xref: /freebsd/sys/net/iflib.c (revision 9aff62dee28239f84b4aa4a3429cf26dbbf770cc)
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/fail.h>
38 #include <sys/kernel.h>
39 #include <sys/lock.h>
40 #include <sys/mutex.h>
41 #include <sys/module.h>
42 #include <sys/kobj.h>
43 #include <sys/proc.h>
44 #include <sys/rman.h>
45 #include <sys/sbuf.h>
46 #include <sys/sched.h>
47 #include <sys/smp.h>
48 #include <sys/socket.h>
49 #include <sys/sockio.h>
50 #include <sys/sysctl.h>
51 #include <sys/syslog.h>
52 #include <sys/taskqueue.h>
53 #include <sys/limits.h>
54 
55 #include <net/if.h>
56 #include <net/if_var.h>
57 #include <net/if_private.h>
58 #include <net/if_types.h>
59 #include <net/if_media.h>
60 #include <net/bpf.h>
61 #include <net/ethernet.h>
62 #include <net/mp_ring.h>
63 #include <net/debugnet.h>
64 #include <net/pfil.h>
65 #include <net/vnet.h>
66 
67 #include <netinet/in.h>
68 #include <netinet/in_pcb.h>
69 #include <netinet/tcp_lro.h>
70 #include <netinet/in_systm.h>
71 #include <netinet/if_ether.h>
72 #include <netinet/ip.h>
73 #include <netinet/ip6.h>
74 #include <netinet/tcp.h>
75 #include <netinet/udp.h>
76 #include <netinet/ip_var.h>
77 #include <netinet6/ip6_var.h>
78 
79 #include <machine/bus.h>
80 #include <machine/in_cksum.h>
81 
82 #include <vm/vm.h>
83 #include <vm/pmap.h>
84 
85 #include <dev/led/led.h>
86 #include <dev/pci/pcireg.h>
87 #include <dev/pci/pcivar.h>
88 #include <dev/pci/pci_private.h>
89 
90 #include <net/iflib.h>
91 
92 #include "ifdi_if.h"
93 
94 #ifdef PCI_IOV
95 #include <dev/pci/pci_iov.h>
96 #endif
97 
98 #include <sys/bitstring.h>
99 /*
100  * enable accounting of every mbuf as it comes in to and goes out of
101  * iflib's software descriptor references
102  */
103 #define MEMORY_LOGGING 0
104 /*
105  * Enable mbuf vectors for compressing long mbuf chains
106  */
107 
108 /*
109  * NB:
110  * - Prefetching in tx cleaning should perhaps be a tunable. The distance ahead
111  *   we prefetch needs to be determined by the time spent in m_free vis a vis
112  *   the cost of a prefetch. This will of course vary based on the workload:
113  *      - NFLX's m_free path is dominated by vm-based M_EXT manipulation which
114  *        is quite expensive, thus suggesting very little prefetch.
115  *      - small packet forwarding which is just returning a single mbuf to
116  *        UMA will typically be very fast vis a vis the cost of a memory
117  *        access.
118  */
119 
120 /*
121  * File organization:
122  *  - private structures
123  *  - iflib private utility functions
124  *  - ifnet functions
125  *  - vlan registry and other exported functions
126  *  - iflib public core functions
127  *
128  *
129  */
130 static MALLOC_DEFINE(M_IFLIB, "iflib", "ifnet library");
131 
132 #define	IFLIB_RXEOF_MORE	(1U << 0)
133 #define	IFLIB_RXEOF_EMPTY	(2U << 0)
134 
135 struct iflib_txq;
136 typedef struct iflib_txq *iflib_txq_t;
137 struct iflib_rxq;
138 typedef struct iflib_rxq *iflib_rxq_t;
139 struct iflib_fl;
140 typedef struct iflib_fl *iflib_fl_t;
141 
142 struct iflib_ctx;
143 
144 static void iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid);
145 static void iflib_timer(void *arg);
146 static void iflib_tqg_detach(if_ctx_t ctx);
147 #ifndef ALTQ
148 static int  iflib_simple_transmit(if_t ifp, struct mbuf *m);
149 #endif
150 
151 typedef struct iflib_filter_info {
152 	driver_filter_t *ifi_filter;
153 	void *ifi_filter_arg;
154 	struct grouptask *ifi_task;
155 	void *ifi_ctx;
156 } *iflib_filter_info_t;
157 
158 struct iflib_ctx {
159 	KOBJ_FIELDS;
160 	/*
161 	 * Pointer to hardware driver's softc
162 	 */
163 	void *ifc_softc;
164 	device_t ifc_dev;
165 	if_t ifc_ifp;
166 
167 	cpuset_t ifc_cpus;
168 	if_shared_ctx_t ifc_sctx;
169 	struct if_softc_ctx ifc_softc_ctx;
170 
171 	struct sx ifc_ctx_sx;
172 	struct mtx ifc_state_mtx;
173 
174 	iflib_txq_t ifc_txqs;
175 	iflib_rxq_t ifc_rxqs;
176 	uint32_t ifc_if_flags;
177 	uint32_t ifc_flags;
178 	uint32_t ifc_max_fl_buf_size;
179 	uint32_t ifc_rx_mbuf_sz;
180 
181 	int ifc_link_state;
182 	uint32_t ifc_tx_watchdog_events;
183 	struct cdev *ifc_led_dev;
184 	int ifc_led_state;
185 	struct resource *ifc_msix_mem;
186 
187 	struct if_irq ifc_legacy_irq;
188 	struct task ifc_admin_task;
189 	struct task ifc_led_task;
190 	struct task ifc_vflr_task;
191 	struct taskqueue *ifc_tq;
192 	struct iflib_filter_info ifc_filter_info;
193 	struct ifmedia	ifc_media;
194 	struct ifmedia	*ifc_mediap;
195 
196 	struct sysctl_ctx_list ifc_sysctl_ctx;
197 	struct sysctl_oid *ifc_sysctl_node;
198 	uint16_t ifc_sysctl_ntxqs;
199 	uint16_t ifc_sysctl_nrxqs;
200 	uint16_t ifc_sysctl_qs_eq_override;
201 	uint16_t ifc_sysctl_rx_budget;
202 	uint16_t ifc_sysctl_tx_abdicate;
203 	uint16_t ifc_sysctl_core_offset;
204 #define	CORE_OFFSET_UNSPECIFIED	0xffff
205 	uint8_t  ifc_sysctl_separate_txrx;
206 	uint8_t  ifc_sysctl_use_logical_cores;
207 	uint16_t ifc_sysctl_extra_msix_vectors;
208 	bool     ifc_cpus_are_physical_cores;
209 	bool     ifc_core_offset_ref;
210 	bool     ifc_sysctl_simple_tx;
211 	bool     ifc_sysctl_tx_defer_mfree;
212 	uint16_t ifc_sysctl_tx_reclaim_thresh;
213 	uint16_t ifc_sysctl_tx_reclaim_ticks;
214 
215 	qidx_t ifc_sysctl_ntxds[8];
216 	qidx_t ifc_sysctl_nrxds[8];
217 	struct if_txrx ifc_txrx;
218 #define isc_txd_encap		ifc_txrx.ift_txd_encap
219 #define isc_txd_flush		ifc_txrx.ift_txd_flush
220 #define isc_txd_credits_update	ifc_txrx.ift_txd_credits_update
221 #define isc_rxd_available	ifc_txrx.ift_rxd_available
222 #define isc_rxd_pkt_get		ifc_txrx.ift_rxd_pkt_get
223 #define isc_rxd_refill		ifc_txrx.ift_rxd_refill
224 #define isc_rxd_flush		ifc_txrx.ift_rxd_flush
225 #define isc_legacy_intr		ifc_txrx.ift_legacy_intr
226 #define isc_txq_select		ifc_txrx.ift_txq_select
227 #define isc_txq_select_v2	ifc_txrx.ift_txq_select_v2
228 
229 	eventhandler_tag ifc_vlan_attach_event;
230 	eventhandler_tag ifc_vlan_detach_event;
231 	struct ether_addr ifc_mac;
232 };
233 
234 void *
235 iflib_get_softc(if_ctx_t ctx)
236 {
237 
238 	return (ctx->ifc_softc);
239 }
240 
241 device_t
242 iflib_get_dev(if_ctx_t ctx)
243 {
244 
245 	return (ctx->ifc_dev);
246 }
247 
248 if_t
249 iflib_get_ifp(if_ctx_t ctx)
250 {
251 
252 	return (ctx->ifc_ifp);
253 }
254 
255 struct ifmedia *
256 iflib_get_media(if_ctx_t ctx)
257 {
258 
259 	return (ctx->ifc_mediap);
260 }
261 
262 void
263 iflib_set_mac(if_ctx_t ctx, uint8_t mac[ETHER_ADDR_LEN])
264 {
265 
266 	bcopy(mac, ctx->ifc_mac.octet, ETHER_ADDR_LEN);
267 }
268 
269 if_softc_ctx_t
270 iflib_get_softc_ctx(if_ctx_t ctx)
271 {
272 
273 	return (&ctx->ifc_softc_ctx);
274 }
275 
276 if_shared_ctx_t
277 iflib_get_sctx(if_ctx_t ctx)
278 {
279 
280 	return (ctx->ifc_sctx);
281 }
282 
283 uint16_t
284 iflib_get_extra_msix_vectors_sysctl(if_ctx_t ctx)
285 {
286 
287 	return (ctx->ifc_sysctl_extra_msix_vectors);
288 }
289 
290 #define IP_ALIGNED(m)		((((uintptr_t)(m)->m_data) & 0x3) == 0x2)
291 #define CACHE_PTR_INCREMENT	(CACHE_LINE_SIZE / sizeof(void *))
292 #define CACHE_PTR_NEXT(ptr)	((void *)(roundup2(ptr, CACHE_LINE_SIZE)))
293 
294 #define LINK_ACTIVE(ctx)	((ctx)->ifc_link_state == LINK_STATE_UP)
295 #define CTX_IS_VF(ctx)		((ctx)->ifc_sctx->isc_flags & IFLIB_IS_VF)
296 
297 typedef struct iflib_sw_rx_desc_array {
298 	bus_dmamap_t	*ifsd_map;	/* bus_dma maps for packet */
299 	struct mbuf	**ifsd_m;	/* pkthdr mbufs */
300 	caddr_t		*ifsd_cl;	/* direct cluster pointer for rx */
301 	bus_addr_t	*ifsd_ba;	/* bus addr of cluster for rx */
302 } iflib_rxsd_array_t;
303 
304 typedef struct iflib_sw_tx_desc_array {
305 	bus_dmamap_t	*ifsd_map;	/* bus_dma maps for packet */
306 	bus_dmamap_t	*ifsd_tso_map;	/* bus_dma maps for TSO packet */
307 	struct mbuf	**ifsd_m;	/* pkthdr mbufs */
308 	struct mbuf	**ifsd_m_defer;	/* deferred mbuf ptr */
309 	struct mbuf	**ifsd_m_deferb;/* deferred mbuf backing ptr */
310 } if_txsd_vec_t;
311 
312 /* magic number that should be high enough for any hardware */
313 #define IFLIB_MAX_TX_SEGS		128
314 #define IFLIB_RX_COPY_THRESH		128
315 #define IFLIB_MAX_RX_REFRESH		32
316 /* The minimum descriptors per second before we start coalescing */
317 #define IFLIB_MIN_DESC_SEC		16384
318 #define IFLIB_DEFAULT_TX_UPDATE_FREQ	16
319 /* maximum number of txqs that can share an rx interrupt */
320 #define IFLIB_MAX_TX_SHARED_INTR	4
321 
322 /* this should really scale with ring size - this is a fairly arbitrary value */
323 #define TX_BATCH_SIZE			32
324 
325 #define IFLIB_RESTART_BUDGET		8
326 
327 
328 /*
329  * Encode TSO or !TSO in the low bits of the tx ifsd_m pointer so as
330  * to avoid defref'ing the mbuf to determine the correct busdma resources
331  * to release
332  */
333 #define IFLIB_TSO		(1ULL << 0)
334 #define IFLIB_NO_TSO		(2ULL << 0)
335 #define IFLIB_FLAGS_MASK	(0x3ULL)
336 #define IFLIB_SAVE_MBUF(mbuf, flags)	((void *)(((uintptr_t)mbuf) | flags))
337 #define IFLIB_GET_FLAGS(a)	((uintptr_t)a & IFLIB_FLAGS_MASK)
338 #define IFLIB_GET_MBUF(a)	((struct mbuf *)((uintptr_t)a & ~IFLIB_FLAGS_MASK))
339 
340 
341 #define	IFC_LEGACY		0x001
342 #define	IFC_QFLUSH		0x002
343 #define	IFC_MULTISEG		0x004
344 #define	IFC_INIT_FAILED		0x008
345 #define	IFC_SC_ALLOCATED	0x010
346 #define	IFC_INIT_DONE		0x020
347 #define	IFC_PREFETCH		0x040
348 #define	IFC_DO_RESET		0x080
349 #define	IFC_DO_WATCHDOG		0x100
350 #define	IFC_DO_RESET_IF_UP	0x200
351 #define	IFC_SPARE2		0x400
352 #define	IFC_IN_DETACH		0x800
353 
354 #define	IFC_NETMAP_TX_IRQ	0x80000000
355 
356 #define CSUM_OFFLOAD		(CSUM_IP_TSO | CSUM_IP6_TSO | CSUM_IP | \
357 				 CSUM_IP_UDP | CSUM_IP_TCP | CSUM_IP_SCTP | \
358 				 CSUM_IP6_UDP | CSUM_IP6_TCP | CSUM_IP6_SCTP)
359 
360 struct iflib_txq {
361 	qidx_t		ift_in_use;
362 	qidx_t		ift_cidx;
363 	qidx_t		ift_cidx_processed;
364 	qidx_t		ift_pidx;
365 	uint8_t		ift_gen;
366 	uint8_t		ift_br_offset:1,
367 			ift_defer_mfree:1,
368 			ift_spare_bits0:6;
369 	uint16_t	ift_npending;
370 	uint16_t	ift_db_pending;
371 	uint16_t	ift_rs_pending;
372 	uint32_t	ift_last_reclaim;
373 	uint16_t	ift_reclaim_thresh;
374 	uint16_t	ift_reclaim_ticks;
375 	uint8_t		ift_txd_size[8];
376 	uint64_t	ift_processed;
377 	uint64_t	ift_cleaned;
378 	uint64_t	ift_processed_prev;
379 #if MEMORY_LOGGING
380 	uint64_t	ift_enqueued;
381 	uint64_t	ift_dequeued;
382 #endif
383 	uint64_t	ift_no_tx_dma_setup;
384 	uint64_t	ift_no_desc_avail;
385 	uint64_t	ift_mbuf_defrag_failed;
386 	uint64_t	ift_mbuf_defrag;
387 	uint64_t	ift_map_failed;
388 	uint64_t	ift_txd_encap_efbig;
389 	uint64_t	ift_pullups;
390 	uint64_t	ift_last_timer_tick;
391 
392 	struct mtx	ift_mtx;
393 	struct mtx	ift_db_mtx;
394 
395 	/* constant values */
396 	if_ctx_t	ift_ctx;
397 	struct ifmp_ring	*ift_br;
398 	struct grouptask	ift_task;
399 	qidx_t		ift_size;
400 	qidx_t		ift_pad;
401 	uint16_t	ift_id;
402 	struct callout	ift_timer;
403 #ifdef DEV_NETMAP
404 	struct callout	ift_netmap_timer;
405 #endif /* DEV_NETMAP */
406 
407 	if_txsd_vec_t	ift_sds;
408 	/*
409 	 * TX watchdog state, updated once per iflib_timer period.  The
410 	 * period count saturates instead of wrapping, and is 16 bits
411 	 * wide so that it still reaches any value
412 	 * net.iflib.tx_watchdog_periods is plausibly set to; an 8-bit
413 	 * counter would silently disable the check for a threshold
414 	 * above 255.
415 	 */
416 	qidx_t		ift_outstanding_prev;
417 	uint16_t	ift_wdog_armed;
418 	uint8_t		ift_closed;
419 	uint8_t		ift_update_freq;
420 	uint8_t		ift_spare0[2];	/* pad to the next pointer boundary */
421 	struct iflib_filter_info ift_filter_info;
422 	bus_dma_tag_t	ift_buf_tag;
423 	bus_dma_tag_t	ift_tso_buf_tag;
424 	iflib_dma_info_t	ift_ifdi;
425 #define	MTX_NAME_LEN	32
426 	char		ift_mtx_name[MTX_NAME_LEN];
427 	bus_dma_segment_t	ift_segs[IFLIB_MAX_TX_SEGS]  __aligned(CACHE_LINE_SIZE);
428 #ifdef IFLIB_DIAGNOSTICS
429 	uint64_t ift_cpu_exec_count[256];
430 #endif
431 } __aligned(CACHE_LINE_SIZE);
432 
433 struct iflib_fl {
434 	qidx_t		ifl_cidx;
435 	qidx_t		ifl_pidx;
436 	qidx_t		ifl_credits;
437 	uint8_t		ifl_gen;
438 	uint8_t		ifl_rxd_size;
439 #if MEMORY_LOGGING
440 	uint64_t	ifl_m_enqueued;
441 	uint64_t	ifl_m_dequeued;
442 	uint64_t	ifl_cl_enqueued;
443 	uint64_t	ifl_cl_dequeued;
444 #endif
445 	/* implicit pad */
446 	bitstr_t	*ifl_rx_bitmap;
447 	qidx_t		ifl_fragidx;
448 	/* constant */
449 	qidx_t		ifl_size;
450 	uint16_t	ifl_buf_size;
451 	uint16_t	ifl_cltype;
452 	uma_zone_t	ifl_zone;
453 	iflib_rxsd_array_t	ifl_sds;
454 	iflib_rxq_t	ifl_rxq;
455 	uint8_t		ifl_id;
456 	bus_dma_tag_t	ifl_buf_tag;
457 	iflib_dma_info_t	ifl_ifdi;
458 	uint64_t	ifl_bus_addrs[IFLIB_MAX_RX_REFRESH] __aligned(CACHE_LINE_SIZE);
459 	qidx_t		ifl_rxd_idxs[IFLIB_MAX_RX_REFRESH];
460 }  __aligned(CACHE_LINE_SIZE);
461 
462 static inline qidx_t
463 get_inuse(int size, qidx_t cidx, qidx_t pidx, uint8_t gen)
464 {
465 	qidx_t used;
466 
467 	if (pidx > cidx)
468 		used = pidx - cidx;
469 	else if (pidx < cidx)
470 		used = size - cidx + pidx;
471 	else if (gen == 0 && pidx == cidx)
472 		used = 0;
473 	else if (gen == 1 && pidx == cidx)
474 		used = size;
475 	else
476 		panic("bad state");
477 
478 	return (used);
479 }
480 
481 #define TXQ_AVAIL(txq) ((txq->ift_size - txq->ift_pad) -\
482 	    get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx, txq->ift_gen))
483 
484 #define	MAX_TX_DESC(ctx) MAX((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max, \
485     (ctx)->ifc_softc_ctx.isc_tx_nsegments)
486 
487 #define IDXDIFF(head, tail, wrap) \
488 	((head) >= (tail) ? (head) - (tail) : (wrap) - (tail) + (head))
489 
490 struct iflib_rxq {
491 	if_ctx_t	ifr_ctx;
492 	iflib_fl_t	ifr_fl;
493 	struct pfil_head	*pfil;
494 	/*
495 	 * If there is a separate completion queue (IFLIB_HAS_RXCQ), this is
496 	 * the completion queue consumer index.  Otherwise it's unused.
497 	 */
498 	qidx_t		ifr_cq_cidx;
499 	uint16_t	ifr_id;
500 	uint8_t		ifr_nfl;
501 	uint8_t		ifr_ntxqirq;
502 	uint8_t		ifr_txqid[IFLIB_MAX_TX_SHARED_INTR];
503 	uint8_t		ifr_fl_offset;
504 	struct lro_ctrl		ifr_lc;
505 	struct grouptask	ifr_task;
506 	struct callout		ifr_watchdog;
507 	struct iflib_filter_info ifr_filter_info;
508 	iflib_dma_info_t	ifr_ifdi;
509 
510 	/* dynamically allocate if any drivers need a value substantially larger than this */
511 	struct if_rxd_frag	ifr_frags[IFLIB_MAX_RX_SEGS] __aligned(CACHE_LINE_SIZE);
512 #ifdef IFLIB_DIAGNOSTICS
513 	uint64_t ifr_cpu_exec_count[256];
514 #endif
515 }  __aligned(CACHE_LINE_SIZE);
516 
517 typedef struct if_rxsd {
518 	caddr_t *ifsd_cl;
519 	iflib_fl_t ifsd_fl;
520 } *if_rxsd_t;
521 
522 /*
523  * Only allow a single packet to take up most 1/nth of the tx ring
524  */
525 #define MAX_SINGLE_PACKET_FRACTION 12
526 #define IF_BAD_DMA	((bus_addr_t)-1)
527 
528 #define CTX_ACTIVE(ctx)	((if_getdrvflags((ctx)->ifc_ifp) & IFF_DRV_RUNNING))
529 
530 #define CTX_LOCK_INIT(_sc)	sx_init(&(_sc)->ifc_ctx_sx, "iflib ctx lock")
531 #define CTX_LOCK(ctx)		sx_xlock(&(ctx)->ifc_ctx_sx)
532 #define CTX_UNLOCK(ctx)		sx_xunlock(&(ctx)->ifc_ctx_sx)
533 #define CTX_LOCK_DESTROY(ctx)	sx_destroy(&(ctx)->ifc_ctx_sx)
534 
535 #define STATE_LOCK_INIT(_sc, _name)	mtx_init(&(_sc)->ifc_state_mtx, _name, "iflib state lock", MTX_DEF)
536 #define STATE_LOCK(ctx)		mtx_lock(&(ctx)->ifc_state_mtx)
537 #define STATE_UNLOCK(ctx)	mtx_unlock(&(ctx)->ifc_state_mtx)
538 #define STATE_LOCK_DESTROY(ctx)	mtx_destroy(&(ctx)->ifc_state_mtx)
539 
540 #define CALLOUT_LOCK(txq)	mtx_lock(&txq->ift_mtx)
541 #define CALLOUT_UNLOCK(txq)	mtx_unlock(&txq->ift_mtx)
542 
543 /* Our boot-time initialization hook */
544 static int	iflib_module_event_handler(module_t, int, void *);
545 
546 static moduledata_t iflib_moduledata = {
547 	"iflib",
548 	iflib_module_event_handler,
549 	NULL
550 };
551 
552 DECLARE_MODULE(iflib, iflib_moduledata, SI_SUB_INIT_IF, SI_ORDER_ANY);
553 MODULE_VERSION(iflib, 1);
554 
555 MODULE_DEPEND(iflib, pci, 1, 1, 1);
556 MODULE_DEPEND(iflib, ether, 1, 1, 1);
557 
558 TASKQGROUP_DEFINE(if_io_tqg, mp_ncpus, 1);
559 TASKQGROUP_DEFINE(if_config_tqg, 1, 1);
560 
561 #ifndef IFLIB_DEBUG_COUNTERS
562 #ifdef INVARIANTS
563 #define IFLIB_DEBUG_COUNTERS 1
564 #else
565 #define IFLIB_DEBUG_COUNTERS 0
566 #endif /* !INVARIANTS */
567 #endif
568 
569 static SYSCTL_NODE(_net, OID_AUTO, iflib, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
570     "iflib driver parameters");
571 
572 static SYSCTL_NODE(_debug_fail_point, OID_AUTO, iflib,
573     CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "iflib fail points");
574 
575 static char iflib_register_fail_device[32];
576 SYSCTL_STRING(_debug_fail_point_iflib, OID_AUTO, register_device,
577     CTLFLAG_RW | CTLFLAG_MPSAFE,
578     iflib_register_fail_device, sizeof(iflib_register_fail_device),
579     "device name eligible for registration fail points");
580 
581 static char iflib_admin_task_fail_device[32];
582 SYSCTL_STRING(_debug_fail_point_iflib, OID_AUTO, admin_task_device,
583     CTLFLAG_RW | CTLFLAG_MPSAFE,
584     iflib_admin_task_fail_device, sizeof(iflib_admin_task_fail_device),
585     "device name eligible for admin task fail points");
586 
587 /*
588  * XXX need to ensure that this can't accidentally cause the head to be moved backwards
589  */
590 static int iflib_min_tx_latency = 0;
591 SYSCTL_INT(_net_iflib, OID_AUTO, min_tx_latency, CTLFLAG_RW,
592     &iflib_min_tx_latency, 0,
593     "minimize transmit latency at the possible expense of throughput");
594 static int iflib_no_tx_batch = 0;
595 SYSCTL_INT(_net_iflib, OID_AUTO, no_tx_batch, CTLFLAG_RW,
596     &iflib_no_tx_batch, 0,
597     "minimize transmit latency at the possible expense of throughput");
598 static int iflib_timer_default = 1000;
599 SYSCTL_INT(_net_iflib, OID_AUTO, timer_default, CTLFLAG_RW,
600     &iflib_timer_default, 0, "number of ticks between iflib_timer calls");
601 /*
602  * Consecutive timer periods a TX queue must stay frozen while demand
603  * persists - see iflib_timer(), which defines those states - before the
604  * hardware is asked whether it has completions pending.  Four periods is
605  * roughly two seconds with the default timer interval: a healthy queue on
606  * hardware that coalesces completion reports (e.g. 8254x, TXDCTL.WTHRESH)
607  * stays frozen for at most two (measured on 82541PI), a wedged one until it
608  * is reset.
609  */
610 static int iflib_tx_watchdog_periods = 4;
611 SYSCTL_INT(_net_iflib, OID_AUTO, tx_watchdog_periods, CTLFLAG_RWTUN,
612     &iflib_tx_watchdog_periods, 0,
613     "consecutive frozen timer periods under demand before a TX queue is "
614     "checked for a hang (0 disables the check)");
615 
616 
617 #if IFLIB_DEBUG_COUNTERS
618 
619 static int iflib_tx_seen;
620 static int iflib_tx_sent;
621 static int iflib_tx_encap;
622 static int iflib_rx_allocs;
623 static int iflib_fl_refills;
624 static int iflib_fl_refills_large;
625 static int iflib_tx_frees;
626 
627 SYSCTL_INT(_net_iflib, OID_AUTO, tx_seen, CTLFLAG_RD, &iflib_tx_seen, 0,
628     "# TX mbufs seen");
629 SYSCTL_INT(_net_iflib, OID_AUTO, tx_sent, CTLFLAG_RD, &iflib_tx_sent, 0,
630     "# TX mbufs sent");
631 SYSCTL_INT(_net_iflib, OID_AUTO, tx_encap, CTLFLAG_RD, &iflib_tx_encap, 0,
632     "# TX mbufs encapped");
633 SYSCTL_INT(_net_iflib, OID_AUTO, tx_frees, CTLFLAG_RD, &iflib_tx_frees, 0,
634     "# TX frees");
635 SYSCTL_INT(_net_iflib, OID_AUTO, rx_allocs, CTLFLAG_RD, &iflib_rx_allocs, 0,
636     "# RX allocations");
637 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills, CTLFLAG_RD, &iflib_fl_refills, 0,
638     "# refills");
639 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills_large, CTLFLAG_RD,
640     &iflib_fl_refills_large, 0, "# large refills");
641 
642 static int iflib_txq_drain_flushing;
643 static int iflib_txq_drain_oactive;
644 static int iflib_txq_drain_notready;
645 
646 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_flushing, CTLFLAG_RD,
647     &iflib_txq_drain_flushing, 0, "# drain flushes");
648 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_oactive, CTLFLAG_RD,
649     &iflib_txq_drain_oactive, 0, "# drain oactives");
650 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_notready, CTLFLAG_RD,
651     &iflib_txq_drain_notready, 0, "# drain notready");
652 
653 static int iflib_encap_load_mbuf_fail;
654 static int iflib_encap_pad_mbuf_fail;
655 static int iflib_encap_txq_avail_fail;
656 static int iflib_encap_txd_encap_fail;
657 
658 SYSCTL_INT(_net_iflib, OID_AUTO, encap_load_mbuf_fail, CTLFLAG_RD,
659     &iflib_encap_load_mbuf_fail, 0, "# busdma load failures");
660 SYSCTL_INT(_net_iflib, OID_AUTO, encap_pad_mbuf_fail, CTLFLAG_RD,
661     &iflib_encap_pad_mbuf_fail, 0, "# runt frame pad failures");
662 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txq_avail_fail, CTLFLAG_RD,
663     &iflib_encap_txq_avail_fail, 0, "# txq avail failures");
664 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txd_encap_fail, CTLFLAG_RD,
665     &iflib_encap_txd_encap_fail, 0, "# driver encap failures");
666 
667 static int iflib_task_fn_rxs;
668 static int iflib_rx_intr_enables;
669 static int iflib_fast_intrs;
670 static int iflib_rx_unavail;
671 static int iflib_rx_ctx_inactive;
672 static int iflib_rx_if_input;
673 static int iflib_rxd_flush;
674 
675 static int iflib_verbose_debug;
676 
677 SYSCTL_INT(_net_iflib, OID_AUTO, task_fn_rx, CTLFLAG_RD, &iflib_task_fn_rxs, 0,
678     "# task_fn_rx calls");
679 SYSCTL_INT(_net_iflib, OID_AUTO, rx_intr_enables, CTLFLAG_RD,
680     &iflib_rx_intr_enables, 0, "# RX intr enables");
681 SYSCTL_INT(_net_iflib, OID_AUTO, fast_intrs, CTLFLAG_RD, &iflib_fast_intrs, 0,
682     "# fast_intr calls");
683 SYSCTL_INT(_net_iflib, OID_AUTO, rx_unavail, CTLFLAG_RD, &iflib_rx_unavail, 0,
684     "# times rxeof called with no available data");
685 SYSCTL_INT(_net_iflib, OID_AUTO, rx_ctx_inactive, CTLFLAG_RD,
686     &iflib_rx_ctx_inactive, 0, "# times rxeof called with inactive context");
687 SYSCTL_INT(_net_iflib, OID_AUTO, rx_if_input, CTLFLAG_RD, &iflib_rx_if_input,
688     0, "# times rxeof called if_input");
689 SYSCTL_INT(_net_iflib, OID_AUTO, rxd_flush, CTLFLAG_RD, &iflib_rxd_flush, 0,
690     "# times rxd_flush called");
691 SYSCTL_INT(_net_iflib, OID_AUTO, verbose_debug, CTLFLAG_RW,
692     &iflib_verbose_debug, 0, "enable verbose debugging");
693 
694 #define DBG_COUNTER_INC(name) atomic_add_int(&(iflib_ ## name), 1)
695 static void
696 iflib_debug_reset(void)
697 {
698 	iflib_tx_seen = iflib_tx_sent = iflib_tx_encap = iflib_rx_allocs =
699 		iflib_fl_refills = iflib_fl_refills_large = iflib_tx_frees =
700 		iflib_txq_drain_flushing = iflib_txq_drain_oactive =
701 		iflib_txq_drain_notready =
702 		iflib_encap_load_mbuf_fail = iflib_encap_pad_mbuf_fail =
703 		iflib_encap_txq_avail_fail = iflib_encap_txd_encap_fail =
704 		iflib_task_fn_rxs = iflib_rx_intr_enables = iflib_fast_intrs =
705 		iflib_rx_unavail =
706 		iflib_rx_ctx_inactive = iflib_rx_if_input =
707 		iflib_rxd_flush = 0;
708 }
709 
710 #else
711 #define DBG_COUNTER_INC(name)
712 static void iflib_debug_reset(void) {}
713 #endif
714 
715 #define IFLIB_DEBUG 0
716 
717 static void iflib_tx_structures_free(if_ctx_t ctx);
718 static void iflib_rx_structures_free(if_ctx_t ctx);
719 static int iflib_queues_alloc(if_ctx_t ctx);
720 static int iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq);
721 static int iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget);
722 static int iflib_qset_structures_setup(if_ctx_t ctx);
723 static int iflib_msix_init(if_ctx_t ctx);
724 static int iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filterarg, int *rid, const char *str);
725 static void iflib_txq_check_drain(iflib_txq_t txq, int budget);
726 static uint32_t iflib_txq_can_drain(struct ifmp_ring *);
727 #ifdef ALTQ
728 static void iflib_altq_if_start(if_t ifp);
729 static int iflib_altq_if_transmit(if_t ifp, struct mbuf *m);
730 #endif
731 static void iflib_register(if_ctx_t);
732 static void iflib_deregister(if_ctx_t);
733 static void iflib_unregister_vlan_handlers(if_ctx_t ctx);
734 static uint16_t iflib_get_mbuf_size_for(unsigned int size);
735 static void iflib_init_locked(if_ctx_t ctx);
736 static void iflib_add_device_sysctl_pre(if_ctx_t ctx);
737 static void iflib_add_device_sysctl_post(if_ctx_t ctx);
738 static void iflib_ifmp_purge(iflib_txq_t txq);
739 static void _iflib_pre_assert(if_softc_ctx_t scctx);
740 static void iflib_stop(if_ctx_t ctx);
741 static void iflib_if_init_locked(if_ctx_t ctx);
742 static void iflib_free_intr_mem(if_ctx_t ctx);
743 #ifndef __NO_STRICT_ALIGNMENT
744 static struct mbuf *iflib_fixup_rx(struct mbuf *m);
745 #endif
746 static __inline int iflib_completed_tx_reclaim(iflib_txq_t txq,
747     struct mbuf **m_defer);
748 static __inline void iflib_completed_tx_reclaim_force(iflib_txq_t txq);
749 
750 static SLIST_HEAD(cpu_offset_list, cpu_offset) cpu_offsets =
751     SLIST_HEAD_INITIALIZER(cpu_offsets);
752 struct cpu_offset {
753 	SLIST_ENTRY(cpu_offset) entries;
754 	cpuset_t	set;
755 	unsigned int	refcount;
756 	uint16_t	next_cpuid;
757 };
758 static struct mtx cpu_offset_mtx;
759 MTX_SYSINIT(iflib_cpu_offset, &cpu_offset_mtx, "iflib_cpu_offset lock",
760     MTX_DEF);
761 
762 DEBUGNET_DEFINE(iflib);
763 
764 static int
765 iflib_num_rx_descs(if_ctx_t ctx)
766 {
767 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
768 	if_shared_ctx_t sctx = ctx->ifc_sctx;
769 	uint16_t first_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
770 
771 	return (scctx->isc_nrxd[first_rxq]);
772 }
773 
774 static int
775 iflib_num_tx_descs(if_ctx_t ctx)
776 {
777 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
778 	if_shared_ctx_t sctx = ctx->ifc_sctx;
779 	uint16_t first_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
780 
781 	return (scctx->isc_ntxd[first_txq]);
782 }
783 
784 #ifdef DEV_NETMAP
785 #include <sys/selinfo.h>
786 #include <net/netmap.h>
787 #include <dev/netmap/netmap_kern.h>
788 
789 MODULE_DEPEND(iflib, netmap, 1, 1, 1);
790 
791 static int netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, bool init);
792 static void iflib_netmap_timer(void *arg);
793 
794 /*
795  * device-specific sysctl variables:
796  *
797  * iflib_crcstrip: 0: keep CRC in rx frames (default), 1: strip it.
798  *	During regular operations the CRC is stripped, but on some
799  *	hardware reception of frames not multiple of 64 is slower,
800  *	so using crcstrip=0 helps in benchmarks.
801  *
802  * iflib_rx_miss, iflib_rx_miss_bufs:
803  *	count packets that might be missed due to lost interrupts.
804  */
805 SYSCTL_DECL(_dev_netmap);
806 /*
807  * The xl driver by default strips CRCs and we do not override it.
808  */
809 
810 int iflib_crcstrip = 1;
811 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_crcstrip,
812     CTLFLAG_RW, &iflib_crcstrip, 1, "strip CRC on RX frames");
813 
814 int iflib_rx_miss, iflib_rx_miss_bufs;
815 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss,
816     CTLFLAG_RW, &iflib_rx_miss, 0, "potentially missed RX intr");
817 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss_bufs,
818     CTLFLAG_RW, &iflib_rx_miss_bufs, 0, "potentially missed RX intr bufs");
819 
820 /*
821  * Register/unregister. We are already under netmap lock.
822  * Only called on the first register or the last unregister.
823  */
824 static int
825 iflib_netmap_register(struct netmap_adapter *na, int onoff)
826 {
827 	if_t ifp = na->ifp;
828 	if_ctx_t ctx = if_getsoftc(ifp);
829 	int status;
830 
831 	CTX_LOCK(ctx);
832 	if (!CTX_IS_VF(ctx))
833 		IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip);
834 
835 	iflib_stop(ctx);
836 
837 	/*
838 	 * Enable (or disable) netmap flags, and intercept (or restore)
839 	 * ifp->if_transmit. This is done once the device has been stopped
840 	 * to prevent race conditions. Also, this must be done after
841 	 * calling netmap_disable_all_rings() and before calling
842 	 * netmap_enable_all_rings(), so that these two functions see the
843 	 * updated state of the NAF_NETMAP_ON bit.
844 	 */
845 	if (onoff) {
846 		nm_set_native_flags(na);
847 	} else {
848 		nm_clear_native_flags(na);
849 	}
850 
851 	iflib_init_locked(ctx);
852 	IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip); // XXX why twice ?
853 	status = if_getdrvflags(ifp) & IFF_DRV_RUNNING ? 0 : 1;
854 	if (status)
855 		nm_clear_native_flags(na);
856 	CTX_UNLOCK(ctx);
857 	return (status);
858 }
859 
860 static int
861 iflib_netmap_config(struct netmap_adapter *na, struct nm_config_info *info)
862 {
863 	if_t ifp = na->ifp;
864 	if_ctx_t ctx = if_getsoftc(ifp);
865 	iflib_rxq_t rxq = &ctx->ifc_rxqs[0];
866 	iflib_fl_t fl = &rxq->ifr_fl[0];
867 
868 	info->num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
869 	info->num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
870 	info->num_tx_descs = iflib_num_tx_descs(ctx);
871 	info->num_rx_descs = iflib_num_rx_descs(ctx);
872 	info->rx_buf_maxsize = fl->ifl_buf_size;
873 	nm_prinf("txr %u rxr %u txd %u rxd %u rbufsz %u",
874 		info->num_tx_rings, info->num_rx_rings, info->num_tx_descs,
875 		info->num_rx_descs, info->rx_buf_maxsize);
876 
877 	return (0);
878 }
879 
880 static int
881 netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, bool init)
882 {
883 	struct netmap_adapter *na = kring->na;
884 	u_int const lim = kring->nkr_num_slots - 1;
885 	struct netmap_ring *ring = kring->ring;
886 	bus_dmamap_t *map;
887 	struct if_rxd_update iru;
888 	if_ctx_t ctx = rxq->ifr_ctx;
889 	iflib_fl_t fl = &rxq->ifr_fl[0];
890 	u_int nic_i_first, nic_i;
891 	u_int nm_i;
892 	int i, n;
893 #if IFLIB_DEBUG_COUNTERS
894 	int rf_count = 0;
895 #endif
896 
897 	/*
898 	 * This function is used both at initialization and in rxsync.
