xref: /freebsd/sys/net/iflib.c (revision b54dcb897a5fa66ff1013d0ea403ed8894e34b8a)
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 #include "opt_inet.h"
30 #include "opt_inet6.h"
31 #include "opt_acpi.h"
32 
33 #include <sys/param.h>
34 #include <sys/types.h>
35 #include <sys/bus.h>
36 #include <sys/eventhandler.h>
37 #include <sys/kernel.h>
38 #include <sys/lock.h>
39 #include <sys/mutex.h>
40 #include <sys/module.h>
41 #include <sys/kobj.h>
42 #include <sys/proc.h>
43 #include <sys/rman.h>
44 #include <sys/sbuf.h>
45 #include <sys/sched.h>
46 #include <sys/smp.h>
47 #include <sys/socket.h>
48 #include <sys/sockio.h>
49 #include <sys/sysctl.h>
50 #include <sys/syslog.h>
51 #include <sys/taskqueue.h>
52 #include <sys/limits.h>
53 
54 #include <net/if.h>
55 #include <net/if_var.h>
56 #include <net/if_private.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/udp.h>
75 #include <netinet/ip_var.h>
76 #include <netinet6/ip6_var.h>
77 
78 #include <machine/bus.h>
79 #include <machine/in_cksum.h>
80 
81 #include <vm/vm.h>
82 #include <vm/pmap.h>
83 
84 #include <dev/led/led.h>
85 #include <dev/pci/pcireg.h>
86 #include <dev/pci/pcivar.h>
87 #include <dev/pci/pci_private.h>
88 
89 #include <net/iflib.h>
90 
91 #include "ifdi_if.h"
92 
93 #ifdef PCI_IOV
94 #include <dev/pci/pci_iov.h>
95 #endif
96 
97 #include <sys/bitstring.h>
98 /*
99  * enable accounting of every mbuf as it comes in to and goes out of
100  * iflib's software descriptor references
101  */
102 #define MEMORY_LOGGING 0
103 /*
104  * Enable mbuf vectors for compressing long mbuf chains
105  */
106 
107 /*
108  * NB:
109  * - Prefetching in tx cleaning should perhaps be a tunable. The distance ahead
110  *   we prefetch needs to be determined by the time spent in m_free vis a vis
111  *   the cost of a prefetch. This will of course vary based on the workload:
112  *      - NFLX's m_free path is dominated by vm-based M_EXT manipulation which
113  *        is quite expensive, thus suggesting very little prefetch.
114  *      - small packet forwarding which is just returning a single mbuf to
115  *        UMA will typically be very fast vis a vis the cost of a memory
116  *        access.
117  */
118 
119 /*
120  * File organization:
121  *  - private structures
122  *  - iflib private utility functions
123  *  - ifnet functions
124  *  - vlan registry and other exported functions
125  *  - iflib public core functions
126  *
127  *
128  */
129 static MALLOC_DEFINE(M_IFLIB, "iflib", "ifnet library");
130 
131 #define	IFLIB_RXEOF_MORE	(1U << 0)
132 #define	IFLIB_RXEOF_EMPTY	(2U << 0)
133 
134 struct iflib_txq;
135 typedef struct iflib_txq *iflib_txq_t;
136 struct iflib_rxq;
137 typedef struct iflib_rxq *iflib_rxq_t;
138 struct iflib_fl;
139 typedef struct iflib_fl *iflib_fl_t;
140 
141 struct iflib_ctx;
142 
143 static void iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid);
144 static void iflib_timer(void *arg);
145 static void iflib_tqg_detach(if_ctx_t ctx);
146 #ifndef ALTQ
147 static int  iflib_simple_transmit(if_t ifp, struct mbuf *m);
148 #endif
149 
150 typedef struct iflib_filter_info {
151 	driver_filter_t *ifi_filter;
152 	void *ifi_filter_arg;
153 	struct grouptask *ifi_task;
154 	void *ifi_ctx;
155 } *iflib_filter_info_t;
156 
157 struct iflib_ctx {
158 	KOBJ_FIELDS;
159 	/*
160 	 * Pointer to hardware driver's softc
161 	 */
162 	void *ifc_softc;
163 	device_t ifc_dev;
164 	if_t ifc_ifp;
165 
166 	cpuset_t ifc_cpus;
167 	if_shared_ctx_t ifc_sctx;
168 	struct if_softc_ctx ifc_softc_ctx;
169 
170 	struct sx ifc_ctx_sx;
171 	struct mtx ifc_state_mtx;
172 
173 	iflib_txq_t ifc_txqs;
174 	iflib_rxq_t ifc_rxqs;
175 	uint32_t ifc_if_flags;
176 	uint32_t ifc_flags;
177 	uint32_t ifc_max_fl_buf_size;
178 	uint32_t ifc_rx_mbuf_sz;
179 
180 	int ifc_link_state;
181 	uint32_t ifc_tx_watchdog_events;
182 	struct cdev *ifc_led_dev;
183 	int ifc_led_state;
184 	struct resource *ifc_msix_mem;
185 
186 	struct if_irq ifc_legacy_irq;
187 	struct task ifc_admin_task;
188 	struct task ifc_led_task;
189 	struct task ifc_vflr_task;
190 	struct taskqueue *ifc_tq;
191 	struct iflib_filter_info ifc_filter_info;
192 	struct ifmedia	ifc_media;
193 	struct ifmedia	*ifc_mediap;
194 
195 	struct sysctl_ctx_list ifc_sysctl_ctx;
196 	struct sysctl_oid *ifc_sysctl_node;
197 	uint16_t ifc_sysctl_ntxqs;
198 	uint16_t ifc_sysctl_nrxqs;
199 	uint16_t ifc_sysctl_qs_eq_override;
200 	uint16_t ifc_sysctl_rx_budget;
201 	uint16_t ifc_sysctl_tx_abdicate;
202 	uint16_t ifc_sysctl_core_offset;
203 #define	CORE_OFFSET_UNSPECIFIED	0xffff
204 	uint8_t  ifc_sysctl_separate_txrx;
205 	uint8_t  ifc_sysctl_use_logical_cores;
206 	uint16_t ifc_sysctl_extra_msix_vectors;
207 	bool     ifc_cpus_are_physical_cores;
208 	bool     ifc_sysctl_simple_tx;
209 	bool     ifc_sysctl_tx_defer_mfree;
210 	uint16_t ifc_sysctl_tx_reclaim_thresh;
211 	uint16_t ifc_sysctl_tx_reclaim_ticks;
212 
213 	qidx_t ifc_sysctl_ntxds[8];
214 	qidx_t ifc_sysctl_nrxds[8];
215 	struct if_txrx ifc_txrx;
216 #define isc_txd_encap		ifc_txrx.ift_txd_encap
217 #define isc_txd_flush		ifc_txrx.ift_txd_flush
218 #define isc_txd_credits_update	ifc_txrx.ift_txd_credits_update
219 #define isc_rxd_available	ifc_txrx.ift_rxd_available
220 #define isc_rxd_pkt_get		ifc_txrx.ift_rxd_pkt_get
221 #define isc_rxd_refill		ifc_txrx.ift_rxd_refill
222 #define isc_rxd_flush		ifc_txrx.ift_rxd_flush
223 #define isc_legacy_intr		ifc_txrx.ift_legacy_intr
224 #define isc_txq_select		ifc_txrx.ift_txq_select
225 #define isc_txq_select_v2	ifc_txrx.ift_txq_select_v2
226 
227 	eventhandler_tag ifc_vlan_attach_event;
228 	eventhandler_tag ifc_vlan_detach_event;
229 	struct ether_addr ifc_mac;
230 };
231 
232 void *
233 iflib_get_softc(if_ctx_t ctx)
234 {
235 
236 	return (ctx->ifc_softc);
237 }
238 
239 device_t
240 iflib_get_dev(if_ctx_t ctx)
241 {
242 
243 	return (ctx->ifc_dev);
244 }
245 
246 if_t
247 iflib_get_ifp(if_ctx_t ctx)
248 {
249 
250 	return (ctx->ifc_ifp);
251 }
252 
253 struct ifmedia *
254 iflib_get_media(if_ctx_t ctx)
255 {
256 
257 	return (ctx->ifc_mediap);
258 }
259 
260 void
261 iflib_set_mac(if_ctx_t ctx, uint8_t mac[ETHER_ADDR_LEN])
262 {
263 
264 	bcopy(mac, ctx->ifc_mac.octet, ETHER_ADDR_LEN);
265 }
266 
267 if_softc_ctx_t
268 iflib_get_softc_ctx(if_ctx_t ctx)
269 {
270 
271 	return (&ctx->ifc_softc_ctx);
272 }
273 
274 if_shared_ctx_t
275 iflib_get_sctx(if_ctx_t ctx)
276 {
277 
278 	return (ctx->ifc_sctx);
279 }
280 
281 uint16_t
282 iflib_get_extra_msix_vectors_sysctl(if_ctx_t ctx)
283 {
284 
285 	return (ctx->ifc_sysctl_extra_msix_vectors);
286 }
287 
288 #define IP_ALIGNED(m)		((((uintptr_t)(m)->m_data) & 0x3) == 0x2)
289 #define CACHE_PTR_INCREMENT	(CACHE_LINE_SIZE / sizeof(void *))
290 #define CACHE_PTR_NEXT(ptr)	((void *)(roundup2(ptr, CACHE_LINE_SIZE)))
291 
292 #define LINK_ACTIVE(ctx)	((ctx)->ifc_link_state == LINK_STATE_UP)
293 #define CTX_IS_VF(ctx)		((ctx)->ifc_sctx->isc_flags & IFLIB_IS_VF)
294 
295 typedef struct iflib_sw_rx_desc_array {
296 	bus_dmamap_t	*ifsd_map;	/* bus_dma maps for packet */
297 	struct mbuf	**ifsd_m;	/* pkthdr mbufs */
298 	caddr_t		*ifsd_cl;	/* direct cluster pointer for rx */
299 	bus_addr_t	*ifsd_ba;	/* bus addr of cluster for rx */
300 } iflib_rxsd_array_t;
301 
302 typedef struct iflib_sw_tx_desc_array {
303 	bus_dmamap_t	*ifsd_map;	/* bus_dma maps for packet */
304 	bus_dmamap_t	*ifsd_tso_map;	/* bus_dma maps for TSO packet */
305 	struct mbuf	**ifsd_m;	/* pkthdr mbufs */
306 	struct mbuf	**ifsd_m_defer;	/* deferred mbuf ptr */
307 	struct mbuf	**ifsd_m_deferb;/* deferred mbuf backing ptr */
308 } if_txsd_vec_t;
309 
310 /* magic number that should be high enough for any hardware */
311 #define IFLIB_MAX_TX_SEGS		128
312 #define IFLIB_RX_COPY_THRESH		128
313 #define IFLIB_MAX_RX_REFRESH		32
314 /* The minimum descriptors per second before we start coalescing */
315 #define IFLIB_MIN_DESC_SEC		16384
316 #define IFLIB_DEFAULT_TX_UPDATE_FREQ	16
317 /* maximum number of txqs that can share an rx interrupt */
318 #define IFLIB_MAX_TX_SHARED_INTR	4
319 
320 /* this should really scale with ring size - this is a fairly arbitrary value */
321 #define TX_BATCH_SIZE			32
322 
323 #define IFLIB_RESTART_BUDGET		8
324 
325 
326 /*
327  * Encode TSO or !TSO in the low bits of the tx ifsd_m pointer so as
328  * to avoid defref'ing the mbuf to determine the correct busdma resources
329  * to release
330  */
331 #define IFLIB_TSO		(1ULL << 0)
332 #define IFLIB_NO_TSO		(2ULL << 0)
333 #define IFLIB_FLAGS_MASK	(0x3ULL)
334 #define IFLIB_SAVE_MBUF(mbuf, flags)	((void *)(((uintptr_t)mbuf) | flags))
335 #define IFLIB_GET_FLAGS(a)	((uintptr_t)a & IFLIB_FLAGS_MASK)
336 #define IFLIB_GET_MBUF(a)	((struct mbuf *)((uintptr_t)a & ~IFLIB_FLAGS_MASK))
337 
338 
339 #define	IFC_LEGACY		0x001
340 #define	IFC_QFLUSH		0x002
341 #define	IFC_MULTISEG		0x004
342 #define	IFC_INIT_FAILED		0x008
343 #define	IFC_SC_ALLOCATED	0x010
344 #define	IFC_INIT_DONE		0x020
345 #define	IFC_PREFETCH		0x040
346 #define	IFC_DO_RESET		0x080
347 #define	IFC_DO_WATCHDOG		0x100
348 #define	IFC_DO_RESET_IF_UP	0x200
349 #define	IFC_SPARE2		0x400
350 #define	IFC_IN_DETACH		0x800
351 
352 #define	IFC_NETMAP_TX_IRQ	0x80000000
353 
354 #define CSUM_OFFLOAD		(CSUM_IP_TSO | CSUM_IP6_TSO | CSUM_IP | \
355 				 CSUM_IP_UDP | CSUM_IP_TCP | CSUM_IP_SCTP | \
356 				 CSUM_IP6_UDP | CSUM_IP6_TCP | CSUM_IP6_SCTP)
357 
358 struct iflib_txq {
359 	qidx_t		ift_in_use;
360 	qidx_t		ift_cidx;
361 	qidx_t		ift_cidx_processed;
362 	qidx_t		ift_pidx;
363 	uint8_t		ift_gen;
364 	uint8_t		ift_br_offset:1,
365 			ift_defer_mfree:1,
366 			ift_spare_bits0:6;
367 	uint16_t	ift_npending;
368 	uint16_t	ift_db_pending;
369 	uint16_t	ift_rs_pending;
370 	uint32_t	ift_last_reclaim;
371 	uint16_t	ift_reclaim_thresh;
372 	uint16_t	ift_reclaim_ticks;
373 	uint8_t		ift_txd_size[8];
374 	uint64_t	ift_processed;
375 	uint64_t	ift_cleaned;
376 	uint64_t	ift_processed_prev;
377 #if MEMORY_LOGGING
378 	uint64_t	ift_enqueued;
379 	uint64_t	ift_dequeued;
380 #endif
381 	uint64_t	ift_no_tx_dma_setup;
382 	uint64_t	ift_no_desc_avail;
383 	uint64_t	ift_mbuf_defrag_failed;
384 	uint64_t	ift_mbuf_defrag;
385 	uint64_t	ift_map_failed;
386 	uint64_t	ift_txd_encap_efbig;
387 	uint64_t	ift_pullups;
388 	uint64_t	ift_last_timer_tick;
389 
390 	struct mtx	ift_mtx;
391 	struct mtx	ift_db_mtx;
392 
393 	/* constant values */
394 	if_ctx_t	ift_ctx;
395 	struct ifmp_ring	*ift_br;
396 	struct grouptask	ift_task;
397 	qidx_t		ift_size;
398 	qidx_t		ift_pad;
399 	uint16_t	ift_id;
400 	struct callout	ift_timer;
401 #ifdef DEV_NETMAP
402 	struct callout	ift_netmap_timer;
403 #endif /* DEV_NETMAP */
404 
405 	if_txsd_vec_t	ift_sds;
406 	/*
407 	 * TX watchdog state, updated once per iflib_timer period.  The
408 	 * period count saturates instead of wrapping, and is 16 bits
409 	 * wide so that it still reaches any value
410 	 * net.iflib.tx_watchdog_periods is plausibly set to; an 8-bit
411 	 * counter would silently disable the check for a threshold
412 	 * above 255.
413 	 */
414 	qidx_t		ift_outstanding_prev;
415 	uint16_t	ift_wdog_armed;
416 	uint8_t		ift_closed;
417 	uint8_t		ift_update_freq;
418 	uint8_t		ift_spare0[2];	/* pad to the next pointer boundary */
419 	struct iflib_filter_info ift_filter_info;
420 	bus_dma_tag_t	ift_buf_tag;
421 	bus_dma_tag_t	ift_tso_buf_tag;
422 	iflib_dma_info_t	ift_ifdi;
423 #define	MTX_NAME_LEN	32
424 	char		ift_mtx_name[MTX_NAME_LEN];
425 	bus_dma_segment_t	ift_segs[IFLIB_MAX_TX_SEGS]  __aligned(CACHE_LINE_SIZE);
426 #ifdef IFLIB_DIAGNOSTICS
427 	uint64_t ift_cpu_exec_count[256];
428 #endif
429 } __aligned(CACHE_LINE_SIZE);
430 
431 struct iflib_fl {
432 	qidx_t		ifl_cidx;
433 	qidx_t		ifl_pidx;
434 	qidx_t		ifl_credits;
435 	uint8_t		ifl_gen;
436 	uint8_t		ifl_rxd_size;
437 #if MEMORY_LOGGING
438 	uint64_t	ifl_m_enqueued;
439 	uint64_t	ifl_m_dequeued;
440 	uint64_t	ifl_cl_enqueued;
441 	uint64_t	ifl_cl_dequeued;
442 #endif
443 	/* implicit pad */
444 	bitstr_t	*ifl_rx_bitmap;
445 	qidx_t		ifl_fragidx;
446 	/* constant */
447 	qidx_t		ifl_size;
448 	uint16_t	ifl_buf_size;
449 	uint16_t	ifl_cltype;
450 	uma_zone_t	ifl_zone;
451 	iflib_rxsd_array_t	ifl_sds;
452 	iflib_rxq_t	ifl_rxq;
453 	uint8_t		ifl_id;
454 	bus_dma_tag_t	ifl_buf_tag;
455 	iflib_dma_info_t	ifl_ifdi;
456 	uint64_t	ifl_bus_addrs[IFLIB_MAX_RX_REFRESH] __aligned(CACHE_LINE_SIZE);
457 	qidx_t		ifl_rxd_idxs[IFLIB_MAX_RX_REFRESH];
458 }  __aligned(CACHE_LINE_SIZE);
459 
460 static inline qidx_t
461 get_inuse(int size, qidx_t cidx, qidx_t pidx, uint8_t gen)
462 {
463 	qidx_t used;
464 
465 	if (pidx > cidx)
466 		used = pidx - cidx;
467 	else if (pidx < cidx)
468 		used = size - cidx + pidx;
469 	else if (gen == 0 && pidx == cidx)
470 		used = 0;
471 	else if (gen == 1 && pidx == cidx)
472 		used = size;
473 	else
474 		panic("bad state");
475 
476 	return (used);
477 }
478 
479 #define TXQ_AVAIL(txq) ((txq->ift_size - txq->ift_pad) -\
480 	    get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx, txq->ift_gen))
481 
482 #define IDXDIFF(head, tail, wrap) \
483 	((head) >= (tail) ? (head) - (tail) : (wrap) - (tail) + (head))
484 
485 struct iflib_rxq {
486 	if_ctx_t	ifr_ctx;
487 	iflib_fl_t	ifr_fl;
488 	struct pfil_head	*pfil;
489 	/*
490 	 * If there is a separate completion queue (IFLIB_HAS_RXCQ), this is
491 	 * the completion queue consumer index.  Otherwise it's unused.
492 	 */
493 	qidx_t		ifr_cq_cidx;
494 	uint16_t	ifr_id;
495 	uint8_t		ifr_nfl;
496 	uint8_t		ifr_ntxqirq;
497 	uint8_t		ifr_txqid[IFLIB_MAX_TX_SHARED_INTR];
498 	uint8_t		ifr_fl_offset;
499 	struct lro_ctrl		ifr_lc;
500 	struct grouptask	ifr_task;
501 	struct callout		ifr_watchdog;
502 	struct iflib_filter_info ifr_filter_info;
503 	iflib_dma_info_t	ifr_ifdi;
504 
505 	/* dynamically allocate if any drivers need a value substantially larger than this */
506 	struct if_rxd_frag	ifr_frags[IFLIB_MAX_RX_SEGS] __aligned(CACHE_LINE_SIZE);
507 #ifdef IFLIB_DIAGNOSTICS
508 	uint64_t ifr_cpu_exec_count[256];
509 #endif
510 }  __aligned(CACHE_LINE_SIZE);
511 
512 typedef struct if_rxsd {
513 	caddr_t *ifsd_cl;
514 	iflib_fl_t ifsd_fl;
515 } *if_rxsd_t;
516 
517 /*
518  * Only allow a single packet to take up most 1/nth of the tx ring
519  */
520 #define MAX_SINGLE_PACKET_FRACTION 12
521 #define IF_BAD_DMA	((bus_addr_t)-1)
522 
523 #define CTX_ACTIVE(ctx)	((if_getdrvflags((ctx)->ifc_ifp) & IFF_DRV_RUNNING))
524 
525 #define CTX_LOCK_INIT(_sc)	sx_init(&(_sc)->ifc_ctx_sx, "iflib ctx lock")
526 #define CTX_LOCK(ctx)		sx_xlock(&(ctx)->ifc_ctx_sx)
527 #define CTX_UNLOCK(ctx)		sx_xunlock(&(ctx)->ifc_ctx_sx)
528 #define CTX_LOCK_DESTROY(ctx)	sx_destroy(&(ctx)->ifc_ctx_sx)
529 
530 #define STATE_LOCK_INIT(_sc, _name)	mtx_init(&(_sc)->ifc_state_mtx, _name, "iflib state lock", MTX_DEF)
531 #define STATE_LOCK(ctx)		mtx_lock(&(ctx)->ifc_state_mtx)
532 #define STATE_UNLOCK(ctx)	mtx_unlock(&(ctx)->ifc_state_mtx)
533 #define STATE_LOCK_DESTROY(ctx)	mtx_destroy(&(ctx)->ifc_state_mtx)
534 
535 #define CALLOUT_LOCK(txq)	mtx_lock(&txq->ift_mtx)
536 #define CALLOUT_UNLOCK(txq)	mtx_unlock(&txq->ift_mtx)
537 
538 /* Our boot-time initialization hook */
539 static int	iflib_module_event_handler(module_t, int, void *);
540 
541 static moduledata_t iflib_moduledata = {
542 	"iflib",
543 	iflib_module_event_handler,
544 	NULL
545 };
546 
547 DECLARE_MODULE(iflib, iflib_moduledata, SI_SUB_INIT_IF, SI_ORDER_ANY);
548 MODULE_VERSION(iflib, 1);
549 
550 MODULE_DEPEND(iflib, pci, 1, 1, 1);
551 MODULE_DEPEND(iflib, ether, 1, 1, 1);
552 
553 TASKQGROUP_DEFINE(if_io_tqg, mp_ncpus, 1);
554 TASKQGROUP_DEFINE(if_config_tqg, 1, 1);
555 
556 #ifndef IFLIB_DEBUG_COUNTERS
557 #ifdef INVARIANTS
558 #define IFLIB_DEBUG_COUNTERS 1
559 #else
560 #define IFLIB_DEBUG_COUNTERS 0
561 #endif /* !INVARIANTS */
562 #endif
563 
564 static SYSCTL_NODE(_net, OID_AUTO, iflib, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
565     "iflib driver parameters");
566 
567 /*
568  * XXX need to ensure that this can't accidentally cause the head to be moved backwards
569  */
570 static int iflib_min_tx_latency = 0;
571 SYSCTL_INT(_net_iflib, OID_AUTO, min_tx_latency, CTLFLAG_RW,
572     &iflib_min_tx_latency, 0,
573     "minimize transmit latency at the possible expense of throughput");
574 static int iflib_no_tx_batch = 0;
575 SYSCTL_INT(_net_iflib, OID_AUTO, no_tx_batch, CTLFLAG_RW,
576     &iflib_no_tx_batch, 0,
577     "minimize transmit latency at the possible expense of throughput");
578 static int iflib_timer_default = 1000;
579 SYSCTL_INT(_net_iflib, OID_AUTO, timer_default, CTLFLAG_RW,
580     &iflib_timer_default, 0, "number of ticks between iflib_timer calls");
581 /*
582  * Consecutive timer periods a TX queue must stay frozen - see
583  * iflib_timer(), which defines that state - before the hardware is
584  * asked whether it has completions pending.  Four periods is roughly
585  * two seconds with the default timer interval: a healthy queue on
586  * hardware that coalesces completion reports (e.g. 8254x,
587  * TXDCTL.WTHRESH) stays frozen for at most two (measured on 82541PI),
588  * a wedged one until it is reset.
589  */
590 static int iflib_tx_watchdog_periods = 4;
591 SYSCTL_INT(_net_iflib, OID_AUTO, tx_watchdog_periods, CTLFLAG_RWTUN,
592     &iflib_tx_watchdog_periods, 0,
593     "consecutive frozen timer periods before a TX queue is checked for "
594     "a hang (0 disables the check)");
595 
596 
597 #if IFLIB_DEBUG_COUNTERS
598 
599 static int iflib_tx_seen;
600 static int iflib_tx_sent;
601 static int iflib_tx_encap;
602 static int iflib_rx_allocs;
603 static int iflib_fl_refills;
604 static int iflib_fl_refills_large;
605 static int iflib_tx_frees;
606 
607 SYSCTL_INT(_net_iflib, OID_AUTO, tx_seen, CTLFLAG_RD, &iflib_tx_seen, 0,
608     "# TX mbufs seen");
609 SYSCTL_INT(_net_iflib, OID_AUTO, tx_sent, CTLFLAG_RD, &iflib_tx_sent, 0,
610     "# TX mbufs sent");
611 SYSCTL_INT(_net_iflib, OID_AUTO, tx_encap, CTLFLAG_RD, &iflib_tx_encap, 0,
612     "# TX mbufs encapped");
613 SYSCTL_INT(_net_iflib, OID_AUTO, tx_frees, CTLFLAG_RD, &iflib_tx_frees, 0,
614     "# TX frees");
615 SYSCTL_INT(_net_iflib, OID_AUTO, rx_allocs, CTLFLAG_RD, &iflib_rx_allocs, 0,
616     "# RX allocations");
617 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills, CTLFLAG_RD, &iflib_fl_refills, 0,
618     "# refills");
619 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills_large, CTLFLAG_RD,
620     &iflib_fl_refills_large, 0, "# large refills");
621 
622 static int iflib_txq_drain_flushing;
623 static int iflib_txq_drain_oactive;
624 static int iflib_txq_drain_notready;
625 
626 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_flushing, CTLFLAG_RD,
627     &iflib_txq_drain_flushing, 0, "# drain flushes");
628 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_oactive, CTLFLAG_RD,
629     &iflib_txq_drain_oactive, 0, "# drain oactives");
630 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_notready, CTLFLAG_RD,
631     &iflib_txq_drain_notready, 0, "# drain notready");
632 
633 static int iflib_encap_load_mbuf_fail;
634 static int iflib_encap_pad_mbuf_fail;
635 static int iflib_encap_txq_avail_fail;
636 static int iflib_encap_txd_encap_fail;
637 
638 SYSCTL_INT(_net_iflib, OID_AUTO, encap_load_mbuf_fail, CTLFLAG_RD,
639     &iflib_encap_load_mbuf_fail, 0, "# busdma load failures");
640 SYSCTL_INT(_net_iflib, OID_AUTO, encap_pad_mbuf_fail, CTLFLAG_RD,
641     &iflib_encap_pad_mbuf_fail, 0, "# runt frame pad failures");
642 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txq_avail_fail, CTLFLAG_RD,
643     &iflib_encap_txq_avail_fail, 0, "# txq avail failures");
644 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txd_encap_fail, CTLFLAG_RD,
645     &iflib_encap_txd_encap_fail, 0, "# driver encap failures");
646 
647 static int iflib_task_fn_rxs;
648 static int iflib_rx_intr_enables;
649 static int iflib_fast_intrs;
650 static int iflib_rx_unavail;
651 static int iflib_rx_ctx_inactive;
652 static int iflib_rx_if_input;
653 static int iflib_rxd_flush;
654 
655 static int iflib_verbose_debug;
656 
657 SYSCTL_INT(_net_iflib, OID_AUTO, task_fn_rx, CTLFLAG_RD, &iflib_task_fn_rxs, 0,
658     "# task_fn_rx calls");
659 SYSCTL_INT(_net_iflib, OID_AUTO, rx_intr_enables, CTLFLAG_RD,
660     &iflib_rx_intr_enables, 0, "# RX intr enables");
661 SYSCTL_INT(_net_iflib, OID_AUTO, fast_intrs, CTLFLAG_RD, &iflib_fast_intrs, 0,
662     "# fast_intr calls");
663 SYSCTL_INT(_net_iflib, OID_AUTO, rx_unavail, CTLFLAG_RD, &iflib_rx_unavail, 0,
664     "# times rxeof called with no available data");
665 SYSCTL_INT(_net_iflib, OID_AUTO, rx_ctx_inactive, CTLFLAG_RD,
666     &iflib_rx_ctx_inactive, 0, "# times rxeof called with inactive context");
667 SYSCTL_INT(_net_iflib, OID_AUTO, rx_if_input, CTLFLAG_RD, &iflib_rx_if_input,
668     0, "# times rxeof called if_input");
669 SYSCTL_INT(_net_iflib, OID_AUTO, rxd_flush, CTLFLAG_RD, &iflib_rxd_flush, 0,
670     "# times rxd_flush called");
671 SYSCTL_INT(_net_iflib, OID_AUTO, verbose_debug, CTLFLAG_RW,
672     &iflib_verbose_debug, 0, "enable verbose debugging");
673 
674 #define DBG_COUNTER_INC(name) atomic_add_int(&(iflib_ ## name), 1)
675 static void
676 iflib_debug_reset(void)
677 {
678 	iflib_tx_seen = iflib_tx_sent = iflib_tx_encap = iflib_rx_allocs =
679 		iflib_fl_refills = iflib_fl_refills_large = iflib_tx_frees =
680 		iflib_txq_drain_flushing = iflib_txq_drain_oactive =
681 		iflib_txq_drain_notready =
682 		iflib_encap_load_mbuf_fail = iflib_encap_pad_mbuf_fail =
683 		iflib_encap_txq_avail_fail = iflib_encap_txd_encap_fail =
684 		iflib_task_fn_rxs = iflib_rx_intr_enables = iflib_fast_intrs =
685 		iflib_rx_unavail =
686 		iflib_rx_ctx_inactive = iflib_rx_if_input =
687 		iflib_rxd_flush = 0;
688 }
689 
690 #else
691 #define DBG_COUNTER_INC(name)
692 static void iflib_debug_reset(void) {}
693 #endif
694 
695 #define IFLIB_DEBUG 0
696 
697 static void iflib_tx_structures_free(if_ctx_t ctx);
698 static void iflib_rx_structures_free(if_ctx_t ctx);
699 static int iflib_queues_alloc(if_ctx_t ctx);
700 static int iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq);
701 static int iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget);
702 static int iflib_qset_structures_setup(if_ctx_t ctx);
703 static int iflib_msix_init(if_ctx_t ctx);
704 static int iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filterarg, int *rid, const char *str);
705 static void iflib_txq_check_drain(iflib_txq_t txq, int budget);
706 static uint32_t iflib_txq_can_drain(struct ifmp_ring *);
707 #ifdef ALTQ
708 static void iflib_altq_if_start(if_t ifp);
709 static int iflib_altq_if_transmit(if_t ifp, struct mbuf *m);
710 #endif
711 static void iflib_register(if_ctx_t);
712 static void iflib_deregister(if_ctx_t);
713 static void iflib_unregister_vlan_handlers(if_ctx_t ctx);
714 static uint16_t iflib_get_mbuf_size_for(unsigned int size);
715 static void iflib_init_locked(if_ctx_t ctx);
716 static void iflib_add_device_sysctl_pre(if_ctx_t ctx);
717 static void iflib_add_device_sysctl_post(if_ctx_t ctx);
718 static void iflib_ifmp_purge(iflib_txq_t txq);
719 static void _iflib_pre_assert(if_softc_ctx_t scctx);
720 static void iflib_stop(if_ctx_t ctx);
721 static void iflib_if_init_locked(if_ctx_t ctx);
722 static void iflib_free_intr_mem(if_ctx_t ctx);
723 #ifndef __NO_STRICT_ALIGNMENT
724 static struct mbuf *iflib_fixup_rx(struct mbuf *m);
725 #endif
726 static __inline int iflib_completed_tx_reclaim(iflib_txq_t txq,
727     struct mbuf **m_defer);
728 static __inline void iflib_completed_tx_reclaim_force(iflib_txq_t txq);
729 
730 static SLIST_HEAD(cpu_offset_list, cpu_offset) cpu_offsets =
731     SLIST_HEAD_INITIALIZER(cpu_offsets);
732 struct cpu_offset {
733 	SLIST_ENTRY(cpu_offset) entries;
734 	cpuset_t	set;
735 	unsigned int	refcount;
736 	uint16_t	next_cpuid;
737 };
738 static struct mtx cpu_offset_mtx;
739 MTX_SYSINIT(iflib_cpu_offset, &cpu_offset_mtx, "iflib_cpu_offset lock",
740     MTX_DEF);
741 
742 DEBUGNET_DEFINE(iflib);
743 
744 static int
745 iflib_num_rx_descs(if_ctx_t ctx)
746 {
747 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
748 	if_shared_ctx_t sctx = ctx->ifc_sctx;
749 	uint16_t first_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
750 
751 	return (scctx->isc_nrxd[first_rxq]);
752 }
753 
754 static int
755 iflib_num_tx_descs(if_ctx_t ctx)
756 {
757 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
758 	if_shared_ctx_t sctx = ctx->ifc_sctx;
759 	uint16_t first_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
760 
761 	return (scctx->isc_ntxd[first_txq]);
762 }
763 
764 #ifdef DEV_NETMAP
765 #include <sys/selinfo.h>
766 #include <net/netmap.h>
767 #include <dev/netmap/netmap_kern.h>
768 
769 MODULE_DEPEND(iflib, netmap, 1, 1, 1);
770 
771 static int netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, bool init);
772 static void iflib_netmap_timer(void *arg);
773 
774 /*
775  * device-specific sysctl variables:
776  *
777  * iflib_crcstrip: 0: keep CRC in rx frames (default), 1: strip it.
778  *	During regular operations the CRC is stripped, but on some
779  *	hardware reception of frames not multiple of 64 is slower,
780  *	so using crcstrip=0 helps in benchmarks.
781  *
782  * iflib_rx_miss, iflib_rx_miss_bufs:
783  *	count packets that might be missed due to lost interrupts.
784  */
785 SYSCTL_DECL(_dev_netmap);
786 /*
787  * The xl driver by default strips CRCs and we do not override it.
788  */
789 
790 int iflib_crcstrip = 1;
791 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_crcstrip,
792     CTLFLAG_RW, &iflib_crcstrip, 1, "strip CRC on RX frames");
793 
794 int iflib_rx_miss, iflib_rx_miss_bufs;
795 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss,
796     CTLFLAG_RW, &iflib_rx_miss, 0, "potentially missed RX intr");
797 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss_bufs,
798     CTLFLAG_RW, &iflib_rx_miss_bufs, 0, "potentially missed RX intr bufs");
799 
800 /*
801  * Register/unregister. We are already under netmap lock.
802  * Only called on the first register or the last unregister.
803  */
804 static int
805 iflib_netmap_register(struct netmap_adapter *na, int onoff)
806 {
807 	if_t ifp = na->ifp;
808 	if_ctx_t ctx = if_getsoftc(ifp);
809 	int status;
810 
811 	CTX_LOCK(ctx);
812 	if (!CTX_IS_VF(ctx))
813 		IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip);
814 
815 	iflib_stop(ctx);
816 
817 	/*
818 	 * Enable (or disable) netmap flags, and intercept (or restore)
819 	 * ifp->if_transmit. This is done once the device has been stopped
820 	 * to prevent race conditions. Also, this must be done after
821 	 * calling netmap_disable_all_rings() and before calling
822 	 * netmap_enable_all_rings(), so that these two functions see the
823 	 * updated state of the NAF_NETMAP_ON bit.
824 	 */
825 	if (onoff) {
826 		nm_set_native_flags(na);
827 	} else {
828 		nm_clear_native_flags(na);
829 	}
830 
831 	iflib_init_locked(ctx);
832 	IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip); // XXX why twice ?
833 	status = if_getdrvflags(ifp) & IFF_DRV_RUNNING ? 0 : 1;
834 	if (status)
835 		nm_clear_native_flags(na);
836 	CTX_UNLOCK(ctx);
837 	return (status);
838 }
839 
840 static int
841 iflib_netmap_config(struct netmap_adapter *na, struct nm_config_info *info)
842 {
843 	if_t ifp = na->ifp;
844 	if_ctx_t ctx = if_getsoftc(ifp);
845 	iflib_rxq_t rxq = &ctx->ifc_rxqs[0];
846 	iflib_fl_t fl = &rxq->ifr_fl[0];
847 
848 	info->num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
849 	info->num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
850 	info->num_tx_descs = iflib_num_tx_descs(ctx);
851 	info->num_rx_descs = iflib_num_rx_descs(ctx);
852 	info->rx_buf_maxsize = fl->ifl_buf_size;
853 	nm_prinf("txr %u rxr %u txd %u rxd %u rbufsz %u",
854 		info->num_tx_rings, info->num_rx_rings, info->num_tx_descs,
855 		info->num_rx_descs, info->rx_buf_maxsize);
856 
857 	return (0);
858 }
859 
860 static int
861 netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, bool init)
862 {
863 	struct netmap_adapter *na = kring->na;
864 	u_int const lim = kring->nkr_num_slots - 1;
865 	struct netmap_ring *ring = kring->ring;
866 	bus_dmamap_t *map;
867 	struct if_rxd_update iru;
868 	if_ctx_t ctx = rxq->ifr_ctx;
869 	iflib_fl_t fl = &rxq->ifr_fl[0];
870 	u_int nic_i_first, nic_i;
871 	u_int nm_i;
872 	int i, n;
873 #if IFLIB_DEBUG_COUNTERS
874 	int rf_count = 0;
875 #endif
876 
877 	/*
878 	 * This function is used both at initialization and in rxsync.
