xref: /freebsd/sys/net/iflib.c (revision 955a3f9ad586f38395e66127f9f2f4afbf3d5a94)
1 /*-
2  * Copyright (c) 2014-2018, Matthew Macy <mmacy@mattmacy.io>
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions are met:
7  *
8  *  1. Redistributions of source code must retain the above copyright notice,
9  *     this list of conditions and the following disclaimer.
10  *
11  *  2. Neither the name of Matthew Macy nor the names of its
12  *     contributors may be used to endorse or promote products derived from
13  *     this software without specific prior written permission.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
16  * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18  * ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
19  * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
20  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
21  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
22  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
23  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
24  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
25  * POSSIBILITY OF SUCH DAMAGE.
26  */
27 
28 #include <sys/cdefs.h>
29 __FBSDID("$FreeBSD$");
30 
31 #include "opt_inet.h"
32 #include "opt_inet6.h"
33 #include "opt_acpi.h"
34 #include "opt_sched.h"
35 
36 #include <sys/param.h>
37 #include <sys/types.h>
38 #include <sys/bus.h>
39 #include <sys/eventhandler.h>
40 #include <sys/kernel.h>
41 #include <sys/lock.h>
42 #include <sys/mutex.h>
43 #include <sys/module.h>
44 #include <sys/kobj.h>
45 #include <sys/rman.h>
46 #include <sys/sbuf.h>
47 #include <sys/smp.h>
48 #include <sys/socket.h>
49 #include <sys/sockio.h>
50 #include <sys/sysctl.h>
51 #include <sys/syslog.h>
52 #include <sys/taskqueue.h>
53 #include <sys/limits.h>
54 
55 #include <net/if.h>
56 #include <net/if_var.h>
57 #include <net/if_types.h>
58 #include <net/if_media.h>
59 #include <net/bpf.h>
60 #include <net/ethernet.h>
61 #include <net/mp_ring.h>
62 #include <net/debugnet.h>
63 #include <net/pfil.h>
64 #include <net/vnet.h>
65 
66 #include <netinet/in.h>
67 #include <netinet/in_pcb.h>
68 #include <netinet/tcp_lro.h>
69 #include <netinet/in_systm.h>
70 #include <netinet/if_ether.h>
71 #include <netinet/ip.h>
72 #include <netinet/ip6.h>
73 #include <netinet/tcp.h>
74 #include <netinet/ip_var.h>
75 #include <netinet6/ip6_var.h>
76 
77 #include <machine/bus.h>
78 #include <machine/in_cksum.h>
79 
80 #include <vm/vm.h>
81 #include <vm/pmap.h>
82 
83 #include <dev/led/led.h>
84 #include <dev/pci/pcireg.h>
85 #include <dev/pci/pcivar.h>
86 #include <dev/pci/pci_private.h>
87 
88 #include <net/iflib.h>
89 #include <net/iflib_private.h>
90 
91 #include "ifdi_if.h"
92 
93 #ifdef PCI_IOV
94 #include <dev/pci/pci_iov.h>
95 #endif
96 
97 #include <sys/bitstring.h>
98 /*
99  * enable accounting of every mbuf as it comes in to and goes out of
100  * iflib's software descriptor references
101  */
102 #define MEMORY_LOGGING 0
103 /*
104  * Enable mbuf vectors for compressing long mbuf chains
105  */
106 
107 /*
108  * NB:
109  * - Prefetching in tx cleaning should perhaps be a tunable. The distance ahead
110  *   we prefetch needs to be determined by the time spent in m_free vis a vis
111  *   the cost of a prefetch. This will of course vary based on the workload:
112  *      - NFLX's m_free path is dominated by vm-based M_EXT manipulation which
113  *        is quite expensive, thus suggesting very little prefetch.
114  *      - small packet forwarding which is just returning a single mbuf to
115  *        UMA will typically be very fast vis a vis the cost of a memory
116  *        access.
117  */
118 
119 /*
120  * File organization:
121  *  - private structures
122  *  - iflib private utility functions
123  *  - ifnet functions
124  *  - vlan registry and other exported functions
125  *  - iflib public core functions
126  *
127  *
128  */
129 MALLOC_DEFINE(M_IFLIB, "iflib", "ifnet library");
130 
131 #define	IFLIB_RXEOF_MORE (1U << 0)
132 #define	IFLIB_RXEOF_EMPTY (2U << 0)
133 
134 struct iflib_txq;
135 typedef struct iflib_txq *iflib_txq_t;
136 struct iflib_rxq;
137 typedef struct iflib_rxq *iflib_rxq_t;
138 struct iflib_fl;
139 typedef struct iflib_fl *iflib_fl_t;
140 
141 struct iflib_ctx;
142 
143 static void iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid);
144 static void iflib_timer(void *arg);
145 static void iflib_tqg_detach(if_ctx_t ctx);
146 
147 typedef struct iflib_filter_info {
148 	driver_filter_t *ifi_filter;
149 	void *ifi_filter_arg;
150 	struct grouptask *ifi_task;
151 	void *ifi_ctx;
152 } *iflib_filter_info_t;
153 
154 struct iflib_ctx {
155 	KOBJ_FIELDS;
156 	/*
157 	 * Pointer to hardware driver's softc
158 	 */
159 	void *ifc_softc;
160 	device_t ifc_dev;
161 	if_t ifc_ifp;
162 
163 	cpuset_t ifc_cpus;
164 	if_shared_ctx_t ifc_sctx;
165 	struct if_softc_ctx ifc_softc_ctx;
166 
167 	struct sx ifc_ctx_sx;
168 	struct mtx ifc_state_mtx;
169 
170 	iflib_txq_t ifc_txqs;
171 	iflib_rxq_t ifc_rxqs;
172 	uint32_t ifc_if_flags;
173 	uint32_t ifc_flags;
174 	uint32_t ifc_max_fl_buf_size;
175 	uint32_t ifc_rx_mbuf_sz;
176 
177 	int ifc_link_state;
178 	int ifc_watchdog_events;
179 	struct cdev *ifc_led_dev;
180 	struct resource *ifc_msix_mem;
181 
182 	struct if_irq ifc_legacy_irq;
183 	struct grouptask ifc_admin_task;
184 	struct grouptask ifc_vflr_task;
185 	struct iflib_filter_info ifc_filter_info;
186 	struct ifmedia	ifc_media;
187 	struct ifmedia	*ifc_mediap;
188 
189 	struct sysctl_oid *ifc_sysctl_node;
190 	uint16_t ifc_sysctl_ntxqs;
191 	uint16_t ifc_sysctl_nrxqs;
192 	uint16_t ifc_sysctl_qs_eq_override;
193 	uint16_t ifc_sysctl_rx_budget;
194 	uint16_t ifc_sysctl_tx_abdicate;
195 	uint16_t ifc_sysctl_core_offset;
196 #define	CORE_OFFSET_UNSPECIFIED	0xffff
197 	uint8_t  ifc_sysctl_separate_txrx;
198 
199 	qidx_t ifc_sysctl_ntxds[8];
200 	qidx_t ifc_sysctl_nrxds[8];
201 	struct if_txrx ifc_txrx;
202 #define isc_txd_encap  ifc_txrx.ift_txd_encap
203 #define isc_txd_flush  ifc_txrx.ift_txd_flush
204 #define isc_txd_credits_update  ifc_txrx.ift_txd_credits_update
205 #define isc_rxd_available ifc_txrx.ift_rxd_available
206 #define isc_rxd_pkt_get ifc_txrx.ift_rxd_pkt_get
207 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
208 #define isc_rxd_flush ifc_txrx.ift_rxd_flush
209 #define isc_legacy_intr ifc_txrx.ift_legacy_intr
210 	eventhandler_tag ifc_vlan_attach_event;
211 	eventhandler_tag ifc_vlan_detach_event;
212 	struct ether_addr ifc_mac;
213 };
214 
215 void *
216 iflib_get_softc(if_ctx_t ctx)
217 {
218 
219 	return (ctx->ifc_softc);
220 }
221 
222 device_t
223 iflib_get_dev(if_ctx_t ctx)
224 {
225 
226 	return (ctx->ifc_dev);
227 }
228 
229 if_t
230 iflib_get_ifp(if_ctx_t ctx)
231 {
232 
233 	return (ctx->ifc_ifp);
234 }
235 
236 struct ifmedia *
237 iflib_get_media(if_ctx_t ctx)
238 {
239 
240 	return (ctx->ifc_mediap);
241 }
242 
243 uint32_t
244 iflib_get_flags(if_ctx_t ctx)
245 {
246 	return (ctx->ifc_flags);
247 }
248 
249 void
250 iflib_set_mac(if_ctx_t ctx, uint8_t mac[ETHER_ADDR_LEN])
251 {
252 
253 	bcopy(mac, ctx->ifc_mac.octet, ETHER_ADDR_LEN);
254 }
255 
256 if_softc_ctx_t
257 iflib_get_softc_ctx(if_ctx_t ctx)
258 {
259 
260 	return (&ctx->ifc_softc_ctx);
261 }
262 
263 if_shared_ctx_t
264 iflib_get_sctx(if_ctx_t ctx)
265 {
266 
267 	return (ctx->ifc_sctx);
268 }
269 
270 #define IP_ALIGNED(m) ((((uintptr_t)(m)->m_data) & 0x3) == 0x2)
271 #define CACHE_PTR_INCREMENT (CACHE_LINE_SIZE/sizeof(void*))
272 #define CACHE_PTR_NEXT(ptr) ((void *)(((uintptr_t)(ptr)+CACHE_LINE_SIZE-1) & (CACHE_LINE_SIZE-1)))
273 
274 #define LINK_ACTIVE(ctx) ((ctx)->ifc_link_state == LINK_STATE_UP)
275 #define CTX_IS_VF(ctx) ((ctx)->ifc_sctx->isc_flags & IFLIB_IS_VF)
276 
277 typedef struct iflib_sw_rx_desc_array {
278 	bus_dmamap_t	*ifsd_map;         /* bus_dma maps for packet */
279 	struct mbuf	**ifsd_m;           /* pkthdr mbufs */
280 	caddr_t		*ifsd_cl;          /* direct cluster pointer for rx */
281 	bus_addr_t	*ifsd_ba;          /* bus addr of cluster for rx */
282 } iflib_rxsd_array_t;
283 
284 typedef struct iflib_sw_tx_desc_array {
285 	bus_dmamap_t    *ifsd_map;         /* bus_dma maps for packet */
286 	bus_dmamap_t	*ifsd_tso_map;     /* bus_dma maps for TSO packet */
287 	struct mbuf    **ifsd_m;           /* pkthdr mbufs */
288 } if_txsd_vec_t;
289 
290 /* magic number that should be high enough for any hardware */
291 #define IFLIB_MAX_TX_SEGS		128
292 #define IFLIB_RX_COPY_THRESH		128
293 #define IFLIB_MAX_RX_REFRESH		32
294 /* The minimum descriptors per second before we start coalescing */
295 #define IFLIB_MIN_DESC_SEC		16384
296 #define IFLIB_DEFAULT_TX_UPDATE_FREQ	16
297 #define IFLIB_QUEUE_IDLE		0
298 #define IFLIB_QUEUE_HUNG		1
299 #define IFLIB_QUEUE_WORKING		2
300 /* maximum number of txqs that can share an rx interrupt */
301 #define IFLIB_MAX_TX_SHARED_INTR	4
302 
303 /* this should really scale with ring size - this is a fairly arbitrary value */
304 #define TX_BATCH_SIZE			32
305 
306 #define IFLIB_RESTART_BUDGET		8
307 
308 #define CSUM_OFFLOAD		(CSUM_IP_TSO|CSUM_IP6_TSO|CSUM_IP| \
309 				 CSUM_IP_UDP|CSUM_IP_TCP|CSUM_IP_SCTP| \
310 				 CSUM_IP6_UDP|CSUM_IP6_TCP|CSUM_IP6_SCTP)
311 
312 struct iflib_txq {
313 	qidx_t		ift_in_use;
314 	qidx_t		ift_cidx;
315 	qidx_t		ift_cidx_processed;
316 	qidx_t		ift_pidx;
317 	uint8_t		ift_gen;
318 	uint8_t		ift_br_offset;
319 	uint16_t	ift_npending;
320 	uint16_t	ift_db_pending;
321 	uint16_t	ift_rs_pending;
322 	/* implicit pad */
323 	uint8_t		ift_txd_size[8];
324 	uint64_t	ift_processed;
325 	uint64_t	ift_cleaned;
326 	uint64_t	ift_cleaned_prev;
327 #if MEMORY_LOGGING
328 	uint64_t	ift_enqueued;
329 	uint64_t	ift_dequeued;
330 #endif
331 	uint64_t	ift_no_tx_dma_setup;
332 	uint64_t	ift_no_desc_avail;
333 	uint64_t	ift_mbuf_defrag_failed;
334 	uint64_t	ift_mbuf_defrag;
335 	uint64_t	ift_map_failed;
336 	uint64_t	ift_txd_encap_efbig;
337 	uint64_t	ift_pullups;
338 	uint64_t	ift_last_timer_tick;
339 
340 	struct mtx	ift_mtx;
341 	struct mtx	ift_db_mtx;
342 
343 	/* constant values */
344 	if_ctx_t	ift_ctx;
345 	struct ifmp_ring        *ift_br;
346 	struct grouptask	ift_task;
347 	qidx_t		ift_size;
348 	uint16_t	ift_id;
349 	struct callout	ift_timer;
350 #ifdef DEV_NETMAP
351 	struct callout	ift_netmap_timer;
352 #endif /* DEV_NETMAP */
353 
354 	if_txsd_vec_t	ift_sds;
355 	uint8_t		ift_qstatus;
356 	uint8_t		ift_closed;
357 	uint8_t		ift_update_freq;
358 	struct iflib_filter_info ift_filter_info;
359 	bus_dma_tag_t	ift_buf_tag;
360 	bus_dma_tag_t	ift_tso_buf_tag;
361 	iflib_dma_info_t	ift_ifdi;
362 #define	MTX_NAME_LEN	32
363 	char                    ift_mtx_name[MTX_NAME_LEN];
364 	bus_dma_segment_t	ift_segs[IFLIB_MAX_TX_SEGS]  __aligned(CACHE_LINE_SIZE);
365 #ifdef IFLIB_DIAGNOSTICS
366 	uint64_t ift_cpu_exec_count[256];
367 #endif
368 } __aligned(CACHE_LINE_SIZE);
369 
370 struct iflib_fl {
371 	qidx_t		ifl_cidx;
372 	qidx_t		ifl_pidx;
373 	qidx_t		ifl_credits;
374 	uint8_t		ifl_gen;
375 	uint8_t		ifl_rxd_size;
376 #if MEMORY_LOGGING
377 	uint64_t	ifl_m_enqueued;
378 	uint64_t	ifl_m_dequeued;
379 	uint64_t	ifl_cl_enqueued;
380 	uint64_t	ifl_cl_dequeued;
381 #endif
382 	/* implicit pad */
383 	bitstr_t 	*ifl_rx_bitmap;
384 	qidx_t		ifl_fragidx;
385 	/* constant */
386 	qidx_t		ifl_size;
387 	uint16_t	ifl_buf_size;
388 	uint16_t	ifl_cltype;
389 	uma_zone_t	ifl_zone;
390 	iflib_rxsd_array_t	ifl_sds;
391 	iflib_rxq_t	ifl_rxq;
392 	uint8_t		ifl_id;
393 	bus_dma_tag_t	ifl_buf_tag;
394 	iflib_dma_info_t	ifl_ifdi;
395 	uint64_t	ifl_bus_addrs[IFLIB_MAX_RX_REFRESH] __aligned(CACHE_LINE_SIZE);
396 	qidx_t		ifl_rxd_idxs[IFLIB_MAX_RX_REFRESH];
397 }  __aligned(CACHE_LINE_SIZE);
398 
399 static inline qidx_t
400 get_inuse(int size, qidx_t cidx, qidx_t pidx, uint8_t gen)
401 {
402 	qidx_t used;
403 
404 	if (pidx > cidx)
405 		used = pidx - cidx;
406 	else if (pidx < cidx)
407 		used = size - cidx + pidx;
408 	else if (gen == 0 && pidx == cidx)
409 		used = 0;
410 	else if (gen == 1 && pidx == cidx)
411 		used = size;
412 	else
413 		panic("bad state");
414 
415 	return (used);
416 }
417 
418 #define TXQ_AVAIL(txq) (txq->ift_size - get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx, txq->ift_gen))
419 
420 #define IDXDIFF(head, tail, wrap) \
421 	((head) >= (tail) ? (head) - (tail) : (wrap) - (tail) + (head))
422 
423 struct iflib_rxq {
424 	if_ctx_t	ifr_ctx;
425 	iflib_fl_t	ifr_fl;
426 	uint64_t	ifr_rx_irq;
427 	struct pfil_head	*pfil;
428 	/*
429 	 * If there is a separate completion queue (IFLIB_HAS_RXCQ), this is
430 	 * the completion queue consumer index.  Otherwise it's unused.
431 	 */
432 	qidx_t		ifr_cq_cidx;
433 	uint16_t	ifr_id;
434 	uint8_t		ifr_nfl;
435 	uint8_t		ifr_ntxqirq;
436 	uint8_t		ifr_txqid[IFLIB_MAX_TX_SHARED_INTR];
437 	uint8_t		ifr_fl_offset;
438 	struct lro_ctrl			ifr_lc;
439 	struct grouptask        ifr_task;
440 	struct callout		ifr_watchdog;
441 	struct iflib_filter_info ifr_filter_info;
442 	iflib_dma_info_t		ifr_ifdi;
443 
444 	/* dynamically allocate if any drivers need a value substantially larger than this */
445 	struct if_rxd_frag	ifr_frags[IFLIB_MAX_RX_SEGS] __aligned(CACHE_LINE_SIZE);
446 #ifdef IFLIB_DIAGNOSTICS
447 	uint64_t ifr_cpu_exec_count[256];
448 #endif
449 }  __aligned(CACHE_LINE_SIZE);
450 
451 typedef struct if_rxsd {
452 	caddr_t *ifsd_cl;
453 	iflib_fl_t ifsd_fl;
454 } *if_rxsd_t;
455 
456 /* multiple of word size */
457 #ifdef __LP64__
458 #define PKT_INFO_SIZE	6
459 #define RXD_INFO_SIZE	5
460 #define PKT_TYPE uint64_t
461 #else
462 #define PKT_INFO_SIZE	11
463 #define RXD_INFO_SIZE	8
464 #define PKT_TYPE uint32_t
465 #endif
466 #define PKT_LOOP_BOUND  ((PKT_INFO_SIZE/3)*3)
467 #define RXD_LOOP_BOUND  ((RXD_INFO_SIZE/4)*4)
468 
469 typedef struct if_pkt_info_pad {
470 	PKT_TYPE pkt_val[PKT_INFO_SIZE];
471 } *if_pkt_info_pad_t;
472 typedef struct if_rxd_info_pad {
473 	PKT_TYPE rxd_val[RXD_INFO_SIZE];
474 } *if_rxd_info_pad_t;
475 
476 CTASSERT(sizeof(struct if_pkt_info_pad) == sizeof(struct if_pkt_info));
477 CTASSERT(sizeof(struct if_rxd_info_pad) == sizeof(struct if_rxd_info));
478 
479 static inline void
480 pkt_info_zero(if_pkt_info_t pi)
481 {
482 	if_pkt_info_pad_t pi_pad;
483 
484 	pi_pad = (if_pkt_info_pad_t)pi;
485 	pi_pad->pkt_val[0] = 0; pi_pad->pkt_val[1] = 0; pi_pad->pkt_val[2] = 0;
486 	pi_pad->pkt_val[3] = 0; pi_pad->pkt_val[4] = 0; pi_pad->pkt_val[5] = 0;
487 #ifndef __LP64__
488 	pi_pad->pkt_val[6] = 0; pi_pad->pkt_val[7] = 0; pi_pad->pkt_val[8] = 0;
489 	pi_pad->pkt_val[9] = 0; pi_pad->pkt_val[10] = 0;
490 #endif
491 }
492 
493 static device_method_t iflib_pseudo_methods[] = {
494 	DEVMETHOD(device_attach, noop_attach),
495 	DEVMETHOD(device_detach, iflib_pseudo_detach),
496 	DEVMETHOD_END
497 };
498 
499 driver_t iflib_pseudodriver = {
500 	"iflib_pseudo", iflib_pseudo_methods, sizeof(struct iflib_ctx),
501 };
502 
503 static inline void
504 rxd_info_zero(if_rxd_info_t ri)
505 {
506 	if_rxd_info_pad_t ri_pad;
507 	int i;
508 
509 	ri_pad = (if_rxd_info_pad_t)ri;
510 	for (i = 0; i < RXD_LOOP_BOUND; i += 4) {
511 		ri_pad->rxd_val[i] = 0;
512 		ri_pad->rxd_val[i+1] = 0;
513 		ri_pad->rxd_val[i+2] = 0;
514 		ri_pad->rxd_val[i+3] = 0;
515 	}
516 #ifdef __LP64__
517 	ri_pad->rxd_val[RXD_INFO_SIZE-1] = 0;
518 #endif
519 }
520 
521 /*
522  * Only allow a single packet to take up most 1/nth of the tx ring
523  */
524 #define MAX_SINGLE_PACKET_FRACTION 12
525 #define IF_BAD_DMA (bus_addr_t)-1
526 
527 #define CTX_ACTIVE(ctx) ((if_getdrvflags((ctx)->ifc_ifp) & IFF_DRV_RUNNING))
528 
529 #define CTX_LOCK_INIT(_sc)  sx_init(&(_sc)->ifc_ctx_sx, "iflib ctx lock")
530 #define CTX_LOCK(ctx) sx_xlock(&(ctx)->ifc_ctx_sx)
531 #define CTX_UNLOCK(ctx) sx_xunlock(&(ctx)->ifc_ctx_sx)
532 #define CTX_LOCK_DESTROY(ctx) sx_destroy(&(ctx)->ifc_ctx_sx)
533 
534 #define STATE_LOCK_INIT(_sc, _name)  mtx_init(&(_sc)->ifc_state_mtx, _name, "iflib state lock", MTX_DEF)
535 #define STATE_LOCK(ctx) mtx_lock(&(ctx)->ifc_state_mtx)
536 #define STATE_UNLOCK(ctx) mtx_unlock(&(ctx)->ifc_state_mtx)
537 #define STATE_LOCK_DESTROY(ctx) mtx_destroy(&(ctx)->ifc_state_mtx)
538 
539 #define CALLOUT_LOCK(txq)	mtx_lock(&txq->ift_mtx)
540 #define CALLOUT_UNLOCK(txq) 	mtx_unlock(&txq->ift_mtx)
541 
542 void
543 iflib_set_detach(if_ctx_t ctx)
544 {
545 	STATE_LOCK(ctx);
546 	ctx->ifc_flags |= IFC_IN_DETACH;
547 	STATE_UNLOCK(ctx);
548 }
549 
550 /* Our boot-time initialization hook */
551 static int	iflib_module_event_handler(module_t, int, void *);
552 
553 static moduledata_t iflib_moduledata = {
554 	"iflib",
555 	iflib_module_event_handler,
556 	NULL
557 };
558 
559 DECLARE_MODULE(iflib, iflib_moduledata, SI_SUB_INIT_IF, SI_ORDER_ANY);
560 MODULE_VERSION(iflib, 1);
561 
562 MODULE_DEPEND(iflib, pci, 1, 1, 1);
563 MODULE_DEPEND(iflib, ether, 1, 1, 1);
564 
565 TASKQGROUP_DEFINE(if_io_tqg, mp_ncpus, 1);
566 TASKQGROUP_DEFINE(if_config_tqg, 1, 1);
567 
568 #ifndef IFLIB_DEBUG_COUNTERS
569 #ifdef INVARIANTS
570 #define IFLIB_DEBUG_COUNTERS 1
571 #else
572 #define IFLIB_DEBUG_COUNTERS 0
573 #endif /* !INVARIANTS */
574 #endif
575 
576 static SYSCTL_NODE(_net, OID_AUTO, iflib, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
577     "iflib driver parameters");
578 
579 /*
580  * XXX need to ensure that this can't accidentally cause the head to be moved backwards
581  */
582 static int iflib_min_tx_latency = 0;
583 SYSCTL_INT(_net_iflib, OID_AUTO, min_tx_latency, CTLFLAG_RW,
584 		   &iflib_min_tx_latency, 0, "minimize transmit latency at the possible expense of throughput");
585 static int iflib_no_tx_batch = 0;
586 SYSCTL_INT(_net_iflib, OID_AUTO, no_tx_batch, CTLFLAG_RW,
587 		   &iflib_no_tx_batch, 0, "minimize transmit latency at the possible expense of throughput");
588 static int iflib_timer_default = 1000;
589 SYSCTL_INT(_net_iflib, OID_AUTO, timer_default, CTLFLAG_RW,
590 		   &iflib_timer_default, 0, "number of ticks between iflib_timer calls");
591 
592 
593 #if IFLIB_DEBUG_COUNTERS
594 
595 static int iflib_tx_seen;
596 static int iflib_tx_sent;
597 static int iflib_tx_encap;
598 static int iflib_rx_allocs;
599 static int iflib_fl_refills;
600 static int iflib_fl_refills_large;
601 static int iflib_tx_frees;
602 
603 SYSCTL_INT(_net_iflib, OID_AUTO, tx_seen, CTLFLAG_RD,
604 		   &iflib_tx_seen, 0, "# TX mbufs seen");
605 SYSCTL_INT(_net_iflib, OID_AUTO, tx_sent, CTLFLAG_RD,
606 		   &iflib_tx_sent, 0, "# TX mbufs sent");
607 SYSCTL_INT(_net_iflib, OID_AUTO, tx_encap, CTLFLAG_RD,
608 		   &iflib_tx_encap, 0, "# TX mbufs encapped");
609 SYSCTL_INT(_net_iflib, OID_AUTO, tx_frees, CTLFLAG_RD,
610 		   &iflib_tx_frees, 0, "# TX frees");
611 SYSCTL_INT(_net_iflib, OID_AUTO, rx_allocs, CTLFLAG_RD,
612 		   &iflib_rx_allocs, 0, "# RX allocations");
613 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills, CTLFLAG_RD,
614 		   &iflib_fl_refills, 0, "# refills");
615 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills_large, CTLFLAG_RD,
616 		   &iflib_fl_refills_large, 0, "# large refills");
617 
618 static int iflib_txq_drain_flushing;
619 static int iflib_txq_drain_oactive;
620 static int iflib_txq_drain_notready;
621 
622 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_flushing, CTLFLAG_RD,
623 		   &iflib_txq_drain_flushing, 0, "# drain flushes");
624 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_oactive, CTLFLAG_RD,
625 		   &iflib_txq_drain_oactive, 0, "# drain oactives");
626 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_notready, CTLFLAG_RD,
627 		   &iflib_txq_drain_notready, 0, "# drain notready");
628 
629 static int iflib_encap_load_mbuf_fail;
630 static int iflib_encap_pad_mbuf_fail;
631 static int iflib_encap_txq_avail_fail;
632 static int iflib_encap_txd_encap_fail;
633 
634 SYSCTL_INT(_net_iflib, OID_AUTO, encap_load_mbuf_fail, CTLFLAG_RD,
635 		   &iflib_encap_load_mbuf_fail, 0, "# busdma load failures");
636 SYSCTL_INT(_net_iflib, OID_AUTO, encap_pad_mbuf_fail, CTLFLAG_RD,
637 		   &iflib_encap_pad_mbuf_fail, 0, "# runt frame pad failures");
638 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txq_avail_fail, CTLFLAG_RD,
639 		   &iflib_encap_txq_avail_fail, 0, "# txq avail failures");
640 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txd_encap_fail, CTLFLAG_RD,
641 		   &iflib_encap_txd_encap_fail, 0, "# driver encap failures");
642 
643 static int iflib_task_fn_rxs;
644 static int iflib_rx_intr_enables;
645 static int iflib_fast_intrs;
646 static int iflib_rx_unavail;
647 static int iflib_rx_ctx_inactive;
648 static int iflib_rx_if_input;
649 static int iflib_rxd_flush;
650 
651 static int iflib_verbose_debug;
652 
653 SYSCTL_INT(_net_iflib, OID_AUTO, task_fn_rx, CTLFLAG_RD,
654 		   &iflib_task_fn_rxs, 0, "# task_fn_rx calls");
655 SYSCTL_INT(_net_iflib, OID_AUTO, rx_intr_enables, CTLFLAG_RD,
656 		   &iflib_rx_intr_enables, 0, "# RX intr enables");
657 SYSCTL_INT(_net_iflib, OID_AUTO, fast_intrs, CTLFLAG_RD,
658 		   &iflib_fast_intrs, 0, "# fast_intr calls");
659 SYSCTL_INT(_net_iflib, OID_AUTO, rx_unavail, CTLFLAG_RD,
660 		   &iflib_rx_unavail, 0, "# times rxeof called with no available data");
661 SYSCTL_INT(_net_iflib, OID_AUTO, rx_ctx_inactive, CTLFLAG_RD,
662 		   &iflib_rx_ctx_inactive, 0, "# times rxeof called with inactive context");
663 SYSCTL_INT(_net_iflib, OID_AUTO, rx_if_input, CTLFLAG_RD,
664 		   &iflib_rx_if_input, 0, "# times rxeof called if_input");
665 SYSCTL_INT(_net_iflib, OID_AUTO, rxd_flush, CTLFLAG_RD,
666 	         &iflib_rxd_flush, 0, "# times rxd_flush called");
667 SYSCTL_INT(_net_iflib, OID_AUTO, verbose_debug, CTLFLAG_RW,
668 		   &iflib_verbose_debug, 0, "enable verbose debugging");
669 
670 #define DBG_COUNTER_INC(name) atomic_add_int(&(iflib_ ## name), 1)
671 static void
672 iflib_debug_reset(void)
673 {
674 	iflib_tx_seen = iflib_tx_sent = iflib_tx_encap = iflib_rx_allocs =
675 		iflib_fl_refills = iflib_fl_refills_large = iflib_tx_frees =
676 		iflib_txq_drain_flushing = iflib_txq_drain_oactive =
677 		iflib_txq_drain_notready =
678 		iflib_encap_load_mbuf_fail = iflib_encap_pad_mbuf_fail =
679 		iflib_encap_txq_avail_fail = iflib_encap_txd_encap_fail =
680 		iflib_task_fn_rxs = iflib_rx_intr_enables = iflib_fast_intrs =
681 		iflib_rx_unavail =
682 		iflib_rx_ctx_inactive = iflib_rx_if_input =
683 		iflib_rxd_flush = 0;
684 }
685 
686 #else
687 #define DBG_COUNTER_INC(name)
688 static void iflib_debug_reset(void) {}
689 #endif
690 
691 #define IFLIB_DEBUG 0
692 
693 static void iflib_tx_structures_free(if_ctx_t ctx);
694 static void iflib_rx_structures_free(if_ctx_t ctx);
695 static int iflib_queues_alloc(if_ctx_t ctx);
696 static int iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq);
697 static int iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget);
698 static int iflib_qset_structures_setup(if_ctx_t ctx);
699 static int iflib_msix_init(if_ctx_t ctx);
700 static int iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filterarg, int *rid, const char *str);
701 static void iflib_txq_check_drain(iflib_txq_t txq, int budget);
702 static uint32_t iflib_txq_can_drain(struct ifmp_ring *);
703 #ifdef ALTQ
704 static void iflib_altq_if_start(if_t ifp);
705 static int iflib_altq_if_transmit(if_t ifp, struct mbuf *m);
706 #endif
707 static int iflib_register(if_ctx_t);
708 static void iflib_deregister(if_ctx_t);
709 static void iflib_unregister_vlan_handlers(if_ctx_t ctx);
710 static uint16_t iflib_get_mbuf_size_for(unsigned int size);
711 static void iflib_init_locked(if_ctx_t ctx);
712 static void iflib_add_device_sysctl_pre(if_ctx_t ctx);
713 static void iflib_add_device_sysctl_post(if_ctx_t ctx);
714 static void iflib_ifmp_purge(iflib_txq_t txq);
715 static void _iflib_pre_assert(if_softc_ctx_t scctx);
716 static void iflib_if_init_locked(if_ctx_t ctx);
717 static void iflib_free_intr_mem(if_ctx_t ctx);
718 #ifndef __NO_STRICT_ALIGNMENT
719 static struct mbuf * iflib_fixup_rx(struct mbuf *m);
720 #endif
721 
722 static SLIST_HEAD(cpu_offset_list, cpu_offset) cpu_offsets =
723     SLIST_HEAD_INITIALIZER(cpu_offsets);
724 struct cpu_offset {
725 	SLIST_ENTRY(cpu_offset) entries;
726 	cpuset_t	set;
727 	unsigned int	refcount;
728 	uint16_t	offset;
729 };
730 static struct mtx cpu_offset_mtx;
731 MTX_SYSINIT(iflib_cpu_offset, &cpu_offset_mtx, "iflib_cpu_offset lock",
732     MTX_DEF);
733 
734 DEBUGNET_DEFINE(iflib);
735 
736 static int
737 iflib_num_rx_descs(if_ctx_t ctx)
738 {
739 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
740 	if_shared_ctx_t sctx = ctx->ifc_sctx;
741 	uint16_t first_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
742 
743 	return scctx->isc_nrxd[first_rxq];
744 }
745 
746 static int
747 iflib_num_tx_descs(if_ctx_t ctx)
748 {
749 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
750 	if_shared_ctx_t sctx = ctx->ifc_sctx;
751 	uint16_t first_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
752 
753 	return scctx->isc_ntxd[first_txq];
754 }
755 
756 #ifdef DEV_NETMAP
757 #include <sys/selinfo.h>
758 #include <net/netmap.h>
759 #include <dev/netmap/netmap_kern.h>
760 
761 MODULE_DEPEND(iflib, netmap, 1, 1, 1);
762 
763 static int netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, bool init);
764 static void iflib_netmap_timer(void *arg);
765 
766 /*
767  * device-specific sysctl variables:
768  *
769  * iflib_crcstrip: 0: keep CRC in rx frames (default), 1: strip it.
770  *	During regular operations the CRC is stripped, but on some
771  *	hardware reception of frames not multiple of 64 is slower,
772  *	so using crcstrip=0 helps in benchmarks.
773  *
774  * iflib_rx_miss, iflib_rx_miss_bufs:
775  *	count packets that might be missed due to lost interrupts.