899 	 * At initialization we need to prepare (with isc_rxd_refill())
900 	 * all the netmap buffers currently owned by the kernel, in
901 	 * such a way to keep fl->ifl_pidx and kring->nr_hwcur in sync
902 	 * (except for kring->nkr_hwofs). These may be less than
903 	 * kring->nkr_num_slots if netmap_reset() was called while
904 	 * an application using the kring that still owned some
905 	 * buffers.
906 	 * At rxsync time, both indexes point to the next buffer to be
907 	 * refilled.
908 	 * In any case we publish (with isc_rxd_flush()) up to
909 	 * (fl->ifl_pidx - 1) % N (included), to avoid the NIC tail/prod
910 	 * pointer to overrun the head/cons pointer, although this is
911 	 * not necessary for some NICs (e.g. vmx).
912 	 */
913 	if (__predict_false(init)) {
914 		n = kring->nkr_num_slots - nm_kr_rxspace(kring);
915 	} else {
916 		n = kring->rhead - kring->nr_hwcur;
917 		if (n == 0)
918 			return (0); /* Nothing to do. */
919 		if (n < 0)
920 			n += kring->nkr_num_slots;
921 	}
922 
923 	iru_init(&iru, rxq, 0 /* flid */);
924 	map = fl->ifl_sds.ifsd_map;
925 	nic_i = fl->ifl_pidx;
926 	nm_i = netmap_idx_n2k(kring, nic_i);
927 	if (__predict_false(init)) {
928 		/*
929 		 * On init/reset, nic_i must be 0, and we must
930 		 * start to refill from hwtail (see netmap_reset()).
931 		 */
932 		MPASS(nic_i == 0);
933 		MPASS(nm_i == kring->nr_hwtail);
934 	} else
935 		MPASS(nm_i == kring->nr_hwcur);
936 	DBG_COUNTER_INC(fl_refills);
937 	while (n > 0) {
938 #if IFLIB_DEBUG_COUNTERS
939 		if (++rf_count == 9)
940 			DBG_COUNTER_INC(fl_refills_large);
941 #endif
942 		nic_i_first = nic_i;
943 		for (i = 0; n > 0 && i < IFLIB_MAX_RX_REFRESH; n--, i++) {
944 			struct netmap_slot *slot = &ring->slot[nm_i];
945 			uint64_t paddr;
946 			void *addr = PNMB(na, slot, &paddr);
947 
948 			MPASS(i < IFLIB_MAX_RX_REFRESH);
949 
950 			if (addr == NETMAP_BUF_BASE(na)) /* bad buf */
951 				return (netmap_ring_reinit(kring));
952 
953 			fl->ifl_bus_addrs[i] = paddr +
954 			    nm_get_offset(kring, slot);
955 			fl->ifl_rxd_idxs[i] = nic_i;
956 
957 			if (__predict_false(init)) {
958 				netmap_load_map(na, fl->ifl_buf_tag,
959 				    map[nic_i], addr);
960 			} else if (slot->flags & NS_BUF_CHANGED) {
961 				/* buffer has changed, reload map */
962 				netmap_reload_map(na, fl->ifl_buf_tag,
963 				    map[nic_i], addr);
964 			}
965 			bus_dmamap_sync(fl->ifl_buf_tag, map[nic_i],
966 			    BUS_DMASYNC_PREREAD);
967 			slot->flags &= ~NS_BUF_CHANGED;
968 
969 			nm_i = nm_next(nm_i, lim);
970 			nic_i = nm_next(nic_i, lim);
971 		}
972 
973 		iru.iru_pidx = nic_i_first;
974 		iru.iru_count = i;
975 		ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
976 	}
977 	fl->ifl_pidx = nic_i;
978 	/*
979 	 * At the end of the loop we must have refilled everything
980 	 * we could possibly refill.
981 	 */
982 	MPASS(nm_i == kring->rhead);
983 	kring->nr_hwcur = nm_i;
984 
985 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
986 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
987 	ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id,
988 	    nm_prev(nic_i, lim));
989 	DBG_COUNTER_INC(rxd_flush);
990 
991 	return (0);
992 }
993 
994 #define NETMAP_TX_TIMER_US	90
995 
996 /*
997  * Reconcile kernel and user view of the transmit ring.
998  *
999  * All information is in the kring.
1000  * Userspace wants to send packets up to the one before kring->rhead,
1001  * kernel knows kring->nr_hwcur is the first unsent packet.
1002  *
1003  * Here we push packets out (as many as possible), and possibly
1004  * reclaim buffers from previously completed transmission.
1005  *
1006  * The caller (netmap) guarantees that there is only one instance
1007  * running at any time. Any interference with other driver
1008  * methods should be handled by the individual drivers.
1009  */
1010 static int
1011 iflib_netmap_txsync(struct netmap_kring *kring, int flags)
1012 {
1013 	struct netmap_adapter *na = kring->na;
1014 	if_t ifp = na->ifp;
1015 	struct netmap_ring *ring = kring->ring;
1016 	u_int nm_i;	/* index into the netmap kring */
1017 	u_int nic_i;	/* index into the NIC ring */
1018 	u_int const lim = kring->nkr_num_slots - 1;
1019 	u_int const head = kring->rhead;
1020 	struct if_pkt_info pi;
1021 	int tx_pkts = 0, tx_bytes = 0;
1022 
1023 	/*
1024 	 * interrupts on every tx packet are expensive so request
1025 	 * them every half ring, or where NS_REPORT is set
1026 	 */
1027 	u_int report_frequency = kring->nkr_num_slots >> 1;
1028 	/* device-specific */
1029 	if_ctx_t ctx = if_getsoftc(ifp);
1030 	iflib_txq_t txq = &ctx->ifc_txqs[kring->ring_id];
1031 
1032 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1033 	    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1034 
1035 	/*
1036 	 * First part: process new packets to send.
1037 	 * nm_i is the current index in the netmap kring,
1038 	 * nic_i is the corresponding index in the NIC ring.
1039 	 *
1040 	 * If we have packets to send (nm_i != head)
1041 	 * iterate over the netmap ring, fetch length and update
1042 	 * the corresponding slot in the NIC ring. Some drivers also
1043 	 * need to update the buffer's physical address in the NIC slot
1044 	 * even NS_BUF_CHANGED is not set (PNMB computes the addresses).
1045 	 *
1046 	 * The netmap_reload_map() calls is especially expensive,
1047 	 * even when (as in this case) the tag is 0, so do only
1048 	 * when the buffer has actually changed.
1049 	 *
1050 	 * If possible do not set the report/intr bit on all slots,
1051 	 * but only a few times per ring or when NS_REPORT is set.
1052 	 *
1053 	 * Finally, on 10G and faster drivers, it might be useful
1054 	 * to prefetch the next slot and txr entry.
1055 	 */
1056 
1057 	nm_i = kring->nr_hwcur;
1058 	if (nm_i != head) {	/* we have new packets to send */
1059 		uint32_t pkt_len = 0, seg_idx = 0;
1060 		int nic_i_start = -1, flags = 0;
1061 		memset(&pi, 0, sizeof(pi));
1062 		pi.ipi_segs = txq->ift_segs;
1063 		pi.ipi_qsidx = kring->ring_id;
1064 		nic_i = netmap_idx_k2n(kring, nm_i);
1065 
1066 		__builtin_prefetch(&ring->slot[nm_i]);
1067 		__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i]);
1068 		__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i]);
1069 
1070 		while (nm_i != head) {
1071 			struct netmap_slot *slot = &ring->slot[nm_i];
1072 			uint64_t offset = nm_get_offset(kring, slot);
1073 			u_int len = slot->len;
1074 			uint64_t paddr;
1075 			void *addr = PNMB(na, slot, &paddr);
1076 
1077 			flags |= (slot->flags & NS_REPORT ||
1078 				nic_i == 0 || nic_i == report_frequency) ?
1079 				IPI_TX_INTR : 0;
1080 
1081 			/*
1082 			 * If this is the first packet fragment, save the
1083 			 * index of the first NIC slot for later.
1084 			 */
1085 			if (nic_i_start < 0)
1086 				nic_i_start = nic_i;
1087 
1088 			pi.ipi_segs[seg_idx].ds_addr = paddr + offset;
1089 			pi.ipi_segs[seg_idx].ds_len = len;
1090 			if (len) {
1091 				pkt_len += len;
1092 				seg_idx++;
1093 			}
1094 
1095 			if (!(slot->flags & NS_MOREFRAG)) {
1096 				pi.ipi_len = pkt_len;
1097 				pi.ipi_nsegs = seg_idx;
1098 				pi.ipi_pidx = nic_i_start;
1099 				pi.ipi_ndescs = 0;
1100 				pi.ipi_flags = flags;
1101 
1102 				/* Prepare the NIC TX ring. */
1103 				ctx->isc_txd_encap(ctx->ifc_softc, &pi);
1104 				DBG_COUNTER_INC(tx_encap);
1105 
1106 				/* Update transmit counters */
1107 				tx_bytes += pi.ipi_len;
1108 				tx_pkts++;
1109 
1110 				/* Reinit per-packet info for the next one. */
1111 				flags = seg_idx = pkt_len = 0;
1112 				nic_i_start = -1;
1113 			}
1114 
1115 			/* prefetch for next round */
1116 			__builtin_prefetch(&ring->slot[nm_i + 1]);
1117 			__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i + 1]);
1118 			__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i + 1]);
1119 
1120 			NM_CHECK_ADDR_LEN_OFF(na, len, offset);
1121 
1122 			if (slot->flags & NS_BUF_CHANGED) {
1123 				/* buffer has changed, reload map */
1124 				netmap_reload_map(na, txq->ift_buf_tag,
1125 				    txq->ift_sds.ifsd_map[nic_i], addr);
1126 			}
1127 			/* make sure changes to the buffer are synced */
1128 			bus_dmamap_sync(txq->ift_buf_tag,
1129 			    txq->ift_sds.ifsd_map[nic_i],
1130 			    BUS_DMASYNC_PREWRITE);
1131 
1132 			slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED | NS_MOREFRAG);
1133 			nm_i = nm_next(nm_i, lim);
1134 			nic_i = nm_next(nic_i, lim);
1135 		}
1136 		kring->nr_hwcur = nm_i;
1137 
1138 		/* synchronize the NIC ring */
1139 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1140 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1141 
1142 		/* (re)start the tx unit up to slot nic_i (excluded) */
1143 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, nic_i);
1144 	}
1145 
1146 	/*
1147 	 * Second part: reclaim buffers for completed transmissions.
1148 	 *
1149 	 * If there are unclaimed buffers, attempt to reclaim them.
1150 	 * If we don't manage to reclaim them all, and TX IRQs are not in use,
1151 	 * trigger a per-tx-queue timer to try again later.
1152 	 */
1153 	if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1154 		if (iflib_tx_credits_update(ctx, txq)) {
1155 			/* some tx completed, increment avail */
1156 			nic_i = txq->ift_cidx_processed;
1157 			kring->nr_hwtail = nm_prev(netmap_idx_n2k(kring, nic_i), lim);
1158 		}
1159 	}
1160 
1161 	if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ))
1162 		if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1163 			callout_reset_sbt_on(&txq->ift_netmap_timer,
1164 			    NETMAP_TX_TIMER_US * SBT_1US, SBT_1US,
1165 			    iflib_netmap_timer, txq,
1166 			    txq->ift_netmap_timer.c_cpu, 0);
1167 		}
1168 
1169 	if_inc_counter(ifp, IFCOUNTER_OBYTES, tx_bytes);
1170 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, tx_pkts);
1171 
1172 	return (0);
1173 }
1174 
1175 /*
1176  * Reconcile kernel and user view of the receive ring.
1177  * Same as for the txsync, this routine must be efficient.
1178  * The caller guarantees a single invocations, but races against
1179  * the rest of the driver should be handled here.
1180  *
1181  * On call, kring->rhead is the first packet that userspace wants
1182  * to keep, and kring->rcur is the wakeup point.
1183  * The kernel has previously reported packets up to kring->rtail.
1184  *
1185  * If (flags & NAF_FORCE_READ) also check for incoming packets irrespective
1186  * of whether or not we received an interrupt.
1187  */
1188 static int
1189 iflib_netmap_rxsync(struct netmap_kring *kring, int flags)
1190 {
1191 	struct netmap_adapter *na = kring->na;
1192 	struct netmap_ring *ring = kring->ring;
1193 	if_t ifp = na->ifp;
1194 	uint32_t nm_i;	/* index into the netmap ring */
1195 	uint32_t nic_i;	/* index into the NIC ring */
1196 	u_int n;
1197 	u_int const lim = kring->nkr_num_slots - 1;
1198 	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
1199 	int i = 0, rx_bytes = 0, rx_pkts = 0;
1200 
1201 	if_ctx_t ctx = if_getsoftc(ifp);
1202 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1203 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1204 	iflib_rxq_t rxq = &ctx->ifc_rxqs[kring->ring_id];
1205 	iflib_fl_t fl = &rxq->ifr_fl[0];
1206 	struct if_rxd_info ri;
1207 	qidx_t *cidxp;
1208 
1209 	/*
1210 	 * netmap only uses free list 0, to avoid out of order consumption
1211 	 * of receive buffers
1212 	 */
1213 
1214 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
1215 	    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1216 
1217 	/*
1218 	 * First part: import newly received packets.
1219 	 *
1220 	 * nm_i is the index of the next free slot in the netmap ring,
1221 	 * nic_i is the index of the next received packet in the NIC ring
1222 	 * (or in the free list 0 if IFLIB_HAS_RXCQ is set), and they may
1223 	 * differ in case if_init() has been called while
1224 	 * in netmap mode. For the receive ring we have
1225 	 *
1226 	 *	nic_i = fl->ifl_cidx;
1227 	 *	nm_i = kring->nr_hwtail (previous)
1228 	 * and
1229 	 *	nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1230 	 *
1231 	 * fl->ifl_cidx is set to 0 on a ring reinit
1232 	 */
1233 	if (netmap_no_pendintr || force_update) {
1234 		uint32_t hwtail_lim = nm_prev(kring->nr_hwcur, lim);
1235 		bool have_rxcq = sctx->isc_flags & IFLIB_HAS_RXCQ;
1236 		int crclen = iflib_crcstrip ? 0 : 4;
1237 		int error, avail;
1238 
1239 		/*
1240 		 * For the free list consumer index, we use the same
1241 		 * logic as in iflib_rxeof().
1242 		 */
1243 		if (have_rxcq)
1244 			cidxp = &rxq->ifr_cq_cidx;
1245 		else
1246 			cidxp = &fl->ifl_cidx;
1247 		avail = ctx->isc_rxd_available(ctx->ifc_softc,
1248 		    rxq->ifr_id, *cidxp, USHRT_MAX);
1249 
1250 		nic_i = fl->ifl_cidx;
1251 		nm_i = netmap_idx_n2k(kring, nic_i);
1252 		MPASS(nm_i == kring->nr_hwtail);
1253 		for (n = 0; avail > 0 && nm_i != hwtail_lim; n++, avail--) {
1254 			memset(&ri, 0, sizeof(ri));
1255 			ri.iri_frags = rxq->ifr_frags;
1256 			ri.iri_qsidx = kring->ring_id;
1257 			ri.iri_ifp = ctx->ifc_ifp;
1258 			ri.iri_cidx = *cidxp;
1259 
1260 			error = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
1261 			for (i = 0; i < ri.iri_nfrags; i++) {
1262 				if (error) {
1263 					ring->slot[nm_i].len = 0;
1264 					ring->slot[nm_i].flags = 0;
1265 				} else {
1266 					ring->slot[nm_i].len = ri.iri_frags[i].irf_len;
1267 					if (i == (ri.iri_nfrags - 1)) {
1268 						ring->slot[nm_i].len -= crclen;
1269 						ring->slot[nm_i].flags = 0;
1270 
1271 						/* Update receive counters */
1272 						rx_bytes += ri.iri_len;
1273 						rx_pkts++;
1274 					} else
1275 						ring->slot[nm_i].flags = NS_MOREFRAG;
1276 				}
1277 
1278 				bus_dmamap_sync(fl->ifl_buf_tag,
1279 				    fl->ifl_sds.ifsd_map[nic_i], BUS_DMASYNC_POSTREAD);
1280 				nm_i = nm_next(nm_i, lim);
1281 				fl->ifl_cidx = nic_i = nm_next(nic_i, lim);
1282 			}
1283 
1284 			if (have_rxcq) {
1285 				*cidxp = ri.iri_cidx;
1286 				while (*cidxp >= scctx->isc_nrxd[0])
1287 					*cidxp -= scctx->isc_nrxd[0];
1288 			}
1289 
1290 		}
1291 		if (n) { /* update the state variables */
1292 			if (netmap_no_pendintr && !force_update) {
1293 				/* diagnostics */
1294 				iflib_rx_miss++;
1295 				iflib_rx_miss_bufs += n;
1296 			}
1297 			kring->nr_hwtail = nm_i;
1298 		}
1299 		kring->nr_kflags &= ~NKR_PENDINTR;
1300 	}
1301 	/*
1302 	 * Second part: skip past packets that userspace has released.
1303 	 * (kring->nr_hwcur to head excluded),
1304 	 * and make the buffers available for reception.
1305 	 * As usual nm_i is the index in the netmap ring,
1306 	 * nic_i is the index in the NIC ring, and
1307 	 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1308 	 */
1309 	netmap_fl_refill(rxq, kring, false);
1310 
1311 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
1312 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
1313 
1314 	return (0);
1315 }
1316 
1317 static void
1318 iflib_netmap_intr(struct netmap_adapter *na, int onoff)
1319 {
1320 	if_ctx_t ctx = if_getsoftc(na->ifp);
1321 
1322 	CTX_LOCK(ctx);
1323 	if (onoff) {
1324 		IFDI_INTR_ENABLE(ctx);
1325 	} else {
1326 		IFDI_INTR_DISABLE(ctx);
1327 	}
1328 	CTX_UNLOCK(ctx);
1329 }
1330 
1331 static int
1332 iflib_netmap_attach(if_ctx_t ctx)
1333 {
1334 	struct netmap_adapter na;
1335 
1336 	bzero(&na, sizeof(na));
1337 
1338 	na.ifp = ctx->ifc_ifp;
1339 	na.na_flags = NAF_BDG_MAYSLEEP | NAF_MOREFRAG | NAF_OFFSETS;
1340 	MPASS(ctx->ifc_softc_ctx.isc_ntxqsets);
1341 	MPASS(ctx->ifc_softc_ctx.isc_nrxqsets);
1342 
1343 	na.num_tx_desc = iflib_num_tx_descs(ctx);
1344 	na.num_rx_desc = iflib_num_rx_descs(ctx);
1345 	na.nm_txsync = iflib_netmap_txsync;
1346 	na.nm_rxsync = iflib_netmap_rxsync;
1347 	na.nm_register = iflib_netmap_register;
1348 	na.nm_intr = iflib_netmap_intr;
1349 	na.nm_config = iflib_netmap_config;
1350 	na.num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
1351 	na.num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
1352 	return (netmap_attach(&na));
1353 }
1354 
1355 static int
1356 iflib_netmap_txq_init(if_ctx_t ctx, iflib_txq_t txq)
1357 {
1358 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1359 	struct netmap_slot *slot;
1360 
1361 	slot = netmap_reset(na, NR_TX, txq->ift_id, 0);
1362 	if (slot == NULL)
1363 		return (0);
1364 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxd[0]; i++) {
1365 		/*
1366 		 * In netmap mode, set the map for the packet buffer.
1367 		 * NOTE: Some drivers (not this one) also need to set
1368 		 * the physical buffer address in the NIC ring.
1369 		 * netmap_idx_n2k() maps a nic index, i, into the corresponding
1370 		 * netmap slot index, si
1371 		 */
1372 		int si = netmap_idx_n2k(na->tx_rings[txq->ift_id], i);
1373 		netmap_load_map(na, txq->ift_buf_tag, txq->ift_sds.ifsd_map[i],
1374 		    NMB(na, slot + si));
1375 	}
1376 	return (1);
1377 }
1378 
1379 static int
1380 iflib_netmap_rxq_init(if_ctx_t ctx, iflib_rxq_t rxq)
1381 {
1382 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1383 	struct netmap_kring *kring;
1384 	struct netmap_slot *slot;
1385 
1386 	slot = netmap_reset(na, NR_RX, rxq->ifr_id, 0);
1387 	if (slot == NULL)
1388 		return (0);
1389 	kring = na->rx_rings[rxq->ifr_id];
1390 	netmap_fl_refill(rxq, kring, true);
1391 	return (1);
1392 }
1393 
1394 static void
1395 iflib_netmap_timer(void *arg)
1396 {
1397 	iflib_txq_t txq = arg;
1398 	if_ctx_t ctx = txq->ift_ctx;
1399 
1400 	/*
1401 	 * Wake up the netmap application, to give it a chance to
1402 	 * call txsync and reclaim more completed TX buffers.
1403 	 */
1404 	netmap_tx_irq(ctx->ifc_ifp, txq->ift_id);
1405 }
1406 
1407 #define iflib_netmap_detach(ifp) netmap_detach(ifp)
1408 
1409 #else
1410 #define iflib_netmap_txq_init(ctx, txq) (0)
1411 #define iflib_netmap_rxq_init(ctx, rxq) (0)
1412 #define iflib_netmap_detach(ifp)
1413 #define netmap_enable_all_rings(ifp)
1414 #define netmap_disable_all_rings(ifp)
1415 
1416 #define iflib_netmap_attach(ctx) (0)
1417 #define netmap_rx_irq(ifp, qid, budget) (0)
1418 #endif
1419 
1420 #if defined(__i386__) || defined(__amd64__)
1421 static __inline void
1422 prefetch(void *x)
1423 {
1424 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1425 }
1426 
1427 static __inline void
1428 prefetch2cachelines(void *x)
1429 {
1430 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1431 #if (CACHE_LINE_SIZE < 128)
1432 	__asm volatile("prefetcht0 %0" :: "m" (*(((unsigned long *)x) + CACHE_LINE_SIZE / (sizeof(unsigned long)))));
1433 #endif
1434 }
1435 #else
1436 static __inline void
1437 prefetch(void *x)
1438 {
1439 }
1440 
1441 static __inline void
1442 prefetch2cachelines(void *x)
1443 {
1444 }
1445 #endif
1446 
1447 static void
1448 iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid)
1449 {
1450 	iflib_fl_t fl;
1451 
1452 	fl = &rxq->ifr_fl[flid];
1453 	iru->iru_paddrs = fl->ifl_bus_addrs;
1454 	iru->iru_idxs = fl->ifl_rxd_idxs;
1455 	iru->iru_qsidx = rxq->ifr_id;
1456 	iru->iru_buf_size = fl->ifl_buf_size;
1457 	iru->iru_flidx = fl->ifl_id;
1458 }
1459 
1460 static void
1461 _iflib_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int err)
1462 {
1463 	if (err)
1464 		return;
1465 	*(bus_addr_t *) arg = segs[0].ds_addr;
1466 }
1467 
1468 #define	DMA_WIDTH_TO_BUS_LOWADDR(width)				\
1469 	(((width) == 0) || (width) == flsll(BUS_SPACE_MAXADDR) ?	\
1470 	    BUS_SPACE_MAXADDR : (1ULL << (width)) - 1ULL)
1471 
1472 int
1473 iflib_dma_alloc_align(if_ctx_t ctx, int size, int align, iflib_dma_info_t dma, int mapflags)
1474 {
1475 	int err;
1476 	device_t dev = ctx->ifc_dev;
1477 	bus_addr_t lowaddr;
1478 
1479 	lowaddr = DMA_WIDTH_TO_BUS_LOWADDR(ctx->ifc_softc_ctx.isc_dma_width);
1480 
1481 	err = bus_dma_tag_create(bus_get_dma_tag(dev),	/* parent */
1482 		    align, 0,		/* alignment, bounds */
1483 		    lowaddr,		/* lowaddr */
1484 		    BUS_SPACE_MAXADDR,	/* highaddr */
1485 		    NULL, NULL,		/* filter, filterarg */
1486 		    size,		/* maxsize */
1487 		    1,			/* nsegments */
1488 		    size,		/* maxsegsize */
1489 		    BUS_DMA_ALLOCNOW,	/* flags */
1490 		    NULL,		/* lockfunc */
1491 		    NULL,		/* lockarg */
1492 		    &dma->idi_tag);
1493 	if (err) {
1494 		device_printf(dev,
1495 		    "%s: bus_dma_tag_create failed: %d (size=%d, align=%d)\n",
1496 		    __func__, err, size, align);
1497 		goto fail_0;
1498 	}
1499 
1500 	err = bus_dmamem_alloc(dma->idi_tag, (void **)&dma->idi_vaddr,
1501 	    BUS_DMA_NOWAIT | BUS_DMA_COHERENT | BUS_DMA_ZERO, &dma->idi_map);
1502 	if (err) {
1503 		device_printf(dev,
1504 		    "%s: bus_dmamem_alloc(%ju) failed: %d\n",
1505 		    __func__, (uintmax_t)size, err);
1506 		goto fail_1;
1507 	}
1508 
1509 	dma->idi_paddr = IF_BAD_DMA;
1510 	err = bus_dmamap_load(dma->idi_tag, dma->idi_map, dma->idi_vaddr,
1511 	    size, _iflib_dmamap_cb, &dma->idi_paddr, mapflags | BUS_DMA_NOWAIT);
1512 	if (err || dma->idi_paddr == IF_BAD_DMA) {
1513 		device_printf(dev,
1514 		    "%s: bus_dmamap_load failed: %d\n",
1515 		    __func__, err);
1516 		goto fail_2;
1517 	}
1518 
1519 	dma->idi_size = size;
1520 	return (0);
1521 
1522 fail_2:
1523 	bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1524 fail_1:
1525 	bus_dma_tag_destroy(dma->idi_tag);
1526 fail_0:
1527 	dma->idi_tag = NULL;
1528 
1529 	return (err);
1530 }
1531 
1532 int
1533 iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags)
1534 {
1535 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1536 
1537 	KASSERT(sctx->isc_q_align != 0, ("alignment value not initialized"));
1538 
1539 	return (iflib_dma_alloc_align(ctx, size, sctx->isc_q_align, dma, mapflags));
1540 }
1541 
1542 int
1543 iflib_dma_alloc_multi(if_ctx_t ctx, int *sizes, iflib_dma_info_t *dmalist, int mapflags, int count)
1544 {
1545 	int i, err;
1546 	iflib_dma_info_t *dmaiter;
1547 
1548 	dmaiter = dmalist;
1549 	for (i = 0; i < count; i++, dmaiter++) {
1550 		if ((err = iflib_dma_alloc(ctx, sizes[i], *dmaiter, mapflags)) != 0)
1551 			break;
1552 	}
1553 	if (err)
1554 		iflib_dma_free_multi(dmalist, i);
1555 	return (err);
1556 }
1557 
1558 void
1559 iflib_dma_free(iflib_dma_info_t dma)
1560 {
1561 	if (dma->idi_tag == NULL)
1562 		return;
1563 	if (dma->idi_paddr != IF_BAD_DMA) {
1564 		bus_dmamap_sync(dma->idi_tag, dma->idi_map,
1565 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1566 		bus_dmamap_unload(dma->idi_tag, dma->idi_map);
1567 		dma->idi_paddr = IF_BAD_DMA;
1568 	}
1569 	if (dma->idi_vaddr != NULL) {
1570 		bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1571 		dma->idi_vaddr = NULL;
1572 	}
1573 	bus_dma_tag_destroy(dma->idi_tag);
1574 	dma->idi_tag = NULL;
1575 }
1576 
1577 void
1578 iflib_dma_free_multi(iflib_dma_info_t *dmalist, int count)
1579 {
1580 	int i;
1581 	iflib_dma_info_t *dmaiter = dmalist;
1582 
1583 	for (i = 0; i < count; i++, dmaiter++)
1584 		iflib_dma_free(*dmaiter);
1585 }
1586 
1587 static int
1588 iflib_fast_intr(void *arg)
1589 {
1590 	iflib_filter_info_t info = arg;
1591 	struct grouptask *gtask = info->ifi_task;
1592 	int result;
1593 
1594 	DBG_COUNTER_INC(fast_intrs);
1595 	if (info->ifi_filter != NULL) {
1596 		result = info->ifi_filter(info->ifi_filter_arg);
1597 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1598 			return (result);
1599 	}
1600 
1601 	GROUPTASK_ENQUEUE(gtask);
1602 	return (FILTER_HANDLED);
1603 }
1604 
1605 static int
1606 iflib_fast_intr_rxtx(void *arg)
1607 {
1608 	iflib_filter_info_t info = arg;
1609 	struct grouptask *gtask = info->ifi_task;
1610 	if_ctx_t ctx;
1611 	iflib_rxq_t rxq = (iflib_rxq_t)info->ifi_ctx;
1612 	iflib_txq_t txq;
1613 	void *sc;
1614 	int i, cidx, result;
1615 	qidx_t txqid;
1616 	bool intr_enable, intr_legacy;
1617 
1618 	DBG_COUNTER_INC(fast_intrs);
1619 	if (info->ifi_filter != NULL) {
1620 		result = info->ifi_filter(info->ifi_filter_arg);
1621 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1622 			return (result);
1623 	}
1624 
1625 	ctx = rxq->ifr_ctx;
1626 	sc = ctx->ifc_softc;
1627 	intr_enable = false;
1628 	intr_legacy = !!(ctx->ifc_flags & IFC_LEGACY);
1629 	MPASS(rxq->ifr_ntxqirq);
1630 	for (i = 0; i < rxq->ifr_ntxqirq; i++) {
1631 		txqid = rxq->ifr_txqid[i];
1632 		txq = &ctx->ifc_txqs[txqid];
1633 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1634 		    BUS_DMASYNC_POSTREAD);
1635 		if (!ctx->isc_txd_credits_update(sc, txqid, false)) {
1636 			if (intr_legacy)
1637 				intr_enable = true;
1638 			else
1639 				IFDI_TX_QUEUE_INTR_ENABLE(ctx, txqid);
1640 			continue;
1641 		}
1642 		GROUPTASK_ENQUEUE(&txq->ift_task);
1643 	}
1644 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_RXCQ)
1645 		cidx = rxq->ifr_cq_cidx;
1646 	else
1647 		cidx = rxq->ifr_fl[0].ifl_cidx;
1648 	if (iflib_rxd_avail(ctx, rxq, cidx, 1))
1649 		GROUPTASK_ENQUEUE(gtask);
1650 	else {
1651 		if (intr_legacy)
1652 			intr_enable = true;
1653 		else
1654 			IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
1655 		DBG_COUNTER_INC(rx_intr_enables);
1656 	}
1657 	if (intr_enable)
1658 		IFDI_INTR_ENABLE(ctx);
1659 	return (FILTER_HANDLED);
1660 }
1661 
1662 static int
1663 iflib_fast_intr_ctx(void *arg)
1664 {
1665 	iflib_filter_info_t info = arg;
1666 	if_ctx_t ctx = info->ifi_ctx;
1667 	int result;
1668 
1669 	DBG_COUNTER_INC(fast_intrs);
1670 	if (info->ifi_filter != NULL) {
1671 		result = info->ifi_filter(info->ifi_filter_arg);
1672 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1673 			return (result);
1674 	}
1675 
1676 	taskqueue_enqueue(ctx->ifc_tq, &ctx->ifc_admin_task);
1677 	return (FILTER_HANDLED);
1678 }
1679 
1680 static int
1681 _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
1682 		 driver_filter_t filter, driver_intr_t handler, void *arg,
1683 		 const char *name)
1684 {
1685 	struct resource *res;
1686 	void *tag = NULL;
1687 	device_t dev = ctx->ifc_dev;
1688 	int flags, i, rc;
1689 
1690 	flags = RF_ACTIVE;
1691 	if (ctx->ifc_flags & IFC_LEGACY)
1692 		flags |= RF_SHAREABLE;
1693 	MPASS(rid < 512);
1694 	i = rid;
1695 	res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &i, flags);
1696 	if (res == NULL) {
1697 		device_printf(dev,
1698 		    "failed to allocate IRQ for rid %d, name %s.\n", rid, name);
1699 		return (ENOMEM);
1700 	}
1701 	irq->ii_res = res;
1702 	KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL"));
1703 	rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET,
1704 		    filter, handler, arg, &tag);
1705 	if (rc != 0) {
1706 		device_printf(dev,
1707 		    "failed to setup interrupt for rid %d, name %s: %d\n",
1708 		    rid, name ? name : "unknown", rc);
1709 		return (rc);
1710 	} else if (name)
1711 		bus_describe_intr(dev, res, tag, "%s", name);
1712 
1713 	irq->ii_tag = tag;
1714 	return (0);
1715 }
1716 
1717 /*********************************************************************
1718  *
1719  *  Allocate DMA resources for TX buffers as well as memory for the TX
1720  *  mbuf map.  TX DMA maps (non-TSO/TSO) and TX mbuf map are kept in a
1721  *  iflib_sw_tx_desc_array structure, storing all the information that
1722  *  is needed to transmit a packet on the wire.  This is called only
1723  *  once at attach, setup is done every reset.
1724  *
1725  **********************************************************************/
1726 static int
1727 iflib_txsd_alloc(iflib_txq_t txq)
1728 {
1729 	if_ctx_t ctx = txq->ift_ctx;
1730 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1731 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1732 	device_t dev = ctx->ifc_dev;
1733 	bus_size_t tsomaxsize;
1734 	bus_addr_t lowaddr;
1735 	int err, nsegments, ntsosegments;
1736 	bool tso;
1737 
1738 	nsegments = scctx->isc_tx_nsegments;
1739 	ntsosegments = scctx->isc_tx_tso_segments_max;
1740 	tsomaxsize = scctx->isc_tx_tso_size_max;
1741 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_VLAN_MTU)
1742 		tsomaxsize += sizeof(struct ether_vlan_header);
1743 	MPASS(scctx->isc_ntxd[0] > 0);
1744 	MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0);
1745 	MPASS(nsegments > 0);
1746 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) {
1747 		MPASS(ntsosegments > 0);
1748 		MPASS(sctx->isc_tso_maxsize >= tsomaxsize);
1749 	}
1750 
1751 	lowaddr = DMA_WIDTH_TO_BUS_LOWADDR(scctx->isc_dma_width);
1752 
1753 	/*
1754 	 * Set up DMA tags for TX buffers.
1755 	 */
1756 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1757 		    1, 0,			/* alignment, bounds */
1758 		    lowaddr,			/* lowaddr */
1759 		    BUS_SPACE_MAXADDR,		/* highaddr */
1760 		    NULL, NULL,			/* filter, filterarg */
1761 		    sctx->isc_tx_maxsize,	/* maxsize */
1762 		    nsegments,			/* nsegments */
1763 		    sctx->isc_tx_maxsegsize,	/* maxsegsize */
1764 		    0,				/* flags */
1765 		    NULL,			/* lockfunc */
1766 		    NULL,			/* lockfuncarg */
1767 		    &txq->ift_buf_tag))) {
1768 		device_printf(dev, "Unable to allocate TX DMA tag: %d\n", err);
1769 		device_printf(dev, "maxsize: %ju nsegments: %d maxsegsize: %ju\n",
1770 		    (uintmax_t)sctx->isc_tx_maxsize, nsegments, (uintmax_t)sctx->isc_tx_maxsegsize);
1771 		goto fail;
1772 	}
1773 	tso = (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) != 0;
1774 	if (tso && (err = bus_dma_tag_create(bus_get_dma_tag(dev),
1775 		    1, 0,			/* alignment, bounds */
1776 		    lowaddr,			/* lowaddr */
1777 		    BUS_SPACE_MAXADDR,		/* highaddr */
1778 		    NULL, NULL,			/* filter, filterarg */
1779 		    tsomaxsize,			/* maxsize */
1780 		    ntsosegments,		/* nsegments */
1781 		    sctx->isc_tso_maxsegsize,	/* maxsegsize */
1782 		    0,				/* flags */
1783 		    NULL,			/* lockfunc */
1784 		    NULL,			/* lockfuncarg */
1785 		    &txq->ift_tso_buf_tag))) {
1786 		device_printf(dev, "Unable to allocate TSO TX DMA tag: %d\n",
1787 		    err);
1788 		goto fail;
1789 	}
1790 
1791 	/* Allocate memory for the TX mbuf map. */
1792 	if (!(txq->ift_sds.ifsd_m =
1793 	    (struct mbuf **) malloc(sizeof(struct mbuf *) *
1794 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1795 		device_printf(dev, "Unable to allocate TX mbuf map memory\n");
1796 		err = ENOMEM;
1797 		goto fail;
1798 	}
1799 	if (ctx->ifc_sysctl_simple_tx) {
1800 		if (!(txq->ift_sds.ifsd_m_defer =
1801 			(struct mbuf **) malloc(sizeof(struct mbuf *) *
1802 			    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1803 			device_printf(dev, "Unable to allocate TX mbuf map memory\n");
1804 			err = ENOMEM;
1805 			goto fail;
1806 		}
1807 	}
1808 	txq->ift_sds.ifsd_m_deferb = txq->ift_sds.ifsd_m_defer;
1809 	/*
1810 	 * Create the DMA maps for TX buffers.