879 	 * At initialization we need to prepare (with isc_rxd_refill())
880 	 * all the netmap buffers currently owned by the kernel, in
881 	 * such a way to keep fl->ifl_pidx and kring->nr_hwcur in sync
882 	 * (except for kring->nkr_hwofs). These may be less than
883 	 * kring->nkr_num_slots if netmap_reset() was called while
884 	 * an application using the kring that still owned some
885 	 * buffers.
886 	 * At rxsync time, both indexes point to the next buffer to be
887 	 * refilled.
888 	 * In any case we publish (with isc_rxd_flush()) up to
889 	 * (fl->ifl_pidx - 1) % N (included), to avoid the NIC tail/prod
890 	 * pointer to overrun the head/cons pointer, although this is
891 	 * not necessary for some NICs (e.g. vmx).
892 	 */
893 	if (__predict_false(init)) {
894 		n = kring->nkr_num_slots - nm_kr_rxspace(kring);
895 	} else {
896 		n = kring->rhead - kring->nr_hwcur;
897 		if (n == 0)
898 			return (0); /* Nothing to do. */
899 		if (n < 0)
900 			n += kring->nkr_num_slots;
901 	}
902 
903 	iru_init(&iru, rxq, 0 /* flid */);
904 	map = fl->ifl_sds.ifsd_map;
905 	nic_i = fl->ifl_pidx;
906 	nm_i = netmap_idx_n2k(kring, nic_i);
907 	if (__predict_false(init)) {
908 		/*
909 		 * On init/reset, nic_i must be 0, and we must
910 		 * start to refill from hwtail (see netmap_reset()).
911 		 */
912 		MPASS(nic_i == 0);
913 		MPASS(nm_i == kring->nr_hwtail);
914 	} else
915 		MPASS(nm_i == kring->nr_hwcur);
916 	DBG_COUNTER_INC(fl_refills);
917 	while (n > 0) {
918 #if IFLIB_DEBUG_COUNTERS
919 		if (++rf_count == 9)
920 			DBG_COUNTER_INC(fl_refills_large);
921 #endif
922 		nic_i_first = nic_i;
923 		for (i = 0; n > 0 && i < IFLIB_MAX_RX_REFRESH; n--, i++) {
924 			struct netmap_slot *slot = &ring->slot[nm_i];
925 			uint64_t paddr;
926 			void *addr = PNMB(na, slot, &paddr);
927 
928 			MPASS(i < IFLIB_MAX_RX_REFRESH);
929 
930 			if (addr == NETMAP_BUF_BASE(na)) /* bad buf */
931 				return (netmap_ring_reinit(kring));
932 
933 			fl->ifl_bus_addrs[i] = paddr +
934 			    nm_get_offset(kring, slot);
935 			fl->ifl_rxd_idxs[i] = nic_i;
936 
937 			if (__predict_false(init)) {
938 				netmap_load_map(na, fl->ifl_buf_tag,
939 				    map[nic_i], addr);
940 			} else if (slot->flags & NS_BUF_CHANGED) {
941 				/* buffer has changed, reload map */
942 				netmap_reload_map(na, fl->ifl_buf_tag,
943 				    map[nic_i], addr);
944 			}
945 			bus_dmamap_sync(fl->ifl_buf_tag, map[nic_i],
946 			    BUS_DMASYNC_PREREAD);
947 			slot->flags &= ~NS_BUF_CHANGED;
948 
949 			nm_i = nm_next(nm_i, lim);
950 			nic_i = nm_next(nic_i, lim);
951 		}
952 
953 		iru.iru_pidx = nic_i_first;
954 		iru.iru_count = i;
955 		ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
956 	}
957 	fl->ifl_pidx = nic_i;
958 	/*
959 	 * At the end of the loop we must have refilled everything
960 	 * we could possibly refill.
961 	 */
962 	MPASS(nm_i == kring->rhead);
963 	kring->nr_hwcur = nm_i;
964 
965 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
966 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
967 	ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id,
968 	    nm_prev(nic_i, lim));
969 	DBG_COUNTER_INC(rxd_flush);
970 
971 	return (0);
972 }
973 
974 #define NETMAP_TX_TIMER_US	90
975 
976 /*
977  * Reconcile kernel and user view of the transmit ring.
978  *
979  * All information is in the kring.
980  * Userspace wants to send packets up to the one before kring->rhead,
981  * kernel knows kring->nr_hwcur is the first unsent packet.
982  *
983  * Here we push packets out (as many as possible), and possibly
984  * reclaim buffers from previously completed transmission.
985  *
986  * The caller (netmap) guarantees that there is only one instance
987  * running at any time. Any interference with other driver
988  * methods should be handled by the individual drivers.
989  */
990 static int
991 iflib_netmap_txsync(struct netmap_kring *kring, int flags)
992 {
993 	struct netmap_adapter *na = kring->na;
994 	if_t ifp = na->ifp;
995 	struct netmap_ring *ring = kring->ring;
996 	u_int nm_i;	/* index into the netmap kring */
997 	u_int nic_i;	/* index into the NIC ring */
998 	u_int const lim = kring->nkr_num_slots - 1;
999 	u_int const head = kring->rhead;
1000 	struct if_pkt_info pi;
1001 	int tx_pkts = 0, tx_bytes = 0;
1002 
1003 	/*
1004 	 * interrupts on every tx packet are expensive so request
1005 	 * them every half ring, or where NS_REPORT is set
1006 	 */
1007 	u_int report_frequency = kring->nkr_num_slots >> 1;
1008 	/* device-specific */
1009 	if_ctx_t ctx = if_getsoftc(ifp);
1010 	iflib_txq_t txq = &ctx->ifc_txqs[kring->ring_id];
1011 
1012 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1013 	    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1014 
1015 	/*
1016 	 * First part: process new packets to send.
1017 	 * nm_i is the current index in the netmap kring,
1018 	 * nic_i is the corresponding index in the NIC ring.
1019 	 *
1020 	 * If we have packets to send (nm_i != head)
1021 	 * iterate over the netmap ring, fetch length and update
1022 	 * the corresponding slot in the NIC ring. Some drivers also
1023 	 * need to update the buffer's physical address in the NIC slot
1024 	 * even NS_BUF_CHANGED is not set (PNMB computes the addresses).
1025 	 *
1026 	 * The netmap_reload_map() calls is especially expensive,
1027 	 * even when (as in this case) the tag is 0, so do only
1028 	 * when the buffer has actually changed.
1029 	 *
1030 	 * If possible do not set the report/intr bit on all slots,
1031 	 * but only a few times per ring or when NS_REPORT is set.
1032 	 *
1033 	 * Finally, on 10G and faster drivers, it might be useful
1034 	 * to prefetch the next slot and txr entry.
1035 	 */
1036 
1037 	nm_i = kring->nr_hwcur;
1038 	if (nm_i != head) {	/* we have new packets to send */
1039 		uint32_t pkt_len = 0, seg_idx = 0;
1040 		int nic_i_start = -1, flags = 0;
1041 		memset(&pi, 0, sizeof(pi));
1042 		pi.ipi_segs = txq->ift_segs;
1043 		pi.ipi_qsidx = kring->ring_id;
1044 		nic_i = netmap_idx_k2n(kring, nm_i);
1045 
1046 		__builtin_prefetch(&ring->slot[nm_i]);
1047 		__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i]);
1048 		__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i]);
1049 
1050 		while (nm_i != head) {
1051 			struct netmap_slot *slot = &ring->slot[nm_i];
1052 			uint64_t offset = nm_get_offset(kring, slot);
1053 			u_int len = slot->len;
1054 			uint64_t paddr;
1055 			void *addr = PNMB(na, slot, &paddr);
1056 
1057 			flags |= (slot->flags & NS_REPORT ||
1058 				nic_i == 0 || nic_i == report_frequency) ?
1059 				IPI_TX_INTR : 0;
1060 
1061 			/*
1062 			 * If this is the first packet fragment, save the
1063 			 * index of the first NIC slot for later.
1064 			 */
1065 			if (nic_i_start < 0)
1066 				nic_i_start = nic_i;
1067 
1068 			pi.ipi_segs[seg_idx].ds_addr = paddr + offset;
1069 			pi.ipi_segs[seg_idx].ds_len = len;
1070 			if (len) {
1071 				pkt_len += len;
1072 				seg_idx++;
1073 			}
1074 
1075 			if (!(slot->flags & NS_MOREFRAG)) {
1076 				pi.ipi_len = pkt_len;
1077 				pi.ipi_nsegs = seg_idx;
1078 				pi.ipi_pidx = nic_i_start;
1079 				pi.ipi_ndescs = 0;
1080 				pi.ipi_flags = flags;
1081 
1082 				/* Prepare the NIC TX ring. */
1083 				ctx->isc_txd_encap(ctx->ifc_softc, &pi);
1084 				DBG_COUNTER_INC(tx_encap);
1085 
1086 				/* Update transmit counters */
1087 				tx_bytes += pi.ipi_len;
1088 				tx_pkts++;
1089 
1090 				/* Reinit per-packet info for the next one. */
1091 				flags = seg_idx = pkt_len = 0;
1092 				nic_i_start = -1;
1093 			}
1094 
1095 			/* prefetch for next round */
1096 			__builtin_prefetch(&ring->slot[nm_i + 1]);
1097 			__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i + 1]);
1098 			__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i + 1]);
1099 
1100 			NM_CHECK_ADDR_LEN_OFF(na, len, offset);
1101 
1102 			if (slot->flags & NS_BUF_CHANGED) {
1103 				/* buffer has changed, reload map */
1104 				netmap_reload_map(na, txq->ift_buf_tag,
1105 				    txq->ift_sds.ifsd_map[nic_i], addr);
1106 			}
1107 			/* make sure changes to the buffer are synced */
1108 			bus_dmamap_sync(txq->ift_buf_tag,
1109 			    txq->ift_sds.ifsd_map[nic_i],
1110 			    BUS_DMASYNC_PREWRITE);
1111 
1112 			slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED | NS_MOREFRAG);
1113 			nm_i = nm_next(nm_i, lim);
1114 			nic_i = nm_next(nic_i, lim);
1115 		}
1116 		kring->nr_hwcur = nm_i;
1117 
1118 		/* synchronize the NIC ring */
1119 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1120 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1121 
1122 		/* (re)start the tx unit up to slot nic_i (excluded) */
1123 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, nic_i);
1124 	}
1125 
1126 	/*
1127 	 * Second part: reclaim buffers for completed transmissions.
1128 	 *
1129 	 * If there are unclaimed buffers, attempt to reclaim them.
1130 	 * If we don't manage to reclaim them all, and TX IRQs are not in use,
1131 	 * trigger a per-tx-queue timer to try again later.
1132 	 */
1133 	if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1134 		if (iflib_tx_credits_update(ctx, txq)) {
1135 			/* some tx completed, increment avail */
1136 			nic_i = txq->ift_cidx_processed;
1137 			kring->nr_hwtail = nm_prev(netmap_idx_n2k(kring, nic_i), lim);
1138 		}
1139 	}
1140 
1141 	if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ))
1142 		if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1143 			callout_reset_sbt_on(&txq->ift_netmap_timer,
1144 			    NETMAP_TX_TIMER_US * SBT_1US, SBT_1US,
1145 			    iflib_netmap_timer, txq,
1146 			    txq->ift_netmap_timer.c_cpu, 0);
1147 		}
1148 
1149 	if_inc_counter(ifp, IFCOUNTER_OBYTES, tx_bytes);
1150 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, tx_pkts);
1151 
1152 	return (0);
1153 }
1154 
1155 /*
1156  * Reconcile kernel and user view of the receive ring.
1157  * Same as for the txsync, this routine must be efficient.
1158  * The caller guarantees a single invocations, but races against
1159  * the rest of the driver should be handled here.
1160  *
1161  * On call, kring->rhead is the first packet that userspace wants
1162  * to keep, and kring->rcur is the wakeup point.
1163  * The kernel has previously reported packets up to kring->rtail.
1164  *
1165  * If (flags & NAF_FORCE_READ) also check for incoming packets irrespective
1166  * of whether or not we received an interrupt.
1167  */
1168 static int
1169 iflib_netmap_rxsync(struct netmap_kring *kring, int flags)
1170 {
1171 	struct netmap_adapter *na = kring->na;
1172 	struct netmap_ring *ring = kring->ring;
1173 	if_t ifp = na->ifp;
1174 	uint32_t nm_i;	/* index into the netmap ring */
1175 	uint32_t nic_i;	/* index into the NIC ring */
1176 	u_int n;
1177 	u_int const lim = kring->nkr_num_slots - 1;
1178 	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
1179 	int i = 0, rx_bytes = 0, rx_pkts = 0;
1180 
1181 	if_ctx_t ctx = if_getsoftc(ifp);
1182 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1183 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1184 	iflib_rxq_t rxq = &ctx->ifc_rxqs[kring->ring_id];
1185 	iflib_fl_t fl = &rxq->ifr_fl[0];
1186 	struct if_rxd_info ri;
1187 	qidx_t *cidxp;
1188 
1189 	/*
1190 	 * netmap only uses free list 0, to avoid out of order consumption
1191 	 * of receive buffers
1192 	 */
1193 
1194 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
1195 	    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1196 
1197 	/*
1198 	 * First part: import newly received packets.
1199 	 *
1200 	 * nm_i is the index of the next free slot in the netmap ring,
1201 	 * nic_i is the index of the next received packet in the NIC ring
1202 	 * (or in the free list 0 if IFLIB_HAS_RXCQ is set), and they may
1203 	 * differ in case if_init() has been called while
1204 	 * in netmap mode. For the receive ring we have
1205 	 *
1206 	 *	nic_i = fl->ifl_cidx;
1207 	 *	nm_i = kring->nr_hwtail (previous)
1208 	 * and
1209 	 *	nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1210 	 *
1211 	 * fl->ifl_cidx is set to 0 on a ring reinit
1212 	 */
1213 	if (netmap_no_pendintr || force_update) {
1214 		uint32_t hwtail_lim = nm_prev(kring->nr_hwcur, lim);
1215 		bool have_rxcq = sctx->isc_flags & IFLIB_HAS_RXCQ;
1216 		int crclen = iflib_crcstrip ? 0 : 4;
1217 		int error, avail;
1218 
1219 		/*
1220 		 * For the free list consumer index, we use the same
1221 		 * logic as in iflib_rxeof().
1222 		 */
1223 		if (have_rxcq)
1224 			cidxp = &rxq->ifr_cq_cidx;
1225 		else
1226 			cidxp = &fl->ifl_cidx;
1227 		avail = ctx->isc_rxd_available(ctx->ifc_softc,
1228 		    rxq->ifr_id, *cidxp, USHRT_MAX);
1229 
1230 		nic_i = fl->ifl_cidx;
1231 		nm_i = netmap_idx_n2k(kring, nic_i);
1232 		MPASS(nm_i == kring->nr_hwtail);
1233 		for (n = 0; avail > 0 && nm_i != hwtail_lim; n++, avail--) {
1234 			memset(&ri, 0, sizeof(ri));
1235 			ri.iri_frags = rxq->ifr_frags;
1236 			ri.iri_qsidx = kring->ring_id;
1237 			ri.iri_ifp = ctx->ifc_ifp;
1238 			ri.iri_cidx = *cidxp;
1239 
1240 			error = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
1241 			for (i = 0; i < ri.iri_nfrags; i++) {
1242 				if (error) {
1243 					ring->slot[nm_i].len = 0;
1244 					ring->slot[nm_i].flags = 0;
1245 				} else {
1246 					ring->slot[nm_i].len = ri.iri_frags[i].irf_len;
1247 					if (i == (ri.iri_nfrags - 1)) {
1248 						ring->slot[nm_i].len -= crclen;
1249 						ring->slot[nm_i].flags = 0;
1250 
1251 						/* Update receive counters */
1252 						rx_bytes += ri.iri_len;
1253 						rx_pkts++;
1254 					} else
1255 						ring->slot[nm_i].flags = NS_MOREFRAG;
1256 				}
1257 
1258 				bus_dmamap_sync(fl->ifl_buf_tag,
1259 				    fl->ifl_sds.ifsd_map[nic_i], BUS_DMASYNC_POSTREAD);
1260 				nm_i = nm_next(nm_i, lim);
1261 				fl->ifl_cidx = nic_i = nm_next(nic_i, lim);
1262 			}
1263 
1264 			if (have_rxcq) {
1265 				*cidxp = ri.iri_cidx;
1266 				while (*cidxp >= scctx->isc_nrxd[0])
1267 					*cidxp -= scctx->isc_nrxd[0];
1268 			}
1269 
1270 		}
1271 		if (n) { /* update the state variables */
1272 			if (netmap_no_pendintr && !force_update) {
1273 				/* diagnostics */
1274 				iflib_rx_miss++;
1275 				iflib_rx_miss_bufs += n;
1276 			}
1277 			kring->nr_hwtail = nm_i;
1278 		}
1279 		kring->nr_kflags &= ~NKR_PENDINTR;
1280 	}
1281 	/*
1282 	 * Second part: skip past packets that userspace has released.
1283 	 * (kring->nr_hwcur to head excluded),
1284 	 * and make the buffers available for reception.
1285 	 * As usual nm_i is the index in the netmap ring,
1286 	 * nic_i is the index in the NIC ring, and
1287 	 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1288 	 */
1289 	netmap_fl_refill(rxq, kring, false);
1290 
1291 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
1292 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
1293 
1294 	return (0);
1295 }
1296 
1297 static void
1298 iflib_netmap_intr(struct netmap_adapter *na, int onoff)
1299 {
1300 	if_ctx_t ctx = if_getsoftc(na->ifp);
1301 
1302 	CTX_LOCK(ctx);
1303 	if (onoff) {
1304 		IFDI_INTR_ENABLE(ctx);
1305 	} else {
1306 		IFDI_INTR_DISABLE(ctx);
1307 	}
1308 	CTX_UNLOCK(ctx);
1309 }
1310 
1311 static int
1312 iflib_netmap_attach(if_ctx_t ctx)
1313 {
1314 	struct netmap_adapter na;
1315 
1316 	bzero(&na, sizeof(na));
1317 
1318 	na.ifp = ctx->ifc_ifp;
1319 	na.na_flags = NAF_BDG_MAYSLEEP | NAF_MOREFRAG | NAF_OFFSETS;
1320 	MPASS(ctx->ifc_softc_ctx.isc_ntxqsets);
1321 	MPASS(ctx->ifc_softc_ctx.isc_nrxqsets);
1322 
1323 	na.num_tx_desc = iflib_num_tx_descs(ctx);
1324 	na.num_rx_desc = iflib_num_rx_descs(ctx);
1325 	na.nm_txsync = iflib_netmap_txsync;
1326 	na.nm_rxsync = iflib_netmap_rxsync;
1327 	na.nm_register = iflib_netmap_register;
1328 	na.nm_intr = iflib_netmap_intr;
1329 	na.nm_config = iflib_netmap_config;
1330 	na.num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
1331 	na.num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
1332 	return (netmap_attach(&na));
1333 }
1334 
1335 static int
1336 iflib_netmap_txq_init(if_ctx_t ctx, iflib_txq_t txq)
1337 {
1338 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1339 	struct netmap_slot *slot;
1340 
1341 	slot = netmap_reset(na, NR_TX, txq->ift_id, 0);
1342 	if (slot == NULL)
1343 		return (0);
1344 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxd[0]; i++) {
1345 		/*
1346 		 * In netmap mode, set the map for the packet buffer.
1347 		 * NOTE: Some drivers (not this one) also need to set
1348 		 * the physical buffer address in the NIC ring.
1349 		 * netmap_idx_n2k() maps a nic index, i, into the corresponding
1350 		 * netmap slot index, si
1351 		 */
1352 		int si = netmap_idx_n2k(na->tx_rings[txq->ift_id], i);
1353 		netmap_load_map(na, txq->ift_buf_tag, txq->ift_sds.ifsd_map[i],
1354 		    NMB(na, slot + si));
1355 	}
1356 	return (1);
1357 }
1358 
1359 static int
1360 iflib_netmap_rxq_init(if_ctx_t ctx, iflib_rxq_t rxq)
1361 {
1362 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1363 	struct netmap_kring *kring;
1364 	struct netmap_slot *slot;
1365 
1366 	slot = netmap_reset(na, NR_RX, rxq->ifr_id, 0);
1367 	if (slot == NULL)
1368 		return (0);
1369 	kring = na->rx_rings[rxq->ifr_id];
1370 	netmap_fl_refill(rxq, kring, true);
1371 	return (1);
1372 }
1373 
1374 static void
1375 iflib_netmap_timer(void *arg)
1376 {
1377 	iflib_txq_t txq = arg;
1378 	if_ctx_t ctx = txq->ift_ctx;
1379 
1380 	/*
1381 	 * Wake up the netmap application, to give it a chance to
1382 	 * call txsync and reclaim more completed TX buffers.
1383 	 */
1384 	netmap_tx_irq(ctx->ifc_ifp, txq->ift_id);
1385 }
1386 
1387 #define iflib_netmap_detach(ifp) netmap_detach(ifp)
1388 
1389 #else
1390 #define iflib_netmap_txq_init(ctx, txq) (0)
1391 #define iflib_netmap_rxq_init(ctx, rxq) (0)
1392 #define iflib_netmap_detach(ifp)
1393 #define netmap_enable_all_rings(ifp)
1394 #define netmap_disable_all_rings(ifp)
1395 
1396 #define iflib_netmap_attach(ctx) (0)
1397 #define netmap_rx_irq(ifp, qid, budget) (0)
1398 #endif
1399 
1400 #if defined(__i386__) || defined(__amd64__)
1401 static __inline void
1402 prefetch(void *x)
1403 {
1404 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1405 }
1406 
1407 static __inline void
1408 prefetch2cachelines(void *x)
1409 {
1410 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1411 #if (CACHE_LINE_SIZE < 128)
1412 	__asm volatile("prefetcht0 %0" :: "m" (*(((unsigned long *)x) + CACHE_LINE_SIZE / (sizeof(unsigned long)))));
1413 #endif
1414 }
1415 #else
1416 static __inline void
1417 prefetch(void *x)
1418 {
1419 }
1420 
1421 static __inline void
1422 prefetch2cachelines(void *x)
1423 {
1424 }
1425 #endif
1426 
1427 static void
1428 iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid)
1429 {
1430 	iflib_fl_t fl;
1431 
1432 	fl = &rxq->ifr_fl[flid];
1433 	iru->iru_paddrs = fl->ifl_bus_addrs;
1434 	iru->iru_idxs = fl->ifl_rxd_idxs;
1435 	iru->iru_qsidx = rxq->ifr_id;
1436 	iru->iru_buf_size = fl->ifl_buf_size;
1437 	iru->iru_flidx = fl->ifl_id;
1438 }
1439 
1440 static void
1441 _iflib_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int err)
1442 {
1443 	if (err)
1444 		return;
1445 	*(bus_addr_t *) arg = segs[0].ds_addr;
1446 }
1447 
1448 #define	DMA_WIDTH_TO_BUS_LOWADDR(width)				\
1449 	(((width) == 0) || (width) == flsll(BUS_SPACE_MAXADDR) ?	\
1450 	    BUS_SPACE_MAXADDR : (1ULL << (width)) - 1ULL)
1451 
1452 int
1453 iflib_dma_alloc_align(if_ctx_t ctx, int size, int align, iflib_dma_info_t dma, int mapflags)
1454 {
1455 	int err;
1456 	device_t dev = ctx->ifc_dev;
1457 	bus_addr_t lowaddr;
1458 
1459 	lowaddr = DMA_WIDTH_TO_BUS_LOWADDR(ctx->ifc_softc_ctx.isc_dma_width);
1460 
1461 	err = bus_dma_tag_create(bus_get_dma_tag(dev),	/* parent */
1462 		    align, 0,		/* alignment, bounds */
1463 		    lowaddr,		/* lowaddr */
1464 		    BUS_SPACE_MAXADDR,	/* highaddr */
1465 		    NULL, NULL,		/* filter, filterarg */
1466 		    size,		/* maxsize */
1467 		    1,			/* nsegments */
1468 		    size,		/* maxsegsize */
1469 		    BUS_DMA_ALLOCNOW,	/* flags */
1470 		    NULL,		/* lockfunc */
1471 		    NULL,		/* lockarg */
1472 		    &dma->idi_tag);
1473 	if (err) {
1474 		device_printf(dev,
1475 		    "%s: bus_dma_tag_create failed: %d (size=%d, align=%d)\n",
1476 		    __func__, err, size, align);
1477 		goto fail_0;
1478 	}
1479 
1480 	err = bus_dmamem_alloc(dma->idi_tag, (void **)&dma->idi_vaddr,
1481 	    BUS_DMA_NOWAIT | BUS_DMA_COHERENT | BUS_DMA_ZERO, &dma->idi_map);
1482 	if (err) {
1483 		device_printf(dev,
1484 		    "%s: bus_dmamem_alloc(%ju) failed: %d\n",
1485 		    __func__, (uintmax_t)size, err);
1486 		goto fail_1;
1487 	}
1488 
1489 	dma->idi_paddr = IF_BAD_DMA;
1490 	err = bus_dmamap_load(dma->idi_tag, dma->idi_map, dma->idi_vaddr,
1491 	    size, _iflib_dmamap_cb, &dma->idi_paddr, mapflags | BUS_DMA_NOWAIT);
1492 	if (err || dma->idi_paddr == IF_BAD_DMA) {
1493 		device_printf(dev,
1494 		    "%s: bus_dmamap_load failed: %d\n",
1495 		    __func__, err);
1496 		goto fail_2;
1497 	}
1498 
1499 	dma->idi_size = size;
1500 	return (0);
1501 
1502 fail_2:
1503 	bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1504 fail_1:
1505 	bus_dma_tag_destroy(dma->idi_tag);
1506 fail_0:
1507 	dma->idi_tag = NULL;
1508 
1509 	return (err);
1510 }
1511 
1512 int
1513 iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags)
1514 {
1515 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1516 
1517 	KASSERT(sctx->isc_q_align != 0, ("alignment value not initialized"));
1518 
1519 	return (iflib_dma_alloc_align(ctx, size, sctx->isc_q_align, dma, mapflags));
1520 }
1521 
1522 int
1523 iflib_dma_alloc_multi(if_ctx_t ctx, int *sizes, iflib_dma_info_t *dmalist, int mapflags, int count)
1524 {
1525 	int i, err;
1526 	iflib_dma_info_t *dmaiter;
1527 
1528 	dmaiter = dmalist;
1529 	for (i = 0; i < count; i++, dmaiter++) {
1530 		if ((err = iflib_dma_alloc(ctx, sizes[i], *dmaiter, mapflags)) != 0)
1531 			break;
1532 	}
1533 	if (err)
1534 		iflib_dma_free_multi(dmalist, i);
1535 	return (err);
1536 }
1537 
1538 void
1539 iflib_dma_free(iflib_dma_info_t dma)
1540 {
1541 	if (dma->idi_tag == NULL)
1542 		return;
1543 	if (dma->idi_paddr != IF_BAD_DMA) {
1544 		bus_dmamap_sync(dma->idi_tag, dma->idi_map,
1545 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1546 		bus_dmamap_unload(dma->idi_tag, dma->idi_map);
1547 		dma->idi_paddr = IF_BAD_DMA;
1548 	}
1549 	if (dma->idi_vaddr != NULL) {
1550 		bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1551 		dma->idi_vaddr = NULL;
1552 	}
1553 	bus_dma_tag_destroy(dma->idi_tag);
1554 	dma->idi_tag = NULL;
1555 }
1556 
1557 void
1558 iflib_dma_free_multi(iflib_dma_info_t *dmalist, int count)
1559 {
1560 	int i;
1561 	iflib_dma_info_t *dmaiter = dmalist;
1562 
1563 	for (i = 0; i < count; i++, dmaiter++)
1564 		iflib_dma_free(*dmaiter);
1565 }
1566 
1567 static int
1568 iflib_fast_intr(void *arg)
1569 {
1570 	iflib_filter_info_t info = arg;
1571 	struct grouptask *gtask = info->ifi_task;
1572 	int result;
1573 
1574 	DBG_COUNTER_INC(fast_intrs);
1575 	if (info->ifi_filter != NULL) {
1576 		result = info->ifi_filter(info->ifi_filter_arg);
1577 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1578 			return (result);
1579 	}
1580 
1581 	GROUPTASK_ENQUEUE(gtask);
1582 	return (FILTER_HANDLED);
1583 }
1584 
1585 static int
1586 iflib_fast_intr_rxtx(void *arg)
1587 {
1588 	iflib_filter_info_t info = arg;
1589 	struct grouptask *gtask = info->ifi_task;
1590 	if_ctx_t ctx;
1591 	iflib_rxq_t rxq = (iflib_rxq_t)info->ifi_ctx;
1592 	iflib_txq_t txq;
1593 	void *sc;
1594 	int i, cidx, result;
1595 	qidx_t txqid;
1596 	bool intr_enable, intr_legacy;
1597 
1598 	DBG_COUNTER_INC(fast_intrs);
1599 	if (info->ifi_filter != NULL) {
1600 		result = info->ifi_filter(info->ifi_filter_arg);
1601 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1602 			return (result);
1603 	}
1604 
1605 	ctx = rxq->ifr_ctx;
1606 	sc = ctx->ifc_softc;
1607 	intr_enable = false;
1608 	intr_legacy = !!(ctx->ifc_flags & IFC_LEGACY);
1609 	MPASS(rxq->ifr_ntxqirq);
1610 	for (i = 0; i < rxq->ifr_ntxqirq; i++) {
1611 		txqid = rxq->ifr_txqid[i];
1612 		txq = &ctx->ifc_txqs[txqid];
1613 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1614 		    BUS_DMASYNC_POSTREAD);
1615 		if (!ctx->isc_txd_credits_update(sc, txqid, false)) {
1616 			if (intr_legacy)
1617 				intr_enable = true;
1618 			else
1619 				IFDI_TX_QUEUE_INTR_ENABLE(ctx, txqid);
1620 			continue;
1621 		}
1622 		GROUPTASK_ENQUEUE(&txq->ift_task);
1623 	}
1624 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_RXCQ)
1625 		cidx = rxq->ifr_cq_cidx;
1626 	else
1627 		cidx = rxq->ifr_fl[0].ifl_cidx;
1628 	if (iflib_rxd_avail(ctx, rxq, cidx, 1))
1629 		GROUPTASK_ENQUEUE(gtask);
1630 	else {
1631 		if (intr_legacy)
1632 			intr_enable = true;
1633 		else
1634 			IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
1635 		DBG_COUNTER_INC(rx_intr_enables);
1636 	}
1637 	if (intr_enable)
1638 		IFDI_INTR_ENABLE(ctx);
1639 	return (FILTER_HANDLED);
1640 }
1641 
1642 static int
1643 iflib_fast_intr_ctx(void *arg)
1644 {
1645 	iflib_filter_info_t info = arg;
1646 	if_ctx_t ctx = info->ifi_ctx;
1647 	int result;
1648 
1649 	DBG_COUNTER_INC(fast_intrs);
1650 	if (info->ifi_filter != NULL) {
1651 		result = info->ifi_filter(info->ifi_filter_arg);
1652 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1653 			return (result);
1654 	}
1655 
1656 	taskqueue_enqueue(ctx->ifc_tq, &ctx->ifc_admin_task);
1657 	return (FILTER_HANDLED);
1658 }
1659 
1660 static int
1661 _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
1662 		 driver_filter_t filter, driver_intr_t handler, void *arg,
1663 		 const char *name)
1664 {
1665 	struct resource *res;
1666 	void *tag = NULL;
1667 	device_t dev = ctx->ifc_dev;
1668 	int flags, i, rc;
1669 
1670 	flags = RF_ACTIVE;
1671 	if (ctx->ifc_flags & IFC_LEGACY)
1672 		flags |= RF_SHAREABLE;
1673 	MPASS(rid < 512);
1674 	i = rid;
1675 	res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &i, flags);
1676 	if (res == NULL) {
1677 		device_printf(dev,
1678 		    "failed to allocate IRQ for rid %d, name %s.\n", rid, name);
1679 		return (ENOMEM);
1680 	}
1681 	irq->ii_res = res;
1682 	KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL"));
1683 	rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET,
1684 		    filter, handler, arg, &tag);
1685 	if (rc != 0) {
1686 		device_printf(dev,
1687 		    "failed to setup interrupt for rid %d, name %s: %d\n",
1688 		    rid, name ? name : "unknown", rc);
1689 		return (rc);
1690 	} else if (name)
1691 		bus_describe_intr(dev, res, tag, "%s", name);
1692 
1693 	irq->ii_tag = tag;
1694 	return (0);
1695 }
1696 
1697 /*********************************************************************
1698  *
1699  *  Allocate DMA resources for TX buffers as well as memory for the TX
1700  *  mbuf map.  TX DMA maps (non-TSO/TSO) and TX mbuf map are kept in a
1701  *  iflib_sw_tx_desc_array structure, storing all the information that
1702  *  is needed to transmit a packet on the wire.  This is called only
1703  *  once at attach, setup is done every reset.
1704  *
1705  **********************************************************************/
1706 static int
1707 iflib_txsd_alloc(iflib_txq_t txq)
1708 {
1709 	if_ctx_t ctx = txq->ift_ctx;
1710 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1711 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1712 	device_t dev = ctx->ifc_dev;
1713 	bus_size_t tsomaxsize;
1714 	bus_addr_t lowaddr;
1715 	int err, nsegments, ntsosegments;
1716 	bool tso;
1717 
1718 	nsegments = scctx->isc_tx_nsegments;
1719 	ntsosegments = scctx->isc_tx_tso_segments_max;
1720 	tsomaxsize = scctx->isc_tx_tso_size_max;
1721 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_VLAN_MTU)
1722 		tsomaxsize += sizeof(struct ether_vlan_header);
1723 	MPASS(scctx->isc_ntxd[0] > 0);
1724 	MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0);
1725 	MPASS(nsegments > 0);
1726 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) {
1727 		MPASS(ntsosegments > 0);
1728 		MPASS(sctx->isc_tso_maxsize >= tsomaxsize);
1729 	}
1730 
1731 	lowaddr = DMA_WIDTH_TO_BUS_LOWADDR(scctx->isc_dma_width);
1732 
1733 	/*
1734 	 * Set up DMA tags for TX buffers.
1735 	 */
1736 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1737 		    1, 0,			/* alignment, bounds */
1738 		    lowaddr,			/* lowaddr */
1739 		    BUS_SPACE_MAXADDR,		/* highaddr */
1740 		    NULL, NULL,			/* filter, filterarg */
1741 		    sctx->isc_tx_maxsize,	/* maxsize */
1742 		    nsegments,			/* nsegments */
1743 		    sctx->isc_tx_maxsegsize,	/* maxsegsize */
1744 		    0,				/* flags */
1745 		    NULL,			/* lockfunc */
1746 		    NULL,			/* lockfuncarg */
1747 		    &txq->ift_buf_tag))) {
1748 		device_printf(dev, "Unable to allocate TX DMA tag: %d\n", err);
1749 		device_printf(dev, "maxsize: %ju nsegments: %d maxsegsize: %ju\n",
1750 		    (uintmax_t)sctx->isc_tx_maxsize, nsegments, (uintmax_t)sctx->isc_tx_maxsegsize);
1751 		goto fail;
1752 	}
1753 	tso = (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) != 0;
1754 	if (tso && (err = bus_dma_tag_create(bus_get_dma_tag(dev),
1755 		    1, 0,			/* alignment, bounds */
1756 		    lowaddr,			/* lowaddr */
1757 		    BUS_SPACE_MAXADDR,		/* highaddr */
1758 		    NULL, NULL,			/* filter, filterarg */
1759 		    tsomaxsize,			/* maxsize */
1760 		    ntsosegments,		/* nsegments */
1761 		    sctx->isc_tso_maxsegsize,	/* maxsegsize */
1762 		    0,				/* flags */
1763 		    NULL,			/* lockfunc */
1764 		    NULL,			/* lockfuncarg */
1765 		    &txq->ift_tso_buf_tag))) {
1766 		device_printf(dev, "Unable to allocate TSO TX DMA tag: %d\n",
1767 		    err);
1768 		goto fail;
1769 	}
1770 
1771 	/* Allocate memory for the TX mbuf map. */
1772 	if (!(txq->ift_sds.ifsd_m =
1773 	    (struct mbuf **) malloc(sizeof(struct mbuf *) *
1774 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1775 		device_printf(dev, "Unable to allocate TX mbuf map memory\n");
1776 		err = ENOMEM;
1777 		goto fail;
1778 	}
1779 	if (ctx->ifc_sysctl_simple_tx) {
1780 		if (!(txq->ift_sds.ifsd_m_defer =
1781 			(struct mbuf **) malloc(sizeof(struct mbuf *) *
1782 			    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1783 			device_printf(dev, "Unable to allocate TX mbuf map memory\n");
1784 			err = ENOMEM;
1785 			goto fail;
1786 		}
1787 	}
1788 	txq->ift_sds.ifsd_m_deferb = txq->ift_sds.ifsd_m_defer;
1789 	/*
1790 	 * Create the DMA maps for TX buffers.
