifr_fl_offset];
5664 			fl[j].ifl_rxd_size = scctx->isc_rxd_size[j];
5665 		}
5666 		/* Allocate receive buffers for the ring */
5667 		if (iflib_rxsd_alloc(rxq)) {
5668 			device_printf(dev,
5669 			    "Critical Failure setting up receive buffers\n");
5670 			err = ENOMEM;
5671 			goto err_rx_desc;
5672 		}
5673 
5674 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5675 			fl->ifl_rx_bitmap = bit_alloc(fl->ifl_size, M_IFLIB,
5676 			    M_WAITOK);
5677 	}
5678 
5679 	/* TXQs */
5680 	vaddrs = malloc(sizeof(caddr_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5681 	paddrs = malloc(sizeof(uint64_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5682 	for (i = 0; i < ntxqsets; i++) {
5683 		iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi;
5684 
5685 		for (j = 0; j < ntxqs; j++, di++) {
5686 			vaddrs[i*ntxqs + j] = di->idi_vaddr;
5687 			paddrs[i*ntxqs + j] = di->idi_paddr;
5688 		}
5689 	}
5690 	if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) {
5691 		device_printf(ctx->ifc_dev,
5692 		    "Unable to allocate device TX queue\n");
5693 		iflib_tx_structures_free(ctx);
5694 		free(vaddrs, M_IFLIB);
5695 		free(paddrs, M_IFLIB);
5696 		goto err_rx_desc;
5697 	}
5698 	free(vaddrs, M_IFLIB);
5699 	free(paddrs, M_IFLIB);
5700 
5701 	/* RXQs */
5702 	vaddrs = malloc(sizeof(caddr_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5703 	paddrs = malloc(sizeof(uint64_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5704 	for (i = 0; i < nrxqsets; i++) {
5705 		iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi;
5706 
5707 		for (j = 0; j < nrxqs; j++, di++) {
5708 			vaddrs[i*nrxqs + j] = di->idi_vaddr;
5709 			paddrs[i*nrxqs + j] = di->idi_paddr;
5710 		}
5711 	}
5712 	if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) {
5713 		device_printf(ctx->ifc_dev,
5714 		    "Unable to allocate device RX queue\n");
5715 		iflib_tx_structures_free(ctx);
5716 		free(vaddrs, M_IFLIB);
5717 		free(paddrs, M_IFLIB);
5718 		goto err_rx_desc;
5719 	}
5720 	free(vaddrs, M_IFLIB);
5721 	free(paddrs, M_IFLIB);
5722 
5723 	return (0);
5724 
5725 /* XXX handle allocation failure changes */
5726 err_rx_desc:
5727 err_tx_desc:
5728 rx_fail:
5729 	if (ctx->ifc_rxqs != NULL)
5730 		free(ctx->ifc_rxqs, M_IFLIB);
5731 	ctx->ifc_rxqs = NULL;
5732 	if (ctx->ifc_txqs != NULL)
5733 		free(ctx->ifc_txqs, M_IFLIB);
5734 	ctx->ifc_txqs = NULL;
5735 fail:
5736 	return (err);
5737 }
5738 
5739 static int
5740 iflib_tx_structures_setup(if_ctx_t ctx)
5741 {
5742 	iflib_txq_t txq = ctx->ifc_txqs;
5743 	int i;
5744 
5745 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
5746 		iflib_txq_setup(txq);
5747 
5748 	return (0);
5749 }
5750 
5751 static void
5752 iflib_tx_structures_free(if_ctx_t ctx)
5753 {
5754 	iflib_txq_t txq = ctx->ifc_txqs;
5755 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5756 	int i, j;
5757 
5758 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
5759 		for (j = 0; j < sctx->isc_ntxqs; j++)
5760 			iflib_dma_free(&txq->ift_ifdi[j]);
5761 		iflib_txq_destroy(txq);
5762 	}
5763 	free(ctx->ifc_txqs, M_IFLIB);
5764 	ctx->ifc_txqs = NULL;
5765 	IFDI_QUEUES_FREE(ctx);
5766 }
5767 
5768 /*********************************************************************
5769  *
5770  *  Initialize all receive rings.
5771  *
5772  **********************************************************************/
5773 static int
5774 iflib_rx_structures_setup(if_ctx_t ctx)
5775 {
5776 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5777 	int q;
5778 #if defined(INET6) || defined(INET)
5779 	int err, i;
5780 #endif
5781 
5782 	for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) {
5783 #if defined(INET6) || defined(INET)
5784 		if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_LRO) {
5785 			err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp,
5786 			    TCP_LRO_ENTRIES, min(1024,
5787 			    ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset]));
5788 			if (err != 0) {
5789 				device_printf(ctx->ifc_dev,
5790 				    "LRO Initialization failed!\n");
5791 				goto fail;
5792 			}
5793 		}
5794 #endif
5795 		IFDI_RXQ_SETUP(ctx, rxq->ifr_id);
5796 	}
5797 	return (0);
5798 #if defined(INET6) || defined(INET)
5799 fail:
5800 	/*
5801 	 * Free LRO resources allocated so far, we will only handle
5802 	 * the rings that completed, the failing case will have
5803 	 * cleaned up for itself.  'q' failed, so its the terminus.
5804 	 */
5805 	rxq = ctx->ifc_rxqs;
5806 	for (i = 0; i < q; ++i, rxq++) {
5807 		if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_LRO)
5808 			tcp_lro_free(&rxq->ifr_lc);
5809 	}
5810 	return (err);
5811 #endif
5812 }
5813 
5814 /*********************************************************************
5815  *
5816  *  Free all receive rings.