776  */
777 SYSCTL_DECL(_dev_netmap);
778 /*
779  * The xl driver by default strips CRCs and we do not override it.
780  */
781 
782 int iflib_crcstrip = 1;
783 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_crcstrip,
784     CTLFLAG_RW, &iflib_crcstrip, 1, "strip CRC on RX frames");
785 
786 int iflib_rx_miss, iflib_rx_miss_bufs;
787 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss,
788     CTLFLAG_RW, &iflib_rx_miss, 0, "potentially missed RX intr");
789 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss_bufs,
790     CTLFLAG_RW, &iflib_rx_miss_bufs, 0, "potentially missed RX intr bufs");
791 
792 /*
793  * Register/unregister. We are already under netmap lock.
794  * Only called on the first register or the last unregister.
795  */
796 static int
797 iflib_netmap_register(struct netmap_adapter *na, int onoff)
798 {
799 	if_t ifp = na->ifp;
800 	if_ctx_t ctx = ifp->if_softc;
801 	int status;
802 
803 	CTX_LOCK(ctx);
804 	if (!CTX_IS_VF(ctx))
805 		IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip);
806 
807 	iflib_stop(ctx);
808 
809 	/*
810 	 * Enable (or disable) netmap flags, and intercept (or restore)
811 	 * ifp->if_transmit. This is done once the device has been stopped
812 	 * to prevent race conditions. Also, this must be done after
813 	 * calling netmap_disable_all_rings() and before calling
814 	 * netmap_enable_all_rings(), so that these two functions see the
815 	 * updated state of the NAF_NETMAP_ON bit.
816 	 */
817 	if (onoff) {
818 		nm_set_native_flags(na);
819 	} else {
820 		nm_clear_native_flags(na);
821 	}
822 
823 	iflib_init_locked(ctx);
824 	IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip); // XXX why twice ?
825 	status = ifp->if_drv_flags & IFF_DRV_RUNNING ? 0 : 1;
826 	if (status)
827 		nm_clear_native_flags(na);
828 	CTX_UNLOCK(ctx);
829 	return (status);
830 }
831 
832 static int
833 netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, bool init)
834 {
835 	struct netmap_adapter *na = kring->na;
836 	u_int const lim = kring->nkr_num_slots - 1;
837 	struct netmap_ring *ring = kring->ring;
838 	bus_dmamap_t *map;
839 	struct if_rxd_update iru;
840 	if_ctx_t ctx = rxq->ifr_ctx;
841 	iflib_fl_t fl = &rxq->ifr_fl[0];
842 	u_int nic_i_first, nic_i;
843 	u_int nm_i;
844 	int i, n;
845 #if IFLIB_DEBUG_COUNTERS
846 	int rf_count = 0;
847 #endif
848 
849 	/*
850 	 * This function is used both at initialization and in rxsync.
851 	 * At initialization we need to prepare (with isc_rxd_refill())
852 	 * all the netmap buffers currently owned by the kernel, in
853 	 * such a way to keep fl->ifl_pidx and kring->nr_hwcur in sync
854 	 * (except for kring->nkr_hwofs). These may be less than
855 	 * kring->nkr_num_slots if netmap_reset() was called while
856 	 * an application using the kring that still owned some
857 	 * buffers.
858 	 * At rxsync time, both indexes point to the next buffer to be
859 	 * refilled.
860 	 * In any case we publish (with isc_rxd_flush()) up to
861 	 * (fl->ifl_pidx - 1) % N (included), to avoid the NIC tail/prod
862 	 * pointer to overrun the head/cons pointer, although this is
863 	 * not necessary for some NICs (e.g. vmx).
864 	 */
865 	if (__predict_false(init)) {
866 		n = kring->nkr_num_slots - nm_kr_rxspace(kring);
867 	} else {
868 		n = kring->rhead - kring->nr_hwcur;
869 		if (n == 0)
870 			return (0); /* Nothing to do. */
871 		if (n < 0)
872 			n += kring->nkr_num_slots;
873 	}
874 
875 	iru_init(&iru, rxq, 0 /* flid */);
876 	map = fl->ifl_sds.ifsd_map;
877 	nic_i = fl->ifl_pidx;
878 	nm_i = netmap_idx_n2k(kring, nic_i);
879 	if (__predict_false(init)) {
880 		/*
881 		 * On init/reset, nic_i must be 0, and we must
882 		 * start to refill from hwtail (see netmap_reset()).
883 		 */
884 		MPASS(nic_i == 0);
885 		MPASS(nm_i == kring->nr_hwtail);
886 	} else
887 		MPASS(nm_i == kring->nr_hwcur);
888 	DBG_COUNTER_INC(fl_refills);
889 	while (n > 0) {
890 #if IFLIB_DEBUG_COUNTERS
891 		if (++rf_count == 9)
892 			DBG_COUNTER_INC(fl_refills_large);
893 #endif
894 		nic_i_first = nic_i;
895 		for (i = 0; n > 0 && i < IFLIB_MAX_RX_REFRESH; n--, i++) {
896 			struct netmap_slot *slot = &ring->slot[nm_i];
897 			void *addr = PNMB(na, slot, &fl->ifl_bus_addrs[i]);
898 
899 			MPASS(i < IFLIB_MAX_RX_REFRESH);
900 
901 			if (addr == NETMAP_BUF_BASE(na)) /* bad buf */
902 			        return netmap_ring_reinit(kring);
903 
904 			fl->ifl_rxd_idxs[i] = nic_i;
905 
906 			if (__predict_false(init)) {
907 				netmap_load_map(na, fl->ifl_buf_tag,
908 				    map[nic_i], addr);
909 			} else if (slot->flags & NS_BUF_CHANGED) {
910 				/* buffer has changed, reload map */
911 				netmap_reload_map(na, fl->ifl_buf_tag,
912 				    map[nic_i], addr);
913 			}
914 			bus_dmamap_sync(fl->ifl_buf_tag, map[nic_i],
915 			    BUS_DMASYNC_PREREAD);
916 			slot->flags &= ~NS_BUF_CHANGED;
917 
918 			nm_i = nm_next(nm_i, lim);
919 			nic_i = nm_next(nic_i, lim);
920 		}
921 
922 		iru.iru_pidx = nic_i_first;
923 		iru.iru_count = i;
924 		ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
925 	}
926 	fl->ifl_pidx = nic_i;
927 	/*
928 	 * At the end of the loop we must have refilled everything
929 	 * we could possibly refill.
930 	 */
931 	MPASS(nm_i == kring->rhead);
932 	kring->nr_hwcur = nm_i;
933 
934 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
935 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
936 	ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id,
937 	    nm_prev(nic_i, lim));
938 	DBG_COUNTER_INC(rxd_flush);
939 
940 	return (0);
941 }
942 
943 #define NETMAP_TX_TIMER_US	90
944 
945 /*
946  * Reconcile kernel and user view of the transmit ring.
947  *
948  * All information is in the kring.
949  * Userspace wants to send packets up to the one before kring->rhead,
950  * kernel knows kring->nr_hwcur is the first unsent packet.
951  *
952  * Here we push packets out (as many as possible), and possibly
953  * reclaim buffers from previously completed transmission.
954  *
955  * The caller (netmap) guarantees that there is only one instance
956  * running at any time. Any interference with other driver
957  * methods should be handled by the individual drivers.
958  */
959 static int
960 iflib_netmap_txsync(struct netmap_kring *kring, int flags)
961 {
962 	struct netmap_adapter *na = kring->na;
963 	if_t ifp = na->ifp;
964 	struct netmap_ring *ring = kring->ring;
965 	u_int nm_i;	/* index into the netmap kring */
966 	u_int nic_i;	/* index into the NIC ring */
967 	u_int n;
968 	u_int const lim = kring->nkr_num_slots - 1;
969 	u_int const head = kring->rhead;
970 	struct if_pkt_info pi;
971 
972 	/*
973 	 * interrupts on every tx packet are expensive so request
974 	 * them every half ring, or where NS_REPORT is set
975 	 */
976 	u_int report_frequency = kring->nkr_num_slots >> 1;
977 	/* device-specific */
978 	if_ctx_t ctx = ifp->if_softc;
979 	iflib_txq_t txq = &ctx->ifc_txqs[kring->ring_id];
980 
981 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
982 	    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
983 
984 	/*
985 	 * First part: process new packets to send.
986 	 * nm_i is the current index in the netmap kring,
987 	 * nic_i is the corresponding index in the NIC ring.
988 	 *
989 	 * If we have packets to send (nm_i != head)
990 	 * iterate over the netmap ring, fetch length and update
991 	 * the corresponding slot in the NIC ring. Some drivers also
992 	 * need to update the buffer's physical address in the NIC slot
993 	 * even NS_BUF_CHANGED is not set (PNMB computes the addresses).
994 	 *
995 	 * The netmap_reload_map() calls is especially expensive,
996 	 * even when (as in this case) the tag is 0, so do only
997 	 * when the buffer has actually changed.
998 	 *
999 	 * If possible do not set the report/intr bit on all slots,
1000 	 * but only a few times per ring or when NS_REPORT is set.
1001 	 *
1002 	 * Finally, on 10G and faster drivers, it might be useful
1003 	 * to prefetch the next slot and txr entry.
1004 	 */
1005 
1006 	nm_i = kring->nr_hwcur;
1007 	if (nm_i != head) {	/* we have new packets to send */
1008 		uint32_t pkt_len = 0, seg_idx = 0;
1009 		int nic_i_start = -1, flags = 0;
1010 		pkt_info_zero(&pi);
1011 		pi.ipi_segs = txq->ift_segs;
1012 		pi.ipi_qsidx = kring->ring_id;
1013 		nic_i = netmap_idx_k2n(kring, nm_i);
1014 
1015 		__builtin_prefetch(&ring->slot[nm_i]);
1016 		__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i]);
1017 		__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i]);
1018 
1019 		for (n = 0; nm_i != head; n++) {
1020 			struct netmap_slot *slot = &ring->slot[nm_i];
1021 			u_int len = slot->len;
1022 			uint64_t paddr;
1023 			void *addr = PNMB(na, slot, &paddr);
1024 
1025 			flags |= (slot->flags & NS_REPORT ||
1026 				nic_i == 0 || nic_i == report_frequency) ?
1027 				IPI_TX_INTR : 0;
1028 
1029 			/*
1030 			 * If this is the first packet fragment, save the
1031 			 * index of the first NIC slot for later.
1032 			 */
1033 			if (nic_i_start < 0)
1034 				nic_i_start = nic_i;
1035 
1036 			pi.ipi_segs[seg_idx].ds_addr = paddr;
1037 			pi.ipi_segs[seg_idx].ds_len = len;
1038 			if (len) {
1039 				pkt_len += len;
1040 				seg_idx++;
1041 			}
1042 
1043 			if (!(slot->flags & NS_MOREFRAG)) {
1044 				pi.ipi_len = pkt_len;
1045 				pi.ipi_nsegs = seg_idx;
1046 				pi.ipi_pidx = nic_i_start;
1047 				pi.ipi_ndescs = 0;
1048 				pi.ipi_flags = flags;
1049 
1050 				/* Prepare the NIC TX ring. */
1051 				ctx->isc_txd_encap(ctx->ifc_softc, &pi);
1052 				DBG_COUNTER_INC(tx_encap);
1053 
1054 				/* Reinit per-packet info for the next one. */
1055 				flags = seg_idx = pkt_len = 0;
1056 				nic_i_start = -1;
1057 			}
1058 
1059 			/* prefetch for next round */
1060 			__builtin_prefetch(&ring->slot[nm_i + 1]);
1061 			__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i + 1]);
1062 			__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i + 1]);
1063 
1064 			NM_CHECK_ADDR_LEN(na, addr, len);
1065 
1066 			if (slot->flags & NS_BUF_CHANGED) {
1067 				/* buffer has changed, reload map */
1068 				netmap_reload_map(na, txq->ift_buf_tag,
1069 				    txq->ift_sds.ifsd_map[nic_i], addr);
1070 			}
1071 			/* make sure changes to the buffer are synced */
1072 			bus_dmamap_sync(txq->ift_buf_tag,
1073 			    txq->ift_sds.ifsd_map[nic_i],
1074 			    BUS_DMASYNC_PREWRITE);
1075 
1076 			slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED | NS_MOREFRAG);
1077 			nm_i = nm_next(nm_i, lim);
1078 			nic_i = nm_next(nic_i, lim);
1079 		}
1080 		kring->nr_hwcur = nm_i;
1081 
1082 		/* synchronize the NIC ring */
1083 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1084 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1085 
1086 		/* (re)start the tx unit up to slot nic_i (excluded) */
1087 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, nic_i);
1088 	}
1089 
1090 	/*
1091 	 * Second part: reclaim buffers for completed transmissions.
1092 	 *
1093 	 * If there are unclaimed buffers, attempt to reclaim them.
1094 	 * If we don't manage to reclaim them all, and TX IRQs are not in use,
1095 	 * trigger a per-tx-queue timer to try again later.
1096 	 */
1097 	if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1098 		if (iflib_tx_credits_update(ctx, txq)) {
1099 			/* some tx completed, increment avail */
1100 			nic_i = txq->ift_cidx_processed;
1101 			kring->nr_hwtail = nm_prev(netmap_idx_n2k(kring, nic_i), lim);
1102 		}
1103 	}
1104 
1105 	if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ))
1106 		if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1107 			callout_reset_sbt_on(&txq->ift_netmap_timer,
1108 			    NETMAP_TX_TIMER_US * SBT_1US, SBT_1US,
1109 			    iflib_netmap_timer, txq,
1110 			    txq->ift_netmap_timer.c_cpu, 0);
1111 		}
1112 	return (0);
1113 }
1114 
1115 /*
1116  * Reconcile kernel and user view of the receive ring.
1117  * Same as for the txsync, this routine must be efficient.
1118  * The caller guarantees a single invocations, but races against
1119  * the rest of the driver should be handled here.
1120  *
1121  * On call, kring->rhead is the first packet that userspace wants
1122  * to keep, and kring->rcur is the wakeup point.
1123  * The kernel has previously reported packets up to kring->rtail.
1124  *
1125  * If (flags & NAF_FORCE_READ) also check for incoming packets irrespective
1126  * of whether or not we received an interrupt.
1127  */
1128 static int
1129 iflib_netmap_rxsync(struct netmap_kring *kring, int flags)
1130 {
1131 	struct netmap_adapter *na = kring->na;
1132 	struct netmap_ring *ring = kring->ring;
1133 	if_t ifp = na->ifp;
1134 	uint32_t nm_i;	/* index into the netmap ring */
1135 	uint32_t nic_i;	/* index into the NIC ring */
1136 	u_int n;
1137 	u_int const lim = kring->nkr_num_slots - 1;
1138 	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
1139 	int i = 0;
1140 
1141 	if_ctx_t ctx = ifp->if_softc;
1142 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1143 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1144 	iflib_rxq_t rxq = &ctx->ifc_rxqs[kring->ring_id];
1145 	iflib_fl_t fl = &rxq->ifr_fl[0];
1146 	struct if_rxd_info ri;
1147 	qidx_t *cidxp;
1148 
1149 	/*
1150 	 * netmap only uses free list 0, to avoid out of order consumption
1151 	 * of receive buffers
1152 	 */
1153 
1154 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
1155 	    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1156 
1157 	/*
1158 	 * First part: import newly received packets.
1159 	 *
1160 	 * nm_i is the index of the next free slot in the netmap ring,
1161 	 * nic_i is the index of the next received packet in the NIC ring
1162 	 * (or in the free list 0 if IFLIB_HAS_RXCQ is set), and they may
1163 	 * differ in case if_init() has been called while
1164 	 * in netmap mode. For the receive ring we have
1165 	 *
1166 	 *	nic_i = fl->ifl_cidx;
1167 	 *	nm_i = kring->nr_hwtail (previous)
1168 	 * and
1169 	 *	nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1170 	 *
1171 	 * fl->ifl_cidx is set to 0 on a ring reinit
1172 	 */
1173 	if (netmap_no_pendintr || force_update) {
1174 		uint32_t hwtail_lim = nm_prev(kring->nr_hwcur, lim);
1175 		bool have_rxcq = sctx->isc_flags & IFLIB_HAS_RXCQ;
1176 		int crclen = iflib_crcstrip ? 0 : 4;
1177 		int error, avail;
1178 
1179 		/*
1180 		 * For the free list consumer index, we use the same
1181 		 * logic as in iflib_rxeof().
1182 		 */
1183 		if (have_rxcq)
1184 			cidxp = &rxq->ifr_cq_cidx;
1185 		else
1186 			cidxp = &fl->ifl_cidx;
1187 		avail = ctx->isc_rxd_available(ctx->ifc_softc,
1188 		    rxq->ifr_id, *cidxp, USHRT_MAX);
1189 
1190 		nic_i = fl->ifl_cidx;
1191 		nm_i = netmap_idx_n2k(kring, nic_i);
1192 		MPASS(nm_i == kring->nr_hwtail);
1193 		for (n = 0; avail > 0 && nm_i != hwtail_lim; n++, avail--) {
1194 			rxd_info_zero(&ri);
1195 			ri.iri_frags = rxq->ifr_frags;
1196 			ri.iri_qsidx = kring->ring_id;
1197 			ri.iri_ifp = ctx->ifc_ifp;
1198 			ri.iri_cidx = *cidxp;
1199 
1200 			error = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
1201 			for (i = 0; i < ri.iri_nfrags; i++) {
1202 				if (error) {
1203 					ring->slot[nm_i].len = 0;
1204 					ring->slot[nm_i].flags = 0;
1205 				} else {
1206 					ring->slot[nm_i].len = ri.iri_frags[i].irf_len;
1207 					if (i == (ri.iri_nfrags - 1)) {
1208 						ring->slot[nm_i].len -= crclen;
1209 						ring->slot[nm_i].flags = 0;
1210 					} else
1211 						ring->slot[nm_i].flags = NS_MOREFRAG;
1212 				}
1213 
1214 				bus_dmamap_sync(fl->ifl_buf_tag,
1215 				    fl->ifl_sds.ifsd_map[nic_i], BUS_DMASYNC_POSTREAD);
1216 				nm_i = nm_next(nm_i, lim);
1217 				fl->ifl_cidx = nic_i = nm_next(nic_i, lim);
1218 			}
1219 
1220 			if (have_rxcq) {
1221 				*cidxp = ri.iri_cidx;
1222 				while (*cidxp >= scctx->isc_nrxd[0])
1223 					*cidxp -= scctx->isc_nrxd[0];
1224 			}
1225 
1226 		}
1227 		if (n) { /* update the state variables */
1228 			if (netmap_no_pendintr && !force_update) {
1229 				/* diagnostics */
1230 				iflib_rx_miss ++;
1231 				iflib_rx_miss_bufs += n;
1232 			}
1233 			kring->nr_hwtail = nm_i;
1234 		}
1235 		kring->nr_kflags &= ~NKR_PENDINTR;
1236 	}
1237 	/*
1238 	 * Second part: skip past packets that userspace has released.
1239 	 * (kring->nr_hwcur to head excluded),
1240 	 * and make the buffers available for reception.
1241 	 * As usual nm_i is the index in the netmap ring,
1242 	 * nic_i is the index in the NIC ring, and
1243 	 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1244 	 */
1245 	netmap_fl_refill(rxq, kring, false);
1246 
1247 	return (0);
1248 }
1249 
1250 static void
1251 iflib_netmap_intr(struct netmap_adapter *na, int onoff)
1252 {
1253 	if_ctx_t ctx = na->ifp->if_softc;
1254 
1255 	CTX_LOCK(ctx);
1256 	if (onoff) {
1257 		IFDI_INTR_ENABLE(ctx);
1258 	} else {
1259 		IFDI_INTR_DISABLE(ctx);
1260 	}
1261 	CTX_UNLOCK(ctx);
1262 }
1263 
1264 static int
1265 iflib_netmap_attach(if_ctx_t ctx)
1266 {
1267 	struct netmap_adapter na;
1268 
1269 	bzero(&na, sizeof(na));
1270 
1271 	na.ifp = ctx->ifc_ifp;
1272 	na.na_flags = NAF_BDG_MAYSLEEP | NAF_MOREFRAG;
1273 	MPASS(ctx->ifc_softc_ctx.isc_ntxqsets);
1274 	MPASS(ctx->ifc_softc_ctx.isc_nrxqsets);
1275 
1276 	na.num_tx_desc = iflib_num_tx_descs(ctx);
1277 	na.num_rx_desc = iflib_num_rx_descs(ctx);
1278 	na.nm_txsync = iflib_netmap_txsync;
1279 	na.nm_rxsync = iflib_netmap_rxsync;
1280 	na.nm_register = iflib_netmap_register;
1281 	na.nm_intr = iflib_netmap_intr;
1282 	na.num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
1283 	na.num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
1284 	return (netmap_attach(&na));
1285 }
1286 
1287 static int
1288 iflib_netmap_txq_init(if_ctx_t ctx, iflib_txq_t txq)
1289 {
1290 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1291 	struct netmap_slot *slot;
1292 
1293 	slot = netmap_reset(na, NR_TX, txq->ift_id, 0);
1294 	if (slot == NULL)
1295 		return (0);
1296 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxd[0]; i++) {
1297 		/*
1298 		 * In netmap mode, set the map for the packet buffer.
1299 		 * NOTE: Some drivers (not this one) also need to set
1300 		 * the physical buffer address in the NIC ring.
1301 		 * netmap_idx_n2k() maps a nic index, i, into the corresponding
1302 		 * netmap slot index, si
1303 		 */
1304 		int si = netmap_idx_n2k(na->tx_rings[txq->ift_id], i);
1305 		netmap_load_map(na, txq->ift_buf_tag, txq->ift_sds.ifsd_map[i],
1306 		    NMB(na, slot + si));
1307 	}
1308 	return (1);
1309 }
1310 
1311 static int
1312 iflib_netmap_rxq_init(if_ctx_t ctx, iflib_rxq_t rxq)
1313 {
1314 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1315 	struct netmap_kring *kring;
1316 	struct netmap_slot *slot;
1317 
1318 	slot = netmap_reset(na, NR_RX, rxq->ifr_id, 0);
1319 	if (slot == NULL)
1320 		return (0);
1321 	kring = na->rx_rings[rxq->ifr_id];
1322 	netmap_fl_refill(rxq, kring, true);
1323 	return (1);
1324 }
1325 
1326 static void
1327 iflib_netmap_timer(void *arg)
1328 {
1329 	iflib_txq_t txq = arg;
1330 	if_ctx_t ctx = txq->ift_ctx;
1331 
1332 	/*
1333 	 * Wake up the netmap application, to give it a chance to
1334 	 * call txsync and reclaim more completed TX buffers.
1335 	 */
1336 	netmap_tx_irq(ctx->ifc_ifp, txq->ift_id);
1337 }
1338 
1339 #define iflib_netmap_detach(ifp) netmap_detach(ifp)
1340 
1341 #else
1342 #define iflib_netmap_txq_init(ctx, txq) (0)
1343 #define iflib_netmap_rxq_init(ctx, rxq) (0)
1344 #define iflib_netmap_detach(ifp)
1345 #define netmap_enable_all_rings(ifp)
1346 #define netmap_disable_all_rings(ifp)
1347 
1348 #define iflib_netmap_attach(ctx) (0)
1349 #define netmap_rx_irq(ifp, qid, budget) (0)
1350 #endif
1351 
1352 #if defined(__i386__) || defined(__amd64__)
1353 static __inline void
1354 prefetch(void *x)
1355 {
1356 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1357 }
1358 static __inline void
1359 prefetch2cachelines(void *x)
1360 {
1361 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1362 #if (CACHE_LINE_SIZE < 128)
1363 	__asm volatile("prefetcht0 %0" :: "m" (*(((unsigned long *)x)+CACHE_LINE_SIZE/(sizeof(unsigned long)))));
1364 #endif
1365 }
1366 #else
1367 #define prefetch(x)
1368 #define prefetch2cachelines(x)
1369 #endif
1370 
1371 static void
1372 iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid)
1373 {
1374 	iflib_fl_t fl;
1375 
1376 	fl = &rxq->ifr_fl[flid];
1377 	iru->iru_paddrs = fl->ifl_bus_addrs;
1378 	iru->iru_idxs = fl->ifl_rxd_idxs;
1379 	iru->iru_qsidx = rxq->ifr_id;
1380 	iru->iru_buf_size = fl->ifl_buf_size;
1381 	iru->iru_flidx = fl->ifl_id;
1382 }
1383 
1384 static void
1385 _iflib_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int err)
1386 {
1387 	if (err)
1388 		return;
1389 	*(bus_addr_t *) arg = segs[0].ds_addr;
1390 }
1391 
1392 #define	DMA_WIDTH_TO_BUS_LOWADDR(width)				\
1393 	(((width) == 0) || (width) == flsll(BUS_SPACE_MAXADDR) ?	\
1394 	    BUS_SPACE_MAXADDR : (1ULL << (width)) - 1ULL)
1395 
1396 int
1397 iflib_dma_alloc_align(if_ctx_t ctx, int size, int align, iflib_dma_info_t dma, int mapflags)
1398 {
1399 	int err;
1400 	device_t dev = ctx->ifc_dev;
1401 	bus_addr_t lowaddr;
1402 
1403 	lowaddr = DMA_WIDTH_TO_BUS_LOWADDR(ctx->ifc_softc_ctx.isc_dma_width);
1404 
1405 	err = bus_dma_tag_create(bus_get_dma_tag(dev),	/* parent */
1406 				align, 0,		/* alignment, bounds */
1407 				lowaddr,		/* lowaddr */
1408 				BUS_SPACE_MAXADDR,	/* highaddr */
1409 				NULL, NULL,		/* filter, filterarg */
1410 				size,			/* maxsize */
1411 				1,			/* nsegments */
1412 				size,			/* maxsegsize */
1413 				BUS_DMA_ALLOCNOW,	/* flags */
1414 				NULL,			/* lockfunc */
1415 				NULL,			/* lockarg */
1416 				&dma->idi_tag);
1417 	if (err) {
1418 		device_printf(dev,
1419 		    "%s: bus_dma_tag_create failed: %d\n",
1420 		    __func__, err);
1421 		goto fail_0;
1422 	}
1423 
1424 	err = bus_dmamem_alloc(dma->idi_tag, (void**) &dma->idi_vaddr,
1425 	    BUS_DMA_NOWAIT | BUS_DMA_COHERENT | BUS_DMA_ZERO, &dma->idi_map);
1426 	if (err) {
1427 		device_printf(dev,
1428 		    "%s: bus_dmamem_alloc(%ju) failed: %d\n",
1429 		    __func__, (uintmax_t)size, err);
1430 		goto fail_1;
1431 	}
1432 
1433 	dma->idi_paddr = IF_BAD_DMA;
1434 	err = bus_dmamap_load(dma->idi_tag, dma->idi_map, dma->idi_vaddr,
1435 	    size, _iflib_dmamap_cb, &dma->idi_paddr, mapflags | BUS_DMA_NOWAIT);
1436 	if (err || dma->idi_paddr == IF_BAD_DMA) {
1437 		device_printf(dev,
1438 		    "%s: bus_dmamap_load failed: %d\n",
1439 		    __func__, err);
1440 		goto fail_2;
1441 	}
1442 
1443 	dma->idi_size = size;
1444 	return (0);
1445 
1446 fail_2:
1447 	bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1448 fail_1:
1449 	bus_dma_tag_destroy(dma->idi_tag);
1450 fail_0:
1451 	dma->idi_tag = NULL;
1452 
1453 	return (err);
1454 }
1455 
1456 int
1457 iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags)
1458 {
1459 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1460 
1461 	KASSERT(sctx->isc_q_align != 0, ("alignment value not initialized"));
1462 
1463 	return (iflib_dma_alloc_align(ctx, size, sctx->isc_q_align, dma, mapflags));
1464 }
1465 
1466 int
1467 iflib_dma_alloc_multi(if_ctx_t ctx, int *sizes, iflib_dma_info_t *dmalist, int mapflags, int count)
1468 {
1469 	int i, err;
1470 	iflib_dma_info_t *dmaiter;
1471 
1472 	dmaiter = dmalist;
1473 	for (i = 0; i < count; i++, dmaiter++) {
1474 		if ((err = iflib_dma_alloc(ctx, sizes[i], *dmaiter, mapflags)) != 0)
1475 			break;
1476 	}
1477 	if (err)
1478 		iflib_dma_free_multi(dmalist, i);
1479 	return (err);
1480 }
1481 
1482 void
1483 iflib_dma_free(iflib_dma_info_t dma)
1484 {
1485 	if (dma->idi_tag == NULL)
1486 		return;
1487 	if (dma->idi_paddr != IF_BAD_DMA) {
1488 		bus_dmamap_sync(dma->idi_tag, dma->idi_map,
1489 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1490 		bus_dmamap_unload(dma->idi_tag, dma->idi_map);
1491 		dma->idi_paddr = IF_BAD_DMA;
1492 	}
1493 	if (dma->idi_vaddr != NULL) {
1494 		bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1495 		dma->idi_vaddr = NULL;
1496 	}
1497 	bus_dma_tag_destroy(dma->idi_tag);
1498 	dma->idi_tag = NULL;
1499 }
1500 
1501 void
1502 iflib_dma_free_multi(iflib_dma_info_t *dmalist, int count)
1503 {
1504 	int i;
1505 	iflib_dma_info_t *dmaiter = dmalist;
1506 
1507 	for (i = 0; i < count; i++, dmaiter++)
1508 		iflib_dma_free(*dmaiter);
1509 }
1510 
1511 static int
1512 iflib_fast_intr(void *arg)
1513 {
1514 	iflib_filter_info_t info = arg;
1515 	struct grouptask *gtask = info->ifi_task;
1516 	int result;
1517 
1518 	DBG_COUNTER_INC(fast_intrs);
1519 	if (info->ifi_filter != NULL) {
1520 		result = info->ifi_filter(info->ifi_filter_arg);
1521 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1522 			return (result);
1523 	}
1524 
1525 	GROUPTASK_ENQUEUE(gtask);
1526 	return (FILTER_HANDLED);
1527 }
1528 
1529 static int
1530 iflib_fast_intr_rxtx(void *arg)
1531 {
1532 	iflib_filter_info_t info = arg;
1533 	struct grouptask *gtask = info->ifi_task;
1534 	if_ctx_t ctx;
1535 	iflib_rxq_t rxq = (iflib_rxq_t)info->ifi_ctx;
1536 	iflib_txq_t txq;
1537 	void *sc;
1538 	int i, cidx, result;
1539 	qidx_t txqid;
1540 	bool intr_enable, intr_legacy;
1541 
1542 	DBG_COUNTER_INC(fast_intrs);
1543 	if (info->ifi_filter != NULL) {
1544 		result = info->ifi_filter(info->ifi_filter_arg);
1545 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1546 			return (result);
1547 	}
1548 
1549 	ctx = rxq->ifr_ctx;
1550 	sc = ctx->ifc_softc;
1551 	intr_enable = false;
1552 	intr_legacy = !!(ctx->ifc_flags & IFC_LEGACY);
1553 	MPASS(rxq->ifr_ntxqirq);
1554 	for (i = 0; i < rxq->ifr_ntxqirq; i++) {
1555 		txqid = rxq->ifr_txqid[i];
1556 		txq = &ctx->ifc_txqs[txqid];
1557 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1558 		    BUS_DMASYNC_POSTREAD);
1559 		if (!ctx->isc_txd_credits_update(sc, txqid, false)) {
1560 			if (intr_legacy)
1561 				intr_enable = true;
1562 			else
1563 				IFDI_TX_QUEUE_INTR_ENABLE(ctx, txqid);
1564 			continue;
1565 		}
1566 		GROUPTASK_ENQUEUE(&txq->ift_task);
1567 	}
1568 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_RXCQ)
1569 		cidx = rxq->ifr_cq_cidx;
1570 	else
1571 		cidx = rxq->ifr_fl[0].ifl_cidx;
1572 	if (iflib_rxd_avail(ctx, rxq, cidx, 1))
1573 		GROUPTASK_ENQUEUE(gtask);
1574 	else {
1575 		if (intr_legacy)
1576 			intr_enable = true;
1577 		else
1578 			IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
1579 		DBG_COUNTER_INC(rx_intr_enables);
1580 	}
1581 	if (intr_enable)
1582 		IFDI_INTR_ENABLE(ctx);
1583 	return (FILTER_HANDLED);
1584 }
1585 
1586 static int
1587 iflib_fast_intr_ctx(void *arg)
1588 {
1589 	iflib_filter_info_t info = arg;
1590 	struct grouptask *gtask = info->ifi_task;
1591 	int result;
1592 
1593 	DBG_COUNTER_INC(fast_intrs);
1594 	if (info->ifi_filter != NULL) {
1595 		result = info->ifi_filter(info->ifi_filter_arg);
1596 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1597 			return (result);
1598 	}
1599 
1600 	GROUPTASK_ENQUEUE(gtask);
1601 	return (FILTER_HANDLED);
1602 }
1603 
1604 static int
1605 _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
1606 		 driver_filter_t filter, driver_intr_t handler, void *arg,
1607 		 const char *name)
1608 {
1609 	struct resource *res;
1610 	void *tag = NULL;
1611 	device_t dev = ctx->ifc_dev;
1612 	int flags, i, rc;
1613 
1614 	flags = RF_ACTIVE;
1615 	if (ctx->ifc_flags & IFC_LEGACY)
1616 		flags |= RF_SHAREABLE;
1617 	MPASS(rid < 512);
1618 	i = rid;
1619 	res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &i, flags);
1620 	if (res == NULL) {
1621 		device_printf(dev,
1622 		    "failed to allocate IRQ for rid %d, name %s.\n", rid, name);
1623 		return (ENOMEM);
1624 	}
1625 	irq->ii_res = res;
1626 	KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL"));
1627 	rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET,
1628 						filter, handler, arg, &tag);
1629 	if (rc != 0) {
1630 		device_printf(dev,
1631 		    "failed to setup interrupt for rid %d, name %s: %d\n",
1632 					  rid, name ? name : "unknown", rc);
1633 		return (rc);
1634 	} else if (name)
1635 		bus_describe_intr(dev, res, tag, "%s", name);
1636 
1637 	irq->ii_tag = tag;
1638 	return (0);
1639 }
1640 
1641 /*********************************************************************
1642  *
1643  *  Allocate DMA resources for TX buffers as well as memory for the TX
1644  *  mbuf map.  TX DMA maps (non-TSO/TSO) and TX mbuf map are kept in a
1645  *  iflib_sw_tx_desc_array structure, storing all the information that
1646  *  is needed to transmit a packet on the wire.  This is called only
1647  *  once at attach, setup is done every reset.
