xref: /freebsd/sys/net/iflib.c (revision 8ddb146abcdf061be9f2c0db7e391697dafad85c)
1 /*-
2  * Copyright (c) 2014-2018, Matthew Macy <mmacy@mattmacy.io>
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions are met:
7  *
8  *  1. Redistributions of source code must retain the above copyright notice,
9  *     this list of conditions and the following disclaimer.
10  *
11  *  2. Neither the name of Matthew Macy nor the names of its
12  *     contributors may be used to endorse or promote products derived from
13  *     this software without specific prior written permission.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
16  * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18  * ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
19  * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
20  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
21  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
22  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
23  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
24  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
25  * POSSIBILITY OF SUCH DAMAGE.
26  */
27 
28 #include <sys/cdefs.h>
29 __FBSDID("$FreeBSD$");
30 
31 #include "opt_inet.h"
32 #include "opt_inet6.h"
33 #include "opt_acpi.h"
34 #include "opt_sched.h"
35 
36 #include <sys/param.h>
37 #include <sys/types.h>
38 #include <sys/bus.h>
39 #include <sys/eventhandler.h>
40 #include <sys/kernel.h>
41 #include <sys/lock.h>
42 #include <sys/mutex.h>
43 #include <sys/module.h>
44 #include <sys/kobj.h>
45 #include <sys/rman.h>
46 #include <sys/sbuf.h>
47 #include <sys/smp.h>
48 #include <sys/socket.h>
49 #include <sys/sockio.h>
50 #include <sys/sysctl.h>
51 #include <sys/syslog.h>
52 #include <sys/taskqueue.h>
53 #include <sys/limits.h>
54 
55 #include <net/if.h>
56 #include <net/if_var.h>
57 #include <net/if_types.h>
58 #include <net/if_media.h>
59 #include <net/bpf.h>
60 #include <net/ethernet.h>
61 #include <net/mp_ring.h>
62 #include <net/debugnet.h>
63 #include <net/pfil.h>
64 #include <net/vnet.h>
65 
66 #include <netinet/in.h>
67 #include <netinet/in_pcb.h>
68 #include <netinet/tcp_lro.h>
69 #include <netinet/in_systm.h>
70 #include <netinet/if_ether.h>
71 #include <netinet/ip.h>
72 #include <netinet/ip6.h>
73 #include <netinet/tcp.h>
74 #include <netinet/ip_var.h>
75 #include <netinet6/ip6_var.h>
76 
77 #include <machine/bus.h>
78 #include <machine/in_cksum.h>
79 
80 #include <vm/vm.h>
81 #include <vm/pmap.h>
82 
83 #include <dev/led/led.h>
84 #include <dev/pci/pcireg.h>
85 #include <dev/pci/pcivar.h>
86 #include <dev/pci/pci_private.h>
87 
88 #include <net/iflib.h>
89 #include <net/iflib_private.h>
90 
91 #include "ifdi_if.h"
92 
93 #ifdef PCI_IOV
94 #include <dev/pci/pci_iov.h>
95 #endif
96 
97 #include <sys/bitstring.h>
98 /*
99  * enable accounting of every mbuf as it comes in to and goes out of
100  * iflib's software descriptor references
101  */
102 #define MEMORY_LOGGING 0
103 /*
104  * Enable mbuf vectors for compressing long mbuf chains
105  */
106 
107 /*
108  * NB:
109  * - Prefetching in tx cleaning should perhaps be a tunable. The distance ahead
110  *   we prefetch needs to be determined by the time spent in m_free vis a vis
111  *   the cost of a prefetch. This will of course vary based on the workload:
112  *      - NFLX's m_free path is dominated by vm-based M_EXT manipulation which
113  *        is quite expensive, thus suggesting very little prefetch.
114  *      - small packet forwarding which is just returning a single mbuf to
115  *        UMA will typically be very fast vis a vis the cost of a memory
116  *        access.
117  */
118 
119 /*
120  * File organization:
121  *  - private structures
122  *  - iflib private utility functions
123  *  - ifnet functions
124  *  - vlan registry and other exported functions
125  *  - iflib public core functions
126  *
127  *
128  */
129 MALLOC_DEFINE(M_IFLIB, "iflib", "ifnet library");
130 
131 #define	IFLIB_RXEOF_MORE (1U << 0)
132 #define	IFLIB_RXEOF_EMPTY (2U << 0)
133 
134 struct iflib_txq;
135 typedef struct iflib_txq *iflib_txq_t;
136 struct iflib_rxq;
137 typedef struct iflib_rxq *iflib_rxq_t;
138 struct iflib_fl;
139 typedef struct iflib_fl *iflib_fl_t;
140 
141 struct iflib_ctx;
142 
143 static void iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid);
144 static void iflib_timer(void *arg);
145 static void iflib_tqg_detach(if_ctx_t ctx);
146 
147 typedef struct iflib_filter_info {
148 	driver_filter_t *ifi_filter;
149 	void *ifi_filter_arg;
150 	struct grouptask *ifi_task;
151 	void *ifi_ctx;
152 } *iflib_filter_info_t;
153 
154 struct iflib_ctx {
155 	KOBJ_FIELDS;
156 	/*
157 	 * Pointer to hardware driver's softc
158 	 */
159 	void *ifc_softc;
160 	device_t ifc_dev;
161 	if_t ifc_ifp;
162 
163 	cpuset_t ifc_cpus;
164 	if_shared_ctx_t ifc_sctx;
165 	struct if_softc_ctx ifc_softc_ctx;
166 
167 	struct sx ifc_ctx_sx;
168 	struct mtx ifc_state_mtx;
169 
170 	iflib_txq_t ifc_txqs;
171 	iflib_rxq_t ifc_rxqs;
172 	uint32_t ifc_if_flags;
173 	uint32_t ifc_flags;
174 	uint32_t ifc_max_fl_buf_size;
175 	uint32_t ifc_rx_mbuf_sz;
176 
177 	int ifc_link_state;
178 	int ifc_watchdog_events;
179 	struct cdev *ifc_led_dev;
180 	struct resource *ifc_msix_mem;
181 
182 	struct if_irq ifc_legacy_irq;
183 	struct grouptask ifc_admin_task;
184 	struct grouptask ifc_vflr_task;
185 	struct iflib_filter_info ifc_filter_info;
186 	struct ifmedia	ifc_media;
187 	struct ifmedia	*ifc_mediap;
188 
189 	struct sysctl_oid *ifc_sysctl_node;
190 	uint16_t ifc_sysctl_ntxqs;
191 	uint16_t ifc_sysctl_nrxqs;
192 	uint16_t ifc_sysctl_qs_eq_override;
193 	uint16_t ifc_sysctl_rx_budget;
194 	uint16_t ifc_sysctl_tx_abdicate;
195 	uint16_t ifc_sysctl_core_offset;
196 #define	CORE_OFFSET_UNSPECIFIED	0xffff
197 	uint8_t  ifc_sysctl_separate_txrx;
198 	uint8_t  ifc_sysctl_use_logical_cores;
199 	bool	 ifc_cpus_are_physical_cores;
200 
201 	qidx_t ifc_sysctl_ntxds[8];
202 	qidx_t ifc_sysctl_nrxds[8];
203 	struct if_txrx ifc_txrx;
204 #define isc_txd_encap  ifc_txrx.ift_txd_encap
205 #define isc_txd_flush  ifc_txrx.ift_txd_flush
206 #define isc_txd_credits_update  ifc_txrx.ift_txd_credits_update
207 #define isc_rxd_available ifc_txrx.ift_rxd_available
208 #define isc_rxd_pkt_get ifc_txrx.ift_rxd_pkt_get
209 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
210 #define isc_rxd_flush ifc_txrx.ift_rxd_flush
211 #define isc_legacy_intr ifc_txrx.ift_legacy_intr
212 #define isc_txq_select ifc_txrx.ift_txq_select
213 	eventhandler_tag ifc_vlan_attach_event;
214 	eventhandler_tag ifc_vlan_detach_event;
215 	struct ether_addr ifc_mac;
216 };
217 
218 void *
219 iflib_get_softc(if_ctx_t ctx)
220 {
221 
222 	return (ctx->ifc_softc);
223 }
224 
225 device_t
226 iflib_get_dev(if_ctx_t ctx)
227 {
228 
229 	return (ctx->ifc_dev);
230 }
231 
232 if_t
233 iflib_get_ifp(if_ctx_t ctx)
234 {
235 
236 	return (ctx->ifc_ifp);
237 }
238 
239 struct ifmedia *
240 iflib_get_media(if_ctx_t ctx)
241 {
242 
243 	return (ctx->ifc_mediap);
244 }
245 
246 uint32_t
247 iflib_get_flags(if_ctx_t ctx)
248 {
249 	return (ctx->ifc_flags);
250 }
251 
252 void
253 iflib_set_mac(if_ctx_t ctx, uint8_t mac[ETHER_ADDR_LEN])
254 {
255 
256 	bcopy(mac, ctx->ifc_mac.octet, ETHER_ADDR_LEN);
257 }
258 
259 if_softc_ctx_t
260 iflib_get_softc_ctx(if_ctx_t ctx)
261 {
262 
263 	return (&ctx->ifc_softc_ctx);
264 }
265 
266 if_shared_ctx_t
267 iflib_get_sctx(if_ctx_t ctx)
268 {
269 
270 	return (ctx->ifc_sctx);
271 }
272 
273 #define IP_ALIGNED(m) ((((uintptr_t)(m)->m_data) & 0x3) == 0x2)
274 #define CACHE_PTR_INCREMENT (CACHE_LINE_SIZE/sizeof(void*))
275 #define CACHE_PTR_NEXT(ptr) ((void *)(((uintptr_t)(ptr)+CACHE_LINE_SIZE-1) & (CACHE_LINE_SIZE-1)))
276 
277 #define LINK_ACTIVE(ctx) ((ctx)->ifc_link_state == LINK_STATE_UP)
278 #define CTX_IS_VF(ctx) ((ctx)->ifc_sctx->isc_flags & IFLIB_IS_VF)
279 
280 typedef struct iflib_sw_rx_desc_array {
281 	bus_dmamap_t	*ifsd_map;         /* bus_dma maps for packet */
282 	struct mbuf	**ifsd_m;           /* pkthdr mbufs */
283 	caddr_t		*ifsd_cl;          /* direct cluster pointer for rx */
284 	bus_addr_t	*ifsd_ba;          /* bus addr of cluster for rx */
285 } iflib_rxsd_array_t;
286 
287 typedef struct iflib_sw_tx_desc_array {
288 	bus_dmamap_t    *ifsd_map;         /* bus_dma maps for packet */
289 	bus_dmamap_t	*ifsd_tso_map;     /* bus_dma maps for TSO packet */
290 	struct mbuf    **ifsd_m;           /* pkthdr mbufs */
291 } if_txsd_vec_t;
292 
293 /* magic number that should be high enough for any hardware */
294 #define IFLIB_MAX_TX_SEGS		128
295 #define IFLIB_RX_COPY_THRESH		128
296 #define IFLIB_MAX_RX_REFRESH		32
297 /* The minimum descriptors per second before we start coalescing */
298 #define IFLIB_MIN_DESC_SEC		16384
299 #define IFLIB_DEFAULT_TX_UPDATE_FREQ	16
300 #define IFLIB_QUEUE_IDLE		0
301 #define IFLIB_QUEUE_HUNG		1
302 #define IFLIB_QUEUE_WORKING		2
303 /* maximum number of txqs that can share an rx interrupt */
304 #define IFLIB_MAX_TX_SHARED_INTR	4
305 
306 /* this should really scale with ring size - this is a fairly arbitrary value */
307 #define TX_BATCH_SIZE			32
308 
309 #define IFLIB_RESTART_BUDGET		8
310 
311 #define CSUM_OFFLOAD		(CSUM_IP_TSO|CSUM_IP6_TSO|CSUM_IP| \
312 				 CSUM_IP_UDP|CSUM_IP_TCP|CSUM_IP_SCTP| \
313 				 CSUM_IP6_UDP|CSUM_IP6_TCP|CSUM_IP6_SCTP)
314 
315 struct iflib_txq {
316 	qidx_t		ift_in_use;
317 	qidx_t		ift_cidx;
318 	qidx_t		ift_cidx_processed;
319 	qidx_t		ift_pidx;
320 	uint8_t		ift_gen;
321 	uint8_t		ift_br_offset;
322 	uint16_t	ift_npending;
323 	uint16_t	ift_db_pending;
324 	uint16_t	ift_rs_pending;
325 	/* implicit pad */
326 	uint8_t		ift_txd_size[8];
327 	uint64_t	ift_processed;
328 	uint64_t	ift_cleaned;
329 	uint64_t	ift_cleaned_prev;
330 #if MEMORY_LOGGING
331 	uint64_t	ift_enqueued;
332 	uint64_t	ift_dequeued;
333 #endif
334 	uint64_t	ift_no_tx_dma_setup;
335 	uint64_t	ift_no_desc_avail;
336 	uint64_t	ift_mbuf_defrag_failed;
337 	uint64_t	ift_mbuf_defrag;
338 	uint64_t	ift_map_failed;
339 	uint64_t	ift_txd_encap_efbig;
340 	uint64_t	ift_pullups;
341 	uint64_t	ift_last_timer_tick;
342 
343 	struct mtx	ift_mtx;
344 	struct mtx	ift_db_mtx;
345 
346 	/* constant values */
347 	if_ctx_t	ift_ctx;
348 	struct ifmp_ring        *ift_br;
349 	struct grouptask	ift_task;
350 	qidx_t		ift_size;
351 	uint16_t	ift_id;
352 	struct callout	ift_timer;
353 #ifdef DEV_NETMAP
354 	struct callout	ift_netmap_timer;
355 #endif /* DEV_NETMAP */
356 
357 	if_txsd_vec_t	ift_sds;
358 	uint8_t		ift_qstatus;
359 	uint8_t		ift_closed;
360 	uint8_t		ift_update_freq;
361 	struct iflib_filter_info ift_filter_info;
362 	bus_dma_tag_t	ift_buf_tag;
363 	bus_dma_tag_t	ift_tso_buf_tag;
364 	iflib_dma_info_t	ift_ifdi;
365 #define	MTX_NAME_LEN	32
366 	char                    ift_mtx_name[MTX_NAME_LEN];
367 	bus_dma_segment_t	ift_segs[IFLIB_MAX_TX_SEGS]  __aligned(CACHE_LINE_SIZE);
368 #ifdef IFLIB_DIAGNOSTICS
369 	uint64_t ift_cpu_exec_count[256];
370 #endif
371 } __aligned(CACHE_LINE_SIZE);
372 
373 struct iflib_fl {
374 	qidx_t		ifl_cidx;
375 	qidx_t		ifl_pidx;
376 	qidx_t		ifl_credits;
377 	uint8_t		ifl_gen;
378 	uint8_t		ifl_rxd_size;
379 #if MEMORY_LOGGING
380 	uint64_t	ifl_m_enqueued;
381 	uint64_t	ifl_m_dequeued;
382 	uint64_t	ifl_cl_enqueued;
383 	uint64_t	ifl_cl_dequeued;
384 #endif
385 	/* implicit pad */
386 	bitstr_t 	*ifl_rx_bitmap;
387 	qidx_t		ifl_fragidx;
388 	/* constant */
389 	qidx_t		ifl_size;
390 	uint16_t	ifl_buf_size;
391 	uint16_t	ifl_cltype;
392 	uma_zone_t	ifl_zone;
393 	iflib_rxsd_array_t	ifl_sds;
394 	iflib_rxq_t	ifl_rxq;
395 	uint8_t		ifl_id;
396 	bus_dma_tag_t	ifl_buf_tag;
397 	iflib_dma_info_t	ifl_ifdi;
398 	uint64_t	ifl_bus_addrs[IFLIB_MAX_RX_REFRESH] __aligned(CACHE_LINE_SIZE);
399 	qidx_t		ifl_rxd_idxs[IFLIB_MAX_RX_REFRESH];
400 }  __aligned(CACHE_LINE_SIZE);
401 
402 static inline qidx_t
403 get_inuse(int size, qidx_t cidx, qidx_t pidx, uint8_t gen)
404 {
405 	qidx_t used;
406 
407 	if (pidx > cidx)
408 		used = pidx - cidx;
409 	else if (pidx < cidx)
410 		used = size - cidx + pidx;
411 	else if (gen == 0 && pidx == cidx)
412 		used = 0;
413 	else if (gen == 1 && pidx == cidx)
414 		used = size;
415 	else
416 		panic("bad state");
417 
418 	return (used);
419 }
420 
421 #define TXQ_AVAIL(txq) (txq->ift_size - get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx, txq->ift_gen))
422 
423 #define IDXDIFF(head, tail, wrap) \
424 	((head) >= (tail) ? (head) - (tail) : (wrap) - (tail) + (head))
425 
426 struct iflib_rxq {
427 	if_ctx_t	ifr_ctx;
428 	iflib_fl_t	ifr_fl;
429 	uint64_t	ifr_rx_irq;
430 	struct pfil_head	*pfil;
431 	/*
432 	 * If there is a separate completion queue (IFLIB_HAS_RXCQ), this is
433 	 * the completion queue consumer index.  Otherwise it's unused.
434 	 */
435 	qidx_t		ifr_cq_cidx;
436 	uint16_t	ifr_id;
437 	uint8_t		ifr_nfl;
438 	uint8_t		ifr_ntxqirq;
439 	uint8_t		ifr_txqid[IFLIB_MAX_TX_SHARED_INTR];
440 	uint8_t		ifr_fl_offset;
441 	struct lro_ctrl			ifr_lc;
442 	struct grouptask        ifr_task;
443 	struct callout		ifr_watchdog;
444 	struct iflib_filter_info ifr_filter_info;
445 	iflib_dma_info_t		ifr_ifdi;
446 
447 	/* dynamically allocate if any drivers need a value substantially larger than this */
448 	struct if_rxd_frag	ifr_frags[IFLIB_MAX_RX_SEGS] __aligned(CACHE_LINE_SIZE);
449 #ifdef IFLIB_DIAGNOSTICS
450 	uint64_t ifr_cpu_exec_count[256];
451 #endif
452 }  __aligned(CACHE_LINE_SIZE);
453 
454 typedef struct if_rxsd {
455 	caddr_t *ifsd_cl;
456 	iflib_fl_t ifsd_fl;
457 } *if_rxsd_t;
458 
459 /* multiple of word size */
460 #ifdef __LP64__
461 #define PKT_INFO_SIZE	6
462 #define RXD_INFO_SIZE	5
463 #define PKT_TYPE uint64_t
464 #else
465 #define PKT_INFO_SIZE	11
466 #define RXD_INFO_SIZE	8
467 #define PKT_TYPE uint32_t
468 #endif
469 #define PKT_LOOP_BOUND  ((PKT_INFO_SIZE/3)*3)
470 #define RXD_LOOP_BOUND  ((RXD_INFO_SIZE/4)*4)
471 
472 typedef struct if_pkt_info_pad {
473 	PKT_TYPE pkt_val[PKT_INFO_SIZE];
474 } *if_pkt_info_pad_t;
475 typedef struct if_rxd_info_pad {
476 	PKT_TYPE rxd_val[RXD_INFO_SIZE];
477 } *if_rxd_info_pad_t;
478 
479 CTASSERT(sizeof(struct if_pkt_info_pad) == sizeof(struct if_pkt_info));
480 CTASSERT(sizeof(struct if_rxd_info_pad) == sizeof(struct if_rxd_info));
481 
482 static inline void
483 pkt_info_zero(if_pkt_info_t pi)
484 {
485 	if_pkt_info_pad_t pi_pad;
486 
487 	pi_pad = (if_pkt_info_pad_t)pi;
488 	pi_pad->pkt_val[0] = 0; pi_pad->pkt_val[1] = 0; pi_pad->pkt_val[2] = 0;
489 	pi_pad->pkt_val[3] = 0; pi_pad->pkt_val[4] = 0; pi_pad->pkt_val[5] = 0;
490 #ifndef __LP64__
491 	pi_pad->pkt_val[6] = 0; pi_pad->pkt_val[7] = 0; pi_pad->pkt_val[8] = 0;
492 	pi_pad->pkt_val[9] = 0; pi_pad->pkt_val[10] = 0;
493 #endif
494 }
495 
496 static device_method_t iflib_pseudo_methods[] = {
497 	DEVMETHOD(device_attach, noop_attach),
498 	DEVMETHOD(device_detach, iflib_pseudo_detach),
499 	DEVMETHOD_END
500 };
501 
502 driver_t iflib_pseudodriver = {
503 	"iflib_pseudo", iflib_pseudo_methods, sizeof(struct iflib_ctx),
504 };
505 
506 static inline void
507 rxd_info_zero(if_rxd_info_t ri)
508 {
509 	if_rxd_info_pad_t ri_pad;
510 	int i;
511 
512 	ri_pad = (if_rxd_info_pad_t)ri;
513 	for (i = 0; i < RXD_LOOP_BOUND; i += 4) {
514 		ri_pad->rxd_val[i] = 0;
515 		ri_pad->rxd_val[i+1] = 0;
516 		ri_pad->rxd_val[i+2] = 0;
517 		ri_pad->rxd_val[i+3] = 0;
518 	}
519 #ifdef __LP64__
520 	ri_pad->rxd_val[RXD_INFO_SIZE-1] = 0;
521 #endif
522 }
523 
524 /*
525  * Only allow a single packet to take up most 1/nth of the tx ring
526  */
527 #define MAX_SINGLE_PACKET_FRACTION 12
528 #define IF_BAD_DMA (bus_addr_t)-1
529 
530 #define CTX_ACTIVE(ctx) ((if_getdrvflags((ctx)->ifc_ifp) & IFF_DRV_RUNNING))
531 
532 #define CTX_LOCK_INIT(_sc)  sx_init(&(_sc)->ifc_ctx_sx, "iflib ctx lock")
533 #define CTX_LOCK(ctx) sx_xlock(&(ctx)->ifc_ctx_sx)
534 #define CTX_UNLOCK(ctx) sx_xunlock(&(ctx)->ifc_ctx_sx)
535 #define CTX_LOCK_DESTROY(ctx) sx_destroy(&(ctx)->ifc_ctx_sx)
536 
537 #define STATE_LOCK_INIT(_sc, _name)  mtx_init(&(_sc)->ifc_state_mtx, _name, "iflib state lock", MTX_DEF)
538 #define STATE_LOCK(ctx) mtx_lock(&(ctx)->ifc_state_mtx)
539 #define STATE_UNLOCK(ctx) mtx_unlock(&(ctx)->ifc_state_mtx)
540 #define STATE_LOCK_DESTROY(ctx) mtx_destroy(&(ctx)->ifc_state_mtx)
541 
542 #define CALLOUT_LOCK(txq)	mtx_lock(&txq->ift_mtx)
543 #define CALLOUT_UNLOCK(txq) 	mtx_unlock(&txq->ift_mtx)
544 
545 void
546 iflib_set_detach(if_ctx_t ctx)
547 {
548 	STATE_LOCK(ctx);
549 	ctx->ifc_flags |= IFC_IN_DETACH;
550 	STATE_UNLOCK(ctx);
551 }
552 
553 /* Our boot-time initialization hook */
554 static int	iflib_module_event_handler(module_t, int, void *);
555 
556 static moduledata_t iflib_moduledata = {
557 	"iflib",
558 	iflib_module_event_handler,
559 	NULL
560 };
561 
562 DECLARE_MODULE(iflib, iflib_moduledata, SI_SUB_INIT_IF, SI_ORDER_ANY);
563 MODULE_VERSION(iflib, 1);
564 
565 MODULE_DEPEND(iflib, pci, 1, 1, 1);
566 MODULE_DEPEND(iflib, ether, 1, 1, 1);
567 
568 TASKQGROUP_DEFINE(if_io_tqg, mp_ncpus, 1);
569 TASKQGROUP_DEFINE(if_config_tqg, 1, 1);
570 
571 #ifndef IFLIB_DEBUG_COUNTERS
572 #ifdef INVARIANTS
573 #define IFLIB_DEBUG_COUNTERS 1
574 #else
575 #define IFLIB_DEBUG_COUNTERS 0
576 #endif /* !INVARIANTS */
577 #endif
578 
579 static SYSCTL_NODE(_net, OID_AUTO, iflib, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
580     "iflib driver parameters");
581 
582 /*
583  * XXX need to ensure that this can't accidentally cause the head to be moved backwards
584  */
585 static int iflib_min_tx_latency = 0;
586 SYSCTL_INT(_net_iflib, OID_AUTO, min_tx_latency, CTLFLAG_RW,
587 		   &iflib_min_tx_latency, 0, "minimize transmit latency at the possible expense of throughput");
588 static int iflib_no_tx_batch = 0;
589 SYSCTL_INT(_net_iflib, OID_AUTO, no_tx_batch, CTLFLAG_RW,
590 		   &iflib_no_tx_batch, 0, "minimize transmit latency at the possible expense of throughput");
591 static int iflib_timer_default = 1000;
592 SYSCTL_INT(_net_iflib, OID_AUTO, timer_default, CTLFLAG_RW,
593 		   &iflib_timer_default, 0, "number of ticks between iflib_timer calls");
594 
595 
596 #if IFLIB_DEBUG_COUNTERS
597 
598 static int iflib_tx_seen;
599 static int iflib_tx_sent;
600 static int iflib_tx_encap;
601 static int iflib_rx_allocs;
602 static int iflib_fl_refills;
603 static int iflib_fl_refills_large;
604 static int iflib_tx_frees;
605 
606 SYSCTL_INT(_net_iflib, OID_AUTO, tx_seen, CTLFLAG_RD,
607 		   &iflib_tx_seen, 0, "# TX mbufs seen");
608 SYSCTL_INT(_net_iflib, OID_AUTO, tx_sent, CTLFLAG_RD,
609 		   &iflib_tx_sent, 0, "# TX mbufs sent");
610 SYSCTL_INT(_net_iflib, OID_AUTO, tx_encap, CTLFLAG_RD,
611 		   &iflib_tx_encap, 0, "# TX mbufs encapped");
612 SYSCTL_INT(_net_iflib, OID_AUTO, tx_frees, CTLFLAG_RD,
613 		   &iflib_tx_frees, 0, "# TX frees");
614 SYSCTL_INT(_net_iflib, OID_AUTO, rx_allocs, CTLFLAG_RD,
615 		   &iflib_rx_allocs, 0, "# RX allocations");
616 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills, CTLFLAG_RD,
617 		   &iflib_fl_refills, 0, "# refills");
618 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills_large, CTLFLAG_RD,
619 		   &iflib_fl_refills_large, 0, "# large refills");
620 
621 static int iflib_txq_drain_flushing;
622 static int iflib_txq_drain_oactive;
623 static int iflib_txq_drain_notready;
624 
625 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_flushing, CTLFLAG_RD,
626 		   &iflib_txq_drain_flushing, 0, "# drain flushes");
627 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_oactive, CTLFLAG_RD,
628 		   &iflib_txq_drain_oactive, 0, "# drain oactives");
629 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_notready, CTLFLAG_RD,
630 		   &iflib_txq_drain_notready, 0, "# drain notready");
631 
632 static int iflib_encap_load_mbuf_fail;
633 static int iflib_encap_pad_mbuf_fail;
634 static int iflib_encap_txq_avail_fail;
635 static int iflib_encap_txd_encap_fail;
636 
637 SYSCTL_INT(_net_iflib, OID_AUTO, encap_load_mbuf_fail, CTLFLAG_RD,
638 		   &iflib_encap_load_mbuf_fail, 0, "# busdma load failures");
639 SYSCTL_INT(_net_iflib, OID_AUTO, encap_pad_mbuf_fail, CTLFLAG_RD,
640 		   &iflib_encap_pad_mbuf_fail, 0, "# runt frame pad failures");
641 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txq_avail_fail, CTLFLAG_RD,
642 		   &iflib_encap_txq_avail_fail, 0, "# txq avail failures");
643 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txd_encap_fail, CTLFLAG_RD,
644 		   &iflib_encap_txd_encap_fail, 0, "# driver encap failures");
645 
646 static int iflib_task_fn_rxs;
647 static int iflib_rx_intr_enables;
648 static int iflib_fast_intrs;
649 static int iflib_rx_unavail;
650 static int iflib_rx_ctx_inactive;
651 static int iflib_rx_if_input;
652 static int iflib_rxd_flush;
653 
654 static int iflib_verbose_debug;
655 
656 SYSCTL_INT(_net_iflib, OID_AUTO, task_fn_rx, CTLFLAG_RD,
657 		   &iflib_task_fn_rxs, 0, "# task_fn_rx calls");
658 SYSCTL_INT(_net_iflib, OID_AUTO, rx_intr_enables, CTLFLAG_RD,
659 		   &iflib_rx_intr_enables, 0, "# RX intr enables");
660 SYSCTL_INT(_net_iflib, OID_AUTO, fast_intrs, CTLFLAG_RD,
661 		   &iflib_fast_intrs, 0, "# fast_intr calls");
662 SYSCTL_INT(_net_iflib, OID_AUTO, rx_unavail, CTLFLAG_RD,
663 		   &iflib_rx_unavail, 0, "# times rxeof called with no available data");
664 SYSCTL_INT(_net_iflib, OID_AUTO, rx_ctx_inactive, CTLFLAG_RD,
665 		   &iflib_rx_ctx_inactive, 0, "# times rxeof called with inactive context");
666 SYSCTL_INT(_net_iflib, OID_AUTO, rx_if_input, CTLFLAG_RD,
667 		   &iflib_rx_if_input, 0, "# times rxeof called if_input");
668 SYSCTL_INT(_net_iflib, OID_AUTO, rxd_flush, CTLFLAG_RD,
669 	         &iflib_rxd_flush, 0, "# times rxd_flush called");
670 SYSCTL_INT(_net_iflib, OID_AUTO, verbose_debug, CTLFLAG_RW,
671 		   &iflib_verbose_debug, 0, "enable verbose debugging");
672 
673 #define DBG_COUNTER_INC(name) atomic_add_int(&(iflib_ ## name), 1)
674 static void
675 iflib_debug_reset(void)
676 {
677 	iflib_tx_seen = iflib_tx_sent = iflib_tx_encap = iflib_rx_allocs =
678 		iflib_fl_refills = iflib_fl_refills_large = iflib_tx_frees =
679 		iflib_txq_drain_flushing = iflib_txq_drain_oactive =
680 		iflib_txq_drain_notready =
681 		iflib_encap_load_mbuf_fail = iflib_encap_pad_mbuf_fail =
682 		iflib_encap_txq_avail_fail = iflib_encap_txd_encap_fail =
683 		iflib_task_fn_rxs = iflib_rx_intr_enables = iflib_fast_intrs =
684 		iflib_rx_unavail =
685 		iflib_rx_ctx_inactive = iflib_rx_if_input =
686 		iflib_rxd_flush = 0;
687 }
688 
689 #else
690 #define DBG_COUNTER_INC(name)
691 static void iflib_debug_reset(void) {}
692 #endif
693 
694 #define IFLIB_DEBUG 0
695 
696 static void iflib_tx_structures_free(if_ctx_t ctx);
697 static void iflib_rx_structures_free(if_ctx_t ctx);
698 static int iflib_queues_alloc(if_ctx_t ctx);
699 static int iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq);
700 static int iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget);
701 static int iflib_qset_structures_setup(if_ctx_t ctx);
702 static int iflib_msix_init(if_ctx_t ctx);
703 static int iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filterarg, int *rid, const char *str);
704 static void iflib_txq_check_drain(iflib_txq_t txq, int budget);
705 static uint32_t iflib_txq_can_drain(struct ifmp_ring *);
706 #ifdef ALTQ
707 static void iflib_altq_if_start(if_t ifp);
708 static int iflib_altq_if_transmit(if_t ifp, struct mbuf *m);
709 #endif
710 static int iflib_register(if_ctx_t);
711 static void iflib_deregister(if_ctx_t);
712 static void iflib_unregister_vlan_handlers(if_ctx_t ctx);
713 static uint16_t iflib_get_mbuf_size_for(unsigned int size);
714 static void iflib_init_locked(if_ctx_t ctx);
715 static void iflib_add_device_sysctl_pre(if_ctx_t ctx);
716 static void iflib_add_device_sysctl_post(if_ctx_t ctx);
717 static void iflib_ifmp_purge(iflib_txq_t txq);
718 static void _iflib_pre_assert(if_softc_ctx_t scctx);
719 static void iflib_if_init_locked(if_ctx_t ctx);
720 static void iflib_free_intr_mem(if_ctx_t ctx);
721 #ifndef __NO_STRICT_ALIGNMENT
722 static struct mbuf * iflib_fixup_rx(struct mbuf *m);
723 #endif
724 
725 static SLIST_HEAD(cpu_offset_list, cpu_offset) cpu_offsets =
726     SLIST_HEAD_INITIALIZER(cpu_offsets);
727 struct cpu_offset {
728 	SLIST_ENTRY(cpu_offset) entries;
729 	cpuset_t	set;
730 	unsigned int	refcount;
731 	uint16_t	next_cpuid;
732 };
733 static struct mtx cpu_offset_mtx;
734 MTX_SYSINIT(iflib_cpu_offset, &cpu_offset_mtx, "iflib_cpu_offset lock",
735     MTX_DEF);
736 
737 DEBUGNET_DEFINE(iflib);
738 
739 static int
740 iflib_num_rx_descs(if_ctx_t ctx)
741 {
742 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
743 	if_shared_ctx_t sctx = ctx->ifc_sctx;
744 	uint16_t first_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
745 
746 	return scctx->isc_nrxd[first_rxq];
747 }
748 
749 static int
750 iflib_num_tx_descs(if_ctx_t ctx)
751 {
752 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
753 	if_shared_ctx_t sctx = ctx->ifc_sctx;
754 	uint16_t first_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
755 
756 	return scctx->isc_ntxd[first_txq];
757 }
758 
759 #ifdef DEV_NETMAP
760 #include <sys/selinfo.h>
761 #include <net/netmap.h>
762 #include <dev/netmap/netmap_kern.h>
763 
764 MODULE_DEPEND(iflib, netmap, 1, 1, 1);
765 
766 static int netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, bool init);
767 static void iflib_netmap_timer(void *arg);
768 
769 /*
770  * device-specific sysctl variables:
771  *
772  * iflib_crcstrip: 0: keep CRC in rx frames (default), 1: strip it.
773  *	During regular operations the CRC is stripped, but on some
774  *	hardware reception of frames not multiple of 64 is slower,
775  *	so using crcstrip=0 helps in benchmarks.
776  *
777  * iflib_rx_miss, iflib_rx_miss_bufs:
778  *	count packets that might be missed due to lost interrupts.
779  */
780 SYSCTL_DECL(_dev_netmap);
781 /*
782  * The xl driver by default strips CRCs and we do not override it.
783  */
784 
785 int iflib_crcstrip = 1;
786 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_crcstrip,
787     CTLFLAG_RW, &iflib_crcstrip, 1, "strip CRC on RX frames");
788 
789 int iflib_rx_miss, iflib_rx_miss_bufs;
790 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss,
791     CTLFLAG_RW, &iflib_rx_miss, 0, "potentially missed RX intr");
792 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss_bufs,
793     CTLFLAG_RW, &iflib_rx_miss_bufs, 0, "potentially missed RX intr bufs");
794 
795 /*
796  * Register/unregister. We are already under netmap lock.
797  * Only called on the first register or the last unregister.
