xref: /freebsd/sys/net/iflib.c (revision b51f459a2098622c31ed54f5c1bf0e03efce403b)
1 /*-
2  * Copyright (c) 2014-2018, Matthew Macy <mmacy@mattmacy.io>
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions are met:
7  *
8  *  1. Redistributions of source code must retain the above copyright notice,
9  *     this list of conditions and the following disclaimer.
10  *
11  *  2. Neither the name of Matthew Macy nor the names of its
12  *     contributors may be used to endorse or promote products derived from
13  *     this software without specific prior written permission.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
16  * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18  * ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
19  * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
20  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
21  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
22  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
23  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
24  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
25  * POSSIBILITY OF SUCH DAMAGE.
26  */
27 
28 #include <sys/cdefs.h>
29 __FBSDID("$FreeBSD$");
30 
31 #include "opt_inet.h"
32 #include "opt_inet6.h"
33 #include "opt_acpi.h"
34 #include "opt_sched.h"
35 
36 #include <sys/param.h>
37 #include <sys/types.h>
38 #include <sys/bus.h>
39 #include <sys/eventhandler.h>
40 #include <sys/kernel.h>
41 #include <sys/lock.h>
42 #include <sys/mutex.h>
43 #include <sys/module.h>
44 #include <sys/kobj.h>
45 #include <sys/rman.h>
46 #include <sys/sbuf.h>
47 #include <sys/smp.h>
48 #include <sys/socket.h>
49 #include <sys/sockio.h>
50 #include <sys/sysctl.h>
51 #include <sys/syslog.h>
52 #include <sys/taskqueue.h>
53 #include <sys/limits.h>
54 
55 #include <net/if.h>
56 #include <net/if_var.h>
57 #include <net/if_types.h>
58 #include <net/if_media.h>
59 #include <net/bpf.h>
60 #include <net/ethernet.h>
61 #include <net/mp_ring.h>
62 #include <net/debugnet.h>
63 #include <net/pfil.h>
64 #include <net/vnet.h>
65 
66 #include <netinet/in.h>
67 #include <netinet/in_pcb.h>
68 #include <netinet/tcp_lro.h>
69 #include <netinet/in_systm.h>
70 #include <netinet/if_ether.h>
71 #include <netinet/ip.h>
72 #include <netinet/ip6.h>
73 #include <netinet/tcp.h>
74 #include <netinet/ip_var.h>
75 #include <netinet6/ip6_var.h>
76 
77 #include <machine/bus.h>
78 #include <machine/in_cksum.h>
79 
80 #include <vm/vm.h>
81 #include <vm/pmap.h>
82 
83 #include <dev/led/led.h>
84 #include <dev/pci/pcireg.h>
85 #include <dev/pci/pcivar.h>
86 #include <dev/pci/pci_private.h>
87 
88 #include <net/iflib.h>
89 #include <net/iflib_private.h>
90 
91 #include "ifdi_if.h"
92 
93 #ifdef PCI_IOV
94 #include <dev/pci/pci_iov.h>
95 #endif
96 
97 #include <sys/bitstring.h>
98 /*
99  * enable accounting of every mbuf as it comes in to and goes out of
100  * iflib's software descriptor references
101  */
102 #define MEMORY_LOGGING 0
103 /*
104  * Enable mbuf vectors for compressing long mbuf chains
105  */
106 
107 /*
108  * NB:
109  * - Prefetching in tx cleaning should perhaps be a tunable. The distance ahead
110  *   we prefetch needs to be determined by the time spent in m_free vis a vis
111  *   the cost of a prefetch. This will of course vary based on the workload:
112  *      - NFLX's m_free path is dominated by vm-based M_EXT manipulation which
113  *        is quite expensive, thus suggesting very little prefetch.
114  *      - small packet forwarding which is just returning a single mbuf to
115  *        UMA will typically be very fast vis a vis the cost of a memory
116  *        access.
117  */
118 
119 /*
120  * File organization:
121  *  - private structures
122  *  - iflib private utility functions
123  *  - ifnet functions
124  *  - vlan registry and other exported functions
125  *  - iflib public core functions
126  *
127  *
128  */
129 MALLOC_DEFINE(M_IFLIB, "iflib", "ifnet library");
130 
131 #define	IFLIB_RXEOF_MORE (1U << 0)
132 #define	IFLIB_RXEOF_EMPTY (2U << 0)
133 
134 struct iflib_txq;
135 typedef struct iflib_txq *iflib_txq_t;
136 struct iflib_rxq;
137 typedef struct iflib_rxq *iflib_rxq_t;
138 struct iflib_fl;
139 typedef struct iflib_fl *iflib_fl_t;
140 
141 struct iflib_ctx;
142 
143 static void iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid);
144 static void iflib_timer(void *arg);
145 static void iflib_tqg_detach(if_ctx_t ctx);
146 
147 typedef struct iflib_filter_info {
148 	driver_filter_t *ifi_filter;
149 	void *ifi_filter_arg;
150 	struct grouptask *ifi_task;
151 	void *ifi_ctx;
152 } *iflib_filter_info_t;
153 
154 struct iflib_ctx {
155 	KOBJ_FIELDS;
156 	/*
157 	 * Pointer to hardware driver's softc
158 	 */
159 	void *ifc_softc;
160 	device_t ifc_dev;
161 	if_t ifc_ifp;
162 
163 	cpuset_t ifc_cpus;
164 	if_shared_ctx_t ifc_sctx;
165 	struct if_softc_ctx ifc_softc_ctx;
166 
167 	struct sx ifc_ctx_sx;
168 	struct mtx ifc_state_mtx;
169 
170 	iflib_txq_t ifc_txqs;
171 	iflib_rxq_t ifc_rxqs;
172 	uint32_t ifc_if_flags;
173 	uint32_t ifc_flags;
174 	uint32_t ifc_max_fl_buf_size;
175 	uint32_t ifc_rx_mbuf_sz;
176 
177 	int ifc_link_state;
178 	int ifc_watchdog_events;
179 	struct cdev *ifc_led_dev;
180 	struct resource *ifc_msix_mem;
181 
182 	struct if_irq ifc_legacy_irq;
183 	struct grouptask ifc_admin_task;
184 	struct grouptask ifc_vflr_task;
185 	struct iflib_filter_info ifc_filter_info;
186 	struct ifmedia	ifc_media;
187 	struct ifmedia	*ifc_mediap;
188 
189 	struct sysctl_oid *ifc_sysctl_node;
190 	uint16_t ifc_sysctl_ntxqs;
191 	uint16_t ifc_sysctl_nrxqs;
192 	uint16_t ifc_sysctl_qs_eq_override;
193 	uint16_t ifc_sysctl_rx_budget;
194 	uint16_t ifc_sysctl_tx_abdicate;
195 	uint16_t ifc_sysctl_core_offset;
196 #define	CORE_OFFSET_UNSPECIFIED	0xffff
197 	uint8_t  ifc_sysctl_separate_txrx;
198 
199 	qidx_t ifc_sysctl_ntxds[8];
200 	qidx_t ifc_sysctl_nrxds[8];
201 	struct if_txrx ifc_txrx;
202 #define isc_txd_encap  ifc_txrx.ift_txd_encap
203 #define isc_txd_flush  ifc_txrx.ift_txd_flush
204 #define isc_txd_credits_update  ifc_txrx.ift_txd_credits_update
205 #define isc_rxd_available ifc_txrx.ift_rxd_available
206 #define isc_rxd_pkt_get ifc_txrx.ift_rxd_pkt_get
207 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
208 #define isc_rxd_flush ifc_txrx.ift_rxd_flush
209 #define isc_legacy_intr ifc_txrx.ift_legacy_intr
210 	eventhandler_tag ifc_vlan_attach_event;
211 	eventhandler_tag ifc_vlan_detach_event;
212 	struct ether_addr ifc_mac;
213 };
214 
215 void *
216 iflib_get_softc(if_ctx_t ctx)
217 {
218 
219 	return (ctx->ifc_softc);
220 }
221 
222 device_t
223 iflib_get_dev(if_ctx_t ctx)
224 {
225 
226 	return (ctx->ifc_dev);
227 }
228 
229 if_t
230 iflib_get_ifp(if_ctx_t ctx)
231 {
232 
233 	return (ctx->ifc_ifp);
234 }
235 
236 struct ifmedia *
237 iflib_get_media(if_ctx_t ctx)
238 {
239 
240 	return (ctx->ifc_mediap);
241 }
242 
243 uint32_t
244 iflib_get_flags(if_ctx_t ctx)
245 {
246 	return (ctx->ifc_flags);
247 }
248 
249 void
250 iflib_set_mac(if_ctx_t ctx, uint8_t mac[ETHER_ADDR_LEN])
251 {
252 
253 	bcopy(mac, ctx->ifc_mac.octet, ETHER_ADDR_LEN);
254 }
255 
256 if_softc_ctx_t
257 iflib_get_softc_ctx(if_ctx_t ctx)
258 {
259 
260 	return (&ctx->ifc_softc_ctx);
261 }
262 
263 if_shared_ctx_t
264 iflib_get_sctx(if_ctx_t ctx)
265 {
266 
267 	return (ctx->ifc_sctx);
268 }
269 
270 #define IP_ALIGNED(m) ((((uintptr_t)(m)->m_data) & 0x3) == 0x2)
271 #define CACHE_PTR_INCREMENT (CACHE_LINE_SIZE/sizeof(void*))
272 #define CACHE_PTR_NEXT(ptr) ((void *)(((uintptr_t)(ptr)+CACHE_LINE_SIZE-1) & (CACHE_LINE_SIZE-1)))
273 
274 #define LINK_ACTIVE(ctx) ((ctx)->ifc_link_state == LINK_STATE_UP)
275 #define CTX_IS_VF(ctx) ((ctx)->ifc_sctx->isc_flags & IFLIB_IS_VF)
276 
277 typedef struct iflib_sw_rx_desc_array {
278 	bus_dmamap_t	*ifsd_map;         /* bus_dma maps for packet */
279 	struct mbuf	**ifsd_m;           /* pkthdr mbufs */
280 	caddr_t		*ifsd_cl;          /* direct cluster pointer for rx */
281 	bus_addr_t	*ifsd_ba;          /* bus addr of cluster for rx */
282 } iflib_rxsd_array_t;
283 
284 typedef struct iflib_sw_tx_desc_array {
285 	bus_dmamap_t    *ifsd_map;         /* bus_dma maps for packet */
286 	bus_dmamap_t	*ifsd_tso_map;     /* bus_dma maps for TSO packet */
287 	struct mbuf    **ifsd_m;           /* pkthdr mbufs */
288 } if_txsd_vec_t;
289 
290 /* magic number that should be high enough for any hardware */
291 #define IFLIB_MAX_TX_SEGS		128
292 #define IFLIB_RX_COPY_THRESH		128
293 #define IFLIB_MAX_RX_REFRESH		32
294 /* The minimum descriptors per second before we start coalescing */
295 #define IFLIB_MIN_DESC_SEC		16384
296 #define IFLIB_DEFAULT_TX_UPDATE_FREQ	16
297 #define IFLIB_QUEUE_IDLE		0
298 #define IFLIB_QUEUE_HUNG		1
299 #define IFLIB_QUEUE_WORKING		2
300 /* maximum number of txqs that can share an rx interrupt */
301 #define IFLIB_MAX_TX_SHARED_INTR	4
302 
303 /* this should really scale with ring size - this is a fairly arbitrary value */
304 #define TX_BATCH_SIZE			32
305 
306 #define IFLIB_RESTART_BUDGET		8
307 
308 #define CSUM_OFFLOAD		(CSUM_IP_TSO|CSUM_IP6_TSO|CSUM_IP| \
309 				 CSUM_IP_UDP|CSUM_IP_TCP|CSUM_IP_SCTP| \
310 				 CSUM_IP6_UDP|CSUM_IP6_TCP|CSUM_IP6_SCTP)
311 
312 struct iflib_txq {
313 	qidx_t		ift_in_use;
314 	qidx_t		ift_cidx;
315 	qidx_t		ift_cidx_processed;
316 	qidx_t		ift_pidx;
317 	uint8_t		ift_gen;
318 	uint8_t		ift_br_offset;
319 	uint16_t	ift_npending;
320 	uint16_t	ift_db_pending;
321 	uint16_t	ift_rs_pending;
322 	/* implicit pad */
323 	uint8_t		ift_txd_size[8];
324 	uint64_t	ift_processed;
325 	uint64_t	ift_cleaned;
326 	uint64_t	ift_cleaned_prev;
327 #if MEMORY_LOGGING
328 	uint64_t	ift_enqueued;
329 	uint64_t	ift_dequeued;
330 #endif
331 	uint64_t	ift_no_tx_dma_setup;
332 	uint64_t	ift_no_desc_avail;
333 	uint64_t	ift_mbuf_defrag_failed;
334 	uint64_t	ift_mbuf_defrag;
335 	uint64_t	ift_map_failed;
336 	uint64_t	ift_txd_encap_efbig;
337 	uint64_t	ift_pullups;
338 	uint64_t	ift_last_timer_tick;
339 
340 	struct mtx	ift_mtx;
341 	struct mtx	ift_db_mtx;
342 
343 	/* constant values */
344 	if_ctx_t	ift_ctx;
345 	struct ifmp_ring        *ift_br;
346 	struct grouptask	ift_task;
347 	qidx_t		ift_size;
348 	uint16_t	ift_id;
349 	struct callout	ift_timer;
350 #ifdef DEV_NETMAP
351 	struct callout	ift_netmap_timer;
352 #endif /* DEV_NETMAP */
353 
354 	if_txsd_vec_t	ift_sds;
355 	uint8_t		ift_qstatus;
356 	uint8_t		ift_closed;
357 	uint8_t		ift_update_freq;
358 	struct iflib_filter_info ift_filter_info;
359 	bus_dma_tag_t	ift_buf_tag;
360 	bus_dma_tag_t	ift_tso_buf_tag;
361 	iflib_dma_info_t	ift_ifdi;
362 #define	MTX_NAME_LEN	32
363 	char                    ift_mtx_name[MTX_NAME_LEN];
364 	bus_dma_segment_t	ift_segs[IFLIB_MAX_TX_SEGS]  __aligned(CACHE_LINE_SIZE);
365 #ifdef IFLIB_DIAGNOSTICS
366 	uint64_t ift_cpu_exec_count[256];
367 #endif
368 } __aligned(CACHE_LINE_SIZE);
369 
370 struct iflib_fl {
371 	qidx_t		ifl_cidx;
372 	qidx_t		ifl_pidx;
373 	qidx_t		ifl_credits;
374 	uint8_t		ifl_gen;
375 	uint8_t		ifl_rxd_size;
376 #if MEMORY_LOGGING
377 	uint64_t	ifl_m_enqueued;
378 	uint64_t	ifl_m_dequeued;
379 	uint64_t	ifl_cl_enqueued;
380 	uint64_t	ifl_cl_dequeued;
381 #endif
382 	/* implicit pad */
383 	bitstr_t 	*ifl_rx_bitmap;
384 	qidx_t		ifl_fragidx;
385 	/* constant */
386 	qidx_t		ifl_size;
387 	uint16_t	ifl_buf_size;
388 	uint16_t	ifl_cltype;
389 	uma_zone_t	ifl_zone;
390 	iflib_rxsd_array_t	ifl_sds;
391 	iflib_rxq_t	ifl_rxq;
392 	uint8_t		ifl_id;
393 	bus_dma_tag_t	ifl_buf_tag;
394 	iflib_dma_info_t	ifl_ifdi;
395 	uint64_t	ifl_bus_addrs[IFLIB_MAX_RX_REFRESH] __aligned(CACHE_LINE_SIZE);
396 	qidx_t		ifl_rxd_idxs[IFLIB_MAX_RX_REFRESH];
397 }  __aligned(CACHE_LINE_SIZE);
398 
399 static inline qidx_t
400 get_inuse(int size, qidx_t cidx, qidx_t pidx, uint8_t gen)
401 {
402 	qidx_t used;
403 
404 	if (pidx > cidx)
405 		used = pidx - cidx;
406 	else if (pidx < cidx)
407 		used = size - cidx + pidx;
408 	else if (gen == 0 && pidx == cidx)
409 		used = 0;
410 	else if (gen == 1 && pidx == cidx)
411 		used = size;
412 	else
413 		panic("bad state");
414 
415 	return (used);
416 }
417 
418 #define TXQ_AVAIL(txq) (txq->ift_size - get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx, txq->ift_gen))
419 
420 #define IDXDIFF(head, tail, wrap) \
421 	((head) >= (tail) ? (head) - (tail) : (wrap) - (tail) + (head))
422 
423 struct iflib_rxq {
424 	if_ctx_t	ifr_ctx;
425 	iflib_fl_t	ifr_fl;
426 	uint64_t	ifr_rx_irq;
427 	struct pfil_head	*pfil;
428 	/*
429 	 * If there is a separate completion queue (IFLIB_HAS_RXCQ), this is
430 	 * the completion queue consumer index.  Otherwise it's unused.
431 	 */
432 	qidx_t		ifr_cq_cidx;
433 	uint16_t	ifr_id;
434 	uint8_t		ifr_nfl;
435 	uint8_t		ifr_ntxqirq;
436 	uint8_t		ifr_txqid[IFLIB_MAX_TX_SHARED_INTR];
437 	uint8_t		ifr_fl_offset;
438 	struct lro_ctrl			ifr_lc;
439 	struct grouptask        ifr_task;
440 	struct callout		ifr_watchdog;
441 	struct iflib_filter_info ifr_filter_info;
442 	iflib_dma_info_t		ifr_ifdi;
443 
444 	/* dynamically allocate if any drivers need a value substantially larger than this */
445 	struct if_rxd_frag	ifr_frags[IFLIB_MAX_RX_SEGS] __aligned(CACHE_LINE_SIZE);
446 #ifdef IFLIB_DIAGNOSTICS
447 	uint64_t ifr_cpu_exec_count[256];
448 #endif
449 }  __aligned(CACHE_LINE_SIZE);
450 
451 typedef struct if_rxsd {
452 	caddr_t *ifsd_cl;
453 	iflib_fl_t ifsd_fl;
454 } *if_rxsd_t;
455 
456 /* multiple of word size */
457 #ifdef __LP64__
458 #define PKT_INFO_SIZE	6
459 #define RXD_INFO_SIZE	5
460 #define PKT_TYPE uint64_t
461 #else
462 #define PKT_INFO_SIZE	11
463 #define RXD_INFO_SIZE	8
464 #define PKT_TYPE uint32_t
465 #endif
466 #define PKT_LOOP_BOUND  ((PKT_INFO_SIZE/3)*3)
467 #define RXD_LOOP_BOUND  ((RXD_INFO_SIZE/4)*4)
468 
469 typedef struct if_pkt_info_pad {
470 	PKT_TYPE pkt_val[PKT_INFO_SIZE];
471 } *if_pkt_info_pad_t;
472 typedef struct if_rxd_info_pad {
473 	PKT_TYPE rxd_val[RXD_INFO_SIZE];
474 } *if_rxd_info_pad_t;
475 
476 CTASSERT(sizeof(struct if_pkt_info_pad) == sizeof(struct if_pkt_info));
477 CTASSERT(sizeof(struct if_rxd_info_pad) == sizeof(struct if_rxd_info));
478 
479 static inline void
480 pkt_info_zero(if_pkt_info_t pi)
481 {
482 	if_pkt_info_pad_t pi_pad;
483 
484 	pi_pad = (if_pkt_info_pad_t)pi;
485 	pi_pad->pkt_val[0] = 0; pi_pad->pkt_val[1] = 0; pi_pad->pkt_val[2] = 0;
486 	pi_pad->pkt_val[3] = 0; pi_pad->pkt_val[4] = 0; pi_pad->pkt_val[5] = 0;
487 #ifndef __LP64__
488 	pi_pad->pkt_val[6] = 0; pi_pad->pkt_val[7] = 0; pi_pad->pkt_val[8] = 0;
489 	pi_pad->pkt_val[9] = 0; pi_pad->pkt_val[10] = 0;
490 #endif
491 }
492 
493 static device_method_t iflib_pseudo_methods[] = {
494 	DEVMETHOD(device_attach, noop_attach),
495 	DEVMETHOD(device_detach, iflib_pseudo_detach),
496 	DEVMETHOD_END
497 };
498 
499 driver_t iflib_pseudodriver = {
500 	"iflib_pseudo", iflib_pseudo_methods, sizeof(struct iflib_ctx),
501 };
502 
503 static inline void
504 rxd_info_zero(if_rxd_info_t ri)
505 {
506 	if_rxd_info_pad_t ri_pad;
507 	int i;
508 
509 	ri_pad = (if_rxd_info_pad_t)ri;
510 	for (i = 0; i < RXD_LOOP_BOUND; i += 4) {
511 		ri_pad->rxd_val[i] = 0;
512 		ri_pad->rxd_val[i+1] = 0;
513 		ri_pad->rxd_val[i+2] = 0;
514 		ri_pad->rxd_val[i+3] = 0;
515 	}
516 #ifdef __LP64__
517 	ri_pad->rxd_val[RXD_INFO_SIZE-1] = 0;
518 #endif
519 }
520 
521 /*
522  * Only allow a single packet to take up most 1/nth of the tx ring
523  */
524 #define MAX_SINGLE_PACKET_FRACTION 12
525 #define IF_BAD_DMA (bus_addr_t)-1
526 
527 #define CTX_ACTIVE(ctx) ((if_getdrvflags((ctx)->ifc_ifp) & IFF_DRV_RUNNING))
528 
529 #define CTX_LOCK_INIT(_sc)  sx_init(&(_sc)->ifc_ctx_sx, "iflib ctx lock")
530 #define CTX_LOCK(ctx) sx_xlock(&(ctx)->ifc_ctx_sx)
531 #define CTX_UNLOCK(ctx) sx_xunlock(&(ctx)->ifc_ctx_sx)
532 #define CTX_LOCK_DESTROY(ctx) sx_destroy(&(ctx)->ifc_ctx_sx)
533 
534 #define STATE_LOCK_INIT(_sc, _name)  mtx_init(&(_sc)->ifc_state_mtx, _name, "iflib state lock", MTX_DEF)
535 #define STATE_LOCK(ctx) mtx_lock(&(ctx)->ifc_state_mtx)
536 #define STATE_UNLOCK(ctx) mtx_unlock(&(ctx)->ifc_state_mtx)
537 #define STATE_LOCK_DESTROY(ctx) mtx_destroy(&(ctx)->ifc_state_mtx)
538 
539 #define CALLOUT_LOCK(txq)	mtx_lock(&txq->ift_mtx)
540 #define CALLOUT_UNLOCK(txq) 	mtx_unlock(&txq->ift_mtx)
541 
542 void
543 iflib_set_detach(if_ctx_t ctx)
544 {
545 	STATE_LOCK(ctx);
546 	ctx->ifc_flags |= IFC_IN_DETACH;
547 	STATE_UNLOCK(ctx);
548 }
549 
550 /* Our boot-time initialization hook */
551 static int	iflib_module_event_handler(module_t, int, void *);
552 
553 static moduledata_t iflib_moduledata = {
554 	"iflib",
555 	iflib_module_event_handler,
556 	NULL
557 };
558 
559 DECLARE_MODULE(iflib, iflib_moduledata, SI_SUB_INIT_IF, SI_ORDER_ANY);
560 MODULE_VERSION(iflib, 1);
561 
562 MODULE_DEPEND(iflib, pci, 1, 1, 1);
563 MODULE_DEPEND(iflib, ether, 1, 1, 1);
564 
565 TASKQGROUP_DEFINE(if_io_tqg, mp_ncpus, 1);
566 TASKQGROUP_DEFINE(if_config_tqg, 1, 1);
567 
568 #ifndef IFLIB_DEBUG_COUNTERS
569 #ifdef INVARIANTS
570 #define IFLIB_DEBUG_COUNTERS 1
571 #else
572 #define IFLIB_DEBUG_COUNTERS 0
573 #endif /* !INVARIANTS */
574 #endif
575 
576 static SYSCTL_NODE(_net, OID_AUTO, iflib, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
577     "iflib driver parameters");
578 
579 /*
580  * XXX need to ensure that this can't accidentally cause the head to be moved backwards
581  */
582 static int iflib_min_tx_latency = 0;
583 SYSCTL_INT(_net_iflib, OID_AUTO, min_tx_latency, CTLFLAG_RW,
584 		   &iflib_min_tx_latency, 0, "minimize transmit latency at the possible expense of throughput");
585 static int iflib_no_tx_batch = 0;
586 SYSCTL_INT(_net_iflib, OID_AUTO, no_tx_batch, CTLFLAG_RW,
587 		   &iflib_no_tx_batch, 0, "minimize transmit latency at the possible expense of throughput");
588 static int iflib_timer_default = 1000;
589 SYSCTL_INT(_net_iflib, OID_AUTO, timer_default, CTLFLAG_RW,
590 		   &iflib_timer_default, 0, "number of ticks between iflib_timer calls");
591 
592 
593 #if IFLIB_DEBUG_COUNTERS
594 
595 static int iflib_tx_seen;
596 static int iflib_tx_sent;
597 static int iflib_tx_encap;
598 static int iflib_rx_allocs;
599 static int iflib_fl_refills;
600 static int iflib_fl_refills_large;
601 static int iflib_tx_frees;
602 
603 SYSCTL_INT(_net_iflib, OID_AUTO, tx_seen, CTLFLAG_RD,
604 		   &iflib_tx_seen, 0, "# TX mbufs seen");
605 SYSCTL_INT(_net_iflib, OID_AUTO, tx_sent, CTLFLAG_RD,
606 		   &iflib_tx_sent, 0, "# TX mbufs sent");
607 SYSCTL_INT(_net_iflib, OID_AUTO, tx_encap, CTLFLAG_RD,
608 		   &iflib_tx_encap, 0, "# TX mbufs encapped");
609 SYSCTL_INT(_net_iflib, OID_AUTO, tx_frees, CTLFLAG_RD,
610 		   &iflib_tx_frees, 0, "# TX frees");
611 SYSCTL_INT(_net_iflib, OID_AUTO, rx_allocs, CTLFLAG_RD,
612 		   &iflib_rx_allocs, 0, "# RX allocations");
613 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills, CTLFLAG_RD,
614 		   &iflib_fl_refills, 0, "# refills");
615 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills_large, CTLFLAG_RD,
616 		   &iflib_fl_refills_large, 0, "# large refills");
617 
618 static int iflib_txq_drain_flushing;
619 static int iflib_txq_drain_oactive;
620 static int iflib_txq_drain_notready;
621 
622 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_flushing, CTLFLAG_RD,
623 		   &iflib_txq_drain_flushing, 0, "# drain flushes");
624 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_oactive, CTLFLAG_RD,
625 		   &iflib_txq_drain_oactive, 0, "# drain oactives");
626 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_notready, CTLFLAG_RD,
627 		   &iflib_txq_drain_notready, 0, "# drain notready");
628 
629 static int iflib_encap_load_mbuf_fail;
630 static int iflib_encap_pad_mbuf_fail;
631 static int iflib_encap_txq_avail_fail;
632 static int iflib_encap_txd_encap_fail;
633 
634 SYSCTL_INT(_net_iflib, OID_AUTO, encap_load_mbuf_fail, CTLFLAG_RD,
635 		   &iflib_encap_load_mbuf_fail, 0, "# busdma load failures");
636 SYSCTL_INT(_net_iflib, OID_AUTO, encap_pad_mbuf_fail, CTLFLAG_RD,
637 		   &iflib_encap_pad_mbuf_fail, 0, "# runt frame pad failures");
638 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txq_avail_fail, CTLFLAG_RD,
639 		   &iflib_encap_txq_avail_fail, 0, "# txq avail failures");
640 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txd_encap_fail, CTLFLAG_RD,
641 		   &iflib_encap_txd_encap_fail, 0, "# driver encap failures");
642 
643 static int iflib_task_fn_rxs;
644 static int iflib_rx_intr_enables;
645 static int iflib_fast_intrs;
646 static int iflib_rx_unavail;
647 static int iflib_rx_ctx_inactive;
648 static int iflib_rx_if_input;
649 static int iflib_rxd_flush;
650 
651 static int iflib_verbose_debug;
652 
653 SYSCTL_INT(_net_iflib, OID_AUTO, task_fn_rx, CTLFLAG_RD,
654 		   &iflib_task_fn_rxs, 0, "# task_fn_rx calls");
655 SYSCTL_INT(_net_iflib, OID_AUTO, rx_intr_enables, CTLFLAG_RD,
656 		   &iflib_rx_intr_enables, 0, "# RX intr enables");
657 SYSCTL_INT(_net_iflib, OID_AUTO, fast_intrs, CTLFLAG_RD,
658 		   &iflib_fast_intrs, 0, "# fast_intr calls");
659 SYSCTL_INT(_net_iflib, OID_AUTO, rx_unavail, CTLFLAG_RD,
660 		   &iflib_rx_unavail, 0, "# times rxeof called with no available data");
661 SYSCTL_INT(_net_iflib, OID_AUTO, rx_ctx_inactive, CTLFLAG_RD,
662 		   &iflib_rx_ctx_inactive, 0, "# times rxeof called with inactive context");
663 SYSCTL_INT(_net_iflib, OID_AUTO, rx_if_input, CTLFLAG_RD,
664 		   &iflib_rx_if_input, 0, "# times rxeof called if_input");
665 SYSCTL_INT(_net_iflib, OID_AUTO, rxd_flush, CTLFLAG_RD,
666 	         &iflib_rxd_flush, 0, "# times rxd_flush called");
667 SYSCTL_INT(_net_iflib, OID_AUTO, verbose_debug, CTLFLAG_RW,
668 		   &iflib_verbose_debug, 0, "enable verbose debugging");
669 
670 #define DBG_COUNTER_INC(name) atomic_add_int(&(iflib_ ## name), 1)
671 static void
672 iflib_debug_reset(void)
673 {
674 	iflib_tx_seen = iflib_tx_sent = iflib_tx_encap = iflib_rx_allocs =
675 		iflib_fl_refills = iflib_fl_refills_large = iflib_tx_frees =
676 		iflib_txq_drain_flushing = iflib_txq_drain_oactive =
677 		iflib_txq_drain_notready =
678 		iflib_encap_load_mbuf_fail = iflib_encap_pad_mbuf_fail =
679 		iflib_encap_txq_avail_fail = iflib_encap_txd_encap_fail =
680 		iflib_task_fn_rxs = iflib_rx_intr_enables = iflib_fast_intrs =
681 		iflib_rx_unavail =
682 		iflib_rx_ctx_inactive = iflib_rx_if_input =
683 		iflib_rxd_flush = 0;
684 }
685 
686 #else
687 #define DBG_COUNTER_INC(name)
688 static void iflib_debug_reset(void) {}
689 #endif
690 
691 #define IFLIB_DEBUG 0
692 
693 static void iflib_tx_structures_free(if_ctx_t ctx);
694 static void iflib_rx_structures_free(if_ctx_t ctx);
695 static int iflib_queues_alloc(if_ctx_t ctx);
696 static int iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq);
697 static int iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget);
698 static int iflib_qset_structures_setup(if_ctx_t ctx);
699 static int iflib_msix_init(if_ctx_t ctx);
700 static int iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filterarg, int *rid, const char *str);
701 static void iflib_txq_check_drain(iflib_txq_t txq, int budget);
702 static uint32_t iflib_txq_can_drain(struct ifmp_ring *);
703 #ifdef ALTQ
704 static void iflib_altq_if_start(if_t ifp);
705 static int iflib_altq_if_transmit(if_t ifp, struct mbuf *m);
706 #endif
707 static int iflib_register(if_ctx_t);
708 static void iflib_deregister(if_ctx_t);
709 static void iflib_unregister_vlan_handlers(if_ctx_t ctx);
710 static uint16_t iflib_get_mbuf_size_for(unsigned int size);
711 static void iflib_init_locked(if_ctx_t ctx);
712 static void iflib_add_device_sysctl_pre(if_ctx_t ctx);
713 static void iflib_add_device_sysctl_post(if_ctx_t ctx);
714 static void iflib_ifmp_purge(iflib_txq_t txq);
715 static void _iflib_pre_assert(if_softc_ctx_t scctx);
716 static void iflib_if_init_locked(if_ctx_t ctx);
717 static void iflib_free_intr_mem(if_ctx_t ctx);
718 #ifndef __NO_STRICT_ALIGNMENT
719 static struct mbuf * iflib_fixup_rx(struct mbuf *m);
720 #endif
721 
722 static SLIST_HEAD(cpu_offset_list, cpu_offset) cpu_offsets =
723     SLIST_HEAD_INITIALIZER(cpu_offsets);
724 struct cpu_offset {
725 	SLIST_ENTRY(cpu_offset) entries;
726 	cpuset_t	set;
727 	unsigned int	refcount;
728 	uint16_t	offset;
729 };
730 static struct mtx cpu_offset_mtx;
731 MTX_SYSINIT(iflib_cpu_offset, &cpu_offset_mtx, "iflib_cpu_offset lock",
732     MTX_DEF);
733 
734 DEBUGNET_DEFINE(iflib);
735 
736 static int
737 iflib_num_rx_descs(if_ctx_t ctx)
738 {
739 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
740 	if_shared_ctx_t sctx = ctx->ifc_sctx;
741 	uint16_t first_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
742 
743 	return scctx->isc_nrxd[first_rxq];
744 }
745 
746 static int
747 iflib_num_tx_descs(if_ctx_t ctx)
748 {
749 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
750 	if_shared_ctx_t sctx = ctx->ifc_sctx;
751 	uint16_t first_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
752 
753 	return scctx->isc_ntxd[first_txq];
754 }
755 
756 #ifdef DEV_NETMAP
757 #include <sys/selinfo.h>
758 #include <net/netmap.h>
759 #include <dev/netmap/netmap_kern.h>
760 
761 MODULE_DEPEND(iflib, netmap, 1, 1, 1);
762 
763 static int netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, bool init);
764 static void iflib_netmap_timer(void *arg);
765 
766 /*
767  * device-specific sysctl variables:
768  *
769  * iflib_crcstrip: 0: keep CRC in rx frames (default), 1: strip it.