1648  *
1649  **********************************************************************/
1650 static int
1651 iflib_txsd_alloc(iflib_txq_t txq)
1652 {
1653 	if_ctx_t ctx = txq->ift_ctx;
1654 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1655 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1656 	device_t dev = ctx->ifc_dev;
1657 	bus_size_t tsomaxsize;
1658 	bus_addr_t lowaddr;
1659 	int err, nsegments, ntsosegments;
1660 	bool tso;
1661 
1662 	nsegments = scctx->isc_tx_nsegments;
1663 	ntsosegments = scctx->isc_tx_tso_segments_max;
1664 	tsomaxsize = scctx->isc_tx_tso_size_max;
1665 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_VLAN_MTU)
1666 		tsomaxsize += sizeof(struct ether_vlan_header);
1667 	MPASS(scctx->isc_ntxd[0] > 0);
1668 	MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0);
1669 	MPASS(nsegments > 0);
1670 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) {
1671 		MPASS(ntsosegments > 0);
1672 		MPASS(sctx->isc_tso_maxsize >= tsomaxsize);
1673 	}
1674 
1675 	lowaddr = DMA_WIDTH_TO_BUS_LOWADDR(scctx->isc_dma_width);
1676 
1677 	/*
1678 	 * Set up DMA tags for TX buffers.
1679 	 */
1680 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1681 			       1, 0,			/* alignment, bounds */
1682 			       lowaddr,			/* lowaddr */
1683 			       BUS_SPACE_MAXADDR,	/* highaddr */
1684 			       NULL, NULL,		/* filter, filterarg */
1685 			       sctx->isc_tx_maxsize,		/* maxsize */
1686 			       nsegments,	/* nsegments */
1687 			       sctx->isc_tx_maxsegsize,	/* maxsegsize */
1688 			       0,			/* flags */
1689 			       NULL,			/* lockfunc */
1690 			       NULL,			/* lockfuncarg */
1691 			       &txq->ift_buf_tag))) {
1692 		device_printf(dev,"Unable to allocate TX DMA tag: %d\n", err);
1693 		device_printf(dev,"maxsize: %ju nsegments: %d maxsegsize: %ju\n",
1694 		    (uintmax_t)sctx->isc_tx_maxsize, nsegments, (uintmax_t)sctx->isc_tx_maxsegsize);
1695 		goto fail;
1696 	}
1697 	tso = (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) != 0;
1698 	if (tso && (err = bus_dma_tag_create(bus_get_dma_tag(dev),
1699 			       1, 0,			/* alignment, bounds */
1700 			       lowaddr,			/* lowaddr */
1701 			       BUS_SPACE_MAXADDR,	/* highaddr */
1702 			       NULL, NULL,		/* filter, filterarg */
1703 			       tsomaxsize,		/* maxsize */
1704 			       ntsosegments,	/* nsegments */
1705 			       sctx->isc_tso_maxsegsize,/* maxsegsize */
1706 			       0,			/* flags */
1707 			       NULL,			/* lockfunc */
1708 			       NULL,			/* lockfuncarg */
1709 			       &txq->ift_tso_buf_tag))) {
1710 		device_printf(dev, "Unable to allocate TSO TX DMA tag: %d\n",
1711 		    err);
1712 		goto fail;
1713 	}
1714 
1715 	/* Allocate memory for the TX mbuf map. */
1716 	if (!(txq->ift_sds.ifsd_m =
1717 	    (struct mbuf **) malloc(sizeof(struct mbuf *) *
1718 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1719 		device_printf(dev, "Unable to allocate TX mbuf map memory\n");
1720 		err = ENOMEM;
1721 		goto fail;
1722 	}
1723 
1724 	/*
1725 	 * Create the DMA maps for TX buffers.
1726 	 */
1727 	if ((txq->ift_sds.ifsd_map = (bus_dmamap_t *)malloc(
1728 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1729 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1730 		device_printf(dev,
1731 		    "Unable to allocate TX buffer DMA map memory\n");
1732 		err = ENOMEM;
1733 		goto fail;
1734 	}
1735 	if (tso && (txq->ift_sds.ifsd_tso_map = (bus_dmamap_t *)malloc(
1736 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1737 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1738 		device_printf(dev,
1739 		    "Unable to allocate TSO TX buffer map memory\n");
1740 		err = ENOMEM;
1741 		goto fail;
1742 	}
1743 	for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) {
1744 		err = bus_dmamap_create(txq->ift_buf_tag, 0,
1745 		    &txq->ift_sds.ifsd_map[i]);
1746 		if (err != 0) {
1747 			device_printf(dev, "Unable to create TX DMA map\n");
1748 			goto fail;
1749 		}
1750 		if (!tso)
1751 			continue;
1752 		err = bus_dmamap_create(txq->ift_tso_buf_tag, 0,
1753 		    &txq->ift_sds.ifsd_tso_map[i]);
1754 		if (err != 0) {
1755 			device_printf(dev, "Unable to create TSO TX DMA map\n");
1756 			goto fail;
1757 		}
1758 	}
1759 	return (0);
1760 fail:
1761 	/* We free all, it handles case where we are in the middle */
1762 	iflib_tx_structures_free(ctx);
1763 	return (err);
1764 }
1765 
1766 static void
1767 iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i)
1768 {
1769 	bus_dmamap_t map;
1770 
1771 	if (txq->ift_sds.ifsd_map != NULL) {
1772 		map = txq->ift_sds.ifsd_map[i];
1773 		bus_dmamap_sync(txq->ift_buf_tag, map, BUS_DMASYNC_POSTWRITE);
1774 		bus_dmamap_unload(txq->ift_buf_tag, map);
1775 		bus_dmamap_destroy(txq->ift_buf_tag, map);
1776 		txq->ift_sds.ifsd_map[i] = NULL;
1777 	}
1778 
1779 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1780 		map = txq->ift_sds.ifsd_tso_map[i];
1781 		bus_dmamap_sync(txq->ift_tso_buf_tag, map,
1782 		    BUS_DMASYNC_POSTWRITE);
1783 		bus_dmamap_unload(txq->ift_tso_buf_tag, map);
1784 		bus_dmamap_destroy(txq->ift_tso_buf_tag, map);
1785 		txq->ift_sds.ifsd_tso_map[i] = NULL;
1786 	}
1787 }
1788 
1789 static void
1790 iflib_txq_destroy(iflib_txq_t txq)
1791 {
1792 	if_ctx_t ctx = txq->ift_ctx;
1793 
1794 	for (int i = 0; i < txq->ift_size; i++)
1795 		iflib_txsd_destroy(ctx, txq, i);
1796 
1797 	if (txq->ift_br != NULL) {
1798 		ifmp_ring_free(txq->ift_br);
1799 		txq->ift_br = NULL;
1800 	}
1801 
1802 	mtx_destroy(&txq->ift_mtx);
1803 
1804 	if (txq->ift_sds.ifsd_map != NULL) {
1805 		free(txq->ift_sds.ifsd_map, M_IFLIB);
1806 		txq->ift_sds.ifsd_map = NULL;
1807 	}
1808 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1809 		free(txq->ift_sds.ifsd_tso_map, M_IFLIB);
1810 		txq->ift_sds.ifsd_tso_map = NULL;
1811 	}
1812 	if (txq->ift_sds.ifsd_m != NULL) {
1813 		free(txq->ift_sds.ifsd_m, M_IFLIB);
1814 		txq->ift_sds.ifsd_m = NULL;
1815 	}
1816 	if (txq->ift_buf_tag != NULL) {
1817 		bus_dma_tag_destroy(txq->ift_buf_tag);
1818 		txq->ift_buf_tag = NULL;
1819 	}
1820 	if (txq->ift_tso_buf_tag != NULL) {
1821 		bus_dma_tag_destroy(txq->ift_tso_buf_tag);
1822 		txq->ift_tso_buf_tag = NULL;
1823 	}
1824 	if (txq->ift_ifdi != NULL) {
1825 		free(txq->ift_ifdi, M_IFLIB);
1826 	}
1827 }
1828 
1829 static void
1830 iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i)
1831 {
1832 	struct mbuf **mp;
1833 
1834 	mp = &txq->ift_sds.ifsd_m[i];
1835 	if (*mp == NULL)
1836 		return;
1837 
1838 	if (txq->ift_sds.ifsd_map != NULL) {
1839 		bus_dmamap_sync(txq->ift_buf_tag,
1840 		    txq->ift_sds.ifsd_map[i], BUS_DMASYNC_POSTWRITE);
1841 		bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[i]);
1842 	}
1843 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1844 		bus_dmamap_sync(txq->ift_tso_buf_tag,
1845 		    txq->ift_sds.ifsd_tso_map[i], BUS_DMASYNC_POSTWRITE);
1846 		bus_dmamap_unload(txq->ift_tso_buf_tag,
1847 		    txq->ift_sds.ifsd_tso_map[i]);
1848 	}
1849 	m_freem(*mp);
1850 	DBG_COUNTER_INC(tx_frees);
1851 	*mp = NULL;
1852 }
1853 
1854 static int
1855 iflib_txq_setup(iflib_txq_t txq)
1856 {
1857 	if_ctx_t ctx = txq->ift_ctx;
1858 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1859 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1860 	iflib_dma_info_t di;
1861 	int i;
1862 
1863 	/* Set number of descriptors available */
1864 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
1865 	/* XXX make configurable */
1866 	txq->ift_update_freq = IFLIB_DEFAULT_TX_UPDATE_FREQ;
1867 
1868 	/* Reset indices */
1869 	txq->ift_cidx_processed = 0;
1870 	txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0;
1871 	txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset];
1872 
1873 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1874 		bzero((void *)di->idi_vaddr, di->idi_size);
1875 
1876 	IFDI_TXQ_SETUP(ctx, txq->ift_id);
1877 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1878 		bus_dmamap_sync(di->idi_tag, di->idi_map,
1879 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1880 	return (0);
1881 }
1882 
1883 /*********************************************************************
1884  *
1885  *  Allocate DMA resources for RX buffers as well as memory for the RX
1886  *  mbuf map, direct RX cluster pointer map and RX cluster bus address
1887  *  map.  RX DMA map, RX mbuf map, direct RX cluster pointer map and
1888  *  RX cluster map are kept in a iflib_sw_rx_desc_array structure.
1889  *  Since we use use one entry in iflib_sw_rx_desc_array per received
1890  *  packet, the maximum number of entries we'll need is equal to the
1891  *  number of hardware receive descriptors that we've allocated.
1892  *
1893  **********************************************************************/
1894 static int
1895 iflib_rxsd_alloc(iflib_rxq_t rxq)
1896 {
1897 	if_ctx_t ctx = rxq->ifr_ctx;
1898 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1899 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1900 	device_t dev = ctx->ifc_dev;
1901 	iflib_fl_t fl;
1902 	bus_addr_t lowaddr;
1903 	int			err;
1904 
1905 	MPASS(scctx->isc_nrxd[0] > 0);
1906 	MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0);
1907 
1908 	lowaddr = DMA_WIDTH_TO_BUS_LOWADDR(scctx->isc_dma_width);
1909 
1910 	fl = rxq->ifr_fl;
1911 	for (int i = 0; i <  rxq->ifr_nfl; i++, fl++) {
1912 		fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */
1913 		/* Set up DMA tag for RX buffers. */
1914 		err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1915 					 1, 0,			/* alignment, bounds */
1916 					 lowaddr,		/* lowaddr */
1917 					 BUS_SPACE_MAXADDR,	/* highaddr */
1918 					 NULL, NULL,		/* filter, filterarg */
1919 					 sctx->isc_rx_maxsize,	/* maxsize */
1920 					 sctx->isc_rx_nsegments,	/* nsegments */
1921 					 sctx->isc_rx_maxsegsize,	/* maxsegsize */
1922 					 0,			/* flags */
1923 					 NULL,			/* lockfunc */
1924 					 NULL,			/* lockarg */
1925 					 &fl->ifl_buf_tag);
1926 		if (err) {
1927 			device_printf(dev,
1928 			    "Unable to allocate RX DMA tag: %d\n", err);
1929 			goto fail;
1930 		}
1931 
1932 		/* Allocate memory for the RX mbuf map. */
1933 		if (!(fl->ifl_sds.ifsd_m =
1934 		      (struct mbuf **) malloc(sizeof(struct mbuf *) *
1935 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1936 			device_printf(dev,
1937 			    "Unable to allocate RX mbuf map memory\n");
1938 			err = ENOMEM;
1939 			goto fail;
1940 		}
1941 
1942 		/* Allocate memory for the direct RX cluster pointer map. */
1943 		if (!(fl->ifl_sds.ifsd_cl =
1944 		      (caddr_t *) malloc(sizeof(caddr_t) *
1945 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1946 			device_printf(dev,
1947 			    "Unable to allocate RX cluster map memory\n");
1948 			err = ENOMEM;
1949 			goto fail;
1950 		}
1951 
1952 		/* Allocate memory for the RX cluster bus address map. */
1953 		if (!(fl->ifl_sds.ifsd_ba =
1954 		      (bus_addr_t *) malloc(sizeof(bus_addr_t) *
1955 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1956 			device_printf(dev,
1957 			    "Unable to allocate RX bus address map memory\n");
1958 			err = ENOMEM;
1959 			goto fail;
1960 		}
1961 
1962 		/*
1963 		 * Create the DMA maps for RX buffers.
1964 		 */
1965 		if (!(fl->ifl_sds.ifsd_map =
1966 		      (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1967 			device_printf(dev,
1968 			    "Unable to allocate RX buffer DMA map memory\n");
1969 			err = ENOMEM;
1970 			goto fail;
1971 		}
1972 		for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++) {
1973 			err = bus_dmamap_create(fl->ifl_buf_tag, 0,
1974 			    &fl->ifl_sds.ifsd_map[i]);
1975 			if (err != 0) {
1976 				device_printf(dev, "Unable to create RX buffer DMA map\n");
1977 				goto fail;
1978 			}
1979 		}
1980 	}
1981 	return (0);
1982 
1983 fail:
1984 	iflib_rx_structures_free(ctx);
1985 	return (err);
1986 }
1987 
1988 /*
1989  * Internal service routines
1990  */
1991 
1992 struct rxq_refill_cb_arg {
1993 	int               error;
1994 	bus_dma_segment_t seg;
1995 	int               nseg;
1996 };
1997 
1998 static void
1999 _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
2000 {
2001 	struct rxq_refill_cb_arg *cb_arg = arg;
2002 
2003 	cb_arg->error = error;
2004 	cb_arg->seg = segs[0];
2005 	cb_arg->nseg = nseg;
2006 }
2007 
2008 /**
2009  * iflib_fl_refill - refill an rxq free-buffer list
2010  * @ctx: the iflib context
2011  * @fl: the free list to refill
2012  * @count: the number of new buffers to allocate
2013  *
2014  * (Re)populate an rxq free-buffer list with up to @count new packet buffers.
2015  * The caller must assure that @count does not exceed the queue's capacity
2016  * minus one (since we always leave a descriptor unavailable).
2017  */
2018 static uint8_t
2019 iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
2020 {
2021 	struct if_rxd_update iru;
2022 	struct rxq_refill_cb_arg cb_arg;
2023 	struct mbuf *m;
2024 	caddr_t cl, *sd_cl;
2025 	struct mbuf **sd_m;
2026 	bus_dmamap_t *sd_map;
2027 	bus_addr_t bus_addr, *sd_ba;
2028 	int err, frag_idx, i, idx, n, pidx;
2029 	qidx_t credits;
2030 
2031 	MPASS(count <= fl->ifl_size - fl->ifl_credits - 1);
2032 
2033 	sd_m = fl->ifl_sds.ifsd_m;
2034 	sd_map = fl->ifl_sds.ifsd_map;
2035 	sd_cl = fl->ifl_sds.ifsd_cl;
2036 	sd_ba = fl->ifl_sds.ifsd_ba;
2037 	pidx = fl->ifl_pidx;
2038 	idx = pidx;
2039 	frag_idx = fl->ifl_fragidx;
2040 	credits = fl->ifl_credits;
2041 
2042 	i = 0;
2043 	n = count;
2044 	MPASS(n > 0);
2045 	MPASS(credits + n <= fl->ifl_size);
2046 
2047 	if (pidx < fl->ifl_cidx)
2048 		MPASS(pidx + n <= fl->ifl_cidx);
2049 	if (pidx == fl->ifl_cidx && (credits < fl->ifl_size))
2050 		MPASS(fl->ifl_gen == 0);
2051 	if (pidx > fl->ifl_cidx)
2052 		MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx);
2053 
2054 	DBG_COUNTER_INC(fl_refills);
2055 	if (n > 8)
2056 		DBG_COUNTER_INC(fl_refills_large);
2057 	iru_init(&iru, fl->ifl_rxq, fl->ifl_id);
2058 	while (n-- > 0) {
2059 		/*
2060 		 * We allocate an uninitialized mbuf + cluster, mbuf is
2061 		 * initialized after rx.
2062 		 *
2063 		 * If the cluster is still set then we know a minimum sized
2064 		 * packet was received
2065 		 */
2066 		bit_ffc_at(fl->ifl_rx_bitmap, frag_idx, fl->ifl_size,
2067 		    &frag_idx);
2068 		if (frag_idx < 0)
2069 			bit_ffc(fl->ifl_rx_bitmap, fl->ifl_size, &frag_idx);
2070 		MPASS(frag_idx >= 0);
2071 		if ((cl = sd_cl[frag_idx]) == NULL) {
2072 			cl = uma_zalloc(fl->ifl_zone, M_NOWAIT);
2073 			if (__predict_false(cl == NULL))
2074 				break;
2075 
2076 			cb_arg.error = 0;
2077 			MPASS(sd_map != NULL);
2078 			err = bus_dmamap_load(fl->ifl_buf_tag, sd_map[frag_idx],
2079 			    cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg,
2080 			    BUS_DMA_NOWAIT);
2081 			if (__predict_false(err != 0 || cb_arg.error)) {
2082 				uma_zfree(fl->ifl_zone, cl);
2083 				break;
2084 			}
2085 
2086 			sd_ba[frag_idx] = bus_addr = cb_arg.seg.ds_addr;
2087 			sd_cl[frag_idx] = cl;
2088 #if MEMORY_LOGGING
2089 			fl->ifl_cl_enqueued++;
2090 #endif
2091 		} else {
2092 			bus_addr = sd_ba[frag_idx];
2093 		}
2094 		bus_dmamap_sync(fl->ifl_buf_tag, sd_map[frag_idx],
2095 		    BUS_DMASYNC_PREREAD);
2096 
2097 		if (sd_m[frag_idx] == NULL) {
2098 			m = m_gethdr(M_NOWAIT, MT_NOINIT);
2099 			if (__predict_false(m == NULL))
2100 				break;
2101 			sd_m[frag_idx] = m;
2102 		}
2103 		bit_set(fl->ifl_rx_bitmap, frag_idx);
2104 #if MEMORY_LOGGING
2105 		fl->ifl_m_enqueued++;
2106 #endif
2107 
2108 		DBG_COUNTER_INC(rx_allocs);
2109 		fl->ifl_rxd_idxs[i] = frag_idx;
2110 		fl->ifl_bus_addrs[i] = bus_addr;
2111 		credits++;
2112 		i++;
2113 		MPASS(credits <= fl->ifl_size);
2114 		if (++idx == fl->ifl_size) {
2115 #ifdef INVARIANTS
2116 			fl->ifl_gen = 1;
2117 #endif
2118 			idx = 0;
2119 		}
2120 		if (n == 0 || i == IFLIB_MAX_RX_REFRESH) {
2121 			iru.iru_pidx = pidx;
2122 			iru.iru_count = i;
2123 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2124 			fl->ifl_pidx = idx;
2125 			fl->ifl_credits = credits;
2126 			pidx = idx;
2127 			i = 0;
2128 		}
2129 	}
2130 
2131 	if (n < count - 1) {
2132 		if (i != 0) {
2133 			iru.iru_pidx = pidx;
2134 			iru.iru_count = i;
2135 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2136 			fl->ifl_pidx = idx;
2137 			fl->ifl_credits = credits;
2138 		}
2139 		DBG_COUNTER_INC(rxd_flush);
2140 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2141 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2142 		ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id,
2143 		    fl->ifl_id, fl->ifl_pidx);
2144 		if (__predict_true(bit_test(fl->ifl_rx_bitmap, frag_idx))) {
2145 			fl->ifl_fragidx = frag_idx + 1;
2146 			if (fl->ifl_fragidx == fl->ifl_size)
2147 				fl->ifl_fragidx = 0;
2148 		} else {
2149 			fl->ifl_fragidx = frag_idx;
2150 		}
2151 	}
2152 
2153 	return (n == -1 ? 0 : IFLIB_RXEOF_EMPTY);
2154 }
2155 
2156 static inline uint8_t
2157 iflib_fl_refill_all(if_ctx_t ctx, iflib_fl_t fl)
2158 {
2159 	/*
2160 	 * We leave an unused descriptor to avoid pidx to catch up with cidx.
2161 	 * This is important as it confuses most NICs. For instance,
2162 	 * Intel NICs have (per receive ring) RDH and RDT registers, where
2163 	 * RDH points to the next receive descriptor to be used by the NIC,
2164 	 * and RDT for the next receive descriptor to be published by the
2165 	 * driver to the NIC (RDT - 1 is thus the last valid one).
2166 	 * The condition RDH == RDT means no descriptors are available to
2167 	 * the NIC, and thus it would be ambiguous if it also meant that
2168 	 * all the descriptors are available to the NIC.
2169 	 */
2170 	int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1;
2171 #ifdef INVARIANTS
2172 	int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1;
2173 #endif
2174 
2175 	MPASS(fl->ifl_credits <= fl->ifl_size);
2176 	MPASS(reclaimable == delta);
2177 
2178 	if (reclaimable > 0)
2179 		return (iflib_fl_refill(ctx, fl, reclaimable));
2180 	return (0);
2181 }
2182 
2183 uint8_t
2184 iflib_in_detach(if_ctx_t ctx)
2185 {
2186 	bool in_detach;
2187 
2188 	STATE_LOCK(ctx);
2189 	in_detach = !!(ctx->ifc_flags & IFC_IN_DETACH);
2190 	STATE_UNLOCK(ctx);
2191 	return (in_detach);
2192 }
2193 
2194 static void
2195 iflib_fl_bufs_free(iflib_fl_t fl)
2196 {
2197 	iflib_dma_info_t idi = fl->ifl_ifdi;
2198 	bus_dmamap_t sd_map;
2199 	uint32_t i;
2200 
2201 	for (i = 0; i < fl->ifl_size; i++) {
2202 		struct mbuf **sd_m = &fl->ifl_sds.ifsd_m[i];
2203 		caddr_t *sd_cl = &fl->ifl_sds.ifsd_cl[i];
2204 
2205 		if (*sd_cl != NULL) {
2206 			sd_map = fl->ifl_sds.ifsd_map[i];
2207 			bus_dmamap_sync(fl->ifl_buf_tag, sd_map,
2208 			    BUS_DMASYNC_POSTREAD);
2209 			bus_dmamap_unload(fl->ifl_buf_tag, sd_map);
2210 			uma_zfree(fl->ifl_zone, *sd_cl);
2211 			*sd_cl = NULL;
2212 			if (*sd_m != NULL) {
2213 				m_init(*sd_m, M_NOWAIT, MT_DATA, 0);
2214 				uma_zfree(zone_mbuf, *sd_m);
2215 				*sd_m = NULL;
2216 			}
2217 		} else {
2218 			MPASS(*sd_m == NULL);
2219 		}
2220 #if MEMORY_LOGGING
2221 		fl->ifl_m_dequeued++;
2222 		fl->ifl_cl_dequeued++;
2223 #endif
2224 	}
2225 #ifdef INVARIANTS
2226 	for (i = 0; i < fl->ifl_size; i++) {
2227 		MPASS(fl->ifl_sds.ifsd_cl[i] == NULL);
2228 		MPASS(fl->ifl_sds.ifsd_m[i] == NULL);
2229 	}
2230 #endif
2231 	/*
2232 	 * Reset free list values
2233 	 */
2234 	fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = fl->ifl_fragidx = 0;
2235 	bzero(idi->idi_vaddr, idi->idi_size);
2236 }
2237 
2238 /*********************************************************************
2239  *
2240  *  Initialize a free list and its buffers.
2241  *
2242  **********************************************************************/
2243 static int
2244 iflib_fl_setup(iflib_fl_t fl)
2245 {
2246 	iflib_rxq_t rxq = fl->ifl_rxq;
2247 	if_ctx_t ctx = rxq->ifr_ctx;
2248 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2249 	int qidx;
2250 
2251 	bit_nclear(fl->ifl_rx_bitmap, 0, fl->ifl_size - 1);
2252 	/*
2253 	** Free current RX buffer structs and their mbufs
2254 	*/
2255 	iflib_fl_bufs_free(fl);
2256 	/* Now replenish the mbufs */
2257 	MPASS(fl->ifl_credits == 0);
2258 	qidx = rxq->ifr_fl_offset + fl->ifl_id;
2259 	if (scctx->isc_rxd_buf_size[qidx] != 0)
2260 		fl->ifl_buf_size = scctx->isc_rxd_buf_size[qidx];
2261 	else
2262 		fl->ifl_buf_size = ctx->ifc_rx_mbuf_sz;
2263 	/*
2264 	 * ifl_buf_size may be a driver-supplied value, so pull it up
2265 	 * to the selected mbuf size.
2266 	 */
2267 	fl->ifl_buf_size = iflib_get_mbuf_size_for(fl->ifl_buf_size);
2268 	if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size)
2269 		ctx->ifc_max_fl_buf_size = fl->ifl_buf_size;
2270 	fl->ifl_cltype = m_gettype(fl->ifl_buf_size);
2271 	fl->ifl_zone = m_getzone(fl->ifl_buf_size);
2272 
2273 	/*
2274 	 * Avoid pre-allocating zillions of clusters to an idle card
2275 	 * potentially speeding up attach. In any case make sure
2276 	 * to leave a descriptor unavailable. See the comment in
2277 	 * iflib_fl_refill_all().
2278 	 */
2279 	MPASS(fl->ifl_size > 0);
2280 	(void)iflib_fl_refill(ctx, fl, min(128, fl->ifl_size - 1));
2281 	if (min(128, fl->ifl_size - 1) != fl->ifl_credits)
2282 		return (ENOBUFS);
2283 	/*
2284 	 * handle failure
2285 	 */
2286 	MPASS(rxq != NULL);
2287 	MPASS(fl->ifl_ifdi != NULL);
2288 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2289 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2290 	return (0);
2291 }
2292 
2293 /*********************************************************************
2294  *
2295  *  Free receive ring data structures
2296  *
2297  **********************************************************************/
2298 static void
2299 iflib_rx_sds_free(iflib_rxq_t rxq)
2300 {
2301 	iflib_fl_t fl;
2302 	int i, j;
2303 
2304 	if (rxq->ifr_fl != NULL) {
2305 		for (i = 0; i < rxq->ifr_nfl; i++) {
2306 			fl = &rxq->ifr_fl[i];
2307 			if (fl->ifl_buf_tag != NULL) {
2308 				if (fl->ifl_sds.ifsd_map != NULL) {
2309 					for (j = 0; j < fl->ifl_size; j++) {
2310 						bus_dmamap_sync(
2311 						    fl->ifl_buf_tag,
2312 						    fl->ifl_sds.ifsd_map[j],
2313 						    BUS_DMASYNC_POSTREAD);
2314 						bus_dmamap_unload(
2315 						    fl->ifl_buf_tag,
2316 						    fl->ifl_sds.ifsd_map[j]);
2317 						bus_dmamap_destroy(
2318 						    fl->ifl_buf_tag,
2319 						    fl->ifl_sds.ifsd_map[j]);
2320 					}
2321 				}
2322 				bus_dma_tag_destroy(fl->ifl_buf_tag);
2323 				fl->ifl_buf_tag = NULL;
2324 			}
2325 			free(fl->ifl_sds.ifsd_m, M_IFLIB);
2326 			free(fl->ifl_sds.ifsd_cl, M_IFLIB);
2327 			free(fl->ifl_sds.ifsd_ba, M_IFLIB);
2328 			free(fl->ifl_sds.ifsd_map, M_IFLIB);
2329 			free(fl->ifl_rx_bitmap, M_IFLIB);
2330 			fl->ifl_sds.ifsd_m = NULL;
2331 			fl->ifl_sds.ifsd_cl = NULL;
2332 			fl->ifl_sds.ifsd_ba = NULL;
2333 			fl->ifl_sds.ifsd_map = NULL;
2334 			fl->ifl_rx_bitmap = NULL;
2335 		}
2336 		free(rxq->ifr_fl, M_IFLIB);
2337 		rxq->ifr_fl = NULL;
2338 		free(rxq->ifr_ifdi, M_IFLIB);
2339 		rxq->ifr_ifdi = NULL;
2340 		rxq->ifr_cq_cidx = 0;
2341 	}
2342 }
2343 
2344 /*
2345  * Timer routine
2346  */
2347 static void
2348 iflib_timer(void *arg)
2349 {
2350 	iflib_txq_t txq = arg;
2351 	if_ctx_t ctx = txq->ift_ctx;
2352 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2353 	uint64_t this_tick = ticks;
2354 
2355 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
2356 		return;
2357 
2358 	/*
2359 	** Check on the state of the TX queue(s), this
2360 	** can be done without the lock because its RO
2361 	** and the HUNG state will be static if set.
2362 	*/
2363 	if (this_tick - txq->ift_last_timer_tick >= iflib_timer_default) {
2364 		txq->ift_last_timer_tick = this_tick;
2365 		IFDI_TIMER(ctx, txq->ift_id);
2366 		if ((txq->ift_qstatus == IFLIB_QUEUE_HUNG) &&
2367 		    ((txq->ift_cleaned_prev == txq->ift_cleaned) ||
2368 		     (sctx->isc_pause_frames == 0)))
2369 			goto hung;
2370 
2371 		if (txq->ift_qstatus != IFLIB_QUEUE_IDLE &&
2372 		    ifmp_ring_is_stalled(txq->ift_br)) {
2373 			KASSERT(ctx->ifc_link_state == LINK_STATE_UP,
2374 			    ("queue can't be marked as hung if interface is down"));
2375 			txq->ift_qstatus = IFLIB_QUEUE_HUNG;
2376 		}
2377 		txq->ift_cleaned_prev = txq->ift_cleaned;
2378 	}
2379 	/* handle any laggards */
2380 	if (txq->ift_db_pending)
2381 		GROUPTASK_ENQUEUE(&txq->ift_task);
2382 
2383 	sctx->isc_pause_frames = 0;
2384 	if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)
2385 		callout_reset_on(&txq->ift_timer, iflib_timer_default, iflib_timer,
2386 		    txq, txq->ift_timer.c_cpu);
2387 	return;
2388 
2389  hung:
2390 	device_printf(ctx->ifc_dev,
2391 	    "Watchdog timeout (TX: %d desc avail: %d pidx: %d) -- resetting\n",
2392 	    txq->ift_id, TXQ_AVAIL(txq), txq->ift_pidx);
2393 	STATE_LOCK(ctx);
2394 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2395 	ctx->ifc_flags |= (IFC_DO_WATCHDOG|IFC_DO_RESET);
2396 	iflib_admin_intr_deferred(ctx);
2397 	STATE_UNLOCK(ctx);
2398 }
2399 
2400 static uint16_t
2401 iflib_get_mbuf_size_for(unsigned int size)
2402 {
2403 
2404 	if (size <= MCLBYTES)
2405 		return (MCLBYTES);
2406 	else
2407 		return (MJUMPAGESIZE);
2408 }
2409 
2410 static void
2411 iflib_calc_rx_mbuf_sz(if_ctx_t ctx)
2412 {
2413 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2414 
2415 	/*
2416 	 * XXX don't set the max_frame_size to larger
2417 	 * than the hardware can handle
2418 	 */
2419 	ctx->ifc_rx_mbuf_sz =
2420 	    iflib_get_mbuf_size_for(sctx->isc_max_frame_size);
2421 }
2422 
2423 uint32_t
2424 iflib_get_rx_mbuf_sz(if_ctx_t ctx)
2425 {
2426 
2427 	return (ctx->ifc_rx_mbuf_sz);
2428 }
2429 
2430 static void
2431 iflib_init_locked(if_ctx_t ctx)
2432 {
2433 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2434 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2435 	if_t ifp = ctx->ifc_ifp;
2436 	iflib_fl_t fl;
2437 	iflib_txq_t txq;
2438 	iflib_rxq_t rxq;
2439 	int i, j, tx_ip_csum_flags, tx_ip6_csum_flags;
2440 
2441 	if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2442 	IFDI_INTR_DISABLE(ctx);
2443 
2444 	/*
2445 	 * See iflib_stop(). Useful in case iflib_init_locked() is
2446 	 * called without first calling iflib_stop().
2447 	 */
2448 	netmap_disable_all_rings(ifp);
2449 
2450 	tx_ip_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP);
2451 	tx_ip6_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | CSUM_IP6_SCTP);
2452 	/* Set hardware offload abilities */
2453 	if_clearhwassist(ifp);
2454 	if (if_getcapenable(ifp) & IFCAP_TXCSUM)
2455 		if_sethwassistbits(ifp, tx_ip_csum_flags, 0);
2456 	if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6)
2457 		if_sethwassistbits(ifp,  tx_ip6_csum_flags, 0);
2458 	if (if_getcapenable(ifp) & IFCAP_TSO4)
2459 		if_sethwassistbits(ifp, CSUM_IP_TSO, 0);
2460 	if (if_getcapenable(ifp) & IFCAP_TSO6)
2461 		if_sethwassistbits(ifp, CSUM_IP6_TSO, 0);
2462 
2463 	for (i = 0, txq = ctx->ifc_txqs; i < sctx->isc_ntxqsets; i++, txq++) {
2464 		CALLOUT_LOCK(txq);
2465 		callout_stop(&txq->ift_timer);
2466 #ifdef DEV_NETMAP
2467 		callout_stop(&txq->ift_netmap_timer);
2468 #endif /* DEV_NETMAP */
2469 		CALLOUT_UNLOCK(txq);
2470 		(void)iflib_netmap_txq_init(ctx, txq);
2471 	}
2472 
2473 	/*
2474 	 * Calculate a suitable Rx mbuf size prior to calling IFDI_INIT, so
2475 	 * that drivers can use the value when setting up the hardware receive
2476 	 * buffers.