1791 	 */
1792 	if ((txq->ift_sds.ifsd_map = (bus_dmamap_t *)malloc(
1793 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1794 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1795 		device_printf(dev,
1796 		    "Unable to allocate TX buffer DMA map memory\n");
1797 		err = ENOMEM;
1798 		goto fail;
1799 	}
1800 	if (tso && (txq->ift_sds.ifsd_tso_map = (bus_dmamap_t *)malloc(
1801 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1802 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1803 		device_printf(dev,
1804 		    "Unable to allocate TSO TX buffer map memory\n");
1805 		err = ENOMEM;
1806 		goto fail;
1807 	}
1808 	for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) {
1809 		err = bus_dmamap_create(txq->ift_buf_tag, 0,
1810 		    &txq->ift_sds.ifsd_map[i]);
1811 		if (err != 0) {
1812 			device_printf(dev, "Unable to create TX DMA map\n");
1813 			goto fail;
1814 		}
1815 		if (!tso)
1816 			continue;
1817 		err = bus_dmamap_create(txq->ift_tso_buf_tag, 0,
1818 		    &txq->ift_sds.ifsd_tso_map[i]);
1819 		if (err != 0) {
1820 			device_printf(dev, "Unable to create TSO TX DMA map\n");
1821 			goto fail;
1822 		}
1823 	}
1824 	return (0);
1825 fail:
1826 	/* We free all, it handles case where we are in the middle */
1827 	iflib_tx_structures_free(ctx);
1828 	return (err);
1829 }
1830 
1831 static void
1832 iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i)
1833 {
1834 	bus_dmamap_t map;
1835 
1836 	if (txq->ift_sds.ifsd_map != NULL) {
1837 		map = txq->ift_sds.ifsd_map[i];
1838 		bus_dmamap_sync(txq->ift_buf_tag, map, BUS_DMASYNC_POSTWRITE);
1839 		bus_dmamap_unload(txq->ift_buf_tag, map);
1840 		bus_dmamap_destroy(txq->ift_buf_tag, map);
1841 		txq->ift_sds.ifsd_map[i] = NULL;
1842 	}
1843 
1844 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1845 		map = txq->ift_sds.ifsd_tso_map[i];
1846 		bus_dmamap_sync(txq->ift_tso_buf_tag, map,
1847 		    BUS_DMASYNC_POSTWRITE);
1848 		bus_dmamap_unload(txq->ift_tso_buf_tag, map);
1849 		bus_dmamap_destroy(txq->ift_tso_buf_tag, map);
1850 		txq->ift_sds.ifsd_tso_map[i] = NULL;
1851 	}
1852 }
1853 
1854 static void
1855 iflib_txq_destroy(iflib_txq_t txq)
1856 {
1857 	if_ctx_t ctx = txq->ift_ctx;
1858 
1859 	for (int i = 0; i < txq->ift_size; i++)
1860 		iflib_txsd_destroy(ctx, txq, i);
1861 
1862 	if (txq->ift_br != NULL) {
1863 		ifmp_ring_free(txq->ift_br);
1864 		txq->ift_br = NULL;
1865 	}
1866 
1867 	mtx_destroy(&txq->ift_mtx);
1868 
1869 	if (txq->ift_sds.ifsd_map != NULL) {
1870 		free(txq->ift_sds.ifsd_map, M_IFLIB);
1871 		txq->ift_sds.ifsd_map = NULL;
1872 	}
1873 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1874 		free(txq->ift_sds.ifsd_tso_map, M_IFLIB);
1875 		txq->ift_sds.ifsd_tso_map = NULL;
1876 	}
1877 	if (txq->ift_sds.ifsd_m != NULL) {
1878 		free(txq->ift_sds.ifsd_m, M_IFLIB);
1879 		txq->ift_sds.ifsd_m = NULL;
1880 	}
1881 	if (txq->ift_sds.ifsd_m_defer != NULL) {
1882 		free(txq->ift_sds.ifsd_m_defer, M_IFLIB);
1883 		txq->ift_sds.ifsd_m_defer = NULL;
1884 	}
1885 	if (txq->ift_buf_tag != NULL) {
1886 		bus_dma_tag_destroy(txq->ift_buf_tag);
1887 		txq->ift_buf_tag = NULL;
1888 	}
1889 	if (txq->ift_tso_buf_tag != NULL) {
1890 		bus_dma_tag_destroy(txq->ift_tso_buf_tag);
1891 		txq->ift_tso_buf_tag = NULL;
1892 	}
1893 	if (txq->ift_ifdi != NULL) {
1894 		free(txq->ift_ifdi, M_IFLIB);
1895 	}
1896 }
1897 
1898 static void
1899 iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i)
1900 {
1901 	struct mbuf *m;
1902 
1903 	m = IFLIB_GET_MBUF(txq->ift_sds.ifsd_m[i]);
1904 	if (m == NULL)
1905 		return;
1906 
1907 	if (txq->ift_sds.ifsd_map != NULL) {
1908 		bus_dmamap_sync(txq->ift_buf_tag,
1909 		    txq->ift_sds.ifsd_map[i], BUS_DMASYNC_POSTWRITE);
1910 		bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[i]);
1911 	}
1912 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1913 		bus_dmamap_sync(txq->ift_tso_buf_tag,
1914 		    txq->ift_sds.ifsd_tso_map[i], BUS_DMASYNC_POSTWRITE);
1915 		bus_dmamap_unload(txq->ift_tso_buf_tag,
1916 		    txq->ift_sds.ifsd_tso_map[i]);
1917 	}
1918 	txq->ift_sds.ifsd_m[i] = NULL;
1919 	m_freem(m);
1920 	DBG_COUNTER_INC(tx_frees);
1921 }
1922 
1923 static int
1924 iflib_txq_setup(iflib_txq_t txq)
1925 {
1926 	if_ctx_t ctx = txq->ift_ctx;
1927 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1928 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1929 	iflib_dma_info_t di;
1930 	int i;
1931 
1932 	/* XXX make configurable */
1933 	txq->ift_update_freq = IFLIB_DEFAULT_TX_UPDATE_FREQ;
1934 
1935 	/* Reset indices */
1936 	txq->ift_cidx_processed = 0;
1937 	txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0;
1938 	txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset];
1939 	txq->ift_pad = scctx->isc_tx_pad;
1940 
1941 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1942 		bzero((void *)di->idi_vaddr, di->idi_size);
1943 
1944 	IFDI_TXQ_SETUP(ctx, txq->ift_id);
1945 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1946 		bus_dmamap_sync(di->idi_tag, di->idi_map,
1947 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1948 	return (0);
1949 }
1950 
1951 /*********************************************************************
1952  *
1953  *  Allocate DMA resources for RX buffers as well as memory for the RX
1954  *  mbuf map, direct RX cluster pointer map and RX cluster bus address
1955  *  map.  RX DMA map, RX mbuf map, direct RX cluster pointer map and
1956  *  RX cluster map are kept in a iflib_sw_rx_desc_array structure.
1957  *  Since we use use one entry in iflib_sw_rx_desc_array per received
1958  *  packet, the maximum number of entries we'll need is equal to the
1959  *  number of hardware receive descriptors that we've allocated.
1960  *
1961  **********************************************************************/
1962 static int
1963 iflib_rxsd_alloc(iflib_rxq_t rxq)
1964 {
1965 	if_ctx_t ctx = rxq->ifr_ctx;
1966 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1967 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1968 	device_t dev = ctx->ifc_dev;
1969 	iflib_fl_t fl;
1970 	bus_addr_t lowaddr;
1971 	int err;
1972 
1973 	MPASS(scctx->isc_nrxd[0] > 0);
1974 	MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0);
1975 
1976 	lowaddr = DMA_WIDTH_TO_BUS_LOWADDR(scctx->isc_dma_width);
1977 
1978 	fl = rxq->ifr_fl;
1979 	for (int i = 0; i < rxq->ifr_nfl; i++, fl++) {
1980 		fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */
1981 		/* Set up DMA tag for RX buffers. */
1982 		err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1983 			    1, 0,			/* alignment, bounds */
1984 			    lowaddr,			/* lowaddr */
1985 			    BUS_SPACE_MAXADDR,		/* highaddr */
1986 			    NULL, NULL,			/* filter, filterarg */
1987 			    sctx->isc_rx_maxsize,	/* maxsize */
1988 			    sctx->isc_rx_nsegments,	/* nsegments */
1989 			    sctx->isc_rx_maxsegsize,	/* maxsegsize */
1990 			    0,				/* flags */
1991 			    NULL,			/* lockfunc */
1992 			    NULL,			/* lockarg */
1993 			    &fl->ifl_buf_tag);
1994 		if (err) {
1995 			device_printf(dev,
1996 			    "Unable to allocate RX DMA tag: %d\n", err);
1997 			goto fail;
1998 		}
1999 
2000 		/* Allocate memory for the RX mbuf map. */
2001 		if (!(fl->ifl_sds.ifsd_m =
2002 		    (struct mbuf **) malloc(sizeof(struct mbuf *) *
2003 			    scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
2004 			device_printf(dev,
2005 			    "Unable to allocate RX mbuf map memory\n");
2006 			err = ENOMEM;
2007 			goto fail;
2008 		}
2009 
2010 		/* Allocate memory for the direct RX cluster pointer map. */
2011 		if (!(fl->ifl_sds.ifsd_cl =
2012 		    (caddr_t *) malloc(sizeof(caddr_t) *
2013 			    scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
2014 			device_printf(dev,
2015 			    "Unable to allocate RX cluster map memory\n");
2016 			err = ENOMEM;
2017 			goto fail;
2018 		}
2019 
2020 		/* Allocate memory for the RX cluster bus address map. */
2021 		if (!(fl->ifl_sds.ifsd_ba =
2022 		    (bus_addr_t *) malloc(sizeof(bus_addr_t) *
2023 			    scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
2024 			device_printf(dev,
2025 			    "Unable to allocate RX bus address map memory\n");
2026 			err = ENOMEM;
2027 			goto fail;
2028 		}
2029 
2030 		/*
2031 		 * Create the DMA maps for RX buffers.
2032 		 */
2033 		if (!(fl->ifl_sds.ifsd_map =
2034 		    (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
2035 			device_printf(dev,
2036 			    "Unable to allocate RX buffer DMA map memory\n");
2037 			err = ENOMEM;
2038 			goto fail;
2039 		}
2040 		for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++) {
2041 			err = bus_dmamap_create(fl->ifl_buf_tag, 0,
2042 			    &fl->ifl_sds.ifsd_map[i]);
2043 			if (err != 0) {
2044 				device_printf(dev, "Unable to create RX buffer DMA map\n");
2045 				goto fail;
2046 			}
2047 		}
2048 	}
2049 	return (0);
2050 
2051 fail:
2052 	iflib_rx_structures_free(ctx);
2053 	return (err);
2054 }
2055 
2056 /*
2057  * Internal service routines
2058  */
2059 
2060 struct rxq_refill_cb_arg {
2061 	int               error;
2062 	bus_dma_segment_t seg;
2063 	int               nseg;
2064 };
2065 
2066 static void
2067 _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
2068 {
2069 	struct rxq_refill_cb_arg *cb_arg = arg;
2070 
2071 	cb_arg->error = error;
2072 	cb_arg->seg = segs[0];
2073 	cb_arg->nseg = nseg;
2074 }
2075 
2076 /**
2077  * iflib_fl_refill - refill an rxq free-buffer list
2078  * @ctx: the iflib context
2079  * @fl: the free list to refill
2080  * @count: the number of new buffers to allocate
2081  *
2082  * (Re)populate an rxq free-buffer list with up to @count new packet buffers.
2083  * The caller must assure that @count does not exceed the queue's capacity
2084  * minus one (since we always leave a descriptor unavailable).
2085  */
2086 static uint8_t
2087 iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
2088 {
2089 	struct if_rxd_update iru;
2090 	struct rxq_refill_cb_arg cb_arg;
2091 	struct mbuf *m;
2092 	caddr_t cl, *sd_cl;
2093 	struct mbuf **sd_m;
2094 	bus_dmamap_t *sd_map;
2095 	bus_addr_t bus_addr, *sd_ba;
2096 	int err, frag_idx, i, idx, n, pidx;
2097 	qidx_t credits;
2098 
2099 	MPASS(count <= fl->ifl_size - fl->ifl_credits - 1);
2100 
2101 	sd_m = fl->ifl_sds.ifsd_m;
2102 	sd_map = fl->ifl_sds.ifsd_map;
2103 	sd_cl = fl->ifl_sds.ifsd_cl;
2104 	sd_ba = fl->ifl_sds.ifsd_ba;
2105 	pidx = fl->ifl_pidx;
2106 	idx = pidx;
2107 	frag_idx = fl->ifl_fragidx;
2108 	credits = fl->ifl_credits;
2109 
2110 	i = 0;
2111 	n = count;
2112 	MPASS(n > 0);
2113 	MPASS(credits + n <= fl->ifl_size);
2114 
2115 	if (pidx < fl->ifl_cidx)
2116 		MPASS(pidx + n <= fl->ifl_cidx);
2117 	if (pidx == fl->ifl_cidx && (credits < fl->ifl_size))
2118 		MPASS(fl->ifl_gen == 0);
2119 	if (pidx > fl->ifl_cidx)
2120 		MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx);
2121 
2122 	DBG_COUNTER_INC(fl_refills);
2123 	if (n > 8)
2124 		DBG_COUNTER_INC(fl_refills_large);
2125 	iru_init(&iru, fl->ifl_rxq, fl->ifl_id);
2126 	while (n-- > 0) {
2127 		/*
2128 		 * We allocate an uninitialized mbuf + cluster, mbuf is
2129 		 * initialized after rx.
2130 		 *
2131 		 * If the cluster is still set then we know a minimum sized
2132 		 * packet was received
2133 		 */
2134 		bit_ffc_at(fl->ifl_rx_bitmap, frag_idx, fl->ifl_size,
2135 		    &frag_idx);
2136 		if (frag_idx < 0)
2137 			bit_ffc(fl->ifl_rx_bitmap, fl->ifl_size, &frag_idx);
2138 		MPASS(frag_idx >= 0);
2139 		if ((cl = sd_cl[frag_idx]) == NULL) {
2140 			cl = uma_zalloc(fl->ifl_zone, M_NOWAIT);
2141 			if (__predict_false(cl == NULL))
2142 				break;
2143 
2144 			cb_arg.error = 0;
2145 			MPASS(sd_map != NULL);
2146 			err = bus_dmamap_load(fl->ifl_buf_tag, sd_map[frag_idx],
2147 			    cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg,
2148 			    BUS_DMA_NOWAIT);
2149 			if (__predict_false(err != 0 || cb_arg.error)) {
2150 				uma_zfree(fl->ifl_zone, cl);
2151 				break;
2152 			}
2153 
2154 			sd_ba[frag_idx] = bus_addr = cb_arg.seg.ds_addr;
2155 			sd_cl[frag_idx] = cl;
2156 #if MEMORY_LOGGING
2157 			fl->ifl_cl_enqueued++;
2158 #endif
2159 		} else {
2160 			bus_addr = sd_ba[frag_idx];
2161 		}
2162 		bus_dmamap_sync(fl->ifl_buf_tag, sd_map[frag_idx],
2163 		    BUS_DMASYNC_PREREAD);
2164 
2165 		if (sd_m[frag_idx] == NULL) {
2166 			m = m_gethdr_raw(M_NOWAIT, 0);
2167 			if (__predict_false(m == NULL))
2168 				break;
2169 			sd_m[frag_idx] = m;
2170 		}
2171 		bit_set(fl->ifl_rx_bitmap, frag_idx);
2172 #if MEMORY_LOGGING
2173 		fl->ifl_m_enqueued++;
2174 #endif
2175 
2176 		DBG_COUNTER_INC(rx_allocs);
2177 		fl->ifl_rxd_idxs[i] = frag_idx;
2178 		fl->ifl_bus_addrs[i] = bus_addr;
2179 		credits++;
2180 		i++;
2181 		MPASS(credits <= fl->ifl_size);
2182 		if (++idx == fl->ifl_size) {
2183 #ifdef INVARIANTS
2184 			fl->ifl_gen = 1;
2185 #endif
2186 			idx = 0;
2187 		}
2188 		if (n == 0 || i == IFLIB_MAX_RX_REFRESH) {
2189 			iru.iru_pidx = pidx;
2190 			iru.iru_count = i;
2191 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2192 			fl->ifl_pidx = idx;
2193 			fl->ifl_credits = credits;
2194 			pidx = idx;
2195 			i = 0;
2196 		}
2197 	}
2198 
2199 	if (n < count - 1) {
2200 		if (i != 0) {
2201 			iru.iru_pidx = pidx;
2202 			iru.iru_count = i;
2203 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2204 			fl->ifl_pidx = idx;
2205 			fl->ifl_credits = credits;
2206 		}
2207 		DBG_COUNTER_INC(rxd_flush);
2208 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2209 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2210 		ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id,
2211 		    fl->ifl_id, fl->ifl_pidx);
2212 		if (__predict_true(bit_test(fl->ifl_rx_bitmap, frag_idx))) {
2213 			fl->ifl_fragidx = frag_idx + 1;
2214 			if (fl->ifl_fragidx == fl->ifl_size)
2215 				fl->ifl_fragidx = 0;
2216 		} else {
2217 			fl->ifl_fragidx = frag_idx;
2218 		}
2219 	}
2220 
2221 	return (n == -1 ? 0 : IFLIB_RXEOF_EMPTY);
2222 }
2223 
2224 static inline uint8_t
2225 iflib_fl_refill_all(if_ctx_t ctx, iflib_fl_t fl)
2226 {
2227 	/*
2228 	 * We leave an unused descriptor to avoid pidx to catch up with cidx.
2229 	 * This is important as it confuses most NICs. For instance,
2230 	 * Intel NICs have (per receive ring) RDH and RDT registers, where
2231 	 * RDH points to the next receive descriptor to be used by the NIC,
2232 	 * and RDT for the next receive descriptor to be published by the
2233 	 * driver to the NIC (RDT - 1 is thus the last valid one).
2234 	 * The condition RDH == RDT means no descriptors are available to
2235 	 * the NIC, and thus it would be ambiguous if it also meant that
2236 	 * all the descriptors are available to the NIC.
2237 	 */
2238 	int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1;
2239 #ifdef INVARIANTS
2240 	int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1;
2241 #endif
2242 
2243 	MPASS(fl->ifl_credits <= fl->ifl_size);
2244 	MPASS(reclaimable == delta);
2245 
2246 	if (reclaimable > 0)
2247 		return (iflib_fl_refill(ctx, fl, reclaimable));
2248 	return (0);
2249 }
2250 
2251 uint8_t
2252 iflib_in_detach(if_ctx_t ctx)
2253 {
2254 	bool in_detach;
2255 
2256 	STATE_LOCK(ctx);
2257 	in_detach = !!(ctx->ifc_flags & IFC_IN_DETACH);
2258 	STATE_UNLOCK(ctx);
2259 	return (in_detach);
2260 }
2261 
2262 static void
2263 iflib_fl_bufs_free(iflib_fl_t fl)
2264 {
2265 	iflib_dma_info_t idi = fl->ifl_ifdi;
2266 	bus_dmamap_t sd_map;
2267 	uint32_t i;
2268 
2269 	for (i = 0; i < fl->ifl_size; i++) {
2270 		struct mbuf **sd_m = &fl->ifl_sds.ifsd_m[i];
2271 		caddr_t *sd_cl = &fl->ifl_sds.ifsd_cl[i];
2272 
2273 		if (*sd_cl != NULL) {
2274 			sd_map = fl->ifl_sds.ifsd_map[i];
2275 			bus_dmamap_sync(fl->ifl_buf_tag, sd_map,
2276 			    BUS_DMASYNC_POSTREAD);
2277 			bus_dmamap_unload(fl->ifl_buf_tag, sd_map);
2278 			uma_zfree(fl->ifl_zone, *sd_cl);
2279 			*sd_cl = NULL;
2280 			if (*sd_m != NULL) {
2281 				m_init(*sd_m, M_NOWAIT, MT_DATA, 0);
2282 				m_free_raw(*sd_m);
2283 				*sd_m = NULL;
2284 			}
2285 		} else {
2286 			MPASS(*sd_m == NULL);
2287 		}
2288 #if MEMORY_LOGGING
2289 		fl->ifl_m_dequeued++;
2290 		fl->ifl_cl_dequeued++;
2291 #endif
2292 	}
2293 #ifdef INVARIANTS
2294 	for (i = 0; i < fl->ifl_size; i++) {
2295 		MPASS(fl->ifl_sds.ifsd_cl[i] == NULL);
2296 		MPASS(fl->ifl_sds.ifsd_m[i] == NULL);
2297 	}
2298 #endif
2299 	/*
2300 	 * Reset free list values
2301 	 */
2302 	fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = fl->ifl_fragidx = 0;
2303 	bzero(idi->idi_vaddr, idi->idi_size);
2304 }
2305 
2306 /*********************************************************************
2307  *
2308  *  Initialize a free list and its buffers.
2309  *
2310  **********************************************************************/
2311 static int
2312 iflib_fl_setup(iflib_fl_t fl)
2313 {
2314 	iflib_rxq_t rxq = fl->ifl_rxq;
2315 	if_ctx_t ctx = rxq->ifr_ctx;
2316 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2317 	int qidx;
2318 
2319 	bit_nclear(fl->ifl_rx_bitmap, 0, fl->ifl_size - 1);
2320 	/*
2321 	 * Free current RX buffer structs and their mbufs
2322 	 */
2323 	iflib_fl_bufs_free(fl);
2324 	/* Now replenish the mbufs */
2325 	MPASS(fl->ifl_credits == 0);
2326 	qidx = rxq->ifr_fl_offset + fl->ifl_id;
2327 	if (scctx->isc_rxd_buf_size[qidx] != 0)
2328 		fl->ifl_buf_size = scctx->isc_rxd_buf_size[qidx];
2329 	else
2330 		fl->ifl_buf_size = ctx->ifc_rx_mbuf_sz;
2331 	/*
2332 	 * ifl_buf_size may be a driver-supplied value, so pull it up
2333 	 * to the selected mbuf size.
2334 	 */
2335 	fl->ifl_buf_size = iflib_get_mbuf_size_for(fl->ifl_buf_size);
2336 	if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size)
2337 		ctx->ifc_max_fl_buf_size = fl->ifl_buf_size;
2338 	fl->ifl_cltype = m_gettype(fl->ifl_buf_size);
2339 	fl->ifl_zone = m_getzone(fl->ifl_buf_size);
2340 
2341 	/*
2342 	 * Avoid pre-allocating zillions of clusters to an idle card
2343 	 * potentially speeding up attach. In any case make sure
2344 	 * to leave a descriptor unavailable. See the comment in
2345 	 * iflib_fl_refill_all().
2346 	 */
2347 	MPASS(fl->ifl_size > 0);
2348 	(void)iflib_fl_refill(ctx, fl, min(128, fl->ifl_size - 1));
2349 	if (min(128, fl->ifl_size - 1) != fl->ifl_credits)
2350 		return (ENOBUFS);
2351 	/*
2352 	 * handle failure
2353 	 */
2354 	MPASS(rxq != NULL);
2355 	MPASS(fl->ifl_ifdi != NULL);
2356 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2357 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2358 	return (0);
2359 }
2360 
2361 /*********************************************************************
2362  *
2363  *  Free receive ring data structures
2364  *
2365  **********************************************************************/
2366 static void
2367 iflib_rx_sds_free(iflib_rxq_t rxq)
2368 {
2369 	iflib_fl_t fl;
2370 	int i, j;
2371 
2372 	if (rxq->ifr_fl != NULL) {
2373 		for (i = 0; i < rxq->ifr_nfl; i++) {
2374 			fl = &rxq->ifr_fl[i];
2375 			if (fl->ifl_buf_tag != NULL) {
2376 				if (fl->ifl_sds.ifsd_map != NULL) {
2377 					for (j = 0; j < fl->ifl_size; j++) {
2378 						bus_dmamap_sync(
2379 						    fl->ifl_buf_tag,
2380 						    fl->ifl_sds.ifsd_map[j],
2381 						    BUS_DMASYNC_POSTREAD);
2382 						bus_dmamap_unload(
2383 						    fl->ifl_buf_tag,
2384 						    fl->ifl_sds.ifsd_map[j]);
2385 						bus_dmamap_destroy(
2386 						    fl->ifl_buf_tag,
2387 						    fl->ifl_sds.ifsd_map[j]);
2388 					}
2389 				}
2390 				bus_dma_tag_destroy(fl->ifl_buf_tag);
2391 				fl->ifl_buf_tag = NULL;
2392 			}
2393 			free(fl->ifl_sds.ifsd_m, M_IFLIB);
2394 			free(fl->ifl_sds.ifsd_cl, M_IFLIB);
2395 			free(fl->ifl_sds.ifsd_ba, M_IFLIB);
2396 			free(fl->ifl_sds.ifsd_map, M_IFLIB);
2397 			free(fl->ifl_rx_bitmap, M_IFLIB);
2398 			fl->ifl_sds.ifsd_m = NULL;
2399 			fl->ifl_sds.ifsd_cl = NULL;
2400 			fl->ifl_sds.ifsd_ba = NULL;
2401 			fl->ifl_sds.ifsd_map = NULL;
2402 			fl->ifl_rx_bitmap = NULL;
2403 		}
2404 		free(rxq->ifr_fl, M_IFLIB);
2405 		rxq->ifr_fl = NULL;
2406 		free(rxq->ifr_ifdi, M_IFLIB);
2407 		rxq->ifr_ifdi = NULL;
2408 		rxq->ifr_cq_cidx = 0;
2409 	}
2410 }
2411 
2412 /*
2413  * Timer routine
2414  */
2415 static void
2416 iflib_timer(void *arg)
2417 {
2418 	iflib_txq_t txq = arg;
2419 	if_ctx_t ctx = txq->ift_ctx;
2420 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2421 	uint64_t this_tick = ticks;
2422 
2423 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
2424 		return;
2425 
2426 	/*
2427 	 * Check on the state of the TX queue(s); this can be done
2428 	 * without the lock: the counters the check reads are only
2429 	 * advanced by the queue's tx task and a stale read just
2430 	 * delays the verdict by one timer period.
2431 	 */
2432 	if (this_tick - txq->ift_last_timer_tick >= iflib_timer_default) {
2433 		qidx_t outstanding;
2434 		bool frozen;
2435 
2436 		txq->ift_last_timer_tick = this_tick;
2437 		IFDI_TIMER(ctx, txq->ift_id);
2438 
2439 		/*
2440 		 * Descriptors the hardware has not reported as
2441 		 * completed: neither harvested as credits
2442 		 * (ift_processed) nor reclaimed (ift_cleaned accounts
2443 		 * the difference to ift_in_use).  The tail whose
2444 		 * report-status request is still deferred is never
2445 		 * reported and must not count (ift_rs_pending
2446 		 * over-counts it by one per packet).
2447 		 */
2448 		outstanding = txq->ift_in_use -
2449 		    (qidx_t)(txq->ift_processed - txq->ift_cleaned);
2450 
2451 		/*
2452 		 * The queue is frozen while it has descriptors the
2453 		 * hardware has not reported as completed and none
2454 		 * were reclaimed over the period; the link must be
2455 		 * up, with no pause frames and no pending doorbell
2456 		 * (the laggard check below rings it).
2457 		 *
2458 		 * Being frozen is not a fault - the hardware may
2459 		 * defer marking descriptors as completed
2460 		 * indefinitely, and 8254x hardware does so for a
2461 		 * quiet queue - therefore the check arms only when a
2462 		 * frozen queue also takes on new work, and acts only
2463 		 * once it has stayed frozen for
2464 		 * net.iflib.tx_watchdog_periods consecutive periods.
2465 		 */
2466 		frozen = outstanding > txq->ift_rs_pending &&
2467 		    txq->ift_processed == txq->ift_processed_prev &&
2468 		    txq->ift_db_pending == 0 &&
2469 		    sctx->isc_pause_frames == 0 &&
2470 		    ctx->ifc_link_state == LINK_STATE_UP;
2471 		if (!frozen)
2472 			txq->ift_wdog_armed = 0;
2473 		else if (txq->ift_wdog_armed > 0 ||
2474 		    outstanding > txq->ift_outstanding_prev) {
2475 			if (txq->ift_wdog_armed < UINT16_MAX)
2476 				txq->ift_wdog_armed++;
2477 		}
2478 
2479 		/*
2480 		 * Frozen long enough: ask the hardware.  Completions
2481 		 * ready but unharvested for this long mean the
2482 		 * completion interrupt went missing - kick the
2483 		 * queue's task.  Nothing ready, although the queue
2484 		 * kept taking on work, means it is hung.
2485 		 */
2486 		if (iflib_tx_watchdog_periods > 0 &&
2487 		    txq->ift_wdog_armed >= iflib_tx_watchdog_periods) {
2488 			bus_dmamap_sync(txq->ift_ifdi->idi_tag,
2489 			    txq->ift_ifdi->idi_map, BUS_DMASYNC_POSTREAD);
2490 			if (ctx->isc_txd_credits_update(ctx->ifc_softc,
2491 			    txq->ift_id, false) == 0) {
2492 				device_printf(ctx->ifc_dev,
2493 				    "Watchdog timeout (TX: %d desc "
2494 				    "avail: %d pidx: %d) -- resetting\n",
2495 				    txq->ift_id, TXQ_AVAIL(txq),
2496 				    txq->ift_pidx);
2497 				STATE_LOCK(ctx);
2498 				if_setdrvflagbits(ctx->ifc_ifp,
2499 				    IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2500 				ctx->ifc_flags |=
2501 				    (IFC_DO_WATCHDOG | IFC_DO_RESET);
2502 				iflib_admin_intr_deferred(ctx);
2503 				STATE_UNLOCK(ctx);
2504 				return;
2505 			}
2506 			GROUPTASK_ENQUEUE(&txq->ift_task);
2507 		}
2508 		txq->ift_outstanding_prev = outstanding;
2509 		txq->ift_processed_prev = txq->ift_processed;
2510 	}
2511 	/* handle any laggards */
2512 	if (txq->ift_db_pending)
2513 		GROUPTASK_ENQUEUE(&txq->ift_task);
2514 
2515 	sctx->isc_pause_frames = 0;
2516 	if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)
2517 		callout_reset_on(&txq->ift_timer, iflib_timer_default, iflib_timer,
2518 		    txq, txq->ift_timer.c_cpu);
2519 }
2520 
2521 static uint16_t
2522 iflib_get_mbuf_size_for(unsigned int size)
2523 {
2524 
2525 	if (size <= MCLBYTES)
2526 		return (MCLBYTES);
2527 	else
2528 		return (MJUMPAGESIZE);
2529 }
2530 
2531 static void
2532 iflib_calc_rx_mbuf_sz(if_ctx_t ctx)
2533 {
2534 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2535 
2536 	/*
2537 	 * XXX don't set the max_frame_size to larger
2538 	 * than the hardware can handle
2539 	 */
2540 	ctx->ifc_rx_mbuf_sz =
2541 	    iflib_get_mbuf_size_for(sctx->isc_max_frame_size);
2542 }
2543 
2544 uint32_t
2545 iflib_get_rx_mbuf_sz(if_ctx_t ctx)
2546 {
2547 
2548 	return (ctx->ifc_rx_mbuf_sz);
2549 }
2550 
2551 static void
2552 iflib_init_locked(if_ctx_t ctx)
2553 {
2554 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2555 	if_t ifp = ctx->ifc_ifp;
2556 	iflib_fl_t fl;
2557 	iflib_txq_t txq;
2558 	iflib_rxq_t rxq;
2559 	int i, j, tx_ip_csum_flags, tx_ip6_csum_flags;
2560 	bool init_failed;
2561 
2562 	if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2563 	IFDI_INTR_DISABLE(ctx);
2564 
2565 	/*
2566 	 * See iflib_stop(). Useful in case iflib_init_locked() is
2567 	 * called without first calling iflib_stop().
2568 	 */
2569 	netmap_disable_all_rings(ifp);
2570 
2571 	tx_ip_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP);
2572 	tx_ip6_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | CSUM_IP6_SCTP);
2573 	/* Set hardware offload abilities */
2574 	if_clearhwassist(ifp);
2575 	if (if_getcapenable(ifp) & IFCAP_TXCSUM)
2576 		if_sethwassistbits(ifp, tx_ip_csum_flags, 0);
2577 	if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6)
2578 		if_sethwassistbits(ifp,  tx_ip6_csum_flags, 0);
2579 	if (if_getcapenable(ifp) & IFCAP_TSO4)
2580 		if_sethwassistbits(ifp, CSUM_IP_TSO, 0);
2581 	if (if_getcapenable(ifp) & IFCAP_TSO6)
2582 		if_sethwassistbits(ifp, CSUM_IP6_TSO, 0);
2583 
2584 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
2585 		CALLOUT_LOCK(txq);
2586 		callout_stop(&txq->ift_timer);
2587 #ifdef DEV_NETMAP
2588 		callout_stop(&txq->ift_netmap_timer);
2589 #endif /* DEV_NETMAP */
2590 		CALLOUT_UNLOCK(txq);
2591 		(void)iflib_netmap_txq_init(ctx, txq);
2592 	}
2593 
2594 	/*
2595 	 * Calculate a suitable Rx mbuf size prior to calling IFDI_INIT, so
2596 	 * that drivers can use the value when setting up the hardware receive
2597 	 * buffers.
2598 	 */
2599 	iflib_calc_rx_mbuf_sz(ctx);
2600 
2601 #ifdef INVARIANTS
2602 	i = if_getdrvflags(ifp);
2603 #endif
2604 	STATE_LOCK(ctx);
2605 	ctx->ifc_flags &= ~IFC_INIT_FAILED;
2606 	STATE_UNLOCK(ctx);
2607 	IFDI_INIT(ctx);
2608 	MPASS(if_getdrvflags(ifp) == i);
2609 	STATE_LOCK(ctx);
2610 	init_failed = (ctx->ifc_flags & IFC_INIT_FAILED) != 0;
2611 	STATE_UNLOCK(ctx);
2612 	if (init_failed)
2613 		return;
2614 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
2615 		if (iflib_netmap_rxq_init(ctx, rxq) > 0) {
2616 			/* This rxq is in netmap mode. Skip normal init. */
2617 			continue;
2618 		}
2619 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
2620 			if (iflib_fl_setup(fl)) {
2621 				device_printf(ctx->ifc_dev,
2622 				    "setting up free list %d failed - "
2623 				    "check cluster settings\n", j);
2624 				goto done;
2625 			}
2626 		}
2627 	}
2628 done:
2629 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
2630 	IFDI_INTR_ENABLE(ctx);
2631 	txq = ctx->ifc_txqs;
2632 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++)
2633 		callout_reset_on(&txq->ift_timer, iflib_timer_default, iflib_timer, txq,
2634 			txq->ift_timer.c_cpu);
2635 
2636 	/* Re-enable txsync/rxsync. */
2637 	netmap_enable_all_rings(ifp);
2638 }
2639 
2640 static int
2641 iflib_media_change(if_t ifp)
2642 {
2643 	if_ctx_t ctx = if_getsoftc(ifp);
2644 	int err;
2645 
2646 	CTX_LOCK(ctx);
2647 	if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0)
2648 		iflib_if_init_locked(ctx);
2649 	CTX_UNLOCK(ctx);
2650 	return (err);
2651 }
2652 
2653 static void
2654 iflib_media_status(if_t ifp, struct ifmediareq *ifmr)
2655 {
2656 	if_ctx_t ctx = if_getsoftc(ifp);
2657 
2658 	CTX_LOCK(ctx);
2659 	IFDI_UPDATE_ADMIN_STATUS(ctx);
2660 	IFDI_MEDIA_STATUS(ctx, ifmr);
2661 	CTX_UNLOCK(ctx);
2662 }
2663 
2664 static void
2665 iflib_stop(if_ctx_t ctx)
2666 {
2667 	iflib_txq_t txq = ctx->ifc_txqs;
2668 	iflib_rxq_t rxq = ctx->ifc_rxqs;
2669 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2670 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2671 	iflib_dma_info_t di;
2672 	iflib_fl_t fl;
2673 	int i, j;
2674 
2675 	/* Tell the stack that the interface is no longer active */
2676 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2677 
2678 	IFDI_INTR_DISABLE(ctx);
2679 	DELAY(1000);
2680 	IFDI_STOP(ctx);
2681 	DELAY(1000);
2682 
2683 	/*
2684 	 * Stop any pending txsync/rxsync and prevent new ones
2685 	 * form starting. Processes blocked in poll() will get
2686 	 * POLLERR.