798  */
799 static int
800 iflib_netmap_register(struct netmap_adapter *na, int onoff)
801 {
802 	if_t ifp = na->ifp;
803 	if_ctx_t ctx = ifp->if_softc;
804 	int status;
805 
806 	CTX_LOCK(ctx);
807 	if (!CTX_IS_VF(ctx))
808 		IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip);
809 
810 	iflib_stop(ctx);
811 
812 	/*
813 	 * Enable (or disable) netmap flags, and intercept (or restore)
814 	 * ifp->if_transmit. This is done once the device has been stopped
815 	 * to prevent race conditions. Also, this must be done after
816 	 * calling netmap_disable_all_rings() and before calling
817 	 * netmap_enable_all_rings(), so that these two functions see the
818 	 * updated state of the NAF_NETMAP_ON bit.
819 	 */
820 	if (onoff) {
821 		nm_set_native_flags(na);
822 	} else {
823 		nm_clear_native_flags(na);
824 	}
825 
826 	iflib_init_locked(ctx);
827 	IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip); // XXX why twice ?
828 	status = ifp->if_drv_flags & IFF_DRV_RUNNING ? 0 : 1;
829 	if (status)
830 		nm_clear_native_flags(na);
831 	CTX_UNLOCK(ctx);
832 	return (status);
833 }
834 
835 static int
836 iflib_netmap_config(struct netmap_adapter *na, struct nm_config_info *info)
837 {
838 	if_t ifp = na->ifp;
839 	if_ctx_t ctx = ifp->if_softc;
840 	iflib_rxq_t rxq = &ctx->ifc_rxqs[0];
841 	iflib_fl_t fl = &rxq->ifr_fl[0];
842 
843 	info->num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
844 	info->num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
845 	info->num_tx_descs = iflib_num_tx_descs(ctx);
846 	info->num_rx_descs = iflib_num_rx_descs(ctx);
847 	info->rx_buf_maxsize = fl->ifl_buf_size;
848 	nm_prinf("txr %u rxr %u txd %u rxd %u rbufsz %u",
849 		info->num_tx_rings, info->num_rx_rings, info->num_tx_descs,
850 		info->num_rx_descs, info->rx_buf_maxsize);
851 
852 	return 0;
853 }
854 
855 static int
856 netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, bool init)
857 {
858 	struct netmap_adapter *na = kring->na;
859 	u_int const lim = kring->nkr_num_slots - 1;
860 	struct netmap_ring *ring = kring->ring;
861 	bus_dmamap_t *map;
862 	struct if_rxd_update iru;
863 	if_ctx_t ctx = rxq->ifr_ctx;
864 	iflib_fl_t fl = &rxq->ifr_fl[0];
865 	u_int nic_i_first, nic_i;
866 	u_int nm_i;
867 	int i, n;
868 #if IFLIB_DEBUG_COUNTERS
869 	int rf_count = 0;
870 #endif
871 
872 	/*
873 	 * This function is used both at initialization and in rxsync.
874 	 * At initialization we need to prepare (with isc_rxd_refill())
875 	 * all the netmap buffers currently owned by the kernel, in
876 	 * such a way to keep fl->ifl_pidx and kring->nr_hwcur in sync
877 	 * (except for kring->nkr_hwofs). These may be less than
878 	 * kring->nkr_num_slots if netmap_reset() was called while
879 	 * an application using the kring that still owned some
880 	 * buffers.
881 	 * At rxsync time, both indexes point to the next buffer to be
882 	 * refilled.
883 	 * In any case we publish (with isc_rxd_flush()) up to
884 	 * (fl->ifl_pidx - 1) % N (included), to avoid the NIC tail/prod
885 	 * pointer to overrun the head/cons pointer, although this is
886 	 * not necessary for some NICs (e.g. vmx).
887 	 */
888 	if (__predict_false(init)) {
889 		n = kring->nkr_num_slots - nm_kr_rxspace(kring);
890 	} else {
891 		n = kring->rhead - kring->nr_hwcur;
892 		if (n == 0)
893 			return (0); /* Nothing to do. */
894 		if (n < 0)
895 			n += kring->nkr_num_slots;
896 	}
897 
898 	iru_init(&iru, rxq, 0 /* flid */);
899 	map = fl->ifl_sds.ifsd_map;
900 	nic_i = fl->ifl_pidx;
901 	nm_i = netmap_idx_n2k(kring, nic_i);
902 	if (__predict_false(init)) {
903 		/*
904 		 * On init/reset, nic_i must be 0, and we must
905 		 * start to refill from hwtail (see netmap_reset()).
906 		 */
907 		MPASS(nic_i == 0);
908 		MPASS(nm_i == kring->nr_hwtail);
909 	} else
910 		MPASS(nm_i == kring->nr_hwcur);
911 	DBG_COUNTER_INC(fl_refills);
912 	while (n > 0) {
913 #if IFLIB_DEBUG_COUNTERS
914 		if (++rf_count == 9)
915 			DBG_COUNTER_INC(fl_refills_large);
916 #endif
917 		nic_i_first = nic_i;
918 		for (i = 0; n > 0 && i < IFLIB_MAX_RX_REFRESH; n--, i++) {
919 			struct netmap_slot *slot = &ring->slot[nm_i];
920 			uint64_t paddr;
921 			void *addr = PNMB(na, slot, &paddr);
922 
923 			MPASS(i < IFLIB_MAX_RX_REFRESH);
924 
925 			if (addr == NETMAP_BUF_BASE(na)) /* bad buf */
926 			        return netmap_ring_reinit(kring);
927 
928 			fl->ifl_bus_addrs[i] = paddr +
929 			    nm_get_offset(kring, slot);
930 			fl->ifl_rxd_idxs[i] = nic_i;
931 
932 			if (__predict_false(init)) {
933 				netmap_load_map(na, fl->ifl_buf_tag,
934 				    map[nic_i], addr);
935 			} else if (slot->flags & NS_BUF_CHANGED) {
936 				/* buffer has changed, reload map */
937 				netmap_reload_map(na, fl->ifl_buf_tag,
938 				    map[nic_i], addr);
939 			}
940 			bus_dmamap_sync(fl->ifl_buf_tag, map[nic_i],
941 			    BUS_DMASYNC_PREREAD);
942 			slot->flags &= ~NS_BUF_CHANGED;
943 
944 			nm_i = nm_next(nm_i, lim);
945 			nic_i = nm_next(nic_i, lim);
946 		}
947 
948 		iru.iru_pidx = nic_i_first;
949 		iru.iru_count = i;
950 		ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
951 	}
952 	fl->ifl_pidx = nic_i;
953 	/*
954 	 * At the end of the loop we must have refilled everything
955 	 * we could possibly refill.
956 	 */
957 	MPASS(nm_i == kring->rhead);
958 	kring->nr_hwcur = nm_i;
959 
960 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
961 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
962 	ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id,
963 	    nm_prev(nic_i, lim));
964 	DBG_COUNTER_INC(rxd_flush);
965 
966 	return (0);
967 }
968 
969 #define NETMAP_TX_TIMER_US	90
970 
971 /*
972  * Reconcile kernel and user view of the transmit ring.
973  *
974  * All information is in the kring.
975  * Userspace wants to send packets up to the one before kring->rhead,
976  * kernel knows kring->nr_hwcur is the first unsent packet.
977  *
978  * Here we push packets out (as many as possible), and possibly
979  * reclaim buffers from previously completed transmission.
980  *
981  * The caller (netmap) guarantees that there is only one instance
982  * running at any time. Any interference with other driver
983  * methods should be handled by the individual drivers.
984  */
985 static int
986 iflib_netmap_txsync(struct netmap_kring *kring, int flags)
987 {
988 	struct netmap_adapter *na = kring->na;
989 	if_t ifp = na->ifp;
990 	struct netmap_ring *ring = kring->ring;
991 	u_int nm_i;	/* index into the netmap kring */
992 	u_int nic_i;	/* index into the NIC ring */
993 	u_int const lim = kring->nkr_num_slots - 1;
994 	u_int const head = kring->rhead;
995 	struct if_pkt_info pi;
996 	int tx_pkts = 0, tx_bytes = 0;
997 
998 	/*
999 	 * interrupts on every tx packet are expensive so request
1000 	 * them every half ring, or where NS_REPORT is set
1001 	 */
1002 	u_int report_frequency = kring->nkr_num_slots >> 1;
1003 	/* device-specific */
1004 	if_ctx_t ctx = ifp->if_softc;
1005 	iflib_txq_t txq = &ctx->ifc_txqs[kring->ring_id];
1006 
1007 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1008 	    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1009 
1010 	/*
1011 	 * First part: process new packets to send.
1012 	 * nm_i is the current index in the netmap kring,
1013 	 * nic_i is the corresponding index in the NIC ring.
1014 	 *
1015 	 * If we have packets to send (nm_i != head)
1016 	 * iterate over the netmap ring, fetch length and update
1017 	 * the corresponding slot in the NIC ring. Some drivers also
1018 	 * need to update the buffer's physical address in the NIC slot
1019 	 * even NS_BUF_CHANGED is not set (PNMB computes the addresses).
1020 	 *
1021 	 * The netmap_reload_map() calls is especially expensive,
1022 	 * even when (as in this case) the tag is 0, so do only
1023 	 * when the buffer has actually changed.
1024 	 *
1025 	 * If possible do not set the report/intr bit on all slots,
1026 	 * but only a few times per ring or when NS_REPORT is set.
1027 	 *
1028 	 * Finally, on 10G and faster drivers, it might be useful
1029 	 * to prefetch the next slot and txr entry.
1030 	 */
1031 
1032 	nm_i = kring->nr_hwcur;
1033 	if (nm_i != head) {	/* we have new packets to send */
1034 		uint32_t pkt_len = 0, seg_idx = 0;
1035 		int nic_i_start = -1, flags = 0;
1036 		pkt_info_zero(&pi);
1037 		pi.ipi_segs = txq->ift_segs;
1038 		pi.ipi_qsidx = kring->ring_id;
1039 		nic_i = netmap_idx_k2n(kring, nm_i);
1040 
1041 		__builtin_prefetch(&ring->slot[nm_i]);
1042 		__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i]);
1043 		__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i]);
1044 
1045 		while (nm_i != head) {
1046 			struct netmap_slot *slot = &ring->slot[nm_i];
1047 			uint64_t offset = nm_get_offset(kring, slot);
1048 			u_int len = slot->len;
1049 			uint64_t paddr;
1050 			void *addr = PNMB(na, slot, &paddr);
1051 
1052 			flags |= (slot->flags & NS_REPORT ||
1053 				nic_i == 0 || nic_i == report_frequency) ?
1054 				IPI_TX_INTR : 0;
1055 
1056 			/*
1057 			 * If this is the first packet fragment, save the
1058 			 * index of the first NIC slot for later.
1059 			 */
1060 			if (nic_i_start < 0)
1061 				nic_i_start = nic_i;
1062 
1063 			pi.ipi_segs[seg_idx].ds_addr = paddr + offset;
1064 			pi.ipi_segs[seg_idx].ds_len = len;
1065 			if (len) {
1066 				pkt_len += len;
1067 				seg_idx++;
1068 			}
1069 
1070 			if (!(slot->flags & NS_MOREFRAG)) {
1071 				pi.ipi_len = pkt_len;
1072 				pi.ipi_nsegs = seg_idx;
1073 				pi.ipi_pidx = nic_i_start;
1074 				pi.ipi_ndescs = 0;
1075 				pi.ipi_flags = flags;
1076 
1077 				/* Prepare the NIC TX ring. */
1078 				ctx->isc_txd_encap(ctx->ifc_softc, &pi);
1079 				DBG_COUNTER_INC(tx_encap);
1080 
1081 				/* Update transmit counters */
1082 				tx_bytes += pi.ipi_len;
1083 				tx_pkts++;
1084 
1085 				/* Reinit per-packet info for the next one. */
1086 				flags = seg_idx = pkt_len = 0;
1087 				nic_i_start = -1;
1088 			}
1089 
1090 			/* prefetch for next round */
1091 			__builtin_prefetch(&ring->slot[nm_i + 1]);
1092 			__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i + 1]);
1093 			__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i + 1]);
1094 
1095 			NM_CHECK_ADDR_LEN_OFF(na, len, offset);
1096 
1097 			if (slot->flags & NS_BUF_CHANGED) {
1098 				/* buffer has changed, reload map */
1099 				netmap_reload_map(na, txq->ift_buf_tag,
1100 				    txq->ift_sds.ifsd_map[nic_i], addr);
1101 			}
1102 			/* make sure changes to the buffer are synced */
1103 			bus_dmamap_sync(txq->ift_buf_tag,
1104 			    txq->ift_sds.ifsd_map[nic_i],
1105 			    BUS_DMASYNC_PREWRITE);
1106 
1107 			slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED | NS_MOREFRAG);
1108 			nm_i = nm_next(nm_i, lim);
1109 			nic_i = nm_next(nic_i, lim);
1110 		}
1111 		kring->nr_hwcur = nm_i;
1112 
1113 		/* synchronize the NIC ring */
1114 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1115 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1116 
1117 		/* (re)start the tx unit up to slot nic_i (excluded) */
1118 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, nic_i);
1119 	}
1120 
1121 	/*
1122 	 * Second part: reclaim buffers for completed transmissions.
1123 	 *
1124 	 * If there are unclaimed buffers, attempt to reclaim them.
1125 	 * If we don't manage to reclaim them all, and TX IRQs are not in use,
1126 	 * trigger a per-tx-queue timer to try again later.
1127 	 */
1128 	if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1129 		if (iflib_tx_credits_update(ctx, txq)) {
1130 			/* some tx completed, increment avail */
1131 			nic_i = txq->ift_cidx_processed;
1132 			kring->nr_hwtail = nm_prev(netmap_idx_n2k(kring, nic_i), lim);
1133 		}
1134 	}
1135 
1136 	if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ))
1137 		if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1138 			callout_reset_sbt_on(&txq->ift_netmap_timer,
1139 			    NETMAP_TX_TIMER_US * SBT_1US, SBT_1US,
1140 			    iflib_netmap_timer, txq,
1141 			    txq->ift_netmap_timer.c_cpu, 0);
1142 		}
1143 
1144 	if_inc_counter(ifp, IFCOUNTER_OBYTES, tx_bytes);
1145 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, tx_pkts);
1146 
1147 	return (0);
1148 }
1149 
1150 /*
1151  * Reconcile kernel and user view of the receive ring.
1152  * Same as for the txsync, this routine must be efficient.
1153  * The caller guarantees a single invocations, but races against
1154  * the rest of the driver should be handled here.
1155  *
1156  * On call, kring->rhead is the first packet that userspace wants
1157  * to keep, and kring->rcur is the wakeup point.
1158  * The kernel has previously reported packets up to kring->rtail.
1159  *
1160  * If (flags & NAF_FORCE_READ) also check for incoming packets irrespective
1161  * of whether or not we received an interrupt.
1162  */
1163 static int
1164 iflib_netmap_rxsync(struct netmap_kring *kring, int flags)
1165 {
1166 	struct netmap_adapter *na = kring->na;
1167 	struct netmap_ring *ring = kring->ring;
1168 	if_t ifp = na->ifp;
1169 	uint32_t nm_i;	/* index into the netmap ring */
1170 	uint32_t nic_i;	/* index into the NIC ring */
1171 	u_int n;
1172 	u_int const lim = kring->nkr_num_slots - 1;
1173 	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
1174 	int i = 0, rx_bytes = 0, rx_pkts = 0;
1175 
1176 	if_ctx_t ctx = ifp->if_softc;
1177 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1178 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1179 	iflib_rxq_t rxq = &ctx->ifc_rxqs[kring->ring_id];
1180 	iflib_fl_t fl = &rxq->ifr_fl[0];
1181 	struct if_rxd_info ri;
1182 	qidx_t *cidxp;
1183 
1184 	/*
1185 	 * netmap only uses free list 0, to avoid out of order consumption
1186 	 * of receive buffers
1187 	 */
1188 
1189 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
1190 	    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1191 
1192 	/*
1193 	 * First part: import newly received packets.
1194 	 *
1195 	 * nm_i is the index of the next free slot in the netmap ring,
1196 	 * nic_i is the index of the next received packet in the NIC ring
1197 	 * (or in the free list 0 if IFLIB_HAS_RXCQ is set), and they may
1198 	 * differ in case if_init() has been called while
1199 	 * in netmap mode. For the receive ring we have
1200 	 *
1201 	 *	nic_i = fl->ifl_cidx;
1202 	 *	nm_i = kring->nr_hwtail (previous)
1203 	 * and
1204 	 *	nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1205 	 *
1206 	 * fl->ifl_cidx is set to 0 on a ring reinit
1207 	 */
1208 	if (netmap_no_pendintr || force_update) {
1209 		uint32_t hwtail_lim = nm_prev(kring->nr_hwcur, lim);
1210 		bool have_rxcq = sctx->isc_flags & IFLIB_HAS_RXCQ;
1211 		int crclen = iflib_crcstrip ? 0 : 4;
1212 		int error, avail;
1213 
1214 		/*
1215 		 * For the free list consumer index, we use the same
1216 		 * logic as in iflib_rxeof().
1217 		 */
1218 		if (have_rxcq)
1219 			cidxp = &rxq->ifr_cq_cidx;
1220 		else
1221 			cidxp = &fl->ifl_cidx;
1222 		avail = ctx->isc_rxd_available(ctx->ifc_softc,
1223 		    rxq->ifr_id, *cidxp, USHRT_MAX);
1224 
1225 		nic_i = fl->ifl_cidx;
1226 		nm_i = netmap_idx_n2k(kring, nic_i);
1227 		MPASS(nm_i == kring->nr_hwtail);
1228 		for (n = 0; avail > 0 && nm_i != hwtail_lim; n++, avail--) {
1229 			rxd_info_zero(&ri);
1230 			ri.iri_frags = rxq->ifr_frags;
1231 			ri.iri_qsidx = kring->ring_id;
1232 			ri.iri_ifp = ctx->ifc_ifp;
1233 			ri.iri_cidx = *cidxp;
1234 
1235 			error = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
1236 			for (i = 0; i < ri.iri_nfrags; i++) {
1237 				if (error) {
1238 					ring->slot[nm_i].len = 0;
1239 					ring->slot[nm_i].flags = 0;
1240 				} else {
1241 					ring->slot[nm_i].len = ri.iri_frags[i].irf_len;
1242 					if (i == (ri.iri_nfrags - 1)) {
1243 						ring->slot[nm_i].len -= crclen;
1244 						ring->slot[nm_i].flags = 0;
1245 
1246 						/* Update receive counters */
1247 						rx_bytes += ri.iri_len;
1248 						rx_pkts++;
1249 					} else
1250 						ring->slot[nm_i].flags = NS_MOREFRAG;
1251 				}
1252 
1253 				bus_dmamap_sync(fl->ifl_buf_tag,
1254 				    fl->ifl_sds.ifsd_map[nic_i], BUS_DMASYNC_POSTREAD);
1255 				nm_i = nm_next(nm_i, lim);
1256 				fl->ifl_cidx = nic_i = nm_next(nic_i, lim);
1257 			}
1258 
1259 			if (have_rxcq) {
1260 				*cidxp = ri.iri_cidx;
1261 				while (*cidxp >= scctx->isc_nrxd[0])
1262 					*cidxp -= scctx->isc_nrxd[0];
1263 			}
1264 
1265 		}
1266 		if (n) { /* update the state variables */
1267 			if (netmap_no_pendintr && !force_update) {
1268 				/* diagnostics */
1269 				iflib_rx_miss ++;
1270 				iflib_rx_miss_bufs += n;
1271 			}
1272 			kring->nr_hwtail = nm_i;
1273 		}
1274 		kring->nr_kflags &= ~NKR_PENDINTR;
1275 	}
1276 	/*
1277 	 * Second part: skip past packets that userspace has released.
1278 	 * (kring->nr_hwcur to head excluded),
1279 	 * and make the buffers available for reception.
1280 	 * As usual nm_i is the index in the netmap ring,
1281 	 * nic_i is the index in the NIC ring, and
1282 	 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1283 	 */
1284 	netmap_fl_refill(rxq, kring, false);
1285 
1286 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
1287 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
1288 
1289 	return (0);
1290 }
1291 
1292 static void
1293 iflib_netmap_intr(struct netmap_adapter *na, int onoff)
1294 {
1295 	if_ctx_t ctx = na->ifp->if_softc;
1296 
1297 	CTX_LOCK(ctx);
1298 	if (onoff) {
1299 		IFDI_INTR_ENABLE(ctx);
1300 	} else {
1301 		IFDI_INTR_DISABLE(ctx);
1302 	}
1303 	CTX_UNLOCK(ctx);
1304 }
1305 
1306 static int
1307 iflib_netmap_attach(if_ctx_t ctx)
1308 {
1309 	struct netmap_adapter na;
1310 
1311 	bzero(&na, sizeof(na));
1312 
1313 	na.ifp = ctx->ifc_ifp;
1314 	na.na_flags = NAF_BDG_MAYSLEEP | NAF_MOREFRAG | NAF_OFFSETS;
1315 	MPASS(ctx->ifc_softc_ctx.isc_ntxqsets);
1316 	MPASS(ctx->ifc_softc_ctx.isc_nrxqsets);
1317 
1318 	na.num_tx_desc = iflib_num_tx_descs(ctx);
1319 	na.num_rx_desc = iflib_num_rx_descs(ctx);
1320 	na.nm_txsync = iflib_netmap_txsync;
1321 	na.nm_rxsync = iflib_netmap_rxsync;
1322 	na.nm_register = iflib_netmap_register;
1323 	na.nm_intr = iflib_netmap_intr;
1324 	na.nm_config = iflib_netmap_config;
1325 	na.num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
1326 	na.num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
1327 	return (netmap_attach(&na));
1328 }
1329 
1330 static int
1331 iflib_netmap_txq_init(if_ctx_t ctx, iflib_txq_t txq)
1332 {
1333 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1334 	struct netmap_slot *slot;
1335 
1336 	slot = netmap_reset(na, NR_TX, txq->ift_id, 0);
1337 	if (slot == NULL)
1338 		return (0);
1339 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxd[0]; i++) {
1340 		/*
1341 		 * In netmap mode, set the map for the packet buffer.
1342 		 * NOTE: Some drivers (not this one) also need to set
1343 		 * the physical buffer address in the NIC ring.
1344 		 * netmap_idx_n2k() maps a nic index, i, into the corresponding
1345 		 * netmap slot index, si
1346 		 */
1347 		int si = netmap_idx_n2k(na->tx_rings[txq->ift_id], i);
1348 		netmap_load_map(na, txq->ift_buf_tag, txq->ift_sds.ifsd_map[i],
1349 		    NMB(na, slot + si));
1350 	}
1351 	return (1);
1352 }
1353 
1354 static int
1355 iflib_netmap_rxq_init(if_ctx_t ctx, iflib_rxq_t rxq)
1356 {
1357 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1358 	struct netmap_kring *kring;
1359 	struct netmap_slot *slot;
1360 
1361 	slot = netmap_reset(na, NR_RX, rxq->ifr_id, 0);
1362 	if (slot == NULL)
1363 		return (0);
1364 	kring = na->rx_rings[rxq->ifr_id];
1365 	netmap_fl_refill(rxq, kring, true);
1366 	return (1);
1367 }
1368 
1369 static void
1370 iflib_netmap_timer(void *arg)
1371 {
1372 	iflib_txq_t txq = arg;
1373 	if_ctx_t ctx = txq->ift_ctx;
1374 
1375 	/*
1376 	 * Wake up the netmap application, to give it a chance to
1377 	 * call txsync and reclaim more completed TX buffers.
1378 	 */
1379 	netmap_tx_irq(ctx->ifc_ifp, txq->ift_id);
1380 }
1381 
1382 #define iflib_netmap_detach(ifp) netmap_detach(ifp)
1383 
1384 #else
1385 #define iflib_netmap_txq_init(ctx, txq) (0)
1386 #define iflib_netmap_rxq_init(ctx, rxq) (0)
1387 #define iflib_netmap_detach(ifp)
1388 #define netmap_enable_all_rings(ifp)
1389 #define netmap_disable_all_rings(ifp)
1390 
1391 #define iflib_netmap_attach(ctx) (0)
1392 #define netmap_rx_irq(ifp, qid, budget) (0)
1393 #endif
1394 
1395 #if defined(__i386__) || defined(__amd64__)
1396 static __inline void
1397 prefetch(void *x)
1398 {
1399 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1400 }
1401 
1402 static __inline void
1403 prefetch2cachelines(void *x)
1404 {
1405 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1406 #if (CACHE_LINE_SIZE < 128)
1407 	__asm volatile("prefetcht0 %0" :: "m" (*(((unsigned long *)x)+CACHE_LINE_SIZE/(sizeof(unsigned long)))));
1408 #endif
1409 }
1410 #else
1411 static __inline void
1412 prefetch(void *x)
1413 {
1414 }
1415 
1416 static __inline void
1417 prefetch2cachelines(void *x)
1418 {
1419 }
1420 #endif
1421 
1422 static void
1423 iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid)
1424 {
1425 	iflib_fl_t fl;
1426 
1427 	fl = &rxq->ifr_fl[flid];
1428 	iru->iru_paddrs = fl->ifl_bus_addrs;
1429 	iru->iru_idxs = fl->ifl_rxd_idxs;
1430 	iru->iru_qsidx = rxq->ifr_id;
1431 	iru->iru_buf_size = fl->ifl_buf_size;
1432 	iru->iru_flidx = fl->ifl_id;
1433 }
1434 
1435 static void
1436 _iflib_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int err)
1437 {
1438 	if (err)
1439 		return;
1440 	*(bus_addr_t *) arg = segs[0].ds_addr;
1441 }
1442 
1443 #define	DMA_WIDTH_TO_BUS_LOWADDR(width)				\
1444 	(((width) == 0) || (width) == flsll(BUS_SPACE_MAXADDR) ?	\
1445 	    BUS_SPACE_MAXADDR : (1ULL << (width)) - 1ULL)
1446 
1447 int
1448 iflib_dma_alloc_align(if_ctx_t ctx, int size, int align, iflib_dma_info_t dma, int mapflags)
1449 {
1450 	int err;
1451 	device_t dev = ctx->ifc_dev;
1452 	bus_addr_t lowaddr;
1453 
1454 	lowaddr = DMA_WIDTH_TO_BUS_LOWADDR(ctx->ifc_softc_ctx.isc_dma_width);
1455 
1456 	err = bus_dma_tag_create(bus_get_dma_tag(dev),	/* parent */
1457 				align, 0,		/* alignment, bounds */
1458 				lowaddr,		/* lowaddr */
1459 				BUS_SPACE_MAXADDR,	/* highaddr */
1460 				NULL, NULL,		/* filter, filterarg */
1461 				size,			/* maxsize */
1462 				1,			/* nsegments */
1463 				size,			/* maxsegsize */
1464 				BUS_DMA_ALLOCNOW,	/* flags */
1465 				NULL,			/* lockfunc */
1466 				NULL,			/* lockarg */
1467 				&dma->idi_tag);
1468 	if (err) {
1469 		device_printf(dev,
1470 		    "%s: bus_dma_tag_create failed: %d\n",
1471 		    __func__, err);
1472 		goto fail_0;
1473 	}
1474 
1475 	err = bus_dmamem_alloc(dma->idi_tag, (void**) &dma->idi_vaddr,
1476 	    BUS_DMA_NOWAIT | BUS_DMA_COHERENT | BUS_DMA_ZERO, &dma->idi_map);
1477 	if (err) {
1478 		device_printf(dev,
1479 		    "%s: bus_dmamem_alloc(%ju) failed: %d\n",
1480 		    __func__, (uintmax_t)size, err);
1481 		goto fail_1;
1482 	}
1483 
1484 	dma->idi_paddr = IF_BAD_DMA;
1485 	err = bus_dmamap_load(dma->idi_tag, dma->idi_map, dma->idi_vaddr,
1486 	    size, _iflib_dmamap_cb, &dma->idi_paddr, mapflags | BUS_DMA_NOWAIT);
1487 	if (err || dma->idi_paddr == IF_BAD_DMA) {
1488 		device_printf(dev,
1489 		    "%s: bus_dmamap_load failed: %d\n",
1490 		    __func__, err);
1491 		goto fail_2;
1492 	}
1493 
1494 	dma->idi_size = size;
1495 	return (0);
1496 
1497 fail_2:
1498 	bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1499 fail_1:
1500 	bus_dma_tag_destroy(dma->idi_tag);
1501 fail_0:
1502 	dma->idi_tag = NULL;
1503 
1504 	return (err);
1505 }
1506 
1507 int
1508 iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags)
1509 {
1510 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1511 
1512 	KASSERT(sctx->isc_q_align != 0, ("alignment value not initialized"));
1513 
1514 	return (iflib_dma_alloc_align(ctx, size, sctx->isc_q_align, dma, mapflags));
1515 }
1516 
1517 int
1518 iflib_dma_alloc_multi(if_ctx_t ctx, int *sizes, iflib_dma_info_t *dmalist, int mapflags, int count)
1519 {
1520 	int i, err;
1521 	iflib_dma_info_t *dmaiter;
1522 
1523 	dmaiter = dmalist;
1524 	for (i = 0; i < count; i++, dmaiter++) {
1525 		if ((err = iflib_dma_alloc(ctx, sizes[i], *dmaiter, mapflags)) != 0)
1526 			break;
1527 	}
1528 	if (err)
1529 		iflib_dma_free_multi(dmalist, i);
1530 	return (err);
1531 }
1532 
1533 void
1534 iflib_dma_free(iflib_dma_info_t dma)
1535 {
1536 	if (dma->idi_tag == NULL)
1537 		return;
1538 	if (dma->idi_paddr != IF_BAD_DMA) {
1539 		bus_dmamap_sync(dma->idi_tag, dma->idi_map,
1540 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1541 		bus_dmamap_unload(dma->idi_tag, dma->idi_map);
1542 		dma->idi_paddr = IF_BAD_DMA;
1543 	}
1544 	if (dma->idi_vaddr != NULL) {
1545 		bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1546 		dma->idi_vaddr = NULL;
1547 	}
1548 	bus_dma_tag_destroy(dma->idi_tag);
1549 	dma->idi_tag = NULL;
1550 }
1551 
1552 void
1553 iflib_dma_free_multi(iflib_dma_info_t *dmalist, int count)
1554 {
1555 	int i;
1556 	iflib_dma_info_t *dmaiter = dmalist;
1557 
1558 	for (i = 0; i < count; i++, dmaiter++)
1559 		iflib_dma_free(*dmaiter);
1560 }
1561 
1562 static int
1563 iflib_fast_intr(void *arg)
1564 {
1565 	iflib_filter_info_t info = arg;
1566 	struct grouptask *gtask = info->ifi_task;
1567 	int result;
1568 
1569 	DBG_COUNTER_INC(fast_intrs);
1570 	if (info->ifi_filter != NULL) {
1571 		result = info->ifi_filter(info->ifi_filter_arg);
1572 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1573 			return (result);
1574 	}
1575 
1576 	GROUPTASK_ENQUEUE(gtask);
1577 	return (FILTER_HANDLED);
1578 }
1579 
1580 static int
1581 iflib_fast_intr_rxtx(void *arg)
1582 {
1583 	iflib_filter_info_t info = arg;
1584 	struct grouptask *gtask = info->ifi_task;
1585 	if_ctx_t ctx;
1586 	iflib_rxq_t rxq = (iflib_rxq_t)info->ifi_ctx;
1587 	iflib_txq_t txq;
1588 	void *sc;
1589 	int i, cidx, result;
1590 	qidx_t txqid;
1591 	bool intr_enable, intr_legacy;
1592 
1593 	DBG_COUNTER_INC(fast_intrs);
1594 	if (info->ifi_filter != NULL) {
1595 		result = info->ifi_filter(info->ifi_filter_arg);
1596 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1597 			return (result);
1598 	}
1599 
1600 	ctx = rxq->ifr_ctx;
1601 	sc = ctx->ifc_softc;
1602 	intr_enable = false;
1603 	intr_legacy = !!(ctx->ifc_flags & IFC_LEGACY);
1604 	MPASS(rxq->ifr_ntxqirq);
1605 	for (i = 0; i < rxq->ifr_ntxqirq; i++) {
1606 		txqid = rxq->ifr_txqid[i];
1607 		txq = &ctx->ifc_txqs[txqid];
1608 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1609 		    BUS_DMASYNC_POSTREAD);
1610 		if (!ctx->isc_txd_credits_update(sc, txqid, false)) {
1611 			if (intr_legacy)
1612 				intr_enable = true;
1613 			else
1614 				IFDI_TX_QUEUE_INTR_ENABLE(ctx, txqid);
1615 			continue;
1616 		}
1617 		GROUPTASK_ENQUEUE(&txq->ift_task);
1618 	}
1619 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_RXCQ)
1620 		cidx = rxq->ifr_cq_cidx;
1621 	else
1622 		cidx = rxq->ifr_fl[0].ifl_cidx;
1623 	if (iflib_rxd_avail(ctx, rxq, cidx, 1))
1624 		GROUPTASK_ENQUEUE(gtask);
1625 	else {
1626 		if (intr_legacy)
1627 			intr_enable = true;
1628 		else
1629 			IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
1630 		DBG_COUNTER_INC(rx_intr_enables);
1631 	}
1632 	if (intr_enable)
1633 		IFDI_INTR_ENABLE(ctx);
1634 	return (FILTER_HANDLED);
1635 }
1636 
1637 static int
1638 iflib_fast_intr_ctx(void *arg)
1639 {
1640 	iflib_filter_info_t info = arg;
1641 	struct grouptask *gtask = info->ifi_task;
1642 	int result;
1643 
1644 	DBG_COUNTER_INC(fast_intrs);
1645 	if (info->ifi_filter != NULL) {
1646 		result = info->ifi_filter(info->ifi_filter_arg);
1647 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1648 			return (result);
1649 	}
1650 
1651 	GROUPTASK_ENQUEUE(gtask);
1652 	return (FILTER_HANDLED);
1653 }
1654 
1655 static int
1656 _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
1657 		 driver_filter_t filter, driver_intr_t handler, void *arg,
1658 		 const char *name)
1659 {
1660 	struct resource *res;
1661 	void *tag = NULL;
1662 	device_t dev = ctx->ifc_dev;
1663 	int flags, i, rc;
1664 
1665 	flags = RF_ACTIVE;
1666 	if (ctx->ifc_flags & IFC_LEGACY)
1667 		flags |= RF_SHAREABLE;
1668 	MPASS(rid < 512);
1669 	i = rid;
1670 	res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &i, flags);
1671 	if (res == NULL) {
1672 		device_printf(dev,
1673 		    "failed to allocate IRQ for rid %d, name %s.\n", rid, name);
1674 		return (ENOMEM);
1675 	}
1676 	irq->ii_res = res;
1677 	KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL"));
1678 	rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET,
1679 						filter, handler, arg, &tag);
1680 	if (rc != 0) {
1681 		device_printf(dev,
1682 		    "failed to setup interrupt for rid %d, name %s: %d\n",
1683 					  rid, name ? name : "unknown", rc);
1684 		return (rc);
1685 	} else if (name)
1686 		bus_describe_intr(dev, res, tag, "%s", name);
1687 
1688 	irq->ii_tag = tag;
1689 	return (0);
1690 }
1691 
1692 /*********************************************************************
1693  *
1694  *  Allocate DMA resources for TX buffers as well as memory for the TX
1695  *  mbuf map.  TX DMA maps (non-TSO/TSO) and TX mbuf map are kept in a
1696  *  iflib_sw_tx_desc_array structure, storing all the information that
1697  *  is needed to transmit a packet on the wire.  This is called only
1698  *  once at attach, setup is done every reset.