5817  *
5818  **********************************************************************/
5819 static void
5820 iflib_rx_structures_free(if_ctx_t ctx)
5821 {
5822 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5823 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5824 	int i, j;
5825 
5826 	for (i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) {
5827 		for (j = 0; j < sctx->isc_nrxqs; j++)
5828 			iflib_dma_free(&rxq->ifr_ifdi[j]);
5829 		iflib_rx_sds_free(rxq);
5830 #if defined(INET6) || defined(INET)
5831 		if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_LRO)
5832 			tcp_lro_free(&rxq->ifr_lc);
5833 #endif
5834 	}
5835 	free(ctx->ifc_rxqs, M_IFLIB);
5836 	ctx->ifc_rxqs = NULL;
5837 }
5838 
5839 static int
5840 iflib_qset_structures_setup(if_ctx_t ctx)
5841 {
5842 	int err;
5843 
5844 	/*
5845 	 * It is expected that the caller takes care of freeing queues if this
5846 	 * fails.
5847 	 */
5848 	if ((err = iflib_tx_structures_setup(ctx)) != 0) {
5849 		device_printf(ctx->ifc_dev, "iflib_tx_structures_setup failed: %d\n", err);
5850 		return (err);
5851 	}
5852 
5853 	if ((err = iflib_rx_structures_setup(ctx)) != 0)
5854 		device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err);
5855 
5856 	return (err);
5857 }
5858 
5859 int
5860 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
5861 		driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, const char *name)
5862 {
5863 
5864 	return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name));
5865 }
5866 
5867 #ifdef SMP
5868 static int
5869 find_nth(if_ctx_t ctx, int qid)
5870 {
5871 	cpuset_t cpus;
5872 	int i, cpuid, eqid, count;
5873 
5874 	CPU_COPY(&ctx->ifc_cpus, &cpus);
5875 	count = CPU_COUNT(&cpus);
5876 	eqid = qid % count;
5877 	/* clear up to the qid'th bit */
5878 	for (i = 0; i < eqid; i++) {
5879 		cpuid = CPU_FFS(&cpus);
5880 		MPASS(cpuid != 0);
5881 		CPU_CLR(cpuid-1, &cpus);
5882 	}
5883 	cpuid = CPU_FFS(&cpus);
5884 	MPASS(cpuid != 0);
5885 	return (cpuid-1);
5886 }
5887 
5888 #ifdef SCHED_ULE
5889 extern struct cpu_group *cpu_top;              /* CPU topology */
5890 
5891 static int
5892 find_child_with_core(int cpu, struct cpu_group *grp)
5893 {
5894 	int i;
5895 
5896 	if (grp->cg_children == 0)
5897 		return -1;
5898 
5899 	MPASS(grp->cg_child);
5900 	for (i = 0; i < grp->cg_children; i++) {
5901 		if (CPU_ISSET(cpu, &grp->cg_child[i].cg_mask))
5902 			return i;
5903 	}
5904 
5905 	return -1;
5906 }
5907 
5908 /*
5909  * Find the nth "close" core to the specified core
5910  * "close" is defined as the deepest level that shares
5911  * at least an L2 cache.  With threads, this will be
5912  * threads on the same core.  If the shared cache is L3
5913  * or higher, simply returns the same core.
5914  */
5915 static int
5916 find_close_core(int cpu, int core_offset)
5917 {
5918 	struct cpu_group *grp;
5919 	int i;
5920 	int fcpu;
5921 	cpuset_t cs;
5922 
5923 	grp = cpu_top;
5924 	if (grp == NULL)
5925 		return cpu;
5926 	i = 0;
5927 	while ((i = find_child_with_core(cpu, grp)) != -1) {
5928 		/* If the child only has one cpu, don't descend */
5929 		if (grp->cg_child[i].cg_count <= 1)
5930 			break;
5931 		grp = &grp->cg_child[i];
5932 	}
5933 
5934 	/* If they don't share at least an L2 cache, use the same CPU */
5935 	if (grp->cg_level > CG_SHARE_L2 || grp->cg_level == CG_SHARE_NONE)
5936 		return cpu;
5937 
5938 	/* Now pick one */
5939 	CPU_COPY(&grp->cg_mask, &cs);
5940 
5941 	/* Add the selected CPU offset to core offset. */
5942 	for (i = 0; (fcpu = CPU_FFS(&cs)) != 0; i++) {
5943 		if (fcpu - 1 == cpu)
5944 			break;
5945 		CPU_CLR(fcpu - 1, &cs);
5946 	}
5947 	MPASS(fcpu);
5948 
5949 	core_offset += i;
5950 
5951 	CPU_COPY(&grp->cg_mask, &cs);
5952 	for (i = core_offset % grp->cg_count; i > 0; i--) {
5953 		MPASS(CPU_FFS(&cs));
5954 		CPU_CLR(CPU_FFS(&cs) - 1, &cs);
5955 	}
5956 	MPASS(CPU_FFS(&cs));
5957 	return CPU_FFS(&cs) - 1;
5958 }
5959 #else
5960 static int
5961 find_close_core(int cpu, int core_offset __unused)
5962 {
5963 	return cpu;
5964 }
5965 #endif
5966 
5967 static int
5968 get_core_offset(if_ctx_t ctx, iflib_intr_type_t type, int qid)
5969 {
5970 	switch (type) {
5971 	case IFLIB_INTR_TX:
5972 		/* TX queues get cores which share at least an L2 cache with the corresponding RX queue */
5973 		/* XXX handle multiple RX threads per core and more than two core per L2 group */
5974 		return qid / CPU_COUNT(&ctx->ifc_cpus) + 1;
5975 	case IFLIB_INTR_RX:
5976 	case IFLIB_INTR_RXTX:
5977 		/* RX queues get the specified core */
5978 		return qid / CPU_COUNT(&ctx->ifc_cpus);
5979 	default:
5980 		return -1;
5981 	}
5982 }
5983 #else
5984 #define get_core_offset(ctx, type, qid)	CPU_FIRST()
5985 #define find_close_core(cpuid, tid)	CPU_FIRST()
5986 #define find_nth(ctx, gid)		CPU_FIRST()
5987 #endif
5988 
5989 /* Just to avoid copy/paste */
5990 static inline int
5991 iflib_irq_set_affinity(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type,
5992     int qid, struct grouptask *gtask, struct taskqgroup *tqg, void *uniq,
5993     const char *name)
5994 {
5995 	device_t dev;
