xref: /freebsd/sys/net/iflib.c (revision 422d05da14fe063e5d187d81a328fa7b362d069f)
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 /*
121  * File organization:
122  *  - private structures
123  *  - iflib private utility functions
124  *  - ifnet functions
125  *  - vlan registry and other exported functions
126  *  - iflib public core functions
127  *
128  *
129  */
130 MALLOC_DEFINE(M_IFLIB, "iflib", "ifnet library");
131 
132 struct iflib_txq;
133 typedef struct iflib_txq *iflib_txq_t;
134 struct iflib_rxq;
135 typedef struct iflib_rxq *iflib_rxq_t;
136 struct iflib_fl;
137 typedef struct iflib_fl *iflib_fl_t;
138 
139 struct iflib_ctx;
140 
141 static void iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid);
142 static void iflib_timer(void *arg);
143 
144 typedef struct iflib_filter_info {
145 	driver_filter_t *ifi_filter;
146 	void *ifi_filter_arg;
147 	struct grouptask *ifi_task;
148 	void *ifi_ctx;
149 } *iflib_filter_info_t;
150 
151 struct iflib_ctx {
152 	KOBJ_FIELDS;
153 	/*
154 	 * Pointer to hardware driver's softc
155 	 */
156 	void *ifc_softc;
157 	device_t ifc_dev;
158 	if_t ifc_ifp;
159 
160 	cpuset_t ifc_cpus;
161 	if_shared_ctx_t ifc_sctx;
162 	struct if_softc_ctx ifc_softc_ctx;
163 
164 	struct sx ifc_ctx_sx;
165 	struct mtx ifc_state_mtx;
166 
167 	iflib_txq_t ifc_txqs;
168 	iflib_rxq_t ifc_rxqs;
169 	uint32_t ifc_if_flags;
170 	uint32_t ifc_flags;
171 	uint32_t ifc_max_fl_buf_size;
172 	uint32_t ifc_rx_mbuf_sz;
173 
174 	int ifc_link_state;
175 	int ifc_watchdog_events;
176 	struct cdev *ifc_led_dev;
177 	struct resource *ifc_msix_mem;
178 
179 	struct if_irq ifc_legacy_irq;
180 	struct grouptask ifc_admin_task;
181 	struct grouptask ifc_vflr_task;
182 	struct iflib_filter_info ifc_filter_info;
183 	struct ifmedia	ifc_media;
184 	struct ifmedia	*ifc_mediap;
185 
186 	struct sysctl_oid *ifc_sysctl_node;
187 	uint16_t ifc_sysctl_ntxqs;
188 	uint16_t ifc_sysctl_nrxqs;
189 	uint16_t ifc_sysctl_qs_eq_override;
190 	uint16_t ifc_sysctl_rx_budget;
191 	uint16_t ifc_sysctl_tx_abdicate;
192 	uint16_t ifc_sysctl_core_offset;
193 #define	CORE_OFFSET_UNSPECIFIED	0xffff
194 	uint8_t  ifc_sysctl_separate_txrx;
195 
196 	qidx_t ifc_sysctl_ntxds[8];
197 	qidx_t ifc_sysctl_nrxds[8];
198 	struct if_txrx ifc_txrx;
199 #define isc_txd_encap  ifc_txrx.ift_txd_encap
200 #define isc_txd_flush  ifc_txrx.ift_txd_flush
201 #define isc_txd_credits_update  ifc_txrx.ift_txd_credits_update
202 #define isc_rxd_available ifc_txrx.ift_rxd_available
203 #define isc_rxd_pkt_get ifc_txrx.ift_rxd_pkt_get
204 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
205 #define isc_rxd_flush ifc_txrx.ift_rxd_flush
206 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
207 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
208 #define isc_legacy_intr ifc_txrx.ift_legacy_intr
209 	eventhandler_tag ifc_vlan_attach_event;
210 	eventhandler_tag ifc_vlan_detach_event;
211 	struct ether_addr ifc_mac;
212 };
213 
214 void *
215 iflib_get_softc(if_ctx_t ctx)
216 {
217 
218 	return (ctx->ifc_softc);
219 }
220 
221 device_t
222 iflib_get_dev(if_ctx_t ctx)
223 {
224 
225 	return (ctx->ifc_dev);
226 }
227 
228 if_t
229 iflib_get_ifp(if_ctx_t ctx)
230 {
231 
232 	return (ctx->ifc_ifp);
233 }
234 
235 struct ifmedia *
236 iflib_get_media(if_ctx_t ctx)
237 {
238 
239 	return (ctx->ifc_mediap);
240 }
241 
242 uint32_t
243 iflib_get_flags(if_ctx_t ctx)
244 {
245 	return (ctx->ifc_flags);
246 }
247 
248 void
249 iflib_set_mac(if_ctx_t ctx, uint8_t mac[ETHER_ADDR_LEN])
250 {
251 
252 	bcopy(mac, ctx->ifc_mac.octet, ETHER_ADDR_LEN);
253 }
254 
255 if_softc_ctx_t
256 iflib_get_softc_ctx(if_ctx_t ctx)
257 {
258 
259 	return (&ctx->ifc_softc_ctx);
260 }
261 
262 if_shared_ctx_t
263 iflib_get_sctx(if_ctx_t ctx)
264 {
265 
266 	return (ctx->ifc_sctx);
267 }
268 
269 #define IP_ALIGNED(m) ((((uintptr_t)(m)->m_data) & 0x3) == 0x2)
270 #define CACHE_PTR_INCREMENT (CACHE_LINE_SIZE/sizeof(void*))
271 #define CACHE_PTR_NEXT(ptr) ((void *)(((uintptr_t)(ptr)+CACHE_LINE_SIZE-1) & (CACHE_LINE_SIZE-1)))
272 
273 #define LINK_ACTIVE(ctx) ((ctx)->ifc_link_state == LINK_STATE_UP)
274 #define CTX_IS_VF(ctx) ((ctx)->ifc_sctx->isc_flags & IFLIB_IS_VF)
275 
276 typedef struct iflib_sw_rx_desc_array {
277 	bus_dmamap_t	*ifsd_map;         /* bus_dma maps for packet */
278 	struct mbuf	**ifsd_m;           /* pkthdr mbufs */
279 	caddr_t		*ifsd_cl;          /* direct cluster pointer for rx */
280 	bus_addr_t	*ifsd_ba;          /* bus addr of cluster for rx */
281 } iflib_rxsd_array_t;
282 
283 typedef struct iflib_sw_tx_desc_array {
284 	bus_dmamap_t    *ifsd_map;         /* bus_dma maps for packet */
285 	bus_dmamap_t	*ifsd_tso_map;     /* bus_dma maps for TSO packet */
286 	struct mbuf    **ifsd_m;           /* pkthdr mbufs */
287 } if_txsd_vec_t;
288 
289 /* magic number that should be high enough for any hardware */
290 #define IFLIB_MAX_TX_SEGS		128
291 #define IFLIB_RX_COPY_THRESH		128
292 #define IFLIB_MAX_RX_REFRESH		32
293 /* The minimum descriptors per second before we start coalescing */
294 #define IFLIB_MIN_DESC_SEC		16384
295 #define IFLIB_DEFAULT_TX_UPDATE_FREQ	16
296 #define IFLIB_QUEUE_IDLE		0
297 #define IFLIB_QUEUE_HUNG		1
298 #define IFLIB_QUEUE_WORKING		2
299 /* maximum number of txqs that can share an rx interrupt */
300 #define IFLIB_MAX_TX_SHARED_INTR	4
301 
302 /* this should really scale with ring size - this is a fairly arbitrary value */
303 #define TX_BATCH_SIZE			32
304 
305 #define IFLIB_RESTART_BUDGET		8
306 
307 #define CSUM_OFFLOAD		(CSUM_IP_TSO|CSUM_IP6_TSO|CSUM_IP| \
308 				 CSUM_IP_UDP|CSUM_IP_TCP|CSUM_IP_SCTP| \
309 				 CSUM_IP6_UDP|CSUM_IP6_TCP|CSUM_IP6_SCTP)
310 
311 struct iflib_txq {
312 	qidx_t		ift_in_use;
313 	qidx_t		ift_cidx;
314 	qidx_t		ift_cidx_processed;
315 	qidx_t		ift_pidx;
316 	uint8_t		ift_gen;
317 	uint8_t		ift_br_offset;
318 	uint16_t	ift_npending;
319 	uint16_t	ift_db_pending;
320 	uint16_t	ift_rs_pending;
321 	/* implicit pad */
322 	uint8_t		ift_txd_size[8];
323 	uint64_t	ift_processed;
324 	uint64_t	ift_cleaned;
325 	uint64_t	ift_cleaned_prev;
326 #if MEMORY_LOGGING
327 	uint64_t	ift_enqueued;
328 	uint64_t	ift_dequeued;
329 #endif
330 	uint64_t	ift_no_tx_dma_setup;
331 	uint64_t	ift_no_desc_avail;
332 	uint64_t	ift_mbuf_defrag_failed;
333 	uint64_t	ift_mbuf_defrag;
334 	uint64_t	ift_map_failed;
335 	uint64_t	ift_txd_encap_efbig;
336 	uint64_t	ift_pullups;
337 	uint64_t	ift_last_timer_tick;
338 
339 	struct mtx	ift_mtx;
340 	struct mtx	ift_db_mtx;
341 
342 	/* constant values */
343 	if_ctx_t	ift_ctx;
344 	struct ifmp_ring        *ift_br;
345 	struct grouptask	ift_task;
346 	qidx_t		ift_size;
347 	uint16_t	ift_id;
348 	struct callout	ift_timer;
349 
350 	if_txsd_vec_t	ift_sds;
351 	uint8_t		ift_qstatus;
352 	uint8_t		ift_closed;
353 	uint8_t		ift_update_freq;
354 	struct iflib_filter_info ift_filter_info;
355 	bus_dma_tag_t	ift_buf_tag;
356 	bus_dma_tag_t	ift_tso_buf_tag;
357 	iflib_dma_info_t	ift_ifdi;
358 #define MTX_NAME_LEN 16
359 	char                    ift_mtx_name[MTX_NAME_LEN];
360 	bus_dma_segment_t	ift_segs[IFLIB_MAX_TX_SEGS]  __aligned(CACHE_LINE_SIZE);
361 #ifdef IFLIB_DIAGNOSTICS
362 	uint64_t ift_cpu_exec_count[256];
363 #endif
364 } __aligned(CACHE_LINE_SIZE);
365 
366 struct iflib_fl {
367 	qidx_t		ifl_cidx;
368 	qidx_t		ifl_pidx;
369 	qidx_t		ifl_credits;
370 	uint8_t		ifl_gen;
371 	uint8_t		ifl_rxd_size;
372 #if MEMORY_LOGGING
373 	uint64_t	ifl_m_enqueued;
374 	uint64_t	ifl_m_dequeued;
375 	uint64_t	ifl_cl_enqueued;
376 	uint64_t	ifl_cl_dequeued;
377 #endif
378 	/* implicit pad */
379 	bitstr_t 	*ifl_rx_bitmap;
380 	qidx_t		ifl_fragidx;
381 	/* constant */
382 	qidx_t		ifl_size;
383 	uint16_t	ifl_buf_size;
384 	uint16_t	ifl_cltype;
385 	uma_zone_t	ifl_zone;
386 	iflib_rxsd_array_t	ifl_sds;
387 	iflib_rxq_t	ifl_rxq;
388 	uint8_t		ifl_id;
389 	bus_dma_tag_t	ifl_buf_tag;
390 	iflib_dma_info_t	ifl_ifdi;
391 	uint64_t	ifl_bus_addrs[IFLIB_MAX_RX_REFRESH] __aligned(CACHE_LINE_SIZE);
392 	caddr_t		ifl_vm_addrs[IFLIB_MAX_RX_REFRESH];
393 	qidx_t	ifl_rxd_idxs[IFLIB_MAX_RX_REFRESH];
394 }  __aligned(CACHE_LINE_SIZE);
395 
396 static inline qidx_t
397 get_inuse(int size, qidx_t cidx, qidx_t pidx, uint8_t gen)
398 {
399 	qidx_t used;
400 
401 	if (pidx > cidx)
402 		used = pidx - cidx;
403 	else if (pidx < cidx)
404 		used = size - cidx + pidx;
405 	else if (gen == 0 && pidx == cidx)
406 		used = 0;
407 	else if (gen == 1 && pidx == cidx)
408 		used = size;
409 	else
410 		panic("bad state");
411 
412 	return (used);
413 }
414 
415 #define TXQ_AVAIL(txq) (txq->ift_size - get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx, txq->ift_gen))
416 
417 #define IDXDIFF(head, tail, wrap) \
418 	((head) >= (tail) ? (head) - (tail) : (wrap) - (tail) + (head))
419 
420 struct iflib_rxq {
421 	if_ctx_t	ifr_ctx;
422 	iflib_fl_t	ifr_fl;
423 	uint64_t	ifr_rx_irq;
424 	struct pfil_head	*pfil;
425 	/*
426 	 * If there is a separate completion queue (IFLIB_HAS_RXCQ), this is
427 	 * the command queue consumer index.  Otherwise it's unused.
428 	 */
429 	qidx_t		ifr_cq_cidx;
430 	uint16_t	ifr_id;
431 	uint8_t		ifr_nfl;
432 	uint8_t		ifr_ntxqirq;
433 	uint8_t		ifr_txqid[IFLIB_MAX_TX_SHARED_INTR];
434 	uint8_t		ifr_fl_offset;
435 	struct lro_ctrl			ifr_lc;
436 	struct grouptask        ifr_task;
437 	struct iflib_filter_info ifr_filter_info;
438 	iflib_dma_info_t		ifr_ifdi;
439 
440 	/* dynamically allocate if any drivers need a value substantially larger than this */
441 	struct if_rxd_frag	ifr_frags[IFLIB_MAX_RX_SEGS] __aligned(CACHE_LINE_SIZE);
442 #ifdef IFLIB_DIAGNOSTICS
443 	uint64_t ifr_cpu_exec_count[256];
444 #endif
445 }  __aligned(CACHE_LINE_SIZE);
446 
447 typedef struct if_rxsd {
448 	caddr_t *ifsd_cl;
449 	iflib_fl_t ifsd_fl;
450 	qidx_t ifsd_cidx;
451 } *if_rxsd_t;
452 
453 /* multiple of word size */
454 #ifdef __LP64__
455 #define PKT_INFO_SIZE	6
456 #define RXD_INFO_SIZE	5
457 #define PKT_TYPE uint64_t
458 #else
459 #define PKT_INFO_SIZE	11
460 #define RXD_INFO_SIZE	8
461 #define PKT_TYPE uint32_t
462 #endif
463 #define PKT_LOOP_BOUND  ((PKT_INFO_SIZE/3)*3)
464 #define RXD_LOOP_BOUND  ((RXD_INFO_SIZE/4)*4)
465 
466 typedef struct if_pkt_info_pad {
467 	PKT_TYPE pkt_val[PKT_INFO_SIZE];
468 } *if_pkt_info_pad_t;
469 typedef struct if_rxd_info_pad {
470 	PKT_TYPE rxd_val[RXD_INFO_SIZE];
471 } *if_rxd_info_pad_t;
472 
473 CTASSERT(sizeof(struct if_pkt_info_pad) == sizeof(struct if_pkt_info));
474 CTASSERT(sizeof(struct if_rxd_info_pad) == sizeof(struct if_rxd_info));
475 
476 
477 static inline void
478 pkt_info_zero(if_pkt_info_t pi)
479 {
480 	if_pkt_info_pad_t pi_pad;
481 
482 	pi_pad = (if_pkt_info_pad_t)pi;
483 	pi_pad->pkt_val[0] = 0; pi_pad->pkt_val[1] = 0; pi_pad->pkt_val[2] = 0;
484 	pi_pad->pkt_val[3] = 0; pi_pad->pkt_val[4] = 0; pi_pad->pkt_val[5] = 0;
485 #ifndef __LP64__
486 	pi_pad->pkt_val[6] = 0; pi_pad->pkt_val[7] = 0; pi_pad->pkt_val[8] = 0;
487 	pi_pad->pkt_val[9] = 0; pi_pad->pkt_val[10] = 0;
488 #endif
489 }
490 
491 static device_method_t iflib_pseudo_methods[] = {
492 	DEVMETHOD(device_attach, noop_attach),
493 	DEVMETHOD(device_detach, iflib_pseudo_detach),
494 	DEVMETHOD_END
495 };
496 
497 driver_t iflib_pseudodriver = {
498 	"iflib_pseudo", iflib_pseudo_methods, sizeof(struct iflib_ctx),
499 };
500 
501 static inline void
502 rxd_info_zero(if_rxd_info_t ri)
503 {
504 	if_rxd_info_pad_t ri_pad;
505 	int i;
506 
507 	ri_pad = (if_rxd_info_pad_t)ri;
508 	for (i = 0; i < RXD_LOOP_BOUND; i += 4) {
509 		ri_pad->rxd_val[i] = 0;
510 		ri_pad->rxd_val[i+1] = 0;
511 		ri_pad->rxd_val[i+2] = 0;
512 		ri_pad->rxd_val[i+3] = 0;
513 	}
514 #ifdef __LP64__
515 	ri_pad->rxd_val[RXD_INFO_SIZE-1] = 0;
516 #endif
517 }
518 
519 /*
520  * Only allow a single packet to take up most 1/nth of the tx ring
521  */
522 #define MAX_SINGLE_PACKET_FRACTION 12
523 #define IF_BAD_DMA (bus_addr_t)-1
524 
525 #define CTX_ACTIVE(ctx) ((if_getdrvflags((ctx)->ifc_ifp) & IFF_DRV_RUNNING))
526 
527 #define CTX_LOCK_INIT(_sc)  sx_init(&(_sc)->ifc_ctx_sx, "iflib ctx lock")
528 #define CTX_LOCK(ctx) sx_xlock(&(ctx)->ifc_ctx_sx)
529 #define CTX_UNLOCK(ctx) sx_xunlock(&(ctx)->ifc_ctx_sx)
530 #define CTX_LOCK_DESTROY(ctx) sx_destroy(&(ctx)->ifc_ctx_sx)
531 
532 #define STATE_LOCK_INIT(_sc, _name)  mtx_init(&(_sc)->ifc_state_mtx, _name, "iflib state lock", MTX_DEF)
533 #define STATE_LOCK(ctx) mtx_lock(&(ctx)->ifc_state_mtx)
534 #define STATE_UNLOCK(ctx) mtx_unlock(&(ctx)->ifc_state_mtx)
535 #define STATE_LOCK_DESTROY(ctx) mtx_destroy(&(ctx)->ifc_state_mtx)
536 
537 #define CALLOUT_LOCK(txq)	mtx_lock(&txq->ift_mtx)
538 #define CALLOUT_UNLOCK(txq) 	mtx_unlock(&txq->ift_mtx)
539 
540 void
541 iflib_set_detach(if_ctx_t ctx)
542 {
543 	STATE_LOCK(ctx);
544 	ctx->ifc_flags |= IFC_IN_DETACH;
545 	STATE_UNLOCK(ctx);
546 }
547 
548 /* Our boot-time initialization hook */
549 static int	iflib_module_event_handler(module_t, int, void *);
550 
551 static moduledata_t iflib_moduledata = {
552 	"iflib",
553 	iflib_module_event_handler,
554 	NULL
555 };
556 
557 DECLARE_MODULE(iflib, iflib_moduledata, SI_SUB_INIT_IF, SI_ORDER_ANY);
558 MODULE_VERSION(iflib, 1);
559 
560 MODULE_DEPEND(iflib, pci, 1, 1, 1);
561 MODULE_DEPEND(iflib, ether, 1, 1, 1);
562 
563 TASKQGROUP_DEFINE(if_io_tqg, mp_ncpus, 1);
564 TASKQGROUP_DEFINE(if_config_tqg, 1, 1);
565 
566 #ifndef IFLIB_DEBUG_COUNTERS
567 #ifdef INVARIANTS
568 #define IFLIB_DEBUG_COUNTERS 1
569 #else
570 #define IFLIB_DEBUG_COUNTERS 0
571 #endif /* !INVARIANTS */
572 #endif
573 
574 static SYSCTL_NODE(_net, OID_AUTO, iflib, CTLFLAG_RD, 0,
575                    "iflib driver parameters");
576 
577 /*
578  * XXX need to ensure that this can't accidentally cause the head to be moved backwards
579  */
580 static int iflib_min_tx_latency = 0;
581 SYSCTL_INT(_net_iflib, OID_AUTO, min_tx_latency, CTLFLAG_RW,
582 		   &iflib_min_tx_latency, 0, "minimize transmit latency at the possible expense of throughput");
583 static int iflib_no_tx_batch = 0;
584 SYSCTL_INT(_net_iflib, OID_AUTO, no_tx_batch, CTLFLAG_RW,
585 		   &iflib_no_tx_batch, 0, "minimize transmit latency at the possible expense of throughput");
586 
587 
588 #if IFLIB_DEBUG_COUNTERS
589 
590 static int iflib_tx_seen;
591 static int iflib_tx_sent;
592 static int iflib_tx_encap;
593 static int iflib_rx_allocs;
594 static int iflib_fl_refills;
595 static int iflib_fl_refills_large;
596 static int iflib_tx_frees;
597 
598 SYSCTL_INT(_net_iflib, OID_AUTO, tx_seen, CTLFLAG_RD,
599 		   &iflib_tx_seen, 0, "# TX mbufs seen");
600 SYSCTL_INT(_net_iflib, OID_AUTO, tx_sent, CTLFLAG_RD,
601 		   &iflib_tx_sent, 0, "# TX mbufs sent");
602 SYSCTL_INT(_net_iflib, OID_AUTO, tx_encap, CTLFLAG_RD,
603 		   &iflib_tx_encap, 0, "# TX mbufs encapped");
604 SYSCTL_INT(_net_iflib, OID_AUTO, tx_frees, CTLFLAG_RD,
605 		   &iflib_tx_frees, 0, "# TX frees");
606 SYSCTL_INT(_net_iflib, OID_AUTO, rx_allocs, CTLFLAG_RD,
607 		   &iflib_rx_allocs, 0, "# RX allocations");
608 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills, CTLFLAG_RD,
609 		   &iflib_fl_refills, 0, "# refills");
610 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills_large, CTLFLAG_RD,
611 		   &iflib_fl_refills_large, 0, "# large refills");
612 
613 
614 static int iflib_txq_drain_flushing;
615 static int iflib_txq_drain_oactive;
616 static int iflib_txq_drain_notready;
617 
618 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_flushing, CTLFLAG_RD,
619 		   &iflib_txq_drain_flushing, 0, "# drain flushes");
620 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_oactive, CTLFLAG_RD,
621 		   &iflib_txq_drain_oactive, 0, "# drain oactives");
622 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_notready, CTLFLAG_RD,
623 		   &iflib_txq_drain_notready, 0, "# drain notready");
624 
625 
626 static int iflib_encap_load_mbuf_fail;
627 static int iflib_encap_pad_mbuf_fail;
628 static int iflib_encap_txq_avail_fail;
629 static int iflib_encap_txd_encap_fail;
630 
631 SYSCTL_INT(_net_iflib, OID_AUTO, encap_load_mbuf_fail, CTLFLAG_RD,
632 		   &iflib_encap_load_mbuf_fail, 0, "# busdma load failures");
633 SYSCTL_INT(_net_iflib, OID_AUTO, encap_pad_mbuf_fail, CTLFLAG_RD,
634 		   &iflib_encap_pad_mbuf_fail, 0, "# runt frame pad failures");
635 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txq_avail_fail, CTLFLAG_RD,
636 		   &iflib_encap_txq_avail_fail, 0, "# txq avail failures");
637 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txd_encap_fail, CTLFLAG_RD,
638 		   &iflib_encap_txd_encap_fail, 0, "# driver encap failures");
639 
640 static int iflib_task_fn_rxs;
641 static int iflib_rx_intr_enables;
642 static int iflib_fast_intrs;
643 static int iflib_rx_unavail;
644 static int iflib_rx_ctx_inactive;
645 static int iflib_rx_if_input;
646 static int iflib_rxd_flush;
647 
648 static int iflib_verbose_debug;
649 
650 SYSCTL_INT(_net_iflib, OID_AUTO, task_fn_rx, CTLFLAG_RD,
651 		   &iflib_task_fn_rxs, 0, "# task_fn_rx calls");
652 SYSCTL_INT(_net_iflib, OID_AUTO, rx_intr_enables, CTLFLAG_RD,
653 		   &iflib_rx_intr_enables, 0, "# RX intr enables");
654 SYSCTL_INT(_net_iflib, OID_AUTO, fast_intrs, CTLFLAG_RD,
655 		   &iflib_fast_intrs, 0, "# fast_intr calls");
656 SYSCTL_INT(_net_iflib, OID_AUTO, rx_unavail, CTLFLAG_RD,
657 		   &iflib_rx_unavail, 0, "# times rxeof called with no available data");
658 SYSCTL_INT(_net_iflib, OID_AUTO, rx_ctx_inactive, CTLFLAG_RD,
659 		   &iflib_rx_ctx_inactive, 0, "# times rxeof called with inactive context");
660 SYSCTL_INT(_net_iflib, OID_AUTO, rx_if_input, CTLFLAG_RD,
661 		   &iflib_rx_if_input, 0, "# times rxeof called if_input");
662 SYSCTL_INT(_net_iflib, OID_AUTO, rxd_flush, CTLFLAG_RD,
663 	         &iflib_rxd_flush, 0, "# times rxd_flush called");
664 SYSCTL_INT(_net_iflib, OID_AUTO, verbose_debug, CTLFLAG_RW,
665 		   &iflib_verbose_debug, 0, "enable verbose debugging");
666 
667 #define DBG_COUNTER_INC(name) atomic_add_int(&(iflib_ ## name), 1)
668 static void
669 iflib_debug_reset(void)
670 {
671 	iflib_tx_seen = iflib_tx_sent = iflib_tx_encap = iflib_rx_allocs =
672 		iflib_fl_refills = iflib_fl_refills_large = iflib_tx_frees =
673 		iflib_txq_drain_flushing = iflib_txq_drain_oactive =
674 		iflib_txq_drain_notready =
675 		iflib_encap_load_mbuf_fail = iflib_encap_pad_mbuf_fail =
676 		iflib_encap_txq_avail_fail = iflib_encap_txd_encap_fail =
677 		iflib_task_fn_rxs = iflib_rx_intr_enables = iflib_fast_intrs =
678 		iflib_rx_unavail =
679 		iflib_rx_ctx_inactive = iflib_rx_if_input =
680 		iflib_rxd_flush = 0;
681 }
682 
683 #else
684 #define DBG_COUNTER_INC(name)
685 static void iflib_debug_reset(void) {}
686 #endif
687 
688 #define IFLIB_DEBUG 0
689 
690 static void iflib_tx_structures_free(if_ctx_t ctx);
691 static void iflib_rx_structures_free(if_ctx_t ctx);
692 static int iflib_queues_alloc(if_ctx_t ctx);
693 static int iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq);
694 static int iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget);
695 static int iflib_qset_structures_setup(if_ctx_t ctx);
696 static int iflib_msix_init(if_ctx_t ctx);
697 static int iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filterarg, int *rid, const char *str);
698 static void iflib_txq_check_drain(iflib_txq_t txq, int budget);
699 static uint32_t iflib_txq_can_drain(struct ifmp_ring *);
700 #ifdef ALTQ
701 static void iflib_altq_if_start(if_t ifp);
702 static int iflib_altq_if_transmit(if_t ifp, struct mbuf *m);
703 #endif
704 static int iflib_register(if_ctx_t);
705 static void iflib_deregister(if_ctx_t);
706 static void iflib_unregister_vlan_handlers(if_ctx_t ctx);
707 static void iflib_init_locked(if_ctx_t ctx);
708 static void iflib_add_device_sysctl_pre(if_ctx_t ctx);
709 static void iflib_add_device_sysctl_post(if_ctx_t ctx);
710 static void iflib_ifmp_purge(iflib_txq_t txq);
711 static void _iflib_pre_assert(if_softc_ctx_t scctx);
712 static void iflib_if_init_locked(if_ctx_t ctx);
713 static void iflib_free_intr_mem(if_ctx_t ctx);
714 #ifndef __NO_STRICT_ALIGNMENT
715 static struct mbuf * iflib_fixup_rx(struct mbuf *m);
716 #endif
717 
718 static SLIST_HEAD(cpu_offset_list, cpu_offset) cpu_offsets =
719     SLIST_HEAD_INITIALIZER(cpu_offsets);
720 struct cpu_offset {
721 	SLIST_ENTRY(cpu_offset) entries;
722 	cpuset_t	set;
723 	unsigned int	refcount;
724 	uint16_t	offset;
725 };
726 static struct mtx cpu_offset_mtx;
727 MTX_SYSINIT(iflib_cpu_offset, &cpu_offset_mtx, "iflib_cpu_offset lock",
728     MTX_DEF);
729 
730 DEBUGNET_DEFINE(iflib);
731 
732 #ifdef DEV_NETMAP
733 #include <sys/selinfo.h>
734 #include <net/netmap.h>
735 #include <dev/netmap/netmap_kern.h>
736 
737 MODULE_DEPEND(iflib, netmap, 1, 1, 1);
738 
739 static int netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, uint32_t nm_i, bool init);
740 
741 /*
742  * device-specific sysctl variables:
743  *
744  * iflib_crcstrip: 0: keep CRC in rx frames (default), 1: strip it.
745  *	During regular operations the CRC is stripped, but on some
746  *	hardware reception of frames not multiple of 64 is slower,
747  *	so using crcstrip=0 helps in benchmarks.
748  *
749  * iflib_rx_miss, iflib_rx_miss_bufs:
750  *	count packets that might be missed due to lost interrupts.
751  */
752 SYSCTL_DECL(_dev_netmap);
753 /*
754  * The xl driver by default strips CRCs and we do not override it.
755  */
756 
757 int iflib_crcstrip = 1;
758 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_crcstrip,
759     CTLFLAG_RW, &iflib_crcstrip, 1, "strip CRC on RX frames");
760 
761 int iflib_rx_miss, iflib_rx_miss_bufs;
762 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss,
763     CTLFLAG_RW, &iflib_rx_miss, 0, "potentially missed RX intr");
764 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss_bufs,
765     CTLFLAG_RW, &iflib_rx_miss_bufs, 0, "potentially missed RX intr bufs");
766 
767 /*
768  * Register/unregister. We are already under netmap lock.
769  * Only called on the first register or the last unregister.