2477 	 */
2478 	iflib_calc_rx_mbuf_sz(ctx);
2479 
2480 #ifdef INVARIANTS
2481 	i = if_getdrvflags(ifp);
2482 #endif
2483 	IFDI_INIT(ctx);
2484 	MPASS(if_getdrvflags(ifp) == i);
2485 	for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
2486 		if (iflib_netmap_rxq_init(ctx, rxq) > 0) {
2487 			/* This rxq is in netmap mode. Skip normal init. */
2488 			continue;
2489 		}
2490 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
2491 			if (iflib_fl_setup(fl)) {
2492 				device_printf(ctx->ifc_dev,
2493 				    "setting up free list %d failed - "
2494 				    "check cluster settings\n", j);
2495 				goto done;
2496 			}
2497 		}
2498 	}
2499 done:
2500 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
2501 	IFDI_INTR_ENABLE(ctx);
2502 	txq = ctx->ifc_txqs;
2503 	for (i = 0; i < sctx->isc_ntxqsets; i++, txq++)
2504 		callout_reset_on(&txq->ift_timer, iflib_timer_default, iflib_timer, txq,
2505 			txq->ift_timer.c_cpu);
2506 
2507         /* Re-enable txsync/rxsync. */
2508 	netmap_enable_all_rings(ifp);
2509 }
2510 
2511 static int
2512 iflib_media_change(if_t ifp)
2513 {
2514 	if_ctx_t ctx = if_getsoftc(ifp);
2515 	int err;
2516 
2517 	CTX_LOCK(ctx);
2518 	if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0)
2519 		iflib_if_init_locked(ctx);
2520 	CTX_UNLOCK(ctx);
2521 	return (err);
2522 }
2523 
2524 static void
2525 iflib_media_status(if_t ifp, struct ifmediareq *ifmr)
2526 {
2527 	if_ctx_t ctx = if_getsoftc(ifp);
2528 
2529 	CTX_LOCK(ctx);
2530 	IFDI_UPDATE_ADMIN_STATUS(ctx);
2531 	IFDI_MEDIA_STATUS(ctx, ifmr);
2532 	CTX_UNLOCK(ctx);
2533 }
2534 
2535 void
2536 iflib_stop(if_ctx_t ctx)
2537 {
2538 	iflib_txq_t txq = ctx->ifc_txqs;
2539 	iflib_rxq_t rxq = ctx->ifc_rxqs;
2540 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2541 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2542 	iflib_dma_info_t di;
2543 	iflib_fl_t fl;
2544 	int i, j;
2545 
2546 	/* Tell the stack that the interface is no longer active */
2547 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2548 
2549 	IFDI_INTR_DISABLE(ctx);
2550 	DELAY(1000);
2551 	IFDI_STOP(ctx);
2552 	DELAY(1000);
2553 
2554 	/*
2555 	 * Stop any pending txsync/rxsync and prevent new ones
2556 	 * form starting. Processes blocked in poll() will get
2557 	 * POLLERR.
2558 	 */
2559 	netmap_disable_all_rings(ctx->ifc_ifp);
2560 
2561 	iflib_debug_reset();
2562 	/* Wait for current tx queue users to exit to disarm watchdog timer. */
2563 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) {
2564 		/* make sure all transmitters have completed before proceeding XXX */
2565 
2566 		CALLOUT_LOCK(txq);
2567 		callout_stop(&txq->ift_timer);
2568 #ifdef DEV_NETMAP
2569 		callout_stop(&txq->ift_netmap_timer);
2570 #endif /* DEV_NETMAP */
2571 		CALLOUT_UNLOCK(txq);
2572 
2573 		/* clean any enqueued buffers */
2574 		iflib_ifmp_purge(txq);
2575 		/* Free any existing tx buffers. */
2576 		for (j = 0; j < txq->ift_size; j++) {
2577 			iflib_txsd_free(ctx, txq, j);
2578 		}
2579 		txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0;
2580 		txq->ift_in_use = txq->ift_gen = txq->ift_cidx = txq->ift_pidx = txq->ift_no_desc_avail = 0;
2581 		txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0;
2582 		txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0;
2583 		txq->ift_pullups = 0;
2584 		ifmp_ring_reset_stats(txq->ift_br);
2585 		for (j = 0, di = txq->ift_ifdi; j < sctx->isc_ntxqs; j++, di++)
2586 			bzero((void *)di->idi_vaddr, di->idi_size);
2587 	}
2588 	for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) {
2589 		/* make sure all transmitters have completed before proceeding XXX */
2590 
2591 		rxq->ifr_cq_cidx = 0;
2592 		for (j = 0, di = rxq->ifr_ifdi; j < sctx->isc_nrxqs; j++, di++)
2593 			bzero((void *)di->idi_vaddr, di->idi_size);
2594 		/* also resets the free lists pidx/cidx */
2595 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
2596 			iflib_fl_bufs_free(fl);
2597 	}
2598 }
2599 
2600 static inline caddr_t
2601 calc_next_rxd(iflib_fl_t fl, int cidx)
2602 {
2603 	qidx_t size;
2604 	int nrxd;
2605 	caddr_t start, end, cur, next;
2606 
2607 	nrxd = fl->ifl_size;
2608 	size = fl->ifl_rxd_size;
2609 	start = fl->ifl_ifdi->idi_vaddr;
2610 
2611 	if (__predict_false(size == 0))
2612 		return (start);
2613 	cur = start + size*cidx;
2614 	end = start + size*nrxd;
2615 	next = CACHE_PTR_NEXT(cur);
2616 	return (next < end ? next : start);
2617 }
2618 
2619 static inline void
2620 prefetch_pkts(iflib_fl_t fl, int cidx)
2621 {
2622 	int nextptr;
2623 	int nrxd = fl->ifl_size;
2624 	caddr_t next_rxd;
2625 
2626 	nextptr = (cidx + CACHE_PTR_INCREMENT) & (nrxd-1);
2627 	prefetch(&fl->ifl_sds.ifsd_m[nextptr]);
2628 	prefetch(&fl->ifl_sds.ifsd_cl[nextptr]);
2629 	next_rxd = calc_next_rxd(fl, cidx);
2630 	prefetch(next_rxd);
2631 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 1) & (nrxd-1)]);
2632 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 2) & (nrxd-1)]);
2633 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 3) & (nrxd-1)]);
2634 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 4) & (nrxd-1)]);
2635 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 1) & (nrxd-1)]);
2636 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 2) & (nrxd-1)]);
2637 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 3) & (nrxd-1)]);
2638 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 4) & (nrxd-1)]);
2639 }
2640 
2641 static struct mbuf *
2642 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, bool unload, if_rxsd_t sd,
2643     int *pf_rv, if_rxd_info_t ri)
2644 {
2645 	bus_dmamap_t map;
2646 	iflib_fl_t fl;
2647 	caddr_t payload;
2648 	struct mbuf *m;
2649 	int flid, cidx, len, next;
2650 
2651 	map = NULL;
2652 	flid = irf->irf_flid;
2653 	cidx = irf->irf_idx;
2654 	fl = &rxq->ifr_fl[flid];
2655 	sd->ifsd_fl = fl;
2656 	m = fl->ifl_sds.ifsd_m[cidx];
2657 	sd->ifsd_cl = &fl->ifl_sds.ifsd_cl[cidx];
2658 	fl->ifl_credits--;
2659 #if MEMORY_LOGGING
2660 	fl->ifl_m_dequeued++;
2661 #endif
2662 	if (rxq->ifr_ctx->ifc_flags & IFC_PREFETCH)
2663 		prefetch_pkts(fl, cidx);
2664 	next = (cidx + CACHE_PTR_INCREMENT) & (fl->ifl_size-1);
2665 	prefetch(&fl->ifl_sds.ifsd_map[next]);
2666 	map = fl->ifl_sds.ifsd_map[cidx];
2667 
2668 	bus_dmamap_sync(fl->ifl_buf_tag, map, BUS_DMASYNC_POSTREAD);
2669 
2670 	if (rxq->pfil != NULL && PFIL_HOOKED_IN(rxq->pfil) && pf_rv != NULL &&
2671 	    irf->irf_len != 0) {
2672 		payload  = *sd->ifsd_cl;
2673 		payload +=  ri->iri_pad;
2674 		len = ri->iri_len - ri->iri_pad;
2675 		*pf_rv = pfil_run_hooks(rxq->pfil, payload, ri->iri_ifp,
2676 		    len | PFIL_MEMPTR | PFIL_IN, NULL);
2677 		switch (*pf_rv) {
2678 		case PFIL_DROPPED:
2679 		case PFIL_CONSUMED:
2680 			/*
2681 			 * The filter ate it.  Everything is recycled.
2682 			 */
2683 			m = NULL;
2684 			unload = 0;
2685 			break;
2686 		case PFIL_REALLOCED:
2687 			/*
2688 			 * The filter copied it.  Everything is recycled.
2689 			 */
2690 			m = pfil_mem2mbuf(payload);
2691 			unload = 0;
2692 			break;
2693 		case PFIL_PASS:
2694 			/*
2695 			 * Filter said it was OK, so receive like
2696 			 * normal
2697 			 */
2698 			fl->ifl_sds.ifsd_m[cidx] = NULL;
2699 			break;
2700 		default:
2701 			MPASS(0);
2702 		}
2703 	} else {
2704 		fl->ifl_sds.ifsd_m[cidx] = NULL;
2705 		if (pf_rv != NULL)
2706 			*pf_rv = PFIL_PASS;
2707 	}
2708 
2709 	if (unload && irf->irf_len != 0)
2710 		bus_dmamap_unload(fl->ifl_buf_tag, map);
2711 	fl->ifl_cidx = (fl->ifl_cidx + 1) & (fl->ifl_size-1);
2712 	if (__predict_false(fl->ifl_cidx == 0))
2713 		fl->ifl_gen = 0;
2714 	bit_clear(fl->ifl_rx_bitmap, cidx);
2715 	return (m);
2716 }
2717 
2718 static struct mbuf *
2719 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri, if_rxsd_t sd, int *pf_rv)
2720 {
2721 	struct mbuf *m, *mh, *mt;
2722 	caddr_t cl;
2723 	int  *pf_rv_ptr, flags, i, padlen;
2724 	bool consumed;
2725 
2726 	i = 0;
2727 	mh = NULL;
2728 	consumed = false;
2729 	*pf_rv = PFIL_PASS;
2730 	pf_rv_ptr = pf_rv;
2731 	do {
2732 		m = rxd_frag_to_sd(rxq, &ri->iri_frags[i], !consumed, sd,
2733 		    pf_rv_ptr, ri);
2734 
2735 		MPASS(*sd->ifsd_cl != NULL);
2736 
2737 		/*
2738 		 * Exclude zero-length frags & frags from
2739 		 * packets the filter has consumed or dropped
2740 		 */
2741 		if (ri->iri_frags[i].irf_len == 0 || consumed ||
2742 		    *pf_rv == PFIL_CONSUMED || *pf_rv == PFIL_DROPPED) {
2743 			if (mh == NULL) {
2744 				/* everything saved here */
2745 				consumed = true;
2746 				pf_rv_ptr = NULL;
2747 				continue;
2748 			}
2749 			/* XXX we can save the cluster here, but not the mbuf */
2750 			m_init(m, M_NOWAIT, MT_DATA, 0);
2751 			m_free(m);
2752 			continue;
2753 		}
2754 		if (mh == NULL) {
2755 			flags = M_PKTHDR|M_EXT;
2756 			mh = mt = m;
2757 			padlen = ri->iri_pad;
2758 		} else {
2759 			flags = M_EXT;
2760 			mt->m_next = m;
2761 			mt = m;
2762 			/* assuming padding is only on the first fragment */
2763 			padlen = 0;
2764 		}
2765 		cl = *sd->ifsd_cl;
2766 		*sd->ifsd_cl = NULL;
2767 
2768 		/* Can these two be made one ? */
2769 		m_init(m, M_NOWAIT, MT_DATA, flags);
2770 		m_cljset(m, cl, sd->ifsd_fl->ifl_cltype);
2771 		/*
2772 		 * These must follow m_init and m_cljset
2773 		 */
2774 		m->m_data += padlen;
2775 		ri->iri_len -= padlen;
2776 		m->m_len = ri->iri_frags[i].irf_len;
2777 	} while (++i < ri->iri_nfrags);
2778 
2779 	return (mh);
2780 }
2781 
2782 /*
2783  * Process one software descriptor
2784  */
2785 static struct mbuf *
2786 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri)
2787 {
2788 	struct if_rxsd sd;
2789 	struct mbuf *m;
2790 	int pf_rv;
2791 
2792 	/* should I merge this back in now that the two paths are basically duplicated? */
2793 	if (ri->iri_nfrags == 1 &&
2794 	    ri->iri_frags[0].irf_len != 0 &&
2795 	    ri->iri_frags[0].irf_len <= MIN(IFLIB_RX_COPY_THRESH, MHLEN)) {
2796 		m = rxd_frag_to_sd(rxq, &ri->iri_frags[0], false, &sd,
2797 		    &pf_rv, ri);
2798 		if (pf_rv != PFIL_PASS && pf_rv != PFIL_REALLOCED)
2799 			return (m);
2800 		if (pf_rv == PFIL_PASS) {
2801 			m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR);
2802 #ifndef __NO_STRICT_ALIGNMENT
2803 			if (!IP_ALIGNED(m))
2804 				m->m_data += 2;
2805 #endif
2806 			memcpy(m->m_data, *sd.ifsd_cl, ri->iri_len);
2807 			m->m_len = ri->iri_frags[0].irf_len;
2808 		}
2809 	} else {
2810 		m = assemble_segments(rxq, ri, &sd, &pf_rv);
2811 		if (m == NULL)
2812 			return (NULL);
2813 		if (pf_rv != PFIL_PASS && pf_rv != PFIL_REALLOCED)
2814 			return (m);
2815 	}
2816 	m->m_pkthdr.len = ri->iri_len;
2817 	m->m_pkthdr.rcvif = ri->iri_ifp;
2818 	m->m_flags |= ri->iri_flags;
2819 	m->m_pkthdr.ether_vtag = ri->iri_vtag;
2820 	m->m_pkthdr.flowid = ri->iri_flowid;
2821 	M_HASHTYPE_SET(m, ri->iri_rsstype);
2822 	m->m_pkthdr.csum_flags = ri->iri_csum_flags;
2823 	m->m_pkthdr.csum_data = ri->iri_csum_data;
2824 	return (m);
2825 }
2826 
2827 #if defined(INET6) || defined(INET)
2828 static void
2829 iflib_get_ip_forwarding(struct lro_ctrl *lc, bool *v4, bool *v6)
2830 {
2831 	CURVNET_SET(lc->ifp->if_vnet);
2832 #if defined(INET6)
2833 	*v6 = V_ip6_forwarding;
2834 #endif
2835 #if defined(INET)
2836 	*v4 = V_ipforwarding;
2837 #endif
2838 	CURVNET_RESTORE();
2839 }
2840 
2841 /*
2842  * Returns true if it's possible this packet could be LROed.
2843  * if it returns false, it is guaranteed that tcp_lro_rx()
2844  * would not return zero.
2845  */
2846 static bool
2847 iflib_check_lro_possible(struct mbuf *m, bool v4_forwarding, bool v6_forwarding)
2848 {
2849 	struct ether_header *eh;
2850 
2851 	eh = mtod(m, struct ether_header *);
2852 	switch (eh->ether_type) {
2853 #if defined(INET6)
2854 		case htons(ETHERTYPE_IPV6):
2855 			return (!v6_forwarding);
2856 #endif
2857 #if defined (INET)
2858 		case htons(ETHERTYPE_IP):
2859 			return (!v4_forwarding);
2860 #endif
2861 	}
2862 
2863 	return false;
2864 }
2865 #else
2866 static void
2867 iflib_get_ip_forwarding(struct lro_ctrl *lc __unused, bool *v4 __unused, bool *v6 __unused)
2868 {
2869 }
2870 #endif
2871 
2872 static void
2873 _task_fn_rx_watchdog(void *context)
2874 {
2875 	iflib_rxq_t rxq = context;
2876 
2877 	GROUPTASK_ENQUEUE(&rxq->ifr_task);
2878 }
2879 
2880 static uint8_t
2881 iflib_rxeof(iflib_rxq_t rxq, qidx_t budget)
2882 {
2883 	if_t ifp;
2884 	if_ctx_t ctx = rxq->ifr_ctx;
2885 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2886 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2887 	int avail, i;
2888 	qidx_t *cidxp;
2889 	struct if_rxd_info ri;
2890 	int err, budget_left, rx_bytes, rx_pkts;
2891 	iflib_fl_t fl;
2892 	int lro_enabled;
2893 	bool v4_forwarding, v6_forwarding, lro_possible;
2894 	uint8_t retval = 0;
2895 
2896 	/*
2897 	 * XXX early demux data packets so that if_input processing only handles
2898 	 * acks in interrupt context
2899 	 */
2900 	struct mbuf *m, *mh, *mt, *mf;
2901 
2902 	NET_EPOCH_ASSERT();
2903 
2904 	lro_possible = v4_forwarding = v6_forwarding = false;
2905 	ifp = ctx->ifc_ifp;
2906 	mh = mt = NULL;
2907 	MPASS(budget > 0);
2908 	rx_pkts	= rx_bytes = 0;
2909 	if (sctx->isc_flags & IFLIB_HAS_RXCQ)
2910 		cidxp = &rxq->ifr_cq_cidx;
2911 	else
2912 		cidxp = &rxq->ifr_fl[0].ifl_cidx;
2913 	if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) {
2914 		for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2915 			retval |= iflib_fl_refill_all(ctx, fl);
2916 		DBG_COUNTER_INC(rx_unavail);
2917 		return (retval);
2918 	}
2919 
2920 	/* pfil needs the vnet to be set */
2921 	CURVNET_SET_QUIET(ifp->if_vnet);
2922 	for (budget_left = budget; budget_left > 0 && avail > 0;) {
2923 		if (__predict_false(!CTX_ACTIVE(ctx))) {
2924 			DBG_COUNTER_INC(rx_ctx_inactive);
2925 			break;
2926 		}
2927 		/*
2928 		 * Reset client set fields to their default values
2929 		 */
2930 		rxd_info_zero(&ri);
2931 		ri.iri_qsidx = rxq->ifr_id;
2932 		ri.iri_cidx = *cidxp;
2933 		ri.iri_ifp = ifp;
2934 		ri.iri_frags = rxq->ifr_frags;
2935 		err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
2936 
2937 		if (err)
2938 			goto err;
2939 		rx_pkts += 1;
2940 		rx_bytes += ri.iri_len;
2941 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
2942 			*cidxp = ri.iri_cidx;
2943 			/* Update our consumer index */
2944 			/* XXX NB: shurd - check if this is still safe */
2945 			while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0])
2946 				rxq->ifr_cq_cidx -= scctx->isc_nrxd[0];
2947 			/* was this only a completion queue message? */
2948 			if (__predict_false(ri.iri_nfrags == 0))
2949 				continue;
2950 		}
2951 		MPASS(ri.iri_nfrags != 0);
2952 		MPASS(ri.iri_len != 0);
2953 
2954 		/* will advance the cidx on the corresponding free lists */
2955 		m = iflib_rxd_pkt_get(rxq, &ri);
2956 		avail--;
2957 		budget_left--;
2958 		if (avail == 0 && budget_left)
2959 			avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left);
2960 
2961 		if (__predict_false(m == NULL))
2962 			continue;
2963 
2964 		/* imm_pkt: -- cxgb */
2965 		if (mh == NULL)
2966 			mh = mt = m;
2967 		else {
2968 			mt->m_nextpkt = m;
2969 			mt = m;
2970 		}
2971 	}
2972 	CURVNET_RESTORE();
2973 	/* make sure that we can refill faster than drain */
2974 	for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2975 		retval |= iflib_fl_refill_all(ctx, fl);
2976 
2977 	lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO);
2978 	if (lro_enabled)
2979 		iflib_get_ip_forwarding(&rxq->ifr_lc, &v4_forwarding, &v6_forwarding);
2980 	mt = mf = NULL;
2981 	while (mh != NULL) {
2982 		m = mh;
2983 		mh = mh->m_nextpkt;
2984 		m->m_nextpkt = NULL;
2985 #ifndef __NO_STRICT_ALIGNMENT
2986 		if (!IP_ALIGNED(m) && (m = iflib_fixup_rx(m)) == NULL)
2987 			continue;
2988 #endif
2989 #if defined(INET6) || defined(INET)
2990 		if (lro_enabled) {
2991 			if (!lro_possible) {
2992 				lro_possible = iflib_check_lro_possible(m, v4_forwarding, v6_forwarding);
2993 				if (lro_possible && mf != NULL) {
2994 					ifp->if_input(ifp, mf);
2995 					DBG_COUNTER_INC(rx_if_input);
2996 					mt = mf = NULL;
2997 				}
2998 			}
2999 			if ((m->m_pkthdr.csum_flags & (CSUM_L4_CALC|CSUM_L4_VALID)) ==
3000 			    (CSUM_L4_CALC|CSUM_L4_VALID)) {
3001 				if (lro_possible && tcp_lro_rx(&rxq->ifr_lc, m, 0) == 0)
3002 					continue;
3003 			}
3004 		}
3005 #endif
3006 		if (lro_possible) {
3007 			ifp->if_input(ifp, m);
3008 			DBG_COUNTER_INC(rx_if_input);
3009 			continue;
3010 		}
3011 
3012 		if (mf == NULL)
3013 			mf = m;
3014 		if (mt != NULL)
3015 			mt->m_nextpkt = m;
3016 		mt = m;
3017 	}
3018 	if (mf != NULL) {
3019 		ifp->if_input(ifp, mf);
3020 		DBG_COUNTER_INC(rx_if_input);
3021 	}
3022 
3023 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
3024 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
3025 
3026 	/*
3027 	 * Flush any outstanding LRO work
3028 	 */
3029 #if defined(INET6) || defined(INET)
3030 	tcp_lro_flush_all(&rxq->ifr_lc);
3031 #endif
3032 	if (avail != 0 || iflib_rxd_avail(ctx, rxq, *cidxp, 1) != 0)
3033 		retval |= IFLIB_RXEOF_MORE;
3034 	return (retval);
3035 err:
3036 	STATE_LOCK(ctx);
3037 	ctx->ifc_flags |= IFC_DO_RESET;
3038 	iflib_admin_intr_deferred(ctx);
3039 	STATE_UNLOCK(ctx);
3040 	return (0);
3041 }
3042 
3043 #define TXD_NOTIFY_COUNT(txq) (((txq)->ift_size / (txq)->ift_update_freq)-1)
3044 static inline qidx_t
3045 txq_max_db_deferred(iflib_txq_t txq, qidx_t in_use)
3046 {
3047 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
3048 	qidx_t minthresh = txq->ift_size / 8;
3049 	if (in_use > 4*minthresh)
3050 		return (notify_count);
3051 	if (in_use > 2*minthresh)
3052 		return (notify_count >> 1);
3053 	if (in_use > minthresh)
3054 		return (notify_count >> 3);
3055 	return (0);
3056 }
3057 
3058 static inline qidx_t
3059 txq_max_rs_deferred(iflib_txq_t txq)
3060 {
3061 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
3062 	qidx_t minthresh = txq->ift_size / 8;
3063 	if (txq->ift_in_use > 4*minthresh)
3064 		return (notify_count);
3065 	if (txq->ift_in_use > 2*minthresh)
3066 		return (notify_count >> 1);
3067 	if (txq->ift_in_use > minthresh)
3068 		return (notify_count >> 2);
3069 	return (2);
3070 }
3071 
3072 #define M_CSUM_FLAGS(m) ((m)->m_pkthdr.csum_flags)
3073 #define M_HAS_VLANTAG(m) (m->m_flags & M_VLANTAG)
3074 
3075 #define TXQ_MAX_DB_DEFERRED(txq, in_use) txq_max_db_deferred((txq), (in_use))
3076 #define TXQ_MAX_RS_DEFERRED(txq) txq_max_rs_deferred(txq)
3077 #define TXQ_MAX_DB_CONSUMED(size) (size >> 4)
3078 
3079 /* forward compatibility for cxgb */
3080 #define FIRST_QSET(ctx) 0
3081 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets)
3082 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets)
3083 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx))
3084 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments))
3085 
3086 /* XXX we should be setting this to something other than zero */
3087 #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh)
3088 #define	MAX_TX_DESC(ctx) MAX((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max, \
3089     (ctx)->ifc_softc_ctx.isc_tx_nsegments)
3090 
3091 static inline bool
3092 iflib_txd_db_check(iflib_txq_t txq, int ring)
3093 {
3094 	if_ctx_t ctx = txq->ift_ctx;
3095 	qidx_t dbval, max;
3096 
3097 	max = TXQ_MAX_DB_DEFERRED(txq, txq->ift_in_use);
3098 
3099 	/* force || threshold exceeded || at the edge of the ring */
3100 	if (ring || (txq->ift_db_pending >= max) || (TXQ_AVAIL(txq) <= MAX_TX_DESC(ctx) + 2)) {
3101 
3102 		/*
3103 		 * 'npending' is used if the card's doorbell is in terms of the number of descriptors
3104 		 * pending flush (BRCM). 'pidx' is used in cases where the card's doorbeel uses the
3105 		 * producer index explicitly (INTC).
3106 		 */
3107 		dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx;
3108 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3109 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
3110 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval);
3111 
3112 		/*
3113 		 * Absent bugs there are zero packets pending so reset pending counts to zero.
3114 		 */
3115 		txq->ift_db_pending = txq->ift_npending = 0;
3116 		return (true);
3117 	}
3118 	return (false);
3119 }
3120 
3121 #ifdef PKT_DEBUG
3122 static void
3123 print_pkt(if_pkt_info_t pi)
3124 {
3125 	printf("pi len:  %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n",
3126 	       pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx);
3127 	printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n",
3128 	       pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag);
3129 	printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n",
3130 	       pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto);
3131 }
3132 #endif
3133 
3134 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO)
3135 #define IS_TX_OFFLOAD4(pi) ((pi)->ipi_csum_flags & (CSUM_IP_TCP | CSUM_IP_TSO))
3136 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO)
3137 #define IS_TX_OFFLOAD6(pi) ((pi)->ipi_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_TSO))
3138 
3139 static int
3140 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp)
3141 {
3142 	if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx;
3143 	struct ether_vlan_header *eh;
3144 	struct mbuf *m;
3145 
3146 	m = *mp;
3147 	if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) &&
3148 	    M_WRITABLE(m) == 0) {
3149 		if ((m = m_dup(m, M_NOWAIT)) == NULL) {
3150 			return (ENOMEM);
3151 		} else {
3152 			m_freem(*mp);
3153 			DBG_COUNTER_INC(tx_frees);
3154 			*mp = m;
3155 		}
3156 	}
3157 
3158 	/*
3159 	 * Determine where frame payload starts.
3160 	 * Jump over vlan headers if already present,
3161 	 * helpful for QinQ too.
3162 	 */
3163 	if (__predict_false(m->m_len < sizeof(*eh))) {
3164 		txq->ift_pullups++;
3165 		if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL))
3166 			return (ENOMEM);
3167 	}
3168 	eh = mtod(m, struct ether_vlan_header *);
3169 	if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
3170 		pi->ipi_etype = ntohs(eh->evl_proto);
3171 		pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
3172 	} else {
3173 		pi->ipi_etype = ntohs(eh->evl_encap_proto);
3174 		pi->ipi_ehdrlen = ETHER_HDR_LEN;
3175 	}
3176 
3177 	switch (pi->ipi_etype) {
3178 #ifdef INET
3179 	case ETHERTYPE_IP:
3180 	{
3181 		struct mbuf *n;
3182 		struct ip *ip = NULL;
3183 		struct tcphdr *th = NULL;
3184 		int minthlen;
3185 
3186 		minthlen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip) + sizeof(*th));
3187 		if (__predict_false(m->m_len < minthlen)) {
3188 			/*
3189 			 * if this code bloat is causing too much of a hit
3190 			 * move it to a separate function and mark it noinline
3191 			 */
3192 			if (m->m_len == pi->ipi_ehdrlen) {
3193 				n = m->m_next;
3194 				MPASS(n);
3195 				if (n->m_len >= sizeof(*ip))  {
3196 					ip = (struct ip *)n->m_data;
3197 					if (n->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3198 						th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3199 				} else {
3200 					txq->ift_pullups++;
3201 					if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
3202 						return (ENOMEM);
3203 					ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3204 				}
3205 			} else {
3206 				txq->ift_pullups++;
3207 				if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
3208 					return (ENOMEM);
3209 				ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3210 				if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3211 					th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3212 			}
3213 		} else {
3214 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3215 			if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3216 				th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3217 		}
3218 		pi->ipi_ip_hlen = ip->ip_hl << 2;
3219 		pi->ipi_ipproto = ip->ip_p;
3220 		pi->ipi_flags |= IPI_TX_IPV4;
3221 
3222 		/* TCP checksum offload may require TCP header length */
3223 		if (IS_TX_OFFLOAD4(pi)) {
3224 			if (__predict_true(pi->ipi_ipproto == IPPROTO_TCP)) {
3225 				if (__predict_false(th == NULL)) {
3226 					txq->ift_pullups++;
3227 					if (__predict_false((m = m_pullup(m, (ip->ip_hl << 2) + sizeof(*th))) == NULL))
3228 						return (ENOMEM);
3229 					th = (struct tcphdr *)((caddr_t)ip + pi->ipi_ip_hlen);
3230 				}
3231 				pi->ipi_tcp_hflags = th->th_flags;
3232 				pi->ipi_tcp_hlen = th->th_off << 2;
3233 				pi->ipi_tcp_seq = th->th_seq;
3234 			}
3235 			if (IS_TSO4(pi)) {
3236 				if (__predict_false(ip->ip_p != IPPROTO_TCP))
3237 					return (ENXIO);
3238 				/*
3239 				 * TSO always requires hardware checksum offload.
3240 				 */
3241 				pi->ipi_csum_flags |= (CSUM_IP_TCP | CSUM_IP);
3242 				th->th_sum = in_pseudo(ip->ip_src.s_addr,
3243 						       ip->ip_dst.s_addr, htons(IPPROTO_TCP));
3244 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3245 				if (sctx->isc_flags & IFLIB_TSO_INIT_IP) {
3246 					ip->ip_sum = 0;
3247 					ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz);
3248 				}
3249 			}
3250 		}
3251 		if ((sctx->isc_flags & IFLIB_NEED_ZERO_CSUM) && (pi->ipi_csum_flags & CSUM_IP))
3252                        ip->ip_sum = 0;
3253 
3254 		break;
3255 	}
3256 #endif
3257 #ifdef INET6
3258 	case ETHERTYPE_IPV6:
3259 	{
3260 		struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
3261 		struct tcphdr *th;
3262 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
3263 
3264 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
3265 			txq->ift_pullups++;
3266 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
3267 				return (ENOMEM);
3268 		}
3269 		th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
3270 
3271 		/* XXX-BZ this will go badly in case of ext hdrs. */
3272 		pi->ipi_ipproto = ip6->ip6_nxt;
3273 		pi->ipi_flags |= IPI_TX_IPV6;
3274 
3275 		/* TCP checksum offload may require TCP header length */
3276 		if (IS_TX_OFFLOAD6(pi)) {
3277 			if (pi->ipi_ipproto == IPPROTO_TCP) {
3278 				if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) {
3279 					txq->ift_pullups++;
3280 					if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL))
3281 						return (ENOMEM);
3282 				}
3283 				pi->ipi_tcp_hflags = th->th_flags;
3284 				pi->ipi_tcp_hlen = th->th_off << 2;
3285 				pi->ipi_tcp_seq = th->th_seq;
3286 			}
3287 			if (IS_TSO6(pi)) {
3288 				if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP))
3289 					return (ENXIO);
3290 				/*
3291 				 * TSO always requires hardware checksum offload.
3292 				 */
3293 				pi->ipi_csum_flags |= CSUM_IP6_TCP;
3294 				th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0);
3295 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3296 			}
3297 		}
3298 		break;
3299 	}
3300 #endif
3301 	default:
3302 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
3303 		pi->ipi_ip_hlen = 0;
3304 		break;
3305 	}
3306 	*mp = m;
3307 
3308 	return (0);
3309 }
3310 
3311 /*
3312  * If dodgy hardware rejects the scatter gather chain we've handed it
3313  * we'll need to remove the mbuf chain from ifsg_m[] before we can add the
3314  * m_defrag'd mbufs
3315  */
3316 static __noinline struct mbuf *
3317 iflib_remove_mbuf(iflib_txq_t txq)
3318 {
3319 	int ntxd, pidx;
3320 	struct mbuf *m, **ifsd_m;
3321 
3322 	ifsd_m = txq->ift_sds.ifsd_m;
3323 	ntxd = txq->ift_size;
3324 	pidx = txq->ift_pidx & (ntxd - 1);
3325 	ifsd_m = txq->ift_sds.ifsd_m;
3326 	m = ifsd_m[pidx];
3327 	ifsd_m[pidx] = NULL;
3328 	bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[pidx]);
3329 	if (txq->ift_sds.ifsd_tso_map != NULL)
3330 		bus_dmamap_unload(txq->ift_tso_buf_tag,
3331 		    txq->ift_sds.ifsd_tso_map[pidx]);
3332 #if MEMORY_LOGGING
3333 	txq->ift_dequeued++;
3334 #endif
3335 	return (m);
3336 }
3337 
3338 static inline caddr_t
3339 calc_next_txd(iflib_txq_t txq, int cidx, uint8_t qid)
3340 {
3341 	qidx_t size;
3342 	int ntxd;
3343 	caddr_t start, end, cur, next;
3344 
3345 	ntxd = txq->ift_size;
3346 	size = txq->ift_txd_size[qid];
3347 	start = txq->ift_ifdi[qid].idi_vaddr;
3348 
3349 	if (__predict_false(size == 0))
3350 		return (start);
3351 	cur = start + size*cidx;
3352 	end = start + size*ntxd;
3353 	next = CACHE_PTR_NEXT(cur);
3354 	return (next < end ? next : start);
3355 }
3356 
3357 /*
3358  * Pad an mbuf to ensure a minimum ethernet frame size.
