xref: /freebsd/sys/net/iflib.c (revision 99db5849f7506e765c43f4e69a7105cc888e8d5e)
1 /*-
2  * Copyright (c) 2014-2018, Matthew Macy <mmacy@mattmacy.io>
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions are met:
7  *
8  *  1. Redistributions of source code must retain the above copyright notice,
9  *     this list of conditions and the following disclaimer.
10  *
11  *  2. Neither the name of Matthew Macy nor the names of its
12  *     contributors may be used to endorse or promote products derived from
13  *     this software without specific prior written permission.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
16  * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18  * ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
19  * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
20  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
21  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
22  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
23  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
24  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
25  * POSSIBILITY OF SUCH DAMAGE.
26  */
27 
28 #include <sys/cdefs.h>
29 __FBSDID("$FreeBSD$");
30 
31 #include "opt_inet.h"
32 #include "opt_inet6.h"
33 #include "opt_acpi.h"
34 #include "opt_sched.h"
35 
36 #include <sys/param.h>
37 #include <sys/types.h>
38 #include <sys/bus.h>
39 #include <sys/eventhandler.h>
40 #include <sys/kernel.h>
41 #include <sys/lock.h>
42 #include <sys/mutex.h>
43 #include <sys/module.h>
44 #include <sys/kobj.h>
45 #include <sys/rman.h>
46 #include <sys/sbuf.h>
47 #include <sys/smp.h>
48 #include <sys/socket.h>
49 #include <sys/sockio.h>
50 #include <sys/sysctl.h>
51 #include <sys/syslog.h>
52 #include <sys/taskqueue.h>
53 #include <sys/limits.h>
54 
55 #include <net/if.h>
56 #include <net/if_var.h>
57 #include <net/if_types.h>
58 #include <net/if_media.h>
59 #include <net/bpf.h>
60 #include <net/ethernet.h>
61 #include <net/mp_ring.h>
62 #include <net/debugnet.h>
63 #include <net/pfil.h>
64 #include <net/vnet.h>
65 
66 #include <netinet/in.h>
67 #include <netinet/in_pcb.h>
68 #include <netinet/tcp_lro.h>
69 #include <netinet/in_systm.h>
70 #include <netinet/if_ether.h>
71 #include <netinet/ip.h>
72 #include <netinet/ip6.h>
73 #include <netinet/tcp.h>
74 #include <netinet/ip_var.h>
75 #include <netinet6/ip6_var.h>
76 
77 #include <machine/bus.h>
78 #include <machine/in_cksum.h>
79 
80 #include <vm/vm.h>
81 #include <vm/pmap.h>
82 
83 #include <dev/led/led.h>
84 #include <dev/pci/pcireg.h>
85 #include <dev/pci/pcivar.h>
86 #include <dev/pci/pci_private.h>
87 
88 #include <net/iflib.h>
89 #include <net/iflib_private.h>
90 
91 #include "ifdi_if.h"
92 
93 #ifdef PCI_IOV
94 #include <dev/pci/pci_iov.h>
95 #endif
96 
97 #include <sys/bitstring.h>
98 /*
99  * enable accounting of every mbuf as it comes in to and goes out of
100  * iflib's software descriptor references
101  */
102 #define MEMORY_LOGGING 0
103 /*
104  * Enable mbuf vectors for compressing long mbuf chains
105  */
106 
107 /*
108  * NB:
109  * - Prefetching in tx cleaning should perhaps be a tunable. The distance ahead
110  *   we prefetch needs to be determined by the time spent in m_free vis a vis
111  *   the cost of a prefetch. This will of course vary based on the workload:
112  *      - NFLX's m_free path is dominated by vm-based M_EXT manipulation which
113  *        is quite expensive, thus suggesting very little prefetch.
114  *      - small packet forwarding which is just returning a single mbuf to
115  *        UMA will typically be very fast vis a vis the cost of a memory
116  *        access.
117  */
118 
119 
120 /*
121  * File organization:
122  *  - private structures
123  *  - iflib private utility functions
124  *  - ifnet functions
125  *  - vlan registry and other exported functions
126  *  - iflib public core functions
127  *
128  *
129  */
130 MALLOC_DEFINE(M_IFLIB, "iflib", "ifnet library");
131 
132 #define	IFLIB_RXEOF_MORE (1U << 0)
133 #define	IFLIB_RXEOF_EMPTY (2U << 0)
134 
135 struct iflib_txq;
136 typedef struct iflib_txq *iflib_txq_t;
137 struct iflib_rxq;
138 typedef struct iflib_rxq *iflib_rxq_t;
139 struct iflib_fl;
140 typedef struct iflib_fl *iflib_fl_t;
141 
142 struct iflib_ctx;
143 
144 static void iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid);
145 static void iflib_timer(void *arg);
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_rxd_refill ifc_txrx.ift_rxd_refill
210 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
211 #define isc_legacy_intr ifc_txrx.ift_legacy_intr
212 	eventhandler_tag ifc_vlan_attach_event;
213 	eventhandler_tag ifc_vlan_detach_event;
214 	struct ether_addr ifc_mac;
215 };
216 
217 void *
218 iflib_get_softc(if_ctx_t ctx)
219 {
220 
221 	return (ctx->ifc_softc);
222 }
223 
224 device_t
225 iflib_get_dev(if_ctx_t ctx)
226 {
227 
228 	return (ctx->ifc_dev);
229 }
230 
231 if_t
232 iflib_get_ifp(if_ctx_t ctx)
233 {
234 
235 	return (ctx->ifc_ifp);
236 }
237 
238 struct ifmedia *
239 iflib_get_media(if_ctx_t ctx)
240 {
241 
242 	return (ctx->ifc_mediap);
243 }
244 
245 uint32_t
246 iflib_get_flags(if_ctx_t ctx)
247 {
248 	return (ctx->ifc_flags);
249 }
250 
251 void
252 iflib_set_mac(if_ctx_t ctx, uint8_t mac[ETHER_ADDR_LEN])
253 {
254 
255 	bcopy(mac, ctx->ifc_mac.octet, ETHER_ADDR_LEN);
256 }
257 
258 if_softc_ctx_t
259 iflib_get_softc_ctx(if_ctx_t ctx)
260 {
261 
262 	return (&ctx->ifc_softc_ctx);
263 }
264 
265 if_shared_ctx_t
266 iflib_get_sctx(if_ctx_t ctx)
267 {
268 
269 	return (ctx->ifc_sctx);
270 }
271 
272 #define IP_ALIGNED(m) ((((uintptr_t)(m)->m_data) & 0x3) == 0x2)
273 #define CACHE_PTR_INCREMENT (CACHE_LINE_SIZE/sizeof(void*))
274 #define CACHE_PTR_NEXT(ptr) ((void *)(((uintptr_t)(ptr)+CACHE_LINE_SIZE-1) & (CACHE_LINE_SIZE-1)))
275 
276 #define LINK_ACTIVE(ctx) ((ctx)->ifc_link_state == LINK_STATE_UP)
277 #define CTX_IS_VF(ctx) ((ctx)->ifc_sctx->isc_flags & IFLIB_IS_VF)
278 
279 typedef struct iflib_sw_rx_desc_array {
280 	bus_dmamap_t	*ifsd_map;         /* bus_dma maps for packet */
281 	struct mbuf	**ifsd_m;           /* pkthdr mbufs */
282 	caddr_t		*ifsd_cl;          /* direct cluster pointer for rx */
283 	bus_addr_t	*ifsd_ba;          /* bus addr of cluster for rx */
284 } iflib_rxsd_array_t;
285 
286 typedef struct iflib_sw_tx_desc_array {
287 	bus_dmamap_t    *ifsd_map;         /* bus_dma maps for packet */
288 	bus_dmamap_t	*ifsd_tso_map;     /* bus_dma maps for TSO packet */
289 	struct mbuf    **ifsd_m;           /* pkthdr mbufs */
290 } if_txsd_vec_t;
291 
292 /* magic number that should be high enough for any hardware */
293 #define IFLIB_MAX_TX_SEGS		128
294 #define IFLIB_RX_COPY_THRESH		128
295 #define IFLIB_MAX_RX_REFRESH		32
296 /* The minimum descriptors per second before we start coalescing */
297 #define IFLIB_MIN_DESC_SEC		16384
298 #define IFLIB_DEFAULT_TX_UPDATE_FREQ	16
299 #define IFLIB_QUEUE_IDLE		0
300 #define IFLIB_QUEUE_HUNG		1
301 #define IFLIB_QUEUE_WORKING		2
302 /* maximum number of txqs that can share an rx interrupt */
303 #define IFLIB_MAX_TX_SHARED_INTR	4
304 
305 /* this should really scale with ring size - this is a fairly arbitrary value */
306 #define TX_BATCH_SIZE			32
307 
308 #define IFLIB_RESTART_BUDGET		8
309 
310 #define CSUM_OFFLOAD		(CSUM_IP_TSO|CSUM_IP6_TSO|CSUM_IP| \
311 				 CSUM_IP_UDP|CSUM_IP_TCP|CSUM_IP_SCTP| \
312 				 CSUM_IP6_UDP|CSUM_IP6_TCP|CSUM_IP6_SCTP)
313 
314 struct iflib_txq {
315 	qidx_t		ift_in_use;
316 	qidx_t		ift_cidx;
317 	qidx_t		ift_cidx_processed;
318 	qidx_t		ift_pidx;
319 	uint8_t		ift_gen;
320 	uint8_t		ift_br_offset;
321 	uint16_t	ift_npending;
322 	uint16_t	ift_db_pending;
323 	uint16_t	ift_rs_pending;
324 	/* implicit pad */
325 	uint8_t		ift_txd_size[8];
326 	uint64_t	ift_processed;
327 	uint64_t	ift_cleaned;
328 	uint64_t	ift_cleaned_prev;
329 #if MEMORY_LOGGING
330 	uint64_t	ift_enqueued;
331 	uint64_t	ift_dequeued;
332 #endif
333 	uint64_t	ift_no_tx_dma_setup;
334 	uint64_t	ift_no_desc_avail;
335 	uint64_t	ift_mbuf_defrag_failed;
336 	uint64_t	ift_mbuf_defrag;
337 	uint64_t	ift_map_failed;
338 	uint64_t	ift_txd_encap_efbig;
339 	uint64_t	ift_pullups;
340 	uint64_t	ift_last_timer_tick;
341 
342 	struct mtx	ift_mtx;
343 	struct mtx	ift_db_mtx;
344 
345 	/* constant values */
346 	if_ctx_t	ift_ctx;
347 	struct ifmp_ring        *ift_br;
348 	struct grouptask	ift_task;
349 	qidx_t		ift_size;
350 	uint16_t	ift_id;
351 	struct callout	ift_timer;
352 
353 	if_txsd_vec_t	ift_sds;
354 	uint8_t		ift_qstatus;
355 	uint8_t		ift_closed;
356 	uint8_t		ift_update_freq;
357 	struct iflib_filter_info ift_filter_info;
358 	bus_dma_tag_t	ift_buf_tag;
359 	bus_dma_tag_t	ift_tso_buf_tag;
360 	iflib_dma_info_t	ift_ifdi;
361 #define	MTX_NAME_LEN	32
362 	char                    ift_mtx_name[MTX_NAME_LEN];
363 	bus_dma_segment_t	ift_segs[IFLIB_MAX_TX_SEGS]  __aligned(CACHE_LINE_SIZE);
364 #ifdef IFLIB_DIAGNOSTICS
365 	uint64_t ift_cpu_exec_count[256];
366 #endif
367 } __aligned(CACHE_LINE_SIZE);
368 
369 struct iflib_fl {
370 	qidx_t		ifl_cidx;
371 	qidx_t		ifl_pidx;
372 	qidx_t		ifl_credits;
373 	uint8_t		ifl_gen;
374 	uint8_t		ifl_rxd_size;
375 #if MEMORY_LOGGING
376 	uint64_t	ifl_m_enqueued;
377 	uint64_t	ifl_m_dequeued;
378 	uint64_t	ifl_cl_enqueued;
379 	uint64_t	ifl_cl_dequeued;
380 #endif
381 	/* implicit pad */
382 	bitstr_t 	*ifl_rx_bitmap;
383 	qidx_t		ifl_fragidx;
384 	/* constant */
385 	qidx_t		ifl_size;
386 	uint16_t	ifl_buf_size;
387 	uint16_t	ifl_cltype;
388 	uma_zone_t	ifl_zone;
389 	iflib_rxsd_array_t	ifl_sds;
390 	iflib_rxq_t	ifl_rxq;
391 	uint8_t		ifl_id;
392 	bus_dma_tag_t	ifl_buf_tag;
393 	iflib_dma_info_t	ifl_ifdi;
394 	uint64_t	ifl_bus_addrs[IFLIB_MAX_RX_REFRESH] __aligned(CACHE_LINE_SIZE);
395 	caddr_t		ifl_vm_addrs[IFLIB_MAX_RX_REFRESH];
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 command 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 
480 static inline void
481 pkt_info_zero(if_pkt_info_t pi)
482 {
483 	if_pkt_info_pad_t pi_pad;
484 
485 	pi_pad = (if_pkt_info_pad_t)pi;
486 	pi_pad->pkt_val[0] = 0; pi_pad->pkt_val[1] = 0; pi_pad->pkt_val[2] = 0;
487 	pi_pad->pkt_val[3] = 0; pi_pad->pkt_val[4] = 0; pi_pad->pkt_val[5] = 0;
488 #ifndef __LP64__
489 	pi_pad->pkt_val[6] = 0; pi_pad->pkt_val[7] = 0; pi_pad->pkt_val[8] = 0;
490 	pi_pad->pkt_val[9] = 0; pi_pad->pkt_val[10] = 0;
491 #endif
492 }
493 
494 static device_method_t iflib_pseudo_methods[] = {
495 	DEVMETHOD(device_attach, noop_attach),
496 	DEVMETHOD(device_detach, iflib_pseudo_detach),
497 	DEVMETHOD_END
498 };
499 
500 driver_t iflib_pseudodriver = {
501 	"iflib_pseudo", iflib_pseudo_methods, sizeof(struct iflib_ctx),
502 };
503 
504 static inline void
505 rxd_info_zero(if_rxd_info_t ri)
506 {
507 	if_rxd_info_pad_t ri_pad;
508 	int i;
509 
510 	ri_pad = (if_rxd_info_pad_t)ri;
511 	for (i = 0; i < RXD_LOOP_BOUND; i += 4) {
512 		ri_pad->rxd_val[i] = 0;
513 		ri_pad->rxd_val[i+1] = 0;
514 		ri_pad->rxd_val[i+2] = 0;
515 		ri_pad->rxd_val[i+3] = 0;
516 	}
517 #ifdef __LP64__
518 	ri_pad->rxd_val[RXD_INFO_SIZE-1] = 0;
519 #endif
520 }
521 
522 /*
523  * Only allow a single packet to take up most 1/nth of the tx ring
524  */
525 #define MAX_SINGLE_PACKET_FRACTION 12
526 #define IF_BAD_DMA (bus_addr_t)-1
527 
528 #define CTX_ACTIVE(ctx) ((if_getdrvflags((ctx)->ifc_ifp) & IFF_DRV_RUNNING))
529 
530 #define CTX_LOCK_INIT(_sc)  sx_init(&(_sc)->ifc_ctx_sx, "iflib ctx lock")
531 #define CTX_LOCK(ctx) sx_xlock(&(ctx)->ifc_ctx_sx)
532 #define CTX_UNLOCK(ctx) sx_xunlock(&(ctx)->ifc_ctx_sx)
533 #define CTX_LOCK_DESTROY(ctx) sx_destroy(&(ctx)->ifc_ctx_sx)
534 
535 #define STATE_LOCK_INIT(_sc, _name)  mtx_init(&(_sc)->ifc_state_mtx, _name, "iflib state lock", MTX_DEF)
536 #define STATE_LOCK(ctx) mtx_lock(&(ctx)->ifc_state_mtx)
537 #define STATE_UNLOCK(ctx) mtx_unlock(&(ctx)->ifc_state_mtx)
538 #define STATE_LOCK_DESTROY(ctx) mtx_destroy(&(ctx)->ifc_state_mtx)
539 
540 #define CALLOUT_LOCK(txq)	mtx_lock(&txq->ift_mtx)
541 #define CALLOUT_UNLOCK(txq) 	mtx_unlock(&txq->ift_mtx)
542 
543 void
544 iflib_set_detach(if_ctx_t ctx)
545 {
546 	STATE_LOCK(ctx);
547 	ctx->ifc_flags |= IFC_IN_DETACH;
548 	STATE_UNLOCK(ctx);
549 }
550 
551 /* Our boot-time initialization hook */
552 static int	iflib_module_event_handler(module_t, int, void *);
553 
554 static moduledata_t iflib_moduledata = {
555 	"iflib",
556 	iflib_module_event_handler,
557 	NULL
558 };
559 
560 DECLARE_MODULE(iflib, iflib_moduledata, SI_SUB_INIT_IF, SI_ORDER_ANY);
561 MODULE_VERSION(iflib, 1);
562 
563 MODULE_DEPEND(iflib, pci, 1, 1, 1);
564 MODULE_DEPEND(iflib, ether, 1, 1, 1);
565 
566 TASKQGROUP_DEFINE(if_io_tqg, mp_ncpus, 1);
567 TASKQGROUP_DEFINE(if_config_tqg, 1, 1);
568 
569 #ifndef IFLIB_DEBUG_COUNTERS
570 #ifdef INVARIANTS
571 #define IFLIB_DEBUG_COUNTERS 1
572 #else
573 #define IFLIB_DEBUG_COUNTERS 0
574 #endif /* !INVARIANTS */
575 #endif
576 
577 static SYSCTL_NODE(_net, OID_AUTO, iflib, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
578     "iflib driver parameters");
579 
580 /*
581  * XXX need to ensure that this can't accidentally cause the head to be moved backwards
582  */
583 static int iflib_min_tx_latency = 0;
584 SYSCTL_INT(_net_iflib, OID_AUTO, min_tx_latency, CTLFLAG_RW,
585 		   &iflib_min_tx_latency, 0, "minimize transmit latency at the possible expense of throughput");
586 static int iflib_no_tx_batch = 0;
587 SYSCTL_INT(_net_iflib, OID_AUTO, no_tx_batch, CTLFLAG_RW,
588 		   &iflib_no_tx_batch, 0, "minimize transmit latency at the possible expense of throughput");
589 
590 
591 #if IFLIB_DEBUG_COUNTERS
592 
593 static int iflib_tx_seen;
594 static int iflib_tx_sent;
595 static int iflib_tx_encap;
596 static int iflib_rx_allocs;
597 static int iflib_fl_refills;
598 static int iflib_fl_refills_large;
599 static int iflib_tx_frees;
600 
601 SYSCTL_INT(_net_iflib, OID_AUTO, tx_seen, CTLFLAG_RD,
602 		   &iflib_tx_seen, 0, "# TX mbufs seen");
603 SYSCTL_INT(_net_iflib, OID_AUTO, tx_sent, CTLFLAG_RD,
604 		   &iflib_tx_sent, 0, "# TX mbufs sent");
605 SYSCTL_INT(_net_iflib, OID_AUTO, tx_encap, CTLFLAG_RD,
606 		   &iflib_tx_encap, 0, "# TX mbufs encapped");
607 SYSCTL_INT(_net_iflib, OID_AUTO, tx_frees, CTLFLAG_RD,
608 		   &iflib_tx_frees, 0, "# TX frees");
609 SYSCTL_INT(_net_iflib, OID_AUTO, rx_allocs, CTLFLAG_RD,
610 		   &iflib_rx_allocs, 0, "# RX allocations");
611 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills, CTLFLAG_RD,
612 		   &iflib_fl_refills, 0, "# refills");
613 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills_large, CTLFLAG_RD,
614 		   &iflib_fl_refills_large, 0, "# large refills");
615 
616 
617 static int iflib_txq_drain_flushing;
618 static int iflib_txq_drain_oactive;
619 static int iflib_txq_drain_notready;
620 
621 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_flushing, CTLFLAG_RD,
622 		   &iflib_txq_drain_flushing, 0, "# drain flushes");
623 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_oactive, CTLFLAG_RD,
624 		   &iflib_txq_drain_oactive, 0, "# drain oactives");
625 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_notready, CTLFLAG_RD,
626 		   &iflib_txq_drain_notready, 0, "# drain notready");
627 
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 #ifdef DEV_NETMAP
737 #include <sys/selinfo.h>
738 #include <net/netmap.h>
739 #include <dev/netmap/netmap_kern.h>
740 
741 MODULE_DEPEND(iflib, netmap, 1, 1, 1);
742 
743 static int netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, uint32_t nm_i, bool init);
744 
745 /*
746  * device-specific sysctl variables:
747  *
748  * iflib_crcstrip: 0: keep CRC in rx frames (default), 1: strip it.
749  *	During regular operations the CRC is stripped, but on some
750  *	hardware reception of frames not multiple of 64 is slower,
751  *	so using crcstrip=0 helps in benchmarks.
752  *
753  * iflib_rx_miss, iflib_rx_miss_bufs:
754  *	count packets that might be missed due to lost interrupts.
755  */
756 SYSCTL_DECL(_dev_netmap);
757 /*
758  * The xl driver by default strips CRCs and we do not override it.
759  */
760 
761 int iflib_crcstrip = 1;
762 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_crcstrip,
763     CTLFLAG_RW, &iflib_crcstrip, 1, "strip CRC on RX frames");
764 
765 int iflib_rx_miss, iflib_rx_miss_bufs;
766 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss,
767     CTLFLAG_RW, &iflib_rx_miss, 0, "potentially missed RX intr");
768 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss_bufs,
769     CTLFLAG_RW, &iflib_rx_miss_bufs, 0, "potentially missed RX intr bufs");
770 
771 /*
772  * Register/unregister. We are already under netmap lock.
773  * Only called on the first register or the last unregister.
774  */
775 static int
776 iflib_netmap_register(struct netmap_adapter *na, int onoff)
777 {
778 	if_t ifp = na->ifp;
779 	if_ctx_t ctx = ifp->if_softc;
780 	int status;
781 
782 	CTX_LOCK(ctx);
783 	IFDI_INTR_DISABLE(ctx);
784 
785 	/* Tell the stack that the interface is no longer active */
786 	ifp->if_drv_flags &= ~(IFF_DRV_RUNNING | IFF_DRV_OACTIVE);
787 
788 	if (!CTX_IS_VF(ctx))
789 		IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip);
790 
791 	iflib_stop(ctx);
792 
793 	/*
794 	 * Enable (or disable) netmap flags, and intercept (or restore)
795 	 * ifp->if_transmit. This is done once the device has been stopped
796 	 * to prevent race conditions.
797 	 */
798 	if (onoff) {
799 		nm_set_native_flags(na);
800 	} else {
801 		nm_clear_native_flags(na);
802 	}
803 
804 	iflib_init_locked(ctx);
805 	IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip); // XXX why twice ?
806 	status = ifp->if_drv_flags & IFF_DRV_RUNNING ? 0 : 1;
807 	if (status)
808 		nm_clear_native_flags(na);
809 	CTX_UNLOCK(ctx);
810 	return (status);
811 }
812 
813 static int
814 netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, uint32_t nm_i, bool init)
815 {
816 	struct netmap_adapter *na = kring->na;
817 	u_int const lim = kring->nkr_num_slots - 1;
818 	u_int head = kring->rhead;
819 	struct netmap_ring *ring = kring->ring;
820 	bus_dmamap_t *map;
821 	struct if_rxd_update iru;
822 	if_ctx_t ctx = rxq->ifr_ctx;
823 	iflib_fl_t fl = &rxq->ifr_fl[0];
824 	uint32_t refill_pidx, nic_i;
825 #if IFLIB_DEBUG_COUNTERS
826 	int rf_count = 0;
827 #endif
828 
829 	if (nm_i == head && __predict_true(!init))
830 		return 0;
831 	iru_init(&iru, rxq, 0 /* flid */);
832 	map = fl->ifl_sds.ifsd_map;
833 	refill_pidx = netmap_idx_k2n(kring, nm_i);
834 	/*
835 	 * IMPORTANT: we must leave one free slot in the ring,
836 	 * so move head back by one unit
837 	 */
838 	head = nm_prev(head, lim);
839 	nic_i = UINT_MAX;
840 	DBG_COUNTER_INC(fl_refills);
841 	while (nm_i != head) {
842 #if IFLIB_DEBUG_COUNTERS
843 		if (++rf_count == 9)
844 			DBG_COUNTER_INC(fl_refills_large);
845 #endif
846 		for (int tmp_pidx = 0; tmp_pidx < IFLIB_MAX_RX_REFRESH && nm_i != head; tmp_pidx++) {
847 			struct netmap_slot *slot = &ring->slot[nm_i];
848 			void *addr = PNMB(na, slot, &fl->ifl_bus_addrs[tmp_pidx]);
849 			uint32_t nic_i_dma = refill_pidx;
850 			nic_i = netmap_idx_k2n(kring, nm_i);
851 
852 			MPASS(tmp_pidx < IFLIB_MAX_RX_REFRESH);
853 
854 			if (addr == NETMAP_BUF_BASE(na)) /* bad buf */
855 			        return netmap_ring_reinit(kring);
856 
857 			fl->ifl_vm_addrs[tmp_pidx] = addr;
858 			if (__predict_false(init)) {
859 				netmap_load_map(na, fl->ifl_buf_tag,
860 				    map[nic_i], addr);
861 			} else if (slot->flags & NS_BUF_CHANGED) {
862 				/* buffer has changed, reload map */
863 				netmap_reload_map(na, fl->ifl_buf_tag,
864 				    map[nic_i], addr);
865 			}
866 			slot->flags &= ~NS_BUF_CHANGED;
867 
868 			nm_i = nm_next(nm_i, lim);
869 			fl->ifl_rxd_idxs[tmp_pidx] = nic_i = nm_next(nic_i, lim);
870 			if (nm_i != head && tmp_pidx < IFLIB_MAX_RX_REFRESH-1)
871 				continue;
872 
873 			iru.iru_pidx = refill_pidx;
874 			iru.iru_count = tmp_pidx+1;
875 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
876 			refill_pidx = nic_i;
877 			for (int n = 0; n < iru.iru_count; n++) {
878 				bus_dmamap_sync(fl->ifl_buf_tag, map[nic_i_dma],
879 						BUS_DMASYNC_PREREAD);
880 				/* XXX - change this to not use the netmap func*/
881 				nic_i_dma = nm_next(nic_i_dma, lim);
882 			}
883 		}
884 	}
885 	kring->nr_hwcur = head;
886 
887 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
888 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
889 	if (__predict_true(nic_i != UINT_MAX)) {
890 		ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id, nic_i);
891 		DBG_COUNTER_INC(rxd_flush);
892 	}
893 	return (0);
894 }
895 
896 /*
897  * Reconcile kernel and user view of the transmit ring.
898  *
899  * All information is in the kring.
900  * Userspace wants to send packets up to the one before kring->rhead,
901  * kernel knows kring->nr_hwcur is the first unsent packet.
902  *
903  * Here we push packets out (as many as possible), and possibly
904  * reclaim buffers from previously completed transmission.
905  *
906  * The caller (netmap) guarantees that there is only one instance
907  * running at any time. Any interference with other driver
908  * methods should be handled by the individual drivers.
909  */
910 static int
911 iflib_netmap_txsync(struct netmap_kring *kring, int flags)
912 {
913 	struct netmap_adapter *na = kring->na;
914 	if_t ifp = na->ifp;
915 	struct netmap_ring *ring = kring->ring;
916 	u_int nm_i;	/* index into the netmap kring */
917 	u_int nic_i;	/* index into the NIC ring */
918 	u_int n;
919 	u_int const lim = kring->nkr_num_slots - 1;
920 	u_int const head = kring->rhead;
921 	struct if_pkt_info pi;
922 
923 	/*
924 	 * interrupts on every tx packet are expensive so request
925 	 * them every half ring, or where NS_REPORT is set
926 	 */
927 	u_int report_frequency = kring->nkr_num_slots >> 1;
928 	/* device-specific */
929 	if_ctx_t ctx = ifp->if_softc;
930 	iflib_txq_t txq = &ctx->ifc_txqs[kring->ring_id];
931 
932 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
933 	    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
934 
935 	/*
936 	 * First part: process new packets to send.
937 	 * nm_i is the current index in the netmap kring,
938 	 * nic_i is the corresponding index in the NIC ring.
939 	 *
940 	 * If we have packets to send (nm_i != head)
941 	 * iterate over the netmap ring, fetch length and update
942 	 * the corresponding slot in the NIC ring. Some drivers also
943 	 * need to update the buffer's physical address in the NIC slot
944 	 * even NS_BUF_CHANGED is not set (PNMB computes the addresses).
945 	 *
946 	 * The netmap_reload_map() calls is especially expensive,
947 	 * even when (as in this case) the tag is 0, so do only
948 	 * when the buffer has actually changed.
949 	 *
950 	 * If possible do not set the report/intr bit on all slots,
951 	 * but only a few times per ring or when NS_REPORT is set.
952 	 *
953 	 * Finally, on 10G and faster drivers, it might be useful
954 	 * to prefetch the next slot and txr entry.
955 	 */
956 
957 	nm_i = kring->nr_hwcur;
958 	if (nm_i != head) {	/* we have new packets to send */
959 		pkt_info_zero(&pi);
960 		pi.ipi_segs = txq->ift_segs;
961 		pi.ipi_qsidx = kring->ring_id;
962 		nic_i = netmap_idx_k2n(kring, nm_i);
963 
964 		__builtin_prefetch(&ring->slot[nm_i]);
965 		__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i]);
966 		__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i]);
967 
968 		for (n = 0; nm_i != head; n++) {
969 			struct netmap_slot *slot = &ring->slot[nm_i];
970 			u_int len = slot->len;
971 			uint64_t paddr;
972 			void *addr = PNMB(na, slot, &paddr);
973 			int flags = (slot->flags & NS_REPORT ||
974 				nic_i == 0 || nic_i == report_frequency) ?
975 				IPI_TX_INTR : 0;
976 
977 			/* device-specific */
978 			pi.ipi_len = len;
979 			pi.ipi_segs[0].ds_addr = paddr;
980 			pi.ipi_segs[0].ds_len = len;
981 			pi.ipi_nsegs = 1;
982 			pi.ipi_ndescs = 0;
983 			pi.ipi_pidx = nic_i;
984 			pi.ipi_flags = flags;
985 
986 			/* Fill the slot in the NIC ring. */
987 			ctx->isc_txd_encap(ctx->ifc_softc, &pi);
988 			DBG_COUNTER_INC(tx_encap);
989 
990 			/* prefetch for next round */
991 			__builtin_prefetch(&ring->slot[nm_i + 1]);
992 			__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i + 1]);
993 			__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i + 1]);
994 
995 			NM_CHECK_ADDR_LEN(na, addr, len);
996 
997 			if (slot->flags & NS_BUF_CHANGED) {
998 				/* buffer has changed, reload map */
999 				netmap_reload_map(na, txq->ift_buf_tag,
1000 				    txq->ift_sds.ifsd_map[nic_i], addr);
1001 			}
1002 			/* make sure changes to the buffer are synced */
1003 			bus_dmamap_sync(txq->ift_buf_tag,
1004 			    txq->ift_sds.ifsd_map[nic_i],
1005 			    BUS_DMASYNC_PREWRITE);
1006 
1007 			slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED);
1008 			nm_i = nm_next(nm_i, lim);
1009 			nic_i = nm_next(nic_i, lim);
1010 		}
1011 		kring->nr_hwcur = nm_i;
1012 
1013 		/* synchronize the NIC ring */
1014 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1015 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1016 
1017 		/* (re)start the tx unit up to slot nic_i (excluded) */
1018 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, nic_i);
1019 	}
1020 
1021 	/*
1022 	 * Second part: reclaim buffers for completed transmissions.
1023 	 *
1024 	 * If there are unclaimed buffers, attempt to reclaim them.
1025 	 * If none are reclaimed, and TX IRQs are not in use, do an initial
1026 	 * minimal delay, then trigger the tx handler which will spin in the
1027 	 * group task queue.