5996 	int co, cpuid, err, tid;
5997 
5998 	dev = ctx->ifc_dev;
5999 	co = ctx->ifc_sysctl_core_offset;
6000 	if (ctx->ifc_sysctl_separate_txrx && type == IFLIB_INTR_TX)
6001 		co += ctx->ifc_softc_ctx.isc_nrxqsets;
6002 	cpuid = find_nth(ctx, qid + co);
6003 	tid = get_core_offset(ctx, type, qid);
6004 	if (tid < 0) {
6005 		device_printf(dev, "get_core_offset failed\n");
6006 		return (EOPNOTSUPP);
6007 	}
6008 	cpuid = find_close_core(cpuid, tid);
6009 	err = taskqgroup_attach_cpu(tqg, gtask, uniq, cpuid, dev, irq->ii_res,
6010 	    name);
6011 	if (err) {
6012 		device_printf(dev, "taskqgroup_attach_cpu failed %d\n", err);
6013 		return (err);
6014 	}
6015 #ifdef notyet
6016 	if (cpuid > ctx->ifc_cpuid_highest)
6017 		ctx->ifc_cpuid_highest = cpuid;
6018 #endif
6019 	return (0);
6020 }
6021 
6022 int
6023 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid,
6024 			iflib_intr_type_t type, driver_filter_t *filter,
6025 			void *filter_arg, int qid, const char *name)
6026 {
6027 	device_t dev;
6028 	struct grouptask *gtask;
6029 	struct taskqgroup *tqg;
6030 	iflib_filter_info_t info;
6031 	gtask_fn_t *fn;
6032 	int tqrid, err;
6033 	driver_filter_t *intr_fast;
6034 	void *q;
6035 
6036 	info = &ctx->ifc_filter_info;
6037 	tqrid = rid;
6038 
6039 	switch (type) {
6040 	/* XXX merge tx/rx for netmap? */
6041 	case IFLIB_INTR_TX:
6042 		q = &ctx->ifc_txqs[qid];
6043 		info = &ctx->ifc_txqs[qid].ift_filter_info;
6044 		gtask = &ctx->ifc_txqs[qid].ift_task;
6045 		tqg = qgroup_if_io_tqg;
6046 		fn = _task_fn_tx;
6047 		intr_fast = iflib_fast_intr;
6048 		GROUPTASK_INIT(gtask, 0, fn, q);
6049 		ctx->ifc_flags |= IFC_NETMAP_TX_IRQ;
6050 		break;
6051 	case IFLIB_INTR_RX:
6052 		q = &ctx->ifc_rxqs[qid];
6053 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
6054 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6055 		tqg = qgroup_if_io_tqg;
6056 		fn = _task_fn_rx;
6057 		intr_fast = iflib_fast_intr;
6058 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6059 		break;
6060 	case IFLIB_INTR_RXTX:
6061 		q = &ctx->ifc_rxqs[qid];
6062 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
6063 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6064 		tqg = qgroup_if_io_tqg;
6065 		fn = _task_fn_rx;
6066 		intr_fast = iflib_fast_intr_rxtx;
6067 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6068 		break;
6069 	case IFLIB_INTR_ADMIN:
6070 		q = ctx;
6071 		tqrid = -1;
6072 		info = &ctx->ifc_filter_info;
6073 		gtask = &ctx->ifc_admin_task;
6074 		tqg = qgroup_if_config_tqg;
6075 		fn = _task_fn_admin;
6076 		intr_fast = iflib_fast_intr_ctx;
6077 		break;
6078 	default:
6079 		device_printf(ctx->ifc_dev, "%s: unknown net intr type\n",
6080 		    __func__);
6081 		return (EINVAL);
6082 	}
6083 
6084 	info->ifi_filter = filter;
6085 	info->ifi_filter_arg = filter_arg;
6086 	info->ifi_task = gtask;
6087 	info->ifi_ctx = q;
6088 
6089 	dev = ctx->ifc_dev;
6090 	err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info,  name);
6091 	if (err != 0) {
6092 		device_printf(dev, "_iflib_irq_alloc failed %d\n", err);
6093 		return (err);
6094 	}
6095 	if (type == IFLIB_INTR_ADMIN)
6096 		return (0);
6097 
6098 	if (tqrid != -1) {
6099 		err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg,
6100 		    q, name);
6101 		if (err)
6102 			return (err);
6103 	} else {
6104 		taskqgroup_attach(tqg, gtask, q, dev, irq->ii_res, name);
6105 	}
6106 
6107 	return (0);
6108 }
6109 
6110 void
6111 iflib_softirq_alloc_generic(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type, void *arg, int qid, const char *name)
6112 {
6113 	struct grouptask *gtask;
6114 	struct taskqgroup *tqg;
6115 	gtask_fn_t *fn;
6116 	void *q;
6117 	int err;
6118 
6119 	switch (type) {
6120 	case IFLIB_INTR_TX:
6121 		q = &ctx->ifc_txqs[qid];
6122 		gtask = &ctx->ifc_txqs[qid].ift_task;
6123 		tqg = qgroup_if_io_tqg;
6124 		fn = _task_fn_tx;
6125 		GROUPTASK_INIT(gtask, 0, fn, q);
6126 		break;
6127 	case IFLIB_INTR_RX:
6128 		q = &ctx->ifc_rxqs[qid];
6129 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6130 		tqg = qgroup_if_io_tqg;
6131 		fn = _task_fn_rx;
6132 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6133 		break;
6134 	case IFLIB_INTR_IOV:
6135 		q = ctx;
6136 		gtask = &ctx->ifc_vflr_task;
6137 		tqg = qgroup_if_config_tqg;
6138 		fn = _task_fn_iov;
6139 		GROUPTASK_INIT(gtask, 0, fn, q);
6140 		break;
6141 	default:
6142 		panic("unknown net intr type");
6143 	}
6144 	if (irq != NULL) {
6145 		err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg,
6146 		    q, name);
6147 		if (err)
6148 			taskqgroup_attach(tqg, gtask, q, ctx->ifc_dev,
6149 			    irq->ii_res, name);
6150 	} else {
6151 		taskqgroup_attach(tqg, gtask, q, NULL, NULL, name);
6152 	}
6153 }
6154 
6155 void
6156 iflib_irq_free(if_ctx_t ctx, if_irq_t irq)
6157 {
6158 
6159 	if (irq->ii_tag)
6160 		bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag);
6161 
6162 	if (irq->ii_res)
6163 		bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ,
6164 		    rman_get_rid(irq->ii_res), irq->ii_res);
6165 }
6166 
6167 static int