1811 	 */
1812 	if ((txq->ift_sds.ifsd_map = (bus_dmamap_t *)malloc(
1813 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1814 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1815 		device_printf(dev,
1816 		    "Unable to allocate TX buffer DMA map memory\n");
1817 		err = ENOMEM;
1818 		goto fail;
1819 	}
1820 	if (tso && (txq->ift_sds.ifsd_tso_map = (bus_dmamap_t *)malloc(
1821 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1822 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1823 		device_printf(dev,
1824 		    "Unable to allocate TSO TX buffer map memory\n");
1825 		err = ENOMEM;
1826 		goto fail;
1827 	}
1828 	for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) {
1829 		err = bus_dmamap_create(txq->ift_buf_tag, 0,
1830 		    &txq->ift_sds.ifsd_map[i]);
1831 		if (err != 0) {
1832 			device_printf(dev, "Unable to create TX DMA map\n");
1833 			goto fail;
1834 		}
1835 		if (!tso)
1836 			continue;
1837 		err = bus_dmamap_create(txq->ift_tso_buf_tag, 0,
1838 		    &txq->ift_sds.ifsd_tso_map[i]);
1839 		if (err != 0) {
1840 			device_printf(dev, "Unable to create TSO TX DMA map\n");
1841 			goto fail;
1842 		}
1843 	}
1844 	return (0);
1845 fail:
1846 	/* We free all, it handles case where we are in the middle */
1847 	iflib_tx_structures_free(ctx);
1848 	return (err);
1849 }
1850 
1851 static void
1852 iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i)
1853 {
1854 	bus_dmamap_t map;
1855 
1856 	if (txq->ift_sds.ifsd_map != NULL) {
1857 		map = txq->ift_sds.ifsd_map[i];
1858 		bus_dmamap_sync(txq->ift_buf_tag, map, BUS_DMASYNC_POSTWRITE);
1859 		bus_dmamap_unload(txq->ift_buf_tag, map);
1860 		bus_dmamap_destroy(txq->ift_buf_tag, map);
1861 		txq->ift_sds.ifsd_map[i] = NULL;
1862 	}
1863 
1864 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1865 		map = txq->ift_sds.ifsd_tso_map[i];
1866 		bus_dmamap_sync(txq->ift_tso_buf_tag, map,
1867 		    BUS_DMASYNC_POSTWRITE);
1868 		bus_dmamap_unload(txq->ift_tso_buf_tag, map);
1869 		bus_dmamap_destroy(txq->ift_tso_buf_tag, map);
1870 		txq->ift_sds.ifsd_tso_map[i] = NULL;
1871 	}
1872 }
1873 
1874 static void
1875 iflib_txq_destroy(iflib_txq_t txq)
1876 {
1877 	if_ctx_t ctx = txq->ift_ctx;
1878 
1879 	for (int i = 0; i < txq->ift_size; i++)
1880 		iflib_txsd_destroy(ctx, txq, i);
1881 
1882 	if (txq->ift_br != NULL) {
1883 		ifmp_ring_free(txq->ift_br);
1884 		txq->ift_br = NULL;
1885 	}
1886 
1887 	mtx_destroy(&txq->ift_mtx);
1888 
1889 	if (txq->ift_sds.ifsd_map != NULL) {
1890 		free(txq->ift_sds.ifsd_map, M_IFLIB);
1891 		txq->ift_sds.ifsd_map = NULL;
1892 	}
1893 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1894 		free(txq->ift_sds.ifsd_tso_map, M_IFLIB);
1895 		txq->ift_sds.ifsd_tso_map = NULL;
1896 	}
1897 	if (txq->ift_sds.ifsd_m != NULL) {
1898 		free(txq->ift_sds.ifsd_m, M_IFLIB);
1899 		txq->ift_sds.ifsd_m = NULL;
1900 	}
1901 	if (txq->ift_sds.ifsd_m_defer != NULL) {
1902 		free(txq->ift_sds.ifsd_m_defer, M_IFLIB);
1903 		txq->ift_sds.ifsd_m_defer = NULL;
1904 	}
1905 	if (txq->ift_buf_tag != NULL) {
1906 		bus_dma_tag_destroy(txq->ift_buf_tag);
1907 		txq->ift_buf_tag = NULL;
1908 	}
1909 	if (txq->ift_tso_buf_tag != NULL) {
1910 		bus_dma_tag_destroy(txq->ift_tso_buf_tag);
1911 		txq->ift_tso_buf_tag = NULL;
1912 	}
1913 	if (txq->ift_ifdi != NULL) {
1914 		free(txq->ift_ifdi, M_IFLIB);
1915 	}
1916 }
1917 
1918 static void
1919 iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i)
1920 {
1921 	struct mbuf *m;
1922 
1923 	m = IFLIB_GET_MBUF(txq->ift_sds.ifsd_m[i]);
1924 	if (m == NULL)
1925 		return;
1926 
1927 	if (txq->ift_sds.ifsd_map != NULL) {
1928 		bus_dmamap_sync(txq->ift_buf_tag,
1929 		    txq->ift_sds.ifsd_map[i], BUS_DMASYNC_POSTWRITE);
1930 		bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[i]);
1931 	}
1932 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1933 		bus_dmamap_sync(txq->ift_tso_buf_tag,
1934 		    txq->ift_sds.ifsd_tso_map[i], BUS_DMASYNC_POSTWRITE);
1935 		bus_dmamap_unload(txq->ift_tso_buf_tag,
1936 		    txq->ift_sds.ifsd_tso_map[i]);
1937 	}
1938 	txq->ift_sds.ifsd_m[i] = NULL;
1939 	m_freem(m);
1940 	DBG_COUNTER_INC(tx_frees);
1941 }
1942 
1943 static int
1944 iflib_txq_setup(iflib_txq_t txq)
1945 {
1946 	if_ctx_t ctx = txq->ift_ctx;
1947 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1948 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1949 	iflib_dma_info_t di;
1950 	int i;
1951 
1952 	/* XXX make configurable */
1953 	txq->ift_update_freq = IFLIB_DEFAULT_TX_UPDATE_FREQ;
1954 
1955 	/* Reset indices */
1956 	txq->ift_cidx_processed = 0;
1957 	txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0;
1958 	txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset];
1959 	txq->ift_pad = scctx->isc_tx_pad;
1960 
1961 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1962 		bzero((void *)di->idi_vaddr, di->idi_size);
1963 
1964 	IFDI_TXQ_SETUP(ctx, txq->ift_id);
1965 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1966 		bus_dmamap_sync(di->idi_tag, di->idi_map,
1967 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1968 	return (0);
1969 }
1970 
1971 /*********************************************************************
1972  *
1973  *  Allocate DMA resources for RX buffers as well as memory for the RX
1974  *  mbuf map, direct RX cluster pointer map and RX cluster bus address
1975  *  map.  RX DMA map, RX mbuf map, direct RX cluster pointer map and
1976  *  RX cluster map are kept in a iflib_sw_rx_desc_array structure.
1977  *  Since we use use one entry in iflib_sw_rx_desc_array per received
1978  *  packet, the maximum number of entries we'll need is equal to the
1979  *  number of hardware receive descriptors that we've allocated.
1980  *
1981  **********************************************************************/
1982 static int
1983 iflib_rxsd_alloc(iflib_rxq_t rxq)
1984 {
1985 	if_ctx_t ctx = rxq->ifr_ctx;
1986 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1987 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1988 	device_t dev = ctx->ifc_dev;
1989 	iflib_fl_t fl;
1990 	bus_addr_t lowaddr;
1991 	int err;
1992 
1993 	MPASS(scctx->isc_nrxd[0] > 0);
1994 	MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0);
1995 
1996 	lowaddr = DMA_WIDTH_TO_BUS_LOWADDR(scctx->isc_dma_width);
1997 
1998 	fl = rxq->ifr_fl;
1999 	for (int i = 0; i < rxq->ifr_nfl; i++, fl++) {
2000 		fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */
2001 		/* Set up DMA tag for RX buffers. */
2002 		err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
2003 			    1, 0,			/* alignment, bounds */
2004 			    lowaddr,			/* lowaddr */
2005 			    BUS_SPACE_MAXADDR,		/* highaddr */
2006 			    NULL, NULL,			/* filter, filterarg */
2007 			    sctx->isc_rx_maxsize,	/* maxsize */
2008 			    sctx->isc_rx_nsegments,	/* nsegments */
2009 			    sctx->isc_rx_maxsegsize,	/* maxsegsize */
2010 			    0,				/* flags */
2011 			    NULL,			/* lockfunc */
2012 			    NULL,			/* lockarg */
2013 			    &fl->ifl_buf_tag);
2014 		if (err) {
2015 			device_printf(dev,
2016 			    "Unable to allocate RX DMA tag: %d\n", err);
2017 			goto fail;
2018 		}
2019 
2020 		/* Allocate memory for the RX mbuf map. */
2021 		if (!(fl->ifl_sds.ifsd_m =
2022 		    (struct mbuf **) malloc(sizeof(struct mbuf *) *
2023 			    scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
2024 			device_printf(dev,
2025 			    "Unable to allocate RX mbuf map memory\n");
2026 			err = ENOMEM;
2027 			goto fail;
2028 		}
2029 
2030 		/* Allocate memory for the direct RX cluster pointer map. */
2031 		if (!(fl->ifl_sds.ifsd_cl =
2032 		    (caddr_t *) malloc(sizeof(caddr_t) *
2033 			    scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
2034 			device_printf(dev,
2035 			    "Unable to allocate RX cluster map memory\n");
2036 			err = ENOMEM;
2037 			goto fail;
2038 		}
2039 
2040 		/* Allocate memory for the RX cluster bus address map. */
2041 		if (!(fl->ifl_sds.ifsd_ba =
2042 		    (bus_addr_t *) malloc(sizeof(bus_addr_t) *
2043 			    scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
2044 			device_printf(dev,
2045 			    "Unable to allocate RX bus address map memory\n");
2046 			err = ENOMEM;
2047 			goto fail;
2048 		}
2049 
2050 		/*
2051 		 * Create the DMA maps for RX buffers.
2052 		 */
2053 		if (!(fl->ifl_sds.ifsd_map =
2054 		    (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
2055 			device_printf(dev,
2056 			    "Unable to allocate RX buffer DMA map memory\n");
2057 			err = ENOMEM;
2058 			goto fail;
2059 		}
2060 		for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++) {
2061 			err = bus_dmamap_create(fl->ifl_buf_tag, 0,
2062 			    &fl->ifl_sds.ifsd_map[i]);
2063 			if (err != 0) {
2064 				device_printf(dev, "Unable to create RX buffer DMA map\n");
2065 				goto fail;
2066 			}
2067 		}
2068 	}
2069 	return (0);
2070 
2071 fail:
2072 	iflib_rx_structures_free(ctx);
2073 	return (err);
2074 }
2075 
2076 /*
2077  * Internal service routines
2078  */
2079 
2080 struct rxq_refill_cb_arg {
2081 	int               error;
2082 	bus_dma_segment_t seg;
2083 	int               nseg;
2084 };
2085 
2086 static void
2087 _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
2088 {
2089 	struct rxq_refill_cb_arg *cb_arg = arg;
2090 
2091 	cb_arg->error = error;
2092 	cb_arg->seg = segs[0];
2093 	cb_arg->nseg = nseg;
2094 }
2095 
2096 /**
2097  * iflib_fl_refill - refill an rxq free-buffer list
2098  * @ctx: the iflib context
2099  * @fl: the free list to refill
2100  * @count: the number of new buffers to allocate
2101  *
2102  * (Re)populate an rxq free-buffer list with up to @count new packet buffers.
2103  * The caller must assure that @count does not exceed the queue's capacity
2104  * minus one (since we always leave a descriptor unavailable).
2105  */
2106 static uint8_t
2107 iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
2108 {
2109 	struct if_rxd_update iru;
2110 	struct rxq_refill_cb_arg cb_arg;
2111 	struct mbuf *m;
2112 	caddr_t cl, *sd_cl;
2113 	struct mbuf **sd_m;
2114 	bus_dmamap_t *sd_map;
2115 	bus_addr_t bus_addr, *sd_ba;
2116 	int err, frag_idx, i, idx, n, pidx;
2117 	qidx_t credits;
2118 
2119 	MPASS(count <= fl->ifl_size - fl->ifl_credits - 1);
2120 
2121 	sd_m = fl->ifl_sds.ifsd_m;
2122 	sd_map = fl->ifl_sds.ifsd_map;
2123 	sd_cl = fl->ifl_sds.ifsd_cl;
2124 	sd_ba = fl->ifl_sds.ifsd_ba;
2125 	pidx = fl->ifl_pidx;
2126 	idx = pidx;
2127 	frag_idx = fl->ifl_fragidx;
2128 	credits = fl->ifl_credits;
2129 
2130 	i = 0;
2131 	n = count;
2132 	MPASS(n > 0);
2133 	MPASS(credits + n <= fl->ifl_size);
2134 
2135 	if (pidx < fl->ifl_cidx)
2136 		MPASS(pidx + n <= fl->ifl_cidx);
2137 	if (pidx == fl->ifl_cidx && (credits < fl->ifl_size))
2138 		MPASS(fl->ifl_gen == 0);
2139 	if (pidx > fl->ifl_cidx)
2140 		MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx);
2141 
2142 	DBG_COUNTER_INC(fl_refills);
2143 	if (n > 8)
2144 		DBG_COUNTER_INC(fl_refills_large);
2145 	iru_init(&iru, fl->ifl_rxq, fl->ifl_id);
2146 	while (n-- > 0) {
2147 		/*
2148 		 * We allocate an uninitialized mbuf + cluster, mbuf is
2149 		 * initialized after rx.
2150 		 *
2151 		 * If the cluster is still set then we know a minimum sized
2152 		 * packet was received
2153 		 */
2154 		bit_ffc_at(fl->ifl_rx_bitmap, frag_idx, fl->ifl_size,
2155 		    &frag_idx);
2156 		if (frag_idx < 0)
2157 			bit_ffc(fl->ifl_rx_bitmap, fl->ifl_size, &frag_idx);
2158 		MPASS(frag_idx >= 0);
2159 		if ((cl = sd_cl[frag_idx]) == NULL) {
2160 			cl = uma_zalloc(fl->ifl_zone, M_NOWAIT);
2161 			if (__predict_false(cl == NULL))
2162 				break;
2163 
2164 			cb_arg.error = 0;
2165 			MPASS(sd_map != NULL);
2166 			err = bus_dmamap_load(fl->ifl_buf_tag, sd_map[frag_idx],
2167 			    cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg,
2168 			    BUS_DMA_NOWAIT);
2169 			if (__predict_false(err != 0 || cb_arg.error)) {
2170 				uma_zfree(fl->ifl_zone, cl);
2171 				break;
2172 			}
2173 
2174 			sd_ba[frag_idx] = bus_addr = cb_arg.seg.ds_addr;
2175 			sd_cl[frag_idx] = cl;
2176 #if MEMORY_LOGGING
2177 			fl->ifl_cl_enqueued++;
2178 #endif
2179 		} else {
2180 			bus_addr = sd_ba[frag_idx];
2181 		}
2182 		bus_dmamap_sync(fl->ifl_buf_tag, sd_map[frag_idx],
2183 		    BUS_DMASYNC_PREREAD);
2184 
2185 		if (sd_m[frag_idx] == NULL) {
2186 			m = m_gethdr_raw(M_NOWAIT, 0);
2187 			if (__predict_false(m == NULL))
2188 				break;
2189 			sd_m[frag_idx] = m;
2190 		}
2191 		bit_set(fl->ifl_rx_bitmap, frag_idx);
2192 #if MEMORY_LOGGING
2193 		fl->ifl_m_enqueued++;
2194 #endif
2195 
2196 		DBG_COUNTER_INC(rx_allocs);
2197 		fl->ifl_rxd_idxs[i] = frag_idx;
2198 		fl->ifl_bus_addrs[i] = bus_addr;
2199 		credits++;
2200 		i++;
2201 		MPASS(credits <= fl->ifl_size);
2202 		if (++idx == fl->ifl_size) {
2203 #ifdef INVARIANTS
2204 			fl->ifl_gen = 1;
2205 #endif
2206 			idx = 0;
2207 		}
2208 		if (n == 0 || i == IFLIB_MAX_RX_REFRESH) {
2209 			iru.iru_pidx = pidx;
2210 			iru.iru_count = i;
2211 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2212 			fl->ifl_pidx = idx;
2213 			fl->ifl_credits = credits;
2214 			pidx = idx;
2215 			i = 0;
2216 		}
2217 	}
2218 
2219 	if (n < count - 1) {
2220 		if (i != 0) {
2221 			iru.iru_pidx = pidx;
2222 			iru.iru_count = i;
2223 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2224 			fl->ifl_pidx = idx;
2225 			fl->ifl_credits = credits;
2226 		}
2227 		DBG_COUNTER_INC(rxd_flush);
2228 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2229 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2230 		ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id,
2231 		    fl->ifl_id, fl->ifl_pidx);
2232 		if (__predict_true(bit_test(fl->ifl_rx_bitmap, frag_idx))) {
2233 			fl->ifl_fragidx = frag_idx + 1;
2234 			if (fl->ifl_fragidx == fl->ifl_size)
2235 				fl->ifl_fragidx = 0;
2236 		} else {
2237 			fl->ifl_fragidx = frag_idx;
2238 		}
2239 	}
2240 
2241 	return (n == -1 ? 0 : IFLIB_RXEOF_EMPTY);
2242 }
2243 
2244 static inline uint8_t
2245 iflib_fl_refill_all(if_ctx_t ctx, iflib_fl_t fl)
2246 {
2247 	/*
2248 	 * We leave an unused descriptor to avoid pidx to catch up with cidx.
2249 	 * This is important as it confuses most NICs. For instance,
2250 	 * Intel NICs have (per receive ring) RDH and RDT registers, where
2251 	 * RDH points to the next receive descriptor to be used by the NIC,
2252 	 * and RDT for the next receive descriptor to be published by the
2253 	 * driver to the NIC (RDT - 1 is thus the last valid one).
2254 	 * The condition RDH == RDT means no descriptors are available to
2255 	 * the NIC, and thus it would be ambiguous if it also meant that
2256 	 * all the descriptors are available to the NIC.
2257 	 */
2258 	int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1;
2259 #ifdef INVARIANTS
2260 	int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1;
2261 #endif
2262 
2263 	MPASS(fl->ifl_credits <= fl->ifl_size);
2264 	MPASS(reclaimable == delta);
2265 
2266 	if (reclaimable > 0)
2267 		return (iflib_fl_refill(ctx, fl, reclaimable));
2268 	return (0);
2269 }
2270 
2271 uint8_t
2272 iflib_in_detach(if_ctx_t ctx)
2273 {
2274 	bool in_detach;
2275 
2276 	STATE_LOCK(ctx);
2277 	in_detach = !!(ctx->ifc_flags & IFC_IN_DETACH);
2278 	STATE_UNLOCK(ctx);
2279 	return (in_detach);
2280 }
2281 
2282 static void
2283 iflib_fl_bufs_free(iflib_fl_t fl)
2284 {
2285 	iflib_dma_info_t idi = fl->ifl_ifdi;
2286 	bus_dmamap_t sd_map;
2287 	uint32_t i;
2288 
2289 	for (i = 0; i < fl->ifl_size; i++) {
2290 		struct mbuf **sd_m = &fl->ifl_sds.ifsd_m[i];
2291 		caddr_t *sd_cl = &fl->ifl_sds.ifsd_cl[i];
2292 
2293 		if (*sd_cl != NULL) {
2294 			sd_map = fl->ifl_sds.ifsd_map[i];
2295 			bus_dmamap_sync(fl->ifl_buf_tag, sd_map,
2296 			    BUS_DMASYNC_POSTREAD);
2297 			bus_dmamap_unload(fl->ifl_buf_tag, sd_map);
2298 			uma_zfree(fl->ifl_zone, *sd_cl);
2299 			*sd_cl = NULL;
2300 			if (*sd_m != NULL) {
2301 				m_init(*sd_m, M_NOWAIT, MT_DATA, 0);
2302 				m_free_raw(*sd_m);
2303 				*sd_m = NULL;
2304 			}
2305 		} else {
2306 			MPASS(*sd_m == NULL);
2307 		}
2308 #if MEMORY_LOGGING
2309 		fl->ifl_m_dequeued++;
2310 		fl->ifl_cl_dequeued++;
2311 #endif
2312 	}
2313 #ifdef INVARIANTS
2314 	for (i = 0; i < fl->ifl_size; i++) {
2315 		MPASS(fl->ifl_sds.ifsd_cl[i] == NULL);
2316 		MPASS(fl->ifl_sds.ifsd_m[i] == NULL);
2317 	}
2318 #endif
2319 	/*
2320 	 * Reset free list values
2321 	 */
2322 	fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = fl->ifl_fragidx = 0;
2323 	bzero(idi->idi_vaddr, idi->idi_size);
2324 }
2325 
2326 /*********************************************************************
2327  *
2328  *  Initialize a free list and its buffers.
2329  *
2330  **********************************************************************/
2331 static int
2332 iflib_fl_setup(iflib_fl_t fl)
2333 {
2334 	iflib_rxq_t rxq = fl->ifl_rxq;
2335 	if_ctx_t ctx = rxq->ifr_ctx;
2336 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2337 	int qidx;
2338 
2339 	bit_nclear(fl->ifl_rx_bitmap, 0, fl->ifl_size - 1);
2340 	/*
2341 	 * Free current RX buffer structs and their mbufs
2342 	 */
2343 	iflib_fl_bufs_free(fl);
2344 	/* Now replenish the mbufs */
2345 	MPASS(fl->ifl_credits == 0);
2346 	qidx = rxq->ifr_fl_offset + fl->ifl_id;
2347 	if (scctx->isc_rxd_buf_size[qidx] != 0)
2348 		fl->ifl_buf_size = scctx->isc_rxd_buf_size[qidx];
2349 	else
2350 		fl->ifl_buf_size = ctx->ifc_rx_mbuf_sz;
2351 	/*
2352 	 * ifl_buf_size may be a driver-supplied value, so pull it up
2353 	 * to the selected mbuf size.
2354 	 */
2355 	fl->ifl_buf_size = iflib_get_mbuf_size_for(fl->ifl_buf_size);
2356 	if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size)
2357 		ctx->ifc_max_fl_buf_size = fl->ifl_buf_size;
2358 	fl->ifl_cltype = m_gettype(fl->ifl_buf_size);
2359 	fl->ifl_zone = m_getzone(fl->ifl_buf_size);
2360 
2361 	/*
2362 	 * Avoid pre-allocating zillions of clusters to an idle card
2363 	 * potentially speeding up attach. In any case make sure
2364 	 * to leave a descriptor unavailable. See the comment in
2365 	 * iflib_fl_refill_all().
2366 	 */
2367 	MPASS(fl->ifl_size > 0);
2368 	(void)iflib_fl_refill(ctx, fl, min(128, fl->ifl_size - 1));
2369 	if (min(128, fl->ifl_size - 1) != fl->ifl_credits)
2370 		return (ENOBUFS);
2371 	/*
2372 	 * handle failure
2373 	 */
2374 	MPASS(rxq != NULL);
2375 	MPASS(fl->ifl_ifdi != NULL);
2376 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2377 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2378 	return (0);
2379 }
2380 
2381 /*********************************************************************
2382  *
2383  *  Free receive ring data structures
2384  *
2385  **********************************************************************/
2386 static void
2387 iflib_rx_sds_free(iflib_rxq_t rxq)
2388 {
2389 	iflib_fl_t fl;
2390 	int i, j;
2391 
2392 	if (rxq->ifr_fl != NULL) {
2393 		for (i = 0; i < rxq->ifr_nfl; i++) {
2394 			fl = &rxq->ifr_fl[i];
2395 			if (fl->ifl_buf_tag != NULL) {
2396 				if (fl->ifl_sds.ifsd_map != NULL) {
2397 					for (j = 0; j < fl->ifl_size; j++) {
2398 						bus_dmamap_sync(
2399 						    fl->ifl_buf_tag,
2400 						    fl->ifl_sds.ifsd_map[j],
2401 						    BUS_DMASYNC_POSTREAD);
2402 						bus_dmamap_unload(
2403 						    fl->ifl_buf_tag,
2404 						    fl->ifl_sds.ifsd_map[j]);
2405 						bus_dmamap_destroy(
2406 						    fl->ifl_buf_tag,
2407 						    fl->ifl_sds.ifsd_map[j]);
2408 					}
2409 				}
2410 				bus_dma_tag_destroy(fl->ifl_buf_tag);
2411 				fl->ifl_buf_tag = NULL;
2412 			}
2413 			free(fl->ifl_sds.ifsd_m, M_IFLIB);
2414 			free(fl->ifl_sds.ifsd_cl, M_IFLIB);
2415 			free(fl->ifl_sds.ifsd_ba, M_IFLIB);
2416 			free(fl->ifl_sds.ifsd_map, M_IFLIB);
2417 			free(fl->ifl_rx_bitmap, M_IFLIB);
2418 			fl->ifl_sds.ifsd_m = NULL;
2419 			fl->ifl_sds.ifsd_cl = NULL;
2420 			fl->ifl_sds.ifsd_ba = NULL;
2421 			fl->ifl_sds.ifsd_map = NULL;
2422 			fl->ifl_rx_bitmap = NULL;
2423 		}
2424 		free(rxq->ifr_fl, M_IFLIB);
2425 		rxq->ifr_fl = NULL;
2426 		free(rxq->ifr_ifdi, M_IFLIB);
2427 		rxq->ifr_ifdi = NULL;
2428 		rxq->ifr_cq_cidx = 0;
2429 	}
2430 }
2431 
2432 /*
2433  * Timer routine
2434  */
2435 static void
2436 iflib_timer(void *arg)
2437 {
2438 	iflib_txq_t txq = arg;
2439 	if_ctx_t ctx = txq->ift_ctx;
2440 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2441 	uint64_t this_tick = ticks;
2442 
2443 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
2444 		return;
2445 
2446 	/*
2447 	 * Check on the state of the TX queue(s); this can be done
2448 	 * without the lock: the counters the check reads are only
2449 	 * advanced by the queue's tx task and a stale read just
2450 	 * delays the verdict by one timer period.
2451 	 */
2452 	if (this_tick - txq->ift_last_timer_tick >= iflib_timer_default) {
2453 		qidx_t in_use, outstanding;
2454 		bool demand, frozen;
2455 
2456 		txq->ift_last_timer_tick = this_tick;
2457 		IFDI_TIMER(ctx, txq->ift_id);
2458 
2459 		/*
2460 		 * Descriptors the hardware has not reported as
2461 		 * completed: neither harvested as credits
2462 		 * (ift_processed) nor reclaimed (ift_cleaned accounts
2463 		 * the difference to ift_in_use).  The tail whose
2464 		 * report-status request is still deferred is never
2465 		 * reported and must not count.
2466 		 */
2467 		in_use = txq->ift_in_use;
2468 		outstanding = in_use -
2469 		    (qidx_t)(txq->ift_processed - txq->ift_cleaned);
2470 
2471 		/*
2472 		 * The queue is frozen while it has descriptors the
2473 		 * hardware has not reported as completed and none
2474 		 * were reclaimed over the period; the link must be
2475 		 * up, with no pause frames and no pending doorbell
2476 		 * (the laggard check below rings it).
2477 		 *
2478 		 * Being frozen is not a fault - the hardware may defer
2479 		 * marking descriptors as completed indefinitely, and
2480 		 * 8254x hardware does so for a quiet queue.  Continue
2481 		 * arming only while demand persists: the outstanding
2482 		 * count grows, the software ring is stalled, or the
2483 		 * hardware ring has reached iflib's backpressure
2484 		 * threshold.  The last condition covers simple-TX, which
2485 		 * does not use the software ring.  This also prevents one
2486 		 * mixed lockless counter sample from arming a quiet queue
2487 		 * until the verdict.  Act only once it has stayed frozen
2488 		 * under demand for
2489 		 * net.iflib.tx_watchdog_periods consecutive periods.
2490 		 */
2491 		frozen = outstanding > txq->ift_rs_pending &&
2492 		    txq->ift_processed == txq->ift_processed_prev &&
2493 		    txq->ift_db_pending == 0 &&
2494 		    sctx->isc_pause_frames == 0 &&
2495 		    ctx->ifc_link_state == LINK_STATE_UP;
2496 		demand = outstanding > txq->ift_outstanding_prev ||
2497 		    ifmp_ring_is_stalled(txq->ift_br) ||
2498 		    in_use + MAX_TX_DESC(ctx) >= txq->ift_size - txq->ift_pad;
2499 		if (!frozen || !demand)
2500 			txq->ift_wdog_armed = 0;
2501 		else {
2502 			if (txq->ift_wdog_armed < UINT16_MAX)
2503 				txq->ift_wdog_armed++;
2504 		}
2505 
2506 		/*
2507 		 * Frozen long enough: ask the hardware.  Completions
2508 		 * ready but unharvested for this long mean the
2509 		 * completion interrupt went missing - kick the
2510 		 * queue's task.  Nothing ready while demand persisted
2511 		 * means it is hung.
2512 		 */
2513 		if (iflib_tx_watchdog_periods > 0 &&
2514 		    txq->ift_wdog_armed >= iflib_tx_watchdog_periods) {
2515 			bus_dmamap_sync(txq->ift_ifdi->idi_tag,
2516 			    txq->ift_ifdi->idi_map, BUS_DMASYNC_POSTREAD);
2517 			if (ctx->isc_txd_credits_update(ctx->ifc_softc,
2518 			    txq->ift_id, false) == 0) {
2519 				device_printf(ctx->ifc_dev,
2520 				    "Watchdog timeout (TX: %d desc "
2521 				    "avail: %d pidx: %d) -- resetting\n",
2522 				    txq->ift_id, TXQ_AVAIL(txq),
2523 				    txq->ift_pidx);
2524 				STATE_LOCK(ctx);
2525 				if_setdrvflagbits(ctx->ifc_ifp,
2526 				    IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2527 				ctx->ifc_flags |=
2528 				    (IFC_DO_WATCHDOG | IFC_DO_RESET);
2529 				iflib_admin_intr_deferred(ctx);
2530 				STATE_UNLOCK(ctx);
2531 				return;
2532 			}
2533 			GROUPTASK_ENQUEUE(&txq->ift_task);
2534 		}
2535 		txq->ift_outstanding_prev = outstanding;
2536 		txq->ift_processed_prev = txq->ift_processed;
2537 	}
2538 	/* handle any laggards */
2539 	if (txq->ift_db_pending)
2540 		GROUPTASK_ENQUEUE(&txq->ift_task);
2541 
2542 	sctx->isc_pause_frames = 0;
2543 	if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)
2544 		callout_reset_on(&txq->ift_timer, iflib_timer_default, iflib_timer,
2545 		    txq, txq->ift_timer.c_cpu);
2546 }
2547 
2548 static uint16_t
2549 iflib_get_mbuf_size_for(unsigned int size)
2550 {
2551 
2552 	if (size <= MCLBYTES)
2553 		return (MCLBYTES);
2554 	else
2555 		return (MJUMPAGESIZE);
2556 }
2557 
2558 static void
2559 iflib_calc_rx_mbuf_sz(if_ctx_t ctx)
2560 {
2561 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2562 
2563 	/*
2564 	 * XXX don't set the max_frame_size to larger
2565 	 * than the hardware can handle
2566 	 */
2567 	ctx->ifc_rx_mbuf_sz =
2568 	    iflib_get_mbuf_size_for(sctx->isc_max_frame_size);
2569 }
2570 
2571 uint32_t
2572 iflib_get_rx_mbuf_sz(if_ctx_t ctx)
2573 {
2574 
2575 	return (ctx->ifc_rx_mbuf_sz);
2576 }
2577 
2578 static void
2579 iflib_init_locked(if_ctx_t ctx)
2580 {
2581 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2582 	if_t ifp = ctx->ifc_ifp;
2583 	iflib_fl_t fl;
2584 	iflib_txq_t txq;
2585 	iflib_rxq_t rxq;
2586 	int i, j, tx_ip_csum_flags, tx_ip6_csum_flags;
2587 	bool init_failed;
2588 
2589 	if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2590 	IFDI_INTR_DISABLE(ctx);
2591 
2592 	/*
2593 	 * See iflib_stop(). Useful in case iflib_init_locked() is
2594 	 * called without first calling iflib_stop().
2595 	 */
2596 	netmap_disable_all_rings(ifp);
2597 
2598 	tx_ip_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP);
2599 	tx_ip6_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | CSUM_IP6_SCTP);
2600 	/* Set hardware offload abilities */
2601 	if_clearhwassist(ifp);
2602 	if (if_getcapenable(ifp) & IFCAP_TXCSUM)
2603 		if_sethwassistbits(ifp, tx_ip_csum_flags, 0);
2604 	if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6)
2605 		if_sethwassistbits(ifp,  tx_ip6_csum_flags, 0);
2606 	if (if_getcapenable(ifp) & IFCAP_TSO4)
2607 		if_sethwassistbits(ifp, CSUM_IP_TSO, 0);
2608 	if (if_getcapenable(ifp) & IFCAP_TSO6)
2609 		if_sethwassistbits(ifp, CSUM_IP6_TSO, 0);
2610 
2611 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
2612 		CALLOUT_LOCK(txq);
2613 		callout_stop(&txq->ift_timer);
2614 #ifdef DEV_NETMAP
2615 		callout_stop(&txq->ift_netmap_timer);
2616 #endif /* DEV_NETMAP */
2617 		CALLOUT_UNLOCK(txq);
2618 		(void)iflib_netmap_txq_init(ctx, txq);
2619 	}
2620 
2621 	/*
2622 	 * Calculate a suitable Rx mbuf size prior to calling IFDI_INIT, so
2623 	 * that drivers can use the value when setting up the hardware receive
2624 	 * buffers.
2625 	 */
2626 	iflib_calc_rx_mbuf_sz(ctx);
2627 
2628 #ifdef INVARIANTS
2629 	i = if_getdrvflags(ifp);
2630 #endif
2631 	STATE_LOCK(ctx);
2632 	ctx->ifc_flags &= ~IFC_INIT_FAILED;
2633 	STATE_UNLOCK(ctx);
2634 	IFDI_INIT(ctx);
2635 	MPASS(if_getdrvflags(ifp) == i);
2636 	STATE_LOCK(ctx);
2637 	init_failed = (ctx->ifc_flags & IFC_INIT_FAILED) != 0;
2638 	STATE_UNLOCK(ctx);
2639 	if (init_failed)
2640 		return;
2641 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
2642 		if (iflib_netmap_rxq_init(ctx, rxq) > 0) {
2643 			/* This rxq is in netmap mode. Skip normal init. */
2644 			continue;
2645 		}
2646 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
2647 			if (iflib_fl_setup(fl)) {
2648 				device_printf(ctx->ifc_dev,
2649 				    "setting up free list %d failed - "
2650 				    "check cluster settings\n", j);
2651 				goto done;
2652 			}
2653 		}
2654 	}
2655 done:
2656 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
2657 	IFDI_INTR_ENABLE(ctx);
2658 	txq = ctx->ifc_txqs;
2659 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++)
2660 		callout_reset_on(&txq->ift_timer, iflib_timer_default, iflib_timer, txq,
2661 			txq->ift_timer.c_cpu);
2662 
2663 	/* Re-enable txsync/rxsync. */
2664 	netmap_enable_all_rings(ifp);
2665 }
2666 
2667 static int
2668 iflib_media_change(if_t ifp)
2669 {
2670 	if_ctx_t ctx = if_getsoftc(ifp);
2671 	int err;
2672 
2673 	CTX_LOCK(ctx);
2674 	if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0)
2675 		iflib_if_init_locked(ctx);
2676 	CTX_UNLOCK(ctx);
2677 	return (err);
2678 }
2679 
2680 static void
2681 iflib_media_status(if_t ifp, struct ifmediareq *ifmr)
2682 {
2683 	if_ctx_t ctx = if_getsoftc(ifp);
2684 	bool oactive, running;
2685 
2686 	STATE_LOCK(ctx);
2687 	running = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING);
2688 	oactive = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE);
2689 	STATE_UNLOCK(ctx);
2690 
2691 	CTX_LOCK(ctx);
2692 	/*
2693 	 * There is no need to update the admin status when it is done regularly by
2694 	 * _task_fn_admin(), so only do it if that's not running. That can be quite
2695 	 * expensive on some drivers.
2696 	 */
2697 	if ((!running && !oactive) &&
2698 	    !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN)) {
2699 		IFDI_UPDATE_ADMIN_STATUS(ctx);
2700 	}
2701 	IFDI_MEDIA_STATUS(ctx, ifmr);
2702 	CTX_UNLOCK(ctx);
2703 }
2704 
2705 static void
2706 iflib_stop(if_ctx_t ctx)
2707 {
2708 	iflib_txq_t txq = ctx->ifc_txqs;
2709 	iflib_rxq_t rxq = ctx->ifc_rxqs;
2710 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2711 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2712 	iflib_dma_info_t di;
2713 	iflib_fl_t fl;
2714 	int i, j;
2715 
2716 	/* Tell the stack that the interface is no longer active */
2717 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2718 
2719 	IFDI_INTR_DISABLE(ctx);
2720 	DELAY(1000);
2721 	IFDI_STOP(ctx);
2722 	DELAY(1000);
2723 
2724 	/*
2725 	 * Stop any pending txsync/rxsync and prevent new ones
2726 	 * form starting. Processes blocked in poll() will get
2727 	 * POLLERR.