770  */
771 static int
772 iflib_netmap_register(struct netmap_adapter *na, int onoff)
773 {
774 	if_t ifp = na->ifp;
775 	if_ctx_t ctx = ifp->if_softc;
776 	int status;
777 
778 	CTX_LOCK(ctx);
779 	IFDI_INTR_DISABLE(ctx);
780 
781 	/* Tell the stack that the interface is no longer active */
782 	ifp->if_drv_flags &= ~(IFF_DRV_RUNNING | IFF_DRV_OACTIVE);
783 
784 	if (!CTX_IS_VF(ctx))
785 		IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip);
786 
787 	/* enable or disable flags and callbacks in na and ifp */
788 	if (onoff) {
789 		nm_set_native_flags(na);
790 	} else {
791 		nm_clear_native_flags(na);
792 	}
793 	iflib_stop(ctx);
794 	iflib_init_locked(ctx);
795 	IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip); // XXX why twice ?
796 	status = ifp->if_drv_flags & IFF_DRV_RUNNING ? 0 : 1;
797 	if (status)
798 		nm_clear_native_flags(na);
799 	CTX_UNLOCK(ctx);
800 	return (status);
801 }
802 
803 static int
804 netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, uint32_t nm_i, bool init)
805 {
806 	struct netmap_adapter *na = kring->na;
807 	u_int const lim = kring->nkr_num_slots - 1;
808 	u_int head = kring->rhead;
809 	struct netmap_ring *ring = kring->ring;
810 	bus_dmamap_t *map;
811 	struct if_rxd_update iru;
812 	if_ctx_t ctx = rxq->ifr_ctx;
813 	iflib_fl_t fl = &rxq->ifr_fl[0];
814 	uint32_t refill_pidx, nic_i;
815 #if IFLIB_DEBUG_COUNTERS
816 	int rf_count = 0;
817 #endif
818 
819 	if (nm_i == head && __predict_true(!init))
820 		return 0;
821 	iru_init(&iru, rxq, 0 /* flid */);
822 	map = fl->ifl_sds.ifsd_map;
823 	refill_pidx = netmap_idx_k2n(kring, nm_i);
824 	/*
825 	 * IMPORTANT: we must leave one free slot in the ring,
826 	 * so move head back by one unit
827 	 */
828 	head = nm_prev(head, lim);
829 	nic_i = UINT_MAX;
830 	DBG_COUNTER_INC(fl_refills);
831 	while (nm_i != head) {
832 #if IFLIB_DEBUG_COUNTERS
833 		if (++rf_count == 9)
834 			DBG_COUNTER_INC(fl_refills_large);
835 #endif
836 		for (int tmp_pidx = 0; tmp_pidx < IFLIB_MAX_RX_REFRESH && nm_i != head; tmp_pidx++) {
837 			struct netmap_slot *slot = &ring->slot[nm_i];
838 			void *addr = PNMB(na, slot, &fl->ifl_bus_addrs[tmp_pidx]);
839 			uint32_t nic_i_dma = refill_pidx;
840 			nic_i = netmap_idx_k2n(kring, nm_i);
841 
842 			MPASS(tmp_pidx < IFLIB_MAX_RX_REFRESH);
843 
844 			if (addr == NETMAP_BUF_BASE(na)) /* bad buf */
845 			        return netmap_ring_reinit(kring);
846 
847 			fl->ifl_vm_addrs[tmp_pidx] = addr;
848 			if (__predict_false(init)) {
849 				netmap_load_map(na, fl->ifl_buf_tag,
850 				    map[nic_i], addr);
851 			} else if (slot->flags & NS_BUF_CHANGED) {
852 				/* buffer has changed, reload map */
853 				netmap_reload_map(na, fl->ifl_buf_tag,
854 				    map[nic_i], addr);
855 			}
856 			slot->flags &= ~NS_BUF_CHANGED;
857 
858 			nm_i = nm_next(nm_i, lim);
859 			fl->ifl_rxd_idxs[tmp_pidx] = nic_i = nm_next(nic_i, lim);
860 			if (nm_i != head && tmp_pidx < IFLIB_MAX_RX_REFRESH-1)
861 				continue;
862 
863 			iru.iru_pidx = refill_pidx;
864 			iru.iru_count = tmp_pidx+1;
865 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
866 			refill_pidx = nic_i;
867 			for (int n = 0; n < iru.iru_count; n++) {
868 				bus_dmamap_sync(fl->ifl_buf_tag, map[nic_i_dma],
869 						BUS_DMASYNC_PREREAD);
870 				/* XXX - change this to not use the netmap func*/
871 				nic_i_dma = nm_next(nic_i_dma, lim);
872 			}
873 		}
874 	}
875 	kring->nr_hwcur = head;
876 
877 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
878 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
879 	if (__predict_true(nic_i != UINT_MAX)) {
880 		ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id, nic_i);
881 		DBG_COUNTER_INC(rxd_flush);
882 	}
883 	return (0);
884 }
885 
886 /*
887  * Reconcile kernel and user view of the transmit ring.
888  *
889  * All information is in the kring.
890  * Userspace wants to send packets up to the one before kring->rhead,
891  * kernel knows kring->nr_hwcur is the first unsent packet.
892  *
893  * Here we push packets out (as many as possible), and possibly
894  * reclaim buffers from previously completed transmission.
895  *
896  * The caller (netmap) guarantees that there is only one instance
897  * running at any time. Any interference with other driver
898  * methods should be handled by the individual drivers.
899  */
900 static int
901 iflib_netmap_txsync(struct netmap_kring *kring, int flags)
902 {
903 	struct netmap_adapter *na = kring->na;
904 	if_t ifp = na->ifp;
905 	struct netmap_ring *ring = kring->ring;
906 	u_int nm_i;	/* index into the netmap kring */
907 	u_int nic_i;	/* index into the NIC ring */
908 	u_int n;
909 	u_int const lim = kring->nkr_num_slots - 1;
910 	u_int const head = kring->rhead;
911 	struct if_pkt_info pi;
912 
913 	/*
914 	 * interrupts on every tx packet are expensive so request
915 	 * them every half ring, or where NS_REPORT is set
916 	 */
917 	u_int report_frequency = kring->nkr_num_slots >> 1;
918 	/* device-specific */
919 	if_ctx_t ctx = ifp->if_softc;
920 	iflib_txq_t txq = &ctx->ifc_txqs[kring->ring_id];
921 
922 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
923 	    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
924 
925 	/*
926 	 * First part: process new packets to send.
927 	 * nm_i is the current index in the netmap kring,
928 	 * nic_i is the corresponding index in the NIC ring.
929 	 *
930 	 * If we have packets to send (nm_i != head)
931 	 * iterate over the netmap ring, fetch length and update
932 	 * the corresponding slot in the NIC ring. Some drivers also
933 	 * need to update the buffer's physical address in the NIC slot
934 	 * even NS_BUF_CHANGED is not set (PNMB computes the addresses).
935 	 *
936 	 * The netmap_reload_map() calls is especially expensive,
937 	 * even when (as in this case) the tag is 0, so do only
938 	 * when the buffer has actually changed.
939 	 *
940 	 * If possible do not set the report/intr bit on all slots,
941 	 * but only a few times per ring or when NS_REPORT is set.
942 	 *
943 	 * Finally, on 10G and faster drivers, it might be useful
944 	 * to prefetch the next slot and txr entry.
945 	 */
946 
947 	nm_i = kring->nr_hwcur;
948 	if (nm_i != head) {	/* we have new packets to send */
949 		pkt_info_zero(&pi);
950 		pi.ipi_segs = txq->ift_segs;
951 		pi.ipi_qsidx = kring->ring_id;
952 		nic_i = netmap_idx_k2n(kring, nm_i);
953 
954 		__builtin_prefetch(&ring->slot[nm_i]);
955 		__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i]);
956 		__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i]);
957 
958 		for (n = 0; nm_i != head; n++) {
959 			struct netmap_slot *slot = &ring->slot[nm_i];
960 			u_int len = slot->len;
961 			uint64_t paddr;
962 			void *addr = PNMB(na, slot, &paddr);
963 			int flags = (slot->flags & NS_REPORT ||
964 				nic_i == 0 || nic_i == report_frequency) ?
965 				IPI_TX_INTR : 0;
966 
967 			/* device-specific */
968 			pi.ipi_len = len;
969 			pi.ipi_segs[0].ds_addr = paddr;
970 			pi.ipi_segs[0].ds_len = len;
971 			pi.ipi_nsegs = 1;
972 			pi.ipi_ndescs = 0;
973 			pi.ipi_pidx = nic_i;
974 			pi.ipi_flags = flags;
975 
976 			/* Fill the slot in the NIC ring. */
977 			ctx->isc_txd_encap(ctx->ifc_softc, &pi);
978 			DBG_COUNTER_INC(tx_encap);
979 
980 			/* prefetch for next round */
981 			__builtin_prefetch(&ring->slot[nm_i + 1]);
982 			__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i + 1]);
983 			__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i + 1]);
984 
985 			NM_CHECK_ADDR_LEN(na, addr, len);
986 
987 			if (slot->flags & NS_BUF_CHANGED) {
988 				/* buffer has changed, reload map */
989 				netmap_reload_map(na, txq->ift_buf_tag,
990 				    txq->ift_sds.ifsd_map[nic_i], addr);
991 			}
992 			/* make sure changes to the buffer are synced */
993 			bus_dmamap_sync(txq->ift_buf_tag,
994 			    txq->ift_sds.ifsd_map[nic_i],
995 			    BUS_DMASYNC_PREWRITE);
996 
997 			slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED);
998 			nm_i = nm_next(nm_i, lim);
999 			nic_i = nm_next(nic_i, lim);
1000 		}
1001 		kring->nr_hwcur = nm_i;
1002 
1003 		/* synchronize the NIC ring */
1004 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1005 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1006 
1007 		/* (re)start the tx unit up to slot nic_i (excluded) */
1008 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, nic_i);
1009 	}
1010 
1011 	/*
1012 	 * Second part: reclaim buffers for completed transmissions.
1013 	 *
1014 	 * If there are unclaimed buffers, attempt to reclaim them.
1015 	 * If none are reclaimed, and TX IRQs are not in use, do an initial
1016 	 * minimal delay, then trigger the tx handler which will spin in the
1017 	 * group task queue.
1018 	 */
1019 	if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1020 		if (iflib_tx_credits_update(ctx, txq)) {
1021 			/* some tx completed, increment avail */
1022 			nic_i = txq->ift_cidx_processed;
1023 			kring->nr_hwtail = nm_prev(netmap_idx_n2k(kring, nic_i), lim);
1024 		}
1025 	}
1026 	if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ))
1027 		if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1028 			callout_reset_on(&txq->ift_timer, hz < 2000 ? 1 : hz / 1000,
1029 			    iflib_timer, txq, txq->ift_timer.c_cpu);
1030 	}
1031 	return (0);
1032 }
1033 
1034 /*
1035  * Reconcile kernel and user view of the receive ring.
1036  * Same as for the txsync, this routine must be efficient.
1037  * The caller guarantees a single invocations, but races against
1038  * the rest of the driver should be handled here.
1039  *
1040  * On call, kring->rhead is the first packet that userspace wants
1041  * to keep, and kring->rcur is the wakeup point.
1042  * The kernel has previously reported packets up to kring->rtail.
1043  *
1044  * If (flags & NAF_FORCE_READ) also check for incoming packets irrespective
1045  * of whether or not we received an interrupt.
1046  */
1047 static int
1048 iflib_netmap_rxsync(struct netmap_kring *kring, int flags)
1049 {
1050 	struct netmap_adapter *na = kring->na;
1051 	struct netmap_ring *ring = kring->ring;
1052 	if_t ifp = na->ifp;
1053 	iflib_fl_t fl;
1054 	uint32_t nm_i;	/* index into the netmap ring */
1055 	uint32_t nic_i;	/* index into the NIC ring */
1056 	u_int i, n;
1057 	u_int const lim = kring->nkr_num_slots - 1;
1058 	u_int const head = kring->rhead;
1059 	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
1060 	struct if_rxd_info ri;
1061 
1062 	if_ctx_t ctx = ifp->if_softc;
1063 	iflib_rxq_t rxq = &ctx->ifc_rxqs[kring->ring_id];
1064 	if (head > lim)
1065 		return netmap_ring_reinit(kring);
1066 
1067 	/*
1068 	 * XXX netmap_fl_refill() only ever (re)fills free list 0 so far.
1069 	 */
1070 
1071 	for (i = 0, fl = rxq->ifr_fl; i < rxq->ifr_nfl; i++, fl++) {
1072 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
1073 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1074 	}
1075 
1076 	/*
1077 	 * First part: import newly received packets.
1078 	 *
1079 	 * nm_i is the index of the next free slot in the netmap ring,
1080 	 * nic_i is the index of the next received packet in the NIC ring,
1081 	 * and they may differ in case if_init() has been called while
1082 	 * in netmap mode. For the receive ring we have
1083 	 *
1084 	 *	nic_i = rxr->next_check;
1085 	 *	nm_i = kring->nr_hwtail (previous)
1086 	 * and
1087 	 *	nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1088 	 *
1089 	 * rxr->next_check is set to 0 on a ring reinit
1090 	 */
1091 	if (netmap_no_pendintr || force_update) {
1092 		int crclen = iflib_crcstrip ? 0 : 4;
1093 		int error, avail;
1094 
1095 		for (i = 0; i < rxq->ifr_nfl; i++) {
1096 			fl = &rxq->ifr_fl[i];
1097 			nic_i = fl->ifl_cidx;
1098 			nm_i = netmap_idx_n2k(kring, nic_i);
1099 			avail = ctx->isc_rxd_available(ctx->ifc_softc,
1100 			    rxq->ifr_id, nic_i, USHRT_MAX);
1101 			for (n = 0; avail > 0; n++, avail--) {
1102 				rxd_info_zero(&ri);
1103 				ri.iri_frags = rxq->ifr_frags;
1104 				ri.iri_qsidx = kring->ring_id;
1105 				ri.iri_ifp = ctx->ifc_ifp;
1106 				ri.iri_cidx = nic_i;
1107 
1108 				error = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
1109 				ring->slot[nm_i].len = error ? 0 : ri.iri_len - crclen;
1110 				ring->slot[nm_i].flags = 0;
1111 				bus_dmamap_sync(fl->ifl_buf_tag,
1112 				    fl->ifl_sds.ifsd_map[nic_i], BUS_DMASYNC_POSTREAD);
1113 				nm_i = nm_next(nm_i, lim);
1114 				nic_i = nm_next(nic_i, lim);
1115 			}
1116 			if (n) { /* update the state variables */
1117 				if (netmap_no_pendintr && !force_update) {
1118 					/* diagnostics */
1119 					iflib_rx_miss ++;
1120 					iflib_rx_miss_bufs += n;
1121 				}
1122 				fl->ifl_cidx = nic_i;
1123 				kring->nr_hwtail = nm_i;
1124 			}
1125 			kring->nr_kflags &= ~NKR_PENDINTR;
1126 		}
1127 	}
1128 	/*
1129 	 * Second part: skip past packets that userspace has released.
1130 	 * (kring->nr_hwcur to head excluded),
1131 	 * and make the buffers available for reception.
1132 	 * As usual nm_i is the index in the netmap ring,
1133 	 * nic_i is the index in the NIC ring, and
1134 	 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1135 	 */
1136 	/* XXX not sure how this will work with multiple free lists */
1137 	nm_i = kring->nr_hwcur;
1138 
1139 	return (netmap_fl_refill(rxq, kring, nm_i, false));
1140 }
1141 
1142 static void
1143 iflib_netmap_intr(struct netmap_adapter *na, int onoff)
1144 {
1145 	if_ctx_t ctx = na->ifp->if_softc;
1146 
1147 	CTX_LOCK(ctx);
1148 	if (onoff) {
1149 		IFDI_INTR_ENABLE(ctx);
1150 	} else {
1151 		IFDI_INTR_DISABLE(ctx);
1152 	}
1153 	CTX_UNLOCK(ctx);
1154 }
1155 
1156 
1157 static int
1158 iflib_netmap_attach(if_ctx_t ctx)
1159 {
1160 	struct netmap_adapter na;
1161 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1162 
1163 	bzero(&na, sizeof(na));
1164 
1165 	na.ifp = ctx->ifc_ifp;
1166 	na.na_flags = NAF_BDG_MAYSLEEP;
1167 	MPASS(ctx->ifc_softc_ctx.isc_ntxqsets);
1168 	MPASS(ctx->ifc_softc_ctx.isc_nrxqsets);
1169 
1170 	na.num_tx_desc = scctx->isc_ntxd[0];
1171 	na.num_rx_desc = scctx->isc_nrxd[0];
1172 	na.nm_txsync = iflib_netmap_txsync;
1173 	na.nm_rxsync = iflib_netmap_rxsync;
1174 	na.nm_register = iflib_netmap_register;
1175 	na.nm_intr = iflib_netmap_intr;
1176 	na.num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
1177 	na.num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
1178 	return (netmap_attach(&na));
1179 }
1180 
1181 static void
1182 iflib_netmap_txq_init(if_ctx_t ctx, iflib_txq_t txq)
1183 {
1184 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1185 	struct netmap_slot *slot;
1186 
1187 	slot = netmap_reset(na, NR_TX, txq->ift_id, 0);
1188 	if (slot == NULL)
1189 		return;
1190 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxd[0]; i++) {
1191 
1192 		/*
1193 		 * In netmap mode, set the map for the packet buffer.
1194 		 * NOTE: Some drivers (not this one) also need to set
1195 		 * the physical buffer address in the NIC ring.
1196 		 * netmap_idx_n2k() maps a nic index, i, into the corresponding
1197 		 * netmap slot index, si
1198 		 */
1199 		int si = netmap_idx_n2k(na->tx_rings[txq->ift_id], i);
1200 		netmap_load_map(na, txq->ift_buf_tag, txq->ift_sds.ifsd_map[i],
1201 		    NMB(na, slot + si));
1202 	}
1203 }
1204 
1205 static void
1206 iflib_netmap_rxq_init(if_ctx_t ctx, iflib_rxq_t rxq)
1207 {
1208 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1209 	struct netmap_kring *kring = na->rx_rings[rxq->ifr_id];
1210 	struct netmap_slot *slot;
1211 	uint32_t nm_i;
1212 
1213 	slot = netmap_reset(na, NR_RX, rxq->ifr_id, 0);
1214 	if (slot == NULL)
1215 		return;
1216 	nm_i = netmap_idx_n2k(kring, 0);
1217 	netmap_fl_refill(rxq, kring, nm_i, true);
1218 }
1219 
1220 static void
1221 iflib_netmap_timer_adjust(if_ctx_t ctx, iflib_txq_t txq, uint32_t *reset_on)
1222 {
1223 	struct netmap_kring *kring;
1224 	uint16_t txqid;
1225 
1226 	txqid = txq->ift_id;
1227 	kring = NA(ctx->ifc_ifp)->tx_rings[txqid];
1228 
1229 	if (kring->nr_hwcur != nm_next(kring->nr_hwtail, kring->nkr_num_slots - 1)) {
1230 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1231 		    BUS_DMASYNC_POSTREAD);
1232 		if (ctx->isc_txd_credits_update(ctx->ifc_softc, txqid, false))
1233 			netmap_tx_irq(ctx->ifc_ifp, txqid);
1234 		if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ)) {
1235 			if (hz < 2000)
1236 				*reset_on = 1;
1237 			else
1238 				*reset_on = hz / 1000;
1239 		}
1240 	}
1241 }
1242 
1243 #define iflib_netmap_detach(ifp) netmap_detach(ifp)
1244 
1245 #else
1246 #define iflib_netmap_txq_init(ctx, txq)
1247 #define iflib_netmap_rxq_init(ctx, rxq)
1248 #define iflib_netmap_detach(ifp)
1249 
1250 #define iflib_netmap_attach(ctx) (0)
1251 #define netmap_rx_irq(ifp, qid, budget) (0)
1252 #define netmap_tx_irq(ifp, qid) do {} while (0)
1253 #define iflib_netmap_timer_adjust(ctx, txq, reset_on)
1254 #endif
1255 
1256 #if defined(__i386__) || defined(__amd64__)
1257 static __inline void
1258 prefetch(void *x)
1259 {
1260 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1261 }
1262 static __inline void
1263 prefetch2cachelines(void *x)
1264 {
1265 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1266 #if (CACHE_LINE_SIZE < 128)
1267 	__asm volatile("prefetcht0 %0" :: "m" (*(((unsigned long *)x)+CACHE_LINE_SIZE/(sizeof(unsigned long)))));
1268 #endif
1269 }
1270 #else
1271 #define prefetch(x)
1272 #define prefetch2cachelines(x)
1273 #endif
1274 
1275 static void
1276 iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid)
1277 {
1278 	iflib_fl_t fl;
1279 
1280 	fl = &rxq->ifr_fl[flid];
1281 	iru->iru_paddrs = fl->ifl_bus_addrs;
1282 	iru->iru_vaddrs = &fl->ifl_vm_addrs[0];
1283 	iru->iru_idxs = fl->ifl_rxd_idxs;
1284 	iru->iru_qsidx = rxq->ifr_id;
1285 	iru->iru_buf_size = fl->ifl_buf_size;
1286 	iru->iru_flidx = fl->ifl_id;
1287 }
1288 
1289 static void
1290 _iflib_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int err)
1291 {
1292 	if (err)
1293 		return;
1294 	*(bus_addr_t *) arg = segs[0].ds_addr;
1295 }
1296 
1297 int
1298 iflib_dma_alloc_align(if_ctx_t ctx, int size, int align, iflib_dma_info_t dma, int mapflags)
1299 {
1300 	int err;
1301 	device_t dev = ctx->ifc_dev;
1302 
1303 	err = bus_dma_tag_create(bus_get_dma_tag(dev),	/* parent */
1304 				align, 0,		/* alignment, bounds */
1305 				BUS_SPACE_MAXADDR,	/* lowaddr */
1306 				BUS_SPACE_MAXADDR,	/* highaddr */
1307 				NULL, NULL,		/* filter, filterarg */
1308 				size,			/* maxsize */
1309 				1,			/* nsegments */
1310 				size,			/* maxsegsize */
1311 				BUS_DMA_ALLOCNOW,	/* flags */
1312 				NULL,			/* lockfunc */
1313 				NULL,			/* lockarg */
1314 				&dma->idi_tag);
1315 	if (err) {
1316 		device_printf(dev,
1317 		    "%s: bus_dma_tag_create failed: %d\n",
1318 		    __func__, err);
1319 		goto fail_0;
1320 	}
1321 
1322 	err = bus_dmamem_alloc(dma->idi_tag, (void**) &dma->idi_vaddr,
1323 	    BUS_DMA_NOWAIT | BUS_DMA_COHERENT | BUS_DMA_ZERO, &dma->idi_map);
1324 	if (err) {
1325 		device_printf(dev,
1326 		    "%s: bus_dmamem_alloc(%ju) failed: %d\n",
1327 		    __func__, (uintmax_t)size, err);
1328 		goto fail_1;
1329 	}
1330 
1331 	dma->idi_paddr = IF_BAD_DMA;
1332 	err = bus_dmamap_load(dma->idi_tag, dma->idi_map, dma->idi_vaddr,
1333 	    size, _iflib_dmamap_cb, &dma->idi_paddr, mapflags | BUS_DMA_NOWAIT);
1334 	if (err || dma->idi_paddr == IF_BAD_DMA) {
1335 		device_printf(dev,
1336 		    "%s: bus_dmamap_load failed: %d\n",
1337 		    __func__, err);
1338 		goto fail_2;
1339 	}
1340 
1341 	dma->idi_size = size;
1342 	return (0);
1343 
1344 fail_2:
1345 	bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1346 fail_1:
1347 	bus_dma_tag_destroy(dma->idi_tag);
1348 fail_0:
1349 	dma->idi_tag = NULL;
1350 
1351 	return (err);
1352 }
1353 
1354 int
1355 iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags)
1356 {
1357 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1358 
1359 	KASSERT(sctx->isc_q_align != 0, ("alignment value not initialized"));
1360 
1361 	return (iflib_dma_alloc_align(ctx, size, sctx->isc_q_align, dma, mapflags));
1362 }
1363 
1364 int
1365 iflib_dma_alloc_multi(if_ctx_t ctx, int *sizes, iflib_dma_info_t *dmalist, int mapflags, int count)
1366 {
1367 	int i, err;
1368 	iflib_dma_info_t *dmaiter;
1369 
1370 	dmaiter = dmalist;
1371 	for (i = 0; i < count; i++, dmaiter++) {
1372 		if ((err = iflib_dma_alloc(ctx, sizes[i], *dmaiter, mapflags)) != 0)
1373 			break;
1374 	}
1375 	if (err)
1376 		iflib_dma_free_multi(dmalist, i);
1377 	return (err);
1378 }
1379 
1380 void
1381 iflib_dma_free(iflib_dma_info_t dma)
1382 {
1383 	if (dma->idi_tag == NULL)
1384 		return;
1385 	if (dma->idi_paddr != IF_BAD_DMA) {
1386 		bus_dmamap_sync(dma->idi_tag, dma->idi_map,
1387 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1388 		bus_dmamap_unload(dma->idi_tag, dma->idi_map);
1389 		dma->idi_paddr = IF_BAD_DMA;
1390 	}
1391 	if (dma->idi_vaddr != NULL) {
1392 		bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1393 		dma->idi_vaddr = NULL;
1394 	}
1395 	bus_dma_tag_destroy(dma->idi_tag);
1396 	dma->idi_tag = NULL;
1397 }
1398 
1399 void
1400 iflib_dma_free_multi(iflib_dma_info_t *dmalist, int count)
1401 {
1402 	int i;
1403 	iflib_dma_info_t *dmaiter = dmalist;
1404 
1405 	for (i = 0; i < count; i++, dmaiter++)
1406 		iflib_dma_free(*dmaiter);
1407 }
1408 
1409 #ifdef EARLY_AP_STARTUP
1410 static const int iflib_started = 1;
1411 #else
1412 /*
1413  * We used to abuse the smp_started flag to decide if the queues have been
1414  * fully initialized (by late taskqgroup_adjust() calls in a SYSINIT()).
1415  * That gave bad races, since the SYSINIT() runs strictly after smp_started
1416  * is set.  Run a SYSINIT() strictly after that to just set a usable
1417  * completion flag.
1418  */
1419 
1420 static int iflib_started;
1421 
1422 static void
1423 iflib_record_started(void *arg)
1424 {
1425 	iflib_started = 1;
1426 }
1427 
1428 SYSINIT(iflib_record_started, SI_SUB_SMP + 1, SI_ORDER_FIRST,
1429 	iflib_record_started, NULL);
1430 #endif
1431 
1432 static int
1433 iflib_fast_intr(void *arg)
1434 {
1435 	iflib_filter_info_t info = arg;
1436 	struct grouptask *gtask = info->ifi_task;
1437 	int result;
1438 
1439 	if (!iflib_started)
1440 		return (FILTER_STRAY);
1441 
1442 	DBG_COUNTER_INC(fast_intrs);
1443 	if (info->ifi_filter != NULL) {
1444 		result = info->ifi_filter(info->ifi_filter_arg);
1445 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1446 			return (result);
1447 	}
1448 
1449 	GROUPTASK_ENQUEUE(gtask);
1450 	return (FILTER_HANDLED);
1451 }
1452 
1453 static int
1454 iflib_fast_intr_rxtx(void *arg)
1455 {
1456 	iflib_filter_info_t info = arg;
1457 	struct grouptask *gtask = info->ifi_task;
1458 	if_ctx_t ctx;
1459 	iflib_rxq_t rxq = (iflib_rxq_t)info->ifi_ctx;
1460 	iflib_txq_t txq;
1461 	void *sc;
1462 	int i, cidx, result;
1463 	qidx_t txqid;
1464 	bool intr_enable, intr_legacy;
1465 
1466 	if (!iflib_started)
1467 		return (FILTER_STRAY);
1468 
1469 	DBG_COUNTER_INC(fast_intrs);
1470 	if (info->ifi_filter != NULL) {
1471 		result = info->ifi_filter(info->ifi_filter_arg);
1472 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1473 			return (result);
1474 	}
1475 
1476 	ctx = rxq->ifr_ctx;
1477 	sc = ctx->ifc_softc;
1478 	intr_enable = false;
1479 	intr_legacy = !!(ctx->ifc_flags & IFC_LEGACY);
1480 	MPASS(rxq->ifr_ntxqirq);
1481 	for (i = 0; i < rxq->ifr_ntxqirq; i++) {
1482 		txqid = rxq->ifr_txqid[i];
1483 		txq = &ctx->ifc_txqs[txqid];
1484 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1485 		    BUS_DMASYNC_POSTREAD);
1486 		if (!ctx->isc_txd_credits_update(sc, txqid, false)) {
1487 			if (intr_legacy)
1488 				intr_enable = true;
1489 			else
1490 				IFDI_TX_QUEUE_INTR_ENABLE(ctx, txqid);
1491 			continue;
1492 		}
1493 		GROUPTASK_ENQUEUE(&txq->ift_task);
1494 	}
1495 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_RXCQ)
1496 		cidx = rxq->ifr_cq_cidx;
1497 	else
1498 		cidx = rxq->ifr_fl[0].ifl_cidx;
1499 	if (iflib_rxd_avail(ctx, rxq, cidx, 1))
1500 		GROUPTASK_ENQUEUE(gtask);
1501 	else {
1502 		if (intr_legacy)
1503 			intr_enable = true;
1504 		else
1505 			IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
1506 		DBG_COUNTER_INC(rx_intr_enables);
1507 	}
1508 	if (intr_enable)
1509 		IFDI_INTR_ENABLE(ctx);
1510 	return (FILTER_HANDLED);
1511 }
1512 
1513 
1514 static int
1515 iflib_fast_intr_ctx(void *arg)
1516 {
1517 	iflib_filter_info_t info = arg;
1518 	struct grouptask *gtask = info->ifi_task;
1519 	int result;
1520 
1521 	if (!iflib_started)
1522 		return (FILTER_STRAY);
1523 
1524 	DBG_COUNTER_INC(fast_intrs);
1525 	if (info->ifi_filter != NULL) {
1526 		result = info->ifi_filter(info->ifi_filter_arg);
1527 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1528 			return (result);
1529 	}
1530 
1531 	GROUPTASK_ENQUEUE(gtask);
1532 	return (FILTER_HANDLED);
1533 }
1534 
1535 static int
1536 _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
1537 		 driver_filter_t filter, driver_intr_t handler, void *arg,
1538 		 const char *name)
1539 {
1540 	struct resource *res;
1541 	void *tag = NULL;
1542 	device_t dev = ctx->ifc_dev;
1543 	int flags, i, rc;
1544 
1545 	flags = RF_ACTIVE;
1546 	if (ctx->ifc_flags & IFC_LEGACY)
1547 		flags |= RF_SHAREABLE;
1548 	MPASS(rid < 512);
1549 	i = rid;
1550 	res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &i, flags);
1551 	if (res == NULL) {
1552 		device_printf(dev,
1553 		    "failed to allocate IRQ for rid %d, name %s.\n", rid, name);
1554 		return (ENOMEM);
1555 	}
1556 	irq->ii_res = res;
1557 	KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL"));
1558 	rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET,
1559 						filter, handler, arg, &tag);
1560 	if (rc != 0) {
1561 		device_printf(dev,
1562 		    "failed to setup interrupt for rid %d, name %s: %d\n",
1563 					  rid, name ? name : "unknown", rc);
1564 		return (rc);
1565 	} else if (name)
1566 		bus_describe_intr(dev, res, tag, "%s", name);
1567 
1568 	irq->ii_tag = tag;
1569 	return (0);
1570 }
1571 
1572 /*********************************************************************
1573  *
1574  *  Allocate DMA resources for TX buffers as well as memory for the TX
1575  *  mbuf map.  TX DMA maps (non-TSO/TSO) and TX mbuf map are kept in a
1576  *  iflib_sw_tx_desc_array structure, storing all the information that
1577  *  is needed to transmit a packet on the wire.  This is called only
1578  *  once at attach, setup is done every reset.
1579  *
1580  **********************************************************************/
1581 static int
1582 iflib_txsd_alloc(iflib_txq_t txq)
1583 {
1584 	if_ctx_t ctx = txq->ift_ctx;
1585 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1586 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1587 	device_t dev = ctx->ifc_dev;
1588 	bus_size_t tsomaxsize;
1589 	int err, nsegments, ntsosegments;
1590 	bool tso;
1591 
1592 	nsegments = scctx->isc_tx_nsegments;
1593 	ntsosegments = scctx->isc_tx_tso_segments_max;
1594 	tsomaxsize = scctx->isc_tx_tso_size_max;
1595 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_VLAN_MTU)
1596 		tsomaxsize += sizeof(struct ether_vlan_header);
1597 	MPASS(scctx->isc_ntxd[0] > 0);
1598 	MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0);
1599 	MPASS(nsegments > 0);
1600 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) {
1601 		MPASS(ntsosegments > 0);
1602 		MPASS(sctx->isc_tso_maxsize >= tsomaxsize);
1603 	}
1604 
1605 	/*
1606 	 * Set up DMA tags for TX buffers.
1607 	 */
1608 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1609 			       1, 0,			/* alignment, bounds */
1610 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1611 			       BUS_SPACE_MAXADDR,	/* highaddr */
1612 			       NULL, NULL,		/* filter, filterarg */
1613 			       sctx->isc_tx_maxsize,		/* maxsize */
1614 			       nsegments,	/* nsegments */
1615 			       sctx->isc_tx_maxsegsize,	/* maxsegsize */
1616 			       0,			/* flags */
1617 			       NULL,			/* lockfunc */
1618 			       NULL,			/* lockfuncarg */
1619 			       &txq->ift_buf_tag))) {
1620 		device_printf(dev,"Unable to allocate TX DMA tag: %d\n", err);
1621 		device_printf(dev,"maxsize: %ju nsegments: %d maxsegsize: %ju\n",
1622 		    (uintmax_t)sctx->isc_tx_maxsize, nsegments, (uintmax_t)sctx->isc_tx_maxsegsize);
1623 		goto fail;
1624 	}
1625 	tso = (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) != 0;
1626 	if (tso && (err = bus_dma_tag_create(bus_get_dma_tag(dev),
1627 			       1, 0,			/* alignment, bounds */
1628 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1629 			       BUS_SPACE_MAXADDR,	/* highaddr */
1630 			       NULL, NULL,		/* filter, filterarg */
1631 			       tsomaxsize,		/* maxsize */
1632 			       ntsosegments,	/* nsegments */
1633 			       sctx->isc_tso_maxsegsize,/* maxsegsize */
1634 			       0,			/* flags */
1635 			       NULL,			/* lockfunc */
1636 			       NULL,			/* lockfuncarg */
1637 			       &txq->ift_tso_buf_tag))) {
1638 		device_printf(dev, "Unable to allocate TSO TX DMA tag: %d\n",
1639 		    err);
1640 		goto fail;
1641 	}
1642 
1643 	/* Allocate memory for the TX mbuf map. */
1644 	if (!(txq->ift_sds.ifsd_m =
1645 	    (struct mbuf **) malloc(sizeof(struct mbuf *) *
1646 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1647 		device_printf(dev, "Unable to allocate TX mbuf map memory\n");
1648 		err = ENOMEM;
1649 		goto fail;
1650 	}
1651 
1652 	/*
1653 	 * Create the DMA maps for TX buffers.