6168 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, const char *name)
6169 {
6170 	iflib_txq_t txq = ctx->ifc_txqs;
6171 	iflib_rxq_t rxq = ctx->ifc_rxqs;
6172 	if_irq_t irq = &ctx->ifc_legacy_irq;
6173 	iflib_filter_info_t info;
6174 	device_t dev;
6175 	struct grouptask *gtask;
6176 	struct resource *res;
6177 	struct taskqgroup *tqg;
6178 	void *q;
6179 	int err, tqrid;
6180 	bool rx_only;
6181 
6182 	q = &ctx->ifc_rxqs[0];
6183 	info = &rxq[0].ifr_filter_info;
6184 	gtask = &rxq[0].ifr_task;
6185 	tqg = qgroup_if_io_tqg;
6186 	tqrid = *rid;
6187 	rx_only = (ctx->ifc_sctx->isc_flags & IFLIB_SINGLE_IRQ_RX_ONLY) != 0;
6188 
6189 	ctx->ifc_flags |= IFC_LEGACY;
6190 	info->ifi_filter = filter;
6191 	info->ifi_filter_arg = filter_arg;
6192 	info->ifi_task = gtask;
6193 	info->ifi_ctx = rx_only ? ctx : q;
6194 
6195 	dev = ctx->ifc_dev;
6196 	/* We allocate a single interrupt resource */
6197 	err = _iflib_irq_alloc(ctx, irq, tqrid, rx_only ? iflib_fast_intr_ctx :
6198 	    iflib_fast_intr_rxtx, NULL, info, name);
6199 	if (err != 0)
6200 		return (err);
6201 	NET_GROUPTASK_INIT(gtask, 0, _task_fn_rx, q);
6202 	res = irq->ii_res;
6203 	taskqgroup_attach(tqg, gtask, q, dev, res, name);
6204 
6205 	GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq);
6206 	taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, dev, res,
6207 	    "tx");
6208 	return (0);
6209 }
6210 
6211 void
6212 iflib_led_create(if_ctx_t ctx)
6213 {
6214 
6215 	ctx->ifc_led_dev = led_create(iflib_led_func, ctx,
6216 	    device_get_nameunit(ctx->ifc_dev));
6217 }
6218 
6219 void
6220 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid)
6221 {
6222 
6223 	GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
6224 }
6225 
6226 void
6227 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid)
6228 {
6229 
6230 	GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task);
6231 }
6232 
6233 void
6234 iflib_admin_intr_deferred(if_ctx_t ctx)
6235 {
6236 
6237 	MPASS(ctx->ifc_admin_task.gt_taskqueue != NULL);
6238 	GROUPTASK_ENQUEUE(&ctx->ifc_admin_task);
6239 }
6240 
6241 void
6242 iflib_iov_intr_deferred(if_ctx_t ctx)
6243 {
6244 
6245 	GROUPTASK_ENQUEUE(&ctx->ifc_vflr_task);
6246 }
6247 
6248 void
6249 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, const char *name)
6250 {
6251 
6252 	taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, NULL, NULL,
6253 	    name);
6254 }
6255 
6256 void
6257 iflib_config_gtask_init(void *ctx, struct grouptask *gtask, gtask_fn_t *fn,
6258 	const char *name)
6259 {
6260 
6261 	GROUPTASK_INIT(gtask, 0, fn, ctx);
6262 	taskqgroup_attach(qgroup_if_config_tqg, gtask, gtask, NULL, NULL,
6263 	    name);
6264 }
6265 
6266 void
6267 iflib_config_gtask_deinit(struct grouptask *gtask)
6268 {
6269 
6270 	taskqgroup_detach(qgroup_if_config_tqg, gtask);
6271 }
6272 
6273 void
6274 iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate)
6275 {
6276 	if_t ifp = ctx->ifc_ifp;
6277 	iflib_txq_t txq = ctx->ifc_txqs;
6278 
6279 	if_setbaudrate(ifp, baudrate);
6280 	if (baudrate >= IF_Gbps(10)) {
6281 		STATE_LOCK(ctx);
6282 		ctx->ifc_flags |= IFC_PREFETCH;
6283 		STATE_UNLOCK(ctx);
6284 	}
6285 	/* If link down, disable watchdog */
6286 	if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) {
6287 		for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++)
6288 			txq->ift_qstatus = IFLIB_QUEUE_IDLE;
6289 	}
6290 	ctx->ifc_link_state = link_state;
6291 	if_link_state_change(ifp, link_state);
6292 }
6293 
6294 static int
6295 iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq)
6296 {
6297 	int credits;
6298 #ifdef INVARIANTS
6299 	int credits_pre = txq->ift_cidx_processed;
6300 #endif
6301 
6302 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
6303 	    BUS_DMASYNC_POSTREAD);
6304 	if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, true)) == 0)
6305 		return (0);
6306 
6307 	txq->ift_processed += credits;
6308 	txq->ift_cidx_processed += credits;
6309 
6310 	MPASS(credits_pre + credits == txq->ift_cidx_processed);
6311 	if (txq->ift_cidx_processed >= txq->ift_size)
6312 		txq->ift_cidx_processed -= txq->ift_size;
6313 	return (credits);
6314 }
6315 
6316 static int
6317 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget)
6318 {
6319 	iflib_fl_t fl;
6320 	u_int i;
6321 
6322 	for (i = 0, fl = &rxq->ifr_fl[0]; i < rxq->ifr_nfl; i++, fl++)
6323 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
6324 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
6325 	return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx,
6326 	    budget));
6327 }
6328 
6329 void
6330 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name,
6331 	const char *description, if_int_delay_info_t info,
6332 	int offset, int value)
6333 {
6334 	info->iidi_ctx = ctx;
6335 	info->iidi_offset = offset;
6336 	info->iidi_value = value;
6337 	SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev),
6338 	    SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)),
6339 	    OID_AUTO, name, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE,
6340 	    info, 0, iflib_sysctl_int_delay, "I", description);
6341 }
6342 
6343 struct sx *
