xref: /freebsd/sys/net/iflib.c (revision b9f654b163bce26de79705e77b872427c9f2afa1)
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/pfil.h>
63 #include <net/vnet.h>
64 
65 #include <netinet/in.h>
66 #include <netinet/in_pcb.h>
67 #include <netinet/tcp_lro.h>
68 #include <netinet/in_systm.h>
69 #include <netinet/if_ether.h>
70 #include <netinet/ip.h>
71 #include <netinet/ip6.h>
72 #include <netinet/tcp.h>
73 #include <netinet/ip_var.h>
74 #include <netinet/netdump/netdump.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_init_locked(if_ctx_t ctx);
706 static void iflib_add_device_sysctl_pre(if_ctx_t ctx);
707 static void iflib_add_device_sysctl_post(if_ctx_t ctx);
708 static void iflib_ifmp_purge(iflib_txq_t txq);
709 static void _iflib_pre_assert(if_softc_ctx_t scctx);
710 static void iflib_if_init_locked(if_ctx_t ctx);
711 static void iflib_free_intr_mem(if_ctx_t ctx);
712 #ifndef __NO_STRICT_ALIGNMENT
713 static struct mbuf * iflib_fixup_rx(struct mbuf *m);
714 #endif
715 
716 static SLIST_HEAD(cpu_offset_list, cpu_offset) cpu_offsets =
717     SLIST_HEAD_INITIALIZER(cpu_offsets);
718 struct cpu_offset {
719 	SLIST_ENTRY(cpu_offset) entries;
720 	cpuset_t	set;
721 	unsigned int	refcount;
722 	uint16_t	offset;
723 };
724 static struct mtx cpu_offset_mtx;
725 MTX_SYSINIT(iflib_cpu_offset, &cpu_offset_mtx, "iflib_cpu_offset lock",
726     MTX_DEF);
727 
728 NETDUMP_DEFINE(iflib);
729 
730 #ifdef DEV_NETMAP
731 #include <sys/selinfo.h>
732 #include <net/netmap.h>
733 #include <dev/netmap/netmap_kern.h>
734 
735 MODULE_DEPEND(iflib, netmap, 1, 1, 1);
736 
737 static int netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, uint32_t nm_i, bool init);
738 
739 /*
740  * device-specific sysctl variables:
741  *
742  * iflib_crcstrip: 0: keep CRC in rx frames (default), 1: strip it.
743  *	During regular operations the CRC is stripped, but on some
744  *	hardware reception of frames not multiple of 64 is slower,
745  *	so using crcstrip=0 helps in benchmarks.
746  *
747  * iflib_rx_miss, iflib_rx_miss_bufs:
748  *	count packets that might be missed due to lost interrupts.
749  */
750 SYSCTL_DECL(_dev_netmap);
751 /*
752  * The xl driver by default strips CRCs and we do not override it.
753  */
754 
755 int iflib_crcstrip = 1;
756 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_crcstrip,
757     CTLFLAG_RW, &iflib_crcstrip, 1, "strip CRC on RX frames");
758 
759 int iflib_rx_miss, iflib_rx_miss_bufs;
760 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss,
761     CTLFLAG_RW, &iflib_rx_miss, 0, "potentially missed RX intr");
762 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss_bufs,
763     CTLFLAG_RW, &iflib_rx_miss_bufs, 0, "potentially missed RX intr bufs");
764 
765 /*
766  * Register/unregister. We are already under netmap lock.
767  * Only called on the first register or the last unregister.
768  */
769 static int
770 iflib_netmap_register(struct netmap_adapter *na, int onoff)
771 {
772 	if_t ifp = na->ifp;
773 	if_ctx_t ctx = ifp->if_softc;
774 	int status;
775 
776 	CTX_LOCK(ctx);
777 	IFDI_INTR_DISABLE(ctx);
778 
779 	/* Tell the stack that the interface is no longer active */
780 	ifp->if_drv_flags &= ~(IFF_DRV_RUNNING | IFF_DRV_OACTIVE);
781 
782 	if (!CTX_IS_VF(ctx))
783 		IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip);
784 
785 	/* enable or disable flags and callbacks in na and ifp */
786 	if (onoff) {
787 		nm_set_native_flags(na);
788 	} else {
789 		nm_clear_native_flags(na);
790 	}
791 	iflib_stop(ctx);
792 	iflib_init_locked(ctx);
793 	IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip); // XXX why twice ?
794 	status = ifp->if_drv_flags & IFF_DRV_RUNNING ? 0 : 1;
795 	if (status)
796 		nm_clear_native_flags(na);
797 	CTX_UNLOCK(ctx);
798 	return (status);
799 }
800 
801 static int
802 netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, uint32_t nm_i, bool init)
803 {
804 	struct netmap_adapter *na = kring->na;
805 	u_int const lim = kring->nkr_num_slots - 1;
806 	u_int head = kring->rhead;
807 	struct netmap_ring *ring = kring->ring;
808 	bus_dmamap_t *map;
809 	struct if_rxd_update iru;
810 	if_ctx_t ctx = rxq->ifr_ctx;
811 	iflib_fl_t fl = &rxq->ifr_fl[0];
812 	uint32_t refill_pidx, nic_i;
813 #if IFLIB_DEBUG_COUNTERS
814 	int rf_count = 0;
815 #endif
816 
817 	if (nm_i == head && __predict_true(!init))
818 		return 0;
819 	iru_init(&iru, rxq, 0 /* flid */);
820 	map = fl->ifl_sds.ifsd_map;
821 	refill_pidx = netmap_idx_k2n(kring, nm_i);
822 	/*
823 	 * IMPORTANT: we must leave one free slot in the ring,
824 	 * so move head back by one unit
825 	 */
826 	head = nm_prev(head, lim);
827 	nic_i = UINT_MAX;
828 	DBG_COUNTER_INC(fl_refills);
829 	while (nm_i != head) {
830 #if IFLIB_DEBUG_COUNTERS
831 		if (++rf_count == 9)
832 			DBG_COUNTER_INC(fl_refills_large);
833 #endif
834 		for (int tmp_pidx = 0; tmp_pidx < IFLIB_MAX_RX_REFRESH && nm_i != head; tmp_pidx++) {
835 			struct netmap_slot *slot = &ring->slot[nm_i];
836 			void *addr = PNMB(na, slot, &fl->ifl_bus_addrs[tmp_pidx]);
837 			uint32_t nic_i_dma = refill_pidx;
838 			nic_i = netmap_idx_k2n(kring, nm_i);
839 
840 			MPASS(tmp_pidx < IFLIB_MAX_RX_REFRESH);
841 
842 			if (addr == NETMAP_BUF_BASE(na)) /* bad buf */
843 			        return netmap_ring_reinit(kring);
844 
845 			fl->ifl_vm_addrs[tmp_pidx] = addr;
846 			if (__predict_false(init)) {
847 				netmap_load_map(na, fl->ifl_buf_tag,
848 				    map[nic_i], addr);
849 			} else if (slot->flags & NS_BUF_CHANGED) {
850 				/* buffer has changed, reload map */
851 				netmap_reload_map(na, fl->ifl_buf_tag,
852 				    map[nic_i], addr);
853 			}
854 			slot->flags &= ~NS_BUF_CHANGED;
855 
856 			nm_i = nm_next(nm_i, lim);
857 			fl->ifl_rxd_idxs[tmp_pidx] = nic_i = nm_next(nic_i, lim);
858 			if (nm_i != head && tmp_pidx < IFLIB_MAX_RX_REFRESH-1)
859 				continue;
860 
861 			iru.iru_pidx = refill_pidx;
862 			iru.iru_count = tmp_pidx+1;
863 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
864 			refill_pidx = nic_i;
865 			for (int n = 0; n < iru.iru_count; n++) {
866 				bus_dmamap_sync(fl->ifl_buf_tag, map[nic_i_dma],
867 						BUS_DMASYNC_PREREAD);
868 				/* XXX - change this to not use the netmap func*/
869 				nic_i_dma = nm_next(nic_i_dma, lim);
870 			}
871 		}
872 	}
873 	kring->nr_hwcur = head;
874 
875 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
876 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
877 	if (__predict_true(nic_i != UINT_MAX)) {
878 		ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id, nic_i);
879 		DBG_COUNTER_INC(rxd_flush);
880 	}
881 	return (0);
882 }
883 
884 /*
885  * Reconcile kernel and user view of the transmit ring.
886  *
887  * All information is in the kring.
888  * Userspace wants to send packets up to the one before kring->rhead,
889  * kernel knows kring->nr_hwcur is the first unsent packet.
890  *
891  * Here we push packets out (as many as possible), and possibly
892  * reclaim buffers from previously completed transmission.
893  *
894  * The caller (netmap) guarantees that there is only one instance
895  * running at any time. Any interference with other driver
896  * methods should be handled by the individual drivers.
897  */
898 static int
899 iflib_netmap_txsync(struct netmap_kring *kring, int flags)
900 {
901 	struct netmap_adapter *na = kring->na;
902 	if_t ifp = na->ifp;
903 	struct netmap_ring *ring = kring->ring;
904 	u_int nm_i;	/* index into the netmap kring */
905 	u_int nic_i;	/* index into the NIC ring */
906 	u_int n;
907 	u_int const lim = kring->nkr_num_slots - 1;
908 	u_int const head = kring->rhead;
909 	struct if_pkt_info pi;
910 
911 	/*
912 	 * interrupts on every tx packet are expensive so request
913 	 * them every half ring, or where NS_REPORT is set
914 	 */
915 	u_int report_frequency = kring->nkr_num_slots >> 1;
916 	/* device-specific */
917 	if_ctx_t ctx = ifp->if_softc;
918 	iflib_txq_t txq = &ctx->ifc_txqs[kring->ring_id];
919 
920 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
921 	    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
922 
923 	/*
924 	 * First part: process new packets to send.
925 	 * nm_i is the current index in the netmap kring,
926 	 * nic_i is the corresponding index in the NIC ring.
927 	 *
928 	 * If we have packets to send (nm_i != head)
929 	 * iterate over the netmap ring, fetch length and update
930 	 * the corresponding slot in the NIC ring. Some drivers also
931 	 * need to update the buffer's physical address in the NIC slot
932 	 * even NS_BUF_CHANGED is not set (PNMB computes the addresses).
933 	 *
934 	 * The netmap_reload_map() calls is especially expensive,
935 	 * even when (as in this case) the tag is 0, so do only
936 	 * when the buffer has actually changed.
937 	 *
938 	 * If possible do not set the report/intr bit on all slots,
939 	 * but only a few times per ring or when NS_REPORT is set.
940 	 *
941 	 * Finally, on 10G and faster drivers, it might be useful
942 	 * to prefetch the next slot and txr entry.
943 	 */
944 
945 	nm_i = kring->nr_hwcur;
946 	if (nm_i != head) {	/* we have new packets to send */
947 		pkt_info_zero(&pi);
948 		pi.ipi_segs = txq->ift_segs;
949 		pi.ipi_qsidx = kring->ring_id;
950 		nic_i = netmap_idx_k2n(kring, nm_i);
951 
952 		__builtin_prefetch(&ring->slot[nm_i]);
953 		__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i]);
954 		__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i]);
955 
956 		for (n = 0; nm_i != head; n++) {
957 			struct netmap_slot *slot = &ring->slot[nm_i];
958 			u_int len = slot->len;
959 			uint64_t paddr;
960 			void *addr = PNMB(na, slot, &paddr);
961 			int flags = (slot->flags & NS_REPORT ||
962 				nic_i == 0 || nic_i == report_frequency) ?
963 				IPI_TX_INTR : 0;
964 
965 			/* device-specific */
966 			pi.ipi_len = len;
967 			pi.ipi_segs[0].ds_addr = paddr;
968 			pi.ipi_segs[0].ds_len = len;
969 			pi.ipi_nsegs = 1;
970 			pi.ipi_ndescs = 0;
971 			pi.ipi_pidx = nic_i;
972 			pi.ipi_flags = flags;
973 
974 			/* Fill the slot in the NIC ring. */
975 			ctx->isc_txd_encap(ctx->ifc_softc, &pi);
976 			DBG_COUNTER_INC(tx_encap);
977 
978 			/* prefetch for next round */
979 			__builtin_prefetch(&ring->slot[nm_i + 1]);
980 			__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i + 1]);
981 			__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i + 1]);
982 
983 			NM_CHECK_ADDR_LEN(na, addr, len);
984 
985 			if (slot->flags & NS_BUF_CHANGED) {
986 				/* buffer has changed, reload map */
987 				netmap_reload_map(na, txq->ift_buf_tag,
988 				    txq->ift_sds.ifsd_map[nic_i], addr);
989 			}
990 			/* make sure changes to the buffer are synced */
991 			bus_dmamap_sync(txq->ift_buf_tag,
992 			    txq->ift_sds.ifsd_map[nic_i],
993 			    BUS_DMASYNC_PREWRITE);
994 
995 			slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED);
996 			nm_i = nm_next(nm_i, lim);
997 			nic_i = nm_next(nic_i, lim);
998 		}
999 		kring->nr_hwcur = nm_i;
1000 
1001 		/* synchronize the NIC ring */
1002 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1003 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1004 
1005 		/* (re)start the tx unit up to slot nic_i (excluded) */
1006 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, nic_i);
1007 	}
1008 
1009 	/*
1010 	 * Second part: reclaim buffers for completed transmissions.
1011 	 *
1012 	 * If there are unclaimed buffers, attempt to reclaim them.
1013 	 * If none are reclaimed, and TX IRQs are not in use, do an initial
1014 	 * minimal delay, then trigger the tx handler which will spin in the
1015 	 * group task queue.
1016 	 */
1017 	if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1018 		if (iflib_tx_credits_update(ctx, txq)) {
1019 			/* some tx completed, increment avail */
1020 			nic_i = txq->ift_cidx_processed;
1021 			kring->nr_hwtail = nm_prev(netmap_idx_n2k(kring, nic_i), lim);
1022 		}
1023 	}
1024 	if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ))
1025 		if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1026 			callout_reset_on(&txq->ift_timer, hz < 2000 ? 1 : hz / 1000,
1027 			    iflib_timer, txq, txq->ift_timer.c_cpu);
1028 	}
1029 	return (0);
1030 }
1031 
1032 /*
1033  * Reconcile kernel and user view of the receive ring.
1034  * Same as for the txsync, this routine must be efficient.
1035  * The caller guarantees a single invocations, but races against
1036  * the rest of the driver should be handled here.
1037  *
1038  * On call, kring->rhead is the first packet that userspace wants
1039  * to keep, and kring->rcur is the wakeup point.
1040  * The kernel has previously reported packets up to kring->rtail.
1041  *
1042  * If (flags & NAF_FORCE_READ) also check for incoming packets irrespective
1043  * of whether or not we received an interrupt.
1044  */
1045 static int
1046 iflib_netmap_rxsync(struct netmap_kring *kring, int flags)
1047 {
1048 	struct netmap_adapter *na = kring->na;
1049 	struct netmap_ring *ring = kring->ring;
1050 	if_t ifp = na->ifp;
1051 	iflib_fl_t fl;
1052 	uint32_t nm_i;	/* index into the netmap ring */
1053 	uint32_t nic_i;	/* index into the NIC ring */
1054 	u_int i, n;
1055 	u_int const lim = kring->nkr_num_slots - 1;
1056 	u_int const head = kring->rhead;
1057 	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
1058 	struct if_rxd_info ri;
1059 
1060 	if_ctx_t ctx = ifp->if_softc;
1061 	iflib_rxq_t rxq = &ctx->ifc_rxqs[kring->ring_id];
1062 	if (head > lim)
1063 		return netmap_ring_reinit(kring);
1064 
1065 	/*
1066 	 * XXX netmap_fl_refill() only ever (re)fills free list 0 so far.
1067 	 */
1068 
1069 	for (i = 0, fl = rxq->ifr_fl; i < rxq->ifr_nfl; i++, fl++) {
1070 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
1071 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1072 	}
1073 
1074 	/*
1075 	 * First part: import newly received packets.
1076 	 *
1077 	 * nm_i is the index of the next free slot in the netmap ring,
1078 	 * nic_i is the index of the next received packet in the NIC ring,
1079 	 * and they may differ in case if_init() has been called while
1080 	 * in netmap mode. For the receive ring we have
1081 	 *
1082 	 *	nic_i = rxr->next_check;
1083 	 *	nm_i = kring->nr_hwtail (previous)
1084 	 * and
1085 	 *	nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1086 	 *
1087 	 * rxr->next_check is set to 0 on a ring reinit
1088 	 */
1089 	if (netmap_no_pendintr || force_update) {
1090 		int crclen = iflib_crcstrip ? 0 : 4;
1091 		int error, avail;
1092 
1093 		for (i = 0; i < rxq->ifr_nfl; i++) {
1094 			fl = &rxq->ifr_fl[i];
1095 			nic_i = fl->ifl_cidx;
1096 			nm_i = netmap_idx_n2k(kring, nic_i);
1097 			avail = ctx->isc_rxd_available(ctx->ifc_softc,
1098 			    rxq->ifr_id, nic_i, USHRT_MAX);
1099 			for (n = 0; avail > 0; n++, avail--) {
1100 				rxd_info_zero(&ri);
1101 				ri.iri_frags = rxq->ifr_frags;
1102 				ri.iri_qsidx = kring->ring_id;
1103 				ri.iri_ifp = ctx->ifc_ifp;
1104 				ri.iri_cidx = nic_i;
1105 
1106 				error = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
1107 				ring->slot[nm_i].len = error ? 0 : ri.iri_len - crclen;
1108 				ring->slot[nm_i].flags = 0;
1109 				bus_dmamap_sync(fl->ifl_buf_tag,
1110 				    fl->ifl_sds.ifsd_map[nic_i], BUS_DMASYNC_POSTREAD);
1111 				nm_i = nm_next(nm_i, lim);
1112 				nic_i = nm_next(nic_i, lim);
1113 			}
1114 			if (n) { /* update the state variables */
1115 				if (netmap_no_pendintr && !force_update) {
1116 					/* diagnostics */
1117 					iflib_rx_miss ++;
1118 					iflib_rx_miss_bufs += n;
1119 				}
1120 				fl->ifl_cidx = nic_i;
1121 				kring->nr_hwtail = nm_i;
1122 			}
1123 			kring->nr_kflags &= ~NKR_PENDINTR;
1124 		}
1125 	}
1126 	/*
1127 	 * Second part: skip past packets that userspace has released.
1128 	 * (kring->nr_hwcur to head excluded),
1129 	 * and make the buffers available for reception.
1130 	 * As usual nm_i is the index in the netmap ring,
1131 	 * nic_i is the index in the NIC ring, and
1132 	 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1133 	 */
1134 	/* XXX not sure how this will work with multiple free lists */
1135 	nm_i = kring->nr_hwcur;
1136 
1137 	return (netmap_fl_refill(rxq, kring, nm_i, false));
1138 }
1139 
1140 static void
1141 iflib_netmap_intr(struct netmap_adapter *na, int onoff)
1142 {
1143 	if_ctx_t ctx = na->ifp->if_softc;
1144 
1145 	CTX_LOCK(ctx);
1146 	if (onoff) {
1147 		IFDI_INTR_ENABLE(ctx);
1148 	} else {
1149 		IFDI_INTR_DISABLE(ctx);
1150 	}
1151 	CTX_UNLOCK(ctx);
1152 }
1153 
1154 
1155 static int
1156 iflib_netmap_attach(if_ctx_t ctx)
1157 {
1158 	struct netmap_adapter na;
1159 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1160 
1161 	bzero(&na, sizeof(na));
1162 
1163 	na.ifp = ctx->ifc_ifp;
1164 	na.na_flags = NAF_BDG_MAYSLEEP;
1165 	MPASS(ctx->ifc_softc_ctx.isc_ntxqsets);
1166 	MPASS(ctx->ifc_softc_ctx.isc_nrxqsets);
1167 
1168 	na.num_tx_desc = scctx->isc_ntxd[0];
1169 	na.num_rx_desc = scctx->isc_nrxd[0];
1170 	na.nm_txsync = iflib_netmap_txsync;
1171 	na.nm_rxsync = iflib_netmap_rxsync;
1172 	na.nm_register = iflib_netmap_register;
1173 	na.nm_intr = iflib_netmap_intr;
1174 	na.num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
1175 	na.num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
1176 	return (netmap_attach(&na));
1177 }
1178 
1179 static void
1180 iflib_netmap_txq_init(if_ctx_t ctx, iflib_txq_t txq)
1181 {
1182 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1183 	struct netmap_slot *slot;
1184 
1185 	slot = netmap_reset(na, NR_TX, txq->ift_id, 0);
1186 	if (slot == NULL)
1187 		return;
1188 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxd[0]; i++) {
1189 
1190 		/*
1191 		 * In netmap mode, set the map for the packet buffer.
1192 		 * NOTE: Some drivers (not this one) also need to set
1193 		 * the physical buffer address in the NIC ring.
1194 		 * netmap_idx_n2k() maps a nic index, i, into the corresponding
1195 		 * netmap slot index, si
1196 		 */
1197 		int si = netmap_idx_n2k(na->tx_rings[txq->ift_id], i);
1198 		netmap_load_map(na, txq->ift_buf_tag, txq->ift_sds.ifsd_map[i],
1199 		    NMB(na, slot + si));
1200 	}
1201 }
1202 
1203 static void
1204 iflib_netmap_rxq_init(if_ctx_t ctx, iflib_rxq_t rxq)
1205 {
1206 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1207 	struct netmap_kring *kring = na->rx_rings[rxq->ifr_id];
1208 	struct netmap_slot *slot;
1209 	uint32_t nm_i;
1210 
1211 	slot = netmap_reset(na, NR_RX, rxq->ifr_id, 0);
1212 	if (slot == NULL)
1213 		return;
1214 	nm_i = netmap_idx_n2k(kring, 0);
1215 	netmap_fl_refill(rxq, kring, nm_i, true);
1216 }
1217 
1218 static void
1219 iflib_netmap_timer_adjust(if_ctx_t ctx, iflib_txq_t txq, uint32_t *reset_on)
1220 {
1221 	struct netmap_kring *kring;
1222 	uint16_t txqid;
1223 
1224 	txqid = txq->ift_id;
1225 	kring = NA(ctx->ifc_ifp)->tx_rings[txqid];
1226 
1227 	if (kring->nr_hwcur != nm_next(kring->nr_hwtail, kring->nkr_num_slots - 1)) {
1228 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1229 		    BUS_DMASYNC_POSTREAD);
1230 		if (ctx->isc_txd_credits_update(ctx->ifc_softc, txqid, false))
1231 			netmap_tx_irq(ctx->ifc_ifp, txqid);
1232 		if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ)) {
1233 			if (hz < 2000)
1234 				*reset_on = 1;
1235 			else
1236 				*reset_on = hz / 1000;
1237 		}
1238 	}
1239 }
1240 
1241 #define iflib_netmap_detach(ifp) netmap_detach(ifp)
1242 
1243 #else
1244 #define iflib_netmap_txq_init(ctx, txq)
1245 #define iflib_netmap_rxq_init(ctx, rxq)
1246 #define iflib_netmap_detach(ifp)
1247 
1248 #define iflib_netmap_attach(ctx) (0)
1249 #define netmap_rx_irq(ifp, qid, budget) (0)
1250 #define netmap_tx_irq(ifp, qid) do {} while (0)
1251 #define iflib_netmap_timer_adjust(ctx, txq, reset_on)
1252 #endif
1253 
1254 #if defined(__i386__) || defined(__amd64__)
1255 static __inline void
1256 prefetch(void *x)
1257 {
1258 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1259 }
1260 static __inline void
1261 prefetch2cachelines(void *x)
1262 {
1263 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1264 #if (CACHE_LINE_SIZE < 128)
1265 	__asm volatile("prefetcht0 %0" :: "m" (*(((unsigned long *)x)+CACHE_LINE_SIZE/(sizeof(unsigned long)))));
1266 #endif
1267 }
1268 #else
1269 #define prefetch(x)
1270 #define prefetch2cachelines(x)
1271 #endif
1272 
1273 static void
1274 iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid)
1275 {
1276 	iflib_fl_t fl;
1277 
1278 	fl = &rxq->ifr_fl[flid];
1279 	iru->iru_paddrs = fl->ifl_bus_addrs;
1280 	iru->iru_vaddrs = &fl->ifl_vm_addrs[0];
1281 	iru->iru_idxs = fl->ifl_rxd_idxs;
1282 	iru->iru_qsidx = rxq->ifr_id;
1283 	iru->iru_buf_size = fl->ifl_buf_size;
1284 	iru->iru_flidx = fl->ifl_id;
1285 }
1286 
1287 static void
1288 _iflib_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int err)
1289 {
1290 	if (err)
1291 		return;
1292 	*(bus_addr_t *) arg = segs[0].ds_addr;
1293 }
1294 
1295 int
1296 iflib_dma_alloc_align(if_ctx_t ctx, int size, int align, iflib_dma_info_t dma, int mapflags)
1297 {
1298 	int err;
1299 	device_t dev = ctx->ifc_dev;
1300 
1301 	err = bus_dma_tag_create(bus_get_dma_tag(dev),	/* parent */
1302 				align, 0,		/* alignment, bounds */
1303 				BUS_SPACE_MAXADDR,	/* lowaddr */
1304 				BUS_SPACE_MAXADDR,	/* highaddr */
1305 				NULL, NULL,		/* filter, filterarg */
1306 				size,			/* maxsize */
1307 				1,			/* nsegments */
1308 				size,			/* maxsegsize */
1309 				BUS_DMA_ALLOCNOW,	/* flags */
1310 				NULL,			/* lockfunc */
1311 				NULL,			/* lockarg */
1312 				&dma->idi_tag);
1313 	if (err) {
1314 		device_printf(dev,
1315 		    "%s: bus_dma_tag_create failed: %d\n",
1316 		    __func__, err);
1317 		goto fail_0;
1318 	}
1319 
1320 	err = bus_dmamem_alloc(dma->idi_tag, (void**) &dma->idi_vaddr,
1321 	    BUS_DMA_NOWAIT | BUS_DMA_COHERENT | BUS_DMA_ZERO, &dma->idi_map);
1322 	if (err) {
1323 		device_printf(dev,
1324 		    "%s: bus_dmamem_alloc(%ju) failed: %d\n",
1325 		    __func__, (uintmax_t)size, err);
1326 		goto fail_1;
1327 	}
1328 
1329 	dma->idi_paddr = IF_BAD_DMA;
1330 	err = bus_dmamap_load(dma->idi_tag, dma->idi_map, dma->idi_vaddr,
1331 	    size, _iflib_dmamap_cb, &dma->idi_paddr, mapflags | BUS_DMA_NOWAIT);
1332 	if (err || dma->idi_paddr == IF_BAD_DMA) {
1333 		device_printf(dev,
1334 		    "%s: bus_dmamap_load failed: %d\n",
1335 		    __func__, err);
1336 		goto fail_2;
1337 	}
1338 
1339 	dma->idi_size = size;
1340 	return (0);
1341 
1342 fail_2:
1343 	bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1344 fail_1:
1345 	bus_dma_tag_destroy(dma->idi_tag);
1346 fail_0:
1347 	dma->idi_tag = NULL;
1348 
1349 	return (err);
1350 }
1351 
1352 int
1353 iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags)
1354 {
1355 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1356 
1357 	KASSERT(sctx->isc_q_align != 0, ("alignment value not initialized"));
1358 
1359 	return (iflib_dma_alloc_align(ctx, size, sctx->isc_q_align, dma, mapflags));
1360 }
1361 
1362 int
1363 iflib_dma_alloc_multi(if_ctx_t ctx, int *sizes, iflib_dma_info_t *dmalist, int mapflags, int count)
1364 {
1365 	int i, err;
1366 	iflib_dma_info_t *dmaiter;
1367 
1368 	dmaiter = dmalist;
1369 	for (i = 0; i < count; i++, dmaiter++) {
1370 		if ((err = iflib_dma_alloc(ctx, sizes[i], *dmaiter, mapflags)) != 0)
1371 			break;
1372 	}
1373 	if (err)
1374 		iflib_dma_free_multi(dmalist, i);
1375 	return (err);
1376 }
1377 
1378 void
1379 iflib_dma_free(iflib_dma_info_t dma)
1380 {
1381 	if (dma->idi_tag == NULL)
1382 		return;
1383 	if (dma->idi_paddr != IF_BAD_DMA) {
1384 		bus_dmamap_sync(dma->idi_tag, dma->idi_map,
1385 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1386 		bus_dmamap_unload(dma->idi_tag, dma->idi_map);
1387 		dma->idi_paddr = IF_BAD_DMA;
1388 	}
1389 	if (dma->idi_vaddr != NULL) {
1390 		bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1391 		dma->idi_vaddr = NULL;
1392 	}
1393 	bus_dma_tag_destroy(dma->idi_tag);
1394 	dma->idi_tag = NULL;
1395 }
1396 
1397 void
1398 iflib_dma_free_multi(iflib_dma_info_t *dmalist, int count)
1399 {
1400 	int i;
1401 	iflib_dma_info_t *dmaiter = dmalist;
1402 
1403 	for (i = 0; i < count; i++, dmaiter++)
1404 		iflib_dma_free(*dmaiter);
1405 }
1406 
1407 #ifdef EARLY_AP_STARTUP
1408 static const int iflib_started = 1;
1409 #else
1410 /*
1411  * We used to abuse the smp_started flag to decide if the queues have been
1412  * fully initialized (by late taskqgroup_adjust() calls in a SYSINIT()).