1654 	 */
1655 	if ((txq->ift_sds.ifsd_map = (bus_dmamap_t *)malloc(
1656 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1657 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1658 		device_printf(dev,
1659 		    "Unable to allocate TX buffer DMA map memory\n");
1660 		err = ENOMEM;
1661 		goto fail;
1662 	}
1663 	if (tso && (txq->ift_sds.ifsd_tso_map = (bus_dmamap_t *)malloc(
1664 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1665 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1666 		device_printf(dev,
1667 		    "Unable to allocate TSO TX buffer map memory\n");
1668 		err = ENOMEM;
1669 		goto fail;
1670 	}
1671 	for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) {
1672 		err = bus_dmamap_create(txq->ift_buf_tag, 0,
1673 		    &txq->ift_sds.ifsd_map[i]);
1674 		if (err != 0) {
1675 			device_printf(dev, "Unable to create TX DMA map\n");
1676 			goto fail;
1677 		}
1678 		if (!tso)
1679 			continue;
1680 		err = bus_dmamap_create(txq->ift_tso_buf_tag, 0,
1681 		    &txq->ift_sds.ifsd_tso_map[i]);
1682 		if (err != 0) {
1683 			device_printf(dev, "Unable to create TSO TX DMA map\n");
1684 			goto fail;
1685 		}
1686 	}
1687 	return (0);
1688 fail:
1689 	/* We free all, it handles case where we are in the middle */
1690 	iflib_tx_structures_free(ctx);
1691 	return (err);
1692 }
1693 
1694 static void
1695 iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i)
1696 {
1697 	bus_dmamap_t map;
1698 
1699 	if (txq->ift_sds.ifsd_map != NULL) {
1700 		map = txq->ift_sds.ifsd_map[i];
1701 		bus_dmamap_sync(txq->ift_buf_tag, map, BUS_DMASYNC_POSTWRITE);
1702 		bus_dmamap_unload(txq->ift_buf_tag, map);
1703 		bus_dmamap_destroy(txq->ift_buf_tag, map);
1704 		txq->ift_sds.ifsd_map[i] = NULL;
1705 	}
1706 
1707 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1708 		map = txq->ift_sds.ifsd_tso_map[i];
1709 		bus_dmamap_sync(txq->ift_tso_buf_tag, map,
1710 		    BUS_DMASYNC_POSTWRITE);
1711 		bus_dmamap_unload(txq->ift_tso_buf_tag, map);
1712 		bus_dmamap_destroy(txq->ift_tso_buf_tag, map);
1713 		txq->ift_sds.ifsd_tso_map[i] = NULL;
1714 	}
1715 }
1716 
1717 static void
1718 iflib_txq_destroy(iflib_txq_t txq)
1719 {
1720 	if_ctx_t ctx = txq->ift_ctx;
1721 
1722 	for (int i = 0; i < txq->ift_size; i++)
1723 		iflib_txsd_destroy(ctx, txq, i);
1724 
1725 	if (txq->ift_br != NULL) {
1726 		ifmp_ring_free(txq->ift_br);
1727 		txq->ift_br = NULL;
1728 	}
1729 
1730 	mtx_destroy(&txq->ift_mtx);
1731 
1732 	if (txq->ift_sds.ifsd_map != NULL) {
1733 		free(txq->ift_sds.ifsd_map, M_IFLIB);
1734 		txq->ift_sds.ifsd_map = NULL;
1735 	}
1736 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1737 		free(txq->ift_sds.ifsd_tso_map, M_IFLIB);
1738 		txq->ift_sds.ifsd_tso_map = NULL;
1739 	}
1740 	if (txq->ift_sds.ifsd_m != NULL) {
1741 		free(txq->ift_sds.ifsd_m, M_IFLIB);
1742 		txq->ift_sds.ifsd_m = NULL;
1743 	}
1744 	if (txq->ift_buf_tag != NULL) {
1745 		bus_dma_tag_destroy(txq->ift_buf_tag);
1746 		txq->ift_buf_tag = NULL;
1747 	}
1748 	if (txq->ift_tso_buf_tag != NULL) {
1749 		bus_dma_tag_destroy(txq->ift_tso_buf_tag);
1750 		txq->ift_tso_buf_tag = NULL;
1751 	}
1752 	if (txq->ift_ifdi != NULL) {
1753 		free(txq->ift_ifdi, M_IFLIB);
1754 	}
1755 }
1756 
1757 static void
1758 iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i)
1759 {
1760 	struct mbuf **mp;
1761 
1762 	mp = &txq->ift_sds.ifsd_m[i];
1763 	if (*mp == NULL)
1764 		return;
1765 
1766 	if (txq->ift_sds.ifsd_map != NULL) {
1767 		bus_dmamap_sync(txq->ift_buf_tag,
1768 		    txq->ift_sds.ifsd_map[i], BUS_DMASYNC_POSTWRITE);
1769 		bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[i]);
1770 	}
1771 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1772 		bus_dmamap_sync(txq->ift_tso_buf_tag,
1773 		    txq->ift_sds.ifsd_tso_map[i], BUS_DMASYNC_POSTWRITE);
1774 		bus_dmamap_unload(txq->ift_tso_buf_tag,
1775 		    txq->ift_sds.ifsd_tso_map[i]);
1776 	}
1777 	m_free(*mp);
1778 	DBG_COUNTER_INC(tx_frees);
1779 	*mp = NULL;
1780 }
1781 
1782 static int
1783 iflib_txq_setup(iflib_txq_t txq)
1784 {
1785 	if_ctx_t ctx = txq->ift_ctx;
1786 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1787 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1788 	iflib_dma_info_t di;
1789 	int i;
1790 
1791 	/* Set number of descriptors available */
1792 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
1793 	/* XXX make configurable */
1794 	txq->ift_update_freq = IFLIB_DEFAULT_TX_UPDATE_FREQ;
1795 
1796 	/* Reset indices */
1797 	txq->ift_cidx_processed = 0;
1798 	txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0;
1799 	txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset];
1800 
1801 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1802 		bzero((void *)di->idi_vaddr, di->idi_size);
1803 
1804 	IFDI_TXQ_SETUP(ctx, txq->ift_id);
1805 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1806 		bus_dmamap_sync(di->idi_tag, di->idi_map,
1807 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1808 	return (0);
1809 }
1810 
1811 /*********************************************************************
1812  *
1813  *  Allocate DMA resources for RX buffers as well as memory for the RX
1814  *  mbuf map, direct RX cluster pointer map and RX cluster bus address
1815  *  map.  RX DMA map, RX mbuf map, direct RX cluster pointer map and
1816  *  RX cluster map are kept in a iflib_sw_rx_desc_array structure.
1817  *  Since we use use one entry in iflib_sw_rx_desc_array per received
1818  *  packet, the maximum number of entries we'll need is equal to the
1819  *  number of hardware receive descriptors that we've allocated.
1820  *
1821  **********************************************************************/
1822 static int
1823 iflib_rxsd_alloc(iflib_rxq_t rxq)
1824 {
1825 	if_ctx_t ctx = rxq->ifr_ctx;
1826 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1827 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1828 	device_t dev = ctx->ifc_dev;
1829 	iflib_fl_t fl;
1830 	int			err;
1831 
1832 	MPASS(scctx->isc_nrxd[0] > 0);
1833 	MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0);
1834 
1835 	fl = rxq->ifr_fl;
1836 	for (int i = 0; i <  rxq->ifr_nfl; i++, fl++) {
1837 		fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */
1838 		/* Set up DMA tag for RX buffers. */
1839 		err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1840 					 1, 0,			/* alignment, bounds */
1841 					 BUS_SPACE_MAXADDR,	/* lowaddr */
1842 					 BUS_SPACE_MAXADDR,	/* highaddr */
1843 					 NULL, NULL,		/* filter, filterarg */
1844 					 sctx->isc_rx_maxsize,	/* maxsize */
1845 					 sctx->isc_rx_nsegments,	/* nsegments */
1846 					 sctx->isc_rx_maxsegsize,	/* maxsegsize */
1847 					 0,			/* flags */
1848 					 NULL,			/* lockfunc */
1849 					 NULL,			/* lockarg */
1850 					 &fl->ifl_buf_tag);
1851 		if (err) {
1852 			device_printf(dev,
1853 			    "Unable to allocate RX DMA tag: %d\n", err);
1854 			goto fail;
1855 		}
1856 
1857 		/* Allocate memory for the RX mbuf map. */
1858 		if (!(fl->ifl_sds.ifsd_m =
1859 		      (struct mbuf **) malloc(sizeof(struct mbuf *) *
1860 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1861 			device_printf(dev,
1862 			    "Unable to allocate RX mbuf map memory\n");
1863 			err = ENOMEM;
1864 			goto fail;
1865 		}
1866 
1867 		/* Allocate memory for the direct RX cluster pointer map. */
1868 		if (!(fl->ifl_sds.ifsd_cl =
1869 		      (caddr_t *) malloc(sizeof(caddr_t) *
1870 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1871 			device_printf(dev,
1872 			    "Unable to allocate RX cluster map memory\n");
1873 			err = ENOMEM;
1874 			goto fail;
1875 		}
1876 
1877 		/* Allocate memory for the RX cluster bus address map. */
1878 		if (!(fl->ifl_sds.ifsd_ba =
1879 		      (bus_addr_t *) malloc(sizeof(bus_addr_t) *
1880 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1881 			device_printf(dev,
1882 			    "Unable to allocate RX bus address map memory\n");
1883 			err = ENOMEM;
1884 			goto fail;
1885 		}
1886 
1887 		/*
1888 		 * Create the DMA maps for RX buffers.
1889 		 */
1890 		if (!(fl->ifl_sds.ifsd_map =
1891 		      (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1892 			device_printf(dev,
1893 			    "Unable to allocate RX buffer DMA map memory\n");
1894 			err = ENOMEM;
1895 			goto fail;
1896 		}
1897 		for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++) {
1898 			err = bus_dmamap_create(fl->ifl_buf_tag, 0,
1899 			    &fl->ifl_sds.ifsd_map[i]);
1900 			if (err != 0) {
1901 				device_printf(dev, "Unable to create RX buffer DMA map\n");
1902 				goto fail;
1903 			}
1904 		}
1905 	}
1906 	return (0);
1907 
1908 fail:
1909 	iflib_rx_structures_free(ctx);
1910 	return (err);
1911 }
1912 
1913 
1914 /*
1915  * Internal service routines
1916  */
1917 
1918 struct rxq_refill_cb_arg {
1919 	int               error;
1920 	bus_dma_segment_t seg;
1921 	int               nseg;
1922 };
1923 
1924 static void
1925 _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
1926 {
1927 	struct rxq_refill_cb_arg *cb_arg = arg;
1928 
1929 	cb_arg->error = error;
1930 	cb_arg->seg = segs[0];
1931 	cb_arg->nseg = nseg;
1932 }
1933 
1934 /**
1935  * _iflib_fl_refill - refill an rxq free-buffer list
1936  * @ctx: the iflib context
1937  * @fl: the free list to refill
1938  * @count: the number of new buffers to allocate
1939  *
1940  * (Re)populate an rxq free-buffer list with up to @count new packet buffers.
1941  * The caller must assure that @count does not exceed the queue's capacity.
1942  */
1943 static void
1944 _iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
1945 {
1946 	struct if_rxd_update iru;
1947 	struct rxq_refill_cb_arg cb_arg;
1948 	struct mbuf *m;
1949 	caddr_t cl, *sd_cl;
1950 	struct mbuf **sd_m;
1951 	bus_dmamap_t *sd_map;
1952 	bus_addr_t bus_addr, *sd_ba;
1953 	int err, frag_idx, i, idx, n, pidx;
1954 	qidx_t credits;
1955 
1956 	sd_m = fl->ifl_sds.ifsd_m;
1957 	sd_map = fl->ifl_sds.ifsd_map;
1958 	sd_cl = fl->ifl_sds.ifsd_cl;
1959 	sd_ba = fl->ifl_sds.ifsd_ba;
1960 	pidx = fl->ifl_pidx;
1961 	idx = pidx;
1962 	frag_idx = fl->ifl_fragidx;
1963 	credits = fl->ifl_credits;
1964 
1965 	i = 0;
1966 	n = count;
1967 	MPASS(n > 0);
1968 	MPASS(credits + n <= fl->ifl_size);
1969 
1970 	if (pidx < fl->ifl_cidx)
1971 		MPASS(pidx + n <= fl->ifl_cidx);
1972 	if (pidx == fl->ifl_cidx && (credits < fl->ifl_size))
1973 		MPASS(fl->ifl_gen == 0);
1974 	if (pidx > fl->ifl_cidx)
1975 		MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx);
1976 
1977 	DBG_COUNTER_INC(fl_refills);
1978 	if (n > 8)
1979 		DBG_COUNTER_INC(fl_refills_large);
1980 	iru_init(&iru, fl->ifl_rxq, fl->ifl_id);
1981 	while (n--) {
1982 		/*
1983 		 * We allocate an uninitialized mbuf + cluster, mbuf is
1984 		 * initialized after rx.
1985 		 *
1986 		 * If the cluster is still set then we know a minimum sized packet was received
1987 		 */
1988 		bit_ffc_at(fl->ifl_rx_bitmap, frag_idx, fl->ifl_size,
1989 		    &frag_idx);
1990 		if (frag_idx < 0)
1991 			bit_ffc(fl->ifl_rx_bitmap, fl->ifl_size, &frag_idx);
1992 		MPASS(frag_idx >= 0);
1993 		if ((cl = sd_cl[frag_idx]) == NULL) {
1994 			if ((cl = m_cljget(NULL, M_NOWAIT, fl->ifl_buf_size)) == NULL)
1995 				break;
1996 
1997 			cb_arg.error = 0;
1998 			MPASS(sd_map != NULL);
1999 			err = bus_dmamap_load(fl->ifl_buf_tag, sd_map[frag_idx],
2000 			    cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg,
2001 			    BUS_DMA_NOWAIT);
2002 			if (err != 0 || cb_arg.error) {
2003 				/*
2004 				 * !zone_pack ?
2005 				 */
2006 				if (fl->ifl_zone == zone_pack)
2007 					uma_zfree(fl->ifl_zone, cl);
2008 				break;
2009 			}
2010 
2011 			sd_ba[frag_idx] =  bus_addr = cb_arg.seg.ds_addr;
2012 			sd_cl[frag_idx] = cl;
2013 #if MEMORY_LOGGING
2014 			fl->ifl_cl_enqueued++;
2015 #endif
2016 		} else {
2017 			bus_addr = sd_ba[frag_idx];
2018 		}
2019 		bus_dmamap_sync(fl->ifl_buf_tag, sd_map[frag_idx],
2020 		    BUS_DMASYNC_PREREAD);
2021 
2022 		if (sd_m[frag_idx] == NULL) {
2023 			if ((m = m_gethdr(M_NOWAIT, MT_NOINIT)) == NULL) {
2024 				break;
2025 			}
2026 			sd_m[frag_idx] = m;
2027 		}
2028 		bit_set(fl->ifl_rx_bitmap, frag_idx);
2029 #if MEMORY_LOGGING
2030 		fl->ifl_m_enqueued++;
2031 #endif
2032 
2033 		DBG_COUNTER_INC(rx_allocs);
2034 		fl->ifl_rxd_idxs[i] = frag_idx;
2035 		fl->ifl_bus_addrs[i] = bus_addr;
2036 		fl->ifl_vm_addrs[i] = cl;
2037 		credits++;
2038 		i++;
2039 		MPASS(credits <= fl->ifl_size);
2040 		if (++idx == fl->ifl_size) {
2041 			fl->ifl_gen = 1;
2042 			idx = 0;
2043 		}
2044 		if (n == 0 || i == IFLIB_MAX_RX_REFRESH) {
2045 			iru.iru_pidx = pidx;
2046 			iru.iru_count = i;
2047 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2048 			i = 0;
2049 			pidx = idx;
2050 			fl->ifl_pidx = idx;
2051 			fl->ifl_credits = credits;
2052 		}
2053 	}
2054 
2055 	if (i) {
2056 		iru.iru_pidx = pidx;
2057 		iru.iru_count = i;
2058 		ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2059 		fl->ifl_pidx = idx;
2060 		fl->ifl_credits = credits;
2061 	}
2062 	DBG_COUNTER_INC(rxd_flush);
2063 	if (fl->ifl_pidx == 0)
2064 		pidx = fl->ifl_size - 1;
2065 	else
2066 		pidx = fl->ifl_pidx - 1;
2067 
2068 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2069 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2070 	ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id, fl->ifl_id, pidx);
2071 	fl->ifl_fragidx = frag_idx;
2072 }
2073 
2074 static __inline void
2075 __iflib_fl_refill_lt(if_ctx_t ctx, iflib_fl_t fl, int max)
2076 {
2077 	/* we avoid allowing pidx to catch up with cidx as it confuses ixl */
2078 	int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1;
2079 #ifdef INVARIANTS
2080 	int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1;
2081 #endif
2082 
2083 	MPASS(fl->ifl_credits <= fl->ifl_size);
2084 	MPASS(reclaimable == delta);
2085 
2086 	if (reclaimable > 0)
2087 		_iflib_fl_refill(ctx, fl, min(max, reclaimable));
2088 }
2089 
2090 uint8_t
2091 iflib_in_detach(if_ctx_t ctx)
2092 {
2093 	bool in_detach;
2094 
2095 	STATE_LOCK(ctx);
2096 	in_detach = !!(ctx->ifc_flags & IFC_IN_DETACH);
2097 	STATE_UNLOCK(ctx);
2098 	return (in_detach);
2099 }
2100 
2101 static void
2102 iflib_fl_bufs_free(iflib_fl_t fl)
2103 {
2104 	iflib_dma_info_t idi = fl->ifl_ifdi;
2105 	bus_dmamap_t sd_map;
2106 	uint32_t i;
2107 
2108 	for (i = 0; i < fl->ifl_size; i++) {
2109 		struct mbuf **sd_m = &fl->ifl_sds.ifsd_m[i];
2110 		caddr_t *sd_cl = &fl->ifl_sds.ifsd_cl[i];
2111 
2112 		if (*sd_cl != NULL) {
2113 			sd_map = fl->ifl_sds.ifsd_map[i];
2114 			bus_dmamap_sync(fl->ifl_buf_tag, sd_map,
2115 			    BUS_DMASYNC_POSTREAD);
2116 			bus_dmamap_unload(fl->ifl_buf_tag, sd_map);
2117 			if (*sd_cl != NULL)
2118 				uma_zfree(fl->ifl_zone, *sd_cl);
2119 			if (*sd_m != NULL) {
2120 				m_init(*sd_m, M_NOWAIT, MT_DATA, 0);
2121 				uma_zfree(zone_mbuf, *sd_m);
2122 			}
2123 		} else {
2124 			MPASS(*sd_cl == NULL);
2125 			MPASS(*sd_m == NULL);
2126 		}
2127 #if MEMORY_LOGGING
2128 		fl->ifl_m_dequeued++;
2129 		fl->ifl_cl_dequeued++;
2130 #endif
2131 		*sd_cl = NULL;
2132 		*sd_m = NULL;
2133 	}
2134 #ifdef INVARIANTS
2135 	for (i = 0; i < fl->ifl_size; i++) {
2136 		MPASS(fl->ifl_sds.ifsd_cl[i] == NULL);
2137 		MPASS(fl->ifl_sds.ifsd_m[i] == NULL);
2138 	}
2139 #endif
2140 	/*
2141 	 * Reset free list values
2142 	 */
2143 	fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = fl->ifl_fragidx = 0;
2144 	bzero(idi->idi_vaddr, idi->idi_size);
2145 }
2146 
2147 /*********************************************************************
2148  *
2149  *  Initialize a free list and its buffers.
2150  *
2151  **********************************************************************/
2152 static int
2153 iflib_fl_setup(iflib_fl_t fl)
2154 {
2155 	iflib_rxq_t rxq = fl->ifl_rxq;
2156 	if_ctx_t ctx = rxq->ifr_ctx;
2157 
2158 	bit_nclear(fl->ifl_rx_bitmap, 0, fl->ifl_size - 1);
2159 	/*
2160 	** Free current RX buffer structs and their mbufs
2161 	*/
2162 	iflib_fl_bufs_free(fl);
2163 	/* Now replenish the mbufs */
2164 	MPASS(fl->ifl_credits == 0);
2165 	fl->ifl_buf_size = ctx->ifc_rx_mbuf_sz;
2166 	if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size)
2167 		ctx->ifc_max_fl_buf_size = fl->ifl_buf_size;
2168 	fl->ifl_cltype = m_gettype(fl->ifl_buf_size);
2169 	fl->ifl_zone = m_getzone(fl->ifl_buf_size);
2170 
2171 
2172 	/* avoid pre-allocating zillions of clusters to an idle card
2173 	 * potentially speeding up attach
2174 	 */
2175 	_iflib_fl_refill(ctx, fl, min(128, fl->ifl_size));
2176 	MPASS(min(128, fl->ifl_size) == fl->ifl_credits);
2177 	if (min(128, fl->ifl_size) != fl->ifl_credits)
2178 		return (ENOBUFS);
2179 	/*
2180 	 * handle failure
2181 	 */
2182 	MPASS(rxq != NULL);
2183 	MPASS(fl->ifl_ifdi != NULL);
2184 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2185 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2186 	return (0);
2187 }
2188 
2189 /*********************************************************************
2190  *
2191  *  Free receive ring data structures
2192  *
2193  **********************************************************************/
2194 static void
2195 iflib_rx_sds_free(iflib_rxq_t rxq)
2196 {
2197 	iflib_fl_t fl;
2198 	int i, j;
2199 
2200 	if (rxq->ifr_fl != NULL) {
2201 		for (i = 0; i < rxq->ifr_nfl; i++) {
2202 			fl = &rxq->ifr_fl[i];
2203 			if (fl->ifl_buf_tag != NULL) {
2204 				if (fl->ifl_sds.ifsd_map != NULL) {
2205 					for (j = 0; j < fl->ifl_size; j++) {
2206 						bus_dmamap_sync(
2207 						    fl->ifl_buf_tag,
2208 						    fl->ifl_sds.ifsd_map[j],
2209 						    BUS_DMASYNC_POSTREAD);
2210 						bus_dmamap_unload(
2211 						    fl->ifl_buf_tag,
2212 						    fl->ifl_sds.ifsd_map[j]);
2213 						bus_dmamap_destroy(
2214 						    fl->ifl_buf_tag,
2215 						    fl->ifl_sds.ifsd_map[j]);
2216 					}
2217 				}
2218 				bus_dma_tag_destroy(fl->ifl_buf_tag);
2219 				fl->ifl_buf_tag = NULL;
2220 			}
2221 			free(fl->ifl_sds.ifsd_m, M_IFLIB);
2222 			free(fl->ifl_sds.ifsd_cl, M_IFLIB);
2223 			free(fl->ifl_sds.ifsd_ba, M_IFLIB);
2224 			free(fl->ifl_sds.ifsd_map, M_IFLIB);
2225 			fl->ifl_sds.ifsd_m = NULL;
2226 			fl->ifl_sds.ifsd_cl = NULL;
2227 			fl->ifl_sds.ifsd_ba = NULL;
2228 			fl->ifl_sds.ifsd_map = NULL;
2229 		}
2230 		free(rxq->ifr_fl, M_IFLIB);
2231 		rxq->ifr_fl = NULL;
2232 		free(rxq->ifr_ifdi, M_IFLIB);
2233 		rxq->ifr_ifdi = NULL;
2234 		rxq->ifr_cq_cidx = 0;
2235 	}
2236 }
2237 
2238 /*
2239  * Timer routine
2240  */
2241 static void
2242 iflib_timer(void *arg)
2243 {
2244 	iflib_txq_t txq = arg;
2245 	if_ctx_t ctx = txq->ift_ctx;
2246 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2247 	uint64_t this_tick = ticks;
2248 	uint32_t reset_on = hz / 2;
2249 
2250 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
2251 		return;
2252 
2253 	/*
2254 	** Check on the state of the TX queue(s), this
2255 	** can be done without the lock because its RO
2256 	** and the HUNG state will be static if set.
2257 	*/
2258 	if (this_tick - txq->ift_last_timer_tick >= hz / 2) {
2259 		txq->ift_last_timer_tick = this_tick;
2260 		IFDI_TIMER(ctx, txq->ift_id);
2261 		if ((txq->ift_qstatus == IFLIB_QUEUE_HUNG) &&
2262 		    ((txq->ift_cleaned_prev == txq->ift_cleaned) ||
2263 		     (sctx->isc_pause_frames == 0)))
2264 			goto hung;
2265 
2266 		if (ifmp_ring_is_stalled(txq->ift_br))
2267 			txq->ift_qstatus = IFLIB_QUEUE_HUNG;
2268 		txq->ift_cleaned_prev = txq->ift_cleaned;
2269 	}
2270 #ifdef DEV_NETMAP
2271 	if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP)
2272 		iflib_netmap_timer_adjust(ctx, txq, &reset_on);
2273 #endif
2274 	/* handle any laggards */
2275 	if (txq->ift_db_pending)
2276 		GROUPTASK_ENQUEUE(&txq->ift_task);
2277 
2278 	sctx->isc_pause_frames = 0;
2279 	if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)
2280 		callout_reset_on(&txq->ift_timer, reset_on, iflib_timer, txq, txq->ift_timer.c_cpu);
2281 	return;
2282 
2283  hung:
2284 	device_printf(ctx->ifc_dev,
2285 	    "Watchdog timeout (TX: %d desc avail: %d pidx: %d) -- resetting\n",
2286 	    txq->ift_id, TXQ_AVAIL(txq), txq->ift_pidx);
2287 	STATE_LOCK(ctx);
2288 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2289 	ctx->ifc_flags |= (IFC_DO_WATCHDOG|IFC_DO_RESET);
2290 	iflib_admin_intr_deferred(ctx);
2291 	STATE_UNLOCK(ctx);
2292 }
2293 
2294 static void
2295 iflib_calc_rx_mbuf_sz(if_ctx_t ctx)
2296 {
2297 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2298 
2299 	/*
2300 	 * XXX don't set the max_frame_size to larger
2301 	 * than the hardware can handle
2302 	 */
2303 	if (sctx->isc_max_frame_size <= MCLBYTES)
2304 		ctx->ifc_rx_mbuf_sz = MCLBYTES;
2305 	else
2306 		ctx->ifc_rx_mbuf_sz = MJUMPAGESIZE;
2307 }
2308 
2309 uint32_t
2310 iflib_get_rx_mbuf_sz(if_ctx_t ctx)
2311 {
2312 
2313 	return (ctx->ifc_rx_mbuf_sz);
2314 }
2315 
2316 static void
2317 iflib_init_locked(if_ctx_t ctx)
2318 {
2319 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2320 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2321 	if_t ifp = ctx->ifc_ifp;
2322 	iflib_fl_t fl;
2323 	iflib_txq_t txq;
2324 	iflib_rxq_t rxq;
2325 	int i, j, tx_ip_csum_flags, tx_ip6_csum_flags;
2326 
2327 	if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2328 	IFDI_INTR_DISABLE(ctx);
2329 
2330 	tx_ip_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP);
2331 	tx_ip6_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | CSUM_IP6_SCTP);
2332 	/* Set hardware offload abilities */
2333 	if_clearhwassist(ifp);
2334 	if (if_getcapenable(ifp) & IFCAP_TXCSUM)
2335 		if_sethwassistbits(ifp, tx_ip_csum_flags, 0);
2336 	if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6)
2337 		if_sethwassistbits(ifp,  tx_ip6_csum_flags, 0);
2338 	if (if_getcapenable(ifp) & IFCAP_TSO4)
2339 		if_sethwassistbits(ifp, CSUM_IP_TSO, 0);
2340 	if (if_getcapenable(ifp) & IFCAP_TSO6)
2341 		if_sethwassistbits(ifp, CSUM_IP6_TSO, 0);
2342 
2343 	for (i = 0, txq = ctx->ifc_txqs; i < sctx->isc_ntxqsets; i++, txq++) {
2344 		CALLOUT_LOCK(txq);
2345 		callout_stop(&txq->ift_timer);
2346 		CALLOUT_UNLOCK(txq);
2347 		iflib_netmap_txq_init(ctx, txq);
2348 	}
2349 
2350 	/*
2351 	 * Calculate a suitable Rx mbuf size prior to calling IFDI_INIT, so
2352 	 * that drivers can use the value when setting up the hardware receive
2353 	 * buffers.
2354 	 */
2355 	iflib_calc_rx_mbuf_sz(ctx);
2356 
2357 #ifdef INVARIANTS
2358 	i = if_getdrvflags(ifp);
2359 #endif
2360 	IFDI_INIT(ctx);
2361 	MPASS(if_getdrvflags(ifp) == i);
2362 	for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
2363 		/* XXX this should really be done on a per-queue basis */
2364 		if (if_getcapenable(ifp) & IFCAP_NETMAP) {
2365 			MPASS(rxq->ifr_id == i);
2366 			iflib_netmap_rxq_init(ctx, rxq);
2367 			continue;
2368 		}
2369 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
2370 			if (iflib_fl_setup(fl)) {
2371 				device_printf(ctx->ifc_dev,
2372 				    "setting up free list %d failed - "
2373 				    "check cluster settings\n", j);
2374 				goto done;
2375 			}
2376 		}
2377 	}
2378 done:
2379 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
2380 	IFDI_INTR_ENABLE(ctx);
2381 	txq = ctx->ifc_txqs;
2382 	for (i = 0; i < sctx->isc_ntxqsets; i++, txq++)
2383 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq,
2384 			txq->ift_timer.c_cpu);
2385 }
2386 
2387 static int
2388 iflib_media_change(if_t ifp)
2389 {
2390 	if_ctx_t ctx = if_getsoftc(ifp);
2391 	int err;
2392 
2393 	CTX_LOCK(ctx);
2394 	if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0)
2395 		iflib_init_locked(ctx);
2396 	CTX_UNLOCK(ctx);
2397 	return (err);
2398 }
2399 
2400 static void
2401 iflib_media_status(if_t ifp, struct ifmediareq *ifmr)
2402 {
2403 	if_ctx_t ctx = if_getsoftc(ifp);
2404 
2405 	CTX_LOCK(ctx);
2406 	IFDI_UPDATE_ADMIN_STATUS(ctx);
2407 	IFDI_MEDIA_STATUS(ctx, ifmr);
2408 	CTX_UNLOCK(ctx);
2409 }
2410 
2411 void
2412 iflib_stop(if_ctx_t ctx)
2413 {
2414 	iflib_txq_t txq = ctx->ifc_txqs;
2415 	iflib_rxq_t rxq = ctx->ifc_rxqs;
2416 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2417 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2418 	iflib_dma_info_t di;
2419 	iflib_fl_t fl;
2420 	int i, j;
2421 
2422 	/* Tell the stack that the interface is no longer active */
2423 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2424 
2425 	IFDI_INTR_DISABLE(ctx);
2426 	DELAY(1000);
2427 	IFDI_STOP(ctx);
2428 	DELAY(1000);
2429 
2430 	iflib_debug_reset();
2431 	/* Wait for current tx queue users to exit to disarm watchdog timer. */
2432 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) {
2433 		/* make sure all transmitters have completed before proceeding XXX */
2434 
2435 		CALLOUT_LOCK(txq);
2436 		callout_stop(&txq->ift_timer);
2437 		CALLOUT_UNLOCK(txq);
2438 
2439 		/* clean any enqueued buffers */
2440 		iflib_ifmp_purge(txq);
2441 		/* Free any existing tx buffers. */
2442 		for (j = 0; j < txq->ift_size; j++) {
2443 			iflib_txsd_free(ctx, txq, j);
2444 		}
2445 		txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0;
2446 		txq->ift_in_use = txq->ift_gen = txq->ift_cidx = txq->ift_pidx = txq->ift_no_desc_avail = 0;
2447 		txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0;
2448 		txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0;
2449 		txq->ift_pullups = 0;
2450 		ifmp_ring_reset_stats(txq->ift_br);
2451 		for (j = 0, di = txq->ift_ifdi; j < sctx->isc_ntxqs; j++, di++)
2452 			bzero((void *)di->idi_vaddr, di->idi_size);
2453 	}
2454 	for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) {
2455 		/* make sure all transmitters have completed before proceeding XXX */
2456 
2457 		rxq->ifr_cq_cidx = 0;
2458 		for (j = 0, di = rxq->ifr_ifdi; j < sctx->isc_nrxqs; j++, di++)
2459 			bzero((void *)di->idi_vaddr, di->idi_size);
2460 		/* also resets the free lists pidx/cidx */
2461 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
2462 			iflib_fl_bufs_free(fl);
2463 	}
2464 }
2465 
2466 static inline caddr_t
2467 calc_next_rxd(iflib_fl_t fl, int cidx)
2468 {
2469 	qidx_t size;
2470 	int nrxd;
2471 	caddr_t start, end, cur, next;
2472 
2473 	nrxd = fl->ifl_size;
2474 	size = fl->ifl_rxd_size;
2475 	start = fl->ifl_ifdi->idi_vaddr;
2476 
2477 	if (__predict_false(size == 0))
2478 		return (start);
2479 	cur = start + size*cidx;
2480 	end = start + size*nrxd;
2481 	next = CACHE_PTR_NEXT(cur);
2482 	return (next < end ? next : start);
2483 }
2484 
2485 static inline void
2486 prefetch_pkts(iflib_fl_t fl, int cidx)
2487 {
2488 	int nextptr;
2489 	int nrxd = fl->ifl_size;
2490 	caddr_t next_rxd;
2491 
2492 
2493 	nextptr = (cidx + CACHE_PTR_INCREMENT) & (nrxd-1);
2494 	prefetch(&fl->ifl_sds.ifsd_m[nextptr]);
2495 	prefetch(&fl->ifl_sds.ifsd_cl[nextptr]);
2496 	next_rxd = calc_next_rxd(fl, cidx);
2497 	prefetch(next_rxd);
2498 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 1) & (nrxd-1)]);
2499 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 2) & (nrxd-1)]);
2500 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 3) & (nrxd-1)]);
2501 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 4) & (nrxd-1)]);
2502 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 1) & (nrxd-1)]);
2503 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 2) & (nrxd-1)]);
2504 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 3) & (nrxd-1)]);
2505 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 4) & (nrxd-1)]);
2506 }
2507 
2508 static struct mbuf *
2509 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, bool unload, if_rxsd_t sd,
2510     int *pf_rv, if_rxd_info_t ri)
2511 {
2512 	bus_dmamap_t map;
2513 	iflib_fl_t fl;
2514 	caddr_t payload;
2515 	struct mbuf *m;
2516 	int flid, cidx, len, next;
2517 
2518 	map = NULL;
2519 	flid = irf->irf_flid;
2520 	cidx = irf->irf_idx;
2521 	fl = &rxq->ifr_fl[flid];
2522 	sd->ifsd_fl = fl;
2523 	sd->ifsd_cidx = cidx;
2524 	m = fl->ifl_sds.ifsd_m[cidx];
2525 	sd->ifsd_cl = &fl->ifl_sds.ifsd_cl[cidx];
2526 	fl->ifl_credits--;
2527 #if MEMORY_LOGGING
2528 	fl->ifl_m_dequeued++;
2529 #endif
2530 	if (rxq->ifr_ctx->ifc_flags & IFC_PREFETCH)
2531 		prefetch_pkts(fl, cidx);
2532 	next = (cidx + CACHE_PTR_INCREMENT) & (fl->ifl_size-1);
2533 	prefetch(&fl->ifl_sds.ifsd_map[next]);
2534 	map = fl->ifl_sds.ifsd_map[cidx];
2535 	next = (cidx + CACHE_LINE_SIZE) & (fl->ifl_size-1);
2536 
2537 	/* not valid assert if bxe really does SGE from non-contiguous elements */
2538 	MPASS(fl->ifl_cidx == cidx);
2539 	bus_dmamap_sync(fl->ifl_buf_tag, map, BUS_DMASYNC_POSTREAD);
2540 
2541 	if (rxq->pfil != NULL && PFIL_HOOKED_IN(rxq->pfil) && pf_rv != NULL) {
2542 		payload  = *sd->ifsd_cl;
2543 		payload +=  ri->iri_pad;
2544 		len = ri->iri_len - ri->iri_pad;
2545 		*pf_rv = pfil_run_hooks(rxq->pfil, payload, ri->iri_ifp,
2546 		    len | PFIL_MEMPTR | PFIL_IN, NULL);
2547 		switch (*pf_rv) {
2548 		case PFIL_DROPPED:
2549 		case PFIL_CONSUMED:
2550 			/*
2551 			 * The filter ate it.  Everything is recycled.