6344 iflib_ctx_lock_get(if_ctx_t ctx)
6345 {
6346 
6347 	return (&ctx->ifc_ctx_sx);
6348 }
6349 
6350 static int
6351 iflib_msix_init(if_ctx_t ctx)
6352 {
6353 	device_t dev = ctx->ifc_dev;
6354 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6355 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6356 	int admincnt, bar, err, iflib_num_rx_queues, iflib_num_tx_queues;
6357 	int msgs, queuemsgs, queues, rx_queues, tx_queues, vectors;
6358 
6359 	iflib_num_tx_queues = ctx->ifc_sysctl_ntxqs;
6360 	iflib_num_rx_queues = ctx->ifc_sysctl_nrxqs;
6361 
6362 	if (bootverbose)
6363 		device_printf(dev, "msix_init qsets capped at %d\n",
6364 		    imax(scctx->isc_ntxqsets, scctx->isc_nrxqsets));
6365 
6366 	/* Override by tuneable */
6367 	if (scctx->isc_disable_msix)
6368 		goto msi;
6369 
6370 	/* First try MSI-X */
6371 	if ((msgs = pci_msix_count(dev)) == 0) {
6372 		if (bootverbose)
6373 			device_printf(dev, "MSI-X not supported or disabled\n");
6374 		goto msi;
6375 	}
6376 
6377 	bar = ctx->ifc_softc_ctx.isc_msix_bar;
6378 	/*
6379 	 * bar == -1 => "trust me I know what I'm doing"
6380 	 * Some drivers are for hardware that is so shoddily
6381 	 * documented that no one knows which bars are which
6382 	 * so the developer has to map all bars. This hack
6383 	 * allows shoddy garbage to use MSI-X in this framework.
6384 	 */
6385 	if (bar != -1) {
6386 		ctx->ifc_msix_mem = bus_alloc_resource_any(dev,
6387 	            SYS_RES_MEMORY, &bar, RF_ACTIVE);
6388 		if (ctx->ifc_msix_mem == NULL) {
6389 			device_printf(dev, "Unable to map MSI-X table\n");
6390 			goto msi;
6391 		}
6392 	}
6393 
6394 	admincnt = sctx->isc_admin_intrcnt;
6395 #if IFLIB_DEBUG
6396 	/* use only 1 qset in debug mode */
6397 	queuemsgs = min(msgs - admincnt, 1);
6398 #else
6399 	queuemsgs = msgs - admincnt;
6400 #endif
6401 #ifdef RSS
6402 	queues = imin(queuemsgs, rss_getnumbuckets());
6403 #else
6404 	queues = queuemsgs;
6405 #endif
6406 	queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues);
6407 	if (bootverbose)
6408 		device_printf(dev,
6409 		    "intr CPUs: %d queue msgs: %d admincnt: %d\n",
6410 		    CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt);
6411 #ifdef  RSS
6412 	/* If we're doing RSS, clamp at the number of RSS buckets */
6413 	if (queues > rss_getnumbuckets())
6414 		queues = rss_getnumbuckets();
6415 #endif
6416 	if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt)
6417 		rx_queues = iflib_num_rx_queues;
6418 	else
6419 		rx_queues = queues;
6420 
6421 	if (rx_queues > scctx->isc_nrxqsets)
6422 		rx_queues = scctx->isc_nrxqsets;
6423 
6424 	/*
6425 	 * We want this to be all logical CPUs by default
6426 	 */
6427 	if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues)
6428 		tx_queues = iflib_num_tx_queues;
6429 	else
6430 		tx_queues = mp_ncpus;
6431 
6432 	if (tx_queues > scctx->isc_ntxqsets)
6433 		tx_queues = scctx->isc_ntxqsets;
6434 
6435 	if (ctx->ifc_sysctl_qs_eq_override == 0) {
6436 #ifdef INVARIANTS
6437 		if (tx_queues != rx_queues)
6438 			device_printf(dev,
6439 			    "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n",
6440 			    min(rx_queues, tx_queues), min(rx_queues, tx_queues));
6441 #endif
6442 		tx_queues = min(rx_queues, tx_queues);
6443 		rx_queues = min(rx_queues, tx_queues);
6444 	}
6445 
6446 	vectors = rx_queues + admincnt;
6447 	if (msgs < vectors) {
6448 		device_printf(dev,
6449 		    "insufficient number of MSI-X vectors "
6450 		    "(supported %d, need %d)\n", msgs, vectors);
6451 		goto msi;
6452 	}
6453 
6454 	device_printf(dev, "Using %d RX queues %d TX queues\n", rx_queues,
6455 	    tx_queues);
6456 	msgs = vectors;
6457 	if ((err = pci_alloc_msix(dev, &vectors)) == 0) {
6458 		if (vectors != msgs) {
6459 			device_printf(dev,
6460 			    "Unable to allocate sufficient MSI-X vectors "
6461 			    "(got %d, need %d)\n", vectors, msgs);
6462 			pci_release_msi(dev);
6463 			if (bar != -1) {
6464 				bus_release_resource(dev, SYS_RES_MEMORY, bar,
6465 				    ctx->ifc_msix_mem);
6466 				ctx->ifc_msix_mem = NULL;
6467 			}
6468 			goto msi;
6469 		}
6470 		device_printf(dev, "Using MSI-X interrupts with %d vectors\n",
6471 		    vectors);
6472 		scctx->isc_vectors = vectors;
6473 		scctx->isc_nrxqsets = rx_queues;
6474 		scctx->isc_ntxqsets = tx_queues;
6475 		scctx->isc_intr = IFLIB_INTR_MSIX;
6476 
6477 		return (vectors);
6478 	} else {
6479 		device_printf(dev,
6480 		    "failed to allocate %d MSI-X vectors, err: %d\n", vectors,
6481 		    err);
6482 		if (bar != -1) {
6483 			bus_release_resource(dev, SYS_RES_MEMORY, bar,
6484 			    ctx->ifc_msix_mem);
6485 			ctx->ifc_msix_mem = NULL;
6486 		}
6487 	}
6488 
6489 msi:
6490 	vectors = pci_msi_count(dev);
6491 	scctx->isc_nrxqsets = 1;
6492 	scctx->isc_ntxqsets = 1;
6493 	scctx->isc_vectors = vectors;
6494 	if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) {
6495 		device_printf(dev,"Using an MSI interrupt\n");
6496 		scctx->isc_intr = IFLIB_INTR_MSI;
6497 	} else {
6498 		scctx->isc_vectors = 1;
6499 		device_printf(dev,"Using a Legacy interrupt\n");
6500 		scctx->isc_intr = IFLIB_INTR_LEGACY;