770  *	During regular operations the CRC is stripped, but on some
771  *	hardware reception of frames not multiple of 64 is slower,
772  *	so using crcstrip=0 helps in benchmarks.
773  *
774  * iflib_rx_miss, iflib_rx_miss_bufs:
775  *	count packets that might be missed due to lost interrupts.
776  */
777 SYSCTL_DECL(_dev_netmap);
778 /*
779  * The xl driver by default strips CRCs and we do not override it.
780  */
781 
782 int iflib_crcstrip = 1;
783 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_crcstrip,
784     CTLFLAG_RW, &iflib_crcstrip, 1, "strip CRC on RX frames");
785 
786 int iflib_rx_miss, iflib_rx_miss_bufs;
787 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss,
788     CTLFLAG_RW, &iflib_rx_miss, 0, "potentially missed RX intr");
789 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss_bufs,
790     CTLFLAG_RW, &iflib_rx_miss_bufs, 0, "potentially missed RX intr bufs");
791 
792 /*
793  * Register/unregister. We are already under netmap lock.
794  * Only called on the first register or the last unregister.
795  */
796 static int
797 iflib_netmap_register(struct netmap_adapter *na, int onoff)
798 {
799 	if_t ifp = na->ifp;
800 	if_ctx_t ctx = ifp->if_softc;
801 	int status;
802 
803 	CTX_LOCK(ctx);
804 	if (!CTX_IS_VF(ctx))
805 		IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip);
806 
807 	iflib_stop(ctx);
808 
809 	/*
810 	 * Enable (or disable) netmap flags, and intercept (or restore)
811 	 * ifp->if_transmit. This is done once the device has been stopped
812 	 * to prevent race conditions. Also, this must be done after
813 	 * calling netmap_disable_all_rings() and before calling
814 	 * netmap_enable_all_rings(), so that these two functions see the
815 	 * updated state of the NAF_NETMAP_ON bit.
816 	 */
817 	if (onoff) {
818 		nm_set_native_flags(na);
819 	} else {
820 		nm_clear_native_flags(na);
821 	}
822 
823 	iflib_init_locked(ctx);
824 	IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip); // XXX why twice ?
825 	status = ifp->if_drv_flags & IFF_DRV_RUNNING ? 0 : 1;
826 	if (status)
827 		nm_clear_native_flags(na);
828 	CTX_UNLOCK(ctx);
829 	return (status);
830 }
831 
832 static int
833 iflib_netmap_config(struct netmap_adapter *na, struct nm_config_info *info)
834 {
835 	if_t ifp = na->ifp;
836 	if_ctx_t ctx = ifp->if_softc;
837 	iflib_rxq_t rxq = &ctx->ifc_rxqs[0];
838 	iflib_fl_t fl = &rxq->ifr_fl[0];
839 
840 	info->num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
841 	info->num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
842 	info->num_tx_descs = iflib_num_tx_descs(ctx);
843 	info->num_rx_descs = iflib_num_rx_descs(ctx);
844 	info->rx_buf_maxsize = fl->ifl_buf_size;
845 	nm_prinf("txr %u rxr %u txd %u rxd %u rbufsz %u",
846 		info->num_tx_rings, info->num_rx_rings, info->num_tx_descs,
847 		info->num_rx_descs, info->rx_buf_maxsize);
848 
849 	return 0;
850 }
851 
852 static int
853 netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, bool init)
854 {
855 	struct netmap_adapter *na = kring->na;
856 	u_int const lim = kring->nkr_num_slots - 1;
857 	struct netmap_ring *ring = kring->ring;
858 	bus_dmamap_t *map;
859 	struct if_rxd_update iru;
860 	if_ctx_t ctx = rxq->ifr_ctx;
861 	iflib_fl_t fl = &rxq->ifr_fl[0];
862 	u_int nic_i_first, nic_i;
863 	u_int nm_i;
864 	int i, n;
865 #if IFLIB_DEBUG_COUNTERS
866 	int rf_count = 0;
867 #endif
868 
869 	/*
870 	 * This function is used both at initialization and in rxsync.
871 	 * At initialization we need to prepare (with isc_rxd_refill())
872 	 * all the netmap buffers currently owned by the kernel, in
873 	 * such a way to keep fl->ifl_pidx and kring->nr_hwcur in sync
874 	 * (except for kring->nkr_hwofs). These may be less than
875 	 * kring->nkr_num_slots if netmap_reset() was called while
876 	 * an application using the kring that still owned some
877 	 * buffers.
878 	 * At rxsync time, both indexes point to the next buffer to be
879 	 * refilled.
880 	 * In any case we publish (with isc_rxd_flush()) up to
881 	 * (fl->ifl_pidx - 1) % N (included), to avoid the NIC tail/prod
882 	 * pointer to overrun the head/cons pointer, although this is
883 	 * not necessary for some NICs (e.g. vmx).
884 	 */
885 	if (__predict_false(init)) {
886 		n = kring->nkr_num_slots - nm_kr_rxspace(kring);
887 	} else {
888 		n = kring->rhead - kring->nr_hwcur;
889 		if (n == 0)
890 			return (0); /* Nothing to do. */
891 		if (n < 0)
892 			n += kring->nkr_num_slots;
893 	}
894 
895 	iru_init(&iru, rxq, 0 /* flid */);
896 	map = fl->ifl_sds.ifsd_map;
897 	nic_i = fl->ifl_pidx;
898 	nm_i = netmap_idx_n2k(kring, nic_i);
899 	if (__predict_false(init)) {
900 		/*
901 		 * On init/reset, nic_i must be 0, and we must
902 		 * start to refill from hwtail (see netmap_reset()).
903 		 */
904 		MPASS(nic_i == 0);
905 		MPASS(nm_i == kring->nr_hwtail);
906 	} else
907 		MPASS(nm_i == kring->nr_hwcur);
908 	DBG_COUNTER_INC(fl_refills);
909 	while (n > 0) {
910 #if IFLIB_DEBUG_COUNTERS
911 		if (++rf_count == 9)
912 			DBG_COUNTER_INC(fl_refills_large);
913 #endif
914 		nic_i_first = nic_i;
915 		for (i = 0; n > 0 && i < IFLIB_MAX_RX_REFRESH; n--, i++) {
916 			struct netmap_slot *slot = &ring->slot[nm_i];
917 			uint64_t paddr;
918 			void *addr = PNMB(na, slot, &paddr);
919 
920 			MPASS(i < IFLIB_MAX_RX_REFRESH);
921 
922 			if (addr == NETMAP_BUF_BASE(na)) /* bad buf */
923 			        return netmap_ring_reinit(kring);
924 
925 			fl->ifl_bus_addrs[i] = paddr +
926 			    nm_get_offset(kring, slot);
927 			fl->ifl_rxd_idxs[i] = nic_i;
928 
929 			if (__predict_false(init)) {
930 				netmap_load_map(na, fl->ifl_buf_tag,
931 				    map[nic_i], addr);
932 			} else if (slot->flags & NS_BUF_CHANGED) {
933 				/* buffer has changed, reload map */
934 				netmap_reload_map(na, fl->ifl_buf_tag,
935 				    map[nic_i], addr);
936 			}
937 			bus_dmamap_sync(fl->ifl_buf_tag, map[nic_i],
938 			    BUS_DMASYNC_PREREAD);
939 			slot->flags &= ~NS_BUF_CHANGED;
940 
941 			nm_i = nm_next(nm_i, lim);
942 			nic_i = nm_next(nic_i, lim);
943 		}
944 
945 		iru.iru_pidx = nic_i_first;
946 		iru.iru_count = i;
947 		ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
948 	}
949 	fl->ifl_pidx = nic_i;
950 	/*
951 	 * At the end of the loop we must have refilled everything
952 	 * we could possibly refill.
953 	 */
954 	MPASS(nm_i == kring->rhead);
955 	kring->nr_hwcur = nm_i;
956 
957 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
958 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
959 	ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id,
960 	    nm_prev(nic_i, lim));
961 	DBG_COUNTER_INC(rxd_flush);
962 
963 	return (0);
964 }
965 
966 #define NETMAP_TX_TIMER_US	90
967 
968 /*
969  * Reconcile kernel and user view of the transmit ring.
970  *
971  * All information is in the kring.
972  * Userspace wants to send packets up to the one before kring->rhead,
973  * kernel knows kring->nr_hwcur is the first unsent packet.
974  *
975  * Here we push packets out (as many as possible), and possibly
976  * reclaim buffers from previously completed transmission.
977  *
978  * The caller (netmap) guarantees that there is only one instance
979  * running at any time. Any interference with other driver
980  * methods should be handled by the individual drivers.
981  */
982 static int
983 iflib_netmap_txsync(struct netmap_kring *kring, int flags)
984 {
985 	struct netmap_adapter *na = kring->na;
986 	if_t ifp = na->ifp;
987 	struct netmap_ring *ring = kring->ring;
988 	u_int nm_i;	/* index into the netmap kring */
989 	u_int nic_i;	/* index into the NIC ring */
990 	u_int n;
991 	u_int const lim = kring->nkr_num_slots - 1;
992 	u_int const head = kring->rhead;
993 	struct if_pkt_info pi;
994 
995 	/*
996 	 * interrupts on every tx packet are expensive so request
997 	 * them every half ring, or where NS_REPORT is set
998 	 */
999 	u_int report_frequency = kring->nkr_num_slots >> 1;
1000 	/* device-specific */
1001 	if_ctx_t ctx = ifp->if_softc;
1002 	iflib_txq_t txq = &ctx->ifc_txqs[kring->ring_id];
1003 
1004 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1005 	    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1006 
1007 	/*
1008 	 * First part: process new packets to send.
1009 	 * nm_i is the current index in the netmap kring,
1010 	 * nic_i is the corresponding index in the NIC ring.
1011 	 *
1012 	 * If we have packets to send (nm_i != head)
1013 	 * iterate over the netmap ring, fetch length and update
1014 	 * the corresponding slot in the NIC ring. Some drivers also
1015 	 * need to update the buffer's physical address in the NIC slot
1016 	 * even NS_BUF_CHANGED is not set (PNMB computes the addresses).
1017 	 *
1018 	 * The netmap_reload_map() calls is especially expensive,
1019 	 * even when (as in this case) the tag is 0, so do only
1020 	 * when the buffer has actually changed.
1021 	 *
1022 	 * If possible do not set the report/intr bit on all slots,
1023 	 * but only a few times per ring or when NS_REPORT is set.
1024 	 *
1025 	 * Finally, on 10G and faster drivers, it might be useful
1026 	 * to prefetch the next slot and txr entry.
1027 	 */
1028 
1029 	nm_i = kring->nr_hwcur;
1030 	if (nm_i != head) {	/* we have new packets to send */
1031 		uint32_t pkt_len = 0, seg_idx = 0;
1032 		int nic_i_start = -1, flags = 0;
1033 		pkt_info_zero(&pi);
1034 		pi.ipi_segs = txq->ift_segs;
1035 		pi.ipi_qsidx = kring->ring_id;
1036 		nic_i = netmap_idx_k2n(kring, nm_i);
1037 
1038 		__builtin_prefetch(&ring->slot[nm_i]);
1039 		__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i]);
1040 		__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i]);
1041 
1042 		for (n = 0; nm_i != head; n++) {
1043 			struct netmap_slot *slot = &ring->slot[nm_i];
1044 			uint64_t offset = nm_get_offset(kring, slot);
1045 			u_int len = slot->len;
1046 			uint64_t paddr;
1047 			void *addr = PNMB(na, slot, &paddr);
1048 
1049 			flags |= (slot->flags & NS_REPORT ||
1050 				nic_i == 0 || nic_i == report_frequency) ?
1051 				IPI_TX_INTR : 0;
1052 
1053 			/*
1054 			 * If this is the first packet fragment, save the
1055 			 * index of the first NIC slot for later.
1056 			 */
1057 			if (nic_i_start < 0)
1058 				nic_i_start = nic_i;
1059 
1060 			pi.ipi_segs[seg_idx].ds_addr = paddr + offset;
1061 			pi.ipi_segs[seg_idx].ds_len = len;
1062 			if (len) {
1063 				pkt_len += len;
1064 				seg_idx++;
1065 			}
1066 
1067 			if (!(slot->flags & NS_MOREFRAG)) {
1068 				pi.ipi_len = pkt_len;
1069 				pi.ipi_nsegs = seg_idx;
1070 				pi.ipi_pidx = nic_i_start;
1071 				pi.ipi_ndescs = 0;
1072 				pi.ipi_flags = flags;
1073 
1074 				/* Prepare the NIC TX ring. */
1075 				ctx->isc_txd_encap(ctx->ifc_softc, &pi);
1076 				DBG_COUNTER_INC(tx_encap);
1077 
1078 				/* Reinit per-packet info for the next one. */
1079 				flags = seg_idx = pkt_len = 0;
1080 				nic_i_start = -1;
1081 			}
1082 
1083 			/* prefetch for next round */
1084 			__builtin_prefetch(&ring->slot[nm_i + 1]);
1085 			__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i + 1]);
1086 			__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i + 1]);
1087 
1088 			NM_CHECK_ADDR_LEN_OFF(na, len, offset);
1089 
1090 			if (slot->flags & NS_BUF_CHANGED) {
1091 				/* buffer has changed, reload map */
1092 				netmap_reload_map(na, txq->ift_buf_tag,
1093 				    txq->ift_sds.ifsd_map[nic_i], addr);
1094 			}
1095 			/* make sure changes to the buffer are synced */
1096 			bus_dmamap_sync(txq->ift_buf_tag,
1097 			    txq->ift_sds.ifsd_map[nic_i],
1098 			    BUS_DMASYNC_PREWRITE);
1099 
1100 			slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED | NS_MOREFRAG);
1101 			nm_i = nm_next(nm_i, lim);
1102 			nic_i = nm_next(nic_i, lim);
1103 		}
1104 		kring->nr_hwcur = nm_i;
1105 
1106 		/* synchronize the NIC ring */
1107 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1108 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1109 
1110 		/* (re)start the tx unit up to slot nic_i (excluded) */
1111 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, nic_i);
1112 	}
1113 
1114 	/*
1115 	 * Second part: reclaim buffers for completed transmissions.
1116 	 *
1117 	 * If there are unclaimed buffers, attempt to reclaim them.
1118 	 * If we don't manage to reclaim them all, and TX IRQs are not in use,
1119 	 * trigger a per-tx-queue timer to try again later.
1120 	 */
1121 	if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1122 		if (iflib_tx_credits_update(ctx, txq)) {
1123 			/* some tx completed, increment avail */
1124 			nic_i = txq->ift_cidx_processed;
1125 			kring->nr_hwtail = nm_prev(netmap_idx_n2k(kring, nic_i), lim);
1126 		}
1127 	}
1128 
1129 	if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ))
1130 		if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1131 			callout_reset_sbt_on(&txq->ift_netmap_timer,
1132 			    NETMAP_TX_TIMER_US * SBT_1US, SBT_1US,
1133 			    iflib_netmap_timer, txq,
1134 			    txq->ift_netmap_timer.c_cpu, 0);
1135 		}
1136 	return (0);
1137 }
1138 
1139 /*
1140  * Reconcile kernel and user view of the receive ring.
1141  * Same as for the txsync, this routine must be efficient.
1142  * The caller guarantees a single invocations, but races against
1143  * the rest of the driver should be handled here.
1144  *
1145  * On call, kring->rhead is the first packet that userspace wants
1146  * to keep, and kring->rcur is the wakeup point.
1147  * The kernel has previously reported packets up to kring->rtail.
1148  *
1149  * If (flags & NAF_FORCE_READ) also check for incoming packets irrespective
1150  * of whether or not we received an interrupt.
1151  */
1152 static int
1153 iflib_netmap_rxsync(struct netmap_kring *kring, int flags)
1154 {
1155 	struct netmap_adapter *na = kring->na;
1156 	struct netmap_ring *ring = kring->ring;
1157 	if_t ifp = na->ifp;
1158 	uint32_t nm_i;	/* index into the netmap ring */
1159 	uint32_t nic_i;	/* index into the NIC ring */
1160 	u_int n;
1161 	u_int const lim = kring->nkr_num_slots - 1;
1162 	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
1163 	int i = 0;
1164 
1165 	if_ctx_t ctx = ifp->if_softc;
1166 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1167 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1168 	iflib_rxq_t rxq = &ctx->ifc_rxqs[kring->ring_id];
1169 	iflib_fl_t fl = &rxq->ifr_fl[0];
1170 	struct if_rxd_info ri;
1171 	qidx_t *cidxp;
1172 
1173 	/*
1174 	 * netmap only uses free list 0, to avoid out of order consumption
1175 	 * of receive buffers
1176 	 */
1177 
1178 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
1179 	    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1180 
1181 	/*
1182 	 * First part: import newly received packets.
1183 	 *
1184 	 * nm_i is the index of the next free slot in the netmap ring,
1185 	 * nic_i is the index of the next received packet in the NIC ring
1186 	 * (or in the free list 0 if IFLIB_HAS_RXCQ is set), and they may
1187 	 * differ in case if_init() has been called while
1188 	 * in netmap mode. For the receive ring we have
1189 	 *
1190 	 *	nic_i = fl->ifl_cidx;
1191 	 *	nm_i = kring->nr_hwtail (previous)
1192 	 * and
1193 	 *	nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1194 	 *
1195 	 * fl->ifl_cidx is set to 0 on a ring reinit
1196 	 */
1197 	if (netmap_no_pendintr || force_update) {
1198 		uint32_t hwtail_lim = nm_prev(kring->nr_hwcur, lim);
1199 		bool have_rxcq = sctx->isc_flags & IFLIB_HAS_RXCQ;
1200 		int crclen = iflib_crcstrip ? 0 : 4;
1201 		int error, avail;
1202 
1203 		/*
1204 		 * For the free list consumer index, we use the same
1205 		 * logic as in iflib_rxeof().
1206 		 */
1207 		if (have_rxcq)
1208 			cidxp = &rxq->ifr_cq_cidx;
1209 		else
1210 			cidxp = &fl->ifl_cidx;
1211 		avail = ctx->isc_rxd_available(ctx->ifc_softc,
1212 		    rxq->ifr_id, *cidxp, USHRT_MAX);
1213 
1214 		nic_i = fl->ifl_cidx;
1215 		nm_i = netmap_idx_n2k(kring, nic_i);
1216 		MPASS(nm_i == kring->nr_hwtail);
1217 		for (n = 0; avail > 0 && nm_i != hwtail_lim; n++, avail--) {
1218 			rxd_info_zero(&ri);
1219 			ri.iri_frags = rxq->ifr_frags;
1220 			ri.iri_qsidx = kring->ring_id;
1221 			ri.iri_ifp = ctx->ifc_ifp;
1222 			ri.iri_cidx = *cidxp;
1223 
1224 			error = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
1225 			for (i = 0; i < ri.iri_nfrags; i++) {
1226 				if (error) {
1227 					ring->slot[nm_i].len = 0;
1228 					ring->slot[nm_i].flags = 0;
1229 				} else {
1230 					ring->slot[nm_i].len = ri.iri_frags[i].irf_len;
1231 					if (i == (ri.iri_nfrags - 1)) {
1232 						ring->slot[nm_i].len -= crclen;
1233 						ring->slot[nm_i].flags = 0;
1234 					} else
1235 						ring->slot[nm_i].flags = NS_MOREFRAG;
1236 				}
1237 
1238 				bus_dmamap_sync(fl->ifl_buf_tag,
1239 				    fl->ifl_sds.ifsd_map[nic_i], BUS_DMASYNC_POSTREAD);
1240 				nm_i = nm_next(nm_i, lim);
1241 				fl->ifl_cidx = nic_i = nm_next(nic_i, lim);
1242 			}
1243 
1244 			if (have_rxcq) {
1245 				*cidxp = ri.iri_cidx;
1246 				while (*cidxp >= scctx->isc_nrxd[0])
1247 					*cidxp -= scctx->isc_nrxd[0];
1248 			}
1249 
1250 		}
1251 		if (n) { /* update the state variables */
1252 			if (netmap_no_pendintr && !force_update) {
1253 				/* diagnostics */
1254 				iflib_rx_miss ++;
1255 				iflib_rx_miss_bufs += n;
1256 			}
1257 			kring->nr_hwtail = nm_i;
1258 		}
1259 		kring->nr_kflags &= ~NKR_PENDINTR;
1260 	}
1261 	/*
1262 	 * Second part: skip past packets that userspace has released.
1263 	 * (kring->nr_hwcur to head excluded),
1264 	 * and make the buffers available for reception.
1265 	 * As usual nm_i is the index in the netmap ring,
1266 	 * nic_i is the index in the NIC ring, and
1267 	 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1268 	 */
1269 	netmap_fl_refill(rxq, kring, false);
1270 
1271 	return (0);
1272 }
1273 
1274 static void
1275 iflib_netmap_intr(struct netmap_adapter *na, int onoff)
1276 {
1277 	if_ctx_t ctx = na->ifp->if_softc;
1278 
1279 	CTX_LOCK(ctx);
1280 	if (onoff) {
1281 		IFDI_INTR_ENABLE(ctx);
1282 	} else {
1283 		IFDI_INTR_DISABLE(ctx);
1284 	}
1285 	CTX_UNLOCK(ctx);
1286 }
1287 
1288 static int
1289 iflib_netmap_attach(if_ctx_t ctx)
1290 {
1291 	struct netmap_adapter na;
1292 
1293 	bzero(&na, sizeof(na));
1294 
1295 	na.ifp = ctx->ifc_ifp;
1296 	na.na_flags = NAF_BDG_MAYSLEEP | NAF_MOREFRAG | NAF_OFFSETS;
1297 	MPASS(ctx->ifc_softc_ctx.isc_ntxqsets);
1298 	MPASS(ctx->ifc_softc_ctx.isc_nrxqsets);
1299 
1300 	na.num_tx_desc = iflib_num_tx_descs(ctx);
1301 	na.num_rx_desc = iflib_num_rx_descs(ctx);
1302 	na.nm_txsync = iflib_netmap_txsync;
1303 	na.nm_rxsync = iflib_netmap_rxsync;
1304 	na.nm_register = iflib_netmap_register;
1305 	na.nm_intr = iflib_netmap_intr;
1306 	na.nm_config = iflib_netmap_config;
1307 	na.num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
1308 	na.num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
1309 	return (netmap_attach(&na));
1310 }
1311 
1312 static int
1313 iflib_netmap_txq_init(if_ctx_t ctx, iflib_txq_t txq)
1314 {
1315 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1316 	struct netmap_slot *slot;
1317 
1318 	slot = netmap_reset(na, NR_TX, txq->ift_id, 0);
1319 	if (slot == NULL)
1320 		return (0);
1321 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxd[0]; i++) {
1322 		/*
1323 		 * In netmap mode, set the map for the packet buffer.
1324 		 * NOTE: Some drivers (not this one) also need to set
1325 		 * the physical buffer address in the NIC ring.
1326 		 * netmap_idx_n2k() maps a nic index, i, into the corresponding
1327 		 * netmap slot index, si
1328 		 */
1329 		int si = netmap_idx_n2k(na->tx_rings[txq->ift_id], i);
1330 		netmap_load_map(na, txq->ift_buf_tag, txq->ift_sds.ifsd_map[i],
1331 		    NMB(na, slot + si));
1332 	}
1333 	return (1);
1334 }
1335 
1336 static int
1337 iflib_netmap_rxq_init(if_ctx_t ctx, iflib_rxq_t rxq)
1338 {
1339 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1340 	struct netmap_kring *kring;
1341 	struct netmap_slot *slot;
1342 
1343 	slot = netmap_reset(na, NR_RX, rxq->ifr_id, 0);
1344 	if (slot == NULL)
1345 		return (0);
1346 	kring = na->rx_rings[rxq->ifr_id];
1347 	netmap_fl_refill(rxq, kring, true);
1348 	return (1);
1349 }
1350 
1351 static void
1352 iflib_netmap_timer(void *arg)
1353 {
1354 	iflib_txq_t txq = arg;
1355 	if_ctx_t ctx = txq->ift_ctx;
1356 
1357 	/*
1358 	 * Wake up the netmap application, to give it a chance to
1359 	 * call txsync and reclaim more completed TX buffers.