2552 			 */
2553 			m = NULL;
2554 			unload = 0;
2555 			break;
2556 		case PFIL_REALLOCED:
2557 			/*
2558 			 * The filter copied it.  Everything is recycled.
2559 			 */
2560 			m = pfil_mem2mbuf(payload);
2561 			unload = 0;
2562 			break;
2563 		case PFIL_PASS:
2564 			/*
2565 			 * Filter said it was OK, so receive like
2566 			 * normal
2567 			 */
2568 			fl->ifl_sds.ifsd_m[cidx] = NULL;
2569 			break;
2570 		default:
2571 			MPASS(0);
2572 		}
2573 	} else {
2574 		fl->ifl_sds.ifsd_m[cidx] = NULL;
2575 		*pf_rv = PFIL_PASS;
2576 	}
2577 
2578 	if (unload)
2579 		bus_dmamap_unload(fl->ifl_buf_tag, map);
2580 	fl->ifl_cidx = (fl->ifl_cidx + 1) & (fl->ifl_size-1);
2581 	if (__predict_false(fl->ifl_cidx == 0))
2582 		fl->ifl_gen = 0;
2583 	bit_clear(fl->ifl_rx_bitmap, cidx);
2584 	return (m);
2585 }
2586 
2587 static struct mbuf *
2588 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri, if_rxsd_t sd, int *pf_rv)
2589 {
2590 	struct mbuf *m, *mh, *mt;
2591 	caddr_t cl;
2592 	int  *pf_rv_ptr, flags, i, padlen;
2593 	bool consumed;
2594 
2595 	i = 0;
2596 	mh = NULL;
2597 	consumed = false;
2598 	*pf_rv = PFIL_PASS;
2599 	pf_rv_ptr = pf_rv;
2600 	do {
2601 		m = rxd_frag_to_sd(rxq, &ri->iri_frags[i], !consumed, sd,
2602 		    pf_rv_ptr, ri);
2603 
2604 		MPASS(*sd->ifsd_cl != NULL);
2605 
2606 		/*
2607 		 * Exclude zero-length frags & frags from
2608 		 * packets the filter has consumed or dropped
2609 		 */
2610 		if (ri->iri_frags[i].irf_len == 0 || consumed ||
2611 		    *pf_rv == PFIL_CONSUMED || *pf_rv == PFIL_DROPPED) {
2612 			if (mh == NULL) {
2613 				/* everything saved here */
2614 				consumed = true;
2615 				pf_rv_ptr = NULL;
2616 				continue;
2617 			}
2618 			/* XXX we can save the cluster here, but not the mbuf */
2619 			m_init(m, M_NOWAIT, MT_DATA, 0);
2620 			m_free(m);
2621 			continue;
2622 		}
2623 		if (mh == NULL) {
2624 			flags = M_PKTHDR|M_EXT;
2625 			mh = mt = m;
2626 			padlen = ri->iri_pad;
2627 		} else {
2628 			flags = M_EXT;
2629 			mt->m_next = m;
2630 			mt = m;
2631 			/* assuming padding is only on the first fragment */
2632 			padlen = 0;
2633 		}
2634 		cl = *sd->ifsd_cl;
2635 		*sd->ifsd_cl = NULL;
2636 
2637 		/* Can these two be made one ? */
2638 		m_init(m, M_NOWAIT, MT_DATA, flags);
2639 		m_cljset(m, cl, sd->ifsd_fl->ifl_cltype);
2640 		/*
2641 		 * These must follow m_init and m_cljset
2642 		 */
2643 		m->m_data += padlen;
2644 		ri->iri_len -= padlen;
2645 		m->m_len = ri->iri_frags[i].irf_len;
2646 	} while (++i < ri->iri_nfrags);
2647 
2648 	return (mh);
2649 }
2650 
2651 /*
2652  * Process one software descriptor
2653  */
2654 static struct mbuf *
2655 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri)
2656 {
2657 	struct if_rxsd sd;
2658 	struct mbuf *m;
2659 	int pf_rv;
2660 
2661 	/* should I merge this back in now that the two paths are basically duplicated? */
2662 	if (ri->iri_nfrags == 1 &&
2663 	    ri->iri_frags[0].irf_len <= MIN(IFLIB_RX_COPY_THRESH, MHLEN)) {
2664 		m = rxd_frag_to_sd(rxq, &ri->iri_frags[0], false, &sd,
2665 		    &pf_rv, ri);
2666 		if (pf_rv != PFIL_PASS && pf_rv != PFIL_REALLOCED)
2667 			return (m);
2668 		if (pf_rv == PFIL_PASS) {
2669 			m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR);
2670 #ifndef __NO_STRICT_ALIGNMENT
2671 			if (!IP_ALIGNED(m))
2672 				m->m_data += 2;
2673 #endif
2674 			memcpy(m->m_data, *sd.ifsd_cl, ri->iri_len);
2675 			m->m_len = ri->iri_frags[0].irf_len;
2676 		}
2677 	} else {
2678 		m = assemble_segments(rxq, ri, &sd, &pf_rv);
2679 		if (pf_rv != PFIL_PASS && pf_rv != PFIL_REALLOCED)
2680 			return (m);
2681 	}
2682 	m->m_pkthdr.len = ri->iri_len;
2683 	m->m_pkthdr.rcvif = ri->iri_ifp;
2684 	m->m_flags |= ri->iri_flags;
2685 	m->m_pkthdr.ether_vtag = ri->iri_vtag;
2686 	m->m_pkthdr.flowid = ri->iri_flowid;
2687 	M_HASHTYPE_SET(m, ri->iri_rsstype);
2688 	m->m_pkthdr.csum_flags = ri->iri_csum_flags;
2689 	m->m_pkthdr.csum_data = ri->iri_csum_data;
2690 	return (m);
2691 }
2692 
2693 #if defined(INET6) || defined(INET)
2694 static void
2695 iflib_get_ip_forwarding(struct lro_ctrl *lc, bool *v4, bool *v6)
2696 {
2697 	CURVNET_SET(lc->ifp->if_vnet);
2698 #if defined(INET6)
2699 	*v6 = V_ip6_forwarding;
2700 #endif
2701 #if defined(INET)
2702 	*v4 = V_ipforwarding;
2703 #endif
2704 	CURVNET_RESTORE();
2705 }
2706 
2707 /*
2708  * Returns true if it's possible this packet could be LROed.
2709  * if it returns false, it is guaranteed that tcp_lro_rx()
2710  * would not return zero.
2711  */
2712 static bool
2713 iflib_check_lro_possible(struct mbuf *m, bool v4_forwarding, bool v6_forwarding)
2714 {
2715 	struct ether_header *eh;
2716 
2717 	eh = mtod(m, struct ether_header *);
2718 	switch (eh->ether_type) {
2719 #if defined(INET6)
2720 		case htons(ETHERTYPE_IPV6):
2721 			return (!v6_forwarding);
2722 #endif
2723 #if defined (INET)
2724 		case htons(ETHERTYPE_IP):
2725 			return (!v4_forwarding);
2726 #endif
2727 	}
2728 
2729 	return false;
2730 }
2731 #else
2732 static void
2733 iflib_get_ip_forwarding(struct lro_ctrl *lc __unused, bool *v4 __unused, bool *v6 __unused)
2734 {
2735 }
2736 #endif
2737 
2738 static bool
2739 iflib_rxeof(iflib_rxq_t rxq, qidx_t budget)
2740 {
2741 	if_t ifp;
2742 	if_ctx_t ctx = rxq->ifr_ctx;
2743 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2744 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2745 	int avail, i;
2746 	qidx_t *cidxp;
2747 	struct if_rxd_info ri;
2748 	int err, budget_left, rx_bytes, rx_pkts;
2749 	iflib_fl_t fl;
2750 	int lro_enabled;
2751 	bool v4_forwarding, v6_forwarding, lro_possible;
2752 
2753 	/*
2754 	 * XXX early demux data packets so that if_input processing only handles
2755 	 * acks in interrupt context
2756 	 */
2757 	struct mbuf *m, *mh, *mt, *mf;
2758 
2759 	lro_possible = v4_forwarding = v6_forwarding = false;
2760 	ifp = ctx->ifc_ifp;
2761 	mh = mt = NULL;
2762 	MPASS(budget > 0);
2763 	rx_pkts	= rx_bytes = 0;
2764 	if (sctx->isc_flags & IFLIB_HAS_RXCQ)
2765 		cidxp = &rxq->ifr_cq_cidx;
2766 	else
2767 		cidxp = &rxq->ifr_fl[0].ifl_cidx;
2768 	if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) {
2769 		for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2770 			__iflib_fl_refill_lt(ctx, fl, budget + 8);
2771 		DBG_COUNTER_INC(rx_unavail);
2772 		return (false);
2773 	}
2774 
2775 	/* pfil needs the vnet to be set */
2776 	CURVNET_SET_QUIET(ifp->if_vnet);
2777 	for (budget_left = budget; budget_left > 0 && avail > 0;) {
2778 		if (__predict_false(!CTX_ACTIVE(ctx))) {
2779 			DBG_COUNTER_INC(rx_ctx_inactive);
2780 			break;
2781 		}
2782 		/*
2783 		 * Reset client set fields to their default values
2784 		 */
2785 		rxd_info_zero(&ri);
2786 		ri.iri_qsidx = rxq->ifr_id;
2787 		ri.iri_cidx = *cidxp;
2788 		ri.iri_ifp = ifp;
2789 		ri.iri_frags = rxq->ifr_frags;
2790 		err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
2791 
2792 		if (err)
2793 			goto err;
2794 		rx_pkts += 1;
2795 		rx_bytes += ri.iri_len;
2796 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
2797 			*cidxp = ri.iri_cidx;
2798 			/* Update our consumer index */
2799 			/* XXX NB: shurd - check if this is still safe */
2800 			while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0])
2801 				rxq->ifr_cq_cidx -= scctx->isc_nrxd[0];
2802 			/* was this only a completion queue message? */
2803 			if (__predict_false(ri.iri_nfrags == 0))
2804 				continue;
2805 		}
2806 		MPASS(ri.iri_nfrags != 0);
2807 		MPASS(ri.iri_len != 0);
2808 
2809 		/* will advance the cidx on the corresponding free lists */
2810 		m = iflib_rxd_pkt_get(rxq, &ri);
2811 		avail--;
2812 		budget_left--;
2813 		if (avail == 0 && budget_left)
2814 			avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left);
2815 
2816 		if (__predict_false(m == NULL))
2817 			continue;
2818 
2819 		/* imm_pkt: -- cxgb */
2820 		if (mh == NULL)
2821 			mh = mt = m;
2822 		else {
2823 			mt->m_nextpkt = m;
2824 			mt = m;
2825 		}
2826 	}
2827 	CURVNET_RESTORE();
2828 	/* make sure that we can refill faster than drain */
2829 	for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2830 		__iflib_fl_refill_lt(ctx, fl, budget + 8);
2831 
2832 	lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO);
2833 	if (lro_enabled)
2834 		iflib_get_ip_forwarding(&rxq->ifr_lc, &v4_forwarding, &v6_forwarding);
2835 	mt = mf = NULL;
2836 	while (mh != NULL) {
2837 		m = mh;
2838 		mh = mh->m_nextpkt;
2839 		m->m_nextpkt = NULL;
2840 #ifndef __NO_STRICT_ALIGNMENT
2841 		if (!IP_ALIGNED(m) && (m = iflib_fixup_rx(m)) == NULL)
2842 			continue;
2843 #endif
2844 		rx_bytes += m->m_pkthdr.len;
2845 		rx_pkts++;
2846 #if defined(INET6) || defined(INET)
2847 		if (lro_enabled) {
2848 			if (!lro_possible) {
2849 				lro_possible = iflib_check_lro_possible(m, v4_forwarding, v6_forwarding);
2850 				if (lro_possible && mf != NULL) {
2851 					ifp->if_input(ifp, mf);
2852 					DBG_COUNTER_INC(rx_if_input);
2853 					mt = mf = NULL;
2854 				}
2855 			}
2856 			if ((m->m_pkthdr.csum_flags & (CSUM_L4_CALC|CSUM_L4_VALID)) ==
2857 			    (CSUM_L4_CALC|CSUM_L4_VALID)) {
2858 				if (lro_possible && tcp_lro_rx(&rxq->ifr_lc, m, 0) == 0)
2859 					continue;
2860 			}
2861 		}
2862 #endif
2863 		if (lro_possible) {
2864 			ifp->if_input(ifp, m);
2865 			DBG_COUNTER_INC(rx_if_input);
2866 			continue;
2867 		}
2868 
2869 		if (mf == NULL)
2870 			mf = m;
2871 		if (mt != NULL)
2872 			mt->m_nextpkt = m;
2873 		mt = m;
2874 	}
2875 	if (mf != NULL) {
2876 		ifp->if_input(ifp, mf);
2877 		DBG_COUNTER_INC(rx_if_input);
2878 	}
2879 
2880 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
2881 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
2882 
2883 	/*
2884 	 * Flush any outstanding LRO work
2885 	 */
2886 #if defined(INET6) || defined(INET)
2887 	tcp_lro_flush_all(&rxq->ifr_lc);
2888 #endif
2889 	if (avail)
2890 		return true;
2891 	return (iflib_rxd_avail(ctx, rxq, *cidxp, 1));
2892 err:
2893 	STATE_LOCK(ctx);
2894 	ctx->ifc_flags |= IFC_DO_RESET;
2895 	iflib_admin_intr_deferred(ctx);
2896 	STATE_UNLOCK(ctx);
2897 	return (false);
2898 }
2899 
2900 #define TXD_NOTIFY_COUNT(txq) (((txq)->ift_size / (txq)->ift_update_freq)-1)
2901 static inline qidx_t
2902 txq_max_db_deferred(iflib_txq_t txq, qidx_t in_use)
2903 {
2904 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2905 	qidx_t minthresh = txq->ift_size / 8;
2906 	if (in_use > 4*minthresh)
2907 		return (notify_count);
2908 	if (in_use > 2*minthresh)
2909 		return (notify_count >> 1);
2910 	if (in_use > minthresh)
2911 		return (notify_count >> 3);
2912 	return (0);
2913 }
2914 
2915 static inline qidx_t
2916 txq_max_rs_deferred(iflib_txq_t txq)
2917 {
2918 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2919 	qidx_t minthresh = txq->ift_size / 8;
2920 	if (txq->ift_in_use > 4*minthresh)
2921 		return (notify_count);
2922 	if (txq->ift_in_use > 2*minthresh)
2923 		return (notify_count >> 1);
2924 	if (txq->ift_in_use > minthresh)
2925 		return (notify_count >> 2);
2926 	return (2);
2927 }
2928 
2929 #define M_CSUM_FLAGS(m) ((m)->m_pkthdr.csum_flags)
2930 #define M_HAS_VLANTAG(m) (m->m_flags & M_VLANTAG)
2931 
2932 #define TXQ_MAX_DB_DEFERRED(txq, in_use) txq_max_db_deferred((txq), (in_use))
2933 #define TXQ_MAX_RS_DEFERRED(txq) txq_max_rs_deferred(txq)
2934 #define TXQ_MAX_DB_CONSUMED(size) (size >> 4)
2935 
2936 /* forward compatibility for cxgb */
2937 #define FIRST_QSET(ctx) 0
2938 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets)
2939 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets)
2940 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx))
2941 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments))
2942 
2943 /* XXX we should be setting this to something other than zero */
2944 #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh)
2945 #define	MAX_TX_DESC(ctx) max((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max, \
2946     (ctx)->ifc_softc_ctx.isc_tx_nsegments)
2947 
2948 static inline bool
2949 iflib_txd_db_check(if_ctx_t ctx, iflib_txq_t txq, int ring, qidx_t in_use)
2950 {
2951 	qidx_t dbval, max;
2952 	bool rang;
2953 
2954 	rang = false;
2955 	max = TXQ_MAX_DB_DEFERRED(txq, in_use);
2956 	if (ring || txq->ift_db_pending >= max) {
2957 		dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx;
2958 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
2959 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2960 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval);
2961 		txq->ift_db_pending = txq->ift_npending = 0;
2962 		rang = true;
2963 	}
2964 	return (rang);
2965 }
2966 
2967 #ifdef PKT_DEBUG
2968 static void
2969 print_pkt(if_pkt_info_t pi)
2970 {
2971 	printf("pi len:  %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n",
2972 	       pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx);
2973 	printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n",
2974 	       pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag);
2975 	printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n",
2976 	       pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto);
2977 }
2978 #endif
2979 
2980 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO)
2981 #define IS_TX_OFFLOAD4(pi) ((pi)->ipi_csum_flags & (CSUM_IP_TCP | CSUM_IP_TSO))
2982 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO)
2983 #define IS_TX_OFFLOAD6(pi) ((pi)->ipi_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_TSO))
2984 
2985 static int
2986 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp)
2987 {
2988 	if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx;
2989 	struct ether_vlan_header *eh;
2990 	struct mbuf *m;
2991 
2992 	m = *mp;
2993 	if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) &&
2994 	    M_WRITABLE(m) == 0) {
2995 		if ((m = m_dup(m, M_NOWAIT)) == NULL) {
2996 			return (ENOMEM);
2997 		} else {
2998 			m_freem(*mp);
2999 			DBG_COUNTER_INC(tx_frees);
3000 			*mp = m;
3001 		}
3002 	}
3003 
3004 	/*
3005 	 * Determine where frame payload starts.
3006 	 * Jump over vlan headers if already present,
3007 	 * helpful for QinQ too.
3008 	 */
3009 	if (__predict_false(m->m_len < sizeof(*eh))) {
3010 		txq->ift_pullups++;
3011 		if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL))
3012 			return (ENOMEM);
3013 	}
3014 	eh = mtod(m, struct ether_vlan_header *);
3015 	if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
3016 		pi->ipi_etype = ntohs(eh->evl_proto);
3017 		pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
3018 	} else {
3019 		pi->ipi_etype = ntohs(eh->evl_encap_proto);
3020 		pi->ipi_ehdrlen = ETHER_HDR_LEN;
3021 	}
3022 
3023 	switch (pi->ipi_etype) {
3024 #ifdef INET
3025 	case ETHERTYPE_IP:
3026 	{
3027 		struct mbuf *n;
3028 		struct ip *ip = NULL;
3029 		struct tcphdr *th = NULL;
3030 		int minthlen;
3031 
3032 		minthlen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip) + sizeof(*th));
3033 		if (__predict_false(m->m_len < minthlen)) {
3034 			/*
3035 			 * if this code bloat is causing too much of a hit
3036 			 * move it to a separate function and mark it noinline
3037 			 */
3038 			if (m->m_len == pi->ipi_ehdrlen) {
3039 				n = m->m_next;
3040 				MPASS(n);
3041 				if (n->m_len >= sizeof(*ip))  {
3042 					ip = (struct ip *)n->m_data;
3043 					if (n->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3044 						th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3045 				} else {
3046 					txq->ift_pullups++;
3047 					if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
3048 						return (ENOMEM);
3049 					ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3050 				}
3051 			} else {
3052 				txq->ift_pullups++;
3053 				if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
3054 					return (ENOMEM);
3055 				ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3056 				if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3057 					th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3058 			}
3059 		} else {
3060 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3061 			if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3062 				th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3063 		}
3064 		pi->ipi_ip_hlen = ip->ip_hl << 2;
3065 		pi->ipi_ipproto = ip->ip_p;
3066 		pi->ipi_flags |= IPI_TX_IPV4;
3067 
3068 		/* TCP checksum offload may require TCP header length */
3069 		if (IS_TX_OFFLOAD4(pi)) {
3070 			if (__predict_true(pi->ipi_ipproto == IPPROTO_TCP)) {
3071 				if (__predict_false(th == NULL)) {
3072 					txq->ift_pullups++;
3073 					if (__predict_false((m = m_pullup(m, (ip->ip_hl << 2) + sizeof(*th))) == NULL))
3074 						return (ENOMEM);
3075 					th = (struct tcphdr *)((caddr_t)ip + pi->ipi_ip_hlen);
3076 				}
3077 				pi->ipi_tcp_hflags = th->th_flags;
3078 				pi->ipi_tcp_hlen = th->th_off << 2;
3079 				pi->ipi_tcp_seq = th->th_seq;
3080 			}
3081 			if (IS_TSO4(pi)) {
3082 				if (__predict_false(ip->ip_p != IPPROTO_TCP))
3083 					return (ENXIO);
3084 				/*
3085 				 * TSO always requires hardware checksum offload.
3086 				 */
3087 				pi->ipi_csum_flags |= (CSUM_IP_TCP | CSUM_IP);
3088 				th->th_sum = in_pseudo(ip->ip_src.s_addr,
3089 						       ip->ip_dst.s_addr, htons(IPPROTO_TCP));
3090 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3091 				if (sctx->isc_flags & IFLIB_TSO_INIT_IP) {
3092 					ip->ip_sum = 0;
3093 					ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz);
3094 				}
3095 			}
3096 		}
3097 		if ((sctx->isc_flags & IFLIB_NEED_ZERO_CSUM) && (pi->ipi_csum_flags & CSUM_IP))
3098                        ip->ip_sum = 0;
3099 
3100 		break;
3101 	}
3102 #endif
3103 #ifdef INET6
3104 	case ETHERTYPE_IPV6:
3105 	{
3106 		struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
3107 		struct tcphdr *th;
3108 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
3109 
3110 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
3111 			txq->ift_pullups++;
3112 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
3113 				return (ENOMEM);
3114 		}
3115 		th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
3116 
3117 		/* XXX-BZ this will go badly in case of ext hdrs. */
3118 		pi->ipi_ipproto = ip6->ip6_nxt;
3119 		pi->ipi_flags |= IPI_TX_IPV6;
3120 
3121 		/* TCP checksum offload may require TCP header length */
3122 		if (IS_TX_OFFLOAD6(pi)) {
3123 			if (pi->ipi_ipproto == IPPROTO_TCP) {
3124 				if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) {
3125 					txq->ift_pullups++;
3126 					if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL))
3127 						return (ENOMEM);
3128 				}
3129 				pi->ipi_tcp_hflags = th->th_flags;
3130 				pi->ipi_tcp_hlen = th->th_off << 2;
3131 				pi->ipi_tcp_seq = th->th_seq;
3132 			}
3133 			if (IS_TSO6(pi)) {
3134 				if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP))
3135 					return (ENXIO);
3136 				/*
3137 				 * TSO always requires hardware checksum offload.
3138 				 */
3139 				pi->ipi_csum_flags |= CSUM_IP6_TCP;
3140 				th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0);
3141 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3142 			}
3143 		}
3144 		break;
3145 	}
3146 #endif
3147 	default:
3148 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
3149 		pi->ipi_ip_hlen = 0;
3150 		break;
3151 	}
3152 	*mp = m;
3153 
3154 	return (0);
3155 }
3156 
3157 /*
3158  * If dodgy hardware rejects the scatter gather chain we've handed it
3159  * we'll need to remove the mbuf chain from ifsg_m[] before we can add the
3160  * m_defrag'd mbufs
3161  */
3162 static __noinline struct mbuf *
3163 iflib_remove_mbuf(iflib_txq_t txq)
3164 {
3165 	int ntxd, pidx;
3166 	struct mbuf *m, **ifsd_m;
3167 
3168 	ifsd_m = txq->ift_sds.ifsd_m;
3169 	ntxd = txq->ift_size;
3170 	pidx = txq->ift_pidx & (ntxd - 1);
3171 	ifsd_m = txq->ift_sds.ifsd_m;
3172 	m = ifsd_m[pidx];
3173 	ifsd_m[pidx] = NULL;
3174 	bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[pidx]);
3175 	if (txq->ift_sds.ifsd_tso_map != NULL)
3176 		bus_dmamap_unload(txq->ift_tso_buf_tag,
3177 		    txq->ift_sds.ifsd_tso_map[pidx]);
3178 #if MEMORY_LOGGING
3179 	txq->ift_dequeued++;
3180 #endif
3181 	return (m);
3182 }
3183 
3184 static inline caddr_t
3185 calc_next_txd(iflib_txq_t txq, int cidx, uint8_t qid)
3186 {
3187 	qidx_t size;
3188 	int ntxd;
3189 	caddr_t start, end, cur, next;
3190 
3191 	ntxd = txq->ift_size;
3192 	size = txq->ift_txd_size[qid];
3193 	start = txq->ift_ifdi[qid].idi_vaddr;
3194 
3195 	if (__predict_false(size == 0))
3196 		return (start);
3197 	cur = start + size*cidx;
3198 	end = start + size*ntxd;
3199 	next = CACHE_PTR_NEXT(cur);
3200 	return (next < end ? next : start);
3201 }
3202 
3203 /*
3204  * Pad an mbuf to ensure a minimum ethernet frame size.
3205  * min_frame_size is the frame size (less CRC) to pad the mbuf to
3206  */
3207 static __noinline int
3208 iflib_ether_pad(device_t dev, struct mbuf **m_head, uint16_t min_frame_size)
3209 {
3210 	/*
3211 	 * 18 is enough bytes to pad an ARP packet to 46 bytes, and
3212 	 * and ARP message is the smallest common payload I can think of
3213 	 */
3214 	static char pad[18];	/* just zeros */
3215 	int n;
3216 	struct mbuf *new_head;
3217 
3218 	if (!M_WRITABLE(*m_head)) {
3219 		new_head = m_dup(*m_head, M_NOWAIT);
3220 		if (new_head == NULL) {
3221 			m_freem(*m_head);
3222 			device_printf(dev, "cannot pad short frame, m_dup() failed");
3223 			DBG_COUNTER_INC(encap_pad_mbuf_fail);
3224 			DBG_COUNTER_INC(tx_frees);
3225 			return ENOMEM;
3226 		}
3227 		m_freem(*m_head);
3228 		*m_head = new_head;
3229 	}
3230 
3231 	for (n = min_frame_size - (*m_head)->m_pkthdr.len;
3232 	     n > 0; n -= sizeof(pad))
3233 		if (!m_append(*m_head, min(n, sizeof(pad)), pad))
3234 			break;
3235 
3236 	if (n > 0) {
3237 		m_freem(*m_head);
3238 		device_printf(dev, "cannot pad short frame\n");
3239 		DBG_COUNTER_INC(encap_pad_mbuf_fail);
3240 		DBG_COUNTER_INC(tx_frees);
3241 		return (ENOBUFS);
3242 	}
3243 
3244 	return 0;
3245 }
3246 
3247 static int
3248 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp)
3249 {
3250 	if_ctx_t		ctx;
3251 	if_shared_ctx_t		sctx;
3252 	if_softc_ctx_t		scctx;
3253 	bus_dma_tag_t		buf_tag;
3254 	bus_dma_segment_t	*segs;
3255 	struct mbuf		*m_head, **ifsd_m;
3256 	void			*next_txd;
3257 	bus_dmamap_t		map;
3258 	struct if_pkt_info	pi;
3259 	int remap = 0;
3260 	int err, nsegs, ndesc, max_segs, pidx, cidx, next, ntxd;
3261 
3262 	ctx = txq->ift_ctx;
3263 	sctx = ctx->ifc_sctx;
3264 	scctx = &ctx->ifc_softc_ctx;
3265 	segs = txq->ift_segs;
3266 	ntxd = txq->ift_size;
3267 	m_head = *m_headp;
3268 	map = NULL;
3269 
3270 	/*
3271 	 * If we're doing TSO the next descriptor to clean may be quite far ahead
3272 	 */
3273 	cidx = txq->ift_cidx;
3274 	pidx = txq->ift_pidx;
3275 	if (ctx->ifc_flags & IFC_PREFETCH) {
3276 		next = (cidx + CACHE_PTR_INCREMENT) & (ntxd-1);
3277 		if (!(ctx->ifc_flags & IFLIB_HAS_TXCQ)) {
3278 			next_txd = calc_next_txd(txq, cidx, 0);
3279 			prefetch(next_txd);
3280 		}
3281 
3282 		/* prefetch the next cache line of mbuf pointers and flags */
3283 		prefetch(&txq->ift_sds.ifsd_m[next]);
3284 		prefetch(&txq->ift_sds.ifsd_map[next]);
3285 		next = (cidx + CACHE_LINE_SIZE) & (ntxd-1);
3286 	}
3287 	map = txq->ift_sds.ifsd_map[pidx];
3288 	ifsd_m = txq->ift_sds.ifsd_m;
3289 
3290 	if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
3291 		buf_tag = txq->ift_tso_buf_tag;
3292 		max_segs = scctx->isc_tx_tso_segments_max;
3293 		map = txq->ift_sds.ifsd_tso_map[pidx];
3294 		MPASS(buf_tag != NULL);
3295 		MPASS(max_segs > 0);
3296 	} else {
3297 		buf_tag = txq->ift_buf_tag;
3298 		max_segs = scctx->isc_tx_nsegments;
3299 		map = txq->ift_sds.ifsd_map[pidx];
3300 	}
3301 	if ((sctx->isc_flags & IFLIB_NEED_ETHER_PAD) &&
3302 	    __predict_false(m_head->m_pkthdr.len < scctx->isc_min_frame_size)) {
3303 		err = iflib_ether_pad(ctx->ifc_dev, m_headp, scctx->isc_min_frame_size);
3304 		if (err) {
3305 			DBG_COUNTER_INC(encap_txd_encap_fail);
3306 			return err;
3307 		}
3308 	}
3309 	m_head = *m_headp;
3310 
3311 	pkt_info_zero(&pi);
3312 	pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG|M_BCAST|M_MCAST));
3313 	pi.ipi_pidx = pidx;
3314 	pi.ipi_qsidx = txq->ift_id;
3315 	pi.ipi_len = m_head->m_pkthdr.len;
3316 	pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags;
3317 	pi.ipi_vtag = M_HAS_VLANTAG(m_head) ? m_head->m_pkthdr.ether_vtag : 0;
3318 
3319 	/* deliberate bitwise OR to make one condition */
3320 	if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) {
3321 		if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0)) {
3322 			DBG_COUNTER_INC(encap_txd_encap_fail);
3323 			return (err);
3324 		}
3325 		m_head = *m_headp;
3326 	}
3327 
3328 retry:
3329 	err = bus_dmamap_load_mbuf_sg(buf_tag, map, m_head, segs, &nsegs,
3330 	    BUS_DMA_NOWAIT);
3331 defrag:
3332 	if (__predict_false(err)) {
3333 		switch (err) {
3334 		case EFBIG:
3335 			/* try collapse once and defrag once */
3336 			if (remap == 0) {
3337 				m_head = m_collapse(*m_headp, M_NOWAIT, max_segs);
3338 				/* try defrag if collapsing fails */
3339 				if (m_head == NULL)
3340 					remap++;
3341 			}
3342 			if (remap == 1) {
3343 				txq->ift_mbuf_defrag++;
3344 				m_head = m_defrag(*m_headp, M_NOWAIT);
3345 			}
3346 			/*
3347 			 * remap should never be >1 unless bus_dmamap_load_mbuf_sg
3348 			 * failed to map an mbuf that was run through m_defrag
3349 			 */
3350 			MPASS(remap <= 1);
3351 			if (__predict_false(m_head == NULL || remap > 1))
3352 				goto defrag_failed;
3353 			remap++;
3354 			*m_headp = m_head;
3355 			goto retry;
3356 			break;
3357 		case ENOMEM:
3358 			txq->ift_no_tx_dma_setup++;
3359 			break;
3360 		default:
3361 			txq->ift_no_tx_dma_setup++;
3362 			m_freem(*m_headp);
3363 			DBG_COUNTER_INC(tx_frees);
3364 			*m_headp = NULL;
3365 			break;
3366 		}
3367 		txq->ift_map_failed++;
3368 		DBG_COUNTER_INC(encap_load_mbuf_fail);
3369 		DBG_COUNTER_INC(encap_txd_encap_fail);
3370 		return (err);
3371 	}
3372 	ifsd_m[pidx] = m_head;
3373 	/*
3374 	 * XXX assumes a 1 to 1 relationship between segments and
3375 	 *        descriptors - this does not hold true on all drivers, e.g.