3359  * min_frame_size is the frame size (less CRC) to pad the mbuf to
3360  */
3361 static __noinline int
3362 iflib_ether_pad(device_t dev, struct mbuf **m_head, uint16_t min_frame_size)
3363 {
3364 	/*
3365 	 * 18 is enough bytes to pad an ARP packet to 46 bytes, and
3366 	 * and ARP message is the smallest common payload I can think of
3367 	 */
3368 	static char pad[18];	/* just zeros */
3369 	int n;
3370 	struct mbuf *new_head;
3371 
3372 	if (!M_WRITABLE(*m_head)) {
3373 		new_head = m_dup(*m_head, M_NOWAIT);
3374 		if (new_head == NULL) {
3375 			m_freem(*m_head);
3376 			device_printf(dev, "cannot pad short frame, m_dup() failed");
3377 			DBG_COUNTER_INC(encap_pad_mbuf_fail);
3378 			DBG_COUNTER_INC(tx_frees);
3379 			return ENOMEM;
3380 		}
3381 		m_freem(*m_head);
3382 		*m_head = new_head;
3383 	}
3384 
3385 	for (n = min_frame_size - (*m_head)->m_pkthdr.len;
3386 	     n > 0; n -= sizeof(pad))
3387 		if (!m_append(*m_head, min(n, sizeof(pad)), pad))
3388 			break;
3389 
3390 	if (n > 0) {
3391 		m_freem(*m_head);
3392 		device_printf(dev, "cannot pad short frame\n");
3393 		DBG_COUNTER_INC(encap_pad_mbuf_fail);
3394 		DBG_COUNTER_INC(tx_frees);
3395 		return (ENOBUFS);
3396 	}
3397 
3398 	return 0;
3399 }
3400 
3401 static int
3402 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp)
3403 {
3404 	if_ctx_t		ctx;
3405 	if_shared_ctx_t		sctx;
3406 	if_softc_ctx_t		scctx;
3407 	bus_dma_tag_t		buf_tag;
3408 	bus_dma_segment_t	*segs;
3409 	struct mbuf		*m_head, **ifsd_m;
3410 	void			*next_txd;
3411 	bus_dmamap_t		map;
3412 	struct if_pkt_info	pi;
3413 	int remap = 0;
3414 	int err, nsegs, ndesc, max_segs, pidx, cidx, next, ntxd;
3415 
3416 	ctx = txq->ift_ctx;
3417 	sctx = ctx->ifc_sctx;
3418 	scctx = &ctx->ifc_softc_ctx;
3419 	segs = txq->ift_segs;
3420 	ntxd = txq->ift_size;
3421 	m_head = *m_headp;
3422 	map = NULL;
3423 
3424 	/*
3425 	 * If we're doing TSO the next descriptor to clean may be quite far ahead
3426 	 */
3427 	cidx = txq->ift_cidx;
3428 	pidx = txq->ift_pidx;
3429 	if (ctx->ifc_flags & IFC_PREFETCH) {
3430 		next = (cidx + CACHE_PTR_INCREMENT) & (ntxd-1);
3431 		if (!(ctx->ifc_flags & IFLIB_HAS_TXCQ)) {
3432 			next_txd = calc_next_txd(txq, cidx, 0);
3433 			prefetch(next_txd);
3434 		}
3435 
3436 		/* prefetch the next cache line of mbuf pointers and flags */
3437 		prefetch(&txq->ift_sds.ifsd_m[next]);
3438 		prefetch(&txq->ift_sds.ifsd_map[next]);
3439 		next = (cidx + CACHE_LINE_SIZE) & (ntxd-1);
3440 	}
3441 	map = txq->ift_sds.ifsd_map[pidx];
3442 	ifsd_m = txq->ift_sds.ifsd_m;
3443 
3444 	if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
3445 		buf_tag = txq->ift_tso_buf_tag;
3446 		max_segs = scctx->isc_tx_tso_segments_max;
3447 		map = txq->ift_sds.ifsd_tso_map[pidx];
3448 		MPASS(buf_tag != NULL);
3449 		MPASS(max_segs > 0);
3450 	} else {
3451 		buf_tag = txq->ift_buf_tag;
3452 		max_segs = scctx->isc_tx_nsegments;
3453 		map = txq->ift_sds.ifsd_map[pidx];
3454 	}
3455 	if ((sctx->isc_flags & IFLIB_NEED_ETHER_PAD) &&
3456 	    __predict_false(m_head->m_pkthdr.len < scctx->isc_min_frame_size)) {
3457 		err = iflib_ether_pad(ctx->ifc_dev, m_headp, scctx->isc_min_frame_size);
3458 		if (err) {
3459 			DBG_COUNTER_INC(encap_txd_encap_fail);
3460 			return err;
3461 		}
3462 	}
3463 	m_head = *m_headp;
3464 
3465 	pkt_info_zero(&pi);
3466 	pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG|M_BCAST|M_MCAST));
3467 	pi.ipi_pidx = pidx;
3468 	pi.ipi_qsidx = txq->ift_id;
3469 	pi.ipi_len = m_head->m_pkthdr.len;
3470 	pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags;
3471 	pi.ipi_vtag = M_HAS_VLANTAG(m_head) ? m_head->m_pkthdr.ether_vtag : 0;
3472 
3473 	/* deliberate bitwise OR to make one condition */
3474 	if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) {
3475 		if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0)) {
3476 			DBG_COUNTER_INC(encap_txd_encap_fail);
3477 			return (err);
3478 		}
3479 		m_head = *m_headp;
3480 	}
3481 
3482 retry:
3483 	err = bus_dmamap_load_mbuf_sg(buf_tag, map, m_head, segs, &nsegs,
3484 	    BUS_DMA_NOWAIT);
3485 defrag:
3486 	if (__predict_false(err)) {
3487 		switch (err) {
3488 		case EFBIG:
3489 			/* try collapse once and defrag once */
3490 			if (remap == 0) {
3491 				m_head = m_collapse(*m_headp, M_NOWAIT, max_segs);
3492 				/* try defrag if collapsing fails */
3493 				if (m_head == NULL)
3494 					remap++;
3495 			}
3496 			if (remap == 1) {
3497 				txq->ift_mbuf_defrag++;
3498 				m_head = m_defrag(*m_headp, M_NOWAIT);
3499 			}
3500 			/*
3501 			 * remap should never be >1 unless bus_dmamap_load_mbuf_sg
3502 			 * failed to map an mbuf that was run through m_defrag
3503 			 */
3504 			MPASS(remap <= 1);
3505 			if (__predict_false(m_head == NULL || remap > 1))
3506 				goto defrag_failed;
3507 			remap++;
3508 			*m_headp = m_head;
3509 			goto retry;
3510 			break;
3511 		case ENOMEM:
3512 			txq->ift_no_tx_dma_setup++;
3513 			break;
3514 		default:
3515 			txq->ift_no_tx_dma_setup++;
3516 			m_freem(*m_headp);
3517 			DBG_COUNTER_INC(tx_frees);
3518 			*m_headp = NULL;
3519 			break;
3520 		}
3521 		txq->ift_map_failed++;
3522 		DBG_COUNTER_INC(encap_load_mbuf_fail);
3523 		DBG_COUNTER_INC(encap_txd_encap_fail);
3524 		return (err);
3525 	}
3526 	ifsd_m[pidx] = m_head;
3527 	/*
3528 	 * XXX assumes a 1 to 1 relationship between segments and
3529 	 *        descriptors - this does not hold true on all drivers, e.g.
3530 	 *        cxgb
3531 	 */
3532 	if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) {
3533 		txq->ift_no_desc_avail++;
3534 		bus_dmamap_unload(buf_tag, map);
3535 		DBG_COUNTER_INC(encap_txq_avail_fail);
3536 		DBG_COUNTER_INC(encap_txd_encap_fail);
3537 		if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0)
3538 			GROUPTASK_ENQUEUE(&txq->ift_task);
3539 		return (ENOBUFS);
3540 	}
3541 	/*
3542 	 * On Intel cards we can greatly reduce the number of TX interrupts
3543 	 * we see by only setting report status on every Nth descriptor.
3544 	 * However, this also means that the driver will need to keep track
3545 	 * of the descriptors that RS was set on to check them for the DD bit.
3546 	 */
3547 	txq->ift_rs_pending += nsegs + 1;
3548 	if (txq->ift_rs_pending > TXQ_MAX_RS_DEFERRED(txq) ||
3549 	     iflib_no_tx_batch || (TXQ_AVAIL(txq) - nsegs) <= MAX_TX_DESC(ctx) + 2) {
3550 		pi.ipi_flags |= IPI_TX_INTR;
3551 		txq->ift_rs_pending = 0;
3552 	}
3553 
3554 	pi.ipi_segs = segs;
3555 	pi.ipi_nsegs = nsegs;
3556 
3557 	MPASS(pidx >= 0 && pidx < txq->ift_size);
3558 #ifdef PKT_DEBUG
3559 	print_pkt(&pi);
3560 #endif
3561 	if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) {
3562 		bus_dmamap_sync(buf_tag, map, BUS_DMASYNC_PREWRITE);
3563 		DBG_COUNTER_INC(tx_encap);
3564 		MPASS(pi.ipi_new_pidx < txq->ift_size);
3565 
3566 		ndesc = pi.ipi_new_pidx - pi.ipi_pidx;
3567 		if (pi.ipi_new_pidx < pi.ipi_pidx) {
3568 			ndesc += txq->ift_size;
3569 			txq->ift_gen = 1;
3570 		}
3571 		/*
3572 		 * drivers can need as many as
3573 		 * two sentinels
3574 		 */
3575 		MPASS(ndesc <= pi.ipi_nsegs + 2);
3576 		MPASS(pi.ipi_new_pidx != pidx);
3577 		MPASS(ndesc > 0);
3578 		txq->ift_in_use += ndesc;
3579 		txq->ift_db_pending += ndesc;
3580 
3581 		/*
3582 		 * We update the last software descriptor again here because there may
3583 		 * be a sentinel and/or there may be more mbufs than segments
3584 		 */
3585 		txq->ift_pidx = pi.ipi_new_pidx;
3586 		txq->ift_npending += pi.ipi_ndescs;
3587 	} else {
3588 		*m_headp = m_head = iflib_remove_mbuf(txq);
3589 		if (err == EFBIG) {
3590 			txq->ift_txd_encap_efbig++;
3591 			if (remap < 2) {
3592 				remap = 1;
3593 				goto defrag;
3594 			}
3595 		}
3596 		goto defrag_failed;
3597 	}
3598 	/*
3599 	 * err can't possibly be non-zero here, so we don't neet to test it
3600 	 * to see if we need to DBG_COUNTER_INC(encap_txd_encap_fail).
3601 	 */
3602 	return (err);
3603 
3604 defrag_failed:
3605 	txq->ift_mbuf_defrag_failed++;
3606 	txq->ift_map_failed++;
3607 	m_freem(*m_headp);
3608 	DBG_COUNTER_INC(tx_frees);
3609 	*m_headp = NULL;
3610 	DBG_COUNTER_INC(encap_txd_encap_fail);
3611 	return (ENOMEM);
3612 }
3613 
3614 static void
3615 iflib_tx_desc_free(iflib_txq_t txq, int n)
3616 {
3617 	uint32_t qsize, cidx, mask, gen;
3618 	struct mbuf *m, **ifsd_m;
3619 	bool do_prefetch;
3620 
3621 	cidx = txq->ift_cidx;
3622 	gen = txq->ift_gen;
3623 	qsize = txq->ift_size;
3624 	mask = qsize-1;
3625 	ifsd_m = txq->ift_sds.ifsd_m;
3626 	do_prefetch = (txq->ift_ctx->ifc_flags & IFC_PREFETCH);
3627 
3628 	while (n-- > 0) {
3629 		if (do_prefetch) {
3630 			prefetch(ifsd_m[(cidx + 3) & mask]);
3631 			prefetch(ifsd_m[(cidx + 4) & mask]);
3632 		}
3633 		if ((m = ifsd_m[cidx]) != NULL) {
3634 			prefetch(&ifsd_m[(cidx + CACHE_PTR_INCREMENT) & mask]);
3635 			if (m->m_pkthdr.csum_flags & CSUM_TSO) {
3636 				bus_dmamap_sync(txq->ift_tso_buf_tag,
3637 				    txq->ift_sds.ifsd_tso_map[cidx],
3638 				    BUS_DMASYNC_POSTWRITE);
3639 				bus_dmamap_unload(txq->ift_tso_buf_tag,
3640 				    txq->ift_sds.ifsd_tso_map[cidx]);
3641 			} else {
3642 				bus_dmamap_sync(txq->ift_buf_tag,
3643 				    txq->ift_sds.ifsd_map[cidx],
3644 				    BUS_DMASYNC_POSTWRITE);
3645 				bus_dmamap_unload(txq->ift_buf_tag,
3646 				    txq->ift_sds.ifsd_map[cidx]);
3647 			}
3648 			/* XXX we don't support any drivers that batch packets yet */
3649 			MPASS(m->m_nextpkt == NULL);
3650 			m_freem(m);
3651 			ifsd_m[cidx] = NULL;
3652 #if MEMORY_LOGGING
3653 			txq->ift_dequeued++;
3654 #endif
3655 			DBG_COUNTER_INC(tx_frees);
3656 		}
3657 		if (__predict_false(++cidx == qsize)) {
3658 			cidx = 0;
3659 			gen = 0;
3660 		}
3661 	}
3662 	txq->ift_cidx = cidx;
3663 	txq->ift_gen = gen;
3664 }
3665 
3666 static __inline int
3667 iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh)
3668 {
3669 	int reclaim;
3670 	if_ctx_t ctx = txq->ift_ctx;
3671 
3672 	KASSERT(thresh >= 0, ("invalid threshold to reclaim"));
3673 	MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size);
3674 
3675 	/*
3676 	 * Need a rate-limiting check so that this isn't called every time
3677 	 */
3678 	iflib_tx_credits_update(ctx, txq);
3679 	reclaim = DESC_RECLAIMABLE(txq);
3680 
3681 	if (reclaim <= thresh /* + MAX_TX_DESC(txq->ift_ctx) */) {
3682 #ifdef INVARIANTS
3683 		if (iflib_verbose_debug) {
3684 			printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __FUNCTION__,
3685 			       txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments,
3686 			       reclaim, thresh);
3687 		}
3688 #endif
3689 		return (0);
3690 	}
3691 	iflib_tx_desc_free(txq, reclaim);
3692 	txq->ift_cleaned += reclaim;
3693 	txq->ift_in_use -= reclaim;
3694 
3695 	return (reclaim);
3696 }
3697 
3698 static struct mbuf **
3699 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset, int remaining)
3700 {
3701 	int next, size;
3702 	struct mbuf **items;
3703 
3704 	size = r->size;
3705 	next = (cidx + CACHE_PTR_INCREMENT) & (size-1);
3706 	items = __DEVOLATILE(struct mbuf **, &r->items[0]);
3707 
3708 	prefetch(items[(cidx + offset) & (size-1)]);
3709 	if (remaining > 1) {
3710 		prefetch2cachelines(&items[next]);
3711 		prefetch2cachelines(items[(cidx + offset + 1) & (size-1)]);
3712 		prefetch2cachelines(items[(cidx + offset + 2) & (size-1)]);
3713 		prefetch2cachelines(items[(cidx + offset + 3) & (size-1)]);
3714 	}
3715 	return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (size-1)]));
3716 }
3717 
3718 static void
3719 iflib_txq_check_drain(iflib_txq_t txq, int budget)
3720 {
3721 
3722 	ifmp_ring_check_drainage(txq->ift_br, budget);
3723 }
3724 
3725 static uint32_t
3726 iflib_txq_can_drain(struct ifmp_ring *r)
3727 {
3728 	iflib_txq_t txq = r->cookie;
3729 	if_ctx_t ctx = txq->ift_ctx;
3730 
3731 	if (TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2)
3732 		return (1);
3733 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3734 	    BUS_DMASYNC_POSTREAD);
3735 	return (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id,
3736 	    false));
3737 }
3738 
3739 static uint32_t
3740 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3741 {
3742 	iflib_txq_t txq = r->cookie;
3743 	if_ctx_t ctx = txq->ift_ctx;
3744 	if_t ifp = ctx->ifc_ifp;
3745 	struct mbuf *m, **mp;
3746 	int avail, bytes_sent, skipped, count, err, i;
3747 	int mcast_sent, pkt_sent, reclaimed;
3748 	bool do_prefetch, rang, ring;
3749 
3750 	if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) ||
3751 			    !LINK_ACTIVE(ctx))) {
3752 		DBG_COUNTER_INC(txq_drain_notready);
3753 		return (0);
3754 	}
3755 	reclaimed = iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
3756 	rang = iflib_txd_db_check(txq, reclaimed && txq->ift_db_pending);
3757 	avail = IDXDIFF(pidx, cidx, r->size);
3758 
3759 	if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) {
3760 		/*
3761 		 * The driver is unloading so we need to free all pending packets.
3762 		 */
3763 		DBG_COUNTER_INC(txq_drain_flushing);
3764 		for (i = 0; i < avail; i++) {
3765 			if (__predict_true(r->items[(cidx + i) & (r->size-1)] != (void *)txq))
3766 				m_freem(r->items[(cidx + i) & (r->size-1)]);
3767 			r->items[(cidx + i) & (r->size-1)] = NULL;
3768 		}
3769 		return (avail);
3770 	}
3771 
3772 	if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3773 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3774 		CALLOUT_LOCK(txq);
3775 		callout_stop(&txq->ift_timer);
3776 		CALLOUT_UNLOCK(txq);
3777 		DBG_COUNTER_INC(txq_drain_oactive);
3778 		return (0);
3779 	}
3780 
3781 	/*
3782 	 * If we've reclaimed any packets this queue cannot be hung.
3783 	 */
3784 	if (reclaimed)
3785 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3786 	skipped = mcast_sent = bytes_sent = pkt_sent = 0;
3787 	count = MIN(avail, TX_BATCH_SIZE);
3788 #ifdef INVARIANTS
3789 	if (iflib_verbose_debug)
3790 		printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __FUNCTION__,
3791 		       avail, ctx->ifc_flags, TXQ_AVAIL(txq));
3792 #endif
3793 	do_prefetch = (ctx->ifc_flags & IFC_PREFETCH);
3794 	err = 0;
3795 	for (i = 0; i < count && TXQ_AVAIL(txq) >= MAX_TX_DESC(ctx) + 2; i++) {
3796 		int rem = do_prefetch ? count - i : 0;
3797 
3798 		mp = _ring_peek_one(r, cidx, i, rem);
3799 		MPASS(mp != NULL && *mp != NULL);
3800 
3801 		/*
3802 		 * Completion interrupts will use the address of the txq
3803 		 * as a sentinel to enqueue _something_ in order to acquire
3804 		 * the lock on the mp_ring (there's no direct lock call).
3805 		 * We obviously whave to check for these sentinel cases
3806 		 * and skip them.
3807 		 */
3808 		if (__predict_false(*mp == (struct mbuf *)txq)) {
3809 			skipped++;
3810 			continue;
3811 		}
3812 		err = iflib_encap(txq, mp);
3813 		if (__predict_false(err)) {
3814 			/* no room - bail out */
3815 			if (err == ENOBUFS)
3816 				break;
3817 			skipped++;
3818 			/* we can't send this packet - skip it */
3819 			continue;
3820 		}
3821 		pkt_sent++;
3822 		m = *mp;
3823 		DBG_COUNTER_INC(tx_sent);
3824 		bytes_sent += m->m_pkthdr.len;
3825 		mcast_sent += !!(m->m_flags & M_MCAST);
3826 
3827 		if (__predict_false(!(ifp->if_drv_flags & IFF_DRV_RUNNING)))
3828 			break;
3829 		ETHER_BPF_MTAP(ifp, m);
3830 		rang = iflib_txd_db_check(txq, false);
3831 	}
3832 
3833 	/* deliberate use of bitwise or to avoid gratuitous short-circuit */
3834 	ring = rang ? false  : (iflib_min_tx_latency | err);
3835 	iflib_txd_db_check(txq, ring);
3836 	if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
3837 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
3838 	if (mcast_sent)
3839 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
3840 #ifdef INVARIANTS
3841 	if (iflib_verbose_debug)
3842 		printf("consumed=%d\n", skipped + pkt_sent);
3843 #endif
3844 	return (skipped + pkt_sent);
3845 }
3846 
3847 static uint32_t
3848 iflib_txq_drain_always(struct ifmp_ring *r)
3849 {
3850 	return (1);
3851 }
3852 
3853 static uint32_t
3854 iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3855 {
3856 	int i, avail;
3857 	struct mbuf **mp;
3858 	iflib_txq_t txq;
3859 
3860 	txq = r->cookie;
3861 
3862 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3863 	CALLOUT_LOCK(txq);
3864 	callout_stop(&txq->ift_timer);
3865 	CALLOUT_UNLOCK(txq);
3866 
3867 	avail = IDXDIFF(pidx, cidx, r->size);
3868 	for (i = 0; i < avail; i++) {
3869 		mp = _ring_peek_one(r, cidx, i, avail - i);
3870 		if (__predict_false(*mp == (struct mbuf *)txq))
3871 			continue;
3872 		m_freem(*mp);
3873 		DBG_COUNTER_INC(tx_frees);
3874 	}
3875 	MPASS(ifmp_ring_is_stalled(r) == 0);
3876 	return (avail);
3877 }
3878 
3879 static void
3880 iflib_ifmp_purge(iflib_txq_t txq)
3881 {
3882 	struct ifmp_ring *r;
3883 
3884 	r = txq->ift_br;
3885 	r->drain = iflib_txq_drain_free;
3886 	r->can_drain = iflib_txq_drain_always;
3887 
3888 	ifmp_ring_check_drainage(r, r->size);
3889 
3890 	r->drain = iflib_txq_drain;
3891 	r->can_drain = iflib_txq_can_drain;
3892 }
3893 
3894 static void
3895 _task_fn_tx(void *context)
3896 {
3897 	iflib_txq_t txq = context;
3898 	if_ctx_t ctx = txq->ift_ctx;
3899 	if_t ifp = ctx->ifc_ifp;
3900 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
3901 
3902 #ifdef IFLIB_DIAGNOSTICS
3903 	txq->ift_cpu_exec_count[curcpu]++;
3904 #endif
3905 	if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
3906 		return;
3907 #ifdef DEV_NETMAP
3908 	if ((if_getcapenable(ifp) & IFCAP_NETMAP) &&
3909 	    netmap_tx_irq(ifp, txq->ift_id))
3910 		goto skip_ifmp;
3911 #endif
3912 #ifdef ALTQ
3913 	if (ALTQ_IS_ENABLED(&ifp->if_snd))
3914 		iflib_altq_if_start(ifp);
3915 #endif
3916 	if (txq->ift_db_pending)
3917 		ifmp_ring_enqueue(txq->ift_br, (void **)&txq, 1, TX_BATCH_SIZE, abdicate);
3918 	else if (!abdicate)
3919 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3920 	/*
3921 	 * When abdicating, we always need to check drainage, not just when we don't enqueue
3922 	 */
3923 	if (abdicate)
3924 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3925 #ifdef DEV_NETMAP
3926 skip_ifmp:
3927 #endif
3928 	if (ctx->ifc_flags & IFC_LEGACY)
3929 		IFDI_INTR_ENABLE(ctx);
3930 	else
3931 		IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3932 }
3933 
3934 static void
3935 _task_fn_rx(void *context)
3936 {
3937 	iflib_rxq_t rxq = context;
3938 	if_ctx_t ctx = rxq->ifr_ctx;
3939 	uint8_t more;
3940 	uint16_t budget;
3941 #ifdef DEV_NETMAP
3942 	u_int work = 0;
3943 	int nmirq;
3944 #endif
3945 
3946 #ifdef IFLIB_DIAGNOSTICS
3947 	rxq->ifr_cpu_exec_count[curcpu]++;
3948 #endif
3949 	DBG_COUNTER_INC(task_fn_rxs);
3950 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3951 		return;
3952 #ifdef DEV_NETMAP
3953 	nmirq = netmap_rx_irq(ctx->ifc_ifp, rxq->ifr_id, &work);
3954 	if (nmirq != NM_IRQ_PASS) {
3955 		more = (nmirq == NM_IRQ_RESCHED) ? IFLIB_RXEOF_MORE : 0;
3956 		goto skip_rxeof;
3957 	}
3958 #endif
3959 	budget = ctx->ifc_sysctl_rx_budget;
3960 	if (budget == 0)
3961 		budget = 16;	/* XXX */
3962 	more = iflib_rxeof(rxq, budget);
3963 #ifdef DEV_NETMAP
3964 skip_rxeof:
3965 #endif
3966 	if ((more & IFLIB_RXEOF_MORE) == 0) {
3967 		if (ctx->ifc_flags & IFC_LEGACY)
3968 			IFDI_INTR_ENABLE(ctx);
3969 		else
3970 			IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
3971 		DBG_COUNTER_INC(rx_intr_enables);
3972 	}
3973 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3974 		return;
3975 
3976 	if (more & IFLIB_RXEOF_MORE)
3977 		GROUPTASK_ENQUEUE(&rxq->ifr_task);
3978 	else if (more & IFLIB_RXEOF_EMPTY)
3979 		callout_reset_curcpu(&rxq->ifr_watchdog, 1, &_task_fn_rx_watchdog, rxq);
3980 }
3981 
3982 static void
3983 _task_fn_admin(void *context)
3984 {
3985 	if_ctx_t ctx = context;
3986 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
3987 	iflib_txq_t txq;
3988 	int i;
3989 	bool oactive, running, do_reset, do_watchdog, in_detach;
3990 
3991 	STATE_LOCK(ctx);
3992 	running = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING);
3993 	oactive = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE);
3994 	do_reset = (ctx->ifc_flags & IFC_DO_RESET);
3995 	do_watchdog = (ctx->ifc_flags & IFC_DO_WATCHDOG);
3996 	in_detach = (ctx->ifc_flags & IFC_IN_DETACH);
3997 	ctx->ifc_flags &= ~(IFC_DO_RESET|IFC_DO_WATCHDOG);
3998 	STATE_UNLOCK(ctx);
3999 
4000 	if ((!running && !oactive) && !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
4001 		return;
4002 	if (in_detach)
4003 		return;
4004 
4005 	CTX_LOCK(ctx);
4006 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
4007 		CALLOUT_LOCK(txq);
4008 		callout_stop(&txq->ift_timer);
4009 		CALLOUT_UNLOCK(txq);
4010 	}
4011 	if (do_watchdog) {
4012 		ctx->ifc_watchdog_events++;
4013 		IFDI_WATCHDOG_RESET(ctx);
4014 	}
4015 	IFDI_UPDATE_ADMIN_STATUS(ctx);
4016 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
4017 		callout_reset_on(&txq->ift_timer, iflib_timer_default, iflib_timer, txq,
4018 		    txq->ift_timer.c_cpu);
4019 	}
4020 	IFDI_LINK_INTR_ENABLE(ctx);
4021 	if (do_reset)
4022 		iflib_if_init_locked(ctx);
4023 	CTX_UNLOCK(ctx);
4024 
4025 	if (LINK_ACTIVE(ctx) == 0)
4026 		return;
4027 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
4028 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
4029 }
4030 
4031 static void
4032 _task_fn_iov(void *context)
4033 {
4034 	if_ctx_t ctx = context;
4035 
4036 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING) &&
4037 	    !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
4038 		return;
4039 
4040 	CTX_LOCK(ctx);
4041 	IFDI_VFLR_HANDLE(ctx);
4042 	CTX_UNLOCK(ctx);
4043 }
4044 
4045 static int
4046 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)
4047 {
4048 	int err;
4049 	if_int_delay_info_t info;
4050 	if_ctx_t ctx;
4051 
4052 	info = (if_int_delay_info_t)arg1;
4053 	ctx = info->iidi_ctx;
4054 	info->iidi_req = req;
4055 	info->iidi_oidp = oidp;
4056 	CTX_LOCK(ctx);
4057 	err = IFDI_SYSCTL_INT_DELAY(ctx, info);
4058 	CTX_UNLOCK(ctx);
4059 	return (err);
4060 }
4061 
4062 /*********************************************************************
4063  *
4064  *  IFNET FUNCTIONS
4065  *
4066  **********************************************************************/
4067 
4068 static void
4069 iflib_if_init_locked(if_ctx_t ctx)
4070 {
4071 	iflib_stop(ctx);
4072 	iflib_init_locked(ctx);
4073 }
4074 
4075 static void
4076 iflib_if_init(void *arg)
4077 {
4078 	if_ctx_t ctx = arg;
4079 
4080 	CTX_LOCK(ctx);
4081 	iflib_if_init_locked(ctx);
4082 	CTX_UNLOCK(ctx);
4083 }
4084 
4085 static int
4086 iflib_if_transmit(if_t ifp, struct mbuf *m)
4087 {
4088 	if_ctx_t	ctx = if_getsoftc(ifp);
4089 
4090 	iflib_txq_t txq;
4091 	int err, qidx;
4092 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
4093 
4094 	if (__predict_false((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) {
4095 		DBG_COUNTER_INC(tx_frees);
4096 		m_freem(m);
4097 		return (ENETDOWN);
4098 	}
4099 
4100 	MPASS(m->m_nextpkt == NULL);
4101 	/* ALTQ-enabled interfaces always use queue 0. */
4102 	qidx = 0;
4103 	if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m) && !ALTQ_IS_ENABLED(&ifp->if_snd))
4104 		qidx = QIDX(ctx, m);
4105 	/*
4106 	 * XXX calculate buf_ring based on flowid (divvy up bits?)
4107 	 */
4108 	txq = &ctx->ifc_txqs[qidx];
4109 
4110 #ifdef DRIVER_BACKPRESSURE
4111 	if (txq->ift_closed) {
4112 		while (m != NULL) {
4113 			next = m->m_nextpkt;
4114 			m->m_nextpkt = NULL;
4115 			m_freem(m);
4116 			DBG_COUNTER_INC(tx_frees);
4117 			m = next;
4118 		}
4119 		return (ENOBUFS);
4120 	}
4121 #endif
4122 #ifdef notyet
4123 	qidx = count = 0;
4124 	mp = marr;
4125 	next = m;
4126 	do {
4127 		count++;
4128 		next = next->m_nextpkt;
4129 	} while (next != NULL);
4130 
4131 	if (count > nitems(marr))
4132 		if ((mp = malloc(count*sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) {
4133 			/* XXX check nextpkt */
4134 			m_freem(m);
4135 			/* XXX simplify for now */
4136 			DBG_COUNTER_INC(tx_frees);
4137 			return (ENOBUFS);
4138 		}
4139 	for (next = m, i = 0; next != NULL; i++) {
4140 		mp[i] = next;
4141 		next = next->m_nextpkt;
4142 		mp[i]->m_nextpkt = NULL;
4143 	}
4144 #endif
4145 	DBG_COUNTER_INC(tx_seen);
4146 	err = ifmp_ring_enqueue(txq->ift_br, (void **)&m, 1, TX_BATCH_SIZE, abdicate);
4147 
4148 	if (abdicate)
4149 		GROUPTASK_ENQUEUE(&txq->ift_task);
4150  	if (err) {
4151 		if (!abdicate)
4152 			GROUPTASK_ENQUEUE(&txq->ift_task);
4153 		/* support forthcoming later */
4154 #ifdef DRIVER_BACKPRESSURE
4155 		txq->ift_closed = TRUE;
4156 #endif
4157 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
4158 		m_freem(m);
4159 		DBG_COUNTER_INC(tx_frees);
4160 	}
4161 
4162 	return (err);
4163 }
4164 
4165 #ifdef ALTQ
4166 /*
4167  * The overall approach to integrating iflib with ALTQ is to continue to use
4168  * the iflib mp_ring machinery between the ALTQ queue(s) and the hardware
4169  * ring.  Technically, when using ALTQ, queueing to an intermediate mp_ring
4170  * is redundant/unnecessary, but doing so minimizes the amount of
4171  * ALTQ-specific code required in iflib.  It is assumed that the overhead of
4172  * redundantly queueing to an intermediate mp_ring is swamped by the
4173  * performance limitations inherent in using ALTQ.
4174  *
4175  * When ALTQ support is compiled in, all iflib drivers will use a transmit
4176  * routine, iflib_altq_if_transmit(), that checks if ALTQ is enabled for the
4177  * given interface.  If ALTQ is enabled for an interface, then all
4178  * transmitted packets for that interface will be submitted to the ALTQ
4179  * subsystem via IFQ_ENQUEUE().  We don't use the legacy if_transmit()
4180  * implementation because it uses IFQ_HANDOFF(), which will duplicatively
4181  * update stats that the iflib machinery handles, and which is sensitve to
4182  * the disused IFF_DRV_OACTIVE flag.  Additionally, iflib_altq_if_start()
4183  * will be installed as the start routine for use by ALTQ facilities that
4184  * need to trigger queue drains on a scheduled basis.
4185  *
4186  */
4187 static void
4188 iflib_altq_if_start(if_t ifp)
4189 {
4190 	struct ifaltq *ifq = &ifp->if_snd;
4191 	struct mbuf *m;
4192 
4193 	IFQ_LOCK(ifq);
4194 	IFQ_DEQUEUE_NOLOCK(ifq, m);
4195 	while (m != NULL) {
4196 		iflib_if_transmit(ifp, m);
4197 		IFQ_DEQUEUE_NOLOCK(ifq, m);
4198 	}
4199 	IFQ_UNLOCK(ifq);
4200 }
4201 
4202 static int
4203 iflib_altq_if_transmit(if_t ifp, struct mbuf *m)
4204 {
4205 	int err;
4206 
4207 	if (ALTQ_IS_ENABLED(&ifp->if_snd)) {
4208 		IFQ_ENQUEUE(&ifp->if_snd, m, err);
4209 		if (err == 0)
4210 			iflib_altq_if_start(ifp);
4211 	} else
4212 		err = iflib_if_transmit(ifp, m);
4213 
4214 	return (err);
4215 }
4216 #endif /* ALTQ */
4217 
4218 static void
4219 iflib_if_qflush(if_t ifp)
4220 {
4221 	if_ctx_t ctx = if_getsoftc(ifp);
4222 	iflib_txq_t txq = ctx->ifc_txqs;
4223 	int i;
4224 
4225 	STATE_LOCK(ctx);
4226 	ctx->ifc_flags |= IFC_QFLUSH;
4227 	STATE_UNLOCK(ctx);
4228 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
4229 		while (!(ifmp_ring_is_idle(txq->ift_br) || ifmp_ring_is_stalled(txq->ift_br)))
4230 			iflib_txq_check_drain(txq, 0);
4231 	STATE_LOCK(ctx);
4232 	ctx->ifc_flags &= ~IFC_QFLUSH;
4233 	STATE_UNLOCK(ctx);
4234 
4235 	/*
4236 	 * When ALTQ is enabled, this will also take care of purging the
4237 	 * ALTQ queue(s).