2687 	 */
2688 	netmap_disable_all_rings(ctx->ifc_ifp);
2689 
2690 	iflib_debug_reset();
2691 	/* Wait for current tx queue users to exit to disarm watchdog timer. */
2692 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) {
2693 		/* make sure all transmitters have completed before proceeding XXX */
2694 
2695 		CALLOUT_LOCK(txq);
2696 		callout_stop(&txq->ift_timer);
2697 #ifdef DEV_NETMAP
2698 		callout_stop(&txq->ift_netmap_timer);
2699 #endif /* DEV_NETMAP */
2700 		CALLOUT_UNLOCK(txq);
2701 
2702 		if (!ctx->ifc_sysctl_simple_tx) {
2703 			/* clean any enqueued buffers */
2704 			iflib_ifmp_purge(txq);
2705 		}
2706 		/* Free any existing tx buffers. */
2707 		for (j = 0; j < txq->ift_size; j++) {
2708 			iflib_txsd_free(ctx, txq, j);
2709 		}
2710 		txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0;
2711 		txq->ift_processed_prev = 0;
2712 		txq->ift_outstanding_prev = 0;
2713 		txq->ift_wdog_armed = 0;
2714 		txq->ift_in_use = txq->ift_gen = txq->ift_no_desc_avail = 0;
2715 		txq->ift_npending = txq->ift_db_pending = 0;
2716 		txq->ift_rs_pending = 0;
2717 		if (sctx->isc_flags & IFLIB_PRESERVE_TX_INDICES)
2718 			txq->ift_cidx = txq->ift_pidx;
2719 		else
2720 			txq->ift_cidx = txq->ift_pidx = 0;
2721 
2722 		txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0;
2723 		txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0;
2724 		txq->ift_pullups = 0;
2725 		ifmp_ring_reset_stats(txq->ift_br);
2726 		for (j = 0, di = txq->ift_ifdi; j < sctx->isc_ntxqs; j++, di++)
2727 			bzero((void *)di->idi_vaddr, di->idi_size);
2728 	}
2729 	for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) {
2730 		if (rxq->ifr_task.gt_taskqueue != NULL)
2731 			gtaskqueue_drain(rxq->ifr_task.gt_taskqueue,
2732 				 &rxq->ifr_task.gt_task);
2733 
2734 		rxq->ifr_cq_cidx = 0;
2735 		for (j = 0, di = rxq->ifr_ifdi; j < sctx->isc_nrxqs; j++, di++)
2736 			bzero((void *)di->idi_vaddr, di->idi_size);
2737 		/* also resets the free lists pidx/cidx */
2738 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
2739 			iflib_fl_bufs_free(fl);
2740 	}
2741 }
2742 
2743 static inline caddr_t
2744 calc_next_rxd(iflib_fl_t fl, int cidx)
2745 {
2746 	qidx_t size;
2747 	int nrxd;
2748 	caddr_t start, end, cur, next;
2749 
2750 	nrxd = fl->ifl_size;
2751 	size = fl->ifl_rxd_size;
2752 	start = fl->ifl_ifdi->idi_vaddr;
2753 
2754 	if (__predict_false(size == 0))
2755 		return (start);
2756 	cur = start + size * cidx;
2757 	end = start + size * nrxd;
2758 	next = CACHE_PTR_NEXT(cur);
2759 	return (next < end ? next : start);
2760 }
2761 
2762 static inline void
2763 prefetch_pkts(iflib_fl_t fl, int cidx)
2764 {
2765 	int nextptr;
2766 	int nrxd = fl->ifl_size;
2767 	caddr_t next_rxd;
2768 
2769 	nextptr = (cidx + CACHE_PTR_INCREMENT) & (nrxd - 1);
2770 	prefetch(&fl->ifl_sds.ifsd_m[nextptr]);
2771 	prefetch(&fl->ifl_sds.ifsd_cl[nextptr]);
2772 	next_rxd = calc_next_rxd(fl, cidx);
2773 	prefetch(next_rxd);
2774 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 1) & (nrxd - 1)]);
2775 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 2) & (nrxd - 1)]);
2776 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 3) & (nrxd - 1)]);
2777 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 4) & (nrxd - 1)]);
2778 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 1) & (nrxd - 1)]);
2779 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 2) & (nrxd - 1)]);
2780 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 3) & (nrxd - 1)]);
2781 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 4) & (nrxd - 1)]);
2782 }
2783 
2784 static struct mbuf *
2785 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, bool unload, if_rxsd_t sd,
2786     int *pf_rv, if_rxd_info_t ri)
2787 {
2788 	bus_dmamap_t map;
2789 	iflib_fl_t fl;
2790 	caddr_t payload;
2791 	struct mbuf *m;
2792 	int flid, cidx, len, next;
2793 
2794 	map = NULL;
2795 	flid = irf->irf_flid;
2796 	cidx = irf->irf_idx;
2797 	fl = &rxq->ifr_fl[flid];
2798 	sd->ifsd_fl = fl;
2799 	sd->ifsd_cl = &fl->ifl_sds.ifsd_cl[cidx];
2800 	fl->ifl_credits--;
2801 #if MEMORY_LOGGING
2802 	fl->ifl_m_dequeued++;
2803 #endif
2804 	if (rxq->ifr_ctx->ifc_flags & IFC_PREFETCH)
2805 		prefetch_pkts(fl, cidx);
2806 	next = (cidx + CACHE_PTR_INCREMENT) & (fl->ifl_size - 1);
2807 	prefetch(&fl->ifl_sds.ifsd_map[next]);
2808 	map = fl->ifl_sds.ifsd_map[cidx];
2809 
2810 	bus_dmamap_sync(fl->ifl_buf_tag, map, BUS_DMASYNC_POSTREAD);
2811 
2812 	if (rxq->pfil != NULL && PFIL_HOOKED_IN(rxq->pfil) && pf_rv != NULL &&
2813 	    irf->irf_len != 0) {
2814 		payload  = *sd->ifsd_cl;
2815 		payload +=  ri->iri_pad;
2816 		len = ri->iri_len - ri->iri_pad;
2817 		*pf_rv = pfil_mem_in(rxq->pfil, payload, len, ri->iri_ifp, &m);
2818 		switch (*pf_rv) {
2819 		case PFIL_DROPPED:
2820 		case PFIL_CONSUMED:
2821 			/*
2822 			 * The filter ate it.  Everything is recycled.
2823 			 */
2824 			m = NULL;
2825 			unload = 0;
2826 			break;
2827 		case PFIL_REALLOCED:
2828 			/*
2829 			 * The filter copied it.  Everything is recycled.
2830 			 * 'm' points at new mbuf.
2831 			 */
2832 			unload = 0;
2833 			break;
2834 		case PFIL_PASS:
2835 			/*
2836 			 * Filter said it was OK, so receive like
2837 			 * normal
2838 			 */
2839 			m = fl->ifl_sds.ifsd_m[cidx];
2840 			fl->ifl_sds.ifsd_m[cidx] = NULL;
2841 			break;
2842 		default:
2843 			MPASS(0);
2844 		}
2845 	} else {
2846 		m = fl->ifl_sds.ifsd_m[cidx];
2847 		fl->ifl_sds.ifsd_m[cidx] = NULL;
2848 		if (pf_rv != NULL)
2849 			*pf_rv = PFIL_PASS;
2850 	}
2851 
2852 	if (unload && irf->irf_len != 0)
2853 		bus_dmamap_unload(fl->ifl_buf_tag, map);
2854 	fl->ifl_cidx = (fl->ifl_cidx + 1) & (fl->ifl_size - 1);
2855 	if (__predict_false(fl->ifl_cidx == 0))
2856 		fl->ifl_gen = 0;
2857 	bit_clear(fl->ifl_rx_bitmap, cidx);
2858 	return (m);
2859 }
2860 
2861 static struct mbuf *
2862 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri, if_rxsd_t sd, int *pf_rv)
2863 {
2864 	struct mbuf *m, *mh, *mt;
2865 	caddr_t cl;
2866 	int  *pf_rv_ptr, flags, i, padlen;
2867 	bool consumed;
2868 
2869 	i = 0;
2870 	mh = NULL;
2871 	consumed = false;
2872 	*pf_rv = PFIL_PASS;
2873 	pf_rv_ptr = pf_rv;
2874 	do {
2875 		m = rxd_frag_to_sd(rxq, &ri->iri_frags[i], !consumed, sd,
2876 		    pf_rv_ptr, ri);
2877 
2878 		MPASS(*sd->ifsd_cl != NULL);
2879 
2880 		/*
2881 		 * Exclude zero-length frags & frags from
2882 		 * packets the filter has consumed or dropped
2883 		 */
2884 		if (ri->iri_frags[i].irf_len == 0 || consumed ||
2885 		    *pf_rv == PFIL_CONSUMED || *pf_rv == PFIL_DROPPED) {
2886 			if (mh == NULL) {
2887 				consumed = true;
2888 				pf_rv_ptr = NULL;
2889 			}
2890 			/* XXX we can save the cluster here, but not the mbuf */
2891 			if (m != NULL) {
2892 				m_init(m, M_NOWAIT, MT_DATA, 0);
2893 				m_free(m);
2894 			}
2895 			continue;
2896 		}
2897 		if (mh == NULL) {
2898 			flags = M_PKTHDR | M_EXT;
2899 			mh = mt = m;
2900 			padlen = ri->iri_pad;
2901 		} else {
2902 			flags = M_EXT;
2903 			mt->m_next = m;
2904 			mt = m;
2905 			/* assuming padding is only on the first fragment */
2906 			padlen = 0;
2907 		}
2908 		cl = *sd->ifsd_cl;
2909 		*sd->ifsd_cl = NULL;
2910 
2911 		/* Can these two be made one ? */
2912 		m_init(m, M_NOWAIT, MT_DATA, flags);
2913 		m_cljset(m, cl, sd->ifsd_fl->ifl_cltype);
2914 		/*
2915 		 * These must follow m_init and m_cljset
2916 		 */
2917 		m->m_data += padlen;
2918 		ri->iri_len -= padlen;
2919 		m->m_len = ri->iri_frags[i].irf_len;
2920 	} while (++i < ri->iri_nfrags);
2921 
2922 	return (mh);
2923 }
2924 
2925 /*
2926  * Process one software descriptor
2927  */
2928 static struct mbuf *
2929 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri)
2930 {
2931 	struct if_rxsd sd;
2932 	struct mbuf *m;
2933 	int pf_rv;
2934 
2935 	/* should I merge this back in now that the two paths are basically duplicated? */
2936 	if (ri->iri_nfrags == 1 &&
2937 	    ri->iri_frags[0].irf_len != 0 &&
2938 	    ri->iri_frags[0].irf_len <= MIN(IFLIB_RX_COPY_THRESH, MHLEN)) {
2939 		m = rxd_frag_to_sd(rxq, &ri->iri_frags[0], false, &sd,
2940 		    &pf_rv, ri);
2941 		if (pf_rv != PFIL_PASS && pf_rv != PFIL_REALLOCED)
2942 			return (m);
2943 		if (pf_rv == PFIL_PASS) {
2944 			m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR);
2945 #ifndef __NO_STRICT_ALIGNMENT
2946 			if (!IP_ALIGNED(m) && ri->iri_pad == 0)
2947 				m->m_data += 2;
2948 #endif
2949 			memcpy(m->m_data, *sd.ifsd_cl, ri->iri_len);
2950 			m->m_len = ri->iri_frags[0].irf_len;
2951 			m->m_data += ri->iri_pad;
2952 			ri->iri_len -= ri->iri_pad;
2953 		}
2954 	} else {
2955 		m = assemble_segments(rxq, ri, &sd, &pf_rv);
2956 		if (m == NULL)
2957 			return (NULL);
2958 		if (pf_rv != PFIL_PASS && pf_rv != PFIL_REALLOCED)
2959 			return (m);
2960 	}
2961 	m->m_pkthdr.len = ri->iri_len;
2962 	m->m_pkthdr.rcvif = ri->iri_ifp;
2963 	m->m_flags |= ri->iri_flags;
2964 	m->m_pkthdr.ether_vtag = ri->iri_vtag;
2965 	m->m_pkthdr.flowid = ri->iri_flowid;
2966 #ifdef NUMA
2967 	m->m_pkthdr.numa_domain = if_getnumadomain(ri->iri_ifp);
2968 #endif
2969 	M_HASHTYPE_SET(m, ri->iri_rsstype);
2970 	m->m_pkthdr.csum_flags = ri->iri_csum_flags;
2971 	m->m_pkthdr.csum_data = ri->iri_csum_data;
2972 	return (m);
2973 }
2974 
2975 static void
2976 _task_fn_rx_watchdog(void *context)
2977 {
2978 	iflib_rxq_t rxq = context;
2979 
2980 	GROUPTASK_ENQUEUE(&rxq->ifr_task);
2981 }
2982 
2983 static uint8_t
2984 iflib_rxeof(iflib_rxq_t rxq, qidx_t budget)
2985 {
2986 	if_t ifp;
2987 	if_ctx_t ctx = rxq->ifr_ctx;
2988 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2989 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2990 	int avail, i;
2991 	qidx_t *cidxp;
2992 	struct if_rxd_info ri;
2993 	int err, budget_left, rx_bytes, rx_pkts;
2994 	iflib_fl_t fl;
2995 #if defined(INET6) || defined(INET)
2996 	int lro_enabled;
2997 #endif
2998 	uint8_t retval = 0;
2999 
3000 	/*
3001 	 * XXX early demux data packets so that if_input processing only handles
3002 	 * acks in interrupt context
3003 	 */
3004 	struct mbuf *m, *mh, *mt;
3005 
3006 	NET_EPOCH_ASSERT();
3007 
3008 	ifp = ctx->ifc_ifp;
3009 	mh = mt = NULL;
3010 	MPASS(budget > 0);
3011 	rx_pkts	= rx_bytes = 0;
3012 	if (sctx->isc_flags & IFLIB_HAS_RXCQ)
3013 		cidxp = &rxq->ifr_cq_cidx;
3014 	else
3015 		cidxp = &rxq->ifr_fl[0].ifl_cidx;
3016 	if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) {
3017 		for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
3018 			retval |= iflib_fl_refill_all(ctx, fl);
3019 		DBG_COUNTER_INC(rx_unavail);
3020 		return (retval);
3021 	}
3022 
3023 #if defined(INET6) || defined(INET)
3024 	lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO);
3025 #endif
3026 
3027 	/* pfil needs the vnet to be set */
3028 	CURVNET_SET_QUIET(if_getvnet(ifp));
3029 	for (budget_left = budget; budget_left > 0 && avail > 0;) {
3030 		if (__predict_false(!CTX_ACTIVE(ctx))) {
3031 			DBG_COUNTER_INC(rx_ctx_inactive);
3032 			break;
3033 		}
3034 		/*
3035 		 * Reset client set fields to their default values
3036 		 */
3037 		memset(&ri, 0, sizeof(ri));
3038 		ri.iri_qsidx = rxq->ifr_id;
3039 		ri.iri_cidx = *cidxp;
3040 		ri.iri_ifp = ifp;
3041 		ri.iri_frags = rxq->ifr_frags;
3042 		err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
3043 
3044 		if (err) {
3045 			CURVNET_RESTORE();
3046 			goto err;
3047 		}
3048 		rx_pkts += 1;
3049 		rx_bytes += ri.iri_len;
3050 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
3051 			*cidxp = ri.iri_cidx;
3052 			/* Update our consumer index */
3053 			/* XXX NB: shurd - check if this is still safe */
3054 			while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0])
3055 				rxq->ifr_cq_cidx -= scctx->isc_nrxd[0];
3056 			/* was this only a completion queue message? */
3057 			if (__predict_false(ri.iri_nfrags == 0))
3058 				continue;
3059 		}
3060 		MPASS(ri.iri_nfrags != 0);
3061 		MPASS(ri.iri_len != 0);
3062 
3063 		/* will advance the cidx on the corresponding free lists */
3064 		m = iflib_rxd_pkt_get(rxq, &ri);
3065 		avail--;
3066 		budget_left--;
3067 		if (avail == 0 && budget_left)
3068 			avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left);
3069 
3070 		if (__predict_false(m == NULL))
3071 			continue;
3072 
3073 #ifndef __NO_STRICT_ALIGNMENT
3074 		if (!IP_ALIGNED(m) && (m = iflib_fixup_rx(m)) == NULL)
3075 			continue;
3076 #endif
3077 #if defined(INET6) || defined(INET)
3078 		if (lro_enabled) {
3079 			tcp_lro_queue_mbuf(&rxq->ifr_lc, m);
3080 			continue;
3081 		}
3082 #endif
3083 
3084 		if (mh == NULL)
3085 			mh = mt = m;
3086 		else {
3087 			mt->m_nextpkt = m;
3088 			mt = m;
3089 		}
3090 	}
3091 	CURVNET_RESTORE();
3092 	/* make sure that we can refill faster than drain */
3093 	for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
3094 		retval |= iflib_fl_refill_all(ctx, fl);
3095 
3096 	if (mh != NULL) {
3097 		if_input(ifp, mh);
3098 		DBG_COUNTER_INC(rx_if_input);
3099 	}
3100 
3101 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
3102 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
3103 
3104 	/*
3105 	 * Flush any outstanding LRO work
3106 	 */
3107 #if defined(INET6) || defined(INET)
3108 	tcp_lro_flush_all(&rxq->ifr_lc);
3109 #endif
3110 	if (avail != 0 || iflib_rxd_avail(ctx, rxq, *cidxp, 1) != 0)
3111 		retval |= IFLIB_RXEOF_MORE;
3112 	return (retval);
3113 err:
3114 	STATE_LOCK(ctx);
3115 	ctx->ifc_flags |= IFC_DO_RESET;
3116 	iflib_admin_intr_deferred(ctx);
3117 	STATE_UNLOCK(ctx);
3118 	return (0);
3119 }
3120 
3121 #define TXD_NOTIFY_COUNT(txq) (((txq)->ift_size / (txq)->ift_update_freq) - 1)
3122 static inline qidx_t
3123 txq_max_db_deferred(iflib_txq_t txq, qidx_t in_use)
3124 {
3125 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
3126 	qidx_t minthresh = txq->ift_size / 8;
3127 	if (in_use > 4 * minthresh)
3128 		return (notify_count);
3129 	if (in_use > 2 * minthresh)
3130 		return (notify_count >> 1);
3131 	if (in_use > minthresh)
3132 		return (notify_count >> 3);
3133 	return (0);
3134 }
3135 
3136 static inline qidx_t
3137 txq_max_rs_deferred(iflib_txq_t txq)
3138 {
3139 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
3140 	qidx_t minthresh = txq->ift_size / 8;
3141 	if (txq->ift_in_use > 4 * minthresh)
3142 		return (notify_count);
3143 	if (txq->ift_in_use > 2 * minthresh)
3144 		return (notify_count >> 1);
3145 	if (txq->ift_in_use > minthresh)
3146 		return (notify_count >> 2);
3147 	return (2);
3148 }
3149 
3150 #define M_CSUM_FLAGS(m)		((m)->m_pkthdr.csum_flags)
3151 #define M_HAS_VLANTAG(m)	(m->m_flags & M_VLANTAG)
3152 
3153 #define TXQ_MAX_DB_DEFERRED(txq, in_use)	txq_max_db_deferred((txq), (in_use))
3154 #define TXQ_MAX_RS_DEFERRED(txq)	txq_max_rs_deferred(txq)
3155 #define TXQ_MAX_DB_CONSUMED(size)	(size >> 4)
3156 
3157 /* forward compatibility for cxgb */
3158 #define FIRST_QSET(ctx) 0
3159 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets)
3160 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets)
3161 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx))
3162 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments))
3163 
3164 #define	MAX_TX_DESC(ctx) MAX((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max, \
3165     (ctx)->ifc_softc_ctx.isc_tx_nsegments)
3166 
3167 static inline bool
3168 iflib_txd_db_check(iflib_txq_t txq, int ring)
3169 {
3170 	if_ctx_t ctx = txq->ift_ctx;
3171 	qidx_t dbval, max;
3172 
3173 	max = TXQ_MAX_DB_DEFERRED(txq, txq->ift_in_use);
3174 
3175 	/* force || threshold exceeded || at the edge of the ring */
3176 	if (ring || (txq->ift_db_pending >= max) || (TXQ_AVAIL(txq) <= MAX_TX_DESC(ctx))) {
3177 
3178 		/*
3179 		 * 'npending' is used if the card's doorbell is in terms of the number of descriptors
3180 		 * pending flush (BRCM). 'pidx' is used in cases where the card's doorbeel uses the
3181 		 * producer index explicitly (INTC).
3182 		 */
3183 		dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx;
3184 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3185 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
3186 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval);
3187 
3188 		/*
3189 		 * Absent bugs there are zero packets pending so reset pending counts to zero.
3190 		 */
3191 		txq->ift_db_pending = txq->ift_npending = 0;
3192 		return (true);
3193 	}
3194 	return (false);
3195 }
3196 
3197 #ifdef PKT_DEBUG
3198 static void
3199 print_pkt(if_pkt_info_t pi)
3200 {
3201 	printf("pi len:  %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n",
3202 	    pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx);
3203 	printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n",
3204 	    pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag);
3205 	printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n",
3206 	    pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto);
3207 }
3208 #endif
3209 
3210 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO)
3211 #define IS_TX_OFFLOAD4(pi) ((pi)->ipi_csum_flags & (CSUM_IP_TCP | CSUM_IP_TSO))
3212 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO)
3213 #define IS_TX_OFFLOAD6(pi) ((pi)->ipi_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_TSO))
3214 
3215 /**
3216  * Parses out ethernet header information in the given mbuf.
3217  * Returns in pi: ipi_etype (EtherType) and ipi_ehdrlen (Ethernet header length)
3218  *
3219  * This will account for the VLAN header if present.
3220  *
3221  * XXX: This doesn't handle QinQ, which could prevent TX offloads for those
3222  * types of packets.
3223  */
3224 static int
3225 iflib_parse_ether_header(if_pkt_info_t pi, struct mbuf **mp, uint64_t *pullups)
3226 {
3227 	struct ether_vlan_header *eh;
3228 	struct mbuf *m;
3229 
3230 	m = *mp;
3231 	if (__predict_false(m->m_len < sizeof(*eh))) {
3232 		(*pullups)++;
3233 		if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL))
3234 			return (ENOMEM);
3235 	}
3236 	eh = mtod(m, struct ether_vlan_header *);
3237 	if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
3238 		pi->ipi_etype = ntohs(eh->evl_proto);
3239 		pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
3240 	} else {
3241 		pi->ipi_etype = ntohs(eh->evl_encap_proto);
3242 		pi->ipi_ehdrlen = ETHER_HDR_LEN;
3243 	}
3244 	*mp = m;
3245 
3246 	return (0);
3247 }
3248 
3249 /**
3250  * Parse up to the L3 header and extract IPv4/IPv6 header information into pi.
3251  * Currently this information includes: IP ToS value, IP header version/presence
3252  *
3253  * This is missing some checks and doesn't edit the packet content as it goes,
3254  * unlike iflib_parse_header(), in order to keep the amount of code here minimal.
3255  */
3256 static int
3257 iflib_parse_header_partial(if_pkt_info_t pi, struct mbuf **mp, uint64_t *pullups)
3258 {
3259 	struct mbuf *m;
3260 	int err;
3261 
3262 	*pullups = 0;
3263 	m = *mp;
3264 	if (!M_WRITABLE(m)) {
3265 		if ((m = m_dup(m, M_NOWAIT)) == NULL) {
3266 			return (ENOMEM);
3267 		} else {
3268 			m_freem(*mp);
3269 			DBG_COUNTER_INC(tx_frees);
3270 			*mp = m;
3271 		}
3272 	}
3273 
3274 	/* Fills out pi->ipi_etype */
3275 	err = iflib_parse_ether_header(pi, mp, pullups);
3276 	if (err)
3277 		return (err);
3278 	m = *mp;
3279 
3280 	switch (pi->ipi_etype) {
3281 #ifdef INET
3282 	case ETHERTYPE_IP:
3283 	{
3284 		struct mbuf *n;
3285 		struct ip *ip = NULL;
3286 		int miniplen;
3287 
3288 		miniplen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip));
3289 		if (__predict_false(m->m_len < miniplen)) {
3290 			/*
3291 			 * Check for common case where the first mbuf only contains
3292 			 * the Ethernet header
3293 			 */
3294 			if (m->m_len == pi->ipi_ehdrlen) {
3295 				n = m->m_next;
3296 				MPASS(n);
3297 				/* If next mbuf contains at least the minimal IP header, then stop */
3298 				if (n->m_len >= sizeof(*ip)) {
3299 					ip = (struct ip *)n->m_data;
3300 				} else {
3301 					(*pullups)++;
3302 					if (__predict_false((m = m_pullup(m, miniplen)) == NULL))
3303 						return (ENOMEM);
3304 					ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3305 				}
3306 			} else {
3307 				(*pullups)++;
3308 				if (__predict_false((m = m_pullup(m, miniplen)) == NULL))
3309 					return (ENOMEM);
3310 				ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3311 			}
3312 		} else {
3313 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3314 		}
3315 
3316 		/* Have the IPv4 header w/ no options here */
3317 		pi->ipi_ip_hlen = ip->ip_hl << 2;
3318 		pi->ipi_ipproto = ip->ip_p;
3319 		pi->ipi_ip_tos = ip->ip_tos;
3320 		pi->ipi_flags |= IPI_TX_IPV4;
3321 
3322 		break;
3323 	}
3324 #endif
3325 #ifdef INET6
3326 	case ETHERTYPE_IPV6:
3327 	{
3328 		struct ip6_hdr *ip6;
3329 
3330 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
3331 			(*pullups)++;
3332 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
3333 				return (ENOMEM);
3334 		}
3335 		ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
3336 
3337 		/* Have the IPv6 fixed header here */
3338 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
3339 		pi->ipi_ipproto = ip6->ip6_nxt;
3340 		pi->ipi_ip_tos = IPV6_TRAFFIC_CLASS(ip6);
3341 		pi->ipi_flags |= IPI_TX_IPV6;
3342 
3343 		break;
3344 	}
3345 #endif
3346 	default:
3347 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
3348 		pi->ipi_ip_hlen = 0;
3349 		break;
3350 	}
3351 	*mp = m;
3352 
3353 	return (0);
3354 
3355 }
3356 
3357 static int
3358 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp)
3359 {
3360 	if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx;
3361 	struct mbuf *m;
3362 	int err;
3363 
3364 	m = *mp;
3365 	if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) &&
3366 	    M_WRITABLE(m) == 0) {
3367 		if ((m = m_dup(m, M_NOWAIT)) == NULL) {
3368 			return (ENOMEM);
3369 		} else {
3370 			m_freem(*mp);
3371 			DBG_COUNTER_INC(tx_frees);
3372 			*mp = m;
3373 		}
3374 	}
3375 
3376 	/* Fills out pi->ipi_etype */
3377 	err = iflib_parse_ether_header(pi, mp, &txq->ift_pullups);
3378 	if (__predict_false(err))
3379 		return (err);
3380 	m = *mp;
3381 
3382 	switch (pi->ipi_etype) {
3383 #ifdef INET
3384 	case ETHERTYPE_IP:
3385 	{
3386 		struct ip *ip;
3387 		struct tcphdr *th;
3388 		uint8_t hlen;
3389 
3390 		hlen = pi->ipi_ehdrlen + sizeof(*ip);
3391 		if (__predict_false(m->m_len < hlen)) {
3392 			txq->ift_pullups++;
3393 			if (__predict_false((m = m_pullup(m, hlen)) == NULL))
3394 				return (ENOMEM);
3395 		}
3396 		ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3397 		hlen = pi->ipi_ehdrlen + (ip->ip_hl << 2);
3398 		if (ip->ip_p == IPPROTO_TCP) {
3399 			hlen += sizeof(*th);
3400 			th = (struct tcphdr *)((char *)ip + (ip->ip_hl << 2));
3401 		} else if (ip->ip_p == IPPROTO_UDP) {
3402 			hlen += sizeof(struct udphdr);
3403 		}
3404 		if (__predict_false(m->m_len < hlen)) {
3405 			txq->ift_pullups++;
3406 			if ((m = m_pullup(m, hlen)) == NULL)
3407 				return (ENOMEM);
3408 		}
3409 		pi->ipi_ip_hlen = ip->ip_hl << 2;
3410 		pi->ipi_ipproto = ip->ip_p;
3411 		pi->ipi_ip_tos = ip->ip_tos;
3412 		pi->ipi_flags |= IPI_TX_IPV4;
3413 
3414 		/* TCP checksum offload may require TCP header length */
3415 		if (IS_TX_OFFLOAD4(pi)) {
3416 			if (__predict_true(pi->ipi_ipproto == IPPROTO_TCP)) {
3417 				pi->ipi_tcp_hflags = tcp_get_flags(th);
3418 				pi->ipi_tcp_hlen = th->th_off << 2;
3419 				pi->ipi_tcp_seq = th->th_seq;
3420 			}
3421 			if (IS_TSO4(pi)) {
3422 				if (__predict_false(ip->ip_p != IPPROTO_TCP))
3423 					return (ENXIO);
3424 				/*
3425 				 * TSO always requires hardware checksum offload.
3426 				 */
3427 				pi->ipi_csum_flags |= (CSUM_IP_TCP | CSUM_IP);
3428 				th->th_sum = in_pseudo(ip->ip_src.s_addr,
3429 						       ip->ip_dst.s_addr, htons(IPPROTO_TCP));
3430 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3431 				if (sctx->isc_flags & IFLIB_TSO_INIT_IP) {
3432 					ip->ip_sum = 0;
3433 					ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz);
3434 				}
3435 			}
3436 		}
3437 		if ((sctx->isc_flags & IFLIB_NEED_ZERO_CSUM) && (pi->ipi_csum_flags & CSUM_IP))
3438 			ip->ip_sum = 0;
3439 
3440 		break;
3441 	}
3442 #endif
3443 #ifdef INET6
3444 	case ETHERTYPE_IPV6:
3445 	{
3446 		struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
3447 		struct tcphdr *th;
3448 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
3449 
3450 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
3451 			txq->ift_pullups++;
3452 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
3453 				return (ENOMEM);
3454 		}
3455 		th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
3456 
3457 		/* XXX-BZ this will go badly in case of ext hdrs. */
3458 		pi->ipi_ipproto = ip6->ip6_nxt;
3459 		pi->ipi_ip_tos = IPV6_TRAFFIC_CLASS(ip6);
3460 		pi->ipi_flags |= IPI_TX_IPV6;
3461 
3462 		/* TCP checksum offload may require TCP header length */
3463 		if (IS_TX_OFFLOAD6(pi)) {
3464 			if (pi->ipi_ipproto == IPPROTO_TCP) {
3465 				if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) {
3466 					txq->ift_pullups++;
3467 					if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL))
3468 						return (ENOMEM);
3469 				}
3470 				pi->ipi_tcp_hflags = tcp_get_flags(th);
3471 				pi->ipi_tcp_hlen = th->th_off << 2;
3472 				pi->ipi_tcp_seq = th->th_seq;
3473 			}
3474 			if (IS_TSO6(pi)) {
3475 				if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP))
3476 					return (ENXIO);
3477 				/*
3478 				 * TSO always requires hardware checksum offload.
3479 				 */
3480 				pi->ipi_csum_flags |= CSUM_IP6_TCP;
3481 				th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0);
3482 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3483 			}
3484 		}
3485 		break;
3486 	}
3487 #endif
3488 	default:
3489 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
3490 		pi->ipi_ip_hlen = 0;
3491 		break;
3492 	}
3493 	*mp = m;
3494 
3495 	return (0);
3496 }
3497 
3498 /*
3499  * If dodgy hardware rejects the scatter gather chain we've handed it
3500  * we'll need to remove the mbuf chain from ifsg_m[] before we can add the
3501  * m_defrag'd mbufs
3502  */
3503 static __noinline struct mbuf *
3504 iflib_remove_mbuf(iflib_txq_t txq)
3505 {
3506 	int ntxd, pidx;
3507 	struct mbuf *m, **ifsd_m;
3508 
3509 	ifsd_m = txq->ift_sds.ifsd_m;
3510 	ntxd = txq->ift_size;
3511 	pidx = txq->ift_pidx & (ntxd - 1);
3512 	ifsd_m = txq->ift_sds.ifsd_m;
3513 	m = IFLIB_GET_MBUF(ifsd_m[pidx]);
3514 	ifsd_m[pidx] = NULL;
3515 	bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[pidx]);
3516 	if (txq->ift_sds.ifsd_tso_map != NULL)
3517 		bus_dmamap_unload(txq->ift_tso_buf_tag,
3518 		    txq->ift_sds.ifsd_tso_map[pidx]);
3519 #if MEMORY_LOGGING
3520 	txq->ift_dequeued++;
3521 #endif
3522 	return (m);
3523 }
3524 
3525 /*
3526  * Pad an mbuf to ensure a minimum ethernet frame size.
3527  * min_frame_size is the frame size (less CRC) to pad the mbuf to
3528  */
3529 static __noinline int
3530 iflib_ether_pad(device_t dev, struct mbuf **m_head, uint16_t min_frame_size)
3531 {
3532 	/*
3533 	 * 18 is enough bytes to pad an ARP packet to 46 bytes, and
3534 	 * and ARP message is the smallest common payload I can think of
3535 	 */
3536 	static char pad[18];	/* just zeros */
3537 	int n;
3538 	struct mbuf *new_head;
3539 
3540 	if (!M_WRITABLE(*m_head)) {
3541 		new_head = m_dup(*m_head, M_NOWAIT);
3542 		if (new_head == NULL) {
3543 			m_freem(*m_head);
3544 			device_printf(dev, "cannot pad short frame, m_dup() failed");
3545 			DBG_COUNTER_INC(encap_pad_mbuf_fail);
3546 			DBG_COUNTER_INC(tx_frees);
3547 			return (ENOMEM);
3548 		}
3549 		m_freem(*m_head);
3550 		*m_head = new_head;
3551 	}
3552 
3553 	for (n = min_frame_size - (*m_head)->m_pkthdr.len;
3554 	     n > 0; n -= sizeof(pad))
3555 		if (!m_append(*m_head, min(n, sizeof(pad)), pad))
3556 			break;
3557 
3558 	if (n > 0) {
3559 		m_freem(*m_head);
3560 		device_printf(dev, "cannot pad short frame\n");
3561 		DBG_COUNTER_INC(encap_pad_mbuf_fail);
3562 		DBG_COUNTER_INC(tx_frees);
3563 		return (ENOBUFS);
3564 	}
3565 
3566 	return (0);
3567 }
3568 
3569 static int
3570 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp, int *obytes, int *opkts)
3571 {
3572 	if_ctx_t		ctx;
3573 	if_shared_ctx_t		sctx;
3574 	if_softc_ctx_t		scctx;
3575 	bus_dma_tag_t		buf_tag;
3576 	bus_dma_segment_t	*segs;
3577 	struct mbuf		*m_head, **ifsd_m;
3578 	bus_dmamap_t		map;
3579 	struct if_pkt_info	pi;
3580 	uintptr_t		flags;
3581 	int remap = 0;
3582 	int err, nsegs, ndesc, max_segs, pidx;
3583 
3584 	ctx = txq->ift_ctx;
3585 	sctx = ctx->ifc_sctx;
3586 	scctx = &ctx->ifc_softc_ctx;
3587 	segs = txq->ift_segs;
3588 	m_head = *m_headp;
3589 	map = NULL;
3590 
3591 	/*
3592 	 * If we're doing TSO the next descriptor to clean may be quite far ahead
3593 	 */
3594 	pidx = txq->ift_pidx;
3595 	map = txq->ift_sds.ifsd_map[pidx];
3596 	ifsd_m = txq->ift_sds.ifsd_m;
3597 
3598 	if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
3599 		buf_tag = txq->ift_tso_buf_tag;
3600 		max_segs = scctx->isc_tx_tso_segments_max;
3601 		map = txq->ift_sds.ifsd_tso_map[pidx];
3602 		MPASS(buf_tag != NULL);
3603 		MPASS(max_segs > 0);
3604 		flags = IFLIB_TSO;
3605 	} else {
3606 		buf_tag = txq->ift_buf_tag;
3607 		max_segs = scctx->isc_tx_nsegments;
3608 		map = txq->ift_sds.ifsd_map[pidx];
3609 		flags = IFLIB_NO_TSO;
3610 	}
3611 	if ((sctx->isc_flags & IFLIB_NEED_ETHER_PAD) &&
3612 	    __predict_false(m_head->m_pkthdr.len < scctx->isc_min_frame_size)) {
3613 		err = iflib_ether_pad(ctx->ifc_dev, m_headp, scctx->isc_min_frame_size);
3614 		if (err) {
3615 			DBG_COUNTER_INC(encap_txd_encap_fail);
3616 			return (err);
3617 		}
3618 	}
3619 	m_head = *m_headp;
3620 
3621 	memset(&pi, 0, sizeof(pi));
3622 	pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG | M_BCAST | M_MCAST));
3623 	pi.ipi_pidx = pidx;
3624 	pi.ipi_qsidx = txq->ift_id;
3625 	pi.ipi_len = m_head->m_pkthdr.len;
3626 	pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags;
3627 	pi.ipi_vtag = M_HAS_VLANTAG(m_head) ? m_head->m_pkthdr.ether_vtag : 0;
3628 
3629 	/* deliberate bitwise OR to make one condition */
3630 	if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) {
3631 		if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0)) {
3632 			DBG_COUNTER_INC(encap_txd_encap_fail);
3633 			return (err);
3634 		}
3635 		m_head = *m_headp;
3636 	}
3637 
3638 retry:
3639 	err = bus_dmamap_load_mbuf_sg(buf_tag, map, m_head, segs, &nsegs,
3640 	    BUS_DMA_NOWAIT);
3641 defrag:
3642 	if (__predict_false(err)) {
3643 		switch (err) {
3644 		case EFBIG:
3645 			/* try collapse once and defrag once */
3646 			if (remap == 0) {
3647 				m_head = m_collapse(*m_headp, M_NOWAIT, max_segs);
3648 				/* try defrag if collapsing fails */
3649 				if (m_head == NULL)
3650 					remap++;
3651 			}
3652 			if (remap == 1) {
3653 				txq->ift_mbuf_defrag++;
3654 				m_head = m_defrag(*m_headp, M_NOWAIT);
3655 			}
3656 			/*
3657 			 * remap should never be >1 unless bus_dmamap_load_mbuf_sg
3658 			 * failed to map an mbuf that was run through m_defrag
3659 			 */
3660 			MPASS(remap <= 1);
3661 			if (__predict_false(m_head == NULL || remap > 1))
3662 				goto defrag_failed;
3663 			remap++;
3664 			*m_headp = m_head;
3665 			goto retry;
3666 			break;
3667 		case ENOMEM:
3668 			txq->ift_no_tx_dma_setup++;
3669 			break;
3670 		default:
3671 			txq->ift_no_tx_dma_setup++;
3672 			m_freem(*m_headp);
3673 			DBG_COUNTER_INC(tx_frees);
3674 			*m_headp = NULL;
3675 			break;
3676 		}
3677 		txq->ift_map_failed++;
3678 		DBG_COUNTER_INC(encap_load_mbuf_fail);
3679 		DBG_COUNTER_INC(encap_txd_encap_fail);
3680 		return (err);
3681 	}
3682 	ifsd_m[pidx] = IFLIB_SAVE_MBUF(m_head, flags);
3683 	if (m_head->m_pkthdr.csum_flags & CSUM_SND_TAG)
3684 		pi.ipi_mbuf = m_head;
3685 	else
3686 		pi.ipi_mbuf = NULL;
3687 	/*
3688 	 * XXX assumes a 1 to 1 relationship between segments and
3689 	 *        descriptors - this does not hold true on all drivers, e.g.