3376 	 *        cxgb
3377 	 */
3378 	if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) {
3379 		txq->ift_no_desc_avail++;
3380 		bus_dmamap_unload(buf_tag, map);
3381 		DBG_COUNTER_INC(encap_txq_avail_fail);
3382 		DBG_COUNTER_INC(encap_txd_encap_fail);
3383 		if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0)
3384 			GROUPTASK_ENQUEUE(&txq->ift_task);
3385 		return (ENOBUFS);
3386 	}
3387 	/*
3388 	 * On Intel cards we can greatly reduce the number of TX interrupts
3389 	 * we see by only setting report status on every Nth descriptor.
3390 	 * However, this also means that the driver will need to keep track
3391 	 * of the descriptors that RS was set on to check them for the DD bit.
3392 	 */
3393 	txq->ift_rs_pending += nsegs + 1;
3394 	if (txq->ift_rs_pending > TXQ_MAX_RS_DEFERRED(txq) ||
3395 	     iflib_no_tx_batch || (TXQ_AVAIL(txq) - nsegs) <= MAX_TX_DESC(ctx) + 2) {
3396 		pi.ipi_flags |= IPI_TX_INTR;
3397 		txq->ift_rs_pending = 0;
3398 	}
3399 
3400 	pi.ipi_segs = segs;
3401 	pi.ipi_nsegs = nsegs;
3402 
3403 	MPASS(pidx >= 0 && pidx < txq->ift_size);
3404 #ifdef PKT_DEBUG
3405 	print_pkt(&pi);
3406 #endif
3407 	if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) {
3408 		bus_dmamap_sync(buf_tag, map, BUS_DMASYNC_PREWRITE);
3409 		DBG_COUNTER_INC(tx_encap);
3410 		MPASS(pi.ipi_new_pidx < txq->ift_size);
3411 
3412 		ndesc = pi.ipi_new_pidx - pi.ipi_pidx;
3413 		if (pi.ipi_new_pidx < pi.ipi_pidx) {
3414 			ndesc += txq->ift_size;
3415 			txq->ift_gen = 1;
3416 		}
3417 		/*
3418 		 * drivers can need as many as
3419 		 * two sentinels
3420 		 */
3421 		MPASS(ndesc <= pi.ipi_nsegs + 2);
3422 		MPASS(pi.ipi_new_pidx != pidx);
3423 		MPASS(ndesc > 0);
3424 		txq->ift_in_use += ndesc;
3425 
3426 		/*
3427 		 * We update the last software descriptor again here because there may
3428 		 * be a sentinel and/or there may be more mbufs than segments
3429 		 */
3430 		txq->ift_pidx = pi.ipi_new_pidx;
3431 		txq->ift_npending += pi.ipi_ndescs;
3432 	} else {
3433 		*m_headp = m_head = iflib_remove_mbuf(txq);
3434 		if (err == EFBIG) {
3435 			txq->ift_txd_encap_efbig++;
3436 			if (remap < 2) {
3437 				remap = 1;
3438 				goto defrag;
3439 			}
3440 		}
3441 		goto defrag_failed;
3442 	}
3443 	/*
3444 	 * err can't possibly be non-zero here, so we don't neet to test it
3445 	 * to see if we need to DBG_COUNTER_INC(encap_txd_encap_fail).
3446 	 */
3447 	return (err);
3448 
3449 defrag_failed:
3450 	txq->ift_mbuf_defrag_failed++;
3451 	txq->ift_map_failed++;
3452 	m_freem(*m_headp);
3453 	DBG_COUNTER_INC(tx_frees);
3454 	*m_headp = NULL;
3455 	DBG_COUNTER_INC(encap_txd_encap_fail);
3456 	return (ENOMEM);
3457 }
3458 
3459 static void
3460 iflib_tx_desc_free(iflib_txq_t txq, int n)
3461 {
3462 	uint32_t qsize, cidx, mask, gen;
3463 	struct mbuf *m, **ifsd_m;
3464 	bool do_prefetch;
3465 
3466 	cidx = txq->ift_cidx;
3467 	gen = txq->ift_gen;
3468 	qsize = txq->ift_size;
3469 	mask = qsize-1;
3470 	ifsd_m = txq->ift_sds.ifsd_m;
3471 	do_prefetch = (txq->ift_ctx->ifc_flags & IFC_PREFETCH);
3472 
3473 	while (n-- > 0) {
3474 		if (do_prefetch) {
3475 			prefetch(ifsd_m[(cidx + 3) & mask]);
3476 			prefetch(ifsd_m[(cidx + 4) & mask]);
3477 		}
3478 		if ((m = ifsd_m[cidx]) != NULL) {
3479 			prefetch(&ifsd_m[(cidx + CACHE_PTR_INCREMENT) & mask]);
3480 			if (m->m_pkthdr.csum_flags & CSUM_TSO) {
3481 				bus_dmamap_sync(txq->ift_tso_buf_tag,
3482 				    txq->ift_sds.ifsd_tso_map[cidx],
3483 				    BUS_DMASYNC_POSTWRITE);
3484 				bus_dmamap_unload(txq->ift_tso_buf_tag,
3485 				    txq->ift_sds.ifsd_tso_map[cidx]);
3486 			} else {
3487 				bus_dmamap_sync(txq->ift_buf_tag,
3488 				    txq->ift_sds.ifsd_map[cidx],
3489 				    BUS_DMASYNC_POSTWRITE);
3490 				bus_dmamap_unload(txq->ift_buf_tag,
3491 				    txq->ift_sds.ifsd_map[cidx]);
3492 			}
3493 			/* XXX we don't support any drivers that batch packets yet */
3494 			MPASS(m->m_nextpkt == NULL);
3495 			m_freem(m);
3496 			ifsd_m[cidx] = NULL;
3497 #if MEMORY_LOGGING
3498 			txq->ift_dequeued++;
3499 #endif
3500 			DBG_COUNTER_INC(tx_frees);
3501 		}
3502 		if (__predict_false(++cidx == qsize)) {
3503 			cidx = 0;
3504 			gen = 0;
3505 		}
3506 	}
3507 	txq->ift_cidx = cidx;
3508 	txq->ift_gen = gen;
3509 }
3510 
3511 static __inline int
3512 iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh)
3513 {
3514 	int reclaim;
3515 	if_ctx_t ctx = txq->ift_ctx;
3516 
3517 	KASSERT(thresh >= 0, ("invalid threshold to reclaim"));
3518 	MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size);
3519 
3520 	/*
3521 	 * Need a rate-limiting check so that this isn't called every time
3522 	 */
3523 	iflib_tx_credits_update(ctx, txq);
3524 	reclaim = DESC_RECLAIMABLE(txq);
3525 
3526 	if (reclaim <= thresh /* + MAX_TX_DESC(txq->ift_ctx) */) {
3527 #ifdef INVARIANTS
3528 		if (iflib_verbose_debug) {
3529 			printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __FUNCTION__,
3530 			       txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments,
3531 			       reclaim, thresh);
3532 
3533 		}
3534 #endif
3535 		return (0);
3536 	}
3537 	iflib_tx_desc_free(txq, reclaim);
3538 	txq->ift_cleaned += reclaim;
3539 	txq->ift_in_use -= reclaim;
3540 
3541 	return (reclaim);
3542 }
3543 
3544 static struct mbuf **
3545 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset, int remaining)
3546 {
3547 	int next, size;
3548 	struct mbuf **items;
3549 
3550 	size = r->size;
3551 	next = (cidx + CACHE_PTR_INCREMENT) & (size-1);
3552 	items = __DEVOLATILE(struct mbuf **, &r->items[0]);
3553 
3554 	prefetch(items[(cidx + offset) & (size-1)]);
3555 	if (remaining > 1) {
3556 		prefetch2cachelines(&items[next]);
3557 		prefetch2cachelines(items[(cidx + offset + 1) & (size-1)]);
3558 		prefetch2cachelines(items[(cidx + offset + 2) & (size-1)]);
3559 		prefetch2cachelines(items[(cidx + offset + 3) & (size-1)]);
3560 	}
3561 	return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (size-1)]));
3562 }
3563 
3564 static void
3565 iflib_txq_check_drain(iflib_txq_t txq, int budget)
3566 {
3567 
3568 	ifmp_ring_check_drainage(txq->ift_br, budget);
3569 }
3570 
3571 static uint32_t
3572 iflib_txq_can_drain(struct ifmp_ring *r)
3573 {
3574 	iflib_txq_t txq = r->cookie;
3575 	if_ctx_t ctx = txq->ift_ctx;
3576 
3577 	if (TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2)
3578 		return (1);
3579 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3580 	    BUS_DMASYNC_POSTREAD);
3581 	return (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id,
3582 	    false));
3583 }
3584 
3585 static uint32_t
3586 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3587 {
3588 	iflib_txq_t txq = r->cookie;
3589 	if_ctx_t ctx = txq->ift_ctx;
3590 	if_t ifp = ctx->ifc_ifp;
3591 	struct mbuf *m, **mp;
3592 	int avail, bytes_sent, consumed, count, err, i, in_use_prev;
3593 	int mcast_sent, pkt_sent, reclaimed, txq_avail;
3594 	bool do_prefetch, rang, ring;
3595 
3596 	if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) ||
3597 			    !LINK_ACTIVE(ctx))) {
3598 		DBG_COUNTER_INC(txq_drain_notready);
3599 		return (0);
3600 	}
3601 	reclaimed = iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
3602 	rang = iflib_txd_db_check(ctx, txq, reclaimed, txq->ift_in_use);
3603 	avail = IDXDIFF(pidx, cidx, r->size);
3604 	if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) {
3605 		DBG_COUNTER_INC(txq_drain_flushing);
3606 		for (i = 0; i < avail; i++) {
3607 			if (__predict_true(r->items[(cidx + i) & (r->size-1)] != (void *)txq))
3608 				m_free(r->items[(cidx + i) & (r->size-1)]);
3609 			r->items[(cidx + i) & (r->size-1)] = NULL;
3610 		}
3611 		return (avail);
3612 	}
3613 
3614 	if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3615 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3616 		CALLOUT_LOCK(txq);
3617 		callout_stop(&txq->ift_timer);
3618 		CALLOUT_UNLOCK(txq);
3619 		DBG_COUNTER_INC(txq_drain_oactive);
3620 		return (0);
3621 	}
3622 	if (reclaimed)
3623 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3624 	consumed = mcast_sent = bytes_sent = pkt_sent = 0;
3625 	count = MIN(avail, TX_BATCH_SIZE);
3626 #ifdef INVARIANTS
3627 	if (iflib_verbose_debug)
3628 		printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __FUNCTION__,
3629 		       avail, ctx->ifc_flags, TXQ_AVAIL(txq));
3630 #endif
3631 	do_prefetch = (ctx->ifc_flags & IFC_PREFETCH);
3632 	txq_avail = TXQ_AVAIL(txq);
3633 	err = 0;
3634 	for (i = 0; i < count && txq_avail > MAX_TX_DESC(ctx) + 2; i++) {
3635 		int rem = do_prefetch ? count - i : 0;
3636 
3637 		mp = _ring_peek_one(r, cidx, i, rem);
3638 		MPASS(mp != NULL && *mp != NULL);
3639 		if (__predict_false(*mp == (struct mbuf *)txq)) {
3640 			consumed++;
3641 			continue;
3642 		}
3643 		in_use_prev = txq->ift_in_use;
3644 		err = iflib_encap(txq, mp);
3645 		if (__predict_false(err)) {
3646 			/* no room - bail out */
3647 			if (err == ENOBUFS)
3648 				break;
3649 			consumed++;
3650 			/* we can't send this packet - skip it */
3651 			continue;
3652 		}
3653 		consumed++;
3654 		pkt_sent++;
3655 		m = *mp;
3656 		DBG_COUNTER_INC(tx_sent);
3657 		bytes_sent += m->m_pkthdr.len;
3658 		mcast_sent += !!(m->m_flags & M_MCAST);
3659 		txq_avail = TXQ_AVAIL(txq);
3660 
3661 		txq->ift_db_pending += (txq->ift_in_use - in_use_prev);
3662 		ETHER_BPF_MTAP(ifp, m);
3663 		if (__predict_false(!(ifp->if_drv_flags & IFF_DRV_RUNNING)))
3664 			break;
3665 		rang = iflib_txd_db_check(ctx, txq, false, in_use_prev);
3666 	}
3667 
3668 	/* deliberate use of bitwise or to avoid gratuitous short-circuit */
3669 	ring = rang ? false  : (iflib_min_tx_latency | err) || (TXQ_AVAIL(txq) < MAX_TX_DESC(ctx));
3670 	iflib_txd_db_check(ctx, txq, ring, txq->ift_in_use);
3671 	if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
3672 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
3673 	if (mcast_sent)
3674 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
3675 #ifdef INVARIANTS
3676 	if (iflib_verbose_debug)
3677 		printf("consumed=%d\n", consumed);
3678 #endif
3679 	return (consumed);
3680 }
3681 
3682 static uint32_t
3683 iflib_txq_drain_always(struct ifmp_ring *r)
3684 {
3685 	return (1);
3686 }
3687 
3688 static uint32_t
3689 iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3690 {
3691 	int i, avail;
3692 	struct mbuf **mp;
3693 	iflib_txq_t txq;
3694 
3695 	txq = r->cookie;
3696 
3697 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3698 	CALLOUT_LOCK(txq);
3699 	callout_stop(&txq->ift_timer);
3700 	CALLOUT_UNLOCK(txq);
3701 
3702 	avail = IDXDIFF(pidx, cidx, r->size);
3703 	for (i = 0; i < avail; i++) {
3704 		mp = _ring_peek_one(r, cidx, i, avail - i);
3705 		if (__predict_false(*mp == (struct mbuf *)txq))
3706 			continue;
3707 		m_freem(*mp);
3708 		DBG_COUNTER_INC(tx_frees);
3709 	}
3710 	MPASS(ifmp_ring_is_stalled(r) == 0);
3711 	return (avail);
3712 }
3713 
3714 static void
3715 iflib_ifmp_purge(iflib_txq_t txq)
3716 {
3717 	struct ifmp_ring *r;
3718 
3719 	r = txq->ift_br;
3720 	r->drain = iflib_txq_drain_free;
3721 	r->can_drain = iflib_txq_drain_always;
3722 
3723 	ifmp_ring_check_drainage(r, r->size);
3724 
3725 	r->drain = iflib_txq_drain;
3726 	r->can_drain = iflib_txq_can_drain;
3727 }
3728 
3729 static void
3730 _task_fn_tx(void *context)
3731 {
3732 	iflib_txq_t txq = context;
3733 	if_ctx_t ctx = txq->ift_ctx;
3734 #if defined(ALTQ) || defined(DEV_NETMAP)
3735 	if_t ifp = ctx->ifc_ifp;
3736 #endif
3737 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
3738 
3739 #ifdef IFLIB_DIAGNOSTICS
3740 	txq->ift_cpu_exec_count[curcpu]++;
3741 #endif
3742 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3743 		return;
3744 #ifdef DEV_NETMAP
3745 	if (if_getcapenable(ifp) & IFCAP_NETMAP) {
3746 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3747 		    BUS_DMASYNC_POSTREAD);
3748 		if (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false))
3749 			netmap_tx_irq(ifp, txq->ift_id);
3750 		if (ctx->ifc_flags & IFC_LEGACY)
3751 			IFDI_INTR_ENABLE(ctx);
3752 		else
3753 			IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3754 		return;
3755 	}
3756 #endif
3757 #ifdef ALTQ
3758 	if (ALTQ_IS_ENABLED(&ifp->if_snd))
3759 		iflib_altq_if_start(ifp);
3760 #endif
3761 	if (txq->ift_db_pending)
3762 		ifmp_ring_enqueue(txq->ift_br, (void **)&txq, 1, TX_BATCH_SIZE, abdicate);
3763 	else if (!abdicate)
3764 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3765 	/*
3766 	 * When abdicating, we always need to check drainage, not just when we don't enqueue
3767 	 */
3768 	if (abdicate)
3769 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3770 	if (ctx->ifc_flags & IFC_LEGACY)
3771 		IFDI_INTR_ENABLE(ctx);
3772 	else
3773 		IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3774 }
3775 
3776 static void
3777 _task_fn_rx(void *context)
3778 {
3779 	iflib_rxq_t rxq = context;
3780 	if_ctx_t ctx = rxq->ifr_ctx;
3781 	bool more;
3782 	uint16_t budget;
3783 
3784 #ifdef IFLIB_DIAGNOSTICS
3785 	rxq->ifr_cpu_exec_count[curcpu]++;
3786 #endif
3787 	DBG_COUNTER_INC(task_fn_rxs);
3788 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3789 		return;
3790 	more = true;
3791 #ifdef DEV_NETMAP
3792 	if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP) {
3793 		u_int work = 0;
3794 		if (netmap_rx_irq(ctx->ifc_ifp, rxq->ifr_id, &work)) {
3795 			more = false;
3796 		}
3797 	}
3798 #endif
3799 	budget = ctx->ifc_sysctl_rx_budget;
3800 	if (budget == 0)
3801 		budget = 16;	/* XXX */
3802 	if (more == false || (more = iflib_rxeof(rxq, budget)) == false) {
3803 		if (ctx->ifc_flags & IFC_LEGACY)
3804 			IFDI_INTR_ENABLE(ctx);
3805 		else
3806 			IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
3807 		DBG_COUNTER_INC(rx_intr_enables);
3808 	}
3809 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3810 		return;
3811 	if (more)
3812 		GROUPTASK_ENQUEUE(&rxq->ifr_task);
3813 }
3814 
3815 static void
3816 _task_fn_admin(void *context)
3817 {
3818 	if_ctx_t ctx = context;
3819 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
3820 	iflib_txq_t txq;
3821 	int i;
3822 	bool oactive, running, do_reset, do_watchdog, in_detach;
3823 	uint32_t reset_on = hz / 2;
3824 
3825 	STATE_LOCK(ctx);
3826 	running = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING);
3827 	oactive = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE);
3828 	do_reset = (ctx->ifc_flags & IFC_DO_RESET);
3829 	do_watchdog = (ctx->ifc_flags & IFC_DO_WATCHDOG);
3830 	in_detach = (ctx->ifc_flags & IFC_IN_DETACH);
3831 	ctx->ifc_flags &= ~(IFC_DO_RESET|IFC_DO_WATCHDOG);
3832 	STATE_UNLOCK(ctx);
3833 
3834 	if ((!running && !oactive) && !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
3835 		return;
3836 	if (in_detach)
3837 		return;
3838 
3839 	CTX_LOCK(ctx);
3840 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3841 		CALLOUT_LOCK(txq);
3842 		callout_stop(&txq->ift_timer);
3843 		CALLOUT_UNLOCK(txq);
3844 	}
3845 	if (do_watchdog) {
3846 		ctx->ifc_watchdog_events++;
3847 		IFDI_WATCHDOG_RESET(ctx);
3848 	}
3849 	IFDI_UPDATE_ADMIN_STATUS(ctx);
3850 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3851 #ifdef DEV_NETMAP
3852 		reset_on = hz / 2;
3853 		if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP)
3854 			iflib_netmap_timer_adjust(ctx, txq, &reset_on);
3855 #endif
3856 		callout_reset_on(&txq->ift_timer, reset_on, iflib_timer, txq, txq->ift_timer.c_cpu);
3857 	}
3858 	IFDI_LINK_INTR_ENABLE(ctx);
3859 	if (do_reset)
3860 		iflib_if_init_locked(ctx);
3861 	CTX_UNLOCK(ctx);
3862 
3863 	if (LINK_ACTIVE(ctx) == 0)
3864 		return;
3865 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3866 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
3867 }
3868 
3869 
3870 static void
3871 _task_fn_iov(void *context)
3872 {
3873 	if_ctx_t ctx = context;
3874 
3875 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING) &&
3876 	    !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
3877 		return;
3878 
3879 	CTX_LOCK(ctx);
3880 	IFDI_VFLR_HANDLE(ctx);
3881 	CTX_UNLOCK(ctx);
3882 }
3883 
3884 static int
3885 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)
3886 {
3887 	int err;
3888 	if_int_delay_info_t info;
3889 	if_ctx_t ctx;
3890 
3891 	info = (if_int_delay_info_t)arg1;
3892 	ctx = info->iidi_ctx;
3893 	info->iidi_req = req;
3894 	info->iidi_oidp = oidp;
3895 	CTX_LOCK(ctx);
3896 	err = IFDI_SYSCTL_INT_DELAY(ctx, info);
3897 	CTX_UNLOCK(ctx);
3898 	return (err);
3899 }
3900 
3901 /*********************************************************************
3902  *
3903  *  IFNET FUNCTIONS
3904  *
3905  **********************************************************************/
3906 
3907 static void
3908 iflib_if_init_locked(if_ctx_t ctx)
3909 {
3910 	iflib_stop(ctx);
3911 	iflib_init_locked(ctx);
3912 }
3913 
3914 
3915 static void
3916 iflib_if_init(void *arg)
3917 {
3918 	if_ctx_t ctx = arg;
3919 
3920 	CTX_LOCK(ctx);
3921 	iflib_if_init_locked(ctx);
3922 	CTX_UNLOCK(ctx);
3923 }
3924 
3925 static int
3926 iflib_if_transmit(if_t ifp, struct mbuf *m)
3927 {
3928 	if_ctx_t	ctx = if_getsoftc(ifp);
3929 
3930 	iflib_txq_t txq;
3931 	int err, qidx;
3932 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
3933 
3934 	if (__predict_false((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) {
3935 		DBG_COUNTER_INC(tx_frees);
3936 		m_freem(m);
3937 		return (ENETDOWN);
3938 	}
3939 
3940 	MPASS(m->m_nextpkt == NULL);
3941 	/* ALTQ-enabled interfaces always use queue 0. */
3942 	qidx = 0;
3943 	if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m) && !ALTQ_IS_ENABLED(&ifp->if_snd))
3944 		qidx = QIDX(ctx, m);
3945 	/*
3946 	 * XXX calculate buf_ring based on flowid (divvy up bits?)
3947 	 */
3948 	txq = &ctx->ifc_txqs[qidx];
3949 
3950 #ifdef DRIVER_BACKPRESSURE
3951 	if (txq->ift_closed) {
3952 		while (m != NULL) {
3953 			next = m->m_nextpkt;
3954 			m->m_nextpkt = NULL;
3955 			m_freem(m);
3956 			DBG_COUNTER_INC(tx_frees);
3957 			m = next;
3958 		}
3959 		return (ENOBUFS);
3960 	}
3961 #endif
3962 #ifdef notyet
3963 	qidx = count = 0;
3964 	mp = marr;
3965 	next = m;
3966 	do {
3967 		count++;
3968 		next = next->m_nextpkt;
3969 	} while (next != NULL);
3970 
3971 	if (count > nitems(marr))
3972 		if ((mp = malloc(count*sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) {
3973 			/* XXX check nextpkt */
3974 			m_freem(m);
3975 			/* XXX simplify for now */
3976 			DBG_COUNTER_INC(tx_frees);
3977 			return (ENOBUFS);
3978 		}
3979 	for (next = m, i = 0; next != NULL; i++) {
3980 		mp[i] = next;
3981 		next = next->m_nextpkt;
3982 		mp[i]->m_nextpkt = NULL;
3983 	}
3984 #endif
3985 	DBG_COUNTER_INC(tx_seen);
3986 	err = ifmp_ring_enqueue(txq->ift_br, (void **)&m, 1, TX_BATCH_SIZE, abdicate);
3987 
3988 	if (abdicate)
3989 		GROUPTASK_ENQUEUE(&txq->ift_task);
3990  	if (err) {
3991 		if (!abdicate)
3992 			GROUPTASK_ENQUEUE(&txq->ift_task);
3993 		/* support forthcoming later */
3994 #ifdef DRIVER_BACKPRESSURE
3995 		txq->ift_closed = TRUE;
3996 #endif
3997 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3998 		m_freem(m);
3999 		DBG_COUNTER_INC(tx_frees);
4000 	}
4001 
4002 	return (err);
4003 }
4004 
4005 #ifdef ALTQ
4006 /*
4007  * The overall approach to integrating iflib with ALTQ is to continue to use
4008  * the iflib mp_ring machinery between the ALTQ queue(s) and the hardware
4009  * ring.  Technically, when using ALTQ, queueing to an intermediate mp_ring
4010  * is redundant/unnecessary, but doing so minimizes the amount of
4011  * ALTQ-specific code required in iflib.  It is assumed that the overhead of
4012  * redundantly queueing to an intermediate mp_ring is swamped by the
4013  * performance limitations inherent in using ALTQ.
4014  *
4015  * When ALTQ support is compiled in, all iflib drivers will use a transmit
4016  * routine, iflib_altq_if_transmit(), that checks if ALTQ is enabled for the
4017  * given interface.  If ALTQ is enabled for an interface, then all
4018  * transmitted packets for that interface will be submitted to the ALTQ
4019  * subsystem via IFQ_ENQUEUE().  We don't use the legacy if_transmit()
4020  * implementation because it uses IFQ_HANDOFF(), which will duplicatively
4021  * update stats that the iflib machinery handles, and which is sensitve to
4022  * the disused IFF_DRV_OACTIVE flag.  Additionally, iflib_altq_if_start()
4023  * will be installed as the start routine for use by ALTQ facilities that
4024  * need to trigger queue drains on a scheduled basis.
4025  *
4026  */
4027 static void
4028 iflib_altq_if_start(if_t ifp)
4029 {
4030 	struct ifaltq *ifq = &ifp->if_snd;
4031 	struct mbuf *m;
4032 
4033 	IFQ_LOCK(ifq);
4034 	IFQ_DEQUEUE_NOLOCK(ifq, m);
4035 	while (m != NULL) {
4036 		iflib_if_transmit(ifp, m);
4037 		IFQ_DEQUEUE_NOLOCK(ifq, m);
4038 	}
4039 	IFQ_UNLOCK(ifq);
4040 }
4041 
4042 static int
4043 iflib_altq_if_transmit(if_t ifp, struct mbuf *m)
4044 {
4045 	int err;
4046 
4047 	if (ALTQ_IS_ENABLED(&ifp->if_snd)) {
4048 		IFQ_ENQUEUE(&ifp->if_snd, m, err);
4049 		if (err == 0)
4050 			iflib_altq_if_start(ifp);
4051 	} else
4052 		err = iflib_if_transmit(ifp, m);
4053 
4054 	return (err);
4055 }
4056 #endif /* ALTQ */
4057 
4058 static void
4059 iflib_if_qflush(if_t ifp)
4060 {
4061 	if_ctx_t ctx = if_getsoftc(ifp);
4062 	iflib_txq_t txq = ctx->ifc_txqs;
4063 	int i;
4064 
4065 	STATE_LOCK(ctx);
4066 	ctx->ifc_flags |= IFC_QFLUSH;
4067 	STATE_UNLOCK(ctx);
4068 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
4069 		while (!(ifmp_ring_is_idle(txq->ift_br) || ifmp_ring_is_stalled(txq->ift_br)))
4070 			iflib_txq_check_drain(txq, 0);
4071 	STATE_LOCK(ctx);
4072 	ctx->ifc_flags &= ~IFC_QFLUSH;
4073 	STATE_UNLOCK(ctx);
4074 
4075 	/*
4076 	 * When ALTQ is enabled, this will also take care of purging the
4077 	 * ALTQ queue(s).
4078 	 */
4079 	if_qflush(ifp);
4080 }
4081 
4082 
4083 #define IFCAP_FLAGS (IFCAP_HWCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \
4084 		     IFCAP_TSO | IFCAP_VLAN_HWTAGGING | IFCAP_HWSTATS | \
4085 		     IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | \
4086 		     IFCAP_VLAN_HWTSO | IFCAP_VLAN_HWCSUM | IFCAP_NOMAP)
4087 
4088 static int
4089 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data)
4090 {
4091 	if_ctx_t ctx = if_getsoftc(ifp);
4092 	struct ifreq	*ifr = (struct ifreq *)data;
4093 #if defined(INET) || defined(INET6)
4094 	struct ifaddr	*ifa = (struct ifaddr *)data;
4095 #endif
4096 	bool		avoid_reset = false;
4097 	int		err = 0, reinit = 0, bits;
4098 
4099 	switch (command) {
4100 	case SIOCSIFADDR:
4101 #ifdef INET
4102 		if (ifa->ifa_addr->sa_family == AF_INET)
4103 			avoid_reset = true;
4104 #endif
4105 #ifdef INET6
4106 		if (ifa->ifa_addr->sa_family == AF_INET6)
4107 			avoid_reset = true;
4108 #endif
4109 		/*
4110 		** Calling init results in link renegotiation,
4111 		** so we avoid doing it when possible.