1699  *
1700  **********************************************************************/
1701 static int
1702 iflib_txsd_alloc(iflib_txq_t txq)
1703 {
1704 	if_ctx_t ctx = txq->ift_ctx;
1705 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1706 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1707 	device_t dev = ctx->ifc_dev;
1708 	bus_size_t tsomaxsize;
1709 	bus_addr_t lowaddr;
1710 	int err, nsegments, ntsosegments;
1711 	bool tso;
1712 
1713 	nsegments = scctx->isc_tx_nsegments;
1714 	ntsosegments = scctx->isc_tx_tso_segments_max;
1715 	tsomaxsize = scctx->isc_tx_tso_size_max;
1716 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_VLAN_MTU)
1717 		tsomaxsize += sizeof(struct ether_vlan_header);
1718 	MPASS(scctx->isc_ntxd[0] > 0);
1719 	MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0);
1720 	MPASS(nsegments > 0);
1721 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) {
1722 		MPASS(ntsosegments > 0);
1723 		MPASS(sctx->isc_tso_maxsize >= tsomaxsize);
1724 	}
1725 
1726 	lowaddr = DMA_WIDTH_TO_BUS_LOWADDR(scctx->isc_dma_width);
1727 
1728 	/*
1729 	 * Set up DMA tags for TX buffers.
1730 	 */
1731 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1732 			       1, 0,			/* alignment, bounds */
1733 			       lowaddr,			/* lowaddr */
1734 			       BUS_SPACE_MAXADDR,	/* highaddr */
1735 			       NULL, NULL,		/* filter, filterarg */
1736 			       sctx->isc_tx_maxsize,		/* maxsize */
1737 			       nsegments,	/* nsegments */
1738 			       sctx->isc_tx_maxsegsize,	/* maxsegsize */
1739 			       0,			/* flags */
1740 			       NULL,			/* lockfunc */
1741 			       NULL,			/* lockfuncarg */
1742 			       &txq->ift_buf_tag))) {
1743 		device_printf(dev,"Unable to allocate TX DMA tag: %d\n", err);
1744 		device_printf(dev,"maxsize: %ju nsegments: %d maxsegsize: %ju\n",
1745 		    (uintmax_t)sctx->isc_tx_maxsize, nsegments, (uintmax_t)sctx->isc_tx_maxsegsize);
1746 		goto fail;
1747 	}
1748 	tso = (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) != 0;
1749 	if (tso && (err = bus_dma_tag_create(bus_get_dma_tag(dev),
1750 			       1, 0,			/* alignment, bounds */
1751 			       lowaddr,			/* lowaddr */
1752 			       BUS_SPACE_MAXADDR,	/* highaddr */
1753 			       NULL, NULL,		/* filter, filterarg */
1754 			       tsomaxsize,		/* maxsize */
1755 			       ntsosegments,	/* nsegments */
1756 			       sctx->isc_tso_maxsegsize,/* maxsegsize */
1757 			       0,			/* flags */
1758 			       NULL,			/* lockfunc */
1759 			       NULL,			/* lockfuncarg */
1760 			       &txq->ift_tso_buf_tag))) {
1761 		device_printf(dev, "Unable to allocate TSO TX DMA tag: %d\n",
1762 		    err);
1763 		goto fail;
1764 	}
1765 
1766 	/* Allocate memory for the TX mbuf map. */
1767 	if (!(txq->ift_sds.ifsd_m =
1768 	    (struct mbuf **) malloc(sizeof(struct mbuf *) *
1769 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1770 		device_printf(dev, "Unable to allocate TX mbuf map memory\n");
1771 		err = ENOMEM;
1772 		goto fail;
1773 	}
1774 
1775 	/*
1776 	 * Create the DMA maps for TX buffers.
1777 	 */
1778 	if ((txq->ift_sds.ifsd_map = (bus_dmamap_t *)malloc(
1779 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1780 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1781 		device_printf(dev,
1782 		    "Unable to allocate TX buffer DMA map memory\n");
1783 		err = ENOMEM;
1784 		goto fail;
1785 	}
1786 	if (tso && (txq->ift_sds.ifsd_tso_map = (bus_dmamap_t *)malloc(
1787 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1788 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1789 		device_printf(dev,
1790 		    "Unable to allocate TSO TX buffer map memory\n");
1791 		err = ENOMEM;
1792 		goto fail;
1793 	}
1794 	for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) {
1795 		err = bus_dmamap_create(txq->ift_buf_tag, 0,
1796 		    &txq->ift_sds.ifsd_map[i]);
1797 		if (err != 0) {
1798 			device_printf(dev, "Unable to create TX DMA map\n");
1799 			goto fail;
1800 		}
1801 		if (!tso)
1802 			continue;
1803 		err = bus_dmamap_create(txq->ift_tso_buf_tag, 0,
1804 		    &txq->ift_sds.ifsd_tso_map[i]);
1805 		if (err != 0) {
1806 			device_printf(dev, "Unable to create TSO TX DMA map\n");
1807 			goto fail;
1808 		}
1809 	}
1810 	return (0);
1811 fail:
1812 	/* We free all, it handles case where we are in the middle */
1813 	iflib_tx_structures_free(ctx);
1814 	return (err);
1815 }
1816 
1817 static void
1818 iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i)
1819 {
1820 	bus_dmamap_t map;
1821 
1822 	if (txq->ift_sds.ifsd_map != NULL) {
1823 		map = txq->ift_sds.ifsd_map[i];
1824 		bus_dmamap_sync(txq->ift_buf_tag, map, BUS_DMASYNC_POSTWRITE);
1825 		bus_dmamap_unload(txq->ift_buf_tag, map);
1826 		bus_dmamap_destroy(txq->ift_buf_tag, map);
1827 		txq->ift_sds.ifsd_map[i] = NULL;
1828 	}
1829 
1830 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1831 		map = txq->ift_sds.ifsd_tso_map[i];
1832 		bus_dmamap_sync(txq->ift_tso_buf_tag, map,
1833 		    BUS_DMASYNC_POSTWRITE);
1834 		bus_dmamap_unload(txq->ift_tso_buf_tag, map);
1835 		bus_dmamap_destroy(txq->ift_tso_buf_tag, map);
1836 		txq->ift_sds.ifsd_tso_map[i] = NULL;
1837 	}
1838 }
1839 
1840 static void
1841 iflib_txq_destroy(iflib_txq_t txq)
1842 {
1843 	if_ctx_t ctx = txq->ift_ctx;
1844 
1845 	for (int i = 0; i < txq->ift_size; i++)
1846 		iflib_txsd_destroy(ctx, txq, i);
1847 
1848 	if (txq->ift_br != NULL) {
1849 		ifmp_ring_free(txq->ift_br);
1850 		txq->ift_br = NULL;
1851 	}
1852 
1853 	mtx_destroy(&txq->ift_mtx);
1854 
1855 	if (txq->ift_sds.ifsd_map != NULL) {
1856 		free(txq->ift_sds.ifsd_map, M_IFLIB);
1857 		txq->ift_sds.ifsd_map = NULL;
1858 	}
1859 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1860 		free(txq->ift_sds.ifsd_tso_map, M_IFLIB);
1861 		txq->ift_sds.ifsd_tso_map = NULL;
1862 	}
1863 	if (txq->ift_sds.ifsd_m != NULL) {
1864 		free(txq->ift_sds.ifsd_m, M_IFLIB);
1865 		txq->ift_sds.ifsd_m = NULL;
1866 	}
1867 	if (txq->ift_buf_tag != NULL) {
1868 		bus_dma_tag_destroy(txq->ift_buf_tag);
1869 		txq->ift_buf_tag = NULL;
1870 	}
1871 	if (txq->ift_tso_buf_tag != NULL) {
1872 		bus_dma_tag_destroy(txq->ift_tso_buf_tag);
1873 		txq->ift_tso_buf_tag = NULL;
1874 	}
1875 	if (txq->ift_ifdi != NULL) {
1876 		free(txq->ift_ifdi, M_IFLIB);
1877 	}
1878 }
1879 
1880 static void
1881 iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i)
1882 {
1883 	struct mbuf **mp;
1884 
1885 	mp = &txq->ift_sds.ifsd_m[i];
1886 	if (*mp == NULL)
1887 		return;
1888 
1889 	if (txq->ift_sds.ifsd_map != NULL) {
1890 		bus_dmamap_sync(txq->ift_buf_tag,
1891 		    txq->ift_sds.ifsd_map[i], BUS_DMASYNC_POSTWRITE);
1892 		bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[i]);
1893 	}
1894 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1895 		bus_dmamap_sync(txq->ift_tso_buf_tag,
1896 		    txq->ift_sds.ifsd_tso_map[i], BUS_DMASYNC_POSTWRITE);
1897 		bus_dmamap_unload(txq->ift_tso_buf_tag,
1898 		    txq->ift_sds.ifsd_tso_map[i]);
1899 	}
1900 	m_freem(*mp);
1901 	DBG_COUNTER_INC(tx_frees);
1902 	*mp = NULL;
1903 }
1904 
1905 static int
1906 iflib_txq_setup(iflib_txq_t txq)
1907 {
1908 	if_ctx_t ctx = txq->ift_ctx;
1909 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1910 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1911 	iflib_dma_info_t di;
1912 	int i;
1913 
1914 	/* Set number of descriptors available */
1915 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
1916 	/* XXX make configurable */
1917 	txq->ift_update_freq = IFLIB_DEFAULT_TX_UPDATE_FREQ;
1918 
1919 	/* Reset indices */
1920 	txq->ift_cidx_processed = 0;
1921 	txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0;
1922 	txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset];
1923 
1924 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1925 		bzero((void *)di->idi_vaddr, di->idi_size);
1926 
1927 	IFDI_TXQ_SETUP(ctx, txq->ift_id);
1928 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1929 		bus_dmamap_sync(di->idi_tag, di->idi_map,
1930 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1931 	return (0);
1932 }
1933 
1934 /*********************************************************************
1935  *
1936  *  Allocate DMA resources for RX buffers as well as memory for the RX
1937  *  mbuf map, direct RX cluster pointer map and RX cluster bus address
1938  *  map.  RX DMA map, RX mbuf map, direct RX cluster pointer map and
1939  *  RX cluster map are kept in a iflib_sw_rx_desc_array structure.
1940  *  Since we use use one entry in iflib_sw_rx_desc_array per received
1941  *  packet, the maximum number of entries we'll need is equal to the
1942  *  number of hardware receive descriptors that we've allocated.
1943  *
1944  **********************************************************************/
1945 static int
1946 iflib_rxsd_alloc(iflib_rxq_t rxq)
1947 {
1948 	if_ctx_t ctx = rxq->ifr_ctx;
1949 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1950 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1951 	device_t dev = ctx->ifc_dev;
1952 	iflib_fl_t fl;
1953 	bus_addr_t lowaddr;
1954 	int			err;
1955 
1956 	MPASS(scctx->isc_nrxd[0] > 0);
1957 	MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0);
1958 
1959 	lowaddr = DMA_WIDTH_TO_BUS_LOWADDR(scctx->isc_dma_width);
1960 
1961 	fl = rxq->ifr_fl;
1962 	for (int i = 0; i <  rxq->ifr_nfl; i++, fl++) {
1963 		fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */
1964 		/* Set up DMA tag for RX buffers. */
1965 		err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1966 					 1, 0,			/* alignment, bounds */
1967 					 lowaddr,		/* lowaddr */
1968 					 BUS_SPACE_MAXADDR,	/* highaddr */
1969 					 NULL, NULL,		/* filter, filterarg */
1970 					 sctx->isc_rx_maxsize,	/* maxsize */
1971 					 sctx->isc_rx_nsegments,	/* nsegments */
1972 					 sctx->isc_rx_maxsegsize,	/* maxsegsize */
1973 					 0,			/* flags */
1974 					 NULL,			/* lockfunc */
1975 					 NULL,			/* lockarg */
1976 					 &fl->ifl_buf_tag);
1977 		if (err) {
1978 			device_printf(dev,
1979 			    "Unable to allocate RX DMA tag: %d\n", err);
1980 			goto fail;
1981 		}
1982 
1983 		/* Allocate memory for the RX mbuf map. */
1984 		if (!(fl->ifl_sds.ifsd_m =
1985 		      (struct mbuf **) malloc(sizeof(struct mbuf *) *
1986 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1987 			device_printf(dev,
1988 			    "Unable to allocate RX mbuf map memory\n");
1989 			err = ENOMEM;
1990 			goto fail;
1991 		}
1992 
1993 		/* Allocate memory for the direct RX cluster pointer map. */
1994 		if (!(fl->ifl_sds.ifsd_cl =
1995 		      (caddr_t *) malloc(sizeof(caddr_t) *
1996 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1997 			device_printf(dev,
1998 			    "Unable to allocate RX cluster map memory\n");
1999 			err = ENOMEM;
2000 			goto fail;
2001 		}
2002 
2003 		/* Allocate memory for the RX cluster bus address map. */
2004 		if (!(fl->ifl_sds.ifsd_ba =
2005 		      (bus_addr_t *) malloc(sizeof(bus_addr_t) *
2006 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
2007 			device_printf(dev,
2008 			    "Unable to allocate RX bus address map memory\n");
2009 			err = ENOMEM;
2010 			goto fail;
2011 		}
2012 
2013 		/*
2014 		 * Create the DMA maps for RX buffers.
2015 		 */
2016 		if (!(fl->ifl_sds.ifsd_map =
2017 		      (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
2018 			device_printf(dev,
2019 			    "Unable to allocate RX buffer DMA map memory\n");
2020 			err = ENOMEM;
2021 			goto fail;
2022 		}
2023 		for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++) {
2024 			err = bus_dmamap_create(fl->ifl_buf_tag, 0,
2025 			    &fl->ifl_sds.ifsd_map[i]);
2026 			if (err != 0) {
2027 				device_printf(dev, "Unable to create RX buffer DMA map\n");
2028 				goto fail;
2029 			}
2030 		}
2031 	}
2032 	return (0);
2033 
2034 fail:
2035 	iflib_rx_structures_free(ctx);
2036 	return (err);
2037 }
2038 
2039 /*
2040  * Internal service routines
2041  */
2042 
2043 struct rxq_refill_cb_arg {
2044 	int               error;
2045 	bus_dma_segment_t seg;
2046 	int               nseg;
2047 };
2048 
2049 static void
2050 _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
2051 {
2052 	struct rxq_refill_cb_arg *cb_arg = arg;
2053 
2054 	cb_arg->error = error;
2055 	cb_arg->seg = segs[0];
2056 	cb_arg->nseg = nseg;
2057 }
2058 
2059 /**
2060  * iflib_fl_refill - refill an rxq free-buffer list
2061  * @ctx: the iflib context
2062  * @fl: the free list to refill
2063  * @count: the number of new buffers to allocate
2064  *
2065  * (Re)populate an rxq free-buffer list with up to @count new packet buffers.
2066  * The caller must assure that @count does not exceed the queue's capacity
2067  * minus one (since we always leave a descriptor unavailable).
2068  */
2069 static uint8_t
2070 iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
2071 {
2072 	struct if_rxd_update iru;
2073 	struct rxq_refill_cb_arg cb_arg;
2074 	struct mbuf *m;
2075 	caddr_t cl, *sd_cl;
2076 	struct mbuf **sd_m;
2077 	bus_dmamap_t *sd_map;
2078 	bus_addr_t bus_addr, *sd_ba;
2079 	int err, frag_idx, i, idx, n, pidx;
2080 	qidx_t credits;
2081 
2082 	MPASS(count <= fl->ifl_size - fl->ifl_credits - 1);
2083 
2084 	sd_m = fl->ifl_sds.ifsd_m;
2085 	sd_map = fl->ifl_sds.ifsd_map;
2086 	sd_cl = fl->ifl_sds.ifsd_cl;
2087 	sd_ba = fl->ifl_sds.ifsd_ba;
2088 	pidx = fl->ifl_pidx;
2089 	idx = pidx;
2090 	frag_idx = fl->ifl_fragidx;
2091 	credits = fl->ifl_credits;
2092 
2093 	i = 0;
2094 	n = count;
2095 	MPASS(n > 0);
2096 	MPASS(credits + n <= fl->ifl_size);
2097 
2098 	if (pidx < fl->ifl_cidx)
2099 		MPASS(pidx + n <= fl->ifl_cidx);
2100 	if (pidx == fl->ifl_cidx && (credits < fl->ifl_size))
2101 		MPASS(fl->ifl_gen == 0);
2102 	if (pidx > fl->ifl_cidx)
2103 		MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx);
2104 
2105 	DBG_COUNTER_INC(fl_refills);
2106 	if (n > 8)
2107 		DBG_COUNTER_INC(fl_refills_large);
2108 	iru_init(&iru, fl->ifl_rxq, fl->ifl_id);
2109 	while (n-- > 0) {
2110 		/*
2111 		 * We allocate an uninitialized mbuf + cluster, mbuf is
2112 		 * initialized after rx.
2113 		 *
2114 		 * If the cluster is still set then we know a minimum sized
2115 		 * packet was received
2116 		 */
2117 		bit_ffc_at(fl->ifl_rx_bitmap, frag_idx, fl->ifl_size,
2118 		    &frag_idx);
2119 		if (frag_idx < 0)
2120 			bit_ffc(fl->ifl_rx_bitmap, fl->ifl_size, &frag_idx);
2121 		MPASS(frag_idx >= 0);
2122 		if ((cl = sd_cl[frag_idx]) == NULL) {
2123 			cl = uma_zalloc(fl->ifl_zone, M_NOWAIT);
2124 			if (__predict_false(cl == NULL))
2125 				break;
2126 
2127 			cb_arg.error = 0;
2128 			MPASS(sd_map != NULL);
2129 			err = bus_dmamap_load(fl->ifl_buf_tag, sd_map[frag_idx],
2130 			    cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg,
2131 			    BUS_DMA_NOWAIT);
2132 			if (__predict_false(err != 0 || cb_arg.error)) {
2133 				uma_zfree(fl->ifl_zone, cl);
2134 				break;
2135 			}
2136 
2137 			sd_ba[frag_idx] = bus_addr = cb_arg.seg.ds_addr;
2138 			sd_cl[frag_idx] = cl;
2139 #if MEMORY_LOGGING
2140 			fl->ifl_cl_enqueued++;
2141 #endif
2142 		} else {
2143 			bus_addr = sd_ba[frag_idx];
2144 		}
2145 		bus_dmamap_sync(fl->ifl_buf_tag, sd_map[frag_idx],
2146 		    BUS_DMASYNC_PREREAD);
2147 
2148 		if (sd_m[frag_idx] == NULL) {
2149 			m = m_gethdr_raw(M_NOWAIT, 0);
2150 			if (__predict_false(m == NULL))
2151 				break;
2152 			sd_m[frag_idx] = m;
2153 		}
2154 		bit_set(fl->ifl_rx_bitmap, frag_idx);
2155 #if MEMORY_LOGGING
2156 		fl->ifl_m_enqueued++;
2157 #endif
2158 
2159 		DBG_COUNTER_INC(rx_allocs);
2160 		fl->ifl_rxd_idxs[i] = frag_idx;
2161 		fl->ifl_bus_addrs[i] = bus_addr;
2162 		credits++;
2163 		i++;
2164 		MPASS(credits <= fl->ifl_size);
2165 		if (++idx == fl->ifl_size) {
2166 #ifdef INVARIANTS
2167 			fl->ifl_gen = 1;
2168 #endif
2169 			idx = 0;
2170 		}
2171 		if (n == 0 || i == IFLIB_MAX_RX_REFRESH) {
2172 			iru.iru_pidx = pidx;
2173 			iru.iru_count = i;
2174 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2175 			fl->ifl_pidx = idx;
2176 			fl->ifl_credits = credits;
2177 			pidx = idx;
2178 			i = 0;
2179 		}
2180 	}
2181 
2182 	if (n < count - 1) {
2183 		if (i != 0) {
2184 			iru.iru_pidx = pidx;
2185 			iru.iru_count = i;
2186 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2187 			fl->ifl_pidx = idx;
2188 			fl->ifl_credits = credits;
2189 		}
2190 		DBG_COUNTER_INC(rxd_flush);
2191 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2192 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2193 		ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id,
2194 		    fl->ifl_id, fl->ifl_pidx);
2195 		if (__predict_true(bit_test(fl->ifl_rx_bitmap, frag_idx))) {
2196 			fl->ifl_fragidx = frag_idx + 1;
2197 			if (fl->ifl_fragidx == fl->ifl_size)
2198 				fl->ifl_fragidx = 0;
2199 		} else {
2200 			fl->ifl_fragidx = frag_idx;
2201 		}
2202 	}
2203 
2204 	return (n == -1 ? 0 : IFLIB_RXEOF_EMPTY);
2205 }
2206 
2207 static inline uint8_t
2208 iflib_fl_refill_all(if_ctx_t ctx, iflib_fl_t fl)
2209 {
2210 	/*
2211 	 * We leave an unused descriptor to avoid pidx to catch up with cidx.
2212 	 * This is important as it confuses most NICs. For instance,
2213 	 * Intel NICs have (per receive ring) RDH and RDT registers, where
2214 	 * RDH points to the next receive descriptor to be used by the NIC,
2215 	 * and RDT for the next receive descriptor to be published by the
2216 	 * driver to the NIC (RDT - 1 is thus the last valid one).
2217 	 * The condition RDH == RDT means no descriptors are available to
2218 	 * the NIC, and thus it would be ambiguous if it also meant that
2219 	 * all the descriptors are available to the NIC.
2220 	 */
2221 	int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1;
2222 #ifdef INVARIANTS
2223 	int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1;
2224 #endif
2225 
2226 	MPASS(fl->ifl_credits <= fl->ifl_size);
2227 	MPASS(reclaimable == delta);
2228 
2229 	if (reclaimable > 0)
2230 		return (iflib_fl_refill(ctx, fl, reclaimable));
2231 	return (0);
2232 }
2233 
2234 uint8_t
2235 iflib_in_detach(if_ctx_t ctx)
2236 {
2237 	bool in_detach;
2238 
2239 	STATE_LOCK(ctx);
2240 	in_detach = !!(ctx->ifc_flags & IFC_IN_DETACH);
2241 	STATE_UNLOCK(ctx);
2242 	return (in_detach);
2243 }
2244 
2245 static void
2246 iflib_fl_bufs_free(iflib_fl_t fl)
2247 {
2248 	iflib_dma_info_t idi = fl->ifl_ifdi;
2249 	bus_dmamap_t sd_map;
2250 	uint32_t i;
2251 
2252 	for (i = 0; i < fl->ifl_size; i++) {
2253 		struct mbuf **sd_m = &fl->ifl_sds.ifsd_m[i];
2254 		caddr_t *sd_cl = &fl->ifl_sds.ifsd_cl[i];
2255 
2256 		if (*sd_cl != NULL) {
2257 			sd_map = fl->ifl_sds.ifsd_map[i];
2258 			bus_dmamap_sync(fl->ifl_buf_tag, sd_map,
2259 			    BUS_DMASYNC_POSTREAD);
2260 			bus_dmamap_unload(fl->ifl_buf_tag, sd_map);
2261 			uma_zfree(fl->ifl_zone, *sd_cl);
2262 			*sd_cl = NULL;
2263 			if (*sd_m != NULL) {
2264 				m_init(*sd_m, M_NOWAIT, MT_DATA, 0);
2265 				m_free_raw(*sd_m);
2266 				*sd_m = NULL;
2267 			}
2268 		} else {
2269 			MPASS(*sd_m == NULL);
2270 		}
2271 #if MEMORY_LOGGING
2272 		fl->ifl_m_dequeued++;
2273 		fl->ifl_cl_dequeued++;
2274 #endif
2275 	}
2276 #ifdef INVARIANTS
2277 	for (i = 0; i < fl->ifl_size; i++) {
2278 		MPASS(fl->ifl_sds.ifsd_cl[i] == NULL);
2279 		MPASS(fl->ifl_sds.ifsd_m[i] == NULL);
2280 	}
2281 #endif
2282 	/*
2283 	 * Reset free list values
2284 	 */
2285 	fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = fl->ifl_fragidx = 0;
2286 	bzero(idi->idi_vaddr, idi->idi_size);
2287 }
2288 
2289 /*********************************************************************
2290  *
2291  *  Initialize a free list and its buffers.
2292  *
2293  **********************************************************************/
2294 static int
2295 iflib_fl_setup(iflib_fl_t fl)
2296 {
2297 	iflib_rxq_t rxq = fl->ifl_rxq;
2298 	if_ctx_t ctx = rxq->ifr_ctx;
2299 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2300 	int qidx;
2301 
2302 	bit_nclear(fl->ifl_rx_bitmap, 0, fl->ifl_size - 1);
2303 	/*
2304 	** Free current RX buffer structs and their mbufs
2305 	*/
2306 	iflib_fl_bufs_free(fl);
2307 	/* Now replenish the mbufs */
2308 	MPASS(fl->ifl_credits == 0);
2309 	qidx = rxq->ifr_fl_offset + fl->ifl_id;
2310 	if (scctx->isc_rxd_buf_size[qidx] != 0)
2311 		fl->ifl_buf_size = scctx->isc_rxd_buf_size[qidx];
2312 	else
2313 		fl->ifl_buf_size = ctx->ifc_rx_mbuf_sz;
2314 	/*
2315 	 * ifl_buf_size may be a driver-supplied value, so pull it up
2316 	 * to the selected mbuf size.
2317 	 */
2318 	fl->ifl_buf_size = iflib_get_mbuf_size_for(fl->ifl_buf_size);
2319 	if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size)
2320 		ctx->ifc_max_fl_buf_size = fl->ifl_buf_size;
2321 	fl->ifl_cltype = m_gettype(fl->ifl_buf_size);
2322 	fl->ifl_zone = m_getzone(fl->ifl_buf_size);
2323 
2324 	/*
2325 	 * Avoid pre-allocating zillions of clusters to an idle card
2326 	 * potentially speeding up attach. In any case make sure
2327 	 * to leave a descriptor unavailable. See the comment in
2328 	 * iflib_fl_refill_all().
2329 	 */
2330 	MPASS(fl->ifl_size > 0);
2331 	(void)iflib_fl_refill(ctx, fl, min(128, fl->ifl_size - 1));
2332 	if (min(128, fl->ifl_size - 1) != fl->ifl_credits)
2333 		return (ENOBUFS);
2334 	/*
2335 	 * handle failure
2336 	 */
2337 	MPASS(rxq != NULL);
2338 	MPASS(fl->ifl_ifdi != NULL);
2339 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2340 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2341 	return (0);
2342 }
2343 
2344 /*********************************************************************
2345  *
2346  *  Free receive ring data structures
2347  *
2348  **********************************************************************/
2349 static void
2350 iflib_rx_sds_free(iflib_rxq_t rxq)
2351 {
2352 	iflib_fl_t fl;
2353 	int i, j;
2354 
2355 	if (rxq->ifr_fl != NULL) {
2356 		for (i = 0; i < rxq->ifr_nfl; i++) {
2357 			fl = &rxq->ifr_fl[i];
2358 			if (fl->ifl_buf_tag != NULL) {
2359 				if (fl->ifl_sds.ifsd_map != NULL) {
2360 					for (j = 0; j < fl->ifl_size; j++) {
2361 						bus_dmamap_sync(
2362 						    fl->ifl_buf_tag,
2363 						    fl->ifl_sds.ifsd_map[j],
2364 						    BUS_DMASYNC_POSTREAD);
2365 						bus_dmamap_unload(
2366 						    fl->ifl_buf_tag,
2367 						    fl->ifl_sds.ifsd_map[j]);
2368 						bus_dmamap_destroy(
2369 						    fl->ifl_buf_tag,
2370 						    fl->ifl_sds.ifsd_map[j]);
2371 					}
2372 				}
2373 				bus_dma_tag_destroy(fl->ifl_buf_tag);
2374 				fl->ifl_buf_tag = NULL;
2375 			}
2376 			free(fl->ifl_sds.ifsd_m, M_IFLIB);
2377 			free(fl->ifl_sds.ifsd_cl, M_IFLIB);
2378 			free(fl->ifl_sds.ifsd_ba, M_IFLIB);
2379 			free(fl->ifl_sds.ifsd_map, M_IFLIB);
2380 			free(fl->ifl_rx_bitmap, M_IFLIB);
2381 			fl->ifl_sds.ifsd_m = NULL;
2382 			fl->ifl_sds.ifsd_cl = NULL;
2383 			fl->ifl_sds.ifsd_ba = NULL;
2384 			fl->ifl_sds.ifsd_map = NULL;
2385 			fl->ifl_rx_bitmap = NULL;
2386 		}
2387 		free(rxq->ifr_fl, M_IFLIB);
2388 		rxq->ifr_fl = NULL;
2389 		free(rxq->ifr_ifdi, M_IFLIB);
2390 		rxq->ifr_ifdi = NULL;
2391 		rxq->ifr_cq_cidx = 0;
2392 	}
2393 }
2394 
2395 /*
2396  * Timer routine
2397  */
2398 static void
2399 iflib_timer(void *arg)
2400 {
2401 	iflib_txq_t txq = arg;
2402 	if_ctx_t ctx = txq->ift_ctx;
2403 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2404 	uint64_t this_tick = ticks;
2405 
2406 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
2407 		return;
2408 
2409 	/*
2410 	** Check on the state of the TX queue(s), this
2411 	** can be done without the lock because its RO
2412 	** and the HUNG state will be static if set.
2413 	*/
2414 	if (this_tick - txq->ift_last_timer_tick >= iflib_timer_default) {
2415 		txq->ift_last_timer_tick = this_tick;
2416 		IFDI_TIMER(ctx, txq->ift_id);
2417 		if ((txq->ift_qstatus == IFLIB_QUEUE_HUNG) &&
2418 		    ((txq->ift_cleaned_prev == txq->ift_cleaned) ||
2419 		     (sctx->isc_pause_frames == 0)))
2420 			goto hung;
2421 
2422 		if (txq->ift_qstatus != IFLIB_QUEUE_IDLE &&
2423 		    ifmp_ring_is_stalled(txq->ift_br)) {
2424 			KASSERT(ctx->ifc_link_state == LINK_STATE_UP,
2425 			    ("queue can't be marked as hung if interface is down"));
2426 			txq->ift_qstatus = IFLIB_QUEUE_HUNG;
2427 		}
2428 		txq->ift_cleaned_prev = txq->ift_cleaned;
2429 	}
2430 	/* handle any laggards */
2431 	if (txq->ift_db_pending)
2432 		GROUPTASK_ENQUEUE(&txq->ift_task);
2433 
2434 	sctx->isc_pause_frames = 0;
2435 	if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)
2436 		callout_reset_on(&txq->ift_timer, iflib_timer_default, iflib_timer,
2437 		    txq, txq->ift_timer.c_cpu);
2438 	return;
2439 
2440  hung:
2441 	device_printf(ctx->ifc_dev,
2442 	    "Watchdog timeout (TX: %d desc avail: %d pidx: %d) -- resetting\n",
2443 	    txq->ift_id, TXQ_AVAIL(txq), txq->ift_pidx);
2444 	STATE_LOCK(ctx);
2445 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2446 	ctx->ifc_flags |= (IFC_DO_WATCHDOG|IFC_DO_RESET);
2447 	iflib_admin_intr_deferred(ctx);
2448 	STATE_UNLOCK(ctx);
2449 }
2450 
2451 static uint16_t
2452 iflib_get_mbuf_size_for(unsigned int size)
2453 {
2454 
2455 	if (size <= MCLBYTES)
2456 		return (MCLBYTES);
2457 	else
2458 		return (MJUMPAGESIZE);
2459 }
2460 
2461 static void
2462 iflib_calc_rx_mbuf_sz(if_ctx_t ctx)
2463 {
2464 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2465 
2466 	/*
2467 	 * XXX don't set the max_frame_size to larger
2468 	 * than the hardware can handle
2469 	 */
2470 	ctx->ifc_rx_mbuf_sz =
2471 	    iflib_get_mbuf_size_for(sctx->isc_max_frame_size);
2472 }
2473 
2474 uint32_t
2475 iflib_get_rx_mbuf_sz(if_ctx_t ctx)
2476 {
2477 
2478 	return (ctx->ifc_rx_mbuf_sz);
2479 }
2480 
2481 static void
2482 iflib_init_locked(if_ctx_t ctx)
2483 {
2484 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2485 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2486 	if_t ifp = ctx->ifc_ifp;
2487 	iflib_fl_t fl;
2488 	iflib_txq_t txq;
2489 	iflib_rxq_t rxq;
2490 	int i, j, tx_ip_csum_flags, tx_ip6_csum_flags;
2491 
2492 	if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2493 	IFDI_INTR_DISABLE(ctx);
2494 
2495 	/*
2496 	 * See iflib_stop(). Useful in case iflib_init_locked() is
2497 	 * called without first calling iflib_stop().
2498 	 */
2499 	netmap_disable_all_rings(ifp);
2500 
2501 	tx_ip_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP);
2502 	tx_ip6_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | CSUM_IP6_SCTP);
2503 	/* Set hardware offload abilities */
2504 	if_clearhwassist(ifp);
2505 	if (if_getcapenable(ifp) & IFCAP_TXCSUM)
2506 		if_sethwassistbits(ifp, tx_ip_csum_flags, 0);
2507 	if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6)
2508 		if_sethwassistbits(ifp,  tx_ip6_csum_flags, 0);
2509 	if (if_getcapenable(ifp) & IFCAP_TSO4)
2510 		if_sethwassistbits(ifp, CSUM_IP_TSO, 0);
2511 	if (if_getcapenable(ifp) & IFCAP_TSO6)
2512 		if_sethwassistbits(ifp, CSUM_IP6_TSO, 0);
2513 
2514 	for (i = 0, txq = ctx->ifc_txqs; i < sctx->isc_ntxqsets; i++, txq++) {
2515 		CALLOUT_LOCK(txq);
2516 		callout_stop(&txq->ift_timer);
2517 #ifdef DEV_NETMAP
2518 		callout_stop(&txq->ift_netmap_timer);
2519 #endif /* DEV_NETMAP */
2520 		CALLOUT_UNLOCK(txq);
2521 		(void)iflib_netmap_txq_init(ctx, txq);
2522 	}
2523 
2524 	/*
2525 	 * Calculate a suitable Rx mbuf size prior to calling IFDI_INIT, so
2526 	 * that drivers can use the value when setting up the hardware receive
2527 	 * buffers.