2728 	 */
2729 	netmap_disable_all_rings(ctx->ifc_ifp);
2730 
2731 	iflib_debug_reset();
2732 	/* Wait for current tx queue users to exit to disarm watchdog timer. */
2733 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) {
2734 		/* make sure all transmitters have completed before proceeding XXX */
2735 
2736 		CALLOUT_LOCK(txq);
2737 		callout_stop(&txq->ift_timer);
2738 #ifdef DEV_NETMAP
2739 		callout_stop(&txq->ift_netmap_timer);
2740 #endif /* DEV_NETMAP */
2741 		CALLOUT_UNLOCK(txq);
2742 
2743 		if (!ctx->ifc_sysctl_simple_tx) {
2744 			/* clean any enqueued buffers */
2745 			iflib_ifmp_purge(txq);
2746 		}
2747 		/* Free any existing tx buffers. */
2748 		for (j = 0; j < txq->ift_size; j++) {
2749 			iflib_txsd_free(ctx, txq, j);
2750 		}
2751 		txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0;
2752 		txq->ift_processed_prev = 0;
2753 		txq->ift_outstanding_prev = 0;
2754 		txq->ift_wdog_armed = 0;
2755 		txq->ift_in_use = txq->ift_gen = txq->ift_no_desc_avail = 0;
2756 		txq->ift_npending = txq->ift_db_pending = 0;
2757 		txq->ift_rs_pending = 0;
2758 		if (sctx->isc_flags & IFLIB_PRESERVE_TX_INDICES)
2759 			txq->ift_cidx = txq->ift_pidx;
2760 		else
2761 			txq->ift_cidx = txq->ift_pidx = 0;
2762 
2763 		txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0;
2764 		txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0;
2765 		txq->ift_pullups = 0;
2766 		ifmp_ring_reset_stats(txq->ift_br);
2767 		for (j = 0, di = txq->ift_ifdi; j < sctx->isc_ntxqs; j++, di++)
2768 			bzero((void *)di->idi_vaddr, di->idi_size);
2769 	}
2770 	for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) {
2771 		if (rxq->ifr_task.gt_taskqueue != NULL)
2772 			gtaskqueue_drain(rxq->ifr_task.gt_taskqueue,
2773 				 &rxq->ifr_task.gt_task);
2774 
2775 		rxq->ifr_cq_cidx = 0;
2776 		for (j = 0, di = rxq->ifr_ifdi; j < sctx->isc_nrxqs; j++, di++)
2777 			bzero((void *)di->idi_vaddr, di->idi_size);
2778 		/* also resets the free lists pidx/cidx */
2779 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
2780 			iflib_fl_bufs_free(fl);
2781 	}
2782 }
2783 
2784 static inline caddr_t
2785 calc_next_rxd(iflib_fl_t fl, int cidx)
2786 {
2787 	qidx_t size;
2788 	int nrxd;
2789 	caddr_t start, end, cur, next;
2790 
2791 	nrxd = fl->ifl_size;
2792 	size = fl->ifl_rxd_size;
2793 	start = fl->ifl_ifdi->idi_vaddr;
2794 
2795 	if (__predict_false(size == 0))
2796 		return (start);
2797 	cur = start + size * cidx;
2798 	end = start + size * nrxd;
2799 	next = CACHE_PTR_NEXT(cur);
2800 	return (next < end ? next : start);
2801 }
2802 
2803 static inline void
2804 prefetch_pkts(iflib_fl_t fl, int cidx)
2805 {
2806 	int nextptr;
2807 	int nrxd = fl->ifl_size;
2808 	caddr_t next_rxd;
2809 
2810 	nextptr = (cidx + CACHE_PTR_INCREMENT) & (nrxd - 1);
2811 	prefetch(&fl->ifl_sds.ifsd_m[nextptr]);
2812 	prefetch(&fl->ifl_sds.ifsd_cl[nextptr]);
2813 	next_rxd = calc_next_rxd(fl, cidx);
2814 	prefetch(next_rxd);
2815 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 1) & (nrxd - 1)]);
2816 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 2) & (nrxd - 1)]);
2817 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 3) & (nrxd - 1)]);
2818 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 4) & (nrxd - 1)]);
2819 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 1) & (nrxd - 1)]);
2820 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 2) & (nrxd - 1)]);
2821 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 3) & (nrxd - 1)]);
2822 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 4) & (nrxd - 1)]);
2823 }
2824 
2825 static struct mbuf *
2826 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, bool unload, if_rxsd_t sd,
2827     int *pf_rv, if_rxd_info_t ri)
2828 {
2829 	bus_dmamap_t map;
2830 	iflib_fl_t fl;
2831 	caddr_t payload;
2832 	struct mbuf *m;
2833 	int flid, cidx, len, next;
2834 
2835 	map = NULL;
2836 	flid = irf->irf_flid;
2837 	cidx = irf->irf_idx;
2838 	fl = &rxq->ifr_fl[flid];
2839 	sd->ifsd_fl = fl;
2840 	sd->ifsd_cl = &fl->ifl_sds.ifsd_cl[cidx];
2841 	fl->ifl_credits--;
2842 #if MEMORY_LOGGING
2843 	fl->ifl_m_dequeued++;
2844 #endif
2845 	if (rxq->ifr_ctx->ifc_flags & IFC_PREFETCH)
2846 		prefetch_pkts(fl, cidx);
2847 	next = (cidx + CACHE_PTR_INCREMENT) & (fl->ifl_size - 1);
2848 	prefetch(&fl->ifl_sds.ifsd_map[next]);
2849 	map = fl->ifl_sds.ifsd_map[cidx];
2850 
2851 	bus_dmamap_sync(fl->ifl_buf_tag, map, BUS_DMASYNC_POSTREAD);
2852 
2853 	if (rxq->pfil != NULL && PFIL_HOOKED_IN(rxq->pfil) && pf_rv != NULL &&
2854 	    irf->irf_len != 0) {
2855 		payload  = *sd->ifsd_cl;
2856 		payload +=  ri->iri_pad;
2857 		len = ri->iri_len - ri->iri_pad;
2858 		*pf_rv = pfil_mem_in(rxq->pfil, payload, len, ri->iri_ifp, &m);
2859 		switch (*pf_rv) {
2860 		case PFIL_DROPPED:
2861 		case PFIL_CONSUMED:
2862 			/*
2863 			 * The filter ate it.  Everything is recycled.
2864 			 */
2865 			m = NULL;
2866 			unload = 0;
2867 			break;
2868 		case PFIL_REALLOCED:
2869 			/*
2870 			 * The filter copied it.  Everything is recycled.
2871 			 * 'm' points at new mbuf.
2872 			 */
2873 			unload = 0;
2874 			break;
2875 		case PFIL_PASS:
2876 			/*
2877 			 * Filter said it was OK, so receive like
2878 			 * normal
2879 			 */
2880 			m = fl->ifl_sds.ifsd_m[cidx];
2881 			fl->ifl_sds.ifsd_m[cidx] = NULL;
2882 			break;
2883 		default:
2884 			MPASS(0);
2885 		}
2886 	} else {
2887 		m = fl->ifl_sds.ifsd_m[cidx];
2888 		fl->ifl_sds.ifsd_m[cidx] = NULL;
2889 		if (pf_rv != NULL)
2890 			*pf_rv = PFIL_PASS;
2891 	}
2892 
2893 	if (unload && irf->irf_len != 0)
2894 		bus_dmamap_unload(fl->ifl_buf_tag, map);
2895 	fl->ifl_cidx = (fl->ifl_cidx + 1) & (fl->ifl_size - 1);
2896 	if (__predict_false(fl->ifl_cidx == 0))
2897 		fl->ifl_gen = 0;
2898 	bit_clear(fl->ifl_rx_bitmap, cidx);
2899 	return (m);
2900 }
2901 
2902 static struct mbuf *
2903 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri, if_rxsd_t sd, int *pf_rv)
2904 {
2905 	struct mbuf *m, *mh, *mt;
2906 	caddr_t cl;
2907 	int  *pf_rv_ptr, flags, i, padlen;
2908 	bool consumed;
2909 
2910 	i = 0;
2911 	mh = NULL;
2912 	consumed = false;
2913 	*pf_rv = PFIL_PASS;
2914 	pf_rv_ptr = pf_rv;
2915 	do {
2916 		m = rxd_frag_to_sd(rxq, &ri->iri_frags[i], !consumed, sd,
2917 		    pf_rv_ptr, ri);
2918 
2919 		MPASS(*sd->ifsd_cl != NULL);
2920 
2921 		/*
2922 		 * Exclude zero-length frags & frags from
2923 		 * packets the filter has consumed or dropped
2924 		 */
2925 		if (ri->iri_frags[i].irf_len == 0 || consumed ||
2926 		    *pf_rv == PFIL_CONSUMED || *pf_rv == PFIL_DROPPED) {
2927 			if (mh == NULL) {
2928 				consumed = true;
2929 				pf_rv_ptr = NULL;
2930 			}
2931 			/* XXX we can save the cluster here, but not the mbuf */
2932 			if (m != NULL) {
2933 				m_init(m, M_NOWAIT, MT_DATA, 0);
2934 				m_free(m);
2935 			}
2936 			continue;
2937 		}
2938 		if (mh == NULL) {
2939 			flags = M_PKTHDR | M_EXT;
2940 			mh = mt = m;
2941 			padlen = ri->iri_pad;
2942 		} else {
2943 			flags = M_EXT;
2944 			mt->m_next = m;
2945 			mt = m;
2946 			/* assuming padding is only on the first fragment */
2947 			padlen = 0;
2948 		}
2949 		cl = *sd->ifsd_cl;
2950 		*sd->ifsd_cl = NULL;
2951 
2952 		/* Can these two be made one ? */
2953 		m_init(m, M_NOWAIT, MT_DATA, flags);
2954 		m_cljset(m, cl, sd->ifsd_fl->ifl_cltype);
2955 		/*
2956 		 * These must follow m_init and m_cljset
2957 		 */
2958 		m->m_data += padlen;
2959 		ri->iri_len -= padlen;
2960 		m->m_len = ri->iri_frags[i].irf_len;
2961 	} while (++i < ri->iri_nfrags);
2962 
2963 	return (mh);
2964 }
2965 
2966 /*
2967  * Process one software descriptor
2968  */
2969 static struct mbuf *
2970 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri)
2971 {
2972 	struct if_rxsd sd;
2973 	struct mbuf *m;
2974 	int pf_rv;
2975 
2976 	/* should I merge this back in now that the two paths are basically duplicated? */
2977 	if (ri->iri_nfrags == 1 &&
2978 	    ri->iri_frags[0].irf_len != 0 &&
2979 	    ri->iri_frags[0].irf_len <= MIN(IFLIB_RX_COPY_THRESH, MHLEN)) {
2980 		m = rxd_frag_to_sd(rxq, &ri->iri_frags[0], false, &sd,
2981 		    &pf_rv, ri);
2982 		if (pf_rv != PFIL_PASS && pf_rv != PFIL_REALLOCED)
2983 			return (m);
2984 		if (pf_rv == PFIL_PASS) {
2985 			m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR);
2986 #ifndef __NO_STRICT_ALIGNMENT
2987 			if (!IP_ALIGNED(m) && ri->iri_pad == 0)
2988 				m->m_data += 2;
2989 #endif
2990 			memcpy(m->m_data, *sd.ifsd_cl, ri->iri_len);
2991 			m->m_len = ri->iri_frags[0].irf_len;
2992 			m->m_data += ri->iri_pad;
2993 			ri->iri_len -= ri->iri_pad;
2994 		}
2995 	} else {
2996 		m = assemble_segments(rxq, ri, &sd, &pf_rv);
2997 		if (m == NULL)
2998 			return (NULL);
2999 		if (pf_rv != PFIL_PASS && pf_rv != PFIL_REALLOCED)
3000 			return (m);
3001 	}
3002 	m->m_pkthdr.len = ri->iri_len;
3003 	m->m_pkthdr.rcvif = ri->iri_ifp;
3004 	m->m_flags |= ri->iri_flags & IFLIB_IRI_VALID_FLAGS;
3005 	m->m_pkthdr.ether_vtag = ri->iri_vtag;
3006 	m->m_pkthdr.flowid = ri->iri_flowid;
3007 #ifdef NUMA
3008 	m->m_pkthdr.numa_domain = if_getnumadomain(ri->iri_ifp);
3009 #endif
3010 	M_HASHTYPE_SET(m, ri->iri_rsstype);
3011 	m->m_pkthdr.csum_flags = ri->iri_csum_flags;
3012 	m->m_pkthdr.csum_data = ri->iri_csum_data;
3013 	m->m_pkthdr.rcv_tstmp = ri->iri_rcv_tstmp;
3014 	return (m);
3015 }
3016 
3017 static void
3018 _task_fn_rx_watchdog(void *context)
3019 {
3020 	iflib_rxq_t rxq = context;
3021 
3022 	GROUPTASK_ENQUEUE(&rxq->ifr_task);
3023 }
3024 
3025 static uint8_t
3026 iflib_rxeof(iflib_rxq_t rxq, qidx_t budget)
3027 {
3028 	if_t ifp;
3029 	if_ctx_t ctx = rxq->ifr_ctx;
3030 	if_shared_ctx_t sctx = ctx->ifc_sctx;
3031 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
3032 	int avail, i;
3033 	qidx_t *cidxp;
3034 	struct if_rxd_info ri;
3035 	int err, budget_left, rx_bytes, rx_pkts;
3036 	iflib_fl_t fl;
3037 #if defined(INET6) || defined(INET)
3038 	int lro_enabled;
3039 #endif
3040 	uint8_t retval = 0;
3041 
3042 	/*
3043 	 * XXX early demux data packets so that if_input processing only handles
3044 	 * acks in interrupt context
3045 	 */
3046 	struct mbuf *m, *mh, *mt;
3047 
3048 	NET_EPOCH_ASSERT();
3049 
3050 	ifp = ctx->ifc_ifp;
3051 	mh = mt = NULL;
3052 	MPASS(budget > 0);
3053 	rx_pkts	= rx_bytes = 0;
3054 	if (sctx->isc_flags & IFLIB_HAS_RXCQ)
3055 		cidxp = &rxq->ifr_cq_cidx;
3056 	else
3057 		cidxp = &rxq->ifr_fl[0].ifl_cidx;
3058 	if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) {
3059 		for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
3060 			retval |= iflib_fl_refill_all(ctx, fl);
3061 		DBG_COUNTER_INC(rx_unavail);
3062 		return (retval);
3063 	}
3064 
3065 #if defined(INET6) || defined(INET)
3066 	lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO);
3067 #endif
3068 
3069 	/* pfil needs the vnet to be set */
3070 	CURVNET_SET_QUIET(if_getvnet(ifp));
3071 	for (budget_left = budget; budget_left > 0 && avail > 0;) {
3072 		if (__predict_false(!CTX_ACTIVE(ctx))) {
3073 			DBG_COUNTER_INC(rx_ctx_inactive);
3074 			break;
3075 		}
3076 		/*
3077 		 * Reset client set fields to their default values
3078 		 */
3079 		memset(&ri, 0, sizeof(ri));
3080 		ri.iri_qsidx = rxq->ifr_id;
3081 		ri.iri_cidx = *cidxp;
3082 		ri.iri_ifp = ifp;
3083 		ri.iri_frags = rxq->ifr_frags;
3084 		err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
3085 
3086 		if (err) {
3087 			CURVNET_RESTORE();
3088 			goto err;
3089 		}
3090 		rx_pkts += 1;
3091 		rx_bytes += ri.iri_len;
3092 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
3093 			*cidxp = ri.iri_cidx;
3094 			/* Update our consumer index */
3095 			/* XXX NB: shurd - check if this is still safe */
3096 			while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0])
3097 				rxq->ifr_cq_cidx -= scctx->isc_nrxd[0];
3098 			/* was this only a completion queue message? */
3099 			if (__predict_false(ri.iri_nfrags == 0))
3100 				continue;
3101 		}
3102 		MPASS(ri.iri_nfrags != 0);
3103 		MPASS(ri.iri_len != 0);
3104 
3105 		/* will advance the cidx on the corresponding free lists */
3106 		m = iflib_rxd_pkt_get(rxq, &ri);
3107 		avail--;
3108 		budget_left--;
3109 		if (avail == 0 && budget_left)
3110 			avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left);
3111 
3112 		if (__predict_false(m == NULL))
3113 			continue;
3114 
3115 #ifndef __NO_STRICT_ALIGNMENT
3116 		if (!IP_ALIGNED(m) && (m = iflib_fixup_rx(m)) == NULL)
3117 			continue;
3118 #endif
3119 #if defined(INET6) || defined(INET)
3120 		if (lro_enabled) {
3121 			tcp_lro_queue_mbuf(&rxq->ifr_lc, m);
3122 			continue;
3123 		}
3124 #endif
3125 
3126 		if (mh == NULL)
3127 			mh = mt = m;
3128 		else {
3129 			mt->m_nextpkt = m;
3130 			mt = m;
3131 		}
3132 	}
3133 	CURVNET_RESTORE();
3134 	/* make sure that we can refill faster than drain */
3135 	for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
3136 		retval |= iflib_fl_refill_all(ctx, fl);
3137 
3138 	if (mh != NULL) {
3139 		if_input(ifp, mh);
3140 		DBG_COUNTER_INC(rx_if_input);
3141 	}
3142 
3143 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
3144 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
3145 
3146 	/*
3147 	 * Flush any outstanding LRO work
3148 	 */
3149 #if defined(INET6) || defined(INET)
3150 	tcp_lro_flush_all(&rxq->ifr_lc);
3151 #endif
3152 	if (avail != 0 || iflib_rxd_avail(ctx, rxq, *cidxp, 1) != 0)
3153 		retval |= IFLIB_RXEOF_MORE;
3154 	return (retval);
3155 err:
3156 	STATE_LOCK(ctx);
3157 	ctx->ifc_flags |= IFC_DO_RESET;
3158 	iflib_admin_intr_deferred(ctx);
3159 	STATE_UNLOCK(ctx);
3160 	return (0);
3161 }
3162 
3163 #define TXD_NOTIFY_COUNT(txq) (((txq)->ift_size / (txq)->ift_update_freq) - 1)
3164 static inline qidx_t
3165 txq_max_db_deferred(iflib_txq_t txq, qidx_t in_use)
3166 {
3167 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
3168 	qidx_t minthresh = txq->ift_size / 8;
3169 	if (in_use > 4 * minthresh)
3170 		return (notify_count);
3171 	if (in_use > 2 * minthresh)
3172 		return (notify_count >> 1);
3173 	if (in_use > minthresh)
3174 		return (notify_count >> 3);
3175 	return (0);
3176 }
3177 
3178 static inline qidx_t
3179 txq_max_rs_deferred(iflib_txq_t txq)
3180 {
3181 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
3182 	qidx_t minthresh = txq->ift_size / 8;
3183 	if (txq->ift_in_use > 4 * minthresh)
3184 		return (notify_count);
3185 	if (txq->ift_in_use > 2 * minthresh)
3186 		return (notify_count >> 1);
3187 	if (txq->ift_in_use > minthresh)
3188 		return (notify_count >> 2);
3189 	return (2);
3190 }
3191 
3192 #define M_CSUM_FLAGS(m)		((m)->m_pkthdr.csum_flags)
3193 #define M_HAS_VLANTAG(m)	(m->m_flags & M_VLANTAG)
3194 
3195 #define TXQ_MAX_DB_DEFERRED(txq, in_use)	txq_max_db_deferred((txq), (in_use))
3196 #define TXQ_MAX_RS_DEFERRED(txq)	txq_max_rs_deferred(txq)
3197 #define TXQ_MAX_DB_CONSUMED(size)	(size >> 4)
3198 
3199 /* forward compatibility for cxgb */
3200 #define FIRST_QSET(ctx) 0
3201 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets)
3202 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets)
3203 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx))
3204 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments))
3205 
3206 static inline bool
3207 iflib_txd_db_check(iflib_txq_t txq, int ring)
3208 {
3209 	if_ctx_t ctx = txq->ift_ctx;
3210 	qidx_t dbval, max;
3211 
3212 	max = TXQ_MAX_DB_DEFERRED(txq, txq->ift_in_use);
3213 
3214 	/* force || threshold exceeded || at the edge of the ring */
3215 	if (ring || (txq->ift_db_pending >= max) || (TXQ_AVAIL(txq) <= MAX_TX_DESC(ctx))) {
3216 
3217 		/*
3218 		 * 'npending' is used if the card's doorbell is in terms of the number of descriptors
3219 		 * pending flush (BRCM). 'pidx' is used in cases where the card's doorbeel uses the
3220 		 * producer index explicitly (INTC).
3221 		 */
3222 		dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx;
3223 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3224 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
3225 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval);
3226 
3227 		/*
3228 		 * Absent bugs there are zero packets pending so reset pending counts to zero.
3229 		 */
3230 		txq->ift_db_pending = txq->ift_npending = 0;
3231 		return (true);
3232 	}
3233 	return (false);
3234 }
3235 
3236 #ifdef PKT_DEBUG
3237 static void
3238 print_pkt(if_pkt_info_t pi)
3239 {
3240 	printf("pi len:  %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n",
3241 	    pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx);
3242 	printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n",
3243 	    pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag);
3244 	printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n",
3245 	    pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto);
3246 }
3247 #endif
3248 
3249 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO)
3250 #define IS_TX_OFFLOAD4(pi) ((pi)->ipi_csum_flags & (CSUM_IP_TCP | CSUM_IP_TSO))
3251 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO)
3252 #define IS_TX_OFFLOAD6(pi) ((pi)->ipi_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_TSO))
3253 
3254 /**
3255  * Parses out ethernet header information in the given mbuf.
3256  * Returns in pi: ipi_etype (EtherType) and ipi_ehdrlen (Ethernet header length)
3257  *
3258  * This will account for the VLAN header if present.
3259  *
3260  * XXX: This doesn't handle QinQ, which could prevent TX offloads for those
3261  * types of packets.
3262  */
3263 static int
3264 iflib_parse_ether_header(if_pkt_info_t pi, struct mbuf **mp, uint64_t *pullups)
3265 {
3266 	struct ether_vlan_header *eh;
3267 	struct mbuf *m;
3268 
3269 	m = *mp;
3270 	if (__predict_false(m->m_len < sizeof(*eh))) {
3271 		(*pullups)++;
3272 		if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL))
3273 			return (ENOMEM);
3274 	}
3275 	eh = mtod(m, struct ether_vlan_header *);
3276 	if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
3277 		pi->ipi_etype = ntohs(eh->evl_proto);
3278 		pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
3279 	} else {
3280 		pi->ipi_etype = ntohs(eh->evl_encap_proto);
3281 		pi->ipi_ehdrlen = ETHER_HDR_LEN;
3282 	}
3283 	*mp = m;
3284 
3285 	return (0);
3286 }
3287 
3288 /**
3289  * Parse up to the L3 header and extract IPv4/IPv6 header information into pi.
3290  * Currently this information includes: IP ToS value, IP header version/presence
3291  *
3292  * This is missing some checks and doesn't edit the packet content as it goes,
3293  * unlike iflib_parse_header(), in order to keep the amount of code here minimal.
3294  */
3295 static int
3296 iflib_parse_header_partial(if_pkt_info_t pi, struct mbuf **mp, uint64_t *pullups)
3297 {
3298 	struct mbuf *m;
3299 	int err;
3300 
3301 	*pullups = 0;
3302 	m = *mp;
3303 	if (!M_WRITABLE(m)) {
3304 		m = m_dup(m, M_NOWAIT);
3305 		m_freem(*mp);
3306 		DBG_COUNTER_INC(tx_frees);
3307 		*mp = m;
3308 		if (m == NULL)
3309 			return (ENOMEM);
3310 	}
3311 
3312 	/* Fills out pi->ipi_etype */
3313 	err = iflib_parse_ether_header(pi, mp, pullups);
3314 	if (err)
3315 		return (err);
3316 	m = *mp;
3317 
3318 	switch (pi->ipi_etype) {
3319 #ifdef INET
3320 	case ETHERTYPE_IP:
3321 	{
3322 		struct mbuf *n;
3323 		struct ip *ip = NULL;
3324 		int miniplen;
3325 
3326 		miniplen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip));
3327 		if (__predict_false(m->m_len < miniplen)) {
3328 			/*
3329 			 * Check for common case where the first mbuf only contains
3330 			 * the Ethernet header
3331 			 */
3332 			if (m->m_len == pi->ipi_ehdrlen) {
3333 				n = m->m_next;
3334 				MPASS(n);
3335 				/* If next mbuf contains at least the minimal IP header, then stop */
3336 				if (n->m_len >= sizeof(*ip)) {
3337 					ip = (struct ip *)n->m_data;
3338 				} else {
3339 					(*pullups)++;
3340 					if (__predict_false((m = m_pullup(m, miniplen)) == NULL))
3341 						return (ENOMEM);
3342 					ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3343 				}
3344 			} else {
3345 				(*pullups)++;
3346 				if (__predict_false((m = m_pullup(m, miniplen)) == NULL))
3347 					return (ENOMEM);
3348 				ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3349 			}
3350 		} else {
3351 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3352 		}
3353 
3354 		/* Have the IPv4 header w/ no options here */
3355 		pi->ipi_ip_hlen = ip->ip_hl << 2;
3356 		pi->ipi_ipproto = ip->ip_p;
3357 		pi->ipi_ip_tos = ip->ip_tos;
3358 		pi->ipi_flags |= IPI_TX_IPV4;
3359 
3360 		break;
3361 	}
3362 #endif
3363 #ifdef INET6
3364 	case ETHERTYPE_IPV6:
3365 	{
3366 		struct ip6_hdr *ip6;
3367 
3368 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
3369 			(*pullups)++;
3370 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
3371 				return (ENOMEM);
3372 		}
3373 		ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
3374 
3375 		/* Have the IPv6 fixed header here */
3376 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
3377 		pi->ipi_ipproto = ip6->ip6_nxt;
3378 		pi->ipi_ip_tos = IPV6_TRAFFIC_CLASS(ip6);
3379 		pi->ipi_flags |= IPI_TX_IPV6;
3380 
3381 		break;
3382 	}
3383 #endif
3384 	default:
3385 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
3386 		pi->ipi_ip_hlen = 0;
3387 		break;
3388 	}
3389 	*mp = m;
3390 
3391 	return (0);
3392 
3393 }
3394 
3395 static int
3396 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp)
3397 {
3398 	if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx;
3399 	struct mbuf *m;
3400 	int err;
3401 
3402 	m = *mp;
3403 	if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) &&
3404 	    M_WRITABLE(m) == 0) {
3405 		m = m_dup(m, M_NOWAIT);
3406 		m_freem(*mp);
3407 		DBG_COUNTER_INC(tx_frees);
3408 		*mp = m;
3409 		if (m == NULL)
3410 			return (ENOMEM);
3411 	}
3412 
3413 	/* Fills out pi->ipi_etype */
3414 	err = iflib_parse_ether_header(pi, mp, &txq->ift_pullups);
3415 	if (__predict_false(err))
3416 		return (err);
3417 	m = *mp;
3418 
3419 	switch (pi->ipi_etype) {
3420 #ifdef INET
3421 	case ETHERTYPE_IP:
3422 	{
3423 		struct ip *ip;
3424 		struct tcphdr *th;
3425 		uint8_t hlen;
3426 
3427 		hlen = pi->ipi_ehdrlen + sizeof(*ip);
3428 		if (__predict_false(m->m_len < hlen)) {
3429 			txq->ift_pullups++;
3430 			if (__predict_false((m = m_pullup(m, hlen)) == NULL))
3431 				return (ENOMEM);
3432 		}
3433 		ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3434 		hlen = pi->ipi_ehdrlen + (ip->ip_hl << 2);
3435 		if (ip->ip_p == IPPROTO_TCP) {
3436 			hlen += sizeof(*th);
3437 			th = (struct tcphdr *)((char *)ip + (ip->ip_hl << 2));
3438 		} else if (ip->ip_p == IPPROTO_UDP) {
3439 			hlen += sizeof(struct udphdr);
3440 		}
3441 		if (__predict_false(m->m_len < hlen)) {
3442 			txq->ift_pullups++;
3443 			if ((m = m_pullup(m, hlen)) == NULL)
3444 				return (ENOMEM);
3445 			/* reset pointers after pullup */
3446 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3447 			th = (struct tcphdr *)((char *)ip + (ip->ip_hl << 2));
3448 		}
3449 		pi->ipi_ip_hlen = ip->ip_hl << 2;
3450 		pi->ipi_ipproto = ip->ip_p;
3451 		pi->ipi_ip_tos = ip->ip_tos;
3452 		pi->ipi_flags |= IPI_TX_IPV4;
3453 
3454 		/* TCP checksum offload may require TCP header length */
3455 		if (IS_TX_OFFLOAD4(pi)) {
3456 			if (__predict_true(pi->ipi_ipproto == IPPROTO_TCP)) {
3457 				pi->ipi_tcp_hflags = tcp_get_flags(th);
3458 				pi->ipi_tcp_hlen = th->th_off << 2;
3459 				pi->ipi_tcp_seq = th->th_seq;
3460 			}
3461 			if (IS_TSO4(pi)) {
3462 				MPASS(ip->ip_p == IPPROTO_TCP);
3463 				/*
3464 				 * TSO always requires hardware checksum offload.
3465 				 */
3466 				pi->ipi_csum_flags |= (CSUM_IP_TCP | CSUM_IP);
3467 				th->th_sum = in_pseudo(ip->ip_src.s_addr,
3468 						       ip->ip_dst.s_addr, htons(IPPROTO_TCP));
3469 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3470 				if (sctx->isc_flags & IFLIB_TSO_INIT_IP) {
3471 					ip->ip_sum = 0;
3472 					ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz);
3473 				}
3474 			}
3475 		}
3476 		if ((sctx->isc_flags & IFLIB_NEED_ZERO_CSUM) && (pi->ipi_csum_flags & CSUM_IP))
3477 			ip->ip_sum = 0;
3478 
3479 		break;
3480 	}
3481 #endif
3482 #ifdef INET6
3483 	case ETHERTYPE_IPV6:
3484 	{
3485 		struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
3486 		struct tcphdr *th;
3487 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
3488 
3489 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
3490 			txq->ift_pullups++;
3491 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
3492 				return (ENOMEM);
3493 			/* reset pointers after pullup */
3494 			ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
3495 		}
3496 		th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
3497 
3498 		/* XXX-BZ this will go badly in case of ext hdrs. */
3499 		pi->ipi_ipproto = ip6->ip6_nxt;
3500 		pi->ipi_ip_tos = IPV6_TRAFFIC_CLASS(ip6);
3501 		pi->ipi_flags |= IPI_TX_IPV6;
3502 
3503 		/* TCP checksum offload may require TCP header length */
3504 		if (IS_TX_OFFLOAD6(pi)) {
3505 			if (pi->ipi_ipproto == IPPROTO_TCP) {
3506 				if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) {
3507 					txq->ift_pullups++;
3508 					if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL))
3509 						return (ENOMEM);
3510 					/* reset pointers after pullup */
3511 					ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
3512 					th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
3513 				}
3514 				pi->ipi_tcp_hflags = tcp_get_flags(th);
3515 				pi->ipi_tcp_hlen = th->th_off << 2;
3516 				pi->ipi_tcp_seq = th->th_seq;
3517 			}
3518 			if (IS_TSO6(pi)) {
3519 				MPASS(ip6->ip6_nxt == IPPROTO_TCP);
3520 				/*
3521 				 * TSO always requires hardware checksum offload.
3522 				 */
3523 				pi->ipi_csum_flags |= CSUM_IP6_TCP;
3524 				th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0);
3525 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3526 			}
3527 		}
3528 		break;
3529 	}
3530 #endif
3531 	default:
3532 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
3533 		pi->ipi_ip_hlen = 0;
3534 		break;
3535 	}
3536 	*mp = m;
3537 
3538 	return (0);
3539 }
3540 
3541 /*
3542  * If dodgy hardware rejects the scatter gather chain we've handed it
3543  * we'll need to remove the mbuf chain from ifsg_m[] before we can add the
3544  * m_defrag'd mbufs
3545  */
3546 static __noinline struct mbuf *
3547 iflib_remove_mbuf(iflib_txq_t txq)
3548 {
3549 	int ntxd, pidx;
3550 	struct mbuf *m, **ifsd_m;
3551 
3552 	ifsd_m = txq->ift_sds.ifsd_m;
3553 	ntxd = txq->ift_size;
3554 	pidx = txq->ift_pidx & (ntxd - 1);
3555 	ifsd_m = txq->ift_sds.ifsd_m;
3556 	m = IFLIB_GET_MBUF(ifsd_m[pidx]);
3557 	ifsd_m[pidx] = NULL;
3558 	bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[pidx]);
3559 	if (txq->ift_sds.ifsd_tso_map != NULL)
3560 		bus_dmamap_unload(txq->ift_tso_buf_tag,
3561 		    txq->ift_sds.ifsd_tso_map[pidx]);
3562 #if MEMORY_LOGGING
3563 	txq->ift_dequeued++;
3564 #endif
3565 	return (m);
3566 }
3567 
3568 /*
3569  * Pad an mbuf to ensure a minimum ethernet frame size.
3570  * min_frame_size is the frame size (less CRC) to pad the mbuf to
3571  */
3572 static __noinline int
3573 iflib_ether_pad(device_t dev, struct mbuf **m_head, uint16_t min_frame_size)
3574 {
3575 	/*
3576 	 * 18 is enough bytes to pad an ARP packet to 46 bytes, and
3577 	 * and ARP message is the smallest common payload I can think of
3578 	 */
3579 	static char pad[18];	/* just zeros */
3580 	int n;
3581 	struct mbuf *new_head;
3582 
3583 	if (!M_WRITABLE(*m_head)) {
3584 		new_head = m_dup(*m_head, M_NOWAIT);
3585 		m_freem(*m_head);
3586 		*m_head = new_head;
3587 		if (new_head == NULL) {
3588 			device_printf(dev, "cannot pad short frame, m_dup() failed");
3589 			DBG_COUNTER_INC(encap_pad_mbuf_fail);
3590 			DBG_COUNTER_INC(tx_frees);
3591 			return (ENOMEM);
3592 		}
3593 	}
3594 
3595 	for (n = min_frame_size - (*m_head)->m_pkthdr.len;
3596 	     n > 0; n -= sizeof(pad))
3597 		if (!m_append(*m_head, min(n, sizeof(pad)), pad))
3598 			break;
3599 
3600 	if (n > 0) {
3601 		m_freem(*m_head);
3602 		*m_head = NULL;
3603 		device_printf(dev, "cannot pad short frame\n");
3604 		DBG_COUNTER_INC(encap_pad_mbuf_fail);
3605 		DBG_COUNTER_INC(tx_frees);
3606 		return (ENOMEM);
3607 	}
3608 
3609 	return (0);
3610 }
3611 
3612 static int
3613 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp, int *obytes, int *opkts)
3614 {
3615 	if_ctx_t		ctx;
3616 	if_shared_ctx_t		sctx;
3617 	if_softc_ctx_t		scctx;
3618 	bus_dma_tag_t		buf_tag;
3619 	bus_dma_segment_t	*segs;
3620 	struct mbuf		*m_head, **ifsd_m;
3621 	bus_dmamap_t		map;
3622 	struct if_pkt_info	pi;
3623 	uintptr_t		flags;
3624 	int remap = 0;
3625 	int err, nsegs, ndesc, max_segs, pidx;
3626 
3627 	ctx = txq->ift_ctx;
3628 	sctx = ctx->ifc_sctx;
3629 	scctx = &ctx->ifc_softc_ctx;
3630 	segs = txq->ift_segs;
3631 	m_head = *m_headp;
3632 	map = NULL;
3633 
3634 	/*
3635 	 * If we're doing TSO the next descriptor to clean may be quite far ahead
3636 	 */
3637 	pidx = txq->ift_pidx;
3638 	map = txq->ift_sds.ifsd_map[pidx];
3639 	ifsd_m = txq->ift_sds.ifsd_m;
3640 
3641 	if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
3642 		buf_tag = txq->ift_tso_buf_tag;
3643 		max_segs = scctx->isc_tx_tso_segments_max;
3644 		map = txq->ift_sds.ifsd_tso_map[pidx];
3645 		MPASS(buf_tag != NULL);
3646 		MPASS(max_segs > 0);
3647 		flags = IFLIB_TSO;
3648 	} else {
3649 		buf_tag = txq->ift_buf_tag;
3650 		max_segs = scctx->isc_tx_nsegments;
3651 		map = txq->ift_sds.ifsd_map[pidx];
3652 		flags = IFLIB_NO_TSO;
3653 	}
3654 	if ((sctx->isc_flags & IFLIB_NEED_ETHER_PAD) &&
3655 	    __predict_false(m_head->m_pkthdr.len < scctx->isc_min_frame_size)) {
3656 		err = iflib_ether_pad(ctx->ifc_dev, m_headp, scctx->isc_min_frame_size);
3657 		if (err) {
3658 			DBG_COUNTER_INC(encap_txd_encap_fail);
3659 			return (err);
3660 		}
3661 	}
3662 	m_head = *m_headp;
3663 
3664 	memset(&pi, 0, sizeof(pi));
3665 	pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG | M_BCAST | M_MCAST));
3666 	pi.ipi_pidx = pidx;
3667 	pi.ipi_qsidx = txq->ift_id;
3668 	pi.ipi_len = m_head->m_pkthdr.len;
3669 	pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags;
3670 	pi.ipi_vtag = M_HAS_VLANTAG(m_head) ? m_head->m_pkthdr.ether_vtag : 0;
3671 
3672 	/* deliberate bitwise OR to make one condition */
3673 	if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) {
3674 		if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0)) {
3675 			DBG_COUNTER_INC(encap_txd_encap_fail);
3676 			return (err);
3677 		}
3678 		m_head = *m_headp;
3679 	}
3680 
3681 retry:
3682 	err = bus_dmamap_load_mbuf_sg(buf_tag, map, m_head, segs, &nsegs,
3683 	    BUS_DMA_NOWAIT);
3684 defrag:
3685 	if (__predict_false(err)) {
3686 		switch (err) {
3687 		case EFBIG:
3688 			/* try collapse once and defrag once */
3689 			if (remap == 0) {
3690 				m_head = m_collapse(*m_headp, M_NOWAIT, max_segs);
3691 				/* try defrag if collapsing fails */
3692 				if (m_head == NULL)
3693 					remap++;
3694 			}
3695 			if (remap == 1) {
3696 				txq->ift_mbuf_defrag++;
3697 				m_head = m_defrag(*m_headp, M_NOWAIT);
3698 			}
3699 			/*
3700 			 * remap should never be >1 unless bus_dmamap_load_mbuf_sg
3701 			 * failed to map an mbuf that was run through m_defrag
3702 			 */
3703 			MPASS(remap <= 1);
3704 			if (__predict_false(m_head == NULL || remap > 1))
3705 				goto defrag_failed;
3706 			remap++;
3707 			*m_headp = m_head;
3708 			goto retry;
3709 			break;
3710 		case ENOMEM:
3711 			/* FALLTHROUGH */
3712 		default:
3713 			txq->ift_no_tx_dma_setup++;
3714 			m_freem(*m_headp);
3715 			DBG_COUNTER_INC(tx_frees);
3716 			*m_headp = NULL;
3717 			break;
3718 		}
3719 		txq->ift_map_failed++;
3720 		DBG_COUNTER_INC(encap_load_mbuf_fail);
3721 		DBG_COUNTER_INC(encap_txd_encap_fail);
3722 		return (err);
3723 	}
3724 	ifsd_m[pidx] = IFLIB_SAVE_MBUF(m_head, flags);
3725 	if (m_head->m_pkthdr.csum_flags & CSUM_SND_TAG)
3726 		pi.ipi_mbuf = m_head;
3727 	else
3728 		pi.ipi_mbuf = NULL;
3729 	/*
3730 	 * XXX assumes a 1 to 1 relationship between segments and
3731 	 *        descriptors - this does not hold true on all drivers, e.g.
3732 	 *        cxgb
3733 	 */
3734 	if (__predict_false(nsegs > TXQ_AVAIL(txq))) {
3735 		iflib_completed_tx_reclaim_force(txq);
3736 		if (__predict_false(nsegs > TXQ_AVAIL(txq))) {
3737 			txq->ift_no_desc_avail++;
3738 			bus_dmamap_unload(buf_tag, map);
3739 			DBG_COUNTER_INC(encap_txq_avail_fail);
3740 			DBG_COUNTER_INC(encap_txd_encap_fail);
3741 			if (ctx->ifc_sysctl_simple_tx) {
3742 				*m_headp = m_head = iflib_remove_mbuf(txq);
3743 				m_freem(*m_headp);
3744 				DBG_COUNTER_INC(tx_frees);
3745 				*m_headp = NULL;
3746 			}
3747 			if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0)
3748 				GROUPTASK_ENQUEUE(&txq->ift_task);
3749 			return (ENOBUFS);
3750 		}
3751 	}
3752 	/*
3753 	 * On Intel cards we can greatly reduce the number of TX interrupts
3754 	 * we see by only setting report status on every Nth descriptor.