4238 	 */
4239 	if_qflush(ifp);
4240 }
4241 
4242 #define IFCAP_FLAGS (IFCAP_HWCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \
4243 		     IFCAP_TSO | IFCAP_VLAN_HWTAGGING | IFCAP_HWSTATS | \
4244 		     IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | \
4245 		     IFCAP_VLAN_HWTSO | IFCAP_VLAN_HWCSUM | IFCAP_MEXTPG)
4246 
4247 static int
4248 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data)
4249 {
4250 	if_ctx_t ctx = if_getsoftc(ifp);
4251 	struct ifreq	*ifr = (struct ifreq *)data;
4252 #if defined(INET) || defined(INET6)
4253 	struct ifaddr	*ifa = (struct ifaddr *)data;
4254 #endif
4255 	bool		avoid_reset = false;
4256 	int		err = 0, reinit = 0, bits;
4257 
4258 	switch (command) {
4259 	case SIOCSIFADDR:
4260 #ifdef INET
4261 		if (ifa->ifa_addr->sa_family == AF_INET)
4262 			avoid_reset = true;
4263 #endif
4264 #ifdef INET6
4265 		if (ifa->ifa_addr->sa_family == AF_INET6)
4266 			avoid_reset = true;
4267 #endif
4268 		/*
4269 		** Calling init results in link renegotiation,
4270 		** so we avoid doing it when possible.
4271 		*/
4272 		if (avoid_reset) {
4273 			if_setflagbits(ifp, IFF_UP,0);
4274 			if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
4275 				reinit = 1;
4276 #ifdef INET
4277 			if (!(if_getflags(ifp) & IFF_NOARP))
4278 				arp_ifinit(ifp, ifa);
4279 #endif
4280 		} else
4281 			err = ether_ioctl(ifp, command, data);
4282 		break;
4283 	case SIOCSIFMTU:
4284 		CTX_LOCK(ctx);
4285 		if (ifr->ifr_mtu == if_getmtu(ifp)) {
4286 			CTX_UNLOCK(ctx);
4287 			break;
4288 		}
4289 		bits = if_getdrvflags(ifp);
4290 		/* stop the driver and free any clusters before proceeding */
4291 		iflib_stop(ctx);
4292 
4293 		if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) {
4294 			STATE_LOCK(ctx);
4295 			if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size)
4296 				ctx->ifc_flags |= IFC_MULTISEG;
4297 			else
4298 				ctx->ifc_flags &= ~IFC_MULTISEG;
4299 			STATE_UNLOCK(ctx);
4300 			err = if_setmtu(ifp, ifr->ifr_mtu);
4301 		}
4302 		iflib_init_locked(ctx);
4303 		STATE_LOCK(ctx);
4304 		if_setdrvflags(ifp, bits);
4305 		STATE_UNLOCK(ctx);
4306 		CTX_UNLOCK(ctx);
4307 		break;
4308 	case SIOCSIFFLAGS:
4309 		CTX_LOCK(ctx);
4310 		if (if_getflags(ifp) & IFF_UP) {
4311 			if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4312 				if ((if_getflags(ifp) ^ ctx->ifc_if_flags) &
4313 				    (IFF_PROMISC | IFF_ALLMULTI)) {
4314 					CTX_UNLOCK(ctx);
4315 					err = IFDI_PROMISC_SET(ctx, if_getflags(ifp));
4316 					CTX_LOCK(ctx);
4317 				}
4318 			} else
4319 				reinit = 1;
4320 		} else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4321 			iflib_stop(ctx);
4322 		}
4323 		ctx->ifc_if_flags = if_getflags(ifp);
4324 		CTX_UNLOCK(ctx);
4325 		break;
4326 	case SIOCADDMULTI:
4327 	case SIOCDELMULTI:
4328 		if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4329 			CTX_LOCK(ctx);
4330 			IFDI_INTR_DISABLE(ctx);
4331 			IFDI_MULTI_SET(ctx);
4332 			IFDI_INTR_ENABLE(ctx);
4333 			CTX_UNLOCK(ctx);
4334 		}
4335 		break;
4336 	case SIOCSIFMEDIA:
4337 		CTX_LOCK(ctx);
4338 		IFDI_MEDIA_SET(ctx);
4339 		CTX_UNLOCK(ctx);
4340 		/* FALLTHROUGH */
4341 	case SIOCGIFMEDIA:
4342 	case SIOCGIFXMEDIA:
4343 		err = ifmedia_ioctl(ifp, ifr, ctx->ifc_mediap, command);
4344 		break;
4345 	case SIOCGI2C:
4346 	{
4347 		struct ifi2creq i2c;
4348 
4349 		err = copyin(ifr_data_get_ptr(ifr), &i2c, sizeof(i2c));
4350 		if (err != 0)
4351 			break;
4352 		if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) {
4353 			err = EINVAL;
4354 			break;
4355 		}
4356 		if (i2c.len > sizeof(i2c.data)) {
4357 			err = EINVAL;
4358 			break;
4359 		}
4360 
4361 		if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0)
4362 			err = copyout(&i2c, ifr_data_get_ptr(ifr),
4363 			    sizeof(i2c));
4364 		break;
4365 	}
4366 	case SIOCSIFCAP:
4367 	{
4368 		int mask, setmask, oldmask;
4369 
4370 		oldmask = if_getcapenable(ifp);
4371 		mask = ifr->ifr_reqcap ^ oldmask;
4372 		mask &= ctx->ifc_softc_ctx.isc_capabilities | IFCAP_MEXTPG;
4373 		setmask = 0;
4374 #ifdef TCP_OFFLOAD
4375 		setmask |= mask & (IFCAP_TOE4|IFCAP_TOE6);
4376 #endif
4377 		setmask |= (mask & IFCAP_FLAGS);
4378 		setmask |= (mask & IFCAP_WOL);
4379 
4380 		/*
4381 		 * If any RX csum has changed, change all the ones that
4382 		 * are supported by the driver.
4383 		 */
4384 		if (setmask & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6)) {
4385 			setmask |= ctx->ifc_softc_ctx.isc_capabilities &
4386 			    (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6);
4387 		}
4388 
4389 		/*
4390 		 * want to ensure that traffic has stopped before we change any of the flags
4391 		 */
4392 		if (setmask) {
4393 			CTX_LOCK(ctx);
4394 			bits = if_getdrvflags(ifp);
4395 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4396 				iflib_stop(ctx);
4397 			STATE_LOCK(ctx);
4398 			if_togglecapenable(ifp, setmask);
4399 			STATE_UNLOCK(ctx);
4400 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4401 				iflib_init_locked(ctx);
4402 			STATE_LOCK(ctx);
4403 			if_setdrvflags(ifp, bits);
4404 			STATE_UNLOCK(ctx);
4405 			CTX_UNLOCK(ctx);
4406 		}
4407 		if_vlancap(ifp);
4408 		break;
4409 	}
4410 	case SIOCGPRIVATE_0:
4411 	case SIOCSDRVSPEC:
4412 	case SIOCGDRVSPEC:
4413 		CTX_LOCK(ctx);
4414 		err = IFDI_PRIV_IOCTL(ctx, command, data);
4415 		CTX_UNLOCK(ctx);
4416 		break;
4417 	default:
4418 		err = ether_ioctl(ifp, command, data);
4419 		break;
4420 	}
4421 	if (reinit)
4422 		iflib_if_init(ctx);
4423 	return (err);
4424 }
4425 
4426 static uint64_t
4427 iflib_if_get_counter(if_t ifp, ift_counter cnt)
4428 {
4429 	if_ctx_t ctx = if_getsoftc(ifp);
4430 
4431 	return (IFDI_GET_COUNTER(ctx, cnt));
4432 }
4433 
4434 /*********************************************************************
4435  *
4436  *  OTHER FUNCTIONS EXPORTED TO THE STACK
4437  *
4438  **********************************************************************/
4439 
4440 static void
4441 iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag)
4442 {
4443 	if_ctx_t ctx = if_getsoftc(ifp);
4444 
4445 	if ((void *)ctx != arg)
4446 		return;
4447 
4448 	if ((vtag == 0) || (vtag > 4095))
4449 		return;
4450 
4451 	if (iflib_in_detach(ctx))
4452 		return;
4453 
4454 	CTX_LOCK(ctx);
4455 	/* Driver may need all untagged packets to be flushed */
4456 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4457 		iflib_stop(ctx);
4458 	IFDI_VLAN_REGISTER(ctx, vtag);
4459 	/* Re-init to load the changes, if required */
4460 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4461 		iflib_init_locked(ctx);
4462 	CTX_UNLOCK(ctx);
4463 }
4464 
4465 static void
4466 iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag)
4467 {
4468 	if_ctx_t ctx = if_getsoftc(ifp);
4469 
4470 	if ((void *)ctx != arg)
4471 		return;
4472 
4473 	if ((vtag == 0) || (vtag > 4095))
4474 		return;
4475 
4476 	CTX_LOCK(ctx);
4477 	/* Driver may need all tagged packets to be flushed */
4478 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4479 		iflib_stop(ctx);
4480 	IFDI_VLAN_UNREGISTER(ctx, vtag);
4481 	/* Re-init to load the changes, if required */
4482 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4483 		iflib_init_locked(ctx);
4484 	CTX_UNLOCK(ctx);
4485 }
4486 
4487 static void
4488 iflib_led_func(void *arg, int onoff)
4489 {
4490 	if_ctx_t ctx = arg;
4491 
4492 	CTX_LOCK(ctx);
4493 	IFDI_LED_FUNC(ctx, onoff);
4494 	CTX_UNLOCK(ctx);
4495 }
4496 
4497 /*********************************************************************
4498  *
4499  *  BUS FUNCTION DEFINITIONS
4500  *
4501  **********************************************************************/
4502 
4503 int
4504 iflib_device_probe(device_t dev)
4505 {
4506 	const pci_vendor_info_t *ent;
4507 	if_shared_ctx_t sctx;
4508 	uint16_t pci_device_id, pci_rev_id, pci_subdevice_id, pci_subvendor_id;
4509 	uint16_t pci_vendor_id;
4510 
4511 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4512 		return (ENOTSUP);
4513 
4514 	pci_vendor_id = pci_get_vendor(dev);
4515 	pci_device_id = pci_get_device(dev);
4516 	pci_subvendor_id = pci_get_subvendor(dev);
4517 	pci_subdevice_id = pci_get_subdevice(dev);
4518 	pci_rev_id = pci_get_revid(dev);
4519 	if (sctx->isc_parse_devinfo != NULL)
4520 		sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id);
4521 
4522 	ent = sctx->isc_vendor_info;
4523 	while (ent->pvi_vendor_id != 0) {
4524 		if (pci_vendor_id != ent->pvi_vendor_id) {
4525 			ent++;
4526 			continue;
4527 		}
4528 		if ((pci_device_id == ent->pvi_device_id) &&
4529 		    ((pci_subvendor_id == ent->pvi_subvendor_id) ||
4530 		     (ent->pvi_subvendor_id == 0)) &&
4531 		    ((pci_subdevice_id == ent->pvi_subdevice_id) ||
4532 		     (ent->pvi_subdevice_id == 0)) &&
4533 		    ((pci_rev_id == ent->pvi_rev_id) ||
4534 		     (ent->pvi_rev_id == 0))) {
4535 			device_set_desc_copy(dev, ent->pvi_name);
4536 			/* this needs to be changed to zero if the bus probing code
4537 			 * ever stops re-probing on best match because the sctx
4538 			 * may have its values over written by register calls
4539 			 * in subsequent probes
4540 			 */
4541 			return (BUS_PROBE_DEFAULT);
4542 		}
4543 		ent++;
4544 	}
4545 	return (ENXIO);
4546 }
4547 
4548 int
4549 iflib_device_probe_vendor(device_t dev)
4550 {
4551 	int probe;
4552 
4553 	probe = iflib_device_probe(dev);
4554 	if (probe == BUS_PROBE_DEFAULT)
4555 		return (BUS_PROBE_VENDOR);
4556 	else
4557 		return (probe);
4558 }
4559 
4560 static void
4561 iflib_reset_qvalues(if_ctx_t ctx)
4562 {
4563 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4564 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4565 	device_t dev = ctx->ifc_dev;
4566 	int i;
4567 
4568 	if (ctx->ifc_sysctl_ntxqs != 0)
4569 		scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs;
4570 	if (ctx->ifc_sysctl_nrxqs != 0)
4571 		scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs;
4572 
4573 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4574 		if (ctx->ifc_sysctl_ntxds[i] != 0)
4575 			scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i];
4576 		else
4577 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4578 	}
4579 
4580 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4581 		if (ctx->ifc_sysctl_nrxds[i] != 0)
4582 			scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i];
4583 		else
4584 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4585 	}
4586 
4587 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4588 		if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) {
4589 			device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n",
4590 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]);
4591 			scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i];
4592 		}
4593 		if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) {
4594 			device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n",
4595 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]);
4596 			scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i];
4597 		}
4598 		if (!powerof2(scctx->isc_nrxd[i])) {
4599 			device_printf(dev, "nrxd%d: %d is not a power of 2 - using default value of %d\n",
4600 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_default[i]);
4601 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4602 		}
4603 	}
4604 
4605 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4606 		if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) {
4607 			device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n",
4608 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]);
4609 			scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i];
4610 		}
4611 		if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) {
4612 			device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n",
4613 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]);
4614 			scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i];
4615 		}
4616 		if (!powerof2(scctx->isc_ntxd[i])) {
4617 			device_printf(dev, "ntxd%d: %d is not a power of 2 - using default value of %d\n",
4618 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_default[i]);
4619 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4620 		}
4621 	}
4622 }
4623 
4624 static void
4625 iflib_add_pfil(if_ctx_t ctx)
4626 {
4627 	struct pfil_head *pfil;
4628 	struct pfil_head_args pa;
4629 	iflib_rxq_t rxq;
4630 	int i;
4631 
4632 	pa.pa_version = PFIL_VERSION;
4633 	pa.pa_flags = PFIL_IN;
4634 	pa.pa_type = PFIL_TYPE_ETHERNET;
4635 	pa.pa_headname = ctx->ifc_ifp->if_xname;
4636 	pfil = pfil_head_register(&pa);
4637 
4638 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4639 		rxq->pfil = pfil;
4640 	}
4641 }
4642 
4643 static void
4644 iflib_rem_pfil(if_ctx_t ctx)
4645 {
4646 	struct pfil_head *pfil;
4647 	iflib_rxq_t rxq;
4648 	int i;
4649 
4650 	rxq = ctx->ifc_rxqs;
4651 	pfil = rxq->pfil;
4652 	for (i = 0; i < NRXQSETS(ctx); i++, rxq++) {
4653 		rxq->pfil = NULL;
4654 	}
4655 	pfil_head_unregister(pfil);
4656 }
4657 
4658 static uint16_t
4659 get_ctx_core_offset(if_ctx_t ctx)
4660 {
4661 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4662 	struct cpu_offset *op;
4663 	uint16_t qc;
4664 	uint16_t ret = ctx->ifc_sysctl_core_offset;
4665 
4666 	if (ret != CORE_OFFSET_UNSPECIFIED)
4667 		return (ret);
4668 
4669 	if (ctx->ifc_sysctl_separate_txrx)
4670 		qc = scctx->isc_ntxqsets + scctx->isc_nrxqsets;
4671 	else
4672 		qc = max(scctx->isc_ntxqsets, scctx->isc_nrxqsets);
4673 
4674 	mtx_lock(&cpu_offset_mtx);
4675 	SLIST_FOREACH(op, &cpu_offsets, entries) {
4676 		if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) {
4677 			ret = op->offset;
4678 			op->offset += qc;
4679 			MPASS(op->refcount < UINT_MAX);
4680 			op->refcount++;
4681 			break;
4682 		}
4683 	}
4684 	if (ret == CORE_OFFSET_UNSPECIFIED) {
4685 		ret = 0;
4686 		op = malloc(sizeof(struct cpu_offset), M_IFLIB,
4687 		    M_NOWAIT | M_ZERO);
4688 		if (op == NULL) {
4689 			device_printf(ctx->ifc_dev,
4690 			    "allocation for cpu offset failed.\n");
4691 		} else {
4692 			op->offset = qc;
4693 			op->refcount = 1;
4694 			CPU_COPY(&ctx->ifc_cpus, &op->set);
4695 			SLIST_INSERT_HEAD(&cpu_offsets, op, entries);
4696 		}
4697 	}
4698 	mtx_unlock(&cpu_offset_mtx);
4699 
4700 	return (ret);
4701 }
4702 
4703 static void
4704 unref_ctx_core_offset(if_ctx_t ctx)
4705 {
4706 	struct cpu_offset *op, *top;
4707 
4708 	mtx_lock(&cpu_offset_mtx);
4709 	SLIST_FOREACH_SAFE(op, &cpu_offsets, entries, top) {
4710 		if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) {
4711 			MPASS(op->refcount > 0);
4712 			op->refcount--;
4713 			if (op->refcount == 0) {
4714 				SLIST_REMOVE(&cpu_offsets, op, cpu_offset, entries);
4715 				free(op, M_IFLIB);
4716 			}
4717 			break;
4718 		}
4719 	}
4720 	mtx_unlock(&cpu_offset_mtx);
4721 }
4722 
4723 int
4724 iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp)
4725 {
4726 	if_ctx_t ctx;
4727 	if_t ifp;
4728 	if_softc_ctx_t scctx;
4729 	kobjop_desc_t kobj_desc;
4730 	kobj_method_t *kobj_method;
4731 	int err, msix, rid;
4732 	int num_txd, num_rxd;
4733 
4734 	ctx = malloc(sizeof(* ctx), M_IFLIB, M_WAITOK|M_ZERO);
4735 
4736 	if (sc == NULL) {
4737 		sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4738 		device_set_softc(dev, ctx);
4739 		ctx->ifc_flags |= IFC_SC_ALLOCATED;
4740 	}
4741 
4742 	ctx->ifc_sctx = sctx;
4743 	ctx->ifc_dev = dev;
4744 	ctx->ifc_softc = sc;
4745 
4746 	if ((err = iflib_register(ctx)) != 0) {
4747 		device_printf(dev, "iflib_register failed %d\n", err);
4748 		goto fail_ctx_free;
4749 	}
4750 	iflib_add_device_sysctl_pre(ctx);
4751 
4752 	scctx = &ctx->ifc_softc_ctx;
4753 	ifp = ctx->ifc_ifp;
4754 
4755 	iflib_reset_qvalues(ctx);
4756 	CTX_LOCK(ctx);
4757 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4758 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4759 		goto fail_unlock;
4760 	}
4761 	_iflib_pre_assert(scctx);
4762 	ctx->ifc_txrx = *scctx->isc_txrx;
4763 
4764 	MPASS(scctx->isc_dma_width <= flsll(BUS_SPACE_MAXADDR));
4765 
4766 	if (sctx->isc_flags & IFLIB_DRIVER_MEDIA)
4767 		ctx->ifc_mediap = scctx->isc_media;
4768 
4769 #ifdef INVARIANTS
4770 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
4771 		MPASS(scctx->isc_tx_csum_flags);
4772 #endif
4773 
4774 	if_setcapabilities(ifp,
4775 	    scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_MEXTPG);
4776 	if_setcapenable(ifp,
4777 	    scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_MEXTPG);
4778 
4779 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4780 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4781 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4782 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4783 
4784 	num_txd = iflib_num_tx_descs(ctx);
4785 	num_rxd = iflib_num_rx_descs(ctx);
4786 
4787 	/* XXX change for per-queue sizes */
4788 	device_printf(dev, "Using %d TX descriptors and %d RX descriptors\n",
4789 	    num_txd, num_rxd);
4790 
4791 	if (scctx->isc_tx_nsegments > num_txd / MAX_SINGLE_PACKET_FRACTION)
4792 		scctx->isc_tx_nsegments = max(1, num_txd /
4793 		    MAX_SINGLE_PACKET_FRACTION);
4794 	if (scctx->isc_tx_tso_segments_max > num_txd /
4795 	    MAX_SINGLE_PACKET_FRACTION)
4796 		scctx->isc_tx_tso_segments_max = max(1,
4797 		    num_txd / MAX_SINGLE_PACKET_FRACTION);
4798 
4799 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4800 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
4801 		/*
4802 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
4803 		 * but some MACs do.
4804 		 */
4805 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
4806 		    IP_MAXPACKET));
4807 		/*
4808 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
4809 		 * into account.  In the worst case, each of these calls will
4810 		 * add another mbuf and, thus, the requirement for another DMA
4811 		 * segment.  So for best performance, it doesn't make sense to
4812 		 * advertize a maximum of TSO segments that typically will
4813 		 * require defragmentation in iflib_encap().
4814 		 */
4815 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
4816 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
4817 	}
4818 	if (scctx->isc_rss_table_size == 0)
4819 		scctx->isc_rss_table_size = 64;
4820 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4821 
4822 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4823 	/* XXX format name */
4824 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx,
4825 	    NULL, NULL, "admin");
4826 
4827 	/* Set up cpu set.  If it fails, use the set of all CPUs. */
4828 	if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) != 0) {
4829 		device_printf(dev, "Unable to fetch CPU list\n");
4830 		CPU_COPY(&all_cpus, &ctx->ifc_cpus);
4831 	}
4832 	MPASS(CPU_COUNT(&ctx->ifc_cpus) > 0);
4833 
4834 	/*
4835 	** Now set up MSI or MSI-X, should return us the number of supported
4836 	** vectors (will be 1 for a legacy interrupt and MSI).
4837 	*/
4838 	if (sctx->isc_flags & IFLIB_SKIP_MSIX) {
4839 		msix = scctx->isc_vectors;
4840 	} else if (scctx->isc_msix_bar != 0)
4841 	       /*
4842 		* The simple fact that isc_msix_bar is not 0 does not mean we
4843 		* we have a good value there that is known to work.
4844 		*/
4845 		msix = iflib_msix_init(ctx);
4846 	else {
4847 		scctx->isc_vectors = 1;
4848 		scctx->isc_ntxqsets = 1;
4849 		scctx->isc_nrxqsets = 1;
4850 		scctx->isc_intr = IFLIB_INTR_LEGACY;
4851 		msix = 0;
4852 	}
4853 	/* Get memory for the station queues */
4854 	if ((err = iflib_queues_alloc(ctx))) {
4855 		device_printf(dev, "Unable to allocate queue memory\n");
4856 		goto fail_intr_free;
4857 	}
4858 
4859 	if ((err = iflib_qset_structures_setup(ctx)))
4860 		goto fail_queues;
4861 
4862 	/*
4863 	 * Now that we know how many queues there are, get the core offset.
4864 	 */
4865 	ctx->ifc_sysctl_core_offset = get_ctx_core_offset(ctx);
4866 
4867 	if (msix > 1) {
4868 		/*
4869 		 * When using MSI-X, ensure that ifdi_{r,t}x_queue_intr_enable
4870 		 * aren't the default NULL implementation.
4871 		 */
4872 		kobj_desc = &ifdi_rx_queue_intr_enable_desc;
4873 		kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL,
4874 		    kobj_desc);
4875 		if (kobj_method == &kobj_desc->deflt) {
4876 			device_printf(dev,
4877 			    "MSI-X requires ifdi_rx_queue_intr_enable method");
4878 			err = EOPNOTSUPP;
4879 			goto fail_queues;
4880 		}
4881 		kobj_desc = &ifdi_tx_queue_intr_enable_desc;
4882 		kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL,
4883 		    kobj_desc);
4884 		if (kobj_method == &kobj_desc->deflt) {
4885 			device_printf(dev,
4886 			    "MSI-X requires ifdi_tx_queue_intr_enable method");
4887 			err = EOPNOTSUPP;
4888 			goto fail_queues;
4889 		}
4890 
4891 		/*
4892 		 * Assign the MSI-X vectors.
4893 		 * Note that the default NULL ifdi_msix_intr_assign method will
4894 		 * fail here, too.
4895 		 */
4896 		err = IFDI_MSIX_INTR_ASSIGN(ctx, msix);
4897 		if (err != 0) {
4898 			device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n",
4899 			    err);
4900 			goto fail_queues;
4901 		}
4902 	} else if (scctx->isc_intr != IFLIB_INTR_MSIX) {
4903 		rid = 0;
4904 		if (scctx->isc_intr == IFLIB_INTR_MSI) {
4905 			MPASS(msix == 1);
4906 			rid = 1;
4907 		}
4908 		if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) {
4909 			device_printf(dev, "iflib_legacy_setup failed %d\n", err);
4910 			goto fail_queues;
4911 		}
4912 	} else {
4913 		device_printf(dev,
4914 		    "Cannot use iflib with only 1 MSI-X interrupt!\n");
4915 		err = ENODEV;
4916 		goto fail_queues;
4917 	}
4918 
4919 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
4920 
4921 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4922 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4923 		goto fail_detach;
4924 	}
4925 
4926 	/*
4927 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4928 	 * This must appear after the call to ether_ifattach() because
4929 	 * ether_ifattach() sets if_hdrlen to the default value.
4930 	 */
4931 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4932 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
4933 
4934 	if ((err = iflib_netmap_attach(ctx))) {
4935 		device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err);
4936 		goto fail_detach;
4937 	}
4938 	*ctxp = ctx;
4939 
4940 	DEBUGNET_SET(ctx->ifc_ifp, iflib);
4941 
4942 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4943 	iflib_add_device_sysctl_post(ctx);
4944 	iflib_add_pfil(ctx);
4945 	ctx->ifc_flags |= IFC_INIT_DONE;
4946 	CTX_UNLOCK(ctx);
4947 
4948 	return (0);
4949 
4950 fail_detach:
4951 	ether_ifdetach(ctx->ifc_ifp);
4952 fail_queues:
4953 	iflib_tqg_detach(ctx);
4954 	iflib_tx_structures_free(ctx);
4955 	iflib_rx_structures_free(ctx);
4956 	IFDI_DETACH(ctx);
4957 	IFDI_QUEUES_FREE(ctx);
4958 fail_intr_free:
4959 	iflib_free_intr_mem(ctx);
4960 fail_unlock:
4961 	CTX_UNLOCK(ctx);
4962 	iflib_deregister(ctx);
4963 fail_ctx_free:
4964 	device_set_softc(ctx->ifc_dev, NULL);
4965         if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4966                 free(ctx->ifc_softc, M_IFLIB);
4967         free(ctx, M_IFLIB);
4968 	return (err);
4969 }
4970 
4971 int
4972 iflib_pseudo_register(device_t dev, if_shared_ctx_t sctx, if_ctx_t *ctxp,
4973 					  struct iflib_cloneattach_ctx *clctx)
4974 {
4975 	int num_txd, num_rxd;
4976 	int err;
4977 	if_ctx_t ctx;
4978 	if_t ifp;
4979 	if_softc_ctx_t scctx;
4980 	int i;
4981 	void *sc;
4982 
4983 	ctx = malloc(sizeof(*ctx), M_IFLIB, M_WAITOK|M_ZERO);
4984 	sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4985 	ctx->ifc_flags |= IFC_SC_ALLOCATED;
4986 	if (sctx->isc_flags & (IFLIB_PSEUDO|IFLIB_VIRTUAL))
4987 		ctx->ifc_flags |= IFC_PSEUDO;
4988 
4989 	ctx->ifc_sctx = sctx;
4990 	ctx->ifc_softc = sc;
4991 	ctx->ifc_dev = dev;
4992 
4993 	if ((err = iflib_register(ctx)) != 0) {
4994 		device_printf(dev, "%s: iflib_register failed %d\n", __func__, err);
4995 		goto fail_ctx_free;
4996 	}
4997 	iflib_add_device_sysctl_pre(ctx);
4998 
4999 	scctx = &ctx->ifc_softc_ctx;
5000 	ifp = ctx->ifc_ifp;
5001 
5002 	iflib_reset_qvalues(ctx);
5003 	CTX_LOCK(ctx);
5004 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
5005 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
5006 		goto fail_unlock;
5007 	}
5008 	if (sctx->isc_flags & IFLIB_GEN_MAC)
5009 		ether_gen_addr(ifp, &ctx->ifc_mac);
5010 	if ((err = IFDI_CLONEATTACH(ctx, clctx->cc_ifc, clctx->cc_name,
5011 								clctx->cc_params)) != 0) {
5012 		device_printf(dev, "IFDI_CLONEATTACH failed %d\n", err);
5013 		goto fail_unlock;
5014 	}
5015 #ifdef INVARIANTS
5016 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
5017 		MPASS(scctx->isc_tx_csum_flags);
5018 #endif
5019 
5020 	if_setcapabilities(ifp, scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_LINKSTATE);
5021 	if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_LINKSTATE);
5022 
5023 	ifp->if_flags |= IFF_NOGROUP;
5024 	if (sctx->isc_flags & IFLIB_PSEUDO) {
5025 		ifmedia_add(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO, 0, NULL);
5026 		ifmedia_set(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO);
5027 		if (sctx->isc_flags & IFLIB_PSEUDO_ETHER) {
5028 			ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
5029 		} else {
5030 			if_attach(ctx->ifc_ifp);
5031 			bpfattach(ctx->ifc_ifp, DLT_NULL, sizeof(u_int32_t));
5032 		}
5033 
5034 		if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
5035 			device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
5036 			goto fail_detach;
5037 		}
5038 		*ctxp = ctx;
5039 
5040 		/*
5041 		 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
5042 		 * This must appear after the call to ether_ifattach() because
5043 		 * ether_ifattach() sets if_hdrlen to the default value.
5044 		 */
5045 		if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
5046 			if_setifheaderlen(ifp,
5047 			    sizeof(struct ether_vlan_header));
5048 
5049 		if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
5050 		iflib_add_device_sysctl_post(ctx);
5051 		ctx->ifc_flags |= IFC_INIT_DONE;
5052 		CTX_UNLOCK(ctx);
5053 		return (0);
5054 	}
5055 	ifmedia_add(ctx->ifc_mediap, IFM_ETHER | IFM_1000_T | IFM_FDX, 0, NULL);
5056 	ifmedia_add(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO, 0, NULL);
5057 	ifmedia_set(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO);
5058 
5059 	_iflib_pre_assert(scctx);
5060 	ctx->ifc_txrx = *scctx->isc_txrx;
5061 
5062 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
5063 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
5064 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
5065 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
5066 
5067 	num_txd = iflib_num_tx_descs(ctx);
5068 	num_rxd = iflib_num_rx_descs(ctx);
5069 
5070 	/* XXX change for per-queue sizes */
5071 	device_printf(dev, "Using %d TX descriptors and %d RX descriptors\n",
5072 	    num_txd, num_rxd);
5073 
5074 	if (scctx->isc_tx_nsegments > num_txd / MAX_SINGLE_PACKET_FRACTION)
5075 		scctx->isc_tx_nsegments = max(1, num_txd /
5076 		    MAX_SINGLE_PACKET_FRACTION);
5077 	if (scctx->isc_tx_tso_segments_max > num_txd /
5078 	    MAX_SINGLE_PACKET_FRACTION)
5079 		scctx->isc_tx_tso_segments_max = max(1,
5080 		    num_txd / MAX_SINGLE_PACKET_FRACTION);
5081 
5082 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
5083 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
5084 		/*
5085 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
5086 		 * but some MACs do.
5087 		 */
5088 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
5089 		    IP_MAXPACKET));
5090 		/*
5091 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
5092 		 * into account.  In the worst case, each of these calls will
5093 		 * add another mbuf and, thus, the requirement for another DMA
5094 		 * segment.  So for best performance, it doesn't make sense to
5095 		 * advertize a maximum of TSO segments that typically will
5096 		 * require defragmentation in iflib_encap().
5097 		 */
5098 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
5099 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
5100 	}
5101 	if (scctx->isc_rss_table_size == 0)
5102 		scctx->isc_rss_table_size = 64;
5103 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
5104 
5105 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
5106 	/* XXX format name */
5107 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx,
5108 	    NULL, NULL, "admin");
5109 
5110 	/* XXX --- can support > 1 -- but keep it simple for now */
5111 	scctx->isc_intr = IFLIB_INTR_LEGACY;
5112 
5113 	/* Get memory for the station queues */
5114 	if ((err = iflib_queues_alloc(ctx))) {
5115 		device_printf(dev, "Unable to allocate queue memory\n");
5116 		goto fail_iflib_detach;
5117 	}
5118 
5119 	if ((err = iflib_qset_structures_setup(ctx))) {
5120 		device_printf(dev, "qset structure setup failed %d\n", err);
5121 		goto fail_queues;
5122 	}
5123 
5124 	/*
5125 	 * XXX What if anything do we want to do about interrupts?
5126 	 */
5127 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
5128 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
5129 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
5130 		goto fail_detach;
5131 	}
5132 
5133 	/*
5134 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
5135 	 * This must appear after the call to ether_ifattach() because
5136 	 * ether_ifattach() sets if_hdrlen to the default value.