1360 	 */
1361 	netmap_tx_irq(ctx->ifc_ifp, txq->ift_id);
1362 }
1363 
1364 #define iflib_netmap_detach(ifp) netmap_detach(ifp)
1365 
1366 #else
1367 #define iflib_netmap_txq_init(ctx, txq) (0)
1368 #define iflib_netmap_rxq_init(ctx, rxq) (0)
1369 #define iflib_netmap_detach(ifp)
1370 #define netmap_enable_all_rings(ifp)
1371 #define netmap_disable_all_rings(ifp)
1372 
1373 #define iflib_netmap_attach(ctx) (0)
1374 #define netmap_rx_irq(ifp, qid, budget) (0)
1375 #endif
1376 
1377 #if defined(__i386__) || defined(__amd64__)
1378 static __inline void
1379 prefetch(void *x)
1380 {
1381 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1382 }
1383 static __inline void
1384 prefetch2cachelines(void *x)
1385 {
1386 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1387 #if (CACHE_LINE_SIZE < 128)
1388 	__asm volatile("prefetcht0 %0" :: "m" (*(((unsigned long *)x)+CACHE_LINE_SIZE/(sizeof(unsigned long)))));
1389 #endif
1390 }
1391 #else
1392 #define prefetch(x)
1393 #define prefetch2cachelines(x)
1394 #endif
1395 
1396 static void
1397 iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid)
1398 {
1399 	iflib_fl_t fl;
1400 
1401 	fl = &rxq->ifr_fl[flid];
1402 	iru->iru_paddrs = fl->ifl_bus_addrs;
1403 	iru->iru_idxs = fl->ifl_rxd_idxs;
1404 	iru->iru_qsidx = rxq->ifr_id;
1405 	iru->iru_buf_size = fl->ifl_buf_size;
1406 	iru->iru_flidx = fl->ifl_id;
1407 }
1408 
1409 static void
1410 _iflib_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int err)
1411 {
1412 	if (err)
1413 		return;
1414 	*(bus_addr_t *) arg = segs[0].ds_addr;
1415 }
1416 
1417 #define	DMA_WIDTH_TO_BUS_LOWADDR(width)				\
1418 	(((width) == 0) || (width) == flsll(BUS_SPACE_MAXADDR) ?	\
1419 	    BUS_SPACE_MAXADDR : (1ULL << (width)) - 1ULL)
1420 
1421 int
1422 iflib_dma_alloc_align(if_ctx_t ctx, int size, int align, iflib_dma_info_t dma, int mapflags)
1423 {
1424 	int err;
1425 	device_t dev = ctx->ifc_dev;
1426 	bus_addr_t lowaddr;
1427 
1428 	lowaddr = DMA_WIDTH_TO_BUS_LOWADDR(ctx->ifc_softc_ctx.isc_dma_width);
1429 
1430 	err = bus_dma_tag_create(bus_get_dma_tag(dev),	/* parent */
1431 				align, 0,		/* alignment, bounds */
1432 				lowaddr,		/* lowaddr */
1433 				BUS_SPACE_MAXADDR,	/* highaddr */
1434 				NULL, NULL,		/* filter, filterarg */
1435 				size,			/* maxsize */
1436 				1,			/* nsegments */
1437 				size,			/* maxsegsize */
1438 				BUS_DMA_ALLOCNOW,	/* flags */
1439 				NULL,			/* lockfunc */
1440 				NULL,			/* lockarg */
1441 				&dma->idi_tag);
1442 	if (err) {
1443 		device_printf(dev,
1444 		    "%s: bus_dma_tag_create failed: %d\n",
1445 		    __func__, err);
1446 		goto fail_0;
1447 	}
1448 
1449 	err = bus_dmamem_alloc(dma->idi_tag, (void**) &dma->idi_vaddr,
1450 	    BUS_DMA_NOWAIT | BUS_DMA_COHERENT | BUS_DMA_ZERO, &dma->idi_map);
1451 	if (err) {
1452 		device_printf(dev,
1453 		    "%s: bus_dmamem_alloc(%ju) failed: %d\n",
1454 		    __func__, (uintmax_t)size, err);
1455 		goto fail_1;
1456 	}
1457 
1458 	dma->idi_paddr = IF_BAD_DMA;
1459 	err = bus_dmamap_load(dma->idi_tag, dma->idi_map, dma->idi_vaddr,
1460 	    size, _iflib_dmamap_cb, &dma->idi_paddr, mapflags | BUS_DMA_NOWAIT);
1461 	if (err || dma->idi_paddr == IF_BAD_DMA) {
1462 		device_printf(dev,
1463 		    "%s: bus_dmamap_load failed: %d\n",
1464 		    __func__, err);
1465 		goto fail_2;
1466 	}
1467 
1468 	dma->idi_size = size;
1469 	return (0);
1470 
1471 fail_2:
1472 	bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1473 fail_1:
1474 	bus_dma_tag_destroy(dma->idi_tag);
1475 fail_0:
1476 	dma->idi_tag = NULL;
1477 
1478 	return (err);
1479 }
1480 
1481 int
1482 iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags)
1483 {
1484 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1485 
1486 	KASSERT(sctx->isc_q_align != 0, ("alignment value not initialized"));
1487 
1488 	return (iflib_dma_alloc_align(ctx, size, sctx->isc_q_align, dma, mapflags));
1489 }
1490 
1491 int
1492 iflib_dma_alloc_multi(if_ctx_t ctx, int *sizes, iflib_dma_info_t *dmalist, int mapflags, int count)
1493 {
1494 	int i, err;
1495 	iflib_dma_info_t *dmaiter;
1496 
1497 	dmaiter = dmalist;
1498 	for (i = 0; i < count; i++, dmaiter++) {
1499 		if ((err = iflib_dma_alloc(ctx, sizes[i], *dmaiter, mapflags)) != 0)
1500 			break;
1501 	}
1502 	if (err)
1503 		iflib_dma_free_multi(dmalist, i);
1504 	return (err);
1505 }
1506 
1507 void
1508 iflib_dma_free(iflib_dma_info_t dma)
1509 {
1510 	if (dma->idi_tag == NULL)
1511 		return;
1512 	if (dma->idi_paddr != IF_BAD_DMA) {
1513 		bus_dmamap_sync(dma->idi_tag, dma->idi_map,
1514 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1515 		bus_dmamap_unload(dma->idi_tag, dma->idi_map);
1516 		dma->idi_paddr = IF_BAD_DMA;
1517 	}
1518 	if (dma->idi_vaddr != NULL) {
1519 		bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1520 		dma->idi_vaddr = NULL;
1521 	}
1522 	bus_dma_tag_destroy(dma->idi_tag);
1523 	dma->idi_tag = NULL;
1524 }
1525 
1526 void
1527 iflib_dma_free_multi(iflib_dma_info_t *dmalist, int count)
1528 {
1529 	int i;
1530 	iflib_dma_info_t *dmaiter = dmalist;
1531 
1532 	for (i = 0; i < count; i++, dmaiter++)
1533 		iflib_dma_free(*dmaiter);
1534 }
1535 
1536 static int
1537 iflib_fast_intr(void *arg)
1538 {
1539 	iflib_filter_info_t info = arg;
1540 	struct grouptask *gtask = info->ifi_task;
1541 	int result;
1542 
1543 	DBG_COUNTER_INC(fast_intrs);
1544 	if (info->ifi_filter != NULL) {
1545 		result = info->ifi_filter(info->ifi_filter_arg);
1546 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1547 			return (result);
1548 	}
1549 
1550 	GROUPTASK_ENQUEUE(gtask);
1551 	return (FILTER_HANDLED);
1552 }
1553 
1554 static int
1555 iflib_fast_intr_rxtx(void *arg)
1556 {
1557 	iflib_filter_info_t info = arg;
1558 	struct grouptask *gtask = info->ifi_task;
1559 	if_ctx_t ctx;
1560 	iflib_rxq_t rxq = (iflib_rxq_t)info->ifi_ctx;
1561 	iflib_txq_t txq;
1562 	void *sc;
1563 	int i, cidx, result;
1564 	qidx_t txqid;
1565 	bool intr_enable, intr_legacy;
1566 
1567 	DBG_COUNTER_INC(fast_intrs);
1568 	if (info->ifi_filter != NULL) {
1569 		result = info->ifi_filter(info->ifi_filter_arg);
1570 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1571 			return (result);
1572 	}
1573 
1574 	ctx = rxq->ifr_ctx;
1575 	sc = ctx->ifc_softc;
1576 	intr_enable = false;
1577 	intr_legacy = !!(ctx->ifc_flags & IFC_LEGACY);
1578 	MPASS(rxq->ifr_ntxqirq);
1579 	for (i = 0; i < rxq->ifr_ntxqirq; i++) {
1580 		txqid = rxq->ifr_txqid[i];
1581 		txq = &ctx->ifc_txqs[txqid];
1582 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1583 		    BUS_DMASYNC_POSTREAD);
1584 		if (!ctx->isc_txd_credits_update(sc, txqid, false)) {
1585 			if (intr_legacy)
1586 				intr_enable = true;
1587 			else
1588 				IFDI_TX_QUEUE_INTR_ENABLE(ctx, txqid);
1589 			continue;
1590 		}
1591 		GROUPTASK_ENQUEUE(&txq->ift_task);
1592 	}
1593 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_RXCQ)
1594 		cidx = rxq->ifr_cq_cidx;
1595 	else
1596 		cidx = rxq->ifr_fl[0].ifl_cidx;
1597 	if (iflib_rxd_avail(ctx, rxq, cidx, 1))
1598 		GROUPTASK_ENQUEUE(gtask);
1599 	else {
1600 		if (intr_legacy)
1601 			intr_enable = true;
1602 		else
1603 			IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
1604 		DBG_COUNTER_INC(rx_intr_enables);
1605 	}
1606 	if (intr_enable)
1607 		IFDI_INTR_ENABLE(ctx);
1608 	return (FILTER_HANDLED);
1609 }
1610 
1611 static int
1612 iflib_fast_intr_ctx(void *arg)
1613 {
1614 	iflib_filter_info_t info = arg;
1615 	struct grouptask *gtask = info->ifi_task;
1616 	int result;
1617 
1618 	DBG_COUNTER_INC(fast_intrs);
1619 	if (info->ifi_filter != NULL) {
1620 		result = info->ifi_filter(info->ifi_filter_arg);
1621 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1622 			return (result);
1623 	}
1624 
1625 	GROUPTASK_ENQUEUE(gtask);
1626 	return (FILTER_HANDLED);
1627 }
1628 
1629 static int
1630 _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
1631 		 driver_filter_t filter, driver_intr_t handler, void *arg,
1632 		 const char *name)
1633 {
1634 	struct resource *res;
1635 	void *tag = NULL;
1636 	device_t dev = ctx->ifc_dev;
1637 	int flags, i, rc;
1638 
1639 	flags = RF_ACTIVE;
1640 	if (ctx->ifc_flags & IFC_LEGACY)
1641 		flags |= RF_SHAREABLE;
1642 	MPASS(rid < 512);
1643 	i = rid;
1644 	res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &i, flags);
1645 	if (res == NULL) {
1646 		device_printf(dev,
1647 		    "failed to allocate IRQ for rid %d, name %s.\n", rid, name);
1648 		return (ENOMEM);
1649 	}
1650 	irq->ii_res = res;
1651 	KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL"));
1652 	rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET,
1653 						filter, handler, arg, &tag);
1654 	if (rc != 0) {
1655 		device_printf(dev,
1656 		    "failed to setup interrupt for rid %d, name %s: %d\n",
1657 					  rid, name ? name : "unknown", rc);
1658 		return (rc);
1659 	} else if (name)
1660 		bus_describe_intr(dev, res, tag, "%s", name);
1661 
1662 	irq->ii_tag = tag;
1663 	return (0);
1664 }
1665 
1666 /*********************************************************************
1667  *
1668  *  Allocate DMA resources for TX buffers as well as memory for the TX
1669  *  mbuf map.  TX DMA maps (non-TSO/TSO) and TX mbuf map are kept in a
1670  *  iflib_sw_tx_desc_array structure, storing all the information that
1671  *  is needed to transmit a packet on the wire.  This is called only
1672  *  once at attach, setup is done every reset.
1673  *
1674  **********************************************************************/
1675 static int
1676 iflib_txsd_alloc(iflib_txq_t txq)
1677 {
1678 	if_ctx_t ctx = txq->ift_ctx;
1679 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1680 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1681 	device_t dev = ctx->ifc_dev;
1682 	bus_size_t tsomaxsize;
1683 	bus_addr_t lowaddr;
1684 	int err, nsegments, ntsosegments;
1685 	bool tso;
1686 
1687 	nsegments = scctx->isc_tx_nsegments;
1688 	ntsosegments = scctx->isc_tx_tso_segments_max;
1689 	tsomaxsize = scctx->isc_tx_tso_size_max;
1690 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_VLAN_MTU)
1691 		tsomaxsize += sizeof(struct ether_vlan_header);
1692 	MPASS(scctx->isc_ntxd[0] > 0);
1693 	MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0);
1694 	MPASS(nsegments > 0);
1695 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) {
1696 		MPASS(ntsosegments > 0);
1697 		MPASS(sctx->isc_tso_maxsize >= tsomaxsize);
1698 	}
1699 
1700 	lowaddr = DMA_WIDTH_TO_BUS_LOWADDR(scctx->isc_dma_width);
1701 
1702 	/*
1703 	 * Set up DMA tags for TX buffers.
1704 	 */
1705 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1706 			       1, 0,			/* alignment, bounds */
1707 			       lowaddr,			/* lowaddr */
1708 			       BUS_SPACE_MAXADDR,	/* highaddr */
1709 			       NULL, NULL,		/* filter, filterarg */
1710 			       sctx->isc_tx_maxsize,		/* maxsize */
1711 			       nsegments,	/* nsegments */
1712 			       sctx->isc_tx_maxsegsize,	/* maxsegsize */
1713 			       0,			/* flags */
1714 			       NULL,			/* lockfunc */
1715 			       NULL,			/* lockfuncarg */
1716 			       &txq->ift_buf_tag))) {
1717 		device_printf(dev,"Unable to allocate TX DMA tag: %d\n", err);
1718 		device_printf(dev,"maxsize: %ju nsegments: %d maxsegsize: %ju\n",
1719 		    (uintmax_t)sctx->isc_tx_maxsize, nsegments, (uintmax_t)sctx->isc_tx_maxsegsize);
1720 		goto fail;
1721 	}
1722 	tso = (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) != 0;
1723 	if (tso && (err = bus_dma_tag_create(bus_get_dma_tag(dev),
1724 			       1, 0,			/* alignment, bounds */
1725 			       lowaddr,			/* lowaddr */
1726 			       BUS_SPACE_MAXADDR,	/* highaddr */
1727 			       NULL, NULL,		/* filter, filterarg */
1728 			       tsomaxsize,		/* maxsize */
1729 			       ntsosegments,	/* nsegments */
1730 			       sctx->isc_tso_maxsegsize,/* maxsegsize */
1731 			       0,			/* flags */
1732 			       NULL,			/* lockfunc */
1733 			       NULL,			/* lockfuncarg */
1734 			       &txq->ift_tso_buf_tag))) {
1735 		device_printf(dev, "Unable to allocate TSO TX DMA tag: %d\n",
1736 		    err);
1737 		goto fail;
1738 	}
1739 
1740 	/* Allocate memory for the TX mbuf map. */
1741 	if (!(txq->ift_sds.ifsd_m =
1742 	    (struct mbuf **) malloc(sizeof(struct mbuf *) *
1743 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1744 		device_printf(dev, "Unable to allocate TX mbuf map memory\n");
1745 		err = ENOMEM;
1746 		goto fail;
1747 	}
1748 
1749 	/*
1750 	 * Create the DMA maps for TX buffers.
1751 	 */
1752 	if ((txq->ift_sds.ifsd_map = (bus_dmamap_t *)malloc(
1753 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1754 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1755 		device_printf(dev,
1756 		    "Unable to allocate TX buffer DMA map memory\n");
1757 		err = ENOMEM;
1758 		goto fail;
1759 	}
1760 	if (tso && (txq->ift_sds.ifsd_tso_map = (bus_dmamap_t *)malloc(
1761 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1762 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1763 		device_printf(dev,
1764 		    "Unable to allocate TSO TX buffer map memory\n");
1765 		err = ENOMEM;
1766 		goto fail;
1767 	}
1768 	for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) {
1769 		err = bus_dmamap_create(txq->ift_buf_tag, 0,
1770 		    &txq->ift_sds.ifsd_map[i]);
1771 		if (err != 0) {
1772 			device_printf(dev, "Unable to create TX DMA map\n");
1773 			goto fail;
1774 		}
1775 		if (!tso)
1776 			continue;
1777 		err = bus_dmamap_create(txq->ift_tso_buf_tag, 0,
1778 		    &txq->ift_sds.ifsd_tso_map[i]);
1779 		if (err != 0) {
1780 			device_printf(dev, "Unable to create TSO TX DMA map\n");
1781 			goto fail;
1782 		}
1783 	}
1784 	return (0);
1785 fail:
1786 	/* We free all, it handles case where we are in the middle */
1787 	iflib_tx_structures_free(ctx);
1788 	return (err);
1789 }
1790 
1791 static void
1792 iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i)
1793 {
1794 	bus_dmamap_t map;
1795 
1796 	if (txq->ift_sds.ifsd_map != NULL) {
1797 		map = txq->ift_sds.ifsd_map[i];
1798 		bus_dmamap_sync(txq->ift_buf_tag, map, BUS_DMASYNC_POSTWRITE);
1799 		bus_dmamap_unload(txq->ift_buf_tag, map);
1800 		bus_dmamap_destroy(txq->ift_buf_tag, map);
1801 		txq->ift_sds.ifsd_map[i] = NULL;
1802 	}
1803 
1804 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1805 		map = txq->ift_sds.ifsd_tso_map[i];
1806 		bus_dmamap_sync(txq->ift_tso_buf_tag, map,
1807 		    BUS_DMASYNC_POSTWRITE);
1808 		bus_dmamap_unload(txq->ift_tso_buf_tag, map);
1809 		bus_dmamap_destroy(txq->ift_tso_buf_tag, map);
1810 		txq->ift_sds.ifsd_tso_map[i] = NULL;
1811 	}
1812 }
1813 
1814 static void
1815 iflib_txq_destroy(iflib_txq_t txq)
1816 {
1817 	if_ctx_t ctx = txq->ift_ctx;
1818 
1819 	for (int i = 0; i < txq->ift_size; i++)
1820 		iflib_txsd_destroy(ctx, txq, i);
1821 
1822 	if (txq->ift_br != NULL) {
1823 		ifmp_ring_free(txq->ift_br);
1824 		txq->ift_br = NULL;
1825 	}
1826 
1827 	mtx_destroy(&txq->ift_mtx);
1828 
1829 	if (txq->ift_sds.ifsd_map != NULL) {
1830 		free(txq->ift_sds.ifsd_map, M_IFLIB);
1831 		txq->ift_sds.ifsd_map = NULL;
1832 	}
1833 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1834 		free(txq->ift_sds.ifsd_tso_map, M_IFLIB);
1835 		txq->ift_sds.ifsd_tso_map = NULL;
1836 	}
1837 	if (txq->ift_sds.ifsd_m != NULL) {
1838 		free(txq->ift_sds.ifsd_m, M_IFLIB);
1839 		txq->ift_sds.ifsd_m = NULL;
1840 	}
1841 	if (txq->ift_buf_tag != NULL) {
1842 		bus_dma_tag_destroy(txq->ift_buf_tag);
1843 		txq->ift_buf_tag = NULL;
1844 	}
1845 	if (txq->ift_tso_buf_tag != NULL) {
1846 		bus_dma_tag_destroy(txq->ift_tso_buf_tag);
1847 		txq->ift_tso_buf_tag = NULL;
1848 	}
1849 	if (txq->ift_ifdi != NULL) {
1850 		free(txq->ift_ifdi, M_IFLIB);
1851 	}
1852 }
1853 
1854 static void
1855 iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i)
1856 {
1857 	struct mbuf **mp;
1858 
1859 	mp = &txq->ift_sds.ifsd_m[i];
1860 	if (*mp == NULL)
1861 		return;
1862 
1863 	if (txq->ift_sds.ifsd_map != NULL) {
1864 		bus_dmamap_sync(txq->ift_buf_tag,
1865 		    txq->ift_sds.ifsd_map[i], BUS_DMASYNC_POSTWRITE);
1866 		bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[i]);
1867 	}
1868 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1869 		bus_dmamap_sync(txq->ift_tso_buf_tag,
1870 		    txq->ift_sds.ifsd_tso_map[i], BUS_DMASYNC_POSTWRITE);
1871 		bus_dmamap_unload(txq->ift_tso_buf_tag,
1872 		    txq->ift_sds.ifsd_tso_map[i]);
1873 	}
1874 	m_freem(*mp);
1875 	DBG_COUNTER_INC(tx_frees);
1876 	*mp = NULL;
1877 }
1878 
1879 static int
1880 iflib_txq_setup(iflib_txq_t txq)
1881 {
1882 	if_ctx_t ctx = txq->ift_ctx;
1883 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1884 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1885 	iflib_dma_info_t di;
1886 	int i;
1887 
1888 	/* Set number of descriptors available */
1889 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
1890 	/* XXX make configurable */
1891 	txq->ift_update_freq = IFLIB_DEFAULT_TX_UPDATE_FREQ;
1892 
1893 	/* Reset indices */
1894 	txq->ift_cidx_processed = 0;
1895 	txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0;
1896 	txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset];
1897 
1898 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1899 		bzero((void *)di->idi_vaddr, di->idi_size);
1900 
1901 	IFDI_TXQ_SETUP(ctx, txq->ift_id);
1902 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1903 		bus_dmamap_sync(di->idi_tag, di->idi_map,
1904 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1905 	return (0);
1906 }
1907 
1908 /*********************************************************************
1909  *
1910  *  Allocate DMA resources for RX buffers as well as memory for the RX
1911  *  mbuf map, direct RX cluster pointer map and RX cluster bus address
1912  *  map.  RX DMA map, RX mbuf map, direct RX cluster pointer map and
1913  *  RX cluster map are kept in a iflib_sw_rx_desc_array structure.
1914  *  Since we use use one entry in iflib_sw_rx_desc_array per received
1915  *  packet, the maximum number of entries we'll need is equal to the
1916  *  number of hardware receive descriptors that we've allocated.
1917  *
1918  **********************************************************************/
1919 static int
1920 iflib_rxsd_alloc(iflib_rxq_t rxq)
1921 {
1922 	if_ctx_t ctx = rxq->ifr_ctx;
1923 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1924 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1925 	device_t dev = ctx->ifc_dev;
1926 	iflib_fl_t fl;
1927 	bus_addr_t lowaddr;
1928 	int			err;
1929 
1930 	MPASS(scctx->isc_nrxd[0] > 0);
1931 	MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0);
1932 
1933 	lowaddr = DMA_WIDTH_TO_BUS_LOWADDR(scctx->isc_dma_width);
1934 
1935 	fl = rxq->ifr_fl;
1936 	for (int i = 0; i <  rxq->ifr_nfl; i++, fl++) {
1937 		fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */
1938 		/* Set up DMA tag for RX buffers. */
1939 		err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1940 					 1, 0,			/* alignment, bounds */
1941 					 lowaddr,		/* lowaddr */
1942 					 BUS_SPACE_MAXADDR,	/* highaddr */
1943 					 NULL, NULL,		/* filter, filterarg */
1944 					 sctx->isc_rx_maxsize,	/* maxsize */
1945 					 sctx->isc_rx_nsegments,	/* nsegments */
1946 					 sctx->isc_rx_maxsegsize,	/* maxsegsize */
1947 					 0,			/* flags */
1948 					 NULL,			/* lockfunc */
1949 					 NULL,			/* lockarg */
1950 					 &fl->ifl_buf_tag);
1951 		if (err) {
1952 			device_printf(dev,
1953 			    "Unable to allocate RX DMA tag: %d\n", err);
1954 			goto fail;
1955 		}
1956 
1957 		/* Allocate memory for the RX mbuf map. */
1958 		if (!(fl->ifl_sds.ifsd_m =
1959 		      (struct mbuf **) malloc(sizeof(struct mbuf *) *
1960 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1961 			device_printf(dev,
1962 			    "Unable to allocate RX mbuf map memory\n");
1963 			err = ENOMEM;
1964 			goto fail;
1965 		}
1966 
1967 		/* Allocate memory for the direct RX cluster pointer map. */
1968 		if (!(fl->ifl_sds.ifsd_cl =
1969 		      (caddr_t *) malloc(sizeof(caddr_t) *
1970 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1971 			device_printf(dev,
1972 			    "Unable to allocate RX cluster map memory\n");
1973 			err = ENOMEM;
1974 			goto fail;
1975 		}
1976 
1977 		/* Allocate memory for the RX cluster bus address map. */
1978 		if (!(fl->ifl_sds.ifsd_ba =
1979 		      (bus_addr_t *) malloc(sizeof(bus_addr_t) *
1980 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1981 			device_printf(dev,
1982 			    "Unable to allocate RX bus address map memory\n");
1983 			err = ENOMEM;
1984 			goto fail;
1985 		}
1986 
1987 		/*
1988 		 * Create the DMA maps for RX buffers.
1989 		 */
1990 		if (!(fl->ifl_sds.ifsd_map =
1991 		      (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1992 			device_printf(dev,
1993 			    "Unable to allocate RX buffer DMA map memory\n");
1994 			err = ENOMEM;
1995 			goto fail;
1996 		}
1997 		for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++) {
1998 			err = bus_dmamap_create(fl->ifl_buf_tag, 0,
1999 			    &fl->ifl_sds.ifsd_map[i]);
2000 			if (err != 0) {
2001 				device_printf(dev, "Unable to create RX buffer DMA map\n");
2002 				goto fail;
2003 			}
2004 		}
2005 	}
2006 	return (0);
2007 
2008 fail:
2009 	iflib_rx_structures_free(ctx);
2010 	return (err);
2011 }
2012 
2013 /*
2014  * Internal service routines
2015  */
2016 
2017 struct rxq_refill_cb_arg {
2018 	int               error;
2019 	bus_dma_segment_t seg;
2020 	int               nseg;
2021 };
2022 
2023 static void
2024 _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
2025 {
2026 	struct rxq_refill_cb_arg *cb_arg = arg;
2027 
2028 	cb_arg->error = error;
2029 	cb_arg->seg = segs[0];
2030 	cb_arg->nseg = nseg;
2031 }
2032 
2033 /**
2034  * iflib_fl_refill - refill an rxq free-buffer list
2035  * @ctx: the iflib context
2036  * @fl: the free list to refill
2037  * @count: the number of new buffers to allocate
2038  *
2039  * (Re)populate an rxq free-buffer list with up to @count new packet buffers.
2040  * The caller must assure that @count does not exceed the queue's capacity
2041  * minus one (since we always leave a descriptor unavailable).
2042  */
2043 static uint8_t
2044 iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
2045 {
2046 	struct if_rxd_update iru;
2047 	struct rxq_refill_cb_arg cb_arg;
2048 	struct mbuf *m;
2049 	caddr_t cl, *sd_cl;
2050 	struct mbuf **sd_m;
2051 	bus_dmamap_t *sd_map;
2052 	bus_addr_t bus_addr, *sd_ba;
2053 	int err, frag_idx, i, idx, n, pidx;
2054 	qidx_t credits;
2055 
2056 	MPASS(count <= fl->ifl_size - fl->ifl_credits - 1);
2057 
2058 	sd_m = fl->ifl_sds.ifsd_m;
2059 	sd_map = fl->ifl_sds.ifsd_map;
2060 	sd_cl = fl->ifl_sds.ifsd_cl;
2061 	sd_ba = fl->ifl_sds.ifsd_ba;
2062 	pidx = fl->ifl_pidx;
2063 	idx = pidx;
2064 	frag_idx = fl->ifl_fragidx;
2065 	credits = fl->ifl_credits;
2066 
2067 	i = 0;
2068 	n = count;
2069 	MPASS(n > 0);
2070 	MPASS(credits + n <= fl->ifl_size);
2071 
2072 	if (pidx < fl->ifl_cidx)
2073 		MPASS(pidx + n <= fl->ifl_cidx);
2074 	if (pidx == fl->ifl_cidx && (credits < fl->ifl_size))
2075 		MPASS(fl->ifl_gen == 0);
2076 	if (pidx > fl->ifl_cidx)
2077 		MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx);
2078 
2079 	DBG_COUNTER_INC(fl_refills);
2080 	if (n > 8)
2081 		DBG_COUNTER_INC(fl_refills_large);
2082 	iru_init(&iru, fl->ifl_rxq, fl->ifl_id);
2083 	while (n-- > 0) {
2084 		/*
2085 		 * We allocate an uninitialized mbuf + cluster, mbuf is
2086 		 * initialized after rx.