1413  * That gave bad races, since the SYSINIT() runs strictly after smp_started
1414  * is set.  Run a SYSINIT() strictly after that to just set a usable
1415  * completion flag.
1416  */
1417 
1418 static int iflib_started;
1419 
1420 static void
1421 iflib_record_started(void *arg)
1422 {
1423 	iflib_started = 1;
1424 }
1425 
1426 SYSINIT(iflib_record_started, SI_SUB_SMP + 1, SI_ORDER_FIRST,
1427 	iflib_record_started, NULL);
1428 #endif
1429 
1430 static int
1431 iflib_fast_intr(void *arg)
1432 {
1433 	iflib_filter_info_t info = arg;
1434 	struct grouptask *gtask = info->ifi_task;
1435 	int result;
1436 
1437 	if (!iflib_started)
1438 		return (FILTER_STRAY);
1439 
1440 	DBG_COUNTER_INC(fast_intrs);
1441 	if (info->ifi_filter != NULL) {
1442 		result = info->ifi_filter(info->ifi_filter_arg);
1443 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1444 			return (result);
1445 	}
1446 
1447 	GROUPTASK_ENQUEUE(gtask);
1448 	return (FILTER_HANDLED);
1449 }
1450 
1451 static int
1452 iflib_fast_intr_rxtx(void *arg)
1453 {
1454 	iflib_filter_info_t info = arg;
1455 	struct grouptask *gtask = info->ifi_task;
1456 	if_ctx_t ctx;
1457 	iflib_rxq_t rxq = (iflib_rxq_t)info->ifi_ctx;
1458 	iflib_txq_t txq;
1459 	void *sc;
1460 	int i, cidx, result;
1461 	qidx_t txqid;
1462 	bool intr_enable, intr_legacy;
1463 
1464 	if (!iflib_started)
1465 		return (FILTER_STRAY);
1466 
1467 	DBG_COUNTER_INC(fast_intrs);
1468 	if (info->ifi_filter != NULL) {
1469 		result = info->ifi_filter(info->ifi_filter_arg);
1470 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1471 			return (result);
1472 	}
1473 
1474 	ctx = rxq->ifr_ctx;
1475 	sc = ctx->ifc_softc;
1476 	intr_enable = false;
1477 	intr_legacy = !!(ctx->ifc_flags & IFC_LEGACY);
1478 	MPASS(rxq->ifr_ntxqirq);
1479 	for (i = 0; i < rxq->ifr_ntxqirq; i++) {
1480 		txqid = rxq->ifr_txqid[i];
1481 		txq = &ctx->ifc_txqs[txqid];
1482 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1483 		    BUS_DMASYNC_POSTREAD);
1484 		if (!ctx->isc_txd_credits_update(sc, txqid, false)) {
1485 			if (intr_legacy)
1486 				intr_enable = true;
1487 			else
1488 				IFDI_TX_QUEUE_INTR_ENABLE(ctx, txqid);
1489 			continue;
1490 		}
1491 		GROUPTASK_ENQUEUE(&txq->ift_task);
1492 	}
1493 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_RXCQ)
1494 		cidx = rxq->ifr_cq_cidx;
1495 	else
1496 		cidx = rxq->ifr_fl[0].ifl_cidx;
1497 	if (iflib_rxd_avail(ctx, rxq, cidx, 1))
1498 		GROUPTASK_ENQUEUE(gtask);
1499 	else {
1500 		if (intr_legacy)
1501 			intr_enable = true;
1502 		else
1503 			IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
1504 		DBG_COUNTER_INC(rx_intr_enables);
1505 	}
1506 	if (intr_enable)
1507 		IFDI_INTR_ENABLE(ctx);
1508 	return (FILTER_HANDLED);
1509 }
1510 
1511 
1512 static int
1513 iflib_fast_intr_ctx(void *arg)
1514 {
1515 	iflib_filter_info_t info = arg;
1516 	struct grouptask *gtask = info->ifi_task;
1517 	int result;
1518 
1519 	if (!iflib_started)
1520 		return (FILTER_STRAY);
1521 
1522 	DBG_COUNTER_INC(fast_intrs);
1523 	if (info->ifi_filter != NULL) {
1524 		result = info->ifi_filter(info->ifi_filter_arg);
1525 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1526 			return (result);
1527 	}
1528 
1529 	GROUPTASK_ENQUEUE(gtask);
1530 	return (FILTER_HANDLED);
1531 }
1532 
1533 static int
1534 _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
1535 		 driver_filter_t filter, driver_intr_t handler, void *arg,
1536 		 const char *name)
1537 {
1538 	struct resource *res;
1539 	void *tag = NULL;
1540 	device_t dev = ctx->ifc_dev;
1541 	int flags, i, rc;
1542 
1543 	flags = RF_ACTIVE;
1544 	if (ctx->ifc_flags & IFC_LEGACY)
1545 		flags |= RF_SHAREABLE;
1546 	MPASS(rid < 512);
1547 	i = rid;
1548 	res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &i, flags);
1549 	if (res == NULL) {
1550 		device_printf(dev,
1551 		    "failed to allocate IRQ for rid %d, name %s.\n", rid, name);
1552 		return (ENOMEM);
1553 	}
1554 	irq->ii_res = res;
1555 	KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL"));
1556 	rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET,
1557 						filter, handler, arg, &tag);
1558 	if (rc != 0) {
1559 		device_printf(dev,
1560 		    "failed to setup interrupt for rid %d, name %s: %d\n",
1561 					  rid, name ? name : "unknown", rc);
1562 		return (rc);
1563 	} else if (name)
1564 		bus_describe_intr(dev, res, tag, "%s", name);
1565 
1566 	irq->ii_tag = tag;
1567 	return (0);
1568 }
1569 
1570 /*********************************************************************
1571  *
1572  *  Allocate DMA resources for TX buffers as well as memory for the TX
1573  *  mbuf map.  TX DMA maps (non-TSO/TSO) and TX mbuf map are kept in a
1574  *  iflib_sw_tx_desc_array structure, storing all the information that
1575  *  is needed to transmit a packet on the wire.  This is called only
1576  *  once at attach, setup is done every reset.
1577  *
1578  **********************************************************************/
1579 static int
1580 iflib_txsd_alloc(iflib_txq_t txq)
1581 {
1582 	if_ctx_t ctx = txq->ift_ctx;
1583 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1584 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1585 	device_t dev = ctx->ifc_dev;
1586 	bus_size_t tsomaxsize;
1587 	int err, nsegments, ntsosegments;
1588 	bool tso;
1589 
1590 	nsegments = scctx->isc_tx_nsegments;
1591 	ntsosegments = scctx->isc_tx_tso_segments_max;
1592 	tsomaxsize = scctx->isc_tx_tso_size_max;
1593 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_VLAN_MTU)
1594 		tsomaxsize += sizeof(struct ether_vlan_header);
1595 	MPASS(scctx->isc_ntxd[0] > 0);
1596 	MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0);
1597 	MPASS(nsegments > 0);
1598 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) {
1599 		MPASS(ntsosegments > 0);
1600 		MPASS(sctx->isc_tso_maxsize >= tsomaxsize);
1601 	}
1602 
1603 	/*
1604 	 * Set up DMA tags for TX buffers.
1605 	 */
1606 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1607 			       1, 0,			/* alignment, bounds */
1608 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1609 			       BUS_SPACE_MAXADDR,	/* highaddr */
1610 			       NULL, NULL,		/* filter, filterarg */
1611 			       sctx->isc_tx_maxsize,		/* maxsize */
1612 			       nsegments,	/* nsegments */
1613 			       sctx->isc_tx_maxsegsize,	/* maxsegsize */
1614 			       0,			/* flags */
1615 			       NULL,			/* lockfunc */
1616 			       NULL,			/* lockfuncarg */
1617 			       &txq->ift_buf_tag))) {
1618 		device_printf(dev,"Unable to allocate TX DMA tag: %d\n", err);
1619 		device_printf(dev,"maxsize: %ju nsegments: %d maxsegsize: %ju\n",
1620 		    (uintmax_t)sctx->isc_tx_maxsize, nsegments, (uintmax_t)sctx->isc_tx_maxsegsize);
1621 		goto fail;
1622 	}
1623 	tso = (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) != 0;
1624 	if (tso && (err = bus_dma_tag_create(bus_get_dma_tag(dev),
1625 			       1, 0,			/* alignment, bounds */
1626 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1627 			       BUS_SPACE_MAXADDR,	/* highaddr */
1628 			       NULL, NULL,		/* filter, filterarg */
1629 			       tsomaxsize,		/* maxsize */
1630 			       ntsosegments,	/* nsegments */
1631 			       sctx->isc_tso_maxsegsize,/* maxsegsize */
1632 			       0,			/* flags */
1633 			       NULL,			/* lockfunc */
1634 			       NULL,			/* lockfuncarg */
1635 			       &txq->ift_tso_buf_tag))) {
1636 		device_printf(dev, "Unable to allocate TSO TX DMA tag: %d\n",
1637 		    err);
1638 		goto fail;
1639 	}
1640 
1641 	/* Allocate memory for the TX mbuf map. */
1642 	if (!(txq->ift_sds.ifsd_m =
1643 	    (struct mbuf **) malloc(sizeof(struct mbuf *) *
1644 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1645 		device_printf(dev, "Unable to allocate TX mbuf map memory\n");
1646 		err = ENOMEM;
1647 		goto fail;
1648 	}
1649 
1650 	/*
1651 	 * Create the DMA maps for TX buffers.
1652 	 */
1653 	if ((txq->ift_sds.ifsd_map = (bus_dmamap_t *)malloc(
1654 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1655 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1656 		device_printf(dev,
1657 		    "Unable to allocate TX buffer DMA map memory\n");
1658 		err = ENOMEM;
1659 		goto fail;
1660 	}
1661 	if (tso && (txq->ift_sds.ifsd_tso_map = (bus_dmamap_t *)malloc(
1662 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1663 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1664 		device_printf(dev,
1665 		    "Unable to allocate TSO TX buffer map memory\n");
1666 		err = ENOMEM;
1667 		goto fail;
1668 	}
1669 	for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) {
1670 		err = bus_dmamap_create(txq->ift_buf_tag, 0,
1671 		    &txq->ift_sds.ifsd_map[i]);
1672 		if (err != 0) {
1673 			device_printf(dev, "Unable to create TX DMA map\n");
1674 			goto fail;
1675 		}
1676 		if (!tso)
1677 			continue;
1678 		err = bus_dmamap_create(txq->ift_tso_buf_tag, 0,
1679 		    &txq->ift_sds.ifsd_tso_map[i]);
1680 		if (err != 0) {
1681 			device_printf(dev, "Unable to create TSO TX DMA map\n");
1682 			goto fail;
1683 		}
1684 	}
1685 	return (0);
1686 fail:
1687 	/* We free all, it handles case where we are in the middle */
1688 	iflib_tx_structures_free(ctx);
1689 	return (err);
1690 }
1691 
1692 static void
1693 iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i)
1694 {
1695 	bus_dmamap_t map;
1696 
1697 	map = NULL;
1698 	if (txq->ift_sds.ifsd_map != NULL)
1699 		map = txq->ift_sds.ifsd_map[i];
1700 	if (map != NULL) {
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 	map = NULL;
1708 	if (txq->ift_sds.ifsd_tso_map != NULL)
1709 		map = txq->ift_sds.ifsd_tso_map[i];
1710 	if (map != NULL) {
1711 		bus_dmamap_sync(txq->ift_tso_buf_tag, map,
1712 		    BUS_DMASYNC_POSTWRITE);
1713 		bus_dmamap_unload(txq->ift_tso_buf_tag, map);
1714 		bus_dmamap_destroy(txq->ift_tso_buf_tag, map);
1715 		txq->ift_sds.ifsd_tso_map[i] = NULL;
1716 	}
1717 }
1718 
1719 static void
1720 iflib_txq_destroy(iflib_txq_t txq)
1721 {
1722 	if_ctx_t ctx = txq->ift_ctx;
1723 
1724 	for (int i = 0; i < txq->ift_size; i++)
1725 		iflib_txsd_destroy(ctx, txq, i);
1726 	if (txq->ift_sds.ifsd_map != NULL) {
1727 		free(txq->ift_sds.ifsd_map, M_IFLIB);
1728 		txq->ift_sds.ifsd_map = NULL;
1729 	}
1730 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1731 		free(txq->ift_sds.ifsd_tso_map, M_IFLIB);
1732 		txq->ift_sds.ifsd_tso_map = NULL;
1733 	}
1734 	if (txq->ift_sds.ifsd_m != NULL) {
1735 		free(txq->ift_sds.ifsd_m, M_IFLIB);
1736 		txq->ift_sds.ifsd_m = NULL;
1737 	}
1738 	if (txq->ift_buf_tag != NULL) {
1739 		bus_dma_tag_destroy(txq->ift_buf_tag);
1740 		txq->ift_buf_tag = NULL;
1741 	}
1742 	if (txq->ift_tso_buf_tag != NULL) {
1743 		bus_dma_tag_destroy(txq->ift_tso_buf_tag);
1744 		txq->ift_tso_buf_tag = NULL;
1745 	}
1746 }
1747 
1748 static void
1749 iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i)
1750 {
1751 	struct mbuf **mp;
1752 
1753 	mp = &txq->ift_sds.ifsd_m[i];
1754 	if (*mp == NULL)
1755 		return;
1756 
1757 	if (txq->ift_sds.ifsd_map != NULL) {
1758 		bus_dmamap_sync(txq->ift_buf_tag,
1759 		    txq->ift_sds.ifsd_map[i], BUS_DMASYNC_POSTWRITE);
1760 		bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[i]);
1761 	}
1762 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1763 		bus_dmamap_sync(txq->ift_tso_buf_tag,
1764 		    txq->ift_sds.ifsd_tso_map[i], BUS_DMASYNC_POSTWRITE);
1765 		bus_dmamap_unload(txq->ift_tso_buf_tag,
1766 		    txq->ift_sds.ifsd_tso_map[i]);
1767 	}
1768 	m_free(*mp);
1769 	DBG_COUNTER_INC(tx_frees);
1770 	*mp = NULL;
1771 }
1772 
1773 static int
1774 iflib_txq_setup(iflib_txq_t txq)
1775 {
1776 	if_ctx_t ctx = txq->ift_ctx;
1777 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1778 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1779 	iflib_dma_info_t di;
1780 	int i;
1781 
1782 	/* Set number of descriptors available */
1783 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
1784 	/* XXX make configurable */
1785 	txq->ift_update_freq = IFLIB_DEFAULT_TX_UPDATE_FREQ;
1786 
1787 	/* Reset indices */
1788 	txq->ift_cidx_processed = 0;
1789 	txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0;
1790 	txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset];
1791 
1792 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1793 		bzero((void *)di->idi_vaddr, di->idi_size);
1794 
1795 	IFDI_TXQ_SETUP(ctx, txq->ift_id);
1796 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1797 		bus_dmamap_sync(di->idi_tag, di->idi_map,
1798 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1799 	return (0);
1800 }
1801 
1802 /*********************************************************************
1803  *
1804  *  Allocate DMA resources for RX buffers as well as memory for the RX
1805  *  mbuf map, direct RX cluster pointer map and RX cluster bus address
1806  *  map.  RX DMA map, RX mbuf map, direct RX cluster pointer map and
1807  *  RX cluster map are kept in a iflib_sw_rx_desc_array structure.
1808  *  Since we use use one entry in iflib_sw_rx_desc_array per received
1809  *  packet, the maximum number of entries we'll need is equal to the
1810  *  number of hardware receive descriptors that we've allocated.
1811  *
1812  **********************************************************************/
1813 static int
1814 iflib_rxsd_alloc(iflib_rxq_t rxq)
1815 {
1816 	if_ctx_t ctx = rxq->ifr_ctx;
1817 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1818 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1819 	device_t dev = ctx->ifc_dev;
1820 	iflib_fl_t fl;
1821 	int			err;
1822 
1823 	MPASS(scctx->isc_nrxd[0] > 0);
1824 	MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0);
1825 
1826 	fl = rxq->ifr_fl;
1827 	for (int i = 0; i <  rxq->ifr_nfl; i++, fl++) {
1828 		fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */
1829 		/* Set up DMA tag for RX buffers. */
1830 		err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1831 					 1, 0,			/* alignment, bounds */
1832 					 BUS_SPACE_MAXADDR,	/* lowaddr */
1833 					 BUS_SPACE_MAXADDR,	/* highaddr */
1834 					 NULL, NULL,		/* filter, filterarg */
1835 					 sctx->isc_rx_maxsize,	/* maxsize */
1836 					 sctx->isc_rx_nsegments,	/* nsegments */
1837 					 sctx->isc_rx_maxsegsize,	/* maxsegsize */
1838 					 0,			/* flags */
1839 					 NULL,			/* lockfunc */
1840 					 NULL,			/* lockarg */
1841 					 &fl->ifl_buf_tag);
1842 		if (err) {
1843 			device_printf(dev,
1844 			    "Unable to allocate RX DMA tag: %d\n", err);
1845 			goto fail;
1846 		}
1847 
1848 		/* Allocate memory for the RX mbuf map. */
1849 		if (!(fl->ifl_sds.ifsd_m =
1850 		      (struct mbuf **) malloc(sizeof(struct mbuf *) *
1851 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1852 			device_printf(dev,
1853 			    "Unable to allocate RX mbuf map memory\n");
1854 			err = ENOMEM;
1855 			goto fail;
1856 		}
1857 
1858 		/* Allocate memory for the direct RX cluster pointer map. */
1859 		if (!(fl->ifl_sds.ifsd_cl =
1860 		      (caddr_t *) malloc(sizeof(caddr_t) *
1861 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1862 			device_printf(dev,
1863 			    "Unable to allocate RX cluster map memory\n");
1864 			err = ENOMEM;
1865 			goto fail;
1866 		}
1867 
1868 		/* Allocate memory for the RX cluster bus address map. */
1869 		if (!(fl->ifl_sds.ifsd_ba =
1870 		      (bus_addr_t *) malloc(sizeof(bus_addr_t) *
1871 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1872 			device_printf(dev,
1873 			    "Unable to allocate RX bus address map memory\n");
1874 			err = ENOMEM;
1875 			goto fail;
1876 		}
1877 
1878 		/*
1879 		 * Create the DMA maps for RX buffers.
1880 		 */
1881 		if (!(fl->ifl_sds.ifsd_map =
1882 		      (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1883 			device_printf(dev,
1884 			    "Unable to allocate RX buffer DMA map memory\n");
1885 			err = ENOMEM;
1886 			goto fail;
1887 		}
1888 		for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++) {
1889 			err = bus_dmamap_create(fl->ifl_buf_tag, 0,
1890 			    &fl->ifl_sds.ifsd_map[i]);
1891 			if (err != 0) {
1892 				device_printf(dev, "Unable to create RX buffer DMA map\n");
1893 				goto fail;
1894 			}
1895 		}
1896 	}
1897 	return (0);
1898 
1899 fail:
1900 	iflib_rx_structures_free(ctx);
1901 	return (err);
1902 }
1903 
1904 
1905 /*
1906  * Internal service routines
1907  */
1908 
1909 struct rxq_refill_cb_arg {
1910 	int               error;
1911 	bus_dma_segment_t seg;
1912 	int               nseg;
1913 };
1914 
1915 static void
1916 _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
1917 {
1918 	struct rxq_refill_cb_arg *cb_arg = arg;
1919 
1920 	cb_arg->error = error;
1921 	cb_arg->seg = segs[0];
1922 	cb_arg->nseg = nseg;
1923 }
1924 
1925 /**
1926  * _iflib_fl_refill - refill an rxq free-buffer list
1927  * @ctx: the iflib context
1928  * @fl: the free list to refill
1929  * @count: the number of new buffers to allocate
1930  *
1931  * (Re)populate an rxq free-buffer list with up to @count new packet buffers.
1932  * The caller must assure that @count does not exceed the queue's capacity.
1933  */
1934 static void
1935 _iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
1936 {
1937 	struct if_rxd_update iru;
1938 	struct rxq_refill_cb_arg cb_arg;
1939 	struct mbuf *m;
1940 	caddr_t cl, *sd_cl;
1941 	struct mbuf **sd_m;
1942 	bus_dmamap_t *sd_map;
1943 	bus_addr_t bus_addr, *sd_ba;
1944 	int err, frag_idx, i, idx, n, pidx;
1945 	qidx_t credits;
1946 
1947 	sd_m = fl->ifl_sds.ifsd_m;
1948 	sd_map = fl->ifl_sds.ifsd_map;
1949 	sd_cl = fl->ifl_sds.ifsd_cl;
1950 	sd_ba = fl->ifl_sds.ifsd_ba;
1951 	pidx = fl->ifl_pidx;
1952 	idx = pidx;
1953 	frag_idx = fl->ifl_fragidx;
1954 	credits = fl->ifl_credits;
1955 
1956 	i = 0;
1957 	n = count;
1958 	MPASS(n > 0);
1959 	MPASS(credits + n <= fl->ifl_size);
1960 
1961 	if (pidx < fl->ifl_cidx)
1962 		MPASS(pidx + n <= fl->ifl_cidx);
1963 	if (pidx == fl->ifl_cidx && (credits < fl->ifl_size))
1964 		MPASS(fl->ifl_gen == 0);
1965 	if (pidx > fl->ifl_cidx)
1966 		MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx);
1967 
1968 	DBG_COUNTER_INC(fl_refills);
1969 	if (n > 8)
1970 		DBG_COUNTER_INC(fl_refills_large);
1971 	iru_init(&iru, fl->ifl_rxq, fl->ifl_id);
1972 	while (n--) {
1973 		/*
1974 		 * We allocate an uninitialized mbuf + cluster, mbuf is
1975 		 * initialized after rx.
1976 		 *
1977 		 * If the cluster is still set then we know a minimum sized packet was received
1978 		 */
1979 		bit_ffc_at(fl->ifl_rx_bitmap, frag_idx, fl->ifl_size,
1980 		    &frag_idx);
1981 		if (frag_idx < 0)
1982 			bit_ffc(fl->ifl_rx_bitmap, fl->ifl_size, &frag_idx);
1983 		MPASS(frag_idx >= 0);
1984 		if ((cl = sd_cl[frag_idx]) == NULL) {
1985 			if ((cl = m_cljget(NULL, M_NOWAIT, fl->ifl_buf_size)) == NULL)
1986 				break;
1987 
1988 			cb_arg.error = 0;
1989 			MPASS(sd_map != NULL);
1990 			err = bus_dmamap_load(fl->ifl_buf_tag, sd_map[frag_idx],
1991 			    cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg,
1992 			    BUS_DMA_NOWAIT);
1993 			if (err != 0 || cb_arg.error) {
1994 				/*
1995 				 * !zone_pack ?
1996 				 */
1997 				if (fl->ifl_zone == zone_pack)
1998 					uma_zfree(fl->ifl_zone, cl);
1999 				break;
2000 			}
2001 
2002 			sd_ba[frag_idx] =  bus_addr = cb_arg.seg.ds_addr;
2003 			sd_cl[frag_idx] = cl;
2004 #if MEMORY_LOGGING
2005 			fl->ifl_cl_enqueued++;
2006 #endif
2007 		} else {
2008 			bus_addr = sd_ba[frag_idx];
2009 		}
2010 		bus_dmamap_sync(fl->ifl_buf_tag, sd_map[frag_idx],
2011 		    BUS_DMASYNC_PREREAD);
2012 
2013 		if (sd_m[frag_idx] == NULL) {
2014 			if ((m = m_gethdr(M_NOWAIT, MT_NOINIT)) == NULL) {
2015 				break;
2016 			}
2017 			sd_m[frag_idx] = m;
2018 		}
2019 		bit_set(fl->ifl_rx_bitmap, frag_idx);
2020 #if MEMORY_LOGGING
2021 		fl->ifl_m_enqueued++;
2022 #endif
2023 
2024 		DBG_COUNTER_INC(rx_allocs);
2025 		fl->ifl_rxd_idxs[i] = frag_idx;
2026 		fl->ifl_bus_addrs[i] = bus_addr;
2027 		fl->ifl_vm_addrs[i] = cl;
2028 		credits++;
2029 		i++;
2030 		MPASS(credits <= fl->ifl_size);
2031 		if (++idx == fl->ifl_size) {
2032 			fl->ifl_gen = 1;
2033 			idx = 0;
2034 		}
2035 		if (n == 0 || i == IFLIB_MAX_RX_REFRESH) {
2036 			iru.iru_pidx = pidx;
2037 			iru.iru_count = i;
2038 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2039 			i = 0;
2040 			pidx = idx;
2041 			fl->ifl_pidx = idx;
2042 			fl->ifl_credits = credits;
2043 		}
2044 	}
2045 
2046 	if (i) {
2047 		iru.iru_pidx = pidx;
2048 		iru.iru_count = i;
2049 		ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2050 		fl->ifl_pidx = idx;
2051 		fl->ifl_credits = credits;
2052 	}
2053 	DBG_COUNTER_INC(rxd_flush);
2054 	if (fl->ifl_pidx == 0)
2055 		pidx = fl->ifl_size - 1;
2056 	else
2057 		pidx = fl->ifl_pidx - 1;
2058 
2059 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2060 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2061 	ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id, fl->ifl_id, pidx);
2062 	fl->ifl_fragidx = frag_idx;
2063 }
2064 
2065 static __inline void
2066 __iflib_fl_refill_lt(if_ctx_t ctx, iflib_fl_t fl, int max)
2067 {
2068 	/* we avoid allowing pidx to catch up with cidx as it confuses ixl */
2069 	int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1;
2070 #ifdef INVARIANTS
2071 	int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1;
2072 #endif
2073 
2074 	MPASS(fl->ifl_credits <= fl->ifl_size);
2075 	MPASS(reclaimable == delta);
2076 
2077 	if (reclaimable > 0)
2078 		_iflib_fl_refill(ctx, fl, min(max, reclaimable));
2079 }
2080 
2081 uint8_t
2082 iflib_in_detach(if_ctx_t ctx)
2083 {
2084 	bool in_detach;
2085 
2086 	STATE_LOCK(ctx);
2087 	in_detach = !!(ctx->ifc_flags & IFC_IN_DETACH);
2088 	STATE_UNLOCK(ctx);
2089 	return (in_detach);
2090 }
2091 
2092 static void
2093 iflib_fl_bufs_free(iflib_fl_t fl)
2094 {
2095 	iflib_dma_info_t idi = fl->ifl_ifdi;
2096 	bus_dmamap_t sd_map;
2097 	uint32_t i;
2098 
2099 	for (i = 0; i < fl->ifl_size; i++) {
2100 		struct mbuf **sd_m = &fl->ifl_sds.ifsd_m[i];
2101 		caddr_t *sd_cl = &fl->ifl_sds.ifsd_cl[i];
2102 
2103 		if (*sd_cl != NULL) {
2104 			sd_map = fl->ifl_sds.ifsd_map[i];
2105 			bus_dmamap_sync(fl->ifl_buf_tag, sd_map,
2106 			    BUS_DMASYNC_POSTREAD);
2107 			bus_dmamap_unload(fl->ifl_buf_tag, sd_map);
2108 			if (*sd_cl != NULL)
2109 				uma_zfree(fl->ifl_zone, *sd_cl);
2110 			// XXX: Should this get moved out?
2111 			if (iflib_in_detach(fl->ifl_rxq->ifr_ctx))
2112 				bus_dmamap_destroy(fl->ifl_buf_tag, sd_map);
2113 			if (*sd_m != NULL) {
2114 				m_init(*sd_m, M_NOWAIT, MT_DATA, 0);
2115 				uma_zfree(zone_mbuf, *sd_m);
2116 			}
2117 		} else {
2118 			MPASS(*sd_cl == NULL);
2119 			MPASS(*sd_m == NULL);
2120 		}
2121 #if MEMORY_LOGGING
2122 		fl->ifl_m_dequeued++;
2123 		fl->ifl_cl_dequeued++;
2124 #endif
2125 		*sd_cl = NULL;
2126 		*sd_m = NULL;
2127 	}
2128 #ifdef INVARIANTS
2129 	for (i = 0; i < fl->ifl_size; i++) {
2130 		MPASS(fl->ifl_sds.ifsd_cl[i] == NULL);
2131 		MPASS(fl->ifl_sds.ifsd_m[i] == NULL);
2132 	}
2133 #endif
2134 	/*
2135 	 * Reset free list values
2136 	 */
2137 	fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = fl->ifl_fragidx = 0;
2138 	bzero(idi->idi_vaddr, idi->idi_size);
2139 }
2140 
2141 /*********************************************************************
2142  *
2143  *  Initialize a free list and its buffers.