5137 	 */
5138 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
5139 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
5140 
5141 	/* XXX handle more than one queue */
5142 	for (i = 0; i < scctx->isc_nrxqsets; i++)
5143 		IFDI_RX_CLSET(ctx, 0, i, ctx->ifc_rxqs[i].ifr_fl[0].ifl_sds.ifsd_cl);
5144 
5145 	*ctxp = ctx;
5146 
5147 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
5148 	iflib_add_device_sysctl_post(ctx);
5149 	ctx->ifc_flags |= IFC_INIT_DONE;
5150 	CTX_UNLOCK(ctx);
5151 
5152 	return (0);
5153 fail_detach:
5154 	ether_ifdetach(ctx->ifc_ifp);
5155 fail_queues:
5156 	iflib_tqg_detach(ctx);
5157 	iflib_tx_structures_free(ctx);
5158 	iflib_rx_structures_free(ctx);
5159 fail_iflib_detach:
5160 	IFDI_DETACH(ctx);
5161 	IFDI_QUEUES_FREE(ctx);
5162 fail_unlock:
5163 	CTX_UNLOCK(ctx);
5164 	iflib_deregister(ctx);
5165 fail_ctx_free:
5166 	free(ctx->ifc_softc, M_IFLIB);
5167 	free(ctx, M_IFLIB);
5168 	return (err);
5169 }
5170 
5171 int
5172 iflib_pseudo_deregister(if_ctx_t ctx)
5173 {
5174 	if_t ifp = ctx->ifc_ifp;
5175 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5176 
5177 	/* Unregister VLAN event handlers early */
5178 	iflib_unregister_vlan_handlers(ctx);
5179 
5180 	if ((sctx->isc_flags & IFLIB_PSEUDO)  &&
5181 		(sctx->isc_flags & IFLIB_PSEUDO_ETHER) == 0) {
5182 		bpfdetach(ifp);
5183 		if_detach(ifp);
5184 	} else {
5185 		ether_ifdetach(ifp);
5186 	}
5187 
5188 	iflib_tqg_detach(ctx);
5189 	iflib_tx_structures_free(ctx);
5190 	iflib_rx_structures_free(ctx);
5191 	IFDI_DETACH(ctx);
5192 	IFDI_QUEUES_FREE(ctx);
5193 
5194 	iflib_deregister(ctx);
5195 
5196 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5197 		free(ctx->ifc_softc, M_IFLIB);
5198 	free(ctx, M_IFLIB);
5199 	return (0);
5200 }
5201 
5202 int
5203 iflib_device_attach(device_t dev)
5204 {
5205 	if_ctx_t ctx;
5206 	if_shared_ctx_t sctx;
5207 
5208 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
5209 		return (ENOTSUP);
5210 
5211 	pci_enable_busmaster(dev);
5212 
5213 	return (iflib_device_register(dev, NULL, sctx, &ctx));
5214 }
5215 
5216 int
5217 iflib_device_deregister(if_ctx_t ctx)
5218 {
5219 	if_t ifp = ctx->ifc_ifp;
5220 	device_t dev = ctx->ifc_dev;
5221 
5222 	/* Make sure VLANS are not using driver */
5223 	if (if_vlantrunkinuse(ifp)) {
5224 		device_printf(dev, "Vlan in use, detach first\n");
5225 		return (EBUSY);
5226 	}
5227 #ifdef PCI_IOV
5228 	if (!CTX_IS_VF(ctx) && pci_iov_detach(dev) != 0) {
5229 		device_printf(dev, "SR-IOV in use; detach first.\n");
5230 		return (EBUSY);
5231 	}
5232 #endif
5233 
5234 	STATE_LOCK(ctx);
5235 	ctx->ifc_flags |= IFC_IN_DETACH;
5236 	STATE_UNLOCK(ctx);
5237 
5238 	/* Unregister VLAN handlers before calling iflib_stop() */
5239 	iflib_unregister_vlan_handlers(ctx);
5240 
5241 	iflib_netmap_detach(ifp);
5242 	ether_ifdetach(ifp);
5243 
5244 	CTX_LOCK(ctx);
5245 	iflib_stop(ctx);
5246 	CTX_UNLOCK(ctx);
5247 
5248 	iflib_rem_pfil(ctx);
5249 	if (ctx->ifc_led_dev != NULL)
5250 		led_destroy(ctx->ifc_led_dev);
5251 
5252 	iflib_tqg_detach(ctx);
5253 	iflib_tx_structures_free(ctx);
5254 	iflib_rx_structures_free(ctx);
5255 
5256 	CTX_LOCK(ctx);
5257 	IFDI_DETACH(ctx);
5258 	IFDI_QUEUES_FREE(ctx);
5259 	CTX_UNLOCK(ctx);
5260 
5261 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
5262 	iflib_free_intr_mem(ctx);
5263 
5264 	bus_generic_detach(dev);
5265 
5266 	iflib_deregister(ctx);
5267 
5268 	device_set_softc(ctx->ifc_dev, NULL);
5269 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5270 		free(ctx->ifc_softc, M_IFLIB);
5271 	unref_ctx_core_offset(ctx);
5272 	free(ctx, M_IFLIB);
5273 	return (0);
5274 }
5275 
5276 static void
5277 iflib_tqg_detach(if_ctx_t ctx)
5278 {
5279 	iflib_txq_t txq;
5280 	iflib_rxq_t rxq;
5281 	int i;
5282 	struct taskqgroup *tqg;
5283 
5284 	/* XXX drain any dependent tasks */
5285 	tqg = qgroup_if_io_tqg;
5286 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
5287 		callout_drain(&txq->ift_timer);
5288 #ifdef DEV_NETMAP
5289 		callout_drain(&txq->ift_netmap_timer);
5290 #endif /* DEV_NETMAP */
5291 		if (txq->ift_task.gt_uniq != NULL)
5292 			taskqgroup_detach(tqg, &txq->ift_task);
5293 	}
5294 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
5295 		if (rxq->ifr_task.gt_uniq != NULL)
5296 			taskqgroup_detach(tqg, &rxq->ifr_task);
5297 	}
5298 	tqg = qgroup_if_config_tqg;
5299 	if (ctx->ifc_admin_task.gt_uniq != NULL)
5300 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
5301 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
5302 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
5303 }
5304 
5305 static void
5306 iflib_free_intr_mem(if_ctx_t ctx)
5307 {
5308 
5309 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) {
5310 		iflib_irq_free(ctx, &ctx->ifc_legacy_irq);
5311 	}
5312 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) {
5313 		pci_release_msi(ctx->ifc_dev);
5314 	}
5315 	if (ctx->ifc_msix_mem != NULL) {
5316 		bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY,
5317 		    rman_get_rid(ctx->ifc_msix_mem), ctx->ifc_msix_mem);
5318 		ctx->ifc_msix_mem = NULL;
5319 	}
5320 }
5321 
5322 int
5323 iflib_device_detach(device_t dev)
5324 {
5325 	if_ctx_t ctx = device_get_softc(dev);
5326 
5327 	return (iflib_device_deregister(ctx));
5328 }
5329 
5330 int
5331 iflib_device_suspend(device_t dev)
5332 {
5333 	if_ctx_t ctx = device_get_softc(dev);
5334 
5335 	CTX_LOCK(ctx);
5336 	IFDI_SUSPEND(ctx);
5337 	CTX_UNLOCK(ctx);
5338 
5339 	return bus_generic_suspend(dev);
5340 }
5341 int
5342 iflib_device_shutdown(device_t dev)
5343 {
5344 	if_ctx_t ctx = device_get_softc(dev);
5345 
5346 	CTX_LOCK(ctx);
5347 	IFDI_SHUTDOWN(ctx);
5348 	CTX_UNLOCK(ctx);
5349 
5350 	return bus_generic_suspend(dev);
5351 }
5352 
5353 int
5354 iflib_device_resume(device_t dev)
5355 {
5356 	if_ctx_t ctx = device_get_softc(dev);
5357 	iflib_txq_t txq = ctx->ifc_txqs;
5358 
5359 	CTX_LOCK(ctx);
5360 	IFDI_RESUME(ctx);
5361 	iflib_if_init_locked(ctx);
5362 	CTX_UNLOCK(ctx);
5363 	for (int i = 0; i < NTXQSETS(ctx); i++, txq++)
5364 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
5365 
5366 	return (bus_generic_resume(dev));
5367 }
5368 
5369 int
5370 iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params)
5371 {
5372 	int error;
5373 	if_ctx_t ctx = device_get_softc(dev);
5374 
5375 	CTX_LOCK(ctx);
5376 	error = IFDI_IOV_INIT(ctx, num_vfs, params);
5377 	CTX_UNLOCK(ctx);
5378 
5379 	return (error);
5380 }
5381 
5382 void
5383 iflib_device_iov_uninit(device_t dev)
5384 {
5385 	if_ctx_t ctx = device_get_softc(dev);
5386 
5387 	CTX_LOCK(ctx);
5388 	IFDI_IOV_UNINIT(ctx);
5389 	CTX_UNLOCK(ctx);
5390 }
5391 
5392 int
5393 iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params)
5394 {
5395 	int error;
5396 	if_ctx_t ctx = device_get_softc(dev);
5397 
5398 	CTX_LOCK(ctx);
5399 	error = IFDI_IOV_VF_ADD(ctx, vfnum, params);
5400 	CTX_UNLOCK(ctx);
5401 
5402 	return (error);
5403 }
5404 
5405 /*********************************************************************
5406  *
5407  *  MODULE FUNCTION DEFINITIONS
5408  *
5409  **********************************************************************/
5410 
5411 /*
5412  * - Start a fast taskqueue thread for each core
5413  * - Start a taskqueue for control operations
5414  */
5415 static int
5416 iflib_module_init(void)
5417 {
5418 	iflib_timer_default = hz / 2;
5419 	return (0);
5420 }
5421 
5422 static int
5423 iflib_module_event_handler(module_t mod, int what, void *arg)
5424 {
5425 	int err;
5426 
5427 	switch (what) {
5428 	case MOD_LOAD:
5429 		if ((err = iflib_module_init()) != 0)
5430 			return (err);
5431 		break;
5432 	case MOD_UNLOAD:
5433 		return (EBUSY);
5434 	default:
5435 		return (EOPNOTSUPP);
5436 	}
5437 
5438 	return (0);
5439 }
5440 
5441 /*********************************************************************
5442  *
5443  *  PUBLIC FUNCTION DEFINITIONS
5444  *     ordered as in iflib.h
5445  *
5446  **********************************************************************/
5447 
5448 static void
5449 _iflib_assert(if_shared_ctx_t sctx)
5450 {
5451 	int i;
5452 
5453 	MPASS(sctx->isc_tx_maxsize);
5454 	MPASS(sctx->isc_tx_maxsegsize);
5455 
5456 	MPASS(sctx->isc_rx_maxsize);
5457 	MPASS(sctx->isc_rx_nsegments);
5458 	MPASS(sctx->isc_rx_maxsegsize);
5459 
5460 	MPASS(sctx->isc_nrxqs >= 1 && sctx->isc_nrxqs <= 8);
5461 	for (i = 0; i < sctx->isc_nrxqs; i++) {
5462 		MPASS(sctx->isc_nrxd_min[i]);
5463 		MPASS(powerof2(sctx->isc_nrxd_min[i]));
5464 		MPASS(sctx->isc_nrxd_max[i]);
5465 		MPASS(powerof2(sctx->isc_nrxd_max[i]));
5466 		MPASS(sctx->isc_nrxd_default[i]);
5467 		MPASS(powerof2(sctx->isc_nrxd_default[i]));
5468 	}
5469 
5470 	MPASS(sctx->isc_ntxqs >= 1 && sctx->isc_ntxqs <= 8);
5471 	for (i = 0; i < sctx->isc_ntxqs; i++) {
5472 		MPASS(sctx->isc_ntxd_min[i]);
5473 		MPASS(powerof2(sctx->isc_ntxd_min[i]));
5474 		MPASS(sctx->isc_ntxd_max[i]);
5475 		MPASS(powerof2(sctx->isc_ntxd_max[i]));
5476 		MPASS(sctx->isc_ntxd_default[i]);
5477 		MPASS(powerof2(sctx->isc_ntxd_default[i]));
5478 	}
5479 }
5480 
5481 static void
5482 _iflib_pre_assert(if_softc_ctx_t scctx)
5483 {
5484 
5485 	MPASS(scctx->isc_txrx->ift_txd_encap);
5486 	MPASS(scctx->isc_txrx->ift_txd_flush);
5487 	MPASS(scctx->isc_txrx->ift_txd_credits_update);
5488 	MPASS(scctx->isc_txrx->ift_rxd_available);
5489 	MPASS(scctx->isc_txrx->ift_rxd_pkt_get);
5490 	MPASS(scctx->isc_txrx->ift_rxd_refill);
5491 	MPASS(scctx->isc_txrx->ift_rxd_flush);
5492 }
5493 
5494 static int
5495 iflib_register(if_ctx_t ctx)
5496 {
5497 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5498 	driver_t *driver = sctx->isc_driver;
5499 	device_t dev = ctx->ifc_dev;
5500 	if_t ifp;
5501 	u_char type;
5502 	int iflags;
5503 
5504 	if ((sctx->isc_flags & IFLIB_PSEUDO) == 0)
5505 		_iflib_assert(sctx);
5506 
5507 	CTX_LOCK_INIT(ctx);
5508 	STATE_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev));
5509 	if (sctx->isc_flags & IFLIB_PSEUDO) {
5510 		if (sctx->isc_flags & IFLIB_PSEUDO_ETHER)
5511 			type = IFT_ETHER;
5512 		else
5513 			type = IFT_PPP;
5514 	} else
5515 		type = IFT_ETHER;
5516 	ifp = ctx->ifc_ifp = if_alloc(type);
5517 	if (ifp == NULL) {
5518 		device_printf(dev, "can not allocate ifnet structure\n");
5519 		return (ENOMEM);
5520 	}
5521 
5522 	/*
5523 	 * Initialize our context's device specific methods
5524 	 */
5525 	kobj_init((kobj_t) ctx, (kobj_class_t) driver);
5526 	kobj_class_compile((kobj_class_t) driver);
5527 
5528 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
5529 	if_setsoftc(ifp, ctx);
5530 	if_setdev(ifp, dev);
5531 	if_setinitfn(ifp, iflib_if_init);
5532 	if_setioctlfn(ifp, iflib_if_ioctl);
5533 #ifdef ALTQ
5534 	if_setstartfn(ifp, iflib_altq_if_start);
5535 	if_settransmitfn(ifp, iflib_altq_if_transmit);
5536 	if_setsendqready(ifp);
5537 #else
5538 	if_settransmitfn(ifp, iflib_if_transmit);
5539 #endif
5540 	if_setqflushfn(ifp, iflib_if_qflush);
5541 	iflags = IFF_MULTICAST | IFF_KNOWSEPOCH;
5542 
5543 	if ((sctx->isc_flags & IFLIB_PSEUDO) &&
5544 		(sctx->isc_flags & IFLIB_PSEUDO_ETHER) == 0)
5545 		iflags |= IFF_POINTOPOINT;
5546 	else
5547 		iflags |= IFF_BROADCAST | IFF_SIMPLEX;
5548 	if_setflags(ifp, iflags);
5549 	ctx->ifc_vlan_attach_event =
5550 		EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx,
5551 							  EVENTHANDLER_PRI_FIRST);
5552 	ctx->ifc_vlan_detach_event =
5553 		EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx,
5554 							  EVENTHANDLER_PRI_FIRST);
5555 
5556 	if ((sctx->isc_flags & IFLIB_DRIVER_MEDIA) == 0) {
5557 		ctx->ifc_mediap = &ctx->ifc_media;
5558 		ifmedia_init(ctx->ifc_mediap, IFM_IMASK,
5559 		    iflib_media_change, iflib_media_status);
5560 	}
5561 	return (0);
5562 }
5563 
5564 static void
5565 iflib_unregister_vlan_handlers(if_ctx_t ctx)
5566 {
5567 	/* Unregister VLAN events */
5568 	if (ctx->ifc_vlan_attach_event != NULL) {
5569 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
5570 		ctx->ifc_vlan_attach_event = NULL;
5571 	}
5572 	if (ctx->ifc_vlan_detach_event != NULL) {
5573 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
5574 		ctx->ifc_vlan_detach_event = NULL;
5575 	}
5576 
5577 }
5578 
5579 static void
5580 iflib_deregister(if_ctx_t ctx)
5581 {
5582 	if_t ifp = ctx->ifc_ifp;
5583 
5584 	/* Remove all media */
5585 	ifmedia_removeall(&ctx->ifc_media);
5586 
5587 	/* Ensure that VLAN event handlers are unregistered */
5588 	iflib_unregister_vlan_handlers(ctx);
5589 
5590 	/* Release kobject reference */
5591 	kobj_delete((kobj_t) ctx, NULL);
5592 
5593 	/* Free the ifnet structure */
5594 	if_free(ifp);
5595 
5596 	STATE_LOCK_DESTROY(ctx);
5597 
5598 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
5599 	CTX_LOCK_DESTROY(ctx);
5600 }
5601 
5602 static int
5603 iflib_queues_alloc(if_ctx_t ctx)
5604 {
5605 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5606 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5607 	device_t dev = ctx->ifc_dev;
5608 	int nrxqsets = scctx->isc_nrxqsets;
5609 	int ntxqsets = scctx->isc_ntxqsets;
5610 	iflib_txq_t txq;
5611 	iflib_rxq_t rxq;
5612 	iflib_fl_t fl = NULL;
5613 	int i, j, cpu, err, txconf, rxconf;
5614 	iflib_dma_info_t ifdip;
5615 	uint32_t *rxqsizes = scctx->isc_rxqsizes;
5616 	uint32_t *txqsizes = scctx->isc_txqsizes;
5617 	uint8_t nrxqs = sctx->isc_nrxqs;
5618 	uint8_t ntxqs = sctx->isc_ntxqs;
5619 	int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1;
5620 	int fl_offset = (sctx->isc_flags & IFLIB_HAS_RXCQ ? 1 : 0);
5621 	caddr_t *vaddrs;
5622 	uint64_t *paddrs;
5623 
5624 	KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1"));
5625 	KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1"));
5626 	KASSERT(nrxqs >= fl_offset + nfree_lists,
5627            ("there must be at least a rxq for each free list"));
5628 
5629 	/* Allocate the TX ring struct memory */
5630 	if (!(ctx->ifc_txqs =
5631 	    (iflib_txq_t) malloc(sizeof(struct iflib_txq) *
5632 	    ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5633 		device_printf(dev, "Unable to allocate TX ring memory\n");
5634 		err = ENOMEM;
5635 		goto fail;
5636 	}
5637 
5638 	/* Now allocate the RX */
5639 	if (!(ctx->ifc_rxqs =
5640 	    (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) *
5641 	    nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5642 		device_printf(dev, "Unable to allocate RX ring memory\n");
5643 		err = ENOMEM;
5644 		goto rx_fail;
5645 	}
5646 
5647 	txq = ctx->ifc_txqs;
5648 	rxq = ctx->ifc_rxqs;
5649 
5650 	/*
5651 	 * XXX handle allocation failure
5652 	 */
5653 	for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) {
5654 		/* Set up some basics */
5655 
5656 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs,
5657 		    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
5658 			device_printf(dev,
5659 			    "Unable to allocate TX DMA info memory\n");
5660 			err = ENOMEM;
5661 			goto err_tx_desc;
5662 		}
5663 		txq->ift_ifdi = ifdip;
5664 		for (j = 0; j < ntxqs; j++, ifdip++) {
5665 			if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, 0)) {
5666 				device_printf(dev,
5667 				    "Unable to allocate TX descriptors\n");
5668 				err = ENOMEM;
5669 				goto err_tx_desc;
5670 			}
5671 			txq->ift_txd_size[j] = scctx->isc_txd_size[j];
5672 			bzero((void *)ifdip->idi_vaddr, txqsizes[j]);
5673 		}
5674 		txq->ift_ctx = ctx;
5675 		txq->ift_id = i;
5676 		if (sctx->isc_flags & IFLIB_HAS_TXCQ) {
5677 			txq->ift_br_offset = 1;
5678 		} else {
5679 			txq->ift_br_offset = 0;
5680 		}
5681 
5682 		if (iflib_txsd_alloc(txq)) {
5683 			device_printf(dev, "Critical Failure setting up TX buffers\n");
5684 			err = ENOMEM;
5685 			goto err_tx_desc;
5686 		}
5687 
5688 		/* Initialize the TX lock */
5689 		snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:TX(%d):callout",
5690 		    device_get_nameunit(dev), txq->ift_id);
5691 		mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF);
5692 		callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0);
5693 		txq->ift_timer.c_cpu = cpu;
5694 #ifdef DEV_NETMAP
5695 		callout_init_mtx(&txq->ift_netmap_timer, &txq->ift_mtx, 0);
5696 		txq->ift_netmap_timer.c_cpu = cpu;
5697 #endif /* DEV_NETMAP */
5698 
5699 		err = ifmp_ring_alloc(&txq->ift_br, 2048, txq, iflib_txq_drain,
5700 				      iflib_txq_can_drain, M_IFLIB, M_WAITOK);
5701 		if (err) {
5702 			/* XXX free any allocated rings */
5703 			device_printf(dev, "Unable to allocate buf_ring\n");
5704 			goto err_tx_desc;
5705 		}
5706 	}
5707 
5708 	for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) {
5709 		/* Set up some basics */
5710 		callout_init(&rxq->ifr_watchdog, 1);
5711 
5712 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs,
5713 		   M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
5714 			device_printf(dev,
5715 			    "Unable to allocate RX DMA info memory\n");
5716 			err = ENOMEM;
5717 			goto err_tx_desc;
5718 		}
5719 
5720 		rxq->ifr_ifdi = ifdip;
5721 		/* XXX this needs to be changed if #rx queues != #tx queues */
5722 		rxq->ifr_ntxqirq = 1;
5723 		rxq->ifr_txqid[0] = i;
5724 		for (j = 0; j < nrxqs; j++, ifdip++) {
5725 			if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, 0)) {
5726 				device_printf(dev,
5727 				    "Unable to allocate RX descriptors\n");
5728 				err = ENOMEM;
5729 				goto err_tx_desc;
5730 			}
5731 			bzero((void *)ifdip->idi_vaddr, rxqsizes[j]);
5732 		}
5733 		rxq->ifr_ctx = ctx;
5734 		rxq->ifr_id = i;
5735 		rxq->ifr_fl_offset = fl_offset;
5736 		rxq->ifr_nfl = nfree_lists;
5737 		if (!(fl =
5738 			  (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) {
5739 			device_printf(dev, "Unable to allocate free list memory\n");
5740 			err = ENOMEM;
5741 			goto err_tx_desc;
5742 		}
5743 		rxq->ifr_fl = fl;
5744 		for (j = 0; j < nfree_lists; j++) {
5745 			fl[j].ifl_rxq = rxq;
5746 			fl[j].ifl_id = j;
5747 			fl[j].ifl_ifdi = &rxq->ifr_ifdi[j + rxq->ifr_fl_offset];
5748 			fl[j].ifl_rxd_size = scctx->isc_rxd_size[j];
5749 		}
5750 		/* Allocate receive buffers for the ring */
5751 		if (iflib_rxsd_alloc(rxq)) {
5752 			device_printf(dev,
5753 			    "Critical Failure setting up receive buffers\n");
5754 			err = ENOMEM;
5755 			goto err_rx_desc;
5756 		}
5757 
5758 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5759 			fl->ifl_rx_bitmap = bit_alloc(fl->ifl_size, M_IFLIB,
5760 			    M_WAITOK);
5761 	}
5762 
5763 	/* TXQs */
5764 	vaddrs = malloc(sizeof(caddr_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5765 	paddrs = malloc(sizeof(uint64_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5766 	for (i = 0; i < ntxqsets; i++) {
5767 		iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi;
5768 
5769 		for (j = 0; j < ntxqs; j++, di++) {
5770 			vaddrs[i*ntxqs + j] = di->idi_vaddr;
5771 			paddrs[i*ntxqs + j] = di->idi_paddr;
5772 		}
5773 	}
5774 	if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) {
5775 		device_printf(ctx->ifc_dev,
5776 		    "Unable to allocate device TX queue\n");
5777 		iflib_tx_structures_free(ctx);
5778 		free(vaddrs, M_IFLIB);
5779 		free(paddrs, M_IFLIB);
5780 		goto err_rx_desc;
5781 	}
5782 	free(vaddrs, M_IFLIB);
5783 	free(paddrs, M_IFLIB);
5784 
5785 	/* RXQs */
5786 	vaddrs = malloc(sizeof(caddr_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5787 	paddrs = malloc(sizeof(uint64_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5788 	for (i = 0; i < nrxqsets; i++) {
5789 		iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi;
5790 
5791 		for (j = 0; j < nrxqs; j++, di++) {
5792 			vaddrs[i*nrxqs + j] = di->idi_vaddr;
5793 			paddrs[i*nrxqs + j] = di->idi_paddr;
5794 		}
5795 	}
5796 	if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) {
5797 		device_printf(ctx->ifc_dev,
5798 		    "Unable to allocate device RX queue\n");
5799 		iflib_tx_structures_free(ctx);
5800 		free(vaddrs, M_IFLIB);
5801 		free(paddrs, M_IFLIB);
5802 		goto err_rx_desc;
5803 	}
5804 	free(vaddrs, M_IFLIB);
5805 	free(paddrs, M_IFLIB);
5806 
5807 	return (0);
5808 
5809 /* XXX handle allocation failure changes */
5810 err_rx_desc:
5811 err_tx_desc:
5812 rx_fail:
5813 	if (ctx->ifc_rxqs != NULL)
5814 		free(ctx->ifc_rxqs, M_IFLIB);
5815 	ctx->ifc_rxqs = NULL;
5816 	if (ctx->ifc_txqs != NULL)
5817 		free(ctx->ifc_txqs, M_IFLIB);
5818 	ctx->ifc_txqs = NULL;
5819 fail:
5820 	return (err);
5821 }
5822 
5823 static int
5824 iflib_tx_structures_setup(if_ctx_t ctx)
5825 {
5826 	iflib_txq_t txq = ctx->ifc_txqs;
5827 	int i;
5828 
5829 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
5830 		iflib_txq_setup(txq);
5831 
5832 	return (0);
5833 }
5834 
5835 static void
5836 iflib_tx_structures_free(if_ctx_t ctx)
5837 {
5838 	iflib_txq_t txq = ctx->ifc_txqs;
5839 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5840 	int i, j;
5841 
5842 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
5843 		for (j = 0; j < sctx->isc_ntxqs; j++)
5844 			iflib_dma_free(&txq->ift_ifdi[j]);
5845 		iflib_txq_destroy(txq);
5846 	}
5847 	free(ctx->ifc_txqs, M_IFLIB);
5848 	ctx->ifc_txqs = NULL;
5849 }
5850 
5851 /*********************************************************************
5852  *
5853  *  Initialize all receive rings.
5854  *
5855  **********************************************************************/
5856 static int
5857 iflib_rx_structures_setup(if_ctx_t ctx)
5858 {
5859 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5860 	int q;
5861 #if defined(INET6) || defined(INET)
5862 	int err, i;
5863 #endif
5864 
5865 	for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) {
5866 #if defined(INET6) || defined(INET)
5867 		if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_LRO) {
5868 			err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp,
5869 			    TCP_LRO_ENTRIES, min(1024,
5870 			    ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset]));
5871 			if (err != 0) {
5872 				device_printf(ctx->ifc_dev,
5873 				    "LRO Initialization failed!\n");
5874 				goto fail;
5875 			}
5876 		}
5877 #endif
5878 		IFDI_RXQ_SETUP(ctx, rxq->ifr_id);
5879 	}
5880 	return (0);
5881 #if defined(INET6) || defined(INET)
5882 fail:
5883 	/*
5884 	 * Free LRO resources allocated so far, we will only handle
5885 	 * the rings that completed, the failing case will have
5886 	 * cleaned up for itself.  'q' failed, so its the terminus.
5887 	 */
5888 	rxq = ctx->ifc_rxqs;
5889 	for (i = 0; i < q; ++i, rxq++) {
5890 		if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_LRO)
5891 			tcp_lro_free(&rxq->ifr_lc);
5892 	}
5893 	return (err);
5894 #endif
5895 }
5896 
5897 /*********************************************************************
5898  *
5899  *  Free all receive rings.
5900  *
5901  **********************************************************************/
5902 static void
5903 iflib_rx_structures_free(if_ctx_t ctx)
5904 {
5905 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5906 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5907 	int i, j;
5908 
5909 	for (i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) {
5910 		for (j = 0; j < sctx->isc_nrxqs; j++)
5911 			iflib_dma_free(&rxq->ifr_ifdi[j]);
5912 		iflib_rx_sds_free(rxq);
5913 #if defined(INET6) || defined(INET)
5914 		if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_LRO)
5915 			tcp_lro_free(&rxq->ifr_lc);
5916 #endif
5917 	}
5918 	free(ctx->ifc_rxqs, M_IFLIB);
5919 	ctx->ifc_rxqs = NULL;
5920 }
5921 
5922 static int
5923 iflib_qset_structures_setup(if_ctx_t ctx)
5924 {
5925 	int err;
5926 
5927 	/*
5928 	 * It is expected that the caller takes care of freeing queues if this
5929 	 * fails.
5930 	 */
5931 	if ((err = iflib_tx_structures_setup(ctx)) != 0) {
5932 		device_printf(ctx->ifc_dev, "iflib_tx_structures_setup failed: %d\n", err);
5933 		return (err);
5934 	}
5935 
5936 	if ((err = iflib_rx_structures_setup(ctx)) != 0)
5937 		device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err);
5938 
5939 	return (err);
5940 }
5941 
5942 int
5943 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
5944 		driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, const char *name)
5945 {
5946 
5947 	return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name));
5948 }
5949 
5950 #ifdef SMP
5951 static int
5952 find_nth(if_ctx_t ctx, int qid)
5953 {
5954 	cpuset_t cpus;
5955 	int i, cpuid, eqid, count;
5956 
5957 	CPU_COPY(&ctx->ifc_cpus, &cpus);
5958 	count = CPU_COUNT(&cpus);
5959 	eqid = qid % count;
5960 	/* clear up to the qid'th bit */
5961 	for (i = 0; i < eqid; i++) {
5962 		cpuid = CPU_FFS(&cpus);
5963 		MPASS(cpuid != 0);
5964 		CPU_CLR(cpuid-1, &cpus);
5965 	}
5966 	cpuid = CPU_FFS(&cpus);
5967 	MPASS(cpuid != 0);
5968 	return (cpuid-1);
5969 }
5970 
5971 #ifdef SCHED_ULE
5972 extern struct cpu_group *cpu_top;              /* CPU topology */
5973 
5974 static int
5975 find_child_with_core(int cpu, struct cpu_group *grp)
5976 {
5977 	int i;
5978 
5979 	if (grp->cg_children == 0)
5980 		return -1;
5981 
5982 	MPASS(grp->cg_child);
5983 	for (i = 0; i < grp->cg_children; i++) {
5984 		if (CPU_ISSET(cpu, &grp->cg_child[i].cg_mask))
5985 			return i;
5986 	}
5987 
5988 	return -1;
5989 }
5990 
5991 /*
5992  * Find the nth "close" core to the specified core
5993  * "close" is defined as the deepest level that shares
5994  * at least an L2 cache.  With threads, this will be
5995  * threads on the same core.  If the shared cache is L3
5996  * or higher, simply returns the same core.