2087 		 *
2088 		 * If the cluster is still set then we know a minimum sized
2089 		 * packet was received
2090 		 */
2091 		bit_ffc_at(fl->ifl_rx_bitmap, frag_idx, fl->ifl_size,
2092 		    &frag_idx);
2093 		if (frag_idx < 0)
2094 			bit_ffc(fl->ifl_rx_bitmap, fl->ifl_size, &frag_idx);
2095 		MPASS(frag_idx >= 0);
2096 		if ((cl = sd_cl[frag_idx]) == NULL) {
2097 			cl = uma_zalloc(fl->ifl_zone, M_NOWAIT);
2098 			if (__predict_false(cl == NULL))
2099 				break;
2100 
2101 			cb_arg.error = 0;
2102 			MPASS(sd_map != NULL);
2103 			err = bus_dmamap_load(fl->ifl_buf_tag, sd_map[frag_idx],
2104 			    cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg,
2105 			    BUS_DMA_NOWAIT);
2106 			if (__predict_false(err != 0 || cb_arg.error)) {
2107 				uma_zfree(fl->ifl_zone, cl);
2108 				break;
2109 			}
2110 
2111 			sd_ba[frag_idx] = bus_addr = cb_arg.seg.ds_addr;
2112 			sd_cl[frag_idx] = cl;
2113 #if MEMORY_LOGGING
2114 			fl->ifl_cl_enqueued++;
2115 #endif
2116 		} else {
2117 			bus_addr = sd_ba[frag_idx];
2118 		}
2119 		bus_dmamap_sync(fl->ifl_buf_tag, sd_map[frag_idx],
2120 		    BUS_DMASYNC_PREREAD);
2121 
2122 		if (sd_m[frag_idx] == NULL) {
2123 			m = m_gethdr(M_NOWAIT, MT_NOINIT);
2124 			if (__predict_false(m == NULL))
2125 				break;
2126 			sd_m[frag_idx] = m;
2127 		}
2128 		bit_set(fl->ifl_rx_bitmap, frag_idx);
2129 #if MEMORY_LOGGING
2130 		fl->ifl_m_enqueued++;
2131 #endif
2132 
2133 		DBG_COUNTER_INC(rx_allocs);
2134 		fl->ifl_rxd_idxs[i] = frag_idx;
2135 		fl->ifl_bus_addrs[i] = bus_addr;
2136 		credits++;
2137 		i++;
2138 		MPASS(credits <= fl->ifl_size);
2139 		if (++idx == fl->ifl_size) {
2140 #ifdef INVARIANTS
2141 			fl->ifl_gen = 1;
2142 #endif
2143 			idx = 0;
2144 		}
2145 		if (n == 0 || i == IFLIB_MAX_RX_REFRESH) {
2146 			iru.iru_pidx = pidx;
2147 			iru.iru_count = i;
2148 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2149 			fl->ifl_pidx = idx;
2150 			fl->ifl_credits = credits;
2151 			pidx = idx;
2152 			i = 0;
2153 		}
2154 	}
2155 
2156 	if (n < count - 1) {
2157 		if (i != 0) {
2158 			iru.iru_pidx = pidx;
2159 			iru.iru_count = i;
2160 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2161 			fl->ifl_pidx = idx;
2162 			fl->ifl_credits = credits;
2163 		}
2164 		DBG_COUNTER_INC(rxd_flush);
2165 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2166 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2167 		ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id,
2168 		    fl->ifl_id, fl->ifl_pidx);
2169 		if (__predict_true(bit_test(fl->ifl_rx_bitmap, frag_idx))) {
2170 			fl->ifl_fragidx = frag_idx + 1;
2171 			if (fl->ifl_fragidx == fl->ifl_size)
2172 				fl->ifl_fragidx = 0;
2173 		} else {
2174 			fl->ifl_fragidx = frag_idx;
2175 		}
2176 	}
2177 
2178 	return (n == -1 ? 0 : IFLIB_RXEOF_EMPTY);
2179 }
2180 
2181 static inline uint8_t
2182 iflib_fl_refill_all(if_ctx_t ctx, iflib_fl_t fl)
2183 {
2184 	/*
2185 	 * We leave an unused descriptor to avoid pidx to catch up with cidx.
2186 	 * This is important as it confuses most NICs. For instance,
2187 	 * Intel NICs have (per receive ring) RDH and RDT registers, where
2188 	 * RDH points to the next receive descriptor to be used by the NIC,
2189 	 * and RDT for the next receive descriptor to be published by the
2190 	 * driver to the NIC (RDT - 1 is thus the last valid one).
2191 	 * The condition RDH == RDT means no descriptors are available to
2192 	 * the NIC, and thus it would be ambiguous if it also meant that
2193 	 * all the descriptors are available to the NIC.
2194 	 */
2195 	int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1;
2196 #ifdef INVARIANTS
2197 	int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1;
2198 #endif
2199 
2200 	MPASS(fl->ifl_credits <= fl->ifl_size);
2201 	MPASS(reclaimable == delta);
2202 
2203 	if (reclaimable > 0)
2204 		return (iflib_fl_refill(ctx, fl, reclaimable));
2205 	return (0);
2206 }
2207 
2208 uint8_t
2209 iflib_in_detach(if_ctx_t ctx)
2210 {
2211 	bool in_detach;
2212 
2213 	STATE_LOCK(ctx);
2214 	in_detach = !!(ctx->ifc_flags & IFC_IN_DETACH);
2215 	STATE_UNLOCK(ctx);
2216 	return (in_detach);
2217 }
2218 
2219 static void
2220 iflib_fl_bufs_free(iflib_fl_t fl)
2221 {
2222 	iflib_dma_info_t idi = fl->ifl_ifdi;
2223 	bus_dmamap_t sd_map;
2224 	uint32_t i;
2225 
2226 	for (i = 0; i < fl->ifl_size; i++) {
2227 		struct mbuf **sd_m = &fl->ifl_sds.ifsd_m[i];
2228 		caddr_t *sd_cl = &fl->ifl_sds.ifsd_cl[i];
2229 
2230 		if (*sd_cl != NULL) {
2231 			sd_map = fl->ifl_sds.ifsd_map[i];
2232 			bus_dmamap_sync(fl->ifl_buf_tag, sd_map,
2233 			    BUS_DMASYNC_POSTREAD);
2234 			bus_dmamap_unload(fl->ifl_buf_tag, sd_map);
2235 			uma_zfree(fl->ifl_zone, *sd_cl);
2236 			*sd_cl = NULL;
2237 			if (*sd_m != NULL) {
2238 				m_init(*sd_m, M_NOWAIT, MT_DATA, 0);
2239 				uma_zfree(zone_mbuf, *sd_m);
2240 				*sd_m = NULL;
2241 			}
2242 		} else {
2243 			MPASS(*sd_m == NULL);
2244 		}
2245 #if MEMORY_LOGGING
2246 		fl->ifl_m_dequeued++;
2247 		fl->ifl_cl_dequeued++;
2248 #endif
2249 	}
2250 #ifdef INVARIANTS
2251 	for (i = 0; i < fl->ifl_size; i++) {
2252 		MPASS(fl->ifl_sds.ifsd_cl[i] == NULL);
2253 		MPASS(fl->ifl_sds.ifsd_m[i] == NULL);
2254 	}
2255 #endif
2256 	/*
2257 	 * Reset free list values
2258 	 */
2259 	fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = fl->ifl_fragidx = 0;
2260 	bzero(idi->idi_vaddr, idi->idi_size);
2261 }
2262 
2263 /*********************************************************************
2264  *
2265  *  Initialize a free list and its buffers.
2266  *
2267  **********************************************************************/
2268 static int
2269 iflib_fl_setup(iflib_fl_t fl)
2270 {
2271 	iflib_rxq_t rxq = fl->ifl_rxq;
2272 	if_ctx_t ctx = rxq->ifr_ctx;
2273 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2274 	int qidx;
2275 
2276 	bit_nclear(fl->ifl_rx_bitmap, 0, fl->ifl_size - 1);
2277 	/*
2278 	** Free current RX buffer structs and their mbufs
2279 	*/
2280 	iflib_fl_bufs_free(fl);
2281 	/* Now replenish the mbufs */
2282 	MPASS(fl->ifl_credits == 0);
2283 	qidx = rxq->ifr_fl_offset + fl->ifl_id;
2284 	if (scctx->isc_rxd_buf_size[qidx] != 0)
2285 		fl->ifl_buf_size = scctx->isc_rxd_buf_size[qidx];
2286 	else
2287 		fl->ifl_buf_size = ctx->ifc_rx_mbuf_sz;
2288 	/*
2289 	 * ifl_buf_size may be a driver-supplied value, so pull it up
2290 	 * to the selected mbuf size.
2291 	 */
2292 	fl->ifl_buf_size = iflib_get_mbuf_size_for(fl->ifl_buf_size);
2293 	if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size)
2294 		ctx->ifc_max_fl_buf_size = fl->ifl_buf_size;
2295 	fl->ifl_cltype = m_gettype(fl->ifl_buf_size);
2296 	fl->ifl_zone = m_getzone(fl->ifl_buf_size);
2297 
2298 	/*
2299 	 * Avoid pre-allocating zillions of clusters to an idle card
2300 	 * potentially speeding up attach. In any case make sure
2301 	 * to leave a descriptor unavailable. See the comment in
2302 	 * iflib_fl_refill_all().
2303 	 */
2304 	MPASS(fl->ifl_size > 0);
2305 	(void)iflib_fl_refill(ctx, fl, min(128, fl->ifl_size - 1));
2306 	if (min(128, fl->ifl_size - 1) != fl->ifl_credits)
2307 		return (ENOBUFS);
2308 	/*
2309 	 * handle failure
2310 	 */
2311 	MPASS(rxq != NULL);
2312 	MPASS(fl->ifl_ifdi != NULL);
2313 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2314 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2315 	return (0);
2316 }
2317 
2318 /*********************************************************************
2319  *
2320  *  Free receive ring data structures
2321  *
2322  **********************************************************************/
2323 static void
2324 iflib_rx_sds_free(iflib_rxq_t rxq)
2325 {
2326 	iflib_fl_t fl;
2327 	int i, j;
2328 
2329 	if (rxq->ifr_fl != NULL) {
2330 		for (i = 0; i < rxq->ifr_nfl; i++) {
2331 			fl = &rxq->ifr_fl[i];
2332 			if (fl->ifl_buf_tag != NULL) {
2333 				if (fl->ifl_sds.ifsd_map != NULL) {
2334 					for (j = 0; j < fl->ifl_size; j++) {
2335 						bus_dmamap_sync(
2336 						    fl->ifl_buf_tag,
2337 						    fl->ifl_sds.ifsd_map[j],
2338 						    BUS_DMASYNC_POSTREAD);
2339 						bus_dmamap_unload(
2340 						    fl->ifl_buf_tag,
2341 						    fl->ifl_sds.ifsd_map[j]);
2342 						bus_dmamap_destroy(
2343 						    fl->ifl_buf_tag,
2344 						    fl->ifl_sds.ifsd_map[j]);
2345 					}
2346 				}
2347 				bus_dma_tag_destroy(fl->ifl_buf_tag);
2348 				fl->ifl_buf_tag = NULL;
2349 			}
2350 			free(fl->ifl_sds.ifsd_m, M_IFLIB);
2351 			free(fl->ifl_sds.ifsd_cl, M_IFLIB);
2352 			free(fl->ifl_sds.ifsd_ba, M_IFLIB);
2353 			free(fl->ifl_sds.ifsd_map, M_IFLIB);
2354 			free(fl->ifl_rx_bitmap, M_IFLIB);
2355 			fl->ifl_sds.ifsd_m = NULL;
2356 			fl->ifl_sds.ifsd_cl = NULL;
2357 			fl->ifl_sds.ifsd_ba = NULL;
2358 			fl->ifl_sds.ifsd_map = NULL;
2359 			fl->ifl_rx_bitmap = NULL;
2360 		}
2361 		free(rxq->ifr_fl, M_IFLIB);
2362 		rxq->ifr_fl = NULL;
2363 		free(rxq->ifr_ifdi, M_IFLIB);
2364 		rxq->ifr_ifdi = NULL;
2365 		rxq->ifr_cq_cidx = 0;
2366 	}
2367 }
2368 
2369 /*
2370  * Timer routine
2371  */
2372 static void
2373 iflib_timer(void *arg)
2374 {
2375 	iflib_txq_t txq = arg;
2376 	if_ctx_t ctx = txq->ift_ctx;
2377 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2378 	uint64_t this_tick = ticks;
2379 
2380 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
2381 		return;
2382 
2383 	/*
2384 	** Check on the state of the TX queue(s), this
2385 	** can be done without the lock because its RO
2386 	** and the HUNG state will be static if set.
2387 	*/
2388 	if (this_tick - txq->ift_last_timer_tick >= iflib_timer_default) {
2389 		txq->ift_last_timer_tick = this_tick;
2390 		IFDI_TIMER(ctx, txq->ift_id);
2391 		if ((txq->ift_qstatus == IFLIB_QUEUE_HUNG) &&
2392 		    ((txq->ift_cleaned_prev == txq->ift_cleaned) ||
2393 		     (sctx->isc_pause_frames == 0)))
2394 			goto hung;
2395 
2396 		if (txq->ift_qstatus != IFLIB_QUEUE_IDLE &&
2397 		    ifmp_ring_is_stalled(txq->ift_br)) {
2398 			KASSERT(ctx->ifc_link_state == LINK_STATE_UP,
2399 			    ("queue can't be marked as hung if interface is down"));
2400 			txq->ift_qstatus = IFLIB_QUEUE_HUNG;
2401 		}
2402 		txq->ift_cleaned_prev = txq->ift_cleaned;
2403 	}
2404 	/* handle any laggards */
2405 	if (txq->ift_db_pending)
2406 		GROUPTASK_ENQUEUE(&txq->ift_task);
2407 
2408 	sctx->isc_pause_frames = 0;
2409 	if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)
2410 		callout_reset_on(&txq->ift_timer, iflib_timer_default, iflib_timer,
2411 		    txq, txq->ift_timer.c_cpu);
2412 	return;
2413 
2414  hung:
2415 	device_printf(ctx->ifc_dev,
2416 	    "Watchdog timeout (TX: %d desc avail: %d pidx: %d) -- resetting\n",
2417 	    txq->ift_id, TXQ_AVAIL(txq), txq->ift_pidx);
2418 	STATE_LOCK(ctx);
2419 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2420 	ctx->ifc_flags |= (IFC_DO_WATCHDOG|IFC_DO_RESET);
2421 	iflib_admin_intr_deferred(ctx);
2422 	STATE_UNLOCK(ctx);
2423 }
2424 
2425 static uint16_t
2426 iflib_get_mbuf_size_for(unsigned int size)
2427 {
2428 
2429 	if (size <= MCLBYTES)
2430 		return (MCLBYTES);
2431 	else
2432 		return (MJUMPAGESIZE);
2433 }
2434 
2435 static void
2436 iflib_calc_rx_mbuf_sz(if_ctx_t ctx)
2437 {
2438 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2439 
2440 	/*
2441 	 * XXX don't set the max_frame_size to larger
2442 	 * than the hardware can handle
2443 	 */
2444 	ctx->ifc_rx_mbuf_sz =
2445 	    iflib_get_mbuf_size_for(sctx->isc_max_frame_size);
2446 }
2447 
2448 uint32_t
2449 iflib_get_rx_mbuf_sz(if_ctx_t ctx)
2450 {
2451 
2452 	return (ctx->ifc_rx_mbuf_sz);
2453 }
2454 
2455 static void
2456 iflib_init_locked(if_ctx_t ctx)
2457 {
2458 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2459 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2460 	if_t ifp = ctx->ifc_ifp;
2461 	iflib_fl_t fl;
2462 	iflib_txq_t txq;
2463 	iflib_rxq_t rxq;
2464 	int i, j, tx_ip_csum_flags, tx_ip6_csum_flags;
2465 
2466 	if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2467 	IFDI_INTR_DISABLE(ctx);
2468 
2469 	/*
2470 	 * See iflib_stop(). Useful in case iflib_init_locked() is
2471 	 * called without first calling iflib_stop().
2472 	 */
2473 	netmap_disable_all_rings(ifp);
2474 
2475 	tx_ip_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP);
2476 	tx_ip6_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | CSUM_IP6_SCTP);
2477 	/* Set hardware offload abilities */
2478 	if_clearhwassist(ifp);
2479 	if (if_getcapenable(ifp) & IFCAP_TXCSUM)
2480 		if_sethwassistbits(ifp, tx_ip_csum_flags, 0);
2481 	if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6)
2482 		if_sethwassistbits(ifp,  tx_ip6_csum_flags, 0);
2483 	if (if_getcapenable(ifp) & IFCAP_TSO4)
2484 		if_sethwassistbits(ifp, CSUM_IP_TSO, 0);
2485 	if (if_getcapenable(ifp) & IFCAP_TSO6)
2486 		if_sethwassistbits(ifp, CSUM_IP6_TSO, 0);
2487 
2488 	for (i = 0, txq = ctx->ifc_txqs; i < sctx->isc_ntxqsets; i++, txq++) {
2489 		CALLOUT_LOCK(txq);
2490 		callout_stop(&txq->ift_timer);
2491 #ifdef DEV_NETMAP
2492 		callout_stop(&txq->ift_netmap_timer);
2493 #endif /* DEV_NETMAP */
2494 		CALLOUT_UNLOCK(txq);
2495 		(void)iflib_netmap_txq_init(ctx, txq);
2496 	}
2497 
2498 	/*
2499 	 * Calculate a suitable Rx mbuf size prior to calling IFDI_INIT, so
2500 	 * that drivers can use the value when setting up the hardware receive
2501 	 * buffers.
2502 	 */
2503 	iflib_calc_rx_mbuf_sz(ctx);
2504 
2505 #ifdef INVARIANTS
2506 	i = if_getdrvflags(ifp);
2507 #endif
2508 	IFDI_INIT(ctx);
2509 	MPASS(if_getdrvflags(ifp) == i);
2510 	for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
2511 		if (iflib_netmap_rxq_init(ctx, rxq) > 0) {
2512 			/* This rxq is in netmap mode. Skip normal init. */
2513 			continue;
2514 		}
2515 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
2516 			if (iflib_fl_setup(fl)) {
2517 				device_printf(ctx->ifc_dev,
2518 				    "setting up free list %d failed - "
2519 				    "check cluster settings\n", j);
2520 				goto done;
2521 			}
2522 		}
2523 	}
2524 done:
2525 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
2526 	IFDI_INTR_ENABLE(ctx);
2527 	txq = ctx->ifc_txqs;
2528 	for (i = 0; i < sctx->isc_ntxqsets; i++, txq++)
2529 		callout_reset_on(&txq->ift_timer, iflib_timer_default, iflib_timer, txq,
2530 			txq->ift_timer.c_cpu);
2531 
2532         /* Re-enable txsync/rxsync. */
2533 	netmap_enable_all_rings(ifp);
2534 }
2535 
2536 static int
2537 iflib_media_change(if_t ifp)
2538 {
2539 	if_ctx_t ctx = if_getsoftc(ifp);
2540 	int err;
2541 
2542 	CTX_LOCK(ctx);
2543 	if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0)
2544 		iflib_if_init_locked(ctx);
2545 	CTX_UNLOCK(ctx);
2546 	return (err);
2547 }
2548 
2549 static void
2550 iflib_media_status(if_t ifp, struct ifmediareq *ifmr)
2551 {
2552 	if_ctx_t ctx = if_getsoftc(ifp);
2553 
2554 	CTX_LOCK(ctx);
2555 	IFDI_UPDATE_ADMIN_STATUS(ctx);
2556 	IFDI_MEDIA_STATUS(ctx, ifmr);
2557 	CTX_UNLOCK(ctx);
2558 }
2559 
2560 void
2561 iflib_stop(if_ctx_t ctx)
2562 {
2563 	iflib_txq_t txq = ctx->ifc_txqs;
2564 	iflib_rxq_t rxq = ctx->ifc_rxqs;
2565 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2566 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2567 	iflib_dma_info_t di;
2568 	iflib_fl_t fl;
2569 	int i, j;
2570 
2571 	/* Tell the stack that the interface is no longer active */
2572 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2573 
2574 	IFDI_INTR_DISABLE(ctx);
2575 	DELAY(1000);
2576 	IFDI_STOP(ctx);
2577 	DELAY(1000);
2578 
2579 	/*
2580 	 * Stop any pending txsync/rxsync and prevent new ones
2581 	 * form starting. Processes blocked in poll() will get
2582 	 * POLLERR.
2583 	 */
2584 	netmap_disable_all_rings(ctx->ifc_ifp);
2585 
2586 	iflib_debug_reset();
2587 	/* Wait for current tx queue users to exit to disarm watchdog timer. */
2588 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) {
2589 		/* make sure all transmitters have completed before proceeding XXX */
2590 
2591 		CALLOUT_LOCK(txq);
2592 		callout_stop(&txq->ift_timer);
2593 #ifdef DEV_NETMAP
2594 		callout_stop(&txq->ift_netmap_timer);
2595 #endif /* DEV_NETMAP */
2596 		CALLOUT_UNLOCK(txq);
2597 
2598 		/* clean any enqueued buffers */
2599 		iflib_ifmp_purge(txq);
2600 		/* Free any existing tx buffers. */
2601 		for (j = 0; j < txq->ift_size; j++) {
2602 			iflib_txsd_free(ctx, txq, j);
2603 		}
2604 		txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0;
2605 		txq->ift_in_use = txq->ift_gen = txq->ift_cidx = txq->ift_pidx = txq->ift_no_desc_avail = 0;
2606 		txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0;
2607 		txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0;
2608 		txq->ift_pullups = 0;
2609 		ifmp_ring_reset_stats(txq->ift_br);
2610 		for (j = 0, di = txq->ift_ifdi; j < sctx->isc_ntxqs; j++, di++)
2611 			bzero((void *)di->idi_vaddr, di->idi_size);
2612 	}
2613 	for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) {
2614 		/* make sure all transmitters have completed before proceeding XXX */
2615 
2616 		rxq->ifr_cq_cidx = 0;
2617 		for (j = 0, di = rxq->ifr_ifdi; j < sctx->isc_nrxqs; j++, di++)
2618 			bzero((void *)di->idi_vaddr, di->idi_size);
2619 		/* also resets the free lists pidx/cidx */
2620 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
2621 			iflib_fl_bufs_free(fl);
2622 	}
2623 }
2624 
2625 static inline caddr_t
2626 calc_next_rxd(iflib_fl_t fl, int cidx)
2627 {
2628 	qidx_t size;
2629 	int nrxd;
2630 	caddr_t start, end, cur, next;
2631 
2632 	nrxd = fl->ifl_size;
2633 	size = fl->ifl_rxd_size;
2634 	start = fl->ifl_ifdi->idi_vaddr;
2635 
2636 	if (__predict_false(size == 0))
2637 		return (start);
2638 	cur = start + size*cidx;
2639 	end = start + size*nrxd;
2640 	next = CACHE_PTR_NEXT(cur);
2641 	return (next < end ? next : start);
2642 }
2643 
2644 static inline void
2645 prefetch_pkts(iflib_fl_t fl, int cidx)
2646 {
2647 	int nextptr;
2648 	int nrxd = fl->ifl_size;
2649 	caddr_t next_rxd;
2650 
2651 	nextptr = (cidx + CACHE_PTR_INCREMENT) & (nrxd-1);
2652 	prefetch(&fl->ifl_sds.ifsd_m[nextptr]);
2653 	prefetch(&fl->ifl_sds.ifsd_cl[nextptr]);
2654 	next_rxd = calc_next_rxd(fl, cidx);
2655 	prefetch(next_rxd);
2656 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 1) & (nrxd-1)]);
2657 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 2) & (nrxd-1)]);
2658 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 3) & (nrxd-1)]);
2659 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 4) & (nrxd-1)]);
2660 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 1) & (nrxd-1)]);
2661 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 2) & (nrxd-1)]);
2662 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 3) & (nrxd-1)]);
2663 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 4) & (nrxd-1)]);
2664 }
2665 
2666 static struct mbuf *
2667 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, bool unload, if_rxsd_t sd,
2668     int *pf_rv, if_rxd_info_t ri)
2669 {
2670 	bus_dmamap_t map;
2671 	iflib_fl_t fl;
2672 	caddr_t payload;
2673 	struct mbuf *m;
2674 	int flid, cidx, len, next;
2675 
2676 	map = NULL;
2677 	flid = irf->irf_flid;
2678 	cidx = irf->irf_idx;
2679 	fl = &rxq->ifr_fl[flid];
2680 	sd->ifsd_fl = fl;
2681 	m = fl->ifl_sds.ifsd_m[cidx];
2682 	sd->ifsd_cl = &fl->ifl_sds.ifsd_cl[cidx];
2683 	fl->ifl_credits--;
2684 #if MEMORY_LOGGING
2685 	fl->ifl_m_dequeued++;
2686 #endif
2687 	if (rxq->ifr_ctx->ifc_flags & IFC_PREFETCH)
2688 		prefetch_pkts(fl, cidx);
2689 	next = (cidx + CACHE_PTR_INCREMENT) & (fl->ifl_size-1);
2690 	prefetch(&fl->ifl_sds.ifsd_map[next]);
2691 	map = fl->ifl_sds.ifsd_map[cidx];
2692 
2693 	bus_dmamap_sync(fl->ifl_buf_tag, map, BUS_DMASYNC_POSTREAD);
2694 
2695 	if (rxq->pfil != NULL && PFIL_HOOKED_IN(rxq->pfil) && pf_rv != NULL &&
2696 	    irf->irf_len != 0) {
2697 		payload  = *sd->ifsd_cl;
2698 		payload +=  ri->iri_pad;
2699 		len = ri->iri_len - ri->iri_pad;
2700 		*pf_rv = pfil_run_hooks(rxq->pfil, payload, ri->iri_ifp,
2701 		    len | PFIL_MEMPTR | PFIL_IN, NULL);
2702 		switch (*pf_rv) {
2703 		case PFIL_DROPPED:
2704 		case PFIL_CONSUMED:
2705 			/*
2706 			 * The filter ate it.  Everything is recycled.
2707 			 */
2708 			m = NULL;
2709 			unload = 0;
2710 			break;
2711 		case PFIL_REALLOCED:
2712 			/*
2713 			 * The filter copied it.  Everything is recycled.
2714 			 */
2715 			m = pfil_mem2mbuf(payload);
2716 			unload = 0;
2717 			break;
2718 		case PFIL_PASS:
2719 			/*
2720 			 * Filter said it was OK, so receive like
2721 			 * normal
2722 			 */
2723 			fl->ifl_sds.ifsd_m[cidx] = NULL;
2724 			break;
2725 		default:
2726 			MPASS(0);
2727 		}
2728 	} else {
2729 		fl->ifl_sds.ifsd_m[cidx] = NULL;
2730 		if (pf_rv != NULL)
2731 			*pf_rv = PFIL_PASS;
2732 	}
2733 
2734 	if (unload && irf->irf_len != 0)
2735 		bus_dmamap_unload(fl->ifl_buf_tag, map);
2736 	fl->ifl_cidx = (fl->ifl_cidx + 1) & (fl->ifl_size-1);
2737 	if (__predict_false(fl->ifl_cidx == 0))
2738 		fl->ifl_gen = 0;
2739 	bit_clear(fl->ifl_rx_bitmap, cidx);
2740 	return (m);
2741 }
2742 
2743 static struct mbuf *
2744 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri, if_rxsd_t sd, int *pf_rv)
2745 {
2746 	struct mbuf *m, *mh, *mt;
2747 	caddr_t cl;
2748 	int  *pf_rv_ptr, flags, i, padlen;
2749 	bool consumed;
2750 
2751 	i = 0;
2752 	mh = NULL;
2753 	consumed = false;
2754 	*pf_rv = PFIL_PASS;
2755 	pf_rv_ptr = pf_rv;
2756 	do {
2757 		m = rxd_frag_to_sd(rxq, &ri->iri_frags[i], !consumed, sd,
2758 		    pf_rv_ptr, ri);
2759 
2760 		MPASS(*sd->ifsd_cl != NULL);
2761 
2762 		/*
2763 		 * Exclude zero-length frags & frags from
2764 		 * packets the filter has consumed or dropped
2765 		 */
2766 		if (ri->iri_frags[i].irf_len == 0 || consumed ||
2767 		    *pf_rv == PFIL_CONSUMED || *pf_rv == PFIL_DROPPED) {
2768 			if (mh == NULL) {
2769 				/* everything saved here */
2770 				consumed = true;
2771 				pf_rv_ptr = NULL;
2772 				continue;
2773 			}
2774 			/* XXX we can save the cluster here, but not the mbuf */
2775 			m_init(m, M_NOWAIT, MT_DATA, 0);
2776 			m_free(m);
2777 			continue;
2778 		}
2779 		if (mh == NULL) {
2780 			flags = M_PKTHDR|M_EXT;
2781 			mh = mt = m;
2782 			padlen = ri->iri_pad;
2783 		} else {
2784 			flags = M_EXT;
2785 			mt->m_next = m;
2786 			mt = m;
2787 			/* assuming padding is only on the first fragment */
2788 			padlen = 0;
2789 		}
2790 		cl = *sd->ifsd_cl;
2791 		*sd->ifsd_cl = NULL;
2792 
2793 		/* Can these two be made one ? */
2794 		m_init(m, M_NOWAIT, MT_DATA, flags);
2795 		m_cljset(m, cl, sd->ifsd_fl->ifl_cltype);
2796 		/*
2797 		 * These must follow m_init and m_cljset
2798 		 */
2799 		m->m_data += padlen;
2800 		ri->iri_len -= padlen;
2801 		m->m_len = ri->iri_frags[i].irf_len;
2802 	} while (++i < ri->iri_nfrags);
2803 
2804 	return (mh);
2805 }
2806 
2807 /*
2808  * Process one software descriptor
2809  */
2810 static struct mbuf *
2811 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri)
2812 {
2813 	struct if_rxsd sd;
2814 	struct mbuf *m;
2815 	int pf_rv;
2816 
2817 	/* should I merge this back in now that the two paths are basically duplicated? */
2818 	if (ri->iri_nfrags == 1 &&
2819 	    ri->iri_frags[0].irf_len != 0 &&
2820 	    ri->iri_frags[0].irf_len <= MIN(IFLIB_RX_COPY_THRESH, MHLEN)) {
2821 		m = rxd_frag_to_sd(rxq, &ri->iri_frags[0], false, &sd,
2822 		    &pf_rv, ri);
2823 		if (pf_rv != PFIL_PASS && pf_rv != PFIL_REALLOCED)
2824 			return (m);
2825 		if (pf_rv == PFIL_PASS) {
2826 			m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR);
2827 #ifndef __NO_STRICT_ALIGNMENT
2828 			if (!IP_ALIGNED(m))
2829 				m->m_data += 2;
2830 #endif
2831 			memcpy(m->m_data, *sd.ifsd_cl, ri->iri_len);
2832 			m->m_len = ri->iri_frags[0].irf_len;
2833 		}
2834 	} else {
2835 		m = assemble_segments(rxq, ri, &sd, &pf_rv);
2836 		if (m == NULL)
2837 			return (NULL);
2838 		if (pf_rv != PFIL_PASS && pf_rv != PFIL_REALLOCED)
2839 			return (m);
2840 	}
2841 	m->m_pkthdr.len = ri->iri_len;
2842 	m->m_pkthdr.rcvif = ri->iri_ifp;
2843 	m->m_flags |= ri->iri_flags;
2844 	m->m_pkthdr.ether_vtag = ri->iri_vtag;
2845 	m->m_pkthdr.flowid = ri->iri_flowid;
2846 	M_HASHTYPE_SET(m, ri->iri_rsstype);
2847 	m->m_pkthdr.csum_flags = ri->iri_csum_flags;
2848 	m->m_pkthdr.csum_data = ri->iri_csum_data;
2849 	return (m);
2850 }
2851 
2852 #if defined(INET6) || defined(INET)
2853 static void
2854 iflib_get_ip_forwarding(struct lro_ctrl *lc, bool *v4, bool *v6)
2855 {
2856 	CURVNET_SET(lc->ifp->if_vnet);
2857 #if defined(INET6)
2858 	*v6 = V_ip6_forwarding;
2859 #endif
2860 #if defined(INET)
2861 	*v4 = V_ipforwarding;
2862 #endif
2863 	CURVNET_RESTORE();
2864 }
2865 
2866 /*
2867  * Returns true if it's possible this packet could be LROed.