6501 	}
6502 
6503 	return (vectors);
6504 }
6505 
6506 static const char *ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" };
6507 
6508 static int
6509 mp_ring_state_handler(SYSCTL_HANDLER_ARGS)
6510 {
6511 	int rc;
6512 	uint16_t *state = ((uint16_t *)oidp->oid_arg1);
6513 	struct sbuf *sb;
6514 	const char *ring_state = "UNKNOWN";
6515 
6516 	/* XXX needed ? */
6517 	rc = sysctl_wire_old_buffer(req, 0);
6518 	MPASS(rc == 0);
6519 	if (rc != 0)
6520 		return (rc);
6521 	sb = sbuf_new_for_sysctl(NULL, NULL, 80, req);
6522 	MPASS(sb != NULL);
6523 	if (sb == NULL)
6524 		return (ENOMEM);
6525 	if (state[3] <= 3)
6526 		ring_state = ring_states[state[3]];
6527 
6528 	sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s",
6529 		    state[0], state[1], state[2], ring_state);
6530 	rc = sbuf_finish(sb);
6531 	sbuf_delete(sb);
6532         return(rc);
6533 }
6534 
6535 enum iflib_ndesc_handler {
6536 	IFLIB_NTXD_HANDLER,
6537 	IFLIB_NRXD_HANDLER,
6538 };
6539 
6540 static int
6541 mp_ndesc_handler(SYSCTL_HANDLER_ARGS)
6542 {
6543 	if_ctx_t ctx = (void *)arg1;
6544 	enum iflib_ndesc_handler type = arg2;
6545 	char buf[256] = {0};
6546 	qidx_t *ndesc;
6547 	char *p, *next;
6548 	int nqs, rc, i;
6549 
6550 	nqs = 8;
6551 	switch(type) {
6552 	case IFLIB_NTXD_HANDLER:
6553 		ndesc = ctx->ifc_sysctl_ntxds;
6554 		if (ctx->ifc_sctx)
6555 			nqs = ctx->ifc_sctx->isc_ntxqs;
6556 		break;
6557 	case IFLIB_NRXD_HANDLER:
6558 		ndesc = ctx->ifc_sysctl_nrxds;
6559 		if (ctx->ifc_sctx)
6560 			nqs = ctx->ifc_sctx->isc_nrxqs;
6561 		break;
6562 	default:
6563 		printf("%s: unhandled type\n", __func__);
6564 		return (EINVAL);
6565 	}
6566 	if (nqs == 0)
6567 		nqs = 8;
6568 
6569 	for (i=0; i<8; i++) {
6570 		if (i >= nqs)
6571 			break;
6572 		if (i)
6573 			strcat(buf, ",");
6574 		sprintf(strchr(buf, 0), "%d", ndesc[i]);
6575 	}
6576 
6577 	rc = sysctl_handle_string(oidp, buf, sizeof(buf), req);
6578 	if (rc || req->newptr == NULL)
6579 		return rc;
6580 
6581 	for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p;
6582 	    i++, p = strsep(&next, " ,")) {
6583 		ndesc[i] = strtoul(p, NULL, 10);
6584 	}
6585 
6586 	return(rc);
6587 }
6588 
6589 #define NAME_BUFLEN 32
6590 static void
6591 iflib_add_device_sysctl_pre(if_ctx_t ctx)
6592 {
6593         device_t dev = iflib_get_dev(ctx);
6594 	struct sysctl_oid_list *child, *oid_list;
6595 	struct sysctl_ctx_list *ctx_list;
6596 	struct sysctl_oid *node;
6597 
6598 	ctx_list = device_get_sysctl_ctx(dev);
6599 	child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev));
6600 	ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, "iflib",
6601 	    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "IFLIB fields");
6602 	oid_list = SYSCTL_CHILDREN(node);
6603 
6604 	SYSCTL_ADD_CONST_STRING(ctx_list, oid_list, OID_AUTO, "driver_version",
6605 		       CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version,
6606 		       "driver version");
6607 
6608 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs",
6609 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0,
6610 			"# of txqs to use, 0 => use default #");
6611 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs",
6612 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0,
6613 			"# of rxqs to use, 0 => use default #");
6614 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_qs_enable",
6615 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0,
6616                        "permit #txq != #rxq");
6617 	SYSCTL_ADD_INT(ctx_list, oid_list, OID_AUTO, "disable_msix",
6618                       CTLFLAG_RWTUN, &ctx->ifc_softc_ctx.isc_disable_msix, 0,
6619                       "disable MSI-X (default 0)");
6620 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "rx_budget",
6621 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_rx_budget, 0,
6622 		       "set the RX budget");
6623 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "tx_abdicate",
6624 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_tx_abdicate, 0,
6625 		       "cause TX to abdicate instead of running to completion");
6626 	ctx->ifc_sysctl_core_offset = CORE_OFFSET_UNSPECIFIED;
6627 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "core_offset",
6628 		       CTLFLAG_RDTUN, &ctx->ifc_sysctl_core_offset, 0,
6629 		       "offset to start using cores at");
6630 	SYSCTL_ADD_U8(ctx_list, oid_list, OID_AUTO, "separate_txrx",
6631 		       CTLFLAG_RDTUN, &ctx->ifc_sysctl_separate_txrx, 0,
6632 		       "use separate cores for TX and RX");
6633 
6634 	/* XXX change for per-queue sizes */
6635 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_ntxds",
6636 	    CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NEEDGIANT, ctx,
6637 	    IFLIB_NTXD_HANDLER, mp_ndesc_handler, "A",
6638 	    "list of # of TX descriptors to use, 0 = use default #");
6639 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_nrxds",
6640 	    CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NEEDGIANT, ctx,
6641 	    IFLIB_NRXD_HANDLER, mp_ndesc_handler, "A",
6642 	    "list of # of RX descriptors to use, 0 = use default #");
6643 }