3755 	 * However, this also means that the driver will need to keep track
3756 	 * of the descriptors that RS was set on to check them for the DD bit.
3757 	 */
3758 	if (txq->ift_rs_pending + nsegs + 1 > TXQ_MAX_RS_DEFERRED(txq) ||
3759 	    iflib_no_tx_batch || (TXQ_AVAIL(txq) - nsegs) <= MAX_TX_DESC(ctx)) {
3760 		pi.ipi_flags |= IPI_TX_INTR;
3761 	}
3762 
3763 	pi.ipi_segs = segs;
3764 	pi.ipi_nsegs = nsegs;
3765 
3766 	MPASS(pidx >= 0 && pidx < txq->ift_size);
3767 #ifdef PKT_DEBUG
3768 	print_pkt(&pi);
3769 #endif
3770 	if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) {
3771 		bus_dmamap_sync(buf_tag, map, BUS_DMASYNC_PREWRITE);
3772 		DBG_COUNTER_INC(tx_encap);
3773 		MPASS(pi.ipi_new_pidx < txq->ift_size);
3774 
3775 		ndesc = pi.ipi_new_pidx - pi.ipi_pidx;
3776 		if (pi.ipi_new_pidx < pi.ipi_pidx) {
3777 			ndesc += txq->ift_size;
3778 			txq->ift_gen = 1;
3779 		}
3780 
3781 		if (pi.ipi_flags & IPI_TX_INTR)
3782 			txq->ift_rs_pending = 0;
3783 		else
3784 			txq->ift_rs_pending += ndesc;
3785 		/*
3786 		 * drivers can need up to ift_pad sentinels
3787 		 */
3788 		MPASS(ndesc <= pi.ipi_nsegs + txq->ift_pad);
3789 		MPASS(pi.ipi_new_pidx != pidx);
3790 		MPASS(ndesc > 0);
3791 		txq->ift_in_use += ndesc;
3792 		txq->ift_db_pending += ndesc;
3793 
3794 		/*
3795 		 * We update the last software descriptor again here because there may
3796 		 * be a sentinel and/or there may be more mbufs than segments
3797 		 */
3798 		txq->ift_pidx = pi.ipi_new_pidx;
3799 		txq->ift_npending += pi.ipi_ndescs;
3800 
3801 		/*
3802 		 * Update packets / bytes sent
3803 		 */
3804 		if (flags & IFLIB_TSO) {
3805 			int hlen = pi.ipi_ehdrlen + pi.ipi_ip_hlen + pi.ipi_tcp_hlen;
3806 			int tsolen = pi.ipi_len - hlen;
3807 			int nsegs = (tsolen + pi.ipi_tso_segsz - 1) / pi.ipi_tso_segsz;
3808 			*obytes += tsolen + nsegs * hlen;
3809 			*opkts += nsegs;
3810 		} else {
3811 			*obytes += pi.ipi_len;
3812 			*opkts += 1;
3813 		}
3814 	} else {
3815 		*m_headp = m_head = iflib_remove_mbuf(txq);
3816 		if (err == EFBIG) {
3817 			txq->ift_txd_encap_efbig++;
3818 			if (remap < 2) {
3819 				remap = 1;
3820 				goto defrag;
3821 			}
3822 			goto defrag_failed;
3823 		}
3824 		/* mp_ring assumes ENOBUFS means we didn't consume the mbuf */
3825 		if (err == ENOBUFS && !ctx->ifc_sysctl_simple_tx)
3826 			err = ENOMEM;
3827 		goto out_with_error;
3828 	}
3829 	/*
3830 	 * err can't possibly be non-zero here, so we don't neet to test it
3831 	 * to see if we need to DBG_COUNTER_INC(encap_txd_encap_fail).
3832 	 */
3833 	return (err);
3834 
3835 defrag_failed:
3836 	err = ENOMEM;
3837 	txq->ift_mbuf_defrag_failed++;
3838 out_with_error:
3839 	txq->ift_map_failed++;
3840 	m_freem(*m_headp);
3841 	DBG_COUNTER_INC(tx_frees);
3842 	*m_headp = NULL;
3843 	DBG_COUNTER_INC(encap_txd_encap_fail);
3844 	return (err);
3845 }
3846 
3847 static void
3848 iflib_tx_desc_free(iflib_txq_t txq, int n, struct mbuf **m_defer)
3849 {
3850 	uint32_t qsize, cidx, gen;
3851 	struct mbuf *m, **ifsd_m;
3852 	uintptr_t flags;
3853 
3854 	cidx = txq->ift_cidx;
3855 	gen = txq->ift_gen;
3856 	qsize = txq->ift_size;
3857 	ifsd_m =txq->ift_sds.ifsd_m;
3858 
3859 	while (n-- > 0) {
3860 		if ((m = IFLIB_GET_MBUF(ifsd_m[cidx])) != NULL) {
3861 			flags = IFLIB_GET_FLAGS(ifsd_m[cidx]);
3862 			MPASS(flags != 0);
3863 			if (flags & IFLIB_TSO) {
3864 				bus_dmamap_sync(txq->ift_tso_buf_tag,
3865 				    txq->ift_sds.ifsd_tso_map[cidx],
3866 				    BUS_DMASYNC_POSTWRITE);
3867 				bus_dmamap_unload(txq->ift_tso_buf_tag,
3868 				    txq->ift_sds.ifsd_tso_map[cidx]);
3869 			} else {
3870 				bus_dmamap_sync(txq->ift_buf_tag,
3871 				    txq->ift_sds.ifsd_map[cidx],
3872 				    BUS_DMASYNC_POSTWRITE);
3873 				bus_dmamap_unload(txq->ift_buf_tag,
3874 				    txq->ift_sds.ifsd_map[cidx]);
3875 			}
3876 			/* XXX we don't support any drivers that batch packets yet */
3877 			MPASS(m->m_nextpkt == NULL);
3878 			if (m_defer == NULL) {
3879 				m_freem(m);
3880 			} else if (m != NULL) {
3881 				*m_defer = m;
3882 				m_defer++;
3883 			}
3884 			ifsd_m[cidx] = NULL;
3885 #if MEMORY_LOGGING
3886 			txq->ift_dequeued++;
3887 #endif
3888 			DBG_COUNTER_INC(tx_frees);
3889 		}
3890 		if (__predict_false(++cidx == qsize)) {
3891 			cidx = 0;
3892 			gen = 0;
3893 		}
3894 	}
3895 	txq->ift_cidx = cidx;
3896 	txq->ift_gen = gen;
3897 }
3898 
3899 static __inline int
3900 iflib_txq_can_reclaim(iflib_txq_t txq)
3901 {
3902 	int reclaim, thresh;
3903 
3904 	thresh = txq->ift_reclaim_thresh;
3905 	KASSERT(thresh >= 0, ("invalid threshold to reclaim"));
3906 	MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size);
3907 
3908 	if (ticks <= (txq->ift_last_reclaim + txq->ift_reclaim_ticks) &&
3909 	    txq->ift_in_use < thresh)
3910 		return (false);
3911 	iflib_tx_credits_update(txq->ift_ctx, txq);
3912 	reclaim = DESC_RECLAIMABLE(txq);
3913 	if (reclaim <= thresh) {
3914 #ifdef INVARIANTS
3915 		if (iflib_verbose_debug) {
3916 			printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __func__,
3917 			    txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments,
3918 			    reclaim, thresh);
3919 		}
3920 #endif
3921 		return (0);
3922 	}
3923 	return (reclaim);
3924 }
3925 
3926 static __inline void
3927 _iflib_completed_tx_reclaim(iflib_txq_t txq, struct mbuf **m_defer, int reclaim)
3928 {
3929 	txq->ift_last_reclaim = ticks;
3930 	iflib_tx_desc_free(txq, reclaim, m_defer);
3931 	txq->ift_cleaned += reclaim;
3932 	txq->ift_in_use -= reclaim;
3933 }
3934 
3935 static __inline int
3936 iflib_completed_tx_reclaim(iflib_txq_t txq, struct mbuf **m_defer)
3937 {
3938 	int reclaim;
3939 
3940 	reclaim = iflib_txq_can_reclaim(txq);
3941 	if (reclaim == 0)
3942 		return (0);
3943 	_iflib_completed_tx_reclaim(txq, m_defer, reclaim);
3944 	return (reclaim);
3945 }
3946 
3947 /*
3948  * Reclaim any transmit descriptors possible, ignoring coalescing
3949  */
3950 static __inline void
3951 iflib_completed_tx_reclaim_force(iflib_txq_t txq)
3952 {
3953 	int reclaim;
3954 
3955 	iflib_tx_credits_update(txq->ift_ctx, txq);
3956 	reclaim = DESC_RECLAIMABLE(txq);
3957 	if (reclaim != 0)
3958 		_iflib_completed_tx_reclaim(txq, NULL, reclaim);
3959 }
3960 
3961 static struct mbuf **
3962 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset, int remaining)
3963 {
3964 	int next, size;
3965 	struct mbuf **items;
3966 
3967 	size = r->size;
3968 	next = (cidx + CACHE_PTR_INCREMENT) & (size - 1);
3969 	items = __DEVOLATILE(struct mbuf **, &r->items[0]);
3970 
3971 	prefetch(items[(cidx + offset) & (size - 1)]);
3972 	if (remaining > 1) {
3973 		prefetch2cachelines(&items[next]);
3974 		prefetch2cachelines(items[(cidx + offset + 1) & (size - 1)]);
3975 		prefetch2cachelines(items[(cidx + offset + 2) & (size - 1)]);
3976 		prefetch2cachelines(items[(cidx + offset + 3) & (size - 1)]);
3977 	}
3978 	return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (size - 1)]));
3979 }
3980 
3981 static void
3982 iflib_txq_check_drain(iflib_txq_t txq, int budget)
3983 {
3984 
3985 	ifmp_ring_check_drainage(txq->ift_br, budget);
3986 }
3987 
3988 static uint32_t
3989 iflib_txq_can_drain(struct ifmp_ring *r)
3990 {
3991 	iflib_txq_t txq = r->cookie;
3992 	if_ctx_t ctx = txq->ift_ctx;
3993 
3994 	if (TXQ_AVAIL(txq) > MAX_TX_DESC(ctx))
3995 		return (1);
3996 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3997 	    BUS_DMASYNC_POSTREAD);
3998 	return (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id,
3999 	    false));
4000 }
4001 
4002 static uint32_t
4003 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
4004 {
4005 	iflib_txq_t txq = r->cookie;
4006 	if_ctx_t ctx = txq->ift_ctx;
4007 	if_t ifp = ctx->ifc_ifp;
4008 	struct mbuf *m, **mp;
4009 	int avail, bytes_sent, consumed, count, err, i;
4010 	int mcast_sent, pkt_sent, reclaimed;
4011 	bool do_prefetch, rang, ring;
4012 
4013 	if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) ||
4014 			    !LINK_ACTIVE(ctx))) {
4015 		DBG_COUNTER_INC(txq_drain_notready);
4016 		return (0);
4017 	}
4018 	reclaimed = iflib_completed_tx_reclaim(txq, NULL);
4019 	rang = iflib_txd_db_check(txq, reclaimed && txq->ift_db_pending);
4020 	avail = IDXDIFF(pidx, cidx, r->size);
4021 
4022 	if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) {
4023 		/*
4024 		 * The driver is unloading so we need to free all pending packets.
4025 		 */
4026 		DBG_COUNTER_INC(txq_drain_flushing);
4027 		for (i = 0; i < avail; i++) {
4028 			if (__predict_true(r->items[(cidx + i) & (r->size - 1)] != (void *)txq))
4029 				m_freem(r->items[(cidx + i) & (r->size - 1)]);
4030 			r->items[(cidx + i) & (r->size - 1)] = NULL;
4031 		}
4032 		return (avail);
4033 	}
4034 
4035 	if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
4036 		CALLOUT_LOCK(txq);
4037 		callout_stop(&txq->ift_timer);
4038 		CALLOUT_UNLOCK(txq);
4039 		DBG_COUNTER_INC(txq_drain_oactive);
4040 		return (0);
4041 	}
4042 
4043 	consumed = mcast_sent = bytes_sent = pkt_sent = 0;
4044 	count = MIN(avail, TX_BATCH_SIZE);
4045 #ifdef INVARIANTS
4046 	if (iflib_verbose_debug)
4047 		printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __func__,
4048 		    avail, ctx->ifc_flags, TXQ_AVAIL(txq));
4049 #endif
4050 	do_prefetch = (ctx->ifc_flags & IFC_PREFETCH);
4051 	err = 0;
4052 	for (i = 0; i < count && TXQ_AVAIL(txq) >= MAX_TX_DESC(ctx); i++) {
4053 		int rem = do_prefetch ? count - i : 0;
4054 
4055 		mp = _ring_peek_one(r, cidx, i, rem);
4056 		MPASS(mp != NULL && *mp != NULL);
4057 
4058 		/*
4059 		 * Completion interrupts will use the address of the txq
4060 		 * as a sentinel to enqueue _something_ in order to acquire
4061 		 * the lock on the mp_ring (there's no direct lock call).
4062 		 * We obviously whave to check for these sentinel cases
4063 		 * and skip them.
4064 		 */
4065 		if (__predict_false(*mp == (struct mbuf *)txq)) {
4066 			consumed++;
4067 			continue;
4068 		}
4069 		err = iflib_encap(txq, mp, &bytes_sent, &pkt_sent);
4070 		if (__predict_false(err)) {
4071 			/* no room - bail out */
4072 			if (err == ENOBUFS)
4073 				break;
4074 			consumed++;
4075 			/* we can't send this packet - skip it */
4076 			continue;
4077 		}
4078 		consumed++;
4079 		m = *mp;
4080 		DBG_COUNTER_INC(tx_sent);
4081 		mcast_sent += !!(m->m_flags & M_MCAST);
4082 
4083 		if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING)))
4084 			break;
4085 		ETHER_BPF_MTAP(ifp, m);
4086 		rang = iflib_txd_db_check(txq, false);
4087 	}
4088 
4089 	/* deliberate use of bitwise or to avoid gratuitous short-circuit */
4090 	ring = rang ? false  : (iflib_min_tx_latency | err | (!!txq->ift_reclaim_thresh));
4091 	iflib_txd_db_check(txq, ring);
4092 	if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
4093 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
4094 	if (mcast_sent)
4095 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
4096 #ifdef INVARIANTS
4097 	if (iflib_verbose_debug)
4098 		printf("consumed=%d\n", consumed);
4099 #endif
4100 	return (consumed);
4101 }
4102 
4103 static uint32_t
4104 iflib_txq_drain_always(struct ifmp_ring *r)
4105 {
4106 	return (1);
4107 }
4108 
4109 static uint32_t
4110 iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
4111 {
4112 	int i, avail;
4113 	struct mbuf **mp;
4114 	iflib_txq_t txq;
4115 
4116 	txq = r->cookie;
4117 
4118 	CALLOUT_LOCK(txq);
4119 	callout_stop(&txq->ift_timer);
4120 	CALLOUT_UNLOCK(txq);
4121 
4122 	avail = IDXDIFF(pidx, cidx, r->size);
4123 	for (i = 0; i < avail; i++) {
4124 		mp = _ring_peek_one(r, cidx, i, avail - i);
4125 		if (__predict_false(*mp == (struct mbuf *)txq))
4126 			continue;
4127 		m_freem(*mp);
4128 		DBG_COUNTER_INC(tx_frees);
4129 	}
4130 	MPASS(ifmp_ring_is_stalled(r) == 0);
4131 	return (avail);
4132 }
4133 
4134 static void
4135 iflib_ifmp_purge(iflib_txq_t txq)
4136 {
4137 	struct ifmp_ring *r;
4138 
4139 	r = txq->ift_br;
4140 	r->drain = iflib_txq_drain_free;
4141 	r->can_drain = iflib_txq_drain_always;
4142 
4143 	ifmp_ring_check_drainage(r, r->size);
4144 
4145 	r->drain = iflib_txq_drain;
4146 	r->can_drain = iflib_txq_can_drain;
4147 }
4148 
4149 static void
4150 _task_fn_tx(void *context)
4151 {
4152 	iflib_txq_t txq = context;
4153 	if_ctx_t ctx = txq->ift_ctx;
4154 	if_t ifp = ctx->ifc_ifp;
4155 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
4156 
4157 #ifdef IFLIB_DIAGNOSTICS
4158 	txq->ift_cpu_exec_count[curcpu]++;
4159 #endif
4160 	if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
4161 		return;
4162 #ifdef DEV_NETMAP
4163 	if ((if_getcapenable(ifp) & IFCAP_NETMAP) &&
4164 	    netmap_tx_irq(ifp, txq->ift_id))
4165 		goto skip_ifmp;
4166 #endif
4167         if (ctx->ifc_sysctl_simple_tx) {
4168                 mtx_lock(&txq->ift_mtx);
4169                 (void)iflib_completed_tx_reclaim(txq, NULL);
4170                 mtx_unlock(&txq->ift_mtx);
4171                 goto skip_ifmp;
4172         }
4173 #ifdef ALTQ
4174 	if (if_altq_is_enabled(ifp))
4175 		iflib_altq_if_start(ifp);
4176 #endif
4177 	if (txq->ift_db_pending)
4178 		ifmp_ring_enqueue(txq->ift_br, (void **)&txq, 1, TX_BATCH_SIZE, abdicate);
4179 	else if (!abdicate)
4180 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
4181 	/*
4182 	 * When abdicating, we always need to check drainage, not just when we don't enqueue
4183 	 */
4184 	if (abdicate)
4185 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
4186 
4187 skip_ifmp:
4188 	if (ctx->ifc_flags & IFC_LEGACY)
4189 		IFDI_INTR_ENABLE(ctx);
4190 	else
4191 		IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
4192 }
4193 
4194 static void
4195 _task_fn_rx(void *context)
4196 {
4197 	iflib_rxq_t rxq = context;
4198 	if_ctx_t ctx = rxq->ifr_ctx;
4199 	uint8_t more;
4200 	uint16_t budget;
4201 #ifdef DEV_NETMAP
4202 	u_int work = 0;
4203 	int nmirq;
4204 #endif
4205 
4206 #ifdef IFLIB_DIAGNOSTICS
4207 	rxq->ifr_cpu_exec_count[curcpu]++;
4208 #endif
4209 	DBG_COUNTER_INC(task_fn_rxs);
4210 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
4211 		return;
4212 #ifdef DEV_NETMAP
4213 	nmirq = netmap_rx_irq(ctx->ifc_ifp, rxq->ifr_id, &work);
4214 	if (nmirq != NM_IRQ_PASS) {
4215 		more = (nmirq == NM_IRQ_RESCHED) ? IFLIB_RXEOF_MORE : 0;
4216 		goto skip_rxeof;
4217 	}
4218 #endif
4219 	budget = ctx->ifc_sysctl_rx_budget;
4220 	if (budget == 0)
4221 		budget = 16;	/* XXX */
4222 	more = iflib_rxeof(rxq, budget);
4223 #ifdef DEV_NETMAP
4224 skip_rxeof:
4225 #endif
4226 	if ((more & IFLIB_RXEOF_MORE) == 0) {
4227 		if (ctx->ifc_flags & IFC_LEGACY)
4228 			IFDI_INTR_ENABLE(ctx);
4229 		else
4230 			IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
4231 		DBG_COUNTER_INC(rx_intr_enables);
4232 	}
4233 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
4234 		return;
4235 
4236 	if (more & IFLIB_RXEOF_MORE)
4237 		GROUPTASK_ENQUEUE(&rxq->ifr_task);
4238 	else if (more & IFLIB_RXEOF_EMPTY)
4239 		callout_reset_curcpu(&rxq->ifr_watchdog, 1, &_task_fn_rx_watchdog, rxq);
4240 }
4241 
4242 static void
4243 _task_fn_admin(void *context, int pending)
4244 {
4245 	if_ctx_t ctx = context;
4246 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
4247 	iflib_txq_t txq;
4248 	int i;
4249 	bool oactive, running, do_reset, do_reset_if_up, do_watchdog;
4250 	bool in_detach;
4251 
4252 	STATE_LOCK(ctx);
4253 	running = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING);
4254 	oactive = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE);
4255 	do_reset = (ctx->ifc_flags & IFC_DO_RESET);
4256 	do_reset_if_up = (ctx->ifc_flags & IFC_DO_RESET_IF_UP);
4257 	do_watchdog = (ctx->ifc_flags & IFC_DO_WATCHDOG);
4258 	in_detach = (ctx->ifc_flags & IFC_IN_DETACH);
4259 	ctx->ifc_flags &= ~(IFC_DO_RESET | IFC_DO_RESET_IF_UP |
4260 	    IFC_DO_WATCHDOG);
4261 	STATE_UNLOCK(ctx);
4262 
4263 	if ((!running && !oactive) && !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
4264 		return;
4265 	if (in_detach)
4266 		return;
4267 	KFAIL_POINT_CODE_COND(_debug_fail_point_iflib,
4268 	    admin_task_after_detach_check,
4269 	    iflib_admin_task_fail_device[0] != '\0' &&
4270 	    strcmp(device_get_nameunit(ctx->ifc_dev),
4271 	    iflib_admin_task_fail_device) == 0, FAIL_POINT_NONSLEEPABLE, {});
4272 
4273 	CTX_LOCK(ctx);
4274 	if (!do_reset && do_reset_if_up &&
4275 	    (if_getflags(ctx->ifc_ifp) & IFF_UP) != 0)
4276 		do_reset = true;
4277 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
4278 		CALLOUT_LOCK(txq);
4279 		callout_stop(&txq->ift_timer);
4280 		CALLOUT_UNLOCK(txq);
4281 	}
4282 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_ADMINCQ)
4283 		IFDI_ADMIN_COMPLETION_HANDLE(ctx);
4284 	if (do_watchdog) {
4285 		ctx->ifc_tx_watchdog_events++;
4286 		IFDI_WATCHDOG_RESET(ctx);
4287 	}
4288 	IFDI_UPDATE_ADMIN_STATUS(ctx);
4289 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
4290 		callout_reset_on(&txq->ift_timer, iflib_timer_default, iflib_timer, txq,
4291 		    txq->ift_timer.c_cpu);
4292 	}
4293 	IFDI_LINK_INTR_ENABLE(ctx);
4294 	if (do_reset)
4295 		iflib_if_init_locked(ctx);
4296 	CTX_UNLOCK(ctx);
4297 
4298 	if (LINK_ACTIVE(ctx) == 0)
4299 		return;
4300 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
4301 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
4302 }
4303 
4304 static void
4305 _task_fn_iov(void *context, int pending)
4306 {
4307 	if_ctx_t ctx = context;
4308 
4309 	if (iflib_in_detach(ctx))
4310 		return;
4311 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING) &&
4312 	    !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
4313 		return;
4314 
4315 	CTX_LOCK(ctx);
4316 	IFDI_VFLR_HANDLE(ctx);
4317 	CTX_UNLOCK(ctx);
4318 }
4319 
4320 static int
4321 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)
4322 {
4323 	int err;
4324 	if_int_delay_info_t info;
4325 	if_ctx_t ctx;
4326 
4327 	info = (if_int_delay_info_t)arg1;
4328 	ctx = info->iidi_ctx;
4329 	info->iidi_req = req;
4330 	info->iidi_oidp = oidp;
4331 	CTX_LOCK(ctx);
4332 	err = IFDI_SYSCTL_INT_DELAY(ctx, info);
4333 	CTX_UNLOCK(ctx);
4334 	return (err);
4335 }
4336 
4337 /*********************************************************************
4338  *
4339  *  IFNET FUNCTIONS
4340  *
4341  **********************************************************************/
4342 
4343 static void
4344 iflib_if_init_locked(if_ctx_t ctx)
4345 {
4346 	iflib_stop(ctx);
4347 	iflib_init_locked(ctx);
4348 }
4349 
4350 static void
4351 iflib_if_init(void *arg)
4352 {
4353 	if_ctx_t ctx = arg;
4354 
4355 	CTX_LOCK(ctx);
4356 	iflib_if_init_locked(ctx);
4357 	CTX_UNLOCK(ctx);
4358 }
4359 
4360 static int
4361 iflib_if_transmit(if_t ifp, struct mbuf *m)
4362 {
4363 	if_ctx_t ctx = if_getsoftc(ifp);
4364 	iflib_txq_t txq;
4365 	int err, qidx;
4366 	int abdicate;
4367 
4368 	if (__predict_false((if_getdrvflags(ifp) & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) {
4369 		DBG_COUNTER_INC(tx_frees);
4370 		m_freem(m);
4371 		return (ENETDOWN);
4372 	}
4373 
4374 	MPASS(m->m_nextpkt == NULL);
4375 	/* ALTQ-enabled interfaces always use queue 0. */
4376 	qidx = 0;
4377 	/* Use driver-supplied queue selection method if it exists */
4378 	if (ctx->isc_txq_select_v2) {
4379 		struct if_pkt_info pi;
4380 		uint64_t early_pullups = 0;
4381 		memset(&pi, 0, sizeof(pi));
4382 
4383 		err = iflib_parse_header_partial(&pi, &m, &early_pullups);
4384 		if (__predict_false(err != 0)) {
4385 			/* Assign pullups for bad pkts to default queue */
4386 			ctx->ifc_txqs[0].ift_pullups += early_pullups;
4387 			DBG_COUNTER_INC(encap_txd_encap_fail);
4388 			return (err);
4389 		}
4390 		/* Let driver make queueing decision */
4391 		qidx = ctx->isc_txq_select_v2(ctx->ifc_softc, m, &pi);
4392 		ctx->ifc_txqs[qidx].ift_pullups += early_pullups;
4393 	}
4394 	/* Backwards compatibility w/ simpler queue select */
4395 	else if (ctx->isc_txq_select)
4396 		qidx = ctx->isc_txq_select(ctx->ifc_softc, m);
4397 	/* If not, use iflib's standard method */
4398 	else if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m) && !if_altq_is_enabled(ifp))
4399 		qidx = QIDX(ctx, m);
4400 
4401 	/* Set TX queue */
4402 	txq = &ctx->ifc_txqs[qidx];
4403 
4404 #ifdef DRIVER_BACKPRESSURE
4405 	if (txq->ift_closed) {
4406 		while (m != NULL) {
4407 			next = m->m_nextpkt;
4408 			m->m_nextpkt = NULL;
4409 			m_freem(m);
4410 			DBG_COUNTER_INC(tx_frees);
4411 			m = next;
4412 		}
4413 		return (ENOBUFS);
4414 	}
4415 #endif
4416 #ifdef notyet
4417 	qidx = count = 0;
4418 	mp = marr;
4419 	next = m;
4420 	do {
4421 		count++;
4422 		next = next->m_nextpkt;
4423 	} while (next != NULL);
4424 
4425 	if (count > nitems(marr))
4426 		if ((mp = malloc(count * sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) {
4427 			/* XXX check nextpkt */
4428 			m_freem(m);
4429 			/* XXX simplify for now */
4430 			DBG_COUNTER_INC(tx_frees);
4431 			return (ENOBUFS);
4432 		}
4433 	for (next = m, i = 0; next != NULL; i++) {
4434 		mp[i] = next;
4435 		next = next->m_nextpkt;
4436 		mp[i]->m_nextpkt = NULL;
4437 	}
4438 #endif
4439 	DBG_COUNTER_INC(tx_seen);
4440 	abdicate = ctx->ifc_sysctl_tx_abdicate;
4441 
4442 	err = ifmp_ring_enqueue(txq->ift_br, (void **)&m, 1, TX_BATCH_SIZE, abdicate);
4443 
4444 	if (abdicate)
4445 		GROUPTASK_ENQUEUE(&txq->ift_task);
4446 	if (err) {
4447 		if (!abdicate)
4448 			GROUPTASK_ENQUEUE(&txq->ift_task);
4449 		/* support forthcoming later */
4450 #ifdef DRIVER_BACKPRESSURE
4451 		txq->ift_closed = TRUE;
4452 #endif
4453 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
4454 		m_freem(m);
4455 		DBG_COUNTER_INC(tx_frees);
4456 		if (err == ENOBUFS)
4457 			if_inc_counter(ifp, IFCOUNTER_OQDROPS, 1);
4458 		else
4459 			if_inc_counter(ifp, IFCOUNTER_OERRORS, 1);
4460 	}
4461 
4462 	return (err);
4463 }
4464 
4465 #ifdef ALTQ
4466 /*
4467  * The overall approach to integrating iflib with ALTQ is to continue to use
4468  * the iflib mp_ring machinery between the ALTQ queue(s) and the hardware
4469  * ring.  Technically, when using ALTQ, queueing to an intermediate mp_ring
4470  * is redundant/unnecessary, but doing so minimizes the amount of
4471  * ALTQ-specific code required in iflib.  It is assumed that the overhead of
4472  * redundantly queueing to an intermediate mp_ring is swamped by the
4473  * performance limitations inherent in using ALTQ.
4474  *
4475  * When ALTQ support is compiled in, all iflib drivers will use a transmit
4476  * routine, iflib_altq_if_transmit(), that checks if ALTQ is enabled for the
4477  * given interface.  If ALTQ is enabled for an interface, then all
4478  * transmitted packets for that interface will be submitted to the ALTQ
4479  * subsystem via IFQ_ENQUEUE().  We don't use the legacy if_transmit()
4480  * implementation because it uses IFQ_HANDOFF(), which will duplicatively
4481  * update stats that the iflib machinery handles, and which is sensitve to
4482  * the disused IFF_DRV_OACTIVE flag.  Additionally, iflib_altq_if_start()
4483  * will be installed as the start routine for use by ALTQ facilities that
4484  * need to trigger queue drains on a scheduled basis.
4485  *
4486  */
4487 static void
4488 iflib_altq_if_start(if_t ifp)
4489 {
4490 	struct ifaltq *ifq = &ifp->if_snd; /* XXX - DRVAPI */
4491 	struct mbuf *m;
4492 
4493 	IFQ_LOCK(ifq);
4494 	IFQ_DEQUEUE_NOLOCK(ifq, m);
4495 	while (m != NULL) {
4496 		iflib_if_transmit(ifp, m);
4497 		IFQ_DEQUEUE_NOLOCK(ifq, m);
4498 	}
4499 	IFQ_UNLOCK(ifq);
4500 }
4501 
4502 static int
4503 iflib_altq_if_transmit(if_t ifp, struct mbuf *m)
4504 {
4505 	int err;
4506 
4507 	if (if_altq_is_enabled(ifp)) {
4508 		IFQ_ENQUEUE(&ifp->if_snd, m, err); /* XXX - DRVAPI */
4509 		if (err == 0)
4510 			iflib_altq_if_start(ifp);
4511 	} else
4512 		err = iflib_if_transmit(ifp, m);
4513 
4514 	return (err);
4515 }
4516 #endif /* ALTQ */
4517 
4518 static void
4519 iflib_if_qflush(if_t ifp)
4520 {
4521 	if_ctx_t ctx = if_getsoftc(ifp);
4522 	iflib_txq_t txq = ctx->ifc_txqs;
4523 	int i;
4524 
4525 	STATE_LOCK(ctx);
4526 	ctx->ifc_flags |= IFC_QFLUSH;
4527 	STATE_UNLOCK(ctx);
4528 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
4529 		while (!(ifmp_ring_is_idle(txq->ift_br) || ifmp_ring_is_stalled(txq->ift_br)))
4530 			iflib_txq_check_drain(txq, 0);
4531 	STATE_LOCK(ctx);
4532 	ctx->ifc_flags &= ~IFC_QFLUSH;
4533 	STATE_UNLOCK(ctx);
4534 
4535 	/*
4536 	 * When ALTQ is enabled, this will also take care of purging the
4537 	 * ALTQ queue(s).
4538 	 */
4539 	if_qflush(ifp);
4540 }
4541 
4542 #define IFCAP_FLAGS (IFCAP_HWCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \
4543 		    IFCAP_TSO | IFCAP_VLAN_HWTAGGING | IFCAP_HWSTATS | \
4544 		    IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | \
4545 		    IFCAP_VLAN_HWTSO | IFCAP_VLAN_HWCSUM | IFCAP_MEXTPG)
4546 
4547 static int
4548 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data)
4549 {
4550 	if_ctx_t ctx = if_getsoftc(ifp);
4551 	struct ifreq	*ifr = (struct ifreq *)data;
4552 #if defined(INET) || defined(INET6)
4553 	struct ifaddr	*ifa = (struct ifaddr *)data;
4554 #endif
4555 	bool		avoid_reset = false;
4556 	int		err = 0, reinit = 0, bits;
4557 
4558 	switch (command) {
4559 	case SIOCSIFADDR:
4560 #ifdef INET
4561 		if (ifa->ifa_addr->sa_family == AF_INET)
4562 			avoid_reset = true;
4563 #endif
4564 #ifdef INET6
4565 		if (ifa->ifa_addr->sa_family == AF_INET6)
4566 			avoid_reset = true;
4567 #endif
4568 		/*
4569 		 * Calling init results in link renegotiation,
4570 		 * so we avoid doing it when possible.
4571 		 */
4572 		if (avoid_reset) {
4573 			if_setflagbits(ifp, IFF_UP, 0);
4574 			if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
4575 				reinit = 1;
4576 #ifdef INET
4577 			if (!(if_getflags(ifp) & IFF_NOARP))
4578 				arp_ifinit(ifp, ifa);
4579 #endif
4580 		} else
4581 			err = ether_ioctl(ifp, command, data);
4582 		break;
4583 	case SIOCSIFMTU:
4584 		CTX_LOCK(ctx);
4585 		if (ifr->ifr_mtu == if_getmtu(ifp)) {
4586 			CTX_UNLOCK(ctx);
4587 			break;
4588 		}
4589 		bits = if_getdrvflags(ifp);
4590 		/* stop the driver and free any clusters before proceeding */
4591 		iflib_stop(ctx);
4592 
4593 		if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) {
4594 			STATE_LOCK(ctx);
4595 			if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size)
4596 				ctx->ifc_flags |= IFC_MULTISEG;
4597 			else
4598 				ctx->ifc_flags &= ~IFC_MULTISEG;
4599 			STATE_UNLOCK(ctx);
4600 			err = if_setmtu(ifp, ifr->ifr_mtu);
4601 		}
4602 		iflib_init_locked(ctx);
4603 		STATE_LOCK(ctx);
4604 		/* Preserve the stopped state reported by iflib_init_failed(). */
4605 		if ((ctx->ifc_flags & IFC_INIT_FAILED) == 0)
4606 			if_setdrvflags(ifp, bits);
4607 		STATE_UNLOCK(ctx);
4608 		CTX_UNLOCK(ctx);
4609 		break;
4610 	case SIOCSIFFLAGS:
4611 		CTX_LOCK(ctx);
4612 		if (if_getflags(ifp) & IFF_UP) {
4613 			if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4614 				if ((if_getflags(ifp) ^ ctx->ifc_if_flags) &
4615 				    (IFF_PROMISC | IFF_ALLMULTI)) {
4616 					CTX_UNLOCK(ctx);
4617 					err = IFDI_PROMISC_SET(ctx, if_getflags(ifp));
4618 					CTX_LOCK(ctx);
4619 				}
4620 			} else
4621 				reinit = 1;
4622 		} else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4623 			iflib_stop(ctx);
4624 		}
4625 		ctx->ifc_if_flags = if_getflags(ifp);
4626 		CTX_UNLOCK(ctx);
4627 		break;
4628 	case SIOCADDMULTI:
4629 	case SIOCDELMULTI:
4630 		if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4631 			CTX_LOCK(ctx);
4632 			IFDI_INTR_DISABLE(ctx);
4633 			IFDI_MULTI_SET(ctx);
4634 			IFDI_INTR_ENABLE(ctx);
4635 			CTX_UNLOCK(ctx);
4636 		}
4637 		break;
4638 	case SIOCSIFMEDIA:
4639 		CTX_LOCK(ctx);
4640 		IFDI_MEDIA_SET(ctx);
4641 		CTX_UNLOCK(ctx);
4642 		/* FALLTHROUGH */
4643 	case SIOCGIFMEDIA:
4644 	case SIOCGIFXMEDIA:
4645 		err = ifmedia_ioctl(ifp, ifr, ctx->ifc_mediap, command);
4646 		break;
4647 	case SIOCGI2C:
4648 		/* FALLTHROUGH */
4649 	case SIOCGI2CPB:
4650 	{
4651 		struct ifi2creq i2c;
4652 		if_shared_ctx_t sctx = ctx->ifc_sctx;
4653 
4654 		err = copyin(ifr_data_get_ptr(ifr), &i2c, sizeof(i2c));
4655 		if (err != 0)
4656 			break;
4657 		if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) {
4658 			err = EINVAL;
4659 			break;
4660 		}
4661 		if (i2c.len > sizeof(i2c.data)) {
4662 			err = EINVAL;
4663 			break;
4664 		}
4665 		if (command == SIOCGI2C) {
4666 			i2c.page = i2c.bank = 0;
4667 		} else if ((sctx->isc_flags & IFLIB_I2C_PAGE_BANK) == 0) {
4668 			err = EINVAL;
4669 			break;
4670 		}
4671 
4672 		if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0)
4673 			err = copyout(&i2c, ifr_data_get_ptr(ifr),
4674 			    sizeof(i2c));
4675 		break;
4676 	}
4677 	case SIOCSIFCAP:
4678 	{
4679 		int mask, setmask, oldmask;
4680 
4681 		oldmask = if_getcapenable(ifp);
4682 		mask = ifr->ifr_reqcap ^ oldmask;
4683 		mask &= ctx->ifc_softc_ctx.isc_capabilities | IFCAP_MEXTPG;
4684 		setmask = 0;
4685 #ifdef TCP_OFFLOAD
4686 		setmask |= mask & (IFCAP_TOE4 | IFCAP_TOE6);
4687 #endif
4688 		setmask |= (mask & IFCAP_FLAGS);
4689 		setmask |= (mask & IFCAP_WOL);
4690 
4691 		/*
4692 		 * If any RX csum has changed, change all the ones that
4693 		 * are supported by the driver.