1028 	 */
1029 	if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1030 		if (iflib_tx_credits_update(ctx, txq)) {
1031 			/* some tx completed, increment avail */
1032 			nic_i = txq->ift_cidx_processed;
1033 			kring->nr_hwtail = nm_prev(netmap_idx_n2k(kring, nic_i), lim);
1034 		}
1035 	}
1036 	if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ))
1037 		if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1038 			callout_reset_on(&txq->ift_timer, hz < 2000 ? 1 : hz / 1000,
1039 			    iflib_timer, txq, txq->ift_timer.c_cpu);
1040 	}
1041 	return (0);
1042 }
1043 
1044 /*
1045  * Reconcile kernel and user view of the receive ring.
1046  * Same as for the txsync, this routine must be efficient.
1047  * The caller guarantees a single invocations, but races against
1048  * the rest of the driver should be handled here.
1049  *
1050  * On call, kring->rhead is the first packet that userspace wants
1051  * to keep, and kring->rcur is the wakeup point.
1052  * The kernel has previously reported packets up to kring->rtail.
1053  *
1054  * If (flags & NAF_FORCE_READ) also check for incoming packets irrespective
1055  * of whether or not we received an interrupt.
1056  */
1057 static int
1058 iflib_netmap_rxsync(struct netmap_kring *kring, int flags)
1059 {
1060 	struct netmap_adapter *na = kring->na;
1061 	struct netmap_ring *ring = kring->ring;
1062 	if_t ifp = na->ifp;
1063 	iflib_fl_t fl;
1064 	uint32_t nm_i;	/* index into the netmap ring */
1065 	uint32_t nic_i;	/* index into the NIC ring */
1066 	u_int i, n;
1067 	u_int const lim = kring->nkr_num_slots - 1;
1068 	u_int const head = kring->rhead;
1069 	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
1070 	struct if_rxd_info ri;
1071 
1072 	if_ctx_t ctx = ifp->if_softc;
1073 	iflib_rxq_t rxq = &ctx->ifc_rxqs[kring->ring_id];
1074 	if (head > lim)
1075 		return netmap_ring_reinit(kring);
1076 
1077 	/*
1078 	 * XXX netmap_fl_refill() only ever (re)fills free list 0 so far.
1079 	 */
1080 
1081 	for (i = 0, fl = rxq->ifr_fl; i < rxq->ifr_nfl; i++, fl++) {
1082 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
1083 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1084 	}
1085 
1086 	/*
1087 	 * First part: import newly received packets.
1088 	 *
1089 	 * nm_i is the index of the next free slot in the netmap ring,
1090 	 * nic_i is the index of the next received packet in the NIC ring,
1091 	 * and they may differ in case if_init() has been called while
1092 	 * in netmap mode. For the receive ring we have
1093 	 *
1094 	 *	nic_i = rxr->next_check;
1095 	 *	nm_i = kring->nr_hwtail (previous)
1096 	 * and
1097 	 *	nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1098 	 *
1099 	 * rxr->next_check is set to 0 on a ring reinit
1100 	 */
1101 	if (netmap_no_pendintr || force_update) {
1102 		uint32_t hwtail_lim = nm_prev(kring->nr_hwcur, lim);
1103 		int crclen = iflib_crcstrip ? 0 : 4;
1104 		int error, avail;
1105 
1106 		for (i = 0; i < rxq->ifr_nfl; i++) {
1107 			fl = &rxq->ifr_fl[i];
1108 			nic_i = fl->ifl_cidx;
1109 			nm_i = netmap_idx_n2k(kring, nic_i);
1110 			avail = ctx->isc_rxd_available(ctx->ifc_softc,
1111 			    rxq->ifr_id, nic_i, USHRT_MAX);
1112 			for (n = 0; avail > 0 && nm_i != hwtail_lim; n++, avail--) {
1113 				rxd_info_zero(&ri);
1114 				ri.iri_frags = rxq->ifr_frags;
1115 				ri.iri_qsidx = kring->ring_id;
1116 				ri.iri_ifp = ctx->ifc_ifp;
1117 				ri.iri_cidx = nic_i;
1118 
1119 				error = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
1120 				ring->slot[nm_i].len = error ? 0 : ri.iri_len - crclen;
1121 				ring->slot[nm_i].flags = 0;
1122 				bus_dmamap_sync(fl->ifl_buf_tag,
1123 				    fl->ifl_sds.ifsd_map[nic_i], BUS_DMASYNC_POSTREAD);
1124 				nm_i = nm_next(nm_i, lim);
1125 				nic_i = nm_next(nic_i, lim);
1126 			}
1127 			if (n) { /* update the state variables */
1128 				if (netmap_no_pendintr && !force_update) {
1129 					/* diagnostics */
1130 					iflib_rx_miss ++;
1131 					iflib_rx_miss_bufs += n;
1132 				}
1133 				fl->ifl_cidx = nic_i;
1134 				kring->nr_hwtail = nm_i;
1135 			}
1136 			kring->nr_kflags &= ~NKR_PENDINTR;
1137 		}
1138 	}
1139 	/*
1140 	 * Second part: skip past packets that userspace has released.
1141 	 * (kring->nr_hwcur to head excluded),
1142 	 * and make the buffers available for reception.
1143 	 * As usual nm_i is the index in the netmap ring,
1144 	 * nic_i is the index in the NIC ring, and
1145 	 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1146 	 */
1147 	/* XXX not sure how this will work with multiple free lists */
1148 	nm_i = kring->nr_hwcur;
1149 
1150 	return (netmap_fl_refill(rxq, kring, nm_i, false));
1151 }
1152 
1153 static void
1154 iflib_netmap_intr(struct netmap_adapter *na, int onoff)
1155 {
1156 	if_ctx_t ctx = na->ifp->if_softc;
1157 
1158 	CTX_LOCK(ctx);
1159 	if (onoff) {
1160 		IFDI_INTR_ENABLE(ctx);
1161 	} else {
1162 		IFDI_INTR_DISABLE(ctx);
1163 	}
1164 	CTX_UNLOCK(ctx);
1165 }
1166 
1167 
1168 static int
1169 iflib_netmap_attach(if_ctx_t ctx)
1170 {
1171 	struct netmap_adapter na;
1172 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1173 
1174 	bzero(&na, sizeof(na));
1175 
1176 	na.ifp = ctx->ifc_ifp;
1177 	na.na_flags = NAF_BDG_MAYSLEEP;
1178 	MPASS(ctx->ifc_softc_ctx.isc_ntxqsets);
1179 	MPASS(ctx->ifc_softc_ctx.isc_nrxqsets);
1180 
1181 	na.num_tx_desc = scctx->isc_ntxd[0];
1182 	na.num_rx_desc = scctx->isc_nrxd[0];
1183 	na.nm_txsync = iflib_netmap_txsync;
1184 	na.nm_rxsync = iflib_netmap_rxsync;
1185 	na.nm_register = iflib_netmap_register;
1186 	na.nm_intr = iflib_netmap_intr;
1187 	na.num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
1188 	na.num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
1189 	return (netmap_attach(&na));
1190 }
1191 
1192 static void
1193 iflib_netmap_txq_init(if_ctx_t ctx, iflib_txq_t txq)
1194 {
1195 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1196 	struct netmap_slot *slot;
1197 
1198 	slot = netmap_reset(na, NR_TX, txq->ift_id, 0);
1199 	if (slot == NULL)
1200 		return;
1201 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxd[0]; i++) {
1202 
1203 		/*
1204 		 * In netmap mode, set the map for the packet buffer.
1205 		 * NOTE: Some drivers (not this one) also need to set
1206 		 * the physical buffer address in the NIC ring.
1207 		 * netmap_idx_n2k() maps a nic index, i, into the corresponding
1208 		 * netmap slot index, si
1209 		 */
1210 		int si = netmap_idx_n2k(na->tx_rings[txq->ift_id], i);
1211 		netmap_load_map(na, txq->ift_buf_tag, txq->ift_sds.ifsd_map[i],
1212 		    NMB(na, slot + si));
1213 	}
1214 }
1215 
1216 static void
1217 iflib_netmap_rxq_init(if_ctx_t ctx, iflib_rxq_t rxq)
1218 {
1219 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1220 	struct netmap_kring *kring = na->rx_rings[rxq->ifr_id];
1221 	struct netmap_slot *slot;
1222 	uint32_t nm_i;
1223 
1224 	slot = netmap_reset(na, NR_RX, rxq->ifr_id, 0);
1225 	if (slot == NULL)
1226 		return;
1227 	nm_i = netmap_idx_n2k(kring, 0);
1228 	netmap_fl_refill(rxq, kring, nm_i, true);
1229 }
1230 
1231 static void
1232 iflib_netmap_timer_adjust(if_ctx_t ctx, iflib_txq_t txq, uint32_t *reset_on)
1233 {
1234 	struct netmap_kring *kring;
1235 	uint16_t txqid;
1236 
1237 	txqid = txq->ift_id;
1238 	kring = NA(ctx->ifc_ifp)->tx_rings[txqid];
1239 
1240 	if (kring->nr_hwcur != nm_next(kring->nr_hwtail, kring->nkr_num_slots - 1)) {
1241 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1242 		    BUS_DMASYNC_POSTREAD);
1243 		if (ctx->isc_txd_credits_update(ctx->ifc_softc, txqid, false))
1244 			netmap_tx_irq(ctx->ifc_ifp, txqid);
1245 		if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ)) {
1246 			if (hz < 2000)
1247 				*reset_on = 1;
1248 			else
1249 				*reset_on = hz / 1000;
1250 		}
1251 	}
1252 }
1253 
1254 #define iflib_netmap_detach(ifp) netmap_detach(ifp)
1255 
1256 #else
1257 #define iflib_netmap_txq_init(ctx, txq)
1258 #define iflib_netmap_rxq_init(ctx, rxq)
1259 #define iflib_netmap_detach(ifp)
1260 
1261 #define iflib_netmap_attach(ctx) (0)
1262 #define netmap_rx_irq(ifp, qid, budget) (0)
1263 #define netmap_tx_irq(ifp, qid) do {} while (0)
1264 #define iflib_netmap_timer_adjust(ctx, txq, reset_on)
1265 #endif
1266 
1267 #if defined(__i386__) || defined(__amd64__)
1268 static __inline void
1269 prefetch(void *x)
1270 {
1271 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1272 }
1273 static __inline void
1274 prefetch2cachelines(void *x)
1275 {
1276 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1277 #if (CACHE_LINE_SIZE < 128)
1278 	__asm volatile("prefetcht0 %0" :: "m" (*(((unsigned long *)x)+CACHE_LINE_SIZE/(sizeof(unsigned long)))));
1279 #endif
1280 }
1281 #else
1282 #define prefetch(x)
1283 #define prefetch2cachelines(x)
1284 #endif
1285 
1286 static void
1287 iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid)
1288 {
1289 	iflib_fl_t fl;
1290 
1291 	fl = &rxq->ifr_fl[flid];
1292 	iru->iru_paddrs = fl->ifl_bus_addrs;
1293 	iru->iru_vaddrs = &fl->ifl_vm_addrs[0];
1294 	iru->iru_idxs = fl->ifl_rxd_idxs;
1295 	iru->iru_qsidx = rxq->ifr_id;
1296 	iru->iru_buf_size = fl->ifl_buf_size;
1297 	iru->iru_flidx = fl->ifl_id;
1298 }
1299 
1300 static void
1301 _iflib_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int err)
1302 {
1303 	if (err)
1304 		return;
1305 	*(bus_addr_t *) arg = segs[0].ds_addr;
1306 }
1307 
1308 int
1309 iflib_dma_alloc_align(if_ctx_t ctx, int size, int align, iflib_dma_info_t dma, int mapflags)
1310 {
1311 	int err;
1312 	device_t dev = ctx->ifc_dev;
1313 
1314 	err = bus_dma_tag_create(bus_get_dma_tag(dev),	/* parent */
1315 				align, 0,		/* alignment, bounds */
1316 				BUS_SPACE_MAXADDR,	/* lowaddr */
1317 				BUS_SPACE_MAXADDR,	/* highaddr */
1318 				NULL, NULL,		/* filter, filterarg */
1319 				size,			/* maxsize */
1320 				1,			/* nsegments */
1321 				size,			/* maxsegsize */
1322 				BUS_DMA_ALLOCNOW,	/* flags */
1323 				NULL,			/* lockfunc */
1324 				NULL,			/* lockarg */
1325 				&dma->idi_tag);
1326 	if (err) {
1327 		device_printf(dev,
1328 		    "%s: bus_dma_tag_create failed: %d\n",
1329 		    __func__, err);
1330 		goto fail_0;
1331 	}
1332 
1333 	err = bus_dmamem_alloc(dma->idi_tag, (void**) &dma->idi_vaddr,
1334 	    BUS_DMA_NOWAIT | BUS_DMA_COHERENT | BUS_DMA_ZERO, &dma->idi_map);
1335 	if (err) {
1336 		device_printf(dev,
1337 		    "%s: bus_dmamem_alloc(%ju) failed: %d\n",
1338 		    __func__, (uintmax_t)size, err);
1339 		goto fail_1;
1340 	}
1341 
1342 	dma->idi_paddr = IF_BAD_DMA;
1343 	err = bus_dmamap_load(dma->idi_tag, dma->idi_map, dma->idi_vaddr,
1344 	    size, _iflib_dmamap_cb, &dma->idi_paddr, mapflags | BUS_DMA_NOWAIT);
1345 	if (err || dma->idi_paddr == IF_BAD_DMA) {
1346 		device_printf(dev,
1347 		    "%s: bus_dmamap_load failed: %d\n",
1348 		    __func__, err);
1349 		goto fail_2;
1350 	}
1351 
1352 	dma->idi_size = size;
1353 	return (0);
1354 
1355 fail_2:
1356 	bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1357 fail_1:
1358 	bus_dma_tag_destroy(dma->idi_tag);
1359 fail_0:
1360 	dma->idi_tag = NULL;
1361 
1362 	return (err);
1363 }
1364 
1365 int
1366 iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags)
1367 {
1368 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1369 
1370 	KASSERT(sctx->isc_q_align != 0, ("alignment value not initialized"));
1371 
1372 	return (iflib_dma_alloc_align(ctx, size, sctx->isc_q_align, dma, mapflags));
1373 }
1374 
1375 int
1376 iflib_dma_alloc_multi(if_ctx_t ctx, int *sizes, iflib_dma_info_t *dmalist, int mapflags, int count)
1377 {
1378 	int i, err;
1379 	iflib_dma_info_t *dmaiter;
1380 
1381 	dmaiter = dmalist;
1382 	for (i = 0; i < count; i++, dmaiter++) {
1383 		if ((err = iflib_dma_alloc(ctx, sizes[i], *dmaiter, mapflags)) != 0)
1384 			break;
1385 	}
1386 	if (err)
1387 		iflib_dma_free_multi(dmalist, i);
1388 	return (err);
1389 }
1390 
1391 void
1392 iflib_dma_free(iflib_dma_info_t dma)
1393 {
1394 	if (dma->idi_tag == NULL)
1395 		return;
1396 	if (dma->idi_paddr != IF_BAD_DMA) {
1397 		bus_dmamap_sync(dma->idi_tag, dma->idi_map,
1398 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1399 		bus_dmamap_unload(dma->idi_tag, dma->idi_map);
1400 		dma->idi_paddr = IF_BAD_DMA;
1401 	}
1402 	if (dma->idi_vaddr != NULL) {
1403 		bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1404 		dma->idi_vaddr = NULL;
1405 	}
1406 	bus_dma_tag_destroy(dma->idi_tag);
1407 	dma->idi_tag = NULL;
1408 }
1409 
1410 void
1411 iflib_dma_free_multi(iflib_dma_info_t *dmalist, int count)
1412 {
1413 	int i;
1414 	iflib_dma_info_t *dmaiter = dmalist;
1415 
1416 	for (i = 0; i < count; i++, dmaiter++)
1417 		iflib_dma_free(*dmaiter);
1418 }
1419 
1420 static int
1421 iflib_fast_intr(void *arg)
1422 {
1423 	iflib_filter_info_t info = arg;
1424 	struct grouptask *gtask = info->ifi_task;
1425 	int result;
1426 
1427 	DBG_COUNTER_INC(fast_intrs);
1428 	if (info->ifi_filter != NULL) {
1429 		result = info->ifi_filter(info->ifi_filter_arg);
1430 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1431 			return (result);
1432 	}
1433 
1434 	GROUPTASK_ENQUEUE(gtask);
1435 	return (FILTER_HANDLED);
1436 }
1437 
1438 static int
1439 iflib_fast_intr_rxtx(void *arg)
1440 {
1441 	iflib_filter_info_t info = arg;
1442 	struct grouptask *gtask = info->ifi_task;
1443 	if_ctx_t ctx;
1444 	iflib_rxq_t rxq = (iflib_rxq_t)info->ifi_ctx;
1445 	iflib_txq_t txq;
1446 	void *sc;
1447 	int i, cidx, result;
1448 	qidx_t txqid;
1449 	bool intr_enable, intr_legacy;
1450 
1451 	DBG_COUNTER_INC(fast_intrs);
1452 	if (info->ifi_filter != NULL) {
1453 		result = info->ifi_filter(info->ifi_filter_arg);
1454 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1455 			return (result);
1456 	}
1457 
1458 	ctx = rxq->ifr_ctx;
1459 	sc = ctx->ifc_softc;
1460 	intr_enable = false;
1461 	intr_legacy = !!(ctx->ifc_flags & IFC_LEGACY);
1462 	MPASS(rxq->ifr_ntxqirq);
1463 	for (i = 0; i < rxq->ifr_ntxqirq; i++) {
1464 		txqid = rxq->ifr_txqid[i];
1465 		txq = &ctx->ifc_txqs[txqid];
1466 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1467 		    BUS_DMASYNC_POSTREAD);
1468 		if (!ctx->isc_txd_credits_update(sc, txqid, false)) {
1469 			if (intr_legacy)
1470 				intr_enable = true;
1471 			else
1472 				IFDI_TX_QUEUE_INTR_ENABLE(ctx, txqid);
1473 			continue;
1474 		}
1475 		GROUPTASK_ENQUEUE(&txq->ift_task);
1476 	}
1477 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_RXCQ)
1478 		cidx = rxq->ifr_cq_cidx;
1479 	else
1480 		cidx = rxq->ifr_fl[0].ifl_cidx;
1481 	if (iflib_rxd_avail(ctx, rxq, cidx, 1))
1482 		GROUPTASK_ENQUEUE(gtask);
1483 	else {
1484 		if (intr_legacy)
1485 			intr_enable = true;
1486 		else
1487 			IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
1488 		DBG_COUNTER_INC(rx_intr_enables);
1489 	}
1490 	if (intr_enable)
1491 		IFDI_INTR_ENABLE(ctx);
1492 	return (FILTER_HANDLED);
1493 }
1494 
1495 
1496 static int
1497 iflib_fast_intr_ctx(void *arg)
1498 {
1499 	iflib_filter_info_t info = arg;
1500 	struct grouptask *gtask = info->ifi_task;
1501 	int result;
1502 
1503 	DBG_COUNTER_INC(fast_intrs);
1504 	if (info->ifi_filter != NULL) {
1505 		result = info->ifi_filter(info->ifi_filter_arg);
1506 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1507 			return (result);
1508 	}
1509 
1510 	GROUPTASK_ENQUEUE(gtask);
1511 	return (FILTER_HANDLED);
1512 }
1513 
1514 static int
1515 _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
1516 		 driver_filter_t filter, driver_intr_t handler, void *arg,
1517 		 const char *name)
1518 {
1519 	struct resource *res;
1520 	void *tag = NULL;
1521 	device_t dev = ctx->ifc_dev;
1522 	int flags, i, rc;
1523 
1524 	flags = RF_ACTIVE;
1525 	if (ctx->ifc_flags & IFC_LEGACY)
1526 		flags |= RF_SHAREABLE;
1527 	MPASS(rid < 512);
1528 	i = rid;
1529 	res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &i, flags);
1530 	if (res == NULL) {
1531 		device_printf(dev,
1532 		    "failed to allocate IRQ for rid %d, name %s.\n", rid, name);
1533 		return (ENOMEM);
1534 	}
1535 	irq->ii_res = res;
1536 	KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL"));
1537 	rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET,
1538 						filter, handler, arg, &tag);
1539 	if (rc != 0) {
1540 		device_printf(dev,
1541 		    "failed to setup interrupt for rid %d, name %s: %d\n",
1542 					  rid, name ? name : "unknown", rc);
1543 		return (rc);
1544 	} else if (name)
1545 		bus_describe_intr(dev, res, tag, "%s", name);
1546 
1547 	irq->ii_tag = tag;
1548 	return (0);
1549 }
1550 
1551 /*********************************************************************
1552  *
1553  *  Allocate DMA resources for TX buffers as well as memory for the TX
1554  *  mbuf map.  TX DMA maps (non-TSO/TSO) and TX mbuf map are kept in a
1555  *  iflib_sw_tx_desc_array structure, storing all the information that
1556  *  is needed to transmit a packet on the wire.  This is called only
1557  *  once at attach, setup is done every reset.
1558  *
1559  **********************************************************************/
1560 static int
1561 iflib_txsd_alloc(iflib_txq_t txq)
1562 {
1563 	if_ctx_t ctx = txq->ift_ctx;
1564 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1565 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1566 	device_t dev = ctx->ifc_dev;
1567 	bus_size_t tsomaxsize;
1568 	int err, nsegments, ntsosegments;
1569 	bool tso;
1570 
1571 	nsegments = scctx->isc_tx_nsegments;
1572 	ntsosegments = scctx->isc_tx_tso_segments_max;
1573 	tsomaxsize = scctx->isc_tx_tso_size_max;
1574 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_VLAN_MTU)
1575 		tsomaxsize += sizeof(struct ether_vlan_header);
1576 	MPASS(scctx->isc_ntxd[0] > 0);
1577 	MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0);
1578 	MPASS(nsegments > 0);
1579 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) {
1580 		MPASS(ntsosegments > 0);
1581 		MPASS(sctx->isc_tso_maxsize >= tsomaxsize);
1582 	}
1583 
1584 	/*
1585 	 * Set up DMA tags for TX buffers.
1586 	 */
1587 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1588 			       1, 0,			/* alignment, bounds */
1589 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1590 			       BUS_SPACE_MAXADDR,	/* highaddr */
1591 			       NULL, NULL,		/* filter, filterarg */
1592 			       sctx->isc_tx_maxsize,		/* maxsize */
1593 			       nsegments,	/* nsegments */
1594 			       sctx->isc_tx_maxsegsize,	/* maxsegsize */
1595 			       0,			/* flags */
1596 			       NULL,			/* lockfunc */
1597 			       NULL,			/* lockfuncarg */
1598 			       &txq->ift_buf_tag))) {
1599 		device_printf(dev,"Unable to allocate TX DMA tag: %d\n", err);
1600 		device_printf(dev,"maxsize: %ju nsegments: %d maxsegsize: %ju\n",
1601 		    (uintmax_t)sctx->isc_tx_maxsize, nsegments, (uintmax_t)sctx->isc_tx_maxsegsize);
1602 		goto fail;
1603 	}
1604 	tso = (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) != 0;
1605 	if (tso && (err = bus_dma_tag_create(bus_get_dma_tag(dev),
1606 			       1, 0,			/* alignment, bounds */
1607 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1608 			       BUS_SPACE_MAXADDR,	/* highaddr */
1609 			       NULL, NULL,		/* filter, filterarg */
1610 			       tsomaxsize,		/* maxsize */
1611 			       ntsosegments,	/* nsegments */
1612 			       sctx->isc_tso_maxsegsize,/* maxsegsize */
1613 			       0,			/* flags */
1614 			       NULL,			/* lockfunc */
1615 			       NULL,			/* lockfuncarg */
1616 			       &txq->ift_tso_buf_tag))) {
1617 		device_printf(dev, "Unable to allocate TSO TX DMA tag: %d\n",
1618 		    err);
1619 		goto fail;
1620 	}
1621 
1622 	/* Allocate memory for the TX mbuf map. */
1623 	if (!(txq->ift_sds.ifsd_m =
1624 	    (struct mbuf **) malloc(sizeof(struct mbuf *) *
1625 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1626 		device_printf(dev, "Unable to allocate TX mbuf map memory\n");
1627 		err = ENOMEM;
1628 		goto fail;
1629 	}
1630 
1631 	/*
1632 	 * Create the DMA maps for TX buffers.
1633 	 */
1634 	if ((txq->ift_sds.ifsd_map = (bus_dmamap_t *)malloc(
1635 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1636 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1637 		device_printf(dev,
1638 		    "Unable to allocate TX buffer DMA map memory\n");
1639 		err = ENOMEM;
1640 		goto fail;
1641 	}
1642 	if (tso && (txq->ift_sds.ifsd_tso_map = (bus_dmamap_t *)malloc(
1643 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1644 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1645 		device_printf(dev,
1646 		    "Unable to allocate TSO TX buffer map memory\n");
1647 		err = ENOMEM;
1648 		goto fail;
1649 	}
1650 	for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) {
1651 		err = bus_dmamap_create(txq->ift_buf_tag, 0,
1652 		    &txq->ift_sds.ifsd_map[i]);
1653 		if (err != 0) {
1654 			device_printf(dev, "Unable to create TX DMA map\n");
1655 			goto fail;
1656 		}
1657 		if (!tso)
1658 			continue;
1659 		err = bus_dmamap_create(txq->ift_tso_buf_tag, 0,
1660 		    &txq->ift_sds.ifsd_tso_map[i]);
1661 		if (err != 0) {
1662 			device_printf(dev, "Unable to create TSO TX DMA map\n");
1663 			goto fail;
1664 		}
1665 	}
1666 	return (0);
1667 fail:
1668 	/* We free all, it handles case where we are in the middle */
1669 	iflib_tx_structures_free(ctx);
1670 	return (err);
1671 }
1672 
1673 static void
1674 iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i)
1675 {
1676 	bus_dmamap_t map;
1677 
1678 	if (txq->ift_sds.ifsd_map != NULL) {
1679 		map = txq->ift_sds.ifsd_map[i];
1680 		bus_dmamap_sync(txq->ift_buf_tag, map, BUS_DMASYNC_POSTWRITE);
1681 		bus_dmamap_unload(txq->ift_buf_tag, map);
1682 		bus_dmamap_destroy(txq->ift_buf_tag, map);
1683 		txq->ift_sds.ifsd_map[i] = NULL;
1684 	}
1685 
1686 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1687 		map = txq->ift_sds.ifsd_tso_map[i];
1688 		bus_dmamap_sync(txq->ift_tso_buf_tag, map,
1689 		    BUS_DMASYNC_POSTWRITE);
1690 		bus_dmamap_unload(txq->ift_tso_buf_tag, map);
1691 		bus_dmamap_destroy(txq->ift_tso_buf_tag, map);
1692 		txq->ift_sds.ifsd_tso_map[i] = NULL;
1693 	}
1694 }
1695 
1696 static void
1697 iflib_txq_destroy(iflib_txq_t txq)
1698 {
1699 	if_ctx_t ctx = txq->ift_ctx;
1700 
1701 	for (int i = 0; i < txq->ift_size; i++)
1702 		iflib_txsd_destroy(ctx, txq, i);
1703 
1704 	if (txq->ift_br != NULL) {
1705 		ifmp_ring_free(txq->ift_br);
1706 		txq->ift_br = NULL;
1707 	}
1708 
1709 	mtx_destroy(&txq->ift_mtx);
1710 
1711 	if (txq->ift_sds.ifsd_map != NULL) {
1712 		free(txq->ift_sds.ifsd_map, M_IFLIB);
1713 		txq->ift_sds.ifsd_map = NULL;
1714 	}
1715 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1716 		free(txq->ift_sds.ifsd_tso_map, M_IFLIB);
1717 		txq->ift_sds.ifsd_tso_map = NULL;
1718 	}
1719 	if (txq->ift_sds.ifsd_m != NULL) {
1720 		free(txq->ift_sds.ifsd_m, M_IFLIB);
1721 		txq->ift_sds.ifsd_m = NULL;
1722 	}
1723 	if (txq->ift_buf_tag != NULL) {
1724 		bus_dma_tag_destroy(txq->ift_buf_tag);
1725 		txq->ift_buf_tag = NULL;
1726 	}
1727 	if (txq->ift_tso_buf_tag != NULL) {
1728 		bus_dma_tag_destroy(txq->ift_tso_buf_tag);
1729 		txq->ift_tso_buf_tag = NULL;
1730 	}
1731 	if (txq->ift_ifdi != NULL) {
1732 		free(txq->ift_ifdi, M_IFLIB);
1733 	}
1734 }
1735 
1736 static void
1737 iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i)
1738 {
1739 	struct mbuf **mp;
1740 
1741 	mp = &txq->ift_sds.ifsd_m[i];
1742 	if (*mp == NULL)
1743 		return;
1744 
1745 	if (txq->ift_sds.ifsd_map != NULL) {
1746 		bus_dmamap_sync(txq->ift_buf_tag,
1747 		    txq->ift_sds.ifsd_map[i], BUS_DMASYNC_POSTWRITE);
1748 		bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[i]);
1749 	}
1750 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1751 		bus_dmamap_sync(txq->ift_tso_buf_tag,
1752 		    txq->ift_sds.ifsd_tso_map[i], BUS_DMASYNC_POSTWRITE);
1753 		bus_dmamap_unload(txq->ift_tso_buf_tag,
1754 		    txq->ift_sds.ifsd_tso_map[i]);
1755 	}
1756 	m_free(*mp);
1757 	DBG_COUNTER_INC(tx_frees);
1758 	*mp = NULL;
1759 }
1760 
1761 static int
1762 iflib_txq_setup(iflib_txq_t txq)
1763 {
1764 	if_ctx_t ctx = txq->ift_ctx;
1765 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1766 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1767 	iflib_dma_info_t di;
1768 	int i;
1769 
1770 	/* Set number of descriptors available */
1771 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
1772 	/* XXX make configurable */
1773 	txq->ift_update_freq = IFLIB_DEFAULT_TX_UPDATE_FREQ;
1774 
1775 	/* Reset indices */
1776 	txq->ift_cidx_processed = 0;
1777 	txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0;
1778 	txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset];
1779 
1780 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1781 		bzero((void *)di->idi_vaddr, di->idi_size);
1782 
1783 	IFDI_TXQ_SETUP(ctx, txq->ift_id);
1784 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1785 		bus_dmamap_sync(di->idi_tag, di->idi_map,
1786 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1787 	return (0);
1788 }
1789 
1790 /*********************************************************************
1791  *
1792  *  Allocate DMA resources for RX buffers as well as memory for the RX
1793  *  mbuf map, direct RX cluster pointer map and RX cluster bus address
1794  *  map.  RX DMA map, RX mbuf map, direct RX cluster pointer map and
1795  *  RX cluster map are kept in a iflib_sw_rx_desc_array structure.
1796  *  Since we use use one entry in iflib_sw_rx_desc_array per received
1797  *  packet, the maximum number of entries we'll need is equal to the
1798  *  number of hardware receive descriptors that we've allocated.