3690 	 *        cxgb
3691 	 */
3692 	if (__predict_false(nsegs > TXQ_AVAIL(txq))) {
3693 		iflib_completed_tx_reclaim_force(txq);
3694 		if (__predict_false(nsegs > TXQ_AVAIL(txq))) {
3695 			txq->ift_no_desc_avail++;
3696 			bus_dmamap_unload(buf_tag, map);
3697 			DBG_COUNTER_INC(encap_txq_avail_fail);
3698 			DBG_COUNTER_INC(encap_txd_encap_fail);
3699 			if (ctx->ifc_sysctl_simple_tx) {
3700 				*m_headp = m_head = iflib_remove_mbuf(txq);
3701 				m_freem(*m_headp);
3702 				DBG_COUNTER_INC(tx_frees);
3703 				*m_headp = NULL;
3704 			}
3705 			if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0)
3706 				GROUPTASK_ENQUEUE(&txq->ift_task);
3707 			return (ENOBUFS);
3708 		}
3709 	}
3710 	/*
3711 	 * On Intel cards we can greatly reduce the number of TX interrupts
3712 	 * we see by only setting report status on every Nth descriptor.
3713 	 * However, this also means that the driver will need to keep track
3714 	 * of the descriptors that RS was set on to check them for the DD bit.
3715 	 */
3716 	txq->ift_rs_pending += nsegs + 1;
3717 	if (txq->ift_rs_pending > TXQ_MAX_RS_DEFERRED(txq) ||
3718 	    iflib_no_tx_batch || (TXQ_AVAIL(txq) - nsegs) <= MAX_TX_DESC(ctx)) {
3719 		pi.ipi_flags |= IPI_TX_INTR;
3720 		txq->ift_rs_pending = 0;
3721 	}
3722 
3723 	pi.ipi_segs = segs;
3724 	pi.ipi_nsegs = nsegs;
3725 
3726 	MPASS(pidx >= 0 && pidx < txq->ift_size);
3727 #ifdef PKT_DEBUG
3728 	print_pkt(&pi);
3729 #endif
3730 	if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) {
3731 		bus_dmamap_sync(buf_tag, map, BUS_DMASYNC_PREWRITE);
3732 		DBG_COUNTER_INC(tx_encap);
3733 		MPASS(pi.ipi_new_pidx < txq->ift_size);
3734 
3735 		ndesc = pi.ipi_new_pidx - pi.ipi_pidx;
3736 		if (pi.ipi_new_pidx < pi.ipi_pidx) {
3737 			ndesc += txq->ift_size;
3738 			txq->ift_gen = 1;
3739 		}
3740 		/*
3741 		 * drivers can need up to ift_pad sentinels
3742 		 */
3743 		MPASS(ndesc <= pi.ipi_nsegs + txq->ift_pad);
3744 		MPASS(pi.ipi_new_pidx != pidx);
3745 		MPASS(ndesc > 0);
3746 		txq->ift_in_use += ndesc;
3747 		txq->ift_db_pending += ndesc;
3748 
3749 		/*
3750 		 * We update the last software descriptor again here because there may
3751 		 * be a sentinel and/or there may be more mbufs than segments
3752 		 */
3753 		txq->ift_pidx = pi.ipi_new_pidx;
3754 		txq->ift_npending += pi.ipi_ndescs;
3755 
3756 		/*
3757 		 * Update packets / bytes sent
3758 		 */
3759 		if (flags & IFLIB_TSO) {
3760 			int hlen = pi.ipi_ehdrlen + pi.ipi_ip_hlen + pi.ipi_tcp_hlen;
3761 			int tsolen = pi.ipi_len - hlen;
3762 			int nsegs = (tsolen + pi.ipi_tso_segsz - 1) / pi.ipi_tso_segsz;
3763 			*obytes += tsolen + nsegs * hlen;
3764 			*opkts += nsegs;
3765 		} else {
3766 			*obytes += pi.ipi_len;
3767 			*opkts += 1;
3768 		}
3769 	} else {
3770 		*m_headp = m_head = iflib_remove_mbuf(txq);
3771 		if (err == EFBIG) {
3772 			txq->ift_txd_encap_efbig++;
3773 			if (remap < 2) {
3774 				remap = 1;
3775 				goto defrag;
3776 			}
3777 			goto defrag_failed;
3778 		}
3779 		goto out_with_error;
3780 	}
3781 	/*
3782 	 * err can't possibly be non-zero here, so we don't neet to test it
3783 	 * to see if we need to DBG_COUNTER_INC(encap_txd_encap_fail).
3784 	 */
3785 	return (err);
3786 
3787 defrag_failed:
3788 	err = ENOMEM;
3789 	txq->ift_mbuf_defrag_failed++;
3790 out_with_error:
3791 	txq->ift_map_failed++;
3792 	m_freem(*m_headp);
3793 	DBG_COUNTER_INC(tx_frees);
3794 	*m_headp = NULL;
3795 	DBG_COUNTER_INC(encap_txd_encap_fail);
3796 	return (err);
3797 }
3798 
3799 static void
3800 iflib_tx_desc_free(iflib_txq_t txq, int n, struct mbuf **m_defer)
3801 {
3802 	uint32_t qsize, cidx, gen;
3803 	struct mbuf *m, **ifsd_m;
3804 	uintptr_t flags;
3805 
3806 	cidx = txq->ift_cidx;
3807 	gen = txq->ift_gen;
3808 	qsize = txq->ift_size;
3809 	ifsd_m =txq->ift_sds.ifsd_m;
3810 
3811 	while (n-- > 0) {
3812 		if ((m = IFLIB_GET_MBUF(ifsd_m[cidx])) != NULL) {
3813 			flags = IFLIB_GET_FLAGS(ifsd_m[cidx]);
3814 			MPASS(flags != 0);
3815 			if (flags & IFLIB_TSO) {
3816 				bus_dmamap_sync(txq->ift_tso_buf_tag,
3817 				    txq->ift_sds.ifsd_tso_map[cidx],
3818 				    BUS_DMASYNC_POSTWRITE);
3819 				bus_dmamap_unload(txq->ift_tso_buf_tag,
3820 				    txq->ift_sds.ifsd_tso_map[cidx]);
3821 			} else {
3822 				bus_dmamap_sync(txq->ift_buf_tag,
3823 				    txq->ift_sds.ifsd_map[cidx],
3824 				    BUS_DMASYNC_POSTWRITE);
3825 				bus_dmamap_unload(txq->ift_buf_tag,
3826 				    txq->ift_sds.ifsd_map[cidx]);
3827 			}
3828 			/* XXX we don't support any drivers that batch packets yet */
3829 			MPASS(m->m_nextpkt == NULL);
3830 			if (m_defer == NULL) {
3831 				m_freem(m);
3832 			} else if (m != NULL) {
3833 				*m_defer = m;
3834 				m_defer++;
3835 			}
3836 			ifsd_m[cidx] = NULL;
3837 #if MEMORY_LOGGING
3838 			txq->ift_dequeued++;
3839 #endif
3840 			DBG_COUNTER_INC(tx_frees);
3841 		}
3842 		if (__predict_false(++cidx == qsize)) {
3843 			cidx = 0;
3844 			gen = 0;
3845 		}
3846 	}
3847 	txq->ift_cidx = cidx;
3848 	txq->ift_gen = gen;
3849 }
3850 
3851 static __inline int
3852 iflib_txq_can_reclaim(iflib_txq_t txq)
3853 {
3854 	int reclaim, thresh;
3855 
3856 	thresh = txq->ift_reclaim_thresh;
3857 	KASSERT(thresh >= 0, ("invalid threshold to reclaim"));
3858 	MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size);
3859 
3860 	if (ticks <= (txq->ift_last_reclaim + txq->ift_reclaim_ticks) &&
3861 	    txq->ift_in_use < thresh)
3862 		return (false);
3863 	iflib_tx_credits_update(txq->ift_ctx, txq);
3864 	reclaim = DESC_RECLAIMABLE(txq);
3865 	if (reclaim <= thresh) {
3866 #ifdef INVARIANTS
3867 		if (iflib_verbose_debug) {
3868 			printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __func__,
3869 			    txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments,
3870 			    reclaim, thresh);
3871 		}
3872 #endif
3873 		return (0);
3874 	}
3875 	return (reclaim);
3876 }
3877 
3878 static __inline void
3879 _iflib_completed_tx_reclaim(iflib_txq_t txq, struct mbuf **m_defer, int reclaim)
3880 {
3881 	txq->ift_last_reclaim = ticks;
3882 	iflib_tx_desc_free(txq, reclaim, m_defer);
3883 	txq->ift_cleaned += reclaim;
3884 	txq->ift_in_use -= reclaim;
3885 }
3886 
3887 static __inline int
3888 iflib_completed_tx_reclaim(iflib_txq_t txq, struct mbuf **m_defer)
3889 {
3890 	int reclaim;
3891 
3892 	reclaim = iflib_txq_can_reclaim(txq);
3893 	if (reclaim == 0)
3894 		return (0);
3895 	_iflib_completed_tx_reclaim(txq, m_defer, reclaim);
3896 	return (reclaim);
3897 }
3898 
3899 /*
3900  * Reclaim any transmit descriptors possible, ignoring coalescing
3901  */
3902 static __inline void
3903 iflib_completed_tx_reclaim_force(iflib_txq_t txq)
3904 {
3905 	int reclaim;
3906 
3907 	iflib_tx_credits_update(txq->ift_ctx, txq);
3908 	reclaim = DESC_RECLAIMABLE(txq);
3909 	if (reclaim != 0)
3910 		_iflib_completed_tx_reclaim(txq, NULL, reclaim);
3911 }
3912 
3913 static struct mbuf **
3914 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset, int remaining)
3915 {
3916 	int next, size;
3917 	struct mbuf **items;
3918 
3919 	size = r->size;
3920 	next = (cidx + CACHE_PTR_INCREMENT) & (size - 1);
3921 	items = __DEVOLATILE(struct mbuf **, &r->items[0]);
3922 
3923 	prefetch(items[(cidx + offset) & (size - 1)]);
3924 	if (remaining > 1) {
3925 		prefetch2cachelines(&items[next]);
3926 		prefetch2cachelines(items[(cidx + offset + 1) & (size - 1)]);
3927 		prefetch2cachelines(items[(cidx + offset + 2) & (size - 1)]);
3928 		prefetch2cachelines(items[(cidx + offset + 3) & (size - 1)]);
3929 	}
3930 	return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (size - 1)]));
3931 }
3932 
3933 static void
3934 iflib_txq_check_drain(iflib_txq_t txq, int budget)
3935 {
3936 
3937 	ifmp_ring_check_drainage(txq->ift_br, budget);
3938 }
3939 
3940 static uint32_t
3941 iflib_txq_can_drain(struct ifmp_ring *r)
3942 {
3943 	iflib_txq_t txq = r->cookie;
3944 	if_ctx_t ctx = txq->ift_ctx;
3945 
3946 	if (TXQ_AVAIL(txq) > MAX_TX_DESC(ctx))
3947 		return (1);
3948 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3949 	    BUS_DMASYNC_POSTREAD);
3950 	return (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id,
3951 	    false));
3952 }
3953 
3954 static uint32_t
3955 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3956 {
3957 	iflib_txq_t txq = r->cookie;
3958 	if_ctx_t ctx = txq->ift_ctx;
3959 	if_t ifp = ctx->ifc_ifp;
3960 	struct mbuf *m, **mp;
3961 	int avail, bytes_sent, consumed, count, err, i;
3962 	int mcast_sent, pkt_sent, reclaimed;
3963 	bool do_prefetch, rang, ring;
3964 
3965 	if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) ||
3966 			    !LINK_ACTIVE(ctx))) {
3967 		DBG_COUNTER_INC(txq_drain_notready);
3968 		return (0);
3969 	}
3970 	reclaimed = iflib_completed_tx_reclaim(txq, NULL);
3971 	rang = iflib_txd_db_check(txq, reclaimed && txq->ift_db_pending);
3972 	avail = IDXDIFF(pidx, cidx, r->size);
3973 
3974 	if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) {
3975 		/*
3976 		 * The driver is unloading so we need to free all pending packets.
3977 		 */
3978 		DBG_COUNTER_INC(txq_drain_flushing);
3979 		for (i = 0; i < avail; i++) {
3980 			if (__predict_true(r->items[(cidx + i) & (r->size - 1)] != (void *)txq))
3981 				m_freem(r->items[(cidx + i) & (r->size - 1)]);
3982 			r->items[(cidx + i) & (r->size - 1)] = NULL;
3983 		}
3984 		return (avail);
3985 	}
3986 
3987 	if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3988 		CALLOUT_LOCK(txq);
3989 		callout_stop(&txq->ift_timer);
3990 		CALLOUT_UNLOCK(txq);
3991 		DBG_COUNTER_INC(txq_drain_oactive);
3992 		return (0);
3993 	}
3994 
3995 	consumed = mcast_sent = bytes_sent = pkt_sent = 0;
3996 	count = MIN(avail, TX_BATCH_SIZE);
3997 #ifdef INVARIANTS
3998 	if (iflib_verbose_debug)
3999 		printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __func__,
4000 		    avail, ctx->ifc_flags, TXQ_AVAIL(txq));
4001 #endif
4002 	do_prefetch = (ctx->ifc_flags & IFC_PREFETCH);
4003 	err = 0;
4004 	for (i = 0; i < count && TXQ_AVAIL(txq) >= MAX_TX_DESC(ctx); i++) {
4005 		int rem = do_prefetch ? count - i : 0;
4006 
4007 		mp = _ring_peek_one(r, cidx, i, rem);
4008 		MPASS(mp != NULL && *mp != NULL);
4009 
4010 		/*
4011 		 * Completion interrupts will use the address of the txq
4012 		 * as a sentinel to enqueue _something_ in order to acquire
4013 		 * the lock on the mp_ring (there's no direct lock call).
4014 		 * We obviously whave to check for these sentinel cases
4015 		 * and skip them.
4016 		 */
4017 		if (__predict_false(*mp == (struct mbuf *)txq)) {
4018 			consumed++;
4019 			continue;
4020 		}
4021 		err = iflib_encap(txq, mp, &bytes_sent, &pkt_sent);
4022 		if (__predict_false(err)) {
4023 			/* no room - bail out */
4024 			if (err == ENOBUFS)
4025 				break;
4026 			consumed++;
4027 			/* we can't send this packet - skip it */
4028 			continue;
4029 		}
4030 		consumed++;
4031 		m = *mp;
4032 		DBG_COUNTER_INC(tx_sent);
4033 		mcast_sent += !!(m->m_flags & M_MCAST);
4034 
4035 		if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING)))
4036 			break;
4037 		ETHER_BPF_MTAP(ifp, m);
4038 		rang = iflib_txd_db_check(txq, false);
4039 	}
4040 
4041 	/* deliberate use of bitwise or to avoid gratuitous short-circuit */
4042 	ring = rang ? false  : (iflib_min_tx_latency | err | (!!txq->ift_reclaim_thresh));
4043 	iflib_txd_db_check(txq, ring);
4044 	if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
4045 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
4046 	if (mcast_sent)
4047 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
4048 #ifdef INVARIANTS
4049 	if (iflib_verbose_debug)
4050 		printf("consumed=%d\n", consumed);
4051 #endif
4052 	return (consumed);
4053 }
4054 
4055 static uint32_t
4056 iflib_txq_drain_always(struct ifmp_ring *r)
4057 {
4058 	return (1);
4059 }
4060 
4061 static uint32_t
4062 iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
4063 {
4064 	int i, avail;
4065 	struct mbuf **mp;
4066 	iflib_txq_t txq;
4067 
4068 	txq = r->cookie;
4069 
4070 	CALLOUT_LOCK(txq);
4071 	callout_stop(&txq->ift_timer);
4072 	CALLOUT_UNLOCK(txq);
4073 
4074 	avail = IDXDIFF(pidx, cidx, r->size);
4075 	for (i = 0; i < avail; i++) {
4076 		mp = _ring_peek_one(r, cidx, i, avail - i);
4077 		if (__predict_false(*mp == (struct mbuf *)txq))
4078 			continue;
4079 		m_freem(*mp);
4080 		DBG_COUNTER_INC(tx_frees);
4081 	}
4082 	MPASS(ifmp_ring_is_stalled(r) == 0);
4083 	return (avail);
4084 }
4085 
4086 static void
4087 iflib_ifmp_purge(iflib_txq_t txq)
4088 {
4089 	struct ifmp_ring *r;
4090 
4091 	r = txq->ift_br;
4092 	r->drain = iflib_txq_drain_free;
4093 	r->can_drain = iflib_txq_drain_always;
4094 
4095 	ifmp_ring_check_drainage(r, r->size);
4096 
4097 	r->drain = iflib_txq_drain;
4098 	r->can_drain = iflib_txq_can_drain;
4099 }
4100 
4101 static void
4102 _task_fn_tx(void *context)
4103 {
4104 	iflib_txq_t txq = context;
4105 	if_ctx_t ctx = txq->ift_ctx;
4106 	if_t ifp = ctx->ifc_ifp;
4107 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
4108 
4109 #ifdef IFLIB_DIAGNOSTICS
4110 	txq->ift_cpu_exec_count[curcpu]++;
4111 #endif
4112 	if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
4113 		return;
4114 #ifdef DEV_NETMAP
4115 	if ((if_getcapenable(ifp) & IFCAP_NETMAP) &&
4116 	    netmap_tx_irq(ifp, txq->ift_id))
4117 		goto skip_ifmp;
4118 #endif
4119         if (ctx->ifc_sysctl_simple_tx) {
4120                 mtx_lock(&txq->ift_mtx);
4121                 (void)iflib_completed_tx_reclaim(txq, NULL);
4122                 mtx_unlock(&txq->ift_mtx);
4123                 goto skip_ifmp;
4124         }
4125 #ifdef ALTQ
4126 	if (if_altq_is_enabled(ifp))
4127 		iflib_altq_if_start(ifp);
4128 #endif
4129 	if (txq->ift_db_pending)
4130 		ifmp_ring_enqueue(txq->ift_br, (void **)&txq, 1, TX_BATCH_SIZE, abdicate);
4131 	else if (!abdicate)
4132 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
4133 	/*
4134 	 * When abdicating, we always need to check drainage, not just when we don't enqueue
4135 	 */
4136 	if (abdicate)
4137 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
4138 
4139 skip_ifmp:
4140 	if (ctx->ifc_flags & IFC_LEGACY)
4141 		IFDI_INTR_ENABLE(ctx);
4142 	else
4143 		IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
4144 }
4145 
4146 static void
4147 _task_fn_rx(void *context)
4148 {
4149 	iflib_rxq_t rxq = context;
4150 	if_ctx_t ctx = rxq->ifr_ctx;
4151 	uint8_t more;
4152 	uint16_t budget;
4153 #ifdef DEV_NETMAP
4154 	u_int work = 0;
4155 	int nmirq;
4156 #endif
4157 
4158 #ifdef IFLIB_DIAGNOSTICS
4159 	rxq->ifr_cpu_exec_count[curcpu]++;
4160 #endif
4161 	DBG_COUNTER_INC(task_fn_rxs);
4162 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
4163 		return;
4164 #ifdef DEV_NETMAP
4165 	nmirq = netmap_rx_irq(ctx->ifc_ifp, rxq->ifr_id, &work);
4166 	if (nmirq != NM_IRQ_PASS) {
4167 		more = (nmirq == NM_IRQ_RESCHED) ? IFLIB_RXEOF_MORE : 0;
4168 		goto skip_rxeof;
4169 	}
4170 #endif
4171 	budget = ctx->ifc_sysctl_rx_budget;
4172 	if (budget == 0)
4173 		budget = 16;	/* XXX */
4174 	more = iflib_rxeof(rxq, budget);
4175 #ifdef DEV_NETMAP
4176 skip_rxeof:
4177 #endif
4178 	if ((more & IFLIB_RXEOF_MORE) == 0) {
4179 		if (ctx->ifc_flags & IFC_LEGACY)
4180 			IFDI_INTR_ENABLE(ctx);
4181 		else
4182 			IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
4183 		DBG_COUNTER_INC(rx_intr_enables);
4184 	}
4185 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
4186 		return;
4187 
4188 	if (more & IFLIB_RXEOF_MORE)
4189 		GROUPTASK_ENQUEUE(&rxq->ifr_task);
4190 	else if (more & IFLIB_RXEOF_EMPTY)
4191 		callout_reset_curcpu(&rxq->ifr_watchdog, 1, &_task_fn_rx_watchdog, rxq);
4192 }
4193 
4194 static void
4195 _task_fn_admin(void *context, int pending)
4196 {
4197 	if_ctx_t ctx = context;
4198 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
4199 	iflib_txq_t txq;
4200 	int i;
4201 	bool oactive, running, do_reset, do_reset_if_up, do_watchdog;
4202 	bool in_detach;
4203 
4204 	STATE_LOCK(ctx);
4205 	running = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING);
4206 	oactive = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE);
4207 	do_reset = (ctx->ifc_flags & IFC_DO_RESET);
4208 	do_reset_if_up = (ctx->ifc_flags & IFC_DO_RESET_IF_UP);
4209 	do_watchdog = (ctx->ifc_flags & IFC_DO_WATCHDOG);
4210 	in_detach = (ctx->ifc_flags & IFC_IN_DETACH);
4211 	ctx->ifc_flags &= ~(IFC_DO_RESET | IFC_DO_RESET_IF_UP |
4212 	    IFC_DO_WATCHDOG);
4213 	STATE_UNLOCK(ctx);
4214 
4215 	if ((!running && !oactive) && !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
4216 		return;
4217 	if (in_detach)
4218 		return;
4219 
4220 	CTX_LOCK(ctx);
4221 	if (!do_reset && do_reset_if_up &&
4222 	    (if_getflags(ctx->ifc_ifp) & IFF_UP) != 0)
4223 		do_reset = true;
4224 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
4225 		CALLOUT_LOCK(txq);
4226 		callout_stop(&txq->ift_timer);
4227 		CALLOUT_UNLOCK(txq);
4228 	}
4229 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_ADMINCQ)
4230 		IFDI_ADMIN_COMPLETION_HANDLE(ctx);
4231 	if (do_watchdog) {
4232 		ctx->ifc_tx_watchdog_events++;
4233 		IFDI_WATCHDOG_RESET(ctx);
4234 	}
4235 	IFDI_UPDATE_ADMIN_STATUS(ctx);
4236 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
4237 		callout_reset_on(&txq->ift_timer, iflib_timer_default, iflib_timer, txq,
4238 		    txq->ift_timer.c_cpu);
4239 	}
4240 	IFDI_LINK_INTR_ENABLE(ctx);
4241 	if (do_reset)
4242 		iflib_if_init_locked(ctx);
4243 	CTX_UNLOCK(ctx);
4244 
4245 	if (LINK_ACTIVE(ctx) == 0)
4246 		return;
4247 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
4248 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
4249 }
4250 
4251 static void
4252 _task_fn_iov(void *context, int pending)
4253 {
4254 	if_ctx_t ctx = context;
4255 
4256 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING) &&
4257 	    !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
4258 		return;
4259 
4260 	CTX_LOCK(ctx);
4261 	IFDI_VFLR_HANDLE(ctx);
4262 	CTX_UNLOCK(ctx);
4263 }
4264 
4265 static int
4266 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)
4267 {
4268 	int err;
4269 	if_int_delay_info_t info;
4270 	if_ctx_t ctx;
4271 
4272 	info = (if_int_delay_info_t)arg1;
4273 	ctx = info->iidi_ctx;
4274 	info->iidi_req = req;
4275 	info->iidi_oidp = oidp;
4276 	CTX_LOCK(ctx);
4277 	err = IFDI_SYSCTL_INT_DELAY(ctx, info);
4278 	CTX_UNLOCK(ctx);
4279 	return (err);
4280 }
4281 
4282 /*********************************************************************
4283  *
4284  *  IFNET FUNCTIONS
4285  *
4286  **********************************************************************/
4287 
4288 static void
4289 iflib_if_init_locked(if_ctx_t ctx)
4290 {
4291 	iflib_stop(ctx);
4292 	iflib_init_locked(ctx);
4293 }
4294 
4295 static void
4296 iflib_if_init(void *arg)
4297 {
4298 	if_ctx_t ctx = arg;
4299 
4300 	CTX_LOCK(ctx);
4301 	iflib_if_init_locked(ctx);
4302 	CTX_UNLOCK(ctx);
4303 }
4304 
4305 static int
4306 iflib_if_transmit(if_t ifp, struct mbuf *m)
4307 {
4308 	if_ctx_t ctx = if_getsoftc(ifp);
4309 	iflib_txq_t txq;
4310 	int err, qidx;
4311 	int abdicate;
4312 
4313 	if (__predict_false((if_getdrvflags(ifp) & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) {
4314 		DBG_COUNTER_INC(tx_frees);
4315 		m_freem(m);
4316 		return (ENETDOWN);
4317 	}
4318 
4319 	MPASS(m->m_nextpkt == NULL);
4320 	/* ALTQ-enabled interfaces always use queue 0. */
4321 	qidx = 0;
4322 	/* Use driver-supplied queue selection method if it exists */
4323 	if (ctx->isc_txq_select_v2) {
4324 		struct if_pkt_info pi;
4325 		uint64_t early_pullups = 0;
4326 		memset(&pi, 0, sizeof(pi));
4327 
4328 		err = iflib_parse_header_partial(&pi, &m, &early_pullups);
4329 		if (__predict_false(err != 0)) {
4330 			/* Assign pullups for bad pkts to default queue */
4331 			ctx->ifc_txqs[0].ift_pullups += early_pullups;
4332 			DBG_COUNTER_INC(encap_txd_encap_fail);
4333 			return (err);
4334 		}
4335 		/* Let driver make queueing decision */
4336 		qidx = ctx->isc_txq_select_v2(ctx->ifc_softc, m, &pi);
4337 		ctx->ifc_txqs[qidx].ift_pullups += early_pullups;
4338 	}
4339 	/* Backwards compatibility w/ simpler queue select */
4340 	else if (ctx->isc_txq_select)
4341 		qidx = ctx->isc_txq_select(ctx->ifc_softc, m);
4342 	/* If not, use iflib's standard method */
4343 	else if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m) && !if_altq_is_enabled(ifp))
4344 		qidx = QIDX(ctx, m);
4345 
4346 	/* Set TX queue */
4347 	txq = &ctx->ifc_txqs[qidx];
4348 
4349 #ifdef DRIVER_BACKPRESSURE
4350 	if (txq->ift_closed) {
4351 		while (m != NULL) {
4352 			next = m->m_nextpkt;
4353 			m->m_nextpkt = NULL;
4354 			m_freem(m);
4355 			DBG_COUNTER_INC(tx_frees);
4356 			m = next;
4357 		}
4358 		return (ENOBUFS);
4359 	}
4360 #endif
4361 #ifdef notyet
4362 	qidx = count = 0;
4363 	mp = marr;
4364 	next = m;
4365 	do {
4366 		count++;
4367 		next = next->m_nextpkt;
4368 	} while (next != NULL);
4369 
4370 	if (count > nitems(marr))
4371 		if ((mp = malloc(count * sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) {
4372 			/* XXX check nextpkt */
4373 			m_freem(m);
4374 			/* XXX simplify for now */
4375 			DBG_COUNTER_INC(tx_frees);
4376 			return (ENOBUFS);
4377 		}
4378 	for (next = m, i = 0; next != NULL; i++) {
4379 		mp[i] = next;
4380 		next = next->m_nextpkt;
4381 		mp[i]->m_nextpkt = NULL;
4382 	}
4383 #endif
4384 	DBG_COUNTER_INC(tx_seen);
4385 	abdicate = ctx->ifc_sysctl_tx_abdicate;
4386 
4387 	err = ifmp_ring_enqueue(txq->ift_br, (void **)&m, 1, TX_BATCH_SIZE, abdicate);
4388 
4389 	if (abdicate)
4390 		GROUPTASK_ENQUEUE(&txq->ift_task);
4391 	if (err) {
4392 		if (!abdicate)
4393 			GROUPTASK_ENQUEUE(&txq->ift_task);
4394 		/* support forthcoming later */
4395 #ifdef DRIVER_BACKPRESSURE
4396 		txq->ift_closed = TRUE;
4397 #endif
4398 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
4399 		m_freem(m);
4400 		DBG_COUNTER_INC(tx_frees);
4401 		if (err == ENOBUFS)
4402 			if_inc_counter(ifp, IFCOUNTER_OQDROPS, 1);
4403 		else
4404 			if_inc_counter(ifp, IFCOUNTER_OERRORS, 1);
4405 	}
4406 
4407 	return (err);
4408 }
4409 
4410 #ifdef ALTQ
4411 /*
4412  * The overall approach to integrating iflib with ALTQ is to continue to use
4413  * the iflib mp_ring machinery between the ALTQ queue(s) and the hardware
4414  * ring.  Technically, when using ALTQ, queueing to an intermediate mp_ring
4415  * is redundant/unnecessary, but doing so minimizes the amount of
4416  * ALTQ-specific code required in iflib.  It is assumed that the overhead of
4417  * redundantly queueing to an intermediate mp_ring is swamped by the
4418  * performance limitations inherent in using ALTQ.
4419  *
4420  * When ALTQ support is compiled in, all iflib drivers will use a transmit
4421  * routine, iflib_altq_if_transmit(), that checks if ALTQ is enabled for the
4422  * given interface.  If ALTQ is enabled for an interface, then all
4423  * transmitted packets for that interface will be submitted to the ALTQ
4424  * subsystem via IFQ_ENQUEUE().  We don't use the legacy if_transmit()
4425  * implementation because it uses IFQ_HANDOFF(), which will duplicatively
4426  * update stats that the iflib machinery handles, and which is sensitve to
4427  * the disused IFF_DRV_OACTIVE flag.  Additionally, iflib_altq_if_start()
4428  * will be installed as the start routine for use by ALTQ facilities that
4429  * need to trigger queue drains on a scheduled basis.
4430  *
4431  */
4432 static void
4433 iflib_altq_if_start(if_t ifp)
4434 {
4435 	struct ifaltq *ifq = &ifp->if_snd; /* XXX - DRVAPI */
4436 	struct mbuf *m;
4437 
4438 	IFQ_LOCK(ifq);
4439 	IFQ_DEQUEUE_NOLOCK(ifq, m);
4440 	while (m != NULL) {
4441 		iflib_if_transmit(ifp, m);
4442 		IFQ_DEQUEUE_NOLOCK(ifq, m);
4443 	}
4444 	IFQ_UNLOCK(ifq);
4445 }
4446 
4447 static int
4448 iflib_altq_if_transmit(if_t ifp, struct mbuf *m)
4449 {
4450 	int err;
4451 
4452 	if (if_altq_is_enabled(ifp)) {
4453 		IFQ_ENQUEUE(&ifp->if_snd, m, err); /* XXX - DRVAPI */
4454 		if (err == 0)
4455 			iflib_altq_if_start(ifp);
4456 	} else
4457 		err = iflib_if_transmit(ifp, m);
4458 
4459 	return (err);
4460 }
4461 #endif /* ALTQ */
4462 
4463 static void
4464 iflib_if_qflush(if_t ifp)
4465 {
4466 	if_ctx_t ctx = if_getsoftc(ifp);
4467 	iflib_txq_t txq = ctx->ifc_txqs;
4468 	int i;
4469 
4470 	STATE_LOCK(ctx);
4471 	ctx->ifc_flags |= IFC_QFLUSH;
4472 	STATE_UNLOCK(ctx);
4473 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
4474 		while (!(ifmp_ring_is_idle(txq->ift_br) || ifmp_ring_is_stalled(txq->ift_br)))
4475 			iflib_txq_check_drain(txq, 0);
4476 	STATE_LOCK(ctx);
4477 	ctx->ifc_flags &= ~IFC_QFLUSH;
4478 	STATE_UNLOCK(ctx);
4479 
4480 	/*
4481 	 * When ALTQ is enabled, this will also take care of purging the
4482 	 * ALTQ queue(s).
4483 	 */
4484 	if_qflush(ifp);
4485 }
4486 
4487 #define IFCAP_FLAGS (IFCAP_HWCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \
4488 		    IFCAP_TSO | IFCAP_VLAN_HWTAGGING | IFCAP_HWSTATS | \
4489 		    IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | \
4490 		    IFCAP_VLAN_HWTSO | IFCAP_VLAN_HWCSUM | IFCAP_MEXTPG)
4491 
4492 static int
4493 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data)
4494 {
4495 	if_ctx_t ctx = if_getsoftc(ifp);
4496 	struct ifreq	*ifr = (struct ifreq *)data;
4497 #if defined(INET) || defined(INET6)
4498 	struct ifaddr	*ifa = (struct ifaddr *)data;
4499 #endif
4500 	bool		avoid_reset = false;
4501 	int		err = 0, reinit = 0, bits;
4502 
4503 	switch (command) {
4504 	case SIOCSIFADDR:
4505 #ifdef INET
4506 		if (ifa->ifa_addr->sa_family == AF_INET)
4507 			avoid_reset = true;
4508 #endif
4509 #ifdef INET6
4510 		if (ifa->ifa_addr->sa_family == AF_INET6)
4511 			avoid_reset = true;
4512 #endif
4513 		/*
4514 		 * Calling init results in link renegotiation,
4515 		 * so we avoid doing it when possible.
4516 		 */
4517 		if (avoid_reset) {
4518 			if_setflagbits(ifp, IFF_UP, 0);
4519 			if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
4520 				reinit = 1;
4521 #ifdef INET
4522 			if (!(if_getflags(ifp) & IFF_NOARP))
4523 				arp_ifinit(ifp, ifa);
4524 #endif
4525 		} else
4526 			err = ether_ioctl(ifp, command, data);
4527 		break;
4528 	case SIOCSIFMTU:
4529 		CTX_LOCK(ctx);
4530 		if (ifr->ifr_mtu == if_getmtu(ifp)) {
4531 			CTX_UNLOCK(ctx);
4532 			break;
4533 		}
4534 		bits = if_getdrvflags(ifp);
4535 		/* stop the driver and free any clusters before proceeding */
4536 		iflib_stop(ctx);
4537 
4538 		if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) {
4539 			STATE_LOCK(ctx);
4540 			if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size)
4541 				ctx->ifc_flags |= IFC_MULTISEG;
4542 			else
4543 				ctx->ifc_flags &= ~IFC_MULTISEG;
4544 			STATE_UNLOCK(ctx);
4545 			err = if_setmtu(ifp, ifr->ifr_mtu);
4546 		}
4547 		iflib_init_locked(ctx);
4548 		STATE_LOCK(ctx);
4549 		/* Preserve the stopped state reported by iflib_init_failed(). */
4550 		if ((ctx->ifc_flags & IFC_INIT_FAILED) == 0)
4551 			if_setdrvflags(ifp, bits);
4552 		STATE_UNLOCK(ctx);
4553 		CTX_UNLOCK(ctx);
4554 		break;
4555 	case SIOCSIFFLAGS:
4556 		CTX_LOCK(ctx);
4557 		if (if_getflags(ifp) & IFF_UP) {
4558 			if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4559 				if ((if_getflags(ifp) ^ ctx->ifc_if_flags) &
4560 				    (IFF_PROMISC | IFF_ALLMULTI)) {
4561 					CTX_UNLOCK(ctx);
4562 					err = IFDI_PROMISC_SET(ctx, if_getflags(ifp));
4563 					CTX_LOCK(ctx);
4564 				}
4565 			} else
4566 				reinit = 1;
4567 		} else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4568 			iflib_stop(ctx);
4569 		}
4570 		ctx->ifc_if_flags = if_getflags(ifp);
4571 		CTX_UNLOCK(ctx);
4572 		break;
4573 	case SIOCADDMULTI:
4574 	case SIOCDELMULTI:
4575 		if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4576 			CTX_LOCK(ctx);
4577 			IFDI_INTR_DISABLE(ctx);
4578 			IFDI_MULTI_SET(ctx);
4579 			IFDI_INTR_ENABLE(ctx);
4580 			CTX_UNLOCK(ctx);
4581 		}
4582 		break;
4583 	case SIOCSIFMEDIA:
4584 		CTX_LOCK(ctx);
4585 		IFDI_MEDIA_SET(ctx);
4586 		CTX_UNLOCK(ctx);
4587 		/* FALLTHROUGH */
4588 	case SIOCGIFMEDIA:
4589 	case SIOCGIFXMEDIA:
4590 		err = ifmedia_ioctl(ifp, ifr, ctx->ifc_mediap, command);
4591 		break;
4592 	case SIOCGI2C:
4593 		/* FALLTHROUGH */
4594 	case SIOCGI2CPB:
4595 	{
4596 		struct ifi2creq i2c;
4597 		if_shared_ctx_t sctx = ctx->ifc_sctx;
4598 
4599 		err = copyin(ifr_data_get_ptr(ifr), &i2c, sizeof(i2c));
4600 		if (err != 0)
4601 			break;
4602 		if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) {
4603 			err = EINVAL;
4604 			break;
4605 		}
4606 		if (i2c.len > sizeof(i2c.data)) {
4607 			err = EINVAL;
4608 			break;
4609 		}
4610 		if (command == SIOCGI2C) {
4611 			i2c.page = i2c.bank = 0;
4612 		} else if ((sctx->isc_flags & IFLIB_I2C_PAGE_BANK) == 0) {
4613 			err = EINVAL;
4614 			break;
4615 		}
4616 
4617 		if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0)
4618 			err = copyout(&i2c, ifr_data_get_ptr(ifr),
4619 			    sizeof(i2c));
4620 		break;
4621 	}
4622 	case SIOCSIFCAP:
4623 	{
4624 		int mask, setmask, oldmask;
4625 
4626 		oldmask = if_getcapenable(ifp);
4627 		mask = ifr->ifr_reqcap ^ oldmask;
4628 		mask &= ctx->ifc_softc_ctx.isc_capabilities | IFCAP_MEXTPG;
4629 		setmask = 0;
4630 #ifdef TCP_OFFLOAD
4631 		setmask |= mask & (IFCAP_TOE4 | IFCAP_TOE6);
4632 #endif
4633 		setmask |= (mask & IFCAP_FLAGS);
4634 		setmask |= (mask & IFCAP_WOL);
4635 
4636 		/*
4637 		 * If any RX csum has changed, change all the ones that
4638 		 * are supported by the driver.