4694 		 */
4695 		if (setmask & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6)) {
4696 			setmask |= ctx->ifc_softc_ctx.isc_capabilities &
4697 			    (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6);
4698 		}
4699 
4700 		/*
4701 		 * want to ensure that traffic has stopped before we change any of the flags
4702 		 */
4703 		if (setmask) {
4704 			CTX_LOCK(ctx);
4705 			bits = if_getdrvflags(ifp);
4706 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4707 				iflib_stop(ctx);
4708 			STATE_LOCK(ctx);
4709 			if_togglecapenable(ifp, setmask);
4710 			ctx->ifc_softc_ctx.isc_capenable ^= setmask;
4711 			STATE_UNLOCK(ctx);
4712 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4713 				iflib_init_locked(ctx);
4714 			STATE_LOCK(ctx);
4715 			if ((ctx->ifc_flags & IFC_INIT_FAILED) == 0)
4716 				if_setdrvflags(ifp, bits);
4717 			STATE_UNLOCK(ctx);
4718 			CTX_UNLOCK(ctx);
4719 		}
4720 		if_vlancap(ifp);
4721 		break;
4722 	}
4723 	case SIOCGPRIVATE_0:
4724 	case SIOCSDRVSPEC:
4725 	case SIOCGDRVSPEC:
4726 		CTX_LOCK(ctx);
4727 		err = IFDI_PRIV_IOCTL(ctx, command, data);
4728 		CTX_UNLOCK(ctx);
4729 		break;
4730 	case SIOCGIFDOWNREASON:
4731 		CTX_LOCK(ctx);
4732 		err = IFDI_GET_DOWNREASON(ctx, (struct ifdownreason *)data);
4733 		CTX_UNLOCK(ctx);
4734 		break;
4735 	default:
4736 		err = ether_ioctl(ifp, command, data);
4737 		break;
4738 	}
4739 	if (reinit)
4740 		iflib_if_init(ctx);
4741 	return (err);
4742 }
4743 
4744 static uint64_t
4745 iflib_if_get_counter(if_t ifp, ift_counter cnt)
4746 {
4747 	if_ctx_t ctx = if_getsoftc(ifp);
4748 
4749 	return (IFDI_GET_COUNTER(ctx, cnt));
4750 }
4751 
4752 /*********************************************************************
4753  *
4754  *  OTHER FUNCTIONS EXPORTED TO THE STACK
4755  *
4756  **********************************************************************/
4757 
4758 static void
4759 iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag)
4760 {
4761 	if_ctx_t ctx = if_getsoftc(ifp);
4762 
4763 	if ((void *)ctx != arg)
4764 		return;
4765 
4766 	if ((vtag == 0) || (vtag > 4095))
4767 		return;
4768 
4769 	if (iflib_in_detach(ctx))
4770 		return;
4771 
4772 	CTX_LOCK(ctx);
4773 	/* Driver may need all untagged packets to be flushed */
4774 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4775 		iflib_stop(ctx);
4776 	IFDI_VLAN_REGISTER(ctx, vtag);
4777 	/* Re-init to load the changes, if required */
4778 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4779 		iflib_init_locked(ctx);
4780 	CTX_UNLOCK(ctx);
4781 }
4782 
4783 static void
4784 iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag)
4785 {
4786 	if_ctx_t ctx = if_getsoftc(ifp);
4787 
4788 	if ((void *)ctx != arg)
4789 		return;
4790 
4791 	if ((vtag == 0) || (vtag > 4095))
4792 		return;
4793 
4794 	CTX_LOCK(ctx);
4795 	/* Driver may need all tagged packets to be flushed */
4796 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4797 		iflib_stop(ctx);
4798 	IFDI_VLAN_UNREGISTER(ctx, vtag);
4799 	/* Re-init to load the changes, if required */
4800 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4801 		iflib_init_locked(ctx);
4802 	CTX_UNLOCK(ctx);
4803 }
4804 
4805 static void
4806 _task_fn_led(void *context, int pending __unused)
4807 {
4808 	if_ctx_t ctx = context;
4809 	bool in_detach;
4810 	int onoff;
4811 
4812 	STATE_LOCK(ctx);
4813 	in_detach = (ctx->ifc_flags & IFC_IN_DETACH) != 0;
4814 	onoff = ctx->ifc_led_state;
4815 	STATE_UNLOCK(ctx);
4816 	if (in_detach)
4817 		return;
4818 
4819 	CTX_LOCK(ctx);
4820 	IFDI_LED_FUNC(ctx, onoff);
4821 	CTX_UNLOCK(ctx);
4822 }
4823 
4824 static void
4825 iflib_led_func(void *arg, int onoff)
4826 {
4827 	if_ctx_t ctx = arg;
4828 	bool in_detach;
4829 
4830 	/* led(4) may invoke this callback from a non-sleepable callout. */
4831 	STATE_LOCK(ctx);
4832 	ctx->ifc_led_state = onoff;
4833 	in_detach = (ctx->ifc_flags & IFC_IN_DETACH) != 0;
4834 	STATE_UNLOCK(ctx);
4835 	if (!in_detach)
4836 		taskqueue_enqueue(ctx->ifc_tq, &ctx->ifc_led_task);
4837 }
4838 
4839 /*********************************************************************
4840  *
4841  *  BUS FUNCTION DEFINITIONS
4842  *
4843  **********************************************************************/
4844 
4845 int
4846 iflib_device_probe(device_t dev)
4847 {
4848 	const pci_vendor_info_t *ent;
4849 	if_shared_ctx_t sctx;
4850 	uint16_t pci_device_id, pci_rev_id, pci_subdevice_id, pci_subvendor_id;
4851 	uint16_t pci_vendor_id;
4852 
4853 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4854 		return (ENOTSUP);
4855 
4856 	pci_vendor_id = pci_get_vendor(dev);
4857 	pci_device_id = pci_get_device(dev);
4858 	pci_subvendor_id = pci_get_subvendor(dev);
4859 	pci_subdevice_id = pci_get_subdevice(dev);
4860 	pci_rev_id = pci_get_revid(dev);
4861 	if (sctx->isc_parse_devinfo != NULL)
4862 		sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id);
4863 
4864 	ent = sctx->isc_vendor_info;
4865 	while (ent->pvi_vendor_id != 0) {
4866 		if (pci_vendor_id != ent->pvi_vendor_id) {
4867 			ent++;
4868 			continue;
4869 		}
4870 		if ((pci_device_id == ent->pvi_device_id) &&
4871 		    ((pci_subvendor_id == ent->pvi_subvendor_id) ||
4872 		     (ent->pvi_subvendor_id == 0)) &&
4873 		    ((pci_subdevice_id == ent->pvi_subdevice_id) ||
4874 		     (ent->pvi_subdevice_id == 0)) &&
4875 		    ((pci_rev_id == ent->pvi_rev_id) ||
4876 		     (ent->pvi_rev_id == 0))) {
4877 			device_set_desc_copy(dev, ent->pvi_name);
4878 			/* this needs to be changed to zero if the bus probing code
4879 			 * ever stops re-probing on best match because the sctx
4880 			 * may have its values over written by register calls
4881 			 * in subsequent probes
4882 			 */
4883 			return (BUS_PROBE_DEFAULT);
4884 		}
4885 		ent++;
4886 	}
4887 	return (ENXIO);
4888 }
4889 
4890 int
4891 iflib_device_probe_vendor(device_t dev)
4892 {
4893 	int probe;
4894 
4895 	probe = iflib_device_probe(dev);
4896 	if (probe == BUS_PROBE_DEFAULT)
4897 		return (BUS_PROBE_VENDOR);
4898 	else
4899 		return (probe);
4900 }
4901 
4902 static void
4903 iflib_reset_qvalues(if_ctx_t ctx)
4904 {
4905 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4906 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4907 	device_t dev = ctx->ifc_dev;
4908 	int i;
4909 
4910 	if (ctx->ifc_sysctl_ntxqs != 0)
4911 		scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs;
4912 	if (ctx->ifc_sysctl_nrxqs != 0)
4913 		scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs;
4914 
4915 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4916 		if (ctx->ifc_sysctl_ntxds[i] != 0)
4917 			scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i];
4918 		else
4919 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4920 	}
4921 
4922 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4923 		if (ctx->ifc_sysctl_nrxds[i] != 0)
4924 			scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i];
4925 		else
4926 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4927 	}
4928 
4929 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4930 		if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) {
4931 			device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n",
4932 			    i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]);
4933 			scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i];
4934 		}
4935 		if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) {
4936 			device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n",
4937 			    i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]);
4938 			scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i];
4939 		}
4940 		if (!powerof2(scctx->isc_nrxd[i])) {
4941 			device_printf(dev, "nrxd%d: %d is not a power of 2 - using default value of %d\n",
4942 			    i, scctx->isc_nrxd[i], sctx->isc_nrxd_default[i]);
4943 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4944 		}
4945 	}
4946 
4947 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4948 		if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) {
4949 			device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n",
4950 			    i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]);
4951 			scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i];
4952 		}
4953 		if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) {
4954 			device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n",
4955 			    i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]);
4956 			scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i];
4957 		}
4958 		if (!powerof2(scctx->isc_ntxd[i])) {
4959 			device_printf(dev, "ntxd%d: %d is not a power of 2 - using default value of %d\n",
4960 			    i, scctx->isc_ntxd[i], sctx->isc_ntxd_default[i]);
4961 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4962 		}
4963 	}
4964 	scctx->isc_tx_pad = 2;
4965 }
4966 
4967 static void
4968 iflib_add_pfil(if_ctx_t ctx)
4969 {
4970 	struct pfil_head *pfil;
4971 	struct pfil_head_args pa;
4972 	iflib_rxq_t rxq;
4973 	int i;
4974 
4975 	pa.pa_version = PFIL_VERSION;
4976 	pa.pa_flags = PFIL_IN;
4977 	pa.pa_type = PFIL_TYPE_ETHERNET;
4978 	pa.pa_headname = if_name(ctx->ifc_ifp);
4979 	pfil = pfil_head_register(&pa);
4980 
4981 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4982 		rxq->pfil = pfil;
4983 	}
4984 }
4985 
4986 static void
4987 iflib_rem_pfil(if_ctx_t ctx)
4988 {
4989 	struct pfil_head *pfil;
4990 	iflib_rxq_t rxq;
4991 	int i;
4992 
4993 	rxq = ctx->ifc_rxqs;
4994 	pfil = rxq->pfil;
4995 	for (i = 0; i < NRXQSETS(ctx); i++, rxq++) {
4996 		rxq->pfil = NULL;
4997 	}
4998 	pfil_head_unregister(pfil);
4999 }
5000 
5001 
5002 /*
5003  * Advance forward by n members of the cpuset ctx->ifc_cpus starting from
5004  * cpuid and wrapping as necessary.
5005  */
5006 static unsigned int
5007 cpuid_advance(if_ctx_t ctx, unsigned int cpuid, unsigned int n)
5008 {
5009 	unsigned int first_valid;
5010 	unsigned int last_valid;
5011 
5012 	/* cpuid should always be in the valid set */
5013 	MPASS(CPU_ISSET(cpuid, &ctx->ifc_cpus));
5014 
5015 	/* valid set should never be empty */
5016 	MPASS(!CPU_EMPTY(&ctx->ifc_cpus));
5017 
5018 	first_valid = CPU_FFS(&ctx->ifc_cpus) - 1;
5019 	last_valid = CPU_FLS(&ctx->ifc_cpus) - 1;
5020 	n = n % CPU_COUNT(&ctx->ifc_cpus);
5021 	while (n > 0) {
5022 		do {
5023 			cpuid++;
5024 			if (cpuid > last_valid)
5025 				cpuid = first_valid;
5026 		} while (!CPU_ISSET(cpuid, &ctx->ifc_cpus));
5027 		n--;
5028 	}
5029 
5030 	return (cpuid);
5031 }
5032 
5033 /*
5034  * CPU mapping behaviors
5035  * ---------------------
5036  * 'separate txrx' refers to the separate_txrx sysctl
5037  * 'use logical' refers to the use_logical_cores sysctl
5038  * 'INTR CPUS' indicates whether bus_get_cpus(INTR_CPUS) succeeded
5039  *
5040  *  separate     use     INTR
5041  *    txrx     logical   CPUS   result
5042  * ---------- --------- ------ ------------------------------------------------
5043  *     -          -       X     RX and TX queues mapped to consecutive physical
5044  *                              cores with RX/TX pairs on same core and excess
5045  *                              of either following
5046  *     -          X       X     RX and TX queues mapped to consecutive cores
5047  *                              of any type with RX/TX pairs on same core and
5048  *                              excess of either following
5049  *     X          -       X     RX and TX queues mapped to consecutive physical
5050  *                              cores; all RX then all TX
5051  *     X          X       X     RX queues mapped to consecutive physical cores
5052  *                              first, then TX queues mapped to L2 neighbor of
5053  *                              the corresponding RX queue if one exists,
5054  *                              otherwise to consecutive physical cores
5055  *     -         n/a      -     RX and TX queues mapped to consecutive cores of
5056  *                              any type with RX/TX pairs on same core and excess
5057  *                              of either following
5058  *     X         n/a      -     RX and TX queues mapped to consecutive cores of
5059  *                              any type; all RX then all TX
5060  */
5061 static unsigned int
5062 get_cpuid_for_queue(if_ctx_t ctx, unsigned int base_cpuid, unsigned int qid,
5063     bool is_tx)
5064 {
5065 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5066 	unsigned int core_index;
5067 
5068 	if (ctx->ifc_sysctl_separate_txrx) {
5069 		/*
5070 		 * When using separate CPUs for TX and RX, the assignment
5071 		 * will always be of a consecutive CPU out of the set of
5072 		 * context CPUs, except for the specific case where the
5073 		 * context CPUs are phsyical cores, the use of logical cores
5074 		 * has been enabled, the assignment is for TX, the TX qid
5075 		 * corresponds to an RX qid, and the CPU assigned to the
5076 		 * corresponding RX queue has an L2 neighbor.
5077 		 */
5078 		if (ctx->ifc_sysctl_use_logical_cores &&
5079 		    ctx->ifc_cpus_are_physical_cores &&
5080 		    is_tx && qid < scctx->isc_nrxqsets) {
5081 			int l2_neighbor;
5082 			unsigned int rx_cpuid;
5083 
5084 			rx_cpuid = cpuid_advance(ctx, base_cpuid, qid);
5085 			l2_neighbor = sched_find_l2_neighbor(rx_cpuid);
5086 			if (l2_neighbor != -1) {
5087 				return (l2_neighbor);
5088 			}
5089 			/*
5090 			 * ... else fall through to the normal
5091 			 * consecutive-after-RX assignment scheme.
5092 			 *
5093 			 * Note that we are assuming that all RX queue CPUs
5094 			 * have an L2 neighbor, or all do not.  If a mixed
5095 			 * scenario is possible, we will have to keep track
5096 			 * separately of how many queues prior to this one
5097 			 * were not able to be assigned to an L2 neighbor.
5098 			 */
5099 		}
5100 		if (is_tx)
5101 			core_index = scctx->isc_nrxqsets + qid;
5102 		else
5103 			core_index = qid;
5104 	} else {
5105 		core_index = qid;
5106 	}
5107 
5108 	return (cpuid_advance(ctx, base_cpuid, core_index));
5109 }
5110 
5111 static uint16_t
5112 get_ctx_core_offset(if_ctx_t ctx)
5113 {
5114 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5115 	struct cpu_offset *op;
5116 	cpuset_t assigned_cpus;
5117 	unsigned int cores_consumed;
5118 	unsigned int base_cpuid = ctx->ifc_sysctl_core_offset;
5119 	unsigned int first_valid;
5120 	unsigned int last_valid;
5121 	unsigned int i;
5122 
5123 	MPASS(!ctx->ifc_core_offset_ref);
5124 	first_valid = CPU_FFS(&ctx->ifc_cpus) - 1;
5125 	last_valid = CPU_FLS(&ctx->ifc_cpus) - 1;
5126 
5127 	if (base_cpuid != CORE_OFFSET_UNSPECIFIED) {
5128 		/*
5129 		 * Align the user-chosen base CPU ID to the next valid CPU
5130 		 * for this device.  If the chosen base CPU ID is smaller
5131 		 * than the first valid CPU or larger than the last valid
5132 		 * CPU, we assume the user does not know what the valid
5133 		 * range is for this device and is thinking in terms of a
5134 		 * zero-based reference frame, and so we shift the given
5135 		 * value into the valid range (and wrap accordingly) so the
5136 		 * intent is translated to the proper frame of reference.
5137 		 * If the base CPU ID is within the valid first/last, but
5138 		 * does not correspond to a valid CPU, it is advanced to the
5139 		 * next valid CPU (wrapping if necessary).
5140 		 */
5141 		if (base_cpuid < first_valid || base_cpuid > last_valid) {
5142 			/* shift from zero-based to first_valid-based */
5143 			base_cpuid += first_valid;
5144 			/* wrap to range [first_valid, last_valid] */
5145 			base_cpuid = (base_cpuid - first_valid) %
5146 			    (last_valid - first_valid + 1);
5147 		}
5148 		if (!CPU_ISSET(base_cpuid, &ctx->ifc_cpus)) {
5149 			/*
5150 			 * base_cpuid is in [first_valid, last_valid], but
5151 			 * not a member of the valid set.  In this case,
5152 			 * there will always be a member of the valid set
5153 			 * with a CPU ID that is greater than base_cpuid,
5154 			 * and we simply advance to it.
5155 			 */
5156 			while (!CPU_ISSET(base_cpuid, &ctx->ifc_cpus))
5157 				base_cpuid++;
5158 		}
5159 		return (base_cpuid);
5160 	}
5161 
5162 	/*
5163 	 * Determine how many cores will be consumed by performing the CPU
5164 	 * assignments and counting how many of the assigned CPUs correspond
5165 	 * to CPUs in the set of context CPUs.  This is done using the CPU
5166 	 * ID first_valid as the base CPU ID, as the base CPU must be within
5167 	 * the set of context CPUs.
5168 	 *
5169 	 * Note not all assigned CPUs will be in the set of context CPUs
5170 	 * when separate CPUs are being allocated to TX and RX queues,
5171 	 * assignment to logical cores has been enabled, the set of context
5172 	 * CPUs contains only physical CPUs, and TX queues are mapped to L2
5173 	 * neighbors of CPUs that RX queues have been mapped to - in this
5174 	 * case we do only want to count how many CPUs in the set of context
5175 	 * CPUs have been consumed, as that determines the next CPU in that
5176 	 * set to start allocating at for the next device for which
5177 	 * core_offset is not set.
5178 	 */
5179 	CPU_ZERO(&assigned_cpus);
5180 	for (i = 0; i < scctx->isc_ntxqsets; i++)
5181 		CPU_SET(get_cpuid_for_queue(ctx, first_valid, i, true),
5182 		    &assigned_cpus);
5183 	for (i = 0; i < scctx->isc_nrxqsets; i++)
5184 		CPU_SET(get_cpuid_for_queue(ctx, first_valid, i, false),
5185 		    &assigned_cpus);
5186 	CPU_AND(&assigned_cpus, &assigned_cpus, &ctx->ifc_cpus);
5187 	cores_consumed = CPU_COUNT(&assigned_cpus);
5188 
5189 	mtx_lock(&cpu_offset_mtx);
5190 	SLIST_FOREACH(op, &cpu_offsets, entries) {
5191 		if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) {
5192 			base_cpuid = op->next_cpuid;
5193 			op->next_cpuid = cpuid_advance(ctx, op->next_cpuid,
5194 			    cores_consumed);
5195 			MPASS(op->refcount < UINT_MAX);
5196 			op->refcount++;
5197 			ctx->ifc_core_offset_ref = true;
5198 			break;
5199 		}
5200 	}
5201 	if (base_cpuid == CORE_OFFSET_UNSPECIFIED) {
5202 		base_cpuid = first_valid;
5203 		op = malloc(sizeof(struct cpu_offset), M_IFLIB,
5204 		    M_NOWAIT | M_ZERO);
5205 		if (op == NULL) {
5206 			device_printf(ctx->ifc_dev,
5207 			    "allocation for cpu offset failed.\n");
5208 		} else {
5209 			op->next_cpuid = cpuid_advance(ctx, base_cpuid,
5210 			    cores_consumed);
5211 			op->refcount = 1;
5212 			CPU_COPY(&ctx->ifc_cpus, &op->set);
5213 			SLIST_INSERT_HEAD(&cpu_offsets, op, entries);
5214 			ctx->ifc_core_offset_ref = true;
5215 		}
5216 	}
5217 	mtx_unlock(&cpu_offset_mtx);
5218 
5219 	return (base_cpuid);
5220 }
5221 
5222 static void
5223 unref_ctx_core_offset(if_ctx_t ctx)
5224 {
5225 	struct cpu_offset *op, *top;
5226 
5227 	if (!ctx->ifc_core_offset_ref)
5228 		return;
5229 
5230 	mtx_lock(&cpu_offset_mtx);
5231 	SLIST_FOREACH_SAFE(op, &cpu_offsets, entries, top) {
5232 		if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) {
5233 			MPASS(op->refcount > 0);
5234 			op->refcount--;
5235 			if (op->refcount == 0) {
5236 				SLIST_REMOVE(&cpu_offsets, op, cpu_offset, entries);
5237 				free(op, M_IFLIB);
5238 			}
5239 			ctx->ifc_core_offset_ref = false;
5240 			break;
5241 		}
5242 	}
5243 	mtx_unlock(&cpu_offset_mtx);
5244 	MPASS(!ctx->ifc_core_offset_ref);
5245 }
5246 
5247 static bool
5248 iflib_register_fail_device_matches(device_t dev)
5249 {
5250 	const char *nameunit;
5251 
5252 	nameunit = device_get_nameunit(dev);
5253 	return (iflib_register_fail_device[0] != '\0' && nameunit != NULL &&
5254 	    strcmp(nameunit, iflib_register_fail_device) == 0);
5255 }
5256 
5257 #define	IFLIB_REGISTER_FAIL_POINT(_dev, _name, _error, _label) do { \
5258 	KFAIL_POINT_CODE_COND(_debug_fail_point_iflib, _name, \
5259 	    iflib_register_fail_device_matches((_dev)), \
5260 	    FAIL_POINT_NONSLEEPABLE, { \
5261 		(_error) = RETURN_VALUE; \
5262 		if ((_error) <= 0) \
5263 			(_error) = EIO; \
5264 		device_printf((_dev), \
5265 		    "injecting iflib registration failure at %s: %d\n", \
5266 		    #_name, (_error)); \
5267 		goto _label; \
5268 	}); \
5269 } while (0)
5270 
5271 int
5272 iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp)
5273 {
5274 	if_ctx_t ctx;
5275 	if_t ifp;
5276 	if_softc_ctx_t scctx;
5277 	kobjop_desc_t kobj_desc;
5278 	kobj_method_t *kobj_method;
5279 	bool attach_pre_succeeded, intr_allocated, queues_allocated;
5280 	int err, msix, rid;
5281 #ifdef PCI_IOV
5282 	int iov_error;
5283 #endif
5284 	int num_txd, num_rxd;
5285 	char namebuf[TASKQUEUE_NAMELEN];
5286 
5287 	attach_pre_succeeded = false;
5288 	intr_allocated = false;
5289 	queues_allocated = false;
5290 	ctx = malloc(sizeof(*ctx), M_IFLIB, M_WAITOK | M_ZERO);
5291 
5292 	if (sc == NULL) {
5293 		sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK | M_ZERO);
5294 		device_set_softc(dev, ctx);
5295 		ctx->ifc_flags |= IFC_SC_ALLOCATED;
5296 	}
5297 
5298 	ctx->ifc_sctx = sctx;
5299 	ctx->ifc_dev = dev;
5300 	ctx->ifc_softc = sc;
5301 
5302 	iflib_register(ctx);
5303 	iflib_add_device_sysctl_pre(ctx);
5304 
5305 	scctx = &ctx->ifc_softc_ctx;
5306 	ifp = ctx->ifc_ifp;
5307 	if (ctx->ifc_sysctl_simple_tx) {
5308 #ifndef ALTQ
5309 		if_settransmitfn(ifp, iflib_simple_transmit);
5310 		device_printf(dev, "using simple if_transmit\n");
5311 #else
5312 		device_printf(dev, "ALTQ prevents using simple if_transmit\n");
5313 		ctx->ifc_sysctl_simple_tx = 0;
5314 #endif
5315 	}
5316 	iflib_reset_qvalues(ctx);
5317 	CTX_LOCK(ctx);
5318 	IFLIB_REGISTER_FAIL_POINT(dev, register_before_attach_pre, err,
5319 	    fail_cleanup);
5320 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
5321 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
5322 		goto fail_cleanup;
5323 	}
5324 	attach_pre_succeeded = true;
5325 	IFLIB_REGISTER_FAIL_POINT(dev, register_after_attach_pre, err,
5326 	    fail_cleanup);
5327 	_iflib_pre_assert(scctx);
5328 	ctx->ifc_txrx = *scctx->isc_txrx;
5329 
5330 	MPASS(scctx->isc_dma_width <= flsll(BUS_SPACE_MAXADDR));
5331 
5332 	if (sctx->isc_flags & IFLIB_DRIVER_MEDIA)
5333 		ctx->ifc_mediap = scctx->isc_media;
5334 
5335 #ifdef INVARIANTS
5336 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
5337 		MPASS(scctx->isc_tx_csum_flags);
5338 #endif
5339 
5340 	if_setcapabilities(ifp,
5341 	    scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_MEXTPG);
5342 	if_setcapenable(ifp,
5343 	    scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_MEXTPG);
5344 
5345 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
5346 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
5347 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
5348 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
5349 
5350 	num_txd = iflib_num_tx_descs(ctx);
5351 	num_rxd = iflib_num_rx_descs(ctx);
5352 
5353 	/* XXX change for per-queue sizes */
5354 	device_printf(dev, "Using %d TX descriptors and %d RX descriptors\n",
5355 	    num_txd, num_rxd);
5356 
5357 	if (scctx->isc_tx_nsegments > num_txd / MAX_SINGLE_PACKET_FRACTION)
5358 		scctx->isc_tx_nsegments = max(1, num_txd /
5359 		    MAX_SINGLE_PACKET_FRACTION);
5360 	if (scctx->isc_tx_tso_segments_max > num_txd /
5361 	    MAX_SINGLE_PACKET_FRACTION)
5362 		scctx->isc_tx_tso_segments_max = max(1,
5363 		    num_txd / MAX_SINGLE_PACKET_FRACTION);
5364 
5365 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
5366 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
5367 		/*
5368 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
5369 		 * but some MACs do.
5370 		 */
5371 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
5372 		    IP_MAXPACKET));
5373 		/*
5374 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
5375 		 * into account.  In the worst case, each of these calls will
5376 		 * add another mbuf and, thus, the requirement for another DMA
5377 		 * segment.  So for best performance, it doesn't make sense to
5378 		 * advertize a maximum of TSO segments that typically will
5379 		 * require defragmentation in iflib_encap().
5380 		 */
5381 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
5382 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
5383 	}
5384 	if (scctx->isc_rss_table_size == 0)
5385 		scctx->isc_rss_table_size = 64;
5386 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size - 1;
5387 
5388 	/* Create and start admin taskqueue */
5389 	snprintf(namebuf, TASKQUEUE_NAMELEN, "if_%s_tq", device_get_nameunit(dev));
5390 	ctx->ifc_tq = taskqueue_create_fast(namebuf, M_NOWAIT,
5391 	    taskqueue_thread_enqueue, &ctx->ifc_tq);
5392 	if (ctx->ifc_tq == NULL) {
5393 		device_printf(dev, "Unable to create admin taskqueue\n");
5394 		err = ENOMEM;
5395 		goto fail_cleanup;
5396 	}
5397 
5398 	err = taskqueue_start_threads(&ctx->ifc_tq, 1, PI_NET, "%s", namebuf);
5399 	if (err) {
5400 		device_printf(dev,
5401 		    "Unable to start admin taskqueue threads error: %d\n",
5402 		    err);
5403 		taskqueue_free(ctx->ifc_tq);
5404 		ctx->ifc_tq = NULL;
5405 		goto fail_cleanup;
5406 	}
5407 
5408 	TASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
5409 	TASK_INIT(&ctx->ifc_led_task, 0, _task_fn_led, ctx);
5410 	TASK_INIT(&ctx->ifc_vflr_task, 0, _task_fn_iov, ctx);
5411 	IFLIB_REGISTER_FAIL_POINT(dev, register_after_taskqueue, err,
5412 	    fail_cleanup);
5413 
5414 	/* Set up cpu set.  If it fails, use the set of all CPUs. */
5415 	if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) != 0) {
5416 		device_printf(dev, "Unable to fetch CPU list\n");
5417 		CPU_COPY(&all_cpus, &ctx->ifc_cpus);
5418 		ctx->ifc_cpus_are_physical_cores = false;
5419 	} else
5420 		ctx->ifc_cpus_are_physical_cores = true;
5421 	MPASS(CPU_COUNT(&ctx->ifc_cpus) > 0);
5422 
5423 	/*
5424 	 * Now set up MSI or MSI-X, should return us the number of supported
5425 	 * vectors (will be 1 for a legacy interrupt and MSI).
5426 	 */
5427 	if (sctx->isc_flags & IFLIB_SKIP_MSIX) {
5428 		msix = scctx->isc_vectors;
5429 	} else if (scctx->isc_msix_bar != 0)
5430 		/*
5431 		 * The simple fact that isc_msix_bar is not 0 does not mean we
5432 		 * we have a good value there that is known to work.
5433 		 */
5434 		msix = iflib_msix_init(ctx);
5435 	else {
5436 		scctx->isc_vectors = 1;
5437 		scctx->isc_ntxqsets = 1;
5438 		scctx->isc_nrxqsets = 1;
5439 		scctx->isc_intr = IFLIB_INTR_LEGACY;
5440 		msix = 0;
5441 	}
5442 	intr_allocated = true;
5443 	IFLIB_REGISTER_FAIL_POINT(dev, register_after_interrupts, err,
5444 	    fail_cleanup);
5445 	/* Get memory for the station queues */
5446 	if ((err = iflib_queues_alloc(ctx))) {
5447 		device_printf(dev, "Unable to allocate queue memory\n");
5448 		goto fail_cleanup;
5449 	}
5450 	queues_allocated = true;
5451 
5452 	if ((err = iflib_qset_structures_setup(ctx)))
5453 		goto fail_cleanup;
5454 
5455 	/*
5456 	 * Now that we know how many queues there are, get the core offset.
5457 	 */
5458 	ctx->ifc_sysctl_core_offset = get_ctx_core_offset(ctx);
5459 	IFLIB_REGISTER_FAIL_POINT(dev, register_after_queues, err,
5460 	    fail_cleanup);
5461 
5462 	if (msix > 1) {
5463 		/*
5464 		 * When using MSI-X, ensure that ifdi_{r,t}x_queue_intr_enable
5465 		 * aren't the default NULL implementation.
5466 		 */
5467 		kobj_desc = &ifdi_rx_queue_intr_enable_desc;
5468 		kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL,
5469 		    kobj_desc);
5470 		if (kobj_method == &kobj_desc->deflt) {
5471 			device_printf(dev,
5472 			    "MSI-X requires ifdi_rx_queue_intr_enable method");
5473 			err = EOPNOTSUPP;
5474 			goto fail_cleanup;
5475 		}
5476 		kobj_desc = &ifdi_tx_queue_intr_enable_desc;
5477 		kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL,
5478 		    kobj_desc);
5479 		if (kobj_method == &kobj_desc->deflt) {
5480 			device_printf(dev,
5481 			    "MSI-X requires ifdi_tx_queue_intr_enable method");
5482 			err = EOPNOTSUPP;
5483 			goto fail_cleanup;
5484 		}
5485 
5486 		/*
5487 		 * Assign the MSI-X vectors.
5488 		 * Note that the default NULL ifdi_msix_intr_assign method will
5489 		 * fail here, too.
5490 		 */
5491 		err = IFDI_MSIX_INTR_ASSIGN(ctx, msix);
5492 		if (err != 0) {
5493 			device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n",
5494 			    err);
5495 			goto fail_cleanup;
5496 		}
5497 	} else if (scctx->isc_intr != IFLIB_INTR_MSIX) {
5498 		rid = 0;
5499 		if (scctx->isc_intr == IFLIB_INTR_MSI) {
5500 			MPASS(msix == 1);
5501 			rid = 1;
5502 		}
5503 		if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) {
5504 			device_printf(dev, "iflib_legacy_setup failed %d\n", err);
5505 			goto fail_cleanup;
5506 		}
5507 	} else {
5508 		device_printf(dev,
5509 		    "Cannot use iflib with only 1 MSI-X interrupt!\n");
5510 		err = ENODEV;
5511 		goto fail_cleanup;
5512 	}
5513 
5514 	/*
5515 	 * It prevents a double-locking panic with iflib_media_status when
5516 	 * the driver loads.
5517 	 */
5518 	CTX_UNLOCK(ctx);
5519 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
5520 	CTX_LOCK(ctx);
5521 
5522 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
5523 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
5524 		goto fail_detach;
5525 	}
5526 	IFLIB_REGISTER_FAIL_POINT(dev, register_after_attach_post, err,
5527 	    fail_detach);
5528 
5529 	/*
5530 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
5531 	 * This must appear after the call to ether_ifattach() because
5532 	 * ether_ifattach() sets if_hdrlen to the default value.
5533 	 */
5534 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
5535 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
5536 
5537 	if ((err = iflib_netmap_attach(ctx))) {
5538 		device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err);
5539 		goto fail_detach;
5540 	}
5541 	*ctxp = ctx;
5542 
5543 	DEBUGNET_SET(ctx->ifc_ifp, iflib);
5544 
5545 	iflib_add_device_sysctl_post(ctx);
5546 	iflib_add_pfil(ctx);
5547 	ctx->ifc_flags |= IFC_INIT_DONE;
5548 	CTX_UNLOCK(ctx);
5549 
5550 	/* Create led(4) devices if the driver defined the method */
5551 	kobj_desc = &ifdi_led_func_desc;
5552 	kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL,
5553 	    kobj_desc);
5554 	if (kobj_method != &kobj_desc->deflt && IFDI_LED_SUPPORTED(ctx))
5555 		iflib_led_create(ctx);
5556 
5557 	return (0);
5558 
5559 fail_detach:
5560 	STATE_LOCK(ctx);
5561 	ctx->ifc_flags |= IFC_IN_DETACH;
5562 	STATE_UNLOCK(ctx);
5563 	/* Tasks may need the context lock; ether_ifdetach() may sleep. */
5564 	CTX_UNLOCK(ctx);
5565 	taskqueue_drain_all(ctx->ifc_tq);
5566 #ifdef PCI_IOV
5567 	/*
5568 	 * IFDI_ATTACH_POST may have registered an SR-IOV schema.  Match the
5569 	 * normal deregistration order so a failed attach cannot leave a stale
5570 	 * /dev/iov node behind.  device_attach() holds Giant throughout this
5571 	 * path, so an IOV configuration cannot race the detach.
5572 	 */
5573 	if (!CTX_IS_VF(ctx)) {
5574 		iov_error = pci_iov_detach(dev);
5575 		if (iov_error != 0)
5576 			device_printf(dev, "Could not detach SR-IOV after "
5577 			    "attach failure: %d\n", iov_error);
5578 	}
5579 #endif
5580 	ether_ifdetach(ctx->ifc_ifp);
5581 	CTX_LOCK(ctx);
5582 	goto fail_cleanup_detaching;
5583 
5584 fail_cleanup:
5585 	STATE_LOCK(ctx);
5586 	ctx->ifc_flags |= IFC_IN_DETACH;
5587 	STATE_UNLOCK(ctx);
5588 
5589 fail_cleanup_detaching:
5590 	/*
5591 	 * The pre-attach sysctls contain pointers into ctx.  Remove them on
5592 	 * every registration failure before iflib_deregister() frees ctx.
5593 	 */
5594 	if (ctx->ifc_sysctl_node != NULL) {
5595 		sysctl_ctx_free(&ctx->ifc_sysctl_ctx);
5596 		ctx->ifc_sysctl_node = NULL;
5597 	}
5598 
5599 	if (ctx->ifc_tq != NULL) {
5600 		/*
5601 		 * Drain without holding the context lock so configuration tasks can
5602 		 * run to completion.  On fail_detach a second drain also catches
5603 		 * tasks queued during the first drain.
5604 		 */
5605 		CTX_UNLOCK(ctx);
5606 		taskqueue_drain_all(ctx->ifc_tq);
5607 		CTX_LOCK(ctx);
5608 	}
5609 
5610 	if (queues_allocated) {
5611 		iflib_tqg_detach(ctx);
5612 		iflib_tx_structures_free(ctx);
5613 		iflib_rx_structures_free(ctx);
5614 	}
5615 
5616 	/*
5617 	 * A successful IFDI_ATTACH_PRE must be matched by IFDI_DETACH, even
5618 	 * when registration fails before queue allocation.  Match
5619 	 * iflib_device_deregister by detaching before taskqueue_free.
5620 	 */
5621 	if (attach_pre_succeeded) {
5622 		IFDI_DETACH(ctx);
5623 		if (queues_allocated)
5624 			IFDI_QUEUES_FREE(ctx);
5625 	}
5626 	if (ctx->ifc_tq != NULL) {
5627 		taskqueue_free(ctx->ifc_tq);
5628 		ctx->ifc_tq = NULL;
5629 	}
5630 	if (intr_allocated)
5631 		iflib_free_intr_mem(ctx);
5632 
5633 	CTX_UNLOCK(ctx);
5634 	iflib_deregister(ctx);
5635 	device_set_softc(ctx->ifc_dev, NULL);
5636 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5637 		free(ctx->ifc_softc, M_IFLIB);
5638 	unref_ctx_core_offset(ctx);
5639 	free(ctx, M_IFLIB);
5640 	return (err);
5641 }
5642 
5643 int
5644 iflib_device_attach(device_t dev)
5645 {
5646 	if_ctx_t ctx;
5647 	if_shared_ctx_t sctx;
5648 
5649 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
5650 		return (ENOTSUP);
5651 
5652 	pci_enable_busmaster(dev);
5653 
5654 	return (iflib_device_register(dev, NULL, sctx, &ctx));
5655 }
5656 
5657 int
5658 iflib_device_deregister(if_ctx_t ctx)
5659 {
5660 	if_t ifp = ctx->ifc_ifp;
5661 	device_t dev = ctx->ifc_dev;
5662 
5663 	sysctl_ctx_free(&ctx->ifc_sysctl_ctx);
5664 	ctx->ifc_sysctl_node = NULL;
5665 
5666 	/* Make sure VLANS are not using driver */
5667 	if (if_vlantrunkinuse(ifp)) {
5668 		device_printf(dev, "Vlan in use, detach first\n");
5669 		return (EBUSY);
5670 	}
5671 #ifdef PCI_IOV
5672 	if (!CTX_IS_VF(ctx) && pci_iov_detach(dev) != 0) {
5673 		device_printf(dev, "SR-IOV in use; detach first.\n");
5674 		return (EBUSY);
5675 	}
5676 #endif
5677 
5678 	STATE_LOCK(ctx);
5679 	ctx->ifc_flags |= IFC_IN_DETACH;
5680 	STATE_UNLOCK(ctx);
5681 
5682 	/* Unregister VLAN handlers before calling iflib_stop() */
5683 	iflib_unregister_vlan_handlers(ctx);
5684 
5685 	iflib_netmap_detach(ifp);
5686 	/*
5687 	 * A task that passed its IFC_IN_DETACH check before the flag was set
5688 	 * can still report a link change.  Drain every private task before
5689 	 * ether_ifdetach() performs the final if_linktask drain.  Drivers may
5690 	 * register their own link-related tasks on this taskqueue.