2144  *
2145  **********************************************************************/
2146 static int
2147 iflib_fl_setup(iflib_fl_t fl)
2148 {
2149 	iflib_rxq_t rxq = fl->ifl_rxq;
2150 	if_ctx_t ctx = rxq->ifr_ctx;
2151 
2152 	bit_nclear(fl->ifl_rx_bitmap, 0, fl->ifl_size - 1);
2153 	/*
2154 	** Free current RX buffer structs and their mbufs
2155 	*/
2156 	iflib_fl_bufs_free(fl);
2157 	/* Now replenish the mbufs */
2158 	MPASS(fl->ifl_credits == 0);
2159 	fl->ifl_buf_size = ctx->ifc_rx_mbuf_sz;
2160 	if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size)
2161 		ctx->ifc_max_fl_buf_size = fl->ifl_buf_size;
2162 	fl->ifl_cltype = m_gettype(fl->ifl_buf_size);
2163 	fl->ifl_zone = m_getzone(fl->ifl_buf_size);
2164 
2165 
2166 	/* avoid pre-allocating zillions of clusters to an idle card
2167 	 * potentially speeding up attach
2168 	 */
2169 	_iflib_fl_refill(ctx, fl, min(128, fl->ifl_size));
2170 	MPASS(min(128, fl->ifl_size) == fl->ifl_credits);
2171 	if (min(128, fl->ifl_size) != fl->ifl_credits)
2172 		return (ENOBUFS);
2173 	/*
2174 	 * handle failure
2175 	 */
2176 	MPASS(rxq != NULL);
2177 	MPASS(fl->ifl_ifdi != NULL);
2178 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2179 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2180 	return (0);
2181 }
2182 
2183 /*********************************************************************
2184  *
2185  *  Free receive ring data structures
2186  *
2187  **********************************************************************/
2188 static void
2189 iflib_rx_sds_free(iflib_rxq_t rxq)
2190 {
2191 	iflib_fl_t fl;
2192 	int i, j;
2193 
2194 	if (rxq->ifr_fl != NULL) {
2195 		for (i = 0; i < rxq->ifr_nfl; i++) {
2196 			fl = &rxq->ifr_fl[i];
2197 			if (fl->ifl_buf_tag != NULL) {
2198 				if (fl->ifl_sds.ifsd_map != NULL) {
2199 					for (j = 0; j < fl->ifl_size; j++) {
2200 						if (fl->ifl_sds.ifsd_map[j] ==
2201 						    NULL)
2202 							continue;
2203 						bus_dmamap_sync(
2204 						    fl->ifl_buf_tag,
2205 						    fl->ifl_sds.ifsd_map[j],
2206 						    BUS_DMASYNC_POSTREAD);
2207 						bus_dmamap_unload(
2208 						    fl->ifl_buf_tag,
2209 						    fl->ifl_sds.ifsd_map[j]);
2210 					}
2211 				}
2212 				bus_dma_tag_destroy(fl->ifl_buf_tag);
2213 				fl->ifl_buf_tag = NULL;
2214 			}
2215 			free(fl->ifl_sds.ifsd_m, M_IFLIB);
2216 			free(fl->ifl_sds.ifsd_cl, M_IFLIB);
2217 			free(fl->ifl_sds.ifsd_ba, M_IFLIB);
2218 			free(fl->ifl_sds.ifsd_map, M_IFLIB);
2219 			fl->ifl_sds.ifsd_m = NULL;
2220 			fl->ifl_sds.ifsd_cl = NULL;
2221 			fl->ifl_sds.ifsd_ba = NULL;
2222 			fl->ifl_sds.ifsd_map = NULL;
2223 		}
2224 		free(rxq->ifr_fl, M_IFLIB);
2225 		rxq->ifr_fl = NULL;
2226 		rxq->ifr_cq_cidx = 0;
2227 	}
2228 }
2229 
2230 /*
2231  * Timer routine
2232  */
2233 static void
2234 iflib_timer(void *arg)
2235 {
2236 	iflib_txq_t txq = arg;
2237 	if_ctx_t ctx = txq->ift_ctx;
2238 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2239 	uint64_t this_tick = ticks;
2240 	uint32_t reset_on = hz / 2;
2241 
2242 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
2243 		return;
2244 
2245 	/*
2246 	** Check on the state of the TX queue(s), this
2247 	** can be done without the lock because its RO
2248 	** and the HUNG state will be static if set.
2249 	*/
2250 	if (this_tick - txq->ift_last_timer_tick >= hz / 2) {
2251 		txq->ift_last_timer_tick = this_tick;
2252 		IFDI_TIMER(ctx, txq->ift_id);
2253 		if ((txq->ift_qstatus == IFLIB_QUEUE_HUNG) &&
2254 		    ((txq->ift_cleaned_prev == txq->ift_cleaned) ||
2255 		     (sctx->isc_pause_frames == 0)))
2256 			goto hung;
2257 
2258 		if (ifmp_ring_is_stalled(txq->ift_br))
2259 			txq->ift_qstatus = IFLIB_QUEUE_HUNG;
2260 		txq->ift_cleaned_prev = txq->ift_cleaned;
2261 	}
2262 #ifdef DEV_NETMAP
2263 	if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP)
2264 		iflib_netmap_timer_adjust(ctx, txq, &reset_on);
2265 #endif
2266 	/* handle any laggards */
2267 	if (txq->ift_db_pending)
2268 		GROUPTASK_ENQUEUE(&txq->ift_task);
2269 
2270 	sctx->isc_pause_frames = 0;
2271 	if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)
2272 		callout_reset_on(&txq->ift_timer, reset_on, iflib_timer, txq, txq->ift_timer.c_cpu);
2273 	return;
2274 
2275  hung:
2276 	device_printf(ctx->ifc_dev,
2277 	    "Watchdog timeout (TX: %d desc avail: %d pidx: %d) -- resetting\n",
2278 	    txq->ift_id, TXQ_AVAIL(txq), txq->ift_pidx);
2279 	STATE_LOCK(ctx);
2280 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2281 	ctx->ifc_flags |= (IFC_DO_WATCHDOG|IFC_DO_RESET);
2282 	iflib_admin_intr_deferred(ctx);
2283 	STATE_UNLOCK(ctx);
2284 }
2285 
2286 static void
2287 iflib_calc_rx_mbuf_sz(if_ctx_t ctx)
2288 {
2289 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2290 
2291 	/*
2292 	 * XXX don't set the max_frame_size to larger
2293 	 * than the hardware can handle
2294 	 */
2295 	if (sctx->isc_max_frame_size <= MCLBYTES)
2296 		ctx->ifc_rx_mbuf_sz = MCLBYTES;
2297 	else
2298 		ctx->ifc_rx_mbuf_sz = MJUMPAGESIZE;
2299 }
2300 
2301 uint32_t
2302 iflib_get_rx_mbuf_sz(if_ctx_t ctx)
2303 {
2304 
2305 	return (ctx->ifc_rx_mbuf_sz);
2306 }
2307 
2308 static void
2309 iflib_init_locked(if_ctx_t ctx)
2310 {
2311 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2312 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2313 	if_t ifp = ctx->ifc_ifp;
2314 	iflib_fl_t fl;
2315 	iflib_txq_t txq;
2316 	iflib_rxq_t rxq;
2317 	int i, j, tx_ip_csum_flags, tx_ip6_csum_flags;
2318 
2319 	if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2320 	IFDI_INTR_DISABLE(ctx);
2321 
2322 	tx_ip_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP);
2323 	tx_ip6_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | CSUM_IP6_SCTP);
2324 	/* Set hardware offload abilities */
2325 	if_clearhwassist(ifp);
2326 	if (if_getcapenable(ifp) & IFCAP_TXCSUM)
2327 		if_sethwassistbits(ifp, tx_ip_csum_flags, 0);
2328 	if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6)
2329 		if_sethwassistbits(ifp,  tx_ip6_csum_flags, 0);
2330 	if (if_getcapenable(ifp) & IFCAP_TSO4)
2331 		if_sethwassistbits(ifp, CSUM_IP_TSO, 0);
2332 	if (if_getcapenable(ifp) & IFCAP_TSO6)
2333 		if_sethwassistbits(ifp, CSUM_IP6_TSO, 0);
2334 
2335 	for (i = 0, txq = ctx->ifc_txqs; i < sctx->isc_ntxqsets; i++, txq++) {
2336 		CALLOUT_LOCK(txq);
2337 		callout_stop(&txq->ift_timer);
2338 		CALLOUT_UNLOCK(txq);
2339 		iflib_netmap_txq_init(ctx, txq);
2340 	}
2341 
2342 	/*
2343 	 * Calculate a suitable Rx mbuf size prior to calling IFDI_INIT, so
2344 	 * that drivers can use the value when setting up the hardware receive
2345 	 * buffers.
2346 	 */
2347 	iflib_calc_rx_mbuf_sz(ctx);
2348 
2349 #ifdef INVARIANTS
2350 	i = if_getdrvflags(ifp);
2351 #endif
2352 	IFDI_INIT(ctx);
2353 	MPASS(if_getdrvflags(ifp) == i);
2354 	for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
2355 		/* XXX this should really be done on a per-queue basis */
2356 		if (if_getcapenable(ifp) & IFCAP_NETMAP) {
2357 			MPASS(rxq->ifr_id == i);
2358 			iflib_netmap_rxq_init(ctx, rxq);
2359 			continue;
2360 		}
2361 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
2362 			if (iflib_fl_setup(fl)) {
2363 				device_printf(ctx->ifc_dev,
2364 				    "setting up free list %d failed - "
2365 				    "check cluster settings\n", j);
2366 				goto done;
2367 			}
2368 		}
2369 	}
2370 done:
2371 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
2372 	IFDI_INTR_ENABLE(ctx);
2373 	txq = ctx->ifc_txqs;
2374 	for (i = 0; i < sctx->isc_ntxqsets; i++, txq++)
2375 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq,
2376 			txq->ift_timer.c_cpu);
2377 }
2378 
2379 static int
2380 iflib_media_change(if_t ifp)
2381 {
2382 	if_ctx_t ctx = if_getsoftc(ifp);
2383 	int err;
2384 
2385 	CTX_LOCK(ctx);
2386 	if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0)
2387 		iflib_init_locked(ctx);
2388 	CTX_UNLOCK(ctx);
2389 	return (err);
2390 }
2391 
2392 static void
2393 iflib_media_status(if_t ifp, struct ifmediareq *ifmr)
2394 {
2395 	if_ctx_t ctx = if_getsoftc(ifp);
2396 
2397 	CTX_LOCK(ctx);
2398 	IFDI_UPDATE_ADMIN_STATUS(ctx);
2399 	IFDI_MEDIA_STATUS(ctx, ifmr);
2400 	CTX_UNLOCK(ctx);
2401 }
2402 
2403 void
2404 iflib_stop(if_ctx_t ctx)
2405 {
2406 	iflib_txq_t txq = ctx->ifc_txqs;
2407 	iflib_rxq_t rxq = ctx->ifc_rxqs;
2408 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2409 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2410 	iflib_dma_info_t di;
2411 	iflib_fl_t fl;
2412 	int i, j;
2413 
2414 	/* Tell the stack that the interface is no longer active */
2415 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2416 
2417 	IFDI_INTR_DISABLE(ctx);
2418 	DELAY(1000);
2419 	IFDI_STOP(ctx);
2420 	DELAY(1000);
2421 
2422 	iflib_debug_reset();
2423 	/* Wait for current tx queue users to exit to disarm watchdog timer. */
2424 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) {
2425 		/* make sure all transmitters have completed before proceeding XXX */
2426 
2427 		CALLOUT_LOCK(txq);
2428 		callout_stop(&txq->ift_timer);
2429 		CALLOUT_UNLOCK(txq);
2430 
2431 		/* clean any enqueued buffers */
2432 		iflib_ifmp_purge(txq);
2433 		/* Free any existing tx buffers. */
2434 		for (j = 0; j < txq->ift_size; j++) {
2435 			iflib_txsd_free(ctx, txq, j);
2436 		}
2437 		txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0;
2438 		txq->ift_in_use = txq->ift_gen = txq->ift_cidx = txq->ift_pidx = txq->ift_no_desc_avail = 0;
2439 		txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0;
2440 		txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0;
2441 		txq->ift_pullups = 0;
2442 		ifmp_ring_reset_stats(txq->ift_br);
2443 		for (j = 0, di = txq->ift_ifdi; j < sctx->isc_ntxqs; j++, di++)
2444 			bzero((void *)di->idi_vaddr, di->idi_size);
2445 	}
2446 	for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) {
2447 		/* make sure all transmitters have completed before proceeding XXX */
2448 
2449 		rxq->ifr_cq_cidx = 0;
2450 		for (j = 0, di = rxq->ifr_ifdi; j < sctx->isc_nrxqs; j++, di++)
2451 			bzero((void *)di->idi_vaddr, di->idi_size);
2452 		/* also resets the free lists pidx/cidx */
2453 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
2454 			iflib_fl_bufs_free(fl);
2455 	}
2456 }
2457 
2458 static inline caddr_t
2459 calc_next_rxd(iflib_fl_t fl, int cidx)
2460 {
2461 	qidx_t size;
2462 	int nrxd;
2463 	caddr_t start, end, cur, next;
2464 
2465 	nrxd = fl->ifl_size;
2466 	size = fl->ifl_rxd_size;
2467 	start = fl->ifl_ifdi->idi_vaddr;
2468 
2469 	if (__predict_false(size == 0))
2470 		return (start);
2471 	cur = start + size*cidx;
2472 	end = start + size*nrxd;
2473 	next = CACHE_PTR_NEXT(cur);
2474 	return (next < end ? next : start);
2475 }
2476 
2477 static inline void
2478 prefetch_pkts(iflib_fl_t fl, int cidx)
2479 {
2480 	int nextptr;
2481 	int nrxd = fl->ifl_size;
2482 	caddr_t next_rxd;
2483 
2484 
2485 	nextptr = (cidx + CACHE_PTR_INCREMENT) & (nrxd-1);
2486 	prefetch(&fl->ifl_sds.ifsd_m[nextptr]);
2487 	prefetch(&fl->ifl_sds.ifsd_cl[nextptr]);
2488 	next_rxd = calc_next_rxd(fl, cidx);
2489 	prefetch(next_rxd);
2490 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 1) & (nrxd-1)]);
2491 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 2) & (nrxd-1)]);
2492 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 3) & (nrxd-1)]);
2493 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 4) & (nrxd-1)]);
2494 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 1) & (nrxd-1)]);
2495 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 2) & (nrxd-1)]);
2496 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 3) & (nrxd-1)]);
2497 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 4) & (nrxd-1)]);
2498 }
2499 
2500 static struct mbuf *
2501 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, bool unload, if_rxsd_t sd,
2502     int *pf_rv, if_rxd_info_t ri)
2503 {
2504 	bus_dmamap_t map;
2505 	iflib_fl_t fl;
2506 	caddr_t payload;
2507 	struct mbuf *m;
2508 	int flid, cidx, len, next;
2509 
2510 	map = NULL;
2511 	flid = irf->irf_flid;
2512 	cidx = irf->irf_idx;
2513 	fl = &rxq->ifr_fl[flid];
2514 	sd->ifsd_fl = fl;
2515 	sd->ifsd_cidx = cidx;
2516 	m = fl->ifl_sds.ifsd_m[cidx];
2517 	sd->ifsd_cl = &fl->ifl_sds.ifsd_cl[cidx];
2518 	fl->ifl_credits--;
2519 #if MEMORY_LOGGING
2520 	fl->ifl_m_dequeued++;
2521 #endif
2522 	if (rxq->ifr_ctx->ifc_flags & IFC_PREFETCH)
2523 		prefetch_pkts(fl, cidx);
2524 	next = (cidx + CACHE_PTR_INCREMENT) & (fl->ifl_size-1);
2525 	prefetch(&fl->ifl_sds.ifsd_map[next]);
2526 	map = fl->ifl_sds.ifsd_map[cidx];
2527 	next = (cidx + CACHE_LINE_SIZE) & (fl->ifl_size-1);
2528 
2529 	/* not valid assert if bxe really does SGE from non-contiguous elements */
2530 	MPASS(fl->ifl_cidx == cidx);
2531 	bus_dmamap_sync(fl->ifl_buf_tag, map, BUS_DMASYNC_POSTREAD);
2532 
2533 	if (rxq->pfil != NULL && PFIL_HOOKED_IN(rxq->pfil) && pf_rv != NULL) {
2534 		payload  = *sd->ifsd_cl;
2535 		payload +=  ri->iri_pad;
2536 		len = ri->iri_len - ri->iri_pad;
2537 		*pf_rv = pfil_run_hooks(rxq->pfil, payload, ri->iri_ifp,
2538 		    len | PFIL_MEMPTR | PFIL_IN, NULL);
2539 		switch (*pf_rv) {
2540 		case PFIL_DROPPED:
2541 		case PFIL_CONSUMED:
2542 			/*
2543 			 * The filter ate it.  Everything is recycled.
2544 			 */
2545 			m = NULL;
2546 			unload = 0;
2547 			break;
2548 		case PFIL_REALLOCED:
2549 			/*
2550 			 * The filter copied it.  Everything is recycled.
2551 			 */
2552 			m = pfil_mem2mbuf(payload);
2553 			unload = 0;
2554 			break;
2555 		case PFIL_PASS:
2556 			/*
2557 			 * Filter said it was OK, so receive like
2558 			 * normal
2559 			 */
2560 			fl->ifl_sds.ifsd_m[cidx] = NULL;
2561 			break;
2562 		default:
2563 			MPASS(0);
2564 		}
2565 	} else {
2566 		fl->ifl_sds.ifsd_m[cidx] = NULL;
2567 		*pf_rv = PFIL_PASS;
2568 	}
2569 
2570 	if (unload)
2571 		bus_dmamap_unload(fl->ifl_buf_tag, map);
2572 	fl->ifl_cidx = (fl->ifl_cidx + 1) & (fl->ifl_size-1);
2573 	if (__predict_false(fl->ifl_cidx == 0))
2574 		fl->ifl_gen = 0;
2575 	bit_clear(fl->ifl_rx_bitmap, cidx);
2576 	return (m);
2577 }
2578 
2579 static struct mbuf *
2580 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri, if_rxsd_t sd, int *pf_rv)
2581 {
2582 	struct mbuf *m, *mh, *mt;
2583 	caddr_t cl;
2584 	int  *pf_rv_ptr, flags, i, padlen;
2585 	bool consumed;
2586 
2587 	i = 0;
2588 	mh = NULL;
2589 	consumed = false;
2590 	*pf_rv = PFIL_PASS;
2591 	pf_rv_ptr = pf_rv;
2592 	do {
2593 		m = rxd_frag_to_sd(rxq, &ri->iri_frags[i], !consumed, sd,
2594 		    pf_rv_ptr, ri);
2595 
2596 		MPASS(*sd->ifsd_cl != NULL);
2597 
2598 		/*
2599 		 * Exclude zero-length frags & frags from
2600 		 * packets the filter has consumed or dropped
2601 		 */
2602 		if (ri->iri_frags[i].irf_len == 0 || consumed ||
2603 		    *pf_rv == PFIL_CONSUMED || *pf_rv == PFIL_DROPPED) {
2604 			if (mh == NULL) {
2605 				/* everything saved here */
2606 				consumed = true;
2607 				pf_rv_ptr = NULL;
2608 				continue;
2609 			}
2610 			/* XXX we can save the cluster here, but not the mbuf */
2611 			m_init(m, M_NOWAIT, MT_DATA, 0);
2612 			m_free(m);
2613 			continue;
2614 		}
2615 		if (mh == NULL) {
2616 			flags = M_PKTHDR|M_EXT;
2617 			mh = mt = m;
2618 			padlen = ri->iri_pad;
2619 		} else {
2620 			flags = M_EXT;
2621 			mt->m_next = m;
2622 			mt = m;
2623 			/* assuming padding is only on the first fragment */
2624 			padlen = 0;
2625 		}
2626 		cl = *sd->ifsd_cl;
2627 		*sd->ifsd_cl = NULL;
2628 
2629 		/* Can these two be made one ? */
2630 		m_init(m, M_NOWAIT, MT_DATA, flags);
2631 		m_cljset(m, cl, sd->ifsd_fl->ifl_cltype);
2632 		/*
2633 		 * These must follow m_init and m_cljset
2634 		 */
2635 		m->m_data += padlen;
2636 		ri->iri_len -= padlen;
2637 		m->m_len = ri->iri_frags[i].irf_len;
2638 	} while (++i < ri->iri_nfrags);
2639 
2640 	return (mh);
2641 }
2642 
2643 /*
2644  * Process one software descriptor
2645  */
2646 static struct mbuf *
2647 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri)
2648 {
2649 	struct if_rxsd sd;
2650 	struct mbuf *m;
2651 	int pf_rv;
2652 
2653 	/* should I merge this back in now that the two paths are basically duplicated? */
2654 	if (ri->iri_nfrags == 1 &&
2655 	    ri->iri_frags[0].irf_len <= MIN(IFLIB_RX_COPY_THRESH, MHLEN)) {
2656 		m = rxd_frag_to_sd(rxq, &ri->iri_frags[0], false, &sd,
2657 		    &pf_rv, ri);
2658 		if (pf_rv != PFIL_PASS && pf_rv != PFIL_REALLOCED)
2659 			return (m);
2660 		if (pf_rv == PFIL_PASS) {
2661 			m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR);
2662 #ifndef __NO_STRICT_ALIGNMENT
2663 			if (!IP_ALIGNED(m))
2664 				m->m_data += 2;
2665 #endif
2666 			memcpy(m->m_data, *sd.ifsd_cl, ri->iri_len);
2667 			m->m_len = ri->iri_frags[0].irf_len;
2668 		}
2669 	} else {
2670 		m = assemble_segments(rxq, ri, &sd, &pf_rv);
2671 		if (pf_rv != PFIL_PASS && pf_rv != PFIL_REALLOCED)
2672 			return (m);
2673 	}
2674 	m->m_pkthdr.len = ri->iri_len;
2675 	m->m_pkthdr.rcvif = ri->iri_ifp;
2676 	m->m_flags |= ri->iri_flags;
2677 	m->m_pkthdr.ether_vtag = ri->iri_vtag;
2678 	m->m_pkthdr.flowid = ri->iri_flowid;
2679 	M_HASHTYPE_SET(m, ri->iri_rsstype);
2680 	m->m_pkthdr.csum_flags = ri->iri_csum_flags;
2681 	m->m_pkthdr.csum_data = ri->iri_csum_data;
2682 	return (m);
2683 }
2684 
2685 #if defined(INET6) || defined(INET)
2686 static void
2687 iflib_get_ip_forwarding(struct lro_ctrl *lc, bool *v4, bool *v6)
2688 {
2689 	CURVNET_SET(lc->ifp->if_vnet);
2690 #if defined(INET6)
2691 	*v6 = VNET(ip6_forwarding);
2692 #endif
2693 #if defined(INET)
2694 	*v4 = VNET(ipforwarding);
2695 #endif
2696 	CURVNET_RESTORE();
2697 }
2698 
2699 /*
2700  * Returns true if it's possible this packet could be LROed.
2701  * if it returns false, it is guaranteed that tcp_lro_rx()
2702  * would not return zero.
2703  */
2704 static bool
2705 iflib_check_lro_possible(struct mbuf *m, bool v4_forwarding, bool v6_forwarding)
2706 {
2707 	struct ether_header *eh;
2708 	uint16_t eh_type;
2709 
2710 	eh = mtod(m, struct ether_header *);
2711 	eh_type = ntohs(eh->ether_type);
2712 	switch (eh_type) {
2713 #if defined(INET6)
2714 		case ETHERTYPE_IPV6:
2715 			return !v6_forwarding;
2716 #endif
2717 #if defined (INET)
2718 		case ETHERTYPE_IP:
2719 			return !v4_forwarding;
2720 #endif
2721 	}
2722 
2723 	return false;
2724 }
2725 #else
2726 static void
2727 iflib_get_ip_forwarding(struct lro_ctrl *lc __unused, bool *v4 __unused, bool *v6 __unused)
2728 {
2729 }
2730 #endif
2731 
2732 static bool
2733 iflib_rxeof(iflib_rxq_t rxq, qidx_t budget)
2734 {
2735 	if_t ifp;
2736 	if_ctx_t ctx = rxq->ifr_ctx;
2737 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2738 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2739 	int avail, i;
2740 	qidx_t *cidxp;
2741 	struct if_rxd_info ri;
2742 	int err, budget_left, rx_bytes, rx_pkts;
2743 	iflib_fl_t fl;
2744 	int lro_enabled;
2745 	bool v4_forwarding, v6_forwarding, lro_possible;
2746 
2747 	/*
2748 	 * XXX early demux data packets so that if_input processing only handles
2749 	 * acks in interrupt context
2750 	 */
2751 	struct mbuf *m, *mh, *mt, *mf;
2752 
2753 	lro_possible = v4_forwarding = v6_forwarding = false;
2754 	ifp = ctx->ifc_ifp;
2755 	mh = mt = NULL;
2756 	MPASS(budget > 0);
2757 	rx_pkts	= rx_bytes = 0;
2758 	if (sctx->isc_flags & IFLIB_HAS_RXCQ)
2759 		cidxp = &rxq->ifr_cq_cidx;
2760 	else
2761 		cidxp = &rxq->ifr_fl[0].ifl_cidx;
2762 	if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) {
2763 		for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2764 			__iflib_fl_refill_lt(ctx, fl, budget + 8);
2765 		DBG_COUNTER_INC(rx_unavail);
2766 		return (false);
2767 	}
2768 
2769 	/* pfil needs the vnet to be set */
2770 	CURVNET_SET_QUIET(ifp->if_vnet);
2771 	for (budget_left = budget; budget_left > 0 && avail > 0;) {
2772 		if (__predict_false(!CTX_ACTIVE(ctx))) {
2773 			DBG_COUNTER_INC(rx_ctx_inactive);
2774 			break;
2775 		}
2776 		/*
2777 		 * Reset client set fields to their default values
2778 		 */
2779 		rxd_info_zero(&ri);
2780 		ri.iri_qsidx = rxq->ifr_id;
2781 		ri.iri_cidx = *cidxp;
2782 		ri.iri_ifp = ifp;
2783 		ri.iri_frags = rxq->ifr_frags;
2784 		err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
2785 
2786 		if (err)
2787 			goto err;
2788 		rx_pkts += 1;
2789 		rx_bytes += ri.iri_len;
2790 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
2791 			*cidxp = ri.iri_cidx;
2792 			/* Update our consumer index */
2793 			/* XXX NB: shurd - check if this is still safe */
2794 			while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0])
2795 				rxq->ifr_cq_cidx -= scctx->isc_nrxd[0];
2796 			/* was this only a completion queue message? */
2797 			if (__predict_false(ri.iri_nfrags == 0))
2798 				continue;
2799 		}
2800 		MPASS(ri.iri_nfrags != 0);
2801 		MPASS(ri.iri_len != 0);
2802 
2803 		/* will advance the cidx on the corresponding free lists */
2804 		m = iflib_rxd_pkt_get(rxq, &ri);
2805 		avail--;
2806 		budget_left--;
2807 		if (avail == 0 && budget_left)
2808 			avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left);
2809 
2810 		if (__predict_false(m == NULL))
2811 			continue;
2812 
2813 		/* imm_pkt: -- cxgb */
2814 		if (mh == NULL)
2815 			mh = mt = m;
2816 		else {
2817 			mt->m_nextpkt = m;
2818 			mt = m;
2819 		}
2820 	}
2821 	CURVNET_RESTORE();
2822 	/* make sure that we can refill faster than drain */
2823 	for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2824 		__iflib_fl_refill_lt(ctx, fl, budget + 8);
2825 
2826 	lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO);
2827 	if (lro_enabled)
2828 		iflib_get_ip_forwarding(&rxq->ifr_lc, &v4_forwarding, &v6_forwarding);
2829 	mt = mf = NULL;
2830 	while (mh != NULL) {
2831 		m = mh;
2832 		mh = mh->m_nextpkt;
2833 		m->m_nextpkt = NULL;
2834 #ifndef __NO_STRICT_ALIGNMENT
2835 		if (!IP_ALIGNED(m) && (m = iflib_fixup_rx(m)) == NULL)
2836 			continue;
2837 #endif
2838 		rx_bytes += m->m_pkthdr.len;
2839 		rx_pkts++;
2840 #if defined(INET6) || defined(INET)
2841 		if (lro_enabled) {
2842 			if (!lro_possible) {
2843 				lro_possible = iflib_check_lro_possible(m, v4_forwarding, v6_forwarding);
2844 				if (lro_possible && mf != NULL) {
2845 					ifp->if_input(ifp, mf);
2846 					DBG_COUNTER_INC(rx_if_input);
2847 					mt = mf = NULL;
2848 				}
2849 			}
2850 			if ((m->m_pkthdr.csum_flags & (CSUM_L4_CALC|CSUM_L4_VALID)) ==
2851 			    (CSUM_L4_CALC|CSUM_L4_VALID)) {
2852 				if (lro_possible && tcp_lro_rx(&rxq->ifr_lc, m, 0) == 0)
2853 					continue;
2854 			}
2855 		}
2856 #endif
2857 		if (lro_possible) {
2858 			ifp->if_input(ifp, m);
2859 			DBG_COUNTER_INC(rx_if_input);
2860 			continue;
2861 		}
2862 
2863 		if (mf == NULL)
2864 			mf = m;
2865 		if (mt != NULL)
2866 			mt->m_nextpkt = m;
2867 		mt = m;
2868 	}
2869 	if (mf != NULL) {
2870 		ifp->if_input(ifp, mf);
2871 		DBG_COUNTER_INC(rx_if_input);
2872 	}
2873 
2874 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
2875 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
2876 
2877 	/*
2878 	 * Flush any outstanding LRO work
2879 	 */
2880 #if defined(INET6) || defined(INET)
2881 	tcp_lro_flush_all(&rxq->ifr_lc);
2882 #endif
2883 	if (avail)
2884 		return true;
2885 	return (iflib_rxd_avail(ctx, rxq, *cidxp, 1));
2886 err:
2887 	STATE_LOCK(ctx);
2888 	ctx->ifc_flags |= IFC_DO_RESET;
2889 	iflib_admin_intr_deferred(ctx);
2890 	STATE_UNLOCK(ctx);
2891 	return (false);
2892 }
2893 
2894 #define TXD_NOTIFY_COUNT(txq) (((txq)->ift_size / (txq)->ift_update_freq)-1)
2895 static inline qidx_t
2896 txq_max_db_deferred(iflib_txq_t txq, qidx_t in_use)
2897 {
2898 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2899 	qidx_t minthresh = txq->ift_size / 8;
2900 	if (in_use > 4*minthresh)
2901 		return (notify_count);
2902 	if (in_use > 2*minthresh)
2903 		return (notify_count >> 1);
2904 	if (in_use > minthresh)
2905 		return (notify_count >> 3);
2906 	return (0);
2907 }
2908 
2909 static inline qidx_t
2910 txq_max_rs_deferred(iflib_txq_t txq)
2911 {
2912 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2913 	qidx_t minthresh = txq->ift_size / 8;
2914 	if (txq->ift_in_use > 4*minthresh)
2915 		return (notify_count);
2916 	if (txq->ift_in_use > 2*minthresh)
2917 		return (notify_count >> 1);
2918 	if (txq->ift_in_use > minthresh)
2919 		return (notify_count >> 2);
2920 	return (2);
2921 }
2922 
2923 #define M_CSUM_FLAGS(m) ((m)->m_pkthdr.csum_flags)
2924 #define M_HAS_VLANTAG(m) (m->m_flags & M_VLANTAG)
2925 
2926 #define TXQ_MAX_DB_DEFERRED(txq, in_use) txq_max_db_deferred((txq), (in_use))
2927 #define TXQ_MAX_RS_DEFERRED(txq) txq_max_rs_deferred(txq)
2928 #define TXQ_MAX_DB_CONSUMED(size) (size >> 4)
2929 
2930 /* forward compatibility for cxgb */
2931 #define FIRST_QSET(ctx) 0
2932 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets)
2933 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets)
2934 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx))
2935 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments))
2936 
2937 /* XXX we should be setting this to something other than zero */
2938 #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh)
2939 #define	MAX_TX_DESC(ctx) max((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max, \
2940     (ctx)->ifc_softc_ctx.isc_tx_nsegments)
2941 
2942 static inline bool
2943 iflib_txd_db_check(if_ctx_t ctx, iflib_txq_t txq, int ring, qidx_t in_use)
2944 {
2945 	qidx_t dbval, max;
2946 	bool rang;
2947 
2948 	rang = false;
2949 	max = TXQ_MAX_DB_DEFERRED(txq, in_use);
2950 	if (ring || txq->ift_db_pending >= max) {
2951 		dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx;
2952 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
2953 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2954 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval);
2955 		txq->ift_db_pending = txq->ift_npending = 0;
2956 		rang = true;
2957 	}
2958 	return (rang);
2959 }
2960 
2961 #ifdef PKT_DEBUG
2962 static void
2963 print_pkt(if_pkt_info_t pi)
2964 {
2965 	printf("pi len:  %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n",
2966 	       pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx);
2967 	printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n",
2968 	       pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag);
2969 	printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n",
2970 	       pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto);
2971 }
2972 #endif
2973 
2974 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO)
2975 #define IS_TX_OFFLOAD4(pi) ((pi)->ipi_csum_flags & (CSUM_IP_TCP | CSUM_IP_TSO))
2976 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO)
2977 #define IS_TX_OFFLOAD6(pi) ((pi)->ipi_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_TSO))
2978 
2979 static int
2980 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp)
2981 {
2982 	if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx;
2983 	struct ether_vlan_header *eh;
2984 	struct mbuf *m;
2985 
2986 	m = *mp;
2987 	if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) &&
2988 	    M_WRITABLE(m) == 0) {
2989 		if ((m = m_dup(m, M_NOWAIT)) == NULL) {
2990 			return (ENOMEM);
2991 		} else {
2992 			m_freem(*mp);
2993 			DBG_COUNTER_INC(tx_frees);
2994 			*mp = m;
2995 		}
2996 	}
2997 
2998 	/*
2999 	 * Determine where frame payload starts.