6644 
6645 static void
6646 iflib_add_device_sysctl_post(if_ctx_t ctx)
6647 {
6648 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6649 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6650         device_t dev = iflib_get_dev(ctx);
6651 	struct sysctl_oid_list *child;
6652 	struct sysctl_ctx_list *ctx_list;
6653 	iflib_fl_t fl;
6654 	iflib_txq_t txq;
6655 	iflib_rxq_t rxq;
6656 	int i, j;
6657 	char namebuf[NAME_BUFLEN];
6658 	char *qfmt;
6659 	struct sysctl_oid *queue_node, *fl_node, *node;
6660 	struct sysctl_oid_list *queue_list, *fl_list;
6661 	ctx_list = device_get_sysctl_ctx(dev);
6662 
6663 	node = ctx->ifc_sysctl_node;
6664 	child = SYSCTL_CHILDREN(node);
6665 
6666 	if (scctx->isc_ntxqsets > 100)
6667 		qfmt = "txq%03d";
6668 	else if (scctx->isc_ntxqsets > 10)
6669 		qfmt = "txq%02d";
6670 	else
6671 		qfmt = "txq%d";
6672 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
6673 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6674 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6675 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "Queue Name");
6676 		queue_list = SYSCTL_CHILDREN(queue_node);
6677 #if MEMORY_LOGGING
6678 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued",
6679 				CTLFLAG_RD,
6680 				&txq->ift_dequeued, "total mbufs freed");
6681 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued",
6682 				CTLFLAG_RD,
6683 				&txq->ift_enqueued, "total mbufs enqueued");
6684 #endif
6685 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag",
6686 				   CTLFLAG_RD,
6687 				   &txq->ift_mbuf_defrag, "# of times m_defrag was called");
6688 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "m_pullups",
6689 				   CTLFLAG_RD,
6690 				   &txq->ift_pullups, "# of times m_pullup was called");
6691 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag_failed",
6692 				   CTLFLAG_RD,
6693 				   &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed");
6694 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_desc_avail",
6695 				   CTLFLAG_RD,
6696 				   &txq->ift_no_desc_avail, "# of times no descriptors were available");
6697 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "tx_map_failed",
6698 				   CTLFLAG_RD,
6699 				   &txq->ift_map_failed, "# of times DMA map failed");
6700 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txd_encap_efbig",
6701 				   CTLFLAG_RD,
6702 				   &txq->ift_txd_encap_efbig, "# of times txd_encap returned EFBIG");
6703 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_tx_dma_setup",
6704 				   CTLFLAG_RD,
6705 				   &txq->ift_no_tx_dma_setup, "# of times map failed for other than EFBIG");
6706 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx",
6707 				   CTLFLAG_RD,
6708 				   &txq->ift_pidx, 1, "Producer Index");
6709 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx",
6710 				   CTLFLAG_RD,
6711 				   &txq->ift_cidx, 1, "Consumer Index");
6712 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx_processed",
6713 				   CTLFLAG_RD,
6714 				   &txq->ift_cidx_processed, 1, "Consumer Index seen by credit update");
6715 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use",
6716 				   CTLFLAG_RD,
6717 				   &txq->ift_in_use, 1, "descriptors in use");
6718 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_processed",
6719 				   CTLFLAG_RD,
6720 				   &txq->ift_processed, "descriptors procesed for clean");
6721 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned",
6722 				   CTLFLAG_RD,
6723 				   &txq->ift_cleaned, "total cleaned");
6724 		SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state",
6725 		    CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_NEEDGIANT,
6726 		    __DEVOLATILE(uint64_t *, &txq->ift_br->state), 0,
6727 		    mp_ring_state_handler, "A", "soft ring state");
6728 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_enqueues",
6729 				       CTLFLAG_RD, &txq->ift_br->enqueues,
6730 				       "# of enqueues to the mp_ring for this queue");
6731 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_drops",
6732 				       CTLFLAG_RD, &txq->ift_br->drops,
6733 				       "# of drops in the mp_ring for this queue");
6734 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_starts",
6735 				       CTLFLAG_RD, &txq->ift_br->starts,
6736 				       "# of normal consumer starts in the mp_ring for this queue");
6737 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_stalls",
6738 				       CTLFLAG_RD, &txq->ift_br->stalls,
6739 					       "# of consumer stalls in the mp_ring for this queue");
6740 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_restarts",
6741 			       CTLFLAG_RD, &txq->ift_br->restarts,
6742 				       "# of consumer restarts in the mp_ring for this queue");
6743 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_abdications",
6744 				       CTLFLAG_RD, &txq->ift_br->abdications,
6745 				       "# of consumer abdications in the mp_ring for this queue");
6746 	}
6747 
6748 	if (scctx->isc_nrxqsets > 100)
6749 		qfmt = "rxq%03d";
6750 	else if (scctx->isc_nrxqsets > 10)
6751 		qfmt = "rxq%02d";
6752 	else