2528 	 */
2529 	iflib_calc_rx_mbuf_sz(ctx);
2530 
2531 #ifdef INVARIANTS
2532 	i = if_getdrvflags(ifp);
2533 #endif
2534 	IFDI_INIT(ctx);
2535 	MPASS(if_getdrvflags(ifp) == i);
2536 	for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
2537 		if (iflib_netmap_rxq_init(ctx, rxq) > 0) {
2538 			/* This rxq is in netmap mode. Skip normal init. */
2539 			continue;
2540 		}
2541 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
2542 			if (iflib_fl_setup(fl)) {
2543 				device_printf(ctx->ifc_dev,
2544 				    "setting up free list %d failed - "
2545 				    "check cluster settings\n", j);
2546 				goto done;
2547 			}
2548 		}
2549 	}
2550 done:
2551 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
2552 	IFDI_INTR_ENABLE(ctx);
2553 	txq = ctx->ifc_txqs;
2554 	for (i = 0; i < sctx->isc_ntxqsets; i++, txq++)
2555 		callout_reset_on(&txq->ift_timer, iflib_timer_default, iflib_timer, txq,
2556 			txq->ift_timer.c_cpu);
2557 
2558         /* Re-enable txsync/rxsync. */
2559 	netmap_enable_all_rings(ifp);
2560 }
2561 
2562 static int
2563 iflib_media_change(if_t ifp)
2564 {
2565 	if_ctx_t ctx = if_getsoftc(ifp);
2566 	int err;
2567 
2568 	CTX_LOCK(ctx);
2569 	if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0)
2570 		iflib_if_init_locked(ctx);
2571 	CTX_UNLOCK(ctx);
2572 	return (err);
2573 }
2574 
2575 static void
2576 iflib_media_status(if_t ifp, struct ifmediareq *ifmr)
2577 {
2578 	if_ctx_t ctx = if_getsoftc(ifp);
2579 
2580 	CTX_LOCK(ctx);
2581 	IFDI_UPDATE_ADMIN_STATUS(ctx);
2582 	IFDI_MEDIA_STATUS(ctx, ifmr);
2583 	CTX_UNLOCK(ctx);
2584 }
2585 
2586 void
2587 iflib_stop(if_ctx_t ctx)
2588 {
2589 	iflib_txq_t txq = ctx->ifc_txqs;
2590 	iflib_rxq_t rxq = ctx->ifc_rxqs;
2591 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2592 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2593 	iflib_dma_info_t di;
2594 	iflib_fl_t fl;
2595 	int i, j;
2596 
2597 	/* Tell the stack that the interface is no longer active */
2598 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2599 
2600 	IFDI_INTR_DISABLE(ctx);
2601 	DELAY(1000);
2602 	IFDI_STOP(ctx);
2603 	DELAY(1000);
2604 
2605 	/*
2606 	 * Stop any pending txsync/rxsync and prevent new ones
2607 	 * form starting. Processes blocked in poll() will get
2608 	 * POLLERR.
2609 	 */
2610 	netmap_disable_all_rings(ctx->ifc_ifp);
2611 
2612 	iflib_debug_reset();
2613 	/* Wait for current tx queue users to exit to disarm watchdog timer. */
2614 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) {
2615 		/* make sure all transmitters have completed before proceeding XXX */
2616 
2617 		CALLOUT_LOCK(txq);
2618 		callout_stop(&txq->ift_timer);
2619 #ifdef DEV_NETMAP
2620 		callout_stop(&txq->ift_netmap_timer);
2621 #endif /* DEV_NETMAP */
2622 		CALLOUT_UNLOCK(txq);
2623 
2624 		/* clean any enqueued buffers */
2625 		iflib_ifmp_purge(txq);
2626 		/* Free any existing tx buffers. */
2627 		for (j = 0; j < txq->ift_size; j++) {
2628 			iflib_txsd_free(ctx, txq, j);
2629 		}
2630 		txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0;
2631 		txq->ift_in_use = txq->ift_gen = txq->ift_no_desc_avail = 0;
2632 		if (sctx->isc_flags & IFLIB_PRESERVE_TX_INDICES)
2633 			txq->ift_cidx = txq->ift_pidx;
2634 		else
2635 			txq->ift_cidx = txq->ift_pidx = 0;
2636 
2637 		txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0;
2638 		txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0;
2639 		txq->ift_pullups = 0;
2640 		ifmp_ring_reset_stats(txq->ift_br);
2641 		for (j = 0, di = txq->ift_ifdi; j < sctx->isc_ntxqs; j++, di++)
2642 			bzero((void *)di->idi_vaddr, di->idi_size);
2643 	}
2644 	for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) {
2645 		gtaskqueue_drain(rxq->ifr_task.gt_taskqueue,
2646 		    &rxq->ifr_task.gt_task);
2647 
2648 		rxq->ifr_cq_cidx = 0;
2649 		for (j = 0, di = rxq->ifr_ifdi; j < sctx->isc_nrxqs; j++, di++)
2650 			bzero((void *)di->idi_vaddr, di->idi_size);
2651 		/* also resets the free lists pidx/cidx */
2652 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
2653 			iflib_fl_bufs_free(fl);
2654 	}
2655 }
2656 
2657 static inline caddr_t
2658 calc_next_rxd(iflib_fl_t fl, int cidx)
2659 {
2660 	qidx_t size;
2661 	int nrxd;
2662 	caddr_t start, end, cur, next;
2663 
2664 	nrxd = fl->ifl_size;
2665 	size = fl->ifl_rxd_size;
2666 	start = fl->ifl_ifdi->idi_vaddr;
2667 
2668 	if (__predict_false(size == 0))
2669 		return (start);
2670 	cur = start + size*cidx;
2671 	end = start + size*nrxd;
2672 	next = CACHE_PTR_NEXT(cur);
2673 	return (next < end ? next : start);
2674 }
2675 
2676 static inline void
2677 prefetch_pkts(iflib_fl_t fl, int cidx)
2678 {
2679 	int nextptr;
2680 	int nrxd = fl->ifl_size;
2681 	caddr_t next_rxd;
2682 
2683 	nextptr = (cidx + CACHE_PTR_INCREMENT) & (nrxd-1);
2684 	prefetch(&fl->ifl_sds.ifsd_m[nextptr]);
2685 	prefetch(&fl->ifl_sds.ifsd_cl[nextptr]);
2686 	next_rxd = calc_next_rxd(fl, cidx);
2687 	prefetch(next_rxd);
2688 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 1) & (nrxd-1)]);
2689 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 2) & (nrxd-1)]);
2690 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 3) & (nrxd-1)]);
2691 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 4) & (nrxd-1)]);
2692 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 1) & (nrxd-1)]);
2693 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 2) & (nrxd-1)]);
2694 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 3) & (nrxd-1)]);
2695 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 4) & (nrxd-1)]);
2696 }
2697 
2698 static struct mbuf *
2699 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, bool unload, if_rxsd_t sd,
2700     int *pf_rv, if_rxd_info_t ri)
2701 {
2702 	bus_dmamap_t map;
2703 	iflib_fl_t fl;
2704 	caddr_t payload;
2705 	struct mbuf *m;
2706 	int flid, cidx, len, next;
2707 
2708 	map = NULL;
2709 	flid = irf->irf_flid;
2710 	cidx = irf->irf_idx;
2711 	fl = &rxq->ifr_fl[flid];
2712 	sd->ifsd_fl = fl;
2713 	m = fl->ifl_sds.ifsd_m[cidx];
2714 	sd->ifsd_cl = &fl->ifl_sds.ifsd_cl[cidx];
2715 	fl->ifl_credits--;
2716 #if MEMORY_LOGGING
2717 	fl->ifl_m_dequeued++;
2718 #endif
2719 	if (rxq->ifr_ctx->ifc_flags & IFC_PREFETCH)
2720 		prefetch_pkts(fl, cidx);
2721 	next = (cidx + CACHE_PTR_INCREMENT) & (fl->ifl_size-1);
2722 	prefetch(&fl->ifl_sds.ifsd_map[next]);
2723 	map = fl->ifl_sds.ifsd_map[cidx];
2724 
2725 	bus_dmamap_sync(fl->ifl_buf_tag, map, BUS_DMASYNC_POSTREAD);
2726 
2727 	if (rxq->pfil != NULL && PFIL_HOOKED_IN(rxq->pfil) && pf_rv != NULL &&
2728 	    irf->irf_len != 0) {
2729 		payload  = *sd->ifsd_cl;
2730 		payload +=  ri->iri_pad;
2731 		len = ri->iri_len - ri->iri_pad;
2732 		*pf_rv = pfil_run_hooks(rxq->pfil, payload, ri->iri_ifp,
2733 		    len | PFIL_MEMPTR | PFIL_IN, NULL);
2734 		switch (*pf_rv) {
2735 		case PFIL_DROPPED:
2736 		case PFIL_CONSUMED:
2737 			/*
2738 			 * The filter ate it.  Everything is recycled.
2739 			 */
2740 			m = NULL;
2741 			unload = 0;
2742 			break;
2743 		case PFIL_REALLOCED:
2744 			/*
2745 			 * The filter copied it.  Everything is recycled.
2746 			 */
2747 			m = pfil_mem2mbuf(payload);
2748 			unload = 0;
2749 			break;
2750 		case PFIL_PASS:
2751 			/*
2752 			 * Filter said it was OK, so receive like
2753 			 * normal
2754 			 */
2755 			fl->ifl_sds.ifsd_m[cidx] = NULL;
2756 			break;
2757 		default:
2758 			MPASS(0);
2759 		}
2760 	} else {
2761 		fl->ifl_sds.ifsd_m[cidx] = NULL;
2762 		if (pf_rv != NULL)
2763 			*pf_rv = PFIL_PASS;
2764 	}
2765 
2766 	if (unload && irf->irf_len != 0)
2767 		bus_dmamap_unload(fl->ifl_buf_tag, map);
2768 	fl->ifl_cidx = (fl->ifl_cidx + 1) & (fl->ifl_size-1);
2769 	if (__predict_false(fl->ifl_cidx == 0))
2770 		fl->ifl_gen = 0;
2771 	bit_clear(fl->ifl_rx_bitmap, cidx);
2772 	return (m);
2773 }
2774 
2775 static struct mbuf *
2776 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri, if_rxsd_t sd, int *pf_rv)
2777 {
2778 	struct mbuf *m, *mh, *mt;
2779 	caddr_t cl;
2780 	int  *pf_rv_ptr, flags, i, padlen;
2781 	bool consumed;
2782 
2783 	i = 0;
2784 	mh = NULL;
2785 	consumed = false;
2786 	*pf_rv = PFIL_PASS;
2787 	pf_rv_ptr = pf_rv;
2788 	do {
2789 		m = rxd_frag_to_sd(rxq, &ri->iri_frags[i], !consumed, sd,
2790 		    pf_rv_ptr, ri);
2791 
2792 		MPASS(*sd->ifsd_cl != NULL);
2793 
2794 		/*
2795 		 * Exclude zero-length frags & frags from
2796 		 * packets the filter has consumed or dropped
2797 		 */
2798 		if (ri->iri_frags[i].irf_len == 0 || consumed ||
2799 		    *pf_rv == PFIL_CONSUMED || *pf_rv == PFIL_DROPPED) {
2800 			if (mh == NULL) {
2801 				/* everything saved here */
2802 				consumed = true;
2803 				pf_rv_ptr = NULL;
2804 				continue;
2805 			}
2806 			/* XXX we can save the cluster here, but not the mbuf */
2807 			m_init(m, M_NOWAIT, MT_DATA, 0);
2808 			m_free(m);
2809 			continue;
2810 		}
2811 		if (mh == NULL) {
2812 			flags = M_PKTHDR|M_EXT;
2813 			mh = mt = m;
2814 			padlen = ri->iri_pad;
2815 		} else {
2816 			flags = M_EXT;
2817 			mt->m_next = m;
2818 			mt = m;
2819 			/* assuming padding is only on the first fragment */
2820 			padlen = 0;
2821 		}
2822 		cl = *sd->ifsd_cl;
2823 		*sd->ifsd_cl = NULL;
2824 
2825 		/* Can these two be made one ? */
2826 		m_init(m, M_NOWAIT, MT_DATA, flags);
2827 		m_cljset(m, cl, sd->ifsd_fl->ifl_cltype);
2828 		/*
2829 		 * These must follow m_init and m_cljset
2830 		 */
2831 		m->m_data += padlen;
2832 		ri->iri_len -= padlen;
2833 		m->m_len = ri->iri_frags[i].irf_len;
2834 	} while (++i < ri->iri_nfrags);
2835 
2836 	return (mh);
2837 }
2838 
2839 /*
2840  * Process one software descriptor
2841  */
2842 static struct mbuf *
2843 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri)
2844 {
2845 	struct if_rxsd sd;
2846 	struct mbuf *m;
2847 	int pf_rv;
2848 
2849 	/* should I merge this back in now that the two paths are basically duplicated? */
2850 	if (ri->iri_nfrags == 1 &&
2851 	    ri->iri_frags[0].irf_len != 0 &&
2852 	    ri->iri_frags[0].irf_len <= MIN(IFLIB_RX_COPY_THRESH, MHLEN)) {
2853 		m = rxd_frag_to_sd(rxq, &ri->iri_frags[0], false, &sd,
2854 		    &pf_rv, ri);
2855 		if (pf_rv != PFIL_PASS && pf_rv != PFIL_REALLOCED)
2856 			return (m);
2857 		if (pf_rv == PFIL_PASS) {
2858 			m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR);
2859 #ifndef __NO_STRICT_ALIGNMENT
2860 			if (!IP_ALIGNED(m) && ri->iri_pad == 0)
2861 				m->m_data += 2;
2862 #endif
2863 			memcpy(m->m_data, *sd.ifsd_cl, ri->iri_len);
2864 			m->m_len = ri->iri_frags[0].irf_len;
2865 			m->m_data += ri->iri_pad;
2866 			ri->iri_len -= ri->iri_pad;
2867 		}
2868 	} else {
2869 		m = assemble_segments(rxq, ri, &sd, &pf_rv);
2870 		if (m == NULL)
2871 			return (NULL);
2872 		if (pf_rv != PFIL_PASS && pf_rv != PFIL_REALLOCED)
2873 			return (m);
2874 	}
2875 	m->m_pkthdr.len = ri->iri_len;
2876 	m->m_pkthdr.rcvif = ri->iri_ifp;
2877 	m->m_flags |= ri->iri_flags;
2878 	m->m_pkthdr.ether_vtag = ri->iri_vtag;
2879 	m->m_pkthdr.flowid = ri->iri_flowid;
2880 	M_HASHTYPE_SET(m, ri->iri_rsstype);
2881 	m->m_pkthdr.csum_flags = ri->iri_csum_flags;
2882 	m->m_pkthdr.csum_data = ri->iri_csum_data;
2883 	return (m);
2884 }
2885 
2886 #if defined(INET6) || defined(INET)
2887 static void
2888 iflib_get_ip_forwarding(struct lro_ctrl *lc, bool *v4, bool *v6)
2889 {
2890 	CURVNET_SET(lc->ifp->if_vnet);
2891 #if defined(INET6)
2892 	*v6 = V_ip6_forwarding;
2893 #endif
2894 #if defined(INET)
2895 	*v4 = V_ipforwarding;
2896 #endif
2897 	CURVNET_RESTORE();
2898 }
2899 
2900 /*
2901  * Returns true if it's possible this packet could be LROed.
2902  * if it returns false, it is guaranteed that tcp_lro_rx()
2903  * would not return zero.
2904  */
2905 static bool
2906 iflib_check_lro_possible(struct mbuf *m, bool v4_forwarding, bool v6_forwarding)
2907 {
2908 	struct ether_header *eh;
2909 
2910 	eh = mtod(m, struct ether_header *);
2911 	switch (eh->ether_type) {
2912 #if defined(INET6)
2913 		case htons(ETHERTYPE_IPV6):
2914 			return (!v6_forwarding);
2915 #endif
2916 #if defined (INET)
2917 		case htons(ETHERTYPE_IP):
2918 			return (!v4_forwarding);
2919 #endif
2920 	}
2921 
2922 	return false;
2923 }
2924 #else
2925 static void
2926 iflib_get_ip_forwarding(struct lro_ctrl *lc __unused, bool *v4 __unused, bool *v6 __unused)
2927 {
2928 }
2929 #endif
2930 
2931 static void
2932 _task_fn_rx_watchdog(void *context)
2933 {
2934 	iflib_rxq_t rxq = context;
2935 
2936 	GROUPTASK_ENQUEUE(&rxq->ifr_task);
2937 }
2938 
2939 static uint8_t
2940 iflib_rxeof(iflib_rxq_t rxq, qidx_t budget)
2941 {
2942 	if_t ifp;
2943 	if_ctx_t ctx = rxq->ifr_ctx;
2944 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2945 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2946 	int avail, i;
2947 	qidx_t *cidxp;
2948 	struct if_rxd_info ri;
2949 	int err, budget_left, rx_bytes, rx_pkts;
2950 	iflib_fl_t fl;
2951 	int lro_enabled;
2952 	bool v4_forwarding, v6_forwarding, lro_possible;
2953 	uint8_t retval = 0;
2954 
2955 	/*
2956 	 * XXX early demux data packets so that if_input processing only handles
2957 	 * acks in interrupt context
2958 	 */
2959 	struct mbuf *m, *mh, *mt, *mf;
2960 
2961 	NET_EPOCH_ASSERT();
2962 
2963 	lro_possible = v4_forwarding = v6_forwarding = false;
2964 	ifp = ctx->ifc_ifp;
2965 	mh = mt = NULL;
2966 	MPASS(budget > 0);
2967 	rx_pkts	= rx_bytes = 0;
2968 	if (sctx->isc_flags & IFLIB_HAS_RXCQ)
2969 		cidxp = &rxq->ifr_cq_cidx;
2970 	else
2971 		cidxp = &rxq->ifr_fl[0].ifl_cidx;
2972 	if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) {
2973 		for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2974 			retval |= iflib_fl_refill_all(ctx, fl);
2975 		DBG_COUNTER_INC(rx_unavail);
2976 		return (retval);
2977 	}
2978 
2979 	/* pfil needs the vnet to be set */
2980 	CURVNET_SET_QUIET(ifp->if_vnet);
2981 	for (budget_left = budget; budget_left > 0 && avail > 0;) {
2982 		if (__predict_false(!CTX_ACTIVE(ctx))) {
2983 			DBG_COUNTER_INC(rx_ctx_inactive);
2984 			break;
2985 		}
2986 		/*
2987 		 * Reset client set fields to their default values
2988 		 */
2989 		rxd_info_zero(&ri);
2990 		ri.iri_qsidx = rxq->ifr_id;
2991 		ri.iri_cidx = *cidxp;
2992 		ri.iri_ifp = ifp;
2993 		ri.iri_frags = rxq->ifr_frags;
2994 		err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
2995 
2996 		if (err)
2997 			goto err;
2998 		rx_pkts += 1;
2999 		rx_bytes += ri.iri_len;
3000 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
3001 			*cidxp = ri.iri_cidx;
3002 			/* Update our consumer index */
3003 			/* XXX NB: shurd - check if this is still safe */
3004 			while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0])
3005 				rxq->ifr_cq_cidx -= scctx->isc_nrxd[0];
3006 			/* was this only a completion queue message? */
3007 			if (__predict_false(ri.iri_nfrags == 0))
3008 				continue;
3009 		}
3010 		MPASS(ri.iri_nfrags != 0);
3011 		MPASS(ri.iri_len != 0);
3012 
3013 		/* will advance the cidx on the corresponding free lists */
3014 		m = iflib_rxd_pkt_get(rxq, &ri);
3015 		avail--;
3016 		budget_left--;
3017 		if (avail == 0 && budget_left)
3018 			avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left);
3019 
3020 		if (__predict_false(m == NULL))
3021 			continue;
3022 
3023 		/* imm_pkt: -- cxgb */
3024 		if (mh == NULL)
3025 			mh = mt = m;
3026 		else {
3027 			mt->m_nextpkt = m;
3028 			mt = m;
3029 		}
3030 	}
3031 	CURVNET_RESTORE();
3032 	/* make sure that we can refill faster than drain */
3033 	for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
3034 		retval |= iflib_fl_refill_all(ctx, fl);
3035 
3036 	lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO);
3037 	if (lro_enabled)
3038 		iflib_get_ip_forwarding(&rxq->ifr_lc, &v4_forwarding, &v6_forwarding);
3039 	mt = mf = NULL;
3040 	while (mh != NULL) {
3041 		m = mh;
3042 		mh = mh->m_nextpkt;
3043 		m->m_nextpkt = NULL;
3044 #ifndef __NO_STRICT_ALIGNMENT
3045 		if (!IP_ALIGNED(m) && (m = iflib_fixup_rx(m)) == NULL)
3046 			continue;
3047 #endif
3048 #if defined(INET6) || defined(INET)
3049 		if (lro_enabled) {
3050 			if (!lro_possible) {
3051 				lro_possible = iflib_check_lro_possible(m, v4_forwarding, v6_forwarding);
3052 				if (lro_possible && mf != NULL) {
3053 					ifp->if_input(ifp, mf);
3054 					DBG_COUNTER_INC(rx_if_input);
3055 					mt = mf = NULL;
3056 				}
3057 			}
3058 			if ((m->m_pkthdr.csum_flags & (CSUM_L4_CALC|CSUM_L4_VALID)) ==
3059 			    (CSUM_L4_CALC|CSUM_L4_VALID)) {
3060 				if (lro_possible && tcp_lro_rx(&rxq->ifr_lc, m, 0) == 0)
3061 					continue;
3062 			}
3063 		}
3064 #endif
3065 		if (lro_possible) {
3066 			ifp->if_input(ifp, m);
3067 			DBG_COUNTER_INC(rx_if_input);
3068 			continue;
3069 		}
3070 
3071 		if (mf == NULL)
3072 			mf = m;
3073 		if (mt != NULL)
3074 			mt->m_nextpkt = m;
3075 		mt = m;
3076 	}
3077 	if (mf != NULL) {
3078 		ifp->if_input(ifp, mf);
3079 		DBG_COUNTER_INC(rx_if_input);
3080 	}
3081 
3082 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
3083 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
3084 
3085 	/*
3086 	 * Flush any outstanding LRO work
3087 	 */
3088 #if defined(INET6) || defined(INET)
3089 	tcp_lro_flush_all(&rxq->ifr_lc);
3090 #endif
3091 	if (avail != 0 || iflib_rxd_avail(ctx, rxq, *cidxp, 1) != 0)
3092 		retval |= IFLIB_RXEOF_MORE;
3093 	return (retval);
3094 err:
3095 	STATE_LOCK(ctx);
3096 	ctx->ifc_flags |= IFC_DO_RESET;
3097 	iflib_admin_intr_deferred(ctx);
3098 	STATE_UNLOCK(ctx);
3099 	return (0);
3100 }
3101 
3102 #define TXD_NOTIFY_COUNT(txq) (((txq)->ift_size / (txq)->ift_update_freq)-1)
3103 static inline qidx_t
3104 txq_max_db_deferred(iflib_txq_t txq, qidx_t in_use)
3105 {
3106 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
3107 	qidx_t minthresh = txq->ift_size / 8;
3108 	if (in_use > 4*minthresh)
3109 		return (notify_count);
3110 	if (in_use > 2*minthresh)
3111 		return (notify_count >> 1);
3112 	if (in_use > minthresh)
3113 		return (notify_count >> 3);
3114 	return (0);
3115 }
3116 
3117 static inline qidx_t
3118 txq_max_rs_deferred(iflib_txq_t txq)
3119 {
3120 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
3121 	qidx_t minthresh = txq->ift_size / 8;
3122 	if (txq->ift_in_use > 4*minthresh)
3123 		return (notify_count);
3124 	if (txq->ift_in_use > 2*minthresh)
3125 		return (notify_count >> 1);
3126 	if (txq->ift_in_use > minthresh)
3127 		return (notify_count >> 2);
3128 	return (2);
3129 }
3130 
3131 #define M_CSUM_FLAGS(m) ((m)->m_pkthdr.csum_flags)
3132 #define M_HAS_VLANTAG(m) (m->m_flags & M_VLANTAG)
3133 
3134 #define TXQ_MAX_DB_DEFERRED(txq, in_use) txq_max_db_deferred((txq), (in_use))
3135 #define TXQ_MAX_RS_DEFERRED(txq) txq_max_rs_deferred(txq)
3136 #define TXQ_MAX_DB_CONSUMED(size) (size >> 4)
3137 
3138 /* forward compatibility for cxgb */
3139 #define FIRST_QSET(ctx) 0
3140 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets)
3141 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets)
3142 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx))
3143 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments))
3144 
3145 /* XXX we should be setting this to something other than zero */
3146 #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh)
3147 #define	MAX_TX_DESC(ctx) MAX((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max, \
3148     (ctx)->ifc_softc_ctx.isc_tx_nsegments)
3149 
3150 static inline bool
3151 iflib_txd_db_check(iflib_txq_t txq, int ring)
3152 {
3153 	if_ctx_t ctx = txq->ift_ctx;
3154 	qidx_t dbval, max;
3155 
3156 	max = TXQ_MAX_DB_DEFERRED(txq, txq->ift_in_use);
3157 
3158 	/* force || threshold exceeded || at the edge of the ring */
3159 	if (ring || (txq->ift_db_pending >= max) || (TXQ_AVAIL(txq) <= MAX_TX_DESC(ctx) + 2)) {
3160 
3161 		/*
3162 		 * 'npending' is used if the card's doorbell is in terms of the number of descriptors
3163 		 * pending flush (BRCM). 'pidx' is used in cases where the card's doorbeel uses the
3164 		 * producer index explicitly (INTC).
3165 		 */
3166 		dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx;
3167 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3168 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
3169 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval);
3170 
3171 		/*
3172 		 * Absent bugs there are zero packets pending so reset pending counts to zero.
3173 		 */
3174 		txq->ift_db_pending = txq->ift_npending = 0;
3175 		return (true);
3176 	}
3177 	return (false);
3178 }
3179 
3180 #ifdef PKT_DEBUG
3181 static void
3182 print_pkt(if_pkt_info_t pi)
3183 {
3184 	printf("pi len:  %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n",
3185 	       pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx);
3186 	printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n",
3187 	       pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag);
3188 	printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n",
3189 	       pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto);
3190 }
3191 #endif
3192 
3193 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO)
3194 #define IS_TX_OFFLOAD4(pi) ((pi)->ipi_csum_flags & (CSUM_IP_TCP | CSUM_IP_TSO))
3195 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO)
3196 #define IS_TX_OFFLOAD6(pi) ((pi)->ipi_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_TSO))
3197 
3198 static int
3199 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp)
3200 {
3201 	if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx;
3202 	struct ether_vlan_header *eh;
3203 	struct mbuf *m;
3204 
3205 	m = *mp;
3206 	if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) &&
3207 	    M_WRITABLE(m) == 0) {
3208 		if ((m = m_dup(m, M_NOWAIT)) == NULL) {
3209 			return (ENOMEM);
3210 		} else {
3211 			m_freem(*mp);
3212 			DBG_COUNTER_INC(tx_frees);
3213 			*mp = m;
3214 		}
3215 	}
3216 
3217 	/*
3218 	 * Determine where frame payload starts.
3219 	 * Jump over vlan headers if already present,
3220 	 * helpful for QinQ too.
3221 	 */
3222 	if (__predict_false(m->m_len < sizeof(*eh))) {
3223 		txq->ift_pullups++;
3224 		if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL))
3225 			return (ENOMEM);
3226 	}
3227 	eh = mtod(m, struct ether_vlan_header *);
3228 	if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
3229 		pi->ipi_etype = ntohs(eh->evl_proto);
3230 		pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
3231 	} else {
3232 		pi->ipi_etype = ntohs(eh->evl_encap_proto);
3233 		pi->ipi_ehdrlen = ETHER_HDR_LEN;
3234 	}
3235 
3236 	switch (pi->ipi_etype) {
3237 #ifdef INET
3238 	case ETHERTYPE_IP:
3239 	{
3240 		struct mbuf *n;
3241 		struct ip *ip = NULL;
3242 		struct tcphdr *th = NULL;
3243 		int minthlen;
3244 
3245 		minthlen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip) + sizeof(*th));
3246 		if (__predict_false(m->m_len < minthlen)) {
3247 			/*
3248 			 * if this code bloat is causing too much of a hit
3249 			 * move it to a separate function and mark it noinline
3250 			 */
3251 			if (m->m_len == pi->ipi_ehdrlen) {
3252 				n = m->m_next;
3253 				MPASS(n);
3254 				if (n->m_len >= sizeof(*ip))  {
3255 					ip = (struct ip *)n->m_data;
3256 					if (n->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3257 						th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3258 				} else {
3259 					txq->ift_pullups++;
3260 					if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
3261 						return (ENOMEM);
3262 					ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3263 				}
3264 			} else {
3265 				txq->ift_pullups++;
3266 				if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
3267 					return (ENOMEM);
3268 				ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3269 				if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3270 					th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3271 			}
3272 		} else {
3273 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3274 			if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3275 				th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3276 		}
3277 		pi->ipi_ip_hlen = ip->ip_hl << 2;
3278 		pi->ipi_ipproto = ip->ip_p;
3279 		pi->ipi_flags |= IPI_TX_IPV4;
3280 
3281 		/* TCP checksum offload may require TCP header length */
3282 		if (IS_TX_OFFLOAD4(pi)) {
3283 			if (__predict_true(pi->ipi_ipproto == IPPROTO_TCP)) {
3284 				if (__predict_false(th == NULL)) {
3285 					txq->ift_pullups++;
3286 					if (__predict_false((m = m_pullup(m, (ip->ip_hl << 2) + sizeof(*th))) == NULL))
3287 						return (ENOMEM);
3288 					th = (struct tcphdr *)((caddr_t)ip + pi->ipi_ip_hlen);
3289 				}
3290 				pi->ipi_tcp_hflags = th->th_flags;
3291 				pi->ipi_tcp_hlen = th->th_off << 2;
3292 				pi->ipi_tcp_seq = th->th_seq;
3293 			}
3294 			if (IS_TSO4(pi)) {
3295 				if (__predict_false(ip->ip_p != IPPROTO_TCP))
3296 					return (ENXIO);
3297 				/*
3298 				 * TSO always requires hardware checksum offload.
3299 				 */
3300 				pi->ipi_csum_flags |= (CSUM_IP_TCP | CSUM_IP);
3301 				th->th_sum = in_pseudo(ip->ip_src.s_addr,
3302 						       ip->ip_dst.s_addr, htons(IPPROTO_TCP));
3303 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3304 				if (sctx->isc_flags & IFLIB_TSO_INIT_IP) {
3305 					ip->ip_sum = 0;
3306 					ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz);
3307 				}
3308 			}
3309 		}
3310 		if ((sctx->isc_flags & IFLIB_NEED_ZERO_CSUM) && (pi->ipi_csum_flags & CSUM_IP))
3311                        ip->ip_sum = 0;
3312 
3313 		break;
3314 	}
3315 #endif
3316 #ifdef INET6
3317 	case ETHERTYPE_IPV6:
3318 	{
3319 		struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
3320 		struct tcphdr *th;
3321 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
3322 
3323 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
3324 			txq->ift_pullups++;
3325 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
3326 				return (ENOMEM);
3327 		}
3328 		th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
3329 
3330 		/* XXX-BZ this will go badly in case of ext hdrs. */
3331 		pi->ipi_ipproto = ip6->ip6_nxt;
3332 		pi->ipi_flags |= IPI_TX_IPV6;
3333 
3334 		/* TCP checksum offload may require TCP header length */
3335 		if (IS_TX_OFFLOAD6(pi)) {
3336 			if (pi->ipi_ipproto == IPPROTO_TCP) {
3337 				if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) {
3338 					txq->ift_pullups++;
3339 					if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL))
3340 						return (ENOMEM);
3341 				}
3342 				pi->ipi_tcp_hflags = th->th_flags;
3343 				pi->ipi_tcp_hlen = th->th_off << 2;
3344 				pi->ipi_tcp_seq = th->th_seq;
3345 			}
3346 			if (IS_TSO6(pi)) {
3347 				if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP))
3348 					return (ENXIO);
3349 				/*
3350 				 * TSO always requires hardware checksum offload.
3351 				 */
3352 				pi->ipi_csum_flags |= CSUM_IP6_TCP;
3353 				th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0);
3354 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3355 			}
3356 		}
3357 		break;
3358 	}
3359 #endif
3360 	default:
3361 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
3362 		pi->ipi_ip_hlen = 0;
3363 		break;
3364 	}
3365 	*mp = m;
3366 
3367 	return (0);
3368 }
3369 
3370 /*
3371  * If dodgy hardware rejects the scatter gather chain we've handed it
3372  * we'll need to remove the mbuf chain from ifsg_m[] before we can add the
3373  * m_defrag'd mbufs
3374  */
3375 static __noinline struct mbuf *
3376 iflib_remove_mbuf(iflib_txq_t txq)
3377 {
3378 	int ntxd, pidx;
3379 	struct mbuf *m, **ifsd_m;
3380 
3381 	ifsd_m = txq->ift_sds.ifsd_m;
3382 	ntxd = txq->ift_size;
3383 	pidx = txq->ift_pidx & (ntxd - 1);
3384 	ifsd_m = txq->ift_sds.ifsd_m;
3385 	m = ifsd_m[pidx];
3386 	ifsd_m[pidx] = NULL;
3387 	bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[pidx]);
3388 	if (txq->ift_sds.ifsd_tso_map != NULL)
3389 		bus_dmamap_unload(txq->ift_tso_buf_tag,
3390 		    txq->ift_sds.ifsd_tso_map[pidx]);
3391 #if MEMORY_LOGGING
3392 	txq->ift_dequeued++;
3393 #endif
3394 	return (m);
3395 }
3396 
3397 static inline caddr_t
3398 calc_next_txd(iflib_txq_t txq, int cidx, uint8_t qid)
3399 {
3400 	qidx_t size;
3401 	int ntxd;
3402 	caddr_t start, end, cur, next;
3403 
3404 	ntxd = txq->ift_size;
3405 	size = txq->ift_txd_size[qid];
3406 	start = txq->ift_ifdi[qid].idi_vaddr;
3407 
3408 	if (__predict_false(size == 0))
3409 		return (start);
3410 	cur = start + size*cidx;
3411 	end = start + size*ntxd;
3412 	next = CACHE_PTR_NEXT(cur);
3413 	return (next < end ? next : start);
3414 }
3415 
3416 /*
3417  * Pad an mbuf to ensure a minimum ethernet frame size.