3000 	 * Jump over vlan headers if already present,
3001 	 * helpful for QinQ too.
3002 	 */
3003 	if (__predict_false(m->m_len < sizeof(*eh))) {
3004 		txq->ift_pullups++;
3005 		if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL))
3006 			return (ENOMEM);
3007 	}
3008 	eh = mtod(m, struct ether_vlan_header *);
3009 	if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
3010 		pi->ipi_etype = ntohs(eh->evl_proto);
3011 		pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
3012 	} else {
3013 		pi->ipi_etype = ntohs(eh->evl_encap_proto);
3014 		pi->ipi_ehdrlen = ETHER_HDR_LEN;
3015 	}
3016 
3017 	switch (pi->ipi_etype) {
3018 #ifdef INET
3019 	case ETHERTYPE_IP:
3020 	{
3021 		struct mbuf *n;
3022 		struct ip *ip = NULL;
3023 		struct tcphdr *th = NULL;
3024 		int minthlen;
3025 
3026 		minthlen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip) + sizeof(*th));
3027 		if (__predict_false(m->m_len < minthlen)) {
3028 			/*
3029 			 * if this code bloat is causing too much of a hit
3030 			 * move it to a separate function and mark it noinline
3031 			 */
3032 			if (m->m_len == pi->ipi_ehdrlen) {
3033 				n = m->m_next;
3034 				MPASS(n);
3035 				if (n->m_len >= sizeof(*ip))  {
3036 					ip = (struct ip *)n->m_data;
3037 					if (n->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3038 						th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3039 				} else {
3040 					txq->ift_pullups++;
3041 					if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
3042 						return (ENOMEM);
3043 					ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3044 				}
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 				if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3051 					th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3052 			}
3053 		} else {
3054 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3055 			if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3056 				th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3057 		}
3058 		pi->ipi_ip_hlen = ip->ip_hl << 2;
3059 		pi->ipi_ipproto = ip->ip_p;
3060 		pi->ipi_flags |= IPI_TX_IPV4;
3061 
3062 		/* TCP checksum offload may require TCP header length */
3063 		if (IS_TX_OFFLOAD4(pi)) {
3064 			if (__predict_true(pi->ipi_ipproto == IPPROTO_TCP)) {
3065 				if (__predict_false(th == NULL)) {
3066 					txq->ift_pullups++;
3067 					if (__predict_false((m = m_pullup(m, (ip->ip_hl << 2) + sizeof(*th))) == NULL))
3068 						return (ENOMEM);
3069 					th = (struct tcphdr *)((caddr_t)ip + pi->ipi_ip_hlen);
3070 				}
3071 				pi->ipi_tcp_hflags = th->th_flags;
3072 				pi->ipi_tcp_hlen = th->th_off << 2;
3073 				pi->ipi_tcp_seq = th->th_seq;
3074 			}
3075 			if (IS_TSO4(pi)) {
3076 				if (__predict_false(ip->ip_p != IPPROTO_TCP))
3077 					return (ENXIO);
3078 				/*
3079 				 * TSO always requires hardware checksum offload.
3080 				 */
3081 				pi->ipi_csum_flags |= (CSUM_IP_TCP | CSUM_IP);
3082 				th->th_sum = in_pseudo(ip->ip_src.s_addr,
3083 						       ip->ip_dst.s_addr, htons(IPPROTO_TCP));
3084 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3085 				if (sctx->isc_flags & IFLIB_TSO_INIT_IP) {
3086 					ip->ip_sum = 0;
3087 					ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz);
3088 				}
3089 			}
3090 		}
3091 		if ((sctx->isc_flags & IFLIB_NEED_ZERO_CSUM) && (pi->ipi_csum_flags & CSUM_IP))
3092                        ip->ip_sum = 0;
3093 
3094 		break;
3095 	}
3096 #endif
3097 #ifdef INET6
3098 	case ETHERTYPE_IPV6:
3099 	{
3100 		struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
3101 		struct tcphdr *th;
3102 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
3103 
3104 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
3105 			txq->ift_pullups++;
3106 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
3107 				return (ENOMEM);
3108 		}
3109 		th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
3110 
3111 		/* XXX-BZ this will go badly in case of ext hdrs. */
3112 		pi->ipi_ipproto = ip6->ip6_nxt;
3113 		pi->ipi_flags |= IPI_TX_IPV6;
3114 
3115 		/* TCP checksum offload may require TCP header length */
3116 		if (IS_TX_OFFLOAD6(pi)) {
3117 			if (pi->ipi_ipproto == IPPROTO_TCP) {
3118 				if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) {
3119 					txq->ift_pullups++;
3120 					if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL))
3121 						return (ENOMEM);
3122 				}
3123 				pi->ipi_tcp_hflags = th->th_flags;
3124 				pi->ipi_tcp_hlen = th->th_off << 2;
3125 				pi->ipi_tcp_seq = th->th_seq;
3126 			}
3127 			if (IS_TSO6(pi)) {
3128 				if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP))
3129 					return (ENXIO);
3130 				/*
3131 				 * TSO always requires hardware checksum offload.
3132 				 */
3133 				pi->ipi_csum_flags |= CSUM_IP6_TCP;
3134 				th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0);
3135 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3136 			}
3137 		}
3138 		break;
3139 	}
3140 #endif
3141 	default:
3142 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
3143 		pi->ipi_ip_hlen = 0;
3144 		break;
3145 	}
3146 	*mp = m;
3147 
3148 	return (0);
3149 }
3150 
3151 /*
3152  * If dodgy hardware rejects the scatter gather chain we've handed it
3153  * we'll need to remove the mbuf chain from ifsg_m[] before we can add the
3154  * m_defrag'd mbufs
3155  */
3156 static __noinline struct mbuf *
3157 iflib_remove_mbuf(iflib_txq_t txq)
3158 {
3159 	int ntxd, pidx;
3160 	struct mbuf *m, **ifsd_m;
3161 
3162 	ifsd_m = txq->ift_sds.ifsd_m;
3163 	ntxd = txq->ift_size;
3164 	pidx = txq->ift_pidx & (ntxd - 1);
3165 	ifsd_m = txq->ift_sds.ifsd_m;
3166 	m = ifsd_m[pidx];
3167 	ifsd_m[pidx] = NULL;
3168 	bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[pidx]);
3169 	if (txq->ift_sds.ifsd_tso_map != NULL)
3170 		bus_dmamap_unload(txq->ift_tso_buf_tag,
3171 		    txq->ift_sds.ifsd_tso_map[pidx]);
3172 #if MEMORY_LOGGING
3173 	txq->ift_dequeued++;
3174 #endif
3175 	return (m);
3176 }
3177 
3178 static inline caddr_t
3179 calc_next_txd(iflib_txq_t txq, int cidx, uint8_t qid)
3180 {
3181 	qidx_t size;
3182 	int ntxd;
3183 	caddr_t start, end, cur, next;
3184 
3185 	ntxd = txq->ift_size;
3186 	size = txq->ift_txd_size[qid];
3187 	start = txq->ift_ifdi[qid].idi_vaddr;
3188 
3189 	if (__predict_false(size == 0))
3190 		return (start);
3191 	cur = start + size*cidx;
3192 	end = start + size*ntxd;
3193 	next = CACHE_PTR_NEXT(cur);
3194 	return (next < end ? next : start);
3195 }
3196 
3197 /*
3198  * Pad an mbuf to ensure a minimum ethernet frame size.
3199  * min_frame_size is the frame size (less CRC) to pad the mbuf to
3200  */
3201 static __noinline int
3202 iflib_ether_pad(device_t dev, struct mbuf **m_head, uint16_t min_frame_size)
3203 {
3204 	/*
3205 	 * 18 is enough bytes to pad an ARP packet to 46 bytes, and
3206 	 * and ARP message is the smallest common payload I can think of
3207 	 */
3208 	static char pad[18];	/* just zeros */
3209 	int n;
3210 	struct mbuf *new_head;
3211 
3212 	if (!M_WRITABLE(*m_head)) {
3213 		new_head = m_dup(*m_head, M_NOWAIT);
3214 		if (new_head == NULL) {
3215 			m_freem(*m_head);
3216 			device_printf(dev, "cannot pad short frame, m_dup() failed");
3217 			DBG_COUNTER_INC(encap_pad_mbuf_fail);
3218 			DBG_COUNTER_INC(tx_frees);
3219 			return ENOMEM;
3220 		}
3221 		m_freem(*m_head);
3222 		*m_head = new_head;
3223 	}
3224 
3225 	for (n = min_frame_size - (*m_head)->m_pkthdr.len;
3226 	     n > 0; n -= sizeof(pad))
3227 		if (!m_append(*m_head, min(n, sizeof(pad)), pad))
3228 			break;
3229 
3230 	if (n > 0) {
3231 		m_freem(*m_head);
3232 		device_printf(dev, "cannot pad short frame\n");
3233 		DBG_COUNTER_INC(encap_pad_mbuf_fail);
3234 		DBG_COUNTER_INC(tx_frees);
3235 		return (ENOBUFS);
3236 	}
3237 
3238 	return 0;
3239 }
3240 
3241 static int
3242 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp)
3243 {
3244 	if_ctx_t		ctx;
3245 	if_shared_ctx_t		sctx;
3246 	if_softc_ctx_t		scctx;
3247 	bus_dma_tag_t		buf_tag;
3248 	bus_dma_segment_t	*segs;
3249 	struct mbuf		*m_head, **ifsd_m;
3250 	void			*next_txd;
3251 	bus_dmamap_t		map;
3252 	struct if_pkt_info	pi;
3253 	int remap = 0;
3254 	int err, nsegs, ndesc, max_segs, pidx, cidx, next, ntxd;
3255 
3256 	ctx = txq->ift_ctx;
3257 	sctx = ctx->ifc_sctx;
3258 	scctx = &ctx->ifc_softc_ctx;
3259 	segs = txq->ift_segs;
3260 	ntxd = txq->ift_size;
3261 	m_head = *m_headp;
3262 	map = NULL;
3263 
3264 	/*
3265 	 * If we're doing TSO the next descriptor to clean may be quite far ahead
3266 	 */
3267 	cidx = txq->ift_cidx;
3268 	pidx = txq->ift_pidx;
3269 	if (ctx->ifc_flags & IFC_PREFETCH) {
3270 		next = (cidx + CACHE_PTR_INCREMENT) & (ntxd-1);
3271 		if (!(ctx->ifc_flags & IFLIB_HAS_TXCQ)) {
3272 			next_txd = calc_next_txd(txq, cidx, 0);
3273 			prefetch(next_txd);
3274 		}
3275 
3276 		/* prefetch the next cache line of mbuf pointers and flags */
3277 		prefetch(&txq->ift_sds.ifsd_m[next]);
3278 		prefetch(&txq->ift_sds.ifsd_map[next]);
3279 		next = (cidx + CACHE_LINE_SIZE) & (ntxd-1);
3280 	}
3281 	map = txq->ift_sds.ifsd_map[pidx];
3282 	ifsd_m = txq->ift_sds.ifsd_m;
3283 
3284 	if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
3285 		buf_tag = txq->ift_tso_buf_tag;
3286 		max_segs = scctx->isc_tx_tso_segments_max;
3287 		map = txq->ift_sds.ifsd_tso_map[pidx];
3288 		MPASS(buf_tag != NULL);
3289 		MPASS(max_segs > 0);
3290 	} else {
3291 		buf_tag = txq->ift_buf_tag;
3292 		max_segs = scctx->isc_tx_nsegments;
3293 		map = txq->ift_sds.ifsd_map[pidx];
3294 	}
3295 	if ((sctx->isc_flags & IFLIB_NEED_ETHER_PAD) &&
3296 	    __predict_false(m_head->m_pkthdr.len < scctx->isc_min_frame_size)) {
3297 		err = iflib_ether_pad(ctx->ifc_dev, m_headp, scctx->isc_min_frame_size);
3298 		if (err) {
3299 			DBG_COUNTER_INC(encap_txd_encap_fail);
3300 			return err;
3301 		}
3302 	}
3303 	m_head = *m_headp;
3304 
3305 	pkt_info_zero(&pi);
3306 	pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG|M_BCAST|M_MCAST));
3307 	pi.ipi_pidx = pidx;
3308 	pi.ipi_qsidx = txq->ift_id;
3309 	pi.ipi_len = m_head->m_pkthdr.len;
3310 	pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags;
3311 	pi.ipi_vtag = M_HAS_VLANTAG(m_head) ? m_head->m_pkthdr.ether_vtag : 0;
3312 
3313 	/* deliberate bitwise OR to make one condition */
3314 	if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) {
3315 		if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0)) {
3316 			DBG_COUNTER_INC(encap_txd_encap_fail);
3317 			return (err);
3318 		}
3319 		m_head = *m_headp;
3320 	}
3321 
3322 retry:
3323 	err = bus_dmamap_load_mbuf_sg(buf_tag, map, m_head, segs, &nsegs,
3324 	    BUS_DMA_NOWAIT);
3325 defrag:
3326 	if (__predict_false(err)) {
3327 		switch (err) {
3328 		case EFBIG:
3329 			/* try collapse once and defrag once */
3330 			if (remap == 0) {
3331 				m_head = m_collapse(*m_headp, M_NOWAIT, max_segs);
3332 				/* try defrag if collapsing fails */
3333 				if (m_head == NULL)
3334 					remap++;
3335 			}
3336 			if (remap == 1) {
3337 				txq->ift_mbuf_defrag++;
3338 				m_head = m_defrag(*m_headp, M_NOWAIT);
3339 			}
3340 			/*
3341 			 * remap should never be >1 unless bus_dmamap_load_mbuf_sg
3342 			 * failed to map an mbuf that was run through m_defrag
3343 			 */
3344 			MPASS(remap <= 1);
3345 			if (__predict_false(m_head == NULL || remap > 1))
3346 				goto defrag_failed;
3347 			remap++;
3348 			*m_headp = m_head;
3349 			goto retry;
3350 			break;
3351 		case ENOMEM:
3352 			txq->ift_no_tx_dma_setup++;
3353 			break;
3354 		default:
3355 			txq->ift_no_tx_dma_setup++;
3356 			m_freem(*m_headp);
3357 			DBG_COUNTER_INC(tx_frees);
3358 			*m_headp = NULL;
3359 			break;
3360 		}
3361 		txq->ift_map_failed++;
3362 		DBG_COUNTER_INC(encap_load_mbuf_fail);
3363 		DBG_COUNTER_INC(encap_txd_encap_fail);
3364 		return (err);
3365 	}
3366 	ifsd_m[pidx] = m_head;
3367 	/*
3368 	 * XXX assumes a 1 to 1 relationship between segments and
3369 	 *        descriptors - this does not hold true on all drivers, e.g.
3370 	 *        cxgb
3371 	 */
3372 	if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) {
3373 		txq->ift_no_desc_avail++;
3374 		bus_dmamap_unload(buf_tag, map);
3375 		DBG_COUNTER_INC(encap_txq_avail_fail);
3376 		DBG_COUNTER_INC(encap_txd_encap_fail);
3377 		if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0)
3378 			GROUPTASK_ENQUEUE(&txq->ift_task);
3379 		return (ENOBUFS);
3380 	}
3381 	/*
3382 	 * On Intel cards we can greatly reduce the number of TX interrupts
3383 	 * we see by only setting report status on every Nth descriptor.
3384 	 * However, this also means that the driver will need to keep track
3385 	 * of the descriptors that RS was set on to check them for the DD bit.
3386 	 */
3387 	txq->ift_rs_pending += nsegs + 1;
3388 	if (txq->ift_rs_pending > TXQ_MAX_RS_DEFERRED(txq) ||
3389 	     iflib_no_tx_batch || (TXQ_AVAIL(txq) - nsegs) <= MAX_TX_DESC(ctx) + 2) {
3390 		pi.ipi_flags |= IPI_TX_INTR;
3391 		txq->ift_rs_pending = 0;
3392 	}
3393 
3394 	pi.ipi_segs = segs;
3395 	pi.ipi_nsegs = nsegs;
3396 
3397 	MPASS(pidx >= 0 && pidx < txq->ift_size);
3398 #ifdef PKT_DEBUG
3399 	print_pkt(&pi);
3400 #endif
3401 	if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) {
3402 		bus_dmamap_sync(buf_tag, map, BUS_DMASYNC_PREWRITE);
3403 		DBG_COUNTER_INC(tx_encap);
3404 		MPASS(pi.ipi_new_pidx < txq->ift_size);
3405 
3406 		ndesc = pi.ipi_new_pidx - pi.ipi_pidx;
3407 		if (pi.ipi_new_pidx < pi.ipi_pidx) {
3408 			ndesc += txq->ift_size;
3409 			txq->ift_gen = 1;
3410 		}
3411 		/*
3412 		 * drivers can need as many as
3413 		 * two sentinels
3414 		 */
3415 		MPASS(ndesc <= pi.ipi_nsegs + 2);
3416 		MPASS(pi.ipi_new_pidx != pidx);
3417 		MPASS(ndesc > 0);
3418 		txq->ift_in_use += ndesc;
3419 
3420 		/*
3421 		 * We update the last software descriptor again here because there may
3422 		 * be a sentinel and/or there may be more mbufs than segments
3423 		 */
3424 		txq->ift_pidx = pi.ipi_new_pidx;
3425 		txq->ift_npending += pi.ipi_ndescs;
3426 	} else {
3427 		*m_headp = m_head = iflib_remove_mbuf(txq);
3428 		if (err == EFBIG) {
3429 			txq->ift_txd_encap_efbig++;
3430 			if (remap < 2) {
3431 				remap = 1;
3432 				goto defrag;
3433 			}
3434 		}
3435 		goto defrag_failed;
3436 	}
3437 	/*
3438 	 * err can't possibly be non-zero here, so we don't neet to test it
3439 	 * to see if we need to DBG_COUNTER_INC(encap_txd_encap_fail).