4639 		 */
4640 		if (setmask & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6)) {
4641 			setmask |= ctx->ifc_softc_ctx.isc_capabilities &
4642 			    (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6);
4643 		}
4644 
4645 		/*
4646 		 * want to ensure that traffic has stopped before we change any of the flags
4647 		 */
4648 		if (setmask) {
4649 			CTX_LOCK(ctx);
4650 			bits = if_getdrvflags(ifp);
4651 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4652 				iflib_stop(ctx);
4653 			STATE_LOCK(ctx);
4654 			if_togglecapenable(ifp, setmask);
4655 			ctx->ifc_softc_ctx.isc_capenable ^= setmask;
4656 			STATE_UNLOCK(ctx);
4657 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4658 				iflib_init_locked(ctx);
4659 			STATE_LOCK(ctx);
4660 			if ((ctx->ifc_flags & IFC_INIT_FAILED) == 0)
4661 				if_setdrvflags(ifp, bits);
4662 			STATE_UNLOCK(ctx);
4663 			CTX_UNLOCK(ctx);
4664 		}
4665 		if_vlancap(ifp);
4666 		break;
4667 	}
4668 	case SIOCGPRIVATE_0:
4669 	case SIOCSDRVSPEC:
4670 	case SIOCGDRVSPEC:
4671 		CTX_LOCK(ctx);
4672 		err = IFDI_PRIV_IOCTL(ctx, command, data);
4673 		CTX_UNLOCK(ctx);
4674 		break;
4675 	case SIOCGIFDOWNREASON:
4676 		CTX_LOCK(ctx);
4677 		err = IFDI_GET_DOWNREASON(ctx, (struct ifdownreason *)data);
4678 		CTX_UNLOCK(ctx);
4679 		break;
4680 	default:
4681 		err = ether_ioctl(ifp, command, data);
4682 		break;
4683 	}
4684 	if (reinit)
4685 		iflib_if_init(ctx);
4686 	return (err);
4687 }
4688 
4689 static uint64_t
4690 iflib_if_get_counter(if_t ifp, ift_counter cnt)
4691 {
4692 	if_ctx_t ctx = if_getsoftc(ifp);
4693 
4694 	return (IFDI_GET_COUNTER(ctx, cnt));
4695 }
4696 
4697 /*********************************************************************
4698  *
4699  *  OTHER FUNCTIONS EXPORTED TO THE STACK
4700  *
4701  **********************************************************************/
4702 
4703 static void
4704 iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag)
4705 {
4706 	if_ctx_t ctx = if_getsoftc(ifp);
4707 
4708 	if ((void *)ctx != arg)
4709 		return;
4710 
4711 	if ((vtag == 0) || (vtag > 4095))
4712 		return;
4713 
4714 	if (iflib_in_detach(ctx))
4715 		return;
4716 
4717 	CTX_LOCK(ctx);
4718 	/* Driver may need all untagged packets to be flushed */
4719 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4720 		iflib_stop(ctx);
4721 	IFDI_VLAN_REGISTER(ctx, vtag);
4722 	/* Re-init to load the changes, if required */
4723 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4724 		iflib_init_locked(ctx);
4725 	CTX_UNLOCK(ctx);
4726 }
4727 
4728 static void
4729 iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag)
4730 {
4731 	if_ctx_t ctx = if_getsoftc(ifp);
4732 
4733 	if ((void *)ctx != arg)
4734 		return;
4735 
4736 	if ((vtag == 0) || (vtag > 4095))
4737 		return;
4738 
4739 	CTX_LOCK(ctx);
4740 	/* Driver may need all tagged packets to be flushed */
4741 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4742 		iflib_stop(ctx);
4743 	IFDI_VLAN_UNREGISTER(ctx, vtag);
4744 	/* Re-init to load the changes, if required */
4745 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4746 		iflib_init_locked(ctx);
4747 	CTX_UNLOCK(ctx);
4748 }
4749 
4750 static void
4751 _task_fn_led(void *context, int pending __unused)
4752 {
4753 	if_ctx_t ctx = context;
4754 	bool in_detach;
4755 	int onoff;
4756 
4757 	STATE_LOCK(ctx);
4758 	in_detach = (ctx->ifc_flags & IFC_IN_DETACH) != 0;
4759 	onoff = ctx->ifc_led_state;
4760 	STATE_UNLOCK(ctx);
4761 	if (in_detach)
4762 		return;
4763 
4764 	CTX_LOCK(ctx);
4765 	IFDI_LED_FUNC(ctx, onoff);
4766 	CTX_UNLOCK(ctx);
4767 }
4768 
4769 static void
4770 iflib_led_func(void *arg, int onoff)
4771 {
4772 	if_ctx_t ctx = arg;
4773 	bool in_detach;
4774 
4775 	/* led(4) may invoke this callback from a non-sleepable callout. */
4776 	STATE_LOCK(ctx);
4777 	ctx->ifc_led_state = onoff;
4778 	in_detach = (ctx->ifc_flags & IFC_IN_DETACH) != 0;
4779 	STATE_UNLOCK(ctx);
4780 	if (!in_detach)
4781 		taskqueue_enqueue(ctx->ifc_tq, &ctx->ifc_led_task);
4782 }
4783 
4784 /*********************************************************************
4785  *
4786  *  BUS FUNCTION DEFINITIONS
4787  *
4788  **********************************************************************/
4789 
4790 int
4791 iflib_device_probe(device_t dev)
4792 {
4793 	const pci_vendor_info_t *ent;
4794 	if_shared_ctx_t sctx;
4795 	uint16_t pci_device_id, pci_rev_id, pci_subdevice_id, pci_subvendor_id;
4796 	uint16_t pci_vendor_id;
4797 
4798 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4799 		return (ENOTSUP);
4800 
4801 	pci_vendor_id = pci_get_vendor(dev);
4802 	pci_device_id = pci_get_device(dev);
4803 	pci_subvendor_id = pci_get_subvendor(dev);
4804 	pci_subdevice_id = pci_get_subdevice(dev);
4805 	pci_rev_id = pci_get_revid(dev);
4806 	if (sctx->isc_parse_devinfo != NULL)
4807 		sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id);
4808 
4809 	ent = sctx->isc_vendor_info;
4810 	while (ent->pvi_vendor_id != 0) {
4811 		if (pci_vendor_id != ent->pvi_vendor_id) {
4812 			ent++;
4813 			continue;
4814 		}
4815 		if ((pci_device_id == ent->pvi_device_id) &&
4816 		    ((pci_subvendor_id == ent->pvi_subvendor_id) ||
4817 		     (ent->pvi_subvendor_id == 0)) &&
4818 		    ((pci_subdevice_id == ent->pvi_subdevice_id) ||
4819 		     (ent->pvi_subdevice_id == 0)) &&
4820 		    ((pci_rev_id == ent->pvi_rev_id) ||
4821 		     (ent->pvi_rev_id == 0))) {
4822 			device_set_desc_copy(dev, ent->pvi_name);
4823 			/* this needs to be changed to zero if the bus probing code
4824 			 * ever stops re-probing on best match because the sctx
4825 			 * may have its values over written by register calls
4826 			 * in subsequent probes
4827 			 */
4828 			return (BUS_PROBE_DEFAULT);
4829 		}
4830 		ent++;
4831 	}
4832 	return (ENXIO);
4833 }
4834 
4835 int
4836 iflib_device_probe_vendor(device_t dev)
4837 {
4838 	int probe;
4839 
4840 	probe = iflib_device_probe(dev);
4841 	if (probe == BUS_PROBE_DEFAULT)
4842 		return (BUS_PROBE_VENDOR);
4843 	else
4844 		return (probe);
4845 }
4846 
4847 static void
4848 iflib_reset_qvalues(if_ctx_t ctx)
4849 {
4850 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4851 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4852 	device_t dev = ctx->ifc_dev;
4853 	int i;
4854 
4855 	if (ctx->ifc_sysctl_ntxqs != 0)
4856 		scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs;
4857 	if (ctx->ifc_sysctl_nrxqs != 0)
4858 		scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs;
4859 
4860 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4861 		if (ctx->ifc_sysctl_ntxds[i] != 0)
4862 			scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i];
4863 		else
4864 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4865 	}
4866 
4867 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4868 		if (ctx->ifc_sysctl_nrxds[i] != 0)
4869 			scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i];
4870 		else
4871 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4872 	}
4873 
4874 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4875 		if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) {
4876 			device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n",
4877 			    i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]);
4878 			scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i];
4879 		}
4880 		if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) {
4881 			device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n",
4882 			    i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]);
4883 			scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i];
4884 		}
4885 		if (!powerof2(scctx->isc_nrxd[i])) {
4886 			device_printf(dev, "nrxd%d: %d is not a power of 2 - using default value of %d\n",
4887 			    i, scctx->isc_nrxd[i], sctx->isc_nrxd_default[i]);
4888 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4889 		}
4890 	}
4891 
4892 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4893 		if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) {
4894 			device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n",
4895 			    i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]);
4896 			scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i];
4897 		}
4898 		if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) {
4899 			device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n",
4900 			    i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]);
4901 			scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i];
4902 		}
4903 		if (!powerof2(scctx->isc_ntxd[i])) {
4904 			device_printf(dev, "ntxd%d: %d is not a power of 2 - using default value of %d\n",
4905 			    i, scctx->isc_ntxd[i], sctx->isc_ntxd_default[i]);
4906 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4907 		}
4908 	}
4909 	scctx->isc_tx_pad = 2;
4910 }
4911 
4912 static void
4913 iflib_add_pfil(if_ctx_t ctx)
4914 {
4915 	struct pfil_head *pfil;
4916 	struct pfil_head_args pa;
4917 	iflib_rxq_t rxq;
4918 	int i;
4919 
4920 	pa.pa_version = PFIL_VERSION;
4921 	pa.pa_flags = PFIL_IN;
4922 	pa.pa_type = PFIL_TYPE_ETHERNET;
4923 	pa.pa_headname = if_name(ctx->ifc_ifp);
4924 	pfil = pfil_head_register(&pa);
4925 
4926 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4927 		rxq->pfil = pfil;
4928 	}
4929 }
4930 
4931 static void
4932 iflib_rem_pfil(if_ctx_t ctx)
4933 {
4934 	struct pfil_head *pfil;
4935 	iflib_rxq_t rxq;
4936 	int i;
4937 
4938 	rxq = ctx->ifc_rxqs;
4939 	pfil = rxq->pfil;
4940 	for (i = 0; i < NRXQSETS(ctx); i++, rxq++) {
4941 		rxq->pfil = NULL;
4942 	}
4943 	pfil_head_unregister(pfil);
4944 }
4945 
4946 
4947 /*
4948  * Advance forward by n members of the cpuset ctx->ifc_cpus starting from
4949  * cpuid and wrapping as necessary.
4950  */
4951 static unsigned int
4952 cpuid_advance(if_ctx_t ctx, unsigned int cpuid, unsigned int n)
4953 {
4954 	unsigned int first_valid;
4955 	unsigned int last_valid;
4956 
4957 	/* cpuid should always be in the valid set */
4958 	MPASS(CPU_ISSET(cpuid, &ctx->ifc_cpus));
4959 
4960 	/* valid set should never be empty */
4961 	MPASS(!CPU_EMPTY(&ctx->ifc_cpus));
4962 
4963 	first_valid = CPU_FFS(&ctx->ifc_cpus) - 1;
4964 	last_valid = CPU_FLS(&ctx->ifc_cpus) - 1;
4965 	n = n % CPU_COUNT(&ctx->ifc_cpus);
4966 	while (n > 0) {
4967 		do {
4968 			cpuid++;
4969 			if (cpuid > last_valid)
4970 				cpuid = first_valid;
4971 		} while (!CPU_ISSET(cpuid, &ctx->ifc_cpus));
4972 		n--;
4973 	}
4974 
4975 	return (cpuid);
4976 }
4977 
4978 /*
4979  * CPU mapping behaviors
4980  * ---------------------
4981  * 'separate txrx' refers to the separate_txrx sysctl
4982  * 'use logical' refers to the use_logical_cores sysctl
4983  * 'INTR CPUS' indicates whether bus_get_cpus(INTR_CPUS) succeeded
4984  *
4985  *  separate     use     INTR
4986  *    txrx     logical   CPUS   result
4987  * ---------- --------- ------ ------------------------------------------------
4988  *     -          -       X     RX and TX queues mapped to consecutive physical
4989  *                              cores with RX/TX pairs on same core and excess
4990  *                              of either following
4991  *     -          X       X     RX and TX queues mapped to consecutive cores
4992  *                              of any type with RX/TX pairs on same core and
4993  *                              excess of either following
4994  *     X          -       X     RX and TX queues mapped to consecutive physical
4995  *                              cores; all RX then all TX
4996  *     X          X       X     RX queues mapped to consecutive physical cores
4997  *                              first, then TX queues mapped to L2 neighbor of
4998  *                              the corresponding RX queue if one exists,
4999  *                              otherwise to consecutive physical cores
5000  *     -         n/a      -     RX and TX queues mapped to consecutive cores of
5001  *                              any type with RX/TX pairs on same core and excess
5002  *                              of either following
5003  *     X         n/a      -     RX and TX queues mapped to consecutive cores of
5004  *                              any type; all RX then all TX
5005  */
5006 static unsigned int
5007 get_cpuid_for_queue(if_ctx_t ctx, unsigned int base_cpuid, unsigned int qid,
5008     bool is_tx)
5009 {
5010 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5011 	unsigned int core_index;
5012 
5013 	if (ctx->ifc_sysctl_separate_txrx) {
5014 		/*
5015 		 * When using separate CPUs for TX and RX, the assignment
5016 		 * will always be of a consecutive CPU out of the set of
5017 		 * context CPUs, except for the specific case where the
5018 		 * context CPUs are phsyical cores, the use of logical cores
5019 		 * has been enabled, the assignment is for TX, the TX qid
5020 		 * corresponds to an RX qid, and the CPU assigned to the
5021 		 * corresponding RX queue has an L2 neighbor.
5022 		 */
5023 		if (ctx->ifc_sysctl_use_logical_cores &&
5024 		    ctx->ifc_cpus_are_physical_cores &&
5025 		    is_tx && qid < scctx->isc_nrxqsets) {
5026 			int l2_neighbor;
5027 			unsigned int rx_cpuid;
5028 
5029 			rx_cpuid = cpuid_advance(ctx, base_cpuid, qid);
5030 			l2_neighbor = sched_find_l2_neighbor(rx_cpuid);
5031 			if (l2_neighbor != -1) {
5032 				return (l2_neighbor);
5033 			}
5034 			/*
5035 			 * ... else fall through to the normal
5036 			 * consecutive-after-RX assignment scheme.
5037 			 *
5038 			 * Note that we are assuming that all RX queue CPUs
5039 			 * have an L2 neighbor, or all do not.  If a mixed
5040 			 * scenario is possible, we will have to keep track
5041 			 * separately of how many queues prior to this one
5042 			 * were not able to be assigned to an L2 neighbor.
5043 			 */
5044 		}
5045 		if (is_tx)
5046 			core_index = scctx->isc_nrxqsets + qid;
5047 		else
5048 			core_index = qid;
5049 	} else {
5050 		core_index = qid;
5051 	}
5052 
5053 	return (cpuid_advance(ctx, base_cpuid, core_index));
5054 }
5055 
5056 static uint16_t
5057 get_ctx_core_offset(if_ctx_t ctx)
5058 {
5059 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5060 	struct cpu_offset *op;
5061 	cpuset_t assigned_cpus;
5062 	unsigned int cores_consumed;
5063 	unsigned int base_cpuid = ctx->ifc_sysctl_core_offset;
5064 	unsigned int first_valid;
5065 	unsigned int last_valid;
5066 	unsigned int i;
5067 
5068 	first_valid = CPU_FFS(&ctx->ifc_cpus) - 1;
5069 	last_valid = CPU_FLS(&ctx->ifc_cpus) - 1;
5070 
5071 	if (base_cpuid != CORE_OFFSET_UNSPECIFIED) {
5072 		/*
5073 		 * Align the user-chosen base CPU ID to the next valid CPU
5074 		 * for this device.  If the chosen base CPU ID is smaller
5075 		 * than the first valid CPU or larger than the last valid
5076 		 * CPU, we assume the user does not know what the valid
5077 		 * range is for this device and is thinking in terms of a
5078 		 * zero-based reference frame, and so we shift the given
5079 		 * value into the valid range (and wrap accordingly) so the
5080 		 * intent is translated to the proper frame of reference.
5081 		 * If the base CPU ID is within the valid first/last, but
5082 		 * does not correspond to a valid CPU, it is advanced to the
5083 		 * next valid CPU (wrapping if necessary).
5084 		 */
5085 		if (base_cpuid < first_valid || base_cpuid > last_valid) {
5086 			/* shift from zero-based to first_valid-based */
5087 			base_cpuid += first_valid;
5088 			/* wrap to range [first_valid, last_valid] */
5089 			base_cpuid = (base_cpuid - first_valid) %
5090 			    (last_valid - first_valid + 1);
5091 		}
5092 		if (!CPU_ISSET(base_cpuid, &ctx->ifc_cpus)) {
5093 			/*
5094 			 * base_cpuid is in [first_valid, last_valid], but
5095 			 * not a member of the valid set.  In this case,
5096 			 * there will always be a member of the valid set
5097 			 * with a CPU ID that is greater than base_cpuid,
5098 			 * and we simply advance to it.
5099 			 */
5100 			while (!CPU_ISSET(base_cpuid, &ctx->ifc_cpus))
5101 				base_cpuid++;
5102 		}
5103 		return (base_cpuid);
5104 	}
5105 
5106 	/*
5107 	 * Determine how many cores will be consumed by performing the CPU
5108 	 * assignments and counting how many of the assigned CPUs correspond
5109 	 * to CPUs in the set of context CPUs.  This is done using the CPU
5110 	 * ID first_valid as the base CPU ID, as the base CPU must be within
5111 	 * the set of context CPUs.
5112 	 *
5113 	 * Note not all assigned CPUs will be in the set of context CPUs
5114 	 * when separate CPUs are being allocated to TX and RX queues,
5115 	 * assignment to logical cores has been enabled, the set of context
5116 	 * CPUs contains only physical CPUs, and TX queues are mapped to L2
5117 	 * neighbors of CPUs that RX queues have been mapped to - in this
5118 	 * case we do only want to count how many CPUs in the set of context
5119 	 * CPUs have been consumed, as that determines the next CPU in that
5120 	 * set to start allocating at for the next device for which
5121 	 * core_offset is not set.
5122 	 */
5123 	CPU_ZERO(&assigned_cpus);
5124 	for (i = 0; i < scctx->isc_ntxqsets; i++)
5125 		CPU_SET(get_cpuid_for_queue(ctx, first_valid, i, true),
5126 		    &assigned_cpus);
5127 	for (i = 0; i < scctx->isc_nrxqsets; i++)
5128 		CPU_SET(get_cpuid_for_queue(ctx, first_valid, i, false),
5129 		    &assigned_cpus);
5130 	CPU_AND(&assigned_cpus, &assigned_cpus, &ctx->ifc_cpus);
5131 	cores_consumed = CPU_COUNT(&assigned_cpus);
5132 
5133 	mtx_lock(&cpu_offset_mtx);
5134 	SLIST_FOREACH(op, &cpu_offsets, entries) {
5135 		if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) {
5136 			base_cpuid = op->next_cpuid;
5137 			op->next_cpuid = cpuid_advance(ctx, op->next_cpuid,
5138 			    cores_consumed);
5139 			MPASS(op->refcount < UINT_MAX);
5140 			op->refcount++;
5141 			break;
5142 		}
5143 	}
5144 	if (base_cpuid == CORE_OFFSET_UNSPECIFIED) {
5145 		base_cpuid = first_valid;
5146 		op = malloc(sizeof(struct cpu_offset), M_IFLIB,
5147 		    M_NOWAIT | M_ZERO);
5148 		if (op == NULL) {
5149 			device_printf(ctx->ifc_dev,
5150 			    "allocation for cpu offset failed.\n");
5151 		} else {
5152 			op->next_cpuid = cpuid_advance(ctx, base_cpuid,
5153 			    cores_consumed);
5154 			op->refcount = 1;
5155 			CPU_COPY(&ctx->ifc_cpus, &op->set);
5156 			SLIST_INSERT_HEAD(&cpu_offsets, op, entries);
5157 		}
5158 	}
5159 	mtx_unlock(&cpu_offset_mtx);
5160 
5161 	return (base_cpuid);
5162 }
5163 
5164 static void
5165 unref_ctx_core_offset(if_ctx_t ctx)
5166 {
5167 	struct cpu_offset *op, *top;
5168 
5169 	mtx_lock(&cpu_offset_mtx);
5170 	SLIST_FOREACH_SAFE(op, &cpu_offsets, entries, top) {
5171 		if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) {
5172 			MPASS(op->refcount > 0);
5173 			op->refcount--;
5174 			if (op->refcount == 0) {
5175 				SLIST_REMOVE(&cpu_offsets, op, cpu_offset, entries);
5176 				free(op, M_IFLIB);
5177 			}
5178 			break;
5179 		}
5180 	}
5181 	mtx_unlock(&cpu_offset_mtx);
5182 }
5183 
5184 int
5185 iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp)
5186 {
5187 	if_ctx_t ctx;
5188 	if_t ifp;
5189 	if_softc_ctx_t scctx;
5190 	kobjop_desc_t kobj_desc;
5191 	kobj_method_t *kobj_method;
5192 	int err, msix, rid;
5193 	int num_txd, num_rxd;
5194 	char namebuf[TASKQUEUE_NAMELEN];
5195 
5196 	ctx = malloc(sizeof(*ctx), M_IFLIB, M_WAITOK | M_ZERO);
5197 
5198 	if (sc == NULL) {
5199 		sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK | M_ZERO);
5200 		device_set_softc(dev, ctx);
5201 		ctx->ifc_flags |= IFC_SC_ALLOCATED;
5202 	}
5203 
5204 	ctx->ifc_sctx = sctx;
5205 	ctx->ifc_dev = dev;
5206 	ctx->ifc_softc = sc;
5207 
5208 	iflib_register(ctx);
5209 	iflib_add_device_sysctl_pre(ctx);
5210 
5211 	scctx = &ctx->ifc_softc_ctx;
5212 	ifp = ctx->ifc_ifp;
5213 	if (ctx->ifc_sysctl_simple_tx) {
5214 #ifndef ALTQ
5215 		if_settransmitfn(ifp, iflib_simple_transmit);
5216 		device_printf(dev, "using simple if_transmit\n");
5217 #else
5218 		device_printf(dev, "ALTQ prevents using simple if_transmit\n");
5219 #endif
5220 	}
5221 	iflib_reset_qvalues(ctx);
5222 	IFNET_WLOCK();
5223 	CTX_LOCK(ctx);
5224 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
5225 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
5226 		goto fail_unlock;
5227 	}
5228 	_iflib_pre_assert(scctx);
5229 	ctx->ifc_txrx = *scctx->isc_txrx;
5230 
5231 	MPASS(scctx->isc_dma_width <= flsll(BUS_SPACE_MAXADDR));
5232 
5233 	if (sctx->isc_flags & IFLIB_DRIVER_MEDIA)
5234 		ctx->ifc_mediap = scctx->isc_media;
5235 
5236 #ifdef INVARIANTS
5237 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
5238 		MPASS(scctx->isc_tx_csum_flags);
5239 #endif
5240 
5241 	if_setcapabilities(ifp,
5242 	    scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_MEXTPG);
5243 	if_setcapenable(ifp,
5244 	    scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_MEXTPG);
5245 
5246 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
5247 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
5248 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
5249 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
5250 
5251 	num_txd = iflib_num_tx_descs(ctx);
5252 	num_rxd = iflib_num_rx_descs(ctx);
5253 
5254 	/* XXX change for per-queue sizes */
5255 	device_printf(dev, "Using %d TX descriptors and %d RX descriptors\n",
5256 	    num_txd, num_rxd);
5257 
5258 	if (scctx->isc_tx_nsegments > num_txd / MAX_SINGLE_PACKET_FRACTION)
5259 		scctx->isc_tx_nsegments = max(1, num_txd /
5260 		    MAX_SINGLE_PACKET_FRACTION);
5261 	if (scctx->isc_tx_tso_segments_max > num_txd /
5262 	    MAX_SINGLE_PACKET_FRACTION)
5263 		scctx->isc_tx_tso_segments_max = max(1,
5264 		    num_txd / MAX_SINGLE_PACKET_FRACTION);
5265 
5266 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
5267 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
5268 		/*
5269 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
5270 		 * but some MACs do.
5271 		 */
5272 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
5273 		    IP_MAXPACKET));
5274 		/*
5275 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
5276 		 * into account.  In the worst case, each of these calls will
5277 		 * add another mbuf and, thus, the requirement for another DMA
5278 		 * segment.  So for best performance, it doesn't make sense to
5279 		 * advertize a maximum of TSO segments that typically will
5280 		 * require defragmentation in iflib_encap().
5281 		 */
5282 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
5283 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
5284 	}
5285 	if (scctx->isc_rss_table_size == 0)
5286 		scctx->isc_rss_table_size = 64;
5287 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size - 1;
5288 
5289 	/* Create and start admin taskqueue */
5290 	snprintf(namebuf, TASKQUEUE_NAMELEN, "if_%s_tq", device_get_nameunit(dev));
5291 	ctx->ifc_tq = taskqueue_create_fast(namebuf, M_NOWAIT,
5292 	    taskqueue_thread_enqueue, &ctx->ifc_tq);
5293 	if (ctx->ifc_tq == NULL) {
5294 		device_printf(dev, "Unable to create admin taskqueue\n");
5295 		return (ENOMEM);
5296 	}
5297 
5298 	err = taskqueue_start_threads(&ctx->ifc_tq, 1, PI_NET, "%s", namebuf);
5299 	if (err) {
5300 		device_printf(dev,
5301 		    "Unable to start admin taskqueue threads error: %d\n",
5302 		    err);
5303 		taskqueue_free(ctx->ifc_tq);
5304 		return (err);
5305 	}
5306 
5307 	TASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
5308 	TASK_INIT(&ctx->ifc_led_task, 0, _task_fn_led, ctx);
5309 
5310 	/* Set up cpu set.  If it fails, use the set of all CPUs. */
5311 	if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) != 0) {
5312 		device_printf(dev, "Unable to fetch CPU list\n");
5313 		CPU_COPY(&all_cpus, &ctx->ifc_cpus);
5314 		ctx->ifc_cpus_are_physical_cores = false;
5315 	} else
5316 		ctx->ifc_cpus_are_physical_cores = true;
5317 	MPASS(CPU_COUNT(&ctx->ifc_cpus) > 0);
5318 
5319 	/*
5320 	 * Now set up MSI or MSI-X, should return us the number of supported
5321 	 * vectors (will be 1 for a legacy interrupt and MSI).
5322 	 */
5323 	if (sctx->isc_flags & IFLIB_SKIP_MSIX) {
5324 		msix = scctx->isc_vectors;
5325 	} else if (scctx->isc_msix_bar != 0)
5326 		/*
5327 		 * The simple fact that isc_msix_bar is not 0 does not mean we
5328 		 * we have a good value there that is known to work.
5329 		 */
5330 		msix = iflib_msix_init(ctx);
5331 	else {
5332 		scctx->isc_vectors = 1;
5333 		scctx->isc_ntxqsets = 1;
5334 		scctx->isc_nrxqsets = 1;
5335 		scctx->isc_intr = IFLIB_INTR_LEGACY;
5336 		msix = 0;
5337 	}
5338 	/* Get memory for the station queues */
5339 	if ((err = iflib_queues_alloc(ctx))) {
5340 		device_printf(dev, "Unable to allocate queue memory\n");
5341 		goto fail_intr_free;
5342 	}
5343 
5344 	if ((err = iflib_qset_structures_setup(ctx)))
5345 		goto fail_queues;
5346 
5347 	/*
5348 	 * Now that we know how many queues there are, get the core offset.
5349 	 */
5350 	ctx->ifc_sysctl_core_offset = get_ctx_core_offset(ctx);
5351 
5352 	if (msix > 1) {
5353 		/*
5354 		 * When using MSI-X, ensure that ifdi_{r,t}x_queue_intr_enable
5355 		 * aren't the default NULL implementation.
5356 		 */
5357 		kobj_desc = &ifdi_rx_queue_intr_enable_desc;
5358 		kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL,
5359 		    kobj_desc);
5360 		if (kobj_method == &kobj_desc->deflt) {
5361 			device_printf(dev,
5362 			    "MSI-X requires ifdi_rx_queue_intr_enable method");
5363 			err = EOPNOTSUPP;
5364 			goto fail_queues;
5365 		}
5366 		kobj_desc = &ifdi_tx_queue_intr_enable_desc;
5367 		kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL,
5368 		    kobj_desc);
5369 		if (kobj_method == &kobj_desc->deflt) {
5370 			device_printf(dev,
5371 			    "MSI-X requires ifdi_tx_queue_intr_enable method");
5372 			err = EOPNOTSUPP;
5373 			goto fail_queues;
5374 		}
5375 
5376 		/*
5377 		 * Assign the MSI-X vectors.
5378 		 * Note that the default NULL ifdi_msix_intr_assign method will
5379 		 * fail here, too.
5380 		 */
5381 		err = IFDI_MSIX_INTR_ASSIGN(ctx, msix);
5382 		if (err != 0) {
5383 			device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n",
5384 			    err);
5385 			goto fail_queues;
5386 		}
5387 	} else if (scctx->isc_intr != IFLIB_INTR_MSIX) {
5388 		rid = 0;
5389 		if (scctx->isc_intr == IFLIB_INTR_MSI) {
5390 			MPASS(msix == 1);
5391 			rid = 1;
5392 		}
5393 		if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) {
5394 			device_printf(dev, "iflib_legacy_setup failed %d\n", err);
5395 			goto fail_queues;
5396 		}
5397 	} else {
5398 		device_printf(dev,
5399 		    "Cannot use iflib with only 1 MSI-X interrupt!\n");
5400 		err = ENODEV;
5401 		goto fail_queues;
5402 	}
5403 
5404 	/*
5405 	 * It prevents a double-locking panic with iflib_media_status when
5406 	 * the driver loads.
5407 	 */
5408 	CTX_UNLOCK(ctx);
5409 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
5410 	CTX_LOCK(ctx);
5411 
5412 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
5413 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
5414 		goto fail_detach;
5415 	}
5416 
5417 	/*
5418 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
5419 	 * This must appear after the call to ether_ifattach() because
5420 	 * ether_ifattach() sets if_hdrlen to the default value.
5421 	 */
5422 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
5423 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
5424 
5425 	if ((err = iflib_netmap_attach(ctx))) {
5426 		device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err);
5427 		goto fail_detach;
5428 	}
5429 	*ctxp = ctx;
5430 
5431 	DEBUGNET_SET(ctx->ifc_ifp, iflib);
5432 
5433 	iflib_add_device_sysctl_post(ctx);
5434 	iflib_add_pfil(ctx);
5435 	ctx->ifc_flags |= IFC_INIT_DONE;
5436 	CTX_UNLOCK(ctx);
5437 	IFNET_WUNLOCK();
5438 
5439 	/* Create led(4) devices if the driver defined the method */
5440 	kobj_desc = &ifdi_led_func_desc;
5441 	kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL,
5442 	    kobj_desc);
5443 	if (kobj_method != &kobj_desc->deflt && IFDI_LED_SUPPORTED(ctx))
5444 		iflib_led_create(ctx);
5445 
5446 	return (0);
5447 
5448 fail_detach:
5449 	CTX_UNLOCK(ctx);
5450 	taskqueue_drain(ctx->ifc_tq, &ctx->ifc_admin_task);
5451 	ether_ifdetach(ctx->ifc_ifp);
5452 	CTX_LOCK(ctx);
5453 fail_queues:
5454 	sysctl_ctx_free(&ctx->ifc_sysctl_ctx);
5455 	ctx->ifc_sysctl_node = NULL;
5456 	/*
5457 	 * Drain without holding CTX_LOCK so _task_fn_admin can run to
5458 	 * completion if it needs the context lock.  On fail_detach we already
5459 	 * drained above; a second drain is a no-op when the queue is empty.
5460 	 */
5461 	CTX_UNLOCK(ctx);
5462 	taskqueue_drain(ctx->ifc_tq, &ctx->ifc_admin_task);
5463 	CTX_LOCK(ctx);
5464 	iflib_tqg_detach(ctx);
5465 	iflib_tx_structures_free(ctx);
5466 	iflib_rx_structures_free(ctx);
5467 	/*
5468 	 * Match iflib_device_deregister: IFDI_DETACH before taskqueue_free.
5469 	 * Avoid IFNET_WLOCK across driver detach (LinuxKPI workqueue drain).
5470 	 */
5471 	IFNET_WUNLOCK();
5472 	IFDI_DETACH(ctx);
5473 	IFDI_QUEUES_FREE(ctx);
5474 	IFNET_WLOCK();
5475 	taskqueue_free(ctx->ifc_tq);
5476 fail_intr_free:
5477 	iflib_free_intr_mem(ctx);
5478 fail_unlock:
5479 	CTX_UNLOCK(ctx);
5480 	IFNET_WUNLOCK();
5481 	iflib_deregister(ctx);
5482 	device_set_softc(ctx->ifc_dev, NULL);
5483 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5484 		free(ctx->ifc_softc, M_IFLIB);
5485 	free(ctx, M_IFLIB);
5486 	return (err);
5487 }
5488 
5489 int
5490 iflib_device_attach(device_t dev)
5491 {
5492 	if_ctx_t ctx;
5493 	if_shared_ctx_t sctx;
5494 
5495 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
5496 		return (ENOTSUP);
5497 
5498 	pci_enable_busmaster(dev);
5499 
5500 	return (iflib_device_register(dev, NULL, sctx, &ctx));
5501 }
5502 
5503 int
5504 iflib_device_deregister(if_ctx_t ctx)
5505 {
5506 	if_t ifp = ctx->ifc_ifp;
5507 	device_t dev = ctx->ifc_dev;
5508 
5509 	sysctl_ctx_free(&ctx->ifc_sysctl_ctx);
5510 	ctx->ifc_sysctl_node = NULL;
5511 
5512 	/* Make sure VLANS are not using driver */
5513 	if (if_vlantrunkinuse(ifp)) {
5514 		device_printf(dev, "Vlan in use, detach first\n");
5515 		return (EBUSY);
5516 	}
5517 #ifdef PCI_IOV
5518 	if (!CTX_IS_VF(ctx) && pci_iov_detach(dev) != 0) {
5519 		device_printf(dev, "SR-IOV in use; detach first.\n");
5520 		return (EBUSY);
5521 	}
5522 #endif
5523 
5524 	STATE_LOCK(ctx);
5525 	ctx->ifc_flags |= IFC_IN_DETACH;
5526 	STATE_UNLOCK(ctx);
5527 
5528 	/* Unregister VLAN handlers before calling iflib_stop() */
5529 	iflib_unregister_vlan_handlers(ctx);
5530 
5531 	iflib_netmap_detach(ifp);
5532 	ether_ifdetach(ifp);
5533 
5534 	CTX_LOCK(ctx);
5535 	iflib_stop(ctx);
5536 	CTX_UNLOCK(ctx);
5537 
5538 	iflib_rem_pfil(ctx);
5539 	if (ctx->ifc_led_dev != NULL) {
5540 		led_destroy(ctx->ifc_led_dev);
5541 		taskqueue_drain(ctx->ifc_tq, &ctx->ifc_led_task);
5542 	}
5543 
5544 	iflib_tqg_detach(ctx);
5545 	iflib_tx_structures_free(ctx);
5546 	iflib_rx_structures_free(ctx);
5547 
5548 	CTX_LOCK(ctx);
5549 	IFDI_DETACH(ctx);
5550 	IFDI_QUEUES_FREE(ctx);
5551 	CTX_UNLOCK(ctx);
5552 
5553 	taskqueue_free(ctx->ifc_tq);
5554 	ctx->ifc_tq = NULL;
5555 
5556 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
5557 	iflib_free_intr_mem(ctx);
5558 
5559 	bus_generic_detach(dev);
5560 
5561 	iflib_deregister(ctx);
5562 
5563 	device_set_softc(ctx->ifc_dev, NULL);
5564 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5565 		free(ctx->ifc_softc, M_IFLIB);
5566 	unref_ctx_core_offset(ctx);
5567 	free(ctx, M_IFLIB);
5568 	return (0);
5569 }
5570 
5571 static void
5572 iflib_tqg_detach(if_ctx_t ctx)
5573 {
5574 	iflib_txq_t txq;
5575 	iflib_rxq_t rxq;
5576 	int i;
5577 	struct taskqgroup *tqg;
5578 
5579 	/* XXX drain any dependent tasks */
5580 	tqg = qgroup_if_io_tqg;
5581 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
5582 		callout_drain(&txq->ift_timer);
5583 #ifdef DEV_NETMAP
5584 		callout_drain(&txq->ift_netmap_timer);
5585 #endif /* DEV_NETMAP */
5586 		if (txq->ift_task.gt_uniq != NULL)
5587 			taskqgroup_detach(tqg, &txq->ift_task);
5588 	}
5589 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
5590 		if (rxq->ifr_task.gt_uniq != NULL)
5591 			taskqgroup_detach(tqg, &rxq->ifr_task);
5592 	}
5593 }
5594 
5595 static void
5596 iflib_free_intr_mem(if_ctx_t ctx)
5597 {
5598 
5599 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) {
5600 		iflib_irq_free(ctx, &ctx->ifc_legacy_irq);
5601 	}
5602 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) {
5603 		pci_release_msi(ctx->ifc_dev);
5604 	}
5605 	if (ctx->ifc_msix_mem != NULL) {
5606 		bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY,
5607 		    rman_get_rid(ctx->ifc_msix_mem), ctx->ifc_msix_mem);
5608 		ctx->ifc_msix_mem = NULL;
5609 	}
5610 }
5611 
5612 int
5613 iflib_device_detach(device_t dev)
5614 {
5615 	if_ctx_t ctx = device_get_softc(dev);
5616 
5617 	return (iflib_device_deregister(ctx));
5618 }
5619 
5620 int
5621 iflib_device_suspend(device_t dev)
5622 {
5623 	if_ctx_t ctx = device_get_softc(dev);
5624 
5625 	CTX_LOCK(ctx);
5626 	IFDI_SUSPEND(ctx);
5627 	CTX_UNLOCK(ctx);
5628 
5629 	return (bus_generic_suspend(dev));
5630 }
5631 int
5632 iflib_device_shutdown(device_t dev)
5633 {
5634 	if_ctx_t ctx = device_get_softc(dev);
5635 
5636 	CTX_LOCK(ctx);
5637 	IFDI_SHUTDOWN(ctx);
5638 	CTX_UNLOCK(ctx);
5639 
5640 	return (bus_generic_suspend(dev));
5641 }
5642 
5643 int
5644 iflib_device_resume(device_t dev)
5645 {
5646 	if_ctx_t ctx = device_get_softc(dev);
5647 	iflib_txq_t txq = ctx->ifc_txqs;
5648 
5649 	CTX_LOCK(ctx);
5650 	IFDI_RESUME(ctx);
5651 	iflib_if_init_locked(ctx);
5652 	CTX_UNLOCK(ctx);
5653 	for (int i = 0; i < NTXQSETS(ctx); i++, txq++)
5654 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
5655 
5656 	return (bus_generic_resume(dev));
5657 }
5658 
5659 int
5660 iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params)
5661 {
5662 	int error;
5663 	if_ctx_t ctx = device_get_softc(dev);
5664 
5665 	CTX_LOCK(ctx);
5666 	error = IFDI_IOV_INIT(ctx, num_vfs, params);
5667 	CTX_UNLOCK(ctx);
5668 
5669 	return (error);
5670 }
5671 
5672 int
5673 iflib_device_iov_init_restart(device_t dev, uint16_t num_vfs,
5674     const nvlist_t *params)
5675 {
5676 	if_ctx_t ctx;
5677 	if_t ifp;
5678 	bool restart, running;
5679 	int error;
5680 
5681 	ctx = device_get_softc(dev);
5682 	ifp = ctx->ifc_ifp;
5683 
5684 	CTX_LOCK(ctx);
5685 	/*
5686 	 * Drivers which change the PF queue layout need the complete iflib
5687 	 * stop/init sequence around their IOV callback when the interface is
5688 	 * active.  An administratively-down interface has no live queues to
5689 	 * quiesce, and must remain down after the new layout is installed.