3418  * min_frame_size is the frame size (less CRC) to pad the mbuf to
3419  */
3420 static __noinline int
3421 iflib_ether_pad(device_t dev, struct mbuf **m_head, uint16_t min_frame_size)
3422 {
3423 	/*
3424 	 * 18 is enough bytes to pad an ARP packet to 46 bytes, and
3425 	 * and ARP message is the smallest common payload I can think of
3426 	 */
3427 	static char pad[18];	/* just zeros */
3428 	int n;
3429 	struct mbuf *new_head;
3430 
3431 	if (!M_WRITABLE(*m_head)) {
3432 		new_head = m_dup(*m_head, M_NOWAIT);
3433 		if (new_head == NULL) {
3434 			m_freem(*m_head);
3435 			device_printf(dev, "cannot pad short frame, m_dup() failed");
3436 			DBG_COUNTER_INC(encap_pad_mbuf_fail);
3437 			DBG_COUNTER_INC(tx_frees);
3438 			return ENOMEM;
3439 		}
3440 		m_freem(*m_head);
3441 		*m_head = new_head;
3442 	}
3443 
3444 	for (n = min_frame_size - (*m_head)->m_pkthdr.len;
3445 	     n > 0; n -= sizeof(pad))
3446 		if (!m_append(*m_head, min(n, sizeof(pad)), pad))
3447 			break;
3448 
3449 	if (n > 0) {
3450 		m_freem(*m_head);
3451 		device_printf(dev, "cannot pad short frame\n");
3452 		DBG_COUNTER_INC(encap_pad_mbuf_fail);
3453 		DBG_COUNTER_INC(tx_frees);
3454 		return (ENOBUFS);
3455 	}
3456 
3457 	return 0;
3458 }
3459 
3460 static int
3461 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp)
3462 {
3463 	if_ctx_t		ctx;
3464 	if_shared_ctx_t		sctx;
3465 	if_softc_ctx_t		scctx;
3466 	bus_dma_tag_t		buf_tag;
3467 	bus_dma_segment_t	*segs;
3468 	struct mbuf		*m_head, **ifsd_m;
3469 	void			*next_txd;
3470 	bus_dmamap_t		map;
3471 	struct if_pkt_info	pi;
3472 	int remap = 0;
3473 	int err, nsegs, ndesc, max_segs, pidx, cidx, next, ntxd;
3474 
3475 	ctx = txq->ift_ctx;
3476 	sctx = ctx->ifc_sctx;
3477 	scctx = &ctx->ifc_softc_ctx;
3478 	segs = txq->ift_segs;
3479 	ntxd = txq->ift_size;
3480 	m_head = *m_headp;
3481 	map = NULL;
3482 
3483 	/*
3484 	 * If we're doing TSO the next descriptor to clean may be quite far ahead
3485 	 */
3486 	cidx = txq->ift_cidx;
3487 	pidx = txq->ift_pidx;
3488 	if (ctx->ifc_flags & IFC_PREFETCH) {
3489 		next = (cidx + CACHE_PTR_INCREMENT) & (ntxd-1);
3490 		if (!(ctx->ifc_flags & IFLIB_HAS_TXCQ)) {
3491 			next_txd = calc_next_txd(txq, cidx, 0);
3492 			prefetch(next_txd);
3493 		}
3494 
3495 		/* prefetch the next cache line of mbuf pointers and flags */
3496 		prefetch(&txq->ift_sds.ifsd_m[next]);
3497 		prefetch(&txq->ift_sds.ifsd_map[next]);
3498 		next = (cidx + CACHE_LINE_SIZE) & (ntxd-1);
3499 	}
3500 	map = txq->ift_sds.ifsd_map[pidx];
3501 	ifsd_m = txq->ift_sds.ifsd_m;
3502 
3503 	if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
3504 		buf_tag = txq->ift_tso_buf_tag;
3505 		max_segs = scctx->isc_tx_tso_segments_max;
3506 		map = txq->ift_sds.ifsd_tso_map[pidx];
3507 		MPASS(buf_tag != NULL);
3508 		MPASS(max_segs > 0);
3509 	} else {
3510 		buf_tag = txq->ift_buf_tag;
3511 		max_segs = scctx->isc_tx_nsegments;
3512 		map = txq->ift_sds.ifsd_map[pidx];
3513 	}
3514 	if ((sctx->isc_flags & IFLIB_NEED_ETHER_PAD) &&
3515 	    __predict_false(m_head->m_pkthdr.len < scctx->isc_min_frame_size)) {
3516 		err = iflib_ether_pad(ctx->ifc_dev, m_headp, scctx->isc_min_frame_size);
3517 		if (err) {
3518 			DBG_COUNTER_INC(encap_txd_encap_fail);
3519 			return err;
3520 		}
3521 	}
3522 	m_head = *m_headp;
3523 
3524 	pkt_info_zero(&pi);
3525 	pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG|M_BCAST|M_MCAST));
3526 	pi.ipi_pidx = pidx;
3527 	pi.ipi_qsidx = txq->ift_id;
3528 	pi.ipi_len = m_head->m_pkthdr.len;
3529 	pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags;
3530 	pi.ipi_vtag = M_HAS_VLANTAG(m_head) ? m_head->m_pkthdr.ether_vtag : 0;
3531 
3532 	/* deliberate bitwise OR to make one condition */
3533 	if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) {
3534 		if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0)) {
3535 			DBG_COUNTER_INC(encap_txd_encap_fail);
3536 			return (err);
3537 		}
3538 		m_head = *m_headp;
3539 	}
3540 
3541 retry:
3542 	err = bus_dmamap_load_mbuf_sg(buf_tag, map, m_head, segs, &nsegs,
3543 	    BUS_DMA_NOWAIT);
3544 defrag:
3545 	if (__predict_false(err)) {
3546 		switch (err) {
3547 		case EFBIG:
3548 			/* try collapse once and defrag once */
3549 			if (remap == 0) {
3550 				m_head = m_collapse(*m_headp, M_NOWAIT, max_segs);
3551 				/* try defrag if collapsing fails */
3552 				if (m_head == NULL)
3553 					remap++;
3554 			}
3555 			if (remap == 1) {
3556 				txq->ift_mbuf_defrag++;
3557 				m_head = m_defrag(*m_headp, M_NOWAIT);
3558 			}
3559 			/*
3560 			 * remap should never be >1 unless bus_dmamap_load_mbuf_sg
3561 			 * failed to map an mbuf that was run through m_defrag
3562 			 */
3563 			MPASS(remap <= 1);
3564 			if (__predict_false(m_head == NULL || remap > 1))
3565 				goto defrag_failed;
3566 			remap++;
3567 			*m_headp = m_head;
3568 			goto retry;
3569 			break;
3570 		case ENOMEM:
3571 			txq->ift_no_tx_dma_setup++;
3572 			break;
3573 		default:
3574 			txq->ift_no_tx_dma_setup++;
3575 			m_freem(*m_headp);
3576 			DBG_COUNTER_INC(tx_frees);
3577 			*m_headp = NULL;
3578 			break;
3579 		}
3580 		txq->ift_map_failed++;
3581 		DBG_COUNTER_INC(encap_load_mbuf_fail);
3582 		DBG_COUNTER_INC(encap_txd_encap_fail);
3583 		return (err);
3584 	}
3585 	ifsd_m[pidx] = m_head;
3586 	/*
3587 	 * XXX assumes a 1 to 1 relationship between segments and
3588 	 *        descriptors - this does not hold true on all drivers, e.g.
3589 	 *        cxgb
3590 	 */
3591 	if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) {
3592 		txq->ift_no_desc_avail++;
3593 		bus_dmamap_unload(buf_tag, map);
3594 		DBG_COUNTER_INC(encap_txq_avail_fail);
3595 		DBG_COUNTER_INC(encap_txd_encap_fail);
3596 		if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0)
3597 			GROUPTASK_ENQUEUE(&txq->ift_task);
3598 		return (ENOBUFS);
3599 	}
3600 	/*
3601 	 * On Intel cards we can greatly reduce the number of TX interrupts
3602 	 * we see by only setting report status on every Nth descriptor.
3603 	 * However, this also means that the driver will need to keep track
3604 	 * of the descriptors that RS was set on to check them for the DD bit.
3605 	 */
3606 	txq->ift_rs_pending += nsegs + 1;
3607 	if (txq->ift_rs_pending > TXQ_MAX_RS_DEFERRED(txq) ||
3608 	     iflib_no_tx_batch || (TXQ_AVAIL(txq) - nsegs) <= MAX_TX_DESC(ctx) + 2) {
3609 		pi.ipi_flags |= IPI_TX_INTR;
3610 		txq->ift_rs_pending = 0;
3611 	}
3612 
3613 	pi.ipi_segs = segs;
3614 	pi.ipi_nsegs = nsegs;
3615 
3616 	MPASS(pidx >= 0 && pidx < txq->ift_size);
3617 #ifdef PKT_DEBUG
3618 	print_pkt(&pi);
3619 #endif
3620 	if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) {
3621 		bus_dmamap_sync(buf_tag, map, BUS_DMASYNC_PREWRITE);
3622 		DBG_COUNTER_INC(tx_encap);
3623 		MPASS(pi.ipi_new_pidx < txq->ift_size);
3624 
3625 		ndesc = pi.ipi_new_pidx - pi.ipi_pidx;
3626 		if (pi.ipi_new_pidx < pi.ipi_pidx) {
3627 			ndesc += txq->ift_size;
3628 			txq->ift_gen = 1;
3629 		}
3630 		/*
3631 		 * drivers can need as many as
3632 		 * two sentinels
3633 		 */
3634 		MPASS(ndesc <= pi.ipi_nsegs + 2);
3635 		MPASS(pi.ipi_new_pidx != pidx);
3636 		MPASS(ndesc > 0);
3637 		txq->ift_in_use += ndesc;
3638 		txq->ift_db_pending += ndesc;
3639 
3640 		/*
3641 		 * We update the last software descriptor again here because there may
3642 		 * be a sentinel and/or there may be more mbufs than segments
3643 		 */
3644 		txq->ift_pidx = pi.ipi_new_pidx;
3645 		txq->ift_npending += pi.ipi_ndescs;
3646 	} else {
3647 		*m_headp = m_head = iflib_remove_mbuf(txq);
3648 		if (err == EFBIG) {
3649 			txq->ift_txd_encap_efbig++;
3650 			if (remap < 2) {
3651 				remap = 1;
3652 				goto defrag;
3653 			}
3654 		}
3655 		goto defrag_failed;
3656 	}
3657 	/*
3658 	 * err can't possibly be non-zero here, so we don't neet to test it
3659 	 * to see if we need to DBG_COUNTER_INC(encap_txd_encap_fail).
3660 	 */
3661 	return (err);
3662 
3663 defrag_failed:
3664 	txq->ift_mbuf_defrag_failed++;
3665 	txq->ift_map_failed++;
3666 	m_freem(*m_headp);
3667 	DBG_COUNTER_INC(tx_frees);
3668 	*m_headp = NULL;
3669 	DBG_COUNTER_INC(encap_txd_encap_fail);
3670 	return (ENOMEM);
3671 }
3672 
3673 static void
3674 iflib_tx_desc_free(iflib_txq_t txq, int n)
3675 {
3676 	uint32_t qsize, cidx, mask, gen;
3677 	struct mbuf *m, **ifsd_m;
3678 	bool do_prefetch;
3679 
3680 	cidx = txq->ift_cidx;
3681 	gen = txq->ift_gen;
3682 	qsize = txq->ift_size;
3683 	mask = qsize-1;
3684 	ifsd_m = txq->ift_sds.ifsd_m;
3685 	do_prefetch = (txq->ift_ctx->ifc_flags & IFC_PREFETCH);
3686 
3687 	while (n-- > 0) {
3688 		if (do_prefetch) {
3689 			prefetch(ifsd_m[(cidx + 3) & mask]);
3690 			prefetch(ifsd_m[(cidx + 4) & mask]);
3691 		}
3692 		if ((m = ifsd_m[cidx]) != NULL) {
3693 			prefetch(&ifsd_m[(cidx + CACHE_PTR_INCREMENT) & mask]);
3694 			if (m->m_pkthdr.csum_flags & CSUM_TSO) {
3695 				bus_dmamap_sync(txq->ift_tso_buf_tag,
3696 				    txq->ift_sds.ifsd_tso_map[cidx],
3697 				    BUS_DMASYNC_POSTWRITE);
3698 				bus_dmamap_unload(txq->ift_tso_buf_tag,
3699 				    txq->ift_sds.ifsd_tso_map[cidx]);
3700 			} else {
3701 				bus_dmamap_sync(txq->ift_buf_tag,
3702 				    txq->ift_sds.ifsd_map[cidx],
3703 				    BUS_DMASYNC_POSTWRITE);
3704 				bus_dmamap_unload(txq->ift_buf_tag,
3705 				    txq->ift_sds.ifsd_map[cidx]);
3706 			}
3707 			/* XXX we don't support any drivers that batch packets yet */
3708 			MPASS(m->m_nextpkt == NULL);
3709 			m_freem(m);
3710 			ifsd_m[cidx] = NULL;
3711 #if MEMORY_LOGGING
3712 			txq->ift_dequeued++;
3713 #endif
3714 			DBG_COUNTER_INC(tx_frees);
3715 		}
3716 		if (__predict_false(++cidx == qsize)) {
3717 			cidx = 0;
3718 			gen = 0;
3719 		}
3720 	}
3721 	txq->ift_cidx = cidx;
3722 	txq->ift_gen = gen;
3723 }
3724 
3725 static __inline int
3726 iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh)
3727 {
3728 	int reclaim;
3729 	if_ctx_t ctx = txq->ift_ctx;
3730 
3731 	KASSERT(thresh >= 0, ("invalid threshold to reclaim"));
3732 	MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size);
3733 
3734 	/*
3735 	 * Need a rate-limiting check so that this isn't called every time
3736 	 */
3737 	iflib_tx_credits_update(ctx, txq);
3738 	reclaim = DESC_RECLAIMABLE(txq);
3739 
3740 	if (reclaim <= thresh /* + MAX_TX_DESC(txq->ift_ctx) */) {
3741 #ifdef INVARIANTS
3742 		if (iflib_verbose_debug) {
3743 			printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __FUNCTION__,
3744 			       txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments,
3745 			       reclaim, thresh);
3746 		}
3747 #endif
3748 		return (0);
3749 	}
3750 	iflib_tx_desc_free(txq, reclaim);
3751 	txq->ift_cleaned += reclaim;
3752 	txq->ift_in_use -= reclaim;
3753 
3754 	return (reclaim);
3755 }
3756 
3757 static struct mbuf **
3758 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset, int remaining)
3759 {
3760 	int next, size;
3761 	struct mbuf **items;
3762 
3763 	size = r->size;
3764 	next = (cidx + CACHE_PTR_INCREMENT) & (size-1);
3765 	items = __DEVOLATILE(struct mbuf **, &r->items[0]);
3766 
3767 	prefetch(items[(cidx + offset) & (size-1)]);
3768 	if (remaining > 1) {
3769 		prefetch2cachelines(&items[next]);
3770 		prefetch2cachelines(items[(cidx + offset + 1) & (size-1)]);
3771 		prefetch2cachelines(items[(cidx + offset + 2) & (size-1)]);
3772 		prefetch2cachelines(items[(cidx + offset + 3) & (size-1)]);
3773 	}
3774 	return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (size-1)]));
3775 }
3776 
3777 static void
3778 iflib_txq_check_drain(iflib_txq_t txq, int budget)
3779 {
3780 
3781 	ifmp_ring_check_drainage(txq->ift_br, budget);
3782 }
3783 
3784 static uint32_t
3785 iflib_txq_can_drain(struct ifmp_ring *r)
3786 {
3787 	iflib_txq_t txq = r->cookie;
3788 	if_ctx_t ctx = txq->ift_ctx;
3789 
3790 	if (TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2)
3791 		return (1);
3792 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3793 	    BUS_DMASYNC_POSTREAD);
3794 	return (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id,
3795 	    false));
3796 }
3797 
3798 static uint32_t
3799 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3800 {
3801 	iflib_txq_t txq = r->cookie;
3802 	if_ctx_t ctx = txq->ift_ctx;
3803 	if_t ifp = ctx->ifc_ifp;
3804 	struct mbuf *m, **mp;
3805 	int avail, bytes_sent, skipped, count, err, i;
3806 	int mcast_sent, pkt_sent, reclaimed;
3807 	bool do_prefetch, rang, ring;
3808 
3809 	if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) ||
3810 			    !LINK_ACTIVE(ctx))) {
3811 		DBG_COUNTER_INC(txq_drain_notready);
3812 		return (0);
3813 	}
3814 	reclaimed = iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
3815 	rang = iflib_txd_db_check(txq, reclaimed && txq->ift_db_pending);
3816 	avail = IDXDIFF(pidx, cidx, r->size);
3817 
3818 	if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) {
3819 		/*
3820 		 * The driver is unloading so we need to free all pending packets.
3821 		 */
3822 		DBG_COUNTER_INC(txq_drain_flushing);
3823 		for (i = 0; i < avail; i++) {
3824 			if (__predict_true(r->items[(cidx + i) & (r->size-1)] != (void *)txq))
3825 				m_freem(r->items[(cidx + i) & (r->size-1)]);
3826 			r->items[(cidx + i) & (r->size-1)] = NULL;
3827 		}
3828 		return (avail);
3829 	}
3830 
3831 	if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3832 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3833 		CALLOUT_LOCK(txq);
3834 		callout_stop(&txq->ift_timer);
3835 		CALLOUT_UNLOCK(txq);
3836 		DBG_COUNTER_INC(txq_drain_oactive);
3837 		return (0);
3838 	}
3839 
3840 	/*
3841 	 * If we've reclaimed any packets this queue cannot be hung.
3842 	 */
3843 	if (reclaimed)
3844 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3845 	skipped = mcast_sent = bytes_sent = pkt_sent = 0;
3846 	count = MIN(avail, TX_BATCH_SIZE);
3847 #ifdef INVARIANTS
3848 	if (iflib_verbose_debug)
3849 		printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __FUNCTION__,
3850 		       avail, ctx->ifc_flags, TXQ_AVAIL(txq));
3851 #endif
3852 	do_prefetch = (ctx->ifc_flags & IFC_PREFETCH);
3853 	err = 0;
3854 	for (i = 0; i < count && TXQ_AVAIL(txq) >= MAX_TX_DESC(ctx) + 2; i++) {
3855 		int rem = do_prefetch ? count - i : 0;
3856 
3857 		mp = _ring_peek_one(r, cidx, i, rem);
3858 		MPASS(mp != NULL && *mp != NULL);
3859 
3860 		/*
3861 		 * Completion interrupts will use the address of the txq
3862 		 * as a sentinel to enqueue _something_ in order to acquire
3863 		 * the lock on the mp_ring (there's no direct lock call).
3864 		 * We obviously whave to check for these sentinel cases
3865 		 * and skip them.
3866 		 */
3867 		if (__predict_false(*mp == (struct mbuf *)txq)) {
3868 			skipped++;
3869 			continue;
3870 		}
3871 		err = iflib_encap(txq, mp);
3872 		if (__predict_false(err)) {
3873 			/* no room - bail out */
3874 			if (err == ENOBUFS)
3875 				break;
3876 			skipped++;
3877 			/* we can't send this packet - skip it */
3878 			continue;
3879 		}
3880 		pkt_sent++;
3881 		m = *mp;
3882 		DBG_COUNTER_INC(tx_sent);
3883 		bytes_sent += m->m_pkthdr.len;
3884 		mcast_sent += !!(m->m_flags & M_MCAST);
3885 
3886 		if (__predict_false(!(ifp->if_drv_flags & IFF_DRV_RUNNING)))
3887 			break;
3888 		ETHER_BPF_MTAP(ifp, m);
3889 		rang = iflib_txd_db_check(txq, false);
3890 	}
3891 
3892 	/* deliberate use of bitwise or to avoid gratuitous short-circuit */
3893 	ring = rang ? false  : (iflib_min_tx_latency | err);
3894 	iflib_txd_db_check(txq, ring);
3895 	if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
3896 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
3897 	if (mcast_sent)
3898 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
3899 #ifdef INVARIANTS
3900 	if (iflib_verbose_debug)
3901 		printf("consumed=%d\n", skipped + pkt_sent);
3902 #endif
3903 	return (skipped + pkt_sent);
3904 }
3905 
3906 static uint32_t
3907 iflib_txq_drain_always(struct ifmp_ring *r)
3908 {
3909 	return (1);
3910 }
3911 
3912 static uint32_t
3913 iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3914 {
3915 	int i, avail;
3916 	struct mbuf **mp;
3917 	iflib_txq_t txq;
3918 
3919 	txq = r->cookie;
3920 
3921 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3922 	CALLOUT_LOCK(txq);
3923 	callout_stop(&txq->ift_timer);
3924 	CALLOUT_UNLOCK(txq);
3925 
3926 	avail = IDXDIFF(pidx, cidx, r->size);
3927 	for (i = 0; i < avail; i++) {
3928 		mp = _ring_peek_one(r, cidx, i, avail - i);
3929 		if (__predict_false(*mp == (struct mbuf *)txq))
3930 			continue;
3931 		m_freem(*mp);
3932 		DBG_COUNTER_INC(tx_frees);
3933 	}
3934 	MPASS(ifmp_ring_is_stalled(r) == 0);
3935 	return (avail);
3936 }
3937 
3938 static void
3939 iflib_ifmp_purge(iflib_txq_t txq)
3940 {
3941 	struct ifmp_ring *r;
3942 
3943 	r = txq->ift_br;
3944 	r->drain = iflib_txq_drain_free;
3945 	r->can_drain = iflib_txq_drain_always;
3946 
3947 	ifmp_ring_check_drainage(r, r->size);
3948 
3949 	r->drain = iflib_txq_drain;
3950 	r->can_drain = iflib_txq_can_drain;
3951 }
3952 
3953 static void
3954 _task_fn_tx(void *context)
3955 {
3956 	iflib_txq_t txq = context;
3957 	if_ctx_t ctx = txq->ift_ctx;
3958 	if_t ifp = ctx->ifc_ifp;
3959 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
3960 
3961 #ifdef IFLIB_DIAGNOSTICS
3962 	txq->ift_cpu_exec_count[curcpu]++;
3963 #endif
3964 	if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
3965 		return;
3966 #ifdef DEV_NETMAP
3967 	if ((if_getcapenable(ifp) & IFCAP_NETMAP) &&
3968 	    netmap_tx_irq(ifp, txq->ift_id))
3969 		goto skip_ifmp;
3970 #endif
3971 #ifdef ALTQ
3972 	if (ALTQ_IS_ENABLED(&ifp->if_snd))
3973 		iflib_altq_if_start(ifp);
3974 #endif
3975 	if (txq->ift_db_pending)
3976 		ifmp_ring_enqueue(txq->ift_br, (void **)&txq, 1, TX_BATCH_SIZE, abdicate);
3977 	else if (!abdicate)
3978 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3979 	/*
3980 	 * When abdicating, we always need to check drainage, not just when we don't enqueue
3981 	 */
3982 	if (abdicate)
3983 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3984 #ifdef DEV_NETMAP
3985 skip_ifmp:
3986 #endif
3987 	if (ctx->ifc_flags & IFC_LEGACY)
3988 		IFDI_INTR_ENABLE(ctx);
3989 	else
3990 		IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3991 }
3992 
3993 static void
3994 _task_fn_rx(void *context)
3995 {
3996 	iflib_rxq_t rxq = context;
3997 	if_ctx_t ctx = rxq->ifr_ctx;
3998 	uint8_t more;
3999 	uint16_t budget;
4000 #ifdef DEV_NETMAP
4001 	u_int work = 0;
4002 	int nmirq;
4003 #endif
4004 
4005 #ifdef IFLIB_DIAGNOSTICS
4006 	rxq->ifr_cpu_exec_count[curcpu]++;
4007 #endif
4008 	DBG_COUNTER_INC(task_fn_rxs);
4009 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
4010 		return;
4011 #ifdef DEV_NETMAP
4012 	nmirq = netmap_rx_irq(ctx->ifc_ifp, rxq->ifr_id, &work);
4013 	if (nmirq != NM_IRQ_PASS) {
4014 		more = (nmirq == NM_IRQ_RESCHED) ? IFLIB_RXEOF_MORE : 0;
4015 		goto skip_rxeof;
4016 	}
4017 #endif
4018 	budget = ctx->ifc_sysctl_rx_budget;
4019 	if (budget == 0)
4020 		budget = 16;	/* XXX */
4021 	more = iflib_rxeof(rxq, budget);
4022 #ifdef DEV_NETMAP
4023 skip_rxeof:
4024 #endif
4025 	if ((more & IFLIB_RXEOF_MORE) == 0) {
4026 		if (ctx->ifc_flags & IFC_LEGACY)
4027 			IFDI_INTR_ENABLE(ctx);
4028 		else
4029 			IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
4030 		DBG_COUNTER_INC(rx_intr_enables);
4031 	}
4032 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
4033 		return;
4034 
4035 	if (more & IFLIB_RXEOF_MORE)
4036 		GROUPTASK_ENQUEUE(&rxq->ifr_task);
4037 	else if (more & IFLIB_RXEOF_EMPTY)
4038 		callout_reset_curcpu(&rxq->ifr_watchdog, 1, &_task_fn_rx_watchdog, rxq);
4039 }
4040 
4041 static void
4042 _task_fn_admin(void *context)
4043 {
4044 	if_ctx_t ctx = context;
4045 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
4046 	iflib_txq_t txq;
4047 	int i;
4048 	bool oactive, running, do_reset, do_watchdog, in_detach;
4049 
4050 	STATE_LOCK(ctx);
4051 	running = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING);
4052 	oactive = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE);
4053 	do_reset = (ctx->ifc_flags & IFC_DO_RESET);
4054 	do_watchdog = (ctx->ifc_flags & IFC_DO_WATCHDOG);
4055 	in_detach = (ctx->ifc_flags & IFC_IN_DETACH);
4056 	ctx->ifc_flags &= ~(IFC_DO_RESET|IFC_DO_WATCHDOG);
4057 	STATE_UNLOCK(ctx);
4058 
4059 	if ((!running && !oactive) && !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
4060 		return;
4061 	if (in_detach)
4062 		return;
4063 
4064 	CTX_LOCK(ctx);
4065 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
4066 		CALLOUT_LOCK(txq);
4067 		callout_stop(&txq->ift_timer);
4068 		CALLOUT_UNLOCK(txq);
4069 	}
4070 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_ADMINCQ)
4071 		IFDI_ADMIN_COMPLETION_HANDLE(ctx);
4072 	if (do_watchdog) {
4073 		ctx->ifc_watchdog_events++;
4074 		IFDI_WATCHDOG_RESET(ctx);
4075 	}
4076 	IFDI_UPDATE_ADMIN_STATUS(ctx);
4077 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
4078 		callout_reset_on(&txq->ift_timer, iflib_timer_default, iflib_timer, txq,
4079 		    txq->ift_timer.c_cpu);
4080 	}
4081 	IFDI_LINK_INTR_ENABLE(ctx);
4082 	if (do_reset)
4083 		iflib_if_init_locked(ctx);
4084 	CTX_UNLOCK(ctx);
4085 
4086 	if (LINK_ACTIVE(ctx) == 0)
4087 		return;
4088 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
4089 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
4090 }
4091 
4092 static void
4093 _task_fn_iov(void *context)
4094 {
4095 	if_ctx_t ctx = context;
4096 
4097 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING) &&
4098 	    !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
4099 		return;
4100 
4101 	CTX_LOCK(ctx);
4102 	IFDI_VFLR_HANDLE(ctx);
4103 	CTX_UNLOCK(ctx);
4104 }
4105 
4106 static int
4107 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)
4108 {
4109 	int err;
4110 	if_int_delay_info_t info;
4111 	if_ctx_t ctx;
4112 
4113 	info = (if_int_delay_info_t)arg1;
4114 	ctx = info->iidi_ctx;
4115 	info->iidi_req = req;
4116 	info->iidi_oidp = oidp;
4117 	CTX_LOCK(ctx);
4118 	err = IFDI_SYSCTL_INT_DELAY(ctx, info);
4119 	CTX_UNLOCK(ctx);
4120 	return (err);
4121 }
4122 
4123 /*********************************************************************
4124  *
4125  *  IFNET FUNCTIONS
4126  *
4127  **********************************************************************/
4128 
4129 static void
4130 iflib_if_init_locked(if_ctx_t ctx)
4131 {
4132 	iflib_stop(ctx);
4133 	iflib_init_locked(ctx);
4134 }
4135 
4136 static void
4137 iflib_if_init(void *arg)
4138 {
4139 	if_ctx_t ctx = arg;
4140 
4141 	CTX_LOCK(ctx);
4142 	iflib_if_init_locked(ctx);
4143 	CTX_UNLOCK(ctx);
4144 }
4145 
4146 static int
4147 iflib_if_transmit(if_t ifp, struct mbuf *m)
4148 {
4149 	if_ctx_t	ctx = if_getsoftc(ifp);
4150 
4151 	iflib_txq_t txq;
4152 	int err, qidx;
4153 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
4154 
4155 	if (__predict_false((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) {
4156 		DBG_COUNTER_INC(tx_frees);
4157 		m_freem(m);
4158 		return (ENETDOWN);
4159 	}
4160 
4161 	MPASS(m->m_nextpkt == NULL);
4162 	/* ALTQ-enabled interfaces always use queue 0. */
4163 	qidx = 0;
4164 	/* Use driver-supplied queue selection method if it exists */
4165 	if (ctx->isc_txq_select)
4166 		qidx = ctx->isc_txq_select(ctx->ifc_softc, m);
4167 	/* If not, use iflib's standard method */
4168 	else if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m) && !ALTQ_IS_ENABLED(&ifp->if_snd))
4169 		qidx = QIDX(ctx, m);
4170 
4171 	/* Set TX queue */
4172 	txq = &ctx->ifc_txqs[qidx];
4173 
4174 #ifdef DRIVER_BACKPRESSURE
4175 	if (txq->ift_closed) {
4176 		while (m != NULL) {
4177 			next = m->m_nextpkt;
4178 			m->m_nextpkt = NULL;
4179 			m_freem(m);
4180 			DBG_COUNTER_INC(tx_frees);
4181 			m = next;
4182 		}
4183 		return (ENOBUFS);
4184 	}
4185 #endif
4186 #ifdef notyet
4187 	qidx = count = 0;
4188 	mp = marr;
4189 	next = m;
4190 	do {
4191 		count++;
4192 		next = next->m_nextpkt;
4193 	} while (next != NULL);
4194 
4195 	if (count > nitems(marr))
4196 		if ((mp = malloc(count*sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) {
4197 			/* XXX check nextpkt */
4198 			m_freem(m);
4199 			/* XXX simplify for now */
4200 			DBG_COUNTER_INC(tx_frees);
4201 			return (ENOBUFS);
4202 		}
4203 	for (next = m, i = 0; next != NULL; i++) {
4204 		mp[i] = next;
4205 		next = next->m_nextpkt;
4206 		mp[i]->m_nextpkt = NULL;
4207 	}
4208 #endif
4209 	DBG_COUNTER_INC(tx_seen);
4210 	err = ifmp_ring_enqueue(txq->ift_br, (void **)&m, 1, TX_BATCH_SIZE, abdicate);
4211 
4212 	if (abdicate)
4213 		GROUPTASK_ENQUEUE(&txq->ift_task);
4214  	if (err) {
4215 		if (!abdicate)
4216 			GROUPTASK_ENQUEUE(&txq->ift_task);
4217 		/* support forthcoming later */
4218 #ifdef DRIVER_BACKPRESSURE
4219 		txq->ift_closed = TRUE;
4220 #endif
4221 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
4222 		m_freem(m);
4223 		DBG_COUNTER_INC(tx_frees);
4224 	}
4225 
4226 	return (err);
4227 }
4228 
4229 #ifdef ALTQ
4230 /*
4231  * The overall approach to integrating iflib with ALTQ is to continue to use
4232  * the iflib mp_ring machinery between the ALTQ queue(s) and the hardware
4233  * ring.  Technically, when using ALTQ, queueing to an intermediate mp_ring
4234  * is redundant/unnecessary, but doing so minimizes the amount of
4235  * ALTQ-specific code required in iflib.  It is assumed that the overhead of
4236  * redundantly queueing to an intermediate mp_ring is swamped by the
4237  * performance limitations inherent in using ALTQ.
4238  *
4239  * When ALTQ support is compiled in, all iflib drivers will use a transmit
4240  * routine, iflib_altq_if_transmit(), that checks if ALTQ is enabled for the
4241  * given interface.  If ALTQ is enabled for an interface, then all
4242  * transmitted packets for that interface will be submitted to the ALTQ
4243  * subsystem via IFQ_ENQUEUE().  We don't use the legacy if_transmit()
4244  * implementation because it uses IFQ_HANDOFF(), which will duplicatively
4245  * update stats that the iflib machinery handles, and which is sensitve to
4246  * the disused IFF_DRV_OACTIVE flag.  Additionally, iflib_altq_if_start()
4247  * will be installed as the start routine for use by ALTQ facilities that
4248  * need to trigger queue drains on a scheduled basis.
4249  *
4250  */
4251 static void
4252 iflib_altq_if_start(if_t ifp)
4253 {
4254 	struct ifaltq *ifq = &ifp->if_snd;
4255 	struct mbuf *m;
4256 
4257 	IFQ_LOCK(ifq);
4258 	IFQ_DEQUEUE_NOLOCK(ifq, m);
4259 	while (m != NULL) {
4260 		iflib_if_transmit(ifp, m);
4261 		IFQ_DEQUEUE_NOLOCK(ifq, m);
4262 	}
4263 	IFQ_UNLOCK(ifq);
4264 }
4265 
4266 static int
4267 iflib_altq_if_transmit(if_t ifp, struct mbuf *m)
4268 {
4269 	int err;
4270 
4271 	if (ALTQ_IS_ENABLED(&ifp->if_snd)) {
4272 		IFQ_ENQUEUE(&ifp->if_snd, m, err);
4273 		if (err == 0)
4274 			iflib_altq_if_start(ifp);
4275 	} else
4276 		err = iflib_if_transmit(ifp, m);
4277 
4278 	return (err);
4279 }
4280 #endif /* ALTQ */
4281 
4282 static void
4283 iflib_if_qflush(if_t ifp)
4284 {
4285 	if_ctx_t ctx = if_getsoftc(ifp);
4286 	iflib_txq_t txq = ctx->ifc_txqs;
4287 	int i;
4288 
4289 	STATE_LOCK(ctx);
4290 	ctx->ifc_flags |= IFC_QFLUSH;
4291 	STATE_UNLOCK(ctx);
4292 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
4293 		while (!(ifmp_ring_is_idle(txq->ift_br) || ifmp_ring_is_stalled(txq->ift_br)))
4294 			iflib_txq_check_drain(txq, 0);
4295 	STATE_LOCK(ctx);
4296 	ctx->ifc_flags &= ~IFC_QFLUSH;
4297 	STATE_UNLOCK(ctx);
4298 
4299 	/*
4300 	 * When ALTQ is enabled, this will also take care of purging the
4301 	 * ALTQ queue(s).
4302 	 */
4303 	if_qflush(ifp);
4304 }
4305 
4306 #define IFCAP_FLAGS (IFCAP_HWCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \
4307 		     IFCAP_TSO | IFCAP_VLAN_HWTAGGING | IFCAP_HWSTATS | \
4308 		     IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | \
4309 		     IFCAP_VLAN_HWTSO | IFCAP_VLAN_HWCSUM | IFCAP_MEXTPG)
4310 
4311 static int
4312 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data)
4313 {
4314 	if_ctx_t ctx = if_getsoftc(ifp);
4315 	struct ifreq	*ifr = (struct ifreq *)data;
4316 #if defined(INET) || defined(INET6)
4317 	struct ifaddr	*ifa = (struct ifaddr *)data;
4318 #endif
4319 	bool		avoid_reset = false;
4320 	int		err = 0, reinit = 0, bits;
4321 
4322 	switch (command) {
4323 	case SIOCSIFADDR:
4324 #ifdef INET
4325 		if (ifa->ifa_addr->sa_family == AF_INET)
4326 			avoid_reset = true;
4327 #endif
4328 #ifdef INET6
4329 		if (ifa->ifa_addr->sa_family == AF_INET6)
4330 			avoid_reset = true;
4331 #endif
4332 		/*
4333 		** Calling init results in link renegotiation,
4334 		** so we avoid doing it when possible.