5691 	 */
5692 	taskqueue_drain_all(ctx->ifc_tq);
5693 	ether_ifdetach(ifp);
5694 
5695 	CTX_LOCK(ctx);
5696 	iflib_stop(ctx);
5697 	CTX_UNLOCK(ctx);
5698 
5699 	iflib_rem_pfil(ctx);
5700 	if (ctx->ifc_led_dev != NULL) {
5701 		led_destroy(ctx->ifc_led_dev);
5702 		taskqueue_drain(ctx->ifc_tq, &ctx->ifc_led_task);
5703 	}
5704 
5705 	iflib_tqg_detach(ctx);
5706 	iflib_tx_structures_free(ctx);
5707 	iflib_rx_structures_free(ctx);
5708 
5709 	CTX_LOCK(ctx);
5710 	IFDI_DETACH(ctx);
5711 	IFDI_QUEUES_FREE(ctx);
5712 	CTX_UNLOCK(ctx);
5713 
5714 	taskqueue_free(ctx->ifc_tq);
5715 	ctx->ifc_tq = NULL;
5716 
5717 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
5718 	iflib_free_intr_mem(ctx);
5719 
5720 	bus_generic_detach(dev);
5721 
5722 	iflib_deregister(ctx);
5723 
5724 	device_set_softc(ctx->ifc_dev, NULL);
5725 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5726 		free(ctx->ifc_softc, M_IFLIB);
5727 	unref_ctx_core_offset(ctx);
5728 	free(ctx, M_IFLIB);
5729 	return (0);
5730 }
5731 
5732 static void
5733 iflib_tqg_detach(if_ctx_t ctx)
5734 {
5735 	iflib_txq_t txq;
5736 	iflib_rxq_t rxq;
5737 	int i;
5738 	struct taskqgroup *tqg;
5739 
5740 	/* XXX drain any dependent tasks */
5741 	tqg = qgroup_if_io_tqg;
5742 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
5743 		callout_drain(&txq->ift_timer);
5744 #ifdef DEV_NETMAP
5745 		callout_drain(&txq->ift_netmap_timer);
5746 #endif /* DEV_NETMAP */
5747 		if (txq->ift_task.gt_uniq != NULL)
5748 			taskqgroup_detach(tqg, &txq->ift_task);
5749 	}
5750 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
5751 		if (rxq->ifr_task.gt_uniq != NULL)
5752 			taskqgroup_detach(tqg, &rxq->ifr_task);
5753 	}
5754 }
5755 
5756 static void
5757 iflib_free_intr_mem(if_ctx_t ctx)
5758 {
5759 
5760 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) {
5761 		iflib_irq_free(ctx, &ctx->ifc_legacy_irq);
5762 	}
5763 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) {
5764 		pci_release_msi(ctx->ifc_dev);
5765 	}
5766 	if (ctx->ifc_msix_mem != NULL) {
5767 		bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY,
5768 		    rman_get_rid(ctx->ifc_msix_mem), ctx->ifc_msix_mem);
5769 		ctx->ifc_msix_mem = NULL;
5770 	}
5771 }
5772 
5773 int
5774 iflib_device_detach(device_t dev)
5775 {
5776 	if_ctx_t ctx = device_get_softc(dev);
5777 
5778 	return (iflib_device_deregister(ctx));
5779 }
5780 
5781 int
5782 iflib_device_suspend(device_t dev)
5783 {
5784 	if_ctx_t ctx = device_get_softc(dev);
5785 
5786 	CTX_LOCK(ctx);
5787 	IFDI_SUSPEND(ctx);
5788 	CTX_UNLOCK(ctx);
5789 
5790 	return (bus_generic_suspend(dev));
5791 }
5792 int
5793 iflib_device_shutdown(device_t dev)
5794 {
5795 	if_ctx_t ctx = device_get_softc(dev);
5796 
5797 	CTX_LOCK(ctx);
5798 	IFDI_SHUTDOWN(ctx);
5799 	CTX_UNLOCK(ctx);
5800 
5801 	return (bus_generic_suspend(dev));
5802 }
5803 
5804 int
5805 iflib_device_resume(device_t dev)
5806 {
5807 	if_ctx_t ctx = device_get_softc(dev);
5808 	iflib_txq_t txq = ctx->ifc_txqs;
5809 
5810 	CTX_LOCK(ctx);
5811 	IFDI_RESUME(ctx);
5812 	iflib_if_init_locked(ctx);
5813 	CTX_UNLOCK(ctx);
5814 	for (int i = 0; i < NTXQSETS(ctx); i++, txq++)
5815 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
5816 
5817 	return (bus_generic_resume(dev));
5818 }
5819 
5820 int
5821 iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params)
5822 {
5823 	int error;
5824 	if_ctx_t ctx = device_get_softc(dev);
5825 
5826 	CTX_LOCK(ctx);
5827 	error = IFDI_IOV_INIT(ctx, num_vfs, params);
5828 	CTX_UNLOCK(ctx);
5829 
5830 	return (error);
5831 }
5832 
5833 int
5834 iflib_device_iov_init_restart(device_t dev, uint16_t num_vfs,
5835     const nvlist_t *params)
5836 {
5837 	if_ctx_t ctx;
5838 	if_t ifp;
5839 	bool restart, running;
5840 	int error;
5841 
5842 	ctx = device_get_softc(dev);
5843 	ifp = ctx->ifc_ifp;
5844 
5845 	CTX_LOCK(ctx);
5846 	/*
5847 	 * Drivers which change the PF queue layout need the complete iflib
5848 	 * stop/init sequence around their IOV callback when the interface is
5849 	 * active.  An administratively-down interface has no live queues to
5850 	 * quiesce, and must remain down after the new layout is installed.
5851 	 * Keep the transition within one context-lock critical section.
5852 	 */
5853 	restart = (if_getflags(ifp) & IFF_UP) != 0;
5854 	running = (if_getdrvflags(ifp) & IFF_DRV_RUNNING) != 0;
5855 	if (restart || running)
5856 		iflib_stop(ctx);
5857 	error = IFDI_IOV_INIT(ctx, num_vfs, params);
5858 	if (restart)
5859 		iflib_init_locked(ctx);
5860 	CTX_UNLOCK(ctx);
5861 	return (error);
5862 }
5863 
5864 void
5865 iflib_device_iov_uninit(device_t dev)
5866 {
5867 	if_ctx_t ctx = device_get_softc(dev);
5868 
5869 	CTX_LOCK(ctx);
5870 	IFDI_IOV_UNINIT(ctx);
5871 	CTX_UNLOCK(ctx);
5872 }
5873 
5874 void
5875 iflib_device_iov_uninit_restart(device_t dev)
5876 {
5877 	if_ctx_t ctx;
5878 	bool restart;
5879 
5880 	ctx = device_get_softc(dev);
5881 
5882 	CTX_LOCK(ctx);
5883 	/*
5884 	 * RUNNING can be clear while a watchdog reset is pending but the
5885 	 * hardware is still live.  Always stop before the driver changes its
5886 	 * queue layout, and use IFF_UP only to preserve administrative state.
5887 	 */
5888 	restart = (if_getflags(ctx->ifc_ifp) & IFF_UP) != 0;
5889 	iflib_stop(ctx);
5890 	IFDI_IOV_UNINIT(ctx);
5891 	if (restart)
5892 		iflib_init_locked(ctx);
5893 	CTX_UNLOCK(ctx);
5894 }
5895 
5896 int
5897 iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params)
5898 {
5899 	int error;
5900 	if_ctx_t ctx = device_get_softc(dev);
5901 
5902 	CTX_LOCK(ctx);
5903 	error = IFDI_IOV_VF_ADD(ctx, vfnum, params);
5904 	CTX_UNLOCK(ctx);
5905 
5906 	return (error);
5907 }
5908 
5909 /*********************************************************************
5910  *
5911  *  MODULE FUNCTION DEFINITIONS
5912  *
5913  **********************************************************************/
5914 
5915 /*
5916  * - Start a fast taskqueue thread for each core
5917  * - Start a taskqueue for control operations
5918  */
5919 static int
5920 iflib_module_init(void)
5921 {
5922 	iflib_timer_default = hz / 2;
5923 	return (0);
5924 }
5925 
5926 static int
5927 iflib_module_event_handler(module_t mod, int what, void *arg)
5928 {
5929 	int err;
5930 
5931 	switch (what) {
5932 	case MOD_LOAD:
5933 		if ((err = iflib_module_init()) != 0)
5934 			return (err);
5935 		break;
5936 	case MOD_UNLOAD:
5937 		return (EBUSY);
5938 	default:
5939 		return (EOPNOTSUPP);
5940 	}
5941 
5942 	return (0);
5943 }
5944 
5945 /*********************************************************************
5946  *
5947  *  PUBLIC FUNCTION DEFINITIONS
5948  *     ordered as in iflib.h
5949  *
5950  **********************************************************************/
5951 
5952 static void
5953 _iflib_assert(if_shared_ctx_t sctx)
5954 {
5955 	int i;
5956 
5957 	MPASS(sctx->isc_tx_maxsize);
5958 	MPASS(sctx->isc_tx_maxsegsize);
5959 
5960 	MPASS(sctx->isc_rx_maxsize);
5961 	MPASS(sctx->isc_rx_nsegments);
5962 	MPASS(sctx->isc_rx_maxsegsize);
5963 
5964 	MPASS(sctx->isc_nrxqs >= 1 && sctx->isc_nrxqs <= 8);
5965 	for (i = 0; i < sctx->isc_nrxqs; i++) {
5966 		MPASS(sctx->isc_nrxd_min[i]);
5967 		MPASS(powerof2(sctx->isc_nrxd_min[i]));
5968 		MPASS(sctx->isc_nrxd_max[i]);
5969 		MPASS(powerof2(sctx->isc_nrxd_max[i]));
5970 		MPASS(sctx->isc_nrxd_default[i]);
5971 		MPASS(powerof2(sctx->isc_nrxd_default[i]));
5972 	}
5973 
5974 	MPASS(sctx->isc_ntxqs >= 1 && sctx->isc_ntxqs <= 8);
5975 	for (i = 0; i < sctx->isc_ntxqs; i++) {
5976 		MPASS(sctx->isc_ntxd_min[i]);
5977 		MPASS(powerof2(sctx->isc_ntxd_min[i]));
5978 		MPASS(sctx->isc_ntxd_max[i]);
5979 		MPASS(powerof2(sctx->isc_ntxd_max[i]));
5980 		MPASS(sctx->isc_ntxd_default[i]);
5981 		MPASS(powerof2(sctx->isc_ntxd_default[i]));
5982 	}
5983 }
5984 
5985 static void
5986 _iflib_pre_assert(if_softc_ctx_t scctx)
5987 {
5988 
5989 	MPASS(scctx->isc_txrx->ift_txd_encap);
5990 	MPASS(scctx->isc_txrx->ift_txd_flush);
5991 	MPASS(scctx->isc_txrx->ift_txd_credits_update);
5992 	MPASS(scctx->isc_txrx->ift_rxd_available);
5993 	MPASS(scctx->isc_txrx->ift_rxd_pkt_get);
5994 	MPASS(scctx->isc_txrx->ift_rxd_refill);
5995 	MPASS(scctx->isc_txrx->ift_rxd_flush);
5996 }
5997 
5998 static void
5999 iflib_register(if_ctx_t ctx)
6000 {
6001 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6002 	driver_t *driver = sctx->isc_driver;
6003 	device_t dev = ctx->ifc_dev;
6004 	if_t ifp;
6005 
6006 	_iflib_assert(sctx);
6007 
6008 	CTX_LOCK_INIT(ctx);
6009 	STATE_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev));
6010 	ifp = ctx->ifc_ifp = if_alloc_dev(IFT_ETHER, dev);
6011 
6012 	/*
6013 	 * Initialize our context's device specific methods
6014 	 */
6015 	kobj_init((kobj_t) ctx, (kobj_class_t) driver);
6016 	kobj_class_compile((kobj_class_t) driver);
6017 
6018 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
6019 	if_setsoftc(ifp, ctx);
6020 	if_setdev(ifp, dev);
6021 	if_setinitfn(ifp, iflib_if_init);
6022 	if_setioctlfn(ifp, iflib_if_ioctl);
6023 #ifdef ALTQ
6024 	if_setstartfn(ifp, iflib_altq_if_start);
6025 	if_settransmitfn(ifp, iflib_altq_if_transmit);
6026 	if_setsendqready(ifp);
6027 #else
6028 	if_settransmitfn(ifp, iflib_if_transmit);
6029 #endif
6030 	if_setqflushfn(ifp, iflib_if_qflush);
6031 	if_setgetcounterfn(ifp, iflib_if_get_counter);
6032 	if_setflags(ifp, IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST);
6033 	ctx->ifc_vlan_attach_event =
6034 	    EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx,
6035 		    EVENTHANDLER_PRI_FIRST);
6036 	ctx->ifc_vlan_detach_event =
6037 	    EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx,
6038 		    EVENTHANDLER_PRI_FIRST);
6039 
6040 	if ((sctx->isc_flags & IFLIB_DRIVER_MEDIA) == 0) {
6041 		ctx->ifc_mediap = &ctx->ifc_media;
6042 		ifmedia_init(ctx->ifc_mediap, IFM_IMASK,
6043 		    iflib_media_change, iflib_media_status);
6044 	}
6045 }
6046 
6047 static void
6048 iflib_unregister_vlan_handlers(if_ctx_t ctx)
6049 {
6050 	/* Unregister VLAN events */
6051 	if (ctx->ifc_vlan_attach_event != NULL) {
6052 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
6053 		ctx->ifc_vlan_attach_event = NULL;
6054 	}
6055 	if (ctx->ifc_vlan_detach_event != NULL) {
6056 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
6057 		ctx->ifc_vlan_detach_event = NULL;
6058 	}
6059 
6060 }
6061 
6062 static void
6063 iflib_deregister(if_ctx_t ctx)
6064 {
6065 	if_t ifp = ctx->ifc_ifp;
6066 
6067 	/* Remove all media */
6068 	ifmedia_removeall(&ctx->ifc_media);
6069 
6070 	/* Ensure that VLAN event handlers are unregistered */
6071 	iflib_unregister_vlan_handlers(ctx);
6072 
6073 	/* Release kobject reference */
6074 	kobj_delete((kobj_t) ctx, NULL);
6075 
6076 	/* Free the ifnet structure */
6077 	if_free(ifp);
6078 
6079 	STATE_LOCK_DESTROY(ctx);
6080 
6081 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
6082 	CTX_LOCK_DESTROY(ctx);
6083 }
6084 
6085 static int
6086 iflib_queues_alloc(if_ctx_t ctx)
6087 {
6088 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6089 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6090 	device_t dev = ctx->ifc_dev;
6091 	int nrxqsets = scctx->isc_nrxqsets;
6092 	int ntxqsets = scctx->isc_ntxqsets;
6093 	iflib_txq_t txq;
6094 	iflib_rxq_t rxq;
6095 	iflib_fl_t fl = NULL;
6096 	int i, j, cpu, err, txconf, rxconf;
6097 	iflib_dma_info_t ifdip;
6098 	uint32_t *rxqsizes = scctx->isc_rxqsizes;
6099 	uint32_t *txqsizes = scctx->isc_txqsizes;
6100 	uint8_t nrxqs = sctx->isc_nrxqs;
6101 	uint8_t ntxqs = sctx->isc_ntxqs;
6102 	int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1;
6103 	int fl_offset = (sctx->isc_flags & IFLIB_HAS_RXCQ ? 1 : 0);
6104 	caddr_t *vaddrs;
6105 	uint64_t *paddrs;
6106 
6107 	KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1"));
6108 	KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1"));
6109 	KASSERT(nrxqs >= fl_offset + nfree_lists,
6110 	    ("there must be at least a rxq for each free list"));
6111 
6112 	/* Allocate the TX ring struct memory */
6113 	if (!(ctx->ifc_txqs =
6114 	    (iflib_txq_t) malloc(sizeof(struct iflib_txq) *
6115 		    ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
6116 		device_printf(dev, "Unable to allocate TX ring memory\n");
6117 		err = ENOMEM;
6118 		goto fail;
6119 	}
6120 
6121 	/* Now allocate the RX */
6122 	if (!(ctx->ifc_rxqs =
6123 	    (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) *
6124 		    nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
6125 		device_printf(dev, "Unable to allocate RX ring memory\n");
6126 		err = ENOMEM;
6127 		goto rx_fail;
6128 	}
6129 
6130 	txq = ctx->ifc_txqs;
6131 	rxq = ctx->ifc_rxqs;
6132 
6133 	/*
6134 	 * XXX handle allocation failure
6135 	 */
6136 	for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) {
6137 		/* Set up some basics */
6138 
6139 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs,
6140 		    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
6141 			device_printf(dev,
6142 			    "Unable to allocate TX DMA info memory\n");
6143 			err = ENOMEM;
6144 			goto err_tx_desc;
6145 		}
6146 		txq->ift_ifdi = ifdip;
6147 		for (j = 0; j < ntxqs; j++, ifdip++) {
6148 			if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, 0)) {
6149 				device_printf(dev,
6150 				    "Unable to allocate TX descriptors\n");
6151 				err = ENOMEM;
6152 				goto err_tx_desc;
6153 			}
6154 			txq->ift_txd_size[j] = scctx->isc_txd_size[j];
6155 			bzero((void *)ifdip->idi_vaddr, txqsizes[j]);
6156 		}
6157 		txq->ift_ctx = ctx;
6158 		txq->ift_id = i;
6159 		if (sctx->isc_flags & IFLIB_HAS_TXCQ) {
6160 			txq->ift_br_offset = 1;
6161 		} else {
6162 			txq->ift_br_offset = 0;
6163 		}
6164 
6165 		if (iflib_txsd_alloc(txq)) {
6166 			device_printf(dev, "Critical Failure setting up TX buffers\n");
6167 			err = ENOMEM;
6168 			goto err_tx_desc;
6169 		}
6170 
6171 		/* Initialize the TX lock */
6172 		snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:TX(%d):callout",
6173 		    device_get_nameunit(dev), txq->ift_id);
6174 		mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF);
6175 		callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0);
6176 		txq->ift_timer.c_cpu = cpu;
6177 #ifdef DEV_NETMAP
6178 		callout_init_mtx(&txq->ift_netmap_timer, &txq->ift_mtx, 0);
6179 		txq->ift_netmap_timer.c_cpu = cpu;
6180 #endif /* DEV_NETMAP */
6181 
6182 		err = ifmp_ring_alloc(&txq->ift_br, 2048, txq, iflib_txq_drain,
6183 		    iflib_txq_can_drain, M_IFLIB, M_WAITOK);
6184 		if (err) {
6185 			/* XXX free any allocated rings */
6186 			device_printf(dev, "Unable to allocate buf_ring\n");
6187 			goto err_tx_desc;
6188 		}
6189 		txq->ift_reclaim_thresh = ctx->ifc_sysctl_tx_reclaim_thresh;
6190 	}
6191 
6192 	for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) {
6193 		/* Set up some basics */
6194 		callout_init(&rxq->ifr_watchdog, 1);
6195 
6196 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs,
6197 		    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
6198 			device_printf(dev,
6199 			    "Unable to allocate RX DMA info memory\n");
6200 			err = ENOMEM;
6201 			goto err_tx_desc;
6202 		}
6203 
6204 		rxq->ifr_ifdi = ifdip;
6205 		/* XXX this needs to be changed if #rx queues != #tx queues */
6206 		rxq->ifr_ntxqirq = 1;
6207 		rxq->ifr_txqid[0] = i;
6208 		for (j = 0; j < nrxqs; j++, ifdip++) {
6209 			if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, 0)) {
6210 				device_printf(dev,
6211 				    "Unable to allocate RX descriptors\n");
6212 				err = ENOMEM;
6213 				goto err_tx_desc;
6214 			}
6215 			bzero((void *)ifdip->idi_vaddr, rxqsizes[j]);
6216 		}
6217 		rxq->ifr_ctx = ctx;
6218 		rxq->ifr_id = i;
6219 		rxq->ifr_fl_offset = fl_offset;
6220 		rxq->ifr_nfl = nfree_lists;
6221 		if (!(fl =
6222 		    (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) {
6223 			device_printf(dev, "Unable to allocate free list memory\n");
6224 			err = ENOMEM;
6225 			goto err_tx_desc;
6226 		}
6227 		rxq->ifr_fl = fl;
6228 		for (j = 0; j < nfree_lists; j++) {
6229 			fl[j].ifl_rxq = rxq;
6230 			fl[j].ifl_id = j;
6231 			fl[j].ifl_ifdi = &rxq->ifr_ifdi[j + rxq->ifr_fl_offset];
6232 			fl[j].ifl_rxd_size = scctx->isc_rxd_size[j];
6233 		}
6234 		/* Allocate receive buffers for the ring */
6235 		if (iflib_rxsd_alloc(rxq)) {
6236 			device_printf(dev,
6237 			    "Critical Failure setting up receive buffers\n");
6238 			err = ENOMEM;
6239 			goto err_rx_desc;
6240 		}
6241 
6242 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
6243 			fl->ifl_rx_bitmap = bit_alloc(fl->ifl_size, M_IFLIB,
6244 			    M_WAITOK);
6245 	}
6246 
6247 	/* TXQs */
6248 	vaddrs = malloc(sizeof(caddr_t)  * ntxqsets * ntxqs, M_IFLIB, M_WAITOK);
6249 	paddrs = malloc(sizeof(uint64_t) * ntxqsets * ntxqs, M_IFLIB, M_WAITOK);
6250 	for (i = 0; i < ntxqsets; i++) {
6251 		iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi;
6252 
6253 		for (j = 0; j < ntxqs; j++, di++) {
6254 			vaddrs[i * ntxqs + j] = di->idi_vaddr;
6255 			paddrs[i * ntxqs + j] = di->idi_paddr;
6256 		}
6257 	}
6258 	if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) {
6259 		device_printf(ctx->ifc_dev,
6260 		    "Unable to allocate device TX queue\n");
6261 		iflib_tx_structures_free(ctx);
6262 		free(vaddrs, M_IFLIB);
6263 		free(paddrs, M_IFLIB);
6264 		goto err_rx_desc;
6265 	}
6266 	free(vaddrs, M_IFLIB);
6267 	free(paddrs, M_IFLIB);
6268 
6269 	/* RXQs */
6270 	vaddrs = malloc(sizeof(caddr_t)  * nrxqsets * nrxqs, M_IFLIB, M_WAITOK);
6271 	paddrs = malloc(sizeof(uint64_t) * nrxqsets * nrxqs, M_IFLIB, M_WAITOK);
6272 	for (i = 0; i < nrxqsets; i++) {
6273 		iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi;
6274 
6275 		for (j = 0; j < nrxqs; j++, di++) {
6276 			vaddrs[i * nrxqs + j] = di->idi_vaddr;
6277 			paddrs[i * nrxqs + j] = di->idi_paddr;
6278 		}
6279 	}
6280 	if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) {
6281 		device_printf(ctx->ifc_dev,
6282 		    "Unable to allocate device RX queue\n");
6283 		iflib_tx_structures_free(ctx);
6284 		free(vaddrs, M_IFLIB);
6285 		free(paddrs, M_IFLIB);
6286 		goto err_rx_desc;
6287 	}
6288 	free(vaddrs, M_IFLIB);
6289 	free(paddrs, M_IFLIB);
6290 
6291 	return (0);
6292 
6293 /* XXX handle allocation failure changes */
6294 err_rx_desc:
6295 err_tx_desc:
6296 rx_fail:
6297 	if (ctx->ifc_rxqs != NULL)
6298 		free(ctx->ifc_rxqs, M_IFLIB);
6299 	ctx->ifc_rxqs = NULL;
6300 	if (ctx->ifc_txqs != NULL)
6301 		free(ctx->ifc_txqs, M_IFLIB);
6302 	ctx->ifc_txqs = NULL;
6303 fail:
6304 	return (err);
6305 }
6306 
6307 static int
6308 iflib_tx_structures_setup(if_ctx_t ctx)
6309 {
6310 	iflib_txq_t txq = ctx->ifc_txqs;
6311 	int i;
6312 
6313 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
6314 		iflib_txq_setup(txq);
6315 
6316 	return (0);
6317 }
6318 
6319 static void
6320 iflib_tx_structures_free(if_ctx_t ctx)
6321 {
6322 	iflib_txq_t txq = ctx->ifc_txqs;
6323 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6324 	int i, j;
6325 
6326 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
6327 		for (j = 0; j < sctx->isc_ntxqs; j++)
6328 			iflib_dma_free(&txq->ift_ifdi[j]);
6329 		iflib_txq_destroy(txq);
6330 	}
6331 	free(ctx->ifc_txqs, M_IFLIB);
6332 	ctx->ifc_txqs = NULL;
6333 }
6334 
6335 /*********************************************************************
6336  *
6337  *  Initialize all receive rings.
6338  *
6339  **********************************************************************/
6340 static int
6341 iflib_rx_structures_setup(if_ctx_t ctx)
6342 {
6343 	iflib_rxq_t rxq = ctx->ifc_rxqs;
6344 	int q;
6345 #if defined(INET6) || defined(INET)
6346 	int err, i;
6347 #endif
6348 
6349 	for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) {
6350 #if defined(INET6) || defined(INET)
6351 		err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp,
6352 		    TCP_LRO_ENTRIES, min(1024,
6353 		    ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset]));
6354 		if (err != 0) {
6355 			device_printf(ctx->ifc_dev,
6356 			    "LRO Initialization failed!\n");
6357 			goto fail;
6358 		}
6359 #endif
6360 		IFDI_RXQ_SETUP(ctx, rxq->ifr_id);
6361 	}
6362 	return (0);
6363 #if defined(INET6) || defined(INET)
6364 fail:
6365 	/*
6366 	 * Free LRO resources allocated so far, we will only handle
6367 	 * the rings that completed, the failing case will have
6368 	 * cleaned up for itself.  'q' failed, so its the terminus.
6369 	 */
6370 	rxq = ctx->ifc_rxqs;
6371 	for (i = 0; i < q; ++i, rxq++) {
6372 		tcp_lro_free(&rxq->ifr_lc);
6373 	}
6374 	return (err);
6375 #endif
6376 }
6377 
6378 /*********************************************************************
6379  *
6380  *  Free all receive rings.
6381  *
6382  **********************************************************************/
6383 static void
6384 iflib_rx_structures_free(if_ctx_t ctx)
6385 {
6386 	iflib_rxq_t rxq = ctx->ifc_rxqs;
6387 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6388 	int i, j;
6389 
6390 	for (i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) {
6391 		for (j = 0; j < sctx->isc_nrxqs; j++)
6392 			iflib_dma_free(&rxq->ifr_ifdi[j]);
6393 		iflib_rx_sds_free(rxq);
6394 #if defined(INET6) || defined(INET)
6395 		tcp_lro_free(&rxq->ifr_lc);
6396 #endif
6397 	}
6398 	free(ctx->ifc_rxqs, M_IFLIB);
6399 	ctx->ifc_rxqs = NULL;
6400 }
6401 
6402 static int
6403 iflib_qset_structures_setup(if_ctx_t ctx)
6404 {
6405 	int err;
6406 
6407 	/*
6408 	 * It is expected that the caller takes care of freeing queues if this
6409 	 * fails.
6410 	 */
6411 	if ((err = iflib_tx_structures_setup(ctx)) != 0) {
6412 		device_printf(ctx->ifc_dev, "iflib_tx_structures_setup failed: %d\n", err);
6413 		return (err);
6414 	}
6415 
6416 	if ((err = iflib_rx_structures_setup(ctx)) != 0)
6417 		device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err);
6418 
6419 	return (err);
6420 }
6421 
6422 int
6423 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
6424 		driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, const char *name)
6425 {
6426 
6427 	return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name));
6428 }
6429 
6430 /* Just to avoid copy/paste */
6431 static inline int
6432 iflib_irq_set_affinity(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type,
6433     int qid, struct grouptask *gtask, struct taskqgroup *tqg, void *uniq,
6434     const char *name)
6435 {
6436 	device_t dev;
6437 	unsigned int base_cpuid, cpuid;
6438 	int err;
6439 
6440 	dev = ctx->ifc_dev;
6441 	base_cpuid = ctx->ifc_sysctl_core_offset;
6442 	cpuid = get_cpuid_for_queue(ctx, base_cpuid, qid, type == IFLIB_INTR_TX);
6443 	err = taskqgroup_attach_cpu(tqg, gtask, uniq, cpuid, dev,
6444 	    irq ? irq->ii_res : NULL, name);
6445 	if (err) {
6446 		device_printf(dev, "taskqgroup_attach_cpu failed %d\n", err);
6447 		return (err);
6448 	}
6449 #ifdef notyet
6450 	if (cpuid > ctx->ifc_cpuid_highest)
6451 		ctx->ifc_cpuid_highest = cpuid;
6452 #endif
6453 	return (0);
6454 }
6455 
6456 /*
6457  * Allocate a hardware interrupt for subctx using the parent (ctx)'s hardware
6458  * resources.
6459  *
6460  * Similar to iflib_irq_alloc_generic(), but for interrupt type IFLIB_INTR_RXTX
6461  * only.
6462  *
6463  * XXX: Could be removed if subctx's dev has its intr resource allocation
6464  * methods replaced with custom ones?
6465  */
6466 int
6467 iflib_irq_alloc_generic_subctx(if_ctx_t ctx, if_ctx_t subctx, if_irq_t irq,
6468 			       int rid, iflib_intr_type_t type,
6469 			       driver_filter_t *filter, void *filter_arg,
6470 			       int qid, const char *name)
6471 {
6472 	device_t dev, subdev;
6473 	struct grouptask *gtask;
6474 	struct taskqgroup *tqg;
6475 	iflib_filter_info_t info;
6476 	gtask_fn_t *fn;
6477 	int tqrid, err;
6478 	driver_filter_t *intr_fast;
6479 	void *q;
6480 
6481 	MPASS(ctx != NULL);
6482 	MPASS(subctx != NULL);
6483 
6484 	tqrid = rid;
6485 	dev = ctx->ifc_dev;
6486 	subdev = subctx->ifc_dev;
6487 
6488 	switch (type) {
6489 	case IFLIB_INTR_RXTX:
6490 		q = &subctx->ifc_rxqs[qid];
6491 		info = &subctx->ifc_rxqs[qid].ifr_filter_info;
6492 		gtask = &subctx->ifc_rxqs[qid].ifr_task;
6493 		tqg = qgroup_if_io_tqg;
6494 		fn = _task_fn_rx;
6495 		intr_fast = iflib_fast_intr_rxtx;
6496 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6497 		break;
6498 	default:
6499 		device_printf(dev, "%s: unknown net intr type for subctx %s (%d)\n",
6500 		    __func__, device_get_nameunit(subdev), type);
6501 		return (EINVAL);
6502 	}
6503 
6504 	info->ifi_filter = filter;
6505 	info->ifi_filter_arg = filter_arg;
6506 	info->ifi_task = gtask;
6507 	info->ifi_ctx = q;
6508 
6509 	NET_GROUPTASK_INIT(gtask, 0, fn, q);
6510 
6511 	/* Allocate interrupts from hardware using parent context */
6512 	err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info, name);
6513 	if (err != 0) {
6514 		device_printf(dev, "_iflib_irq_alloc failed for subctx %s: %d\n",
6515 		    device_get_nameunit(subdev), err);
6516 		return (err);
6517 	}
6518 
6519 	if (tqrid != -1) {
6520 		err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg, q,
6521 		    name);
6522 		if (err)
6523 			return (err);
6524 	} else {
6525 		taskqgroup_attach(tqg, gtask, q, dev, irq->ii_res, name);
6526 	}
6527 
6528 	return (0);
6529 }
6530 
6531 int
6532 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid,
6533 			iflib_intr_type_t type, driver_filter_t *filter,
6534 			void *filter_arg, int qid, const char *name)
6535 {
6536 	device_t dev;
6537 	struct grouptask *gtask;
6538 	struct taskqgroup *tqg;
6539 	iflib_filter_info_t info;
6540 	gtask_fn_t *fn;
6541 	int tqrid, err;
6542 	driver_filter_t *intr_fast;
6543 	void *q;
6544 
6545 	info = &ctx->ifc_filter_info;
6546 	tqrid = rid;
6547 
6548 	switch (type) {
6549 	/* XXX merge tx/rx for netmap? */
6550 	case IFLIB_INTR_TX:
6551 		q = &ctx->ifc_txqs[qid];
6552 		info = &ctx->ifc_txqs[qid].ift_filter_info;
6553 		gtask = &ctx->ifc_txqs[qid].ift_task;
6554 		tqg = qgroup_if_io_tqg;
6555 		fn = _task_fn_tx;
6556 		intr_fast = iflib_fast_intr;
6557 		GROUPTASK_INIT(gtask, 0, fn, q);
6558 		ctx->ifc_flags |= IFC_NETMAP_TX_IRQ;
6559 		break;
6560 	case IFLIB_INTR_RX:
6561 		q = &ctx->ifc_rxqs[qid];
6562 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
6563 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6564 		tqg = qgroup_if_io_tqg;
6565 		fn = _task_fn_rx;
6566 		intr_fast = iflib_fast_intr;
6567 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6568 		break;
6569 	case IFLIB_INTR_RXTX:
6570 		q = &ctx->ifc_rxqs[qid];
6571 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
6572 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6573 		tqg = qgroup_if_io_tqg;
6574 		fn = _task_fn_rx;
6575 		intr_fast = iflib_fast_intr_rxtx;
6576 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6577 		break;
6578 	case IFLIB_INTR_ADMIN:
6579 		q = ctx;
6580 		tqrid = -1;
6581 		info = &ctx->ifc_filter_info;
6582 		gtask = NULL;
6583 		intr_fast = iflib_fast_intr_ctx;
6584 		break;
6585 	default:
6586 		device_printf(ctx->ifc_dev, "%s: unknown net intr type\n",
6587 		    __func__);
6588 		return (EINVAL);
6589 	}
6590 
6591 	info->ifi_filter = filter;
6592 	info->ifi_filter_arg = filter_arg;
6593 	info->ifi_task = gtask;
6594 	info->ifi_ctx = q;
6595 
6596 	dev = ctx->ifc_dev;
6597 	err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info,  name);
6598 	if (err != 0) {
6599 		device_printf(dev, "_iflib_irq_alloc failed %d\n", err);
6600 		return (err);
6601 	}
6602 	if (type == IFLIB_INTR_ADMIN)
6603 		return (0);
6604 
6605 	if (tqrid != -1) {
6606 		err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg, q,
6607 		    name);
6608 		if (err)
6609 			return (err);
6610 	} else {
6611 		taskqgroup_attach(tqg, gtask, q, dev, irq->ii_res, name);
6612 	}
6613 
6614 	return (0);
6615 }
6616 
6617 void
6618 iflib_softirq_alloc_generic(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type,
6619 			    void *arg, int qid, const char *name)
6620 {
6621 	device_t dev;
6622 	struct grouptask *gtask;
6623 	struct taskqgroup *tqg;
6624 	gtask_fn_t *fn;
6625 	void *q;
6626 	int err;
6627 
6628 	switch (type) {
6629 	case IFLIB_INTR_TX:
6630 		q = &ctx->ifc_txqs[qid];
6631 		gtask = &ctx->ifc_txqs[qid].ift_task;
6632 		tqg = qgroup_if_io_tqg;
6633 		fn = _task_fn_tx;
6634 		GROUPTASK_INIT(gtask, 0, fn, q);
6635 		break;
6636 	case IFLIB_INTR_RX:
6637 		q = &ctx->ifc_rxqs[qid];
6638 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6639 		tqg = qgroup_if_io_tqg;
6640 		fn = _task_fn_rx;
6641 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6642 		break;
6643 	case IFLIB_INTR_IOV:
6644 		return;
6645 	default:
6646 		panic("unknown net intr type");
6647 	}
6648 	err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg, q, name);
6649 	if (err) {
6650 		dev = ctx->ifc_dev;
6651 		taskqgroup_attach(tqg, gtask, q, dev, irq ? irq->ii_res : NULL,
6652 		    name);
6653 	}
6654 }
6655 
6656 void
6657 iflib_irq_free(if_ctx_t ctx, if_irq_t irq)
6658 {
6659 
6660 	if (irq->ii_tag)
6661 		bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag);
6662 
6663 	if (irq->ii_res)
6664 		bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ,
6665 		    rman_get_rid(irq->ii_res), irq->ii_res);
6666 }
6667 
6668 static int
6669 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, const char *name)
6670 {
6671 	iflib_txq_t txq = ctx->ifc_txqs;
6672 	iflib_rxq_t rxq = ctx->ifc_rxqs;
6673 	if_irq_t irq = &ctx->ifc_legacy_irq;
6674 	iflib_filter_info_t info;
6675 	device_t dev;
6676 	struct grouptask *gtask;
6677 	struct resource *res;
6678 	int err, tqrid;
6679 	bool rx_only;
6680 
6681 	info = &rxq->ifr_filter_info;
6682 	gtask = &rxq->ifr_task;
6683 	tqrid = *rid;
6684 	rx_only = (ctx->ifc_sctx->isc_flags & IFLIB_SINGLE_IRQ_RX_ONLY) != 0;
6685 
6686 	ctx->ifc_flags |= IFC_LEGACY;
6687 	info->ifi_filter = filter;
6688 	info->ifi_filter_arg = filter_arg;
6689 	info->ifi_task = gtask;
6690 	info->ifi_ctx = rxq;
6691 
6692 	dev = ctx->ifc_dev;
6693 	/* We allocate a single interrupt resource */
6694 	err = _iflib_irq_alloc(ctx, irq, tqrid, rx_only ? iflib_fast_intr :
6695 	    iflib_fast_intr_rxtx, NULL, info, name);
6696 	if (err != 0)
6697 		return (err);
6698 	NET_GROUPTASK_INIT(gtask, 0, _task_fn_rx, rxq);
6699 	res = irq->ii_res;
6700 	taskqgroup_attach(qgroup_if_io_tqg, gtask, rxq, dev, res, name);
6701 
6702 	GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq);
6703 	taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, dev, res,
6704 	    "tx");
6705 	return (0);
6706 }
6707 
6708 void
6709 iflib_led_create(if_ctx_t ctx)
6710 {
6711 
6712 	ctx->ifc_led_dev = led_create(iflib_led_func, ctx,
6713 	    device_get_nameunit(ctx->ifc_dev));
6714 }
6715 
6716 void
6717 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid)
6718 {
6719 
6720 	GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
6721 }
6722 
6723 void
6724 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid)
6725 {
6726 
6727 	GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task);
6728 }
6729 
6730 void
6731 iflib_admin_intr_deferred(if_ctx_t ctx)
6732 {
6733 
6734 	taskqueue_enqueue(ctx->ifc_tq, &ctx->ifc_admin_task);
6735 }
6736 
6737 void
6738 iflib_iov_intr_deferred(if_ctx_t ctx)
6739 {
6740 
6741 	taskqueue_enqueue(ctx->ifc_tq, &ctx->ifc_vflr_task);
6742 }
6743 
6744 void
6745 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, const char *name)
6746 {
6747 
6748 	taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, NULL, NULL,
6749 	    name);
6750 }
6751 
6752 void
6753 iflib_config_task_init(if_ctx_t ctx, struct task *config_task, task_fn_t *fn)
6754 {
6755 	TASK_INIT(config_task, 0, fn, ctx);
6756 }
6757 
6758 void
6759 iflib_config_task_enqueue(if_ctx_t ctx, struct task *config_task)
6760 {
6761 	taskqueue_enqueue(ctx->ifc_tq, config_task);
6762 }
6763 
6764 void
6765 iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate)
6766 {
6767 	if_t ifp = ctx->ifc_ifp;
6768 
6769 	if_setbaudrate(ifp, baudrate);
6770 	if (baudrate >= IF_Gbps(10)) {
6771 		STATE_LOCK(ctx);
6772 		ctx->ifc_flags |= IFC_PREFETCH;
6773 		STATE_UNLOCK(ctx);
6774 	}
6775 	ctx->ifc_link_state = link_state;
6776 	if_link_state_change(ifp, link_state);
6777 }
6778 
6779 static int
6780 iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq)
6781 {
6782 	int credits;
6783 #ifdef INVARIANTS
6784 	int credits_pre = txq->ift_cidx_processed;
6785 #endif
6786 
6787 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
6788 	    BUS_DMASYNC_POSTREAD);
6789 	if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, true)) == 0)
6790 		return (0);
6791 
6792 	txq->ift_processed += credits;
6793 	txq->ift_cidx_processed += credits;
6794 
6795 	MPASS(credits_pre + credits == txq->ift_cidx_processed);
6796 	if (txq->ift_cidx_processed >= txq->ift_size)
6797 		txq->ift_cidx_processed -= txq->ift_size;
6798 	return (credits);
6799 }
6800 
6801 static int
6802 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget)
6803 {
6804 	iflib_fl_t fl;
6805 	u_int i;
6806 
6807 	for (i = 0, fl = &rxq->ifr_fl[0]; i < rxq->ifr_nfl; i++, fl++)
6808 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
6809 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
6810 	return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx,
6811 	    budget));
6812 }
6813 
6814 void
6815 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name,
6816 	const char *description, if_int_delay_info_t info,
6817 	int offset, int value)
6818 {
6819 	info->iidi_ctx = ctx;
6820 	info->iidi_offset = offset;
6821 	info->iidi_value = value;
6822 	SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev),
6823 	    SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)),
6824 	    OID_AUTO, name, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE,
6825 	    info, 0, iflib_sysctl_int_delay, "I", description);
6826 }
6827 
6828 struct sx *
6829 iflib_ctx_lock_get(if_ctx_t ctx)
6830 {
6831 
6832 	return (&ctx->ifc_ctx_sx);
6833 }
6834 
6835 static int
6836 iflib_msix_init(if_ctx_t ctx)
6837 {
6838 	device_t dev = ctx->ifc_dev;
6839 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6840 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6841 	int admincnt, bar, err, iflib_num_rx_queues, iflib_num_tx_queues;
6842 	int msgs, queuemsgs, queues, rx_queues, tx_queues, vectors;
6843 
6844 	iflib_num_tx_queues = ctx->ifc_sysctl_ntxqs;
6845 	iflib_num_rx_queues = ctx->ifc_sysctl_nrxqs;
6846 
6847 	if (bootverbose)
6848 		device_printf(dev, "msix_init qsets capped at %d\n",
6849 		    imax(scctx->isc_ntxqsets, scctx->isc_nrxqsets));
6850 
6851 	/* Override by tuneable */
6852 	if (scctx->isc_disable_msix)
6853 		goto msi;
6854 
6855 	/* First try MSI-X */
6856 	if ((msgs = pci_msix_count(dev)) == 0) {
6857 		if (bootverbose)
6858 			device_printf(dev, "MSI-X not supported or disabled\n");
6859 		goto msi;
6860 	}
6861 
6862 	bar = ctx->ifc_softc_ctx.isc_msix_bar;
6863 	/*
6864 	 * bar == -1 => "trust me I know what I'm doing"
6865 	 * Some drivers are for hardware that is so shoddily
6866 	 * documented that no one knows which bars are which
6867 	 * so the developer has to map all bars. This hack
6868 	 * allows shoddy garbage to use MSI-X in this framework.