5997  */
5998 static int
5999 find_close_core(int cpu, int core_offset)
6000 {
6001 	struct cpu_group *grp;
6002 	int i;
6003 	int fcpu;
6004 	cpuset_t cs;
6005 
6006 	grp = cpu_top;
6007 	if (grp == NULL)
6008 		return cpu;
6009 	i = 0;
6010 	while ((i = find_child_with_core(cpu, grp)) != -1) {
6011 		/* If the child only has one cpu, don't descend */
6012 		if (grp->cg_child[i].cg_count <= 1)
6013 			break;
6014 		grp = &grp->cg_child[i];
6015 	}
6016 
6017 	/* If they don't share at least an L2 cache, use the same CPU */
6018 	if (grp->cg_level > CG_SHARE_L2 || grp->cg_level == CG_SHARE_NONE)
6019 		return cpu;
6020 
6021 	/* Now pick one */
6022 	CPU_COPY(&grp->cg_mask, &cs);
6023 
6024 	/* Add the selected CPU offset to core offset. */
6025 	for (i = 0; (fcpu = CPU_FFS(&cs)) != 0; i++) {
6026 		if (fcpu - 1 == cpu)
6027 			break;
6028 		CPU_CLR(fcpu - 1, &cs);
6029 	}
6030 	MPASS(fcpu);
6031 
6032 	core_offset += i;
6033 
6034 	CPU_COPY(&grp->cg_mask, &cs);
6035 	for (i = core_offset % grp->cg_count; i > 0; i--) {
6036 		MPASS(CPU_FFS(&cs));
6037 		CPU_CLR(CPU_FFS(&cs) - 1, &cs);
6038 	}
6039 	MPASS(CPU_FFS(&cs));
6040 	return CPU_FFS(&cs) - 1;
6041 }
6042 #else
6043 static int
6044 find_close_core(int cpu, int core_offset __unused)
6045 {
6046 	return cpu;
6047 }
6048 #endif
6049 
6050 static int
6051 get_core_offset(if_ctx_t ctx, iflib_intr_type_t type, int qid)
6052 {
6053 	switch (type) {
6054 	case IFLIB_INTR_TX:
6055 		/* TX queues get cores which share at least an L2 cache with the corresponding RX queue */
6056 		/* XXX handle multiple RX threads per core and more than two core per L2 group */
6057 		return qid / CPU_COUNT(&ctx->ifc_cpus) + 1;
6058 	case IFLIB_INTR_RX:
6059 	case IFLIB_INTR_RXTX:
6060 		/* RX queues get the specified core */
6061 		return qid / CPU_COUNT(&ctx->ifc_cpus);
6062 	default:
6063 		return -1;
6064 	}
6065 }
6066 #else
6067 #define get_core_offset(ctx, type, qid)	CPU_FIRST()
6068 #define find_close_core(cpuid, tid)	CPU_FIRST()
6069 #define find_nth(ctx, gid)		CPU_FIRST()
6070 #endif
6071 
6072 /* Just to avoid copy/paste */
6073 static inline int
6074 iflib_irq_set_affinity(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type,
6075     int qid, struct grouptask *gtask, struct taskqgroup *tqg, void *uniq,
6076     const char *name)
6077 {
6078 	device_t dev;
6079 	int co, cpuid, err, tid;
6080 
6081 	dev = ctx->ifc_dev;
6082 	co = ctx->ifc_sysctl_core_offset;
6083 	if (ctx->ifc_sysctl_separate_txrx && type == IFLIB_INTR_TX)
6084 		co += ctx->ifc_softc_ctx.isc_nrxqsets;
6085 	cpuid = find_nth(ctx, qid + co);
6086 	tid = get_core_offset(ctx, type, qid);
6087 	if (tid < 0) {
6088 		device_printf(dev, "get_core_offset failed\n");
6089 		return (EOPNOTSUPP);
6090 	}
6091 	cpuid = find_close_core(cpuid, tid);
6092 	err = taskqgroup_attach_cpu(tqg, gtask, uniq, cpuid, dev, irq->ii_res,
6093 	    name);
6094 	if (err) {
6095 		device_printf(dev, "taskqgroup_attach_cpu failed %d\n", err);
6096 		return (err);
6097 	}
6098 #ifdef notyet
6099 	if (cpuid > ctx->ifc_cpuid_highest)
6100 		ctx->ifc_cpuid_highest = cpuid;
6101 #endif
6102 	return (0);
6103 }
6104 
6105 int
6106 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid,
6107 			iflib_intr_type_t type, driver_filter_t *filter,
6108 			void *filter_arg, int qid, const char *name)
6109 {
6110 	device_t dev;
6111 	struct grouptask *gtask;
6112 	struct taskqgroup *tqg;
6113 	iflib_filter_info_t info;
6114 	gtask_fn_t *fn;
6115 	int tqrid, err;
6116 	driver_filter_t *intr_fast;
6117 	void *q;
6118 
6119 	info = &ctx->ifc_filter_info;
6120 	tqrid = rid;
6121 
6122 	switch (type) {
6123 	/* XXX merge tx/rx for netmap? */
6124 	case IFLIB_INTR_TX:
6125 		q = &ctx->ifc_txqs[qid];
6126 		info = &ctx->ifc_txqs[qid].ift_filter_info;
6127 		gtask = &ctx->ifc_txqs[qid].ift_task;
6128 		tqg = qgroup_if_io_tqg;
6129 		fn = _task_fn_tx;
6130 		intr_fast = iflib_fast_intr;
6131 		GROUPTASK_INIT(gtask, 0, fn, q);
6132 		ctx->ifc_flags |= IFC_NETMAP_TX_IRQ;
6133 		break;
6134 	case IFLIB_INTR_RX:
6135 		q = &ctx->ifc_rxqs[qid];
6136 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
6137 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6138 		tqg = qgroup_if_io_tqg;
6139 		fn = _task_fn_rx;
6140 		intr_fast = iflib_fast_intr;
6141 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6142 		break;
6143 	case IFLIB_INTR_RXTX:
6144 		q = &ctx->ifc_rxqs[qid];
6145 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
6146 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6147 		tqg = qgroup_if_io_tqg;
6148 		fn = _task_fn_rx;
6149 		intr_fast = iflib_fast_intr_rxtx;
6150 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6151 		break;
6152 	case IFLIB_INTR_ADMIN:
6153 		q = ctx;
6154 		tqrid = -1;
6155 		info = &ctx->ifc_filter_info;
6156 		gtask = &ctx->ifc_admin_task;
6157 		tqg = qgroup_if_config_tqg;
6158 		fn = _task_fn_admin;
6159 		intr_fast = iflib_fast_intr_ctx;
6160 		break;
6161 	default:
6162 		device_printf(ctx->ifc_dev, "%s: unknown net intr type\n",
6163 		    __func__);
6164 		return (EINVAL);
6165 	}
6166 
6167 	info->ifi_filter = filter;
6168 	info->ifi_filter_arg = filter_arg;
6169 	info->ifi_task = gtask;
6170 	info->ifi_ctx = q;
6171 
6172 	dev = ctx->ifc_dev;
6173 	err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info,  name);
6174 	if (err != 0) {
6175 		device_printf(dev, "_iflib_irq_alloc failed %d\n", err);
6176 		return (err);
6177 	}
6178 	if (type == IFLIB_INTR_ADMIN)
6179 		return (0);
6180 
6181 	if (tqrid != -1) {
6182 		err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg,
6183 		    q, name);
6184 		if (err)
6185 			return (err);
6186 	} else {
6187 		taskqgroup_attach(tqg, gtask, q, dev, irq->ii_res, name);
6188 	}
6189 
6190 	return (0);
6191 }
6192 
6193 void
6194 iflib_softirq_alloc_generic(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type, void *arg, int qid, const char *name)
6195 {
6196 	struct grouptask *gtask;
6197 	struct taskqgroup *tqg;
6198 	gtask_fn_t *fn;
6199 	void *q;
6200 	int err;
6201 
6202 	switch (type) {
6203 	case IFLIB_INTR_TX:
6204 		q = &ctx->ifc_txqs[qid];
6205 		gtask = &ctx->ifc_txqs[qid].ift_task;
6206 		tqg = qgroup_if_io_tqg;
6207 		fn = _task_fn_tx;
6208 		GROUPTASK_INIT(gtask, 0, fn, q);
6209 		break;
6210 	case IFLIB_INTR_RX:
6211 		q = &ctx->ifc_rxqs[qid];
6212 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6213 		tqg = qgroup_if_io_tqg;
6214 		fn = _task_fn_rx;
6215 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6216 		break;
6217 	case IFLIB_INTR_IOV:
6218 		q = ctx;
6219 		gtask = &ctx->ifc_vflr_task;
6220 		tqg = qgroup_if_config_tqg;
6221 		fn = _task_fn_iov;
6222 		GROUPTASK_INIT(gtask, 0, fn, q);
6223 		break;
6224 	default:
6225 		panic("unknown net intr type");
6226 	}
6227 	if (irq != NULL) {
6228 		err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg,
6229 		    q, name);
6230 		if (err)
6231 			taskqgroup_attach(tqg, gtask, q, ctx->ifc_dev,
6232 			    irq->ii_res, name);
6233 	} else {
6234 		taskqgroup_attach(tqg, gtask, q, NULL, NULL, name);
6235 	}
6236 }
6237 
6238 void
6239 iflib_irq_free(if_ctx_t ctx, if_irq_t irq)
6240 {
6241 
6242 	if (irq->ii_tag)
6243 		bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag);
6244 
6245 	if (irq->ii_res)
6246 		bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ,
6247 		    rman_get_rid(irq->ii_res), irq->ii_res);
6248 }
6249 
6250 static int
6251 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, const char *name)
6252 {
6253 	iflib_txq_t txq = ctx->ifc_txqs;
6254 	iflib_rxq_t rxq = ctx->ifc_rxqs;
6255 	if_irq_t irq = &ctx->ifc_legacy_irq;
6256 	iflib_filter_info_t info;
6257 	device_t dev;
6258 	struct grouptask *gtask;
6259 	struct resource *res;
6260 	struct taskqgroup *tqg;
6261 	void *q;
6262 	int err, tqrid;
6263 	bool rx_only;
6264 
6265 	q = &ctx->ifc_rxqs[0];
6266 	info = &rxq[0].ifr_filter_info;
6267 	gtask = &rxq[0].ifr_task;
6268 	tqg = qgroup_if_io_tqg;
6269 	tqrid = *rid;
6270 	rx_only = (ctx->ifc_sctx->isc_flags & IFLIB_SINGLE_IRQ_RX_ONLY) != 0;
6271 
6272 	ctx->ifc_flags |= IFC_LEGACY;
6273 	info->ifi_filter = filter;
6274 	info->ifi_filter_arg = filter_arg;
6275 	info->ifi_task = gtask;
6276 	info->ifi_ctx = rx_only ? ctx : q;
6277 
6278 	dev = ctx->ifc_dev;
6279 	/* We allocate a single interrupt resource */
6280 	err = _iflib_irq_alloc(ctx, irq, tqrid, rx_only ? iflib_fast_intr_ctx :
6281 	    iflib_fast_intr_rxtx, NULL, info, name);
6282 	if (err != 0)
6283 		return (err);
6284 	NET_GROUPTASK_INIT(gtask, 0, _task_fn_rx, q);
6285 	res = irq->ii_res;
6286 	taskqgroup_attach(tqg, gtask, q, dev, res, name);
6287 
6288 	GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq);
6289 	taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, dev, res,
6290 	    "tx");
6291 	return (0);
6292 }
6293 
6294 void
6295 iflib_led_create(if_ctx_t ctx)
6296 {
6297 
6298 	ctx->ifc_led_dev = led_create(iflib_led_func, ctx,
6299 	    device_get_nameunit(ctx->ifc_dev));
6300 }
6301 
6302 void
6303 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid)
6304 {
6305 
6306 	GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
6307 }
6308 
6309 void
6310 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid)
6311 {
6312 
6313 	GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task);
6314 }
6315 
6316 void
6317 iflib_admin_intr_deferred(if_ctx_t ctx)
6318 {
6319 
6320 	MPASS(ctx->ifc_admin_task.gt_taskqueue != NULL);
6321 	GROUPTASK_ENQUEUE(&ctx->ifc_admin_task);
6322 }
6323 
6324 void
6325 iflib_iov_intr_deferred(if_ctx_t ctx)
6326 {
6327 
6328 	GROUPTASK_ENQUEUE(&ctx->ifc_vflr_task);
6329 }
6330 
6331 void
6332 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, const char *name)
6333 {
6334 
6335 	taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, NULL, NULL,
6336 	    name);
6337 }
6338 
6339 void
6340 iflib_config_gtask_init(void *ctx, struct grouptask *gtask, gtask_fn_t *fn,
6341 	const char *name)
6342 {
6343 
6344 	GROUPTASK_INIT(gtask, 0, fn, ctx);
6345 	taskqgroup_attach(qgroup_if_config_tqg, gtask, gtask, NULL, NULL,
6346 	    name);
6347 }
6348 
6349 void
6350 iflib_config_gtask_deinit(struct grouptask *gtask)
6351 {
6352 
6353 	taskqgroup_detach(qgroup_if_config_tqg, gtask);
6354 }
6355 
6356 void
6357 iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate)
6358 {
6359 	if_t ifp = ctx->ifc_ifp;
6360 	iflib_txq_t txq = ctx->ifc_txqs;
6361 
6362 	if_setbaudrate(ifp, baudrate);
6363 	if (baudrate >= IF_Gbps(10)) {
6364 		STATE_LOCK(ctx);
6365 		ctx->ifc_flags |= IFC_PREFETCH;
6366 		STATE_UNLOCK(ctx);
6367 	}
6368 	/* If link down, disable watchdog */
6369 	if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) {
6370 		for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++)
6371 			txq->ift_qstatus = IFLIB_QUEUE_IDLE;
6372 	}
6373 	ctx->ifc_link_state = link_state;
6374 	if_link_state_change(ifp, link_state);
6375 }
6376 
6377 static int
6378 iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq)
6379 {
6380 	int credits;
6381 #ifdef INVARIANTS
6382 	int credits_pre = txq->ift_cidx_processed;
6383 #endif
6384 
6385 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
6386 	    BUS_DMASYNC_POSTREAD);
6387 	if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, true)) == 0)
6388 		return (0);
6389 
6390 	txq->ift_processed += credits;
6391 	txq->ift_cidx_processed += credits;
6392 
6393 	MPASS(credits_pre + credits == txq->ift_cidx_processed);
6394 	if (txq->ift_cidx_processed >= txq->ift_size)
6395 		txq->ift_cidx_processed -= txq->ift_size;
6396 	return (credits);
6397 }
6398 
6399 static int
6400 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget)
6401 {
6402 	iflib_fl_t fl;
6403 	u_int i;
6404 
6405 	for (i = 0, fl = &rxq->ifr_fl[0]; i < rxq->ifr_nfl; i++, fl++)
6406 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
6407 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
6408 	return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx,
6409 	    budget));
6410 }
6411 
6412 void
6413 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name,
6414 	const char *description, if_int_delay_info_t info,
6415 	int offset, int value)
6416 {
6417 	info->iidi_ctx = ctx;
6418 	info->iidi_offset = offset;
6419 	info->iidi_value = value;
6420 	SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev),
6421 	    SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)),
6422 	    OID_AUTO, name, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE,
6423 	    info, 0, iflib_sysctl_int_delay, "I", description);
6424 }
6425 
6426 struct sx *
6427 iflib_ctx_lock_get(if_ctx_t ctx)
6428 {
6429 
6430 	return (&ctx->ifc_ctx_sx);
6431 }
6432 
6433 static int
6434 iflib_msix_init(if_ctx_t ctx)
6435 {
6436 	device_t dev = ctx->ifc_dev;
6437 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6438 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6439 	int admincnt, bar, err, iflib_num_rx_queues, iflib_num_tx_queues;
6440 	int msgs, queuemsgs, queues, rx_queues, tx_queues, vectors;
6441 
6442 	iflib_num_tx_queues = ctx->ifc_sysctl_ntxqs;
6443 	iflib_num_rx_queues = ctx->ifc_sysctl_nrxqs;
6444 
6445 	if (bootverbose)
6446 		device_printf(dev, "msix_init qsets capped at %d\n",
6447 		    imax(scctx->isc_ntxqsets, scctx->isc_nrxqsets));
6448 
6449 	/* Override by tuneable */
6450 	if (scctx->isc_disable_msix)
6451 		goto msi;
6452 
6453 	/* First try MSI-X */
6454 	if ((msgs = pci_msix_count(dev)) == 0) {
6455 		if (bootverbose)
6456 			device_printf(dev, "MSI-X not supported or disabled\n");
6457 		goto msi;
6458 	}
6459 
6460 	bar = ctx->ifc_softc_ctx.isc_msix_bar;
6461 	/*
6462 	 * bar == -1 => "trust me I know what I'm doing"
6463 	 * Some drivers are for hardware that is so shoddily
6464 	 * documented that no one knows which bars are which
6465 	 * so the developer has to map all bars. This hack
6466 	 * allows shoddy garbage to use MSI-X in this framework.
6467 	 */
6468 	if (bar != -1) {
6469 		ctx->ifc_msix_mem = bus_alloc_resource_any(dev,
6470 	            SYS_RES_MEMORY, &bar, RF_ACTIVE);
6471 		if (ctx->ifc_msix_mem == NULL) {
6472 			device_printf(dev, "Unable to map MSI-X table\n");
6473 			goto msi;
6474 		}
6475 	}
6476 
6477 	admincnt = sctx->isc_admin_intrcnt;
6478 #if IFLIB_DEBUG
6479 	/* use only 1 qset in debug mode */
6480 	queuemsgs = min(msgs - admincnt, 1);
6481 #else
6482 	queuemsgs = msgs - admincnt;
6483 #endif
6484 #ifdef RSS
6485 	queues = imin(queuemsgs, rss_getnumbuckets());
6486 #else
6487 	queues = queuemsgs;
6488 #endif
6489 	queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues);
6490 	if (bootverbose)
6491 		device_printf(dev,
6492 		    "intr CPUs: %d queue msgs: %d admincnt: %d\n",
6493 		    CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt);
6494 #ifdef  RSS
6495 	/* If we're doing RSS, clamp at the number of RSS buckets */
6496 	if (queues > rss_getnumbuckets())
6497 		queues = rss_getnumbuckets();
6498 #endif
6499 	if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt)
6500 		rx_queues = iflib_num_rx_queues;
6501 	else
6502 		rx_queues = queues;
6503 
6504 	if (rx_queues > scctx->isc_nrxqsets)
6505 		rx_queues = scctx->isc_nrxqsets;
6506 
6507 	/*
6508 	 * We want this to be all logical CPUs by default
6509 	 */
6510 	if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues)
6511 		tx_queues = iflib_num_tx_queues;
6512 	else
6513 		tx_queues = mp_ncpus;
6514 
6515 	if (tx_queues > scctx->isc_ntxqsets)
6516 		tx_queues = scctx->isc_ntxqsets;
6517 
6518 	if (ctx->ifc_sysctl_qs_eq_override == 0) {
6519 #ifdef INVARIANTS
6520 		if (tx_queues != rx_queues)
6521 			device_printf(dev,
6522 			    "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n",
6523 			    min(rx_queues, tx_queues), min(rx_queues, tx_queues));
6524 #endif
6525 		tx_queues = min(rx_queues, tx_queues);
6526 		rx_queues = min(rx_queues, tx_queues);
6527 	}
6528 
6529 	vectors = rx_queues + admincnt;
6530 	if (msgs < vectors) {
6531 		device_printf(dev,
6532 		    "insufficient number of MSI-X vectors "
6533 		    "(supported %d, need %d)\n", msgs, vectors);
6534 		goto msi;
6535 	}
6536 
6537 	device_printf(dev, "Using %d RX queues %d TX queues\n", rx_queues,
6538 	    tx_queues);
6539 	msgs = vectors;
6540 	if ((err = pci_alloc_msix(dev, &vectors)) == 0) {
6541 		if (vectors != msgs) {
6542 			device_printf(dev,
6543 			    "Unable to allocate sufficient MSI-X vectors "
6544 			    "(got %d, need %d)\n", vectors, msgs);
6545 			pci_release_msi(dev);
6546 			if (bar != -1) {
6547 				bus_release_resource(dev, SYS_RES_MEMORY, bar,
6548 				    ctx->ifc_msix_mem);
6549 				ctx->ifc_msix_mem = NULL;
6550 			}
6551 			goto msi;
6552 		}
6553 		device_printf(dev, "Using MSI-X interrupts with %d vectors\n",
6554 		    vectors);
6555 		scctx->isc_vectors = vectors;
6556 		scctx->isc_nrxqsets = rx_queues;
6557 		scctx->isc_ntxqsets = tx_queues;
6558 		scctx->isc_intr = IFLIB_INTR_MSIX;
6559 
6560 		return (vectors);
6561 	} else {
6562 		device_printf(dev,
6563 		    "failed to allocate %d MSI-X vectors, err: %d\n", vectors,
6564 		    err);
6565 		if (bar != -1) {
6566 			bus_release_resource(dev, SYS_RES_MEMORY, bar,
6567 			    ctx->ifc_msix_mem);
6568 			ctx->ifc_msix_mem = NULL;
6569 		}
6570 	}
6571 
6572 msi:
6573 	vectors = pci_msi_count(dev);
6574 	scctx->isc_nrxqsets = 1;
6575 	scctx->isc_ntxqsets = 1;
6576 	scctx->isc_vectors = vectors;
6577 	if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) {
6578 		device_printf(dev,"Using an MSI interrupt\n");
6579 		scctx->isc_intr = IFLIB_INTR_MSI;
6580 	} else {
6581 		scctx->isc_vectors = 1;
6582 		device_printf(dev,"Using a Legacy interrupt\n");
6583 		scctx->isc_intr = IFLIB_INTR_LEGACY;
6584 	}
6585 
6586 	return (vectors);
6587 }
6588 
6589 static const char *ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" };
6590 
6591 static int
6592 mp_ring_state_handler(SYSCTL_HANDLER_ARGS)
6593 {
6594 	int rc;
6595 	uint16_t *state = ((uint16_t *)oidp->oid_arg1);
6596 	struct sbuf *sb;
6597 	const char *ring_state = "UNKNOWN";
6598 
6599 	/* XXX needed ? */
6600 	rc = sysctl_wire_old_buffer(req, 0);
6601 	MPASS(rc == 0);
6602 	if (rc != 0)
6603 		return (rc);
6604 	sb = sbuf_new_for_sysctl(NULL, NULL, 80, req);
6605 	MPASS(sb != NULL);
6606 	if (sb == NULL)
6607 		return (ENOMEM);
6608 	if (state[3] <= 3)
6609 		ring_state = ring_states[state[3]];
6610 
6611 	sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s",
6612 		    state[0], state[1], state[2], ring_state);
6613 	rc = sbuf_finish(sb);
6614 	sbuf_delete(sb);
6615         return(rc);
6616 }
6617 
6618 enum iflib_ndesc_handler {
6619 	IFLIB_NTXD_HANDLER,
6620 	IFLIB_NRXD_HANDLER,
6621 };
6622 
6623 static int
6624 mp_ndesc_handler(SYSCTL_HANDLER_ARGS)
6625 {
6626 	if_ctx_t ctx = (void *)arg1;
6627 	enum iflib_ndesc_handler type = arg2;
6628 	char buf[256] = {0};
6629 	qidx_t *ndesc;
6630 	char *p, *next;
6631 	int nqs, rc, i;
6632 
6633 	nqs = 8;
6634 	switch(type) {
6635 	case IFLIB_NTXD_HANDLER:
6636 		ndesc = ctx->ifc_sysctl_ntxds;
6637 		if (ctx->ifc_sctx)
6638 			nqs = ctx->ifc_sctx->isc_ntxqs;
6639 		break;
6640 	case IFLIB_NRXD_HANDLER:
6641 		ndesc = ctx->ifc_sysctl_nrxds;
6642 		if (ctx->ifc_sctx)
6643 			nqs = ctx->ifc_sctx->isc_nrxqs;
6644 		break;
6645 	default:
6646 		printf("%s: unhandled type\n", __func__);
6647 		return (EINVAL);
6648 	}
6649 	if (nqs == 0)
6650 		nqs = 8;
6651 
6652 	for (i=0; i<8; i++) {
6653 		if (i >= nqs)
6654 			break;
6655 		if (i)
6656 			strcat(buf, ",");
6657 		sprintf(strchr(buf, 0), "%d", ndesc[i]);
6658 	}
6659 
6660 	rc = sysctl_handle_string(oidp, buf, sizeof(buf), req);
6661 	if (rc || req->newptr == NULL)
6662 		return rc;
6663 
6664 	for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p;
6665 	    i++, p = strsep(&next, " ,")) {
6666 		ndesc[i] = strtoul(p, NULL, 10);
6667 	}
6668 
6669 	return(rc);
6670 }
6671 
6672 #define NAME_BUFLEN 32
6673 static void
6674 iflib_add_device_sysctl_pre(if_ctx_t ctx)
6675 {
6676         device_t dev = iflib_get_dev(ctx);
6677 	struct sysctl_oid_list *child, *oid_list;
6678 	struct sysctl_ctx_list *ctx_list;
6679 	struct sysctl_oid *node;
6680 
6681 	ctx_list = device_get_sysctl_ctx(dev);
6682 	child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev));
6683 	ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, "iflib",
6684 	    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "IFLIB fields");
6685 	oid_list = SYSCTL_CHILDREN(node);
6686 
6687 	SYSCTL_ADD_CONST_STRING(ctx_list, oid_list, OID_AUTO, "driver_version",
6688 		       CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version,
6689 		       "driver version");
6690 
6691 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs",
6692 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0,
6693 			"# of txqs to use, 0 => use default #");
6694 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs",
6695 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0,
6696 			"# of rxqs to use, 0 => use default #");
6697 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_qs_enable",
6698 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0,
6699                        "permit #txq != #rxq");
6700 	SYSCTL_ADD_INT(ctx_list, oid_list, OID_AUTO, "disable_msix",
6701                       CTLFLAG_RWTUN, &ctx->ifc_softc_ctx.isc_disable_msix, 0,
6702                       "disable MSI-X (default 0)");
6703 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "rx_budget",
6704 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_rx_budget, 0,
6705 		       "set the RX budget");
6706 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "tx_abdicate",
6707 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_tx_abdicate, 0,
6708 		       "cause TX to abdicate instead of running to completion");
6709 	ctx->ifc_sysctl_core_offset = CORE_OFFSET_UNSPECIFIED;
6710 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "core_offset",
6711 		       CTLFLAG_RDTUN, &ctx->ifc_sysctl_core_offset, 0,
6712 		       "offset to start using cores at");
6713 	SYSCTL_ADD_U8(ctx_list, oid_list, OID_AUTO, "separate_txrx",
6714 		       CTLFLAG_RDTUN, &ctx->ifc_sysctl_separate_txrx, 0,
6715 		       "use separate cores for TX and RX");
6716 
6717 	/* XXX change for per-queue sizes */
6718 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_ntxds",
6719 	    CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NEEDGIANT, ctx,
6720 	    IFLIB_NTXD_HANDLER, mp_ndesc_handler, "A",
6721 	    "list of # of TX descriptors to use, 0 = use default #");
6722 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_nrxds",
6723 	    CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NEEDGIANT, ctx,
6724 	    IFLIB_NRXD_HANDLER, mp_ndesc_handler, "A",
6725 	    "list of # of RX descriptors to use, 0 = use default #");
6726 }
6727 
6728 static void
6729 iflib_add_device_sysctl_post(if_ctx_t ctx)
6730 {
6731 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6732 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6733         device_t dev = iflib_get_dev(ctx);
6734 	struct sysctl_oid_list *child;
6735 	struct sysctl_ctx_list *ctx_list;
6736 	iflib_fl_t fl;
6737 	iflib_txq_t txq;
6738 	iflib_rxq_t rxq;
6739 	int i, j;
6740 	char namebuf[NAME_BUFLEN];
6741 	char *qfmt;
6742 	struct sysctl_oid *queue_node, *fl_node, *node;
6743 	struct sysctl_oid_list *queue_list, *fl_list;
6744 	ctx_list = device_get_sysctl_ctx(dev);
6745 
6746 	node = ctx->ifc_sysctl_node;
6747 	child = SYSCTL_CHILDREN(node);
6748 
6749 	if (scctx->isc_ntxqsets > 100)
6750 		qfmt = "txq%03d";
6751 	else if (scctx->isc_ntxqsets > 10)
6752 		qfmt = "txq%02d";
6753 	else
6754 		qfmt = "txq%d";
6755 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
6756 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6757 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6758 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "Queue Name");
6759 		queue_list = SYSCTL_CHILDREN(queue_node);
6760 #if MEMORY_LOGGING
6761 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued",
6762 				CTLFLAG_RD,
6763 				&txq->ift_dequeued, "total mbufs freed");
6764 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued",
6765 				CTLFLAG_RD,
6766 				&txq->ift_enqueued, "total mbufs enqueued");
6767 #endif
6768 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag",
6769 				   CTLFLAG_RD,
6770 				   &txq->ift_mbuf_defrag, "# of times m_defrag was called");
6771 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "m_pullups",
6772 				   CTLFLAG_RD,
6773 				   &txq->ift_pullups, "# of times m_pullup was called");
6774 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag_failed",
6775 				   CTLFLAG_RD,
6776 				   &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed");
6777 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_desc_avail",
6778 				   CTLFLAG_RD,
6779 				   &txq->ift_no_desc_avail, "# of times no descriptors were available");
6780 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "tx_map_failed",
6781 				   CTLFLAG_RD,
6782 				   &txq->ift_map_failed, "# of times DMA map failed");
6783 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txd_encap_efbig",
6784 				   CTLFLAG_RD,
6785 				   &txq->ift_txd_encap_efbig, "# of times txd_encap returned EFBIG");
6786 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_tx_dma_setup",
6787 				   CTLFLAG_RD,
6788 				   &txq->ift_no_tx_dma_setup, "# of times map failed for other than EFBIG");
6789 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx",
6790 				   CTLFLAG_RD,
6791 				   &txq->ift_pidx, 1, "Producer Index");
6792 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx",
6793 				   CTLFLAG_RD,
6794 				   &txq->ift_cidx, 1, "Consumer Index");
6795 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx_processed",
6796 				   CTLFLAG_RD,
6797 				   &txq->ift_cidx_processed, 1, "Consumer Index seen by credit update");
6798 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use",
6799 				   CTLFLAG_RD,
6800 				   &txq->ift_in_use, 1, "descriptors in use");
6801 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_processed",
6802 				   CTLFLAG_RD,
6803 				   &txq->ift_processed, "descriptors procesed for clean");
6804 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned",
6805 				   CTLFLAG_RD,
6806 				   &txq->ift_cleaned, "total cleaned");
6807 		SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state",
6808 		    CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_NEEDGIANT,
6809 		    __DEVOLATILE(uint64_t *, &txq->ift_br->state), 0,
6810 		    mp_ring_state_handler, "A", "soft ring state");
6811 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_enqueues",
6812 				       CTLFLAG_RD, &txq->ift_br->enqueues,
6813 				       "# of enqueues to the mp_ring for this queue");
6814 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_drops",
6815 				       CTLFLAG_RD, &txq->ift_br->drops,
6816 				       "# of drops in the mp_ring for this queue");
6817 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_starts",
6818 				       CTLFLAG_RD, &txq->ift_br->starts,
6819 				       "# of normal consumer starts in the mp_ring for this queue");
6820 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_stalls",
6821 				       CTLFLAG_RD, &txq->ift_br->stalls,
6822 					       "# of consumer stalls in the mp_ring for this queue");
6823 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_restarts",
6824 			       CTLFLAG_RD, &txq->ift_br->restarts,
6825 				       "# of consumer restarts in the mp_ring for this queue");
6826 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_abdications",
6827 				       CTLFLAG_RD, &txq->ift_br->abdications,
6828 				       "# of consumer abdications in the mp_ring for this queue");
6829 	}
6830 
6831 	if (scctx->isc_nrxqsets > 100)
6832 		qfmt = "rxq%03d";
6833 	else if (scctx->isc_nrxqsets > 10)
6834 		qfmt = "rxq%02d";
6835 	else
6836 		qfmt = "rxq%d";
6837 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
6838 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6839 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6840 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "Queue Name");
6841 		queue_list = SYSCTL_CHILDREN(queue_node);
6842 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
6843 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_cidx",
6844 				       CTLFLAG_RD,
6845 				       &rxq->ifr_cq_cidx, 1, "Consumer Index");
6846 		}
6847 
6848 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
6849 			snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j);
6850 			fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list, OID_AUTO, namebuf,
6851 			    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "freelist Name");
6852 			fl_list = SYSCTL_CHILDREN(fl_node);
6853 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx",
6854 				       CTLFLAG_RD,
6855 				       &fl->ifl_pidx, 1, "Producer Index");
6856 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx",
6857 				       CTLFLAG_RD,
6858 				       &fl->ifl_cidx, 1, "Consumer Index");
6859 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits",
6860 				       CTLFLAG_RD,
6861 				       &fl->ifl_credits, 1, "credits available");
6862 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "buf_size",
6863 				       CTLFLAG_RD,
6864 				       &fl->ifl_buf_size, 1, "buffer size");
6865 #if MEMORY_LOGGING
6866 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_enqueued",
6867 					CTLFLAG_RD,
6868 					&fl->ifl_m_enqueued, "mbufs allocated");
6869 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_dequeued",
6870 					CTLFLAG_RD,
6871 					&fl->ifl_m_dequeued, "mbufs freed");
6872 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_enqueued",
6873 					CTLFLAG_RD,
6874 					&fl->ifl_cl_enqueued, "clusters allocated");
6875 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_dequeued",
6876 					CTLFLAG_RD,
6877 					&fl->ifl_cl_dequeued, "clusters freed");
6878 #endif
6879 		}
6880 	}
6881 
6882 }
6883 
6884 void
6885 iflib_request_reset(if_ctx_t ctx)
6886 {
6887 
6888 	STATE_LOCK(ctx);
6889 	ctx->ifc_flags |= IFC_DO_RESET;
6890 	STATE_UNLOCK(ctx);
6891 }
6892 
6893 #ifndef __NO_STRICT_ALIGNMENT
6894 static struct mbuf *
6895 iflib_fixup_rx(struct mbuf *m)
6896 {
6897 	struct mbuf *n;
6898 
6899 	if (m->m_len <= (MCLBYTES - ETHER_HDR_LEN)) {
6900 		bcopy(m->m_data, m->m_data + ETHER_HDR_LEN, m->m_len);
6901 		m->m_data += ETHER_HDR_LEN;
6902 		n = m;
6903 	} else {
6904 		MGETHDR(n, M_NOWAIT, MT_DATA);
6905 		if (n == NULL) {
6906 			m_freem(m);
6907 			return (NULL);
6908 		}
6909 		bcopy(m->m_data, n->m_data, ETHER_HDR_LEN);
6910 		m->m_data += ETHER_HDR_LEN;
6911 		m->m_len -= ETHER_HDR_LEN;
6912 		n->m_len = ETHER_HDR_LEN;
6913 		M_MOVE_PKTHDR(n, m);
6914 		n->m_next = m;
6915 	}
6916 	return (n);
6917 }
6918 #endif
6919 
6920 #ifdef DEBUGNET
6921 static void
6922 iflib_debugnet_init(if_t ifp, int *nrxr, int *ncl, int *clsize)
6923 {
6924 	if_ctx_t ctx;
6925 
6926 	ctx = if_getsoftc(ifp);
6927 	CTX_LOCK(ctx);
6928 	*nrxr = NRXQSETS(ctx);
6929 	*ncl = ctx->ifc_rxqs[0].ifr_fl->ifl_size;
6930 	*clsize = ctx->ifc_rxqs[0].ifr_fl->ifl_buf_size;
6931 	CTX_UNLOCK(ctx);
6932 }
6933 
6934 static void
6935 iflib_debugnet_event(if_t ifp, enum debugnet_ev event)
6936 {
6937 	if_ctx_t ctx;
6938 	if_softc_ctx_t scctx;
6939 	iflib_fl_t fl;
6940 	iflib_rxq_t rxq;
6941 	int i, j;
6942 
6943 	ctx = if_getsoftc(ifp);
6944 	scctx = &ctx->ifc_softc_ctx;
6945 
6946 	switch (event) {
6947 	case DEBUGNET_START:
6948 		for (i = 0; i < scctx->isc_nrxqsets; i++) {
6949 			rxq = &ctx->ifc_rxqs[i];
6950 			for (j = 0; j < rxq->ifr_nfl; j++) {
6951 				fl = rxq->ifr_fl;
6952 				fl->ifl_zone = m_getzone(fl->ifl_buf_size);
6953 			}
6954 		}
6955 		iflib_no_tx_batch = 1;
6956 		break;
6957 	default:
6958 		break;
6959 	}
6960 }
6961 
6962 static int
6963 iflib_debugnet_transmit(if_t ifp, struct mbuf *m)
6964 {
6965 	if_ctx_t ctx;
6966 	iflib_txq_t txq;
6967 	int error;
6968 
6969 	ctx = if_getsoftc(ifp);
6970 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
6971 	    IFF_DRV_RUNNING)
6972 		return (EBUSY);
6973 
6974 	txq = &ctx->ifc_txqs[0];
6975 	error = iflib_encap(txq, &m);
6976 	if (error == 0)
6977 		(void)iflib_txd_db_check(txq, true);
6978 	return (error);
6979 }
6980 
6981 static int
6982 iflib_debugnet_poll(if_t ifp, int count)
6983 {
6984 	struct epoch_tracker et;
6985 	if_ctx_t ctx;
6986 	if_softc_ctx_t scctx;
6987 	iflib_txq_t txq;
6988 	int i;
6989 
6990 	ctx = if_getsoftc(ifp);
6991 	scctx = &ctx->ifc_softc_ctx;
6992 
6993 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
6994 	    IFF_DRV_RUNNING)
6995 		return (EBUSY);
6996 
6997 	txq = &ctx->ifc_txqs[0];
6998 	(void)iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
6999 
7000 	NET_EPOCH_ENTER(et);
7001 	for (i = 0; i < scctx->isc_nrxqsets; i++)
7002 		(void)iflib_rxeof(&ctx->ifc_rxqs[i], 16 /* XXX */);
7003 	NET_EPOCH_EXIT(et);
7004 	return (0);
7005 }
7006 #endif /* DEBUGNET */
7007