4335 		*/
4336 		if (avoid_reset) {
4337 			if_setflagbits(ifp, IFF_UP,0);
4338 			if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
4339 				reinit = 1;
4340 #ifdef INET
4341 			if (!(if_getflags(ifp) & IFF_NOARP))
4342 				arp_ifinit(ifp, ifa);
4343 #endif
4344 		} else
4345 			err = ether_ioctl(ifp, command, data);
4346 		break;
4347 	case SIOCSIFMTU:
4348 		CTX_LOCK(ctx);
4349 		if (ifr->ifr_mtu == if_getmtu(ifp)) {
4350 			CTX_UNLOCK(ctx);
4351 			break;
4352 		}
4353 		bits = if_getdrvflags(ifp);
4354 		/* stop the driver and free any clusters before proceeding */
4355 		iflib_stop(ctx);
4356 
4357 		if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) {
4358 			STATE_LOCK(ctx);
4359 			if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size)
4360 				ctx->ifc_flags |= IFC_MULTISEG;
4361 			else
4362 				ctx->ifc_flags &= ~IFC_MULTISEG;
4363 			STATE_UNLOCK(ctx);
4364 			err = if_setmtu(ifp, ifr->ifr_mtu);
4365 		}
4366 		iflib_init_locked(ctx);
4367 		STATE_LOCK(ctx);
4368 		if_setdrvflags(ifp, bits);
4369 		STATE_UNLOCK(ctx);
4370 		CTX_UNLOCK(ctx);
4371 		break;
4372 	case SIOCSIFFLAGS:
4373 		CTX_LOCK(ctx);
4374 		if (if_getflags(ifp) & IFF_UP) {
4375 			if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4376 				if ((if_getflags(ifp) ^ ctx->ifc_if_flags) &
4377 				    (IFF_PROMISC | IFF_ALLMULTI)) {
4378 					CTX_UNLOCK(ctx);
4379 					err = IFDI_PROMISC_SET(ctx, if_getflags(ifp));
4380 					CTX_LOCK(ctx);
4381 				}
4382 			} else
4383 				reinit = 1;
4384 		} else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4385 			iflib_stop(ctx);
4386 		}
4387 		ctx->ifc_if_flags = if_getflags(ifp);
4388 		CTX_UNLOCK(ctx);
4389 		break;
4390 	case SIOCADDMULTI:
4391 	case SIOCDELMULTI:
4392 		if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4393 			CTX_LOCK(ctx);
4394 			IFDI_INTR_DISABLE(ctx);
4395 			IFDI_MULTI_SET(ctx);
4396 			IFDI_INTR_ENABLE(ctx);
4397 			CTX_UNLOCK(ctx);
4398 		}
4399 		break;
4400 	case SIOCSIFMEDIA:
4401 		CTX_LOCK(ctx);
4402 		IFDI_MEDIA_SET(ctx);
4403 		CTX_UNLOCK(ctx);
4404 		/* FALLTHROUGH */
4405 	case SIOCGIFMEDIA:
4406 	case SIOCGIFXMEDIA:
4407 		err = ifmedia_ioctl(ifp, ifr, ctx->ifc_mediap, command);
4408 		break;
4409 	case SIOCGI2C:
4410 	{
4411 		struct ifi2creq i2c;
4412 
4413 		err = copyin(ifr_data_get_ptr(ifr), &i2c, sizeof(i2c));
4414 		if (err != 0)
4415 			break;
4416 		if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) {
4417 			err = EINVAL;
4418 			break;
4419 		}
4420 		if (i2c.len > sizeof(i2c.data)) {
4421 			err = EINVAL;
4422 			break;
4423 		}
4424 
4425 		if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0)
4426 			err = copyout(&i2c, ifr_data_get_ptr(ifr),
4427 			    sizeof(i2c));
4428 		break;
4429 	}
4430 	case SIOCSIFCAP:
4431 	{
4432 		int mask, setmask, oldmask;
4433 
4434 		oldmask = if_getcapenable(ifp);
4435 		mask = ifr->ifr_reqcap ^ oldmask;
4436 		mask &= ctx->ifc_softc_ctx.isc_capabilities | IFCAP_MEXTPG;
4437 		setmask = 0;
4438 #ifdef TCP_OFFLOAD
4439 		setmask |= mask & (IFCAP_TOE4|IFCAP_TOE6);
4440 #endif
4441 		setmask |= (mask & IFCAP_FLAGS);
4442 		setmask |= (mask & IFCAP_WOL);
4443 
4444 		/*
4445 		 * If any RX csum has changed, change all the ones that
4446 		 * are supported by the driver.
4447 		 */
4448 		if (setmask & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6)) {
4449 			setmask |= ctx->ifc_softc_ctx.isc_capabilities &
4450 			    (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6);
4451 		}
4452 
4453 		/*
4454 		 * want to ensure that traffic has stopped before we change any of the flags
4455 		 */
4456 		if (setmask) {
4457 			CTX_LOCK(ctx);
4458 			bits = if_getdrvflags(ifp);
4459 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4460 				iflib_stop(ctx);
4461 			STATE_LOCK(ctx);
4462 			if_togglecapenable(ifp, setmask);
4463 			ctx->ifc_softc_ctx.isc_capenable ^= setmask;
4464 			STATE_UNLOCK(ctx);
4465 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4466 				iflib_init_locked(ctx);
4467 			STATE_LOCK(ctx);
4468 			if_setdrvflags(ifp, bits);
4469 			STATE_UNLOCK(ctx);
4470 			CTX_UNLOCK(ctx);
4471 		}
4472 		if_vlancap(ifp);
4473 		break;
4474 	}
4475 	case SIOCGPRIVATE_0:
4476 	case SIOCSDRVSPEC:
4477 	case SIOCGDRVSPEC:
4478 		CTX_LOCK(ctx);
4479 		err = IFDI_PRIV_IOCTL(ctx, command, data);
4480 		CTX_UNLOCK(ctx);
4481 		break;
4482 	default:
4483 		err = ether_ioctl(ifp, command, data);
4484 		break;
4485 	}
4486 	if (reinit)
4487 		iflib_if_init(ctx);
4488 	return (err);
4489 }
4490 
4491 static uint64_t
4492 iflib_if_get_counter(if_t ifp, ift_counter cnt)
4493 {
4494 	if_ctx_t ctx = if_getsoftc(ifp);
4495 
4496 	return (IFDI_GET_COUNTER(ctx, cnt));
4497 }
4498 
4499 /*********************************************************************
4500  *
4501  *  OTHER FUNCTIONS EXPORTED TO THE STACK
4502  *
4503  **********************************************************************/
4504 
4505 static void
4506 iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag)
4507 {
4508 	if_ctx_t ctx = if_getsoftc(ifp);
4509 
4510 	if ((void *)ctx != arg)
4511 		return;
4512 
4513 	if ((vtag == 0) || (vtag > 4095))
4514 		return;
4515 
4516 	if (iflib_in_detach(ctx))
4517 		return;
4518 
4519 	CTX_LOCK(ctx);
4520 	/* Driver may need all untagged packets to be flushed */
4521 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4522 		iflib_stop(ctx);
4523 	IFDI_VLAN_REGISTER(ctx, vtag);
4524 	/* Re-init to load the changes, if required */
4525 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4526 		iflib_init_locked(ctx);
4527 	CTX_UNLOCK(ctx);
4528 }
4529 
4530 static void
4531 iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag)
4532 {
4533 	if_ctx_t ctx = if_getsoftc(ifp);
4534 
4535 	if ((void *)ctx != arg)
4536 		return;
4537 
4538 	if ((vtag == 0) || (vtag > 4095))
4539 		return;
4540 
4541 	CTX_LOCK(ctx);
4542 	/* Driver may need all tagged packets to be flushed */
4543 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4544 		iflib_stop(ctx);
4545 	IFDI_VLAN_UNREGISTER(ctx, vtag);
4546 	/* Re-init to load the changes, if required */
4547 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4548 		iflib_init_locked(ctx);
4549 	CTX_UNLOCK(ctx);
4550 }
4551 
4552 static void
4553 iflib_led_func(void *arg, int onoff)
4554 {
4555 	if_ctx_t ctx = arg;
4556 
4557 	CTX_LOCK(ctx);
4558 	IFDI_LED_FUNC(ctx, onoff);
4559 	CTX_UNLOCK(ctx);
4560 }
4561 
4562 /*********************************************************************
4563  *
4564  *  BUS FUNCTION DEFINITIONS
4565  *
4566  **********************************************************************/
4567 
4568 int
4569 iflib_device_probe(device_t dev)
4570 {
4571 	const pci_vendor_info_t *ent;
4572 	if_shared_ctx_t sctx;
4573 	uint16_t pci_device_id, pci_rev_id, pci_subdevice_id, pci_subvendor_id;
4574 	uint16_t pci_vendor_id;
4575 
4576 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4577 		return (ENOTSUP);
4578 
4579 	pci_vendor_id = pci_get_vendor(dev);
4580 	pci_device_id = pci_get_device(dev);
4581 	pci_subvendor_id = pci_get_subvendor(dev);
4582 	pci_subdevice_id = pci_get_subdevice(dev);
4583 	pci_rev_id = pci_get_revid(dev);
4584 	if (sctx->isc_parse_devinfo != NULL)
4585 		sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id);
4586 
4587 	ent = sctx->isc_vendor_info;
4588 	while (ent->pvi_vendor_id != 0) {
4589 		if (pci_vendor_id != ent->pvi_vendor_id) {
4590 			ent++;
4591 			continue;
4592 		}
4593 		if ((pci_device_id == ent->pvi_device_id) &&
4594 		    ((pci_subvendor_id == ent->pvi_subvendor_id) ||
4595 		     (ent->pvi_subvendor_id == 0)) &&
4596 		    ((pci_subdevice_id == ent->pvi_subdevice_id) ||
4597 		     (ent->pvi_subdevice_id == 0)) &&
4598 		    ((pci_rev_id == ent->pvi_rev_id) ||
4599 		     (ent->pvi_rev_id == 0))) {
4600 			device_set_desc_copy(dev, ent->pvi_name);
4601 			/* this needs to be changed to zero if the bus probing code
4602 			 * ever stops re-probing on best match because the sctx
4603 			 * may have its values over written by register calls
4604 			 * in subsequent probes
4605 			 */
4606 			return (BUS_PROBE_DEFAULT);
4607 		}
4608 		ent++;
4609 	}
4610 	return (ENXIO);
4611 }
4612 
4613 int
4614 iflib_device_probe_vendor(device_t dev)
4615 {
4616 	int probe;
4617 
4618 	probe = iflib_device_probe(dev);
4619 	if (probe == BUS_PROBE_DEFAULT)
4620 		return (BUS_PROBE_VENDOR);
4621 	else
4622 		return (probe);
4623 }
4624 
4625 static void
4626 iflib_reset_qvalues(if_ctx_t ctx)
4627 {
4628 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4629 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4630 	device_t dev = ctx->ifc_dev;
4631 	int i;
4632 
4633 	if (ctx->ifc_sysctl_ntxqs != 0)
4634 		scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs;
4635 	if (ctx->ifc_sysctl_nrxqs != 0)
4636 		scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs;
4637 
4638 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4639 		if (ctx->ifc_sysctl_ntxds[i] != 0)
4640 			scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i];
4641 		else
4642 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4643 	}
4644 
4645 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4646 		if (ctx->ifc_sysctl_nrxds[i] != 0)
4647 			scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i];
4648 		else
4649 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4650 	}
4651 
4652 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4653 		if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) {
4654 			device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n",
4655 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]);
4656 			scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i];
4657 		}
4658 		if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) {
4659 			device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n",
4660 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]);
4661 			scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i];
4662 		}
4663 		if (!powerof2(scctx->isc_nrxd[i])) {
4664 			device_printf(dev, "nrxd%d: %d is not a power of 2 - using default value of %d\n",
4665 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_default[i]);
4666 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4667 		}
4668 	}
4669 
4670 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4671 		if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) {
4672 			device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n",
4673 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]);
4674 			scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i];
4675 		}
4676 		if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) {
4677 			device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n",
4678 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]);
4679 			scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i];
4680 		}
4681 		if (!powerof2(scctx->isc_ntxd[i])) {
4682 			device_printf(dev, "ntxd%d: %d is not a power of 2 - using default value of %d\n",
4683 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_default[i]);
4684 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4685 		}
4686 	}
4687 }
4688 
4689 static void
4690 iflib_add_pfil(if_ctx_t ctx)
4691 {
4692 	struct pfil_head *pfil;
4693 	struct pfil_head_args pa;
4694 	iflib_rxq_t rxq;
4695 	int i;
4696 
4697 	pa.pa_version = PFIL_VERSION;
4698 	pa.pa_flags = PFIL_IN;
4699 	pa.pa_type = PFIL_TYPE_ETHERNET;
4700 	pa.pa_headname = ctx->ifc_ifp->if_xname;
4701 	pfil = pfil_head_register(&pa);
4702 
4703 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4704 		rxq->pfil = pfil;
4705 	}
4706 }
4707 
4708 static void
4709 iflib_rem_pfil(if_ctx_t ctx)
4710 {
4711 	struct pfil_head *pfil;
4712 	iflib_rxq_t rxq;
4713 	int i;
4714 
4715 	rxq = ctx->ifc_rxqs;
4716 	pfil = rxq->pfil;
4717 	for (i = 0; i < NRXQSETS(ctx); i++, rxq++) {
4718 		rxq->pfil = NULL;
4719 	}
4720 	pfil_head_unregister(pfil);
4721 }
4722 
4723 
4724 /*
4725  * Advance forward by n members of the cpuset ctx->ifc_cpus starting from
4726  * cpuid and wrapping as necessary.
4727  */
4728 static unsigned int
4729 cpuid_advance(if_ctx_t ctx, unsigned int cpuid, unsigned int n)
4730 {
4731 	unsigned int first_valid;
4732 	unsigned int last_valid;
4733 
4734 	/* cpuid should always be in the valid set */
4735 	MPASS(CPU_ISSET(cpuid, &ctx->ifc_cpus));
4736 
4737 	/* valid set should never be empty */
4738 	MPASS(!CPU_EMPTY(&ctx->ifc_cpus));
4739 
4740 	first_valid = CPU_FFS(&ctx->ifc_cpus) - 1;
4741 	last_valid = CPU_FLS(&ctx->ifc_cpus) - 1;
4742 	n = n % CPU_COUNT(&ctx->ifc_cpus);
4743 	while (n > 0) {
4744 		do {
4745 			cpuid++;
4746 			if (cpuid > last_valid)
4747 				cpuid = first_valid;
4748 		} while (!CPU_ISSET(cpuid, &ctx->ifc_cpus));
4749 		n--;
4750 	}
4751 
4752 	return (cpuid);
4753 }
4754 
4755 #if defined(SMP) && defined(SCHED_ULE)
4756 extern struct cpu_group *cpu_top;              /* CPU topology */
4757 
4758 static int
4759 find_child_with_core(int cpu, struct cpu_group *grp)
4760 {
4761 	int i;
4762 
4763 	if (grp->cg_children == 0)
4764 		return -1;
4765 
4766 	MPASS(grp->cg_child);
4767 	for (i = 0; i < grp->cg_children; i++) {
4768 		if (CPU_ISSET(cpu, &grp->cg_child[i].cg_mask))
4769 			return i;
4770 	}
4771 
4772 	return -1;
4773 }
4774 
4775 
4776 /*
4777  * Find an L2 neighbor of the given CPU or return -1 if none found.  This
4778  * does not distinguish among multiple L2 neighbors if the given CPU has
4779  * more than one (it will always return the same result in that case).
4780  */
4781 static int
4782 find_l2_neighbor(int cpu)
4783 {
4784 	struct cpu_group *grp;
4785 	int i;
4786 
4787 	grp = cpu_top;
4788 	if (grp == NULL)
4789 		return -1;
4790 
4791 	/*
4792 	 * Find the smallest CPU group that contains the given core.
4793 	 */
4794 	i = 0;
4795 	while ((i = find_child_with_core(cpu, grp)) != -1) {
4796 		/*
4797 		 * If the smallest group containing the given CPU has less
4798 		 * than two members, we conclude the given CPU has no
4799 		 * L2 neighbor.
4800 		 */
4801 		if (grp->cg_child[i].cg_count <= 1)
4802 			return (-1);
4803 		grp = &grp->cg_child[i];
4804 	}
4805 
4806 	/* Must share L2. */
4807 	if (grp->cg_level > CG_SHARE_L2 || grp->cg_level == CG_SHARE_NONE)
4808 		return -1;
4809 
4810 	/*
4811 	 * Select the first member of the set that isn't the reference
4812 	 * CPU, which at this point is guaranteed to exist.
4813 	 */
4814 	for (i = 0; i < CPU_SETSIZE; i++) {
4815 		if (CPU_ISSET(i, &grp->cg_mask) && i != cpu)
4816 			return (i);
4817 	}
4818 
4819 	/* Should never be reached */
4820 	return (-1);
4821 }
4822 
4823 #else
4824 static int
4825 find_l2_neighbor(int cpu)
4826 {
4827 
4828 	return (-1);
4829 }
4830 #endif
4831 
4832 /*
4833  * CPU mapping behaviors
4834  * ---------------------
4835  * 'separate txrx' refers to the separate_txrx sysctl
4836  * 'use logical' refers to the use_logical_cores sysctl
4837  * 'INTR CPUS' indicates whether bus_get_cpus(INTR_CPUS) succeeded
4838  *
4839  *  separate     use     INTR
4840  *    txrx     logical   CPUS   result
4841  * ---------- --------- ------ ------------------------------------------------
4842  *     -          -       X     RX and TX queues mapped to consecutive physical
4843  *                              cores with RX/TX pairs on same core and excess
4844  *                              of either following
4845  *     -          X       X     RX and TX queues mapped to consecutive cores
4846  *                              of any type with RX/TX pairs on same core and
4847  *                              excess of either following
4848  *     X          -       X     RX and TX queues mapped to consecutive physical
4849  *                              cores; all RX then all TX
4850  *     X          X       X     RX queues mapped to consecutive physical cores
4851  *                              first, then TX queues mapped to L2 neighbor of
4852  *                              the corresponding RX queue if one exists,
4853  *                              otherwise to consecutive physical cores
4854  *     -         n/a      -     RX and TX queues mapped to consecutive cores of
4855  *                              any type with RX/TX pairs on same core and excess
4856  *                              of either following
4857  *     X         n/a      -     RX and TX queues mapped to consecutive cores of
4858  *                              any type; all RX then all TX
4859  */
4860 static unsigned int
4861 get_cpuid_for_queue(if_ctx_t ctx, unsigned int base_cpuid, unsigned int qid,
4862     bool is_tx)
4863 {
4864 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4865 	unsigned int core_index;
4866 
4867 	if (ctx->ifc_sysctl_separate_txrx) {
4868 		/*
4869 		 * When using separate CPUs for TX and RX, the assignment
4870 		 * will always be of a consecutive CPU out of the set of
4871 		 * context CPUs, except for the specific case where the
4872 		 * context CPUs are phsyical cores, the use of logical cores
4873 		 * has been enabled, the assignment is for TX, the TX qid
4874 		 * corresponds to an RX qid, and the CPU assigned to the
4875 		 * corresponding RX queue has an L2 neighbor.
4876 		 */
4877 		if (ctx->ifc_sysctl_use_logical_cores &&
4878 		    ctx->ifc_cpus_are_physical_cores &&
4879 		    is_tx && qid < scctx->isc_nrxqsets) {
4880 			int l2_neighbor;
4881 			unsigned int rx_cpuid;
4882 
4883 			rx_cpuid = cpuid_advance(ctx, base_cpuid, qid);
4884 			l2_neighbor = find_l2_neighbor(rx_cpuid);
4885 			if (l2_neighbor != -1) {
4886 				return (l2_neighbor);
4887 			}
4888 			/*
4889 			 * ... else fall through to the normal
4890 			 * consecutive-after-RX assignment scheme.
4891 			 *
4892 			 * Note that we are assuming that all RX queue CPUs
4893 			 * have an L2 neighbor, or all do not.  If a mixed
4894 			 * scenario is possible, we will have to keep track
4895 			 * separately of how many queues prior to this one
4896 			 * were not able to be assigned to an L2 neighbor.
4897 			 */
4898 		}
4899 		if (is_tx)
4900 			core_index = scctx->isc_nrxqsets + qid;
4901 		else
4902 			core_index = qid;
4903 	} else {
4904 		core_index = qid;
4905 	}
4906 
4907 	return (cpuid_advance(ctx, base_cpuid, core_index));
4908 }
4909 
4910 static uint16_t
4911 get_ctx_core_offset(if_ctx_t ctx)
4912 {
4913 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4914 	struct cpu_offset *op;
4915 	cpuset_t assigned_cpus;
4916 	unsigned int cores_consumed;
4917 	unsigned int base_cpuid = ctx->ifc_sysctl_core_offset;
4918 	unsigned int first_valid;
4919 	unsigned int last_valid;
4920 	unsigned int i;
4921 
4922 	first_valid = CPU_FFS(&ctx->ifc_cpus) - 1;
4923 	last_valid = CPU_FLS(&ctx->ifc_cpus) - 1;
4924 
4925 	if (base_cpuid != CORE_OFFSET_UNSPECIFIED) {
4926 		/*
4927 		 * Align the user-chosen base CPU ID to the next valid CPU
4928 		 * for this device.  If the chosen base CPU ID is smaller
4929 		 * than the first valid CPU or larger than the last valid
4930 		 * CPU, we assume the user does not know what the valid
4931 		 * range is for this device and is thinking in terms of a
4932 		 * zero-based reference frame, and so we shift the given
4933 		 * value into the valid range (and wrap accordingly) so the
4934 		 * intent is translated to the proper frame of reference.
4935 		 * If the base CPU ID is within the valid first/last, but
4936 		 * does not correspond to a valid CPU, it is advanced to the
4937 		 * next valid CPU (wrapping if necessary).
4938 		 */
4939 		if (base_cpuid < first_valid || base_cpuid > last_valid) {
4940 			/* shift from zero-based to first_valid-based */
4941 			base_cpuid += first_valid;
4942 			/* wrap to range [first_valid, last_valid] */
4943 			base_cpuid = (base_cpuid - first_valid) %
4944 			    (last_valid - first_valid + 1);
4945 		}
4946 		if (!CPU_ISSET(base_cpuid, &ctx->ifc_cpus)) {
4947 			/*
4948 			 * base_cpuid is in [first_valid, last_valid], but
4949 			 * not a member of the valid set.  In this case,
4950 			 * there will always be a member of the valid set
4951 			 * with a CPU ID that is greater than base_cpuid,
4952 			 * and we simply advance to it.
4953 			 */
4954 			while (!CPU_ISSET(base_cpuid, &ctx->ifc_cpus))
4955 				base_cpuid++;
4956 		}
4957 		return (base_cpuid);
4958 	}
4959 
4960 	/*
4961 	 * Determine how many cores will be consumed by performing the CPU
4962 	 * assignments and counting how many of the assigned CPUs correspond
4963 	 * to CPUs in the set of context CPUs.  This is done using the CPU
4964 	 * ID first_valid as the base CPU ID, as the base CPU must be within
4965 	 * the set of context CPUs.
4966 	 *
4967 	 * Note not all assigned CPUs will be in the set of context CPUs
4968 	 * when separate CPUs are being allocated to TX and RX queues,
4969 	 * assignment to logical cores has been enabled, the set of context
4970 	 * CPUs contains only physical CPUs, and TX queues are mapped to L2
4971 	 * neighbors of CPUs that RX queues have been mapped to - in this
4972 	 * case we do only want to count how many CPUs in the set of context
4973 	 * CPUs have been consumed, as that determines the next CPU in that
4974 	 * set to start allocating at for the next device for which
4975 	 * core_offset is not set.
4976 	 */
4977 	CPU_ZERO(&assigned_cpus);
4978 	for (i = 0; i < scctx->isc_ntxqsets; i++)
4979 		CPU_SET(get_cpuid_for_queue(ctx, first_valid, i, true),
4980 		    &assigned_cpus);
4981 	for (i = 0; i < scctx->isc_nrxqsets; i++)
4982 		CPU_SET(get_cpuid_for_queue(ctx, first_valid, i, false),
4983 		    &assigned_cpus);
4984 	CPU_AND(&assigned_cpus, &assigned_cpus, &ctx->ifc_cpus);
4985 	cores_consumed = CPU_COUNT(&assigned_cpus);
4986 
4987 	mtx_lock(&cpu_offset_mtx);
4988 	SLIST_FOREACH(op, &cpu_offsets, entries) {
4989 		if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) {
4990 			base_cpuid = op->next_cpuid;
4991 			op->next_cpuid = cpuid_advance(ctx, op->next_cpuid,
4992 			    cores_consumed);
4993 			MPASS(op->refcount < UINT_MAX);
4994 			op->refcount++;
4995 			break;
4996 		}
4997 	}
4998 	if (base_cpuid == CORE_OFFSET_UNSPECIFIED) {
4999 		base_cpuid = first_valid;
5000 		op = malloc(sizeof(struct cpu_offset), M_IFLIB,
5001 		    M_NOWAIT | M_ZERO);
5002 		if (op == NULL) {
5003 			device_printf(ctx->ifc_dev,
5004 			    "allocation for cpu offset failed.\n");
5005 		} else {
5006 			op->next_cpuid = cpuid_advance(ctx, base_cpuid,
5007 			    cores_consumed);
5008 			op->refcount = 1;
5009 			CPU_COPY(&ctx->ifc_cpus, &op->set);
5010 			SLIST_INSERT_HEAD(&cpu_offsets, op, entries);
5011 		}
5012 	}
5013 	mtx_unlock(&cpu_offset_mtx);
5014 
5015 	return (base_cpuid);
5016 }
5017 
5018 static void
5019 unref_ctx_core_offset(if_ctx_t ctx)
5020 {
5021 	struct cpu_offset *op, *top;
5022 
5023 	mtx_lock(&cpu_offset_mtx);
5024 	SLIST_FOREACH_SAFE(op, &cpu_offsets, entries, top) {
5025 		if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) {
5026 			MPASS(op->refcount > 0);
5027 			op->refcount--;
5028 			if (op->refcount == 0) {
5029 				SLIST_REMOVE(&cpu_offsets, op, cpu_offset, entries);
5030 				free(op, M_IFLIB);
5031 			}
5032 			break;
5033 		}
5034 	}
5035 	mtx_unlock(&cpu_offset_mtx);
5036 }
5037 
5038 int
5039 iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp)
5040 {
5041 	if_ctx_t ctx;
5042 	if_t ifp;
5043 	if_softc_ctx_t scctx;
5044 	kobjop_desc_t kobj_desc;
5045 	kobj_method_t *kobj_method;
5046 	int err, msix, rid;
5047 	int num_txd, num_rxd;
5048 
5049 	ctx = malloc(sizeof(* ctx), M_IFLIB, M_WAITOK|M_ZERO);
5050 
5051 	if (sc == NULL) {
5052 		sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
5053 		device_set_softc(dev, ctx);
5054 		ctx->ifc_flags |= IFC_SC_ALLOCATED;
5055 	}
5056 
5057 	ctx->ifc_sctx = sctx;
5058 	ctx->ifc_dev = dev;
5059 	ctx->ifc_softc = sc;
5060 
5061 	if ((err = iflib_register(ctx)) != 0) {
5062 		device_printf(dev, "iflib_register failed %d\n", err);
5063 		goto fail_ctx_free;
5064 	}
5065 	iflib_add_device_sysctl_pre(ctx);
5066 
5067 	scctx = &ctx->ifc_softc_ctx;
5068 	ifp = ctx->ifc_ifp;
5069 
5070 	iflib_reset_qvalues(ctx);
5071 	IFNET_WLOCK();
5072 	CTX_LOCK(ctx);
5073 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
5074 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
5075 		goto fail_unlock;
5076 	}
5077 	_iflib_pre_assert(scctx);
5078 	ctx->ifc_txrx = *scctx->isc_txrx;
5079 
5080 	MPASS(scctx->isc_dma_width <= flsll(BUS_SPACE_MAXADDR));
5081 
5082 	if (sctx->isc_flags & IFLIB_DRIVER_MEDIA)
5083 		ctx->ifc_mediap = scctx->isc_media;
5084 
5085 #ifdef INVARIANTS
5086 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
5087 		MPASS(scctx->isc_tx_csum_flags);
5088 #endif
5089 
5090 	if_setcapabilities(ifp,
5091 	    scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_MEXTPG);
5092 	if_setcapenable(ifp,
5093 	    scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_MEXTPG);
5094 
5095 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
5096 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
5097 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
5098 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
5099 
5100 	num_txd = iflib_num_tx_descs(ctx);
5101 	num_rxd = iflib_num_rx_descs(ctx);
5102 
5103 	/* XXX change for per-queue sizes */
5104 	device_printf(dev, "Using %d TX descriptors and %d RX descriptors\n",
5105 	    num_txd, num_rxd);
5106 
5107 	if (scctx->isc_tx_nsegments > num_txd / MAX_SINGLE_PACKET_FRACTION)
5108 		scctx->isc_tx_nsegments = max(1, num_txd /
5109 		    MAX_SINGLE_PACKET_FRACTION);
5110 	if (scctx->isc_tx_tso_segments_max > num_txd /
5111 	    MAX_SINGLE_PACKET_FRACTION)
5112 		scctx->isc_tx_tso_segments_max = max(1,
5113 		    num_txd / MAX_SINGLE_PACKET_FRACTION);
5114 
5115 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
5116 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
5117 		/*
5118 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
5119 		 * but some MACs do.
5120 		 */
5121 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
5122 		    IP_MAXPACKET));
5123 		/*
5124 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
5125 		 * into account.  In the worst case, each of these calls will
5126 		 * add another mbuf and, thus, the requirement for another DMA
5127 		 * segment.  So for best performance, it doesn't make sense to
5128 		 * advertize a maximum of TSO segments that typically will
5129 		 * require defragmentation in iflib_encap().
5130 		 */
5131 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
5132 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
5133 	}
5134 	if (scctx->isc_rss_table_size == 0)
5135 		scctx->isc_rss_table_size = 64;
5136 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
5137 
5138 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
5139 	/* XXX format name */
5140 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx,
5141 	    NULL, NULL, "admin");
5142 
5143 	/* Set up cpu set.  If it fails, use the set of all CPUs. */
5144 	if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) != 0) {
5145 		device_printf(dev, "Unable to fetch CPU list\n");
5146 		CPU_COPY(&all_cpus, &ctx->ifc_cpus);
5147 		ctx->ifc_cpus_are_physical_cores = false;
5148 	} else
5149 		ctx->ifc_cpus_are_physical_cores = true;
5150 	MPASS(CPU_COUNT(&ctx->ifc_cpus) > 0);
5151 
5152 	/*
5153 	** Now set up MSI or MSI-X, should return us the number of supported
5154 	** vectors (will be 1 for a legacy interrupt and MSI).
5155 	*/
5156 	if (sctx->isc_flags & IFLIB_SKIP_MSIX) {
5157 		msix = scctx->isc_vectors;
5158 	} else if (scctx->isc_msix_bar != 0)
5159 	       /*
5160 		* The simple fact that isc_msix_bar is not 0 does not mean we
5161 		* we have a good value there that is known to work.
5162 		*/
5163 		msix = iflib_msix_init(ctx);
5164 	else {
5165 		scctx->isc_vectors = 1;
5166 		scctx->isc_ntxqsets = 1;
5167 		scctx->isc_nrxqsets = 1;
5168 		scctx->isc_intr = IFLIB_INTR_LEGACY;
5169 		msix = 0;
5170 	}
5171 	/* Get memory for the station queues */
5172 	if ((err = iflib_queues_alloc(ctx))) {
5173 		device_printf(dev, "Unable to allocate queue memory\n");
5174 		goto fail_intr_free;
5175 	}
5176 
5177 	if ((err = iflib_qset_structures_setup(ctx)))
5178 		goto fail_queues;
5179 
5180 	/*
5181 	 * Now that we know how many queues there are, get the core offset.
5182 	 */
5183 	ctx->ifc_sysctl_core_offset = get_ctx_core_offset(ctx);
5184 
5185 	if (msix > 1) {
5186 		/*
5187 		 * When using MSI-X, ensure that ifdi_{r,t}x_queue_intr_enable
5188 		 * aren't the default NULL implementation.
5189 		 */
5190 		kobj_desc = &ifdi_rx_queue_intr_enable_desc;
5191 		kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL,
5192 		    kobj_desc);
5193 		if (kobj_method == &kobj_desc->deflt) {
5194 			device_printf(dev,
5195 			    "MSI-X requires ifdi_rx_queue_intr_enable method");
5196 			err = EOPNOTSUPP;
5197 			goto fail_queues;
5198 		}
5199 		kobj_desc = &ifdi_tx_queue_intr_enable_desc;
5200 		kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL,
5201 		    kobj_desc);
5202 		if (kobj_method == &kobj_desc->deflt) {
5203 			device_printf(dev,
5204 			    "MSI-X requires ifdi_tx_queue_intr_enable method");
5205 			err = EOPNOTSUPP;
5206 			goto fail_queues;
5207 		}
5208 
5209 		/*
5210 		 * Assign the MSI-X vectors.
5211 		 * Note that the default NULL ifdi_msix_intr_assign method will
5212 		 * fail here, too.
5213 		 */
5214 		err = IFDI_MSIX_INTR_ASSIGN(ctx, msix);
5215 		if (err != 0) {
5216 			device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n",
5217 			    err);
5218 			goto fail_queues;
5219 		}
5220 	} else if (scctx->isc_intr != IFLIB_INTR_MSIX) {
5221 		rid = 0;
5222 		if (scctx->isc_intr == IFLIB_INTR_MSI) {
5223 			MPASS(msix == 1);
5224 			rid = 1;
5225 		}
5226 		if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) {
5227 			device_printf(dev, "iflib_legacy_setup failed %d\n", err);
5228 			goto fail_queues;
5229 		}
5230 	} else {
5231 		device_printf(dev,
5232 		    "Cannot use iflib with only 1 MSI-X interrupt!\n");
5233 		err = ENODEV;
5234 		goto fail_queues;
5235 	}
5236 
5237 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
5238 
5239 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
5240 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
5241 		goto fail_detach;
5242 	}
5243 
5244 	/*
5245 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
5246 	 * This must appear after the call to ether_ifattach() because
5247 	 * ether_ifattach() sets if_hdrlen to the default value.
5248 	 */
5249 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
5250 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
5251 
5252 	if ((err = iflib_netmap_attach(ctx))) {
5253 		device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err);
5254 		goto fail_detach;
5255 	}
5256 	*ctxp = ctx;
5257 
5258 	DEBUGNET_SET(ctx->ifc_ifp, iflib);
5259 
5260 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
5261 	iflib_add_device_sysctl_post(ctx);
5262 	iflib_add_pfil(ctx);
5263 	ctx->ifc_flags |= IFC_INIT_DONE;
5264 	CTX_UNLOCK(ctx);
5265 	IFNET_WUNLOCK();
5266 
5267 	return (0);
5268 
5269 fail_detach:
5270 	ether_ifdetach(ctx->ifc_ifp);
5271 fail_queues:
5272 	iflib_tqg_detach(ctx);
5273 	iflib_tx_structures_free(ctx);
5274 	iflib_rx_structures_free(ctx);
5275 	IFDI_DETACH(ctx);
5276 	IFDI_QUEUES_FREE(ctx);
5277 fail_intr_free:
5278 	iflib_free_intr_mem(ctx);
5279 fail_unlock:
5280 	CTX_UNLOCK(ctx);
5281 	IFNET_WUNLOCK();
5282 	iflib_deregister(ctx);
5283 fail_ctx_free:
5284 	device_set_softc(ctx->ifc_dev, NULL);
5285         if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5286                 free(ctx->ifc_softc, M_IFLIB);
5287         free(ctx, M_IFLIB);
5288 	return (err);
5289 }
5290 
5291 int
5292 iflib_pseudo_register(device_t dev, if_shared_ctx_t sctx, if_ctx_t *ctxp,
5293 					  struct iflib_cloneattach_ctx *clctx)
5294 {
5295 	int num_txd, num_rxd;
5296 	int err;
5297 	if_ctx_t ctx;
5298 	if_t ifp;
5299 	if_softc_ctx_t scctx;
5300 	int i;
5301 	void *sc;
5302 
5303 	ctx = malloc(sizeof(*ctx), M_IFLIB, M_WAITOK|M_ZERO);
5304 	sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
5305 	ctx->ifc_flags |= IFC_SC_ALLOCATED;
5306 	if (sctx->isc_flags & (IFLIB_PSEUDO|IFLIB_VIRTUAL))
5307 		ctx->ifc_flags |= IFC_PSEUDO;
5308 
5309 	ctx->ifc_sctx = sctx;
5310 	ctx->ifc_softc = sc;
5311 	ctx->ifc_dev = dev;
5312 
5313 	if ((err = iflib_register(ctx)) != 0) {
5314 		device_printf(dev, "%s: iflib_register failed %d\n", __func__, err);
5315 		goto fail_ctx_free;
5316 	}
5317 	iflib_add_device_sysctl_pre(ctx);
5318 
5319 	scctx = &ctx->ifc_softc_ctx;
5320 	ifp = ctx->ifc_ifp;
5321 
5322 	iflib_reset_qvalues(ctx);
5323 	CTX_LOCK(ctx);
5324 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
5325 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
5326 		goto fail_unlock;
5327 	}
5328 	if (sctx->isc_flags & IFLIB_GEN_MAC)
5329 		ether_gen_addr(ifp, &ctx->ifc_mac);
5330 	if ((err = IFDI_CLONEATTACH(ctx, clctx->cc_ifc, clctx->cc_name,
5331 								clctx->cc_params)) != 0) {
5332 		device_printf(dev, "IFDI_CLONEATTACH failed %d\n", err);
5333 		goto fail_unlock;
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, scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_LINKSTATE);
5341 	if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_LINKSTATE);
5342 
5343 	ifp->if_flags |= IFF_NOGROUP;
5344 	if (sctx->isc_flags & IFLIB_PSEUDO) {
5345 		ifmedia_add(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO, 0, NULL);
5346 		ifmedia_set(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO);
5347 		if (sctx->isc_flags & IFLIB_PSEUDO_ETHER) {
5348 			ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
5349 		} else {
5350 			if_attach(ctx->ifc_ifp);
5351 			bpfattach(ctx->ifc_ifp, DLT_NULL, sizeof(u_int32_t));
5352 		}
5353 
5354 		if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
5355 			device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
5356 			goto fail_detach;
5357 		}
5358 		*ctxp = ctx;
5359 
5360 		/*
5361 		 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
5362 		 * This must appear after the call to ether_ifattach() because
5363 		 * ether_ifattach() sets if_hdrlen to the default value.