6869 	 */
6870 	if (bar != -1) {
6871 		ctx->ifc_msix_mem = bus_alloc_resource_any(dev,
6872 		    SYS_RES_MEMORY, &bar, RF_ACTIVE);
6873 		if (ctx->ifc_msix_mem == NULL) {
6874 			device_printf(dev, "Unable to map MSI-X table\n");
6875 			goto msi;
6876 		}
6877 	}
6878 
6879 	admincnt = sctx->isc_admin_intrcnt;
6880 #if IFLIB_DEBUG
6881 	/* use only 1 qset in debug mode */
6882 	queuemsgs = min(msgs - admincnt, 1);
6883 #else
6884 	queuemsgs = msgs - admincnt;
6885 #endif
6886 #ifdef RSS
6887 	queues = imin(queuemsgs, rss_getnumbuckets());
6888 #else
6889 	queues = queuemsgs;
6890 #endif
6891 	queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues);
6892 	if (bootverbose)
6893 		device_printf(dev,
6894 		    "intr CPUs: %d queue msgs: %d admincnt: %d\n",
6895 		    CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt);
6896 #ifdef  RSS
6897 	/* If we're doing RSS, clamp at the number of RSS buckets */
6898 	if (queues > rss_getnumbuckets())
6899 		queues = rss_getnumbuckets();
6900 #endif
6901 	if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt)
6902 		rx_queues = iflib_num_rx_queues;
6903 	else
6904 		rx_queues = queues;
6905 
6906 	if (rx_queues > scctx->isc_nrxqsets)
6907 		rx_queues = scctx->isc_nrxqsets;
6908 
6909 	/*
6910 	 * We want this to be all logical CPUs by default
6911 	 */
6912 	if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues)
6913 		tx_queues = iflib_num_tx_queues;
6914 	else
6915 		tx_queues = mp_ncpus;
6916 
6917 	if (tx_queues > scctx->isc_ntxqsets)
6918 		tx_queues = scctx->isc_ntxqsets;
6919 
6920 	if (ctx->ifc_sysctl_qs_eq_override == 0) {
6921 #ifdef INVARIANTS
6922 		if (tx_queues != rx_queues)
6923 			device_printf(dev,
6924 			    "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n",
6925 			    min(rx_queues, tx_queues), min(rx_queues, tx_queues));
6926 #endif
6927 		tx_queues = min(rx_queues, tx_queues);
6928 		rx_queues = min(rx_queues, tx_queues);
6929 	}
6930 
6931 	vectors = rx_queues + admincnt;
6932 	if (msgs < vectors) {
6933 		device_printf(dev,
6934 		    "insufficient number of MSI-X vectors "
6935 		    "(supported %d, need %d)\n", msgs, vectors);
6936 		goto msi;
6937 	}
6938 
6939 	device_printf(dev, "Using %d RX queues %d TX queues\n", rx_queues,
6940 	    tx_queues);
6941 	msgs = vectors;
6942 	if ((err = pci_alloc_msix(dev, &vectors)) == 0) {
6943 		if (vectors != msgs) {
6944 			device_printf(dev,
6945 			    "Unable to allocate sufficient MSI-X vectors "
6946 			    "(got %d, need %d)\n", vectors, msgs);
6947 			pci_release_msi(dev);
6948 			if (bar != -1) {
6949 				bus_release_resource(dev, SYS_RES_MEMORY, bar,
6950 				    ctx->ifc_msix_mem);
6951 				ctx->ifc_msix_mem = NULL;
6952 			}
6953 			goto msi;
6954 		}
6955 		device_printf(dev, "Using MSI-X interrupts with %d vectors\n",
6956 		    vectors);
6957 		scctx->isc_vectors = vectors;
6958 		scctx->isc_nrxqsets = rx_queues;
6959 		scctx->isc_ntxqsets = tx_queues;
6960 		scctx->isc_intr = IFLIB_INTR_MSIX;
6961 
6962 		return (vectors);
6963 	} else {
6964 		device_printf(dev,
6965 		    "failed to allocate %d MSI-X vectors, err: %d\n", vectors,
6966 		    err);
6967 		if (bar != -1) {
6968 			bus_release_resource(dev, SYS_RES_MEMORY, bar,
6969 			    ctx->ifc_msix_mem);
6970 			ctx->ifc_msix_mem = NULL;
6971 		}
6972 	}
6973 
6974 msi:
6975 	vectors = pci_msi_count(dev);
6976 	scctx->isc_nrxqsets = 1;
6977 	scctx->isc_ntxqsets = 1;
6978 	scctx->isc_vectors = vectors;
6979 	if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) {
6980 		device_printf(dev, "Using an MSI interrupt\n");
6981 		scctx->isc_intr = IFLIB_INTR_MSI;
6982 	} else {
6983 		scctx->isc_vectors = 1;
6984 		device_printf(dev, "Using a Legacy interrupt\n");
6985 		scctx->isc_intr = IFLIB_INTR_LEGACY;
6986 	}
6987 
6988 	return (vectors);
6989 }
6990 
6991 static const char *ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" };
6992 
6993 static int
6994 mp_ring_state_handler(SYSCTL_HANDLER_ARGS)
6995 {
6996 	int rc;
6997 	uint16_t *state = ((uint16_t *)oidp->oid_arg1);
6998 	struct sbuf *sb;
6999 	const char *ring_state = "UNKNOWN";
7000 
7001 	/* XXX needed ? */
7002 	rc = sysctl_wire_old_buffer(req, 0);
7003 	MPASS(rc == 0);
7004 	if (rc != 0)
7005 		return (rc);
7006 	sb = sbuf_new_for_sysctl(NULL, NULL, 80, req);
7007 	MPASS(sb != NULL);
7008 	if (sb == NULL)
7009 		return (ENOMEM);
7010 	if (state[3] <= 3)
7011 		ring_state = ring_states[state[3]];
7012 
7013 	sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s",
7014 		    state[0], state[1], state[2], ring_state);
7015 	rc = sbuf_finish(sb);
7016 	sbuf_delete(sb);
7017 	return (rc);
7018 }
7019 
7020 enum iflib_ndesc_handler {
7021 	IFLIB_NTXD_HANDLER,
7022 	IFLIB_NRXD_HANDLER,
7023 };
7024 
7025 static int
7026 mp_ndesc_handler(SYSCTL_HANDLER_ARGS)
7027 {
7028 	if_ctx_t ctx = (void *)arg1;
7029 	enum iflib_ndesc_handler type = arg2;
7030 	char buf[256] = {0};
7031 	qidx_t *ndesc;
7032 	char *p, *next;
7033 	int nqs, rc, i;
7034 
7035 	nqs = 8;
7036 	switch (type) {
7037 	case IFLIB_NTXD_HANDLER:
7038 		ndesc = ctx->ifc_sysctl_ntxds;
7039 		if (ctx->ifc_sctx)
7040 			nqs = ctx->ifc_sctx->isc_ntxqs;
7041 		break;
7042 	case IFLIB_NRXD_HANDLER:
7043 		ndesc = ctx->ifc_sysctl_nrxds;
7044 		if (ctx->ifc_sctx)
7045 			nqs = ctx->ifc_sctx->isc_nrxqs;
7046 		break;
7047 	default:
7048 		printf("%s: unhandled type\n", __func__);
7049 		return (EINVAL);
7050 	}
7051 	if (nqs == 0)
7052 		nqs = 8;
7053 
7054 	for (i = 0; i < 8; i++) {
7055 		if (i >= nqs)
7056 			break;
7057 		if (i)
7058 			strcat(buf, ",");
7059 		sprintf(strchr(buf, 0), "%d", ndesc[i]);
7060 	}
7061 
7062 	rc = sysctl_handle_string(oidp, buf, sizeof(buf), req);
7063 	if (rc || req->newptr == NULL)
7064 		return (rc);
7065 
7066 	for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p;
7067 	    i++, p = strsep(&next, " ,")) {
7068 		ndesc[i] = strtoul(p, NULL, 10);
7069 	}
7070 
7071 	return (rc);
7072 }
7073 
7074 static int
7075 iflib_handle_tx_reclaim_thresh(SYSCTL_HANDLER_ARGS)
7076 {
7077 	if_ctx_t ctx = (void *)arg1;
7078 	iflib_txq_t txq;
7079 	int i, err;
7080 	int thresh;
7081 
7082 	thresh = ctx->ifc_sysctl_tx_reclaim_thresh;
7083 	err = sysctl_handle_int(oidp, &thresh, arg2, req);
7084 	if (err != 0) {
7085 		return err;
7086 	}
7087 
7088 	if (thresh == ctx->ifc_sysctl_tx_reclaim_thresh)
7089 		return 0;
7090 
7091 	if (thresh > ctx->ifc_softc_ctx.isc_ntxd[0] / 2) {
7092 		device_printf(ctx->ifc_dev, "TX Reclaim thresh must be <= %d\n",
7093 		    ctx->ifc_softc_ctx.isc_ntxd[0] / 2);
7094 		return (EINVAL);
7095 	}
7096 
7097 	ctx->ifc_sysctl_tx_reclaim_thresh = thresh;
7098 	if (ctx->ifc_txqs == NULL)
7099 		return (err);
7100 
7101 	txq = &ctx->ifc_txqs[0];
7102 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
7103 		txq->ift_reclaim_thresh = thresh;
7104 	}
7105 	return (err);
7106 }
7107 
7108 static int
7109 iflib_handle_tx_reclaim_ticks(SYSCTL_HANDLER_ARGS)
7110 {
7111 	if_ctx_t ctx = (void *)arg1;
7112 	iflib_txq_t txq;
7113 	int i, err;
7114 	int ticks;
7115 
7116 	ticks = ctx->ifc_sysctl_tx_reclaim_ticks;
7117 	err = sysctl_handle_int(oidp, &ticks, arg2, req);
7118 	if (err != 0) {
7119 		return err;
7120 	}
7121 
7122 	if (ticks == ctx->ifc_sysctl_tx_reclaim_ticks)
7123 		return 0;
7124 
7125 	if (ticks > hz) {
7126 		device_printf(ctx->ifc_dev,
7127 		    "TX Reclaim ticks must be <= hz (%d)\n", hz);
7128 		return (EINVAL);
7129 	}
7130 
7131 	ctx->ifc_sysctl_tx_reclaim_ticks = ticks;
7132 	if (ctx->ifc_txqs == NULL)
7133 		return (err);
7134 
7135 	txq = &ctx->ifc_txqs[0];
7136 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
7137 		txq->ift_reclaim_ticks = ticks;
7138 	}
7139 	return (err);
7140 }
7141 
7142 static int
7143 iflib_handle_tx_defer_mfree(SYSCTL_HANDLER_ARGS)
7144 {
7145 	if_ctx_t ctx = (void *)arg1;
7146 	iflib_txq_t txq;
7147 	int i, err;
7148 	int defer;
7149 
7150 	defer = ctx->ifc_sysctl_tx_defer_mfree;
7151 	err = sysctl_handle_int(oidp, &defer, arg2, req);
7152 	if (err != 0) {
7153 		return err;
7154 	}
7155 
7156 	if (defer == ctx->ifc_sysctl_tx_defer_mfree)
7157 		return 0;
7158 
7159 	ctx->ifc_sysctl_tx_defer_mfree = defer;
7160 	if (ctx->ifc_txqs == NULL)
7161 		return (err);
7162 
7163 	txq = &ctx->ifc_txqs[0];
7164 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
7165 		txq->ift_defer_mfree = defer;
7166 	}
7167 	return (err);
7168 }
7169 
7170 #define NAME_BUFLEN 32
7171 static void
7172 iflib_add_device_sysctl_pre(if_ctx_t ctx)
7173 {
7174 	device_t dev = iflib_get_dev(ctx);
7175 	struct sysctl_oid_list *child, *oid_list;
7176 	struct sysctl_oid *node;
7177 
7178 	sysctl_ctx_init(&ctx->ifc_sysctl_ctx);
7179 	child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev));
7180 	ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(&ctx->ifc_sysctl_ctx, child,
7181 	    OID_AUTO, "iflib", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL,
7182 	    "IFLIB fields");
7183 	oid_list = SYSCTL_CHILDREN(node);
7184 
7185 	SYSCTL_ADD_CONST_STRING(&ctx->ifc_sysctl_ctx, oid_list, OID_AUTO, "driver_version",
7186 	    CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version, "driver version");
7187 	SYSCTL_ADD_U32(&ctx->ifc_sysctl_ctx, oid_list, OID_AUTO,
7188 	    "tx_watchdog_events", CTLFLAG_RD, &ctx->ifc_tx_watchdog_events, 0,
7189 	    "TX watchdog resets initiated by iflib");
7190 
7191 	SYSCTL_ADD_BOOL(&ctx->ifc_sysctl_ctx, oid_list, OID_AUTO, "simple_tx",
7192 	    CTLFLAG_RDTUN, &ctx->ifc_sysctl_simple_tx, 0,
7193 	    "use simple tx ring");
7194 	SYSCTL_ADD_U16(&ctx->ifc_sysctl_ctx, oid_list, OID_AUTO, "override_ntxqs",
7195 	    CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0,
7196 	    "# of txqs to use, 0 => use default #");
7197 	SYSCTL_ADD_U16(&ctx->ifc_sysctl_ctx, oid_list, OID_AUTO, "override_nrxqs",
7198 	    CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0,
7199 	    "# of rxqs to use, 0 => use default #");
7200 	SYSCTL_ADD_U16(&ctx->ifc_sysctl_ctx, oid_list, OID_AUTO, "override_qs_enable",
7201 	    CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0,
7202 	    "permit #txq != #rxq");
7203 	SYSCTL_ADD_INT(&ctx->ifc_sysctl_ctx, oid_list, OID_AUTO, "disable_msix",
7204 	    CTLFLAG_RWTUN, &ctx->ifc_softc_ctx.isc_disable_msix, 0,
7205 	    "disable MSI-X (default 0)");
7206 	SYSCTL_ADD_U16(&ctx->ifc_sysctl_ctx, oid_list, OID_AUTO, "rx_budget",
7207 	    CTLFLAG_RWTUN, &ctx->ifc_sysctl_rx_budget, 0, "set the RX budget");
7208 	SYSCTL_ADD_U16(&ctx->ifc_sysctl_ctx, oid_list, OID_AUTO, "tx_abdicate",
7209 	    CTLFLAG_RWTUN, &ctx->ifc_sysctl_tx_abdicate, 0,
7210 	    "cause TX to abdicate instead of running to completion");
7211 	ctx->ifc_sysctl_core_offset = CORE_OFFSET_UNSPECIFIED;
7212 	SYSCTL_ADD_U16(&ctx->ifc_sysctl_ctx, oid_list, OID_AUTO, "core_offset",
7213 	    CTLFLAG_RDTUN, &ctx->ifc_sysctl_core_offset, 0,
7214 	    "offset to start using cores at");
7215 	SYSCTL_ADD_U8(&ctx->ifc_sysctl_ctx, oid_list, OID_AUTO, "separate_txrx",
7216 	    CTLFLAG_RDTUN, &ctx->ifc_sysctl_separate_txrx, 0,
7217 	    "use separate cores for TX and RX");
7218 	SYSCTL_ADD_U8(&ctx->ifc_sysctl_ctx, oid_list, OID_AUTO, "use_logical_cores",
7219 	    CTLFLAG_RDTUN, &ctx->ifc_sysctl_use_logical_cores, 0,
7220 	    "try to make use of logical cores for TX and RX");
7221 	SYSCTL_ADD_U16(&ctx->ifc_sysctl_ctx, oid_list, OID_AUTO, "use_extra_msix_vectors",
7222 	    CTLFLAG_RDTUN, &ctx->ifc_sysctl_extra_msix_vectors, 0,
7223 	    "attempt to reserve the given number of extra MSI-X vectors during driver load for the creation of additional interfaces later");
7224 	SYSCTL_ADD_INT(&ctx->ifc_sysctl_ctx, oid_list, OID_AUTO, "allocated_msix_vectors",
7225 	    CTLFLAG_RDTUN, &ctx->ifc_softc_ctx.isc_vectors, 0,
7226 	    "total # of MSI-X vectors allocated by driver");
7227 
7228 	/* XXX change for per-queue sizes */
7229 	SYSCTL_ADD_PROC(&ctx->ifc_sysctl_ctx, oid_list, OID_AUTO, "override_ntxds",
7230 	    CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NEEDGIANT, ctx,
7231 	    IFLIB_NTXD_HANDLER, mp_ndesc_handler, "A",
7232 	    "list of # of TX descriptors to use, 0 = use default #");
7233 	SYSCTL_ADD_PROC(&ctx->ifc_sysctl_ctx, oid_list, OID_AUTO, "override_nrxds",
7234 	    CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NEEDGIANT, ctx,
7235 	    IFLIB_NRXD_HANDLER, mp_ndesc_handler, "A",
7236 	    "list of # of RX descriptors to use, 0 = use default #");
7237 }
7238 
7239 static void
7240 iflib_add_device_sysctl_post(if_ctx_t ctx)
7241 {
7242 	if_shared_ctx_t sctx = ctx->ifc_sctx;
7243 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
7244 	struct sysctl_oid_list *child;
7245 	struct sysctl_ctx_list *ctx_list = &ctx->ifc_sysctl_ctx;
7246 	iflib_fl_t fl;
7247 	iflib_txq_t txq;
7248 	iflib_rxq_t rxq;
7249 	int i, j;
7250 	char namebuf[NAME_BUFLEN];
7251 	char *qfmt;
7252 	struct sysctl_oid *queue_node, *fl_node, *node;
7253 	struct sysctl_oid_list *queue_list, *fl_list;
7254 
7255 	node = ctx->ifc_sysctl_node;
7256 	child = SYSCTL_CHILDREN(node);
7257 
7258        SYSCTL_ADD_PROC(ctx_list, child, OID_AUTO, "tx_reclaim_thresh",
7259            CTLTYPE_INT | CTLFLAG_RWTUN, ctx,
7260            0, iflib_handle_tx_reclaim_thresh, "I",
7261            "Number of TX descs outstanding before reclaim is called");
7262 
7263        SYSCTL_ADD_PROC(ctx_list, child, OID_AUTO, "tx_reclaim_ticks",
7264            CTLTYPE_INT | CTLFLAG_RWTUN, ctx,
7265            0, iflib_handle_tx_reclaim_ticks, "I",
7266            "Number of ticks before a TX reclaim is forced");
7267 
7268        SYSCTL_ADD_PROC(ctx_list, child, OID_AUTO, "tx_defer_mfree",
7269            CTLTYPE_INT | CTLFLAG_RWTUN, ctx,
7270            0, iflib_handle_tx_defer_mfree, "I",
7271            "Free completed transmits outside of TX ring lock");
7272 
7273 	if (scctx->isc_ntxqsets > 100)
7274 		qfmt = "txq%03d";
7275 	else if (scctx->isc_ntxqsets > 10)
7276 		qfmt = "txq%02d";
7277 	else
7278 		qfmt = "txq%d";
7279 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
7280 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
7281 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
7282 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "Queue Name");
7283 		queue_list = SYSCTL_CHILDREN(queue_node);
7284 		SYSCTL_ADD_INT(ctx_list, queue_list, OID_AUTO, "cpu",
7285 		    CTLFLAG_RD, &txq->ift_task.gt_cpu, 0,
7286 		    "cpu this queue is bound to");
7287 #if MEMORY_LOGGING
7288 		SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued",
7289 		    CTLFLAG_RD, &txq->ift_dequeued, "total mbufs freed");
7290 		SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued",
7291 		    CTLFLAG_RD, &txq->ift_enqueued, "total mbufs enqueued");
7292 #endif
7293 		SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag",
7294 		    CTLFLAG_RD, &txq->ift_mbuf_defrag,
7295 		    "# of times m_defrag was called");
7296 		SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO, "m_pullups",
7297 		    CTLFLAG_RD, &txq->ift_pullups,
7298 		    "# of times m_pullup was called");
7299 		SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO,
7300 		    "mbuf_defrag_failed", CTLFLAG_RD,
7301 		    &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed");
7302 		SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO,
7303 		    "no_desc_avail", CTLFLAG_RD, &txq->ift_no_desc_avail,
7304 		    "# of times no descriptors were available");
7305 		SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO,
7306 		    "tx_map_failed", CTLFLAG_RD, &txq->ift_map_failed,
7307 		    "# of times DMA map failed");
7308 		SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO,
7309 		    "txd_encap_efbig", CTLFLAG_RD, &txq->ift_txd_encap_efbig,
7310 		    "# of times txd_encap returned EFBIG");
7311 		SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO,
7312 		    "no_tx_dma_setup", CTLFLAG_RD, &txq->ift_no_tx_dma_setup,
7313 		    "# of times map failed for other than EFBIG");
7314 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx",
7315 		    CTLFLAG_RD, &txq->ift_pidx, 1, "Producer Index");
7316 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx",
7317 		    CTLFLAG_RD, &txq->ift_cidx, 1, "Consumer Index");
7318 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO,
7319 		    "txq_cidx_processed", CTLFLAG_RD, &txq->ift_cidx_processed,
7320 		    1, "Consumer Index seen by credit update");
7321 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use",
7322 		    CTLFLAG_RD, &txq->ift_in_use, 1, "descriptors in use");
7323 		SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO,
7324 		    "txq_processed", CTLFLAG_RD, &txq->ift_processed,
7325 		    "descriptors procesed for clean");
7326 		SYSCTL_ADD_UQUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned",
7327 		    CTLFLAG_RD, &txq->ift_cleaned, "total cleaned");
7328 		SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state",
7329 		    CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_NEEDGIANT,
7330 		    __DEVOLATILE(uint64_t *, &txq->ift_br->state), 0,
7331 		    mp_ring_state_handler, "A", "soft ring state");
7332 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO,
7333 		    "r_enqueues", CTLFLAG_RD, &txq->ift_br->enqueues,
7334 		    "# of enqueues to the mp_ring for this queue");
7335 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO,
7336 		    "r_drops", CTLFLAG_RD, &txq->ift_br->drops,
7337 		    "# of drops in the mp_ring for this queue");
7338 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO,
7339 		    "r_starts", CTLFLAG_RD, &txq->ift_br->starts,
7340 		    "# of normal consumer starts in mp_ring for this queue");
7341 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO,
7342 		    "r_stalls", CTLFLAG_RD, &txq->ift_br->stalls,
7343 		    "# of consumer stalls in the mp_ring for this queue");
7344 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO,
7345 		    "r_restarts", CTLFLAG_RD, &txq->ift_br->restarts,
7346 		    "# of consumer restarts in the mp_ring for this queue");
7347 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO,
7348 		    "r_abdications", CTLFLAG_RD, &txq->ift_br->abdications,
7349 		    "# of consumer abdications in the mp_ring for this queue");
7350 	}
7351 
7352 	if (scctx->isc_nrxqsets > 100)
7353 		qfmt = "rxq%03d";
7354 	else if (scctx->isc_nrxqsets > 10)
7355 		qfmt = "rxq%02d";
7356 	else
7357 		qfmt = "rxq%d";
7358 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
7359 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
7360 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
7361 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "Queue Name");
7362 		queue_list = SYSCTL_CHILDREN(queue_node);
7363 		SYSCTL_ADD_INT(ctx_list, queue_list, OID_AUTO, "cpu",
7364 		    CTLFLAG_RD, &rxq->ifr_task.gt_cpu, 0,
7365 		    "cpu this queue is bound to");
7366 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
7367 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO,
7368 			    "rxq_cq_cidx", CTLFLAG_RD, &rxq->ifr_cq_cidx, 1,
7369 			    "Consumer Index");
7370 		}
7371 
7372 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
7373 			snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j);
7374 			fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list,
7375 			    OID_AUTO, namebuf, CTLFLAG_RD | CTLFLAG_MPSAFE,
7376 			    NULL, "freelist Name");
7377 			fl_list = SYSCTL_CHILDREN(fl_node);
7378 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx",
7379 			    CTLFLAG_RD, &fl->ifl_pidx, 1, "Producer Index");
7380 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx",
7381 			    CTLFLAG_RD, &fl->ifl_cidx, 1, "Consumer Index");
7382 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits",
7383 			    CTLFLAG_RD, &fl->ifl_credits, 1,
7384 			    "credits available");
7385 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "buf_size",
7386 			    CTLFLAG_RD, &fl->ifl_buf_size, 1, "buffer size");
7387 #if MEMORY_LOGGING
7388 			SYSCTL_ADD_UQUAD(ctx_list, fl_list, OID_AUTO,
7389 			    "fl_m_enqueued", CTLFLAG_RD, &fl->ifl_m_enqueued,
7390 			    "mbufs allocated");
7391 			SYSCTL_ADD_UQUAD(ctx_list, fl_list, OID_AUTO,
7392 			    "fl_m_dequeued", CTLFLAG_RD, &fl->ifl_m_dequeued,
7393 			    "mbufs freed");
7394 			SYSCTL_ADD_UQUAD(ctx_list, fl_list, OID_AUTO,
7395 			    "fl_cl_enqueued", CTLFLAG_RD, &fl->ifl_cl_enqueued,
7396 			    "clusters allocated");
7397 			SYSCTL_ADD_UQUAD(ctx_list, fl_list, OID_AUTO,
7398 			    "fl_cl_dequeued", CTLFLAG_RD, &fl->ifl_cl_dequeued,
7399 			    "clusters freed");
7400 #endif
7401 		}
7402 	}
7403 
7404 }
7405 
7406 void
7407 iflib_request_reset(if_ctx_t ctx)
7408 {
7409 
7410 	STATE_LOCK(ctx);
7411 	ctx->ifc_flags |= IFC_DO_RESET;
7412 	STATE_UNLOCK(ctx);
7413 }
7414 
7415 void
7416 iflib_request_reset_if_up(if_ctx_t ctx)
7417 {
7418 
7419 	STATE_LOCK(ctx);
7420 	ctx->ifc_flags |= IFC_DO_RESET_IF_UP;
7421 	STATE_UNLOCK(ctx);
7422 }
7423 
7424 void
7425 iflib_init_failed(if_ctx_t ctx)
7426 {
7427 
7428 	sx_assert(&ctx->ifc_ctx_sx, SA_XLOCKED);
7429 	STATE_LOCK(ctx);
7430 	ctx->ifc_flags |= IFC_INIT_FAILED;
7431 	STATE_UNLOCK(ctx);
7432 }
7433 
7434 #ifndef __NO_STRICT_ALIGNMENT
7435 static struct mbuf *
7436 iflib_fixup_rx(struct mbuf *m)
7437 {
7438 	struct mbuf *n;
7439 
7440 	if (m->m_len <= (MCLBYTES - ETHER_HDR_LEN)) {
7441 		bcopy(m->m_data, m->m_data + ETHER_HDR_LEN, m->m_len);
7442 		m->m_data += ETHER_HDR_LEN;
7443 		n = m;
7444 	} else {
7445 		MGETHDR(n, M_NOWAIT, MT_DATA);
7446 		if (n == NULL) {
7447 			m_freem(m);
7448 			return (NULL);
7449 		}
7450 		bcopy(m->m_data, n->m_data, ETHER_HDR_LEN);
7451 		m->m_data += ETHER_HDR_LEN;
7452 		m->m_len -= ETHER_HDR_LEN;
7453 		n->m_len = ETHER_HDR_LEN;
7454 		M_MOVE_PKTHDR(n, m);
7455 		n->m_next = m;
7456 	}
7457 	return (n);
7458 }
7459 #endif
7460 
7461 #ifdef DEBUGNET
7462 static void
7463 iflib_debugnet_init(if_t ifp, int *nrxr, int *ncl, int *clsize)
7464 {
7465 	if_ctx_t ctx;
7466 
7467 	ctx = if_getsoftc(ifp);
7468 	CTX_LOCK(ctx);
7469 	*nrxr = NRXQSETS(ctx);
7470 	*ncl = ctx->ifc_rxqs[0].ifr_fl->ifl_size;
7471 	*clsize = ctx->ifc_rxqs[0].ifr_fl->ifl_buf_size;
7472 	CTX_UNLOCK(ctx);
7473 }
7474 
7475 static void
7476 iflib_debugnet_event(if_t ifp, enum debugnet_ev event)
7477 {
7478 	if_ctx_t ctx;
7479 	if_softc_ctx_t scctx;
7480 	iflib_fl_t fl;
7481 	iflib_rxq_t rxq;
7482 	int i, j;
7483 
7484 	ctx = if_getsoftc(ifp);
7485 	scctx = &ctx->ifc_softc_ctx;
7486 
7487 	switch (event) {
7488 	case DEBUGNET_START:
7489 		for (i = 0; i < scctx->isc_nrxqsets; i++) {
7490 			rxq = &ctx->ifc_rxqs[i];
7491 			for (j = 0; j < rxq->ifr_nfl; j++) {
7492 				fl = rxq->ifr_fl;
7493 				fl->ifl_zone = m_getzone(fl->ifl_buf_size);
7494 			}
7495 		}
7496 		iflib_no_tx_batch = 1;
7497 		break;
7498 	default:
7499 		break;
7500 	}
7501 }
7502 
7503 static int
7504 iflib_debugnet_transmit(if_t ifp, struct mbuf *m)
7505 {
7506 	if_ctx_t ctx;
7507 	iflib_txq_t txq;
7508 	int error;
7509 	int bytes_sent = 0;
7510 	int pkt_sent = 0;
7511 
7512 	ctx = if_getsoftc(ifp);
7513 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
7514 	    IFF_DRV_RUNNING)
7515 		return (EBUSY);
7516 
7517 	txq = &ctx->ifc_txqs[0];
7518 	error = iflib_encap(txq, &m, &bytes_sent, &pkt_sent);
7519 	if (error == 0)
7520 		(void)iflib_txd_db_check(txq, true);
7521 	return (error);
7522 }
7523 
7524 static int
7525 iflib_debugnet_poll(if_t ifp, int count)
7526 {
7527 	struct epoch_tracker et;
7528 	if_ctx_t ctx;
7529 	if_softc_ctx_t scctx;
7530 	iflib_txq_t txq;
7531 	int i;
7532 
7533 	ctx = if_getsoftc(ifp);
7534 	scctx = &ctx->ifc_softc_ctx;
7535 
7536 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
7537 	    IFF_DRV_RUNNING)
7538 		return (EBUSY);
7539 
7540 	txq = &ctx->ifc_txqs[0];
7541 	(void)iflib_completed_tx_reclaim(txq, NULL);
7542 
7543 	NET_EPOCH_ENTER(et);
7544 	for (i = 0; i < scctx->isc_nrxqsets; i++)
7545 		(void)iflib_rxeof(&ctx->ifc_rxqs[i], 16 /* XXX */);
7546 	NET_EPOCH_EXIT(et);
7547 	return (0);
7548 }
7549 #endif /* DEBUGNET */
7550 
7551 #ifndef ALTQ
7552 static inline iflib_txq_t
7553 iflib_simple_select_queue(if_ctx_t ctx, struct mbuf *m)
7554 {
7555 	int qidx;
7556 
7557 	if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m))
7558 		qidx = QIDX(ctx, m);
7559 	else
7560 		qidx = NTXQSETS(ctx) + FIRST_QSET(ctx) - 1;
7561 	return (&ctx->ifc_txqs[qidx]);
7562 }
7563 
7564 static int
7565 iflib_simple_transmit(if_t ifp, struct mbuf *m)
7566 {
7567 	if_ctx_t ctx;
7568 	iflib_txq_t txq;
7569 	struct mbuf **m_defer;
7570 	int error, i, reclaimable;
7571 	int bytes_sent = 0, pkt_sent = 0, mcast_sent = 0;
7572 
7573 
7574 	ctx = if_getsoftc(ifp);
7575 	if (__predict_false((if_getdrvflags(ifp) & IFF_DRV_RUNNING) == 0
7576 		|| !LINK_ACTIVE(ctx))) {
7577 		DBG_COUNTER_INC(tx_frees);
7578 		m_freem(m);
7579 		return (ENETDOWN);
7580 	}
7581 
7582 	txq = iflib_simple_select_queue(ctx, m);
7583 	mtx_lock(&txq->ift_mtx);
7584 	error = iflib_encap(txq, &m, &bytes_sent, &pkt_sent);
7585 	if (error == 0) {
7586 		mcast_sent += !!(m->m_flags & M_MCAST);
7587 		(void)iflib_txd_db_check(txq, true);
7588 	} else {
7589 		if (error == ENOBUFS)
7590 			if_inc_counter(ifp, IFCOUNTER_OQDROPS, 1);
7591 		else
7592 			if_inc_counter(ifp, IFCOUNTER_OERRORS, 1);
7593 	}
7594 	m_defer = NULL;
7595 	reclaimable = iflib_txq_can_reclaim(txq);
7596 	if (reclaimable != 0) {
7597 		/*
7598 		 * Try to set m_defer to the deferred mbuf reclaim array.  If
7599 		 * we can, the frees will happen outside the tx lock.  If we
7600 		 * can't, it means another thread is still proccessing frees.
7601 		 */
7602 		if (txq->ift_defer_mfree &&
7603 		    atomic_cmpset_acq_ptr((uintptr_t *)&txq->ift_sds.ifsd_m_defer,
7604 			(uintptr_t )txq->ift_sds.ifsd_m_deferb, 0)) {
7605 			m_defer = txq->ift_sds.ifsd_m_deferb;
7606 		}
7607 		_iflib_completed_tx_reclaim(txq, m_defer, reclaimable);
7608 	}
7609 	mtx_unlock(&txq->ift_mtx);
7610 
7611 	/*
7612 	 * Process mbuf frees outside the tx lock
7613 	 */
7614 	if (m_defer != NULL) {
7615 		for (i = 0; m_defer[i] != NULL; i++) {
7616 			m_freem(m_defer[i]);
7617 			m_defer[i] = NULL;
7618 		}
7619 		atomic_store_rel_ptr((uintptr_t *)&txq->ift_sds.ifsd_m_defer,
7620 		    (uintptr_t)m_defer);
7621 	}
7622 	if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
7623 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
7624 	if (mcast_sent)
7625 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
7626 
7627 	return (error);
7628 }
7629 #endif
7630