3440 	 */
3441 	return (err);
3442 
3443 defrag_failed:
3444 	txq->ift_mbuf_defrag_failed++;
3445 	txq->ift_map_failed++;
3446 	m_freem(*m_headp);
3447 	DBG_COUNTER_INC(tx_frees);
3448 	*m_headp = NULL;
3449 	DBG_COUNTER_INC(encap_txd_encap_fail);
3450 	return (ENOMEM);
3451 }
3452 
3453 static void
3454 iflib_tx_desc_free(iflib_txq_t txq, int n)
3455 {
3456 	uint32_t qsize, cidx, mask, gen;
3457 	struct mbuf *m, **ifsd_m;
3458 	bool do_prefetch;
3459 
3460 	cidx = txq->ift_cidx;
3461 	gen = txq->ift_gen;
3462 	qsize = txq->ift_size;
3463 	mask = qsize-1;
3464 	ifsd_m = txq->ift_sds.ifsd_m;
3465 	do_prefetch = (txq->ift_ctx->ifc_flags & IFC_PREFETCH);
3466 
3467 	while (n-- > 0) {
3468 		if (do_prefetch) {
3469 			prefetch(ifsd_m[(cidx + 3) & mask]);
3470 			prefetch(ifsd_m[(cidx + 4) & mask]);
3471 		}
3472 		if ((m = ifsd_m[cidx]) != NULL) {
3473 			prefetch(&ifsd_m[(cidx + CACHE_PTR_INCREMENT) & mask]);
3474 			if (m->m_pkthdr.csum_flags & CSUM_TSO) {
3475 				bus_dmamap_sync(txq->ift_tso_buf_tag,
3476 				    txq->ift_sds.ifsd_tso_map[cidx],
3477 				    BUS_DMASYNC_POSTWRITE);
3478 				bus_dmamap_unload(txq->ift_tso_buf_tag,
3479 				    txq->ift_sds.ifsd_tso_map[cidx]);
3480 			} else {
3481 				bus_dmamap_sync(txq->ift_buf_tag,
3482 				    txq->ift_sds.ifsd_map[cidx],
3483 				    BUS_DMASYNC_POSTWRITE);
3484 				bus_dmamap_unload(txq->ift_buf_tag,
3485 				    txq->ift_sds.ifsd_map[cidx]);
3486 			}
3487 			/* XXX we don't support any drivers that batch packets yet */
3488 			MPASS(m->m_nextpkt == NULL);
3489 			m_freem(m);
3490 			ifsd_m[cidx] = NULL;
3491 #if MEMORY_LOGGING
3492 			txq->ift_dequeued++;
3493 #endif
3494 			DBG_COUNTER_INC(tx_frees);
3495 		}
3496 		if (__predict_false(++cidx == qsize)) {
3497 			cidx = 0;
3498 			gen = 0;
3499 		}
3500 	}
3501 	txq->ift_cidx = cidx;
3502 	txq->ift_gen = gen;
3503 }
3504 
3505 static __inline int
3506 iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh)
3507 {
3508 	int reclaim;
3509 	if_ctx_t ctx = txq->ift_ctx;
3510 
3511 	KASSERT(thresh >= 0, ("invalid threshold to reclaim"));
3512 	MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size);
3513 
3514 	/*
3515 	 * Need a rate-limiting check so that this isn't called every time
3516 	 */
3517 	iflib_tx_credits_update(ctx, txq);
3518 	reclaim = DESC_RECLAIMABLE(txq);
3519 
3520 	if (reclaim <= thresh /* + MAX_TX_DESC(txq->ift_ctx) */) {
3521 #ifdef INVARIANTS
3522 		if (iflib_verbose_debug) {
3523 			printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __FUNCTION__,
3524 			       txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments,
3525 			       reclaim, thresh);
3526 
3527 		}
3528 #endif
3529 		return (0);
3530 	}
3531 	iflib_tx_desc_free(txq, reclaim);
3532 	txq->ift_cleaned += reclaim;
3533 	txq->ift_in_use -= reclaim;
3534 
3535 	return (reclaim);
3536 }
3537 
3538 static struct mbuf **
3539 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset, int remaining)
3540 {
3541 	int next, size;
3542 	struct mbuf **items;
3543 
3544 	size = r->size;
3545 	next = (cidx + CACHE_PTR_INCREMENT) & (size-1);
3546 	items = __DEVOLATILE(struct mbuf **, &r->items[0]);
3547 
3548 	prefetch(items[(cidx + offset) & (size-1)]);
3549 	if (remaining > 1) {
3550 		prefetch2cachelines(&items[next]);
3551 		prefetch2cachelines(items[(cidx + offset + 1) & (size-1)]);
3552 		prefetch2cachelines(items[(cidx + offset + 2) & (size-1)]);
3553 		prefetch2cachelines(items[(cidx + offset + 3) & (size-1)]);
3554 	}
3555 	return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (size-1)]));
3556 }
3557 
3558 static void
3559 iflib_txq_check_drain(iflib_txq_t txq, int budget)
3560 {
3561 
3562 	ifmp_ring_check_drainage(txq->ift_br, budget);
3563 }
3564 
3565 static uint32_t
3566 iflib_txq_can_drain(struct ifmp_ring *r)
3567 {
3568 	iflib_txq_t txq = r->cookie;
3569 	if_ctx_t ctx = txq->ift_ctx;
3570 
3571 	if (TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2)
3572 		return (1);
3573 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3574 	    BUS_DMASYNC_POSTREAD);
3575 	return (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id,
3576 	    false));
3577 }
3578 
3579 static uint32_t
3580 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3581 {
3582 	iflib_txq_t txq = r->cookie;
3583 	if_ctx_t ctx = txq->ift_ctx;
3584 	if_t ifp = ctx->ifc_ifp;
3585 	struct mbuf **mp, *m;
3586 	int i, count, consumed, pkt_sent, bytes_sent, mcast_sent, avail;
3587 	int reclaimed, err, in_use_prev, desc_used;
3588 	bool do_prefetch, ring, rang;
3589 
3590 	if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) ||
3591 			    !LINK_ACTIVE(ctx))) {
3592 		DBG_COUNTER_INC(txq_drain_notready);
3593 		return (0);
3594 	}
3595 	reclaimed = iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
3596 	rang = iflib_txd_db_check(ctx, txq, reclaimed, txq->ift_in_use);
3597 	avail = IDXDIFF(pidx, cidx, r->size);
3598 	if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) {
3599 		DBG_COUNTER_INC(txq_drain_flushing);
3600 		for (i = 0; i < avail; i++) {
3601 			if (__predict_true(r->items[(cidx + i) & (r->size-1)] != (void *)txq))
3602 				m_free(r->items[(cidx + i) & (r->size-1)]);
3603 			r->items[(cidx + i) & (r->size-1)] = NULL;
3604 		}
3605 		return (avail);
3606 	}
3607 
3608 	if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3609 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3610 		CALLOUT_LOCK(txq);
3611 		callout_stop(&txq->ift_timer);
3612 		CALLOUT_UNLOCK(txq);
3613 		DBG_COUNTER_INC(txq_drain_oactive);
3614 		return (0);
3615 	}
3616 	if (reclaimed)
3617 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3618 	consumed = mcast_sent = bytes_sent = pkt_sent = 0;
3619 	count = MIN(avail, TX_BATCH_SIZE);
3620 #ifdef INVARIANTS
3621 	if (iflib_verbose_debug)
3622 		printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __FUNCTION__,
3623 		       avail, ctx->ifc_flags, TXQ_AVAIL(txq));
3624 #endif
3625 	do_prefetch = (ctx->ifc_flags & IFC_PREFETCH);
3626 	avail = TXQ_AVAIL(txq);
3627 	err = 0;
3628 	for (desc_used = i = 0; i < count && avail > MAX_TX_DESC(ctx) + 2; i++) {
3629 		int rem = do_prefetch ? count - i : 0;
3630 
3631 		mp = _ring_peek_one(r, cidx, i, rem);
3632 		MPASS(mp != NULL && *mp != NULL);
3633 		if (__predict_false(*mp == (struct mbuf *)txq)) {
3634 			consumed++;
3635 			reclaimed++;
3636 			continue;
3637 		}
3638 		in_use_prev = txq->ift_in_use;
3639 		err = iflib_encap(txq, mp);
3640 		if (__predict_false(err)) {
3641 			/* no room - bail out */
3642 			if (err == ENOBUFS)
3643 				break;
3644 			consumed++;
3645 			/* we can't send this packet - skip it */
3646 			continue;
3647 		}
3648 		consumed++;
3649 		pkt_sent++;
3650 		m = *mp;
3651 		DBG_COUNTER_INC(tx_sent);
3652 		bytes_sent += m->m_pkthdr.len;
3653 		mcast_sent += !!(m->m_flags & M_MCAST);
3654 		avail = TXQ_AVAIL(txq);
3655 
3656 		txq->ift_db_pending += (txq->ift_in_use - in_use_prev);
3657 		desc_used += (txq->ift_in_use - in_use_prev);
3658 		ETHER_BPF_MTAP(ifp, m);
3659 		if (__predict_false(!(ifp->if_drv_flags & IFF_DRV_RUNNING)))
3660 			break;
3661 		rang = iflib_txd_db_check(ctx, txq, false, in_use_prev);
3662 	}
3663 
3664 	/* deliberate use of bitwise or to avoid gratuitous short-circuit */
3665 	ring = rang ? false  : (iflib_min_tx_latency | err) || (TXQ_AVAIL(txq) < MAX_TX_DESC(ctx));
3666 	iflib_txd_db_check(ctx, txq, ring, txq->ift_in_use);
3667 	if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
3668 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
3669 	if (mcast_sent)
3670 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
3671 #ifdef INVARIANTS
3672 	if (iflib_verbose_debug)
3673 		printf("consumed=%d\n", consumed);
3674 #endif
3675 	return (consumed);
3676 }
3677 
3678 static uint32_t
3679 iflib_txq_drain_always(struct ifmp_ring *r)
3680 {
3681 	return (1);
3682 }
3683 
3684 static uint32_t
3685 iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3686 {
3687 	int i, avail;
3688 	struct mbuf **mp;
3689 	iflib_txq_t txq;
3690 
3691 	txq = r->cookie;
3692 
3693 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3694 	CALLOUT_LOCK(txq);
3695 	callout_stop(&txq->ift_timer);
3696 	CALLOUT_UNLOCK(txq);
3697 
3698 	avail = IDXDIFF(pidx, cidx, r->size);
3699 	for (i = 0; i < avail; i++) {
3700 		mp = _ring_peek_one(r, cidx, i, avail - i);
3701 		if (__predict_false(*mp == (struct mbuf *)txq))
3702 			continue;
3703 		m_freem(*mp);
3704 		DBG_COUNTER_INC(tx_frees);
3705 	}
3706 	MPASS(ifmp_ring_is_stalled(r) == 0);
3707 	return (avail);
3708 }
3709 
3710 static void
3711 iflib_ifmp_purge(iflib_txq_t txq)
3712 {
3713 	struct ifmp_ring *r;
3714 
3715 	r = txq->ift_br;
3716 	r->drain = iflib_txq_drain_free;
3717 	r->can_drain = iflib_txq_drain_always;
3718 
3719 	ifmp_ring_check_drainage(r, r->size);
3720 
3721 	r->drain = iflib_txq_drain;
3722 	r->can_drain = iflib_txq_can_drain;
3723 }
3724 
3725 static void
3726 _task_fn_tx(void *context)
3727 {
3728 	iflib_txq_t txq = context;
3729 	if_ctx_t ctx = txq->ift_ctx;
3730 #if defined(ALTQ) || defined(DEV_NETMAP)
3731 	if_t ifp = ctx->ifc_ifp;
3732 #endif
3733 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
3734 
3735 #ifdef IFLIB_DIAGNOSTICS
3736 	txq->ift_cpu_exec_count[curcpu]++;
3737 #endif
3738 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3739 		return;
3740 #ifdef DEV_NETMAP
3741 	if (if_getcapenable(ifp) & IFCAP_NETMAP) {
3742 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3743 		    BUS_DMASYNC_POSTREAD);
3744 		if (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false))
3745 			netmap_tx_irq(ifp, txq->ift_id);
3746 		if (ctx->ifc_flags & IFC_LEGACY)
3747 			IFDI_INTR_ENABLE(ctx);
3748 		else
3749 			IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3750 		return;
3751 	}
3752 #endif
3753 #ifdef ALTQ
3754 	if (ALTQ_IS_ENABLED(&ifp->if_snd))
3755 		iflib_altq_if_start(ifp);
3756 #endif
3757 	if (txq->ift_db_pending)
3758 		ifmp_ring_enqueue(txq->ift_br, (void **)&txq, 1, TX_BATCH_SIZE, abdicate);
3759 	else if (!abdicate)
3760 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3761 	/*
3762 	 * When abdicating, we always need to check drainage, not just when we don't enqueue
3763 	 */
3764 	if (abdicate)
3765 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3766 	if (ctx->ifc_flags & IFC_LEGACY)
3767 		IFDI_INTR_ENABLE(ctx);
3768 	else
3769 		IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3770 }
3771 
3772 static void
3773 _task_fn_rx(void *context)
3774 {
3775 	iflib_rxq_t rxq = context;
3776 	if_ctx_t ctx = rxq->ifr_ctx;
3777 	bool more;
3778 	uint16_t budget;
3779 
3780 #ifdef IFLIB_DIAGNOSTICS
3781 	rxq->ifr_cpu_exec_count[curcpu]++;
3782 #endif
3783 	DBG_COUNTER_INC(task_fn_rxs);
3784 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3785 		return;
3786 	more = true;
3787 #ifdef DEV_NETMAP
3788 	if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP) {
3789 		u_int work = 0;
3790 		if (netmap_rx_irq(ctx->ifc_ifp, rxq->ifr_id, &work)) {
3791 			more = false;
3792 		}
3793 	}
3794 #endif
3795 	budget = ctx->ifc_sysctl_rx_budget;
3796 	if (budget == 0)
3797 		budget = 16;	/* XXX */
3798 	if (more == false || (more = iflib_rxeof(rxq, budget)) == false) {
3799 		if (ctx->ifc_flags & IFC_LEGACY)
3800 			IFDI_INTR_ENABLE(ctx);
3801 		else
3802 			IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
3803 		DBG_COUNTER_INC(rx_intr_enables);
3804 	}
3805 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3806 		return;
3807 	if (more)
3808 		GROUPTASK_ENQUEUE(&rxq->ifr_task);
3809 }
3810 
3811 static void
3812 _task_fn_admin(void *context)
3813 {
3814 	if_ctx_t ctx = context;
3815 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
3816 	iflib_txq_t txq;
3817 	int i;
3818 	bool oactive, running, do_reset, do_watchdog, in_detach;
3819 	uint32_t reset_on = hz / 2;
3820 
3821 	STATE_LOCK(ctx);
3822 	running = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING);
3823 	oactive = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE);
3824 	do_reset = (ctx->ifc_flags & IFC_DO_RESET);
3825 	do_watchdog = (ctx->ifc_flags & IFC_DO_WATCHDOG);
3826 	in_detach = (ctx->ifc_flags & IFC_IN_DETACH);
3827 	ctx->ifc_flags &= ~(IFC_DO_RESET|IFC_DO_WATCHDOG);
3828 	STATE_UNLOCK(ctx);
3829 
3830 	if ((!running && !oactive) && !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
3831 		return;
3832 	if (in_detach)
3833 		return;
3834 
3835 	CTX_LOCK(ctx);
3836 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3837 		CALLOUT_LOCK(txq);
3838 		callout_stop(&txq->ift_timer);
3839 		CALLOUT_UNLOCK(txq);
3840 	}
3841 	if (do_watchdog) {
3842 		ctx->ifc_watchdog_events++;
3843 		IFDI_WATCHDOG_RESET(ctx);
3844 	}
3845 	IFDI_UPDATE_ADMIN_STATUS(ctx);
3846 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3847 #ifdef DEV_NETMAP
3848 		reset_on = hz / 2;
3849 		if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP)
3850 			iflib_netmap_timer_adjust(ctx, txq, &reset_on);
3851 #endif
3852 		callout_reset_on(&txq->ift_timer, reset_on, iflib_timer, txq, txq->ift_timer.c_cpu);
3853 	}
3854 	IFDI_LINK_INTR_ENABLE(ctx);
3855 	if (do_reset)
3856 		iflib_if_init_locked(ctx);
3857 	CTX_UNLOCK(ctx);
3858 
3859 	if (LINK_ACTIVE(ctx) == 0)
3860 		return;
3861 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3862 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
3863 }
3864 
3865 
3866 static void
3867 _task_fn_iov(void *context)
3868 {
3869 	if_ctx_t ctx = context;
3870 
3871 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING) &&
3872 	    !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
3873 		return;
3874 
3875 	CTX_LOCK(ctx);
3876 	IFDI_VFLR_HANDLE(ctx);
3877 	CTX_UNLOCK(ctx);
3878 }
3879 
3880 static int
3881 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)
3882 {
3883 	int err;
3884 	if_int_delay_info_t info;
3885 	if_ctx_t ctx;
3886 
3887 	info = (if_int_delay_info_t)arg1;
3888 	ctx = info->iidi_ctx;
3889 	info->iidi_req = req;
3890 	info->iidi_oidp = oidp;
3891 	CTX_LOCK(ctx);
3892 	err = IFDI_SYSCTL_INT_DELAY(ctx, info);
3893 	CTX_UNLOCK(ctx);
3894 	return (err);
3895 }
3896 
3897 /*********************************************************************
3898  *
3899  *  IFNET FUNCTIONS
3900  *
3901  **********************************************************************/
3902 
3903 static void
3904 iflib_if_init_locked(if_ctx_t ctx)
3905 {
3906 	iflib_stop(ctx);
3907 	iflib_init_locked(ctx);
3908 }
3909 
3910 
3911 static void
3912 iflib_if_init(void *arg)
3913 {
3914 	if_ctx_t ctx = arg;
3915 
3916 	CTX_LOCK(ctx);
3917 	iflib_if_init_locked(ctx);
3918 	CTX_UNLOCK(ctx);
3919 }
3920 
3921 static int
3922 iflib_if_transmit(if_t ifp, struct mbuf *m)
3923 {
3924 	if_ctx_t	ctx = if_getsoftc(ifp);
3925 
3926 	iflib_txq_t txq;
3927 	int err, qidx;
3928 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
3929 
3930 	if (__predict_false((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) {
3931 		DBG_COUNTER_INC(tx_frees);
3932 		m_freem(m);
3933 		return (ENETDOWN);
3934 	}
3935 
3936 	MPASS(m->m_nextpkt == NULL);
3937 	/* ALTQ-enabled interfaces always use queue 0. */
3938 	qidx = 0;
3939 	if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m) && !ALTQ_IS_ENABLED(&ifp->if_snd))
3940 		qidx = QIDX(ctx, m);
3941 	/*
3942 	 * XXX calculate buf_ring based on flowid (divvy up bits?)
3943 	 */
3944 	txq = &ctx->ifc_txqs[qidx];
3945 
3946 #ifdef DRIVER_BACKPRESSURE
3947 	if (txq->ift_closed) {
3948 		while (m != NULL) {
3949 			next = m->m_nextpkt;
3950 			m->m_nextpkt = NULL;
3951 			m_freem(m);
3952 			DBG_COUNTER_INC(tx_frees);
3953 			m = next;
3954 		}
3955 		return (ENOBUFS);
3956 	}
3957 #endif
3958 #ifdef notyet
3959 	qidx = count = 0;
3960 	mp = marr;
3961 	next = m;
3962 	do {
3963 		count++;
3964 		next = next->m_nextpkt;
3965 	} while (next != NULL);
3966 
3967 	if (count > nitems(marr))
3968 		if ((mp = malloc(count*sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) {
3969 			/* XXX check nextpkt */
3970 			m_freem(m);
3971 			/* XXX simplify for now */
3972 			DBG_COUNTER_INC(tx_frees);
3973 			return (ENOBUFS);
3974 		}
3975 	for (next = m, i = 0; next != NULL; i++) {
3976 		mp[i] = next;
3977 		next = next->m_nextpkt;
3978 		mp[i]->m_nextpkt = NULL;
3979 	}
3980 #endif
3981 	DBG_COUNTER_INC(tx_seen);
3982 	err = ifmp_ring_enqueue(txq->ift_br, (void **)&m, 1, TX_BATCH_SIZE, abdicate);
3983 
3984 	if (abdicate)
3985 		GROUPTASK_ENQUEUE(&txq->ift_task);
3986  	if (err) {
3987 		if (!abdicate)
3988 			GROUPTASK_ENQUEUE(&txq->ift_task);
3989 		/* support forthcoming later */
3990 #ifdef DRIVER_BACKPRESSURE
3991 		txq->ift_closed = TRUE;
3992 #endif
3993 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3994 		m_freem(m);
3995 		DBG_COUNTER_INC(tx_frees);
3996 	}
3997 
3998 	return (err);
3999 }
4000 
4001 #ifdef ALTQ
4002 /*
4003  * The overall approach to integrating iflib with ALTQ is to continue to use
4004  * the iflib mp_ring machinery between the ALTQ queue(s) and the hardware
4005  * ring.  Technically, when using ALTQ, queueing to an intermediate mp_ring
4006  * is redundant/unnecessary, but doing so minimizes the amount of
4007  * ALTQ-specific code required in iflib.  It is assumed that the overhead of
4008  * redundantly queueing to an intermediate mp_ring is swamped by the
4009  * performance limitations inherent in using ALTQ.
4010  *
4011  * When ALTQ support is compiled in, all iflib drivers will use a transmit
4012  * routine, iflib_altq_if_transmit(), that checks if ALTQ is enabled for the
4013  * given interface.  If ALTQ is enabled for an interface, then all
4014  * transmitted packets for that interface will be submitted to the ALTQ
4015  * subsystem via IFQ_ENQUEUE().  We don't use the legacy if_transmit()
4016  * implementation because it uses IFQ_HANDOFF(), which will duplicatively
4017  * update stats that the iflib machinery handles, and which is sensitve to
4018  * the disused IFF_DRV_OACTIVE flag.  Additionally, iflib_altq_if_start()
4019  * will be installed as the start routine for use by ALTQ facilities that
4020  * need to trigger queue drains on a scheduled basis.
4021  *
4022  */
4023 static void
4024 iflib_altq_if_start(if_t ifp)
4025 {
4026 	struct ifaltq *ifq = &ifp->if_snd;
4027 	struct mbuf *m;
4028 
4029 	IFQ_LOCK(ifq);
4030 	IFQ_DEQUEUE_NOLOCK(ifq, m);
4031 	while (m != NULL) {
4032 		iflib_if_transmit(ifp, m);
4033 		IFQ_DEQUEUE_NOLOCK(ifq, m);
4034 	}
4035 	IFQ_UNLOCK(ifq);
4036 }
4037 
4038 static int
4039 iflib_altq_if_transmit(if_t ifp, struct mbuf *m)
4040 {
4041 	int err;
4042 
4043 	if (ALTQ_IS_ENABLED(&ifp->if_snd)) {
4044 		IFQ_ENQUEUE(&ifp->if_snd, m, err);
4045 		if (err == 0)
4046 			iflib_altq_if_start(ifp);
4047 	} else
4048 		err = iflib_if_transmit(ifp, m);
4049 
4050 	return (err);
4051 }
4052 #endif /* ALTQ */
4053 
4054 static void
4055 iflib_if_qflush(if_t ifp)
4056 {
4057 	if_ctx_t ctx = if_getsoftc(ifp);
4058 	iflib_txq_t txq = ctx->ifc_txqs;
4059 	int i;
4060 
4061 	STATE_LOCK(ctx);
4062 	ctx->ifc_flags |= IFC_QFLUSH;
4063 	STATE_UNLOCK(ctx);
4064 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
4065 		while (!(ifmp_ring_is_idle(txq->ift_br) || ifmp_ring_is_stalled(txq->ift_br)))
4066 			iflib_txq_check_drain(txq, 0);
4067 	STATE_LOCK(ctx);
4068 	ctx->ifc_flags &= ~IFC_QFLUSH;
4069 	STATE_UNLOCK(ctx);
4070 
4071 	/*
4072 	 * When ALTQ is enabled, this will also take care of purging the
4073 	 * ALTQ queue(s).
4074 	 */
4075 	if_qflush(ifp);
4076 }
4077 
4078 
4079 #define IFCAP_FLAGS (IFCAP_HWCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \
4080 		     IFCAP_TSO | IFCAP_VLAN_HWTAGGING | IFCAP_HWSTATS | \
4081 		     IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | \
4082 		     IFCAP_VLAN_HWTSO | IFCAP_VLAN_HWCSUM)
4083 
4084 static int
4085 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data)
4086 {
4087 	if_ctx_t ctx = if_getsoftc(ifp);
4088 	struct ifreq	*ifr = (struct ifreq *)data;
4089 #if defined(INET) || defined(INET6)
4090 	struct ifaddr	*ifa = (struct ifaddr *)data;
4091 #endif
4092 	bool		avoid_reset = false;
4093 	int		err = 0, reinit = 0, bits;
4094 
4095 	switch (command) {
4096 	case SIOCSIFADDR:
4097 #ifdef INET
4098 		if (ifa->ifa_addr->sa_family == AF_INET)
4099 			avoid_reset = true;
4100 #endif
4101 #ifdef INET6
4102 		if (ifa->ifa_addr->sa_family == AF_INET6)
4103 			avoid_reset = true;
4104 #endif
4105 		/*
4106 		** Calling init results in link renegotiation,
4107 		** so we avoid doing it when possible.
4108 		*/
4109 		if (avoid_reset) {
4110 			if_setflagbits(ifp, IFF_UP,0);
4111 			if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
4112 				reinit = 1;
4113 #ifdef INET
4114 			if (!(if_getflags(ifp) & IFF_NOARP))
4115 				arp_ifinit(ifp, ifa);
4116 #endif
4117 		} else
4118 			err = ether_ioctl(ifp, command, data);
4119 		break;
4120 	case SIOCSIFMTU:
4121 		CTX_LOCK(ctx);
4122 		if (ifr->ifr_mtu == if_getmtu(ifp)) {
4123 			CTX_UNLOCK(ctx);
4124 			break;
4125 		}
4126 		bits = if_getdrvflags(ifp);
4127 		/* stop the driver and free any clusters before proceeding */
4128 		iflib_stop(ctx);
4129 
4130 		if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) {
4131 			STATE_LOCK(ctx);
4132 			if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size)
4133 				ctx->ifc_flags |= IFC_MULTISEG;
4134 			else
4135 				ctx->ifc_flags &= ~IFC_MULTISEG;
4136 			STATE_UNLOCK(ctx);
4137 			err = if_setmtu(ifp, ifr->ifr_mtu);
4138 		}
4139 		iflib_init_locked(ctx);
4140 		STATE_LOCK(ctx);
4141 		if_setdrvflags(ifp, bits);
4142 		STATE_UNLOCK(ctx);
4143 		CTX_UNLOCK(ctx);
4144 		break;
4145 	case SIOCSIFFLAGS:
4146 		CTX_LOCK(ctx);
4147 		if (if_getflags(ifp) & IFF_UP) {
4148 			if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4149 				if ((if_getflags(ifp) ^ ctx->ifc_if_flags) &
4150 				    (IFF_PROMISC | IFF_ALLMULTI)) {
4151 					err = IFDI_PROMISC_SET(ctx, if_getflags(ifp));
4152 				}
4153 			} else
4154 				reinit = 1;
4155 		} else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4156 			iflib_stop(ctx);
4157 		}
4158 		ctx->ifc_if_flags = if_getflags(ifp);
4159 		CTX_UNLOCK(ctx);
4160 		break;
4161 	case SIOCADDMULTI:
4162 	case SIOCDELMULTI:
4163 		if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4164 			CTX_LOCK(ctx);
4165 			IFDI_INTR_DISABLE(ctx);
4166 			IFDI_MULTI_SET(ctx);
4167 			IFDI_INTR_ENABLE(ctx);
4168 			CTX_UNLOCK(ctx);
4169 		}
4170 		break;
4171 	case SIOCSIFMEDIA:
4172 		CTX_LOCK(ctx);
4173 		IFDI_MEDIA_SET(ctx);
4174 		CTX_UNLOCK(ctx);
4175 		/* FALLTHROUGH */
4176 	case SIOCGIFMEDIA:
4177 	case SIOCGIFXMEDIA:
4178 		err = ifmedia_ioctl(ifp, ifr, ctx->ifc_mediap, command);
4179 		break;
4180 	case SIOCGI2C:
4181 	{
4182 		struct ifi2creq i2c;
4183 
4184 		err = copyin(ifr_data_get_ptr(ifr), &i2c, sizeof(i2c));
4185 		if (err != 0)
4186 			break;
4187 		if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) {
4188 			err = EINVAL;
4189 			break;
4190 		}
4191 		if (i2c.len > sizeof(i2c.data)) {
4192 			err = EINVAL;
4193 			break;
4194 		}
4195 
4196 		if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0)
4197 			err = copyout(&i2c, ifr_data_get_ptr(ifr),
4198 			    sizeof(i2c));
4199 		break;
4200 	}
4201 	case SIOCSIFCAP:
4202 	{
4203 		int mask, setmask, oldmask;
4204 
4205 		oldmask = if_getcapenable(ifp);
4206 		mask = ifr->ifr_reqcap ^ oldmask;
4207 		mask &= ctx->ifc_softc_ctx.isc_capabilities;
4208 		setmask = 0;
4209 #ifdef TCP_OFFLOAD
4210 		setmask |= mask & (IFCAP_TOE4|IFCAP_TOE6);
4211 #endif
4212 		setmask |= (mask & IFCAP_FLAGS);
4213 		setmask |= (mask & IFCAP_WOL);
4214 
4215 		/*
4216 		 * If any RX csum has changed, change all the ones that
4217 		 * are supported by the driver.