2868  * if it returns false, it is guaranteed that tcp_lro_rx()
2869  * would not return zero.
2870  */
2871 static bool
2872 iflib_check_lro_possible(struct mbuf *m, bool v4_forwarding, bool v6_forwarding)
2873 {
2874 	struct ether_header *eh;
2875 
2876 	eh = mtod(m, struct ether_header *);
2877 	switch (eh->ether_type) {
2878 #if defined(INET6)
2879 		case htons(ETHERTYPE_IPV6):
2880 			return (!v6_forwarding);
2881 #endif
2882 #if defined (INET)
2883 		case htons(ETHERTYPE_IP):
2884 			return (!v4_forwarding);
2885 #endif
2886 	}
2887 
2888 	return false;
2889 }
2890 #else
2891 static void
2892 iflib_get_ip_forwarding(struct lro_ctrl *lc __unused, bool *v4 __unused, bool *v6 __unused)
2893 {
2894 }
2895 #endif
2896 
2897 static void
2898 _task_fn_rx_watchdog(void *context)
2899 {
2900 	iflib_rxq_t rxq = context;
2901 
2902 	GROUPTASK_ENQUEUE(&rxq->ifr_task);
2903 }
2904 
2905 static uint8_t
2906 iflib_rxeof(iflib_rxq_t rxq, qidx_t budget)
2907 {
2908 	if_t ifp;
2909 	if_ctx_t ctx = rxq->ifr_ctx;
2910 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2911 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2912 	int avail, i;
2913 	qidx_t *cidxp;
2914 	struct if_rxd_info ri;
2915 	int err, budget_left, rx_bytes, rx_pkts;
2916 	iflib_fl_t fl;
2917 	int lro_enabled;
2918 	bool v4_forwarding, v6_forwarding, lro_possible;
2919 	uint8_t retval = 0;
2920 
2921 	/*
2922 	 * XXX early demux data packets so that if_input processing only handles
2923 	 * acks in interrupt context
2924 	 */
2925 	struct mbuf *m, *mh, *mt, *mf;
2926 
2927 	NET_EPOCH_ASSERT();
2928 
2929 	lro_possible = v4_forwarding = v6_forwarding = false;
2930 	ifp = ctx->ifc_ifp;
2931 	mh = mt = NULL;
2932 	MPASS(budget > 0);
2933 	rx_pkts	= rx_bytes = 0;
2934 	if (sctx->isc_flags & IFLIB_HAS_RXCQ)
2935 		cidxp = &rxq->ifr_cq_cidx;
2936 	else
2937 		cidxp = &rxq->ifr_fl[0].ifl_cidx;
2938 	if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) {
2939 		for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2940 			retval |= iflib_fl_refill_all(ctx, fl);
2941 		DBG_COUNTER_INC(rx_unavail);
2942 		return (retval);
2943 	}
2944 
2945 	/* pfil needs the vnet to be set */
2946 	CURVNET_SET_QUIET(ifp->if_vnet);
2947 	for (budget_left = budget; budget_left > 0 && avail > 0;) {
2948 		if (__predict_false(!CTX_ACTIVE(ctx))) {
2949 			DBG_COUNTER_INC(rx_ctx_inactive);
2950 			break;
2951 		}
2952 		/*
2953 		 * Reset client set fields to their default values
2954 		 */
2955 		rxd_info_zero(&ri);
2956 		ri.iri_qsidx = rxq->ifr_id;
2957 		ri.iri_cidx = *cidxp;
2958 		ri.iri_ifp = ifp;
2959 		ri.iri_frags = rxq->ifr_frags;
2960 		err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
2961 
2962 		if (err)
2963 			goto err;
2964 		rx_pkts += 1;
2965 		rx_bytes += ri.iri_len;
2966 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
2967 			*cidxp = ri.iri_cidx;
2968 			/* Update our consumer index */
2969 			/* XXX NB: shurd - check if this is still safe */
2970 			while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0])
2971 				rxq->ifr_cq_cidx -= scctx->isc_nrxd[0];
2972 			/* was this only a completion queue message? */
2973 			if (__predict_false(ri.iri_nfrags == 0))
2974 				continue;
2975 		}
2976 		MPASS(ri.iri_nfrags != 0);
2977 		MPASS(ri.iri_len != 0);
2978 
2979 		/* will advance the cidx on the corresponding free lists */
2980 		m = iflib_rxd_pkt_get(rxq, &ri);
2981 		avail--;
2982 		budget_left--;
2983 		if (avail == 0 && budget_left)
2984 			avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left);
2985 
2986 		if (__predict_false(m == NULL))
2987 			continue;
2988 
2989 		/* imm_pkt: -- cxgb */
2990 		if (mh == NULL)
2991 			mh = mt = m;
2992 		else {
2993 			mt->m_nextpkt = m;
2994 			mt = m;
2995 		}
2996 	}
2997 	CURVNET_RESTORE();
2998 	/* make sure that we can refill faster than drain */
2999 	for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
3000 		retval |= iflib_fl_refill_all(ctx, fl);
3001 
3002 	lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO);
3003 	if (lro_enabled)
3004 		iflib_get_ip_forwarding(&rxq->ifr_lc, &v4_forwarding, &v6_forwarding);
3005 	mt = mf = NULL;
3006 	while (mh != NULL) {
3007 		m = mh;
3008 		mh = mh->m_nextpkt;
3009 		m->m_nextpkt = NULL;
3010 #ifndef __NO_STRICT_ALIGNMENT
3011 		if (!IP_ALIGNED(m) && (m = iflib_fixup_rx(m)) == NULL)
3012 			continue;
3013 #endif
3014 #if defined(INET6) || defined(INET)
3015 		if (lro_enabled) {
3016 			if (!lro_possible) {
3017 				lro_possible = iflib_check_lro_possible(m, v4_forwarding, v6_forwarding);
3018 				if (lro_possible && mf != NULL) {
3019 					ifp->if_input(ifp, mf);
3020 					DBG_COUNTER_INC(rx_if_input);
3021 					mt = mf = NULL;
3022 				}
3023 			}
3024 			if ((m->m_pkthdr.csum_flags & (CSUM_L4_CALC|CSUM_L4_VALID)) ==
3025 			    (CSUM_L4_CALC|CSUM_L4_VALID)) {
3026 				if (lro_possible && tcp_lro_rx(&rxq->ifr_lc, m, 0) == 0)
3027 					continue;
3028 			}
3029 		}
3030 #endif
3031 		if (lro_possible) {
3032 			ifp->if_input(ifp, m);
3033 			DBG_COUNTER_INC(rx_if_input);
3034 			continue;
3035 		}
3036 
3037 		if (mf == NULL)
3038 			mf = m;
3039 		if (mt != NULL)
3040 			mt->m_nextpkt = m;
3041 		mt = m;
3042 	}
3043 	if (mf != NULL) {
3044 		ifp->if_input(ifp, mf);
3045 		DBG_COUNTER_INC(rx_if_input);
3046 	}
3047 
3048 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
3049 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
3050 
3051 	/*
3052 	 * Flush any outstanding LRO work
3053 	 */
3054 #if defined(INET6) || defined(INET)
3055 	tcp_lro_flush_all(&rxq->ifr_lc);
3056 #endif
3057 	if (avail != 0 || iflib_rxd_avail(ctx, rxq, *cidxp, 1) != 0)
3058 		retval |= IFLIB_RXEOF_MORE;
3059 	return (retval);
3060 err:
3061 	STATE_LOCK(ctx);
3062 	ctx->ifc_flags |= IFC_DO_RESET;
3063 	iflib_admin_intr_deferred(ctx);
3064 	STATE_UNLOCK(ctx);
3065 	return (0);
3066 }
3067 
3068 #define TXD_NOTIFY_COUNT(txq) (((txq)->ift_size / (txq)->ift_update_freq)-1)
3069 static inline qidx_t
3070 txq_max_db_deferred(iflib_txq_t txq, qidx_t in_use)
3071 {
3072 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
3073 	qidx_t minthresh = txq->ift_size / 8;
3074 	if (in_use > 4*minthresh)
3075 		return (notify_count);
3076 	if (in_use > 2*minthresh)
3077 		return (notify_count >> 1);
3078 	if (in_use > minthresh)
3079 		return (notify_count >> 3);
3080 	return (0);
3081 }
3082 
3083 static inline qidx_t
3084 txq_max_rs_deferred(iflib_txq_t txq)
3085 {
3086 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
3087 	qidx_t minthresh = txq->ift_size / 8;
3088 	if (txq->ift_in_use > 4*minthresh)
3089 		return (notify_count);
3090 	if (txq->ift_in_use > 2*minthresh)
3091 		return (notify_count >> 1);
3092 	if (txq->ift_in_use > minthresh)
3093 		return (notify_count >> 2);
3094 	return (2);
3095 }
3096 
3097 #define M_CSUM_FLAGS(m) ((m)->m_pkthdr.csum_flags)
3098 #define M_HAS_VLANTAG(m) (m->m_flags & M_VLANTAG)
3099 
3100 #define TXQ_MAX_DB_DEFERRED(txq, in_use) txq_max_db_deferred((txq), (in_use))
3101 #define TXQ_MAX_RS_DEFERRED(txq) txq_max_rs_deferred(txq)
3102 #define TXQ_MAX_DB_CONSUMED(size) (size >> 4)
3103 
3104 /* forward compatibility for cxgb */
3105 #define FIRST_QSET(ctx) 0
3106 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets)
3107 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets)
3108 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx))
3109 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments))
3110 
3111 /* XXX we should be setting this to something other than zero */
3112 #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh)
3113 #define	MAX_TX_DESC(ctx) MAX((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max, \
3114     (ctx)->ifc_softc_ctx.isc_tx_nsegments)
3115 
3116 static inline bool
3117 iflib_txd_db_check(iflib_txq_t txq, int ring)
3118 {
3119 	if_ctx_t ctx = txq->ift_ctx;
3120 	qidx_t dbval, max;
3121 
3122 	max = TXQ_MAX_DB_DEFERRED(txq, txq->ift_in_use);
3123 
3124 	/* force || threshold exceeded || at the edge of the ring */
3125 	if (ring || (txq->ift_db_pending >= max) || (TXQ_AVAIL(txq) <= MAX_TX_DESC(ctx) + 2)) {
3126 
3127 		/*
3128 		 * 'npending' is used if the card's doorbell is in terms of the number of descriptors
3129 		 * pending flush (BRCM). 'pidx' is used in cases where the card's doorbeel uses the
3130 		 * producer index explicitly (INTC).
3131 		 */
3132 		dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx;
3133 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3134 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
3135 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval);
3136 
3137 		/*
3138 		 * Absent bugs there are zero packets pending so reset pending counts to zero.
3139 		 */
3140 		txq->ift_db_pending = txq->ift_npending = 0;
3141 		return (true);
3142 	}
3143 	return (false);
3144 }
3145 
3146 #ifdef PKT_DEBUG
3147 static void
3148 print_pkt(if_pkt_info_t pi)
3149 {
3150 	printf("pi len:  %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n",
3151 	       pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx);
3152 	printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n",
3153 	       pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag);
3154 	printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n",
3155 	       pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto);
3156 }
3157 #endif
3158 
3159 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO)
3160 #define IS_TX_OFFLOAD4(pi) ((pi)->ipi_csum_flags & (CSUM_IP_TCP | CSUM_IP_TSO))
3161 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO)
3162 #define IS_TX_OFFLOAD6(pi) ((pi)->ipi_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_TSO))
3163 
3164 static int
3165 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp)
3166 {
3167 	if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx;
3168 	struct ether_vlan_header *eh;
3169 	struct mbuf *m;
3170 
3171 	m = *mp;
3172 	if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) &&
3173 	    M_WRITABLE(m) == 0) {
3174 		if ((m = m_dup(m, M_NOWAIT)) == NULL) {
3175 			return (ENOMEM);
3176 		} else {
3177 			m_freem(*mp);
3178 			DBG_COUNTER_INC(tx_frees);
3179 			*mp = m;
3180 		}
3181 	}
3182 
3183 	/*
3184 	 * Determine where frame payload starts.
3185 	 * Jump over vlan headers if already present,
3186 	 * helpful for QinQ too.
3187 	 */
3188 	if (__predict_false(m->m_len < sizeof(*eh))) {
3189 		txq->ift_pullups++;
3190 		if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL))
3191 			return (ENOMEM);
3192 	}
3193 	eh = mtod(m, struct ether_vlan_header *);
3194 	if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
3195 		pi->ipi_etype = ntohs(eh->evl_proto);
3196 		pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
3197 	} else {
3198 		pi->ipi_etype = ntohs(eh->evl_encap_proto);
3199 		pi->ipi_ehdrlen = ETHER_HDR_LEN;
3200 	}
3201 
3202 	switch (pi->ipi_etype) {
3203 #ifdef INET
3204 	case ETHERTYPE_IP:
3205 	{
3206 		struct mbuf *n;
3207 		struct ip *ip = NULL;
3208 		struct tcphdr *th = NULL;
3209 		int minthlen;
3210 
3211 		minthlen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip) + sizeof(*th));
3212 		if (__predict_false(m->m_len < minthlen)) {
3213 			/*
3214 			 * if this code bloat is causing too much of a hit
3215 			 * move it to a separate function and mark it noinline
3216 			 */
3217 			if (m->m_len == pi->ipi_ehdrlen) {
3218 				n = m->m_next;
3219 				MPASS(n);
3220 				if (n->m_len >= sizeof(*ip))  {
3221 					ip = (struct ip *)n->m_data;
3222 					if (n->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3223 						th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3224 				} else {
3225 					txq->ift_pullups++;
3226 					if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
3227 						return (ENOMEM);
3228 					ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3229 				}
3230 			} else {
3231 				txq->ift_pullups++;
3232 				if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
3233 					return (ENOMEM);
3234 				ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3235 				if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3236 					th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3237 			}
3238 		} else {
3239 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3240 			if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3241 				th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3242 		}
3243 		pi->ipi_ip_hlen = ip->ip_hl << 2;
3244 		pi->ipi_ipproto = ip->ip_p;
3245 		pi->ipi_flags |= IPI_TX_IPV4;
3246 
3247 		/* TCP checksum offload may require TCP header length */
3248 		if (IS_TX_OFFLOAD4(pi)) {
3249 			if (__predict_true(pi->ipi_ipproto == IPPROTO_TCP)) {
3250 				if (__predict_false(th == NULL)) {
3251 					txq->ift_pullups++;
3252 					if (__predict_false((m = m_pullup(m, (ip->ip_hl << 2) + sizeof(*th))) == NULL))
3253 						return (ENOMEM);
3254 					th = (struct tcphdr *)((caddr_t)ip + pi->ipi_ip_hlen);
3255 				}
3256 				pi->ipi_tcp_hflags = th->th_flags;
3257 				pi->ipi_tcp_hlen = th->th_off << 2;
3258 				pi->ipi_tcp_seq = th->th_seq;
3259 			}
3260 			if (IS_TSO4(pi)) {
3261 				if (__predict_false(ip->ip_p != IPPROTO_TCP))
3262 					return (ENXIO);
3263 				/*
3264 				 * TSO always requires hardware checksum offload.
3265 				 */
3266 				pi->ipi_csum_flags |= (CSUM_IP_TCP | CSUM_IP);
3267 				th->th_sum = in_pseudo(ip->ip_src.s_addr,
3268 						       ip->ip_dst.s_addr, htons(IPPROTO_TCP));
3269 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3270 				if (sctx->isc_flags & IFLIB_TSO_INIT_IP) {
3271 					ip->ip_sum = 0;
3272 					ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz);
3273 				}
3274 			}
3275 		}
3276 		if ((sctx->isc_flags & IFLIB_NEED_ZERO_CSUM) && (pi->ipi_csum_flags & CSUM_IP))
3277                        ip->ip_sum = 0;
3278 
3279 		break;
3280 	}
3281 #endif
3282 #ifdef INET6
3283 	case ETHERTYPE_IPV6:
3284 	{
3285 		struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
3286 		struct tcphdr *th;
3287 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
3288 
3289 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
3290 			txq->ift_pullups++;
3291 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
3292 				return (ENOMEM);
3293 		}
3294 		th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
3295 
3296 		/* XXX-BZ this will go badly in case of ext hdrs. */
3297 		pi->ipi_ipproto = ip6->ip6_nxt;
3298 		pi->ipi_flags |= IPI_TX_IPV6;
3299 
3300 		/* TCP checksum offload may require TCP header length */
3301 		if (IS_TX_OFFLOAD6(pi)) {
3302 			if (pi->ipi_ipproto == IPPROTO_TCP) {
3303 				if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) {
3304 					txq->ift_pullups++;
3305 					if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL))
3306 						return (ENOMEM);
3307 				}
3308 				pi->ipi_tcp_hflags = th->th_flags;
3309 				pi->ipi_tcp_hlen = th->th_off << 2;
3310 				pi->ipi_tcp_seq = th->th_seq;
3311 			}
3312 			if (IS_TSO6(pi)) {
3313 				if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP))
3314 					return (ENXIO);
3315 				/*
3316 				 * TSO always requires hardware checksum offload.
3317 				 */
3318 				pi->ipi_csum_flags |= CSUM_IP6_TCP;
3319 				th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0);
3320 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3321 			}
3322 		}
3323 		break;
3324 	}
3325 #endif
3326 	default:
3327 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
3328 		pi->ipi_ip_hlen = 0;
3329 		break;
3330 	}
3331 	*mp = m;
3332 
3333 	return (0);
3334 }
3335 
3336 /*
3337  * If dodgy hardware rejects the scatter gather chain we've handed it
3338  * we'll need to remove the mbuf chain from ifsg_m[] before we can add the
3339  * m_defrag'd mbufs
3340  */
3341 static __noinline struct mbuf *
3342 iflib_remove_mbuf(iflib_txq_t txq)
3343 {
3344 	int ntxd, pidx;
3345 	struct mbuf *m, **ifsd_m;
3346 
3347 	ifsd_m = txq->ift_sds.ifsd_m;
3348 	ntxd = txq->ift_size;
3349 	pidx = txq->ift_pidx & (ntxd - 1);
3350 	ifsd_m = txq->ift_sds.ifsd_m;
3351 	m = ifsd_m[pidx];
3352 	ifsd_m[pidx] = NULL;
3353 	bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[pidx]);
3354 	if (txq->ift_sds.ifsd_tso_map != NULL)
3355 		bus_dmamap_unload(txq->ift_tso_buf_tag,
3356 		    txq->ift_sds.ifsd_tso_map[pidx]);
3357 #if MEMORY_LOGGING
3358 	txq->ift_dequeued++;
3359 #endif
3360 	return (m);
3361 }
3362 
3363 static inline caddr_t
3364 calc_next_txd(iflib_txq_t txq, int cidx, uint8_t qid)
3365 {
3366 	qidx_t size;
3367 	int ntxd;
3368 	caddr_t start, end, cur, next;
3369 
3370 	ntxd = txq->ift_size;
3371 	size = txq->ift_txd_size[qid];
3372 	start = txq->ift_ifdi[qid].idi_vaddr;
3373 
3374 	if (__predict_false(size == 0))
3375 		return (start);
3376 	cur = start + size*cidx;
3377 	end = start + size*ntxd;
3378 	next = CACHE_PTR_NEXT(cur);
3379 	return (next < end ? next : start);
3380 }
3381 
3382 /*
3383  * Pad an mbuf to ensure a minimum ethernet frame size.
3384  * min_frame_size is the frame size (less CRC) to pad the mbuf to
3385  */
3386 static __noinline int
3387 iflib_ether_pad(device_t dev, struct mbuf **m_head, uint16_t min_frame_size)
3388 {
3389 	/*
3390 	 * 18 is enough bytes to pad an ARP packet to 46 bytes, and
3391 	 * and ARP message is the smallest common payload I can think of
3392 	 */
3393 	static char pad[18];	/* just zeros */
3394 	int n;
3395 	struct mbuf *new_head;
3396 
3397 	if (!M_WRITABLE(*m_head)) {
3398 		new_head = m_dup(*m_head, M_NOWAIT);
3399 		if (new_head == NULL) {
3400 			m_freem(*m_head);
3401 			device_printf(dev, "cannot pad short frame, m_dup() failed");
3402 			DBG_COUNTER_INC(encap_pad_mbuf_fail);
3403 			DBG_COUNTER_INC(tx_frees);
3404 			return ENOMEM;
3405 		}
3406 		m_freem(*m_head);
3407 		*m_head = new_head;
3408 	}
3409 
3410 	for (n = min_frame_size - (*m_head)->m_pkthdr.len;
3411 	     n > 0; n -= sizeof(pad))
3412 		if (!m_append(*m_head, min(n, sizeof(pad)), pad))
3413 			break;
3414 
3415 	if (n > 0) {
3416 		m_freem(*m_head);
3417 		device_printf(dev, "cannot pad short frame\n");
3418 		DBG_COUNTER_INC(encap_pad_mbuf_fail);
3419 		DBG_COUNTER_INC(tx_frees);
3420 		return (ENOBUFS);
3421 	}
3422 
3423 	return 0;
3424 }
3425 
3426 static int
3427 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp)
3428 {
3429 	if_ctx_t		ctx;
3430 	if_shared_ctx_t		sctx;
3431 	if_softc_ctx_t		scctx;
3432 	bus_dma_tag_t		buf_tag;
3433 	bus_dma_segment_t	*segs;
3434 	struct mbuf		*m_head, **ifsd_m;
3435 	void			*next_txd;
3436 	bus_dmamap_t		map;
3437 	struct if_pkt_info	pi;
3438 	int remap = 0;
3439 	int err, nsegs, ndesc, max_segs, pidx, cidx, next, ntxd;
3440 
3441 	ctx = txq->ift_ctx;
3442 	sctx = ctx->ifc_sctx;
3443 	scctx = &ctx->ifc_softc_ctx;
3444 	segs = txq->ift_segs;
3445 	ntxd = txq->ift_size;
3446 	m_head = *m_headp;
3447 	map = NULL;
3448 
3449 	/*
3450 	 * If we're doing TSO the next descriptor to clean may be quite far ahead
3451 	 */
3452 	cidx = txq->ift_cidx;
3453 	pidx = txq->ift_pidx;
3454 	if (ctx->ifc_flags & IFC_PREFETCH) {
3455 		next = (cidx + CACHE_PTR_INCREMENT) & (ntxd-1);
3456 		if (!(ctx->ifc_flags & IFLIB_HAS_TXCQ)) {
3457 			next_txd = calc_next_txd(txq, cidx, 0);
3458 			prefetch(next_txd);
3459 		}
3460 
3461 		/* prefetch the next cache line of mbuf pointers and flags */
3462 		prefetch(&txq->ift_sds.ifsd_m[next]);
3463 		prefetch(&txq->ift_sds.ifsd_map[next]);
3464 		next = (cidx + CACHE_LINE_SIZE) & (ntxd-1);
3465 	}
3466 	map = txq->ift_sds.ifsd_map[pidx];
3467 	ifsd_m = txq->ift_sds.ifsd_m;
3468 
3469 	if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
3470 		buf_tag = txq->ift_tso_buf_tag;
3471 		max_segs = scctx->isc_tx_tso_segments_max;
3472 		map = txq->ift_sds.ifsd_tso_map[pidx];
3473 		MPASS(buf_tag != NULL);
3474 		MPASS(max_segs > 0);
3475 	} else {
3476 		buf_tag = txq->ift_buf_tag;
3477 		max_segs = scctx->isc_tx_nsegments;
3478 		map = txq->ift_sds.ifsd_map[pidx];
3479 	}
3480 	if ((sctx->isc_flags & IFLIB_NEED_ETHER_PAD) &&
3481 	    __predict_false(m_head->m_pkthdr.len < scctx->isc_min_frame_size)) {
3482 		err = iflib_ether_pad(ctx->ifc_dev, m_headp, scctx->isc_min_frame_size);
3483 		if (err) {
3484 			DBG_COUNTER_INC(encap_txd_encap_fail);
3485 			return err;
3486 		}
3487 	}
3488 	m_head = *m_headp;
3489 
3490 	pkt_info_zero(&pi);
3491 	pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG|M_BCAST|M_MCAST));
3492 	pi.ipi_pidx = pidx;
3493 	pi.ipi_qsidx = txq->ift_id;
3494 	pi.ipi_len = m_head->m_pkthdr.len;
3495 	pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags;
3496 	pi.ipi_vtag = M_HAS_VLANTAG(m_head) ? m_head->m_pkthdr.ether_vtag : 0;
3497 
3498 	/* deliberate bitwise OR to make one condition */
3499 	if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) {
3500 		if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0)) {
3501 			DBG_COUNTER_INC(encap_txd_encap_fail);
3502 			return (err);
3503 		}
3504 		m_head = *m_headp;
3505 	}
3506 
3507 retry:
3508 	err = bus_dmamap_load_mbuf_sg(buf_tag, map, m_head, segs, &nsegs,
3509 	    BUS_DMA_NOWAIT);
3510 defrag:
3511 	if (__predict_false(err)) {
3512 		switch (err) {
3513 		case EFBIG:
3514 			/* try collapse once and defrag once */
3515 			if (remap == 0) {
3516 				m_head = m_collapse(*m_headp, M_NOWAIT, max_segs);
3517 				/* try defrag if collapsing fails */
3518 				if (m_head == NULL)
3519 					remap++;
3520 			}
3521 			if (remap == 1) {
3522 				txq->ift_mbuf_defrag++;
3523 				m_head = m_defrag(*m_headp, M_NOWAIT);
3524 			}
3525 			/*
3526 			 * remap should never be >1 unless bus_dmamap_load_mbuf_sg
3527 			 * failed to map an mbuf that was run through m_defrag
3528 			 */
3529 			MPASS(remap <= 1);
3530 			if (__predict_false(m_head == NULL || remap > 1))
3531 				goto defrag_failed;
3532 			remap++;
3533 			*m_headp = m_head;
3534 			goto retry;
3535 			break;
3536 		case ENOMEM:
3537 			txq->ift_no_tx_dma_setup++;
3538 			break;
3539 		default:
3540 			txq->ift_no_tx_dma_setup++;
3541 			m_freem(*m_headp);
3542 			DBG_COUNTER_INC(tx_frees);
3543 			*m_headp = NULL;
3544 			break;
3545 		}
3546 		txq->ift_map_failed++;
3547 		DBG_COUNTER_INC(encap_load_mbuf_fail);
3548 		DBG_COUNTER_INC(encap_txd_encap_fail);
3549 		return (err);
3550 	}
3551 	ifsd_m[pidx] = m_head;
3552 	/*
3553 	 * XXX assumes a 1 to 1 relationship between segments and
3554 	 *        descriptors - this does not hold true on all drivers, e.g.
3555 	 *        cxgb
3556 	 */
3557 	if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) {
3558 		txq->ift_no_desc_avail++;
3559 		bus_dmamap_unload(buf_tag, map);
3560 		DBG_COUNTER_INC(encap_txq_avail_fail);
3561 		DBG_COUNTER_INC(encap_txd_encap_fail);
3562 		if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0)
3563 			GROUPTASK_ENQUEUE(&txq->ift_task);
3564 		return (ENOBUFS);
3565 	}
3566 	/*
3567 	 * On Intel cards we can greatly reduce the number of TX interrupts
3568 	 * we see by only setting report status on every Nth descriptor.
3569 	 * However, this also means that the driver will need to keep track
3570 	 * of the descriptors that RS was set on to check them for the DD bit.