4112 		*/
4113 		if (avoid_reset) {
4114 			if_setflagbits(ifp, IFF_UP,0);
4115 			if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
4116 				reinit = 1;
4117 #ifdef INET
4118 			if (!(if_getflags(ifp) & IFF_NOARP))
4119 				arp_ifinit(ifp, ifa);
4120 #endif
4121 		} else
4122 			err = ether_ioctl(ifp, command, data);
4123 		break;
4124 	case SIOCSIFMTU:
4125 		CTX_LOCK(ctx);
4126 		if (ifr->ifr_mtu == if_getmtu(ifp)) {
4127 			CTX_UNLOCK(ctx);
4128 			break;
4129 		}
4130 		bits = if_getdrvflags(ifp);
4131 		/* stop the driver and free any clusters before proceeding */
4132 		iflib_stop(ctx);
4133 
4134 		if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) {
4135 			STATE_LOCK(ctx);
4136 			if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size)
4137 				ctx->ifc_flags |= IFC_MULTISEG;
4138 			else
4139 				ctx->ifc_flags &= ~IFC_MULTISEG;
4140 			STATE_UNLOCK(ctx);
4141 			err = if_setmtu(ifp, ifr->ifr_mtu);
4142 		}
4143 		iflib_init_locked(ctx);
4144 		STATE_LOCK(ctx);
4145 		if_setdrvflags(ifp, bits);
4146 		STATE_UNLOCK(ctx);
4147 		CTX_UNLOCK(ctx);
4148 		break;
4149 	case SIOCSIFFLAGS:
4150 		CTX_LOCK(ctx);
4151 		if (if_getflags(ifp) & IFF_UP) {
4152 			if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4153 				if ((if_getflags(ifp) ^ ctx->ifc_if_flags) &
4154 				    (IFF_PROMISC | IFF_ALLMULTI)) {
4155 					err = IFDI_PROMISC_SET(ctx, if_getflags(ifp));
4156 				}
4157 			} else
4158 				reinit = 1;
4159 		} else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4160 			iflib_stop(ctx);
4161 		}
4162 		ctx->ifc_if_flags = if_getflags(ifp);
4163 		CTX_UNLOCK(ctx);
4164 		break;
4165 	case SIOCADDMULTI:
4166 	case SIOCDELMULTI:
4167 		if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4168 			CTX_LOCK(ctx);
4169 			IFDI_INTR_DISABLE(ctx);
4170 			IFDI_MULTI_SET(ctx);
4171 			IFDI_INTR_ENABLE(ctx);
4172 			CTX_UNLOCK(ctx);
4173 		}
4174 		break;
4175 	case SIOCSIFMEDIA:
4176 		CTX_LOCK(ctx);
4177 		IFDI_MEDIA_SET(ctx);
4178 		CTX_UNLOCK(ctx);
4179 		/* FALLTHROUGH */
4180 	case SIOCGIFMEDIA:
4181 	case SIOCGIFXMEDIA:
4182 		err = ifmedia_ioctl(ifp, ifr, ctx->ifc_mediap, command);
4183 		break;
4184 	case SIOCGI2C:
4185 	{
4186 		struct ifi2creq i2c;
4187 
4188 		err = copyin(ifr_data_get_ptr(ifr), &i2c, sizeof(i2c));
4189 		if (err != 0)
4190 			break;
4191 		if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) {
4192 			err = EINVAL;
4193 			break;
4194 		}
4195 		if (i2c.len > sizeof(i2c.data)) {
4196 			err = EINVAL;
4197 			break;
4198 		}
4199 
4200 		if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0)
4201 			err = copyout(&i2c, ifr_data_get_ptr(ifr),
4202 			    sizeof(i2c));
4203 		break;
4204 	}
4205 	case SIOCSIFCAP:
4206 	{
4207 		int mask, setmask, oldmask;
4208 
4209 		oldmask = if_getcapenable(ifp);
4210 		mask = ifr->ifr_reqcap ^ oldmask;
4211 		mask &= ctx->ifc_softc_ctx.isc_capabilities | IFCAP_NOMAP;
4212 		setmask = 0;
4213 #ifdef TCP_OFFLOAD
4214 		setmask |= mask & (IFCAP_TOE4|IFCAP_TOE6);
4215 #endif
4216 		setmask |= (mask & IFCAP_FLAGS);
4217 		setmask |= (mask & IFCAP_WOL);
4218 
4219 		/*
4220 		 * If any RX csum has changed, change all the ones that
4221 		 * are supported by the driver.
4222 		 */
4223 		if (setmask & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6)) {
4224 			setmask |= ctx->ifc_softc_ctx.isc_capabilities &
4225 			    (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6);
4226 		}
4227 
4228 		/*
4229 		 * want to ensure that traffic has stopped before we change any of the flags
4230 		 */
4231 		if (setmask) {
4232 			CTX_LOCK(ctx);
4233 			bits = if_getdrvflags(ifp);
4234 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4235 				iflib_stop(ctx);
4236 			STATE_LOCK(ctx);
4237 			if_togglecapenable(ifp, setmask);
4238 			STATE_UNLOCK(ctx);
4239 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4240 				iflib_init_locked(ctx);
4241 			STATE_LOCK(ctx);
4242 			if_setdrvflags(ifp, bits);
4243 			STATE_UNLOCK(ctx);
4244 			CTX_UNLOCK(ctx);
4245 		}
4246 		if_vlancap(ifp);
4247 		break;
4248 	}
4249 	case SIOCGPRIVATE_0:
4250 	case SIOCSDRVSPEC:
4251 	case SIOCGDRVSPEC:
4252 		CTX_LOCK(ctx);
4253 		err = IFDI_PRIV_IOCTL(ctx, command, data);
4254 		CTX_UNLOCK(ctx);
4255 		break;
4256 	default:
4257 		err = ether_ioctl(ifp, command, data);
4258 		break;
4259 	}
4260 	if (reinit)
4261 		iflib_if_init(ctx);
4262 	return (err);
4263 }
4264 
4265 static uint64_t
4266 iflib_if_get_counter(if_t ifp, ift_counter cnt)
4267 {
4268 	if_ctx_t ctx = if_getsoftc(ifp);
4269 
4270 	return (IFDI_GET_COUNTER(ctx, cnt));
4271 }
4272 
4273 /*********************************************************************
4274  *
4275  *  OTHER FUNCTIONS EXPORTED TO THE STACK
4276  *
4277  **********************************************************************/
4278 
4279 static void
4280 iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag)
4281 {
4282 	if_ctx_t ctx = if_getsoftc(ifp);
4283 
4284 	if ((void *)ctx != arg)
4285 		return;
4286 
4287 	if ((vtag == 0) || (vtag > 4095))
4288 		return;
4289 
4290 	if (iflib_in_detach(ctx))
4291 		return;
4292 
4293 	CTX_LOCK(ctx);
4294 	IFDI_VLAN_REGISTER(ctx, vtag);
4295 	/* Re-init to load the changes */
4296 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
4297 		iflib_if_init_locked(ctx);
4298 	CTX_UNLOCK(ctx);
4299 }
4300 
4301 static void
4302 iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag)
4303 {
4304 	if_ctx_t ctx = if_getsoftc(ifp);
4305 
4306 	if ((void *)ctx != arg)
4307 		return;
4308 
4309 	if ((vtag == 0) || (vtag > 4095))
4310 		return;
4311 
4312 	CTX_LOCK(ctx);
4313 	IFDI_VLAN_UNREGISTER(ctx, vtag);
4314 	/* Re-init to load the changes */
4315 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
4316 		iflib_if_init_locked(ctx);
4317 	CTX_UNLOCK(ctx);
4318 }
4319 
4320 static void
4321 iflib_led_func(void *arg, int onoff)
4322 {
4323 	if_ctx_t ctx = arg;
4324 
4325 	CTX_LOCK(ctx);
4326 	IFDI_LED_FUNC(ctx, onoff);
4327 	CTX_UNLOCK(ctx);
4328 }
4329 
4330 /*********************************************************************
4331  *
4332  *  BUS FUNCTION DEFINITIONS
4333  *
4334  **********************************************************************/
4335 
4336 int
4337 iflib_device_probe(device_t dev)
4338 {
4339 	const pci_vendor_info_t *ent;
4340 	if_shared_ctx_t sctx;
4341 	uint16_t pci_device_id, pci_rev_id, pci_subdevice_id, pci_subvendor_id;
4342 	uint16_t pci_vendor_id;
4343 
4344 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4345 		return (ENOTSUP);
4346 
4347 	pci_vendor_id = pci_get_vendor(dev);
4348 	pci_device_id = pci_get_device(dev);
4349 	pci_subvendor_id = pci_get_subvendor(dev);
4350 	pci_subdevice_id = pci_get_subdevice(dev);
4351 	pci_rev_id = pci_get_revid(dev);
4352 	if (sctx->isc_parse_devinfo != NULL)
4353 		sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id);
4354 
4355 	ent = sctx->isc_vendor_info;
4356 	while (ent->pvi_vendor_id != 0) {
4357 		if (pci_vendor_id != ent->pvi_vendor_id) {
4358 			ent++;
4359 			continue;
4360 		}
4361 		if ((pci_device_id == ent->pvi_device_id) &&
4362 		    ((pci_subvendor_id == ent->pvi_subvendor_id) ||
4363 		     (ent->pvi_subvendor_id == 0)) &&
4364 		    ((pci_subdevice_id == ent->pvi_subdevice_id) ||
4365 		     (ent->pvi_subdevice_id == 0)) &&
4366 		    ((pci_rev_id == ent->pvi_rev_id) ||
4367 		     (ent->pvi_rev_id == 0))) {
4368 
4369 			device_set_desc_copy(dev, ent->pvi_name);
4370 			/* this needs to be changed to zero if the bus probing code
4371 			 * ever stops re-probing on best match because the sctx
4372 			 * may have its values over written by register calls
4373 			 * in subsequent probes
4374 			 */
4375 			return (BUS_PROBE_DEFAULT);
4376 		}
4377 		ent++;
4378 	}
4379 	return (ENXIO);
4380 }
4381 
4382 int
4383 iflib_device_probe_vendor(device_t dev)
4384 {
4385 	int probe;
4386 
4387 	probe = iflib_device_probe(dev);
4388 	if (probe == BUS_PROBE_DEFAULT)
4389 		return (BUS_PROBE_VENDOR);
4390 	else
4391 		return (probe);
4392 }
4393 
4394 static void
4395 iflib_reset_qvalues(if_ctx_t ctx)
4396 {
4397 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4398 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4399 	device_t dev = ctx->ifc_dev;
4400 	int i;
4401 
4402 	if (ctx->ifc_sysctl_ntxqs != 0)
4403 		scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs;
4404 	if (ctx->ifc_sysctl_nrxqs != 0)
4405 		scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs;
4406 
4407 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4408 		if (ctx->ifc_sysctl_ntxds[i] != 0)
4409 			scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i];
4410 		else
4411 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4412 	}
4413 
4414 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4415 		if (ctx->ifc_sysctl_nrxds[i] != 0)
4416 			scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i];
4417 		else
4418 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4419 	}
4420 
4421 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4422 		if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) {
4423 			device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n",
4424 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]);
4425 			scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i];
4426 		}
4427 		if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) {
4428 			device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n",
4429 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]);
4430 			scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i];
4431 		}
4432 		if (!powerof2(scctx->isc_nrxd[i])) {
4433 			device_printf(dev, "nrxd%d: %d is not a power of 2 - using default value of %d\n",
4434 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_default[i]);
4435 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4436 		}
4437 	}
4438 
4439 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4440 		if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) {
4441 			device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n",
4442 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]);
4443 			scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i];
4444 		}
4445 		if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) {
4446 			device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n",
4447 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]);
4448 			scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i];
4449 		}
4450 		if (!powerof2(scctx->isc_ntxd[i])) {
4451 			device_printf(dev, "ntxd%d: %d is not a power of 2 - using default value of %d\n",
4452 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_default[i]);
4453 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4454 		}
4455 	}
4456 }
4457 
4458 static void
4459 iflib_add_pfil(if_ctx_t ctx)
4460 {
4461 	struct pfil_head *pfil;
4462 	struct pfil_head_args pa;
4463 	iflib_rxq_t rxq;
4464 	int i;
4465 
4466 	pa.pa_version = PFIL_VERSION;
4467 	pa.pa_flags = PFIL_IN;
4468 	pa.pa_type = PFIL_TYPE_ETHERNET;
4469 	pa.pa_headname = ctx->ifc_ifp->if_xname;
4470 	pfil = pfil_head_register(&pa);
4471 
4472 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4473 		rxq->pfil = pfil;
4474 	}
4475 }
4476 
4477 static void
4478 iflib_rem_pfil(if_ctx_t ctx)
4479 {
4480 	struct pfil_head *pfil;
4481 	iflib_rxq_t rxq;
4482 	int i;
4483 
4484 	rxq = ctx->ifc_rxqs;
4485 	pfil = rxq->pfil;
4486 	for (i = 0; i < NRXQSETS(ctx); i++, rxq++) {
4487 		rxq->pfil = NULL;
4488 	}
4489 	pfil_head_unregister(pfil);
4490 }
4491 
4492 static uint16_t
4493 get_ctx_core_offset(if_ctx_t ctx)
4494 {
4495 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4496 	struct cpu_offset *op;
4497 	uint16_t qc;
4498 	uint16_t ret = ctx->ifc_sysctl_core_offset;
4499 
4500 	if (ret != CORE_OFFSET_UNSPECIFIED)
4501 		return (ret);
4502 
4503 	if (ctx->ifc_sysctl_separate_txrx)
4504 		qc = scctx->isc_ntxqsets + scctx->isc_nrxqsets;
4505 	else
4506 		qc = max(scctx->isc_ntxqsets, scctx->isc_nrxqsets);
4507 
4508 	mtx_lock(&cpu_offset_mtx);
4509 	SLIST_FOREACH(op, &cpu_offsets, entries) {
4510 		if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) {
4511 			ret = op->offset;
4512 			op->offset += qc;
4513 			MPASS(op->refcount < UINT_MAX);
4514 			op->refcount++;
4515 			break;
4516 		}
4517 	}
4518 	if (ret == CORE_OFFSET_UNSPECIFIED) {
4519 		ret = 0;
4520 		op = malloc(sizeof(struct cpu_offset), M_IFLIB,
4521 		    M_NOWAIT | M_ZERO);
4522 		if (op == NULL) {
4523 			device_printf(ctx->ifc_dev,
4524 			    "allocation for cpu offset failed.\n");
4525 		} else {
4526 			op->offset = qc;
4527 			op->refcount = 1;
4528 			CPU_COPY(&ctx->ifc_cpus, &op->set);
4529 			SLIST_INSERT_HEAD(&cpu_offsets, op, entries);
4530 		}
4531 	}
4532 	mtx_unlock(&cpu_offset_mtx);
4533 
4534 	return (ret);
4535 }
4536 
4537 static void
4538 unref_ctx_core_offset(if_ctx_t ctx)
4539 {
4540 	struct cpu_offset *op, *top;
4541 
4542 	mtx_lock(&cpu_offset_mtx);
4543 	SLIST_FOREACH_SAFE(op, &cpu_offsets, entries, top) {
4544 		if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) {
4545 			MPASS(op->refcount > 0);
4546 			op->refcount--;
4547 			if (op->refcount == 0) {
4548 				SLIST_REMOVE(&cpu_offsets, op, cpu_offset, entries);
4549 				free(op, M_IFLIB);
4550 			}
4551 			break;
4552 		}
4553 	}
4554 	mtx_unlock(&cpu_offset_mtx);
4555 }
4556 
4557 int
4558 iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp)
4559 {
4560 	if_ctx_t ctx;
4561 	if_t ifp;
4562 	if_softc_ctx_t scctx;
4563 	kobjop_desc_t kobj_desc;
4564 	kobj_method_t *kobj_method;
4565 	int err, msix, rid;
4566 	uint16_t main_rxq, main_txq;
4567 
4568 	ctx = malloc(sizeof(* ctx), M_IFLIB, M_WAITOK|M_ZERO);
4569 
4570 	if (sc == NULL) {
4571 		sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4572 		device_set_softc(dev, ctx);
4573 		ctx->ifc_flags |= IFC_SC_ALLOCATED;
4574 	}
4575 
4576 	ctx->ifc_sctx = sctx;
4577 	ctx->ifc_dev = dev;
4578 	ctx->ifc_softc = sc;
4579 
4580 	if ((err = iflib_register(ctx)) != 0) {
4581 		device_printf(dev, "iflib_register failed %d\n", err);
4582 		goto fail_ctx_free;
4583 	}
4584 	iflib_add_device_sysctl_pre(ctx);
4585 
4586 	scctx = &ctx->ifc_softc_ctx;
4587 	ifp = ctx->ifc_ifp;
4588 
4589 	iflib_reset_qvalues(ctx);
4590 	CTX_LOCK(ctx);
4591 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4592 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4593 		goto fail_unlock;
4594 	}
4595 	_iflib_pre_assert(scctx);
4596 	ctx->ifc_txrx = *scctx->isc_txrx;
4597 
4598 	if (sctx->isc_flags & IFLIB_DRIVER_MEDIA)
4599 		ctx->ifc_mediap = scctx->isc_media;
4600 
4601 #ifdef INVARIANTS
4602 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
4603 		MPASS(scctx->isc_tx_csum_flags);
4604 #endif
4605 
4606 	if_setcapabilities(ifp,
4607 	    scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_NOMAP);
4608 	if_setcapenable(ifp,
4609 	    scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_NOMAP);
4610 
4611 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4612 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4613 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4614 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4615 
4616 	main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4617 	main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4618 
4619 	/* XXX change for per-queue sizes */
4620 	device_printf(dev, "Using %d TX descriptors and %d RX descriptors\n",
4621 	    scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4622 
4623 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4624 	    MAX_SINGLE_PACKET_FRACTION)
4625 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4626 		    MAX_SINGLE_PACKET_FRACTION);
4627 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4628 	    MAX_SINGLE_PACKET_FRACTION)
4629 		scctx->isc_tx_tso_segments_max = max(1,
4630 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4631 
4632 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4633 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
4634 		/*
4635 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
4636 		 * but some MACs do.
4637 		 */
4638 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
4639 		    IP_MAXPACKET));
4640 		/*
4641 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
4642 		 * into account.  In the worst case, each of these calls will
4643 		 * add another mbuf and, thus, the requirement for another DMA
4644 		 * segment.  So for best performance, it doesn't make sense to
4645 		 * advertize a maximum of TSO segments that typically will
4646 		 * require defragmentation in iflib_encap().
4647 		 */
4648 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
4649 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
4650 	}
4651 	if (scctx->isc_rss_table_size == 0)
4652 		scctx->isc_rss_table_size = 64;
4653 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4654 
4655 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4656 	/* XXX format name */
4657 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx,
4658 	    NULL, NULL, "admin");
4659 
4660 	/* Set up cpu set.  If it fails, use the set of all CPUs. */
4661 	if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) != 0) {
4662 		device_printf(dev, "Unable to fetch CPU list\n");
4663 		CPU_COPY(&all_cpus, &ctx->ifc_cpus);
4664 	}
4665 	MPASS(CPU_COUNT(&ctx->ifc_cpus) > 0);
4666 
4667 	/*
4668 	** Now set up MSI or MSI-X, should return us the number of supported
4669 	** vectors (will be 1 for a legacy interrupt and MSI).
4670 	*/
4671 	if (sctx->isc_flags & IFLIB_SKIP_MSIX) {
4672 		msix = scctx->isc_vectors;
4673 	} else if (scctx->isc_msix_bar != 0)
4674 	       /*
4675 		* The simple fact that isc_msix_bar is not 0 does not mean we
4676 		* we have a good value there that is known to work.
4677 		*/
4678 		msix = iflib_msix_init(ctx);
4679 	else {
4680 		scctx->isc_vectors = 1;
4681 		scctx->isc_ntxqsets = 1;
4682 		scctx->isc_nrxqsets = 1;
4683 		scctx->isc_intr = IFLIB_INTR_LEGACY;
4684 		msix = 0;
4685 	}
4686 	/* Get memory for the station queues */
4687 	if ((err = iflib_queues_alloc(ctx))) {
4688 		device_printf(dev, "Unable to allocate queue memory\n");
4689 		goto fail_intr_free;
4690 	}
4691 
4692 	if ((err = iflib_qset_structures_setup(ctx)))
4693 		goto fail_queues;
4694 
4695 	/*
4696 	 * Now that we know how many queues there are, get the core offset.
4697 	 */
4698 	ctx->ifc_sysctl_core_offset = get_ctx_core_offset(ctx);
4699 
4700 	/*
4701 	 * Group taskqueues aren't properly set up until SMP is started,
4702 	 * so we disable interrupts until we can handle them post
4703 	 * SI_SUB_SMP.
4704 	 *
4705 	 * XXX: disabling interrupts doesn't actually work, at least for
4706 	 * the non-MSI case.  When they occur before SI_SUB_SMP completes,
4707 	 * we do null handling and depend on this not causing too large an
4708 	 * interrupt storm.
4709 	 */
4710 	IFDI_INTR_DISABLE(ctx);
4711 
4712 	if (msix > 1) {
4713 		/*
4714 		 * When using MSI-X, ensure that ifdi_{r,t}x_queue_intr_enable
4715 		 * aren't the default NULL implementation.
4716 		 */
4717 		kobj_desc = &ifdi_rx_queue_intr_enable_desc;
4718 		kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL,
4719 		    kobj_desc);
4720 		if (kobj_method == &kobj_desc->deflt) {
4721 			device_printf(dev,
4722 			    "MSI-X requires ifdi_rx_queue_intr_enable method");
4723 			err = EOPNOTSUPP;
4724 			goto fail_queues;
4725 		}
4726 		kobj_desc = &ifdi_tx_queue_intr_enable_desc;
4727 		kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL,
4728 		    kobj_desc);
4729 		if (kobj_method == &kobj_desc->deflt) {
4730 			device_printf(dev,
4731 			    "MSI-X requires ifdi_tx_queue_intr_enable method");
4732 			err = EOPNOTSUPP;
4733 			goto fail_queues;
4734 		}
4735 
4736 		/*
4737 		 * Assign the MSI-X vectors.
4738 		 * Note that the default NULL ifdi_msix_intr_assign method will
4739 		 * fail here, too.
4740 		 */
4741 		err = IFDI_MSIX_INTR_ASSIGN(ctx, msix);
4742 		if (err != 0) {
4743 			device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n",
4744 			    err);
4745 			goto fail_queues;
4746 		}
4747 	} else if (scctx->isc_intr != IFLIB_INTR_MSIX) {
4748 		rid = 0;
4749 		if (scctx->isc_intr == IFLIB_INTR_MSI) {
4750 			MPASS(msix == 1);
4751 			rid = 1;
4752 		}
4753 		if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) {
4754 			device_printf(dev, "iflib_legacy_setup failed %d\n", err);
4755 			goto fail_queues;
4756 		}
4757 	} else {
4758 		device_printf(dev,
4759 		    "Cannot use iflib with only 1 MSI-X interrupt!\n");
4760 		err = ENODEV;
4761 		goto fail_intr_free;
4762 	}
4763 
4764 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
4765 
4766 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4767 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4768 		goto fail_detach;
4769 	}
4770 
4771 	/*
4772 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4773 	 * This must appear after the call to ether_ifattach() because
4774 	 * ether_ifattach() sets if_hdrlen to the default value.
4775 	 */
4776 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4777 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
4778 
4779 	if ((err = iflib_netmap_attach(ctx))) {
4780 		device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err);
4781 		goto fail_detach;
4782 	}
4783 	*ctxp = ctx;
4784 
4785 	DEBUGNET_SET(ctx->ifc_ifp, iflib);
4786 
4787 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4788 	iflib_add_device_sysctl_post(ctx);
4789 	iflib_add_pfil(ctx);
4790 	ctx->ifc_flags |= IFC_INIT_DONE;
4791 	CTX_UNLOCK(ctx);
4792 
4793 	return (0);
4794 
4795 fail_detach:
4796 	ether_ifdetach(ctx->ifc_ifp);
4797 fail_intr_free:
4798 	iflib_free_intr_mem(ctx);
4799 fail_queues:
4800 	iflib_tx_structures_free(ctx);
4801 	iflib_rx_structures_free(ctx);
4802 	taskqgroup_detach(qgroup_if_config_tqg, &ctx->ifc_admin_task);
4803 	IFDI_DETACH(ctx);
4804 fail_unlock:
4805 	CTX_UNLOCK(ctx);
4806 	iflib_deregister(ctx);
4807 fail_ctx_free:
4808 	device_set_softc(ctx->ifc_dev, NULL);
4809         if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4810                 free(ctx->ifc_softc, M_IFLIB);
4811         free(ctx, M_IFLIB);
4812 	return (err);
4813 }
4814 
4815 int
4816 iflib_pseudo_register(device_t dev, if_shared_ctx_t sctx, if_ctx_t *ctxp,
4817 					  struct iflib_cloneattach_ctx *clctx)
4818 {
4819 	int err;
4820 	if_ctx_t ctx;
4821 	if_t ifp;
4822 	if_softc_ctx_t scctx;
4823 	int i;
4824 	void *sc;
4825 	uint16_t main_txq;
4826 	uint16_t main_rxq;
4827 
4828 	ctx = malloc(sizeof(*ctx), M_IFLIB, M_WAITOK|M_ZERO);
4829 	sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4830 	ctx->ifc_flags |= IFC_SC_ALLOCATED;
4831 	if (sctx->isc_flags & (IFLIB_PSEUDO|IFLIB_VIRTUAL))
4832 		ctx->ifc_flags |= IFC_PSEUDO;
4833 
4834 	ctx->ifc_sctx = sctx;
4835 	ctx->ifc_softc = sc;
4836 	ctx->ifc_dev = dev;
4837 
4838 	if ((err = iflib_register(ctx)) != 0) {
4839 		device_printf(dev, "%s: iflib_register failed %d\n", __func__, err);
4840 		goto fail_ctx_free;
4841 	}
4842 	iflib_add_device_sysctl_pre(ctx);
4843 
4844 	scctx = &ctx->ifc_softc_ctx;
4845 	ifp = ctx->ifc_ifp;
4846 
4847 	iflib_reset_qvalues(ctx);
4848 	CTX_LOCK(ctx);
4849 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4850 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4851 		goto fail_unlock;
4852 	}
4853 	if (sctx->isc_flags & IFLIB_GEN_MAC)
4854 		ether_gen_addr(ifp, &ctx->ifc_mac);
4855 	if ((err = IFDI_CLONEATTACH(ctx, clctx->cc_ifc, clctx->cc_name,
4856 								clctx->cc_params)) != 0) {
4857 		device_printf(dev, "IFDI_CLONEATTACH failed %d\n", err);
4858 		goto fail_ctx_free;
4859 	}
4860 	ifmedia_add(ctx->ifc_mediap, IFM_ETHER | IFM_1000_T | IFM_FDX, 0, NULL);
4861 	ifmedia_add(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO, 0, NULL);
4862 	ifmedia_set(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO);
4863 
4864 #ifdef INVARIANTS
4865 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
4866 		MPASS(scctx->isc_tx_csum_flags);
4867 #endif
4868 
4869 	if_setcapabilities(ifp, scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_LINKSTATE);
4870 	if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_LINKSTATE);
4871 
4872 	ifp->if_flags |= IFF_NOGROUP;
4873 	if (sctx->isc_flags & IFLIB_PSEUDO) {
4874 		ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
4875 
4876 		if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4877 			device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4878 			goto fail_detach;
4879 		}
4880 		*ctxp = ctx;
4881 
4882 		/*
4883 		 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4884 		 * This must appear after the call to ether_ifattach() because
4885 		 * ether_ifattach() sets if_hdrlen to the default value.
4886 		 */
4887 		if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4888 			if_setifheaderlen(ifp,
4889 			    sizeof(struct ether_vlan_header));
4890 
4891 		if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4892 		iflib_add_device_sysctl_post(ctx);
4893 		ctx->ifc_flags |= IFC_INIT_DONE;
4894 		return (0);
4895 	}
4896 	_iflib_pre_assert(scctx);
4897 	ctx->ifc_txrx = *scctx->isc_txrx;
4898 
4899 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4900 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4901 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4902 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4903 
4904 	main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4905 	main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4906 
4907 	/* XXX change for per-queue sizes */
4908 	device_printf(dev, "Using %d TX descriptors and %d RX descriptors\n",
4909 	    scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4910 
4911 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4912 	    MAX_SINGLE_PACKET_FRACTION)
4913 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4914 		    MAX_SINGLE_PACKET_FRACTION);
4915 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4916 	    MAX_SINGLE_PACKET_FRACTION)
4917 		scctx->isc_tx_tso_segments_max = max(1,
4918 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4919 
4920 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4921 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
4922 		/*
4923 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
4924 		 * but some MACs do.
4925 		 */
4926 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
4927 		    IP_MAXPACKET));
4928 		/*
4929 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
4930 		 * into account.  In the worst case, each of these calls will
4931 		 * add another mbuf and, thus, the requirement for another DMA
4932 		 * segment.  So for best performance, it doesn't make sense to
4933 		 * advertize a maximum of TSO segments that typically will
4934 		 * require defragmentation in iflib_encap().
4935 		 */
4936 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
4937 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
4938 	}
4939 	if (scctx->isc_rss_table_size == 0)
4940 		scctx->isc_rss_table_size = 64;
4941 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4942 
4943 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4944 	/* XXX format name */
4945 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx,
4946 	    NULL, NULL, "admin");
4947 
4948 	/* XXX --- can support > 1 -- but keep it simple for now */
4949 	scctx->isc_intr = IFLIB_INTR_LEGACY;
4950 
4951 	/* Get memory for the station queues */
4952 	if ((err = iflib_queues_alloc(ctx))) {
4953 		device_printf(dev, "Unable to allocate queue memory\n");
4954 		goto fail_iflib_detach;
4955 	}
4956 
4957 	if ((err = iflib_qset_structures_setup(ctx))) {
4958 		device_printf(dev, "qset structure setup failed %d\n", err);
4959 		goto fail_queues;
4960 	}
4961 
4962 	/*
4963 	 * XXX What if anything do we want to do about interrupts?
4964 	 */
4965 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
4966 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4967 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4968 		goto fail_detach;
4969 	}
4970 
4971 	/*
4972 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4973 	 * This must appear after the call to ether_ifattach() because
4974 	 * ether_ifattach() sets if_hdrlen to the default value.