4218 		 */
4219 		if (setmask & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6)) {
4220 			setmask |= ctx->ifc_softc_ctx.isc_capabilities &
4221 			    (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6);
4222 		}
4223 
4224 		/*
4225 		 * want to ensure that traffic has stopped before we change any of the flags
4226 		 */
4227 		if (setmask) {
4228 			CTX_LOCK(ctx);
4229 			bits = if_getdrvflags(ifp);
4230 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4231 				iflib_stop(ctx);
4232 			STATE_LOCK(ctx);
4233 			if_togglecapenable(ifp, setmask);
4234 			STATE_UNLOCK(ctx);
4235 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4236 				iflib_init_locked(ctx);
4237 			STATE_LOCK(ctx);
4238 			if_setdrvflags(ifp, bits);
4239 			STATE_UNLOCK(ctx);
4240 			CTX_UNLOCK(ctx);
4241 		}
4242 		if_vlancap(ifp);
4243 		break;
4244 	}
4245 	case SIOCGPRIVATE_0:
4246 	case SIOCSDRVSPEC:
4247 	case SIOCGDRVSPEC:
4248 		CTX_LOCK(ctx);
4249 		err = IFDI_PRIV_IOCTL(ctx, command, data);
4250 		CTX_UNLOCK(ctx);
4251 		break;
4252 	default:
4253 		err = ether_ioctl(ifp, command, data);
4254 		break;
4255 	}
4256 	if (reinit)
4257 		iflib_if_init(ctx);
4258 	return (err);
4259 }
4260 
4261 static uint64_t
4262 iflib_if_get_counter(if_t ifp, ift_counter cnt)
4263 {
4264 	if_ctx_t ctx = if_getsoftc(ifp);
4265 
4266 	return (IFDI_GET_COUNTER(ctx, cnt));
4267 }
4268 
4269 /*********************************************************************
4270  *
4271  *  OTHER FUNCTIONS EXPORTED TO THE STACK
4272  *
4273  **********************************************************************/
4274 
4275 static void
4276 iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag)
4277 {
4278 	if_ctx_t ctx = if_getsoftc(ifp);
4279 
4280 	if ((void *)ctx != arg)
4281 		return;
4282 
4283 	if ((vtag == 0) || (vtag > 4095))
4284 		return;
4285 
4286 	CTX_LOCK(ctx);
4287 	IFDI_VLAN_REGISTER(ctx, vtag);
4288 	/* Re-init to load the changes */
4289 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
4290 		iflib_if_init_locked(ctx);
4291 	CTX_UNLOCK(ctx);
4292 }
4293 
4294 static void
4295 iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag)
4296 {
4297 	if_ctx_t ctx = if_getsoftc(ifp);
4298 
4299 	if ((void *)ctx != arg)
4300 		return;
4301 
4302 	if ((vtag == 0) || (vtag > 4095))
4303 		return;
4304 
4305 	CTX_LOCK(ctx);
4306 	IFDI_VLAN_UNREGISTER(ctx, vtag);
4307 	/* Re-init to load the changes */
4308 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
4309 		iflib_if_init_locked(ctx);
4310 	CTX_UNLOCK(ctx);
4311 }
4312 
4313 static void
4314 iflib_led_func(void *arg, int onoff)
4315 {
4316 	if_ctx_t ctx = arg;
4317 
4318 	CTX_LOCK(ctx);
4319 	IFDI_LED_FUNC(ctx, onoff);
4320 	CTX_UNLOCK(ctx);
4321 }
4322 
4323 /*********************************************************************
4324  *
4325  *  BUS FUNCTION DEFINITIONS
4326  *
4327  **********************************************************************/
4328 
4329 int
4330 iflib_device_probe(device_t dev)
4331 {
4332 	const pci_vendor_info_t *ent;
4333 	if_shared_ctx_t sctx;
4334 	uint16_t pci_device_id, pci_rev_id, pci_subdevice_id, pci_subvendor_id;
4335 	uint16_t pci_vendor_id;
4336 
4337 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4338 		return (ENOTSUP);
4339 
4340 	pci_vendor_id = pci_get_vendor(dev);
4341 	pci_device_id = pci_get_device(dev);
4342 	pci_subvendor_id = pci_get_subvendor(dev);
4343 	pci_subdevice_id = pci_get_subdevice(dev);
4344 	pci_rev_id = pci_get_revid(dev);
4345 	if (sctx->isc_parse_devinfo != NULL)
4346 		sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id);
4347 
4348 	ent = sctx->isc_vendor_info;
4349 	while (ent->pvi_vendor_id != 0) {
4350 		if (pci_vendor_id != ent->pvi_vendor_id) {
4351 			ent++;
4352 			continue;
4353 		}
4354 		if ((pci_device_id == ent->pvi_device_id) &&
4355 		    ((pci_subvendor_id == ent->pvi_subvendor_id) ||
4356 		     (ent->pvi_subvendor_id == 0)) &&
4357 		    ((pci_subdevice_id == ent->pvi_subdevice_id) ||
4358 		     (ent->pvi_subdevice_id == 0)) &&
4359 		    ((pci_rev_id == ent->pvi_rev_id) ||
4360 		     (ent->pvi_rev_id == 0))) {
4361 
4362 			device_set_desc_copy(dev, ent->pvi_name);
4363 			/* this needs to be changed to zero if the bus probing code
4364 			 * ever stops re-probing on best match because the sctx
4365 			 * may have its values over written by register calls
4366 			 * in subsequent probes
4367 			 */
4368 			return (BUS_PROBE_DEFAULT);
4369 		}
4370 		ent++;
4371 	}
4372 	return (ENXIO);
4373 }
4374 
4375 int
4376 iflib_device_probe_vendor(device_t dev)
4377 {
4378 	int probe;
4379 
4380 	probe = iflib_device_probe(dev);
4381 	if (probe == BUS_PROBE_DEFAULT)
4382 		return (BUS_PROBE_VENDOR);
4383 	else
4384 		return (probe);
4385 }
4386 
4387 static void
4388 iflib_reset_qvalues(if_ctx_t ctx)
4389 {
4390 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4391 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4392 	device_t dev = ctx->ifc_dev;
4393 	int i;
4394 
4395 	if (ctx->ifc_sysctl_ntxqs != 0)
4396 		scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs;
4397 	if (ctx->ifc_sysctl_nrxqs != 0)
4398 		scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs;
4399 
4400 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4401 		if (ctx->ifc_sysctl_ntxds[i] != 0)
4402 			scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i];
4403 		else
4404 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4405 	}
4406 
4407 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4408 		if (ctx->ifc_sysctl_nrxds[i] != 0)
4409 			scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i];
4410 		else
4411 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4412 	}
4413 
4414 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4415 		if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) {
4416 			device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n",
4417 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]);
4418 			scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i];
4419 		}
4420 		if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) {
4421 			device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n",
4422 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]);
4423 			scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i];
4424 		}
4425 		if (!powerof2(scctx->isc_nrxd[i])) {
4426 			device_printf(dev, "nrxd%d: %d is not a power of 2 - using default value of %d\n",
4427 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_default[i]);
4428 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4429 		}
4430 	}
4431 
4432 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4433 		if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) {
4434 			device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n",
4435 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]);
4436 			scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i];
4437 		}
4438 		if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) {
4439 			device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n",
4440 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]);
4441 			scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i];
4442 		}
4443 		if (!powerof2(scctx->isc_ntxd[i])) {
4444 			device_printf(dev, "ntxd%d: %d is not a power of 2 - using default value of %d\n",
4445 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_default[i]);
4446 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4447 		}
4448 	}
4449 }
4450 
4451 static void
4452 iflib_add_pfil(if_ctx_t ctx)
4453 {
4454 	struct pfil_head *pfil;
4455 	struct pfil_head_args pa;
4456 	iflib_rxq_t rxq;
4457 	int i;
4458 
4459 	pa.pa_version = PFIL_VERSION;
4460 	pa.pa_flags = PFIL_IN;
4461 	pa.pa_type = PFIL_TYPE_ETHERNET;
4462 	pa.pa_headname = ctx->ifc_ifp->if_xname;
4463 	pfil = pfil_head_register(&pa);
4464 
4465 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4466 		rxq->pfil = pfil;
4467 	}
4468 }
4469 
4470 static void
4471 iflib_rem_pfil(if_ctx_t ctx)
4472 {
4473 	struct pfil_head *pfil;
4474 	iflib_rxq_t rxq;
4475 	int i;
4476 
4477 	rxq = ctx->ifc_rxqs;
4478 	pfil = rxq->pfil;
4479 	for (i = 0; i < NRXQSETS(ctx); i++, rxq++) {
4480 		rxq->pfil = NULL;
4481 	}
4482 	pfil_head_unregister(pfil);
4483 }
4484 
4485 static uint16_t
4486 get_ctx_core_offset(if_ctx_t ctx)
4487 {
4488 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4489 	struct cpu_offset *op;
4490 	uint16_t qc;
4491 	uint16_t ret = ctx->ifc_sysctl_core_offset;
4492 
4493 	if (ret != CORE_OFFSET_UNSPECIFIED)
4494 		return (ret);
4495 
4496 	if (ctx->ifc_sysctl_separate_txrx)
4497 		qc = scctx->isc_ntxqsets + scctx->isc_nrxqsets;
4498 	else
4499 		qc = max(scctx->isc_ntxqsets, scctx->isc_nrxqsets);
4500 
4501 	mtx_lock(&cpu_offset_mtx);
4502 	SLIST_FOREACH(op, &cpu_offsets, entries) {
4503 		if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) {
4504 			ret = op->offset;
4505 			op->offset += qc;
4506 			MPASS(op->refcount < UINT_MAX);
4507 			op->refcount++;
4508 			break;
4509 		}
4510 	}
4511 	if (ret == CORE_OFFSET_UNSPECIFIED) {
4512 		ret = 0;
4513 		op = malloc(sizeof(struct cpu_offset), M_IFLIB,
4514 		    M_NOWAIT | M_ZERO);
4515 		if (op == NULL) {
4516 			device_printf(ctx->ifc_dev,
4517 			    "allocation for cpu offset failed.\n");
4518 		} else {
4519 			op->offset = qc;
4520 			op->refcount = 1;
4521 			CPU_COPY(&ctx->ifc_cpus, &op->set);
4522 			SLIST_INSERT_HEAD(&cpu_offsets, op, entries);
4523 		}
4524 	}
4525 	mtx_unlock(&cpu_offset_mtx);
4526 
4527 	return (ret);
4528 }
4529 
4530 static void
4531 unref_ctx_core_offset(if_ctx_t ctx)
4532 {
4533 	struct cpu_offset *op, *top;
4534 
4535 	mtx_lock(&cpu_offset_mtx);
4536 	SLIST_FOREACH_SAFE(op, &cpu_offsets, entries, top) {
4537 		if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) {
4538 			MPASS(op->refcount > 0);
4539 			op->refcount--;
4540 			if (op->refcount == 0) {
4541 				SLIST_REMOVE(&cpu_offsets, op, cpu_offset, entries);
4542 				free(op, M_IFLIB);
4543 			}
4544 			break;
4545 		}
4546 	}
4547 	mtx_unlock(&cpu_offset_mtx);
4548 }
4549 
4550 int
4551 iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp)
4552 {
4553 	if_ctx_t ctx;
4554 	if_t ifp;
4555 	if_softc_ctx_t scctx;
4556 	kobjop_desc_t kobj_desc;
4557 	kobj_method_t *kobj_method;
4558 	int err, msix, rid;
4559 	uint16_t main_rxq, main_txq;
4560 
4561 	ctx = malloc(sizeof(* ctx), M_IFLIB, M_WAITOK|M_ZERO);
4562 
4563 	if (sc == NULL) {
4564 		sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4565 		device_set_softc(dev, ctx);
4566 		ctx->ifc_flags |= IFC_SC_ALLOCATED;
4567 	}
4568 
4569 	ctx->ifc_sctx = sctx;
4570 	ctx->ifc_dev = dev;
4571 	ctx->ifc_softc = sc;
4572 
4573 	if ((err = iflib_register(ctx)) != 0) {
4574 		device_printf(dev, "iflib_register failed %d\n", err);
4575 		goto fail_ctx_free;
4576 	}
4577 	iflib_add_device_sysctl_pre(ctx);
4578 
4579 	scctx = &ctx->ifc_softc_ctx;
4580 	ifp = ctx->ifc_ifp;
4581 
4582 	iflib_reset_qvalues(ctx);
4583 	CTX_LOCK(ctx);
4584 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4585 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4586 		goto fail_unlock;
4587 	}
4588 	_iflib_pre_assert(scctx);
4589 	ctx->ifc_txrx = *scctx->isc_txrx;
4590 
4591 	if (sctx->isc_flags & IFLIB_DRIVER_MEDIA)
4592 		ctx->ifc_mediap = scctx->isc_media;
4593 
4594 #ifdef INVARIANTS
4595 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
4596 		MPASS(scctx->isc_tx_csum_flags);
4597 #endif
4598 
4599 	if_setcapabilities(ifp, scctx->isc_capabilities | IFCAP_HWSTATS);
4600 	if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS);
4601 
4602 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4603 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4604 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4605 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4606 
4607 	main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4608 	main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4609 
4610 	/* XXX change for per-queue sizes */
4611 	device_printf(dev, "Using %d TX descriptors and %d RX descriptors\n",
4612 	    scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4613 
4614 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4615 	    MAX_SINGLE_PACKET_FRACTION)
4616 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4617 		    MAX_SINGLE_PACKET_FRACTION);
4618 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4619 	    MAX_SINGLE_PACKET_FRACTION)
4620 		scctx->isc_tx_tso_segments_max = max(1,
4621 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4622 
4623 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4624 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
4625 		/*
4626 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
4627 		 * but some MACs do.
4628 		 */
4629 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
4630 		    IP_MAXPACKET));
4631 		/*
4632 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
4633 		 * into account.  In the worst case, each of these calls will
4634 		 * add another mbuf and, thus, the requirement for another DMA
4635 		 * segment.  So for best performance, it doesn't make sense to
4636 		 * advertize a maximum of TSO segments that typically will
4637 		 * require defragmentation in iflib_encap().
4638 		 */
4639 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
4640 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
4641 	}
4642 	if (scctx->isc_rss_table_size == 0)
4643 		scctx->isc_rss_table_size = 64;
4644 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4645 
4646 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4647 	/* XXX format name */
4648 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx,
4649 	    NULL, NULL, "admin");
4650 
4651 	/* Set up cpu set.  If it fails, use the set of all CPUs. */
4652 	if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) != 0) {
4653 		device_printf(dev, "Unable to fetch CPU list\n");
4654 		CPU_COPY(&all_cpus, &ctx->ifc_cpus);
4655 	}
4656 	MPASS(CPU_COUNT(&ctx->ifc_cpus) > 0);
4657 
4658 	/*
4659 	** Now set up MSI or MSI-X, should return us the number of supported
4660 	** vectors (will be 1 for a legacy interrupt and MSI).
4661 	*/
4662 	if (sctx->isc_flags & IFLIB_SKIP_MSIX) {
4663 		msix = scctx->isc_vectors;
4664 	} else if (scctx->isc_msix_bar != 0)
4665 	       /*
4666 		* The simple fact that isc_msix_bar is not 0 does not mean we
4667 		* we have a good value there that is known to work.
4668 		*/
4669 		msix = iflib_msix_init(ctx);
4670 	else {
4671 		scctx->isc_vectors = 1;
4672 		scctx->isc_ntxqsets = 1;
4673 		scctx->isc_nrxqsets = 1;
4674 		scctx->isc_intr = IFLIB_INTR_LEGACY;
4675 		msix = 0;
4676 	}
4677 	/* Get memory for the station queues */
4678 	if ((err = iflib_queues_alloc(ctx))) {
4679 		device_printf(dev, "Unable to allocate queue memory\n");
4680 		goto fail_intr_free;
4681 	}
4682 
4683 	if ((err = iflib_qset_structures_setup(ctx)))
4684 		goto fail_queues;
4685 
4686 	/*
4687 	 * Now that we know how many queues there are, get the core offset.
4688 	 */
4689 	ctx->ifc_sysctl_core_offset = get_ctx_core_offset(ctx);
4690 
4691 	/*
4692 	 * Group taskqueues aren't properly set up until SMP is started,
4693 	 * so we disable interrupts until we can handle them post
4694 	 * SI_SUB_SMP.
4695 	 *
4696 	 * XXX: disabling interrupts doesn't actually work, at least for
4697 	 * the non-MSI case.  When they occur before SI_SUB_SMP completes,
4698 	 * we do null handling and depend on this not causing too large an
4699 	 * interrupt storm.
4700 	 */
4701 	IFDI_INTR_DISABLE(ctx);
4702 
4703 	if (msix > 1) {
4704 		/*
4705 		 * When using MSI-X, ensure that ifdi_{r,t}x_queue_intr_enable
4706 		 * aren't the default NULL implementation.
4707 		 */
4708 		kobj_desc = &ifdi_rx_queue_intr_enable_desc;
4709 		kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL,
4710 		    kobj_desc);
4711 		if (kobj_method == &kobj_desc->deflt) {
4712 			device_printf(dev,
4713 			    "MSI-X requires ifdi_rx_queue_intr_enable method");
4714 			err = EOPNOTSUPP;
4715 			goto fail_queues;
4716 		}
4717 		kobj_desc = &ifdi_tx_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_tx_queue_intr_enable method");
4723 			err = EOPNOTSUPP;
4724 			goto fail_queues;
4725 		}
4726 
4727 		/*
4728 		 * Assign the MSI-X vectors.
4729 		 * Note that the default NULL ifdi_msix_intr_assign method will
4730 		 * fail here, too.
4731 		 */
4732 		err = IFDI_MSIX_INTR_ASSIGN(ctx, msix);
4733 		if (err != 0) {
4734 			device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n",
4735 			    err);
4736 			goto fail_queues;
4737 		}
4738 	} else {
4739 		rid = 0;
4740 		if (scctx->isc_intr == IFLIB_INTR_MSI) {
4741 			MPASS(msix == 1);
4742 			rid = 1;
4743 		}
4744 		if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) {
4745 			device_printf(dev, "iflib_legacy_setup failed %d\n", err);
4746 			goto fail_queues;
4747 		}
4748 	}
4749 
4750 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
4751 
4752 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4753 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4754 		goto fail_detach;
4755 	}
4756 
4757 	/*
4758 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4759 	 * This must appear after the call to ether_ifattach() because
4760 	 * ether_ifattach() sets if_hdrlen to the default value.
4761 	 */
4762 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4763 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
4764 
4765 	if ((err = iflib_netmap_attach(ctx))) {
4766 		device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err);
4767 		goto fail_detach;
4768 	}
4769 	*ctxp = ctx;
4770 
4771 	NETDUMP_SET(ctx->ifc_ifp, iflib);
4772 
4773 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4774 	iflib_add_device_sysctl_post(ctx);
4775 	iflib_add_pfil(ctx);
4776 	ctx->ifc_flags |= IFC_INIT_DONE;
4777 	CTX_UNLOCK(ctx);
4778 
4779 	return (0);
4780 
4781 fail_detach:
4782 	ether_ifdetach(ctx->ifc_ifp);
4783 fail_intr_free:
4784 	iflib_free_intr_mem(ctx);
4785 fail_queues:
4786 	iflib_tx_structures_free(ctx);
4787 	iflib_rx_structures_free(ctx);
4788 	IFDI_DETACH(ctx);
4789 fail_unlock:
4790 	CTX_UNLOCK(ctx);
4791 fail_ctx_free:
4792         if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4793                 free(ctx->ifc_softc, M_IFLIB);
4794         free(ctx, M_IFLIB);
4795 	return (err);
4796 }
4797 
4798 int
4799 iflib_pseudo_register(device_t dev, if_shared_ctx_t sctx, if_ctx_t *ctxp,
4800 					  struct iflib_cloneattach_ctx *clctx)
4801 {
4802 	int err;
4803 	if_ctx_t ctx;
4804 	if_t ifp;
4805 	if_softc_ctx_t scctx;
4806 	int i;
4807 	void *sc;
4808 	uint16_t main_txq;
4809 	uint16_t main_rxq;
4810 
4811 	ctx = malloc(sizeof(*ctx), M_IFLIB, M_WAITOK|M_ZERO);
4812 	sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4813 	ctx->ifc_flags |= IFC_SC_ALLOCATED;
4814 	if (sctx->isc_flags & (IFLIB_PSEUDO|IFLIB_VIRTUAL))
4815 		ctx->ifc_flags |= IFC_PSEUDO;
4816 
4817 	ctx->ifc_sctx = sctx;
4818 	ctx->ifc_softc = sc;
4819 	ctx->ifc_dev = dev;
4820 
4821 	if ((err = iflib_register(ctx)) != 0) {
4822 		device_printf(dev, "%s: iflib_register failed %d\n", __func__, err);
4823 		goto fail_ctx_free;
4824 	}
4825 	iflib_add_device_sysctl_pre(ctx);
4826 
4827 	scctx = &ctx->ifc_softc_ctx;
4828 	ifp = ctx->ifc_ifp;
4829 
4830 	iflib_reset_qvalues(ctx);
4831 	CTX_LOCK(ctx);
4832 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4833 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4834 		goto fail_unlock;
4835 	}
4836 	if (sctx->isc_flags & IFLIB_GEN_MAC)
4837 		ether_gen_addr(ifp, &ctx->ifc_mac);
4838 	if ((err = IFDI_CLONEATTACH(ctx, clctx->cc_ifc, clctx->cc_name,
4839 								clctx->cc_params)) != 0) {
4840 		device_printf(dev, "IFDI_CLONEATTACH failed %d\n", err);
4841 		goto fail_ctx_free;
4842 	}
4843 	ifmedia_add(ctx->ifc_mediap, IFM_ETHER | IFM_1000_T | IFM_FDX, 0, NULL);
4844 	ifmedia_add(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO, 0, NULL);
4845 	ifmedia_set(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO);
4846 
4847 #ifdef INVARIANTS
4848 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
4849 		MPASS(scctx->isc_tx_csum_flags);
4850 #endif
4851 
4852 	if_setcapabilities(ifp, scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_LINKSTATE);
4853 	if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_LINKSTATE);
4854 
4855 	ifp->if_flags |= IFF_NOGROUP;
4856 	if (sctx->isc_flags & IFLIB_PSEUDO) {
4857 		ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
4858 
4859 		if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4860 			device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4861 			goto fail_detach;
4862 		}
4863 		*ctxp = ctx;
4864 
4865 		/*
4866 		 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4867 		 * This must appear after the call to ether_ifattach() because
4868 		 * ether_ifattach() sets if_hdrlen to the default value.
4869 		 */
4870 		if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4871 			if_setifheaderlen(ifp,
4872 			    sizeof(struct ether_vlan_header));
4873 
4874 		if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4875 		iflib_add_device_sysctl_post(ctx);
4876 		ctx->ifc_flags |= IFC_INIT_DONE;
4877 		return (0);
4878 	}
4879 	_iflib_pre_assert(scctx);
4880 	ctx->ifc_txrx = *scctx->isc_txrx;
4881 
4882 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4883 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4884 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4885 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4886 
4887 	main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4888 	main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4889 
4890 	/* XXX change for per-queue sizes */
4891 	device_printf(dev, "Using %d TX descriptors and %d RX descriptors\n",
4892 	    scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4893 
4894 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4895 	    MAX_SINGLE_PACKET_FRACTION)
4896 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4897 		    MAX_SINGLE_PACKET_FRACTION);
4898 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4899 	    MAX_SINGLE_PACKET_FRACTION)
4900 		scctx->isc_tx_tso_segments_max = max(1,
4901 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4902 
4903 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4904 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
4905 		/*
4906 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
4907 		 * but some MACs do.
4908 		 */
4909 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
4910 		    IP_MAXPACKET));
4911 		/*
4912 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
4913 		 * into account.  In the worst case, each of these calls will
4914 		 * add another mbuf and, thus, the requirement for another DMA
4915 		 * segment.  So for best performance, it doesn't make sense to
4916 		 * advertize a maximum of TSO segments that typically will
4917 		 * require defragmentation in iflib_encap().
4918 		 */
4919 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
4920 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
4921 	}
4922 	if (scctx->isc_rss_table_size == 0)
4923 		scctx->isc_rss_table_size = 64;
4924 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4925 
4926 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4927 	/* XXX format name */
4928 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx,
4929 	    NULL, NULL, "admin");
4930 
4931 	/* XXX --- can support > 1 -- but keep it simple for now */
4932 	scctx->isc_intr = IFLIB_INTR_LEGACY;
4933 
4934 	/* Get memory for the station queues */
4935 	if ((err = iflib_queues_alloc(ctx))) {
4936 		device_printf(dev, "Unable to allocate queue memory\n");
4937 		goto fail_iflib_detach;
4938 	}
4939 
4940 	if ((err = iflib_qset_structures_setup(ctx))) {
4941 		device_printf(dev, "qset structure setup failed %d\n", err);
4942 		goto fail_queues;
4943 	}
4944 
4945 	/*
4946 	 * XXX What if anything do we want to do about interrupts?
4947 	 */
4948 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
4949 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4950 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4951 		goto fail_detach;
4952 	}
4953 
4954 	/*
4955 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4956 	 * This must appear after the call to ether_ifattach() because
4957 	 * ether_ifattach() sets if_hdrlen to the default value.
4958 	 */
4959 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4960 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
4961 
4962 	/* XXX handle more than one queue */
4963 	for (i = 0; i < scctx->isc_nrxqsets; i++)
4964 		IFDI_RX_CLSET(ctx, 0, i, ctx->ifc_rxqs[i].ifr_fl[0].ifl_sds.ifsd_cl);
4965 
4966 	*ctxp = ctx;
4967 
4968 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4969 	iflib_add_device_sysctl_post(ctx);
4970 	ctx->ifc_flags |= IFC_INIT_DONE;
4971 	CTX_UNLOCK(ctx);
4972 
4973 	return (0);
4974 fail_detach:
4975 	ether_ifdetach(ctx->ifc_ifp);
4976 fail_queues:
4977 	iflib_tx_structures_free(ctx);
4978 	iflib_rx_structures_free(ctx);
4979 fail_iflib_detach:
4980 	IFDI_DETACH(ctx);
4981 fail_unlock:
4982 	CTX_UNLOCK(ctx);
4983 fail_ctx_free:
4984 	free(ctx->ifc_softc, M_IFLIB);
4985 	free(ctx, M_IFLIB);
4986 	return (err);
4987 }
4988 
4989 int
4990 iflib_pseudo_deregister(if_ctx_t ctx)
4991 {
4992 	if_t ifp = ctx->ifc_ifp;
4993 	iflib_txq_t txq;
4994 	iflib_rxq_t rxq;
4995 	int i, j;
4996 	struct taskqgroup *tqg;
4997 	iflib_fl_t fl;
4998 
4999 	/* Unregister VLAN events */
5000 	if (ctx->ifc_vlan_attach_event != NULL)
5001 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
5002 	if (ctx->ifc_vlan_detach_event != NULL)
5003 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
5004 
5005 	ether_ifdetach(ifp);
5006 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
5007 	CTX_LOCK_DESTROY(ctx);
5008 	/* XXX drain any dependent tasks */
5009 	tqg = qgroup_if_io_tqg;
5010 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
5011 		callout_drain(&txq->ift_timer);
5012 		if (txq->ift_task.gt_uniq != NULL)
5013 			taskqgroup_detach(tqg, &txq->ift_task);
5014 	}
5015 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
5016 		if (rxq->ifr_task.gt_uniq != NULL)
5017 			taskqgroup_detach(tqg, &rxq->ifr_task);
5018 
5019 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5020 			free(fl->ifl_rx_bitmap, M_IFLIB);
5021 	}
5022 	tqg = qgroup_if_config_tqg;
5023 	if (ctx->ifc_admin_task.gt_uniq != NULL)
5024 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
5025 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
5026 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
5027 
5028 	if_free(ifp);
5029 
5030 	iflib_tx_structures_free(ctx);
5031 	iflib_rx_structures_free(ctx);
5032 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5033 		free(ctx->ifc_softc, M_IFLIB);
5034 	free(ctx, M_IFLIB);
5035 	return (0);
5036 }
5037 
5038 int
5039 iflib_device_attach(device_t dev)
5040 {
5041 	if_ctx_t ctx;
5042 	if_shared_ctx_t sctx;
5043 
5044 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
5045 		return (ENOTSUP);
5046 
5047 	pci_enable_busmaster(dev);
5048 
5049 	return (iflib_device_register(dev, NULL, sctx, &ctx));
5050 }
5051 
5052 int
5053 iflib_device_deregister(if_ctx_t ctx)
5054 {
5055 	if_t ifp = ctx->ifc_ifp;
5056 	iflib_txq_t txq;
5057 	iflib_rxq_t rxq;
5058 	device_t dev = ctx->ifc_dev;
5059 	int i, j;
5060 	struct taskqgroup *tqg;
5061 	iflib_fl_t fl;
5062 
5063 	/* Make sure VLANS are not using driver */
5064 	if (if_vlantrunkinuse(ifp)) {
5065 		device_printf(dev, "Vlan in use, detach first\n");
5066 		return (EBUSY);
5067 	}
5068 #ifdef PCI_IOV
5069 	if (!CTX_IS_VF(ctx) && pci_iov_detach(dev) != 0) {
5070 		device_printf(dev, "SR-IOV in use; detach first.\n");
5071 		return (EBUSY);
5072 	}
5073 #endif
5074 
5075 	STATE_LOCK(ctx);
5076 	ctx->ifc_flags |= IFC_IN_DETACH;
5077 	STATE_UNLOCK(ctx);
5078 
5079 	CTX_LOCK(ctx);
5080 	iflib_stop(ctx);
5081 	CTX_UNLOCK(ctx);
5082 
5083 	/* Unregister VLAN events */
5084 	if (ctx->ifc_vlan_attach_event != NULL)
5085 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
5086 	if (ctx->ifc_vlan_detach_event != NULL)
5087 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
5088 
5089 	iflib_netmap_detach(ifp);
5090 	ether_ifdetach(ifp);
5091 	iflib_rem_pfil(ctx);
5092 	if (ctx->ifc_led_dev != NULL)
5093 		led_destroy(ctx->ifc_led_dev);
5094 	/* XXX drain any dependent tasks */
5095 	tqg = qgroup_if_io_tqg;
5096 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
5097 		callout_drain(&txq->ift_timer);
5098 		if (txq->ift_task.gt_uniq != NULL)
5099 			taskqgroup_detach(tqg, &txq->ift_task);
5100 	}
5101 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
5102 		if (rxq->ifr_task.gt_uniq != NULL)
5103 			taskqgroup_detach(tqg, &rxq->ifr_task);
5104 
5105 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5106 			free(fl->ifl_rx_bitmap, M_IFLIB);
5107 	}
5108 	tqg = qgroup_if_config_tqg;
5109 	if (ctx->ifc_admin_task.gt_uniq != NULL)
5110 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
5111 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
5112 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
5113 	CTX_LOCK(ctx);
5114 	IFDI_DETACH(ctx);
5115 	CTX_UNLOCK(ctx);
5116 
5117 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
5118 	CTX_LOCK_DESTROY(ctx);
5119 	device_set_softc(ctx->ifc_dev, NULL);
5120 	iflib_free_intr_mem(ctx);
5121 
5122 	bus_generic_detach(dev);
5123 	if_free(ifp);
5124 
5125 	iflib_tx_structures_free(ctx);
5126 	iflib_rx_structures_free(ctx);
5127 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5128 		free(ctx->ifc_softc, M_IFLIB);
5129 	unref_ctx_core_offset(ctx);
5130 	STATE_LOCK_DESTROY(ctx);
5131 	free(ctx, M_IFLIB);
5132 	return (0);
5133 }
5134 
5135 static void
5136 iflib_free_intr_mem(if_ctx_t ctx)
5137 {
5138 
5139 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) {
5140 		iflib_irq_free(ctx, &ctx->ifc_legacy_irq);
5141 	}
5142 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) {
5143 		pci_release_msi(ctx->ifc_dev);
5144 	}
5145 	if (ctx->ifc_msix_mem != NULL) {
5146 		bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY,
5147 		    rman_get_rid(ctx->ifc_msix_mem), ctx->ifc_msix_mem);
5148 		ctx->ifc_msix_mem = NULL;
5149 	}
5150 }
5151 
5152 int
5153 iflib_device_detach(device_t dev)
5154 {
5155 	if_ctx_t ctx = device_get_softc(dev);
5156 
5157 	return (iflib_device_deregister(ctx));
5158 }
5159 
5160 int
5161 iflib_device_suspend(device_t dev)
5162 {
5163 	if_ctx_t ctx = device_get_softc(dev);
5164 
5165 	CTX_LOCK(ctx);
5166 	IFDI_SUSPEND(ctx);
5167 	CTX_UNLOCK(ctx);
5168 
5169 	return bus_generic_suspend(dev);
5170 }
5171 int
5172 iflib_device_shutdown(device_t dev)
5173 {
5174 	if_ctx_t ctx = device_get_softc(dev);
5175 
5176 	CTX_LOCK(ctx);
5177 	IFDI_SHUTDOWN(ctx);
5178 	CTX_UNLOCK(ctx);
5179 
5180 	return bus_generic_suspend(dev);
5181 }
5182 
5183 
5184 int
5185 iflib_device_resume(device_t dev)
5186 {
5187 	if_ctx_t ctx = device_get_softc(dev);
5188 	iflib_txq_t txq = ctx->ifc_txqs;
5189 
5190 	CTX_LOCK(ctx);
5191 	IFDI_RESUME(ctx);
5192 	iflib_if_init_locked(ctx);
5193 	CTX_UNLOCK(ctx);
5194 	for (int i = 0; i < NTXQSETS(ctx); i++, txq++)