5690 	 * Keep the transition within one context-lock critical section.
5691 	 */
5692 	restart = (if_getflags(ifp) & IFF_UP) != 0;
5693 	running = (if_getdrvflags(ifp) & IFF_DRV_RUNNING) != 0;
5694 	if (restart || running)
5695 		iflib_stop(ctx);
5696 	error = IFDI_IOV_INIT(ctx, num_vfs, params);
5697 	if (restart)
5698 		iflib_init_locked(ctx);
5699 	CTX_UNLOCK(ctx);
5700 	return (error);
5701 }
5702 
5703 void
5704 iflib_device_iov_uninit(device_t dev)
5705 {
5706 	if_ctx_t ctx = device_get_softc(dev);
5707 
5708 	CTX_LOCK(ctx);
5709 	IFDI_IOV_UNINIT(ctx);
5710 	CTX_UNLOCK(ctx);
5711 }
5712 
5713 void
5714 iflib_device_iov_uninit_restart(device_t dev)
5715 {
5716 	if_ctx_t ctx;
5717 	bool restart;
5718 
5719 	ctx = device_get_softc(dev);
5720 
5721 	CTX_LOCK(ctx);
5722 	/*
5723 	 * RUNNING can be clear while a watchdog reset is pending but the
5724 	 * hardware is still live.  Always stop before the driver changes its
5725 	 * queue layout, and use IFF_UP only to preserve administrative state.
5726 	 */
5727 	restart = (if_getflags(ctx->ifc_ifp) & IFF_UP) != 0;
5728 	iflib_stop(ctx);
5729 	IFDI_IOV_UNINIT(ctx);
5730 	if (restart)
5731 		iflib_init_locked(ctx);
5732 	CTX_UNLOCK(ctx);
5733 }
5734 
5735 int
5736 iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params)
5737 {
5738 	int error;
5739 	if_ctx_t ctx = device_get_softc(dev);
5740 
5741 	CTX_LOCK(ctx);
5742 	error = IFDI_IOV_VF_ADD(ctx, vfnum, params);
5743 	CTX_UNLOCK(ctx);
5744 
5745 	return (error);
5746 }
5747 
5748 /*********************************************************************
5749  *
5750  *  MODULE FUNCTION DEFINITIONS
5751  *
5752  **********************************************************************/
5753 
5754 /*
5755  * - Start a fast taskqueue thread for each core
5756  * - Start a taskqueue for control operations
5757  */
5758 static int
5759 iflib_module_init(void)
5760 {
5761 	iflib_timer_default = hz / 2;
5762 	return (0);
5763 }
5764 
5765 static int
5766 iflib_module_event_handler(module_t mod, int what, void *arg)
5767 {
5768 	int err;
5769 
5770 	switch (what) {
5771 	case MOD_LOAD:
5772 		if ((err = iflib_module_init()) != 0)
5773 			return (err);
5774 		break;
5775 	case MOD_UNLOAD:
5776 		return (EBUSY);
5777 	default:
5778 		return (EOPNOTSUPP);
5779 	}
5780 
5781 	return (0);
5782 }
5783 
5784 /*********************************************************************
5785  *
5786  *  PUBLIC FUNCTION DEFINITIONS
5787  *     ordered as in iflib.h
5788  *
5789  **********************************************************************/
5790 
5791 static void
5792 _iflib_assert(if_shared_ctx_t sctx)
5793 {
5794 	int i;
5795 
5796 	MPASS(sctx->isc_tx_maxsize);
5797 	MPASS(sctx->isc_tx_maxsegsize);
5798 
5799 	MPASS(sctx->isc_rx_maxsize);
5800 	MPASS(sctx->isc_rx_nsegments);
5801 	MPASS(sctx->isc_rx_maxsegsize);
5802 
5803 	MPASS(sctx->isc_nrxqs >= 1 && sctx->isc_nrxqs <= 8);
5804 	for (i = 0; i < sctx->isc_nrxqs; i++) {
5805 		MPASS(sctx->isc_nrxd_min[i]);
5806 		MPASS(powerof2(sctx->isc_nrxd_min[i]));
5807 		MPASS(sctx->isc_nrxd_max[i]);
5808 		MPASS(powerof2(sctx->isc_nrxd_max[i]));
5809 		MPASS(sctx->isc_nrxd_default[i]);
5810 		MPASS(powerof2(sctx->isc_nrxd_default[i]));
5811 	}
5812 
5813 	MPASS(sctx->isc_ntxqs >= 1 && sctx->isc_ntxqs <= 8);
5814 	for (i = 0; i < sctx->isc_ntxqs; i++) {
5815 		MPASS(sctx->isc_ntxd_min[i]);
5816 		MPASS(powerof2(sctx->isc_ntxd_min[i]));
5817 		MPASS(sctx->isc_ntxd_max[i]);
5818 		MPASS(powerof2(sctx->isc_ntxd_max[i]));
5819 		MPASS(sctx->isc_ntxd_default[i]);
5820 		MPASS(powerof2(sctx->isc_ntxd_default[i]));
5821 	}
5822 }
5823 
5824 static void
5825 _iflib_pre_assert(if_softc_ctx_t scctx)
5826 {
5827 
5828 	MPASS(scctx->isc_txrx->ift_txd_encap);
5829 	MPASS(scctx->isc_txrx->ift_txd_flush);
5830 	MPASS(scctx->isc_txrx->ift_txd_credits_update);
5831 	MPASS(scctx->isc_txrx->ift_rxd_available);
5832 	MPASS(scctx->isc_txrx->ift_rxd_pkt_get);
5833 	MPASS(scctx->isc_txrx->ift_rxd_refill);
5834 	MPASS(scctx->isc_txrx->ift_rxd_flush);
5835 }
5836 
5837 static void
5838 iflib_register(if_ctx_t ctx)
5839 {
5840 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5841 	driver_t *driver = sctx->isc_driver;
5842 	device_t dev = ctx->ifc_dev;
5843 	if_t ifp;
5844 
5845 	_iflib_assert(sctx);
5846 
5847 	CTX_LOCK_INIT(ctx);
5848 	STATE_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev));
5849 	ifp = ctx->ifc_ifp = if_alloc_dev(IFT_ETHER, dev);
5850 
5851 	/*
5852 	 * Initialize our context's device specific methods
5853 	 */
5854 	kobj_init((kobj_t) ctx, (kobj_class_t) driver);
5855 	kobj_class_compile((kobj_class_t) driver);
5856 
5857 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
5858 	if_setsoftc(ifp, ctx);
5859 	if_setdev(ifp, dev);
5860 	if_setinitfn(ifp, iflib_if_init);
5861 	if_setioctlfn(ifp, iflib_if_ioctl);
5862 #ifdef ALTQ
5863 	if_setstartfn(ifp, iflib_altq_if_start);
5864 	if_settransmitfn(ifp, iflib_altq_if_transmit);
5865 	if_setsendqready(ifp);
5866 #else
5867 	if_settransmitfn(ifp, iflib_if_transmit);
5868 #endif
5869 	if_setqflushfn(ifp, iflib_if_qflush);
5870 	if_setgetcounterfn(ifp, iflib_if_get_counter);
5871 	if_setflags(ifp, IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST);
5872 	ctx->ifc_vlan_attach_event =
5873 	    EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx,
5874 		    EVENTHANDLER_PRI_FIRST);
5875 	ctx->ifc_vlan_detach_event =
5876 	    EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx,
5877 		    EVENTHANDLER_PRI_FIRST);
5878 
5879 	if ((sctx->isc_flags & IFLIB_DRIVER_MEDIA) == 0) {
5880 		ctx->ifc_mediap = &ctx->ifc_media;
5881 		ifmedia_init(ctx->ifc_mediap, IFM_IMASK,
5882 		    iflib_media_change, iflib_media_status);
5883 	}
5884 }
5885 
5886 static void
5887 iflib_unregister_vlan_handlers(if_ctx_t ctx)
5888 {
5889 	/* Unregister VLAN events */
5890 	if (ctx->ifc_vlan_attach_event != NULL) {
5891 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
5892 		ctx->ifc_vlan_attach_event = NULL;
5893 	}
5894 	if (ctx->ifc_vlan_detach_event != NULL) {
5895 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
5896 		ctx->ifc_vlan_detach_event = NULL;
5897 	}
5898 
5899 }
5900 
5901 static void
5902 iflib_deregister(if_ctx_t ctx)
5903 {
5904 	if_t ifp = ctx->ifc_ifp;
5905 
5906 	/* Remove all media */
5907 	ifmedia_removeall(&ctx->ifc_media);
5908 
5909 	/* Ensure that VLAN event handlers are unregistered */
5910 	iflib_unregister_vlan_handlers(ctx);
5911 
5912 	/* Release kobject reference */
5913 	kobj_delete((kobj_t) ctx, NULL);
5914 
5915 	/* Free the ifnet structure */
5916 	if_free(ifp);
5917 
5918 	STATE_LOCK_DESTROY(ctx);
5919 
5920 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
5921 	CTX_LOCK_DESTROY(ctx);
5922 }
5923 
5924 static int
5925 iflib_queues_alloc(if_ctx_t ctx)
5926 {
5927 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5928 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5929 	device_t dev = ctx->ifc_dev;
5930 	int nrxqsets = scctx->isc_nrxqsets;
5931 	int ntxqsets = scctx->isc_ntxqsets;
5932 	iflib_txq_t txq;
5933 	iflib_rxq_t rxq;
5934 	iflib_fl_t fl = NULL;
5935 	int i, j, cpu, err, txconf, rxconf;
5936 	iflib_dma_info_t ifdip;
5937 	uint32_t *rxqsizes = scctx->isc_rxqsizes;
5938 	uint32_t *txqsizes = scctx->isc_txqsizes;
5939 	uint8_t nrxqs = sctx->isc_nrxqs;
5940 	uint8_t ntxqs = sctx->isc_ntxqs;
5941 	int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1;
5942 	int fl_offset = (sctx->isc_flags & IFLIB_HAS_RXCQ ? 1 : 0);
5943 	caddr_t *vaddrs;
5944 	uint64_t *paddrs;
5945 
5946 	KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1"));
5947 	KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1"));
5948 	KASSERT(nrxqs >= fl_offset + nfree_lists,
5949 	    ("there must be at least a rxq for each free list"));
5950 
5951 	/* Allocate the TX ring struct memory */
5952 	if (!(ctx->ifc_txqs =
5953 	    (iflib_txq_t) malloc(sizeof(struct iflib_txq) *
5954 		    ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5955 		device_printf(dev, "Unable to allocate TX ring memory\n");
5956 		err = ENOMEM;
5957 		goto fail;
5958 	}
5959 
5960 	/* Now allocate the RX */
5961 	if (!(ctx->ifc_rxqs =
5962 	    (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) *
5963 		    nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5964 		device_printf(dev, "Unable to allocate RX ring memory\n");
5965 		err = ENOMEM;
5966 		goto rx_fail;
5967 	}
5968 
5969 	txq = ctx->ifc_txqs;
5970 	rxq = ctx->ifc_rxqs;
5971 
5972 	/*
5973 	 * XXX handle allocation failure
5974 	 */
5975 	for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) {
5976 		/* Set up some basics */
5977 
5978 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs,
5979 		    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
5980 			device_printf(dev,
5981 			    "Unable to allocate TX DMA info memory\n");
5982 			err = ENOMEM;
5983 			goto err_tx_desc;
5984 		}
5985 		txq->ift_ifdi = ifdip;
5986 		for (j = 0; j < ntxqs; j++, ifdip++) {
5987 			if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, 0)) {
5988 				device_printf(dev,
5989 				    "Unable to allocate TX descriptors\n");
5990 				err = ENOMEM;
5991 				goto err_tx_desc;
5992 			}
5993 			txq->ift_txd_size[j] = scctx->isc_txd_size[j];
5994 			bzero((void *)ifdip->idi_vaddr, txqsizes[j]);
5995 		}
5996 		txq->ift_ctx = ctx;
5997 		txq->ift_id = i;
5998 		if (sctx->isc_flags & IFLIB_HAS_TXCQ) {
5999 			txq->ift_br_offset = 1;
6000 		} else {
6001 			txq->ift_br_offset = 0;
6002 		}
6003 
6004 		if (iflib_txsd_alloc(txq)) {
6005 			device_printf(dev, "Critical Failure setting up TX buffers\n");
6006 			err = ENOMEM;
6007 			goto err_tx_desc;
6008 		}
6009 
6010 		/* Initialize the TX lock */
6011 		snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:TX(%d):callout",
6012 		    device_get_nameunit(dev), txq->ift_id);
6013 		mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF);
6014 		callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0);
6015 		txq->ift_timer.c_cpu = cpu;
6016 #ifdef DEV_NETMAP
6017 		callout_init_mtx(&txq->ift_netmap_timer, &txq->ift_mtx, 0);
6018 		txq->ift_netmap_timer.c_cpu = cpu;
6019 #endif /* DEV_NETMAP */
6020 
6021 		err = ifmp_ring_alloc(&txq->ift_br, 2048, txq, iflib_txq_drain,
6022 		    iflib_txq_can_drain, M_IFLIB, M_WAITOK);
6023 		if (err) {
6024 			/* XXX free any allocated rings */
6025 			device_printf(dev, "Unable to allocate buf_ring\n");
6026 			goto err_tx_desc;
6027 		}
6028 		txq->ift_reclaim_thresh = ctx->ifc_sysctl_tx_reclaim_thresh;
6029 	}
6030 
6031 	for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) {
6032 		/* Set up some basics */
6033 		callout_init(&rxq->ifr_watchdog, 1);
6034 
6035 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs,
6036 		    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
6037 			device_printf(dev,
6038 			    "Unable to allocate RX DMA info memory\n");
6039 			err = ENOMEM;
6040 			goto err_tx_desc;
6041 		}
6042 
6043 		rxq->ifr_ifdi = ifdip;
6044 		/* XXX this needs to be changed if #rx queues != #tx queues */
6045 		rxq->ifr_ntxqirq = 1;
6046 		rxq->ifr_txqid[0] = i;
6047 		for (j = 0; j < nrxqs; j++, ifdip++) {
6048 			if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, 0)) {
6049 				device_printf(dev,
6050 				    "Unable to allocate RX descriptors\n");
6051 				err = ENOMEM;
6052 				goto err_tx_desc;
6053 			}
6054 			bzero((void *)ifdip->idi_vaddr, rxqsizes[j]);
6055 		}
6056 		rxq->ifr_ctx = ctx;
6057 		rxq->ifr_id = i;
6058 		rxq->ifr_fl_offset = fl_offset;
6059 		rxq->ifr_nfl = nfree_lists;
6060 		if (!(fl =
6061 		    (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) {
6062 			device_printf(dev, "Unable to allocate free list memory\n");
6063 			err = ENOMEM;
6064 			goto err_tx_desc;
6065 		}
6066 		rxq->ifr_fl = fl;
6067 		for (j = 0; j < nfree_lists; j++) {
6068 			fl[j].ifl_rxq = rxq;
6069 			fl[j].ifl_id = j;
6070 			fl[j].ifl_ifdi = &rxq->ifr_ifdi[j + rxq->ifr_fl_offset];
6071 			fl[j].ifl_rxd_size = scctx->isc_rxd_size[j];
6072 		}
6073 		/* Allocate receive buffers for the ring */
6074 		if (iflib_rxsd_alloc(rxq)) {
6075 			device_printf(dev,
6076 			    "Critical Failure setting up receive buffers\n");
6077 			err = ENOMEM;
6078 			goto err_rx_desc;
6079 		}
6080 
6081 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
6082 			fl->ifl_rx_bitmap = bit_alloc(fl->ifl_size, M_IFLIB,
6083 			    M_WAITOK);
6084 	}
6085 
6086 	/* TXQs */
6087 	vaddrs = malloc(sizeof(caddr_t)  * ntxqsets * ntxqs, M_IFLIB, M_WAITOK);
6088 	paddrs = malloc(sizeof(uint64_t) * ntxqsets * ntxqs, M_IFLIB, M_WAITOK);
6089 	for (i = 0; i < ntxqsets; i++) {
6090 		iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi;
6091 
6092 		for (j = 0; j < ntxqs; j++, di++) {
6093 			vaddrs[i * ntxqs + j] = di->idi_vaddr;
6094 			paddrs[i * ntxqs + j] = di->idi_paddr;
6095 		}
6096 	}
6097 	if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) {
6098 		device_printf(ctx->ifc_dev,
6099 		    "Unable to allocate device TX queue\n");
6100 		iflib_tx_structures_free(ctx);
6101 		free(vaddrs, M_IFLIB);
6102 		free(paddrs, M_IFLIB);
6103 		goto err_rx_desc;
6104 	}
6105 	free(vaddrs, M_IFLIB);
6106 	free(paddrs, M_IFLIB);
6107 
6108 	/* RXQs */
6109 	vaddrs = malloc(sizeof(caddr_t)  * nrxqsets * nrxqs, M_IFLIB, M_WAITOK);
6110 	paddrs = malloc(sizeof(uint64_t) * nrxqsets * nrxqs, M_IFLIB, M_WAITOK);
6111 	for (i = 0; i < nrxqsets; i++) {
6112 		iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi;
6113 
6114 		for (j = 0; j < nrxqs; j++, di++) {
6115 			vaddrs[i * nrxqs + j] = di->idi_vaddr;
6116 			paddrs[i * nrxqs + j] = di->idi_paddr;
6117 		}
6118 	}
6119 	if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) {
6120 		device_printf(ctx->ifc_dev,
6121 		    "Unable to allocate device RX queue\n");
6122 		iflib_tx_structures_free(ctx);
6123 		free(vaddrs, M_IFLIB);
6124 		free(paddrs, M_IFLIB);
6125 		goto err_rx_desc;
6126 	}
6127 	free(vaddrs, M_IFLIB);
6128 	free(paddrs, M_IFLIB);
6129 
6130 	return (0);
6131 
6132 /* XXX handle allocation failure changes */
6133 err_rx_desc:
6134 err_tx_desc:
6135 rx_fail:
6136 	if (ctx->ifc_rxqs != NULL)
6137 		free(ctx->ifc_rxqs, M_IFLIB);
6138 	ctx->ifc_rxqs = NULL;
6139 	if (ctx->ifc_txqs != NULL)
6140 		free(ctx->ifc_txqs, M_IFLIB);
6141 	ctx->ifc_txqs = NULL;
6142 fail:
6143 	return (err);
6144 }
6145 
6146 static int
6147 iflib_tx_structures_setup(if_ctx_t ctx)
6148 {
6149 	iflib_txq_t txq = ctx->ifc_txqs;
6150 	int i;
6151 
6152 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
6153 		iflib_txq_setup(txq);
6154 
6155 	return (0);
6156 }
6157 
6158 static void
6159 iflib_tx_structures_free(if_ctx_t ctx)
6160 {
6161 	iflib_txq_t txq = ctx->ifc_txqs;
6162 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6163 	int i, j;
6164 
6165 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
6166 		for (j = 0; j < sctx->isc_ntxqs; j++)
6167 			iflib_dma_free(&txq->ift_ifdi[j]);
6168 		iflib_txq_destroy(txq);
6169 	}
6170 	free(ctx->ifc_txqs, M_IFLIB);
6171 	ctx->ifc_txqs = NULL;
6172 }
6173 
6174 /*********************************************************************
6175  *
6176  *  Initialize all receive rings.
6177  *
6178  **********************************************************************/
6179 static int
6180 iflib_rx_structures_setup(if_ctx_t ctx)
6181 {
6182 	iflib_rxq_t rxq = ctx->ifc_rxqs;
6183 	int q;
6184 #if defined(INET6) || defined(INET)
6185 	int err, i;
6186 #endif
6187 
6188 	for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) {
6189 #if defined(INET6) || defined(INET)
6190 		err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp,
6191 		    TCP_LRO_ENTRIES, min(1024,
6192 		    ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset]));
6193 		if (err != 0) {
6194 			device_printf(ctx->ifc_dev,
6195 			    "LRO Initialization failed!\n");
6196 			goto fail;
6197 		}
6198 #endif
6199 		IFDI_RXQ_SETUP(ctx, rxq->ifr_id);
6200 	}
6201 	return (0);
6202 #if defined(INET6) || defined(INET)
6203 fail:
6204 	/*
6205 	 * Free LRO resources allocated so far, we will only handle
6206 	 * the rings that completed, the failing case will have
6207 	 * cleaned up for itself.  'q' failed, so its the terminus.
6208 	 */
6209 	rxq = ctx->ifc_rxqs;
6210 	for (i = 0; i < q; ++i, rxq++) {
6211 		tcp_lro_free(&rxq->ifr_lc);
6212 	}
6213 	return (err);
6214 #endif
6215 }
6216 
6217 /*********************************************************************
6218  *
6219  *  Free all receive rings.
6220  *
6221  **********************************************************************/
6222 static void
6223 iflib_rx_structures_free(if_ctx_t ctx)
6224 {
6225 	iflib_rxq_t rxq = ctx->ifc_rxqs;
6226 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6227 	int i, j;
6228 
6229 	for (i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) {
6230 		for (j = 0; j < sctx->isc_nrxqs; j++)
6231 			iflib_dma_free(&rxq->ifr_ifdi[j]);
6232 		iflib_rx_sds_free(rxq);
6233 #if defined(INET6) || defined(INET)
6234 		tcp_lro_free(&rxq->ifr_lc);
6235 #endif
6236 	}
6237 	free(ctx->ifc_rxqs, M_IFLIB);
6238 	ctx->ifc_rxqs = NULL;
6239 }
6240 
6241 static int
6242 iflib_qset_structures_setup(if_ctx_t ctx)
6243 {
6244 	int err;
6245 
6246 	/*
6247 	 * It is expected that the caller takes care of freeing queues if this
6248 	 * fails.
6249 	 */
6250 	if ((err = iflib_tx_structures_setup(ctx)) != 0) {
6251 		device_printf(ctx->ifc_dev, "iflib_tx_structures_setup failed: %d\n", err);
6252 		return (err);
6253 	}
6254 
6255 	if ((err = iflib_rx_structures_setup(ctx)) != 0)
6256 		device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err);
6257 
6258 	return (err);
6259 }
6260 
6261 int
6262 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
6263 		driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, const char *name)
6264 {
6265 
6266 	return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name));
6267 }
6268 
6269 /* Just to avoid copy/paste */
6270 static inline int
6271 iflib_irq_set_affinity(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type,
6272     int qid, struct grouptask *gtask, struct taskqgroup *tqg, void *uniq,
6273     const char *name)
6274 {
6275 	device_t dev;
6276 	unsigned int base_cpuid, cpuid;
6277 	int err;
6278 
6279 	dev = ctx->ifc_dev;
6280 	base_cpuid = ctx->ifc_sysctl_core_offset;
6281 	cpuid = get_cpuid_for_queue(ctx, base_cpuid, qid, type == IFLIB_INTR_TX);
6282 	err = taskqgroup_attach_cpu(tqg, gtask, uniq, cpuid, dev,
6283 	    irq ? irq->ii_res : NULL, name);
6284 	if (err) {
6285 		device_printf(dev, "taskqgroup_attach_cpu failed %d\n", err);
6286 		return (err);
6287 	}
6288 #ifdef notyet
6289 	if (cpuid > ctx->ifc_cpuid_highest)
6290 		ctx->ifc_cpuid_highest = cpuid;
6291 #endif
6292 	return (0);
6293 }
6294 
6295 /*
6296  * Allocate a hardware interrupt for subctx using the parent (ctx)'s hardware
6297  * resources.
6298  *
6299  * Similar to iflib_irq_alloc_generic(), but for interrupt type IFLIB_INTR_RXTX
6300  * only.
6301  *
6302  * XXX: Could be removed if subctx's dev has its intr resource allocation
6303  * methods replaced with custom ones?
6304  */
6305 int
6306 iflib_irq_alloc_generic_subctx(if_ctx_t ctx, if_ctx_t subctx, if_irq_t irq,
6307 			       int rid, iflib_intr_type_t type,
6308 			       driver_filter_t *filter, void *filter_arg,
6309 			       int qid, const char *name)
6310 {
6311 	device_t dev, subdev;
6312 	struct grouptask *gtask;
6313 	struct taskqgroup *tqg;
6314 	iflib_filter_info_t info;
6315 	gtask_fn_t *fn;
6316 	int tqrid, err;
6317 	driver_filter_t *intr_fast;
6318 	void *q;
6319 
6320 	MPASS(ctx != NULL);
6321 	MPASS(subctx != NULL);
6322 
6323 	tqrid = rid;
6324 	dev = ctx->ifc_dev;
6325 	subdev = subctx->ifc_dev;
6326 
6327 	switch (type) {
6328 	case IFLIB_INTR_RXTX:
6329 		q = &subctx->ifc_rxqs[qid];
6330 		info = &subctx->ifc_rxqs[qid].ifr_filter_info;
6331 		gtask = &subctx->ifc_rxqs[qid].ifr_task;
6332 		tqg = qgroup_if_io_tqg;
6333 		fn = _task_fn_rx;
6334 		intr_fast = iflib_fast_intr_rxtx;
6335 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6336 		break;
6337 	default:
6338 		device_printf(dev, "%s: unknown net intr type for subctx %s (%d)\n",
6339 		    __func__, device_get_nameunit(subdev), type);
6340 		return (EINVAL);
6341 	}
6342 
6343 	info->ifi_filter = filter;
6344 	info->ifi_filter_arg = filter_arg;
6345 	info->ifi_task = gtask;
6346 	info->ifi_ctx = q;
6347 
6348 	NET_GROUPTASK_INIT(gtask, 0, fn, q);
6349 
6350 	/* Allocate interrupts from hardware using parent context */
6351 	err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info, name);
6352 	if (err != 0) {
6353 		device_printf(dev, "_iflib_irq_alloc failed for subctx %s: %d\n",
6354 		    device_get_nameunit(subdev), err);
6355 		return (err);
6356 	}
6357 
6358 	if (tqrid != -1) {
6359 		err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg, q,
6360 		    name);
6361 		if (err)
6362 			return (err);
6363 	} else {
6364 		taskqgroup_attach(tqg, gtask, q, dev, irq->ii_res, name);
6365 	}
6366 
6367 	return (0);
6368 }
6369 
6370 int
6371 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid,
6372 			iflib_intr_type_t type, driver_filter_t *filter,
6373 			void *filter_arg, int qid, const char *name)
6374 {
6375 	device_t dev;
6376 	struct grouptask *gtask;
6377 	struct taskqgroup *tqg;
6378 	iflib_filter_info_t info;
6379 	gtask_fn_t *fn;
6380 	int tqrid, err;
6381 	driver_filter_t *intr_fast;
6382 	void *q;
6383 
6384 	info = &ctx->ifc_filter_info;
6385 	tqrid = rid;
6386 
6387 	switch (type) {
6388 	/* XXX merge tx/rx for netmap? */
6389 	case IFLIB_INTR_TX:
6390 		q = &ctx->ifc_txqs[qid];
6391 		info = &ctx->ifc_txqs[qid].ift_filter_info;
6392 		gtask = &ctx->ifc_txqs[qid].ift_task;
6393 		tqg = qgroup_if_io_tqg;
6394 		fn = _task_fn_tx;
6395 		intr_fast = iflib_fast_intr;
6396 		GROUPTASK_INIT(gtask, 0, fn, q);
6397 		ctx->ifc_flags |= IFC_NETMAP_TX_IRQ;
6398 		break;
6399 	case IFLIB_INTR_RX:
6400 		q = &ctx->ifc_rxqs[qid];
6401 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
6402 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6403 		tqg = qgroup_if_io_tqg;
6404 		fn = _task_fn_rx;
6405 		intr_fast = iflib_fast_intr;
6406 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6407 		break;
6408 	case IFLIB_INTR_RXTX:
6409 		q = &ctx->ifc_rxqs[qid];
6410 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
6411 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6412 		tqg = qgroup_if_io_tqg;
6413 		fn = _task_fn_rx;
6414 		intr_fast = iflib_fast_intr_rxtx;
6415 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6416 		break;
6417 	case IFLIB_INTR_ADMIN:
6418 		q = ctx;
6419 		tqrid = -1;
6420 		info = &ctx->ifc_filter_info;
6421 		gtask = NULL;
6422 		intr_fast = iflib_fast_intr_ctx;
6423 		break;
6424 	default:
6425 		device_printf(ctx->ifc_dev, "%s: unknown net intr type\n",
6426 		    __func__);
6427 		return (EINVAL);
6428 	}
6429 
6430 	info->ifi_filter = filter;
6431 	info->ifi_filter_arg = filter_arg;
6432 	info->ifi_task = gtask;
6433 	info->ifi_ctx = q;
6434 
6435 	dev = ctx->ifc_dev;
6436 	err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info,  name);
6437 	if (err != 0) {
6438 		device_printf(dev, "_iflib_irq_alloc failed %d\n", err);
6439 		return (err);
6440 	}
6441 	if (type == IFLIB_INTR_ADMIN)
6442 		return (0);
6443 
6444 	if (tqrid != -1) {
6445 		err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg, q,
6446 		    name);
6447 		if (err)
6448 			return (err);
6449 	} else {
6450 		taskqgroup_attach(tqg, gtask, q, dev, irq->ii_res, name);
6451 	}
6452 
6453 	return (0);
6454 }
6455 
6456 void
6457 iflib_softirq_alloc_generic(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type,
6458 			    void *arg, int qid, const char *name)
6459 {
6460 	device_t dev;
6461 	struct grouptask *gtask;
6462 	struct taskqgroup *tqg;
6463 	gtask_fn_t *fn;
6464 	void *q;
6465 	int err;
6466 
6467 	switch (type) {
6468 	case IFLIB_INTR_TX:
6469 		q = &ctx->ifc_txqs[qid];
6470 		gtask = &ctx->ifc_txqs[qid].ift_task;
6471 		tqg = qgroup_if_io_tqg;
6472 		fn = _task_fn_tx;
6473 		GROUPTASK_INIT(gtask, 0, fn, q);
6474 		break;
6475 	case IFLIB_INTR_RX:
6476 		q = &ctx->ifc_rxqs[qid];
6477 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6478 		tqg = qgroup_if_io_tqg;
6479 		fn = _task_fn_rx;
6480 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6481 		break;
6482 	case IFLIB_INTR_IOV:
6483 		TASK_INIT(&ctx->ifc_vflr_task, 0, _task_fn_iov, ctx);
6484 		return;
6485 	default:
6486 		panic("unknown net intr type");
6487 	}
6488 	err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg, q, name);
6489 	if (err) {
6490 		dev = ctx->ifc_dev;
6491 		taskqgroup_attach(tqg, gtask, q, dev, irq ? irq->ii_res : NULL,
6492 		    name);
6493 	}
6494 }
6495 
6496 void
6497 iflib_irq_free(if_ctx_t ctx, if_irq_t irq)
6498 {
6499 
6500 	if (irq->ii_tag)
6501 		bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag);
6502 
6503 	if (irq->ii_res)
6504 		bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ,
6505 		    rman_get_rid(irq->ii_res), irq->ii_res);
6506 }
6507 
6508 static int
6509 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, const char *name)
6510 {
6511 	iflib_txq_t txq = ctx->ifc_txqs;
6512 	iflib_rxq_t rxq = ctx->ifc_rxqs;
6513 	if_irq_t irq = &ctx->ifc_legacy_irq;
6514 	iflib_filter_info_t info;
6515 	device_t dev;
6516 	struct grouptask *gtask;
6517 	struct resource *res;
6518 	int err, tqrid;
6519 	bool rx_only;
6520 
6521 	info = &rxq->ifr_filter_info;
6522 	gtask = &rxq->ifr_task;
6523 	tqrid = *rid;
6524 	rx_only = (ctx->ifc_sctx->isc_flags & IFLIB_SINGLE_IRQ_RX_ONLY) != 0;
6525 
6526 	ctx->ifc_flags |= IFC_LEGACY;
6527 	info->ifi_filter = filter;
6528 	info->ifi_filter_arg = filter_arg;
6529 	info->ifi_task = gtask;
6530 	info->ifi_ctx = rxq;
6531 
6532 	dev = ctx->ifc_dev;
6533 	/* We allocate a single interrupt resource */
6534 	err = _iflib_irq_alloc(ctx, irq, tqrid, rx_only ? iflib_fast_intr :
6535 	    iflib_fast_intr_rxtx, NULL, info, name);
6536 	if (err != 0)
6537 		return (err);
6538 	NET_GROUPTASK_INIT(gtask, 0, _task_fn_rx, rxq);
6539 	res = irq->ii_res;
6540 	taskqgroup_attach(qgroup_if_io_tqg, gtask, rxq, dev, res, name);
6541 
6542 	GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq);
6543 	taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, dev, res,
6544 	    "tx");
6545 	return (0);
6546 }
6547 
6548 void
6549 iflib_led_create(if_ctx_t ctx)
6550 {
6551 
6552 	ctx->ifc_led_dev = led_create(iflib_led_func, ctx,
6553 	    device_get_nameunit(ctx->ifc_dev));
6554 }
6555 
6556 void
6557 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid)
6558 {
6559 
6560 	GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
6561 }
6562 
6563 void
6564 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid)
6565 {
6566 
6567 	GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task);
6568 }
6569 
6570 void
6571 iflib_admin_intr_deferred(if_ctx_t ctx)
6572 {
6573 
6574 	taskqueue_enqueue(ctx->ifc_tq, &ctx->ifc_admin_task);
6575 }
6576 
6577 void
6578 iflib_iov_intr_deferred(if_ctx_t ctx)
6579 {
6580 
6581 	taskqueue_enqueue(ctx->ifc_tq, &ctx->ifc_vflr_task);
6582 }
6583 
6584 void
6585 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, const char *name)
6586 {
6587 
6588 	taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, NULL, NULL,
6589 	    name);
6590 }
6591 
6592 void
6593 iflib_config_task_init(if_ctx_t ctx, struct task *config_task, task_fn_t *fn)
6594 {
6595 	TASK_INIT(config_task, 0, fn, ctx);
6596 }
6597 
6598 void
6599 iflib_config_task_enqueue(if_ctx_t ctx, struct task *config_task)
6600 {
6601 	taskqueue_enqueue(ctx->ifc_tq, config_task);
6602 }
6603 
6604 void
6605 iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate)
6606 {
6607 	if_t ifp = ctx->ifc_ifp;
6608 
6609 	if_setbaudrate(ifp, baudrate);
6610 	if (baudrate >= IF_Gbps(10)) {
6611 		STATE_LOCK(ctx);
6612 		ctx->ifc_flags |= IFC_PREFETCH;
6613 		STATE_UNLOCK(ctx);
6614 	}
6615 	ctx->ifc_link_state = link_state;
6616 	if_link_state_change(ifp, link_state);
6617 }
6618 
6619 static int
6620 iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq)
6621 {
6622 	int credits;
6623 #ifdef INVARIANTS
6624 	int credits_pre = txq->ift_cidx_processed;
6625 #endif
6626 
6627 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
6628 	    BUS_DMASYNC_POSTREAD);
6629 	if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, true)) == 0)
6630 		return (0);
6631 
6632 	txq->ift_processed += credits;
6633 	txq->ift_cidx_processed += credits;
6634 
6635 	MPASS(credits_pre + credits == txq->ift_cidx_processed);
6636 	if (txq->ift_cidx_processed >= txq->ift_size)
6637 		txq->ift_cidx_processed -= txq->ift_size;
6638 	return (credits);
6639 }
6640 
6641 static int
6642 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget)
6643 {
6644 	iflib_fl_t fl;
6645 	u_int i;
6646 
6647 	for (i = 0, fl = &rxq->ifr_fl[0]; i < rxq->ifr_nfl; i++, fl++)
6648 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
6649 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
6650 	return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx,
6651 	    budget));
6652 }
6653 
6654 void
6655 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name,
6656 	const char *description, if_int_delay_info_t info,
6657 	int offset, int value)
6658 {
6659 	info->iidi_ctx = ctx;
6660 	info->iidi_offset = offset;
6661 	info->iidi_value = value;
6662 	SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev),
6663 	    SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)),
6664 	    OID_AUTO, name, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE,
6665 	    info, 0, iflib_sysctl_int_delay, "I", description);
6666 }
6667 
6668 struct sx *
6669 iflib_ctx_lock_get(if_ctx_t ctx)
6670 {
6671 
6672 	return (&ctx->ifc_ctx_sx);
6673 }
6674 
6675 static int
6676 iflib_msix_init(if_ctx_t ctx)
6677 {
6678 	device_t dev = ctx->ifc_dev;
6679 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6680 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6681 	int admincnt, bar, err, iflib_num_rx_queues, iflib_num_tx_queues;
6682 	int msgs, queuemsgs, queues, rx_queues, tx_queues, vectors;
6683 
6684 	iflib_num_tx_queues = ctx->ifc_sysctl_ntxqs;
6685 	iflib_num_rx_queues = ctx->ifc_sysctl_nrxqs;
6686 
6687 	if (bootverbose)
6688 		device_printf(dev, "msix_init qsets capped at %d\n",
6689 		    imax(scctx->isc_ntxqsets, scctx->isc_nrxqsets));
6690 
6691 	/* Override by tuneable */
6692 	if (scctx->isc_disable_msix)
6693 		goto msi;
6694 
6695 	/* First try MSI-X */
6696 	if ((msgs = pci_msix_count(dev)) == 0) {
6697 		if (bootverbose)
6698 			device_printf(dev, "MSI-X not supported or disabled\n");
6699 		goto msi;
6700 	}
6701 
6702 	bar = ctx->ifc_softc_ctx.isc_msix_bar;
6703 	/*
6704 	 * bar == -1 => "trust me I know what I'm doing"
6705 	 * Some drivers are for hardware that is so shoddily
6706 	 * documented that no one knows which bars are which
6707 	 * so the developer has to map all bars. This hack
6708 	 * allows shoddy garbage to use MSI-X in this framework.