4975 	 */
4976 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4977 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
4978 
4979 	/* XXX handle more than one queue */
4980 	for (i = 0; i < scctx->isc_nrxqsets; i++)
4981 		IFDI_RX_CLSET(ctx, 0, i, ctx->ifc_rxqs[i].ifr_fl[0].ifl_sds.ifsd_cl);
4982 
4983 	*ctxp = ctx;
4984 
4985 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4986 	iflib_add_device_sysctl_post(ctx);
4987 	ctx->ifc_flags |= IFC_INIT_DONE;
4988 	CTX_UNLOCK(ctx);
4989 
4990 	return (0);
4991 fail_detach:
4992 	ether_ifdetach(ctx->ifc_ifp);
4993 fail_queues:
4994 	iflib_tx_structures_free(ctx);
4995 	iflib_rx_structures_free(ctx);
4996 fail_iflib_detach:
4997 	IFDI_DETACH(ctx);
4998 fail_unlock:
4999 	CTX_UNLOCK(ctx);
5000 	iflib_deregister(ctx);
5001 fail_ctx_free:
5002 	free(ctx->ifc_softc, M_IFLIB);
5003 	free(ctx, M_IFLIB);
5004 	return (err);
5005 }
5006 
5007 int
5008 iflib_pseudo_deregister(if_ctx_t ctx)
5009 {
5010 	if_t ifp = ctx->ifc_ifp;
5011 	iflib_txq_t txq;
5012 	iflib_rxq_t rxq;
5013 	int i, j;
5014 	struct taskqgroup *tqg;
5015 	iflib_fl_t fl;
5016 
5017 	/* Unregister VLAN event handlers early */
5018 	iflib_unregister_vlan_handlers(ctx);
5019 
5020 	ether_ifdetach(ifp);
5021 	/* XXX drain any dependent tasks */
5022 	tqg = qgroup_if_io_tqg;
5023 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
5024 		callout_drain(&txq->ift_timer);
5025 		if (txq->ift_task.gt_uniq != NULL)
5026 			taskqgroup_detach(tqg, &txq->ift_task);
5027 	}
5028 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
5029 		if (rxq->ifr_task.gt_uniq != NULL)
5030 			taskqgroup_detach(tqg, &rxq->ifr_task);
5031 
5032 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5033 			free(fl->ifl_rx_bitmap, M_IFLIB);
5034 	}
5035 	tqg = qgroup_if_config_tqg;
5036 	if (ctx->ifc_admin_task.gt_uniq != NULL)
5037 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
5038 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
5039 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
5040 
5041 	iflib_tx_structures_free(ctx);
5042 	iflib_rx_structures_free(ctx);
5043 
5044 	iflib_deregister(ctx);
5045 
5046 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5047 		free(ctx->ifc_softc, M_IFLIB);
5048 	free(ctx, M_IFLIB);
5049 	return (0);
5050 }
5051 
5052 int
5053 iflib_device_attach(device_t dev)
5054 {
5055 	if_ctx_t ctx;
5056 	if_shared_ctx_t sctx;
5057 
5058 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
5059 		return (ENOTSUP);
5060 
5061 	pci_enable_busmaster(dev);
5062 
5063 	return (iflib_device_register(dev, NULL, sctx, &ctx));
5064 }
5065 
5066 int
5067 iflib_device_deregister(if_ctx_t ctx)
5068 {
5069 	if_t ifp = ctx->ifc_ifp;
5070 	iflib_txq_t txq;
5071 	iflib_rxq_t rxq;
5072 	device_t dev = ctx->ifc_dev;
5073 	int i, j;
5074 	struct taskqgroup *tqg;
5075 	iflib_fl_t fl;
5076 
5077 	/* Make sure VLANS are not using driver */
5078 	if (if_vlantrunkinuse(ifp)) {
5079 		device_printf(dev, "Vlan in use, detach first\n");
5080 		return (EBUSY);
5081 	}
5082 #ifdef PCI_IOV
5083 	if (!CTX_IS_VF(ctx) && pci_iov_detach(dev) != 0) {
5084 		device_printf(dev, "SR-IOV in use; detach first.\n");
5085 		return (EBUSY);
5086 	}
5087 #endif
5088 
5089 	STATE_LOCK(ctx);
5090 	ctx->ifc_flags |= IFC_IN_DETACH;
5091 	STATE_UNLOCK(ctx);
5092 
5093 	/* Unregister VLAN handlers before calling iflib_stop() */
5094 	iflib_unregister_vlan_handlers(ctx);
5095 
5096 	iflib_netmap_detach(ifp);
5097 	ether_ifdetach(ifp);
5098 
5099 	CTX_LOCK(ctx);
5100 	iflib_stop(ctx);
5101 	CTX_UNLOCK(ctx);
5102 
5103 	iflib_rem_pfil(ctx);
5104 	if (ctx->ifc_led_dev != NULL)
5105 		led_destroy(ctx->ifc_led_dev);
5106 	/* XXX drain any dependent tasks */
5107 	tqg = qgroup_if_io_tqg;
5108 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
5109 		callout_drain(&txq->ift_timer);
5110 		if (txq->ift_task.gt_uniq != NULL)
5111 			taskqgroup_detach(tqg, &txq->ift_task);
5112 	}
5113 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
5114 		if (rxq->ifr_task.gt_uniq != NULL)
5115 			taskqgroup_detach(tqg, &rxq->ifr_task);
5116 
5117 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5118 			free(fl->ifl_rx_bitmap, M_IFLIB);
5119 	}
5120 	tqg = qgroup_if_config_tqg;
5121 	if (ctx->ifc_admin_task.gt_uniq != NULL)
5122 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
5123 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
5124 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
5125 	CTX_LOCK(ctx);
5126 	IFDI_DETACH(ctx);
5127 	CTX_UNLOCK(ctx);
5128 
5129 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
5130 	iflib_free_intr_mem(ctx);
5131 
5132 	bus_generic_detach(dev);
5133 
5134 	iflib_tx_structures_free(ctx);
5135 	iflib_rx_structures_free(ctx);
5136 
5137 	iflib_deregister(ctx);
5138 
5139 	device_set_softc(ctx->ifc_dev, NULL);
5140 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5141 		free(ctx->ifc_softc, M_IFLIB);
5142 	unref_ctx_core_offset(ctx);
5143 	free(ctx, M_IFLIB);
5144 	return (0);
5145 }
5146 
5147 static void
5148 iflib_free_intr_mem(if_ctx_t ctx)
5149 {
5150 
5151 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) {
5152 		iflib_irq_free(ctx, &ctx->ifc_legacy_irq);
5153 	}
5154 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) {
5155 		pci_release_msi(ctx->ifc_dev);
5156 	}
5157 	if (ctx->ifc_msix_mem != NULL) {
5158 		bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY,
5159 		    rman_get_rid(ctx->ifc_msix_mem), ctx->ifc_msix_mem);
5160 		ctx->ifc_msix_mem = NULL;
5161 	}
5162 }
5163 
5164 int
5165 iflib_device_detach(device_t dev)
5166 {
5167 	if_ctx_t ctx = device_get_softc(dev);
5168 
5169 	return (iflib_device_deregister(ctx));
5170 }
5171 
5172 int
5173 iflib_device_suspend(device_t dev)
5174 {
5175 	if_ctx_t ctx = device_get_softc(dev);
5176 
5177 	CTX_LOCK(ctx);
5178 	IFDI_SUSPEND(ctx);
5179 	CTX_UNLOCK(ctx);
5180 
5181 	return bus_generic_suspend(dev);
5182 }
5183 int
5184 iflib_device_shutdown(device_t dev)
5185 {
5186 	if_ctx_t ctx = device_get_softc(dev);
5187 
5188 	CTX_LOCK(ctx);
5189 	IFDI_SHUTDOWN(ctx);
5190 	CTX_UNLOCK(ctx);
5191 
5192 	return bus_generic_suspend(dev);
5193 }
5194 
5195 
5196 int
5197 iflib_device_resume(device_t dev)
5198 {
5199 	if_ctx_t ctx = device_get_softc(dev);
5200 	iflib_txq_t txq = ctx->ifc_txqs;
5201 
5202 	CTX_LOCK(ctx);
5203 	IFDI_RESUME(ctx);
5204 	iflib_if_init_locked(ctx);
5205 	CTX_UNLOCK(ctx);
5206 	for (int i = 0; i < NTXQSETS(ctx); i++, txq++)
5207 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
5208 
5209 	return (bus_generic_resume(dev));
5210 }
5211 
5212 int
5213 iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params)
5214 {
5215 	int error;
5216 	if_ctx_t ctx = device_get_softc(dev);
5217 
5218 	CTX_LOCK(ctx);
5219 	error = IFDI_IOV_INIT(ctx, num_vfs, params);
5220 	CTX_UNLOCK(ctx);
5221 
5222 	return (error);
5223 }
5224 
5225 void
5226 iflib_device_iov_uninit(device_t dev)
5227 {
5228 	if_ctx_t ctx = device_get_softc(dev);
5229 
5230 	CTX_LOCK(ctx);
5231 	IFDI_IOV_UNINIT(ctx);
5232 	CTX_UNLOCK(ctx);
5233 }
5234 
5235 int
5236 iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params)
5237 {
5238 	int error;
5239 	if_ctx_t ctx = device_get_softc(dev);
5240 
5241 	CTX_LOCK(ctx);
5242 	error = IFDI_IOV_VF_ADD(ctx, vfnum, params);
5243 	CTX_UNLOCK(ctx);
5244 
5245 	return (error);
5246 }
5247 
5248 /*********************************************************************
5249  *
5250  *  MODULE FUNCTION DEFINITIONS
5251  *
5252  **********************************************************************/
5253 
5254 /*
5255  * - Start a fast taskqueue thread for each core
5256  * - Start a taskqueue for control operations
5257  */
5258 static int
5259 iflib_module_init(void)
5260 {
5261 	return (0);
5262 }
5263 
5264 static int
5265 iflib_module_event_handler(module_t mod, int what, void *arg)
5266 {
5267 	int err;
5268 
5269 	switch (what) {
5270 	case MOD_LOAD:
5271 		if ((err = iflib_module_init()) != 0)
5272 			return (err);
5273 		break;
5274 	case MOD_UNLOAD:
5275 		return (EBUSY);
5276 	default:
5277 		return (EOPNOTSUPP);
5278 	}
5279 
5280 	return (0);
5281 }
5282 
5283 /*********************************************************************
5284  *
5285  *  PUBLIC FUNCTION DEFINITIONS
5286  *     ordered as in iflib.h
5287  *
5288  **********************************************************************/
5289 
5290 
5291 static void
5292 _iflib_assert(if_shared_ctx_t sctx)
5293 {
5294 	int i;
5295 
5296 	MPASS(sctx->isc_tx_maxsize);
5297 	MPASS(sctx->isc_tx_maxsegsize);
5298 
5299 	MPASS(sctx->isc_rx_maxsize);
5300 	MPASS(sctx->isc_rx_nsegments);
5301 	MPASS(sctx->isc_rx_maxsegsize);
5302 
5303 	MPASS(sctx->isc_nrxqs >= 1 && sctx->isc_nrxqs <= 8);
5304 	for (i = 0; i < sctx->isc_nrxqs; i++) {
5305 		MPASS(sctx->isc_nrxd_min[i]);
5306 		MPASS(powerof2(sctx->isc_nrxd_min[i]));
5307 		MPASS(sctx->isc_nrxd_max[i]);
5308 		MPASS(powerof2(sctx->isc_nrxd_max[i]));
5309 		MPASS(sctx->isc_nrxd_default[i]);
5310 		MPASS(powerof2(sctx->isc_nrxd_default[i]));
5311 	}
5312 
5313 	MPASS(sctx->isc_ntxqs >= 1 && sctx->isc_ntxqs <= 8);
5314 	for (i = 0; i < sctx->isc_ntxqs; i++) {
5315 		MPASS(sctx->isc_ntxd_min[i]);
5316 		MPASS(powerof2(sctx->isc_ntxd_min[i]));
5317 		MPASS(sctx->isc_ntxd_max[i]);
5318 		MPASS(powerof2(sctx->isc_ntxd_max[i]));
5319 		MPASS(sctx->isc_ntxd_default[i]);
5320 		MPASS(powerof2(sctx->isc_ntxd_default[i]));
5321 	}
5322 }
5323 
5324 static void
5325 _iflib_pre_assert(if_softc_ctx_t scctx)
5326 {
5327 
5328 	MPASS(scctx->isc_txrx->ift_txd_encap);
5329 	MPASS(scctx->isc_txrx->ift_txd_flush);
5330 	MPASS(scctx->isc_txrx->ift_txd_credits_update);
5331 	MPASS(scctx->isc_txrx->ift_rxd_available);
5332 	MPASS(scctx->isc_txrx->ift_rxd_pkt_get);
5333 	MPASS(scctx->isc_txrx->ift_rxd_refill);
5334 	MPASS(scctx->isc_txrx->ift_rxd_flush);
5335 }
5336 
5337 static int
5338 iflib_register(if_ctx_t ctx)
5339 {
5340 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5341 	driver_t *driver = sctx->isc_driver;
5342 	device_t dev = ctx->ifc_dev;
5343 	if_t ifp;
5344 
5345 	_iflib_assert(sctx);
5346 
5347 	CTX_LOCK_INIT(ctx);
5348 	STATE_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev));
5349 	ifp = ctx->ifc_ifp = if_alloc(IFT_ETHER);
5350 	if (ifp == NULL) {
5351 		device_printf(dev, "can not allocate ifnet structure\n");
5352 		return (ENOMEM);
5353 	}
5354 
5355 	/*
5356 	 * Initialize our context's device specific methods
5357 	 */
5358 	kobj_init((kobj_t) ctx, (kobj_class_t) driver);
5359 	kobj_class_compile((kobj_class_t) driver);
5360 
5361 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
5362 	if_setsoftc(ifp, ctx);
5363 	if_setdev(ifp, dev);
5364 	if_setinitfn(ifp, iflib_if_init);
5365 	if_setioctlfn(ifp, iflib_if_ioctl);
5366 #ifdef ALTQ
5367 	if_setstartfn(ifp, iflib_altq_if_start);
5368 	if_settransmitfn(ifp, iflib_altq_if_transmit);
5369 	if_setsendqready(ifp);
5370 #else
5371 	if_settransmitfn(ifp, iflib_if_transmit);
5372 #endif
5373 	if_setqflushfn(ifp, iflib_if_qflush);
5374 	if_setflags(ifp, IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST);
5375 
5376 	ctx->ifc_vlan_attach_event =
5377 		EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx,
5378 							  EVENTHANDLER_PRI_FIRST);
5379 	ctx->ifc_vlan_detach_event =
5380 		EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx,
5381 							  EVENTHANDLER_PRI_FIRST);
5382 
5383 	if ((sctx->isc_flags & IFLIB_DRIVER_MEDIA) == 0) {
5384 		ctx->ifc_mediap = &ctx->ifc_media;
5385 		ifmedia_init(ctx->ifc_mediap, IFM_IMASK,
5386 		    iflib_media_change, iflib_media_status);
5387 	}
5388 	return (0);
5389 }
5390 
5391 static void
5392 iflib_unregister_vlan_handlers(if_ctx_t ctx)
5393 {
5394 	/* Unregister VLAN events */
5395 	if (ctx->ifc_vlan_attach_event != NULL) {
5396 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
5397 		ctx->ifc_vlan_attach_event = NULL;
5398 	}
5399 	if (ctx->ifc_vlan_detach_event != NULL) {
5400 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
5401 		ctx->ifc_vlan_detach_event = NULL;
5402 	}
5403 
5404 }
5405 
5406 static void
5407 iflib_deregister(if_ctx_t ctx)
5408 {
5409 	if_t ifp = ctx->ifc_ifp;
5410 
5411 	/* Remove all media */
5412 	ifmedia_removeall(&ctx->ifc_media);
5413 
5414 	/* Ensure that VLAN event handlers are unregistered */
5415 	iflib_unregister_vlan_handlers(ctx);
5416 
5417 	/* Release kobject reference */
5418 	kobj_delete((kobj_t) ctx, NULL);
5419 
5420 	/* Free the ifnet structure */
5421 	if_free(ifp);
5422 
5423 	STATE_LOCK_DESTROY(ctx);
5424 
5425 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
5426 	CTX_LOCK_DESTROY(ctx);
5427 }
5428 
5429 static int
5430 iflib_queues_alloc(if_ctx_t ctx)
5431 {
5432 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5433 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5434 	device_t dev = ctx->ifc_dev;
5435 	int nrxqsets = scctx->isc_nrxqsets;
5436 	int ntxqsets = scctx->isc_ntxqsets;
5437 	iflib_txq_t txq;
5438 	iflib_rxq_t rxq;
5439 	iflib_fl_t fl = NULL;
5440 	int i, j, cpu, err, txconf, rxconf;
5441 	iflib_dma_info_t ifdip;
5442 	uint32_t *rxqsizes = scctx->isc_rxqsizes;
5443 	uint32_t *txqsizes = scctx->isc_txqsizes;
5444 	uint8_t nrxqs = sctx->isc_nrxqs;
5445 	uint8_t ntxqs = sctx->isc_ntxqs;
5446 	int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1;
5447 	caddr_t *vaddrs;
5448 	uint64_t *paddrs;
5449 
5450 	KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1"));
5451 	KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1"));
5452 
5453 	/* Allocate the TX ring struct memory */
5454 	if (!(ctx->ifc_txqs =
5455 	    (iflib_txq_t) malloc(sizeof(struct iflib_txq) *
5456 	    ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5457 		device_printf(dev, "Unable to allocate TX ring memory\n");
5458 		err = ENOMEM;
5459 		goto fail;
5460 	}
5461 
5462 	/* Now allocate the RX */
5463 	if (!(ctx->ifc_rxqs =
5464 	    (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) *
5465 	    nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5466 		device_printf(dev, "Unable to allocate RX ring memory\n");
5467 		err = ENOMEM;
5468 		goto rx_fail;
5469 	}
5470 
5471 	txq = ctx->ifc_txqs;
5472 	rxq = ctx->ifc_rxqs;
5473 
5474 	/*
5475 	 * XXX handle allocation failure
5476 	 */
5477 	for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) {
5478 		/* Set up some basics */
5479 
5480 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs,
5481 		    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
5482 			device_printf(dev,
5483 			    "Unable to allocate TX DMA info memory\n");
5484 			err = ENOMEM;
5485 			goto err_tx_desc;
5486 		}
5487 		txq->ift_ifdi = ifdip;
5488 		for (j = 0; j < ntxqs; j++, ifdip++) {
5489 			if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, 0)) {
5490 				device_printf(dev,
5491 				    "Unable to allocate TX descriptors\n");
5492 				err = ENOMEM;
5493 				goto err_tx_desc;
5494 			}
5495 			txq->ift_txd_size[j] = scctx->isc_txd_size[j];
5496 			bzero((void *)ifdip->idi_vaddr, txqsizes[j]);
5497 		}
5498 		txq->ift_ctx = ctx;
5499 		txq->ift_id = i;
5500 		if (sctx->isc_flags & IFLIB_HAS_TXCQ) {
5501 			txq->ift_br_offset = 1;
5502 		} else {
5503 			txq->ift_br_offset = 0;
5504 		}
5505 		/* XXX fix this */
5506 		txq->ift_timer.c_cpu = cpu;
5507 
5508 		if (iflib_txsd_alloc(txq)) {
5509 			device_printf(dev, "Critical Failure setting up TX buffers\n");
5510 			err = ENOMEM;
5511 			goto err_tx_desc;
5512 		}
5513 
5514 		/* Initialize the TX lock */
5515 		snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:TX(%d):callout",
5516 		    device_get_nameunit(dev), txq->ift_id);
5517 		mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF);
5518 		callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0);
5519 
5520 		err = ifmp_ring_alloc(&txq->ift_br, 2048, txq, iflib_txq_drain,
5521 				      iflib_txq_can_drain, M_IFLIB, M_WAITOK);
5522 		if (err) {
5523 			/* XXX free any allocated rings */
5524 			device_printf(dev, "Unable to allocate buf_ring\n");
5525 			goto err_tx_desc;
5526 		}
5527 	}
5528 
5529 	for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) {
5530 		/* Set up some basics */
5531 
5532 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs,
5533 		   M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
5534 			device_printf(dev,
5535 			    "Unable to allocate RX DMA info memory\n");
5536 			err = ENOMEM;
5537 			goto err_tx_desc;
5538 		}
5539 
5540 		rxq->ifr_ifdi = ifdip;
5541 		/* XXX this needs to be changed if #rx queues != #tx queues */
5542 		rxq->ifr_ntxqirq = 1;
5543 		rxq->ifr_txqid[0] = i;
5544 		for (j = 0; j < nrxqs; j++, ifdip++) {
5545 			if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, 0)) {
5546 				device_printf(dev,
5547 				    "Unable to allocate RX descriptors\n");
5548 				err = ENOMEM;
5549 				goto err_tx_desc;
5550 			}
5551 			bzero((void *)ifdip->idi_vaddr, rxqsizes[j]);
5552 		}
5553 		rxq->ifr_ctx = ctx;
5554 		rxq->ifr_id = i;
5555 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
5556 			rxq->ifr_fl_offset = 1;
5557 		} else {
5558 			rxq->ifr_fl_offset = 0;
5559 		}
5560 		rxq->ifr_nfl = nfree_lists;
5561 		if (!(fl =
5562 			  (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) {
5563 			device_printf(dev, "Unable to allocate free list memory\n");
5564 			err = ENOMEM;
5565 			goto err_tx_desc;
5566 		}
5567 		rxq->ifr_fl = fl;
5568 		for (j = 0; j < nfree_lists; j++) {
5569 			fl[j].ifl_rxq = rxq;
5570 			fl[j].ifl_id = j;
5571 			fl[j].ifl_ifdi = &rxq->ifr_ifdi[j + rxq->ifr_fl_offset];
5572 			fl[j].ifl_rxd_size = scctx->isc_rxd_size[j];
5573 		}
5574 		/* Allocate receive buffers for the ring */
5575 		if (iflib_rxsd_alloc(rxq)) {
5576 			device_printf(dev,
5577 			    "Critical Failure setting up receive buffers\n");
5578 			err = ENOMEM;
5579 			goto err_rx_desc;
5580 		}
5581 
5582 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5583 			fl->ifl_rx_bitmap = bit_alloc(fl->ifl_size, M_IFLIB,
5584 			    M_WAITOK);
5585 	}
5586 
5587 	/* TXQs */
5588 	vaddrs = malloc(sizeof(caddr_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5589 	paddrs = malloc(sizeof(uint64_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5590 	for (i = 0; i < ntxqsets; i++) {
5591 		iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi;
5592 
5593 		for (j = 0; j < ntxqs; j++, di++) {
5594 			vaddrs[i*ntxqs + j] = di->idi_vaddr;
5595 			paddrs[i*ntxqs + j] = di->idi_paddr;
5596 		}
5597 	}
5598 	if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) {
5599 		device_printf(ctx->ifc_dev,
5600 		    "Unable to allocate device TX queue\n");
5601 		iflib_tx_structures_free(ctx);
5602 		free(vaddrs, M_IFLIB);
5603 		free(paddrs, M_IFLIB);
5604 		goto err_rx_desc;
5605 	}
5606 	free(vaddrs, M_IFLIB);
5607 	free(paddrs, M_IFLIB);
5608 
5609 	/* RXQs */
5610 	vaddrs = malloc(sizeof(caddr_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5611 	paddrs = malloc(sizeof(uint64_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5612 	for (i = 0; i < nrxqsets; i++) {
5613 		iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi;
5614 
5615 		for (j = 0; j < nrxqs; j++, di++) {
5616 			vaddrs[i*nrxqs + j] = di->idi_vaddr;
5617 			paddrs[i*nrxqs + j] = di->idi_paddr;
5618 		}
5619 	}
5620 	if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) {
5621 		device_printf(ctx->ifc_dev,
5622 		    "Unable to allocate device RX queue\n");
5623 		iflib_tx_structures_free(ctx);
5624 		free(vaddrs, M_IFLIB);
5625 		free(paddrs, M_IFLIB);
5626 		goto err_rx_desc;
5627 	}
5628 	free(vaddrs, M_IFLIB);
5629 	free(paddrs, M_IFLIB);
5630 
5631 	return (0);
5632 
5633 /* XXX handle allocation failure changes */
5634 err_rx_desc:
5635 err_tx_desc:
5636 rx_fail:
5637 	if (ctx->ifc_rxqs != NULL)
5638 		free(ctx->ifc_rxqs, M_IFLIB);
5639 	ctx->ifc_rxqs = NULL;
5640 	if (ctx->ifc_txqs != NULL)
5641 		free(ctx->ifc_txqs, M_IFLIB);
5642 	ctx->ifc_txqs = NULL;
5643 fail:
5644 	return (err);
5645 }
5646 
5647 static int
5648 iflib_tx_structures_setup(if_ctx_t ctx)
5649 {
5650 	iflib_txq_t txq = ctx->ifc_txqs;
5651 	int i;
5652 
5653 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
5654 		iflib_txq_setup(txq);
5655 
5656 	return (0);
5657 }
5658 
5659 static void
5660 iflib_tx_structures_free(if_ctx_t ctx)
5661 {
5662 	iflib_txq_t txq = ctx->ifc_txqs;
5663 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5664 	int i, j;
5665 
5666 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
5667 		for (j = 0; j < sctx->isc_ntxqs; j++)
5668 			iflib_dma_free(&txq->ift_ifdi[j]);
5669 		iflib_txq_destroy(txq);
5670 	}
5671 	free(ctx->ifc_txqs, M_IFLIB);
5672 	ctx->ifc_txqs = NULL;
5673 	IFDI_QUEUES_FREE(ctx);
5674 }
5675 
5676 /*********************************************************************
5677  *
5678  *  Initialize all receive rings.
5679  *
5680  **********************************************************************/
5681 static int
5682 iflib_rx_structures_setup(if_ctx_t ctx)
5683 {
5684 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5685 	int q;
5686 #if defined(INET6) || defined(INET)
5687 	int err, i;
5688 #endif
5689 
5690 	for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) {
5691 #if defined(INET6) || defined(INET)
5692 		if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_LRO) {
5693 			err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp,
5694 			    TCP_LRO_ENTRIES, min(1024,
5695 			    ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset]));
5696 			if (err != 0) {
5697 				device_printf(ctx->ifc_dev,
5698 				    "LRO Initialization failed!\n");
5699 				goto fail;
5700 			}
5701 		}
5702 #endif
5703 		IFDI_RXQ_SETUP(ctx, rxq->ifr_id);
5704 	}
5705 	return (0);
5706 #if defined(INET6) || defined(INET)
5707 fail:
5708 	/*
5709 	 * Free LRO resources allocated so far, we will only handle
5710 	 * the rings that completed, the failing case will have
5711 	 * cleaned up for itself.  'q' failed, so its the terminus.
5712 	 */
5713 	rxq = ctx->ifc_rxqs;
5714 	for (i = 0; i < q; ++i, rxq++) {
5715 		if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_LRO)
5716 			tcp_lro_free(&rxq->ifr_lc);
5717 	}
5718 	return (err);
5719 #endif
5720 }
5721 
5722 /*********************************************************************
5723  *
5724  *  Free all receive rings.
5725  *
5726  **********************************************************************/
5727 static void
5728 iflib_rx_structures_free(if_ctx_t ctx)
5729 {
5730 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5731 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5732 	int i, j;
5733 
5734 	for (i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) {
5735 		for (j = 0; j < sctx->isc_nrxqs; j++)
5736 			iflib_dma_free(&rxq->ifr_ifdi[j]);
5737 		iflib_rx_sds_free(rxq);
5738 #if defined(INET6) || defined(INET)
5739 		if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_LRO)
5740 			tcp_lro_free(&rxq->ifr_lc);
5741 #endif
5742 	}
5743 	free(ctx->ifc_rxqs, M_IFLIB);
5744 	ctx->ifc_rxqs = NULL;
5745 }
5746 
5747 static int
5748 iflib_qset_structures_setup(if_ctx_t ctx)
5749 {
5750 	int err;
5751 
5752 	/*
5753 	 * It is expected that the caller takes care of freeing queues if this
5754 	 * fails.
5755 	 */
5756 	if ((err = iflib_tx_structures_setup(ctx)) != 0) {
5757 		device_printf(ctx->ifc_dev, "iflib_tx_structures_setup failed: %d\n", err);
5758 		return (err);
5759 	}
5760 
5761 	if ((err = iflib_rx_structures_setup(ctx)) != 0)
5762 		device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err);
5763 
5764 	return (err);
5765 }
5766 
5767 int
5768 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
5769 		driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, const char *name)
5770 {
5771 
5772 	return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name));
5773 }
5774 
5775 #ifdef SMP
5776 static int
5777 find_nth(if_ctx_t ctx, int qid)
5778 {
5779 	cpuset_t cpus;
5780 	int i, cpuid, eqid, count;
5781 
5782 	CPU_COPY(&ctx->ifc_cpus, &cpus);
5783 	count = CPU_COUNT(&cpus);
5784 	eqid = qid % count;
5785 	/* clear up to the qid'th bit */
5786 	for (i = 0; i < eqid; i++) {
5787 		cpuid = CPU_FFS(&cpus);
5788 		MPASS(cpuid != 0);
5789 		CPU_CLR(cpuid-1, &cpus);
5790 	}
5791 	cpuid = CPU_FFS(&cpus);
5792 	MPASS(cpuid != 0);
5793 	return (cpuid-1);
5794 }
5795 
5796 #ifdef SCHED_ULE
5797 extern struct cpu_group *cpu_top;              /* CPU topology */
5798 
5799 static int
5800 find_child_with_core(int cpu, struct cpu_group *grp)
5801 {
5802 	int i;
5803 
5804 	if (grp->cg_children == 0)
5805 		return -1;
5806 
5807 	MPASS(grp->cg_child);
5808 	for (i = 0; i < grp->cg_children; i++) {
5809 		if (CPU_ISSET(cpu, &grp->cg_child[i].cg_mask))
5810 			return i;
5811 	}
5812 
5813 	return -1;
5814 }
5815 
5816 /*
5817  * Find the nth "close" core to the specified core
5818  * "close" is defined as the deepest level that shares
5819  * at least an L2 cache.  With threads, this will be
5820  * threads on the same core.  If the shared cache is L3
5821  * or higher, simply returns the same core.