1799  *
1800  **********************************************************************/
1801 static int
1802 iflib_rxsd_alloc(iflib_rxq_t rxq)
1803 {
1804 	if_ctx_t ctx = rxq->ifr_ctx;
1805 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1806 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1807 	device_t dev = ctx->ifc_dev;
1808 	iflib_fl_t fl;
1809 	int			err;
1810 
1811 	MPASS(scctx->isc_nrxd[0] > 0);
1812 	MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0);
1813 
1814 	fl = rxq->ifr_fl;
1815 	for (int i = 0; i <  rxq->ifr_nfl; i++, fl++) {
1816 		fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */
1817 		/* Set up DMA tag for RX buffers. */
1818 		err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1819 					 1, 0,			/* alignment, bounds */
1820 					 BUS_SPACE_MAXADDR,	/* lowaddr */
1821 					 BUS_SPACE_MAXADDR,	/* highaddr */
1822 					 NULL, NULL,		/* filter, filterarg */
1823 					 sctx->isc_rx_maxsize,	/* maxsize */
1824 					 sctx->isc_rx_nsegments,	/* nsegments */
1825 					 sctx->isc_rx_maxsegsize,	/* maxsegsize */
1826 					 0,			/* flags */
1827 					 NULL,			/* lockfunc */
1828 					 NULL,			/* lockarg */
1829 					 &fl->ifl_buf_tag);
1830 		if (err) {
1831 			device_printf(dev,
1832 			    "Unable to allocate RX DMA tag: %d\n", err);
1833 			goto fail;
1834 		}
1835 
1836 		/* Allocate memory for the RX mbuf map. */
1837 		if (!(fl->ifl_sds.ifsd_m =
1838 		      (struct mbuf **) malloc(sizeof(struct mbuf *) *
1839 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1840 			device_printf(dev,
1841 			    "Unable to allocate RX mbuf map memory\n");
1842 			err = ENOMEM;
1843 			goto fail;
1844 		}
1845 
1846 		/* Allocate memory for the direct RX cluster pointer map. */
1847 		if (!(fl->ifl_sds.ifsd_cl =
1848 		      (caddr_t *) malloc(sizeof(caddr_t) *
1849 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1850 			device_printf(dev,
1851 			    "Unable to allocate RX cluster map memory\n");
1852 			err = ENOMEM;
1853 			goto fail;
1854 		}
1855 
1856 		/* Allocate memory for the RX cluster bus address map. */
1857 		if (!(fl->ifl_sds.ifsd_ba =
1858 		      (bus_addr_t *) malloc(sizeof(bus_addr_t) *
1859 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1860 			device_printf(dev,
1861 			    "Unable to allocate RX bus address map memory\n");
1862 			err = ENOMEM;
1863 			goto fail;
1864 		}
1865 
1866 		/*
1867 		 * Create the DMA maps for RX buffers.
1868 		 */
1869 		if (!(fl->ifl_sds.ifsd_map =
1870 		      (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1871 			device_printf(dev,
1872 			    "Unable to allocate RX buffer DMA map memory\n");
1873 			err = ENOMEM;
1874 			goto fail;
1875 		}
1876 		for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++) {
1877 			err = bus_dmamap_create(fl->ifl_buf_tag, 0,
1878 			    &fl->ifl_sds.ifsd_map[i]);
1879 			if (err != 0) {
1880 				device_printf(dev, "Unable to create RX buffer DMA map\n");
1881 				goto fail;
1882 			}
1883 		}
1884 	}
1885 	return (0);
1886 
1887 fail:
1888 	iflib_rx_structures_free(ctx);
1889 	return (err);
1890 }
1891 
1892 
1893 /*
1894  * Internal service routines
1895  */
1896 
1897 struct rxq_refill_cb_arg {
1898 	int               error;
1899 	bus_dma_segment_t seg;
1900 	int               nseg;
1901 };
1902 
1903 static void
1904 _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
1905 {
1906 	struct rxq_refill_cb_arg *cb_arg = arg;
1907 
1908 	cb_arg->error = error;
1909 	cb_arg->seg = segs[0];
1910 	cb_arg->nseg = nseg;
1911 }
1912 
1913 /**
1914  * _iflib_fl_refill - refill an rxq free-buffer list
1915  * @ctx: the iflib context
1916  * @fl: the free list to refill
1917  * @count: the number of new buffers to allocate
1918  *
1919  * (Re)populate an rxq free-buffer list with up to @count new packet buffers.
1920  * The caller must assure that @count does not exceed the queue's capacity.
1921  */
1922 static uint8_t
1923 _iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
1924 {
1925 	struct if_rxd_update iru;
1926 	struct rxq_refill_cb_arg cb_arg;
1927 	struct mbuf *m;
1928 	caddr_t cl, *sd_cl;
1929 	struct mbuf **sd_m;
1930 	bus_dmamap_t *sd_map;
1931 	bus_addr_t bus_addr, *sd_ba;
1932 	int err, frag_idx, i, idx, n, pidx;
1933 	qidx_t credits;
1934 
1935 	sd_m = fl->ifl_sds.ifsd_m;
1936 	sd_map = fl->ifl_sds.ifsd_map;
1937 	sd_cl = fl->ifl_sds.ifsd_cl;
1938 	sd_ba = fl->ifl_sds.ifsd_ba;
1939 	pidx = fl->ifl_pidx;
1940 	idx = pidx;
1941 	frag_idx = fl->ifl_fragidx;
1942 	credits = fl->ifl_credits;
1943 
1944 	i = 0;
1945 	n = count;
1946 	MPASS(n > 0);
1947 	MPASS(credits + n <= fl->ifl_size);
1948 
1949 	if (pidx < fl->ifl_cidx)
1950 		MPASS(pidx + n <= fl->ifl_cidx);
1951 	if (pidx == fl->ifl_cidx && (credits < fl->ifl_size))
1952 		MPASS(fl->ifl_gen == 0);
1953 	if (pidx > fl->ifl_cidx)
1954 		MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx);
1955 
1956 	DBG_COUNTER_INC(fl_refills);
1957 	if (n > 8)
1958 		DBG_COUNTER_INC(fl_refills_large);
1959 	iru_init(&iru, fl->ifl_rxq, fl->ifl_id);
1960 	while (n--) {
1961 		/*
1962 		 * We allocate an uninitialized mbuf + cluster, mbuf is
1963 		 * initialized after rx.
1964 		 *
1965 		 * If the cluster is still set then we know a minimum sized packet was received
1966 		 */
1967 		bit_ffc_at(fl->ifl_rx_bitmap, frag_idx, fl->ifl_size,
1968 		    &frag_idx);
1969 		if (frag_idx < 0)
1970 			bit_ffc(fl->ifl_rx_bitmap, fl->ifl_size, &frag_idx);
1971 		MPASS(frag_idx >= 0);
1972 		if ((cl = sd_cl[frag_idx]) == NULL) {
1973 			if ((cl = m_cljget(NULL, M_NOWAIT, fl->ifl_buf_size)) == NULL)
1974 				break;
1975 
1976 			cb_arg.error = 0;
1977 			MPASS(sd_map != NULL);
1978 			err = bus_dmamap_load(fl->ifl_buf_tag, sd_map[frag_idx],
1979 			    cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg,
1980 			    BUS_DMA_NOWAIT);
1981 			if (err != 0 || cb_arg.error) {
1982 				/*
1983 				 * !zone_pack ?
1984 				 */
1985 				if (fl->ifl_zone == zone_pack)
1986 					uma_zfree(fl->ifl_zone, cl);
1987 				break;
1988 			}
1989 
1990 			sd_ba[frag_idx] =  bus_addr = cb_arg.seg.ds_addr;
1991 			sd_cl[frag_idx] = cl;
1992 #if MEMORY_LOGGING
1993 			fl->ifl_cl_enqueued++;
1994 #endif
1995 		} else {
1996 			bus_addr = sd_ba[frag_idx];
1997 		}
1998 		bus_dmamap_sync(fl->ifl_buf_tag, sd_map[frag_idx],
1999 		    BUS_DMASYNC_PREREAD);
2000 
2001 		if (sd_m[frag_idx] == NULL) {
2002 			if ((m = m_gethdr(M_NOWAIT, MT_NOINIT)) == NULL) {
2003 				break;
2004 			}
2005 			sd_m[frag_idx] = m;
2006 		}
2007 		bit_set(fl->ifl_rx_bitmap, frag_idx);
2008 #if MEMORY_LOGGING
2009 		fl->ifl_m_enqueued++;
2010 #endif
2011 
2012 		DBG_COUNTER_INC(rx_allocs);
2013 		fl->ifl_rxd_idxs[i] = frag_idx;
2014 		fl->ifl_bus_addrs[i] = bus_addr;
2015 		fl->ifl_vm_addrs[i] = cl;
2016 		credits++;
2017 		i++;
2018 		MPASS(credits <= fl->ifl_size);
2019 		if (++idx == fl->ifl_size) {
2020 			fl->ifl_gen = 1;
2021 			idx = 0;
2022 		}
2023 		if (n == 0 || i == IFLIB_MAX_RX_REFRESH) {
2024 			iru.iru_pidx = pidx;
2025 			iru.iru_count = i;
2026 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2027 			i = 0;
2028 			pidx = idx;
2029 			fl->ifl_pidx = idx;
2030 			fl->ifl_credits = credits;
2031 		}
2032 	}
2033 
2034 	if (i) {
2035 		iru.iru_pidx = pidx;
2036 		iru.iru_count = i;
2037 		ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2038 		fl->ifl_pidx = idx;
2039 		fl->ifl_credits = credits;
2040 	}
2041 	DBG_COUNTER_INC(rxd_flush);
2042 	if (fl->ifl_pidx == 0)
2043 		pidx = fl->ifl_size - 1;
2044 	else
2045 		pidx = fl->ifl_pidx - 1;
2046 
2047 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2048 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2049 	ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id, fl->ifl_id, pidx);
2050 	fl->ifl_fragidx = frag_idx + 1;
2051 	if (fl->ifl_fragidx == fl->ifl_size)
2052 		fl->ifl_fragidx = 0;
2053 
2054 	return (n == -1 ? 0 : IFLIB_RXEOF_EMPTY);
2055 }
2056 
2057 static __inline uint8_t
2058 __iflib_fl_refill_all(if_ctx_t ctx, iflib_fl_t fl)
2059 {
2060 	/* we avoid allowing pidx to catch up with cidx as it confuses ixl */
2061 	int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1;
2062 #ifdef INVARIANTS
2063 	int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1;
2064 #endif
2065 
2066 	MPASS(fl->ifl_credits <= fl->ifl_size);
2067 	MPASS(reclaimable == delta);
2068 
2069 	if (reclaimable > 0)
2070 		return (_iflib_fl_refill(ctx, fl, reclaimable));
2071 	return (0);
2072 }
2073 
2074 uint8_t
2075 iflib_in_detach(if_ctx_t ctx)
2076 {
2077 	bool in_detach;
2078 
2079 	STATE_LOCK(ctx);
2080 	in_detach = !!(ctx->ifc_flags & IFC_IN_DETACH);
2081 	STATE_UNLOCK(ctx);
2082 	return (in_detach);
2083 }
2084 
2085 static void
2086 iflib_fl_bufs_free(iflib_fl_t fl)
2087 {
2088 	iflib_dma_info_t idi = fl->ifl_ifdi;
2089 	bus_dmamap_t sd_map;
2090 	uint32_t i;
2091 
2092 	for (i = 0; i < fl->ifl_size; i++) {
2093 		struct mbuf **sd_m = &fl->ifl_sds.ifsd_m[i];
2094 		caddr_t *sd_cl = &fl->ifl_sds.ifsd_cl[i];
2095 
2096 		if (*sd_cl != NULL) {
2097 			sd_map = fl->ifl_sds.ifsd_map[i];
2098 			bus_dmamap_sync(fl->ifl_buf_tag, sd_map,
2099 			    BUS_DMASYNC_POSTREAD);
2100 			bus_dmamap_unload(fl->ifl_buf_tag, sd_map);
2101 			if (*sd_cl != NULL)
2102 				uma_zfree(fl->ifl_zone, *sd_cl);
2103 			if (*sd_m != NULL) {
2104 				m_init(*sd_m, M_NOWAIT, MT_DATA, 0);
2105 				uma_zfree(zone_mbuf, *sd_m);
2106 			}
2107 		} else {
2108 			MPASS(*sd_cl == NULL);
2109 			MPASS(*sd_m == NULL);
2110 		}
2111 #if MEMORY_LOGGING
2112 		fl->ifl_m_dequeued++;
2113 		fl->ifl_cl_dequeued++;
2114 #endif
2115 		*sd_cl = NULL;
2116 		*sd_m = NULL;
2117 	}
2118 #ifdef INVARIANTS
2119 	for (i = 0; i < fl->ifl_size; i++) {
2120 		MPASS(fl->ifl_sds.ifsd_cl[i] == NULL);
2121 		MPASS(fl->ifl_sds.ifsd_m[i] == NULL);
2122 	}
2123 #endif
2124 	/*
2125 	 * Reset free list values
2126 	 */
2127 	fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = fl->ifl_fragidx = 0;
2128 	bzero(idi->idi_vaddr, idi->idi_size);
2129 }
2130 
2131 /*********************************************************************
2132  *
2133  *  Initialize a free list and its buffers.
2134  *
2135  **********************************************************************/
2136 static int
2137 iflib_fl_setup(iflib_fl_t fl)
2138 {
2139 	iflib_rxq_t rxq = fl->ifl_rxq;
2140 	if_ctx_t ctx = rxq->ifr_ctx;
2141 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2142 	int qidx;
2143 
2144 	bit_nclear(fl->ifl_rx_bitmap, 0, fl->ifl_size - 1);
2145 	/*
2146 	** Free current RX buffer structs and their mbufs
2147 	*/
2148 	iflib_fl_bufs_free(fl);
2149 	/* Now replenish the mbufs */
2150 	MPASS(fl->ifl_credits == 0);
2151 	qidx = rxq->ifr_fl_offset + fl->ifl_id;
2152 	if (scctx->isc_rxd_buf_size[qidx] != 0)
2153 		fl->ifl_buf_size = scctx->isc_rxd_buf_size[qidx];
2154 	else
2155 		fl->ifl_buf_size = ctx->ifc_rx_mbuf_sz;
2156 	/*
2157 	 * ifl_buf_size may be a driver-supplied value, so pull it up
2158 	 * to the selected mbuf size.
2159 	 */
2160 	fl->ifl_buf_size = iflib_get_mbuf_size_for(fl->ifl_buf_size);
2161 	if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size)
2162 		ctx->ifc_max_fl_buf_size = fl->ifl_buf_size;
2163 	fl->ifl_cltype = m_gettype(fl->ifl_buf_size);
2164 	fl->ifl_zone = m_getzone(fl->ifl_buf_size);
2165 
2166 
2167 	/* avoid pre-allocating zillions of clusters to an idle card
2168 	 * potentially speeding up attach
2169 	 */
2170 	(void) _iflib_fl_refill(ctx, fl, min(128, fl->ifl_size));
2171 	MPASS(min(128, fl->ifl_size) == fl->ifl_credits);
2172 	if (min(128, fl->ifl_size) != fl->ifl_credits)
2173 		return (ENOBUFS);
2174 	/*
2175 	 * handle failure
2176 	 */
2177 	MPASS(rxq != NULL);
2178 	MPASS(fl->ifl_ifdi != NULL);
2179 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2180 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2181 	return (0);
2182 }
2183 
2184 /*********************************************************************
2185  *
2186  *  Free receive ring data structures
2187  *
2188  **********************************************************************/
2189 static void
2190 iflib_rx_sds_free(iflib_rxq_t rxq)
2191 {
2192 	iflib_fl_t fl;
2193 	int i, j;
2194 
2195 	if (rxq->ifr_fl != NULL) {
2196 		for (i = 0; i < rxq->ifr_nfl; i++) {
2197 			fl = &rxq->ifr_fl[i];
2198 			if (fl->ifl_buf_tag != NULL) {
2199 				if (fl->ifl_sds.ifsd_map != NULL) {
2200 					for (j = 0; j < fl->ifl_size; j++) {
2201 						bus_dmamap_sync(
2202 						    fl->ifl_buf_tag,
2203 						    fl->ifl_sds.ifsd_map[j],
2204 						    BUS_DMASYNC_POSTREAD);
2205 						bus_dmamap_unload(
2206 						    fl->ifl_buf_tag,
2207 						    fl->ifl_sds.ifsd_map[j]);
2208 						bus_dmamap_destroy(
2209 						    fl->ifl_buf_tag,
2210 						    fl->ifl_sds.ifsd_map[j]);
2211 					}
2212 				}
2213 				bus_dma_tag_destroy(fl->ifl_buf_tag);
2214 				fl->ifl_buf_tag = NULL;
2215 			}
2216 			free(fl->ifl_sds.ifsd_m, M_IFLIB);
2217 			free(fl->ifl_sds.ifsd_cl, M_IFLIB);
2218 			free(fl->ifl_sds.ifsd_ba, M_IFLIB);
2219 			free(fl->ifl_sds.ifsd_map, M_IFLIB);
2220 			fl->ifl_sds.ifsd_m = NULL;
2221 			fl->ifl_sds.ifsd_cl = NULL;
2222 			fl->ifl_sds.ifsd_ba = NULL;
2223 			fl->ifl_sds.ifsd_map = NULL;
2224 		}
2225 		free(rxq->ifr_fl, M_IFLIB);
2226 		rxq->ifr_fl = NULL;
2227 		free(rxq->ifr_ifdi, M_IFLIB);
2228 		rxq->ifr_ifdi = NULL;
2229 		rxq->ifr_cq_cidx = 0;
2230 	}
2231 }
2232 
2233 /*
2234  * Timer routine
2235  */
2236 static void
2237 iflib_timer(void *arg)
2238 {
2239 	iflib_txq_t txq = arg;
2240 	if_ctx_t ctx = txq->ift_ctx;
2241 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2242 	uint64_t this_tick = ticks;
2243 	uint32_t reset_on = hz / 2;
2244 
2245 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
2246 		return;
2247 
2248 	/*
2249 	** Check on the state of the TX queue(s), this
2250 	** can be done without the lock because its RO
2251 	** and the HUNG state will be static if set.
2252 	*/
2253 	if (this_tick - txq->ift_last_timer_tick >= hz / 2) {
2254 		txq->ift_last_timer_tick = this_tick;
2255 		IFDI_TIMER(ctx, txq->ift_id);
2256 		if ((txq->ift_qstatus == IFLIB_QUEUE_HUNG) &&
2257 		    ((txq->ift_cleaned_prev == txq->ift_cleaned) ||
2258 		     (sctx->isc_pause_frames == 0)))
2259 			goto hung;
2260 
2261 		if (txq->ift_qstatus != IFLIB_QUEUE_IDLE &&
2262 		    ifmp_ring_is_stalled(txq->ift_br)) {
2263 			KASSERT(ctx->ifc_link_state == LINK_STATE_UP, ("queue can't be marked as hung if interface is down"));
2264 			txq->ift_qstatus = IFLIB_QUEUE_HUNG;
2265 		}
2266 		txq->ift_cleaned_prev = txq->ift_cleaned;
2267 	}
2268 #ifdef DEV_NETMAP
2269 	if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP)
2270 		iflib_netmap_timer_adjust(ctx, txq, &reset_on);
2271 #endif
2272 	/* handle any laggards */
2273 	if (txq->ift_db_pending)
2274 		GROUPTASK_ENQUEUE(&txq->ift_task);
2275 
2276 	sctx->isc_pause_frames = 0;
2277 	if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)
2278 		callout_reset_on(&txq->ift_timer, reset_on, iflib_timer, txq, txq->ift_timer.c_cpu);
2279 	return;
2280 
2281  hung:
2282 	device_printf(ctx->ifc_dev,
2283 	    "Watchdog timeout (TX: %d desc avail: %d pidx: %d) -- resetting\n",
2284 	    txq->ift_id, TXQ_AVAIL(txq), txq->ift_pidx);
2285 	STATE_LOCK(ctx);
2286 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2287 	ctx->ifc_flags |= (IFC_DO_WATCHDOG|IFC_DO_RESET);
2288 	iflib_admin_intr_deferred(ctx);
2289 	STATE_UNLOCK(ctx);
2290 }
2291 
2292 static uint16_t
2293 iflib_get_mbuf_size_for(unsigned int size)
2294 {
2295 
2296 	if (size <= MCLBYTES)
2297 		return (MCLBYTES);
2298 	else
2299 		return (MJUMPAGESIZE);
2300 }
2301 
2302 static void
2303 iflib_calc_rx_mbuf_sz(if_ctx_t ctx)
2304 {
2305 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2306 
2307 	/*
2308 	 * XXX don't set the max_frame_size to larger
2309 	 * than the hardware can handle
2310 	 */
2311 	ctx->ifc_rx_mbuf_sz =
2312 	    iflib_get_mbuf_size_for(sctx->isc_max_frame_size);
2313 }
2314 
2315 uint32_t
2316 iflib_get_rx_mbuf_sz(if_ctx_t ctx)
2317 {
2318 
2319 	return (ctx->ifc_rx_mbuf_sz);
2320 }
2321 
2322 static void
2323 iflib_init_locked(if_ctx_t ctx)
2324 {
2325 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2326 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2327 	if_t ifp = ctx->ifc_ifp;
2328 	iflib_fl_t fl;
2329 	iflib_txq_t txq;
2330 	iflib_rxq_t rxq;
2331 	int i, j, tx_ip_csum_flags, tx_ip6_csum_flags;
2332 
2333 	if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2334 	IFDI_INTR_DISABLE(ctx);
2335 
2336 	tx_ip_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP);
2337 	tx_ip6_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | CSUM_IP6_SCTP);
2338 	/* Set hardware offload abilities */
2339 	if_clearhwassist(ifp);
2340 	if (if_getcapenable(ifp) & IFCAP_TXCSUM)
2341 		if_sethwassistbits(ifp, tx_ip_csum_flags, 0);
2342 	if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6)
2343 		if_sethwassistbits(ifp,  tx_ip6_csum_flags, 0);
2344 	if (if_getcapenable(ifp) & IFCAP_TSO4)
2345 		if_sethwassistbits(ifp, CSUM_IP_TSO, 0);
2346 	if (if_getcapenable(ifp) & IFCAP_TSO6)
2347 		if_sethwassistbits(ifp, CSUM_IP6_TSO, 0);
2348 
2349 	for (i = 0, txq = ctx->ifc_txqs; i < sctx->isc_ntxqsets; i++, txq++) {
2350 		CALLOUT_LOCK(txq);
2351 		callout_stop(&txq->ift_timer);
2352 		CALLOUT_UNLOCK(txq);
2353 		iflib_netmap_txq_init(ctx, txq);
2354 	}
2355 
2356 	/*
2357 	 * Calculate a suitable Rx mbuf size prior to calling IFDI_INIT, so
2358 	 * that drivers can use the value when setting up the hardware receive
2359 	 * buffers.
2360 	 */
2361 	iflib_calc_rx_mbuf_sz(ctx);
2362 
2363 #ifdef INVARIANTS
2364 	i = if_getdrvflags(ifp);
2365 #endif
2366 	IFDI_INIT(ctx);
2367 	MPASS(if_getdrvflags(ifp) == i);
2368 	for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
2369 		/* XXX this should really be done on a per-queue basis */
2370 		if (if_getcapenable(ifp) & IFCAP_NETMAP) {
2371 			MPASS(rxq->ifr_id == i);
2372 			iflib_netmap_rxq_init(ctx, rxq);
2373 			continue;
2374 		}
2375 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
2376 			if (iflib_fl_setup(fl)) {
2377 				device_printf(ctx->ifc_dev,
2378 				    "setting up free list %d failed - "
2379 				    "check cluster settings\n", j);
2380 				goto done;
2381 			}
2382 		}
2383 	}
2384 done:
2385 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
2386 	IFDI_INTR_ENABLE(ctx);
2387 	txq = ctx->ifc_txqs;
2388 	for (i = 0; i < sctx->isc_ntxqsets; i++, txq++)
2389 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq,
2390 			txq->ift_timer.c_cpu);
2391 }
2392 
2393 static int
2394 iflib_media_change(if_t ifp)
2395 {
2396 	if_ctx_t ctx = if_getsoftc(ifp);
2397 	int err;
2398 
2399 	CTX_LOCK(ctx);
2400 	if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0)
2401 		iflib_init_locked(ctx);
2402 	CTX_UNLOCK(ctx);
2403 	return (err);
2404 }
2405 
2406 static void
2407 iflib_media_status(if_t ifp, struct ifmediareq *ifmr)
2408 {
2409 	if_ctx_t ctx = if_getsoftc(ifp);
2410 
2411 	CTX_LOCK(ctx);
2412 	IFDI_UPDATE_ADMIN_STATUS(ctx);
2413 	IFDI_MEDIA_STATUS(ctx, ifmr);
2414 	CTX_UNLOCK(ctx);
2415 }
2416 
2417 void
2418 iflib_stop(if_ctx_t ctx)
2419 {
2420 	iflib_txq_t txq = ctx->ifc_txqs;
2421 	iflib_rxq_t rxq = ctx->ifc_rxqs;
2422 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2423 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2424 	iflib_dma_info_t di;
2425 	iflib_fl_t fl;
2426 	int i, j;
2427 
2428 	/* Tell the stack that the interface is no longer active */
2429 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2430 
2431 	IFDI_INTR_DISABLE(ctx);
2432 	DELAY(1000);
2433 	IFDI_STOP(ctx);
2434 	DELAY(1000);
2435 
2436 	iflib_debug_reset();
2437 	/* Wait for current tx queue users to exit to disarm watchdog timer. */
2438 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) {
2439 		/* make sure all transmitters have completed before proceeding XXX */
2440 
2441 		CALLOUT_LOCK(txq);
2442 		callout_stop(&txq->ift_timer);
2443 		CALLOUT_UNLOCK(txq);
2444 
2445 		/* clean any enqueued buffers */
2446 		iflib_ifmp_purge(txq);
2447 		/* Free any existing tx buffers. */
2448 		for (j = 0; j < txq->ift_size; j++) {
2449 			iflib_txsd_free(ctx, txq, j);
2450 		}
2451 		txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0;
2452 		txq->ift_in_use = txq->ift_gen = txq->ift_cidx = txq->ift_pidx = txq->ift_no_desc_avail = 0;
2453 		txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0;
2454 		txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0;
2455 		txq->ift_pullups = 0;
2456 		ifmp_ring_reset_stats(txq->ift_br);
2457 		for (j = 0, di = txq->ift_ifdi; j < sctx->isc_ntxqs; j++, di++)
2458 			bzero((void *)di->idi_vaddr, di->idi_size);
2459 	}
2460 	for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) {
2461 		/* make sure all transmitters have completed before proceeding XXX */
2462 
2463 		rxq->ifr_cq_cidx = 0;
2464 		for (j = 0, di = rxq->ifr_ifdi; j < sctx->isc_nrxqs; j++, di++)
2465 			bzero((void *)di->idi_vaddr, di->idi_size);
2466 		/* also resets the free lists pidx/cidx */
2467 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
2468 			iflib_fl_bufs_free(fl);
2469 	}
2470 }
2471 
2472 static inline caddr_t
2473 calc_next_rxd(iflib_fl_t fl, int cidx)
2474 {
2475 	qidx_t size;
2476 	int nrxd;
2477 	caddr_t start, end, cur, next;
2478 
2479 	nrxd = fl->ifl_size;
2480 	size = fl->ifl_rxd_size;
2481 	start = fl->ifl_ifdi->idi_vaddr;
2482 
2483 	if (__predict_false(size == 0))
2484 		return (start);
2485 	cur = start + size*cidx;
2486 	end = start + size*nrxd;
2487 	next = CACHE_PTR_NEXT(cur);
2488 	return (next < end ? next : start);
2489 }
2490 
2491 static inline void
2492 prefetch_pkts(iflib_fl_t fl, int cidx)
2493 {
2494 	int nextptr;
2495 	int nrxd = fl->ifl_size;
2496 	caddr_t next_rxd;
2497 
2498 
2499 	nextptr = (cidx + CACHE_PTR_INCREMENT) & (nrxd-1);
2500 	prefetch(&fl->ifl_sds.ifsd_m[nextptr]);
2501 	prefetch(&fl->ifl_sds.ifsd_cl[nextptr]);
2502 	next_rxd = calc_next_rxd(fl, cidx);
2503 	prefetch(next_rxd);
2504 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 1) & (nrxd-1)]);
2505 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 2) & (nrxd-1)]);
2506 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 3) & (nrxd-1)]);
2507 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 4) & (nrxd-1)]);
2508 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 1) & (nrxd-1)]);
2509 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 2) & (nrxd-1)]);
2510 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 3) & (nrxd-1)]);
2511 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 4) & (nrxd-1)]);
2512 }
2513 
2514 static struct mbuf *
2515 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, bool unload, if_rxsd_t sd,
2516     int *pf_rv, if_rxd_info_t ri)
2517 {
2518 	bus_dmamap_t map;
2519 	iflib_fl_t fl;
2520 	caddr_t payload;
2521 	struct mbuf *m;
2522 	int flid, cidx, len, next;
2523 
2524 	map = NULL;
2525 	flid = irf->irf_flid;
2526 	cidx = irf->irf_idx;
2527 	fl = &rxq->ifr_fl[flid];
2528 	sd->ifsd_fl = fl;
2529 	m = fl->ifl_sds.ifsd_m[cidx];
2530 	sd->ifsd_cl = &fl->ifl_sds.ifsd_cl[cidx];
2531 	fl->ifl_credits--;
2532 #if MEMORY_LOGGING
2533 	fl->ifl_m_dequeued++;
2534 #endif
2535 	if (rxq->ifr_ctx->ifc_flags & IFC_PREFETCH)
2536 		prefetch_pkts(fl, cidx);
2537 	next = (cidx + CACHE_PTR_INCREMENT) & (fl->ifl_size-1);
2538 	prefetch(&fl->ifl_sds.ifsd_map[next]);
2539 	map = fl->ifl_sds.ifsd_map[cidx];
2540 
2541 	bus_dmamap_sync(fl->ifl_buf_tag, map, BUS_DMASYNC_POSTREAD);
2542 
2543 	if (rxq->pfil != NULL && PFIL_HOOKED_IN(rxq->pfil) && pf_rv != NULL &&
2544 	    irf->irf_len != 0) {
2545 		payload  = *sd->ifsd_cl;
2546 		payload +=  ri->iri_pad;
2547 		len = ri->iri_len - ri->iri_pad;
2548 		*pf_rv = pfil_run_hooks(rxq->pfil, payload, ri->iri_ifp,
2549 		    len | PFIL_MEMPTR | PFIL_IN, NULL);
2550 		switch (*pf_rv) {
2551 		case PFIL_DROPPED:
2552 		case PFIL_CONSUMED:
2553 			/*
2554 			 * The filter ate it.  Everything is recycled.
2555 			 */
2556 			m = NULL;
2557 			unload = 0;
2558 			break;
2559 		case PFIL_REALLOCED:
2560 			/*
2561 			 * The filter copied it.  Everything is recycled.