6709 	 */
6710 	if (bar != -1) {
6711 		ctx->ifc_msix_mem = bus_alloc_resource_any(dev,
6712 		    SYS_RES_MEMORY, &bar, RF_ACTIVE);
6713 		if (ctx->ifc_msix_mem == NULL) {
6714 			device_printf(dev, "Unable to map MSI-X table\n");
6715 			goto msi;
6716 		}
6717 	}
6718 
6719 	admincnt = sctx->isc_admin_intrcnt;
6720 #if IFLIB_DEBUG
6721 	/* use only 1 qset in debug mode */
6722 	queuemsgs = min(msgs - admincnt, 1);
6723 #else
6724 	queuemsgs = msgs - admincnt;
6725 #endif
6726 #ifdef RSS
6727 	queues = imin(queuemsgs, rss_getnumbuckets());
6728 #else
6729 	queues = queuemsgs;
6730 #endif
6731 	queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues);
6732 	if (bootverbose)
6733 		device_printf(dev,
6734 		    "intr CPUs: %d queue msgs: %d admincnt: %d\n",
6735 		    CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt);
6736 #ifdef  RSS
6737 	/* If we're doing RSS, clamp at the number of RSS buckets */
6738 	if (queues > rss_getnumbuckets())
6739 		queues = rss_getnumbuckets();
6740 #endif
6741 	if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt)
6742 		rx_queues = iflib_num_rx_queues;
6743 	else
6744 		rx_queues = queues;
6745 
6746 	if (rx_queues > scctx->isc_nrxqsets)
6747 		rx_queues = scctx->isc_nrxqsets;
6748 
6749 	/*
6750 	 * We want this to be all logical CPUs by default
6751 	 */
6752 	if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues)
6753 		tx_queues = iflib_num_tx_queues;
6754 	else
6755 		tx_queues = mp_ncpus;
6756 
6757 	if (tx_queues > scctx->isc_ntxqsets)
6758 		tx_queues = scctx->isc_ntxqsets;
6759 
6760 	if (ctx->ifc_sysctl_qs_eq_override == 0) {
6761 #ifdef INVARIANTS
6762 		if (tx_queues != rx_queues)
6763 			device_printf(dev,
6764 			    "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n",
6765 			    min(rx_queues, tx_queues), min(rx_queues, tx_queues));
6766 #endif
6767 		tx_queues = min(rx_queues, tx_queues);
6768 		rx_queues = min(rx_queues, tx_queues);
6769 	}
6770 
6771 	vectors = rx_queues + admincnt;
6772 	if (msgs < vectors) {
6773 		device_printf(dev,
6774 		    "insufficient number of MSI-X vectors "
6775 		    "(supported %d, need %d)\n", msgs, vectors);
6776 		goto msi;
6777 	}
6778 
6779 	device_printf(dev, "Using %d RX queues %d TX queues\n", rx_queues,
6780 	    tx_queues);
6781 	msgs = vectors;
6782 	if ((err = pci_alloc_msix(dev, &vectors)) == 0) {
6783 		if (vectors != msgs) {
6784 			device_printf(dev,
6785 			    "Unable to allocate sufficient MSI-X vectors "
6786 			    "(got %d, need %d)\n", vectors, msgs);
6787 			pci_release_msi(dev);
6788 			if (bar != -1) {
6789 				bus_release_resource(dev, SYS_RES_MEMORY, bar,
6790 				    ctx->ifc_msix_mem);
6791 				ctx->ifc_msix_mem = NULL;
6792 			}
6793 			goto msi;
6794 		}
6795 		device_printf(dev, "Using MSI-X interrupts with %d vectors\n",
6796 		    vectors);
6797 		scctx->isc_vectors = vectors;
6798 		scctx->isc_nrxqsets = rx_queues;
6799 		scctx->isc_ntxqsets = tx_queues;
6800 		scctx->isc_intr = IFLIB_INTR_MSIX;
6801 
6802 		return (vectors);
6803 	} else {
6804 		device_printf(dev,
6805 		    "failed to allocate %d MSI-X vectors, err: %d\n", vectors,
6806 		    err);
6807 		if (bar != -1) {
6808 			bus_release_resource(dev, SYS_RES_MEMORY, bar,
6809 			    ctx->ifc_msix_mem);
6810 			ctx->ifc_msix_mem = NULL;
6811 		}
6812 	}
6813 
6814 msi:
6815 	vectors = pci_msi_count(dev);
6816 	scctx->isc_nrxqsets = 1;
6817 	scctx->isc_ntxqsets = 1;
6818 	scctx->isc_vectors = vectors;
6819 	if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) {
6820 		device_printf(dev, "Using an MSI interrupt\n");
6821 		scctx->isc_intr = IFLIB_INTR_MSI;
6822 	} else {
6823 		scctx->isc_vectors = 1;
6824 		device_printf(dev, "Using a Legacy interrupt\n");
6825 		scctx->isc_intr = IFLIB_INTR_LEGACY;
6826 	}
6827 
6828 	return (vectors);
6829 }
6830 
6831 static const char *ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" };
6832 
6833 static int
6834 mp_ring_state_handler(SYSCTL_HANDLER_ARGS)
6835 {
6836 	int rc;
6837 	uint16_t *state = ((uint16_t *)oidp->oid_arg1);
6838 	struct sbuf *sb;
6839 	const char *ring_state = "UNKNOWN";
6840 
6841 	/* XXX needed ? */
6842 	rc = sysctl_wire_old_buffer(req, 0);
6843 	MPASS(rc == 0);
6844 	if (rc != 0)
6845 		return (rc);
6846 	sb = sbuf_new_for_sysctl(NULL, NULL, 80, req);
6847 	MPASS(sb != NULL);
6848 	if (sb == NULL)
6849 		return (ENOMEM);
6850 	if (state[3] <= 3)
6851 		ring_state = ring_states[state[3]];
6852 
6853 	sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s",
6854 		    state[0], state[1], state[2], ring_state);
6855 	rc = sbuf_finish(sb);
6856 	sbuf_delete(sb);
6857 	return (rc);
6858 }
6859 
6860 enum iflib_ndesc_handler {
6861 	IFLIB_NTXD_HANDLER,
6862 	IFLIB_NRXD_HANDLER,
6863 };
6864 
6865 static int
6866 mp_ndesc_handler(SYSCTL_HANDLER_ARGS)
6867 {
6868 	if_ctx_t ctx = (void *)arg1;
6869 	enum iflib_ndesc_handler type = arg2;
6870 	char buf[256] = {0};
6871 	qidx_t *ndesc;
6872 	char *p, *next;
6873 	int nqs, rc, i;
6874 
6875 	nqs = 8;
6876 	switch (type) {
6877 	case IFLIB_NTXD_HANDLER:
6878 		ndesc = ctx->ifc_sysctl_ntxds;
6879 		if (ctx->ifc_sctx)
6880 			nqs = ctx->ifc_sctx->isc_ntxqs;
6881 		break;
6882 	case IFLIB_NRXD_HANDLER:
6883 		ndesc = ctx->ifc_sysctl_nrxds;
6884 		if (ctx->ifc_sctx)
6885 			nqs = ctx->ifc_sctx->isc_nrxqs;
6886 		break;
6887 	default:
6888 		printf("%s: unhandled type\n", __func__);
6889 		return (EINVAL);
6890 	}
6891 	if (nqs == 0)
6892 		nqs = 8;
6893 
6894 	for (i = 0; i < 8; i++) {
6895 		if (i >= nqs)
6896 			break;
6897 		if (i)
6898 			strcat(buf, ",");
6899 		sprintf(strchr(buf, 0), "%d", ndesc[i]);
6900 	}
6901 
6902 	rc = sysctl_handle_string(oidp, buf, sizeof(buf), req);
6903 	if (rc || req->newptr == NULL)
6904 		return (rc);
6905 
6906 	for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p;
6907 	    i++, p = strsep(&next, " ,")) {
6908 		ndesc[i] = strtoul(p, NULL, 10);
6909 	}
6910 
6911 	return (rc);
6912 }
6913 
6914 static int
6915 iflib_handle_tx_reclaim_thresh(SYSCTL_HANDLER_ARGS)
6916 {
6917 	if_ctx_t ctx = (void *)arg1;
6918 	iflib_txq_t txq;
6919 	int i, err;
6920 	int thresh;
6921 
6922 	thresh = ctx->ifc_sysctl_tx_reclaim_thresh;
6923 	err = sysctl_handle_int(oidp, &thresh, arg2, req);
6924 	if (err != 0) {
6925 		return err;
6926 	}
6927 
6928 	if (thresh == ctx->ifc_sysctl_tx_reclaim_thresh)
6929 		return 0;
6930 
6931 	if (thresh > ctx->ifc_softc_ctx.isc_ntxd[0] / 2) {
6932 		device_printf(ctx->ifc_dev, "TX Reclaim thresh must be <= %d\n",
6933 		    ctx->ifc_softc_ctx.isc_ntxd[0] / 2);
6934 		return (EINVAL);
6935 	}
6936 
6937 	ctx->ifc_sysctl_tx_reclaim_thresh = thresh;
6938 	if (ctx->ifc_txqs == NULL)
6939 		return (err);
6940 
6941 	txq = &ctx->ifc_txqs[0];
6942 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
6943 		txq->ift_reclaim_thresh = thresh;
6944 	}
6945 	return (err);
6946 }
6947 
6948 static int
6949 iflib_handle_tx_reclaim_ticks(SYSCTL_HANDLER_ARGS)
6950 {
6951 	if_ctx_t ctx = (void *)arg1;
6952 	iflib_txq_t txq;
6953 	int i, err;
6954 	int ticks;
6955 
6956 	ticks = ctx->ifc_sysctl_tx_reclaim_ticks;
6957 	err = sysctl_handle_int(oidp, &ticks, arg2, req);
6958 	if (err != 0) {
6959 		return err;
6960 	}
6961 
6962 	if (ticks == ctx->ifc_sysctl_tx_reclaim_ticks)
6963 		return 0;
6964 
6965 	if (ticks > hz) {
6966 		device_printf(ctx->ifc_dev,
6967 		    "TX Reclaim ticks must be <= hz (%d)\n", hz);
6968 		return (EINVAL);
6969 	}
6970 
6971 	ctx->ifc_sysctl_tx_reclaim_ticks = ticks;
6972 	if (ctx->ifc_txqs == NULL)
6973 		return (err);
6974 
6975 	txq = &ctx->ifc_txqs[0];
6976 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
6977 		txq->ift_reclaim_ticks = ticks;
6978 	}
6979 	return (err);
6980 }
6981 
6982 static int
6983 iflib_handle_tx_defer_mfree(SYSCTL_HANDLER_ARGS)
6984 {
6985 	if_ctx_t ctx = (void *)arg1;
6986 	iflib_txq_t txq;
6987 	int i, err;
6988 	int defer;
6989 
6990 	defer = ctx->ifc_sysctl_tx_defer_mfree;
6991 	err = sysctl_handle_int(oidp, &defer, arg2, req);
6992 	if (err != 0) {
6993 		return err;
6994 	}
6995 
6996 	if (defer == ctx->ifc_sysctl_tx_defer_mfree)
6997 		return 0;
6998 
6999 	ctx->ifc_sysctl_tx_defer_mfree = defer;
7000 	if (ctx->ifc_txqs == NULL)
7001 		return (err);
7002 
7003 	txq = &ctx->ifc_txqs[0];
7004 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
7005 		txq->ift_defer_mfree = defer;
7006 	}
7007 	return (err);
7008 }
7009 
7010 #define NAME_BUFLEN 32
7011 static void
7012 iflib_add_device_sysctl_pre(if_ctx_t ctx)
7013 {
7014 	device_t dev = iflib_get_dev(ctx);
7015 	struct sysctl_oid_list *child, *oid_list;
7016 	struct sysctl_oid *node;
7017 
7018 	sysctl_ctx_init(&ctx->ifc_sysctl_ctx);
7019 	child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev));
7020 	ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(&ctx->ifc_sysctl_ctx, child,
7021 	    OID_AUTO, "iflib", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL,
7022 	    "IFLIB fields");
7023 	oid_list = SYSCTL_CHILDREN(node);
7024 
7025 	SYSCTL_ADD_CONST_STRING(&ctx->ifc_sysctl_ctx, oid_list, OID_AUTO, "driver_version",
7026 	    CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version, "driver version");
7027 	SYSCTL_ADD_U32(&ctx->ifc_sysctl_ctx, oid_list, OID_AUTO,
7028 	    "tx_watchdog_events", CTLFLAG_RD, &ctx->ifc_tx_watchdog_events, 0,
7029 	    "TX watchdog resets initiated by iflib");
7030 
7031 	SYSCTL_ADD_BOOL(&ctx->ifc_sysctl_ctx, oid_list, OID_AUTO, "simple_tx",
7032 	    CTLFLAG_RDTUN, &ctx->ifc_sysctl_simple_tx, 0,
7033 	    "use simple tx ring");
7034 	SYSCTL_ADD_U16(&ctx->ifc_sysctl_ctx, oid_list, OID_AUTO, "override_ntxqs",
7035 	    CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0,
7036 	    "# of txqs to use, 0 => use default #");
7037 	SYSCTL_ADD_U16(&ctx->ifc_sysctl_ctx, oid_list, OID_AUTO, "override_nrxqs",
7038 	    CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0,
7039 	    "# of rxqs to use, 0 => use default #");
7040 	SYSCTL_ADD_U16(&ctx->ifc_sysctl_ctx, oid_list, OID_AUTO, "override_qs_enable",
7041 	    CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0,
7042 	    "permit #txq != #rxq");
7043 	SYSCTL_ADD_INT(&ctx->ifc_sysctl_ctx, oid_list, OID_AUTO, "disable_msix",
7044 	    CTLFLAG_RWTUN, &ctx->ifc_softc_ctx.isc_disable_msix, 0,
7045 	    "disable MSI-X (default 0)");
7046 	SYSCTL_ADD_U16(&ctx->ifc_sysctl_ctx, oid_list, OID_AUTO, "rx_budget",
7047 	    CTLFLAG_RWTUN, &ctx->ifc_sysctl_rx_budget, 0, "set the RX budget");
7048 	SYSCTL_ADD_U16(&ctx->ifc_sysctl_ctx, oid_list, OID_AUTO, "tx_abdicate",
7049 	    CTLFLAG_RWTUN, &ctx->ifc_sysctl_tx_abdicate, 0,
7050 	    "cause TX to abdicate instead of running to completion");
7051 	ctx->ifc_sysctl_core_offset = CORE_OFFSET_UNSPECIFIED;
7052 	SYSCTL_ADD_U16(&ctx->ifc_sysctl_ctx, oid_list, OID_AUTO, "core_offset",
7053 	    CTLFLAG_RDTUN, &ctx->ifc_sysctl_core_offset, 0,
7054 	    "offset to start using cores at");
7055 	SYSCTL_ADD_U8(&ctx->ifc_sysctl_ctx, oid_list, OID_AUTO, "separate_txrx",
7056 	    CTLFLAG_RDTUN, &ctx->ifc_sysctl_separate_txrx, 0,
7057 	    "use separate cores for TX and RX");
7058 	SYSCTL_ADD_U8(&ctx->ifc_sysctl_ctx, oid_list, OID_AUTO, "use_logical_cores",
7059 	    CTLFLAG_RDTUN, &ctx->ifc_sysctl_use_logical_cores, 0,
7060 	    "try to make use of logical cores for TX and RX");
7061 	SYSCTL_ADD_U16(&ctx->ifc_sysctl_ctx, oid_list, OID_AUTO, "use_extra_msix_vectors",
7062 	    CTLFLAG_RDTUN, &ctx->ifc_sysctl_extra_msix_vectors, 0,
7063 	    "attempt to reserve the given number of extra MSI-X vectors during driver load for the creation of additional interfaces later");
7064 	SYSCTL_ADD_INT(&ctx->ifc_sysctl_ctx, oid_list, OID_AUTO, "allocated_msix_vectors",
7065 	    CTLFLAG_RDTUN, &ctx->ifc_softc_ctx.isc_vectors, 0,
7066 	    "total # of MSI-X vectors allocated by driver");
7067 
7068 	/* XXX change for per-queue sizes */
7069 	SYSCTL_ADD_PROC(&ctx->ifc_sysctl_ctx, oid_list, OID_AUTO, "override_ntxds",
7070 	    CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NEEDGIANT, ctx,
7071 	    IFLIB_NTXD_HANDLER, mp_ndesc_handler, "A",
7072 	    "list of # of TX descriptors to use, 0 = use default #");
7073 	SYSCTL_ADD_PROC(&ctx->ifc_sysctl_ctx, oid_list, OID_AUTO, "override_nrxds",
7074 	    CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NEEDGIANT, ctx,
7075 	    IFLIB_NRXD_HANDLER, mp_ndesc_handler, "A",
7076 	    "list of # of RX descriptors to use, 0 = use default #");
7077 }
7078 
7079 static void
7080 iflib_add_device_sysctl_post(if_ctx_t ctx)
7081 {
7082 	if_shared_ctx_t sctx = ctx->ifc_sctx;
7083 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
7084 	struct sysctl_oid_list *child;
7085 	struct sysctl_ctx_list *ctx_list = &ctx->ifc_sysctl_ctx;
7086 	iflib_fl_t fl;
7087 	iflib_txq_t txq;
7088 	iflib_rxq_t rxq;
7089 	int i, j;
7090 	char namebuf[NAME_BUFLEN];
7091 	char *qfmt;
7092 	struct sysctl_oid *queue_node, *fl_node, *node;
7093 	struct sysctl_oid_list *queue_list, *fl_list;
7094 
7095 	node = ctx->ifc_sysctl_node;
7096 	child = SYSCTL_CHILDREN(node);
7097 
7098        SYSCTL_ADD_PROC(ctx_list, child, OID_AUTO, "tx_reclaim_thresh",
7099            CTLTYPE_INT | CTLFLAG_RWTUN, ctx,
7100            0, iflib_handle_tx_reclaim_thresh, "I",
7101            "Number of TX descs outstanding before reclaim is called");
7102 
7103        SYSCTL_ADD_PROC(ctx_list, child, OID_AUTO, "tx_reclaim_ticks",
7104            CTLTYPE_INT | CTLFLAG_RWTUN, ctx,
7105            0, iflib_handle_tx_reclaim_ticks, "I",
7106            "Number of ticks before a TX reclaim is forced");
7107 
7108        SYSCTL_ADD_PROC(ctx_list, child, OID_AUTO, "tx_defer_mfree",
7109            CTLTYPE_INT | CTLFLAG_RWTUN, ctx,
7110            0, iflib_handle_tx_defer_mfree, "I",
7111            "Free completed transmits outside of TX ring lock");
7112 
7113 	if (scctx->isc_ntxqsets > 100)
7114 		qfmt = "txq%03d";
7115 	else if (scctx->isc_ntxqsets > 10)
7116 		qfmt = "txq%02d";
7117 	else
7118 		qfmt = "txq%d";
7119 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
7120 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
7121 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
7122 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "Queue Name");
7123 		queue_list = SYSCTL_CHILDREN(queue_node);
7124 		SYSCTL_ADD_INT(ctx_list, queue_list, OID_AUTO, "cpu",
7125 		    CTLFLAG_RD, &txq->ift_task.gt_cpu, 0,
7126 		    "cpu this queue is bound to");
7127 #if MEMORY_LOGGING
7128 		SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued",
7129 		    CTLFLAG_RD, &txq->ift_dequeued, "total mbufs freed");
7130 		SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued",
7131 		    CTLFLAG_RD, &txq->ift_enqueued, "total mbufs enqueued");
7132 #endif
7133 		SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag",
7134 		    CTLFLAG_RD, &txq->ift_mbuf_defrag,
7135 		    "# of times m_defrag was called");
7136 		SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO, "m_pullups",
7137 		    CTLFLAG_RD, &txq->ift_pullups,
7138 		    "# of times m_pullup was called");
7139 		SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO,
7140 		    "mbuf_defrag_failed", CTLFLAG_RD,
7141 		    &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed");
7142 		SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO,
7143 		    "no_desc_avail", CTLFLAG_RD, &txq->ift_no_desc_avail,
7144 		    "# of times no descriptors were available");
7145 		SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO,
7146 		    "tx_map_failed", CTLFLAG_RD, &txq->ift_map_failed,
7147 		    "# of times DMA map failed");
7148 		SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO,
7149 		    "txd_encap_efbig", CTLFLAG_RD, &txq->ift_txd_encap_efbig,
7150 		    "# of times txd_encap returned EFBIG");
7151 		SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO,
7152 		    "no_tx_dma_setup", CTLFLAG_RD, &txq->ift_no_tx_dma_setup,
7153 		    "# of times map failed for other than EFBIG");
7154 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx",
7155 		    CTLFLAG_RD, &txq->ift_pidx, 1, "Producer Index");
7156 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx",
7157 		    CTLFLAG_RD, &txq->ift_cidx, 1, "Consumer Index");
7158 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO,
7159 		    "txq_cidx_processed", CTLFLAG_RD, &txq->ift_cidx_processed,
7160 		    1, "Consumer Index seen by credit update");
7161 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use",
7162 		    CTLFLAG_RD, &txq->ift_in_use, 1, "descriptors in use");
7163 		SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO,
7164 		    "txq_processed", CTLFLAG_RD, &txq->ift_processed,
7165 		    "descriptors procesed for clean");
7166 		SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned",
7167 		    CTLFLAG_RD, &txq->ift_cleaned, "total cleaned");
7168 		SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state",
7169 		    CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_NEEDGIANT,
7170 		    __DEVOLATILE(uint64_t *, &txq->ift_br->state), 0,
7171 		    mp_ring_state_handler, "A", "soft ring state");
7172 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO,
7173 		    "r_enqueues", CTLFLAG_RD, &txq->ift_br->enqueues,
7174 		    "# of enqueues to the mp_ring for this queue");
7175 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO,
7176 		    "r_drops", CTLFLAG_RD, &txq->ift_br->drops,
7177 		    "# of drops in the mp_ring for this queue");
7178 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO,
7179 		    "r_starts", CTLFLAG_RD, &txq->ift_br->starts,
7180 		    "# of normal consumer starts in mp_ring for this queue");
7181 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO,
7182 		    "r_stalls", CTLFLAG_RD, &txq->ift_br->stalls,
7183 		    "# of consumer stalls in the mp_ring for this queue");
7184 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO,
7185 		    "r_restarts", CTLFLAG_RD, &txq->ift_br->restarts,
7186 		    "# of consumer restarts in the mp_ring for this queue");
7187 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO,
7188 		    "r_abdications", CTLFLAG_RD, &txq->ift_br->abdications,
7189 		    "# of consumer abdications in the mp_ring for this queue");
7190 	}
7191 
7192 	if (scctx->isc_nrxqsets > 100)
7193 		qfmt = "rxq%03d";
7194 	else if (scctx->isc_nrxqsets > 10)
7195 		qfmt = "rxq%02d";
7196 	else
7197 		qfmt = "rxq%d";
7198 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
7199 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
7200 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
7201 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "Queue Name");
7202 		queue_list = SYSCTL_CHILDREN(queue_node);
7203 		SYSCTL_ADD_INT(ctx_list, queue_list, OID_AUTO, "cpu",
7204 		    CTLFLAG_RD, &rxq->ifr_task.gt_cpu, 0,
7205 		    "cpu this queue is bound to");
7206 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
7207 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO,
7208 			    "rxq_cq_cidx", CTLFLAG_RD, &rxq->ifr_cq_cidx, 1,
7209 			    "Consumer Index");
7210 		}
7211 
7212 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
7213 			snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j);
7214 			fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list,
7215 			    OID_AUTO, namebuf, CTLFLAG_RD | CTLFLAG_MPSAFE,
7216 			    NULL, "freelist Name");
7217 			fl_list = SYSCTL_CHILDREN(fl_node);
7218 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx",
7219 			    CTLFLAG_RD, &fl->ifl_pidx, 1, "Producer Index");
7220 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx",
7221 			    CTLFLAG_RD, &fl->ifl_cidx, 1, "Consumer Index");
7222 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits",
7223 			    CTLFLAG_RD, &fl->ifl_credits, 1,
7224 			    "credits available");
7225 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "buf_size",
7226 			    CTLFLAG_RD, &fl->ifl_buf_size, 1, "buffer size");
7227 #if MEMORY_LOGGING
7228 			SYSCTL_ADD_UQUAD(ctx_list, fl_list, OID_AUTO,
7229 			    "fl_m_enqueued", CTLFLAG_RD, &fl->ifl_m_enqueued,
7230 			    "mbufs allocated");
7231 			SYSCTL_ADD_UQUAD(ctx_list, fl_list, OID_AUTO,
7232 			    "fl_m_dequeued", CTLFLAG_RD, &fl->ifl_m_dequeued,
7233 			    "mbufs freed");
7234 			SYSCTL_ADD_UQUAD(ctx_list, fl_list, OID_AUTO,
7235 			    "fl_cl_enqueued", CTLFLAG_RD, &fl->ifl_cl_enqueued,
7236 			    "clusters allocated");
7237 			SYSCTL_ADD_UQUAD(ctx_list, fl_list, OID_AUTO,
7238 			    "fl_cl_dequeued", CTLFLAG_RD, &fl->ifl_cl_dequeued,
7239 			    "clusters freed");
7240 #endif
7241 		}
7242 	}
7243 
7244 }
7245 
7246 void
7247 iflib_request_reset(if_ctx_t ctx)
7248 {
7249 
7250 	STATE_LOCK(ctx);
7251 	ctx->ifc_flags |= IFC_DO_RESET;
7252 	STATE_UNLOCK(ctx);
7253 }
7254 
7255 void
7256 iflib_request_reset_if_up(if_ctx_t ctx)
7257 {
7258 
7259 	STATE_LOCK(ctx);
7260 	ctx->ifc_flags |= IFC_DO_RESET_IF_UP;
7261 	STATE_UNLOCK(ctx);
7262 }
7263 
7264 void
7265 iflib_init_failed(if_ctx_t ctx)
7266 {
7267 
7268 	sx_assert(&ctx->ifc_ctx_sx, SA_XLOCKED);
7269 	STATE_LOCK(ctx);
7270 	ctx->ifc_flags |= IFC_INIT_FAILED;
7271 	STATE_UNLOCK(ctx);
7272 }
7273 
7274 #ifndef __NO_STRICT_ALIGNMENT
7275 static struct mbuf *
7276 iflib_fixup_rx(struct mbuf *m)
7277 {
7278 	struct mbuf *n;
7279 
7280 	if (m->m_len <= (MCLBYTES - ETHER_HDR_LEN)) {
7281 		bcopy(m->m_data, m->m_data + ETHER_HDR_LEN, m->m_len);
7282 		m->m_data += ETHER_HDR_LEN;
7283 		n = m;
7284 	} else {
7285 		MGETHDR(n, M_NOWAIT, MT_DATA);
7286 		if (n == NULL) {
7287 			m_freem(m);
7288 			return (NULL);
7289 		}
7290 		bcopy(m->m_data, n->m_data, ETHER_HDR_LEN);
7291 		m->m_data += ETHER_HDR_LEN;
7292 		m->m_len -= ETHER_HDR_LEN;
7293 		n->m_len = ETHER_HDR_LEN;
7294 		M_MOVE_PKTHDR(n, m);
7295 		n->m_next = m;
7296 	}
7297 	return (n);
7298 }
7299 #endif
7300 
7301 #ifdef DEBUGNET
7302 static void
7303 iflib_debugnet_init(if_t ifp, int *nrxr, int *ncl, int *clsize)
7304 {
7305 	if_ctx_t ctx;
7306 
7307 	ctx = if_getsoftc(ifp);
7308 	CTX_LOCK(ctx);
7309 	*nrxr = NRXQSETS(ctx);
7310 	*ncl = ctx->ifc_rxqs[0].ifr_fl->ifl_size;
7311 	*clsize = ctx->ifc_rxqs[0].ifr_fl->ifl_buf_size;
7312 	CTX_UNLOCK(ctx);
7313 }
7314 
7315 static void
7316 iflib_debugnet_event(if_t ifp, enum debugnet_ev event)
7317 {
7318 	if_ctx_t ctx;
7319 	if_softc_ctx_t scctx;
7320 	iflib_fl_t fl;
7321 	iflib_rxq_t rxq;
7322 	int i, j;
7323 
7324 	ctx = if_getsoftc(ifp);
7325 	scctx = &ctx->ifc_softc_ctx;
7326 
7327 	switch (event) {
7328 	case DEBUGNET_START:
7329 		for (i = 0; i < scctx->isc_nrxqsets; i++) {
7330 			rxq = &ctx->ifc_rxqs[i];
7331 			for (j = 0; j < rxq->ifr_nfl; j++) {
7332 				fl = rxq->ifr_fl;
7333 				fl->ifl_zone = m_getzone(fl->ifl_buf_size);
7334 			}
7335 		}
7336 		iflib_no_tx_batch = 1;
7337 		break;
7338 	default:
7339 		break;
7340 	}
7341 }
7342 
7343 static int
7344 iflib_debugnet_transmit(if_t ifp, struct mbuf *m)
7345 {
7346 	if_ctx_t ctx;
7347 	iflib_txq_t txq;
7348 	int error;
7349 	int bytes_sent = 0;
7350 	int pkt_sent = 0;
7351 
7352 	ctx = if_getsoftc(ifp);
7353 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
7354 	    IFF_DRV_RUNNING)
7355 		return (EBUSY);
7356 
7357 	txq = &ctx->ifc_txqs[0];
7358 	error = iflib_encap(txq, &m, &bytes_sent, &pkt_sent);
7359 	if (error == 0)
7360 		(void)iflib_txd_db_check(txq, true);
7361 	return (error);
7362 }
7363 
7364 static int
7365 iflib_debugnet_poll(if_t ifp, int count)
7366 {
7367 	struct epoch_tracker et;
7368 	if_ctx_t ctx;
7369 	if_softc_ctx_t scctx;
7370 	iflib_txq_t txq;
7371 	int i;
7372 
7373 	ctx = if_getsoftc(ifp);
7374 	scctx = &ctx->ifc_softc_ctx;
7375 
7376 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
7377 	    IFF_DRV_RUNNING)
7378 		return (EBUSY);
7379 
7380 	txq = &ctx->ifc_txqs[0];
7381 	(void)iflib_completed_tx_reclaim(txq, NULL);
7382 
7383 	NET_EPOCH_ENTER(et);
7384 	for (i = 0; i < scctx->isc_nrxqsets; i++)
7385 		(void)iflib_rxeof(&ctx->ifc_rxqs[i], 16 /* XXX */);
7386 	NET_EPOCH_EXIT(et);
7387 	return (0);
7388 }
7389 #endif /* DEBUGNET */
7390 
7391 #ifndef ALTQ
7392 static inline iflib_txq_t
7393 iflib_simple_select_queue(if_ctx_t ctx, struct mbuf *m)
7394 {
7395 	int qidx;
7396 
7397 	if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m))
7398 		qidx = QIDX(ctx, m);
7399 	else
7400 		qidx = NTXQSETS(ctx) + FIRST_QSET(ctx) - 1;
7401 	return (&ctx->ifc_txqs[qidx]);
7402 }
7403 
7404 static int
7405 iflib_simple_transmit(if_t ifp, struct mbuf *m)
7406 {
7407 	if_ctx_t ctx;
7408 	iflib_txq_t txq;
7409 	struct mbuf **m_defer;
7410 	int error, i, reclaimable;
7411 	int bytes_sent = 0, pkt_sent = 0, mcast_sent = 0;
7412 
7413 
7414 	ctx = if_getsoftc(ifp);
7415 	if (__predict_false((if_getdrvflags(ifp) & IFF_DRV_RUNNING) == 0
7416 		|| !LINK_ACTIVE(ctx))) {
7417 		DBG_COUNTER_INC(tx_frees);
7418 		m_freem(m);
7419 		return (ENETDOWN);
7420 	}
7421 
7422 	txq = iflib_simple_select_queue(ctx, m);
7423 	mtx_lock(&txq->ift_mtx);
7424 	error = iflib_encap(txq, &m, &bytes_sent, &pkt_sent);
7425 	if (error == 0) {
7426 		mcast_sent += !!(m->m_flags & M_MCAST);
7427 		(void)iflib_txd_db_check(txq, true);
7428 	} else {
7429 		if (error == ENOBUFS)
7430 			if_inc_counter(ifp, IFCOUNTER_OQDROPS, 1);
7431 		else
7432 			if_inc_counter(ifp, IFCOUNTER_OERRORS, 1);
7433 	}
7434 	m_defer = NULL;
7435 	reclaimable = iflib_txq_can_reclaim(txq);
7436 	if (reclaimable != 0) {
7437 		/*
7438 		 * Try to set m_defer to the deferred mbuf reclaim array.  If
7439 		 * we can, the frees will happen outside the tx lock.  If we
7440 		 * can't, it means another thread is still proccessing frees.
7441 		 */
7442 		if (txq->ift_defer_mfree &&
7443 		    atomic_cmpset_acq_ptr((uintptr_t *)&txq->ift_sds.ifsd_m_defer,
7444 			(uintptr_t )txq->ift_sds.ifsd_m_deferb, 0)) {
7445 			m_defer = txq->ift_sds.ifsd_m_deferb;
7446 		}
7447 		_iflib_completed_tx_reclaim(txq, m_defer, reclaimable);
7448 	}
7449 	mtx_unlock(&txq->ift_mtx);
7450 
7451 	/*
7452 	 * Process mbuf frees outside the tx lock
7453 	 */
7454 	if (m_defer != NULL) {
7455 		for (i = 0; m_defer[i] != NULL; i++) {
7456 			m_freem(m_defer[i]);
7457 			m_defer[i] = NULL;
7458 		}
7459 		atomic_store_rel_ptr((uintptr_t *)&txq->ift_sds.ifsd_m_defer,
7460 		    (uintptr_t)m_defer);
7461 	}
7462 	if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
7463 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
7464 	if (mcast_sent)
7465 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
7466 
7467 	return (error);
7468 }
7469 #endif
7470