3571 	 */
3572 	txq->ift_rs_pending += nsegs + 1;
3573 	if (txq->ift_rs_pending > TXQ_MAX_RS_DEFERRED(txq) ||
3574 	     iflib_no_tx_batch || (TXQ_AVAIL(txq) - nsegs) <= MAX_TX_DESC(ctx) + 2) {
3575 		pi.ipi_flags |= IPI_TX_INTR;
3576 		txq->ift_rs_pending = 0;
3577 	}
3578 
3579 	pi.ipi_segs = segs;
3580 	pi.ipi_nsegs = nsegs;
3581 
3582 	MPASS(pidx >= 0 && pidx < txq->ift_size);
3583 #ifdef PKT_DEBUG
3584 	print_pkt(&pi);
3585 #endif
3586 	if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) {
3587 		bus_dmamap_sync(buf_tag, map, BUS_DMASYNC_PREWRITE);
3588 		DBG_COUNTER_INC(tx_encap);
3589 		MPASS(pi.ipi_new_pidx < txq->ift_size);
3590 
3591 		ndesc = pi.ipi_new_pidx - pi.ipi_pidx;
3592 		if (pi.ipi_new_pidx < pi.ipi_pidx) {
3593 			ndesc += txq->ift_size;
3594 			txq->ift_gen = 1;
3595 		}
3596 		/*
3597 		 * drivers can need as many as
3598 		 * two sentinels
3599 		 */
3600 		MPASS(ndesc <= pi.ipi_nsegs + 2);
3601 		MPASS(pi.ipi_new_pidx != pidx);
3602 		MPASS(ndesc > 0);
3603 		txq->ift_in_use += ndesc;
3604 		txq->ift_db_pending += ndesc;
3605 
3606 		/*
3607 		 * We update the last software descriptor again here because there may
3608 		 * be a sentinel and/or there may be more mbufs than segments
3609 		 */
3610 		txq->ift_pidx = pi.ipi_new_pidx;
3611 		txq->ift_npending += pi.ipi_ndescs;
3612 	} else {
3613 		*m_headp = m_head = iflib_remove_mbuf(txq);
3614 		if (err == EFBIG) {
3615 			txq->ift_txd_encap_efbig++;
3616 			if (remap < 2) {
3617 				remap = 1;
3618 				goto defrag;
3619 			}
3620 		}
3621 		goto defrag_failed;
3622 	}
3623 	/*
3624 	 * err can't possibly be non-zero here, so we don't neet to test it
3625 	 * to see if we need to DBG_COUNTER_INC(encap_txd_encap_fail).
3626 	 */
3627 	return (err);
3628 
3629 defrag_failed:
3630 	txq->ift_mbuf_defrag_failed++;
3631 	txq->ift_map_failed++;
3632 	m_freem(*m_headp);
3633 	DBG_COUNTER_INC(tx_frees);
3634 	*m_headp = NULL;
3635 	DBG_COUNTER_INC(encap_txd_encap_fail);
3636 	return (ENOMEM);
3637 }
3638 
3639 static void
3640 iflib_tx_desc_free(iflib_txq_t txq, int n)
3641 {
3642 	uint32_t qsize, cidx, mask, gen;
3643 	struct mbuf *m, **ifsd_m;
3644 	bool do_prefetch;
3645 
3646 	cidx = txq->ift_cidx;
3647 	gen = txq->ift_gen;
3648 	qsize = txq->ift_size;
3649 	mask = qsize-1;
3650 	ifsd_m = txq->ift_sds.ifsd_m;
3651 	do_prefetch = (txq->ift_ctx->ifc_flags & IFC_PREFETCH);
3652 
3653 	while (n-- > 0) {
3654 		if (do_prefetch) {
3655 			prefetch(ifsd_m[(cidx + 3) & mask]);
3656 			prefetch(ifsd_m[(cidx + 4) & mask]);
3657 		}
3658 		if ((m = ifsd_m[cidx]) != NULL) {
3659 			prefetch(&ifsd_m[(cidx + CACHE_PTR_INCREMENT) & mask]);
3660 			if (m->m_pkthdr.csum_flags & CSUM_TSO) {
3661 				bus_dmamap_sync(txq->ift_tso_buf_tag,
3662 				    txq->ift_sds.ifsd_tso_map[cidx],
3663 				    BUS_DMASYNC_POSTWRITE);
3664 				bus_dmamap_unload(txq->ift_tso_buf_tag,
3665 				    txq->ift_sds.ifsd_tso_map[cidx]);
3666 			} else {
3667 				bus_dmamap_sync(txq->ift_buf_tag,
3668 				    txq->ift_sds.ifsd_map[cidx],
3669 				    BUS_DMASYNC_POSTWRITE);
3670 				bus_dmamap_unload(txq->ift_buf_tag,
3671 				    txq->ift_sds.ifsd_map[cidx]);
3672 			}
3673 			/* XXX we don't support any drivers that batch packets yet */
3674 			MPASS(m->m_nextpkt == NULL);
3675 			m_freem(m);
3676 			ifsd_m[cidx] = NULL;
3677 #if MEMORY_LOGGING
3678 			txq->ift_dequeued++;
3679 #endif
3680 			DBG_COUNTER_INC(tx_frees);
3681 		}
3682 		if (__predict_false(++cidx == qsize)) {
3683 			cidx = 0;
3684 			gen = 0;
3685 		}
3686 	}
3687 	txq->ift_cidx = cidx;
3688 	txq->ift_gen = gen;
3689 }
3690 
3691 static __inline int
3692 iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh)
3693 {
3694 	int reclaim;
3695 	if_ctx_t ctx = txq->ift_ctx;
3696 
3697 	KASSERT(thresh >= 0, ("invalid threshold to reclaim"));
3698 	MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size);
3699 
3700 	/*
3701 	 * Need a rate-limiting check so that this isn't called every time
3702 	 */
3703 	iflib_tx_credits_update(ctx, txq);
3704 	reclaim = DESC_RECLAIMABLE(txq);
3705 
3706 	if (reclaim <= thresh /* + MAX_TX_DESC(txq->ift_ctx) */) {
3707 #ifdef INVARIANTS
3708 		if (iflib_verbose_debug) {
3709 			printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __FUNCTION__,
3710 			       txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments,
3711 			       reclaim, thresh);
3712 		}
3713 #endif
3714 		return (0);
3715 	}
3716 	iflib_tx_desc_free(txq, reclaim);
3717 	txq->ift_cleaned += reclaim;
3718 	txq->ift_in_use -= reclaim;
3719 
3720 	return (reclaim);
3721 }
3722 
3723 static struct mbuf **
3724 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset, int remaining)
3725 {
3726 	int next, size;
3727 	struct mbuf **items;
3728 
3729 	size = r->size;
3730 	next = (cidx + CACHE_PTR_INCREMENT) & (size-1);
3731 	items = __DEVOLATILE(struct mbuf **, &r->items[0]);
3732 
3733 	prefetch(items[(cidx + offset) & (size-1)]);
3734 	if (remaining > 1) {
3735 		prefetch2cachelines(&items[next]);
3736 		prefetch2cachelines(items[(cidx + offset + 1) & (size-1)]);
3737 		prefetch2cachelines(items[(cidx + offset + 2) & (size-1)]);
3738 		prefetch2cachelines(items[(cidx + offset + 3) & (size-1)]);
3739 	}
3740 	return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (size-1)]));
3741 }
3742 
3743 static void
3744 iflib_txq_check_drain(iflib_txq_t txq, int budget)
3745 {
3746 
3747 	ifmp_ring_check_drainage(txq->ift_br, budget);
3748 }
3749 
3750 static uint32_t
3751 iflib_txq_can_drain(struct ifmp_ring *r)
3752 {
3753 	iflib_txq_t txq = r->cookie;
3754 	if_ctx_t ctx = txq->ift_ctx;
3755 
3756 	if (TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2)
3757 		return (1);
3758 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3759 	    BUS_DMASYNC_POSTREAD);
3760 	return (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id,
3761 	    false));
3762 }
3763 
3764 static uint32_t
3765 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3766 {
3767 	iflib_txq_t txq = r->cookie;
3768 	if_ctx_t ctx = txq->ift_ctx;
3769 	if_t ifp = ctx->ifc_ifp;
3770 	struct mbuf *m, **mp;
3771 	int avail, bytes_sent, skipped, count, err, i;
3772 	int mcast_sent, pkt_sent, reclaimed;
3773 	bool do_prefetch, rang, ring;
3774 
3775 	if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) ||
3776 			    !LINK_ACTIVE(ctx))) {
3777 		DBG_COUNTER_INC(txq_drain_notready);
3778 		return (0);
3779 	}
3780 	reclaimed = iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
3781 	rang = iflib_txd_db_check(txq, reclaimed && txq->ift_db_pending);
3782 	avail = IDXDIFF(pidx, cidx, r->size);
3783 
3784 	if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) {
3785 		/*
3786 		 * The driver is unloading so we need to free all pending packets.
3787 		 */
3788 		DBG_COUNTER_INC(txq_drain_flushing);
3789 		for (i = 0; i < avail; i++) {
3790 			if (__predict_true(r->items[(cidx + i) & (r->size-1)] != (void *)txq))
3791 				m_freem(r->items[(cidx + i) & (r->size-1)]);
3792 			r->items[(cidx + i) & (r->size-1)] = NULL;
3793 		}
3794 		return (avail);
3795 	}
3796 
3797 	if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3798 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3799 		CALLOUT_LOCK(txq);
3800 		callout_stop(&txq->ift_timer);
3801 		CALLOUT_UNLOCK(txq);
3802 		DBG_COUNTER_INC(txq_drain_oactive);
3803 		return (0);
3804 	}
3805 
3806 	/*
3807 	 * If we've reclaimed any packets this queue cannot be hung.
3808 	 */
3809 	if (reclaimed)
3810 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3811 	skipped = mcast_sent = bytes_sent = pkt_sent = 0;
3812 	count = MIN(avail, TX_BATCH_SIZE);
3813 #ifdef INVARIANTS
3814 	if (iflib_verbose_debug)
3815 		printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __FUNCTION__,
3816 		       avail, ctx->ifc_flags, TXQ_AVAIL(txq));
3817 #endif
3818 	do_prefetch = (ctx->ifc_flags & IFC_PREFETCH);
3819 	err = 0;
3820 	for (i = 0; i < count && TXQ_AVAIL(txq) >= MAX_TX_DESC(ctx) + 2; i++) {
3821 		int rem = do_prefetch ? count - i : 0;
3822 
3823 		mp = _ring_peek_one(r, cidx, i, rem);
3824 		MPASS(mp != NULL && *mp != NULL);
3825 
3826 		/*
3827 		 * Completion interrupts will use the address of the txq
3828 		 * as a sentinel to enqueue _something_ in order to acquire
3829 		 * the lock on the mp_ring (there's no direct lock call).
3830 		 * We obviously whave to check for these sentinel cases
3831 		 * and skip them.
3832 		 */
3833 		if (__predict_false(*mp == (struct mbuf *)txq)) {
3834 			skipped++;
3835 			continue;
3836 		}
3837 		err = iflib_encap(txq, mp);
3838 		if (__predict_false(err)) {
3839 			/* no room - bail out */
3840 			if (err == ENOBUFS)
3841 				break;
3842 			skipped++;
3843 			/* we can't send this packet - skip it */
3844 			continue;
3845 		}
3846 		pkt_sent++;
3847 		m = *mp;
3848 		DBG_COUNTER_INC(tx_sent);
3849 		bytes_sent += m->m_pkthdr.len;
3850 		mcast_sent += !!(m->m_flags & M_MCAST);
3851 
3852 		if (__predict_false(!(ifp->if_drv_flags & IFF_DRV_RUNNING)))
3853 			break;
3854 		ETHER_BPF_MTAP(ifp, m);
3855 		rang = iflib_txd_db_check(txq, false);
3856 	}
3857 
3858 	/* deliberate use of bitwise or to avoid gratuitous short-circuit */
3859 	ring = rang ? false  : (iflib_min_tx_latency | err);
3860 	iflib_txd_db_check(txq, ring);
3861 	if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
3862 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
3863 	if (mcast_sent)
3864 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
3865 #ifdef INVARIANTS
3866 	if (iflib_verbose_debug)
3867 		printf("consumed=%d\n", skipped + pkt_sent);
3868 #endif
3869 	return (skipped + pkt_sent);
3870 }
3871 
3872 static uint32_t
3873 iflib_txq_drain_always(struct ifmp_ring *r)
3874 {
3875 	return (1);
3876 }
3877 
3878 static uint32_t
3879 iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3880 {
3881 	int i, avail;
3882 	struct mbuf **mp;
3883 	iflib_txq_t txq;
3884 
3885 	txq = r->cookie;
3886 
3887 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3888 	CALLOUT_LOCK(txq);
3889 	callout_stop(&txq->ift_timer);
3890 	CALLOUT_UNLOCK(txq);
3891 
3892 	avail = IDXDIFF(pidx, cidx, r->size);
3893 	for (i = 0; i < avail; i++) {
3894 		mp = _ring_peek_one(r, cidx, i, avail - i);
3895 		if (__predict_false(*mp == (struct mbuf *)txq))
3896 			continue;
3897 		m_freem(*mp);
3898 		DBG_COUNTER_INC(tx_frees);
3899 	}
3900 	MPASS(ifmp_ring_is_stalled(r) == 0);
3901 	return (avail);
3902 }
3903 
3904 static void
3905 iflib_ifmp_purge(iflib_txq_t txq)
3906 {
3907 	struct ifmp_ring *r;
3908 
3909 	r = txq->ift_br;
3910 	r->drain = iflib_txq_drain_free;
3911 	r->can_drain = iflib_txq_drain_always;
3912 
3913 	ifmp_ring_check_drainage(r, r->size);
3914 
3915 	r->drain = iflib_txq_drain;
3916 	r->can_drain = iflib_txq_can_drain;
3917 }
3918 
3919 static void
3920 _task_fn_tx(void *context)
3921 {
3922 	iflib_txq_t txq = context;
3923 	if_ctx_t ctx = txq->ift_ctx;
3924 	if_t ifp = ctx->ifc_ifp;
3925 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
3926 
3927 #ifdef IFLIB_DIAGNOSTICS
3928 	txq->ift_cpu_exec_count[curcpu]++;
3929 #endif
3930 	if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
3931 		return;
3932 #ifdef DEV_NETMAP
3933 	if ((if_getcapenable(ifp) & IFCAP_NETMAP) &&
3934 	    netmap_tx_irq(ifp, txq->ift_id))
3935 		goto skip_ifmp;
3936 #endif
3937 #ifdef ALTQ
3938 	if (ALTQ_IS_ENABLED(&ifp->if_snd))
3939 		iflib_altq_if_start(ifp);
3940 #endif
3941 	if (txq->ift_db_pending)
3942 		ifmp_ring_enqueue(txq->ift_br, (void **)&txq, 1, TX_BATCH_SIZE, abdicate);
3943 	else if (!abdicate)
3944 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3945 	/*
3946 	 * When abdicating, we always need to check drainage, not just when we don't enqueue
3947 	 */
3948 	if (abdicate)
3949 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3950 #ifdef DEV_NETMAP
3951 skip_ifmp:
3952 #endif
3953 	if (ctx->ifc_flags & IFC_LEGACY)
3954 		IFDI_INTR_ENABLE(ctx);
3955 	else
3956 		IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3957 }
3958 
3959 static void
3960 _task_fn_rx(void *context)
3961 {
3962 	iflib_rxq_t rxq = context;
3963 	if_ctx_t ctx = rxq->ifr_ctx;
3964 	uint8_t more;
3965 	uint16_t budget;
3966 #ifdef DEV_NETMAP
3967 	u_int work = 0;
3968 	int nmirq;
3969 #endif
3970 
3971 #ifdef IFLIB_DIAGNOSTICS
3972 	rxq->ifr_cpu_exec_count[curcpu]++;
3973 #endif
3974 	DBG_COUNTER_INC(task_fn_rxs);
3975 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3976 		return;
3977 #ifdef DEV_NETMAP
3978 	nmirq = netmap_rx_irq(ctx->ifc_ifp, rxq->ifr_id, &work);
3979 	if (nmirq != NM_IRQ_PASS) {
3980 		more = (nmirq == NM_IRQ_RESCHED) ? IFLIB_RXEOF_MORE : 0;
3981 		goto skip_rxeof;
3982 	}
3983 #endif
3984 	budget = ctx->ifc_sysctl_rx_budget;
3985 	if (budget == 0)
3986 		budget = 16;	/* XXX */
3987 	more = iflib_rxeof(rxq, budget);
3988 #ifdef DEV_NETMAP
3989 skip_rxeof:
3990 #endif
3991 	if ((more & IFLIB_RXEOF_MORE) == 0) {
3992 		if (ctx->ifc_flags & IFC_LEGACY)
3993 			IFDI_INTR_ENABLE(ctx);
3994 		else
3995 			IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
3996 		DBG_COUNTER_INC(rx_intr_enables);
3997 	}
3998 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3999 		return;
4000 
4001 	if (more & IFLIB_RXEOF_MORE)
4002 		GROUPTASK_ENQUEUE(&rxq->ifr_task);
4003 	else if (more & IFLIB_RXEOF_EMPTY)
4004 		callout_reset_curcpu(&rxq->ifr_watchdog, 1, &_task_fn_rx_watchdog, rxq);
4005 }
4006 
4007 static void
4008 _task_fn_admin(void *context)
4009 {
4010 	if_ctx_t ctx = context;
4011 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
4012 	iflib_txq_t txq;
4013 	int i;
4014 	bool oactive, running, do_reset, do_watchdog, in_detach;
4015 
4016 	STATE_LOCK(ctx);
4017 	running = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING);
4018 	oactive = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE);
4019 	do_reset = (ctx->ifc_flags & IFC_DO_RESET);
4020 	do_watchdog = (ctx->ifc_flags & IFC_DO_WATCHDOG);
4021 	in_detach = (ctx->ifc_flags & IFC_IN_DETACH);
4022 	ctx->ifc_flags &= ~(IFC_DO_RESET|IFC_DO_WATCHDOG);
4023 	STATE_UNLOCK(ctx);
4024 
4025 	if ((!running && !oactive) && !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
4026 		return;
4027 	if (in_detach)
4028 		return;
4029 
4030 	CTX_LOCK(ctx);
4031 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
4032 		CALLOUT_LOCK(txq);
4033 		callout_stop(&txq->ift_timer);
4034 		CALLOUT_UNLOCK(txq);
4035 	}
4036 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_ADMINCQ)
4037 		IFDI_ADMIN_COMPLETION_HANDLE(ctx);
4038 	if (do_watchdog) {
4039 		ctx->ifc_watchdog_events++;
4040 		IFDI_WATCHDOG_RESET(ctx);
4041 	}
4042 	IFDI_UPDATE_ADMIN_STATUS(ctx);
4043 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
4044 		callout_reset_on(&txq->ift_timer, iflib_timer_default, iflib_timer, txq,
4045 		    txq->ift_timer.c_cpu);
4046 	}
4047 	IFDI_LINK_INTR_ENABLE(ctx);
4048 	if (do_reset)
4049 		iflib_if_init_locked(ctx);
4050 	CTX_UNLOCK(ctx);
4051 
4052 	if (LINK_ACTIVE(ctx) == 0)
4053 		return;
4054 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
4055 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
4056 }
4057 
4058 static void
4059 _task_fn_iov(void *context)
4060 {
4061 	if_ctx_t ctx = context;
4062 
4063 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING) &&
4064 	    !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
4065 		return;
4066 
4067 	CTX_LOCK(ctx);
4068 	IFDI_VFLR_HANDLE(ctx);
4069 	CTX_UNLOCK(ctx);
4070 }
4071 
4072 static int
4073 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)
4074 {
4075 	int err;
4076 	if_int_delay_info_t info;
4077 	if_ctx_t ctx;
4078 
4079 	info = (if_int_delay_info_t)arg1;
4080 	ctx = info->iidi_ctx;
4081 	info->iidi_req = req;
4082 	info->iidi_oidp = oidp;
4083 	CTX_LOCK(ctx);
4084 	err = IFDI_SYSCTL_INT_DELAY(ctx, info);
4085 	CTX_UNLOCK(ctx);
4086 	return (err);
4087 }
4088 
4089 /*********************************************************************
4090  *
4091  *  IFNET FUNCTIONS
4092  *
4093  **********************************************************************/
4094 
4095 static void
4096 iflib_if_init_locked(if_ctx_t ctx)
4097 {
4098 	iflib_stop(ctx);
4099 	iflib_init_locked(ctx);
4100 }
4101 
4102 static void
4103 iflib_if_init(void *arg)
4104 {
4105 	if_ctx_t ctx = arg;
4106 
4107 	CTX_LOCK(ctx);
4108 	iflib_if_init_locked(ctx);
4109 	CTX_UNLOCK(ctx);
4110 }
4111 
4112 static int
4113 iflib_if_transmit(if_t ifp, struct mbuf *m)
4114 {
4115 	if_ctx_t	ctx = if_getsoftc(ifp);
4116 
4117 	iflib_txq_t txq;
4118 	int err, qidx;
4119 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
4120 
4121 	if (__predict_false((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) {
4122 		DBG_COUNTER_INC(tx_frees);
4123 		m_freem(m);
4124 		return (ENETDOWN);
4125 	}
4126 
4127 	MPASS(m->m_nextpkt == NULL);
4128 	/* ALTQ-enabled interfaces always use queue 0. */
4129 	qidx = 0;
4130 	if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m) && !ALTQ_IS_ENABLED(&ifp->if_snd))
4131 		qidx = QIDX(ctx, m);
4132 	/*
4133 	 * XXX calculate buf_ring based on flowid (divvy up bits?)
4134 	 */
4135 	txq = &ctx->ifc_txqs[qidx];
4136 
4137 #ifdef DRIVER_BACKPRESSURE
4138 	if (txq->ift_closed) {
4139 		while (m != NULL) {
4140 			next = m->m_nextpkt;
4141 			m->m_nextpkt = NULL;
4142 			m_freem(m);
4143 			DBG_COUNTER_INC(tx_frees);
4144 			m = next;
4145 		}
4146 		return (ENOBUFS);
4147 	}
4148 #endif
4149 #ifdef notyet
4150 	qidx = count = 0;
4151 	mp = marr;
4152 	next = m;
4153 	do {
4154 		count++;
4155 		next = next->m_nextpkt;
4156 	} while (next != NULL);
4157 
4158 	if (count > nitems(marr))
4159 		if ((mp = malloc(count*sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) {
4160 			/* XXX check nextpkt */
4161 			m_freem(m);
4162 			/* XXX simplify for now */
4163 			DBG_COUNTER_INC(tx_frees);
4164 			return (ENOBUFS);
4165 		}
4166 	for (next = m, i = 0; next != NULL; i++) {
4167 		mp[i] = next;
4168 		next = next->m_nextpkt;
4169 		mp[i]->m_nextpkt = NULL;
4170 	}
4171 #endif
4172 	DBG_COUNTER_INC(tx_seen);
4173 	err = ifmp_ring_enqueue(txq->ift_br, (void **)&m, 1, TX_BATCH_SIZE, abdicate);
4174 
4175 	if (abdicate)
4176 		GROUPTASK_ENQUEUE(&txq->ift_task);
4177  	if (err) {
4178 		if (!abdicate)
4179 			GROUPTASK_ENQUEUE(&txq->ift_task);
4180 		/* support forthcoming later */
4181 #ifdef DRIVER_BACKPRESSURE
4182 		txq->ift_closed = TRUE;
4183 #endif
4184 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
4185 		m_freem(m);
4186 		DBG_COUNTER_INC(tx_frees);
4187 	}
4188 
4189 	return (err);
4190 }
4191 
4192 #ifdef ALTQ
4193 /*
4194  * The overall approach to integrating iflib with ALTQ is to continue to use
4195  * the iflib mp_ring machinery between the ALTQ queue(s) and the hardware
4196  * ring.  Technically, when using ALTQ, queueing to an intermediate mp_ring
4197  * is redundant/unnecessary, but doing so minimizes the amount of
4198  * ALTQ-specific code required in iflib.  It is assumed that the overhead of
4199  * redundantly queueing to an intermediate mp_ring is swamped by the
4200  * performance limitations inherent in using ALTQ.
4201  *
4202  * When ALTQ support is compiled in, all iflib drivers will use a transmit
4203  * routine, iflib_altq_if_transmit(), that checks if ALTQ is enabled for the
4204  * given interface.  If ALTQ is enabled for an interface, then all
4205  * transmitted packets for that interface will be submitted to the ALTQ
4206  * subsystem via IFQ_ENQUEUE().  We don't use the legacy if_transmit()
4207  * implementation because it uses IFQ_HANDOFF(), which will duplicatively
4208  * update stats that the iflib machinery handles, and which is sensitve to
4209  * the disused IFF_DRV_OACTIVE flag.  Additionally, iflib_altq_if_start()
4210  * will be installed as the start routine for use by ALTQ facilities that
4211  * need to trigger queue drains on a scheduled basis.
4212  *
4213  */
4214 static void
4215 iflib_altq_if_start(if_t ifp)
4216 {
4217 	struct ifaltq *ifq = &ifp->if_snd;
4218 	struct mbuf *m;
4219 
4220 	IFQ_LOCK(ifq);
4221 	IFQ_DEQUEUE_NOLOCK(ifq, m);
4222 	while (m != NULL) {
4223 		iflib_if_transmit(ifp, m);
4224 		IFQ_DEQUEUE_NOLOCK(ifq, m);
4225 	}
4226 	IFQ_UNLOCK(ifq);
4227 }
4228 
4229 static int
4230 iflib_altq_if_transmit(if_t ifp, struct mbuf *m)
4231 {
4232 	int err;
4233 
4234 	if (ALTQ_IS_ENABLED(&ifp->if_snd)) {
4235 		IFQ_ENQUEUE(&ifp->if_snd, m, err);
4236 		if (err == 0)
4237 			iflib_altq_if_start(ifp);
4238 	} else
4239 		err = iflib_if_transmit(ifp, m);
4240 
4241 	return (err);
4242 }
4243 #endif /* ALTQ */
4244 
4245 static void
4246 iflib_if_qflush(if_t ifp)
4247 {
4248 	if_ctx_t ctx = if_getsoftc(ifp);
4249 	iflib_txq_t txq = ctx->ifc_txqs;
4250 	int i;
4251 
4252 	STATE_LOCK(ctx);
4253 	ctx->ifc_flags |= IFC_QFLUSH;
4254 	STATE_UNLOCK(ctx);
4255 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
4256 		while (!(ifmp_ring_is_idle(txq->ift_br) || ifmp_ring_is_stalled(txq->ift_br)))
4257 			iflib_txq_check_drain(txq, 0);
4258 	STATE_LOCK(ctx);
4259 	ctx->ifc_flags &= ~IFC_QFLUSH;
4260 	STATE_UNLOCK(ctx);
4261 
4262 	/*
4263 	 * When ALTQ is enabled, this will also take care of purging the
4264 	 * ALTQ queue(s).
4265 	 */
4266 	if_qflush(ifp);
4267 }
4268 
4269 #define IFCAP_FLAGS (IFCAP_HWCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \
4270 		     IFCAP_TSO | IFCAP_VLAN_HWTAGGING | IFCAP_HWSTATS | \
4271 		     IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | \
4272 		     IFCAP_VLAN_HWTSO | IFCAP_VLAN_HWCSUM | IFCAP_MEXTPG)
4273 
4274 static int
4275 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data)
4276 {
4277 	if_ctx_t ctx = if_getsoftc(ifp);
4278 	struct ifreq	*ifr = (struct ifreq *)data;
4279 #if defined(INET) || defined(INET6)
4280 	struct ifaddr	*ifa = (struct ifaddr *)data;
4281 #endif
4282 	bool		avoid_reset = false;
4283 	int		err = 0, reinit = 0, bits;
4284 
4285 	switch (command) {
4286 	case SIOCSIFADDR:
4287 #ifdef INET
4288 		if (ifa->ifa_addr->sa_family == AF_INET)
4289 			avoid_reset = true;
4290 #endif
4291 #ifdef INET6
4292 		if (ifa->ifa_addr->sa_family == AF_INET6)
4293 			avoid_reset = true;
4294 #endif
4295 		/*
4296 		** Calling init results in link renegotiation,
4297 		** so we avoid doing it when possible.