2562 			 */
2563 			m = pfil_mem2mbuf(payload);
2564 			unload = 0;
2565 			break;
2566 		case PFIL_PASS:
2567 			/*
2568 			 * Filter said it was OK, so receive like
2569 			 * normal
2570 			 */
2571 			fl->ifl_sds.ifsd_m[cidx] = NULL;
2572 			break;
2573 		default:
2574 			MPASS(0);
2575 		}
2576 	} else {
2577 		fl->ifl_sds.ifsd_m[cidx] = NULL;
2578 		*pf_rv = PFIL_PASS;
2579 	}
2580 
2581 	if (unload && irf->irf_len != 0)
2582 		bus_dmamap_unload(fl->ifl_buf_tag, map);
2583 	fl->ifl_cidx = (fl->ifl_cidx + 1) & (fl->ifl_size-1);
2584 	if (__predict_false(fl->ifl_cidx == 0))
2585 		fl->ifl_gen = 0;
2586 	bit_clear(fl->ifl_rx_bitmap, cidx);
2587 	return (m);
2588 }
2589 
2590 static struct mbuf *
2591 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri, if_rxsd_t sd, int *pf_rv)
2592 {
2593 	struct mbuf *m, *mh, *mt;
2594 	caddr_t cl;
2595 	int  *pf_rv_ptr, flags, i, padlen;
2596 	bool consumed;
2597 
2598 	i = 0;
2599 	mh = NULL;
2600 	consumed = false;
2601 	*pf_rv = PFIL_PASS;
2602 	pf_rv_ptr = pf_rv;
2603 	do {
2604 		m = rxd_frag_to_sd(rxq, &ri->iri_frags[i], !consumed, sd,
2605 		    pf_rv_ptr, ri);
2606 
2607 		MPASS(*sd->ifsd_cl != NULL);
2608 
2609 		/*
2610 		 * Exclude zero-length frags & frags from
2611 		 * packets the filter has consumed or dropped
2612 		 */
2613 		if (ri->iri_frags[i].irf_len == 0 || consumed ||
2614 		    *pf_rv == PFIL_CONSUMED || *pf_rv == PFIL_DROPPED) {
2615 			if (mh == NULL) {
2616 				/* everything saved here */
2617 				consumed = true;
2618 				pf_rv_ptr = NULL;
2619 				continue;
2620 			}
2621 			/* XXX we can save the cluster here, but not the mbuf */
2622 			m_init(m, M_NOWAIT, MT_DATA, 0);
2623 			m_free(m);
2624 			continue;
2625 		}
2626 		if (mh == NULL) {
2627 			flags = M_PKTHDR|M_EXT;
2628 			mh = mt = m;
2629 			padlen = ri->iri_pad;
2630 		} else {
2631 			flags = M_EXT;
2632 			mt->m_next = m;
2633 			mt = m;
2634 			/* assuming padding is only on the first fragment */
2635 			padlen = 0;
2636 		}
2637 		cl = *sd->ifsd_cl;
2638 		*sd->ifsd_cl = NULL;
2639 
2640 		/* Can these two be made one ? */
2641 		m_init(m, M_NOWAIT, MT_DATA, flags);
2642 		m_cljset(m, cl, sd->ifsd_fl->ifl_cltype);
2643 		/*
2644 		 * These must follow m_init and m_cljset
2645 		 */
2646 		m->m_data += padlen;
2647 		ri->iri_len -= padlen;
2648 		m->m_len = ri->iri_frags[i].irf_len;
2649 	} while (++i < ri->iri_nfrags);
2650 
2651 	return (mh);
2652 }
2653 
2654 /*
2655  * Process one software descriptor
2656  */
2657 static struct mbuf *
2658 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri)
2659 {
2660 	struct if_rxsd sd;
2661 	struct mbuf *m;
2662 	int pf_rv;
2663 
2664 	/* should I merge this back in now that the two paths are basically duplicated? */
2665 	if (ri->iri_nfrags == 1 &&
2666 	    ri->iri_frags[0].irf_len != 0 &&
2667 	    ri->iri_frags[0].irf_len <= MIN(IFLIB_RX_COPY_THRESH, MHLEN)) {
2668 		m = rxd_frag_to_sd(rxq, &ri->iri_frags[0], false, &sd,
2669 		    &pf_rv, ri);
2670 		if (pf_rv != PFIL_PASS && pf_rv != PFIL_REALLOCED)
2671 			return (m);
2672 		if (pf_rv == PFIL_PASS) {
2673 			m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR);
2674 #ifndef __NO_STRICT_ALIGNMENT
2675 			if (!IP_ALIGNED(m))
2676 				m->m_data += 2;
2677 #endif
2678 			memcpy(m->m_data, *sd.ifsd_cl, ri->iri_len);
2679 			m->m_len = ri->iri_frags[0].irf_len;
2680 		}
2681 	} else {
2682 		m = assemble_segments(rxq, ri, &sd, &pf_rv);
2683 		if (m == NULL)
2684 			return (NULL);
2685 		if (pf_rv != PFIL_PASS && pf_rv != PFIL_REALLOCED)
2686 			return (m);
2687 	}
2688 	m->m_pkthdr.len = ri->iri_len;
2689 	m->m_pkthdr.rcvif = ri->iri_ifp;
2690 	m->m_flags |= ri->iri_flags;
2691 	m->m_pkthdr.ether_vtag = ri->iri_vtag;
2692 	m->m_pkthdr.flowid = ri->iri_flowid;
2693 	M_HASHTYPE_SET(m, ri->iri_rsstype);
2694 	m->m_pkthdr.csum_flags = ri->iri_csum_flags;
2695 	m->m_pkthdr.csum_data = ri->iri_csum_data;
2696 	return (m);
2697 }
2698 
2699 #if defined(INET6) || defined(INET)
2700 static void
2701 iflib_get_ip_forwarding(struct lro_ctrl *lc, bool *v4, bool *v6)
2702 {
2703 	CURVNET_SET(lc->ifp->if_vnet);
2704 #if defined(INET6)
2705 	*v6 = V_ip6_forwarding;
2706 #endif
2707 #if defined(INET)
2708 	*v4 = V_ipforwarding;
2709 #endif
2710 	CURVNET_RESTORE();
2711 }
2712 
2713 /*
2714  * Returns true if it's possible this packet could be LROed.
2715  * if it returns false, it is guaranteed that tcp_lro_rx()
2716  * would not return zero.
2717  */
2718 static bool
2719 iflib_check_lro_possible(struct mbuf *m, bool v4_forwarding, bool v6_forwarding)
2720 {
2721 	struct ether_header *eh;
2722 
2723 	eh = mtod(m, struct ether_header *);
2724 	switch (eh->ether_type) {
2725 #if defined(INET6)
2726 		case htons(ETHERTYPE_IPV6):
2727 			return (!v6_forwarding);
2728 #endif
2729 #if defined (INET)
2730 		case htons(ETHERTYPE_IP):
2731 			return (!v4_forwarding);
2732 #endif
2733 	}
2734 
2735 	return false;
2736 }
2737 #else
2738 static void
2739 iflib_get_ip_forwarding(struct lro_ctrl *lc __unused, bool *v4 __unused, bool *v6 __unused)
2740 {
2741 }
2742 #endif
2743 
2744 static void
2745 _task_fn_rx_watchdog(void *context)
2746 {
2747 	iflib_rxq_t rxq = context;
2748 
2749 	GROUPTASK_ENQUEUE(&rxq->ifr_task);
2750 }
2751 
2752 static uint8_t
2753 iflib_rxeof(iflib_rxq_t rxq, qidx_t budget)
2754 {
2755 	if_t ifp;
2756 	if_ctx_t ctx = rxq->ifr_ctx;
2757 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2758 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2759 	int avail, i;
2760 	qidx_t *cidxp;
2761 	struct if_rxd_info ri;
2762 	int err, budget_left, rx_bytes, rx_pkts;
2763 	iflib_fl_t fl;
2764 	int lro_enabled;
2765 	bool v4_forwarding, v6_forwarding, lro_possible;
2766 	uint8_t retval = 0;
2767 
2768 	/*
2769 	 * XXX early demux data packets so that if_input processing only handles
2770 	 * acks in interrupt context
2771 	 */
2772 	struct mbuf *m, *mh, *mt, *mf;
2773 
2774 	NET_EPOCH_ASSERT();
2775 
2776 	lro_possible = v4_forwarding = v6_forwarding = false;
2777 	ifp = ctx->ifc_ifp;
2778 	mh = mt = NULL;
2779 	MPASS(budget > 0);
2780 	rx_pkts	= rx_bytes = 0;
2781 	if (sctx->isc_flags & IFLIB_HAS_RXCQ)
2782 		cidxp = &rxq->ifr_cq_cidx;
2783 	else
2784 		cidxp = &rxq->ifr_fl[0].ifl_cidx;
2785 	if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) {
2786 		for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2787 			retval |= __iflib_fl_refill_all(ctx, fl);
2788 		DBG_COUNTER_INC(rx_unavail);
2789 		return (retval);
2790 	}
2791 
2792 	/* pfil needs the vnet to be set */
2793 	CURVNET_SET_QUIET(ifp->if_vnet);
2794 	for (budget_left = budget; budget_left > 0 && avail > 0;) {
2795 		if (__predict_false(!CTX_ACTIVE(ctx))) {
2796 			DBG_COUNTER_INC(rx_ctx_inactive);
2797 			break;
2798 		}
2799 		/*
2800 		 * Reset client set fields to their default values
2801 		 */
2802 		rxd_info_zero(&ri);
2803 		ri.iri_qsidx = rxq->ifr_id;
2804 		ri.iri_cidx = *cidxp;
2805 		ri.iri_ifp = ifp;
2806 		ri.iri_frags = rxq->ifr_frags;
2807 		err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
2808 
2809 		if (err)
2810 			goto err;
2811 		rx_pkts += 1;
2812 		rx_bytes += ri.iri_len;
2813 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
2814 			*cidxp = ri.iri_cidx;
2815 			/* Update our consumer index */
2816 			/* XXX NB: shurd - check if this is still safe */
2817 			while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0])
2818 				rxq->ifr_cq_cidx -= scctx->isc_nrxd[0];
2819 			/* was this only a completion queue message? */
2820 			if (__predict_false(ri.iri_nfrags == 0))
2821 				continue;
2822 		}
2823 		MPASS(ri.iri_nfrags != 0);
2824 		MPASS(ri.iri_len != 0);
2825 
2826 		/* will advance the cidx on the corresponding free lists */
2827 		m = iflib_rxd_pkt_get(rxq, &ri);
2828 		avail--;
2829 		budget_left--;
2830 		if (avail == 0 && budget_left)
2831 			avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left);
2832 
2833 		if (__predict_false(m == NULL))
2834 			continue;
2835 
2836 		/* imm_pkt: -- cxgb */
2837 		if (mh == NULL)
2838 			mh = mt = m;
2839 		else {
2840 			mt->m_nextpkt = m;
2841 			mt = m;
2842 		}
2843 	}
2844 	CURVNET_RESTORE();
2845 	/* make sure that we can refill faster than drain */
2846 	for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2847 		retval |= __iflib_fl_refill_all(ctx, fl);
2848 
2849 	lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO);
2850 	if (lro_enabled)
2851 		iflib_get_ip_forwarding(&rxq->ifr_lc, &v4_forwarding, &v6_forwarding);
2852 	mt = mf = NULL;
2853 	while (mh != NULL) {
2854 		m = mh;
2855 		mh = mh->m_nextpkt;
2856 		m->m_nextpkt = NULL;
2857 #ifndef __NO_STRICT_ALIGNMENT
2858 		if (!IP_ALIGNED(m) && (m = iflib_fixup_rx(m)) == NULL)
2859 			continue;
2860 #endif
2861 		rx_bytes += m->m_pkthdr.len;
2862 		rx_pkts++;
2863 #if defined(INET6) || defined(INET)
2864 		if (lro_enabled) {
2865 			if (!lro_possible) {
2866 				lro_possible = iflib_check_lro_possible(m, v4_forwarding, v6_forwarding);
2867 				if (lro_possible && mf != NULL) {
2868 					ifp->if_input(ifp, mf);
2869 					DBG_COUNTER_INC(rx_if_input);
2870 					mt = mf = NULL;
2871 				}
2872 			}
2873 			if ((m->m_pkthdr.csum_flags & (CSUM_L4_CALC|CSUM_L4_VALID)) ==
2874 			    (CSUM_L4_CALC|CSUM_L4_VALID)) {
2875 				if (lro_possible && tcp_lro_rx(&rxq->ifr_lc, m, 0) == 0)
2876 					continue;
2877 			}
2878 		}
2879 #endif
2880 		if (lro_possible) {
2881 			ifp->if_input(ifp, m);
2882 			DBG_COUNTER_INC(rx_if_input);
2883 			continue;
2884 		}
2885 
2886 		if (mf == NULL)
2887 			mf = m;
2888 		if (mt != NULL)
2889 			mt->m_nextpkt = m;
2890 		mt = m;
2891 	}
2892 	if (mf != NULL) {
2893 		ifp->if_input(ifp, mf);
2894 		DBG_COUNTER_INC(rx_if_input);
2895 	}
2896 
2897 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
2898 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
2899 
2900 	/*
2901 	 * Flush any outstanding LRO work
2902 	 */
2903 #if defined(INET6) || defined(INET)
2904 	tcp_lro_flush_all(&rxq->ifr_lc);
2905 #endif
2906 	if (avail != 0 || iflib_rxd_avail(ctx, rxq, *cidxp, 1) != 0)
2907 		retval |= IFLIB_RXEOF_MORE;
2908 	return (retval);
2909 err:
2910 	STATE_LOCK(ctx);
2911 	ctx->ifc_flags |= IFC_DO_RESET;
2912 	iflib_admin_intr_deferred(ctx);
2913 	STATE_UNLOCK(ctx);
2914 	return (0);
2915 }
2916 
2917 #define TXD_NOTIFY_COUNT(txq) (((txq)->ift_size / (txq)->ift_update_freq)-1)
2918 static inline qidx_t
2919 txq_max_db_deferred(iflib_txq_t txq, qidx_t in_use)
2920 {
2921 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2922 	qidx_t minthresh = txq->ift_size / 8;
2923 	if (in_use > 4*minthresh)
2924 		return (notify_count);
2925 	if (in_use > 2*minthresh)
2926 		return (notify_count >> 1);
2927 	if (in_use > minthresh)
2928 		return (notify_count >> 3);
2929 	return (0);
2930 }
2931 
2932 static inline qidx_t
2933 txq_max_rs_deferred(iflib_txq_t txq)
2934 {
2935 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2936 	qidx_t minthresh = txq->ift_size / 8;
2937 	if (txq->ift_in_use > 4*minthresh)
2938 		return (notify_count);
2939 	if (txq->ift_in_use > 2*minthresh)
2940 		return (notify_count >> 1);
2941 	if (txq->ift_in_use > minthresh)
2942 		return (notify_count >> 2);
2943 	return (2);
2944 }
2945 
2946 #define M_CSUM_FLAGS(m) ((m)->m_pkthdr.csum_flags)
2947 #define M_HAS_VLANTAG(m) (m->m_flags & M_VLANTAG)
2948 
2949 #define TXQ_MAX_DB_DEFERRED(txq, in_use) txq_max_db_deferred((txq), (in_use))
2950 #define TXQ_MAX_RS_DEFERRED(txq) txq_max_rs_deferred(txq)
2951 #define TXQ_MAX_DB_CONSUMED(size) (size >> 4)
2952 
2953 /* forward compatibility for cxgb */
2954 #define FIRST_QSET(ctx) 0
2955 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets)
2956 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets)
2957 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx))
2958 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments))
2959 
2960 /* XXX we should be setting this to something other than zero */
2961 #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh)
2962 #define	MAX_TX_DESC(ctx) max((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max, \
2963     (ctx)->ifc_softc_ctx.isc_tx_nsegments)
2964 
2965 static inline bool
2966 iflib_txd_db_check(if_ctx_t ctx, iflib_txq_t txq, int ring, qidx_t in_use)
2967 {
2968 	qidx_t dbval, max;
2969 	bool rang;
2970 
2971 	rang = false;
2972 	max = TXQ_MAX_DB_DEFERRED(txq, in_use);
2973 	if (ring || txq->ift_db_pending >= max) {
2974 		dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx;
2975 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
2976 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2977 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval);
2978 		txq->ift_db_pending = txq->ift_npending = 0;
2979 		rang = true;
2980 	}
2981 	return (rang);
2982 }
2983 
2984 #ifdef PKT_DEBUG
2985 static void
2986 print_pkt(if_pkt_info_t pi)
2987 {
2988 	printf("pi len:  %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n",
2989 	       pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx);
2990 	printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n",
2991 	       pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag);
2992 	printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n",
2993 	       pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto);
2994 }
2995 #endif
2996 
2997 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO)
2998 #define IS_TX_OFFLOAD4(pi) ((pi)->ipi_csum_flags & (CSUM_IP_TCP | CSUM_IP_TSO))
2999 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO)
3000 #define IS_TX_OFFLOAD6(pi) ((pi)->ipi_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_TSO))
3001 
3002 static int
3003 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp)
3004 {
3005 	if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx;
3006 	struct ether_vlan_header *eh;
3007 	struct mbuf *m;
3008 
3009 	m = *mp;
3010 	if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) &&
3011 	    M_WRITABLE(m) == 0) {
3012 		if ((m = m_dup(m, M_NOWAIT)) == NULL) {
3013 			return (ENOMEM);
3014 		} else {
3015 			m_freem(*mp);
3016 			DBG_COUNTER_INC(tx_frees);
3017 			*mp = m;
3018 		}
3019 	}
3020 
3021 	/*
3022 	 * Determine where frame payload starts.
3023 	 * Jump over vlan headers if already present,
3024 	 * helpful for QinQ too.
3025 	 */
3026 	if (__predict_false(m->m_len < sizeof(*eh))) {
3027 		txq->ift_pullups++;
3028 		if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL))
3029 			return (ENOMEM);
3030 	}
3031 	eh = mtod(m, struct ether_vlan_header *);
3032 	if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
3033 		pi->ipi_etype = ntohs(eh->evl_proto);
3034 		pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
3035 	} else {
3036 		pi->ipi_etype = ntohs(eh->evl_encap_proto);
3037 		pi->ipi_ehdrlen = ETHER_HDR_LEN;
3038 	}
3039 
3040 	switch (pi->ipi_etype) {
3041 #ifdef INET
3042 	case ETHERTYPE_IP:
3043 	{
3044 		struct mbuf *n;
3045 		struct ip *ip = NULL;
3046 		struct tcphdr *th = NULL;
3047 		int minthlen;
3048 
3049 		minthlen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip) + sizeof(*th));
3050 		if (__predict_false(m->m_len < minthlen)) {
3051 			/*
3052 			 * if this code bloat is causing too much of a hit
3053 			 * move it to a separate function and mark it noinline
3054 			 */
3055 			if (m->m_len == pi->ipi_ehdrlen) {
3056 				n = m->m_next;
3057 				MPASS(n);
3058 				if (n->m_len >= sizeof(*ip))  {
3059 					ip = (struct ip *)n->m_data;
3060 					if (n->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3061 						th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3062 				} else {
3063 					txq->ift_pullups++;
3064 					if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
3065 						return (ENOMEM);
3066 					ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3067 				}
3068 			} else {
3069 				txq->ift_pullups++;
3070 				if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
3071 					return (ENOMEM);
3072 				ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3073 				if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3074 					th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3075 			}
3076 		} else {
3077 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3078 			if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3079 				th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3080 		}
3081 		pi->ipi_ip_hlen = ip->ip_hl << 2;
3082 		pi->ipi_ipproto = ip->ip_p;
3083 		pi->ipi_flags |= IPI_TX_IPV4;
3084 
3085 		/* TCP checksum offload may require TCP header length */
3086 		if (IS_TX_OFFLOAD4(pi)) {
3087 			if (__predict_true(pi->ipi_ipproto == IPPROTO_TCP)) {
3088 				if (__predict_false(th == NULL)) {
3089 					txq->ift_pullups++;
3090 					if (__predict_false((m = m_pullup(m, (ip->ip_hl << 2) + sizeof(*th))) == NULL))
3091 						return (ENOMEM);
3092 					th = (struct tcphdr *)((caddr_t)ip + pi->ipi_ip_hlen);
3093 				}
3094 				pi->ipi_tcp_hflags = th->th_flags;
3095 				pi->ipi_tcp_hlen = th->th_off << 2;
3096 				pi->ipi_tcp_seq = th->th_seq;
3097 			}
3098 			if (IS_TSO4(pi)) {
3099 				if (__predict_false(ip->ip_p != IPPROTO_TCP))
3100 					return (ENXIO);
3101 				/*
3102 				 * TSO always requires hardware checksum offload.
3103 				 */
3104 				pi->ipi_csum_flags |= (CSUM_IP_TCP | CSUM_IP);
3105 				th->th_sum = in_pseudo(ip->ip_src.s_addr,
3106 						       ip->ip_dst.s_addr, htons(IPPROTO_TCP));
3107 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3108 				if (sctx->isc_flags & IFLIB_TSO_INIT_IP) {
3109 					ip->ip_sum = 0;
3110 					ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz);
3111 				}
3112 			}
3113 		}
3114 		if ((sctx->isc_flags & IFLIB_NEED_ZERO_CSUM) && (pi->ipi_csum_flags & CSUM_IP))
3115                        ip->ip_sum = 0;
3116 
3117 		break;
3118 	}
3119 #endif
3120 #ifdef INET6
3121 	case ETHERTYPE_IPV6:
3122 	{
3123 		struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
3124 		struct tcphdr *th;
3125 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
3126 
3127 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
3128 			txq->ift_pullups++;
3129 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
3130 				return (ENOMEM);
3131 		}
3132 		th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
3133 
3134 		/* XXX-BZ this will go badly in case of ext hdrs. */
3135 		pi->ipi_ipproto = ip6->ip6_nxt;
3136 		pi->ipi_flags |= IPI_TX_IPV6;
3137 
3138 		/* TCP checksum offload may require TCP header length */
3139 		if (IS_TX_OFFLOAD6(pi)) {
3140 			if (pi->ipi_ipproto == IPPROTO_TCP) {
3141 				if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) {
3142 					txq->ift_pullups++;
3143 					if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL))
3144 						return (ENOMEM);
3145 				}
3146 				pi->ipi_tcp_hflags = th->th_flags;
3147 				pi->ipi_tcp_hlen = th->th_off << 2;
3148 				pi->ipi_tcp_seq = th->th_seq;
3149 			}
3150 			if (IS_TSO6(pi)) {
3151 				if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP))
3152 					return (ENXIO);
3153 				/*
3154 				 * TSO always requires hardware checksum offload.
3155 				 */
3156 				pi->ipi_csum_flags |= CSUM_IP6_TCP;
3157 				th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0);
3158 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3159 			}
3160 		}
3161 		break;
3162 	}
3163 #endif
3164 	default:
3165 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
3166 		pi->ipi_ip_hlen = 0;
3167 		break;
3168 	}
3169 	*mp = m;
3170 
3171 	return (0);
3172 }
3173 
3174 /*
3175  * If dodgy hardware rejects the scatter gather chain we've handed it
3176  * we'll need to remove the mbuf chain from ifsg_m[] before we can add the
3177  * m_defrag'd mbufs
3178  */
3179 static __noinline struct mbuf *
3180 iflib_remove_mbuf(iflib_txq_t txq)
3181 {
3182 	int ntxd, pidx;
3183 	struct mbuf *m, **ifsd_m;
3184 
3185 	ifsd_m = txq->ift_sds.ifsd_m;
3186 	ntxd = txq->ift_size;
3187 	pidx = txq->ift_pidx & (ntxd - 1);
3188 	ifsd_m = txq->ift_sds.ifsd_m;
3189 	m = ifsd_m[pidx];
3190 	ifsd_m[pidx] = NULL;
3191 	bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[pidx]);
3192 	if (txq->ift_sds.ifsd_tso_map != NULL)
3193 		bus_dmamap_unload(txq->ift_tso_buf_tag,
3194 		    txq->ift_sds.ifsd_tso_map[pidx]);
3195 #if MEMORY_LOGGING
3196 	txq->ift_dequeued++;
3197 #endif
3198 	return (m);
3199 }
3200 
3201 static inline caddr_t
3202 calc_next_txd(iflib_txq_t txq, int cidx, uint8_t qid)
3203 {
3204 	qidx_t size;
3205 	int ntxd;
3206 	caddr_t start, end, cur, next;
3207 
3208 	ntxd = txq->ift_size;
3209 	size = txq->ift_txd_size[qid];
3210 	start = txq->ift_ifdi[qid].idi_vaddr;
3211 
3212 	if (__predict_false(size == 0))
3213 		return (start);
3214 	cur = start + size*cidx;
3215 	end = start + size*ntxd;
3216 	next = CACHE_PTR_NEXT(cur);
3217 	return (next < end ? next : start);
3218 }
3219 
3220 /*
3221  * Pad an mbuf to ensure a minimum ethernet frame size.
3222  * min_frame_size is the frame size (less CRC) to pad the mbuf to
3223  */
3224 static __noinline int
3225 iflib_ether_pad(device_t dev, struct mbuf **m_head, uint16_t min_frame_size)
3226 {
3227 	/*
3228 	 * 18 is enough bytes to pad an ARP packet to 46 bytes, and
3229 	 * and ARP message is the smallest common payload I can think of
3230 	 */
3231 	static char pad[18];	/* just zeros */
3232 	int n;
3233 	struct mbuf *new_head;
3234 
3235 	if (!M_WRITABLE(*m_head)) {
3236 		new_head = m_dup(*m_head, M_NOWAIT);
3237 		if (new_head == NULL) {
3238 			m_freem(*m_head);
3239 			device_printf(dev, "cannot pad short frame, m_dup() failed");
3240 			DBG_COUNTER_INC(encap_pad_mbuf_fail);
3241 			DBG_COUNTER_INC(tx_frees);
3242 			return ENOMEM;
3243 		}
3244 		m_freem(*m_head);
3245 		*m_head = new_head;
3246 	}
3247 
3248 	for (n = min_frame_size - (*m_head)->m_pkthdr.len;
3249 	     n > 0; n -= sizeof(pad))
3250 		if (!m_append(*m_head, min(n, sizeof(pad)), pad))
3251 			break;
3252 
3253 	if (n > 0) {
3254 		m_freem(*m_head);
3255 		device_printf(dev, "cannot pad short frame\n");
3256 		DBG_COUNTER_INC(encap_pad_mbuf_fail);
3257 		DBG_COUNTER_INC(tx_frees);
3258 		return (ENOBUFS);
3259 	}
3260 
3261 	return 0;
3262 }
3263 
3264 static int
3265 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp)
3266 {
3267 	if_ctx_t		ctx;
3268 	if_shared_ctx_t		sctx;
3269 	if_softc_ctx_t		scctx;
3270 	bus_dma_tag_t		buf_tag;
3271 	bus_dma_segment_t	*segs;
3272 	struct mbuf		*m_head, **ifsd_m;
3273 	void			*next_txd;
3274 	bus_dmamap_t		map;
3275 	struct if_pkt_info	pi;
3276 	int remap = 0;
3277 	int err, nsegs, ndesc, max_segs, pidx, cidx, next, ntxd;
3278 
3279 	ctx = txq->ift_ctx;
3280 	sctx = ctx->ifc_sctx;
3281 	scctx = &ctx->ifc_softc_ctx;
3282 	segs = txq->ift_segs;
3283 	ntxd = txq->ift_size;
3284 	m_head = *m_headp;
3285 	map = NULL;
3286 
3287 	/*
3288 	 * If we're doing TSO the next descriptor to clean may be quite far ahead
3289 	 */
3290 	cidx = txq->ift_cidx;
3291 	pidx = txq->ift_pidx;
3292 	if (ctx->ifc_flags & IFC_PREFETCH) {
3293 		next = (cidx + CACHE_PTR_INCREMENT) & (ntxd-1);
3294 		if (!(ctx->ifc_flags & IFLIB_HAS_TXCQ)) {
3295 			next_txd = calc_next_txd(txq, cidx, 0);
3296 			prefetch(next_txd);
3297 		}
3298 
3299 		/* prefetch the next cache line of mbuf pointers and flags */
3300 		prefetch(&txq->ift_sds.ifsd_m[next]);
3301 		prefetch(&txq->ift_sds.ifsd_map[next]);
3302 		next = (cidx + CACHE_LINE_SIZE) & (ntxd-1);
3303 	}
3304 	map = txq->ift_sds.ifsd_map[pidx];
3305 	ifsd_m = txq->ift_sds.ifsd_m;
3306 
3307 	if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
3308 		buf_tag = txq->ift_tso_buf_tag;
3309 		max_segs = scctx->isc_tx_tso_segments_max;
3310 		map = txq->ift_sds.ifsd_tso_map[pidx];
3311 		MPASS(buf_tag != NULL);
3312 		MPASS(max_segs > 0);
3313 	} else {
3314 		buf_tag = txq->ift_buf_tag;
3315 		max_segs = scctx->isc_tx_nsegments;
3316 		map = txq->ift_sds.ifsd_map[pidx];
3317 	}
3318 	if ((sctx->isc_flags & IFLIB_NEED_ETHER_PAD) &&
3319 	    __predict_false(m_head->m_pkthdr.len < scctx->isc_min_frame_size)) {
3320 		err = iflib_ether_pad(ctx->ifc_dev, m_headp, scctx->isc_min_frame_size);
3321 		if (err) {
3322 			DBG_COUNTER_INC(encap_txd_encap_fail);
3323 			return err;
3324 		}
3325 	}
3326 	m_head = *m_headp;
3327 
3328 	pkt_info_zero(&pi);
3329 	pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG|M_BCAST|M_MCAST));
3330 	pi.ipi_pidx = pidx;
3331 	pi.ipi_qsidx = txq->ift_id;
3332 	pi.ipi_len = m_head->m_pkthdr.len;
3333 	pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags;
3334 	pi.ipi_vtag = M_HAS_VLANTAG(m_head) ? m_head->m_pkthdr.ether_vtag : 0;
3335 
3336 	/* deliberate bitwise OR to make one condition */
3337 	if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) {
3338 		if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0)) {
3339 			DBG_COUNTER_INC(encap_txd_encap_fail);
3340 			return (err);
3341 		}
3342 		m_head = *m_headp;
3343 	}
3344 
3345 retry:
3346 	err = bus_dmamap_load_mbuf_sg(buf_tag, map, m_head, segs, &nsegs,
3347 	    BUS_DMA_NOWAIT);
3348 defrag:
3349 	if (__predict_false(err)) {
3350 		switch (err) {
3351 		case EFBIG:
3352 			/* try collapse once and defrag once */
3353 			if (remap == 0) {
3354 				m_head = m_collapse(*m_headp, M_NOWAIT, max_segs);
3355 				/* try defrag if collapsing fails */
3356 				if (m_head == NULL)
3357 					remap++;
3358 			}
3359 			if (remap == 1) {
3360 				txq->ift_mbuf_defrag++;
3361 				m_head = m_defrag(*m_headp, M_NOWAIT);
3362 			}
3363 			/*
3364 			 * remap should never be >1 unless bus_dmamap_load_mbuf_sg
3365 			 * failed to map an mbuf that was run through m_defrag
3366 			 */
3367 			MPASS(remap <= 1);
3368 			if (__predict_false(m_head == NULL || remap > 1))
3369 				goto defrag_failed;
3370 			remap++;
3371 			*m_headp = m_head;
3372 			goto retry;
3373 			break;
3374 		case ENOMEM:
3375 			txq->ift_no_tx_dma_setup++;
3376 			break;
3377 		default:
3378 			txq->ift_no_tx_dma_setup++;
3379 			m_freem(*m_headp);
3380 			DBG_COUNTER_INC(tx_frees);
3381 			*m_headp = NULL;
3382 			break;
3383 		}
3384 		txq->ift_map_failed++;
3385 		DBG_COUNTER_INC(encap_load_mbuf_fail);
3386 		DBG_COUNTER_INC(encap_txd_encap_fail);
3387 		return (err);
3388 	}
3389 	ifsd_m[pidx] = m_head;
3390 	/*
3391 	 * XXX assumes a 1 to 1 relationship between segments and
3392 	 *        descriptors - this does not hold true on all drivers, e.g.
3393 	 *        cxgb
3394 	 */
3395 	if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) {
3396 		txq->ift_no_desc_avail++;
3397 		bus_dmamap_unload(buf_tag, map);
3398 		DBG_COUNTER_INC(encap_txq_avail_fail);
3399 		DBG_COUNTER_INC(encap_txd_encap_fail);
3400 		if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0)
3401 			GROUPTASK_ENQUEUE(&txq->ift_task);
3402 		return (ENOBUFS);
3403 	}
3404 	/*
3405 	 * On Intel cards we can greatly reduce the number of TX interrupts
3406 	 * we see by only setting report status on every Nth descriptor.
3407 	 * However, this also means that the driver will need to keep track
3408 	 * of the descriptors that RS was set on to check them for the DD bit.