6753 		qfmt = "rxq%d";
6754 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
6755 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6756 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6757 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "Queue Name");
6758 		queue_list = SYSCTL_CHILDREN(queue_node);
6759 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
6760 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_cidx",
6761 				       CTLFLAG_RD,
6762 				       &rxq->ifr_cq_cidx, 1, "Consumer Index");
6763 		}
6764 
6765 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
6766 			snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j);
6767 			fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list, OID_AUTO, namebuf,
6768 			    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "freelist Name");
6769 			fl_list = SYSCTL_CHILDREN(fl_node);
6770 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx",
6771 				       CTLFLAG_RD,
6772 				       &fl->ifl_pidx, 1, "Producer Index");
6773 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx",
6774 				       CTLFLAG_RD,
6775 				       &fl->ifl_cidx, 1, "Consumer Index");
6776 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits",
6777 				       CTLFLAG_RD,
6778 				       &fl->ifl_credits, 1, "credits available");
6779 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "buf_size",
6780 				       CTLFLAG_RD,
6781 				       &fl->ifl_buf_size, 1, "buffer size");
6782 #if MEMORY_LOGGING
6783 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_enqueued",
6784 					CTLFLAG_RD,
6785 					&fl->ifl_m_enqueued, "mbufs allocated");
6786 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_dequeued",
6787 					CTLFLAG_RD,
6788 					&fl->ifl_m_dequeued, "mbufs freed");
6789 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_enqueued",
6790 					CTLFLAG_RD,
6791 					&fl->ifl_cl_enqueued, "clusters allocated");
6792 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_dequeued",
6793 					CTLFLAG_RD,
6794 					&fl->ifl_cl_dequeued, "clusters freed");
6795 #endif
6796 		}
6797 	}
6798 
6799 }
6800 
6801 void
6802 iflib_request_reset(if_ctx_t ctx)
6803 {
6804 
6805 	STATE_LOCK(ctx);
6806 	ctx->ifc_flags |= IFC_DO_RESET;
6807 	STATE_UNLOCK(ctx);
6808 }
6809 
6810 #ifndef __NO_STRICT_ALIGNMENT
6811 static struct mbuf *
6812 iflib_fixup_rx(struct mbuf *m)
6813 {
6814 	struct mbuf *n;
6815 
6816 	if (m->m_len <= (MCLBYTES - ETHER_HDR_LEN)) {
6817 		bcopy(m->m_data, m->m_data + ETHER_HDR_LEN, m->m_len);
6818 		m->m_data += ETHER_HDR_LEN;
6819 		n = m;
6820 	} else {
6821 		MGETHDR(n, M_NOWAIT, MT_DATA);
6822 		if (n == NULL) {
6823 			m_freem(m);
6824 			return (NULL);
6825 		}
6826 		bcopy(m->m_data, n->m_data, ETHER_HDR_LEN);
6827 		m->m_data += ETHER_HDR_LEN;
6828 		m->m_len -= ETHER_HDR_LEN;
6829 		n->m_len = ETHER_HDR_LEN;
6830 		M_MOVE_PKTHDR(n, m);
6831 		n->m_next = m;
6832 	}
6833 	return (n);
6834 }
6835 #endif
6836 
6837 #ifdef DEBUGNET
6838 static void
6839 iflib_debugnet_init(if_t ifp, int *nrxr, int *ncl, int *clsize)
6840 {
6841 	if_ctx_t ctx;
6842 
6843 	ctx = if_getsoftc(ifp);
6844 	CTX_LOCK(ctx);
6845 	*nrxr = NRXQSETS(ctx);
6846 	*ncl = ctx->ifc_rxqs[0].ifr_fl->ifl_size;
6847 	*clsize = ctx->ifc_rxqs[0].ifr_fl->ifl_buf_size;
6848 	CTX_UNLOCK(ctx);
6849 }
6850 
6851 static void
6852 iflib_debugnet_event(if_t ifp, enum debugnet_ev event)
6853 {
6854 	if_ctx_t ctx;
6855 	if_softc_ctx_t scctx;
6856 	iflib_fl_t fl;
6857 	iflib_rxq_t rxq;
6858 	int i, j;
6859 
6860 	ctx = if_getsoftc(ifp);
6861 	scctx = &ctx->ifc_softc_ctx;
6862 
6863 	switch (event) {
6864 	case DEBUGNET_START:
6865 		for (i = 0; i < scctx->isc_nrxqsets; i++) {
6866 			rxq = &ctx->ifc_rxqs[i];
6867 			for (j = 0; j < rxq->ifr_nfl; j++) {
6868 				fl = rxq->ifr_fl;
6869 				fl->ifl_zone = m_getzone(fl->ifl_buf_size);
6870 			}
6871 		}
6872 		iflib_no_tx_batch = 1;
6873 		break;
6874 	default:
6875 		break;
6876 	}
6877 }
6878 
6879 static int
6880 iflib_debugnet_transmit(if_t ifp, struct mbuf *m)
6881 {
6882 	if_ctx_t ctx;
6883 	iflib_txq_t txq;
6884 	int error;
6885 
6886 	ctx = if_getsoftc(ifp);
6887 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
6888 	    IFF_DRV_RUNNING)
6889 		return (EBUSY);
6890 
6891 	txq = &ctx->ifc_txqs[0];
6892 	error = iflib_encap(txq, &m);
6893 	if (error == 0)
6894 		(void)iflib_txd_db_check(txq, true);
6895 	return (error);
6896 }
6897 
6898 static int
6899 iflib_debugnet_poll(if_t ifp, int count)
6900 {
6901 	struct epoch_tracker et;
6902 	if_ctx_t ctx;
6903 	if_softc_ctx_t scctx;
6904 	iflib_txq_t txq;
6905 	int i;
6906 
6907 	ctx = if_getsoftc(ifp);
6908 	scctx = &ctx->ifc_softc_ctx;
6909 
6910 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
6911 	    IFF_DRV_RUNNING)
6912 		return (EBUSY);
6913 
6914 	txq = &ctx->ifc_txqs[0];
6915 	(void)iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
6916 
6917 	NET_EPOCH_ENTER(et);
6918 	for (i = 0; i < scctx->isc_nrxqsets; i++)
6919 		(void)iflib_rxeof(&ctx->ifc_rxqs[i], 16 /* XXX */);
6920 	NET_EPOCH_EXIT(et);
6921 	return (0);
6922 }
6923 #endif /* DEBUGNET */
6924