4298 		*/
4299 		if (avoid_reset) {
4300 			if_setflagbits(ifp, IFF_UP,0);
4301 			if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
4302 				reinit = 1;
4303 #ifdef INET
4304 			if (!(if_getflags(ifp) & IFF_NOARP))
4305 				arp_ifinit(ifp, ifa);
4306 #endif
4307 		} else
4308 			err = ether_ioctl(ifp, command, data);
4309 		break;
4310 	case SIOCSIFMTU:
4311 		CTX_LOCK(ctx);
4312 		if (ifr->ifr_mtu == if_getmtu(ifp)) {
4313 			CTX_UNLOCK(ctx);
4314 			break;
4315 		}
4316 		bits = if_getdrvflags(ifp);
4317 		/* stop the driver and free any clusters before proceeding */
4318 		iflib_stop(ctx);
4319 
4320 		if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) {
4321 			STATE_LOCK(ctx);
4322 			if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size)
4323 				ctx->ifc_flags |= IFC_MULTISEG;
4324 			else
4325 				ctx->ifc_flags &= ~IFC_MULTISEG;
4326 			STATE_UNLOCK(ctx);
4327 			err = if_setmtu(ifp, ifr->ifr_mtu);
4328 		}
4329 		iflib_init_locked(ctx);
4330 		STATE_LOCK(ctx);
4331 		if_setdrvflags(ifp, bits);
4332 		STATE_UNLOCK(ctx);
4333 		CTX_UNLOCK(ctx);
4334 		break;
4335 	case SIOCSIFFLAGS:
4336 		CTX_LOCK(ctx);
4337 		if (if_getflags(ifp) & IFF_UP) {
4338 			if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4339 				if ((if_getflags(ifp) ^ ctx->ifc_if_flags) &
4340 				    (IFF_PROMISC | IFF_ALLMULTI)) {
4341 					CTX_UNLOCK(ctx);
4342 					err = IFDI_PROMISC_SET(ctx, if_getflags(ifp));
4343 					CTX_LOCK(ctx);
4344 				}
4345 			} else
4346 				reinit = 1;
4347 		} else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4348 			iflib_stop(ctx);
4349 		}
4350 		ctx->ifc_if_flags = if_getflags(ifp);
4351 		CTX_UNLOCK(ctx);
4352 		break;
4353 	case SIOCADDMULTI:
4354 	case SIOCDELMULTI:
4355 		if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4356 			CTX_LOCK(ctx);
4357 			IFDI_INTR_DISABLE(ctx);
4358 			IFDI_MULTI_SET(ctx);
4359 			IFDI_INTR_ENABLE(ctx);
4360 			CTX_UNLOCK(ctx);
4361 		}
4362 		break;
4363 	case SIOCSIFMEDIA:
4364 		CTX_LOCK(ctx);
4365 		IFDI_MEDIA_SET(ctx);
4366 		CTX_UNLOCK(ctx);
4367 		/* FALLTHROUGH */
4368 	case SIOCGIFMEDIA:
4369 	case SIOCGIFXMEDIA:
4370 		err = ifmedia_ioctl(ifp, ifr, ctx->ifc_mediap, command);
4371 		break;
4372 	case SIOCGI2C:
4373 	{
4374 		struct ifi2creq i2c;
4375 
4376 		err = copyin(ifr_data_get_ptr(ifr), &i2c, sizeof(i2c));
4377 		if (err != 0)
4378 			break;
4379 		if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) {
4380 			err = EINVAL;
4381 			break;
4382 		}
4383 		if (i2c.len > sizeof(i2c.data)) {
4384 			err = EINVAL;
4385 			break;
4386 		}
4387 
4388 		if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0)
4389 			err = copyout(&i2c, ifr_data_get_ptr(ifr),
4390 			    sizeof(i2c));
4391 		break;
4392 	}
4393 	case SIOCSIFCAP:
4394 	{
4395 		int mask, setmask, oldmask;
4396 
4397 		oldmask = if_getcapenable(ifp);
4398 		mask = ifr->ifr_reqcap ^ oldmask;
4399 		mask &= ctx->ifc_softc_ctx.isc_capabilities | IFCAP_MEXTPG;
4400 		setmask = 0;
4401 #ifdef TCP_OFFLOAD
4402 		setmask |= mask & (IFCAP_TOE4|IFCAP_TOE6);
4403 #endif
4404 		setmask |= (mask & IFCAP_FLAGS);
4405 		setmask |= (mask & IFCAP_WOL);
4406 
4407 		/*
4408 		 * If any RX csum has changed, change all the ones that
4409 		 * are supported by the driver.
4410 		 */
4411 		if (setmask & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6)) {
4412 			setmask |= ctx->ifc_softc_ctx.isc_capabilities &
4413 			    (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6);
4414 		}
4415 
4416 		/*
4417 		 * want to ensure that traffic has stopped before we change any of the flags
4418 		 */
4419 		if (setmask) {
4420 			CTX_LOCK(ctx);
4421 			bits = if_getdrvflags(ifp);
4422 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4423 				iflib_stop(ctx);
4424 			STATE_LOCK(ctx);
4425 			if_togglecapenable(ifp, setmask);
4426 			STATE_UNLOCK(ctx);
4427 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4428 				iflib_init_locked(ctx);
4429 			STATE_LOCK(ctx);
4430 			if_setdrvflags(ifp, bits);
4431 			STATE_UNLOCK(ctx);
4432 			CTX_UNLOCK(ctx);
4433 		}
4434 		if_vlancap(ifp);
4435 		break;
4436 	}
4437 	case SIOCGPRIVATE_0:
4438 	case SIOCSDRVSPEC:
4439 	case SIOCGDRVSPEC:
4440 		CTX_LOCK(ctx);
4441 		err = IFDI_PRIV_IOCTL(ctx, command, data);
4442 		CTX_UNLOCK(ctx);
4443 		break;
4444 	default:
4445 		err = ether_ioctl(ifp, command, data);
4446 		break;
4447 	}
4448 	if (reinit)
4449 		iflib_if_init(ctx);
4450 	return (err);
4451 }
4452 
4453 static uint64_t
4454 iflib_if_get_counter(if_t ifp, ift_counter cnt)
4455 {
4456 	if_ctx_t ctx = if_getsoftc(ifp);
4457 
4458 	return (IFDI_GET_COUNTER(ctx, cnt));
4459 }
4460 
4461 /*********************************************************************
4462  *
4463  *  OTHER FUNCTIONS EXPORTED TO THE STACK
4464  *
4465  **********************************************************************/
4466 
4467 static void
4468 iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag)
4469 {
4470 	if_ctx_t ctx = if_getsoftc(ifp);
4471 
4472 	if ((void *)ctx != arg)
4473 		return;
4474 
4475 	if ((vtag == 0) || (vtag > 4095))
4476 		return;
4477 
4478 	if (iflib_in_detach(ctx))
4479 		return;
4480 
4481 	CTX_LOCK(ctx);
4482 	/* Driver may need all untagged packets to be flushed */
4483 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4484 		iflib_stop(ctx);
4485 	IFDI_VLAN_REGISTER(ctx, vtag);
4486 	/* Re-init to load the changes, if required */
4487 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4488 		iflib_init_locked(ctx);
4489 	CTX_UNLOCK(ctx);
4490 }
4491 
4492 static void
4493 iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag)
4494 {
4495 	if_ctx_t ctx = if_getsoftc(ifp);
4496 
4497 	if ((void *)ctx != arg)
4498 		return;
4499 
4500 	if ((vtag == 0) || (vtag > 4095))
4501 		return;
4502 
4503 	CTX_LOCK(ctx);
4504 	/* Driver may need all tagged packets to be flushed */
4505 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4506 		iflib_stop(ctx);
4507 	IFDI_VLAN_UNREGISTER(ctx, vtag);
4508 	/* Re-init to load the changes, if required */
4509 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4510 		iflib_init_locked(ctx);
4511 	CTX_UNLOCK(ctx);
4512 }
4513 
4514 static void
4515 iflib_led_func(void *arg, int onoff)
4516 {
4517 	if_ctx_t ctx = arg;
4518 
4519 	CTX_LOCK(ctx);
4520 	IFDI_LED_FUNC(ctx, onoff);
4521 	CTX_UNLOCK(ctx);
4522 }
4523 
4524 /*********************************************************************
4525  *
4526  *  BUS FUNCTION DEFINITIONS
4527  *
4528  **********************************************************************/
4529 
4530 int
4531 iflib_device_probe(device_t dev)
4532 {
4533 	const pci_vendor_info_t *ent;
4534 	if_shared_ctx_t sctx;
4535 	uint16_t pci_device_id, pci_rev_id, pci_subdevice_id, pci_subvendor_id;
4536 	uint16_t pci_vendor_id;
4537 
4538 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4539 		return (ENOTSUP);
4540 
4541 	pci_vendor_id = pci_get_vendor(dev);
4542 	pci_device_id = pci_get_device(dev);
4543 	pci_subvendor_id = pci_get_subvendor(dev);
4544 	pci_subdevice_id = pci_get_subdevice(dev);
4545 	pci_rev_id = pci_get_revid(dev);
4546 	if (sctx->isc_parse_devinfo != NULL)
4547 		sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id);
4548 
4549 	ent = sctx->isc_vendor_info;
4550 	while (ent->pvi_vendor_id != 0) {
4551 		if (pci_vendor_id != ent->pvi_vendor_id) {
4552 			ent++;
4553 			continue;
4554 		}
4555 		if ((pci_device_id == ent->pvi_device_id) &&
4556 		    ((pci_subvendor_id == ent->pvi_subvendor_id) ||
4557 		     (ent->pvi_subvendor_id == 0)) &&
4558 		    ((pci_subdevice_id == ent->pvi_subdevice_id) ||
4559 		     (ent->pvi_subdevice_id == 0)) &&
4560 		    ((pci_rev_id == ent->pvi_rev_id) ||
4561 		     (ent->pvi_rev_id == 0))) {
4562 			device_set_desc_copy(dev, ent->pvi_name);
4563 			/* this needs to be changed to zero if the bus probing code
4564 			 * ever stops re-probing on best match because the sctx
4565 			 * may have its values over written by register calls
4566 			 * in subsequent probes
4567 			 */
4568 			return (BUS_PROBE_DEFAULT);
4569 		}
4570 		ent++;
4571 	}
4572 	return (ENXIO);
4573 }
4574 
4575 int
4576 iflib_device_probe_vendor(device_t dev)
4577 {
4578 	int probe;
4579 
4580 	probe = iflib_device_probe(dev);
4581 	if (probe == BUS_PROBE_DEFAULT)
4582 		return (BUS_PROBE_VENDOR);
4583 	else
4584 		return (probe);
4585 }
4586 
4587 static void
4588 iflib_reset_qvalues(if_ctx_t ctx)
4589 {
4590 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4591 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4592 	device_t dev = ctx->ifc_dev;
4593 	int i;
4594 
4595 	if (ctx->ifc_sysctl_ntxqs != 0)
4596 		scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs;
4597 	if (ctx->ifc_sysctl_nrxqs != 0)
4598 		scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs;
4599 
4600 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4601 		if (ctx->ifc_sysctl_ntxds[i] != 0)
4602 			scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i];
4603 		else
4604 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4605 	}
4606 
4607 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4608 		if (ctx->ifc_sysctl_nrxds[i] != 0)
4609 			scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i];
4610 		else
4611 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4612 	}
4613 
4614 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4615 		if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) {
4616 			device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n",
4617 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]);
4618 			scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i];
4619 		}
4620 		if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) {
4621 			device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n",
4622 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]);
4623 			scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i];
4624 		}
4625 		if (!powerof2(scctx->isc_nrxd[i])) {
4626 			device_printf(dev, "nrxd%d: %d is not a power of 2 - using default value of %d\n",
4627 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_default[i]);
4628 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4629 		}
4630 	}
4631 
4632 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4633 		if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) {
4634 			device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n",
4635 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]);
4636 			scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i];
4637 		}
4638 		if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) {
4639 			device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n",
4640 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]);
4641 			scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i];
4642 		}
4643 		if (!powerof2(scctx->isc_ntxd[i])) {
4644 			device_printf(dev, "ntxd%d: %d is not a power of 2 - using default value of %d\n",
4645 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_default[i]);
4646 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4647 		}
4648 	}
4649 }
4650 
4651 static void
4652 iflib_add_pfil(if_ctx_t ctx)
4653 {
4654 	struct pfil_head *pfil;
4655 	struct pfil_head_args pa;
4656 	iflib_rxq_t rxq;
4657 	int i;
4658 
4659 	pa.pa_version = PFIL_VERSION;
4660 	pa.pa_flags = PFIL_IN;
4661 	pa.pa_type = PFIL_TYPE_ETHERNET;
4662 	pa.pa_headname = ctx->ifc_ifp->if_xname;
4663 	pfil = pfil_head_register(&pa);
4664 
4665 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4666 		rxq->pfil = pfil;
4667 	}
4668 }
4669 
4670 static void
4671 iflib_rem_pfil(if_ctx_t ctx)
4672 {
4673 	struct pfil_head *pfil;
4674 	iflib_rxq_t rxq;
4675 	int i;
4676 
4677 	rxq = ctx->ifc_rxqs;
4678 	pfil = rxq->pfil;
4679 	for (i = 0; i < NRXQSETS(ctx); i++, rxq++) {
4680 		rxq->pfil = NULL;
4681 	}
4682 	pfil_head_unregister(pfil);
4683 }
4684 
4685 static uint16_t
4686 get_ctx_core_offset(if_ctx_t ctx)
4687 {
4688 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4689 	struct cpu_offset *op;
4690 	uint16_t qc;
4691 	uint16_t ret = ctx->ifc_sysctl_core_offset;
4692 
4693 	if (ret != CORE_OFFSET_UNSPECIFIED)
4694 		return (ret);
4695 
4696 	if (ctx->ifc_sysctl_separate_txrx)
4697 		qc = scctx->isc_ntxqsets + scctx->isc_nrxqsets;
4698 	else
4699 		qc = max(scctx->isc_ntxqsets, scctx->isc_nrxqsets);
4700 
4701 	mtx_lock(&cpu_offset_mtx);
4702 	SLIST_FOREACH(op, &cpu_offsets, entries) {
4703 		if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) {
4704 			ret = op->offset;
4705 			op->offset += qc;
4706 			MPASS(op->refcount < UINT_MAX);
4707 			op->refcount++;
4708 			break;
4709 		}
4710 	}
4711 	if (ret == CORE_OFFSET_UNSPECIFIED) {
4712 		ret = 0;
4713 		op = malloc(sizeof(struct cpu_offset), M_IFLIB,
4714 		    M_NOWAIT | M_ZERO);
4715 		if (op == NULL) {
4716 			device_printf(ctx->ifc_dev,
4717 			    "allocation for cpu offset failed.\n");
4718 		} else {
4719 			op->offset = qc;
4720 			op->refcount = 1;
4721 			CPU_COPY(&ctx->ifc_cpus, &op->set);
4722 			SLIST_INSERT_HEAD(&cpu_offsets, op, entries);
4723 		}
4724 	}
4725 	mtx_unlock(&cpu_offset_mtx);
4726 
4727 	return (ret);
4728 }
4729 
4730 static void
4731 unref_ctx_core_offset(if_ctx_t ctx)
4732 {
4733 	struct cpu_offset *op, *top;
4734 
4735 	mtx_lock(&cpu_offset_mtx);
4736 	SLIST_FOREACH_SAFE(op, &cpu_offsets, entries, top) {
4737 		if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) {
4738 			MPASS(op->refcount > 0);
4739 			op->refcount--;
4740 			if (op->refcount == 0) {
4741 				SLIST_REMOVE(&cpu_offsets, op, cpu_offset, entries);
4742 				free(op, M_IFLIB);
4743 			}
4744 			break;
4745 		}
4746 	}
4747 	mtx_unlock(&cpu_offset_mtx);
4748 }
4749 
4750 int
4751 iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp)
4752 {
4753 	if_ctx_t ctx;
4754 	if_t ifp;
4755 	if_softc_ctx_t scctx;
4756 	kobjop_desc_t kobj_desc;
4757 	kobj_method_t *kobj_method;
4758 	int err, msix, rid;
4759 	int num_txd, num_rxd;
4760 
4761 	ctx = malloc(sizeof(* ctx), M_IFLIB, M_WAITOK|M_ZERO);
4762 
4763 	if (sc == NULL) {
4764 		sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4765 		device_set_softc(dev, ctx);
4766 		ctx->ifc_flags |= IFC_SC_ALLOCATED;
4767 	}
4768 
4769 	ctx->ifc_sctx = sctx;
4770 	ctx->ifc_dev = dev;
4771 	ctx->ifc_softc = sc;
4772 
4773 	if ((err = iflib_register(ctx)) != 0) {
4774 		device_printf(dev, "iflib_register failed %d\n", err);
4775 		goto fail_ctx_free;
4776 	}
4777 	iflib_add_device_sysctl_pre(ctx);
4778 
4779 	scctx = &ctx->ifc_softc_ctx;
4780 	ifp = ctx->ifc_ifp;
4781 
4782 	iflib_reset_qvalues(ctx);
4783 	CTX_LOCK(ctx);
4784 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4785 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4786 		goto fail_unlock;
4787 	}
4788 	_iflib_pre_assert(scctx);
4789 	ctx->ifc_txrx = *scctx->isc_txrx;
4790 
4791 	MPASS(scctx->isc_dma_width <= flsll(BUS_SPACE_MAXADDR));
4792 
4793 	if (sctx->isc_flags & IFLIB_DRIVER_MEDIA)
4794 		ctx->ifc_mediap = scctx->isc_media;
4795 
4796 #ifdef INVARIANTS
4797 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
4798 		MPASS(scctx->isc_tx_csum_flags);
4799 #endif
4800 
4801 	if_setcapabilities(ifp,
4802 	    scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_MEXTPG);
4803 	if_setcapenable(ifp,
4804 	    scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_MEXTPG);
4805 
4806 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4807 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4808 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4809 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4810 
4811 	num_txd = iflib_num_tx_descs(ctx);
4812 	num_rxd = iflib_num_rx_descs(ctx);
4813 
4814 	/* XXX change for per-queue sizes */
4815 	device_printf(dev, "Using %d TX descriptors and %d RX descriptors\n",
4816 	    num_txd, num_rxd);
4817 
4818 	if (scctx->isc_tx_nsegments > num_txd / MAX_SINGLE_PACKET_FRACTION)
4819 		scctx->isc_tx_nsegments = max(1, num_txd /
4820 		    MAX_SINGLE_PACKET_FRACTION);
4821 	if (scctx->isc_tx_tso_segments_max > num_txd /
4822 	    MAX_SINGLE_PACKET_FRACTION)
4823 		scctx->isc_tx_tso_segments_max = max(1,
4824 		    num_txd / MAX_SINGLE_PACKET_FRACTION);
4825 
4826 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4827 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
4828 		/*
4829 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
4830 		 * but some MACs do.
4831 		 */
4832 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
4833 		    IP_MAXPACKET));
4834 		/*
4835 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
4836 		 * into account.  In the worst case, each of these calls will
4837 		 * add another mbuf and, thus, the requirement for another DMA
4838 		 * segment.  So for best performance, it doesn't make sense to
4839 		 * advertize a maximum of TSO segments that typically will
4840 		 * require defragmentation in iflib_encap().
4841 		 */
4842 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
4843 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
4844 	}
4845 	if (scctx->isc_rss_table_size == 0)
4846 		scctx->isc_rss_table_size = 64;
4847 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4848 
4849 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4850 	/* XXX format name */
4851 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx,
4852 	    NULL, NULL, "admin");
4853 
4854 	/* Set up cpu set.  If it fails, use the set of all CPUs. */
4855 	if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) != 0) {
4856 		device_printf(dev, "Unable to fetch CPU list\n");
4857 		CPU_COPY(&all_cpus, &ctx->ifc_cpus);
4858 	}
4859 	MPASS(CPU_COUNT(&ctx->ifc_cpus) > 0);
4860 
4861 	/*
4862 	** Now set up MSI or MSI-X, should return us the number of supported
4863 	** vectors (will be 1 for a legacy interrupt and MSI).
4864 	*/
4865 	if (sctx->isc_flags & IFLIB_SKIP_MSIX) {
4866 		msix = scctx->isc_vectors;
4867 	} else if (scctx->isc_msix_bar != 0)
4868 	       /*
4869 		* The simple fact that isc_msix_bar is not 0 does not mean we
4870 		* we have a good value there that is known to work.
4871 		*/
4872 		msix = iflib_msix_init(ctx);
4873 	else {
4874 		scctx->isc_vectors = 1;
4875 		scctx->isc_ntxqsets = 1;
4876 		scctx->isc_nrxqsets = 1;
4877 		scctx->isc_intr = IFLIB_INTR_LEGACY;
4878 		msix = 0;
4879 	}
4880 	/* Get memory for the station queues */
4881 	if ((err = iflib_queues_alloc(ctx))) {
4882 		device_printf(dev, "Unable to allocate queue memory\n");
4883 		goto fail_intr_free;
4884 	}
4885 
4886 	if ((err = iflib_qset_structures_setup(ctx)))
4887 		goto fail_queues;
4888 
4889 	/*
4890 	 * Now that we know how many queues there are, get the core offset.
4891 	 */
4892 	ctx->ifc_sysctl_core_offset = get_ctx_core_offset(ctx);
4893 
4894 	if (msix > 1) {
4895 		/*
4896 		 * When using MSI-X, ensure that ifdi_{r,t}x_queue_intr_enable
4897 		 * aren't the default NULL implementation.
4898 		 */
4899 		kobj_desc = &ifdi_rx_queue_intr_enable_desc;
4900 		kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL,
4901 		    kobj_desc);
4902 		if (kobj_method == &kobj_desc->deflt) {
4903 			device_printf(dev,
4904 			    "MSI-X requires ifdi_rx_queue_intr_enable method");
4905 			err = EOPNOTSUPP;
4906 			goto fail_queues;
4907 		}
4908 		kobj_desc = &ifdi_tx_queue_intr_enable_desc;
4909 		kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL,
4910 		    kobj_desc);
4911 		if (kobj_method == &kobj_desc->deflt) {
4912 			device_printf(dev,
4913 			    "MSI-X requires ifdi_tx_queue_intr_enable method");
4914 			err = EOPNOTSUPP;
4915 			goto fail_queues;
4916 		}
4917 
4918 		/*
4919 		 * Assign the MSI-X vectors.
4920 		 * Note that the default NULL ifdi_msix_intr_assign method will
4921 		 * fail here, too.
4922 		 */
4923 		err = IFDI_MSIX_INTR_ASSIGN(ctx, msix);
4924 		if (err != 0) {
4925 			device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n",
4926 			    err);
4927 			goto fail_queues;
4928 		}
4929 	} else if (scctx->isc_intr != IFLIB_INTR_MSIX) {
4930 		rid = 0;
4931 		if (scctx->isc_intr == IFLIB_INTR_MSI) {
4932 			MPASS(msix == 1);
4933 			rid = 1;
4934 		}
4935 		if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) {
4936 			device_printf(dev, "iflib_legacy_setup failed %d\n", err);
4937 			goto fail_queues;
4938 		}
4939 	} else {
4940 		device_printf(dev,
4941 		    "Cannot use iflib with only 1 MSI-X interrupt!\n");
4942 		err = ENODEV;
4943 		goto fail_queues;
4944 	}
4945 
4946 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
4947 
4948 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4949 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4950 		goto fail_detach;
4951 	}
4952 
4953 	/*
4954 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4955 	 * This must appear after the call to ether_ifattach() because
4956 	 * ether_ifattach() sets if_hdrlen to the default value.
4957 	 */
4958 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4959 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
4960 
4961 	if ((err = iflib_netmap_attach(ctx))) {
4962 		device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err);
4963 		goto fail_detach;
4964 	}
4965 	*ctxp = ctx;
4966 
4967 	DEBUGNET_SET(ctx->ifc_ifp, iflib);
4968 
4969 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4970 	iflib_add_device_sysctl_post(ctx);
4971 	iflib_add_pfil(ctx);
4972 	ctx->ifc_flags |= IFC_INIT_DONE;
4973 	CTX_UNLOCK(ctx);
4974 
4975 	return (0);
4976 
4977 fail_detach:
4978 	ether_ifdetach(ctx->ifc_ifp);
4979 fail_queues:
4980 	iflib_tqg_detach(ctx);
4981 	iflib_tx_structures_free(ctx);
4982 	iflib_rx_structures_free(ctx);
4983 	IFDI_DETACH(ctx);
4984 	IFDI_QUEUES_FREE(ctx);
4985 fail_intr_free:
4986 	iflib_free_intr_mem(ctx);
4987 fail_unlock:
4988 	CTX_UNLOCK(ctx);
4989 	iflib_deregister(ctx);
4990 fail_ctx_free:
4991 	device_set_softc(ctx->ifc_dev, NULL);
4992         if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4993                 free(ctx->ifc_softc, M_IFLIB);
4994         free(ctx, M_IFLIB);
4995 	return (err);
4996 }
4997 
4998 int
4999 iflib_pseudo_register(device_t dev, if_shared_ctx_t sctx, if_ctx_t *ctxp,
5000 					  struct iflib_cloneattach_ctx *clctx)
5001 {
5002 	int num_txd, num_rxd;
5003 	int err;
5004 	if_ctx_t ctx;
5005 	if_t ifp;
5006 	if_softc_ctx_t scctx;
5007 	int i;
5008 	void *sc;
5009 
5010 	ctx = malloc(sizeof(*ctx), M_IFLIB, M_WAITOK|M_ZERO);
5011 	sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
5012 	ctx->ifc_flags |= IFC_SC_ALLOCATED;
5013 	if (sctx->isc_flags & (IFLIB_PSEUDO|IFLIB_VIRTUAL))
5014 		ctx->ifc_flags |= IFC_PSEUDO;
5015 
5016 	ctx->ifc_sctx = sctx;
5017 	ctx->ifc_softc = sc;
5018 	ctx->ifc_dev = dev;
5019 
5020 	if ((err = iflib_register(ctx)) != 0) {
5021 		device_printf(dev, "%s: iflib_register failed %d\n", __func__, err);
5022 		goto fail_ctx_free;
5023 	}
5024 	iflib_add_device_sysctl_pre(ctx);
5025 
5026 	scctx = &ctx->ifc_softc_ctx;
5027 	ifp = ctx->ifc_ifp;
5028 
5029 	iflib_reset_qvalues(ctx);
5030 	CTX_LOCK(ctx);
5031 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
5032 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
5033 		goto fail_unlock;
5034 	}
5035 	if (sctx->isc_flags & IFLIB_GEN_MAC)
5036 		ether_gen_addr(ifp, &ctx->ifc_mac);
5037 	if ((err = IFDI_CLONEATTACH(ctx, clctx->cc_ifc, clctx->cc_name,
5038 								clctx->cc_params)) != 0) {
5039 		device_printf(dev, "IFDI_CLONEATTACH failed %d\n", err);
5040 		goto fail_unlock;
5041 	}
5042 #ifdef INVARIANTS
5043 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
5044 		MPASS(scctx->isc_tx_csum_flags);
5045 #endif
5046 
5047 	if_setcapabilities(ifp, scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_LINKSTATE);
5048 	if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_LINKSTATE);
5049 
5050 	ifp->if_flags |= IFF_NOGROUP;
5051 	if (sctx->isc_flags & IFLIB_PSEUDO) {
5052 		ifmedia_add(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO, 0, NULL);
5053 		ifmedia_set(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO);
5054 		if (sctx->isc_flags & IFLIB_PSEUDO_ETHER) {
5055 			ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
5056 		} else {
5057 			if_attach(ctx->ifc_ifp);
5058 			bpfattach(ctx->ifc_ifp, DLT_NULL, sizeof(u_int32_t));
5059 		}
5060 
5061 		if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
5062 			device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
5063 			goto fail_detach;
5064 		}
5065 		*ctxp = ctx;
5066 
5067 		/*
5068 		 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
5069 		 * This must appear after the call to ether_ifattach() because
5070 		 * ether_ifattach() sets if_hdrlen to the default value.
5071 		 */
5072 		if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
5073 			if_setifheaderlen(ifp,
5074 			    sizeof(struct ether_vlan_header));
5075 
5076 		if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
5077 		iflib_add_device_sysctl_post(ctx);
5078 		ctx->ifc_flags |= IFC_INIT_DONE;
5079 		CTX_UNLOCK(ctx);
5080 		return (0);
5081 	}
5082 	ifmedia_add(ctx->ifc_mediap, IFM_ETHER | IFM_1000_T | IFM_FDX, 0, NULL);
5083 	ifmedia_add(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO, 0, NULL);
5084 	ifmedia_set(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO);
5085 
5086 	_iflib_pre_assert(scctx);
5087 	ctx->ifc_txrx = *scctx->isc_txrx;
5088 
5089 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
5090 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
5091 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
5092 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
5093 
5094 	num_txd = iflib_num_tx_descs(ctx);
5095 	num_rxd = iflib_num_rx_descs(ctx);
5096 
5097 	/* XXX change for per-queue sizes */
5098 	device_printf(dev, "Using %d TX descriptors and %d RX descriptors\n",
5099 	    num_txd, num_rxd);
5100 
5101 	if (scctx->isc_tx_nsegments > num_txd / MAX_SINGLE_PACKET_FRACTION)
5102 		scctx->isc_tx_nsegments = max(1, num_txd /
5103 		    MAX_SINGLE_PACKET_FRACTION);
5104 	if (scctx->isc_tx_tso_segments_max > num_txd /
5105 	    MAX_SINGLE_PACKET_FRACTION)
5106 		scctx->isc_tx_tso_segments_max = max(1,
5107 		    num_txd / MAX_SINGLE_PACKET_FRACTION);
5108 
5109 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
5110 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
5111 		/*
5112 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
5113 		 * but some MACs do.
5114 		 */
5115 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
5116 		    IP_MAXPACKET));
5117 		/*
5118 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
5119 		 * into account.  In the worst case, each of these calls will
5120 		 * add another mbuf and, thus, the requirement for another DMA
5121 		 * segment.  So for best performance, it doesn't make sense to
5122 		 * advertize a maximum of TSO segments that typically will
5123 		 * require defragmentation in iflib_encap().
5124 		 */
5125 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
5126 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
5127 	}
5128 	if (scctx->isc_rss_table_size == 0)
5129 		scctx->isc_rss_table_size = 64;
5130 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
5131 
5132 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
5133 	/* XXX format name */
5134 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx,
5135 	    NULL, NULL, "admin");
5136 
5137 	/* XXX --- can support > 1 -- but keep it simple for now */
5138 	scctx->isc_intr = IFLIB_INTR_LEGACY;
5139 
5140 	/* Get memory for the station queues */
5141 	if ((err = iflib_queues_alloc(ctx))) {
5142 		device_printf(dev, "Unable to allocate queue memory\n");
5143 		goto fail_iflib_detach;
5144 	}
5145 
5146 	if ((err = iflib_qset_structures_setup(ctx))) {
5147 		device_printf(dev, "qset structure setup failed %d\n", err);
5148 		goto fail_queues;
5149 	}
5150 
5151 	/*
5152 	 * XXX What if anything do we want to do about interrupts?