5195 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
5196 
5197 	return (bus_generic_resume(dev));
5198 }
5199 
5200 int
5201 iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params)
5202 {
5203 	int error;
5204 	if_ctx_t ctx = device_get_softc(dev);
5205 
5206 	CTX_LOCK(ctx);
5207 	error = IFDI_IOV_INIT(ctx, num_vfs, params);
5208 	CTX_UNLOCK(ctx);
5209 
5210 	return (error);
5211 }
5212 
5213 void
5214 iflib_device_iov_uninit(device_t dev)
5215 {
5216 	if_ctx_t ctx = device_get_softc(dev);
5217 
5218 	CTX_LOCK(ctx);
5219 	IFDI_IOV_UNINIT(ctx);
5220 	CTX_UNLOCK(ctx);
5221 }
5222 
5223 int
5224 iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params)
5225 {
5226 	int error;
5227 	if_ctx_t ctx = device_get_softc(dev);
5228 
5229 	CTX_LOCK(ctx);
5230 	error = IFDI_IOV_VF_ADD(ctx, vfnum, params);
5231 	CTX_UNLOCK(ctx);
5232 
5233 	return (error);
5234 }
5235 
5236 /*********************************************************************
5237  *
5238  *  MODULE FUNCTION DEFINITIONS
5239  *
5240  **********************************************************************/
5241 
5242 /*
5243  * - Start a fast taskqueue thread for each core
5244  * - Start a taskqueue for control operations
5245  */
5246 static int
5247 iflib_module_init(void)
5248 {
5249 	return (0);
5250 }
5251 
5252 static int
5253 iflib_module_event_handler(module_t mod, int what, void *arg)
5254 {
5255 	int err;
5256 
5257 	switch (what) {
5258 	case MOD_LOAD:
5259 		if ((err = iflib_module_init()) != 0)
5260 			return (err);
5261 		break;
5262 	case MOD_UNLOAD:
5263 		return (EBUSY);
5264 	default:
5265 		return (EOPNOTSUPP);
5266 	}
5267 
5268 	return (0);
5269 }
5270 
5271 /*********************************************************************
5272  *
5273  *  PUBLIC FUNCTION DEFINITIONS
5274  *     ordered as in iflib.h
5275  *
5276  **********************************************************************/
5277 
5278 
5279 static void
5280 _iflib_assert(if_shared_ctx_t sctx)
5281 {
5282 	int i;
5283 
5284 	MPASS(sctx->isc_tx_maxsize);
5285 	MPASS(sctx->isc_tx_maxsegsize);
5286 
5287 	MPASS(sctx->isc_rx_maxsize);
5288 	MPASS(sctx->isc_rx_nsegments);
5289 	MPASS(sctx->isc_rx_maxsegsize);
5290 
5291 	MPASS(sctx->isc_nrxqs >= 1 && sctx->isc_nrxqs <= 8);
5292 	for (i = 0; i < sctx->isc_nrxqs; i++) {
5293 		MPASS(sctx->isc_nrxd_min[i]);
5294 		MPASS(powerof2(sctx->isc_nrxd_min[i]));
5295 		MPASS(sctx->isc_nrxd_max[i]);
5296 		MPASS(powerof2(sctx->isc_nrxd_max[i]));
5297 		MPASS(sctx->isc_nrxd_default[i]);
5298 		MPASS(powerof2(sctx->isc_nrxd_default[i]));
5299 	}
5300 
5301 	MPASS(sctx->isc_ntxqs >= 1 && sctx->isc_ntxqs <= 8);
5302 	for (i = 0; i < sctx->isc_ntxqs; i++) {
5303 		MPASS(sctx->isc_ntxd_min[i]);
5304 		MPASS(powerof2(sctx->isc_ntxd_min[i]));
5305 		MPASS(sctx->isc_ntxd_max[i]);
5306 		MPASS(powerof2(sctx->isc_ntxd_max[i]));
5307 		MPASS(sctx->isc_ntxd_default[i]);
5308 		MPASS(powerof2(sctx->isc_ntxd_default[i]));
5309 	}
5310 }
5311 
5312 static void
5313 _iflib_pre_assert(if_softc_ctx_t scctx)
5314 {
5315 
5316 	MPASS(scctx->isc_txrx->ift_txd_encap);
5317 	MPASS(scctx->isc_txrx->ift_txd_flush);
5318 	MPASS(scctx->isc_txrx->ift_txd_credits_update);
5319 	MPASS(scctx->isc_txrx->ift_rxd_available);
5320 	MPASS(scctx->isc_txrx->ift_rxd_pkt_get);
5321 	MPASS(scctx->isc_txrx->ift_rxd_refill);
5322 	MPASS(scctx->isc_txrx->ift_rxd_flush);
5323 }
5324 
5325 static int
5326 iflib_register(if_ctx_t ctx)
5327 {
5328 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5329 	driver_t *driver = sctx->isc_driver;
5330 	device_t dev = ctx->ifc_dev;
5331 	if_t ifp;
5332 
5333 	_iflib_assert(sctx);
5334 
5335 	CTX_LOCK_INIT(ctx);
5336 	STATE_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev));
5337 	ifp = ctx->ifc_ifp = if_alloc(IFT_ETHER);
5338 	if (ifp == NULL) {
5339 		device_printf(dev, "can not allocate ifnet structure\n");
5340 		return (ENOMEM);
5341 	}
5342 
5343 	/*
5344 	 * Initialize our context's device specific methods
5345 	 */
5346 	kobj_init((kobj_t) ctx, (kobj_class_t) driver);
5347 	kobj_class_compile((kobj_class_t) driver);
5348 	driver->refs++;
5349 
5350 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
5351 	if_setsoftc(ifp, ctx);
5352 	if_setdev(ifp, dev);
5353 	if_setinitfn(ifp, iflib_if_init);
5354 	if_setioctlfn(ifp, iflib_if_ioctl);
5355 #ifdef ALTQ
5356 	if_setstartfn(ifp, iflib_altq_if_start);
5357 	if_settransmitfn(ifp, iflib_altq_if_transmit);
5358 	if_setsendqready(ifp);
5359 #else
5360 	if_settransmitfn(ifp, iflib_if_transmit);
5361 #endif
5362 	if_setqflushfn(ifp, iflib_if_qflush);
5363 	if_setflags(ifp, IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST);
5364 
5365 	ctx->ifc_vlan_attach_event =
5366 		EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx,
5367 							  EVENTHANDLER_PRI_FIRST);
5368 	ctx->ifc_vlan_detach_event =
5369 		EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx,
5370 							  EVENTHANDLER_PRI_FIRST);
5371 
5372 	if ((sctx->isc_flags & IFLIB_DRIVER_MEDIA) == 0) {
5373 		ctx->ifc_mediap = &ctx->ifc_media;
5374 		ifmedia_init(ctx->ifc_mediap, IFM_IMASK,
5375 		    iflib_media_change, iflib_media_status);
5376 	}
5377 	return (0);
5378 }
5379 
5380 static int
5381 iflib_queues_alloc(if_ctx_t ctx)
5382 {
5383 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5384 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5385 	device_t dev = ctx->ifc_dev;
5386 	int nrxqsets = scctx->isc_nrxqsets;
5387 	int ntxqsets = scctx->isc_ntxqsets;
5388 	iflib_txq_t txq;
5389 	iflib_rxq_t rxq;
5390 	iflib_fl_t fl = NULL;
5391 	int i, j, cpu, err, txconf, rxconf;
5392 	iflib_dma_info_t ifdip;
5393 	uint32_t *rxqsizes = scctx->isc_rxqsizes;
5394 	uint32_t *txqsizes = scctx->isc_txqsizes;
5395 	uint8_t nrxqs = sctx->isc_nrxqs;
5396 	uint8_t ntxqs = sctx->isc_ntxqs;
5397 	int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1;
5398 	caddr_t *vaddrs;
5399 	uint64_t *paddrs;
5400 
5401 	KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1"));
5402 	KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1"));
5403 
5404 	/* Allocate the TX ring struct memory */
5405 	if (!(ctx->ifc_txqs =
5406 	    (iflib_txq_t) malloc(sizeof(struct iflib_txq) *
5407 	    ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5408 		device_printf(dev, "Unable to allocate TX ring memory\n");
5409 		err = ENOMEM;
5410 		goto fail;
5411 	}
5412 
5413 	/* Now allocate the RX */
5414 	if (!(ctx->ifc_rxqs =
5415 	    (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) *
5416 	    nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5417 		device_printf(dev, "Unable to allocate RX ring memory\n");
5418 		err = ENOMEM;
5419 		goto rx_fail;
5420 	}
5421 
5422 	txq = ctx->ifc_txqs;
5423 	rxq = ctx->ifc_rxqs;
5424 
5425 	/*
5426 	 * XXX handle allocation failure
5427 	 */
5428 	for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) {
5429 		/* Set up some basics */
5430 
5431 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs,
5432 		    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
5433 			device_printf(dev,
5434 			    "Unable to allocate TX DMA info memory\n");
5435 			err = ENOMEM;
5436 			goto err_tx_desc;
5437 		}
5438 		txq->ift_ifdi = ifdip;
5439 		for (j = 0; j < ntxqs; j++, ifdip++) {
5440 			if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, 0)) {
5441 				device_printf(dev,
5442 				    "Unable to allocate TX descriptors\n");
5443 				err = ENOMEM;
5444 				goto err_tx_desc;
5445 			}
5446 			txq->ift_txd_size[j] = scctx->isc_txd_size[j];
5447 			bzero((void *)ifdip->idi_vaddr, txqsizes[j]);
5448 		}
5449 		txq->ift_ctx = ctx;
5450 		txq->ift_id = i;
5451 		if (sctx->isc_flags & IFLIB_HAS_TXCQ) {
5452 			txq->ift_br_offset = 1;
5453 		} else {
5454 			txq->ift_br_offset = 0;
5455 		}
5456 		/* XXX fix this */
5457 		txq->ift_timer.c_cpu = cpu;
5458 
5459 		if (iflib_txsd_alloc(txq)) {
5460 			device_printf(dev, "Critical Failure setting up TX buffers\n");
5461 			err = ENOMEM;
5462 			goto err_tx_desc;
5463 		}
5464 
5465 		/* Initialize the TX lock */
5466 		snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:TX(%d):callout",
5467 		    device_get_nameunit(dev), txq->ift_id);
5468 		mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF);
5469 		callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0);
5470 
5471 		err = ifmp_ring_alloc(&txq->ift_br, 2048, txq, iflib_txq_drain,
5472 				      iflib_txq_can_drain, M_IFLIB, M_WAITOK);
5473 		if (err) {
5474 			/* XXX free any allocated rings */
5475 			device_printf(dev, "Unable to allocate buf_ring\n");
5476 			goto err_tx_desc;
5477 		}
5478 	}
5479 
5480 	for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) {
5481 		/* Set up some basics */
5482 
5483 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs,
5484 		   M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
5485 			device_printf(dev,
5486 			    "Unable to allocate RX DMA info memory\n");
5487 			err = ENOMEM;
5488 			goto err_tx_desc;
5489 		}
5490 
5491 		rxq->ifr_ifdi = ifdip;
5492 		/* XXX this needs to be changed if #rx queues != #tx queues */
5493 		rxq->ifr_ntxqirq = 1;
5494 		rxq->ifr_txqid[0] = i;
5495 		for (j = 0; j < nrxqs; j++, ifdip++) {
5496 			if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, 0)) {
5497 				device_printf(dev,
5498 				    "Unable to allocate RX descriptors\n");
5499 				err = ENOMEM;
5500 				goto err_tx_desc;
5501 			}
5502 			bzero((void *)ifdip->idi_vaddr, rxqsizes[j]);
5503 		}
5504 		rxq->ifr_ctx = ctx;
5505 		rxq->ifr_id = i;
5506 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
5507 			rxq->ifr_fl_offset = 1;
5508 		} else {
5509 			rxq->ifr_fl_offset = 0;
5510 		}
5511 		rxq->ifr_nfl = nfree_lists;
5512 		if (!(fl =
5513 			  (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) {
5514 			device_printf(dev, "Unable to allocate free list memory\n");
5515 			err = ENOMEM;
5516 			goto err_tx_desc;
5517 		}
5518 		rxq->ifr_fl = fl;
5519 		for (j = 0; j < nfree_lists; j++) {
5520 			fl[j].ifl_rxq = rxq;
5521 			fl[j].ifl_id = j;
5522 			fl[j].ifl_ifdi = &rxq->ifr_ifdi[j + rxq->ifr_fl_offset];
5523 			fl[j].ifl_rxd_size = scctx->isc_rxd_size[j];
5524 		}
5525 		/* Allocate receive buffers for the ring */
5526 		if (iflib_rxsd_alloc(rxq)) {
5527 			device_printf(dev,
5528 			    "Critical Failure setting up receive buffers\n");
5529 			err = ENOMEM;
5530 			goto err_rx_desc;
5531 		}
5532 
5533 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5534 			fl->ifl_rx_bitmap = bit_alloc(fl->ifl_size, M_IFLIB,
5535 			    M_WAITOK);
5536 	}
5537 
5538 	/* TXQs */
5539 	vaddrs = malloc(sizeof(caddr_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5540 	paddrs = malloc(sizeof(uint64_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5541 	for (i = 0; i < ntxqsets; i++) {
5542 		iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi;
5543 
5544 		for (j = 0; j < ntxqs; j++, di++) {
5545 			vaddrs[i*ntxqs + j] = di->idi_vaddr;
5546 			paddrs[i*ntxqs + j] = di->idi_paddr;
5547 		}
5548 	}
5549 	if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) {
5550 		device_printf(ctx->ifc_dev,
5551 		    "Unable to allocate device TX queue\n");
5552 		iflib_tx_structures_free(ctx);
5553 		free(vaddrs, M_IFLIB);
5554 		free(paddrs, M_IFLIB);
5555 		goto err_rx_desc;
5556 	}
5557 	free(vaddrs, M_IFLIB);
5558 	free(paddrs, M_IFLIB);
5559 
5560 	/* RXQs */
5561 	vaddrs = malloc(sizeof(caddr_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5562 	paddrs = malloc(sizeof(uint64_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5563 	for (i = 0; i < nrxqsets; i++) {
5564 		iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi;
5565 
5566 		for (j = 0; j < nrxqs; j++, di++) {
5567 			vaddrs[i*nrxqs + j] = di->idi_vaddr;
5568 			paddrs[i*nrxqs + j] = di->idi_paddr;
5569 		}
5570 	}
5571 	if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) {
5572 		device_printf(ctx->ifc_dev,
5573 		    "Unable to allocate device RX queue\n");
5574 		iflib_tx_structures_free(ctx);
5575 		free(vaddrs, M_IFLIB);
5576 		free(paddrs, M_IFLIB);
5577 		goto err_rx_desc;
5578 	}
5579 	free(vaddrs, M_IFLIB);
5580 	free(paddrs, M_IFLIB);
5581 
5582 	return (0);
5583 
5584 /* XXX handle allocation failure changes */
5585 err_rx_desc:
5586 err_tx_desc:
5587 rx_fail:
5588 	if (ctx->ifc_rxqs != NULL)
5589 		free(ctx->ifc_rxqs, M_IFLIB);
5590 	ctx->ifc_rxqs = NULL;
5591 	if (ctx->ifc_txqs != NULL)
5592 		free(ctx->ifc_txqs, M_IFLIB);
5593 	ctx->ifc_txqs = NULL;
5594 fail:
5595 	return (err);
5596 }
5597 
5598 static int
5599 iflib_tx_structures_setup(if_ctx_t ctx)
5600 {
5601 	iflib_txq_t txq = ctx->ifc_txqs;
5602 	int i;
5603 
5604 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
5605 		iflib_txq_setup(txq);
5606 
5607 	return (0);
5608 }
5609 
5610 static void
5611 iflib_tx_structures_free(if_ctx_t ctx)
5612 {
5613 	iflib_txq_t txq = ctx->ifc_txqs;
5614 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5615 	int i, j;
5616 
5617 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
5618 		iflib_txq_destroy(txq);
5619 		for (j = 0; j < sctx->isc_ntxqs; j++)
5620 			iflib_dma_free(&txq->ift_ifdi[j]);
5621 	}
5622 	free(ctx->ifc_txqs, M_IFLIB);
5623 	ctx->ifc_txqs = NULL;
5624 	IFDI_QUEUES_FREE(ctx);
5625 }
5626 
5627 /*********************************************************************
5628  *
5629  *  Initialize all receive rings.
5630  *
5631  **********************************************************************/
5632 static int
5633 iflib_rx_structures_setup(if_ctx_t ctx)
5634 {
5635 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5636 	int q;
5637 #if defined(INET6) || defined(INET)
5638 	int err, i;
5639 #endif
5640 
5641 	for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) {
5642 #if defined(INET6) || defined(INET)
5643 		if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_LRO) {
5644 			err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp,
5645 			    TCP_LRO_ENTRIES, min(1024,
5646 			    ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset]));
5647 			if (err != 0) {
5648 				device_printf(ctx->ifc_dev,
5649 				    "LRO Initialization failed!\n");
5650 				goto fail;
5651 			}
5652 		}
5653 #endif
5654 		IFDI_RXQ_SETUP(ctx, rxq->ifr_id);
5655 	}
5656 	return (0);
5657 #if defined(INET6) || defined(INET)
5658 fail:
5659 	/*
5660 	 * Free LRO resources allocated so far, we will only handle
5661 	 * the rings that completed, the failing case will have
5662 	 * cleaned up for itself.  'q' failed, so its the terminus.
5663 	 */
5664 	rxq = ctx->ifc_rxqs;
5665 	for (i = 0; i < q; ++i, rxq++) {
5666 		if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_LRO)
5667 			tcp_lro_free(&rxq->ifr_lc);
5668 	}
5669 	return (err);
5670 #endif
5671 }
5672 
5673 /*********************************************************************
5674  *
5675  *  Free all receive rings.
5676  *
5677  **********************************************************************/
5678 static void
5679 iflib_rx_structures_free(if_ctx_t ctx)
5680 {
5681 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5682 	int i;
5683 
5684 	for (i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) {
5685 		iflib_rx_sds_free(rxq);
5686 #if defined(INET6) || defined(INET)
5687 		if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_LRO)
5688 			tcp_lro_free(&rxq->ifr_lc);
5689 #endif
5690 	}
5691 	free(ctx->ifc_rxqs, M_IFLIB);
5692 	ctx->ifc_rxqs = NULL;
5693 }
5694 
5695 static int
5696 iflib_qset_structures_setup(if_ctx_t ctx)
5697 {
5698 	int err;
5699 
5700 	/*
5701 	 * It is expected that the caller takes care of freeing queues if this
5702 	 * fails.
5703 	 */
5704 	if ((err = iflib_tx_structures_setup(ctx)) != 0) {
5705 		device_printf(ctx->ifc_dev, "iflib_tx_structures_setup failed: %d\n", err);
5706 		return (err);
5707 	}
5708 
5709 	if ((err = iflib_rx_structures_setup(ctx)) != 0)
5710 		device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err);
5711 
5712 	return (err);
5713 }
5714 
5715 int
5716 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
5717 		driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, const char *name)
5718 {
5719 
5720 	return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name));
5721 }
5722 
5723 #ifdef SMP
5724 static int
5725 find_nth(if_ctx_t ctx, int qid)
5726 {
5727 	cpuset_t cpus;
5728 	int i, cpuid, eqid, count;
5729 
5730 	CPU_COPY(&ctx->ifc_cpus, &cpus);
5731 	count = CPU_COUNT(&cpus);
5732 	eqid = qid % count;
5733 	/* clear up to the qid'th bit */
5734 	for (i = 0; i < eqid; i++) {
5735 		cpuid = CPU_FFS(&cpus);
5736 		MPASS(cpuid != 0);
5737 		CPU_CLR(cpuid-1, &cpus);
5738 	}
5739 	cpuid = CPU_FFS(&cpus);
5740 	MPASS(cpuid != 0);
5741 	return (cpuid-1);
5742 }
5743 
5744 #ifdef SCHED_ULE
5745 extern struct cpu_group *cpu_top;              /* CPU topology */
5746 
5747 static int
5748 find_child_with_core(int cpu, struct cpu_group *grp)
5749 {
5750 	int i;
5751 
5752 	if (grp->cg_children == 0)
5753 		return -1;
5754 
5755 	MPASS(grp->cg_child);
5756 	for (i = 0; i < grp->cg_children; i++) {
5757 		if (CPU_ISSET(cpu, &grp->cg_child[i].cg_mask))
5758 			return i;
5759 	}
5760 
5761 	return -1;
5762 }
5763 
5764 /*
5765  * Find the nth "close" core to the specified core
5766  * "close" is defined as the deepest level that shares
5767  * at least an L2 cache.  With threads, this will be
5768  * threads on the same core.  If the shared cache is L3
5769  * or higher, simply returns the same core.
5770  */
5771 static int
5772 find_close_core(int cpu, int core_offset)
5773 {
5774 	struct cpu_group *grp;
5775 	int i;
5776 	int fcpu;
5777 	cpuset_t cs;
5778 
5779 	grp = cpu_top;
5780 	if (grp == NULL)
5781 		return cpu;
5782 	i = 0;
5783 	while ((i = find_child_with_core(cpu, grp)) != -1) {
5784 		/* If the child only has one cpu, don't descend */
5785 		if (grp->cg_child[i].cg_count <= 1)
5786 			break;
5787 		grp = &grp->cg_child[i];
5788 	}
5789 
5790 	/* If they don't share at least an L2 cache, use the same CPU */
5791 	if (grp->cg_level > CG_SHARE_L2 || grp->cg_level == CG_SHARE_NONE)
5792 		return cpu;
5793 
5794 	/* Now pick one */
5795 	CPU_COPY(&grp->cg_mask, &cs);
5796 
5797 	/* Add the selected CPU offset to core offset. */
5798 	for (i = 0; (fcpu = CPU_FFS(&cs)) != 0; i++) {
5799 		if (fcpu - 1 == cpu)
5800 			break;
5801 		CPU_CLR(fcpu - 1, &cs);
5802 	}
5803 	MPASS(fcpu);
5804 
5805 	core_offset += i;
5806 
5807 	CPU_COPY(&grp->cg_mask, &cs);
5808 	for (i = core_offset % grp->cg_count; i > 0; i--) {
5809 		MPASS(CPU_FFS(&cs));
5810 		CPU_CLR(CPU_FFS(&cs) - 1, &cs);
5811 	}
5812 	MPASS(CPU_FFS(&cs));
5813 	return CPU_FFS(&cs) - 1;
5814 }
5815 #else
5816 static int
5817 find_close_core(int cpu, int core_offset __unused)
5818 {
5819 	return cpu;
5820 }
5821 #endif
5822 
5823 static int
5824 get_core_offset(if_ctx_t ctx, iflib_intr_type_t type, int qid)
5825 {
5826 	switch (type) {
5827 	case IFLIB_INTR_TX:
5828 		/* TX queues get cores which share at least an L2 cache with the corresponding RX queue */
5829 		/* XXX handle multiple RX threads per core and more than two core per L2 group */
5830 		return qid / CPU_COUNT(&ctx->ifc_cpus) + 1;
5831 	case IFLIB_INTR_RX:
5832 	case IFLIB_INTR_RXTX:
5833 		/* RX queues get the specified core */
5834 		return qid / CPU_COUNT(&ctx->ifc_cpus);
5835 	default:
5836 		return -1;
5837 	}
5838 }
5839 #else
5840 #define get_core_offset(ctx, type, qid)	CPU_FIRST()
5841 #define find_close_core(cpuid, tid)	CPU_FIRST()
5842 #define find_nth(ctx, gid)		CPU_FIRST()
5843 #endif
5844 
5845 /* Just to avoid copy/paste */
5846 static inline int
5847 iflib_irq_set_affinity(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type,
5848     int qid, struct grouptask *gtask, struct taskqgroup *tqg, void *uniq,
5849     const char *name)
5850 {
5851 	device_t dev;
5852 	int co, cpuid, err, tid;
5853 
5854 	dev = ctx->ifc_dev;
5855 	co = ctx->ifc_sysctl_core_offset;
5856 	if (ctx->ifc_sysctl_separate_txrx && type == IFLIB_INTR_TX)
5857 		co += ctx->ifc_softc_ctx.isc_nrxqsets;
5858 	cpuid = find_nth(ctx, qid + co);
5859 	tid = get_core_offset(ctx, type, qid);
5860 	if (tid < 0) {
5861 		device_printf(dev, "get_core_offset failed\n");
5862 		return (EOPNOTSUPP);
5863 	}
5864 	cpuid = find_close_core(cpuid, tid);
5865 	err = taskqgroup_attach_cpu(tqg, gtask, uniq, cpuid, dev, irq->ii_res,
5866 	    name);
5867 	if (err) {
5868 		device_printf(dev, "taskqgroup_attach_cpu failed %d\n", err);
5869 		return (err);
5870 	}
5871 #ifdef notyet
5872 	if (cpuid > ctx->ifc_cpuid_highest)
5873 		ctx->ifc_cpuid_highest = cpuid;
5874 #endif
5875 	return (0);
5876 }
5877 
5878 int
5879 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid,
5880 			iflib_intr_type_t type, driver_filter_t *filter,
5881 			void *filter_arg, int qid, const char *name)
5882 {
5883 	device_t dev;
5884 	struct grouptask *gtask;
5885 	struct taskqgroup *tqg;
5886 	iflib_filter_info_t info;
5887 	gtask_fn_t *fn;
5888 	int tqrid, err;
5889 	driver_filter_t *intr_fast;
5890 	void *q;
5891 
5892 	info = &ctx->ifc_filter_info;
5893 	tqrid = rid;
5894 
5895 	switch (type) {
5896 	/* XXX merge tx/rx for netmap? */
5897 	case IFLIB_INTR_TX:
5898 		q = &ctx->ifc_txqs[qid];
5899 		info = &ctx->ifc_txqs[qid].ift_filter_info;
5900 		gtask = &ctx->ifc_txqs[qid].ift_task;
5901 		tqg = qgroup_if_io_tqg;
5902 		fn = _task_fn_tx;
5903 		intr_fast = iflib_fast_intr;
5904 		GROUPTASK_INIT(gtask, 0, fn, q);
5905 		ctx->ifc_flags |= IFC_NETMAP_TX_IRQ;
5906 		break;
5907 	case IFLIB_INTR_RX:
5908 		q = &ctx->ifc_rxqs[qid];
5909 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
5910 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5911 		tqg = qgroup_if_io_tqg;
5912 		fn = _task_fn_rx;
5913 		intr_fast = iflib_fast_intr;
5914 		GROUPTASK_INIT(gtask, 0, fn, q);
5915 		break;
5916 	case IFLIB_INTR_RXTX:
5917 		q = &ctx->ifc_rxqs[qid];
5918 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
5919 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5920 		tqg = qgroup_if_io_tqg;
5921 		fn = _task_fn_rx;
5922 		intr_fast = iflib_fast_intr_rxtx;
5923 		GROUPTASK_INIT(gtask, 0, fn, q);
5924 		break;
5925 	case IFLIB_INTR_ADMIN:
5926 		q = ctx;
5927 		tqrid = -1;
5928 		info = &ctx->ifc_filter_info;
5929 		gtask = &ctx->ifc_admin_task;
5930 		tqg = qgroup_if_config_tqg;
5931 		fn = _task_fn_admin;
5932 		intr_fast = iflib_fast_intr_ctx;
5933 		break;
5934 	default:
5935 		device_printf(ctx->ifc_dev, "%s: unknown net intr type\n",
5936 		    __func__);
5937 		return (EINVAL);
5938 	}
5939 
5940 	info->ifi_filter = filter;
5941 	info->ifi_filter_arg = filter_arg;
5942 	info->ifi_task = gtask;
5943 	info->ifi_ctx = q;
5944 
5945 	dev = ctx->ifc_dev;
5946 	err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info,  name);
5947 	if (err != 0) {
5948 		device_printf(dev, "_iflib_irq_alloc failed %d\n", err);
5949 		return (err);
5950 	}
5951 	if (type == IFLIB_INTR_ADMIN)
5952 		return (0);
5953 
5954 	if (tqrid != -1) {
5955 		err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg,
5956 		    q, name);
5957 		if (err)
5958 			return (err);
5959 	} else {
5960 		taskqgroup_attach(tqg, gtask, q, dev, irq->ii_res, name);
5961 	}
5962 
5963 	return (0);
5964 }
5965 
5966 void
5967 iflib_softirq_alloc_generic(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type, void *arg, int qid, const char *name)
5968 {
5969 	struct grouptask *gtask;
5970 	struct taskqgroup *tqg;
5971 	gtask_fn_t *fn;
5972 	void *q;
5973 	int err;
5974 
5975 	switch (type) {
5976 	case IFLIB_INTR_TX:
5977 		q = &ctx->ifc_txqs[qid];
5978 		gtask = &ctx->ifc_txqs[qid].ift_task;
5979 		tqg = qgroup_if_io_tqg;
5980 		fn = _task_fn_tx;
5981 		break;
5982 	case IFLIB_INTR_RX:
5983 		q = &ctx->ifc_rxqs[qid];
5984 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5985 		tqg = qgroup_if_io_tqg;
5986 		fn = _task_fn_rx;
5987 		break;
5988 	case IFLIB_INTR_IOV:
5989 		q = ctx;
5990 		gtask = &ctx->ifc_vflr_task;
5991 		tqg = qgroup_if_config_tqg;
5992 		fn = _task_fn_iov;
5993 		break;
5994 	default:
5995 		panic("unknown net intr type");
5996 	}
5997 	GROUPTASK_INIT(gtask, 0, fn, q);
5998 	if (irq != NULL) {
5999 		err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg,
6000 		    q, name);
6001 		if (err)
6002 			taskqgroup_attach(tqg, gtask, q, ctx->ifc_dev,
6003 			    irq->ii_res, name);
6004 	} else {
6005 		taskqgroup_attach(tqg, gtask, q, NULL, NULL, name);
6006 	}
6007 }
6008 
6009 void
6010 iflib_irq_free(if_ctx_t ctx, if_irq_t irq)
6011 {
6012 
6013 	if (irq->ii_tag)
6014 		bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag);
6015 
6016 	if (irq->ii_res)
6017 		bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ,
6018 		    rman_get_rid(irq->ii_res), irq->ii_res);
6019 }
6020 
6021 static int
6022 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, const char *name)
6023 {
6024 	iflib_txq_t txq = ctx->ifc_txqs;
6025 	iflib_rxq_t rxq = ctx->ifc_rxqs;
6026 	if_irq_t irq = &ctx->ifc_legacy_irq;
6027 	iflib_filter_info_t info;
6028 	device_t dev;
6029 	struct grouptask *gtask;
6030 	struct resource *res;
6031 	struct taskqgroup *tqg;
6032 	gtask_fn_t *fn;
6033 	void *q;
6034 	int err, tqrid;
6035 
6036 	q = &ctx->ifc_rxqs[0];
6037 	info = &rxq[0].ifr_filter_info;
6038 	gtask = &rxq[0].ifr_task;
6039 	tqg = qgroup_if_io_tqg;
6040 	tqrid = *rid;
6041 	fn = _task_fn_rx;
6042 
6043 	ctx->ifc_flags |= IFC_LEGACY;
6044 	info->ifi_filter = filter;
6045 	info->ifi_filter_arg = filter_arg;
6046 	info->ifi_task = gtask;
6047 	info->ifi_ctx = q;
6048 
6049 	dev = ctx->ifc_dev;
6050 	/* We allocate a single interrupt resource */
6051 	if ((err = _iflib_irq_alloc(ctx, irq, tqrid, iflib_fast_intr_rxtx,
6052 	    NULL, info, name)) != 0)
6053 		return (err);
6054 	GROUPTASK_INIT(gtask, 0, fn, q);
6055 	res = irq->ii_res;
6056 	taskqgroup_attach(tqg, gtask, q, dev, res, name);
6057 
6058 	GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq);
6059 	taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, dev, res,
6060 	    "tx");
6061 	return (0);
6062 }
6063 
6064 void
6065 iflib_led_create(if_ctx_t ctx)
6066 {
6067 
6068 	ctx->ifc_led_dev = led_create(iflib_led_func, ctx,
6069 	    device_get_nameunit(ctx->ifc_dev));
6070 }
6071 
6072 void
6073 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid)
6074 {
6075 
6076 	GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
6077 }
6078 
6079 void
6080 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid)
6081 {
6082 
6083 	GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task);
6084 }
6085 
6086 void
6087 iflib_admin_intr_deferred(if_ctx_t ctx)
6088 {
6089 #ifdef INVARIANTS
6090 	struct grouptask *gtask;
6091 
6092 	gtask = &ctx->ifc_admin_task;
6093 	MPASS(gtask != NULL && gtask->gt_taskqueue != NULL);
6094 #endif
6095 
6096 	GROUPTASK_ENQUEUE(&ctx->ifc_admin_task);
6097 }
6098 
6099 void
6100 iflib_iov_intr_deferred(if_ctx_t ctx)
6101 {
6102 
6103 	GROUPTASK_ENQUEUE(&ctx->ifc_vflr_task);
6104 }
6105 
6106 void
6107 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, const char *name)
6108 {
6109 
6110 	taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, NULL, NULL,
6111 	    name);
6112 }
6113 
6114 void
6115 iflib_config_gtask_init(void *ctx, struct grouptask *gtask, gtask_fn_t *fn,
6116 	const char *name)
6117 {
6118 
6119 	GROUPTASK_INIT(gtask, 0, fn, ctx);
6120 	taskqgroup_attach(qgroup_if_config_tqg, gtask, gtask, NULL, NULL,
6121 	    name);
6122 }
6123 
6124 void
6125 iflib_config_gtask_deinit(struct grouptask *gtask)
6126 {
6127 
6128 	taskqgroup_detach(qgroup_if_config_tqg, gtask);
6129 }
6130 
6131 void
6132 iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate)
6133 {
6134 	if_t ifp = ctx->ifc_ifp;
6135 	iflib_txq_t txq = ctx->ifc_txqs;
6136 
6137 	if_setbaudrate(ifp, baudrate);
6138 	if (baudrate >= IF_Gbps(10)) {
6139 		STATE_LOCK(ctx);
6140 		ctx->ifc_flags |= IFC_PREFETCH;
6141 		STATE_UNLOCK(ctx);
6142 	}
6143 	/* If link down, disable watchdog */
6144 	if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) {
6145 		for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++)
6146 			txq->ift_qstatus = IFLIB_QUEUE_IDLE;
6147 	}
6148 	ctx->ifc_link_state = link_state;
6149 	if_link_state_change(ifp, link_state);
6150 }
6151 
6152 static int
6153 iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq)
6154 {
6155 	int credits;
6156 #ifdef INVARIANTS
6157 	int credits_pre = txq->ift_cidx_processed;
6158 #endif
6159 
6160 	if (ctx->isc_txd_credits_update == NULL)
6161 		return (0);
6162 
6163 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
6164 	    BUS_DMASYNC_POSTREAD);
6165 	if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, true)) == 0)
6166 		return (0);
6167 
6168 	txq->ift_processed += credits;
6169 	txq->ift_cidx_processed += credits;
6170 
6171 	MPASS(credits_pre + credits == txq->ift_cidx_processed);
6172 	if (txq->ift_cidx_processed >= txq->ift_size)
6173 		txq->ift_cidx_processed -= txq->ift_size;
6174 	return (credits);
6175 }
6176 
6177 static int
6178 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget)
6179 {
6180 	iflib_fl_t fl;
6181 	u_int i;
6182 
6183 	for (i = 0, fl = &rxq->ifr_fl[0]; i < rxq->ifr_nfl; i++, fl++)
6184 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
6185 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
6186 	return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx,
6187 	    budget));
6188 }
6189 
6190 void
6191 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name,
6192 	const char *description, if_int_delay_info_t info,
6193 	int offset, int value)
6194 {
6195 	info->iidi_ctx = ctx;
6196 	info->iidi_offset = offset;
6197 	info->iidi_value = value;
6198 	SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev),
6199 	    SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)),
6200 	    OID_AUTO, name, CTLTYPE_INT|CTLFLAG_RW,
6201 	    info, 0, iflib_sysctl_int_delay, "I", description);
6202 }
6203 
6204 struct sx *
6205 iflib_ctx_lock_get(if_ctx_t ctx)
6206 {
6207 
6208 	return (&ctx->ifc_ctx_sx);
6209 }
6210 
6211 static int
6212 iflib_msix_init(if_ctx_t ctx)
6213 {
6214 	device_t dev = ctx->ifc_dev;
6215 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6216 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6217 	int admincnt, bar, err, iflib_num_rx_queues, iflib_num_tx_queues;
6218 	int msgs, queuemsgs, queues, rx_queues, tx_queues, vectors;
6219 
6220 	iflib_num_tx_queues = ctx->ifc_sysctl_ntxqs;
6221 	iflib_num_rx_queues = ctx->ifc_sysctl_nrxqs;
6222 
6223 	if (bootverbose)
6224 		device_printf(dev, "msix_init qsets capped at %d\n",
6225 		    imax(scctx->isc_ntxqsets, scctx->isc_nrxqsets));
6226 
6227 	/* Override by tuneable */
6228 	if (scctx->isc_disable_msix)
6229 		goto msi;
6230 
6231 	/* First try MSI-X */
6232 	if ((msgs = pci_msix_count(dev)) == 0) {
6233 		if (bootverbose)
6234 			device_printf(dev, "MSI-X not supported or disabled\n");
6235 		goto msi;
6236 	}
6237 
6238 	bar = ctx->ifc_softc_ctx.isc_msix_bar;
6239 	/*
6240 	 * bar == -1 => "trust me I know what I'm doing"
6241 	 * Some drivers are for hardware that is so shoddily
6242 	 * documented that no one knows which bars are which
6243 	 * so the developer has to map all bars. This hack
6244 	 * allows shoddy garbage to use MSI-X in this framework.