5364 		 */
5365 		if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
5366 			if_setifheaderlen(ifp,
5367 			    sizeof(struct ether_vlan_header));
5368 
5369 		if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
5370 		iflib_add_device_sysctl_post(ctx);
5371 		ctx->ifc_flags |= IFC_INIT_DONE;
5372 		CTX_UNLOCK(ctx);
5373 		return (0);
5374 	}
5375 	ifmedia_add(ctx->ifc_mediap, IFM_ETHER | IFM_1000_T | IFM_FDX, 0, NULL);
5376 	ifmedia_add(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO, 0, NULL);
5377 	ifmedia_set(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO);
5378 
5379 	_iflib_pre_assert(scctx);
5380 	ctx->ifc_txrx = *scctx->isc_txrx;
5381 
5382 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
5383 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
5384 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
5385 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
5386 
5387 	num_txd = iflib_num_tx_descs(ctx);
5388 	num_rxd = iflib_num_rx_descs(ctx);
5389 
5390 	/* XXX change for per-queue sizes */
5391 	device_printf(dev, "Using %d TX descriptors and %d RX descriptors\n",
5392 	    num_txd, num_rxd);
5393 
5394 	if (scctx->isc_tx_nsegments > num_txd / MAX_SINGLE_PACKET_FRACTION)
5395 		scctx->isc_tx_nsegments = max(1, num_txd /
5396 		    MAX_SINGLE_PACKET_FRACTION);
5397 	if (scctx->isc_tx_tso_segments_max > num_txd /
5398 	    MAX_SINGLE_PACKET_FRACTION)
5399 		scctx->isc_tx_tso_segments_max = max(1,
5400 		    num_txd / MAX_SINGLE_PACKET_FRACTION);
5401 
5402 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
5403 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
5404 		/*
5405 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
5406 		 * but some MACs do.
5407 		 */
5408 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
5409 		    IP_MAXPACKET));
5410 		/*
5411 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
5412 		 * into account.  In the worst case, each of these calls will
5413 		 * add another mbuf and, thus, the requirement for another DMA
5414 		 * segment.  So for best performance, it doesn't make sense to
5415 		 * advertize a maximum of TSO segments that typically will
5416 		 * require defragmentation in iflib_encap().
5417 		 */
5418 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
5419 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
5420 	}
5421 	if (scctx->isc_rss_table_size == 0)
5422 		scctx->isc_rss_table_size = 64;
5423 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
5424 
5425 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
5426 	/* XXX format name */
5427 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx,
5428 	    NULL, NULL, "admin");
5429 
5430 	/* XXX --- can support > 1 -- but keep it simple for now */
5431 	scctx->isc_intr = IFLIB_INTR_LEGACY;
5432 
5433 	/* Get memory for the station queues */
5434 	if ((err = iflib_queues_alloc(ctx))) {
5435 		device_printf(dev, "Unable to allocate queue memory\n");
5436 		goto fail_iflib_detach;
5437 	}
5438 
5439 	if ((err = iflib_qset_structures_setup(ctx))) {
5440 		device_printf(dev, "qset structure setup failed %d\n", err);
5441 		goto fail_queues;
5442 	}
5443 
5444 	/*
5445 	 * XXX What if anything do we want to do about interrupts?
5446 	 */
5447 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
5448 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
5449 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
5450 		goto fail_detach;
5451 	}
5452 
5453 	/*
5454 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
5455 	 * This must appear after the call to ether_ifattach() because
5456 	 * ether_ifattach() sets if_hdrlen to the default value.
5457 	 */
5458 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
5459 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
5460 
5461 	/* XXX handle more than one queue */
5462 	for (i = 0; i < scctx->isc_nrxqsets; i++)
5463 		IFDI_RX_CLSET(ctx, 0, i, ctx->ifc_rxqs[i].ifr_fl[0].ifl_sds.ifsd_cl);
5464 
5465 	*ctxp = ctx;
5466 
5467 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
5468 	iflib_add_device_sysctl_post(ctx);
5469 	ctx->ifc_flags |= IFC_INIT_DONE;
5470 	CTX_UNLOCK(ctx);
5471 
5472 	return (0);
5473 fail_detach:
5474 	ether_ifdetach(ctx->ifc_ifp);
5475 fail_queues:
5476 	iflib_tqg_detach(ctx);
5477 	iflib_tx_structures_free(ctx);
5478 	iflib_rx_structures_free(ctx);
5479 fail_iflib_detach:
5480 	IFDI_DETACH(ctx);
5481 	IFDI_QUEUES_FREE(ctx);
5482 fail_unlock:
5483 	CTX_UNLOCK(ctx);
5484 	iflib_deregister(ctx);
5485 fail_ctx_free:
5486 	free(ctx->ifc_softc, M_IFLIB);
5487 	free(ctx, M_IFLIB);
5488 	return (err);
5489 }
5490 
5491 int
5492 iflib_pseudo_deregister(if_ctx_t ctx)
5493 {
5494 	if_t ifp = ctx->ifc_ifp;
5495 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5496 
5497 	/* Unregister VLAN event handlers early */
5498 	iflib_unregister_vlan_handlers(ctx);
5499 
5500 	if ((sctx->isc_flags & IFLIB_PSEUDO)  &&
5501 		(sctx->isc_flags & IFLIB_PSEUDO_ETHER) == 0) {
5502 		bpfdetach(ifp);
5503 		if_detach(ifp);
5504 	} else {
5505 		ether_ifdetach(ifp);
5506 	}
5507 
5508 	iflib_tqg_detach(ctx);
5509 	iflib_tx_structures_free(ctx);
5510 	iflib_rx_structures_free(ctx);
5511 	IFDI_DETACH(ctx);
5512 	IFDI_QUEUES_FREE(ctx);
5513 
5514 	iflib_deregister(ctx);
5515 
5516 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5517 		free(ctx->ifc_softc, M_IFLIB);
5518 	free(ctx, M_IFLIB);
5519 	return (0);
5520 }
5521 
5522 int
5523 iflib_device_attach(device_t dev)
5524 {
5525 	if_ctx_t ctx;
5526 	if_shared_ctx_t sctx;
5527 
5528 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
5529 		return (ENOTSUP);
5530 
5531 	pci_enable_busmaster(dev);
5532 
5533 	return (iflib_device_register(dev, NULL, sctx, &ctx));
5534 }
5535 
5536 int
5537 iflib_device_deregister(if_ctx_t ctx)
5538 {
5539 	if_t ifp = ctx->ifc_ifp;
5540 	device_t dev = ctx->ifc_dev;
5541 
5542 	/* Make sure VLANS are not using driver */
5543 	if (if_vlantrunkinuse(ifp)) {
5544 		device_printf(dev, "Vlan in use, detach first\n");
5545 		return (EBUSY);
5546 	}
5547 #ifdef PCI_IOV
5548 	if (!CTX_IS_VF(ctx) && pci_iov_detach(dev) != 0) {
5549 		device_printf(dev, "SR-IOV in use; detach first.\n");
5550 		return (EBUSY);
5551 	}
5552 #endif
5553 
5554 	STATE_LOCK(ctx);
5555 	ctx->ifc_flags |= IFC_IN_DETACH;
5556 	STATE_UNLOCK(ctx);
5557 
5558 	/* Unregister VLAN handlers before calling iflib_stop() */
5559 	iflib_unregister_vlan_handlers(ctx);
5560 
5561 	iflib_netmap_detach(ifp);
5562 	ether_ifdetach(ifp);
5563 
5564 	CTX_LOCK(ctx);
5565 	iflib_stop(ctx);
5566 	CTX_UNLOCK(ctx);
5567 
5568 	iflib_rem_pfil(ctx);
5569 	if (ctx->ifc_led_dev != NULL)
5570 		led_destroy(ctx->ifc_led_dev);
5571 
5572 	iflib_tqg_detach(ctx);
5573 	iflib_tx_structures_free(ctx);
5574 	iflib_rx_structures_free(ctx);
5575 
5576 	CTX_LOCK(ctx);
5577 	IFDI_DETACH(ctx);
5578 	IFDI_QUEUES_FREE(ctx);
5579 	CTX_UNLOCK(ctx);
5580 
5581 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
5582 	iflib_free_intr_mem(ctx);
5583 
5584 	bus_generic_detach(dev);
5585 
5586 	iflib_deregister(ctx);
5587 
5588 	device_set_softc(ctx->ifc_dev, NULL);
5589 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5590 		free(ctx->ifc_softc, M_IFLIB);
5591 	unref_ctx_core_offset(ctx);
5592 	free(ctx, M_IFLIB);
5593 	return (0);
5594 }
5595 
5596 static void
5597 iflib_tqg_detach(if_ctx_t ctx)
5598 {
5599 	iflib_txq_t txq;
5600 	iflib_rxq_t rxq;
5601 	int i;
5602 	struct taskqgroup *tqg;
5603 
5604 	/* XXX drain any dependent tasks */
5605 	tqg = qgroup_if_io_tqg;
5606 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
5607 		callout_drain(&txq->ift_timer);
5608 #ifdef DEV_NETMAP
5609 		callout_drain(&txq->ift_netmap_timer);
5610 #endif /* DEV_NETMAP */
5611 		if (txq->ift_task.gt_uniq != NULL)
5612 			taskqgroup_detach(tqg, &txq->ift_task);
5613 	}
5614 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
5615 		if (rxq->ifr_task.gt_uniq != NULL)
5616 			taskqgroup_detach(tqg, &rxq->ifr_task);
5617 	}
5618 	tqg = qgroup_if_config_tqg;
5619 	if (ctx->ifc_admin_task.gt_uniq != NULL)
5620 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
5621 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
5622 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
5623 }
5624 
5625 static void
5626 iflib_free_intr_mem(if_ctx_t ctx)
5627 {
5628 
5629 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) {
5630 		iflib_irq_free(ctx, &ctx->ifc_legacy_irq);
5631 	}
5632 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) {
5633 		pci_release_msi(ctx->ifc_dev);
5634 	}
5635 	if (ctx->ifc_msix_mem != NULL) {
5636 		bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY,
5637 		    rman_get_rid(ctx->ifc_msix_mem), ctx->ifc_msix_mem);
5638 		ctx->ifc_msix_mem = NULL;
5639 	}
5640 }
5641 
5642 int
5643 iflib_device_detach(device_t dev)
5644 {
5645 	if_ctx_t ctx = device_get_softc(dev);
5646 
5647 	return (iflib_device_deregister(ctx));
5648 }
5649 
5650 int
5651 iflib_device_suspend(device_t dev)
5652 {
5653 	if_ctx_t ctx = device_get_softc(dev);
5654 
5655 	CTX_LOCK(ctx);
5656 	IFDI_SUSPEND(ctx);
5657 	CTX_UNLOCK(ctx);
5658 
5659 	return bus_generic_suspend(dev);
5660 }
5661 int
5662 iflib_device_shutdown(device_t dev)
5663 {
5664 	if_ctx_t ctx = device_get_softc(dev);
5665 
5666 	CTX_LOCK(ctx);
5667 	IFDI_SHUTDOWN(ctx);
5668 	CTX_UNLOCK(ctx);
5669 
5670 	return bus_generic_suspend(dev);
5671 }
5672 
5673 int
5674 iflib_device_resume(device_t dev)
5675 {
5676 	if_ctx_t ctx = device_get_softc(dev);
5677 	iflib_txq_t txq = ctx->ifc_txqs;
5678 
5679 	CTX_LOCK(ctx);
5680 	IFDI_RESUME(ctx);
5681 	iflib_if_init_locked(ctx);
5682 	CTX_UNLOCK(ctx);
5683 	for (int i = 0; i < NTXQSETS(ctx); i++, txq++)
5684 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
5685 
5686 	return (bus_generic_resume(dev));
5687 }
5688 
5689 int
5690 iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params)
5691 {
5692 	int error;
5693 	if_ctx_t ctx = device_get_softc(dev);
5694 
5695 	CTX_LOCK(ctx);
5696 	error = IFDI_IOV_INIT(ctx, num_vfs, params);
5697 	CTX_UNLOCK(ctx);
5698 
5699 	return (error);
5700 }
5701 
5702 void
5703 iflib_device_iov_uninit(device_t dev)
5704 {
5705 	if_ctx_t ctx = device_get_softc(dev);
5706 
5707 	CTX_LOCK(ctx);
5708 	IFDI_IOV_UNINIT(ctx);
5709 	CTX_UNLOCK(ctx);
5710 }
5711 
5712 int
5713 iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params)
5714 {
5715 	int error;
5716 	if_ctx_t ctx = device_get_softc(dev);
5717 
5718 	CTX_LOCK(ctx);
5719 	error = IFDI_IOV_VF_ADD(ctx, vfnum, params);
5720 	CTX_UNLOCK(ctx);
5721 
5722 	return (error);
5723 }
5724 
5725 /*********************************************************************
5726  *
5727  *  MODULE FUNCTION DEFINITIONS
5728  *
5729  **********************************************************************/
5730 
5731 /*
5732  * - Start a fast taskqueue thread for each core
5733  * - Start a taskqueue for control operations
5734  */
5735 static int
5736 iflib_module_init(void)
5737 {
5738 	iflib_timer_default = hz / 2;
5739 	return (0);
5740 }
5741 
5742 static int
5743 iflib_module_event_handler(module_t mod, int what, void *arg)
5744 {
5745 	int err;
5746 
5747 	switch (what) {
5748 	case MOD_LOAD:
5749 		if ((err = iflib_module_init()) != 0)
5750 			return (err);
5751 		break;
5752 	case MOD_UNLOAD:
5753 		return (EBUSY);
5754 	default:
5755 		return (EOPNOTSUPP);
5756 	}
5757 
5758 	return (0);
5759 }
5760 
5761 /*********************************************************************
5762  *
5763  *  PUBLIC FUNCTION DEFINITIONS
5764  *     ordered as in iflib.h
5765  *
5766  **********************************************************************/
5767 
5768 static void
5769 _iflib_assert(if_shared_ctx_t sctx)
5770 {
5771 	int i;
5772 
5773 	MPASS(sctx->isc_tx_maxsize);
5774 	MPASS(sctx->isc_tx_maxsegsize);
5775 
5776 	MPASS(sctx->isc_rx_maxsize);
5777 	MPASS(sctx->isc_rx_nsegments);
5778 	MPASS(sctx->isc_rx_maxsegsize);
5779 
5780 	MPASS(sctx->isc_nrxqs >= 1 && sctx->isc_nrxqs <= 8);
5781 	for (i = 0; i < sctx->isc_nrxqs; i++) {
5782 		MPASS(sctx->isc_nrxd_min[i]);
5783 		MPASS(powerof2(sctx->isc_nrxd_min[i]));
5784 		MPASS(sctx->isc_nrxd_max[i]);
5785 		MPASS(powerof2(sctx->isc_nrxd_max[i]));
5786 		MPASS(sctx->isc_nrxd_default[i]);
5787 		MPASS(powerof2(sctx->isc_nrxd_default[i]));
5788 	}
5789 
5790 	MPASS(sctx->isc_ntxqs >= 1 && sctx->isc_ntxqs <= 8);
5791 	for (i = 0; i < sctx->isc_ntxqs; i++) {
5792 		MPASS(sctx->isc_ntxd_min[i]);
5793 		MPASS(powerof2(sctx->isc_ntxd_min[i]));
5794 		MPASS(sctx->isc_ntxd_max[i]);
5795 		MPASS(powerof2(sctx->isc_ntxd_max[i]));
5796 		MPASS(sctx->isc_ntxd_default[i]);
5797 		MPASS(powerof2(sctx->isc_ntxd_default[i]));
5798 	}
5799 }
5800 
5801 static void
5802 _iflib_pre_assert(if_softc_ctx_t scctx)
5803 {
5804 
5805 	MPASS(scctx->isc_txrx->ift_txd_encap);
5806 	MPASS(scctx->isc_txrx->ift_txd_flush);
5807 	MPASS(scctx->isc_txrx->ift_txd_credits_update);
5808 	MPASS(scctx->isc_txrx->ift_rxd_available);
5809 	MPASS(scctx->isc_txrx->ift_rxd_pkt_get);
5810 	MPASS(scctx->isc_txrx->ift_rxd_refill);
5811 	MPASS(scctx->isc_txrx->ift_rxd_flush);
5812 }
5813 
5814 static int
5815 iflib_register(if_ctx_t ctx)
5816 {
5817 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5818 	driver_t *driver = sctx->isc_driver;
5819 	device_t dev = ctx->ifc_dev;
5820 	if_t ifp;
5821 	u_char type;
5822 	int iflags;
5823 
5824 	if ((sctx->isc_flags & IFLIB_PSEUDO) == 0)
5825 		_iflib_assert(sctx);
5826 
5827 	CTX_LOCK_INIT(ctx);
5828 	STATE_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev));
5829 	if (sctx->isc_flags & IFLIB_PSEUDO) {
5830 		if (sctx->isc_flags & IFLIB_PSEUDO_ETHER)
5831 			type = IFT_ETHER;
5832 		else
5833 			type = IFT_PPP;
5834 	} else
5835 		type = IFT_ETHER;
5836 	ifp = ctx->ifc_ifp = if_alloc(type);
5837 	if (ifp == NULL) {
5838 		device_printf(dev, "can not allocate ifnet structure\n");
5839 		return (ENOMEM);
5840 	}
5841 
5842 	/*
5843 	 * Initialize our context's device specific methods
5844 	 */
5845 	kobj_init((kobj_t) ctx, (kobj_class_t) driver);
5846 	kobj_class_compile((kobj_class_t) driver);
5847 
5848 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
5849 	if_setsoftc(ifp, ctx);
5850 	if_setdev(ifp, dev);
5851 	if_setinitfn(ifp, iflib_if_init);
5852 	if_setioctlfn(ifp, iflib_if_ioctl);
5853 #ifdef ALTQ
5854 	if_setstartfn(ifp, iflib_altq_if_start);
5855 	if_settransmitfn(ifp, iflib_altq_if_transmit);
5856 	if_setsendqready(ifp);
5857 #else
5858 	if_settransmitfn(ifp, iflib_if_transmit);
5859 #endif
5860 	if_setqflushfn(ifp, iflib_if_qflush);
5861 	iflags = IFF_MULTICAST | IFF_KNOWSEPOCH;
5862 
5863 	if ((sctx->isc_flags & IFLIB_PSEUDO) &&
5864 		(sctx->isc_flags & IFLIB_PSEUDO_ETHER) == 0)
5865 		iflags |= IFF_POINTOPOINT;
5866 	else
5867 		iflags |= IFF_BROADCAST | IFF_SIMPLEX;
5868 	if_setflags(ifp, iflags);
5869 	ctx->ifc_vlan_attach_event =
5870 		EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx,
5871 							  EVENTHANDLER_PRI_FIRST);
5872 	ctx->ifc_vlan_detach_event =
5873 		EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx,
5874 							  EVENTHANDLER_PRI_FIRST);
5875 
5876 	if ((sctx->isc_flags & IFLIB_DRIVER_MEDIA) == 0) {
5877 		ctx->ifc_mediap = &ctx->ifc_media;
5878 		ifmedia_init(ctx->ifc_mediap, IFM_IMASK,
5879 		    iflib_media_change, iflib_media_status);
5880 	}
5881 	return (0);
5882 }
5883 
5884 static void
5885 iflib_unregister_vlan_handlers(if_ctx_t ctx)
5886 {
5887 	/* Unregister VLAN events */
5888 	if (ctx->ifc_vlan_attach_event != NULL) {
5889 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
5890 		ctx->ifc_vlan_attach_event = NULL;
5891 	}
5892 	if (ctx->ifc_vlan_detach_event != NULL) {
5893 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
5894 		ctx->ifc_vlan_detach_event = NULL;
5895 	}
5896 
5897 }
5898 
5899 static void
5900 iflib_deregister(if_ctx_t ctx)
5901 {
5902 	if_t ifp = ctx->ifc_ifp;
5903 
5904 	/* Remove all media */
5905 	ifmedia_removeall(&ctx->ifc_media);
5906 
5907 	/* Ensure that VLAN event handlers are unregistered */
5908 	iflib_unregister_vlan_handlers(ctx);
5909 
5910 	/* Release kobject reference */
5911 	kobj_delete((kobj_t) ctx, NULL);
5912 
5913 	/* Free the ifnet structure */
5914 	if_free(ifp);
5915 
5916 	STATE_LOCK_DESTROY(ctx);
5917 
5918 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
5919 	CTX_LOCK_DESTROY(ctx);
5920 }
5921 
5922 static int
5923 iflib_queues_alloc(if_ctx_t ctx)
5924 {
5925 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5926 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5927 	device_t dev = ctx->ifc_dev;
5928 	int nrxqsets = scctx->isc_nrxqsets;
5929 	int ntxqsets = scctx->isc_ntxqsets;
5930 	iflib_txq_t txq;
5931 	iflib_rxq_t rxq;
5932 	iflib_fl_t fl = NULL;
5933 	int i, j, cpu, err, txconf, rxconf;
5934 	iflib_dma_info_t ifdip;
5935 	uint32_t *rxqsizes = scctx->isc_rxqsizes;
5936 	uint32_t *txqsizes = scctx->isc_txqsizes;
5937 	uint8_t nrxqs = sctx->isc_nrxqs;
5938 	uint8_t ntxqs = sctx->isc_ntxqs;
5939 	int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1;
5940 	int fl_offset = (sctx->isc_flags & IFLIB_HAS_RXCQ ? 1 : 0);
5941 	caddr_t *vaddrs;
5942 	uint64_t *paddrs;
5943 
5944 	KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1"));
5945 	KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1"));
5946 	KASSERT(nrxqs >= fl_offset + nfree_lists,
5947            ("there must be at least a rxq for each free list"));
5948 
5949 	/* Allocate the TX ring struct memory */
5950 	if (!(ctx->ifc_txqs =
5951 	    (iflib_txq_t) malloc(sizeof(struct iflib_txq) *
5952 	    ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5953 		device_printf(dev, "Unable to allocate TX ring memory\n");
5954 		err = ENOMEM;
5955 		goto fail;
5956 	}
5957 
5958 	/* Now allocate the RX */
5959 	if (!(ctx->ifc_rxqs =
5960 	    (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) *
5961 	    nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5962 		device_printf(dev, "Unable to allocate RX ring memory\n");
5963 		err = ENOMEM;
5964 		goto rx_fail;
5965 	}
5966 
5967 	txq = ctx->ifc_txqs;
5968 	rxq = ctx->ifc_rxqs;
5969 
5970 	/*
5971 	 * XXX handle allocation failure
5972 	 */
5973 	for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) {
5974 		/* Set up some basics */
5975 
5976 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs,
5977 		    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
5978 			device_printf(dev,
5979 			    "Unable to allocate TX DMA info memory\n");
5980 			err = ENOMEM;
5981 			goto err_tx_desc;
5982 		}
5983 		txq->ift_ifdi = ifdip;
5984 		for (j = 0; j < ntxqs; j++, ifdip++) {
5985 			if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, 0)) {
5986 				device_printf(dev,
5987 				    "Unable to allocate TX descriptors\n");
5988 				err = ENOMEM;
5989 				goto err_tx_desc;
5990 			}
5991 			txq->ift_txd_size[j] = scctx->isc_txd_size[j];
5992 			bzero((void *)ifdip->idi_vaddr, txqsizes[j]);
5993 		}
5994 		txq->ift_ctx = ctx;
5995 		txq->ift_id = i;
5996 		if (sctx->isc_flags & IFLIB_HAS_TXCQ) {
5997 			txq->ift_br_offset = 1;
5998 		} else {
5999 			txq->ift_br_offset = 0;
6000 		}
6001 
6002 		if (iflib_txsd_alloc(txq)) {
6003 			device_printf(dev, "Critical Failure setting up TX buffers\n");
6004 			err = ENOMEM;
6005 			goto err_tx_desc;
6006 		}
6007 
6008 		/* Initialize the TX lock */
6009 		snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:TX(%d):callout",
6010 		    device_get_nameunit(dev), txq->ift_id);
6011 		mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF);
6012 		callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0);
6013 		txq->ift_timer.c_cpu = cpu;
6014 #ifdef DEV_NETMAP
6015 		callout_init_mtx(&txq->ift_netmap_timer, &txq->ift_mtx, 0);
6016 		txq->ift_netmap_timer.c_cpu = cpu;
6017 #endif /* DEV_NETMAP */
6018 
6019 		err = ifmp_ring_alloc(&txq->ift_br, 2048, txq, iflib_txq_drain,
6020 				      iflib_txq_can_drain, M_IFLIB, M_WAITOK);
6021 		if (err) {
6022 			/* XXX free any allocated rings */
6023 			device_printf(dev, "Unable to allocate buf_ring\n");
6024 			goto err_tx_desc;
6025 		}
6026 	}
6027 
6028 	for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) {
6029 		/* Set up some basics */
6030 		callout_init(&rxq->ifr_watchdog, 1);
6031 
6032 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs,
6033 		   M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
6034 			device_printf(dev,
6035 			    "Unable to allocate RX DMA info memory\n");
6036 			err = ENOMEM;
6037 			goto err_tx_desc;
6038 		}
6039 
6040 		rxq->ifr_ifdi = ifdip;
6041 		/* XXX this needs to be changed if #rx queues != #tx queues */
6042 		rxq->ifr_ntxqirq = 1;
6043 		rxq->ifr_txqid[0] = i;
6044 		for (j = 0; j < nrxqs; j++, ifdip++) {
6045 			if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, 0)) {
6046 				device_printf(dev,
6047 				    "Unable to allocate RX descriptors\n");
6048 				err = ENOMEM;
6049 				goto err_tx_desc;
6050 			}
6051 			bzero((void *)ifdip->idi_vaddr, rxqsizes[j]);
6052 		}
6053 		rxq->ifr_ctx = ctx;
6054 		rxq->ifr_id = i;
6055 		rxq->ifr_fl_offset = fl_offset;
6056 		rxq->ifr_nfl = nfree_lists;
6057 		if (!(fl =
6058 			  (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) {
6059 			device_printf(dev, "Unable to allocate free list memory\n");
6060 			err = ENOMEM;
6061 			goto err_tx_desc;
6062 		}
6063 		rxq->ifr_fl = fl;
6064 		for (j = 0; j < nfree_lists; j++) {
6065 			fl[j].ifl_rxq = rxq;
6066 			fl[j].ifl_id = j;
6067 			fl[j].ifl_ifdi = &rxq->ifr_ifdi[j + rxq->ifr_fl_offset];
6068 			fl[j].ifl_rxd_size = scctx->isc_rxd_size[j];
6069 		}
6070 		/* Allocate receive buffers for the ring */
6071 		if (iflib_rxsd_alloc(rxq)) {
6072 			device_printf(dev,
6073 			    "Critical Failure setting up receive buffers\n");
6074 			err = ENOMEM;
6075 			goto err_rx_desc;
6076 		}
6077 
6078 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
6079 			fl->ifl_rx_bitmap = bit_alloc(fl->ifl_size, M_IFLIB,
6080 			    M_WAITOK);
6081 	}
6082 
6083 	/* TXQs */
6084 	vaddrs = malloc(sizeof(caddr_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
6085 	paddrs = malloc(sizeof(uint64_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
6086 	for (i = 0; i < ntxqsets; i++) {
6087 		iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi;
6088 
6089 		for (j = 0; j < ntxqs; j++, di++) {
6090 			vaddrs[i*ntxqs + j] = di->idi_vaddr;
6091 			paddrs[i*ntxqs + j] = di->idi_paddr;
6092 		}
6093 	}
6094 	if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) {
6095 		device_printf(ctx->ifc_dev,
6096 		    "Unable to allocate device TX queue\n");
6097 		iflib_tx_structures_free(ctx);
6098 		free(vaddrs, M_IFLIB);
6099 		free(paddrs, M_IFLIB);
6100 		goto err_rx_desc;
6101 	}
6102 	free(vaddrs, M_IFLIB);
6103 	free(paddrs, M_IFLIB);
6104 
6105 	/* RXQs */
6106 	vaddrs = malloc(sizeof(caddr_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
6107 	paddrs = malloc(sizeof(uint64_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
6108 	for (i = 0; i < nrxqsets; i++) {
6109 		iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi;
6110 
6111 		for (j = 0; j < nrxqs; j++, di++) {
6112 			vaddrs[i*nrxqs + j] = di->idi_vaddr;
6113 			paddrs[i*nrxqs + j] = di->idi_paddr;
6114 		}
6115 	}
6116 	if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) {
6117 		device_printf(ctx->ifc_dev,
6118 		    "Unable to allocate device RX queue\n");
6119 		iflib_tx_structures_free(ctx);
6120 		free(vaddrs, M_IFLIB);
6121 		free(paddrs, M_IFLIB);
6122 		goto err_rx_desc;
6123 	}
6124 	free(vaddrs, M_IFLIB);
6125 	free(paddrs, M_IFLIB);
6126 
6127 	return (0);
6128 
6129 /* XXX handle allocation failure changes */
6130 err_rx_desc:
6131 err_tx_desc:
6132 rx_fail:
6133 	if (ctx->ifc_rxqs != NULL)
6134 		free(ctx->ifc_rxqs, M_IFLIB);
6135 	ctx->ifc_rxqs = NULL;
6136 	if (ctx->ifc_txqs != NULL)
6137 		free(ctx->ifc_txqs, M_IFLIB);
6138 	ctx->ifc_txqs = NULL;
6139 fail:
6140 	return (err);
6141 }
6142 
6143 static int
6144 iflib_tx_structures_setup(if_ctx_t ctx)
6145 {
6146 	iflib_txq_t txq = ctx->ifc_txqs;
6147 	int i;
6148 
6149 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
6150 		iflib_txq_setup(txq);
6151 
6152 	return (0);
6153 }
6154 
6155 static void
6156 iflib_tx_structures_free(if_ctx_t ctx)
6157 {
6158 	iflib_txq_t txq = ctx->ifc_txqs;
6159 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6160 	int i, j;
6161 
6162 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
6163 		for (j = 0; j < sctx->isc_ntxqs; j++)
6164 			iflib_dma_free(&txq->ift_ifdi[j]);
6165 		iflib_txq_destroy(txq);
6166 	}
6167 	free(ctx->ifc_txqs, M_IFLIB);
6168 	ctx->ifc_txqs = NULL;
6169 }
6170 
6171 /*********************************************************************
6172  *
6173  *  Initialize all receive rings.
6174  *
6175  **********************************************************************/
6176 static int
6177 iflib_rx_structures_setup(if_ctx_t ctx)
6178 {
6179 	iflib_rxq_t rxq = ctx->ifc_rxqs;
6180 	int q;
6181 #if defined(INET6) || defined(INET)
6182 	int err, i;
6183 #endif
6184 
6185 	for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) {
6186 #if defined(INET6) || defined(INET)
6187 		err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp,
6188 		    TCP_LRO_ENTRIES, min(1024,
6189 		    ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset]));
6190 		if (err != 0) {
6191 			device_printf(ctx->ifc_dev,
6192 			    "LRO Initialization failed!\n");
6193 			goto fail;
6194 		}
6195 #endif
6196 		IFDI_RXQ_SETUP(ctx, rxq->ifr_id);
6197 	}
6198 	return (0);
6199 #if defined(INET6) || defined(INET)
6200 fail:
6201 	/*
6202 	 * Free LRO resources allocated so far, we will only handle
6203 	 * the rings that completed, the failing case will have
6204 	 * cleaned up for itself.  'q' failed, so its the terminus.
6205 	 */
6206 	rxq = ctx->ifc_rxqs;
6207 	for (i = 0; i < q; ++i, rxq++) {
6208 		tcp_lro_free(&rxq->ifr_lc);
6209 	}
6210 	return (err);
6211 #endif
6212 }
6213 
6214 /*********************************************************************
6215  *
6216  *  Free all receive rings.
6217  *
6218  **********************************************************************/
6219 static void
6220 iflib_rx_structures_free(if_ctx_t ctx)
6221 {
6222 	iflib_rxq_t rxq = ctx->ifc_rxqs;
6223 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6224 	int i, j;
6225 
6226 	for (i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) {
6227 		for (j = 0; j < sctx->isc_nrxqs; j++)
6228 			iflib_dma_free(&rxq->ifr_ifdi[j]);
6229 		iflib_rx_sds_free(rxq);
6230 #if defined(INET6) || defined(INET)
6231 		tcp_lro_free(&rxq->ifr_lc);
6232 #endif
6233 	}
6234 	free(ctx->ifc_rxqs, M_IFLIB);
6235 	ctx->ifc_rxqs = NULL;
6236 }
6237 
6238 static int
6239 iflib_qset_structures_setup(if_ctx_t ctx)
6240 {
6241 	int err;
6242 
6243 	/*
6244 	 * It is expected that the caller takes care of freeing queues if this
6245 	 * fails.