5822  */
5823 static int
5824 find_close_core(int cpu, int core_offset)
5825 {
5826 	struct cpu_group *grp;
5827 	int i;
5828 	int fcpu;
5829 	cpuset_t cs;
5830 
5831 	grp = cpu_top;
5832 	if (grp == NULL)
5833 		return cpu;
5834 	i = 0;
5835 	while ((i = find_child_with_core(cpu, grp)) != -1) {
5836 		/* If the child only has one cpu, don't descend */
5837 		if (grp->cg_child[i].cg_count <= 1)
5838 			break;
5839 		grp = &grp->cg_child[i];
5840 	}
5841 
5842 	/* If they don't share at least an L2 cache, use the same CPU */
5843 	if (grp->cg_level > CG_SHARE_L2 || grp->cg_level == CG_SHARE_NONE)
5844 		return cpu;
5845 
5846 	/* Now pick one */
5847 	CPU_COPY(&grp->cg_mask, &cs);
5848 
5849 	/* Add the selected CPU offset to core offset. */
5850 	for (i = 0; (fcpu = CPU_FFS(&cs)) != 0; i++) {
5851 		if (fcpu - 1 == cpu)
5852 			break;
5853 		CPU_CLR(fcpu - 1, &cs);
5854 	}
5855 	MPASS(fcpu);
5856 
5857 	core_offset += i;
5858 
5859 	CPU_COPY(&grp->cg_mask, &cs);
5860 	for (i = core_offset % grp->cg_count; i > 0; i--) {
5861 		MPASS(CPU_FFS(&cs));
5862 		CPU_CLR(CPU_FFS(&cs) - 1, &cs);
5863 	}
5864 	MPASS(CPU_FFS(&cs));
5865 	return CPU_FFS(&cs) - 1;
5866 }
5867 #else
5868 static int
5869 find_close_core(int cpu, int core_offset __unused)
5870 {
5871 	return cpu;
5872 }
5873 #endif
5874 
5875 static int
5876 get_core_offset(if_ctx_t ctx, iflib_intr_type_t type, int qid)
5877 {
5878 	switch (type) {
5879 	case IFLIB_INTR_TX:
5880 		/* TX queues get cores which share at least an L2 cache with the corresponding RX queue */
5881 		/* XXX handle multiple RX threads per core and more than two core per L2 group */
5882 		return qid / CPU_COUNT(&ctx->ifc_cpus) + 1;
5883 	case IFLIB_INTR_RX:
5884 	case IFLIB_INTR_RXTX:
5885 		/* RX queues get the specified core */
5886 		return qid / CPU_COUNT(&ctx->ifc_cpus);
5887 	default:
5888 		return -1;
5889 	}
5890 }
5891 #else
5892 #define get_core_offset(ctx, type, qid)	CPU_FIRST()
5893 #define find_close_core(cpuid, tid)	CPU_FIRST()
5894 #define find_nth(ctx, gid)		CPU_FIRST()
5895 #endif
5896 
5897 /* Just to avoid copy/paste */
5898 static inline int
5899 iflib_irq_set_affinity(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type,
5900     int qid, struct grouptask *gtask, struct taskqgroup *tqg, void *uniq,
5901     const char *name)
5902 {
5903 	device_t dev;
5904 	int co, cpuid, err, tid;
5905 
5906 	dev = ctx->ifc_dev;
5907 	co = ctx->ifc_sysctl_core_offset;
5908 	if (ctx->ifc_sysctl_separate_txrx && type == IFLIB_INTR_TX)
5909 		co += ctx->ifc_softc_ctx.isc_nrxqsets;
5910 	cpuid = find_nth(ctx, qid + co);
5911 	tid = get_core_offset(ctx, type, qid);
5912 	if (tid < 0) {
5913 		device_printf(dev, "get_core_offset failed\n");
5914 		return (EOPNOTSUPP);
5915 	}
5916 	cpuid = find_close_core(cpuid, tid);
5917 	err = taskqgroup_attach_cpu(tqg, gtask, uniq, cpuid, dev, irq->ii_res,
5918 	    name);
5919 	if (err) {
5920 		device_printf(dev, "taskqgroup_attach_cpu failed %d\n", err);
5921 		return (err);
5922 	}
5923 #ifdef notyet
5924 	if (cpuid > ctx->ifc_cpuid_highest)
5925 		ctx->ifc_cpuid_highest = cpuid;
5926 #endif
5927 	return (0);
5928 }
5929 
5930 int
5931 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid,
5932 			iflib_intr_type_t type, driver_filter_t *filter,
5933 			void *filter_arg, int qid, const char *name)
5934 {
5935 	device_t dev;
5936 	struct grouptask *gtask;
5937 	struct taskqgroup *tqg;
5938 	iflib_filter_info_t info;
5939 	gtask_fn_t *fn;
5940 	int tqrid, err;
5941 	driver_filter_t *intr_fast;
5942 	void *q;
5943 
5944 	info = &ctx->ifc_filter_info;
5945 	tqrid = rid;
5946 
5947 	switch (type) {
5948 	/* XXX merge tx/rx for netmap? */
5949 	case IFLIB_INTR_TX:
5950 		q = &ctx->ifc_txqs[qid];
5951 		info = &ctx->ifc_txqs[qid].ift_filter_info;
5952 		gtask = &ctx->ifc_txqs[qid].ift_task;
5953 		tqg = qgroup_if_io_tqg;
5954 		fn = _task_fn_tx;
5955 		intr_fast = iflib_fast_intr;
5956 		GROUPTASK_INIT(gtask, 0, fn, q);
5957 		ctx->ifc_flags |= IFC_NETMAP_TX_IRQ;
5958 		break;
5959 	case IFLIB_INTR_RX:
5960 		q = &ctx->ifc_rxqs[qid];
5961 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
5962 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5963 		tqg = qgroup_if_io_tqg;
5964 		fn = _task_fn_rx;
5965 		intr_fast = iflib_fast_intr;
5966 		GROUPTASK_INIT(gtask, 0, fn, q);
5967 		break;
5968 	case IFLIB_INTR_RXTX:
5969 		q = &ctx->ifc_rxqs[qid];
5970 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
5971 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5972 		tqg = qgroup_if_io_tqg;
5973 		fn = _task_fn_rx;
5974 		intr_fast = iflib_fast_intr_rxtx;
5975 		GROUPTASK_INIT(gtask, 0, fn, q);
5976 		break;
5977 	case IFLIB_INTR_ADMIN:
5978 		q = ctx;
5979 		tqrid = -1;
5980 		info = &ctx->ifc_filter_info;
5981 		gtask = &ctx->ifc_admin_task;
5982 		tqg = qgroup_if_config_tqg;
5983 		fn = _task_fn_admin;
5984 		intr_fast = iflib_fast_intr_ctx;
5985 		break;
5986 	default:
5987 		device_printf(ctx->ifc_dev, "%s: unknown net intr type\n",
5988 		    __func__);
5989 		return (EINVAL);
5990 	}
5991 
5992 	info->ifi_filter = filter;
5993 	info->ifi_filter_arg = filter_arg;
5994 	info->ifi_task = gtask;
5995 	info->ifi_ctx = q;
5996 
5997 	dev = ctx->ifc_dev;
5998 	err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info,  name);
5999 	if (err != 0) {
6000 		device_printf(dev, "_iflib_irq_alloc failed %d\n", err);
6001 		return (err);
6002 	}
6003 	if (type == IFLIB_INTR_ADMIN)
6004 		return (0);
6005 
6006 	if (tqrid != -1) {
6007 		err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg,
6008 		    q, name);
6009 		if (err)
6010 			return (err);
6011 	} else {
6012 		taskqgroup_attach(tqg, gtask, q, dev, irq->ii_res, name);
6013 	}
6014 
6015 	return (0);
6016 }
6017 
6018 void
6019 iflib_softirq_alloc_generic(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type, void *arg, int qid, const char *name)
6020 {
6021 	struct grouptask *gtask;
6022 	struct taskqgroup *tqg;
6023 	gtask_fn_t *fn;
6024 	void *q;
6025 	int err;
6026 
6027 	switch (type) {
6028 	case IFLIB_INTR_TX:
6029 		q = &ctx->ifc_txqs[qid];
6030 		gtask = &ctx->ifc_txqs[qid].ift_task;
6031 		tqg = qgroup_if_io_tqg;
6032 		fn = _task_fn_tx;
6033 		break;
6034 	case IFLIB_INTR_RX:
6035 		q = &ctx->ifc_rxqs[qid];
6036 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6037 		tqg = qgroup_if_io_tqg;
6038 		fn = _task_fn_rx;
6039 		break;
6040 	case IFLIB_INTR_IOV:
6041 		q = ctx;
6042 		gtask = &ctx->ifc_vflr_task;
6043 		tqg = qgroup_if_config_tqg;
6044 		fn = _task_fn_iov;
6045 		break;
6046 	default:
6047 		panic("unknown net intr type");
6048 	}
6049 	GROUPTASK_INIT(gtask, 0, fn, q);
6050 	if (irq != NULL) {
6051 		err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg,
6052 		    q, name);
6053 		if (err)
6054 			taskqgroup_attach(tqg, gtask, q, ctx->ifc_dev,
6055 			    irq->ii_res, name);
6056 	} else {
6057 		taskqgroup_attach(tqg, gtask, q, NULL, NULL, name);
6058 	}
6059 }
6060 
6061 void
6062 iflib_irq_free(if_ctx_t ctx, if_irq_t irq)
6063 {
6064 
6065 	if (irq->ii_tag)
6066 		bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag);
6067 
6068 	if (irq->ii_res)
6069 		bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ,
6070 		    rman_get_rid(irq->ii_res), irq->ii_res);
6071 }
6072 
6073 static int
6074 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, const char *name)
6075 {
6076 	iflib_txq_t txq = ctx->ifc_txqs;
6077 	iflib_rxq_t rxq = ctx->ifc_rxqs;
6078 	if_irq_t irq = &ctx->ifc_legacy_irq;
6079 	iflib_filter_info_t info;
6080 	device_t dev;
6081 	struct grouptask *gtask;
6082 	struct resource *res;
6083 	struct taskqgroup *tqg;
6084 	gtask_fn_t *fn;
6085 	void *q;
6086 	int err, tqrid;
6087 	bool rx_only;
6088 
6089 	q = &ctx->ifc_rxqs[0];
6090 	info = &rxq[0].ifr_filter_info;
6091 	gtask = &rxq[0].ifr_task;
6092 	tqg = qgroup_if_io_tqg;
6093 	tqrid = *rid;
6094 	fn = _task_fn_rx;
6095 	rx_only = (ctx->ifc_sctx->isc_flags & IFLIB_SINGLE_IRQ_RX_ONLY) != 0;
6096 
6097 	ctx->ifc_flags |= IFC_LEGACY;
6098 	info->ifi_filter = filter;
6099 	info->ifi_filter_arg = filter_arg;
6100 	info->ifi_task = gtask;
6101 	info->ifi_ctx = rx_only ? ctx : q;
6102 
6103 	dev = ctx->ifc_dev;
6104 	/* We allocate a single interrupt resource */
6105 	err = _iflib_irq_alloc(ctx, irq, tqrid, rx_only ? iflib_fast_intr_ctx :
6106 	    iflib_fast_intr_rxtx, NULL, info, name);
6107 	if (err != 0)
6108 		return (err);
6109 	GROUPTASK_INIT(gtask, 0, fn, q);
6110 	res = irq->ii_res;
6111 	taskqgroup_attach(tqg, gtask, q, dev, res, name);
6112 
6113 	GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq);
6114 	taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, dev, res,
6115 	    "tx");
6116 	return (0);
6117 }
6118 
6119 void
6120 iflib_led_create(if_ctx_t ctx)
6121 {
6122 
6123 	ctx->ifc_led_dev = led_create(iflib_led_func, ctx,
6124 	    device_get_nameunit(ctx->ifc_dev));
6125 }
6126 
6127 void
6128 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid)
6129 {
6130 
6131 	GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
6132 }
6133 
6134 void
6135 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid)
6136 {
6137 
6138 	GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task);
6139 }
6140 
6141 void
6142 iflib_admin_intr_deferred(if_ctx_t ctx)
6143 {
6144 #ifdef INVARIANTS
6145 	struct grouptask *gtask;
6146 
6147 	gtask = &ctx->ifc_admin_task;
6148 	MPASS(gtask != NULL && gtask->gt_taskqueue != NULL);
6149 #endif
6150 
6151 	GROUPTASK_ENQUEUE(&ctx->ifc_admin_task);
6152 }
6153 
6154 void
6155 iflib_iov_intr_deferred(if_ctx_t ctx)
6156 {
6157 
6158 	GROUPTASK_ENQUEUE(&ctx->ifc_vflr_task);
6159 }
6160 
6161 void
6162 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, const char *name)
6163 {
6164 
6165 	taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, NULL, NULL,
6166 	    name);
6167 }
6168 
6169 void
6170 iflib_config_gtask_init(void *ctx, struct grouptask *gtask, gtask_fn_t *fn,
6171 	const char *name)
6172 {
6173 
6174 	GROUPTASK_INIT(gtask, 0, fn, ctx);
6175 	taskqgroup_attach(qgroup_if_config_tqg, gtask, gtask, NULL, NULL,
6176 	    name);
6177 }
6178 
6179 void
6180 iflib_config_gtask_deinit(struct grouptask *gtask)
6181 {
6182 
6183 	taskqgroup_detach(qgroup_if_config_tqg, gtask);
6184 }
6185 
6186 void
6187 iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate)
6188 {
6189 	if_t ifp = ctx->ifc_ifp;
6190 	iflib_txq_t txq = ctx->ifc_txqs;
6191 
6192 	if_setbaudrate(ifp, baudrate);
6193 	if (baudrate >= IF_Gbps(10)) {
6194 		STATE_LOCK(ctx);
6195 		ctx->ifc_flags |= IFC_PREFETCH;
6196 		STATE_UNLOCK(ctx);
6197 	}
6198 	/* If link down, disable watchdog */
6199 	if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) {
6200 		for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++)
6201 			txq->ift_qstatus = IFLIB_QUEUE_IDLE;
6202 	}
6203 	ctx->ifc_link_state = link_state;
6204 	if_link_state_change(ifp, link_state);
6205 }
6206 
6207 static int
6208 iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq)
6209 {
6210 	int credits;
6211 #ifdef INVARIANTS
6212 	int credits_pre = txq->ift_cidx_processed;
6213 #endif
6214 
6215 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
6216 	    BUS_DMASYNC_POSTREAD);
6217 	if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, true)) == 0)
6218 		return (0);
6219 
6220 	txq->ift_processed += credits;
6221 	txq->ift_cidx_processed += credits;
6222 
6223 	MPASS(credits_pre + credits == txq->ift_cidx_processed);
6224 	if (txq->ift_cidx_processed >= txq->ift_size)
6225 		txq->ift_cidx_processed -= txq->ift_size;
6226 	return (credits);
6227 }
6228 
6229 static int
6230 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget)
6231 {
6232 	iflib_fl_t fl;
6233 	u_int i;
6234 
6235 	for (i = 0, fl = &rxq->ifr_fl[0]; i < rxq->ifr_nfl; i++, fl++)
6236 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
6237 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
6238 	return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx,
6239 	    budget));
6240 }
6241 
6242 void
6243 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name,
6244 	const char *description, if_int_delay_info_t info,
6245 	int offset, int value)
6246 {
6247 	info->iidi_ctx = ctx;
6248 	info->iidi_offset = offset;
6249 	info->iidi_value = value;
6250 	SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev),
6251 	    SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)),
6252 	    OID_AUTO, name, CTLTYPE_INT|CTLFLAG_RW,
6253 	    info, 0, iflib_sysctl_int_delay, "I", description);
6254 }
6255 
6256 struct sx *
6257 iflib_ctx_lock_get(if_ctx_t ctx)
6258 {
6259 
6260 	return (&ctx->ifc_ctx_sx);
6261 }
6262 
6263 static int
6264 iflib_msix_init(if_ctx_t ctx)
6265 {
6266 	device_t dev = ctx->ifc_dev;
6267 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6268 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6269 	int admincnt, bar, err, iflib_num_rx_queues, iflib_num_tx_queues;
6270 	int msgs, queuemsgs, queues, rx_queues, tx_queues, vectors;
6271 
6272 	iflib_num_tx_queues = ctx->ifc_sysctl_ntxqs;
6273 	iflib_num_rx_queues = ctx->ifc_sysctl_nrxqs;
6274 
6275 	if (bootverbose)
6276 		device_printf(dev, "msix_init qsets capped at %d\n",
6277 		    imax(scctx->isc_ntxqsets, scctx->isc_nrxqsets));
6278 
6279 	/* Override by tuneable */
6280 	if (scctx->isc_disable_msix)
6281 		goto msi;
6282 
6283 	/* First try MSI-X */
6284 	if ((msgs = pci_msix_count(dev)) == 0) {
6285 		if (bootverbose)
6286 			device_printf(dev, "MSI-X not supported or disabled\n");
6287 		goto msi;
6288 	}
6289 
6290 	bar = ctx->ifc_softc_ctx.isc_msix_bar;
6291 	/*
6292 	 * bar == -1 => "trust me I know what I'm doing"
6293 	 * Some drivers are for hardware that is so shoddily
6294 	 * documented that no one knows which bars are which
6295 	 * so the developer has to map all bars. This hack
6296 	 * allows shoddy garbage to use MSI-X in this framework.
6297 	 */
6298 	if (bar != -1) {
6299 		ctx->ifc_msix_mem = bus_alloc_resource_any(dev,
6300 	            SYS_RES_MEMORY, &bar, RF_ACTIVE);
6301 		if (ctx->ifc_msix_mem == NULL) {
6302 			device_printf(dev, "Unable to map MSI-X table\n");
6303 			goto msi;
6304 		}
6305 	}
6306 
6307 	admincnt = sctx->isc_admin_intrcnt;
6308 #if IFLIB_DEBUG
6309 	/* use only 1 qset in debug mode */
6310 	queuemsgs = min(msgs - admincnt, 1);
6311 #else
6312 	queuemsgs = msgs - admincnt;
6313 #endif
6314 #ifdef RSS
6315 	queues = imin(queuemsgs, rss_getnumbuckets());
6316 #else
6317 	queues = queuemsgs;
6318 #endif
6319 	queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues);
6320 	if (bootverbose)
6321 		device_printf(dev,
6322 		    "intr CPUs: %d queue msgs: %d admincnt: %d\n",
6323 		    CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt);
6324 #ifdef  RSS
6325 	/* If we're doing RSS, clamp at the number of RSS buckets */
6326 	if (queues > rss_getnumbuckets())
6327 		queues = rss_getnumbuckets();
6328 #endif
6329 	if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt)
6330 		rx_queues = iflib_num_rx_queues;
6331 	else
6332 		rx_queues = queues;
6333 
6334 	if (rx_queues > scctx->isc_nrxqsets)
6335 		rx_queues = scctx->isc_nrxqsets;
6336 
6337 	/*
6338 	 * We want this to be all logical CPUs by default
6339 	 */
6340 	if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues)
6341 		tx_queues = iflib_num_tx_queues;
6342 	else
6343 		tx_queues = mp_ncpus;
6344 
6345 	if (tx_queues > scctx->isc_ntxqsets)
6346 		tx_queues = scctx->isc_ntxqsets;
6347 
6348 	if (ctx->ifc_sysctl_qs_eq_override == 0) {
6349 #ifdef INVARIANTS
6350 		if (tx_queues != rx_queues)
6351 			device_printf(dev,
6352 			    "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n",
6353 			    min(rx_queues, tx_queues), min(rx_queues, tx_queues));
6354 #endif
6355 		tx_queues = min(rx_queues, tx_queues);
6356 		rx_queues = min(rx_queues, tx_queues);
6357 	}
6358 
6359 	vectors = rx_queues + admincnt;
6360 	if (msgs < vectors) {
6361 		device_printf(dev,
6362 		    "insufficient number of MSI-X vectors "
6363 		    "(supported %d, need %d)\n", msgs, vectors);
6364 		goto msi;
6365 	}
6366 
6367 	device_printf(dev, "Using %d RX queues %d TX queues\n", rx_queues,
6368 	    tx_queues);
6369 	msgs = vectors;
6370 	if ((err = pci_alloc_msix(dev, &vectors)) == 0) {
6371 		if (vectors != msgs) {
6372 			device_printf(dev,
6373 			    "Unable to allocate sufficient MSI-X vectors "
6374 			    "(got %d, need %d)\n", vectors, msgs);
6375 			pci_release_msi(dev);
6376 			if (bar != -1) {
6377 				bus_release_resource(dev, SYS_RES_MEMORY, bar,
6378 				    ctx->ifc_msix_mem);
6379 				ctx->ifc_msix_mem = NULL;
6380 			}
6381 			goto msi;
6382 		}
6383 		device_printf(dev, "Using MSI-X interrupts with %d vectors\n",
6384 		    vectors);
6385 		scctx->isc_vectors = vectors;
6386 		scctx->isc_nrxqsets = rx_queues;
6387 		scctx->isc_ntxqsets = tx_queues;
6388 		scctx->isc_intr = IFLIB_INTR_MSIX;
6389 
6390 		return (vectors);
6391 	} else {
6392 		device_printf(dev,
6393 		    "failed to allocate %d MSI-X vectors, err: %d\n", vectors,
6394 		    err);
6395 		if (bar != -1) {
6396 			bus_release_resource(dev, SYS_RES_MEMORY, bar,
6397 			    ctx->ifc_msix_mem);
6398 			ctx->ifc_msix_mem = NULL;
6399 		}
6400 	}
6401 
6402 msi:
6403 	vectors = pci_msi_count(dev);
6404 	scctx->isc_nrxqsets = 1;
6405 	scctx->isc_ntxqsets = 1;
6406 	scctx->isc_vectors = vectors;
6407 	if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) {
6408 		device_printf(dev,"Using an MSI interrupt\n");
6409 		scctx->isc_intr = IFLIB_INTR_MSI;
6410 	} else {
6411 		scctx->isc_vectors = 1;
6412 		device_printf(dev,"Using a Legacy interrupt\n");
6413 		scctx->isc_intr = IFLIB_INTR_LEGACY;
6414 	}
6415 
6416 	return (vectors);
6417 }
6418 
6419 static const char *ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" };
6420 
6421 static int
6422 mp_ring_state_handler(SYSCTL_HANDLER_ARGS)
6423 {
6424 	int rc;
6425 	uint16_t *state = ((uint16_t *)oidp->oid_arg1);
6426 	struct sbuf *sb;
6427 	const char *ring_state = "UNKNOWN";
6428 
6429 	/* XXX needed ? */
6430 	rc = sysctl_wire_old_buffer(req, 0);
6431 	MPASS(rc == 0);
6432 	if (rc != 0)
6433 		return (rc);
6434 	sb = sbuf_new_for_sysctl(NULL, NULL, 80, req);
6435 	MPASS(sb != NULL);
6436 	if (sb == NULL)
6437 		return (ENOMEM);
6438 	if (state[3] <= 3)
6439 		ring_state = ring_states[state[3]];
6440 
6441 	sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s",
6442 		    state[0], state[1], state[2], ring_state);
6443 	rc = sbuf_finish(sb);
6444 	sbuf_delete(sb);
6445         return(rc);
6446 }
6447 
6448 enum iflib_ndesc_handler {
6449 	IFLIB_NTXD_HANDLER,
6450 	IFLIB_NRXD_HANDLER,
6451 };
6452 
6453 static int
6454 mp_ndesc_handler(SYSCTL_HANDLER_ARGS)
6455 {
6456 	if_ctx_t ctx = (void *)arg1;
6457 	enum iflib_ndesc_handler type = arg2;
6458 	char buf[256] = {0};
6459 	qidx_t *ndesc;
6460 	char *p, *next;
6461 	int nqs, rc, i;
6462 
6463 	nqs = 8;
6464 	switch(type) {
6465 	case IFLIB_NTXD_HANDLER:
6466 		ndesc = ctx->ifc_sysctl_ntxds;
6467 		if (ctx->ifc_sctx)
6468 			nqs = ctx->ifc_sctx->isc_ntxqs;
6469 		break;
6470 	case IFLIB_NRXD_HANDLER:
6471 		ndesc = ctx->ifc_sysctl_nrxds;
6472 		if (ctx->ifc_sctx)
6473 			nqs = ctx->ifc_sctx->isc_nrxqs;
6474 		break;
6475 	default:
6476 		printf("%s: unhandled type\n", __func__);
6477 		return (EINVAL);
6478 	}
6479 	if (nqs == 0)
6480 		nqs = 8;
6481 
6482 	for (i=0; i<8; i++) {
6483 		if (i >= nqs)
6484 			break;
6485 		if (i)
6486 			strcat(buf, ",");
6487 		sprintf(strchr(buf, 0), "%d", ndesc[i]);
6488 	}
6489 
6490 	rc = sysctl_handle_string(oidp, buf, sizeof(buf), req);
6491 	if (rc || req->newptr == NULL)
6492 		return rc;
6493 
6494 	for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p;
6495 	    i++, p = strsep(&next, " ,")) {
6496 		ndesc[i] = strtoul(p, NULL, 10);
6497 	}
6498 
6499 	return(rc);
6500 }
6501 
6502 #define NAME_BUFLEN 32
6503 static void
6504 iflib_add_device_sysctl_pre(if_ctx_t ctx)
6505 {
6506         device_t dev = iflib_get_dev(ctx);
6507 	struct sysctl_oid_list *child, *oid_list;
6508 	struct sysctl_ctx_list *ctx_list;
6509 	struct sysctl_oid *node;
6510 
6511 	ctx_list = device_get_sysctl_ctx(dev);
6512 	child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev));
6513 	ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, "iflib",
6514 						      CTLFLAG_RD, NULL, "IFLIB fields");
6515 	oid_list = SYSCTL_CHILDREN(node);
6516 
6517 	SYSCTL_ADD_CONST_STRING(ctx_list, oid_list, OID_AUTO, "driver_version",
6518 		       CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version,
6519 		       "driver version");
6520 
6521 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs",
6522 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0,
6523 			"# of txqs to use, 0 => use default #");
6524 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs",
6525 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0,
6526 			"# of rxqs to use, 0 => use default #");
6527 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_qs_enable",
6528 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0,
6529                        "permit #txq != #rxq");
6530 	SYSCTL_ADD_INT(ctx_list, oid_list, OID_AUTO, "disable_msix",
6531                       CTLFLAG_RWTUN, &ctx->ifc_softc_ctx.isc_disable_msix, 0,
6532                       "disable MSI-X (default 0)");
6533 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "rx_budget",
6534 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_rx_budget, 0,
6535 		       "set the RX budget");
6536 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "tx_abdicate",
6537 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_tx_abdicate, 0,
6538 		       "cause TX to abdicate instead of running to completion");
6539 	ctx->ifc_sysctl_core_offset = CORE_OFFSET_UNSPECIFIED;
6540 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "core_offset",
6541 		       CTLFLAG_RDTUN, &ctx->ifc_sysctl_core_offset, 0,
6542 		       "offset to start using cores at");
6543 	SYSCTL_ADD_U8(ctx_list, oid_list, OID_AUTO, "separate_txrx",
6544 		       CTLFLAG_RDTUN, &ctx->ifc_sysctl_separate_txrx, 0,
6545 		       "use separate cores for TX and RX");
6546 
6547 	/* XXX change for per-queue sizes */
6548 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_ntxds",
6549 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NTXD_HANDLER,
6550                        mp_ndesc_handler, "A",
6551 		       "list of # of TX descriptors to use, 0 = use default #");
6552 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_nrxds",
6553 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NRXD_HANDLER,
6554                        mp_ndesc_handler, "A",
6555 		       "list of # of RX descriptors to use, 0 = use default #");
6556 }
6557 
6558 static void
6559 iflib_add_device_sysctl_post(if_ctx_t ctx)
6560 {
6561 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6562 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6563         device_t dev = iflib_get_dev(ctx);
6564 	struct sysctl_oid_list *child;
6565 	struct sysctl_ctx_list *ctx_list;
6566 	iflib_fl_t fl;
6567 	iflib_txq_t txq;
6568 	iflib_rxq_t rxq;
6569 	int i, j;
6570 	char namebuf[NAME_BUFLEN];
6571 	char *qfmt;
6572 	struct sysctl_oid *queue_node, *fl_node, *node;
6573 	struct sysctl_oid_list *queue_list, *fl_list;
6574 	ctx_list = device_get_sysctl_ctx(dev);
6575 
6576 	node = ctx->ifc_sysctl_node;
6577 	child = SYSCTL_CHILDREN(node);
6578 
6579 	if (scctx->isc_ntxqsets > 100)
6580 		qfmt = "txq%03d";
6581 	else if (scctx->isc_ntxqsets > 10)
6582 		qfmt = "txq%02d";
6583 	else
6584 		qfmt = "txq%d";
6585 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
6586 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6587 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6588 					     CTLFLAG_RD, NULL, "Queue Name");
6589 		queue_list = SYSCTL_CHILDREN(queue_node);
6590 #if MEMORY_LOGGING
6591 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued",
6592 				CTLFLAG_RD,
6593 				&txq->ift_dequeued, "total mbufs freed");
6594 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued",
6595 				CTLFLAG_RD,
6596 				&txq->ift_enqueued, "total mbufs enqueued");
6597 #endif
6598 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag",
6599 				   CTLFLAG_RD,
6600 				   &txq->ift_mbuf_defrag, "# of times m_defrag was called");
6601 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "m_pullups",
6602 				   CTLFLAG_RD,
6603 				   &txq->ift_pullups, "# of times m_pullup was called");
6604 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag_failed",
6605 				   CTLFLAG_RD,
6606 				   &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed");
6607 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_desc_avail",
6608 				   CTLFLAG_RD,
6609 				   &txq->ift_no_desc_avail, "# of times no descriptors were available");
6610 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "tx_map_failed",
6611 				   CTLFLAG_RD,
6612 				   &txq->ift_map_failed, "# of times DMA map failed");
6613 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txd_encap_efbig",
6614 				   CTLFLAG_RD,
6615 				   &txq->ift_txd_encap_efbig, "# of times txd_encap returned EFBIG");
6616 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_tx_dma_setup",
6617 				   CTLFLAG_RD,
6618 				   &txq->ift_no_tx_dma_setup, "# of times map failed for other than EFBIG");
6619 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx",
6620 				   CTLFLAG_RD,
6621 				   &txq->ift_pidx, 1, "Producer Index");
6622 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx",
6623 				   CTLFLAG_RD,
6624 				   &txq->ift_cidx, 1, "Consumer Index");
6625 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx_processed",
6626 				   CTLFLAG_RD,
6627 				   &txq->ift_cidx_processed, 1, "Consumer Index seen by credit update");
6628 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use",
6629 				   CTLFLAG_RD,
6630 				   &txq->ift_in_use, 1, "descriptors in use");
6631 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_processed",
6632 				   CTLFLAG_RD,
6633 				   &txq->ift_processed, "descriptors procesed for clean");
6634 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned",
6635 				   CTLFLAG_RD,
6636 				   &txq->ift_cleaned, "total cleaned");
6637 		SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state",
6638 				CTLTYPE_STRING | CTLFLAG_RD, __DEVOLATILE(uint64_t *, &txq->ift_br->state),
6639 				0, mp_ring_state_handler, "A", "soft ring state");
6640 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_enqueues",
6641 				       CTLFLAG_RD, &txq->ift_br->enqueues,
6642 				       "# of enqueues to the mp_ring for this queue");
6643 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_drops",
6644 				       CTLFLAG_RD, &txq->ift_br->drops,
6645 				       "# of drops in the mp_ring for this queue");
6646 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_starts",
6647 				       CTLFLAG_RD, &txq->ift_br->starts,
6648 				       "# of normal consumer starts in the mp_ring for this queue");
6649 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_stalls",
6650 				       CTLFLAG_RD, &txq->ift_br->stalls,
6651 					       "# of consumer stalls in the mp_ring for this queue");
6652 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_restarts",
6653 			       CTLFLAG_RD, &txq->ift_br->restarts,
6654 				       "# of consumer restarts in the mp_ring for this queue");
6655 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_abdications",
6656 				       CTLFLAG_RD, &txq->ift_br->abdications,
6657 				       "# of consumer abdications in the mp_ring for this queue");
6658 	}
6659 
6660 	if (scctx->isc_nrxqsets > 100)
6661 		qfmt = "rxq%03d";
6662 	else if (scctx->isc_nrxqsets > 10)
6663 		qfmt = "rxq%02d";
6664 	else
6665 		qfmt = "rxq%d";
6666 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
6667 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6668 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6669 					     CTLFLAG_RD, NULL, "Queue Name");
6670 		queue_list = SYSCTL_CHILDREN(queue_node);
6671 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
6672 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_cidx",
6673 				       CTLFLAG_RD,
6674 				       &rxq->ifr_cq_cidx, 1, "Consumer Index");
6675 		}
6676 
6677 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
6678 			snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j);
6679 			fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list, OID_AUTO, namebuf,
6680 						     CTLFLAG_RD, NULL, "freelist Name");
6681 			fl_list = SYSCTL_CHILDREN(fl_node);
6682 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx",
6683 				       CTLFLAG_RD,
6684 				       &fl->ifl_pidx, 1, "Producer Index");
6685 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx",
6686 				       CTLFLAG_RD,
6687 				       &fl->ifl_cidx, 1, "Consumer Index");
6688 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits",
6689 				       CTLFLAG_RD,
6690 				       &fl->ifl_credits, 1, "credits available");
6691 #if MEMORY_LOGGING
6692 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_enqueued",
6693 					CTLFLAG_RD,
6694 					&fl->ifl_m_enqueued, "mbufs allocated");
6695 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_dequeued",
6696 					CTLFLAG_RD,
6697 					&fl->ifl_m_dequeued, "mbufs freed");
6698 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_enqueued",
6699 					CTLFLAG_RD,
6700 					&fl->ifl_cl_enqueued, "clusters allocated");
6701 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_dequeued",
6702 					CTLFLAG_RD,
6703 					&fl->ifl_cl_dequeued, "clusters freed");
6704 #endif
6705 
6706 		}
6707 	}
6708 
6709 }
6710 
6711 void
6712 iflib_request_reset(if_ctx_t ctx)
6713 {
6714 
6715 	STATE_LOCK(ctx);
6716 	ctx->ifc_flags |= IFC_DO_RESET;
6717 	STATE_UNLOCK(ctx);
6718 }
6719 
6720 #ifndef __NO_STRICT_ALIGNMENT
6721 static struct mbuf *
6722 iflib_fixup_rx(struct mbuf *m)
6723 {
6724 	struct mbuf *n;
6725 
6726 	if (m->m_len <= (MCLBYTES - ETHER_HDR_LEN)) {
6727 		bcopy(m->m_data, m->m_data + ETHER_HDR_LEN, m->m_len);
6728 		m->m_data += ETHER_HDR_LEN;
6729 		n = m;
6730 	} else {
6731 		MGETHDR(n, M_NOWAIT, MT_DATA);
6732 		if (n == NULL) {
6733 			m_freem(m);
6734 			return (NULL);
6735 		}
6736 		bcopy(m->m_data, n->m_data, ETHER_HDR_LEN);
6737 		m->m_data += ETHER_HDR_LEN;
6738 		m->m_len -= ETHER_HDR_LEN;
6739 		n->m_len = ETHER_HDR_LEN;
6740 		M_MOVE_PKTHDR(n, m);
6741 		n->m_next = m;
6742 	}
6743 	return (n);
6744 }
6745 #endif
6746 
6747 #ifdef DEBUGNET
6748 static void
6749 iflib_debugnet_init(if_t ifp, int *nrxr, int *ncl, int *clsize)
6750 {
6751 	if_ctx_t ctx;
6752 
6753 	ctx = if_getsoftc(ifp);
6754 	CTX_LOCK(ctx);
6755 	*nrxr = NRXQSETS(ctx);
6756 	*ncl = ctx->ifc_rxqs[0].ifr_fl->ifl_size;
6757 	*clsize = ctx->ifc_rxqs[0].ifr_fl->ifl_buf_size;
6758 	CTX_UNLOCK(ctx);
6759 }
6760 
6761 static void
6762 iflib_debugnet_event(if_t ifp, enum debugnet_ev event)
6763 {
6764 	if_ctx_t ctx;
6765 	if_softc_ctx_t scctx;
6766 	iflib_fl_t fl;
6767 	iflib_rxq_t rxq;
6768 	int i, j;
6769 
6770 	ctx = if_getsoftc(ifp);
6771 	scctx = &ctx->ifc_softc_ctx;
6772 
6773 	switch (event) {
6774 	case DEBUGNET_START:
6775 		for (i = 0; i < scctx->isc_nrxqsets; i++) {
6776 			rxq = &ctx->ifc_rxqs[i];
6777 			for (j = 0; j < rxq->ifr_nfl; j++) {
6778 				fl = rxq->ifr_fl;
6779 				fl->ifl_zone = m_getzone(fl->ifl_buf_size);
6780 			}
6781 		}
6782 		iflib_no_tx_batch = 1;
6783 		break;
6784 	default:
6785 		break;
6786 	}
6787 }
6788 
6789 static int
6790 iflib_debugnet_transmit(if_t ifp, struct mbuf *m)
6791 {
6792 	if_ctx_t ctx;
6793 	iflib_txq_t txq;
6794 	int error;
6795 
6796 	ctx = if_getsoftc(ifp);
6797 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
6798 	    IFF_DRV_RUNNING)
6799 		return (EBUSY);
6800 
6801 	txq = &ctx->ifc_txqs[0];
6802 	error = iflib_encap(txq, &m);
6803 	if (error == 0)
6804 		(void)iflib_txd_db_check(ctx, txq, true, txq->ift_in_use);
6805 	return (error);
6806 }
6807 
6808 static int
6809 iflib_debugnet_poll(if_t ifp, int count)
6810 {
6811 	if_ctx_t ctx;
6812 	if_softc_ctx_t scctx;
6813 	iflib_txq_t txq;
6814 	int i;
6815 
6816 	ctx = if_getsoftc(ifp);
6817 	scctx = &ctx->ifc_softc_ctx;
6818 
6819 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
6820 	    IFF_DRV_RUNNING)
6821 		return (EBUSY);
6822 
6823 	txq = &ctx->ifc_txqs[0];
6824 	(void)iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
6825 
6826 	for (i = 0; i < scctx->isc_nrxqsets; i++)
6827 		(void)iflib_rxeof(&ctx->ifc_rxqs[i], 16 /* XXX */);
6828 	return (0);
6829 }
6830 #endif /* DEBUGNET */
6831