3409 	 */
3410 	txq->ift_rs_pending += nsegs + 1;
3411 	if (txq->ift_rs_pending > TXQ_MAX_RS_DEFERRED(txq) ||
3412 	     iflib_no_tx_batch || (TXQ_AVAIL(txq) - nsegs) <= MAX_TX_DESC(ctx) + 2) {
3413 		pi.ipi_flags |= IPI_TX_INTR;
3414 		txq->ift_rs_pending = 0;
3415 	}
3416 
3417 	pi.ipi_segs = segs;
3418 	pi.ipi_nsegs = nsegs;
3419 
3420 	MPASS(pidx >= 0 && pidx < txq->ift_size);
3421 #ifdef PKT_DEBUG
3422 	print_pkt(&pi);
3423 #endif
3424 	if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) {
3425 		bus_dmamap_sync(buf_tag, map, BUS_DMASYNC_PREWRITE);
3426 		DBG_COUNTER_INC(tx_encap);
3427 		MPASS(pi.ipi_new_pidx < txq->ift_size);
3428 
3429 		ndesc = pi.ipi_new_pidx - pi.ipi_pidx;
3430 		if (pi.ipi_new_pidx < pi.ipi_pidx) {
3431 			ndesc += txq->ift_size;
3432 			txq->ift_gen = 1;
3433 		}
3434 		/*
3435 		 * drivers can need as many as
3436 		 * two sentinels
3437 		 */
3438 		MPASS(ndesc <= pi.ipi_nsegs + 2);
3439 		MPASS(pi.ipi_new_pidx != pidx);
3440 		MPASS(ndesc > 0);
3441 		txq->ift_in_use += ndesc;
3442 
3443 		/*
3444 		 * We update the last software descriptor again here because there may
3445 		 * be a sentinel and/or there may be more mbufs than segments
3446 		 */
3447 		txq->ift_pidx = pi.ipi_new_pidx;
3448 		txq->ift_npending += pi.ipi_ndescs;
3449 	} else {
3450 		*m_headp = m_head = iflib_remove_mbuf(txq);
3451 		if (err == EFBIG) {
3452 			txq->ift_txd_encap_efbig++;
3453 			if (remap < 2) {
3454 				remap = 1;
3455 				goto defrag;
3456 			}
3457 		}
3458 		goto defrag_failed;
3459 	}
3460 	/*
3461 	 * err can't possibly be non-zero here, so we don't neet to test it
3462 	 * to see if we need to DBG_COUNTER_INC(encap_txd_encap_fail).
3463 	 */
3464 	return (err);
3465 
3466 defrag_failed:
3467 	txq->ift_mbuf_defrag_failed++;
3468 	txq->ift_map_failed++;
3469 	m_freem(*m_headp);
3470 	DBG_COUNTER_INC(tx_frees);
3471 	*m_headp = NULL;
3472 	DBG_COUNTER_INC(encap_txd_encap_fail);
3473 	return (ENOMEM);
3474 }
3475 
3476 static void
3477 iflib_tx_desc_free(iflib_txq_t txq, int n)
3478 {
3479 	uint32_t qsize, cidx, mask, gen;
3480 	struct mbuf *m, **ifsd_m;
3481 	bool do_prefetch;
3482 
3483 	cidx = txq->ift_cidx;
3484 	gen = txq->ift_gen;
3485 	qsize = txq->ift_size;
3486 	mask = qsize-1;
3487 	ifsd_m = txq->ift_sds.ifsd_m;
3488 	do_prefetch = (txq->ift_ctx->ifc_flags & IFC_PREFETCH);
3489 
3490 	while (n-- > 0) {
3491 		if (do_prefetch) {
3492 			prefetch(ifsd_m[(cidx + 3) & mask]);
3493 			prefetch(ifsd_m[(cidx + 4) & mask]);
3494 		}
3495 		if ((m = ifsd_m[cidx]) != NULL) {
3496 			prefetch(&ifsd_m[(cidx + CACHE_PTR_INCREMENT) & mask]);
3497 			if (m->m_pkthdr.csum_flags & CSUM_TSO) {
3498 				bus_dmamap_sync(txq->ift_tso_buf_tag,
3499 				    txq->ift_sds.ifsd_tso_map[cidx],
3500 				    BUS_DMASYNC_POSTWRITE);
3501 				bus_dmamap_unload(txq->ift_tso_buf_tag,
3502 				    txq->ift_sds.ifsd_tso_map[cidx]);
3503 			} else {
3504 				bus_dmamap_sync(txq->ift_buf_tag,
3505 				    txq->ift_sds.ifsd_map[cidx],
3506 				    BUS_DMASYNC_POSTWRITE);
3507 				bus_dmamap_unload(txq->ift_buf_tag,
3508 				    txq->ift_sds.ifsd_map[cidx]);
3509 			}
3510 			/* XXX we don't support any drivers that batch packets yet */
3511 			MPASS(m->m_nextpkt == NULL);
3512 			m_freem(m);
3513 			ifsd_m[cidx] = NULL;
3514 #if MEMORY_LOGGING
3515 			txq->ift_dequeued++;
3516 #endif
3517 			DBG_COUNTER_INC(tx_frees);
3518 		}
3519 		if (__predict_false(++cidx == qsize)) {
3520 			cidx = 0;
3521 			gen = 0;
3522 		}
3523 	}
3524 	txq->ift_cidx = cidx;
3525 	txq->ift_gen = gen;
3526 }
3527 
3528 static __inline int
3529 iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh)
3530 {
3531 	int reclaim;
3532 	if_ctx_t ctx = txq->ift_ctx;
3533 
3534 	KASSERT(thresh >= 0, ("invalid threshold to reclaim"));
3535 	MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size);
3536 
3537 	/*
3538 	 * Need a rate-limiting check so that this isn't called every time
3539 	 */
3540 	iflib_tx_credits_update(ctx, txq);
3541 	reclaim = DESC_RECLAIMABLE(txq);
3542 
3543 	if (reclaim <= thresh /* + MAX_TX_DESC(txq->ift_ctx) */) {
3544 #ifdef INVARIANTS
3545 		if (iflib_verbose_debug) {
3546 			printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __FUNCTION__,
3547 			       txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments,
3548 			       reclaim, thresh);
3549 
3550 		}
3551 #endif
3552 		return (0);
3553 	}
3554 	iflib_tx_desc_free(txq, reclaim);
3555 	txq->ift_cleaned += reclaim;
3556 	txq->ift_in_use -= reclaim;
3557 
3558 	return (reclaim);
3559 }
3560 
3561 static struct mbuf **
3562 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset, int remaining)
3563 {
3564 	int next, size;
3565 	struct mbuf **items;
3566 
3567 	size = r->size;
3568 	next = (cidx + CACHE_PTR_INCREMENT) & (size-1);
3569 	items = __DEVOLATILE(struct mbuf **, &r->items[0]);
3570 
3571 	prefetch(items[(cidx + offset) & (size-1)]);
3572 	if (remaining > 1) {
3573 		prefetch2cachelines(&items[next]);
3574 		prefetch2cachelines(items[(cidx + offset + 1) & (size-1)]);
3575 		prefetch2cachelines(items[(cidx + offset + 2) & (size-1)]);
3576 		prefetch2cachelines(items[(cidx + offset + 3) & (size-1)]);
3577 	}
3578 	return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (size-1)]));
3579 }
3580 
3581 static void
3582 iflib_txq_check_drain(iflib_txq_t txq, int budget)
3583 {
3584 
3585 	ifmp_ring_check_drainage(txq->ift_br, budget);
3586 }
3587 
3588 static uint32_t
3589 iflib_txq_can_drain(struct ifmp_ring *r)
3590 {
3591 	iflib_txq_t txq = r->cookie;
3592 	if_ctx_t ctx = txq->ift_ctx;
3593 
3594 	if (TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2)
3595 		return (1);
3596 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3597 	    BUS_DMASYNC_POSTREAD);
3598 	return (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id,
3599 	    false));
3600 }
3601 
3602 static uint32_t
3603 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3604 {
3605 	iflib_txq_t txq = r->cookie;
3606 	if_ctx_t ctx = txq->ift_ctx;
3607 	if_t ifp = ctx->ifc_ifp;
3608 	struct mbuf *m, **mp;
3609 	int avail, bytes_sent, consumed, count, err, i, in_use_prev;
3610 	int mcast_sent, pkt_sent, reclaimed, txq_avail;
3611 	bool do_prefetch, rang, ring;
3612 
3613 	if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) ||
3614 			    !LINK_ACTIVE(ctx))) {
3615 		DBG_COUNTER_INC(txq_drain_notready);
3616 		return (0);
3617 	}
3618 	reclaimed = iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
3619 	rang = iflib_txd_db_check(ctx, txq, reclaimed, txq->ift_in_use);
3620 	avail = IDXDIFF(pidx, cidx, r->size);
3621 	if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) {
3622 		DBG_COUNTER_INC(txq_drain_flushing);
3623 		for (i = 0; i < avail; i++) {
3624 			if (__predict_true(r->items[(cidx + i) & (r->size-1)] != (void *)txq))
3625 				m_free(r->items[(cidx + i) & (r->size-1)]);
3626 			r->items[(cidx + i) & (r->size-1)] = NULL;
3627 		}
3628 		return (avail);
3629 	}
3630 
3631 	if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3632 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3633 		CALLOUT_LOCK(txq);
3634 		callout_stop(&txq->ift_timer);
3635 		CALLOUT_UNLOCK(txq);
3636 		DBG_COUNTER_INC(txq_drain_oactive);
3637 		return (0);
3638 	}
3639 	if (reclaimed)
3640 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3641 	consumed = mcast_sent = bytes_sent = pkt_sent = 0;
3642 	count = MIN(avail, TX_BATCH_SIZE);
3643 #ifdef INVARIANTS
3644 	if (iflib_verbose_debug)
3645 		printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __FUNCTION__,
3646 		       avail, ctx->ifc_flags, TXQ_AVAIL(txq));
3647 #endif
3648 	do_prefetch = (ctx->ifc_flags & IFC_PREFETCH);
3649 	txq_avail = TXQ_AVAIL(txq);
3650 	err = 0;
3651 	for (i = 0; i < count && txq_avail > MAX_TX_DESC(ctx) + 2; i++) {
3652 		int rem = do_prefetch ? count - i : 0;
3653 
3654 		mp = _ring_peek_one(r, cidx, i, rem);
3655 		MPASS(mp != NULL && *mp != NULL);
3656 		if (__predict_false(*mp == (struct mbuf *)txq)) {
3657 			consumed++;
3658 			continue;
3659 		}
3660 		in_use_prev = txq->ift_in_use;
3661 		err = iflib_encap(txq, mp);
3662 		if (__predict_false(err)) {
3663 			/* no room - bail out */
3664 			if (err == ENOBUFS)
3665 				break;
3666 			consumed++;
3667 			/* we can't send this packet - skip it */
3668 			continue;
3669 		}
3670 		consumed++;
3671 		pkt_sent++;
3672 		m = *mp;
3673 		DBG_COUNTER_INC(tx_sent);
3674 		bytes_sent += m->m_pkthdr.len;
3675 		mcast_sent += !!(m->m_flags & M_MCAST);
3676 		txq_avail = TXQ_AVAIL(txq);
3677 
3678 		txq->ift_db_pending += (txq->ift_in_use - in_use_prev);
3679 		ETHER_BPF_MTAP(ifp, m);
3680 		if (__predict_false(!(ifp->if_drv_flags & IFF_DRV_RUNNING)))
3681 			break;
3682 		rang = iflib_txd_db_check(ctx, txq, false, in_use_prev);
3683 	}
3684 
3685 	/* deliberate use of bitwise or to avoid gratuitous short-circuit */
3686 	ring = rang ? false  : (iflib_min_tx_latency | err) || (TXQ_AVAIL(txq) < MAX_TX_DESC(ctx));
3687 	iflib_txd_db_check(ctx, txq, ring, txq->ift_in_use);
3688 	if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
3689 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
3690 	if (mcast_sent)
3691 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
3692 #ifdef INVARIANTS
3693 	if (iflib_verbose_debug)
3694 		printf("consumed=%d\n", consumed);
3695 #endif
3696 	return (consumed);
3697 }
3698 
3699 static uint32_t
3700 iflib_txq_drain_always(struct ifmp_ring *r)
3701 {
3702 	return (1);
3703 }
3704 
3705 static uint32_t
3706 iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3707 {
3708 	int i, avail;
3709 	struct mbuf **mp;
3710 	iflib_txq_t txq;
3711 
3712 	txq = r->cookie;
3713 
3714 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3715 	CALLOUT_LOCK(txq);
3716 	callout_stop(&txq->ift_timer);
3717 	CALLOUT_UNLOCK(txq);
3718 
3719 	avail = IDXDIFF(pidx, cidx, r->size);
3720 	for (i = 0; i < avail; i++) {
3721 		mp = _ring_peek_one(r, cidx, i, avail - i);
3722 		if (__predict_false(*mp == (struct mbuf *)txq))
3723 			continue;
3724 		m_freem(*mp);
3725 		DBG_COUNTER_INC(tx_frees);
3726 	}
3727 	MPASS(ifmp_ring_is_stalled(r) == 0);
3728 	return (avail);
3729 }
3730 
3731 static void
3732 iflib_ifmp_purge(iflib_txq_t txq)
3733 {
3734 	struct ifmp_ring *r;
3735 
3736 	r = txq->ift_br;
3737 	r->drain = iflib_txq_drain_free;
3738 	r->can_drain = iflib_txq_drain_always;
3739 
3740 	ifmp_ring_check_drainage(r, r->size);
3741 
3742 	r->drain = iflib_txq_drain;
3743 	r->can_drain = iflib_txq_can_drain;
3744 }
3745 
3746 static void
3747 _task_fn_tx(void *context)
3748 {
3749 	iflib_txq_t txq = context;
3750 	if_ctx_t ctx = txq->ift_ctx;
3751 #if defined(ALTQ) || defined(DEV_NETMAP)
3752 	if_t ifp = ctx->ifc_ifp;
3753 #endif
3754 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
3755 
3756 #ifdef IFLIB_DIAGNOSTICS
3757 	txq->ift_cpu_exec_count[curcpu]++;
3758 #endif
3759 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3760 		return;
3761 #ifdef DEV_NETMAP
3762 	if (if_getcapenable(ifp) & IFCAP_NETMAP) {
3763 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3764 		    BUS_DMASYNC_POSTREAD);
3765 		if (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false))
3766 			netmap_tx_irq(ifp, txq->ift_id);
3767 		if (ctx->ifc_flags & IFC_LEGACY)
3768 			IFDI_INTR_ENABLE(ctx);
3769 		else
3770 			IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3771 		return;
3772 	}
3773 #endif
3774 #ifdef ALTQ
3775 	if (ALTQ_IS_ENABLED(&ifp->if_snd))
3776 		iflib_altq_if_start(ifp);
3777 #endif
3778 	if (txq->ift_db_pending)
3779 		ifmp_ring_enqueue(txq->ift_br, (void **)&txq, 1, TX_BATCH_SIZE, abdicate);
3780 	else if (!abdicate)
3781 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3782 	/*
3783 	 * When abdicating, we always need to check drainage, not just when we don't enqueue
3784 	 */
3785 	if (abdicate)
3786 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3787 	if (ctx->ifc_flags & IFC_LEGACY)
3788 		IFDI_INTR_ENABLE(ctx);
3789 	else
3790 		IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3791 }
3792 
3793 static void
3794 _task_fn_rx(void *context)
3795 {
3796 	iflib_rxq_t rxq = context;
3797 	if_ctx_t ctx = rxq->ifr_ctx;
3798 	uint8_t more;
3799 	uint16_t budget;
3800 #ifdef DEV_NETMAP
3801 	u_int work = 0;
3802 	int nmirq;
3803 #endif
3804 
3805 #ifdef IFLIB_DIAGNOSTICS
3806 	rxq->ifr_cpu_exec_count[curcpu]++;
3807 #endif
3808 	DBG_COUNTER_INC(task_fn_rxs);
3809 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3810 		return;
3811 #ifdef DEV_NETMAP
3812 	nmirq = netmap_rx_irq(ctx->ifc_ifp, rxq->ifr_id, &work);
3813 	if (nmirq != NM_IRQ_PASS) {
3814 		more = (nmirq == NM_IRQ_RESCHED) ? IFLIB_RXEOF_MORE : 0;
3815 		goto skip_rxeof;
3816 	}
3817 #endif
3818 	budget = ctx->ifc_sysctl_rx_budget;
3819 	if (budget == 0)
3820 		budget = 16;	/* XXX */
3821 	more = iflib_rxeof(rxq, budget);
3822 #ifdef DEV_NETMAP
3823 skip_rxeof:
3824 #endif
3825 	if ((more & IFLIB_RXEOF_MORE) == 0) {
3826 		if (ctx->ifc_flags & IFC_LEGACY)
3827 			IFDI_INTR_ENABLE(ctx);
3828 		else
3829 			IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
3830 		DBG_COUNTER_INC(rx_intr_enables);
3831 	}
3832 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3833 		return;
3834 
3835 	if (more & IFLIB_RXEOF_MORE)
3836 		GROUPTASK_ENQUEUE(&rxq->ifr_task);
3837 	else if (more & IFLIB_RXEOF_EMPTY)
3838 		callout_reset_curcpu(&rxq->ifr_watchdog, 1, &_task_fn_rx_watchdog, rxq);
3839 }
3840 
3841 static void
3842 _task_fn_admin(void *context)
3843 {
3844 	if_ctx_t ctx = context;
3845 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
3846 	iflib_txq_t txq;
3847 	int i;
3848 	bool oactive, running, do_reset, do_watchdog, in_detach;
3849 	uint32_t reset_on = hz / 2;
3850 
3851 	STATE_LOCK(ctx);
3852 	running = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING);
3853 	oactive = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE);
3854 	do_reset = (ctx->ifc_flags & IFC_DO_RESET);
3855 	do_watchdog = (ctx->ifc_flags & IFC_DO_WATCHDOG);
3856 	in_detach = (ctx->ifc_flags & IFC_IN_DETACH);
3857 	ctx->ifc_flags &= ~(IFC_DO_RESET|IFC_DO_WATCHDOG);
3858 	STATE_UNLOCK(ctx);
3859 
3860 	if ((!running && !oactive) && !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
3861 		return;
3862 	if (in_detach)
3863 		return;
3864 
3865 	CTX_LOCK(ctx);
3866 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3867 		CALLOUT_LOCK(txq);
3868 		callout_stop(&txq->ift_timer);
3869 		CALLOUT_UNLOCK(txq);
3870 	}
3871 	if (do_watchdog) {
3872 		ctx->ifc_watchdog_events++;
3873 		IFDI_WATCHDOG_RESET(ctx);
3874 	}
3875 	IFDI_UPDATE_ADMIN_STATUS(ctx);
3876 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3877 #ifdef DEV_NETMAP
3878 		reset_on = hz / 2;
3879 		if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP)
3880 			iflib_netmap_timer_adjust(ctx, txq, &reset_on);
3881 #endif
3882 		callout_reset_on(&txq->ift_timer, reset_on, iflib_timer, txq, txq->ift_timer.c_cpu);
3883 	}
3884 	IFDI_LINK_INTR_ENABLE(ctx);
3885 	if (do_reset)
3886 		iflib_if_init_locked(ctx);
3887 	CTX_UNLOCK(ctx);
3888 
3889 	if (LINK_ACTIVE(ctx) == 0)
3890 		return;
3891 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3892 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
3893 }
3894 
3895 
3896 static void
3897 _task_fn_iov(void *context)
3898 {
3899 	if_ctx_t ctx = context;
3900 
3901 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING) &&
3902 	    !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
3903 		return;
3904 
3905 	CTX_LOCK(ctx);
3906 	IFDI_VFLR_HANDLE(ctx);
3907 	CTX_UNLOCK(ctx);
3908 }
3909 
3910 static int
3911 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)
3912 {
3913 	int err;
3914 	if_int_delay_info_t info;
3915 	if_ctx_t ctx;
3916 
3917 	info = (if_int_delay_info_t)arg1;
3918 	ctx = info->iidi_ctx;
3919 	info->iidi_req = req;
3920 	info->iidi_oidp = oidp;
3921 	CTX_LOCK(ctx);
3922 	err = IFDI_SYSCTL_INT_DELAY(ctx, info);
3923 	CTX_UNLOCK(ctx);
3924 	return (err);
3925 }
3926 
3927 /*********************************************************************
3928  *
3929  *  IFNET FUNCTIONS
3930  *
3931  **********************************************************************/
3932 
3933 static void
3934 iflib_if_init_locked(if_ctx_t ctx)
3935 {
3936 	iflib_stop(ctx);
3937 	iflib_init_locked(ctx);
3938 }
3939 
3940 
3941 static void
3942 iflib_if_init(void *arg)
3943 {
3944 	if_ctx_t ctx = arg;
3945 
3946 	CTX_LOCK(ctx);
3947 	iflib_if_init_locked(ctx);
3948 	CTX_UNLOCK(ctx);
3949 }
3950 
3951 static int
3952 iflib_if_transmit(if_t ifp, struct mbuf *m)
3953 {
3954 	if_ctx_t	ctx = if_getsoftc(ifp);
3955 
3956 	iflib_txq_t txq;
3957 	int err, qidx;
3958 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
3959 
3960 	if (__predict_false((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) {
3961 		DBG_COUNTER_INC(tx_frees);
3962 		m_freem(m);
3963 		return (ENETDOWN);
3964 	}
3965 
3966 	MPASS(m->m_nextpkt == NULL);
3967 	/* ALTQ-enabled interfaces always use queue 0. */
3968 	qidx = 0;
3969 	if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m) && !ALTQ_IS_ENABLED(&ifp->if_snd))
3970 		qidx = QIDX(ctx, m);
3971 	/*
3972 	 * XXX calculate buf_ring based on flowid (divvy up bits?)
3973 	 */
3974 	txq = &ctx->ifc_txqs[qidx];
3975 
3976 #ifdef DRIVER_BACKPRESSURE
3977 	if (txq->ift_closed) {
3978 		while (m != NULL) {
3979 			next = m->m_nextpkt;
3980 			m->m_nextpkt = NULL;
3981 			m_freem(m);
3982 			DBG_COUNTER_INC(tx_frees);
3983 			m = next;
3984 		}
3985 		return (ENOBUFS);
3986 	}
3987 #endif
3988 #ifdef notyet
3989 	qidx = count = 0;
3990 	mp = marr;
3991 	next = m;
3992 	do {
3993 		count++;
3994 		next = next->m_nextpkt;
3995 	} while (next != NULL);
3996 
3997 	if (count > nitems(marr))
3998 		if ((mp = malloc(count*sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) {
3999 			/* XXX check nextpkt */
4000 			m_freem(m);
4001 			/* XXX simplify for now */
4002 			DBG_COUNTER_INC(tx_frees);
4003 			return (ENOBUFS);
4004 		}
4005 	for (next = m, i = 0; next != NULL; i++) {
4006 		mp[i] = next;
4007 		next = next->m_nextpkt;
4008 		mp[i]->m_nextpkt = NULL;
4009 	}
4010 #endif
4011 	DBG_COUNTER_INC(tx_seen);
4012 	err = ifmp_ring_enqueue(txq->ift_br, (void **)&m, 1, TX_BATCH_SIZE, abdicate);
4013 
4014 	if (abdicate)
4015 		GROUPTASK_ENQUEUE(&txq->ift_task);
4016  	if (err) {
4017 		if (!abdicate)
4018 			GROUPTASK_ENQUEUE(&txq->ift_task);
4019 		/* support forthcoming later */
4020 #ifdef DRIVER_BACKPRESSURE
4021 		txq->ift_closed = TRUE;
4022 #endif
4023 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
4024 		m_freem(m);
4025 		DBG_COUNTER_INC(tx_frees);
4026 	}
4027 
4028 	return (err);
4029 }
4030 
4031 #ifdef ALTQ
4032 /*
4033  * The overall approach to integrating iflib with ALTQ is to continue to use
4034  * the iflib mp_ring machinery between the ALTQ queue(s) and the hardware
4035  * ring.  Technically, when using ALTQ, queueing to an intermediate mp_ring
4036  * is redundant/unnecessary, but doing so minimizes the amount of
4037  * ALTQ-specific code required in iflib.  It is assumed that the overhead of
4038  * redundantly queueing to an intermediate mp_ring is swamped by the
4039  * performance limitations inherent in using ALTQ.
4040  *
4041  * When ALTQ support is compiled in, all iflib drivers will use a transmit
4042  * routine, iflib_altq_if_transmit(), that checks if ALTQ is enabled for the
4043  * given interface.  If ALTQ is enabled for an interface, then all
4044  * transmitted packets for that interface will be submitted to the ALTQ
4045  * subsystem via IFQ_ENQUEUE().  We don't use the legacy if_transmit()
4046  * implementation because it uses IFQ_HANDOFF(), which will duplicatively
4047  * update stats that the iflib machinery handles, and which is sensitve to
4048  * the disused IFF_DRV_OACTIVE flag.  Additionally, iflib_altq_if_start()
4049  * will be installed as the start routine for use by ALTQ facilities that
4050  * need to trigger queue drains on a scheduled basis.
4051  *
4052  */
4053 static void
4054 iflib_altq_if_start(if_t ifp)
4055 {
4056 	struct ifaltq *ifq = &ifp->if_snd;
4057 	struct mbuf *m;
4058 
4059 	IFQ_LOCK(ifq);
4060 	IFQ_DEQUEUE_NOLOCK(ifq, m);
4061 	while (m != NULL) {
4062 		iflib_if_transmit(ifp, m);
4063 		IFQ_DEQUEUE_NOLOCK(ifq, m);
4064 	}
4065 	IFQ_UNLOCK(ifq);
4066 }
4067 
4068 static int
4069 iflib_altq_if_transmit(if_t ifp, struct mbuf *m)
4070 {
4071 	int err;
4072 
4073 	if (ALTQ_IS_ENABLED(&ifp->if_snd)) {
4074 		IFQ_ENQUEUE(&ifp->if_snd, m, err);
4075 		if (err == 0)
4076 			iflib_altq_if_start(ifp);
4077 	} else
4078 		err = iflib_if_transmit(ifp, m);
4079 
4080 	return (err);
4081 }
4082 #endif /* ALTQ */
4083 
4084 static void
4085 iflib_if_qflush(if_t ifp)
4086 {
4087 	if_ctx_t ctx = if_getsoftc(ifp);
4088 	iflib_txq_t txq = ctx->ifc_txqs;
4089 	int i;
4090 
4091 	STATE_LOCK(ctx);
4092 	ctx->ifc_flags |= IFC_QFLUSH;
4093 	STATE_UNLOCK(ctx);
4094 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
4095 		while (!(ifmp_ring_is_idle(txq->ift_br) || ifmp_ring_is_stalled(txq->ift_br)))
4096 			iflib_txq_check_drain(txq, 0);
4097 	STATE_LOCK(ctx);
4098 	ctx->ifc_flags &= ~IFC_QFLUSH;
4099 	STATE_UNLOCK(ctx);
4100 
4101 	/*
4102 	 * When ALTQ is enabled, this will also take care of purging the
4103 	 * ALTQ queue(s).
4104 	 */
4105 	if_qflush(ifp);
4106 }
4107 
4108 
4109 #define IFCAP_FLAGS (IFCAP_HWCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \
4110 		     IFCAP_TSO | IFCAP_VLAN_HWTAGGING | IFCAP_HWSTATS | \
4111 		     IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | \
4112 		     IFCAP_VLAN_HWTSO | IFCAP_VLAN_HWCSUM | IFCAP_NOMAP)
4113 
4114 static int
4115 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data)
4116 {
4117 	if_ctx_t ctx = if_getsoftc(ifp);
4118 	struct ifreq	*ifr = (struct ifreq *)data;
4119 #if defined(INET) || defined(INET6)
4120 	struct ifaddr	*ifa = (struct ifaddr *)data;
4121 #endif
4122 	bool		avoid_reset = false;
4123 	int		err = 0, reinit = 0, bits;
4124 
4125 	switch (command) {
4126 	case SIOCSIFADDR:
4127 #ifdef INET
4128 		if (ifa->ifa_addr->sa_family == AF_INET)
4129 			avoid_reset = true;
4130 #endif
4131 #ifdef INET6
4132 		if (ifa->ifa_addr->sa_family == AF_INET6)
4133 			avoid_reset = true;
4134 #endif
4135 		/*
4136 		** Calling init results in link renegotiation,
4137 		** so we avoid doing it when possible.
4138 		*/
4139 		if (avoid_reset) {
4140 			if_setflagbits(ifp, IFF_UP,0);
4141 			if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
4142 				reinit = 1;
4143 #ifdef INET
4144 			if (!(if_getflags(ifp) & IFF_NOARP))
4145 				arp_ifinit(ifp, ifa);
4146 #endif
4147 		} else
4148 			err = ether_ioctl(ifp, command, data);
4149 		break;
4150 	case SIOCSIFMTU:
4151 		CTX_LOCK(ctx);
4152 		if (ifr->ifr_mtu == if_getmtu(ifp)) {
4153 			CTX_UNLOCK(ctx);
4154 			break;
4155 		}
4156 		bits = if_getdrvflags(ifp);
4157 		/* stop the driver and free any clusters before proceeding */
4158 		iflib_stop(ctx);
4159 
4160 		if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) {
4161 			STATE_LOCK(ctx);
4162 			if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size)
4163 				ctx->ifc_flags |= IFC_MULTISEG;
4164 			else
4165 				ctx->ifc_flags &= ~IFC_MULTISEG;
4166 			STATE_UNLOCK(ctx);
4167 			err = if_setmtu(ifp, ifr->ifr_mtu);
4168 		}
4169 		iflib_init_locked(ctx);
4170 		STATE_LOCK(ctx);
4171 		if_setdrvflags(ifp, bits);
4172 		STATE_UNLOCK(ctx);
4173 		CTX_UNLOCK(ctx);
4174 		break;
4175 	case SIOCSIFFLAGS:
4176 		CTX_LOCK(ctx);
4177 		if (if_getflags(ifp) & IFF_UP) {
4178 			if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4179 				if ((if_getflags(ifp) ^ ctx->ifc_if_flags) &
4180 				    (IFF_PROMISC | IFF_ALLMULTI)) {
4181 					err = IFDI_PROMISC_SET(ctx, if_getflags(ifp));
4182 				}
4183 			} else
4184 				reinit = 1;
4185 		} else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4186 			iflib_stop(ctx);
4187 		}
4188 		ctx->ifc_if_flags = if_getflags(ifp);
4189 		CTX_UNLOCK(ctx);
4190 		break;
4191 	case SIOCADDMULTI:
4192 	case SIOCDELMULTI:
4193 		if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4194 			CTX_LOCK(ctx);
4195 			IFDI_INTR_DISABLE(ctx);
4196 			IFDI_MULTI_SET(ctx);
4197 			IFDI_INTR_ENABLE(ctx);
4198 			CTX_UNLOCK(ctx);
4199 		}
4200 		break;
4201 	case SIOCSIFMEDIA:
4202 		CTX_LOCK(ctx);
4203 		IFDI_MEDIA_SET(ctx);
4204 		CTX_UNLOCK(ctx);
4205 		/* FALLTHROUGH */
4206 	case SIOCGIFMEDIA:
4207 	case SIOCGIFXMEDIA:
4208 		err = ifmedia_ioctl(ifp, ifr, ctx->ifc_mediap, command);
4209 		break;
4210 	case SIOCGI2C:
4211 	{
4212 		struct ifi2creq i2c;
4213 
4214 		err = copyin(ifr_data_get_ptr(ifr), &i2c, sizeof(i2c));
4215 		if (err != 0)
4216 			break;
4217 		if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) {
4218 			err = EINVAL;
4219 			break;
4220 		}
4221 		if (i2c.len > sizeof(i2c.data)) {
4222 			err = EINVAL;
4223 			break;
4224 		}
4225 
4226 		if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0)
4227 			err = copyout(&i2c, ifr_data_get_ptr(ifr),
4228 			    sizeof(i2c));
4229 		break;
4230 	}
4231 	case SIOCSIFCAP:
4232 	{
4233 		int mask, setmask, oldmask;
4234 
4235 		oldmask = if_getcapenable(ifp);
4236 		mask = ifr->ifr_reqcap ^ oldmask;
4237 		mask &= ctx->ifc_softc_ctx.isc_capabilities | IFCAP_NOMAP;
4238 		setmask = 0;
4239 #ifdef TCP_OFFLOAD
4240 		setmask |= mask & (IFCAP_TOE4|IFCAP_TOE6);
4241 #endif
4242 		setmask |= (mask & IFCAP_FLAGS);
4243 		setmask |= (mask & IFCAP_WOL);
4244 
4245 		/*
4246 		 * If any RX csum has changed, change all the ones that
4247 		 * are supported by the driver.