6246 	 */
6247 	if ((err = iflib_tx_structures_setup(ctx)) != 0) {
6248 		device_printf(ctx->ifc_dev, "iflib_tx_structures_setup failed: %d\n", err);
6249 		return (err);
6250 	}
6251 
6252 	if ((err = iflib_rx_structures_setup(ctx)) != 0)
6253 		device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err);
6254 
6255 	return (err);
6256 }
6257 
6258 int
6259 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
6260 		driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, const char *name)
6261 {
6262 
6263 	return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name));
6264 }
6265 
6266 /* Just to avoid copy/paste */
6267 static inline int
6268 iflib_irq_set_affinity(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type,
6269     int qid, struct grouptask *gtask, struct taskqgroup *tqg, void *uniq,
6270     const char *name)
6271 {
6272 	device_t dev;
6273 	unsigned int base_cpuid, cpuid;
6274 	int err;
6275 
6276 	dev = ctx->ifc_dev;
6277 	base_cpuid = ctx->ifc_sysctl_core_offset;
6278 	cpuid = get_cpuid_for_queue(ctx, base_cpuid, qid, type == IFLIB_INTR_TX);
6279 	err = taskqgroup_attach_cpu(tqg, gtask, uniq, cpuid, dev,
6280 	    irq ? irq->ii_res : NULL, name);
6281 	if (err) {
6282 		device_printf(dev, "taskqgroup_attach_cpu failed %d\n", err);
6283 		return (err);
6284 	}
6285 #ifdef notyet
6286 	if (cpuid > ctx->ifc_cpuid_highest)
6287 		ctx->ifc_cpuid_highest = cpuid;
6288 #endif
6289 	return (0);
6290 }
6291 
6292 int
6293 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid,
6294 			iflib_intr_type_t type, driver_filter_t *filter,
6295 			void *filter_arg, int qid, const char *name)
6296 {
6297 	device_t dev;
6298 	struct grouptask *gtask;
6299 	struct taskqgroup *tqg;
6300 	iflib_filter_info_t info;
6301 	gtask_fn_t *fn;
6302 	int tqrid, err;
6303 	driver_filter_t *intr_fast;
6304 	void *q;
6305 
6306 	info = &ctx->ifc_filter_info;
6307 	tqrid = rid;
6308 
6309 	switch (type) {
6310 	/* XXX merge tx/rx for netmap? */
6311 	case IFLIB_INTR_TX:
6312 		q = &ctx->ifc_txqs[qid];
6313 		info = &ctx->ifc_txqs[qid].ift_filter_info;
6314 		gtask = &ctx->ifc_txqs[qid].ift_task;
6315 		tqg = qgroup_if_io_tqg;
6316 		fn = _task_fn_tx;
6317 		intr_fast = iflib_fast_intr;
6318 		GROUPTASK_INIT(gtask, 0, fn, q);
6319 		ctx->ifc_flags |= IFC_NETMAP_TX_IRQ;
6320 		break;
6321 	case IFLIB_INTR_RX:
6322 		q = &ctx->ifc_rxqs[qid];
6323 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
6324 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6325 		tqg = qgroup_if_io_tqg;
6326 		fn = _task_fn_rx;
6327 		intr_fast = iflib_fast_intr;
6328 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6329 		break;
6330 	case IFLIB_INTR_RXTX:
6331 		q = &ctx->ifc_rxqs[qid];
6332 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
6333 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6334 		tqg = qgroup_if_io_tqg;
6335 		fn = _task_fn_rx;
6336 		intr_fast = iflib_fast_intr_rxtx;
6337 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6338 		break;
6339 	case IFLIB_INTR_ADMIN:
6340 		q = ctx;
6341 		tqrid = -1;
6342 		info = &ctx->ifc_filter_info;
6343 		gtask = &ctx->ifc_admin_task;
6344 		tqg = qgroup_if_config_tqg;
6345 		fn = _task_fn_admin;
6346 		intr_fast = iflib_fast_intr_ctx;
6347 		break;
6348 	default:
6349 		device_printf(ctx->ifc_dev, "%s: unknown net intr type\n",
6350 		    __func__);
6351 		return (EINVAL);
6352 	}
6353 
6354 	info->ifi_filter = filter;
6355 	info->ifi_filter_arg = filter_arg;
6356 	info->ifi_task = gtask;
6357 	info->ifi_ctx = q;
6358 
6359 	dev = ctx->ifc_dev;
6360 	err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info,  name);
6361 	if (err != 0) {
6362 		device_printf(dev, "_iflib_irq_alloc failed %d\n", err);
6363 		return (err);
6364 	}
6365 	if (type == IFLIB_INTR_ADMIN)
6366 		return (0);
6367 
6368 	if (tqrid != -1) {
6369 		err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg, q,
6370 		    name);
6371 		if (err)
6372 			return (err);
6373 	} else {
6374 		taskqgroup_attach(tqg, gtask, q, dev, irq->ii_res, name);
6375 	}
6376 
6377 	return (0);
6378 }
6379 
6380 void
6381 iflib_softirq_alloc_generic(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type, void *arg, int qid, const char *name)
6382 {
6383 	device_t dev;
6384 	struct grouptask *gtask;
6385 	struct taskqgroup *tqg;
6386 	gtask_fn_t *fn;
6387 	void *q;
6388 	int err;
6389 
6390 	switch (type) {
6391 	case IFLIB_INTR_TX:
6392 		q = &ctx->ifc_txqs[qid];
6393 		gtask = &ctx->ifc_txqs[qid].ift_task;
6394 		tqg = qgroup_if_io_tqg;
6395 		fn = _task_fn_tx;
6396 		GROUPTASK_INIT(gtask, 0, fn, q);
6397 		break;
6398 	case IFLIB_INTR_RX:
6399 		q = &ctx->ifc_rxqs[qid];
6400 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6401 		tqg = qgroup_if_io_tqg;
6402 		fn = _task_fn_rx;
6403 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6404 		break;
6405 	case IFLIB_INTR_IOV:
6406 		q = ctx;
6407 		gtask = &ctx->ifc_vflr_task;
6408 		tqg = qgroup_if_config_tqg;
6409 		fn = _task_fn_iov;
6410 		GROUPTASK_INIT(gtask, 0, fn, q);
6411 		break;
6412 	default:
6413 		panic("unknown net intr type");
6414 	}
6415 	err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg, q, name);
6416 	if (err) {
6417 		dev = ctx->ifc_dev;
6418 		taskqgroup_attach(tqg, gtask, q, dev, irq ? irq->ii_res : NULL,
6419 		    name);
6420 	}
6421 }
6422 
6423 void
6424 iflib_irq_free(if_ctx_t ctx, if_irq_t irq)
6425 {
6426 
6427 	if (irq->ii_tag)
6428 		bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag);
6429 
6430 	if (irq->ii_res)
6431 		bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ,
6432 		    rman_get_rid(irq->ii_res), irq->ii_res);
6433 }
6434 
6435 static int
6436 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, const char *name)
6437 {
6438 	iflib_txq_t txq = ctx->ifc_txqs;
6439 	iflib_rxq_t rxq = ctx->ifc_rxqs;
6440 	if_irq_t irq = &ctx->ifc_legacy_irq;
6441 	iflib_filter_info_t info;
6442 	device_t dev;
6443 	struct grouptask *gtask;
6444 	struct resource *res;
6445 	struct taskqgroup *tqg;
6446 	void *q;
6447 	int err, tqrid;
6448 	bool rx_only;
6449 
6450 	q = &ctx->ifc_rxqs[0];
6451 	info = &rxq[0].ifr_filter_info;
6452 	gtask = &rxq[0].ifr_task;
6453 	tqg = qgroup_if_io_tqg;
6454 	tqrid = *rid;
6455 	rx_only = (ctx->ifc_sctx->isc_flags & IFLIB_SINGLE_IRQ_RX_ONLY) != 0;
6456 
6457 	ctx->ifc_flags |= IFC_LEGACY;
6458 	info->ifi_filter = filter;
6459 	info->ifi_filter_arg = filter_arg;
6460 	info->ifi_task = gtask;
6461 	info->ifi_ctx = rx_only ? ctx : q;
6462 
6463 	dev = ctx->ifc_dev;
6464 	/* We allocate a single interrupt resource */
6465 	err = _iflib_irq_alloc(ctx, irq, tqrid, rx_only ? iflib_fast_intr_ctx :
6466 	    iflib_fast_intr_rxtx, NULL, info, name);
6467 	if (err != 0)
6468 		return (err);
6469 	NET_GROUPTASK_INIT(gtask, 0, _task_fn_rx, q);
6470 	res = irq->ii_res;
6471 	taskqgroup_attach(tqg, gtask, q, dev, res, name);
6472 
6473 	GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq);
6474 	taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, dev, res,
6475 	    "tx");
6476 	return (0);
6477 }
6478 
6479 void
6480 iflib_led_create(if_ctx_t ctx)
6481 {
6482 
6483 	ctx->ifc_led_dev = led_create(iflib_led_func, ctx,
6484 	    device_get_nameunit(ctx->ifc_dev));
6485 }
6486 
6487 void
6488 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid)
6489 {
6490 
6491 	GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
6492 }
6493 
6494 void
6495 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid)
6496 {
6497 
6498 	GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task);
6499 }
6500 
6501 void
6502 iflib_admin_intr_deferred(if_ctx_t ctx)
6503 {
6504 
6505 	MPASS(ctx->ifc_admin_task.gt_taskqueue != NULL);
6506 	GROUPTASK_ENQUEUE(&ctx->ifc_admin_task);
6507 }
6508 
6509 void
6510 iflib_iov_intr_deferred(if_ctx_t ctx)
6511 {
6512 
6513 	GROUPTASK_ENQUEUE(&ctx->ifc_vflr_task);
6514 }
6515 
6516 void
6517 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, const char *name)
6518 {
6519 
6520 	taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, NULL, NULL,
6521 	    name);
6522 }
6523 
6524 void
6525 iflib_config_gtask_init(void *ctx, struct grouptask *gtask, gtask_fn_t *fn,
6526 	const char *name)
6527 {
6528 
6529 	GROUPTASK_INIT(gtask, 0, fn, ctx);
6530 	taskqgroup_attach(qgroup_if_config_tqg, gtask, gtask, NULL, NULL,
6531 	    name);
6532 }
6533 
6534 void
6535 iflib_config_gtask_deinit(struct grouptask *gtask)
6536 {
6537 
6538 	taskqgroup_detach(qgroup_if_config_tqg, gtask);
6539 }
6540 
6541 void
6542 iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate)
6543 {
6544 	if_t ifp = ctx->ifc_ifp;
6545 	iflib_txq_t txq = ctx->ifc_txqs;
6546 
6547 	if_setbaudrate(ifp, baudrate);
6548 	if (baudrate >= IF_Gbps(10)) {
6549 		STATE_LOCK(ctx);
6550 		ctx->ifc_flags |= IFC_PREFETCH;
6551 		STATE_UNLOCK(ctx);
6552 	}
6553 	/* If link down, disable watchdog */
6554 	if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) {
6555 		for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++)
6556 			txq->ift_qstatus = IFLIB_QUEUE_IDLE;
6557 	}
6558 	ctx->ifc_link_state = link_state;
6559 	if_link_state_change(ifp, link_state);
6560 }
6561 
6562 static int
6563 iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq)
6564 {
6565 	int credits;
6566 #ifdef INVARIANTS
6567 	int credits_pre = txq->ift_cidx_processed;
6568 #endif
6569 
6570 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
6571 	    BUS_DMASYNC_POSTREAD);
6572 	if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, true)) == 0)
6573 		return (0);
6574 
6575 	txq->ift_processed += credits;
6576 	txq->ift_cidx_processed += credits;
6577 
6578 	MPASS(credits_pre + credits == txq->ift_cidx_processed);
6579 	if (txq->ift_cidx_processed >= txq->ift_size)
6580 		txq->ift_cidx_processed -= txq->ift_size;
6581 	return (credits);
6582 }
6583 
6584 static int
6585 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget)
6586 {
6587 	iflib_fl_t fl;
6588 	u_int i;
6589 
6590 	for (i = 0, fl = &rxq->ifr_fl[0]; i < rxq->ifr_nfl; i++, fl++)
6591 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
6592 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
6593 	return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx,
6594 	    budget));
6595 }
6596 
6597 void
6598 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name,
6599 	const char *description, if_int_delay_info_t info,
6600 	int offset, int value)
6601 {
6602 	info->iidi_ctx = ctx;
6603 	info->iidi_offset = offset;
6604 	info->iidi_value = value;
6605 	SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev),
6606 	    SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)),
6607 	    OID_AUTO, name, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE,
6608 	    info, 0, iflib_sysctl_int_delay, "I", description);
6609 }
6610 
6611 struct sx *
6612 iflib_ctx_lock_get(if_ctx_t ctx)
6613 {
6614 
6615 	return (&ctx->ifc_ctx_sx);
6616 }
6617 
6618 static int
6619 iflib_msix_init(if_ctx_t ctx)
6620 {
6621 	device_t dev = ctx->ifc_dev;
6622 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6623 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6624 	int admincnt, bar, err, iflib_num_rx_queues, iflib_num_tx_queues;
6625 	int msgs, queuemsgs, queues, rx_queues, tx_queues, vectors;
6626 
6627 	iflib_num_tx_queues = ctx->ifc_sysctl_ntxqs;
6628 	iflib_num_rx_queues = ctx->ifc_sysctl_nrxqs;
6629 
6630 	if (bootverbose)
6631 		device_printf(dev, "msix_init qsets capped at %d\n",
6632 		    imax(scctx->isc_ntxqsets, scctx->isc_nrxqsets));
6633 
6634 	/* Override by tuneable */
6635 	if (scctx->isc_disable_msix)
6636 		goto msi;
6637 
6638 	/* First try MSI-X */
6639 	if ((msgs = pci_msix_count(dev)) == 0) {
6640 		if (bootverbose)
6641 			device_printf(dev, "MSI-X not supported or disabled\n");
6642 		goto msi;
6643 	}
6644 
6645 	bar = ctx->ifc_softc_ctx.isc_msix_bar;
6646 	/*
6647 	 * bar == -1 => "trust me I know what I'm doing"
6648 	 * Some drivers are for hardware that is so shoddily
6649 	 * documented that no one knows which bars are which
6650 	 * so the developer has to map all bars. This hack
6651 	 * allows shoddy garbage to use MSI-X in this framework.
6652 	 */
6653 	if (bar != -1) {
6654 		ctx->ifc_msix_mem = bus_alloc_resource_any(dev,
6655 	            SYS_RES_MEMORY, &bar, RF_ACTIVE);
6656 		if (ctx->ifc_msix_mem == NULL) {
6657 			device_printf(dev, "Unable to map MSI-X table\n");
6658 			goto msi;
6659 		}
6660 	}
6661 
6662 	admincnt = sctx->isc_admin_intrcnt;
6663 #if IFLIB_DEBUG
6664 	/* use only 1 qset in debug mode */
6665 	queuemsgs = min(msgs - admincnt, 1);
6666 #else
6667 	queuemsgs = msgs - admincnt;
6668 #endif
6669 #ifdef RSS
6670 	queues = imin(queuemsgs, rss_getnumbuckets());
6671 #else
6672 	queues = queuemsgs;
6673 #endif
6674 	queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues);
6675 	if (bootverbose)
6676 		device_printf(dev,
6677 		    "intr CPUs: %d queue msgs: %d admincnt: %d\n",
6678 		    CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt);
6679 #ifdef  RSS
6680 	/* If we're doing RSS, clamp at the number of RSS buckets */
6681 	if (queues > rss_getnumbuckets())
6682 		queues = rss_getnumbuckets();
6683 #endif
6684 	if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt)
6685 		rx_queues = iflib_num_rx_queues;
6686 	else
6687 		rx_queues = queues;
6688 
6689 	if (rx_queues > scctx->isc_nrxqsets)
6690 		rx_queues = scctx->isc_nrxqsets;
6691 
6692 	/*
6693 	 * We want this to be all logical CPUs by default
6694 	 */
6695 	if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues)
6696 		tx_queues = iflib_num_tx_queues;
6697 	else
6698 		tx_queues = mp_ncpus;
6699 
6700 	if (tx_queues > scctx->isc_ntxqsets)
6701 		tx_queues = scctx->isc_ntxqsets;
6702 
6703 	if (ctx->ifc_sysctl_qs_eq_override == 0) {
6704 #ifdef INVARIANTS
6705 		if (tx_queues != rx_queues)
6706 			device_printf(dev,
6707 			    "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n",
6708 			    min(rx_queues, tx_queues), min(rx_queues, tx_queues));
6709 #endif
6710 		tx_queues = min(rx_queues, tx_queues);
6711 		rx_queues = min(rx_queues, tx_queues);
6712 	}
6713 
6714 	vectors = rx_queues + admincnt;
6715 	if (msgs < vectors) {
6716 		device_printf(dev,
6717 		    "insufficient number of MSI-X vectors "
6718 		    "(supported %d, need %d)\n", msgs, vectors);
6719 		goto msi;
6720 	}
6721 
6722 	device_printf(dev, "Using %d RX queues %d TX queues\n", rx_queues,
6723 	    tx_queues);
6724 	msgs = vectors;
6725 	if ((err = pci_alloc_msix(dev, &vectors)) == 0) {
6726 		if (vectors != msgs) {
6727 			device_printf(dev,
6728 			    "Unable to allocate sufficient MSI-X vectors "
6729 			    "(got %d, need %d)\n", vectors, msgs);
6730 			pci_release_msi(dev);
6731 			if (bar != -1) {
6732 				bus_release_resource(dev, SYS_RES_MEMORY, bar,
6733 				    ctx->ifc_msix_mem);
6734 				ctx->ifc_msix_mem = NULL;
6735 			}
6736 			goto msi;
6737 		}
6738 		device_printf(dev, "Using MSI-X interrupts with %d vectors\n",
6739 		    vectors);
6740 		scctx->isc_vectors = vectors;
6741 		scctx->isc_nrxqsets = rx_queues;
6742 		scctx->isc_ntxqsets = tx_queues;
6743 		scctx->isc_intr = IFLIB_INTR_MSIX;
6744 
6745 		return (vectors);
6746 	} else {
6747 		device_printf(dev,
6748 		    "failed to allocate %d MSI-X vectors, err: %d\n", vectors,
6749 		    err);
6750 		if (bar != -1) {
6751 			bus_release_resource(dev, SYS_RES_MEMORY, bar,
6752 			    ctx->ifc_msix_mem);
6753 			ctx->ifc_msix_mem = NULL;
6754 		}
6755 	}
6756 
6757 msi:
6758 	vectors = pci_msi_count(dev);
6759 	scctx->isc_nrxqsets = 1;
6760 	scctx->isc_ntxqsets = 1;
6761 	scctx->isc_vectors = vectors;
6762 	if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) {
6763 		device_printf(dev,"Using an MSI interrupt\n");
6764 		scctx->isc_intr = IFLIB_INTR_MSI;
6765 	} else {
6766 		scctx->isc_vectors = 1;
6767 		device_printf(dev,"Using a Legacy interrupt\n");
6768 		scctx->isc_intr = IFLIB_INTR_LEGACY;
6769 	}
6770 
6771 	return (vectors);
6772 }
6773 
6774 static const char *ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" };
6775 
6776 static int
6777 mp_ring_state_handler(SYSCTL_HANDLER_ARGS)
6778 {
6779 	int rc;
6780 	uint16_t *state = ((uint16_t *)oidp->oid_arg1);
6781 	struct sbuf *sb;
6782 	const char *ring_state = "UNKNOWN";
6783 
6784 	/* XXX needed ? */
6785 	rc = sysctl_wire_old_buffer(req, 0);
6786 	MPASS(rc == 0);
6787 	if (rc != 0)
6788 		return (rc);
6789 	sb = sbuf_new_for_sysctl(NULL, NULL, 80, req);
6790 	MPASS(sb != NULL);
6791 	if (sb == NULL)
6792 		return (ENOMEM);
6793 	if (state[3] <= 3)
6794 		ring_state = ring_states[state[3]];
6795 
6796 	sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s",
6797 		    state[0], state[1], state[2], ring_state);
6798 	rc = sbuf_finish(sb);
6799 	sbuf_delete(sb);
6800         return(rc);
6801 }
6802 
6803 enum iflib_ndesc_handler {
6804 	IFLIB_NTXD_HANDLER,
6805 	IFLIB_NRXD_HANDLER,
6806 };
6807 
6808 static int
6809 mp_ndesc_handler(SYSCTL_HANDLER_ARGS)
6810 {
6811 	if_ctx_t ctx = (void *)arg1;
6812 	enum iflib_ndesc_handler type = arg2;
6813 	char buf[256] = {0};
6814 	qidx_t *ndesc;
6815 	char *p, *next;
6816 	int nqs, rc, i;
6817 
6818 	nqs = 8;
6819 	switch(type) {
6820 	case IFLIB_NTXD_HANDLER:
6821 		ndesc = ctx->ifc_sysctl_ntxds;
6822 		if (ctx->ifc_sctx)
6823 			nqs = ctx->ifc_sctx->isc_ntxqs;
6824 		break;
6825 	case IFLIB_NRXD_HANDLER:
6826 		ndesc = ctx->ifc_sysctl_nrxds;
6827 		if (ctx->ifc_sctx)
6828 			nqs = ctx->ifc_sctx->isc_nrxqs;
6829 		break;
6830 	default:
6831 		printf("%s: unhandled type\n", __func__);
6832 		return (EINVAL);
6833 	}
6834 	if (nqs == 0)
6835 		nqs = 8;
6836 
6837 	for (i=0; i<8; i++) {
6838 		if (i >= nqs)
6839 			break;
6840 		if (i)
6841 			strcat(buf, ",");
6842 		sprintf(strchr(buf, 0), "%d", ndesc[i]);
6843 	}
6844 
6845 	rc = sysctl_handle_string(oidp, buf, sizeof(buf), req);
6846 	if (rc || req->newptr == NULL)
6847 		return rc;
6848 
6849 	for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p;
6850 	    i++, p = strsep(&next, " ,")) {
6851 		ndesc[i] = strtoul(p, NULL, 10);
6852 	}
6853 
6854 	return(rc);
6855 }
6856 
6857 #define NAME_BUFLEN 32
6858 static void
6859 iflib_add_device_sysctl_pre(if_ctx_t ctx)
6860 {
6861         device_t dev = iflib_get_dev(ctx);
6862 	struct sysctl_oid_list *child, *oid_list;
6863 	struct sysctl_ctx_list *ctx_list;
6864 	struct sysctl_oid *node;
6865 
6866 	ctx_list = device_get_sysctl_ctx(dev);
6867 	child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev));
6868 	ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, "iflib",
6869 	    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "IFLIB fields");
6870 	oid_list = SYSCTL_CHILDREN(node);
6871 
6872 	SYSCTL_ADD_CONST_STRING(ctx_list, oid_list, OID_AUTO, "driver_version",
6873 		       CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version,
6874 		       "driver version");
6875 
6876 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs",
6877 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0,
6878 			"# of txqs to use, 0 => use default #");
6879 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs",
6880 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0,
6881 			"# of rxqs to use, 0 => use default #");
6882 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_qs_enable",
6883 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0,
6884                        "permit #txq != #rxq");
6885 	SYSCTL_ADD_INT(ctx_list, oid_list, OID_AUTO, "disable_msix",
6886                       CTLFLAG_RWTUN, &ctx->ifc_softc_ctx.isc_disable_msix, 0,
6887                       "disable MSI-X (default 0)");
6888 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "rx_budget",
6889 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_rx_budget, 0,
6890 		       "set the RX budget");
6891 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "tx_abdicate",
6892 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_tx_abdicate, 0,
6893 		       "cause TX to abdicate instead of running to completion");
6894 	ctx->ifc_sysctl_core_offset = CORE_OFFSET_UNSPECIFIED;
6895 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "core_offset",
6896 		       CTLFLAG_RDTUN, &ctx->ifc_sysctl_core_offset, 0,
6897 		       "offset to start using cores at");
6898 	SYSCTL_ADD_U8(ctx_list, oid_list, OID_AUTO, "separate_txrx",
6899 		       CTLFLAG_RDTUN, &ctx->ifc_sysctl_separate_txrx, 0,
6900 		       "use separate cores for TX and RX");
6901 	SYSCTL_ADD_U8(ctx_list, oid_list, OID_AUTO, "use_logical_cores",
6902 		      CTLFLAG_RDTUN, &ctx->ifc_sysctl_use_logical_cores, 0,
6903 		      "try to make use of logical cores for TX and RX");
6904 
6905 	/* XXX change for per-queue sizes */
6906 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_ntxds",
6907 	    CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NEEDGIANT, ctx,
6908 	    IFLIB_NTXD_HANDLER, mp_ndesc_handler, "A",
6909 	    "list of # of TX descriptors to use, 0 = use default #");
6910 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_nrxds",
6911 	    CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NEEDGIANT, ctx,
6912 	    IFLIB_NRXD_HANDLER, mp_ndesc_handler, "A",
6913 	    "list of # of RX descriptors to use, 0 = use default #");
6914 }
6915 
6916 static void
6917 iflib_add_device_sysctl_post(if_ctx_t ctx)
6918 {
6919 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6920 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6921         device_t dev = iflib_get_dev(ctx);
6922 	struct sysctl_oid_list *child;
6923 	struct sysctl_ctx_list *ctx_list;
6924 	iflib_fl_t fl;
6925 	iflib_txq_t txq;
6926 	iflib_rxq_t rxq;
6927 	int i, j;
6928 	char namebuf[NAME_BUFLEN];
6929 	char *qfmt;
6930 	struct sysctl_oid *queue_node, *fl_node, *node;
6931 	struct sysctl_oid_list *queue_list, *fl_list;
6932 	ctx_list = device_get_sysctl_ctx(dev);
6933 
6934 	node = ctx->ifc_sysctl_node;
6935 	child = SYSCTL_CHILDREN(node);
6936 
6937 	if (scctx->isc_ntxqsets > 100)
6938 		qfmt = "txq%03d";
6939 	else if (scctx->isc_ntxqsets > 10)
6940 		qfmt = "txq%02d";
6941 	else
6942 		qfmt = "txq%d";
6943 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
6944 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6945 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6946 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "Queue Name");
6947 		queue_list = SYSCTL_CHILDREN(queue_node);
6948 		SYSCTL_ADD_INT(ctx_list, queue_list, OID_AUTO, "cpu",
6949 			       CTLFLAG_RD,
6950 			       &txq->ift_task.gt_cpu, 0, "cpu this queue is bound to");
6951 #if MEMORY_LOGGING
6952 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued",
6953 				CTLFLAG_RD,
6954 				&txq->ift_dequeued, "total mbufs freed");
6955 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued",
6956 				CTLFLAG_RD,
6957 				&txq->ift_enqueued, "total mbufs enqueued");
6958 #endif
6959 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag",
6960 				   CTLFLAG_RD,
6961 				   &txq->ift_mbuf_defrag, "# of times m_defrag was called");
6962 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "m_pullups",
6963 				   CTLFLAG_RD,
6964 				   &txq->ift_pullups, "# of times m_pullup was called");
6965 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag_failed",
6966 				   CTLFLAG_RD,
6967 				   &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed");
6968 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_desc_avail",
6969 				   CTLFLAG_RD,
6970 				   &txq->ift_no_desc_avail, "# of times no descriptors were available");
6971 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "tx_map_failed",
6972 				   CTLFLAG_RD,
6973 				   &txq->ift_map_failed, "# of times DMA map failed");
6974 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txd_encap_efbig",
6975 				   CTLFLAG_RD,
6976 				   &txq->ift_txd_encap_efbig, "# of times txd_encap returned EFBIG");
6977 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_tx_dma_setup",
6978 				   CTLFLAG_RD,
6979 				   &txq->ift_no_tx_dma_setup, "# of times map failed for other than EFBIG");
6980 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx",
6981 				   CTLFLAG_RD,
6982 				   &txq->ift_pidx, 1, "Producer Index");
6983 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx",
6984 				   CTLFLAG_RD,
6985 				   &txq->ift_cidx, 1, "Consumer Index");
6986 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx_processed",
6987 				   CTLFLAG_RD,
6988 				   &txq->ift_cidx_processed, 1, "Consumer Index seen by credit update");
6989 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use",
6990 				   CTLFLAG_RD,
6991 				   &txq->ift_in_use, 1, "descriptors in use");
6992 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_processed",
6993 				   CTLFLAG_RD,
6994 				   &txq->ift_processed, "descriptors procesed for clean");
6995 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned",
6996 				   CTLFLAG_RD,
6997 				   &txq->ift_cleaned, "total cleaned");
6998 		SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state",
6999 		    CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_NEEDGIANT,
7000 		    __DEVOLATILE(uint64_t *, &txq->ift_br->state), 0,
7001 		    mp_ring_state_handler, "A", "soft ring state");
7002 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_enqueues",
7003 				       CTLFLAG_RD, &txq->ift_br->enqueues,
7004 				       "# of enqueues to the mp_ring for this queue");
7005 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_drops",
7006 				       CTLFLAG_RD, &txq->ift_br->drops,
7007 				       "# of drops in the mp_ring for this queue");
7008 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_starts",
7009 				       CTLFLAG_RD, &txq->ift_br->starts,
7010 				       "# of normal consumer starts in the mp_ring for this queue");
7011 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_stalls",
7012 				       CTLFLAG_RD, &txq->ift_br->stalls,
7013 					       "# of consumer stalls in the mp_ring for this queue");
7014 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_restarts",
7015 			       CTLFLAG_RD, &txq->ift_br->restarts,
7016 				       "# of consumer restarts in the mp_ring for this queue");
7017 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_abdications",
7018 				       CTLFLAG_RD, &txq->ift_br->abdications,
7019 				       "# of consumer abdications in the mp_ring for this queue");
7020 	}
7021 
7022 	if (scctx->isc_nrxqsets > 100)
7023 		qfmt = "rxq%03d";
7024 	else if (scctx->isc_nrxqsets > 10)
7025 		qfmt = "rxq%02d";
7026 	else
7027 		qfmt = "rxq%d";
7028 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
7029 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
7030 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
7031 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "Queue Name");
7032 		queue_list = SYSCTL_CHILDREN(queue_node);
7033 		SYSCTL_ADD_INT(ctx_list, queue_list, OID_AUTO, "cpu",
7034 			       CTLFLAG_RD,
7035 			       &rxq->ifr_task.gt_cpu, 0, "cpu this queue is bound to");
7036 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
7037 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_cidx",
7038 				       CTLFLAG_RD,
7039 				       &rxq->ifr_cq_cidx, 1, "Consumer Index");
7040 		}
7041 
7042 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
7043 			snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j);
7044 			fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list, OID_AUTO, namebuf,
7045 			    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "freelist Name");
7046 			fl_list = SYSCTL_CHILDREN(fl_node);
7047 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx",
7048 				       CTLFLAG_RD,
7049 				       &fl->ifl_pidx, 1, "Producer Index");
7050 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx",
7051 				       CTLFLAG_RD,
7052 				       &fl->ifl_cidx, 1, "Consumer Index");
7053 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits",
7054 				       CTLFLAG_RD,
7055 				       &fl->ifl_credits, 1, "credits available");
7056 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "buf_size",
7057 				       CTLFLAG_RD,
7058 				       &fl->ifl_buf_size, 1, "buffer size");
7059 #if MEMORY_LOGGING
7060 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_enqueued",
7061 					CTLFLAG_RD,
7062 					&fl->ifl_m_enqueued, "mbufs allocated");
7063 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_dequeued",
7064 					CTLFLAG_RD,
7065 					&fl->ifl_m_dequeued, "mbufs freed");
7066 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_enqueued",
7067 					CTLFLAG_RD,
7068 					&fl->ifl_cl_enqueued, "clusters allocated");
7069 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_dequeued",
7070 					CTLFLAG_RD,
7071 					&fl->ifl_cl_dequeued, "clusters freed");
7072 #endif
7073 		}
7074 	}
7075 
7076 }
7077 
7078 void
7079 iflib_request_reset(if_ctx_t ctx)
7080 {
7081 
7082 	STATE_LOCK(ctx);
7083 	ctx->ifc_flags |= IFC_DO_RESET;
7084 	STATE_UNLOCK(ctx);
7085 }
7086 
7087 #ifndef __NO_STRICT_ALIGNMENT
7088 static struct mbuf *
7089 iflib_fixup_rx(struct mbuf *m)
7090 {
7091 	struct mbuf *n;
7092 
7093 	if (m->m_len <= (MCLBYTES - ETHER_HDR_LEN)) {
7094 		bcopy(m->m_data, m->m_data + ETHER_HDR_LEN, m->m_len);
7095 		m->m_data += ETHER_HDR_LEN;
7096 		n = m;
7097 	} else {
7098 		MGETHDR(n, M_NOWAIT, MT_DATA);
7099 		if (n == NULL) {
7100 			m_freem(m);
7101 			return (NULL);
7102 		}
7103 		bcopy(m->m_data, n->m_data, ETHER_HDR_LEN);
7104 		m->m_data += ETHER_HDR_LEN;
7105 		m->m_len -= ETHER_HDR_LEN;
7106 		n->m_len = ETHER_HDR_LEN;
7107 		M_MOVE_PKTHDR(n, m);
7108 		n->m_next = m;
7109 	}
7110 	return (n);
7111 }
7112 #endif
7113 
7114 #ifdef DEBUGNET
7115 static void
7116 iflib_debugnet_init(if_t ifp, int *nrxr, int *ncl, int *clsize)
7117 {
7118 	if_ctx_t ctx;
7119 
7120 	ctx = if_getsoftc(ifp);
7121 	CTX_LOCK(ctx);
7122 	*nrxr = NRXQSETS(ctx);
7123 	*ncl = ctx->ifc_rxqs[0].ifr_fl->ifl_size;
7124 	*clsize = ctx->ifc_rxqs[0].ifr_fl->ifl_buf_size;
7125 	CTX_UNLOCK(ctx);
7126 }
7127 
7128 static void
7129 iflib_debugnet_event(if_t ifp, enum debugnet_ev event)
7130 {
7131 	if_ctx_t ctx;
7132 	if_softc_ctx_t scctx;
7133 	iflib_fl_t fl;
7134 	iflib_rxq_t rxq;
7135 	int i, j;
7136 
7137 	ctx = if_getsoftc(ifp);
7138 	scctx = &ctx->ifc_softc_ctx;
7139 
7140 	switch (event) {
7141 	case DEBUGNET_START:
7142 		for (i = 0; i < scctx->isc_nrxqsets; i++) {
7143 			rxq = &ctx->ifc_rxqs[i];
7144 			for (j = 0; j < rxq->ifr_nfl; j++) {
7145 				fl = rxq->ifr_fl;
7146 				fl->ifl_zone = m_getzone(fl->ifl_buf_size);
7147 			}
7148 		}
7149 		iflib_no_tx_batch = 1;
7150 		break;
7151 	default:
7152 		break;
7153 	}
7154 }
7155 
7156 static int
7157 iflib_debugnet_transmit(if_t ifp, struct mbuf *m)
7158 {
7159 	if_ctx_t ctx;
7160 	iflib_txq_t txq;
7161 	int error;
7162 
7163 	ctx = if_getsoftc(ifp);
7164 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
7165 	    IFF_DRV_RUNNING)
7166 		return (EBUSY);
7167 
7168 	txq = &ctx->ifc_txqs[0];
7169 	error = iflib_encap(txq, &m);
7170 	if (error == 0)
7171 		(void)iflib_txd_db_check(txq, true);
7172 	return (error);
7173 }
7174 
7175 static int
7176 iflib_debugnet_poll(if_t ifp, int count)
7177 {
7178 	struct epoch_tracker et;
7179 	if_ctx_t ctx;
7180 	if_softc_ctx_t scctx;
7181 	iflib_txq_t txq;
7182 	int i;
7183 
7184 	ctx = if_getsoftc(ifp);
7185 	scctx = &ctx->ifc_softc_ctx;
7186 
7187 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
7188 	    IFF_DRV_RUNNING)
7189 		return (EBUSY);
7190 
7191 	txq = &ctx->ifc_txqs[0];
7192 	(void)iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
7193 
7194 	NET_EPOCH_ENTER(et);
7195 	for (i = 0; i < scctx->isc_nrxqsets; i++)
7196 		(void)iflib_rxeof(&ctx->ifc_rxqs[i], 16 /* XXX */);
7197 	NET_EPOCH_EXIT(et);
7198 	return (0);
7199 }
7200 #endif /* DEBUGNET */
7201