6245 	 */
6246 	if (bar != -1) {
6247 		ctx->ifc_msix_mem = bus_alloc_resource_any(dev,
6248 	            SYS_RES_MEMORY, &bar, RF_ACTIVE);
6249 		if (ctx->ifc_msix_mem == NULL) {
6250 			device_printf(dev, "Unable to map MSI-X table\n");
6251 			goto msi;
6252 		}
6253 	}
6254 
6255 	admincnt = sctx->isc_admin_intrcnt;
6256 #if IFLIB_DEBUG
6257 	/* use only 1 qset in debug mode */
6258 	queuemsgs = min(msgs - admincnt, 1);
6259 #else
6260 	queuemsgs = msgs - admincnt;
6261 #endif
6262 #ifdef RSS
6263 	queues = imin(queuemsgs, rss_getnumbuckets());
6264 #else
6265 	queues = queuemsgs;
6266 #endif
6267 	queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues);
6268 	if (bootverbose)
6269 		device_printf(dev,
6270 		    "intr CPUs: %d queue msgs: %d admincnt: %d\n",
6271 		    CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt);
6272 #ifdef  RSS
6273 	/* If we're doing RSS, clamp at the number of RSS buckets */
6274 	if (queues > rss_getnumbuckets())
6275 		queues = rss_getnumbuckets();
6276 #endif
6277 	if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt)
6278 		rx_queues = iflib_num_rx_queues;
6279 	else
6280 		rx_queues = queues;
6281 
6282 	if (rx_queues > scctx->isc_nrxqsets)
6283 		rx_queues = scctx->isc_nrxqsets;
6284 
6285 	/*
6286 	 * We want this to be all logical CPUs by default
6287 	 */
6288 	if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues)
6289 		tx_queues = iflib_num_tx_queues;
6290 	else
6291 		tx_queues = mp_ncpus;
6292 
6293 	if (tx_queues > scctx->isc_ntxqsets)
6294 		tx_queues = scctx->isc_ntxqsets;
6295 
6296 	if (ctx->ifc_sysctl_qs_eq_override == 0) {
6297 #ifdef INVARIANTS
6298 		if (tx_queues != rx_queues)
6299 			device_printf(dev,
6300 			    "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n",
6301 			    min(rx_queues, tx_queues), min(rx_queues, tx_queues));
6302 #endif
6303 		tx_queues = min(rx_queues, tx_queues);
6304 		rx_queues = min(rx_queues, tx_queues);
6305 	}
6306 
6307 	vectors = rx_queues + admincnt;
6308 	if (msgs < vectors) {
6309 		device_printf(dev,
6310 		    "insufficient number of MSI-X vectors "
6311 		    "(supported %d, need %d)\n", msgs, vectors);
6312 		goto msi;
6313 	}
6314 
6315 	device_printf(dev, "Using %d RX queues %d TX queues\n", rx_queues,
6316 	    tx_queues);
6317 	msgs = vectors;
6318 	if ((err = pci_alloc_msix(dev, &vectors)) == 0) {
6319 		if (vectors != msgs) {
6320 			device_printf(dev,
6321 			    "Unable to allocate sufficient MSI-X vectors "
6322 			    "(got %d, need %d)\n", vectors, msgs);
6323 			pci_release_msi(dev);
6324 			if (bar != -1) {
6325 				bus_release_resource(dev, SYS_RES_MEMORY, bar,
6326 				    ctx->ifc_msix_mem);
6327 				ctx->ifc_msix_mem = NULL;
6328 			}
6329 			goto msi;
6330 		}
6331 		device_printf(dev, "Using MSI-X interrupts with %d vectors\n",
6332 		    vectors);
6333 		scctx->isc_vectors = vectors;
6334 		scctx->isc_nrxqsets = rx_queues;
6335 		scctx->isc_ntxqsets = tx_queues;
6336 		scctx->isc_intr = IFLIB_INTR_MSIX;
6337 
6338 		return (vectors);
6339 	} else {
6340 		device_printf(dev,
6341 		    "failed to allocate %d MSI-X vectors, err: %d\n", vectors,
6342 		    err);
6343 		if (bar != -1) {
6344 			bus_release_resource(dev, SYS_RES_MEMORY, bar,
6345 			    ctx->ifc_msix_mem);
6346 			ctx->ifc_msix_mem = NULL;
6347 		}
6348 	}
6349 
6350 msi:
6351 	vectors = pci_msi_count(dev);
6352 	scctx->isc_nrxqsets = 1;
6353 	scctx->isc_ntxqsets = 1;
6354 	scctx->isc_vectors = vectors;
6355 	if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) {
6356 		device_printf(dev,"Using an MSI interrupt\n");
6357 		scctx->isc_intr = IFLIB_INTR_MSI;
6358 	} else {
6359 		scctx->isc_vectors = 1;
6360 		device_printf(dev,"Using a Legacy interrupt\n");
6361 		scctx->isc_intr = IFLIB_INTR_LEGACY;
6362 	}
6363 
6364 	return (vectors);
6365 }
6366 
6367 static const char *ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" };
6368 
6369 static int
6370 mp_ring_state_handler(SYSCTL_HANDLER_ARGS)
6371 {
6372 	int rc;
6373 	uint16_t *state = ((uint16_t *)oidp->oid_arg1);
6374 	struct sbuf *sb;
6375 	const char *ring_state = "UNKNOWN";
6376 
6377 	/* XXX needed ? */
6378 	rc = sysctl_wire_old_buffer(req, 0);
6379 	MPASS(rc == 0);
6380 	if (rc != 0)
6381 		return (rc);
6382 	sb = sbuf_new_for_sysctl(NULL, NULL, 80, req);
6383 	MPASS(sb != NULL);
6384 	if (sb == NULL)
6385 		return (ENOMEM);
6386 	if (state[3] <= 3)
6387 		ring_state = ring_states[state[3]];
6388 
6389 	sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s",
6390 		    state[0], state[1], state[2], ring_state);
6391 	rc = sbuf_finish(sb);
6392 	sbuf_delete(sb);
6393         return(rc);
6394 }
6395 
6396 enum iflib_ndesc_handler {
6397 	IFLIB_NTXD_HANDLER,
6398 	IFLIB_NRXD_HANDLER,
6399 };
6400 
6401 static int
6402 mp_ndesc_handler(SYSCTL_HANDLER_ARGS)
6403 {
6404 	if_ctx_t ctx = (void *)arg1;
6405 	enum iflib_ndesc_handler type = arg2;
6406 	char buf[256] = {0};
6407 	qidx_t *ndesc;
6408 	char *p, *next;
6409 	int nqs, rc, i;
6410 
6411 	nqs = 8;
6412 	switch(type) {
6413 	case IFLIB_NTXD_HANDLER:
6414 		ndesc = ctx->ifc_sysctl_ntxds;
6415 		if (ctx->ifc_sctx)
6416 			nqs = ctx->ifc_sctx->isc_ntxqs;
6417 		break;
6418 	case IFLIB_NRXD_HANDLER:
6419 		ndesc = ctx->ifc_sysctl_nrxds;
6420 		if (ctx->ifc_sctx)
6421 			nqs = ctx->ifc_sctx->isc_nrxqs;
6422 		break;
6423 	default:
6424 		printf("%s: unhandled type\n", __func__);
6425 		return (EINVAL);
6426 	}
6427 	if (nqs == 0)
6428 		nqs = 8;
6429 
6430 	for (i=0; i<8; i++) {
6431 		if (i >= nqs)
6432 			break;
6433 		if (i)
6434 			strcat(buf, ",");
6435 		sprintf(strchr(buf, 0), "%d", ndesc[i]);
6436 	}
6437 
6438 	rc = sysctl_handle_string(oidp, buf, sizeof(buf), req);
6439 	if (rc || req->newptr == NULL)
6440 		return rc;
6441 
6442 	for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p;
6443 	    i++, p = strsep(&next, " ,")) {
6444 		ndesc[i] = strtoul(p, NULL, 10);
6445 	}
6446 
6447 	return(rc);
6448 }
6449 
6450 #define NAME_BUFLEN 32
6451 static void
6452 iflib_add_device_sysctl_pre(if_ctx_t ctx)
6453 {
6454         device_t dev = iflib_get_dev(ctx);
6455 	struct sysctl_oid_list *child, *oid_list;
6456 	struct sysctl_ctx_list *ctx_list;
6457 	struct sysctl_oid *node;
6458 
6459 	ctx_list = device_get_sysctl_ctx(dev);
6460 	child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev));
6461 	ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, "iflib",
6462 						      CTLFLAG_RD, NULL, "IFLIB fields");
6463 	oid_list = SYSCTL_CHILDREN(node);
6464 
6465 	SYSCTL_ADD_CONST_STRING(ctx_list, oid_list, OID_AUTO, "driver_version",
6466 		       CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version,
6467 		       "driver version");
6468 
6469 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs",
6470 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0,
6471 			"# of txqs to use, 0 => use default #");
6472 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs",
6473 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0,
6474 			"# of rxqs to use, 0 => use default #");
6475 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_qs_enable",
6476 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0,
6477                        "permit #txq != #rxq");
6478 	SYSCTL_ADD_INT(ctx_list, oid_list, OID_AUTO, "disable_msix",
6479                       CTLFLAG_RWTUN, &ctx->ifc_softc_ctx.isc_disable_msix, 0,
6480                       "disable MSI-X (default 0)");
6481 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "rx_budget",
6482 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_rx_budget, 0,
6483 		       "set the RX budget");
6484 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "tx_abdicate",
6485 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_tx_abdicate, 0,
6486 		       "cause TX to abdicate instead of running to completion");
6487 	ctx->ifc_sysctl_core_offset = CORE_OFFSET_UNSPECIFIED;
6488 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "core_offset",
6489 		       CTLFLAG_RDTUN, &ctx->ifc_sysctl_core_offset, 0,
6490 		       "offset to start using cores at");
6491 	SYSCTL_ADD_U8(ctx_list, oid_list, OID_AUTO, "separate_txrx",
6492 		       CTLFLAG_RDTUN, &ctx->ifc_sysctl_separate_txrx, 0,
6493 		       "use separate cores for TX and RX");
6494 
6495 	/* XXX change for per-queue sizes */
6496 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_ntxds",
6497 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NTXD_HANDLER,
6498                        mp_ndesc_handler, "A",
6499 		       "list of # of TX descriptors to use, 0 = use default #");
6500 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_nrxds",
6501 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NRXD_HANDLER,
6502                        mp_ndesc_handler, "A",
6503 		       "list of # of RX descriptors to use, 0 = use default #");
6504 }
6505 
6506 static void
6507 iflib_add_device_sysctl_post(if_ctx_t ctx)
6508 {
6509 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6510 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6511         device_t dev = iflib_get_dev(ctx);
6512 	struct sysctl_oid_list *child;
6513 	struct sysctl_ctx_list *ctx_list;
6514 	iflib_fl_t fl;
6515 	iflib_txq_t txq;
6516 	iflib_rxq_t rxq;
6517 	int i, j;
6518 	char namebuf[NAME_BUFLEN];
6519 	char *qfmt;
6520 	struct sysctl_oid *queue_node, *fl_node, *node;
6521 	struct sysctl_oid_list *queue_list, *fl_list;
6522 	ctx_list = device_get_sysctl_ctx(dev);
6523 
6524 	node = ctx->ifc_sysctl_node;
6525 	child = SYSCTL_CHILDREN(node);
6526 
6527 	if (scctx->isc_ntxqsets > 100)
6528 		qfmt = "txq%03d";
6529 	else if (scctx->isc_ntxqsets > 10)
6530 		qfmt = "txq%02d";
6531 	else
6532 		qfmt = "txq%d";
6533 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
6534 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6535 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6536 					     CTLFLAG_RD, NULL, "Queue Name");
6537 		queue_list = SYSCTL_CHILDREN(queue_node);
6538 #if MEMORY_LOGGING
6539 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued",
6540 				CTLFLAG_RD,
6541 				&txq->ift_dequeued, "total mbufs freed");
6542 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued",
6543 				CTLFLAG_RD,
6544 				&txq->ift_enqueued, "total mbufs enqueued");
6545 #endif
6546 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag",
6547 				   CTLFLAG_RD,
6548 				   &txq->ift_mbuf_defrag, "# of times m_defrag was called");
6549 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "m_pullups",
6550 				   CTLFLAG_RD,
6551 				   &txq->ift_pullups, "# of times m_pullup was called");
6552 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag_failed",
6553 				   CTLFLAG_RD,
6554 				   &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed");
6555 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_desc_avail",
6556 				   CTLFLAG_RD,
6557 				   &txq->ift_no_desc_avail, "# of times no descriptors were available");
6558 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "tx_map_failed",
6559 				   CTLFLAG_RD,
6560 				   &txq->ift_map_failed, "# of times DMA map failed");
6561 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txd_encap_efbig",
6562 				   CTLFLAG_RD,
6563 				   &txq->ift_txd_encap_efbig, "# of times txd_encap returned EFBIG");
6564 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_tx_dma_setup",
6565 				   CTLFLAG_RD,
6566 				   &txq->ift_no_tx_dma_setup, "# of times map failed for other than EFBIG");
6567 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx",
6568 				   CTLFLAG_RD,
6569 				   &txq->ift_pidx, 1, "Producer Index");
6570 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx",
6571 				   CTLFLAG_RD,
6572 				   &txq->ift_cidx, 1, "Consumer Index");
6573 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx_processed",
6574 				   CTLFLAG_RD,
6575 				   &txq->ift_cidx_processed, 1, "Consumer Index seen by credit update");
6576 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use",
6577 				   CTLFLAG_RD,
6578 				   &txq->ift_in_use, 1, "descriptors in use");
6579 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_processed",
6580 				   CTLFLAG_RD,
6581 				   &txq->ift_processed, "descriptors procesed for clean");
6582 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned",
6583 				   CTLFLAG_RD,
6584 				   &txq->ift_cleaned, "total cleaned");
6585 		SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state",
6586 				CTLTYPE_STRING | CTLFLAG_RD, __DEVOLATILE(uint64_t *, &txq->ift_br->state),
6587 				0, mp_ring_state_handler, "A", "soft ring state");
6588 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_enqueues",
6589 				       CTLFLAG_RD, &txq->ift_br->enqueues,
6590 				       "# of enqueues to the mp_ring for this queue");
6591 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_drops",
6592 				       CTLFLAG_RD, &txq->ift_br->drops,
6593 				       "# of drops in the mp_ring for this queue");
6594 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_starts",
6595 				       CTLFLAG_RD, &txq->ift_br->starts,
6596 				       "# of normal consumer starts in the mp_ring for this queue");
6597 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_stalls",
6598 				       CTLFLAG_RD, &txq->ift_br->stalls,
6599 					       "# of consumer stalls in the mp_ring for this queue");
6600 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_restarts",
6601 			       CTLFLAG_RD, &txq->ift_br->restarts,
6602 				       "# of consumer restarts in the mp_ring for this queue");
6603 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_abdications",
6604 				       CTLFLAG_RD, &txq->ift_br->abdications,
6605 				       "# of consumer abdications in the mp_ring for this queue");
6606 	}
6607 
6608 	if (scctx->isc_nrxqsets > 100)
6609 		qfmt = "rxq%03d";
6610 	else if (scctx->isc_nrxqsets > 10)
6611 		qfmt = "rxq%02d";
6612 	else
6613 		qfmt = "rxq%d";
6614 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
6615 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6616 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6617 					     CTLFLAG_RD, NULL, "Queue Name");
6618 		queue_list = SYSCTL_CHILDREN(queue_node);
6619 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
6620 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_cidx",
6621 				       CTLFLAG_RD,
6622 				       &rxq->ifr_cq_cidx, 1, "Consumer Index");
6623 		}
6624 
6625 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
6626 			snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j);
6627 			fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list, OID_AUTO, namebuf,
6628 						     CTLFLAG_RD, NULL, "freelist Name");
6629 			fl_list = SYSCTL_CHILDREN(fl_node);
6630 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx",
6631 				       CTLFLAG_RD,
6632 				       &fl->ifl_pidx, 1, "Producer Index");
6633 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx",
6634 				       CTLFLAG_RD,
6635 				       &fl->ifl_cidx, 1, "Consumer Index");
6636 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits",
6637 				       CTLFLAG_RD,
6638 				       &fl->ifl_credits, 1, "credits available");
6639 #if MEMORY_LOGGING
6640 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_enqueued",
6641 					CTLFLAG_RD,
6642 					&fl->ifl_m_enqueued, "mbufs allocated");
6643 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_dequeued",
6644 					CTLFLAG_RD,
6645 					&fl->ifl_m_dequeued, "mbufs freed");
6646 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_enqueued",
6647 					CTLFLAG_RD,
6648 					&fl->ifl_cl_enqueued, "clusters allocated");
6649 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_dequeued",
6650 					CTLFLAG_RD,
6651 					&fl->ifl_cl_dequeued, "clusters freed");
6652 #endif
6653 
6654 		}
6655 	}
6656 
6657 }
6658 
6659 void
6660 iflib_request_reset(if_ctx_t ctx)
6661 {
6662 
6663 	STATE_LOCK(ctx);
6664 	ctx->ifc_flags |= IFC_DO_RESET;
6665 	STATE_UNLOCK(ctx);
6666 }
6667 
6668 #ifndef __NO_STRICT_ALIGNMENT
6669 static struct mbuf *
6670 iflib_fixup_rx(struct mbuf *m)
6671 {
6672 	struct mbuf *n;
6673 
6674 	if (m->m_len <= (MCLBYTES - ETHER_HDR_LEN)) {
6675 		bcopy(m->m_data, m->m_data + ETHER_HDR_LEN, m->m_len);
6676 		m->m_data += ETHER_HDR_LEN;
6677 		n = m;
6678 	} else {
6679 		MGETHDR(n, M_NOWAIT, MT_DATA);
6680 		if (n == NULL) {
6681 			m_freem(m);
6682 			return (NULL);
6683 		}
6684 		bcopy(m->m_data, n->m_data, ETHER_HDR_LEN);
6685 		m->m_data += ETHER_HDR_LEN;
6686 		m->m_len -= ETHER_HDR_LEN;
6687 		n->m_len = ETHER_HDR_LEN;
6688 		M_MOVE_PKTHDR(n, m);
6689 		n->m_next = m;
6690 	}
6691 	return (n);
6692 }
6693 #endif
6694 
6695 #ifdef NETDUMP
6696 static void
6697 iflib_netdump_init(if_t ifp, int *nrxr, int *ncl, int *clsize)
6698 {
6699 	if_ctx_t ctx;
6700 
6701 	ctx = if_getsoftc(ifp);
6702 	CTX_LOCK(ctx);
6703 	*nrxr = NRXQSETS(ctx);
6704 	*ncl = ctx->ifc_rxqs[0].ifr_fl->ifl_size;
6705 	*clsize = ctx->ifc_rxqs[0].ifr_fl->ifl_buf_size;
6706 	CTX_UNLOCK(ctx);
6707 }
6708 
6709 static void
6710 iflib_netdump_event(if_t ifp, enum netdump_ev event)
6711 {
6712 	if_ctx_t ctx;
6713 	if_softc_ctx_t scctx;
6714 	iflib_fl_t fl;
6715 	iflib_rxq_t rxq;
6716 	int i, j;
6717 
6718 	ctx = if_getsoftc(ifp);
6719 	scctx = &ctx->ifc_softc_ctx;
6720 
6721 	switch (event) {
6722 	case NETDUMP_START:
6723 		for (i = 0; i < scctx->isc_nrxqsets; i++) {
6724 			rxq = &ctx->ifc_rxqs[i];
6725 			for (j = 0; j < rxq->ifr_nfl; j++) {
6726 				fl = rxq->ifr_fl;
6727 				fl->ifl_zone = m_getzone(fl->ifl_buf_size);
6728 			}
6729 		}
6730 		iflib_no_tx_batch = 1;
6731 		break;
6732 	default:
6733 		break;
6734 	}
6735 }
6736 
6737 static int
6738 iflib_netdump_transmit(if_t ifp, struct mbuf *m)
6739 {
6740 	if_ctx_t ctx;
6741 	iflib_txq_t txq;
6742 	int error;
6743 
6744 	ctx = if_getsoftc(ifp);
6745 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
6746 	    IFF_DRV_RUNNING)
6747 		return (EBUSY);
6748 
6749 	txq = &ctx->ifc_txqs[0];
6750 	error = iflib_encap(txq, &m);
6751 	if (error == 0)
6752 		(void)iflib_txd_db_check(ctx, txq, true, txq->ift_in_use);
6753 	return (error);
6754 }
6755 
6756 static int
6757 iflib_netdump_poll(if_t ifp, int count)
6758 {
6759 	if_ctx_t ctx;
6760 	if_softc_ctx_t scctx;
6761 	iflib_txq_t txq;
6762 	int i;
6763 
6764 	ctx = if_getsoftc(ifp);
6765 	scctx = &ctx->ifc_softc_ctx;
6766 
6767 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
6768 	    IFF_DRV_RUNNING)
6769 		return (EBUSY);
6770 
6771 	txq = &ctx->ifc_txqs[0];
6772 	(void)iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
6773 
6774 	for (i = 0; i < scctx->isc_nrxqsets; i++)
6775 		(void)iflib_rxeof(&ctx->ifc_rxqs[i], 16 /* XXX */);
6776 	return (0);
6777 }
6778 #endif /* NETDUMP */
6779