4248 		 */
4249 		if (setmask & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6)) {
4250 			setmask |= ctx->ifc_softc_ctx.isc_capabilities &
4251 			    (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6);
4252 		}
4253 
4254 		/*
4255 		 * want to ensure that traffic has stopped before we change any of the flags
4256 		 */
4257 		if (setmask) {
4258 			CTX_LOCK(ctx);
4259 			bits = if_getdrvflags(ifp);
4260 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4261 				iflib_stop(ctx);
4262 			STATE_LOCK(ctx);
4263 			if_togglecapenable(ifp, setmask);
4264 			STATE_UNLOCK(ctx);
4265 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4266 				iflib_init_locked(ctx);
4267 			STATE_LOCK(ctx);
4268 			if_setdrvflags(ifp, bits);
4269 			STATE_UNLOCK(ctx);
4270 			CTX_UNLOCK(ctx);
4271 		}
4272 		if_vlancap(ifp);
4273 		break;
4274 	}
4275 	case SIOCGPRIVATE_0:
4276 	case SIOCSDRVSPEC:
4277 	case SIOCGDRVSPEC:
4278 		CTX_LOCK(ctx);
4279 		err = IFDI_PRIV_IOCTL(ctx, command, data);
4280 		CTX_UNLOCK(ctx);
4281 		break;
4282 	default:
4283 		err = ether_ioctl(ifp, command, data);
4284 		break;
4285 	}
4286 	if (reinit)
4287 		iflib_if_init(ctx);
4288 	return (err);
4289 }
4290 
4291 static uint64_t
4292 iflib_if_get_counter(if_t ifp, ift_counter cnt)
4293 {
4294 	if_ctx_t ctx = if_getsoftc(ifp);
4295 
4296 	return (IFDI_GET_COUNTER(ctx, cnt));
4297 }
4298 
4299 /*********************************************************************
4300  *
4301  *  OTHER FUNCTIONS EXPORTED TO THE STACK
4302  *
4303  **********************************************************************/
4304 
4305 static void
4306 iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag)
4307 {
4308 	if_ctx_t ctx = if_getsoftc(ifp);
4309 
4310 	if ((void *)ctx != arg)
4311 		return;
4312 
4313 	if ((vtag == 0) || (vtag > 4095))
4314 		return;
4315 
4316 	if (iflib_in_detach(ctx))
4317 		return;
4318 
4319 	CTX_LOCK(ctx);
4320 	/* Driver may need all untagged packets to be flushed */
4321 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4322 		iflib_stop(ctx);
4323 	IFDI_VLAN_REGISTER(ctx, vtag);
4324 	/* Re-init to load the changes, if required */
4325 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4326 		iflib_init_locked(ctx);
4327 	CTX_UNLOCK(ctx);
4328 }
4329 
4330 static void
4331 iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag)
4332 {
4333 	if_ctx_t ctx = if_getsoftc(ifp);
4334 
4335 	if ((void *)ctx != arg)
4336 		return;
4337 
4338 	if ((vtag == 0) || (vtag > 4095))
4339 		return;
4340 
4341 	CTX_LOCK(ctx);
4342 	/* Driver may need all tagged packets to be flushed */
4343 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4344 		iflib_stop(ctx);
4345 	IFDI_VLAN_UNREGISTER(ctx, vtag);
4346 	/* Re-init to load the changes, if required */
4347 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4348 		iflib_init_locked(ctx);
4349 	CTX_UNLOCK(ctx);
4350 }
4351 
4352 static void
4353 iflib_led_func(void *arg, int onoff)
4354 {
4355 	if_ctx_t ctx = arg;
4356 
4357 	CTX_LOCK(ctx);
4358 	IFDI_LED_FUNC(ctx, onoff);
4359 	CTX_UNLOCK(ctx);
4360 }
4361 
4362 /*********************************************************************
4363  *
4364  *  BUS FUNCTION DEFINITIONS
4365  *
4366  **********************************************************************/
4367 
4368 int
4369 iflib_device_probe(device_t dev)
4370 {
4371 	const pci_vendor_info_t *ent;
4372 	if_shared_ctx_t sctx;
4373 	uint16_t pci_device_id, pci_rev_id, pci_subdevice_id, pci_subvendor_id;
4374 	uint16_t pci_vendor_id;
4375 
4376 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4377 		return (ENOTSUP);
4378 
4379 	pci_vendor_id = pci_get_vendor(dev);
4380 	pci_device_id = pci_get_device(dev);
4381 	pci_subvendor_id = pci_get_subvendor(dev);
4382 	pci_subdevice_id = pci_get_subdevice(dev);
4383 	pci_rev_id = pci_get_revid(dev);
4384 	if (sctx->isc_parse_devinfo != NULL)
4385 		sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id);
4386 
4387 	ent = sctx->isc_vendor_info;
4388 	while (ent->pvi_vendor_id != 0) {
4389 		if (pci_vendor_id != ent->pvi_vendor_id) {
4390 			ent++;
4391 			continue;
4392 		}
4393 		if ((pci_device_id == ent->pvi_device_id) &&
4394 		    ((pci_subvendor_id == ent->pvi_subvendor_id) ||
4395 		     (ent->pvi_subvendor_id == 0)) &&
4396 		    ((pci_subdevice_id == ent->pvi_subdevice_id) ||
4397 		     (ent->pvi_subdevice_id == 0)) &&
4398 		    ((pci_rev_id == ent->pvi_rev_id) ||
4399 		     (ent->pvi_rev_id == 0))) {
4400 
4401 			device_set_desc_copy(dev, ent->pvi_name);
4402 			/* this needs to be changed to zero if the bus probing code
4403 			 * ever stops re-probing on best match because the sctx
4404 			 * may have its values over written by register calls
4405 			 * in subsequent probes
4406 			 */
4407 			return (BUS_PROBE_DEFAULT);
4408 		}
4409 		ent++;
4410 	}
4411 	return (ENXIO);
4412 }
4413 
4414 int
4415 iflib_device_probe_vendor(device_t dev)
4416 {
4417 	int probe;
4418 
4419 	probe = iflib_device_probe(dev);
4420 	if (probe == BUS_PROBE_DEFAULT)
4421 		return (BUS_PROBE_VENDOR);
4422 	else
4423 		return (probe);
4424 }
4425 
4426 static void
4427 iflib_reset_qvalues(if_ctx_t ctx)
4428 {
4429 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4430 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4431 	device_t dev = ctx->ifc_dev;
4432 	int i;
4433 
4434 	if (ctx->ifc_sysctl_ntxqs != 0)
4435 		scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs;
4436 	if (ctx->ifc_sysctl_nrxqs != 0)
4437 		scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs;
4438 
4439 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4440 		if (ctx->ifc_sysctl_ntxds[i] != 0)
4441 			scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i];
4442 		else
4443 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4444 	}
4445 
4446 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4447 		if (ctx->ifc_sysctl_nrxds[i] != 0)
4448 			scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i];
4449 		else
4450 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4451 	}
4452 
4453 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4454 		if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) {
4455 			device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n",
4456 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]);
4457 			scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i];
4458 		}
4459 		if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) {
4460 			device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n",
4461 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]);
4462 			scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i];
4463 		}
4464 		if (!powerof2(scctx->isc_nrxd[i])) {
4465 			device_printf(dev, "nrxd%d: %d is not a power of 2 - using default value of %d\n",
4466 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_default[i]);
4467 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4468 		}
4469 	}
4470 
4471 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4472 		if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) {
4473 			device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n",
4474 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]);
4475 			scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i];
4476 		}
4477 		if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) {
4478 			device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n",
4479 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]);
4480 			scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i];
4481 		}
4482 		if (!powerof2(scctx->isc_ntxd[i])) {
4483 			device_printf(dev, "ntxd%d: %d is not a power of 2 - using default value of %d\n",
4484 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_default[i]);
4485 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4486 		}
4487 	}
4488 }
4489 
4490 static void
4491 iflib_add_pfil(if_ctx_t ctx)
4492 {
4493 	struct pfil_head *pfil;
4494 	struct pfil_head_args pa;
4495 	iflib_rxq_t rxq;
4496 	int i;
4497 
4498 	pa.pa_version = PFIL_VERSION;
4499 	pa.pa_flags = PFIL_IN;
4500 	pa.pa_type = PFIL_TYPE_ETHERNET;
4501 	pa.pa_headname = ctx->ifc_ifp->if_xname;
4502 	pfil = pfil_head_register(&pa);
4503 
4504 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4505 		rxq->pfil = pfil;
4506 	}
4507 }
4508 
4509 static void
4510 iflib_rem_pfil(if_ctx_t ctx)
4511 {
4512 	struct pfil_head *pfil;
4513 	iflib_rxq_t rxq;
4514 	int i;
4515 
4516 	rxq = ctx->ifc_rxqs;
4517 	pfil = rxq->pfil;
4518 	for (i = 0; i < NRXQSETS(ctx); i++, rxq++) {
4519 		rxq->pfil = NULL;
4520 	}
4521 	pfil_head_unregister(pfil);
4522 }
4523 
4524 static uint16_t
4525 get_ctx_core_offset(if_ctx_t ctx)
4526 {
4527 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4528 	struct cpu_offset *op;
4529 	uint16_t qc;
4530 	uint16_t ret = ctx->ifc_sysctl_core_offset;
4531 
4532 	if (ret != CORE_OFFSET_UNSPECIFIED)
4533 		return (ret);
4534 
4535 	if (ctx->ifc_sysctl_separate_txrx)
4536 		qc = scctx->isc_ntxqsets + scctx->isc_nrxqsets;
4537 	else
4538 		qc = max(scctx->isc_ntxqsets, scctx->isc_nrxqsets);
4539 
4540 	mtx_lock(&cpu_offset_mtx);
4541 	SLIST_FOREACH(op, &cpu_offsets, entries) {
4542 		if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) {
4543 			ret = op->offset;
4544 			op->offset += qc;
4545 			MPASS(op->refcount < UINT_MAX);
4546 			op->refcount++;
4547 			break;
4548 		}
4549 	}
4550 	if (ret == CORE_OFFSET_UNSPECIFIED) {
4551 		ret = 0;
4552 		op = malloc(sizeof(struct cpu_offset), M_IFLIB,
4553 		    M_NOWAIT | M_ZERO);
4554 		if (op == NULL) {
4555 			device_printf(ctx->ifc_dev,
4556 			    "allocation for cpu offset failed.\n");
4557 		} else {
4558 			op->offset = qc;
4559 			op->refcount = 1;
4560 			CPU_COPY(&ctx->ifc_cpus, &op->set);
4561 			SLIST_INSERT_HEAD(&cpu_offsets, op, entries);
4562 		}
4563 	}
4564 	mtx_unlock(&cpu_offset_mtx);
4565 
4566 	return (ret);
4567 }
4568 
4569 static void
4570 unref_ctx_core_offset(if_ctx_t ctx)
4571 {
4572 	struct cpu_offset *op, *top;
4573 
4574 	mtx_lock(&cpu_offset_mtx);
4575 	SLIST_FOREACH_SAFE(op, &cpu_offsets, entries, top) {
4576 		if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) {
4577 			MPASS(op->refcount > 0);
4578 			op->refcount--;
4579 			if (op->refcount == 0) {
4580 				SLIST_REMOVE(&cpu_offsets, op, cpu_offset, entries);
4581 				free(op, M_IFLIB);
4582 			}
4583 			break;
4584 		}
4585 	}
4586 	mtx_unlock(&cpu_offset_mtx);
4587 }
4588 
4589 int
4590 iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp)
4591 {
4592 	if_ctx_t ctx;
4593 	if_t ifp;
4594 	if_softc_ctx_t scctx;
4595 	kobjop_desc_t kobj_desc;
4596 	kobj_method_t *kobj_method;
4597 	int err, msix, rid;
4598 	uint16_t main_rxq, main_txq;
4599 
4600 	ctx = malloc(sizeof(* ctx), M_IFLIB, M_WAITOK|M_ZERO);
4601 
4602 	if (sc == NULL) {
4603 		sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4604 		device_set_softc(dev, ctx);
4605 		ctx->ifc_flags |= IFC_SC_ALLOCATED;
4606 	}
4607 
4608 	ctx->ifc_sctx = sctx;
4609 	ctx->ifc_dev = dev;
4610 	ctx->ifc_softc = sc;
4611 
4612 	if ((err = iflib_register(ctx)) != 0) {
4613 		device_printf(dev, "iflib_register failed %d\n", err);
4614 		goto fail_ctx_free;
4615 	}
4616 	iflib_add_device_sysctl_pre(ctx);
4617 
4618 	scctx = &ctx->ifc_softc_ctx;
4619 	ifp = ctx->ifc_ifp;
4620 
4621 	iflib_reset_qvalues(ctx);
4622 	CTX_LOCK(ctx);
4623 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4624 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4625 		goto fail_unlock;
4626 	}
4627 	_iflib_pre_assert(scctx);
4628 	ctx->ifc_txrx = *scctx->isc_txrx;
4629 
4630 	if (sctx->isc_flags & IFLIB_DRIVER_MEDIA)
4631 		ctx->ifc_mediap = scctx->isc_media;
4632 
4633 #ifdef INVARIANTS
4634 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
4635 		MPASS(scctx->isc_tx_csum_flags);
4636 #endif
4637 
4638 	if_setcapabilities(ifp,
4639 	    scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_NOMAP);
4640 	if_setcapenable(ifp,
4641 	    scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_NOMAP);
4642 
4643 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4644 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4645 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4646 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4647 
4648 	main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4649 	main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4650 
4651 	/* XXX change for per-queue sizes */
4652 	device_printf(dev, "Using %d TX descriptors and %d RX descriptors\n",
4653 	    scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4654 
4655 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4656 	    MAX_SINGLE_PACKET_FRACTION)
4657 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4658 		    MAX_SINGLE_PACKET_FRACTION);
4659 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4660 	    MAX_SINGLE_PACKET_FRACTION)
4661 		scctx->isc_tx_tso_segments_max = max(1,
4662 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4663 
4664 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4665 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
4666 		/*
4667 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
4668 		 * but some MACs do.
4669 		 */
4670 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
4671 		    IP_MAXPACKET));
4672 		/*
4673 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
4674 		 * into account.  In the worst case, each of these calls will
4675 		 * add another mbuf and, thus, the requirement for another DMA
4676 		 * segment.  So for best performance, it doesn't make sense to
4677 		 * advertize a maximum of TSO segments that typically will
4678 		 * require defragmentation in iflib_encap().
4679 		 */
4680 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
4681 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
4682 	}
4683 	if (scctx->isc_rss_table_size == 0)
4684 		scctx->isc_rss_table_size = 64;
4685 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4686 
4687 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4688 	/* XXX format name */
4689 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx,
4690 	    NULL, NULL, "admin");
4691 
4692 	/* Set up cpu set.  If it fails, use the set of all CPUs. */
4693 	if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) != 0) {
4694 		device_printf(dev, "Unable to fetch CPU list\n");
4695 		CPU_COPY(&all_cpus, &ctx->ifc_cpus);
4696 	}
4697 	MPASS(CPU_COUNT(&ctx->ifc_cpus) > 0);
4698 
4699 	/*
4700 	** Now set up MSI or MSI-X, should return us the number of supported
4701 	** vectors (will be 1 for a legacy interrupt and MSI).
4702 	*/
4703 	if (sctx->isc_flags & IFLIB_SKIP_MSIX) {
4704 		msix = scctx->isc_vectors;
4705 	} else if (scctx->isc_msix_bar != 0)
4706 	       /*
4707 		* The simple fact that isc_msix_bar is not 0 does not mean we
4708 		* we have a good value there that is known to work.
4709 		*/
4710 		msix = iflib_msix_init(ctx);
4711 	else {
4712 		scctx->isc_vectors = 1;
4713 		scctx->isc_ntxqsets = 1;
4714 		scctx->isc_nrxqsets = 1;
4715 		scctx->isc_intr = IFLIB_INTR_LEGACY;
4716 		msix = 0;
4717 	}
4718 	/* Get memory for the station queues */
4719 	if ((err = iflib_queues_alloc(ctx))) {
4720 		device_printf(dev, "Unable to allocate queue memory\n");
4721 		goto fail_intr_free;
4722 	}
4723 
4724 	if ((err = iflib_qset_structures_setup(ctx)))
4725 		goto fail_queues;
4726 
4727 	/*
4728 	 * Now that we know how many queues there are, get the core offset.
4729 	 */
4730 	ctx->ifc_sysctl_core_offset = get_ctx_core_offset(ctx);
4731 
4732 	if (msix > 1) {
4733 		/*
4734 		 * When using MSI-X, ensure that ifdi_{r,t}x_queue_intr_enable
4735 		 * aren't the default NULL implementation.
4736 		 */
4737 		kobj_desc = &ifdi_rx_queue_intr_enable_desc;
4738 		kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL,
4739 		    kobj_desc);
4740 		if (kobj_method == &kobj_desc->deflt) {
4741 			device_printf(dev,
4742 			    "MSI-X requires ifdi_rx_queue_intr_enable method");
4743 			err = EOPNOTSUPP;
4744 			goto fail_queues;
4745 		}
4746 		kobj_desc = &ifdi_tx_queue_intr_enable_desc;
4747 		kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL,
4748 		    kobj_desc);
4749 		if (kobj_method == &kobj_desc->deflt) {
4750 			device_printf(dev,
4751 			    "MSI-X requires ifdi_tx_queue_intr_enable method");
4752 			err = EOPNOTSUPP;
4753 			goto fail_queues;
4754 		}
4755 
4756 		/*
4757 		 * Assign the MSI-X vectors.
4758 		 * Note that the default NULL ifdi_msix_intr_assign method will
4759 		 * fail here, too.
4760 		 */
4761 		err = IFDI_MSIX_INTR_ASSIGN(ctx, msix);
4762 		if (err != 0) {
4763 			device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n",
4764 			    err);
4765 			goto fail_queues;
4766 		}
4767 	} else if (scctx->isc_intr != IFLIB_INTR_MSIX) {
4768 		rid = 0;
4769 		if (scctx->isc_intr == IFLIB_INTR_MSI) {
4770 			MPASS(msix == 1);
4771 			rid = 1;
4772 		}
4773 		if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) {
4774 			device_printf(dev, "iflib_legacy_setup failed %d\n", err);
4775 			goto fail_queues;
4776 		}
4777 	} else {
4778 		device_printf(dev,
4779 		    "Cannot use iflib with only 1 MSI-X interrupt!\n");
4780 		err = ENODEV;
4781 		goto fail_intr_free;
4782 	}
4783 
4784 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
4785 
4786 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4787 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4788 		goto fail_detach;
4789 	}
4790 
4791 	/*
4792 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4793 	 * This must appear after the call to ether_ifattach() because
4794 	 * ether_ifattach() sets if_hdrlen to the default value.
4795 	 */
4796 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4797 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
4798 
4799 	if ((err = iflib_netmap_attach(ctx))) {
4800 		device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err);
4801 		goto fail_detach;
4802 	}
4803 	*ctxp = ctx;
4804 
4805 	DEBUGNET_SET(ctx->ifc_ifp, iflib);
4806 
4807 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4808 	iflib_add_device_sysctl_post(ctx);
4809 	iflib_add_pfil(ctx);
4810 	ctx->ifc_flags |= IFC_INIT_DONE;
4811 	CTX_UNLOCK(ctx);
4812 
4813 	return (0);
4814 
4815 fail_detach:
4816 	ether_ifdetach(ctx->ifc_ifp);
4817 fail_intr_free:
4818 	iflib_free_intr_mem(ctx);
4819 fail_queues:
4820 	iflib_tx_structures_free(ctx);
4821 	iflib_rx_structures_free(ctx);
4822 	taskqgroup_detach(qgroup_if_config_tqg, &ctx->ifc_admin_task);
4823 	IFDI_DETACH(ctx);
4824 fail_unlock:
4825 	CTX_UNLOCK(ctx);
4826 	iflib_deregister(ctx);
4827 fail_ctx_free:
4828 	device_set_softc(ctx->ifc_dev, NULL);
4829         if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4830                 free(ctx->ifc_softc, M_IFLIB);
4831         free(ctx, M_IFLIB);
4832 	return (err);
4833 }
4834 
4835 int
4836 iflib_pseudo_register(device_t dev, if_shared_ctx_t sctx, if_ctx_t *ctxp,
4837 					  struct iflib_cloneattach_ctx *clctx)
4838 {
4839 	int err;
4840 	if_ctx_t ctx;
4841 	if_t ifp;
4842 	if_softc_ctx_t scctx;
4843 	int i;
4844 	void *sc;
4845 	uint16_t main_txq;
4846 	uint16_t main_rxq;
4847 
4848 	ctx = malloc(sizeof(*ctx), M_IFLIB, M_WAITOK|M_ZERO);
4849 	sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4850 	ctx->ifc_flags |= IFC_SC_ALLOCATED;
4851 	if (sctx->isc_flags & (IFLIB_PSEUDO|IFLIB_VIRTUAL))
4852 		ctx->ifc_flags |= IFC_PSEUDO;
4853 
4854 	ctx->ifc_sctx = sctx;
4855 	ctx->ifc_softc = sc;
4856 	ctx->ifc_dev = dev;
4857 
4858 	if ((err = iflib_register(ctx)) != 0) {
4859 		device_printf(dev, "%s: iflib_register failed %d\n", __func__, err);
4860 		goto fail_ctx_free;
4861 	}
4862 	iflib_add_device_sysctl_pre(ctx);
4863 
4864 	scctx = &ctx->ifc_softc_ctx;
4865 	ifp = ctx->ifc_ifp;
4866 
4867 	iflib_reset_qvalues(ctx);
4868 	CTX_LOCK(ctx);
4869 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4870 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4871 		goto fail_unlock;
4872 	}
4873 	if (sctx->isc_flags & IFLIB_GEN_MAC)
4874 		ether_gen_addr(ifp, &ctx->ifc_mac);
4875 	if ((err = IFDI_CLONEATTACH(ctx, clctx->cc_ifc, clctx->cc_name,
4876 								clctx->cc_params)) != 0) {
4877 		device_printf(dev, "IFDI_CLONEATTACH failed %d\n", err);
4878 		goto fail_unlock;
4879 	}
4880 #ifdef INVARIANTS
4881 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
4882 		MPASS(scctx->isc_tx_csum_flags);
4883 #endif
4884 
4885 	if_setcapabilities(ifp, scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_LINKSTATE);
4886 	if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_LINKSTATE);
4887 
4888 	ifp->if_flags |= IFF_NOGROUP;
4889 	if (sctx->isc_flags & IFLIB_PSEUDO) {
4890 		ifmedia_add(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO, 0, NULL);
4891 		ifmedia_set(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO);
4892 		if (sctx->isc_flags & IFLIB_PSEUDO_ETHER) {
4893 			ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
4894 		} else {
4895 			if_attach(ctx->ifc_ifp);
4896 			bpfattach(ctx->ifc_ifp, DLT_NULL, sizeof(u_int32_t));
4897 		}
4898 
4899 		if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4900 			device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4901 			goto fail_detach;
4902 		}
4903 		*ctxp = ctx;
4904 
4905 		/*
4906 		 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4907 		 * This must appear after the call to ether_ifattach() because
4908 		 * ether_ifattach() sets if_hdrlen to the default value.
4909 		 */
4910 		if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4911 			if_setifheaderlen(ifp,
4912 			    sizeof(struct ether_vlan_header));
4913 
4914 		if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4915 		iflib_add_device_sysctl_post(ctx);
4916 		ctx->ifc_flags |= IFC_INIT_DONE;
4917 		CTX_UNLOCK(ctx);
4918 		return (0);
4919 	}
4920 	ifmedia_add(ctx->ifc_mediap, IFM_ETHER | IFM_1000_T | IFM_FDX, 0, NULL);
4921 	ifmedia_add(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO, 0, NULL);
4922 	ifmedia_set(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO);
4923 
4924 	_iflib_pre_assert(scctx);
4925 	ctx->ifc_txrx = *scctx->isc_txrx;
4926 
4927 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4928 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4929 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4930 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4931 
4932 	main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4933 	main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4934 
4935 	/* XXX change for per-queue sizes */
4936 	device_printf(dev, "Using %d TX descriptors and %d RX descriptors\n",
4937 	    scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4938 
4939 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4940 	    MAX_SINGLE_PACKET_FRACTION)
4941 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4942 		    MAX_SINGLE_PACKET_FRACTION);
4943 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4944 	    MAX_SINGLE_PACKET_FRACTION)
4945 		scctx->isc_tx_tso_segments_max = max(1,
4946 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4947 
4948 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4949 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
4950 		/*
4951 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
4952 		 * but some MACs do.
4953 		 */
4954 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
4955 		    IP_MAXPACKET));
4956 		/*
4957 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
4958 		 * into account.  In the worst case, each of these calls will
4959 		 * add another mbuf and, thus, the requirement for another DMA
4960 		 * segment.  So for best performance, it doesn't make sense to
4961 		 * advertize a maximum of TSO segments that typically will
4962 		 * require defragmentation in iflib_encap().
4963 		 */
4964 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
4965 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
4966 	}
4967 	if (scctx->isc_rss_table_size == 0)
4968 		scctx->isc_rss_table_size = 64;
4969 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4970 
4971 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4972 	/* XXX format name */
4973 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx,
4974 	    NULL, NULL, "admin");
4975 
4976 	/* XXX --- can support > 1 -- but keep it simple for now */
4977 	scctx->isc_intr = IFLIB_INTR_LEGACY;
4978 
4979 	/* Get memory for the station queues */
4980 	if ((err = iflib_queues_alloc(ctx))) {
4981 		device_printf(dev, "Unable to allocate queue memory\n");
4982 		goto fail_iflib_detach;
4983 	}
4984 
4985 	if ((err = iflib_qset_structures_setup(ctx))) {
4986 		device_printf(dev, "qset structure setup failed %d\n", err);
4987 		goto fail_queues;
4988 	}
4989 
4990 	/*
4991 	 * XXX What if anything do we want to do about interrupts?
4992 	 */
4993 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
4994 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4995 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4996 		goto fail_detach;
4997 	}
4998 
4999 	/*
5000 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
5001 	 * This must appear after the call to ether_ifattach() because
5002 	 * ether_ifattach() sets if_hdrlen to the default value.
5003 	 */
5004 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
5005 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
5006 
5007 	/* XXX handle more than one queue */
5008 	for (i = 0; i < scctx->isc_nrxqsets; i++)
5009 		IFDI_RX_CLSET(ctx, 0, i, ctx->ifc_rxqs[i].ifr_fl[0].ifl_sds.ifsd_cl);
5010 
5011 	*ctxp = ctx;
5012 
5013 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
5014 	iflib_add_device_sysctl_post(ctx);
5015 	ctx->ifc_flags |= IFC_INIT_DONE;
5016 	CTX_UNLOCK(ctx);
5017 
5018 	return (0);
5019 fail_detach:
5020 	ether_ifdetach(ctx->ifc_ifp);
5021 fail_queues:
5022 	iflib_tx_structures_free(ctx);
5023 	iflib_rx_structures_free(ctx);
5024 fail_iflib_detach:
5025 	IFDI_DETACH(ctx);
5026 fail_unlock:
5027 	CTX_UNLOCK(ctx);
5028 	iflib_deregister(ctx);
5029 fail_ctx_free:
5030 	free(ctx->ifc_softc, M_IFLIB);
5031 	free(ctx, M_IFLIB);
5032 	return (err);
5033 }
5034 
5035 int
5036 iflib_pseudo_deregister(if_ctx_t ctx)
5037 {
5038 	if_t ifp = ctx->ifc_ifp;
5039 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5040 	iflib_txq_t txq;
5041 	iflib_rxq_t rxq;
5042 	int i, j;
5043 	struct taskqgroup *tqg;
5044 	iflib_fl_t fl;
5045 
5046 	/* Unregister VLAN event handlers early */
5047 	iflib_unregister_vlan_handlers(ctx);
5048 
5049 	if ((sctx->isc_flags & IFLIB_PSEUDO)  &&
5050 		(sctx->isc_flags & IFLIB_PSEUDO_ETHER) == 0) {
5051 		bpfdetach(ifp);
5052 		if_detach(ifp);
5053 	} else {
5054 		ether_ifdetach(ifp);
5055 	}
5056 	/* XXX drain any dependent tasks */
5057 	tqg = qgroup_if_io_tqg;
5058 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
5059 		callout_drain(&txq->ift_timer);
5060 		if (txq->ift_task.gt_uniq != NULL)
5061 			taskqgroup_detach(tqg, &txq->ift_task);
5062 	}
5063 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
5064 		callout_drain(&rxq->ifr_watchdog);
5065 		if (rxq->ifr_task.gt_uniq != NULL)
5066 			taskqgroup_detach(tqg, &rxq->ifr_task);
5067 
5068 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5069 			free(fl->ifl_rx_bitmap, M_IFLIB);
5070 	}
5071 	tqg = qgroup_if_config_tqg;
5072 	if (ctx->ifc_admin_task.gt_uniq != NULL)
5073 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
5074 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
5075 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
5076 
5077 	iflib_tx_structures_free(ctx);
5078 	iflib_rx_structures_free(ctx);
5079 
5080 	iflib_deregister(ctx);
5081 
5082 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5083 		free(ctx->ifc_softc, M_IFLIB);
5084 	free(ctx, M_IFLIB);
5085 	return (0);
5086 }
5087 
5088 int
5089 iflib_device_attach(device_t dev)
5090 {
5091 	if_ctx_t ctx;
5092 	if_shared_ctx_t sctx;
5093 
5094 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
5095 		return (ENOTSUP);
5096 
5097 	pci_enable_busmaster(dev);
5098 
5099 	return (iflib_device_register(dev, NULL, sctx, &ctx));
5100 }
5101 
5102 int
5103 iflib_device_deregister(if_ctx_t ctx)
5104 {
5105 	if_t ifp = ctx->ifc_ifp;
5106 	iflib_txq_t txq;
5107 	iflib_rxq_t rxq;
5108 	device_t dev = ctx->ifc_dev;
5109 	int i, j;
5110 	struct taskqgroup *tqg;
5111 	iflib_fl_t fl;
5112 
5113 	/* Make sure VLANS are not using driver */
5114 	if (if_vlantrunkinuse(ifp)) {
5115 		device_printf(dev, "Vlan in use, detach first\n");
5116 		return (EBUSY);
5117 	}
5118 #ifdef PCI_IOV
5119 	if (!CTX_IS_VF(ctx) && pci_iov_detach(dev) != 0) {
5120 		device_printf(dev, "SR-IOV in use; detach first.\n");
5121 		return (EBUSY);
5122 	}
5123 #endif
5124 
5125 	STATE_LOCK(ctx);
5126 	ctx->ifc_flags |= IFC_IN_DETACH;
5127 	STATE_UNLOCK(ctx);
5128 
5129 	/* Unregister VLAN handlers before calling iflib_stop() */
5130 	iflib_unregister_vlan_handlers(ctx);
5131 
5132 	iflib_netmap_detach(ifp);
5133 	ether_ifdetach(ifp);
5134 
5135 	CTX_LOCK(ctx);
5136 	iflib_stop(ctx);
5137 	CTX_UNLOCK(ctx);
5138 
5139 	iflib_rem_pfil(ctx);
5140 	if (ctx->ifc_led_dev != NULL)
5141 		led_destroy(ctx->ifc_led_dev);
5142 	/* XXX drain any dependent tasks */
5143 	tqg = qgroup_if_io_tqg;
5144 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
5145 		callout_drain(&txq->ift_timer);
5146 		if (txq->ift_task.gt_uniq != NULL)
5147 			taskqgroup_detach(tqg, &txq->ift_task);
5148 	}
5149 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
5150 		if (rxq->ifr_task.gt_uniq != NULL)
5151 			taskqgroup_detach(tqg, &rxq->ifr_task);
5152 
5153 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5154 			free(fl->ifl_rx_bitmap, M_IFLIB);
5155 	}
5156 	tqg = qgroup_if_config_tqg;
5157 	if (ctx->ifc_admin_task.gt_uniq != NULL)
5158 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
5159 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
5160 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
5161 	CTX_LOCK(ctx);
5162 	IFDI_DETACH(ctx);
5163 	CTX_UNLOCK(ctx);
5164 
5165 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
5166 	iflib_free_intr_mem(ctx);
5167 
5168 	bus_generic_detach(dev);
5169 
5170 	iflib_tx_structures_free(ctx);
5171 	iflib_rx_structures_free(ctx);
5172 
5173 	iflib_deregister(ctx);
5174 
5175 	device_set_softc(ctx->ifc_dev, NULL);
5176 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5177 		free(ctx->ifc_softc, M_IFLIB);
5178 	unref_ctx_core_offset(ctx);
5179 	free(ctx, M_IFLIB);
5180 	return (0);
5181 }
5182 
5183 static void
5184 iflib_free_intr_mem(if_ctx_t ctx)
5185 {
5186 
5187 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) {
5188 		iflib_irq_free(ctx, &ctx->ifc_legacy_irq);
5189 	}
5190 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) {
5191 		pci_release_msi(ctx->ifc_dev);
5192 	}
5193 	if (ctx->ifc_msix_mem != NULL) {
5194 		bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY,
5195 		    rman_get_rid(ctx->ifc_msix_mem), ctx->ifc_msix_mem);
5196 		ctx->ifc_msix_mem = NULL;
5197 	}
5198 }
5199 
5200 int
5201 iflib_device_detach(device_t dev)
5202 {
5203 	if_ctx_t ctx = device_get_softc(dev);
5204 
5205 	return (iflib_device_deregister(ctx));
5206 }
5207 
5208 int
5209 iflib_device_suspend(device_t dev)
5210 {
5211 	if_ctx_t ctx = device_get_softc(dev);
5212 
5213 	CTX_LOCK(ctx);
5214 	IFDI_SUSPEND(ctx);
5215 	CTX_UNLOCK(ctx);
5216 
5217 	return bus_generic_suspend(dev);
5218 }
5219 int
5220 iflib_device_shutdown(device_t dev)
5221 {
5222 	if_ctx_t ctx = device_get_softc(dev);
5223 
5224 	CTX_LOCK(ctx);
5225 	IFDI_SHUTDOWN(ctx);
5226 	CTX_UNLOCK(ctx);
5227 
5228 	return bus_generic_suspend(dev);
5229 }
5230 
5231 
5232 int
5233 iflib_device_resume(device_t dev)
5234 {
5235 	if_ctx_t ctx = device_get_softc(dev);
5236 	iflib_txq_t txq = ctx->ifc_txqs;
5237 
5238 	CTX_LOCK(ctx);
5239 	IFDI_RESUME(ctx);
5240 	iflib_if_init_locked(ctx);
5241 	CTX_UNLOCK(ctx);
5242 	for (int i = 0; i < NTXQSETS(ctx); i++, txq++)
5243 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
5244 
5245 	return (bus_generic_resume(dev));
5246 }
5247 
5248 int
5249 iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params)
5250 {
5251 	int error;
5252 	if_ctx_t ctx = device_get_softc(dev);
5253 
5254 	CTX_LOCK(ctx);
5255 	error = IFDI_IOV_INIT(ctx, num_vfs, params);
5256 	CTX_UNLOCK(ctx);
5257 
5258 	return (error);
5259 }
5260 
5261 void
5262 iflib_device_iov_uninit(device_t dev)
5263 {
5264 	if_ctx_t ctx = device_get_softc(dev);
5265 
5266 	CTX_LOCK(ctx);
5267 	IFDI_IOV_UNINIT(ctx);
5268 	CTX_UNLOCK(ctx);
5269 }
5270 
5271 int
5272 iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params)
5273 {
5274 	int error;
5275 	if_ctx_t ctx = device_get_softc(dev);
5276 
5277 	CTX_LOCK(ctx);
5278 	error = IFDI_IOV_VF_ADD(ctx, vfnum, params);
5279 	CTX_UNLOCK(ctx);
5280 
5281 	return (error);
5282 }
5283 
5284 /*********************************************************************
5285  *
5286  *  MODULE FUNCTION DEFINITIONS
5287  *
5288  **********************************************************************/
5289 
5290 /*
5291  * - Start a fast taskqueue thread for each core
5292  * - Start a taskqueue for control operations
5293  */
5294 static int
5295 iflib_module_init(void)
5296 {
5297 	return (0);
5298 }
5299 
5300 static int
5301 iflib_module_event_handler(module_t mod, int what, void *arg)
5302 {
5303 	int err;
5304 
5305 	switch (what) {
5306 	case MOD_LOAD:
5307 		if ((err = iflib_module_init()) != 0)
5308 			return (err);
5309 		break;
5310 	case MOD_UNLOAD:
5311 		return (EBUSY);
5312 	default:
5313 		return (EOPNOTSUPP);
5314 	}
5315 
5316 	return (0);
5317 }
5318 
5319 /*********************************************************************
5320  *
5321  *  PUBLIC FUNCTION DEFINITIONS
5322  *     ordered as in iflib.h
5323  *
5324  **********************************************************************/
5325 
5326 
5327 static void
5328 _iflib_assert(if_shared_ctx_t sctx)
5329 {
5330 	int i;
5331 
5332 	MPASS(sctx->isc_tx_maxsize);
5333 	MPASS(sctx->isc_tx_maxsegsize);
5334 
5335 	MPASS(sctx->isc_rx_maxsize);
5336 	MPASS(sctx->isc_rx_nsegments);
5337 	MPASS(sctx->isc_rx_maxsegsize);
5338 
5339 	MPASS(sctx->isc_nrxqs >= 1 && sctx->isc_nrxqs <= 8);
5340 	for (i = 0; i < sctx->isc_nrxqs; i++) {
5341 		MPASS(sctx->isc_nrxd_min[i]);
5342 		MPASS(powerof2(sctx->isc_nrxd_min[i]));
5343 		MPASS(sctx->isc_nrxd_max[i]);
5344 		MPASS(powerof2(sctx->isc_nrxd_max[i]));
5345 		MPASS(sctx->isc_nrxd_default[i]);
5346 		MPASS(powerof2(sctx->isc_nrxd_default[i]));
5347 	}
5348 
5349 	MPASS(sctx->isc_ntxqs >= 1 && sctx->isc_ntxqs <= 8);
5350 	for (i = 0; i < sctx->isc_ntxqs; i++) {
5351 		MPASS(sctx->isc_ntxd_min[i]);
5352 		MPASS(powerof2(sctx->isc_ntxd_min[i]));
5353 		MPASS(sctx->isc_ntxd_max[i]);
5354 		MPASS(powerof2(sctx->isc_ntxd_max[i]));
5355 		MPASS(sctx->isc_ntxd_default[i]);
5356 		MPASS(powerof2(sctx->isc_ntxd_default[i]));
5357 	}
5358 }
5359 
5360 static void
5361 _iflib_pre_assert(if_softc_ctx_t scctx)
5362 {
5363 
5364 	MPASS(scctx->isc_txrx->ift_txd_encap);
5365 	MPASS(scctx->isc_txrx->ift_txd_flush);
5366 	MPASS(scctx->isc_txrx->ift_txd_credits_update);
5367 	MPASS(scctx->isc_txrx->ift_rxd_available);
5368 	MPASS(scctx->isc_txrx->ift_rxd_pkt_get);
5369 	MPASS(scctx->isc_txrx->ift_rxd_refill);
5370 	MPASS(scctx->isc_txrx->ift_rxd_flush);
5371 }
5372 
5373 static int
5374 iflib_register(if_ctx_t ctx)
5375 {
5376 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5377 	driver_t *driver = sctx->isc_driver;
5378 	device_t dev = ctx->ifc_dev;
5379 	if_t ifp;
5380 	u_char type;
5381 	int iflags;
5382 
5383 	if ((sctx->isc_flags & IFLIB_PSEUDO) == 0)
5384 		_iflib_assert(sctx);
5385 
5386 	CTX_LOCK_INIT(ctx);
5387 	STATE_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev));
5388 	if (sctx->isc_flags & IFLIB_PSEUDO) {
5389 		if (sctx->isc_flags & IFLIB_PSEUDO_ETHER)
5390 			type = IFT_ETHER;
5391 		else
5392 			type = IFT_PPP;
5393 	} else
5394 		type = IFT_ETHER;
5395 	ifp = ctx->ifc_ifp = if_alloc(type);
5396 	if (ifp == NULL) {
5397 		device_printf(dev, "can not allocate ifnet structure\n");
5398 		return (ENOMEM);
5399 	}
5400 
5401 	/*
5402 	 * Initialize our context's device specific methods
5403 	 */
5404 	kobj_init((kobj_t) ctx, (kobj_class_t) driver);
5405 	kobj_class_compile((kobj_class_t) driver);
5406 
5407 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
5408 	if_setsoftc(ifp, ctx);
5409 	if_setdev(ifp, dev);
5410 	if_setinitfn(ifp, iflib_if_init);
5411 	if_setioctlfn(ifp, iflib_if_ioctl);
5412 #ifdef ALTQ
5413 	if_setstartfn(ifp, iflib_altq_if_start);
5414 	if_settransmitfn(ifp, iflib_altq_if_transmit);
5415 	if_setsendqready(ifp);
5416 #else
5417 	if_settransmitfn(ifp, iflib_if_transmit);
5418 #endif
5419 	if_setqflushfn(ifp, iflib_if_qflush);
5420 	iflags = IFF_MULTICAST | IFF_KNOWSEPOCH;
5421 
5422 	if ((sctx->isc_flags & IFLIB_PSEUDO) &&
5423 		(sctx->isc_flags & IFLIB_PSEUDO_ETHER) == 0)
5424 		iflags |= IFF_POINTOPOINT;
5425 	else
5426 		iflags |= IFF_BROADCAST | IFF_SIMPLEX;
5427 	if_setflags(ifp, iflags);
5428 	ctx->ifc_vlan_attach_event =
5429 		EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx,
5430 							  EVENTHANDLER_PRI_FIRST);
5431 	ctx->ifc_vlan_detach_event =
5432 		EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx,
5433 							  EVENTHANDLER_PRI_FIRST);
5434 
5435 	if ((sctx->isc_flags & IFLIB_DRIVER_MEDIA) == 0) {
5436 		ctx->ifc_mediap = &ctx->ifc_media;
5437 		ifmedia_init(ctx->ifc_mediap, IFM_IMASK,
5438 		    iflib_media_change, iflib_media_status);
5439 	}
5440 	return (0);
5441 }
5442 
5443 static void
5444 iflib_unregister_vlan_handlers(if_ctx_t ctx)
5445 {
5446 	/* Unregister VLAN events */
5447 	if (ctx->ifc_vlan_attach_event != NULL) {
5448 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
5449 		ctx->ifc_vlan_attach_event = NULL;
5450 	}
5451 	if (ctx->ifc_vlan_detach_event != NULL) {
5452 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
5453 		ctx->ifc_vlan_detach_event = NULL;
5454 	}
5455 
5456 }
5457 
5458 static void
5459 iflib_deregister(if_ctx_t ctx)
5460 {
5461 	if_t ifp = ctx->ifc_ifp;
5462 
5463 	/* Remove all media */
5464 	ifmedia_removeall(&ctx->ifc_media);
5465 
5466 	/* Ensure that VLAN event handlers are unregistered */
5467 	iflib_unregister_vlan_handlers(ctx);
5468 
5469 	/* Release kobject reference */
5470 	kobj_delete((kobj_t) ctx, NULL);
5471 
5472 	/* Free the ifnet structure */
5473 	if_free(ifp);
5474 
5475 	STATE_LOCK_DESTROY(ctx);
5476 
5477 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
5478 	CTX_LOCK_DESTROY(ctx);
5479 }
5480 
5481 static int
5482 iflib_queues_alloc(if_ctx_t ctx)
5483 {
5484 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5485 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5486 	device_t dev = ctx->ifc_dev;
5487 	int nrxqsets = scctx->isc_nrxqsets;
5488 	int ntxqsets = scctx->isc_ntxqsets;
5489 	iflib_txq_t txq;
5490 	iflib_rxq_t rxq;
5491 	iflib_fl_t fl = NULL;
5492 	int i, j, cpu, err, txconf, rxconf;
5493 	iflib_dma_info_t ifdip;
5494 	uint32_t *rxqsizes = scctx->isc_rxqsizes;
5495 	uint32_t *txqsizes = scctx->isc_txqsizes;
5496 	uint8_t nrxqs = sctx->isc_nrxqs;
5497 	uint8_t ntxqs = sctx->isc_ntxqs;
5498 	int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1;
5499 	caddr_t *vaddrs;
5500 	uint64_t *paddrs;
5501 
5502 	KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1"));
5503 	KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1"));
5504 
5505 	/* Allocate the TX ring struct memory */
5506 	if (!(ctx->ifc_txqs =
5507 	    (iflib_txq_t) malloc(sizeof(struct iflib_txq) *
5508 	    ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5509 		device_printf(dev, "Unable to allocate TX ring memory\n");
5510 		err = ENOMEM;
5511 		goto fail;
5512 	}
5513 
5514 	/* Now allocate the RX */
5515 	if (!(ctx->ifc_rxqs =
5516 	    (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) *
5517 	    nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5518 		device_printf(dev, "Unable to allocate RX ring memory\n");
5519 		err = ENOMEM;
5520 		goto rx_fail;
5521 	}
5522 
5523 	txq = ctx->ifc_txqs;
5524 	rxq = ctx->ifc_rxqs;
5525 
5526 	/*
5527 	 * XXX handle allocation failure
5528 	 */
5529 	for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) {
5530 		/* Set up some basics */
5531 
5532 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs,
5533 		    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
5534 			device_printf(dev,
5535 			    "Unable to allocate TX DMA info memory\n");
5536 			err = ENOMEM;
5537 			goto err_tx_desc;
5538 		}
5539 		txq->ift_ifdi = ifdip;
5540 		for (j = 0; j < ntxqs; j++, ifdip++) {
5541 			if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, 0)) {
5542 				device_printf(dev,
5543 				    "Unable to allocate TX descriptors\n");
5544 				err = ENOMEM;
5545 				goto err_tx_desc;
5546 			}
5547 			txq->ift_txd_size[j] = scctx->isc_txd_size[j];
5548 			bzero((void *)ifdip->idi_vaddr, txqsizes[j]);
5549 		}
5550 		txq->ift_ctx = ctx;
5551 		txq->ift_id = i;
5552 		if (sctx->isc_flags & IFLIB_HAS_TXCQ) {
5553 			txq->ift_br_offset = 1;
5554 		} else {
5555 			txq->ift_br_offset = 0;
5556 		}
5557 		/* XXX fix this */
5558 		txq->ift_timer.c_cpu = cpu;
5559 
5560 		if (iflib_txsd_alloc(txq)) {
5561 			device_printf(dev, "Critical Failure setting up TX buffers\n");
5562 			err = ENOMEM;
5563 			goto err_tx_desc;
5564 		}
5565 
5566 		/* Initialize the TX lock */
5567 		snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:TX(%d):callout",
5568 		    device_get_nameunit(dev), txq->ift_id);
5569 		mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF);
5570 		callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0);
5571 
5572 		err = ifmp_ring_alloc(&txq->ift_br, 2048, txq, iflib_txq_drain,
5573 				      iflib_txq_can_drain, M_IFLIB, M_WAITOK);
5574 		if (err) {
5575 			/* XXX free any allocated rings */
5576 			device_printf(dev, "Unable to allocate buf_ring\n");
5577 			goto err_tx_desc;
5578 		}
5579 	}
5580 
5581 	for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) {
5582 		/* Set up some basics */
5583 		callout_init(&rxq->ifr_watchdog, 1);
5584 
5585 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs,
5586 		   M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
5587 			device_printf(dev,
5588 			    "Unable to allocate RX DMA info memory\n");
5589 			err = ENOMEM;
5590 			goto err_tx_desc;
5591 		}
5592 
5593 		rxq->ifr_ifdi = ifdip;
5594 		/* XXX this needs to be changed if #rx queues != #tx queues */
5595 		rxq->ifr_ntxqirq = 1;
5596 		rxq->ifr_txqid[0] = i;
5597 		for (j = 0; j < nrxqs; j++, ifdip++) {
5598 			if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, 0)) {
5599 				device_printf(dev,
5600 				    "Unable to allocate RX descriptors\n");
5601 				err = ENOMEM;
5602 				goto err_tx_desc;
5603 			}
5604 			bzero((void *)ifdip->idi_vaddr, rxqsizes[j]);
5605 		}
5606 		rxq->ifr_ctx = ctx;
5607 		rxq->ifr_id = i;
5608 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
5609 			rxq->ifr_fl_offset = 1;
5610 		} else {
5611 			rxq->ifr_fl_offset = 0;
5612 		}
5613 		rxq->ifr_nfl = nfree_lists;
5614 		if (!(fl =
5615 			  (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) {
5616 			device_printf(dev, "Unable to allocate free list memory\n");
5617 			err = ENOMEM;
5618 			goto err_tx_desc;
5619 		}
5620 		rxq->ifr_fl = fl;
5621 		for (j = 0; j < nfree_lists; j++) {
5622 			fl[j].ifl_rxq = rxq;
5623 			fl[j].ifl_id = j;
5624 			fl[j].ifl_ifdi = &rxq->ifr_ifdi[j + rxq->ifr_fl_offset];
5625 			fl[j].ifl_rxd_size = scctx->isc_rxd_size[j];
5626 		}
5627 		/* Allocate receive buffers for the ring */
5628 		if (iflib_rxsd_alloc(rxq)) {
5629 			device_printf(dev,
5630 			    "Critical Failure setting up receive buffers\n");
5631 			err = ENOMEM;
5632 			goto err_rx_desc;
5633 		}
5634 
5635 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5636 			fl->ifl_rx_bitmap = bit_alloc(fl->ifl_size, M_IFLIB,
5637 			    M_WAITOK);
5638 	}
5639 
5640 	/* TXQs */
5641 	vaddrs = malloc(sizeof(caddr_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5642 	paddrs = malloc(sizeof(uint64_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5643 	for (i = 0; i < ntxqsets; i++) {
5644 		iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi;
5645 
5646 		for (j = 0; j < ntxqs; j++, di++) {
5647 			vaddrs[i*ntxqs + j] = di->idi_vaddr;
5648 			paddrs[i*ntxqs + j] = di->idi_paddr;
5649 		}
5650 	}
5651 	if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) {
5652 		device_printf(ctx->ifc_dev,
5653 		    "Unable to allocate device TX queue\n");
5654 		iflib_tx_structures_free(ctx);
5655 		free(vaddrs, M_IFLIB);
5656 		free(paddrs, M_IFLIB);
5657 		goto err_rx_desc;
5658 	}
5659 	free(vaddrs, M_IFLIB);
5660 	free(paddrs, M_IFLIB);
5661 
5662 	/* RXQs */
5663 	vaddrs = malloc(sizeof(caddr_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5664 	paddrs = malloc(sizeof(uint64_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5665 	for (i = 0; i < nrxqsets; i++) {
5666 		iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi;
5667 
5668 		for (j = 0; j < nrxqs; j++, di++) {
5669 			vaddrs[i*nrxqs + j] = di->idi_vaddr;
5670 			paddrs[i*nrxqs + j] = di->idi_paddr;
5671 		}
5672 	}
5673 	if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) {
5674 		device_printf(ctx->ifc_dev,
5675 		    "Unable to allocate device RX queue\n");
5676 		iflib_tx_structures_free(ctx);
5677 		free(vaddrs, M_IFLIB);
5678 		free(paddrs, M_IFLIB);
5679 		goto err_rx_desc;
5680 	}
5681 	free(vaddrs, M_IFLIB);
5682 	free(paddrs, M_IFLIB);
5683 
5684 	return (0);
5685 
5686 /* XXX handle allocation failure changes */
5687 err_rx_desc:
5688 err_tx_desc:
5689 rx_fail:
5690 	if (ctx->ifc_rxqs != NULL)
5691 		free(ctx->ifc_rxqs, M_IFLIB);
5692 	ctx->ifc_rxqs = NULL;
5693 	if (ctx->ifc_txqs != NULL)
5694 		free(ctx->ifc_txqs, M_IFLIB);
5695 	ctx->ifc_txqs = NULL;
5696 fail:
5697 	return (err);
5698 }
5699 
5700 static int
5701 iflib_tx_structures_setup(if_ctx_t ctx)
5702 {
5703 	iflib_txq_t txq = ctx->ifc_txqs;
5704 	int i;
5705 
5706 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
5707 		iflib_txq_setup(txq);
5708 
5709 	return (0);
5710 }
5711 
5712 static void
5713 iflib_tx_structures_free(if_ctx_t ctx)
5714 {
5715 	iflib_txq_t txq = ctx->ifc_txqs;
5716 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5717 	int i, j;
5718 
5719 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
5720 		for (j = 0; j < sctx->isc_ntxqs; j++)
5721 			iflib_dma_free(&txq->ift_ifdi[j]);
5722 		iflib_txq_destroy(txq);
5723 	}
5724 	free(ctx->ifc_txqs, M_IFLIB);
5725 	ctx->ifc_txqs = NULL;
5726 	IFDI_QUEUES_FREE(ctx);
5727 }
5728 
5729 /*********************************************************************
5730  *
5731  *  Initialize all receive rings.
5732  *
5733  **********************************************************************/
5734 static int
5735 iflib_rx_structures_setup(if_ctx_t ctx)
5736 {
5737 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5738 	int q;
5739 #if defined(INET6) || defined(INET)
5740 	int err, i;
5741 #endif
5742 
5743 	for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) {
5744 #if defined(INET6) || defined(INET)
5745 		if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_LRO) {
5746 			err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp,
5747 			    TCP_LRO_ENTRIES, min(1024,
5748 			    ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset]));
5749 			if (err != 0) {
5750 				device_printf(ctx->ifc_dev,
5751 				    "LRO Initialization failed!\n");
5752 				goto fail;
5753 			}
5754 		}
5755 #endif
5756 		IFDI_RXQ_SETUP(ctx, rxq->ifr_id);
5757 	}
5758 	return (0);
5759 #if defined(INET6) || defined(INET)
5760 fail:
5761 	/*
5762 	 * Free LRO resources allocated so far, we will only handle
5763 	 * the rings that completed, the failing case will have
5764 	 * cleaned up for itself.  'q' failed, so its the terminus.
5765 	 */
5766 	rxq = ctx->ifc_rxqs;
5767 	for (i = 0; i < q; ++i, rxq++) {
5768 		if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_LRO)
5769 			tcp_lro_free(&rxq->ifr_lc);
5770 	}
5771 	return (err);
5772 #endif
5773 }
5774 
5775 /*********************************************************************
5776  *
5777  *  Free all receive rings.
5778  *
5779  **********************************************************************/
5780 static void
5781 iflib_rx_structures_free(if_ctx_t ctx)
5782 {
5783 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5784 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5785 	int i, j;
5786 
5787 	for (i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) {
5788 		for (j = 0; j < sctx->isc_nrxqs; j++)
5789 			iflib_dma_free(&rxq->ifr_ifdi[j]);
5790 		iflib_rx_sds_free(rxq);
5791 #if defined(INET6) || defined(INET)
5792 		if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_LRO)
5793 			tcp_lro_free(&rxq->ifr_lc);
5794 #endif
5795 	}
5796 	free(ctx->ifc_rxqs, M_IFLIB);
5797 	ctx->ifc_rxqs = NULL;
5798 }
5799 
5800 static int
5801 iflib_qset_structures_setup(if_ctx_t ctx)
5802 {
5803 	int err;
5804 
5805 	/*
5806 	 * It is expected that the caller takes care of freeing queues if this
5807 	 * fails.
5808 	 */
5809 	if ((err = iflib_tx_structures_setup(ctx)) != 0) {
5810 		device_printf(ctx->ifc_dev, "iflib_tx_structures_setup failed: %d\n", err);
5811 		return (err);
5812 	}
5813 
5814 	if ((err = iflib_rx_structures_setup(ctx)) != 0)
5815 		device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err);
5816 
5817 	return (err);
5818 }
5819 
5820 int
5821 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
5822 		driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, const char *name)
5823 {
5824 
5825 	return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name));
5826 }
5827 
5828 #ifdef SMP
5829 static int
5830 find_nth(if_ctx_t ctx, int qid)
5831 {
5832 	cpuset_t cpus;
5833 	int i, cpuid, eqid, count;
5834 
5835 	CPU_COPY(&ctx->ifc_cpus, &cpus);
5836 	count = CPU_COUNT(&cpus);
5837 	eqid = qid % count;
5838 	/* clear up to the qid'th bit */
5839 	for (i = 0; i < eqid; i++) {
5840 		cpuid = CPU_FFS(&cpus);
5841 		MPASS(cpuid != 0);
5842 		CPU_CLR(cpuid-1, &cpus);
5843 	}
5844 	cpuid = CPU_FFS(&cpus);
5845 	MPASS(cpuid != 0);
5846 	return (cpuid-1);
5847 }
5848 
5849 #ifdef SCHED_ULE
5850 extern struct cpu_group *cpu_top;              /* CPU topology */
5851 
5852 static int
5853 find_child_with_core(int cpu, struct cpu_group *grp)
5854 {
5855 	int i;
5856 
5857 	if (grp->cg_children == 0)
5858 		return -1;
5859 
5860 	MPASS(grp->cg_child);
5861 	for (i = 0; i < grp->cg_children; i++) {
5862 		if (CPU_ISSET(cpu, &grp->cg_child[i].cg_mask))
5863 			return i;
5864 	}
5865 
5866 	return -1;
5867 }
5868 
5869 /*
5870  * Find the nth "close" core to the specified core
5871  * "close" is defined as the deepest level that shares
5872  * at least an L2 cache.  With threads, this will be
5873  * threads on the same core.  If the shared cache is L3
5874  * or higher, simply returns the same core.
5875  */
5876 static int
5877 find_close_core(int cpu, int core_offset)
5878 {
5879 	struct cpu_group *grp;
5880 	int i;
5881 	int fcpu;
5882 	cpuset_t cs;
5883 
5884 	grp = cpu_top;
5885 	if (grp == NULL)
5886 		return cpu;
5887 	i = 0;
5888 	while ((i = find_child_with_core(cpu, grp)) != -1) {
5889 		/* If the child only has one cpu, don't descend */
5890 		if (grp->cg_child[i].cg_count <= 1)
5891 			break;
5892 		grp = &grp->cg_child[i];
5893 	}
5894 
5895 	/* If they don't share at least an L2 cache, use the same CPU */
5896 	if (grp->cg_level > CG_SHARE_L2 || grp->cg_level == CG_SHARE_NONE)
5897 		return cpu;
5898 
5899 	/* Now pick one */
5900 	CPU_COPY(&grp->cg_mask, &cs);
5901 
5902 	/* Add the selected CPU offset to core offset. */
5903 	for (i = 0; (fcpu = CPU_FFS(&cs)) != 0; i++) {
5904 		if (fcpu - 1 == cpu)
5905 			break;
5906 		CPU_CLR(fcpu - 1, &cs);
5907 	}
5908 	MPASS(fcpu);
5909 
5910 	core_offset += i;
5911 
5912 	CPU_COPY(&grp->cg_mask, &cs);
5913 	for (i = core_offset % grp->cg_count; i > 0; i--) {
5914 		MPASS(CPU_FFS(&cs));
5915 		CPU_CLR(CPU_FFS(&cs) - 1, &cs);
5916 	}
5917 	MPASS(CPU_FFS(&cs));
5918 	return CPU_FFS(&cs) - 1;
5919 }
5920 #else
5921 static int
5922 find_close_core(int cpu, int core_offset __unused)
5923 {
5924 	return cpu;
5925 }
5926 #endif
5927 
5928 static int
5929 get_core_offset(if_ctx_t ctx, iflib_intr_type_t type, int qid)
5930 {
5931 	switch (type) {
5932 	case IFLIB_INTR_TX:
5933 		/* TX queues get cores which share at least an L2 cache with the corresponding RX queue */
5934 		/* XXX handle multiple RX threads per core and more than two core per L2 group */
5935 		return qid / CPU_COUNT(&ctx->ifc_cpus) + 1;
5936 	case IFLIB_INTR_RX:
5937 	case IFLIB_INTR_RXTX:
5938 		/* RX queues get the specified core */
5939 		return qid / CPU_COUNT(&ctx->ifc_cpus);
5940 	default:
5941 		return -1;
5942 	}
5943 }
5944 #else
5945 #define get_core_offset(ctx, type, qid)	CPU_FIRST()
5946 #define find_close_core(cpuid, tid)	CPU_FIRST()
5947 #define find_nth(ctx, gid)		CPU_FIRST()
5948 #endif
5949 
5950 /* Just to avoid copy/paste */
5951 static inline int
5952 iflib_irq_set_affinity(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type,
5953     int qid, struct grouptask *gtask, struct taskqgroup *tqg, void *uniq,
5954     const char *name)
5955 {
5956 	device_t dev;
5957 	int co, cpuid, err, tid;
5958 
5959 	dev = ctx->ifc_dev;
5960 	co = ctx->ifc_sysctl_core_offset;
5961 	if (ctx->ifc_sysctl_separate_txrx && type == IFLIB_INTR_TX)
5962 		co += ctx->ifc_softc_ctx.isc_nrxqsets;
5963 	cpuid = find_nth(ctx, qid + co);
5964 	tid = get_core_offset(ctx, type, qid);
5965 	if (tid < 0) {
5966 		device_printf(dev, "get_core_offset failed\n");
5967 		return (EOPNOTSUPP);
5968 	}
5969 	cpuid = find_close_core(cpuid, tid);
5970 	err = taskqgroup_attach_cpu(tqg, gtask, uniq, cpuid, dev, irq->ii_res,
5971 	    name);
5972 	if (err) {
5973 		device_printf(dev, "taskqgroup_attach_cpu failed %d\n", err);
5974 		return (err);
5975 	}
5976 #ifdef notyet
5977 	if (cpuid > ctx->ifc_cpuid_highest)
5978 		ctx->ifc_cpuid_highest = cpuid;
5979 #endif
5980 	return (0);
5981 }
5982 
5983 int
5984 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid,
5985 			iflib_intr_type_t type, driver_filter_t *filter,
5986 			void *filter_arg, int qid, const char *name)
5987 {
5988 	device_t dev;
5989 	struct grouptask *gtask;
5990 	struct taskqgroup *tqg;
5991 	iflib_filter_info_t info;
5992 	gtask_fn_t *fn;
5993 	int tqrid, err;
5994 	driver_filter_t *intr_fast;
5995 	void *q;
5996 
5997 	info = &ctx->ifc_filter_info;
5998 	tqrid = rid;
5999 
6000 	switch (type) {
6001 	/* XXX merge tx/rx for netmap? */
6002 	case IFLIB_INTR_TX:
6003 		q = &ctx->ifc_txqs[qid];
6004 		info = &ctx->ifc_txqs[qid].ift_filter_info;
6005 		gtask = &ctx->ifc_txqs[qid].ift_task;
6006 		tqg = qgroup_if_io_tqg;
6007 		fn = _task_fn_tx;
6008 		intr_fast = iflib_fast_intr;
6009 		GROUPTASK_INIT(gtask, 0, fn, q);
6010 		ctx->ifc_flags |= IFC_NETMAP_TX_IRQ;
6011 		break;
6012 	case IFLIB_INTR_RX:
6013 		q = &ctx->ifc_rxqs[qid];
6014 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
6015 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6016 		tqg = qgroup_if_io_tqg;
6017 		fn = _task_fn_rx;
6018 		intr_fast = iflib_fast_intr;
6019 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6020 		break;
6021 	case IFLIB_INTR_RXTX:
6022 		q = &ctx->ifc_rxqs[qid];
6023 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
6024 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6025 		tqg = qgroup_if_io_tqg;
6026 		fn = _task_fn_rx;
6027 		intr_fast = iflib_fast_intr_rxtx;
6028 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6029 		break;
6030 	case IFLIB_INTR_ADMIN:
6031 		q = ctx;
6032 		tqrid = -1;
6033 		info = &ctx->ifc_filter_info;
6034 		gtask = &ctx->ifc_admin_task;
6035 		tqg = qgroup_if_config_tqg;
6036 		fn = _task_fn_admin;
6037 		intr_fast = iflib_fast_intr_ctx;
6038 		break;
6039 	default:
6040 		device_printf(ctx->ifc_dev, "%s: unknown net intr type\n",
6041 		    __func__);
6042 		return (EINVAL);
6043 	}
6044 
6045 	info->ifi_filter = filter;
6046 	info->ifi_filter_arg = filter_arg;
6047 	info->ifi_task = gtask;
6048 	info->ifi_ctx = q;
6049 
6050 	dev = ctx->ifc_dev;
6051 	err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info,  name);
6052 	if (err != 0) {
6053 		device_printf(dev, "_iflib_irq_alloc failed %d\n", err);
6054 		return (err);
6055 	}
6056 	if (type == IFLIB_INTR_ADMIN)
6057 		return (0);
6058 
6059 	if (tqrid != -1) {
6060 		err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg,
6061 		    q, name);
6062 		if (err)
6063 			return (err);
6064 	} else {
6065 		taskqgroup_attach(tqg, gtask, q, dev, irq->ii_res, name);
6066 	}
6067 
6068 	return (0);
6069 }
6070 
6071 void
6072 iflib_softirq_alloc_generic(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type, void *arg, int qid, const char *name)
6073 {
6074 	struct grouptask *gtask;
6075 	struct taskqgroup *tqg;
6076 	gtask_fn_t *fn;
6077 	void *q;
6078 	int err;
6079 
6080 	switch (type) {
6081 	case IFLIB_INTR_TX:
6082 		q = &ctx->ifc_txqs[qid];
6083 		gtask = &ctx->ifc_txqs[qid].ift_task;
6084 		tqg = qgroup_if_io_tqg;
6085 		fn = _task_fn_tx;
6086 		GROUPTASK_INIT(gtask, 0, fn, q);
6087 		break;
6088 	case IFLIB_INTR_RX:
6089 		q = &ctx->ifc_rxqs[qid];
6090 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6091 		tqg = qgroup_if_io_tqg;
6092 		fn = _task_fn_rx;
6093 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6094 		break;
6095 	case IFLIB_INTR_IOV:
6096 		q = ctx;
6097 		gtask = &ctx->ifc_vflr_task;
6098 		tqg = qgroup_if_config_tqg;
6099 		fn = _task_fn_iov;
6100 		GROUPTASK_INIT(gtask, 0, fn, q);
6101 		break;
6102 	default:
6103 		panic("unknown net intr type");
6104 	}
6105 	if (irq != NULL) {
6106 		err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg,
6107 		    q, name);
6108 		if (err)
6109 			taskqgroup_attach(tqg, gtask, q, ctx->ifc_dev,
6110 			    irq->ii_res, name);
6111 	} else {
6112 		taskqgroup_attach(tqg, gtask, q, NULL, NULL, name);
6113 	}
6114 }
6115 
6116 void
6117 iflib_irq_free(if_ctx_t ctx, if_irq_t irq)
6118 {
6119 
6120 	if (irq->ii_tag)
6121 		bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag);
6122 
6123 	if (irq->ii_res)
6124 		bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ,
6125 		    rman_get_rid(irq->ii_res), irq->ii_res);
6126 }
6127 
6128 static int
6129 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, const char *name)
6130 {
6131 	iflib_txq_t txq = ctx->ifc_txqs;
6132 	iflib_rxq_t rxq = ctx->ifc_rxqs;
6133 	if_irq_t irq = &ctx->ifc_legacy_irq;
6134 	iflib_filter_info_t info;
6135 	device_t dev;
6136 	struct grouptask *gtask;
6137 	struct resource *res;
6138 	struct taskqgroup *tqg;
6139 	void *q;
6140 	int err, tqrid;
6141 	bool rx_only;
6142 
6143 	q = &ctx->ifc_rxqs[0];
6144 	info = &rxq[0].ifr_filter_info;
6145 	gtask = &rxq[0].ifr_task;
6146 	tqg = qgroup_if_io_tqg;
6147 	tqrid = *rid;
6148 	rx_only = (ctx->ifc_sctx->isc_flags & IFLIB_SINGLE_IRQ_RX_ONLY) != 0;
6149 
6150 	ctx->ifc_flags |= IFC_LEGACY;
6151 	info->ifi_filter = filter;
6152 	info->ifi_filter_arg = filter_arg;
6153 	info->ifi_task = gtask;
6154 	info->ifi_ctx = rx_only ? ctx : q;
6155 
6156 	dev = ctx->ifc_dev;
6157 	/* We allocate a single interrupt resource */
6158 	err = _iflib_irq_alloc(ctx, irq, tqrid, rx_only ? iflib_fast_intr_ctx :
6159 	    iflib_fast_intr_rxtx, NULL, info, name);
6160 	if (err != 0)
6161 		return (err);
6162 	NET_GROUPTASK_INIT(gtask, 0, _task_fn_rx, q);
6163 	res = irq->ii_res;
6164 	taskqgroup_attach(tqg, gtask, q, dev, res, name);
6165 
6166 	GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq);
6167 	taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, dev, res,
6168 	    "tx");
6169 	return (0);
6170 }
6171 
6172 void
6173 iflib_led_create(if_ctx_t ctx)
6174 {
6175 
6176 	ctx->ifc_led_dev = led_create(iflib_led_func, ctx,
6177 	    device_get_nameunit(ctx->ifc_dev));
6178 }
6179 
6180 void
6181 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid)
6182 {
6183 
6184 	GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
6185 }
6186 
6187 void
6188 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid)
6189 {
6190 
6191 	GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task);
6192 }
6193 
6194 void
6195 iflib_admin_intr_deferred(if_ctx_t ctx)
6196 {
6197 
6198 	MPASS(ctx->ifc_admin_task.gt_taskqueue != NULL);
6199 	GROUPTASK_ENQUEUE(&ctx->ifc_admin_task);
6200 }
6201 
6202 void
6203 iflib_iov_intr_deferred(if_ctx_t ctx)
6204 {
6205 
6206 	GROUPTASK_ENQUEUE(&ctx->ifc_vflr_task);
6207 }
6208 
6209 void
6210 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, const char *name)
6211 {
6212 
6213 	taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, NULL, NULL,
6214 	    name);
6215 }
6216 
6217 void
6218 iflib_config_gtask_init(void *ctx, struct grouptask *gtask, gtask_fn_t *fn,
6219 	const char *name)
6220 {
6221 
6222 	GROUPTASK_INIT(gtask, 0, fn, ctx);
6223 	taskqgroup_attach(qgroup_if_config_tqg, gtask, gtask, NULL, NULL,
6224 	    name);
6225 }
6226 
6227 void
6228 iflib_config_gtask_deinit(struct grouptask *gtask)
6229 {
6230 
6231 	taskqgroup_detach(qgroup_if_config_tqg, gtask);
6232 }
6233 
6234 void
6235 iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate)
6236 {
6237 	if_t ifp = ctx->ifc_ifp;
6238 	iflib_txq_t txq = ctx->ifc_txqs;
6239 
6240 	if_setbaudrate(ifp, baudrate);
6241 	if (baudrate >= IF_Gbps(10)) {
6242 		STATE_LOCK(ctx);
6243 		ctx->ifc_flags |= IFC_PREFETCH;
6244 		STATE_UNLOCK(ctx);
6245 	}
6246 	/* If link down, disable watchdog */
6247 	if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) {
6248 		for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++)
6249 			txq->ift_qstatus = IFLIB_QUEUE_IDLE;
6250 	}
6251 	ctx->ifc_link_state = link_state;
6252 	if_link_state_change(ifp, link_state);
6253 }
6254 
6255 static int
6256 iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq)
6257 {
6258 	int credits;
6259 #ifdef INVARIANTS
6260 	int credits_pre = txq->ift_cidx_processed;
6261 #endif
6262 
6263 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
6264 	    BUS_DMASYNC_POSTREAD);
6265 	if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, true)) == 0)
6266 		return (0);
6267 
6268 	txq->ift_processed += credits;
6269 	txq->ift_cidx_processed += credits;
6270 
6271 	MPASS(credits_pre + credits == txq->ift_cidx_processed);
6272 	if (txq->ift_cidx_processed >= txq->ift_size)
6273 		txq->ift_cidx_processed -= txq->ift_size;
6274 	return (credits);
6275 }
6276 
6277 static int
6278 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget)
6279 {
6280 	iflib_fl_t fl;
6281 	u_int i;
6282 
6283 	for (i = 0, fl = &rxq->ifr_fl[0]; i < rxq->ifr_nfl; i++, fl++)
6284 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
6285 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
6286 	return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx,
6287 	    budget));
6288 }
6289 
6290 void
6291 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name,
6292 	const char *description, if_int_delay_info_t info,
6293 	int offset, int value)
6294 {
6295 	info->iidi_ctx = ctx;
6296 	info->iidi_offset = offset;
6297 	info->iidi_value = value;
6298 	SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev),
6299 	    SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)),
6300 	    OID_AUTO, name, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE,
6301 	    info, 0, iflib_sysctl_int_delay, "I", description);
6302 }
6303 
6304 struct sx *
6305 iflib_ctx_lock_get(if_ctx_t ctx)
6306 {
6307 
6308 	return (&ctx->ifc_ctx_sx);
6309 }
6310 
6311 static int
6312 iflib_msix_init(if_ctx_t ctx)
6313 {
6314 	device_t dev = ctx->ifc_dev;
6315 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6316 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6317 	int admincnt, bar, err, iflib_num_rx_queues, iflib_num_tx_queues;
6318 	int msgs, queuemsgs, queues, rx_queues, tx_queues, vectors;
6319 
6320 	iflib_num_tx_queues = ctx->ifc_sysctl_ntxqs;
6321 	iflib_num_rx_queues = ctx->ifc_sysctl_nrxqs;
6322 
6323 	if (bootverbose)
6324 		device_printf(dev, "msix_init qsets capped at %d\n",
6325 		    imax(scctx->isc_ntxqsets, scctx->isc_nrxqsets));
6326 
6327 	/* Override by tuneable */
6328 	if (scctx->isc_disable_msix)
6329 		goto msi;
6330 
6331 	/* First try MSI-X */
6332 	if ((msgs = pci_msix_count(dev)) == 0) {
6333 		if (bootverbose)
6334 			device_printf(dev, "MSI-X not supported or disabled\n");
6335 		goto msi;
6336 	}
6337 
6338 	bar = ctx->ifc_softc_ctx.isc_msix_bar;
6339 	/*
6340 	 * bar == -1 => "trust me I know what I'm doing"
6341 	 * Some drivers are for hardware that is so shoddily
6342 	 * documented that no one knows which bars are which
6343 	 * so the developer has to map all bars. This hack
6344 	 * allows shoddy garbage to use MSI-X in this framework.
6345 	 */
6346 	if (bar != -1) {
6347 		ctx->ifc_msix_mem = bus_alloc_resource_any(dev,
6348 	            SYS_RES_MEMORY, &bar, RF_ACTIVE);
6349 		if (ctx->ifc_msix_mem == NULL) {
6350 			device_printf(dev, "Unable to map MSI-X table\n");
6351 			goto msi;
6352 		}
6353 	}
6354 
6355 	admincnt = sctx->isc_admin_intrcnt;
6356 #if IFLIB_DEBUG
6357 	/* use only 1 qset in debug mode */
6358 	queuemsgs = min(msgs - admincnt, 1);
6359 #else
6360 	queuemsgs = msgs - admincnt;
6361 #endif
6362 #ifdef RSS
6363 	queues = imin(queuemsgs, rss_getnumbuckets());
6364 #else
6365 	queues = queuemsgs;
6366 #endif
6367 	queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues);
6368 	if (bootverbose)
6369 		device_printf(dev,
6370 		    "intr CPUs: %d queue msgs: %d admincnt: %d\n",
6371 		    CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt);
6372 #ifdef  RSS
6373 	/* If we're doing RSS, clamp at the number of RSS buckets */
6374 	if (queues > rss_getnumbuckets())
6375 		queues = rss_getnumbuckets();
6376 #endif
6377 	if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt)
6378 		rx_queues = iflib_num_rx_queues;
6379 	else
6380 		rx_queues = queues;
6381 
6382 	if (rx_queues > scctx->isc_nrxqsets)
6383 		rx_queues = scctx->isc_nrxqsets;
6384 
6385 	/*
6386 	 * We want this to be all logical CPUs by default
6387 	 */
6388 	if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues)
6389 		tx_queues = iflib_num_tx_queues;
6390 	else
6391 		tx_queues = mp_ncpus;
6392 
6393 	if (tx_queues > scctx->isc_ntxqsets)
6394 		tx_queues = scctx->isc_ntxqsets;
6395 
6396 	if (ctx->ifc_sysctl_qs_eq_override == 0) {
6397 #ifdef INVARIANTS
6398 		if (tx_queues != rx_queues)
6399 			device_printf(dev,
6400 			    "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n",
6401 			    min(rx_queues, tx_queues), min(rx_queues, tx_queues));
6402 #endif
6403 		tx_queues = min(rx_queues, tx_queues);
6404 		rx_queues = min(rx_queues, tx_queues);
6405 	}
6406 
6407 	vectors = rx_queues + admincnt;
6408 	if (msgs < vectors) {
6409 		device_printf(dev,
6410 		    "insufficient number of MSI-X vectors "
6411 		    "(supported %d, need %d)\n", msgs, vectors);
6412 		goto msi;
6413 	}
6414 
6415 	device_printf(dev, "Using %d RX queues %d TX queues\n", rx_queues,
6416 	    tx_queues);
6417 	msgs = vectors;
6418 	if ((err = pci_alloc_msix(dev, &vectors)) == 0) {
6419 		if (vectors != msgs) {
6420 			device_printf(dev,
6421 			    "Unable to allocate sufficient MSI-X vectors "
6422 			    "(got %d, need %d)\n", vectors, msgs);
6423 			pci_release_msi(dev);
6424 			if (bar != -1) {
6425 				bus_release_resource(dev, SYS_RES_MEMORY, bar,
6426 				    ctx->ifc_msix_mem);
6427 				ctx->ifc_msix_mem = NULL;
6428 			}
6429 			goto msi;
6430 		}
6431 		device_printf(dev, "Using MSI-X interrupts with %d vectors\n",
6432 		    vectors);
6433 		scctx->isc_vectors = vectors;
6434 		scctx->isc_nrxqsets = rx_queues;
6435 		scctx->isc_ntxqsets = tx_queues;
6436 		scctx->isc_intr = IFLIB_INTR_MSIX;
6437 
6438 		return (vectors);
6439 	} else {
6440 		device_printf(dev,
6441 		    "failed to allocate %d MSI-X vectors, err: %d\n", vectors,
6442 		    err);
6443 		if (bar != -1) {
6444 			bus_release_resource(dev, SYS_RES_MEMORY, bar,
6445 			    ctx->ifc_msix_mem);
6446 			ctx->ifc_msix_mem = NULL;
6447 		}
6448 	}
6449 
6450 msi:
6451 	vectors = pci_msi_count(dev);
6452 	scctx->isc_nrxqsets = 1;
6453 	scctx->isc_ntxqsets = 1;
6454 	scctx->isc_vectors = vectors;
6455 	if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) {
6456 		device_printf(dev,"Using an MSI interrupt\n");
6457 		scctx->isc_intr = IFLIB_INTR_MSI;
6458 	} else {
6459 		scctx->isc_vectors = 1;
6460 		device_printf(dev,"Using a Legacy interrupt\n");
6461 		scctx->isc_intr = IFLIB_INTR_LEGACY;
6462 	}
6463 
6464 	return (vectors);
6465 }
6466 
6467 static const char *ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" };
6468 
6469 static int
6470 mp_ring_state_handler(SYSCTL_HANDLER_ARGS)
6471 {
6472 	int rc;
6473 	uint16_t *state = ((uint16_t *)oidp->oid_arg1);
6474 	struct sbuf *sb;
6475 	const char *ring_state = "UNKNOWN";
6476 
6477 	/* XXX needed ? */
6478 	rc = sysctl_wire_old_buffer(req, 0);
6479 	MPASS(rc == 0);
6480 	if (rc != 0)
6481 		return (rc);
6482 	sb = sbuf_new_for_sysctl(NULL, NULL, 80, req);
6483 	MPASS(sb != NULL);
6484 	if (sb == NULL)
6485 		return (ENOMEM);
6486 	if (state[3] <= 3)
6487 		ring_state = ring_states[state[3]];
6488 
6489 	sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s",
6490 		    state[0], state[1], state[2], ring_state);
6491 	rc = sbuf_finish(sb);
6492 	sbuf_delete(sb);
6493         return(rc);
6494 }
6495 
6496 enum iflib_ndesc_handler {
6497 	IFLIB_NTXD_HANDLER,
6498 	IFLIB_NRXD_HANDLER,
6499 };
6500 
6501 static int
6502 mp_ndesc_handler(SYSCTL_HANDLER_ARGS)
6503 {
6504 	if_ctx_t ctx = (void *)arg1;
6505 	enum iflib_ndesc_handler type = arg2;
6506 	char buf[256] = {0};
6507 	qidx_t *ndesc;
6508 	char *p, *next;
6509 	int nqs, rc, i;
6510 
6511 	nqs = 8;
6512 	switch(type) {
6513 	case IFLIB_NTXD_HANDLER:
6514 		ndesc = ctx->ifc_sysctl_ntxds;
6515 		if (ctx->ifc_sctx)
6516 			nqs = ctx->ifc_sctx->isc_ntxqs;
6517 		break;
6518 	case IFLIB_NRXD_HANDLER:
6519 		ndesc = ctx->ifc_sysctl_nrxds;
6520 		if (ctx->ifc_sctx)
6521 			nqs = ctx->ifc_sctx->isc_nrxqs;
6522 		break;
6523 	default:
6524 		printf("%s: unhandled type\n", __func__);
6525 		return (EINVAL);
6526 	}
6527 	if (nqs == 0)
6528 		nqs = 8;
6529 
6530 	for (i=0; i<8; i++) {
6531 		if (i >= nqs)
6532 			break;
6533 		if (i)
6534 			strcat(buf, ",");
6535 		sprintf(strchr(buf, 0), "%d", ndesc[i]);
6536 	}
6537 
6538 	rc = sysctl_handle_string(oidp, buf, sizeof(buf), req);
6539 	if (rc || req->newptr == NULL)
6540 		return rc;
6541 
6542 	for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p;
6543 	    i++, p = strsep(&next, " ,")) {
6544 		ndesc[i] = strtoul(p, NULL, 10);
6545 	}
6546 
6547 	return(rc);
6548 }
6549 
6550 #define NAME_BUFLEN 32
6551 static void
6552 iflib_add_device_sysctl_pre(if_ctx_t ctx)
6553 {
6554         device_t dev = iflib_get_dev(ctx);
6555 	struct sysctl_oid_list *child, *oid_list;
6556 	struct sysctl_ctx_list *ctx_list;
6557 	struct sysctl_oid *node;
6558 
6559 	ctx_list = device_get_sysctl_ctx(dev);
6560 	child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev));
6561 	ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, "iflib",
6562 	    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "IFLIB fields");
6563 	oid_list = SYSCTL_CHILDREN(node);
6564 
6565 	SYSCTL_ADD_CONST_STRING(ctx_list, oid_list, OID_AUTO, "driver_version",
6566 		       CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version,
6567 		       "driver version");
6568 
6569 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs",
6570 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0,
6571 			"# of txqs to use, 0 => use default #");
6572 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs",
6573 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0,
6574 			"# of rxqs to use, 0 => use default #");
6575 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_qs_enable",
6576 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0,
6577                        "permit #txq != #rxq");
6578 	SYSCTL_ADD_INT(ctx_list, oid_list, OID_AUTO, "disable_msix",
6579                       CTLFLAG_RWTUN, &ctx->ifc_softc_ctx.isc_disable_msix, 0,
6580                       "disable MSI-X (default 0)");
6581 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "rx_budget",
6582 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_rx_budget, 0,
6583 		       "set the RX budget");
6584 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "tx_abdicate",
6585 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_tx_abdicate, 0,
6586 		       "cause TX to abdicate instead of running to completion");
6587 	ctx->ifc_sysctl_core_offset = CORE_OFFSET_UNSPECIFIED;
6588 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "core_offset",
6589 		       CTLFLAG_RDTUN, &ctx->ifc_sysctl_core_offset, 0,
6590 		       "offset to start using cores at");
6591 	SYSCTL_ADD_U8(ctx_list, oid_list, OID_AUTO, "separate_txrx",
6592 		       CTLFLAG_RDTUN, &ctx->ifc_sysctl_separate_txrx, 0,
6593 		       "use separate cores for TX and RX");
6594 
6595 	/* XXX change for per-queue sizes */
6596 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_ntxds",
6597 	    CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NEEDGIANT, ctx,
6598 	    IFLIB_NTXD_HANDLER, mp_ndesc_handler, "A",
6599 	    "list of # of TX descriptors to use, 0 = use default #");
6600 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_nrxds",
6601 	    CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NEEDGIANT, ctx,
6602 	    IFLIB_NRXD_HANDLER, mp_ndesc_handler, "A",
6603 	    "list of # of RX descriptors to use, 0 = use default #");
6604 }
6605 
6606 static void
6607 iflib_add_device_sysctl_post(if_ctx_t ctx)
6608 {
6609 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6610 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6611         device_t dev = iflib_get_dev(ctx);
6612 	struct sysctl_oid_list *child;
6613 	struct sysctl_ctx_list *ctx_list;
6614 	iflib_fl_t fl;
6615 	iflib_txq_t txq;
6616 	iflib_rxq_t rxq;
6617 	int i, j;
6618 	char namebuf[NAME_BUFLEN];
6619 	char *qfmt;
6620 	struct sysctl_oid *queue_node, *fl_node, *node;
6621 	struct sysctl_oid_list *queue_list, *fl_list;
6622 	ctx_list = device_get_sysctl_ctx(dev);
6623 
6624 	node = ctx->ifc_sysctl_node;
6625 	child = SYSCTL_CHILDREN(node);
6626 
6627 	if (scctx->isc_ntxqsets > 100)
6628 		qfmt = "txq%03d";
6629 	else if (scctx->isc_ntxqsets > 10)
6630 		qfmt = "txq%02d";
6631 	else
6632 		qfmt = "txq%d";
6633 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
6634 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6635 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6636 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "Queue Name");
6637 		queue_list = SYSCTL_CHILDREN(queue_node);
6638 #if MEMORY_LOGGING
6639 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued",
6640 				CTLFLAG_RD,
6641 				&txq->ift_dequeued, "total mbufs freed");
6642 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued",
6643 				CTLFLAG_RD,
6644 				&txq->ift_enqueued, "total mbufs enqueued");
6645 #endif
6646 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag",
6647 				   CTLFLAG_RD,
6648 				   &txq->ift_mbuf_defrag, "# of times m_defrag was called");
6649 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "m_pullups",
6650 				   CTLFLAG_RD,
6651 				   &txq->ift_pullups, "# of times m_pullup was called");
6652 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag_failed",
6653 				   CTLFLAG_RD,
6654 				   &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed");
6655 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_desc_avail",
6656 				   CTLFLAG_RD,
6657 				   &txq->ift_no_desc_avail, "# of times no descriptors were available");
6658 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "tx_map_failed",
6659 				   CTLFLAG_RD,
6660 				   &txq->ift_map_failed, "# of times DMA map failed");
6661 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txd_encap_efbig",
6662 				   CTLFLAG_RD,
6663 				   &txq->ift_txd_encap_efbig, "# of times txd_encap returned EFBIG");
6664 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_tx_dma_setup",
6665 				   CTLFLAG_RD,
6666 				   &txq->ift_no_tx_dma_setup, "# of times map failed for other than EFBIG");
6667 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx",
6668 				   CTLFLAG_RD,
6669 				   &txq->ift_pidx, 1, "Producer Index");
6670 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx",
6671 				   CTLFLAG_RD,
6672 				   &txq->ift_cidx, 1, "Consumer Index");
6673 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx_processed",
6674 				   CTLFLAG_RD,
6675 				   &txq->ift_cidx_processed, 1, "Consumer Index seen by credit update");
6676 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use",
6677 				   CTLFLAG_RD,
6678 				   &txq->ift_in_use, 1, "descriptors in use");
6679 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_processed",
6680 				   CTLFLAG_RD,
6681 				   &txq->ift_processed, "descriptors procesed for clean");
6682 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned",
6683 				   CTLFLAG_RD,
6684 				   &txq->ift_cleaned, "total cleaned");
6685 		SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state",
6686 		    CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_NEEDGIANT,
6687 		    __DEVOLATILE(uint64_t *, &txq->ift_br->state), 0,
6688 		    mp_ring_state_handler, "A", "soft ring state");
6689 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_enqueues",
6690 				       CTLFLAG_RD, &txq->ift_br->enqueues,
6691 				       "# of enqueues to the mp_ring for this queue");
6692 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_drops",
6693 				       CTLFLAG_RD, &txq->ift_br->drops,
6694 				       "# of drops in the mp_ring for this queue");
6695 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_starts",
6696 				       CTLFLAG_RD, &txq->ift_br->starts,
6697 				       "# of normal consumer starts in the mp_ring for this queue");
6698 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_stalls",
6699 				       CTLFLAG_RD, &txq->ift_br->stalls,
6700 					       "# of consumer stalls in the mp_ring for this queue");
6701 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_restarts",
6702 			       CTLFLAG_RD, &txq->ift_br->restarts,
6703 				       "# of consumer restarts in the mp_ring for this queue");
6704 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_abdications",
6705 				       CTLFLAG_RD, &txq->ift_br->abdications,
6706 				       "# of consumer abdications in the mp_ring for this queue");
6707 	}
6708 
6709 	if (scctx->isc_nrxqsets > 100)
6710 		qfmt = "rxq%03d";
6711 	else if (scctx->isc_nrxqsets > 10)
6712 		qfmt = "rxq%02d";
6713 	else
6714 		qfmt = "rxq%d";
6715 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
6716 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6717 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6718 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "Queue Name");
6719 		queue_list = SYSCTL_CHILDREN(queue_node);
6720 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
6721 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_cidx",
6722 				       CTLFLAG_RD,
6723 				       &rxq->ifr_cq_cidx, 1, "Consumer Index");
6724 		}
6725 
6726 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
6727 			snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j);
6728 			fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list, OID_AUTO, namebuf,
6729 			    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "freelist Name");
6730 			fl_list = SYSCTL_CHILDREN(fl_node);
6731 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx",
6732 				       CTLFLAG_RD,
6733 				       &fl->ifl_pidx, 1, "Producer Index");
6734 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx",
6735 				       CTLFLAG_RD,
6736 				       &fl->ifl_cidx, 1, "Consumer Index");
6737 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits",
6738 				       CTLFLAG_RD,
6739 				       &fl->ifl_credits, 1, "credits available");
6740 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "buf_size",
6741 				       CTLFLAG_RD,
6742 				       &fl->ifl_buf_size, 1, "buffer size");
6743 #if MEMORY_LOGGING
6744 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_enqueued",
6745 					CTLFLAG_RD,
6746 					&fl->ifl_m_enqueued, "mbufs allocated");
6747 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_dequeued",
6748 					CTLFLAG_RD,
6749 					&fl->ifl_m_dequeued, "mbufs freed");
6750 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_enqueued",
6751 					CTLFLAG_RD,
6752 					&fl->ifl_cl_enqueued, "clusters allocated");
6753 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_dequeued",
6754 					CTLFLAG_RD,
6755 					&fl->ifl_cl_dequeued, "clusters freed");
6756 #endif
6757 
6758 		}
6759 	}
6760 
6761 }
6762 
6763 void
6764 iflib_request_reset(if_ctx_t ctx)
6765 {
6766 
6767 	STATE_LOCK(ctx);
6768 	ctx->ifc_flags |= IFC_DO_RESET;
6769 	STATE_UNLOCK(ctx);
6770 }
6771 
6772 #ifndef __NO_STRICT_ALIGNMENT
6773 static struct mbuf *
6774 iflib_fixup_rx(struct mbuf *m)
6775 {
6776 	struct mbuf *n;
6777 
6778 	if (m->m_len <= (MCLBYTES - ETHER_HDR_LEN)) {
6779 		bcopy(m->m_data, m->m_data + ETHER_HDR_LEN, m->m_len);
6780 		m->m_data += ETHER_HDR_LEN;
6781 		n = m;
6782 	} else {
6783 		MGETHDR(n, M_NOWAIT, MT_DATA);
6784 		if (n == NULL) {
6785 			m_freem(m);
6786 			return (NULL);
6787 		}
6788 		bcopy(m->m_data, n->m_data, ETHER_HDR_LEN);
6789 		m->m_data += ETHER_HDR_LEN;
6790 		m->m_len -= ETHER_HDR_LEN;
6791 		n->m_len = ETHER_HDR_LEN;
6792 		M_MOVE_PKTHDR(n, m);
6793 		n->m_next = m;
6794 	}
6795 	return (n);
6796 }
6797 #endif
6798 
6799 #ifdef DEBUGNET
6800 static void
6801 iflib_debugnet_init(if_t ifp, int *nrxr, int *ncl, int *clsize)
6802 {
6803 	if_ctx_t ctx;
6804 
6805 	ctx = if_getsoftc(ifp);
6806 	CTX_LOCK(ctx);
6807 	*nrxr = NRXQSETS(ctx);
6808 	*ncl = ctx->ifc_rxqs[0].ifr_fl->ifl_size;
6809 	*clsize = ctx->ifc_rxqs[0].ifr_fl->ifl_buf_size;
6810 	CTX_UNLOCK(ctx);
6811 }
6812 
6813 static void
6814 iflib_debugnet_event(if_t ifp, enum debugnet_ev event)
6815 {
6816 	if_ctx_t ctx;
6817 	if_softc_ctx_t scctx;
6818 	iflib_fl_t fl;
6819 	iflib_rxq_t rxq;
6820 	int i, j;
6821 
6822 	ctx = if_getsoftc(ifp);
6823 	scctx = &ctx->ifc_softc_ctx;
6824 
6825 	switch (event) {
6826 	case DEBUGNET_START:
6827 		for (i = 0; i < scctx->isc_nrxqsets; i++) {
6828 			rxq = &ctx->ifc_rxqs[i];
6829 			for (j = 0; j < rxq->ifr_nfl; j++) {
6830 				fl = rxq->ifr_fl;
6831 				fl->ifl_zone = m_getzone(fl->ifl_buf_size);
6832 			}
6833 		}
6834 		iflib_no_tx_batch = 1;
6835 		break;
6836 	default:
6837 		break;
6838 	}
6839 }
6840 
6841 static int
6842 iflib_debugnet_transmit(if_t ifp, struct mbuf *m)
6843 {
6844 	if_ctx_t ctx;
6845 	iflib_txq_t txq;
6846 	int error;
6847 
6848 	ctx = if_getsoftc(ifp);
6849 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
6850 	    IFF_DRV_RUNNING)
6851 		return (EBUSY);
6852 
6853 	txq = &ctx->ifc_txqs[0];
6854 	error = iflib_encap(txq, &m);
6855 	if (error == 0)
6856 		(void)iflib_txd_db_check(ctx, txq, true, txq->ift_in_use);
6857 	return (error);
6858 }
6859 
6860 static int
6861 iflib_debugnet_poll(if_t ifp, int count)
6862 {
6863 	struct epoch_tracker et;
6864 	if_ctx_t ctx;
6865 	if_softc_ctx_t scctx;
6866 	iflib_txq_t txq;
6867 	int i;
6868 
6869 	ctx = if_getsoftc(ifp);
6870 	scctx = &ctx->ifc_softc_ctx;
6871 
6872 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
6873 	    IFF_DRV_RUNNING)
6874 		return (EBUSY);
6875 
6876 	txq = &ctx->ifc_txqs[0];
6877 	(void)iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
6878 
6879 	NET_EPOCH_ENTER(et);
6880 	for (i = 0; i < scctx->isc_nrxqsets; i++)
6881 		(void)iflib_rxeof(&ctx->ifc_rxqs[i], 16 /* XXX */);
6882 	NET_EPOCH_EXIT(et);
6883 	return (0);
6884 }
6885 #endif /* DEBUGNET */
6886