5153 	 */
5154 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
5155 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
5156 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
5157 		goto fail_detach;
5158 	}
5159 
5160 	/*
5161 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
5162 	 * This must appear after the call to ether_ifattach() because
5163 	 * ether_ifattach() sets if_hdrlen to the default value.
5164 	 */
5165 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
5166 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
5167 
5168 	/* XXX handle more than one queue */
5169 	for (i = 0; i < scctx->isc_nrxqsets; i++)
5170 		IFDI_RX_CLSET(ctx, 0, i, ctx->ifc_rxqs[i].ifr_fl[0].ifl_sds.ifsd_cl);
5171 
5172 	*ctxp = ctx;
5173 
5174 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
5175 	iflib_add_device_sysctl_post(ctx);
5176 	ctx->ifc_flags |= IFC_INIT_DONE;
5177 	CTX_UNLOCK(ctx);
5178 
5179 	return (0);
5180 fail_detach:
5181 	ether_ifdetach(ctx->ifc_ifp);
5182 fail_queues:
5183 	iflib_tqg_detach(ctx);
5184 	iflib_tx_structures_free(ctx);
5185 	iflib_rx_structures_free(ctx);
5186 fail_iflib_detach:
5187 	IFDI_DETACH(ctx);
5188 	IFDI_QUEUES_FREE(ctx);
5189 fail_unlock:
5190 	CTX_UNLOCK(ctx);
5191 	iflib_deregister(ctx);
5192 fail_ctx_free:
5193 	free(ctx->ifc_softc, M_IFLIB);
5194 	free(ctx, M_IFLIB);
5195 	return (err);
5196 }
5197 
5198 int
5199 iflib_pseudo_deregister(if_ctx_t ctx)
5200 {
5201 	if_t ifp = ctx->ifc_ifp;
5202 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5203 
5204 	/* Unregister VLAN event handlers early */
5205 	iflib_unregister_vlan_handlers(ctx);
5206 
5207 	if ((sctx->isc_flags & IFLIB_PSEUDO)  &&
5208 		(sctx->isc_flags & IFLIB_PSEUDO_ETHER) == 0) {
5209 		bpfdetach(ifp);
5210 		if_detach(ifp);
5211 	} else {
5212 		ether_ifdetach(ifp);
5213 	}
5214 
5215 	iflib_tqg_detach(ctx);
5216 	iflib_tx_structures_free(ctx);
5217 	iflib_rx_structures_free(ctx);
5218 	IFDI_DETACH(ctx);
5219 	IFDI_QUEUES_FREE(ctx);
5220 
5221 	iflib_deregister(ctx);
5222 
5223 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5224 		free(ctx->ifc_softc, M_IFLIB);
5225 	free(ctx, M_IFLIB);
5226 	return (0);
5227 }
5228 
5229 int
5230 iflib_device_attach(device_t dev)
5231 {
5232 	if_ctx_t ctx;
5233 	if_shared_ctx_t sctx;
5234 
5235 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
5236 		return (ENOTSUP);
5237 
5238 	pci_enable_busmaster(dev);
5239 
5240 	return (iflib_device_register(dev, NULL, sctx, &ctx));
5241 }
5242 
5243 int
5244 iflib_device_deregister(if_ctx_t ctx)
5245 {
5246 	if_t ifp = ctx->ifc_ifp;
5247 	device_t dev = ctx->ifc_dev;
5248 
5249 	/* Make sure VLANS are not using driver */
5250 	if (if_vlantrunkinuse(ifp)) {
5251 		device_printf(dev, "Vlan in use, detach first\n");
5252 		return (EBUSY);
5253 	}
5254 #ifdef PCI_IOV
5255 	if (!CTX_IS_VF(ctx) && pci_iov_detach(dev) != 0) {
5256 		device_printf(dev, "SR-IOV in use; detach first.\n");
5257 		return (EBUSY);
5258 	}
5259 #endif
5260 
5261 	STATE_LOCK(ctx);
5262 	ctx->ifc_flags |= IFC_IN_DETACH;
5263 	STATE_UNLOCK(ctx);
5264 
5265 	/* Unregister VLAN handlers before calling iflib_stop() */
5266 	iflib_unregister_vlan_handlers(ctx);
5267 
5268 	iflib_netmap_detach(ifp);
5269 	ether_ifdetach(ifp);
5270 
5271 	CTX_LOCK(ctx);
5272 	iflib_stop(ctx);
5273 	CTX_UNLOCK(ctx);
5274 
5275 	iflib_rem_pfil(ctx);
5276 	if (ctx->ifc_led_dev != NULL)
5277 		led_destroy(ctx->ifc_led_dev);
5278 
5279 	iflib_tqg_detach(ctx);
5280 	iflib_tx_structures_free(ctx);
5281 	iflib_rx_structures_free(ctx);
5282 
5283 	CTX_LOCK(ctx);
5284 	IFDI_DETACH(ctx);
5285 	IFDI_QUEUES_FREE(ctx);
5286 	CTX_UNLOCK(ctx);
5287 
5288 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
5289 	iflib_free_intr_mem(ctx);
5290 
5291 	bus_generic_detach(dev);
5292 
5293 	iflib_deregister(ctx);
5294 
5295 	device_set_softc(ctx->ifc_dev, NULL);
5296 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5297 		free(ctx->ifc_softc, M_IFLIB);
5298 	unref_ctx_core_offset(ctx);
5299 	free(ctx, M_IFLIB);
5300 	return (0);
5301 }
5302 
5303 static void
5304 iflib_tqg_detach(if_ctx_t ctx)
5305 {
5306 	iflib_txq_t txq;
5307 	iflib_rxq_t rxq;
5308 	int i;
5309 	struct taskqgroup *tqg;
5310 
5311 	/* XXX drain any dependent tasks */
5312 	tqg = qgroup_if_io_tqg;
5313 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
5314 		callout_drain(&txq->ift_timer);
5315 #ifdef DEV_NETMAP
5316 		callout_drain(&txq->ift_netmap_timer);
5317 #endif /* DEV_NETMAP */
5318 		if (txq->ift_task.gt_uniq != NULL)
5319 			taskqgroup_detach(tqg, &txq->ift_task);
5320 	}
5321 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
5322 		if (rxq->ifr_task.gt_uniq != NULL)
5323 			taskqgroup_detach(tqg, &rxq->ifr_task);
5324 	}
5325 	tqg = qgroup_if_config_tqg;
5326 	if (ctx->ifc_admin_task.gt_uniq != NULL)
5327 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
5328 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
5329 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
5330 }
5331 
5332 static void
5333 iflib_free_intr_mem(if_ctx_t ctx)
5334 {
5335 
5336 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) {
5337 		iflib_irq_free(ctx, &ctx->ifc_legacy_irq);
5338 	}
5339 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) {
5340 		pci_release_msi(ctx->ifc_dev);
5341 	}
5342 	if (ctx->ifc_msix_mem != NULL) {
5343 		bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY,
5344 		    rman_get_rid(ctx->ifc_msix_mem), ctx->ifc_msix_mem);
5345 		ctx->ifc_msix_mem = NULL;
5346 	}
5347 }
5348 
5349 int
5350 iflib_device_detach(device_t dev)
5351 {
5352 	if_ctx_t ctx = device_get_softc(dev);
5353 
5354 	return (iflib_device_deregister(ctx));
5355 }
5356 
5357 int
5358 iflib_device_suspend(device_t dev)
5359 {
5360 	if_ctx_t ctx = device_get_softc(dev);
5361 
5362 	CTX_LOCK(ctx);
5363 	IFDI_SUSPEND(ctx);
5364 	CTX_UNLOCK(ctx);
5365 
5366 	return bus_generic_suspend(dev);
5367 }
5368 int
5369 iflib_device_shutdown(device_t dev)
5370 {
5371 	if_ctx_t ctx = device_get_softc(dev);
5372 
5373 	CTX_LOCK(ctx);
5374 	IFDI_SHUTDOWN(ctx);
5375 	CTX_UNLOCK(ctx);
5376 
5377 	return bus_generic_suspend(dev);
5378 }
5379 
5380 int
5381 iflib_device_resume(device_t dev)
5382 {
5383 	if_ctx_t ctx = device_get_softc(dev);
5384 	iflib_txq_t txq = ctx->ifc_txqs;
5385 
5386 	CTX_LOCK(ctx);
5387 	IFDI_RESUME(ctx);
5388 	iflib_if_init_locked(ctx);
5389 	CTX_UNLOCK(ctx);
5390 	for (int i = 0; i < NTXQSETS(ctx); i++, txq++)
5391 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
5392 
5393 	return (bus_generic_resume(dev));
5394 }
5395 
5396 int
5397 iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params)
5398 {
5399 	int error;
5400 	if_ctx_t ctx = device_get_softc(dev);
5401 
5402 	CTX_LOCK(ctx);
5403 	error = IFDI_IOV_INIT(ctx, num_vfs, params);
5404 	CTX_UNLOCK(ctx);
5405 
5406 	return (error);
5407 }
5408 
5409 void
5410 iflib_device_iov_uninit(device_t dev)
5411 {
5412 	if_ctx_t ctx = device_get_softc(dev);
5413 
5414 	CTX_LOCK(ctx);
5415 	IFDI_IOV_UNINIT(ctx);
5416 	CTX_UNLOCK(ctx);
5417 }
5418 
5419 int
5420 iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params)
5421 {
5422 	int error;
5423 	if_ctx_t ctx = device_get_softc(dev);
5424 
5425 	CTX_LOCK(ctx);
5426 	error = IFDI_IOV_VF_ADD(ctx, vfnum, params);
5427 	CTX_UNLOCK(ctx);
5428 
5429 	return (error);
5430 }
5431 
5432 /*********************************************************************
5433  *
5434  *  MODULE FUNCTION DEFINITIONS
5435  *
5436  **********************************************************************/
5437 
5438 /*
5439  * - Start a fast taskqueue thread for each core
5440  * - Start a taskqueue for control operations
5441  */
5442 static int
5443 iflib_module_init(void)
5444 {
5445 	iflib_timer_default = hz / 2;
5446 	return (0);
5447 }
5448 
5449 static int
5450 iflib_module_event_handler(module_t mod, int what, void *arg)
5451 {
5452 	int err;
5453 
5454 	switch (what) {
5455 	case MOD_LOAD:
5456 		if ((err = iflib_module_init()) != 0)
5457 			return (err);
5458 		break;
5459 	case MOD_UNLOAD:
5460 		return (EBUSY);
5461 	default:
5462 		return (EOPNOTSUPP);
5463 	}
5464 
5465 	return (0);
5466 }
5467 
5468 /*********************************************************************
5469  *
5470  *  PUBLIC FUNCTION DEFINITIONS
5471  *     ordered as in iflib.h
5472  *
5473  **********************************************************************/
5474 
5475 static void
5476 _iflib_assert(if_shared_ctx_t sctx)
5477 {
5478 	int i;
5479 
5480 	MPASS(sctx->isc_tx_maxsize);
5481 	MPASS(sctx->isc_tx_maxsegsize);
5482 
5483 	MPASS(sctx->isc_rx_maxsize);
5484 	MPASS(sctx->isc_rx_nsegments);
5485 	MPASS(sctx->isc_rx_maxsegsize);
5486 
5487 	MPASS(sctx->isc_nrxqs >= 1 && sctx->isc_nrxqs <= 8);
5488 	for (i = 0; i < sctx->isc_nrxqs; i++) {
5489 		MPASS(sctx->isc_nrxd_min[i]);
5490 		MPASS(powerof2(sctx->isc_nrxd_min[i]));
5491 		MPASS(sctx->isc_nrxd_max[i]);
5492 		MPASS(powerof2(sctx->isc_nrxd_max[i]));
5493 		MPASS(sctx->isc_nrxd_default[i]);
5494 		MPASS(powerof2(sctx->isc_nrxd_default[i]));
5495 	}
5496 
5497 	MPASS(sctx->isc_ntxqs >= 1 && sctx->isc_ntxqs <= 8);
5498 	for (i = 0; i < sctx->isc_ntxqs; i++) {
5499 		MPASS(sctx->isc_ntxd_min[i]);
5500 		MPASS(powerof2(sctx->isc_ntxd_min[i]));
5501 		MPASS(sctx->isc_ntxd_max[i]);
5502 		MPASS(powerof2(sctx->isc_ntxd_max[i]));
5503 		MPASS(sctx->isc_ntxd_default[i]);
5504 		MPASS(powerof2(sctx->isc_ntxd_default[i]));
5505 	}
5506 }
5507 
5508 static void
5509 _iflib_pre_assert(if_softc_ctx_t scctx)
5510 {
5511 
5512 	MPASS(scctx->isc_txrx->ift_txd_encap);
5513 	MPASS(scctx->isc_txrx->ift_txd_flush);
5514 	MPASS(scctx->isc_txrx->ift_txd_credits_update);
5515 	MPASS(scctx->isc_txrx->ift_rxd_available);
5516 	MPASS(scctx->isc_txrx->ift_rxd_pkt_get);
5517 	MPASS(scctx->isc_txrx->ift_rxd_refill);
5518 	MPASS(scctx->isc_txrx->ift_rxd_flush);
5519 }
5520 
5521 static int
5522 iflib_register(if_ctx_t ctx)
5523 {
5524 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5525 	driver_t *driver = sctx->isc_driver;
5526 	device_t dev = ctx->ifc_dev;
5527 	if_t ifp;
5528 	u_char type;
5529 	int iflags;
5530 
5531 	if ((sctx->isc_flags & IFLIB_PSEUDO) == 0)
5532 		_iflib_assert(sctx);
5533 
5534 	CTX_LOCK_INIT(ctx);
5535 	STATE_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev));
5536 	if (sctx->isc_flags & IFLIB_PSEUDO) {
5537 		if (sctx->isc_flags & IFLIB_PSEUDO_ETHER)
5538 			type = IFT_ETHER;
5539 		else
5540 			type = IFT_PPP;
5541 	} else
5542 		type = IFT_ETHER;
5543 	ifp = ctx->ifc_ifp = if_alloc(type);
5544 	if (ifp == NULL) {
5545 		device_printf(dev, "can not allocate ifnet structure\n");
5546 		return (ENOMEM);
5547 	}
5548 
5549 	/*
5550 	 * Initialize our context's device specific methods
5551 	 */
5552 	kobj_init((kobj_t) ctx, (kobj_class_t) driver);
5553 	kobj_class_compile((kobj_class_t) driver);
5554 
5555 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
5556 	if_setsoftc(ifp, ctx);
5557 	if_setdev(ifp, dev);
5558 	if_setinitfn(ifp, iflib_if_init);
5559 	if_setioctlfn(ifp, iflib_if_ioctl);
5560 #ifdef ALTQ
5561 	if_setstartfn(ifp, iflib_altq_if_start);
5562 	if_settransmitfn(ifp, iflib_altq_if_transmit);
5563 	if_setsendqready(ifp);
5564 #else
5565 	if_settransmitfn(ifp, iflib_if_transmit);
5566 #endif
5567 	if_setqflushfn(ifp, iflib_if_qflush);
5568 	iflags = IFF_MULTICAST | IFF_KNOWSEPOCH;
5569 
5570 	if ((sctx->isc_flags & IFLIB_PSEUDO) &&
5571 		(sctx->isc_flags & IFLIB_PSEUDO_ETHER) == 0)
5572 		iflags |= IFF_POINTOPOINT;
5573 	else
5574 		iflags |= IFF_BROADCAST | IFF_SIMPLEX;
5575 	if_setflags(ifp, iflags);
5576 	ctx->ifc_vlan_attach_event =
5577 		EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx,
5578 							  EVENTHANDLER_PRI_FIRST);
5579 	ctx->ifc_vlan_detach_event =
5580 		EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx,
5581 							  EVENTHANDLER_PRI_FIRST);
5582 
5583 	if ((sctx->isc_flags & IFLIB_DRIVER_MEDIA) == 0) {
5584 		ctx->ifc_mediap = &ctx->ifc_media;
5585 		ifmedia_init(ctx->ifc_mediap, IFM_IMASK,
5586 		    iflib_media_change, iflib_media_status);
5587 	}
5588 	return (0);
5589 }
5590 
5591 static void
5592 iflib_unregister_vlan_handlers(if_ctx_t ctx)
5593 {
5594 	/* Unregister VLAN events */
5595 	if (ctx->ifc_vlan_attach_event != NULL) {
5596 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
5597 		ctx->ifc_vlan_attach_event = NULL;
5598 	}
5599 	if (ctx->ifc_vlan_detach_event != NULL) {
5600 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
5601 		ctx->ifc_vlan_detach_event = NULL;
5602 	}
5603 
5604 }
5605 
5606 static void
5607 iflib_deregister(if_ctx_t ctx)
5608 {
5609 	if_t ifp = ctx->ifc_ifp;
5610 
5611 	/* Remove all media */
5612 	ifmedia_removeall(&ctx->ifc_media);
5613 
5614 	/* Ensure that VLAN event handlers are unregistered */
5615 	iflib_unregister_vlan_handlers(ctx);
5616 
5617 	/* Release kobject reference */
5618 	kobj_delete((kobj_t) ctx, NULL);
5619 
5620 	/* Free the ifnet structure */
5621 	if_free(ifp);
5622 
5623 	STATE_LOCK_DESTROY(ctx);
5624 
5625 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
5626 	CTX_LOCK_DESTROY(ctx);
5627 }
5628 
5629 static int
5630 iflib_queues_alloc(if_ctx_t ctx)
5631 {
5632 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5633 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5634 	device_t dev = ctx->ifc_dev;
5635 	int nrxqsets = scctx->isc_nrxqsets;
5636 	int ntxqsets = scctx->isc_ntxqsets;
5637 	iflib_txq_t txq;
5638 	iflib_rxq_t rxq;
5639 	iflib_fl_t fl = NULL;
5640 	int i, j, cpu, err, txconf, rxconf;
5641 	iflib_dma_info_t ifdip;
5642 	uint32_t *rxqsizes = scctx->isc_rxqsizes;
5643 	uint32_t *txqsizes = scctx->isc_txqsizes;
5644 	uint8_t nrxqs = sctx->isc_nrxqs;
5645 	uint8_t ntxqs = sctx->isc_ntxqs;
5646 	int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1;
5647 	int fl_offset = (sctx->isc_flags & IFLIB_HAS_RXCQ ? 1 : 0);
5648 	caddr_t *vaddrs;
5649 	uint64_t *paddrs;
5650 
5651 	KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1"));
5652 	KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1"));
5653 	KASSERT(nrxqs >= fl_offset + nfree_lists,
5654            ("there must be at least a rxq for each free list"));
5655 
5656 	/* Allocate the TX ring struct memory */
5657 	if (!(ctx->ifc_txqs =
5658 	    (iflib_txq_t) malloc(sizeof(struct iflib_txq) *
5659 	    ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5660 		device_printf(dev, "Unable to allocate TX ring memory\n");
5661 		err = ENOMEM;
5662 		goto fail;
5663 	}
5664 
5665 	/* Now allocate the RX */
5666 	if (!(ctx->ifc_rxqs =
5667 	    (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) *
5668 	    nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5669 		device_printf(dev, "Unable to allocate RX ring memory\n");
5670 		err = ENOMEM;
5671 		goto rx_fail;
5672 	}
5673 
5674 	txq = ctx->ifc_txqs;
5675 	rxq = ctx->ifc_rxqs;
5676 
5677 	/*
5678 	 * XXX handle allocation failure
5679 	 */
5680 	for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) {
5681 		/* Set up some basics */
5682 
5683 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs,
5684 		    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
5685 			device_printf(dev,
5686 			    "Unable to allocate TX DMA info memory\n");
5687 			err = ENOMEM;
5688 			goto err_tx_desc;
5689 		}
5690 		txq->ift_ifdi = ifdip;
5691 		for (j = 0; j < ntxqs; j++, ifdip++) {
5692 			if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, 0)) {
5693 				device_printf(dev,
5694 				    "Unable to allocate TX descriptors\n");
5695 				err = ENOMEM;
5696 				goto err_tx_desc;
5697 			}
5698 			txq->ift_txd_size[j] = scctx->isc_txd_size[j];
5699 			bzero((void *)ifdip->idi_vaddr, txqsizes[j]);
5700 		}
5701 		txq->ift_ctx = ctx;
5702 		txq->ift_id = i;
5703 		if (sctx->isc_flags & IFLIB_HAS_TXCQ) {
5704 			txq->ift_br_offset = 1;
5705 		} else {
5706 			txq->ift_br_offset = 0;
5707 		}
5708 
5709 		if (iflib_txsd_alloc(txq)) {
5710 			device_printf(dev, "Critical Failure setting up TX buffers\n");
5711 			err = ENOMEM;
5712 			goto err_tx_desc;
5713 		}
5714 
5715 		/* Initialize the TX lock */
5716 		snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:TX(%d):callout",
5717 		    device_get_nameunit(dev), txq->ift_id);
5718 		mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF);
5719 		callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0);
5720 		txq->ift_timer.c_cpu = cpu;
5721 #ifdef DEV_NETMAP
5722 		callout_init_mtx(&txq->ift_netmap_timer, &txq->ift_mtx, 0);
5723 		txq->ift_netmap_timer.c_cpu = cpu;
5724 #endif /* DEV_NETMAP */
5725 
5726 		err = ifmp_ring_alloc(&txq->ift_br, 2048, txq, iflib_txq_drain,
5727 				      iflib_txq_can_drain, M_IFLIB, M_WAITOK);
5728 		if (err) {
5729 			/* XXX free any allocated rings */
5730 			device_printf(dev, "Unable to allocate buf_ring\n");
5731 			goto err_tx_desc;
5732 		}
5733 	}
5734 
5735 	for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) {
5736 		/* Set up some basics */
5737 		callout_init(&rxq->ifr_watchdog, 1);
5738 
5739 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs,
5740 		   M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
5741 			device_printf(dev,
5742 			    "Unable to allocate RX DMA info memory\n");
5743 			err = ENOMEM;
5744 			goto err_tx_desc;
5745 		}
5746 
5747 		rxq->ifr_ifdi = ifdip;
5748 		/* XXX this needs to be changed if #rx queues != #tx queues */
5749 		rxq->ifr_ntxqirq = 1;
5750 		rxq->ifr_txqid[0] = i;
5751 		for (j = 0; j < nrxqs; j++, ifdip++) {
5752 			if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, 0)) {
5753 				device_printf(dev,
5754 				    "Unable to allocate RX descriptors\n");
5755 				err = ENOMEM;
5756 				goto err_tx_desc;
5757 			}
5758 			bzero((void *)ifdip->idi_vaddr, rxqsizes[j]);
5759 		}
5760 		rxq->ifr_ctx = ctx;
5761 		rxq->ifr_id = i;
5762 		rxq->ifr_fl_offset = fl_offset;
5763 		rxq->ifr_nfl = nfree_lists;
5764 		if (!(fl =
5765 			  (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) {
5766 			device_printf(dev, "Unable to allocate free list memory\n");
5767 			err = ENOMEM;
5768 			goto err_tx_desc;
5769 		}
5770 		rxq->ifr_fl = fl;
5771 		for (j = 0; j < nfree_lists; j++) {
5772 			fl[j].ifl_rxq = rxq;
5773 			fl[j].ifl_id = j;
5774 			fl[j].ifl_ifdi = &rxq->ifr_ifdi[j + rxq->ifr_fl_offset];
5775 			fl[j].ifl_rxd_size = scctx->isc_rxd_size[j];
5776 		}
5777 		/* Allocate receive buffers for the ring */
5778 		if (iflib_rxsd_alloc(rxq)) {
5779 			device_printf(dev,
5780 			    "Critical Failure setting up receive buffers\n");
5781 			err = ENOMEM;
5782 			goto err_rx_desc;
5783 		}
5784 
5785 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5786 			fl->ifl_rx_bitmap = bit_alloc(fl->ifl_size, M_IFLIB,
5787 			    M_WAITOK);
5788 	}
5789 
5790 	/* TXQs */
5791 	vaddrs = malloc(sizeof(caddr_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5792 	paddrs = malloc(sizeof(uint64_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5793 	for (i = 0; i < ntxqsets; i++) {
5794 		iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi;
5795 
5796 		for (j = 0; j < ntxqs; j++, di++) {
5797 			vaddrs[i*ntxqs + j] = di->idi_vaddr;
5798 			paddrs[i*ntxqs + j] = di->idi_paddr;
5799 		}
5800 	}
5801 	if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) {
5802 		device_printf(ctx->ifc_dev,
5803 		    "Unable to allocate device TX queue\n");
5804 		iflib_tx_structures_free(ctx);
5805 		free(vaddrs, M_IFLIB);
5806 		free(paddrs, M_IFLIB);
5807 		goto err_rx_desc;
5808 	}
5809 	free(vaddrs, M_IFLIB);
5810 	free(paddrs, M_IFLIB);
5811 
5812 	/* RXQs */
5813 	vaddrs = malloc(sizeof(caddr_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5814 	paddrs = malloc(sizeof(uint64_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5815 	for (i = 0; i < nrxqsets; i++) {
5816 		iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi;
5817 
5818 		for (j = 0; j < nrxqs; j++, di++) {
5819 			vaddrs[i*nrxqs + j] = di->idi_vaddr;
5820 			paddrs[i*nrxqs + j] = di->idi_paddr;
5821 		}
5822 	}
5823 	if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) {
5824 		device_printf(ctx->ifc_dev,
5825 		    "Unable to allocate device RX queue\n");
5826 		iflib_tx_structures_free(ctx);
5827 		free(vaddrs, M_IFLIB);
5828 		free(paddrs, M_IFLIB);
5829 		goto err_rx_desc;
5830 	}
5831 	free(vaddrs, M_IFLIB);
5832 	free(paddrs, M_IFLIB);
5833 
5834 	return (0);
5835 
5836 /* XXX handle allocation failure changes */
5837 err_rx_desc:
5838 err_tx_desc:
5839 rx_fail:
5840 	if (ctx->ifc_rxqs != NULL)
5841 		free(ctx->ifc_rxqs, M_IFLIB);
5842 	ctx->ifc_rxqs = NULL;
5843 	if (ctx->ifc_txqs != NULL)
5844 		free(ctx->ifc_txqs, M_IFLIB);
5845 	ctx->ifc_txqs = NULL;
5846 fail:
5847 	return (err);
5848 }
5849 
5850 static int
5851 iflib_tx_structures_setup(if_ctx_t ctx)
5852 {
5853 	iflib_txq_t txq = ctx->ifc_txqs;
5854 	int i;
5855 
5856 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
5857 		iflib_txq_setup(txq);
5858 
5859 	return (0);
5860 }
5861 
5862 static void
5863 iflib_tx_structures_free(if_ctx_t ctx)
5864 {
5865 	iflib_txq_t txq = ctx->ifc_txqs;
5866 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5867 	int i, j;
5868 
5869 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
5870 		for (j = 0; j < sctx->isc_ntxqs; j++)
5871 			iflib_dma_free(&txq->ift_ifdi[j]);
5872 		iflib_txq_destroy(txq);
5873 	}
5874 	free(ctx->ifc_txqs, M_IFLIB);
5875 	ctx->ifc_txqs = NULL;
5876 }
5877 
5878 /*********************************************************************
5879  *
5880  *  Initialize all receive rings.
5881  *
5882  **********************************************************************/
5883 static int
5884 iflib_rx_structures_setup(if_ctx_t ctx)
5885 {
5886 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5887 	int q;
5888 #if defined(INET6) || defined(INET)
5889 	int err, i;
5890 #endif
5891 
5892 	for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) {
5893 #if defined(INET6) || defined(INET)
5894 		if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_LRO) {
5895 			err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp,
5896 			    TCP_LRO_ENTRIES, min(1024,
5897 			    ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset]));
5898 			if (err != 0) {
5899 				device_printf(ctx->ifc_dev,
5900 				    "LRO Initialization failed!\n");
5901 				goto fail;
5902 			}
5903 		}
5904 #endif
5905 		IFDI_RXQ_SETUP(ctx, rxq->ifr_id);
5906 	}
5907 	return (0);
5908 #if defined(INET6) || defined(INET)
5909 fail:
5910 	/*
5911 	 * Free LRO resources allocated so far, we will only handle
5912 	 * the rings that completed, the failing case will have
5913 	 * cleaned up for itself.  'q' failed, so its the terminus.
5914 	 */
5915 	rxq = ctx->ifc_rxqs;
5916 	for (i = 0; i < q; ++i, rxq++) {
5917 		if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_LRO)
5918 			tcp_lro_free(&rxq->ifr_lc);
5919 	}
5920 	return (err);
5921 #endif
5922 }
5923 
5924 /*********************************************************************
5925  *
5926  *  Free all receive rings.
5927  *
5928  **********************************************************************/
5929 static void
5930 iflib_rx_structures_free(if_ctx_t ctx)
5931 {
5932 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5933 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5934 	int i, j;
5935 
5936 	for (i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) {
5937 		for (j = 0; j < sctx->isc_nrxqs; j++)
5938 			iflib_dma_free(&rxq->ifr_ifdi[j]);
5939 		iflib_rx_sds_free(rxq);
5940 #if defined(INET6) || defined(INET)
5941 		if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_LRO)
5942 			tcp_lro_free(&rxq->ifr_lc);
5943 #endif
5944 	}
5945 	free(ctx->ifc_rxqs, M_IFLIB);
5946 	ctx->ifc_rxqs = NULL;
5947 }
5948 
5949 static int
5950 iflib_qset_structures_setup(if_ctx_t ctx)
5951 {
5952 	int err;
5953 
5954 	/*
5955 	 * It is expected that the caller takes care of freeing queues if this
5956 	 * fails.
5957 	 */
5958 	if ((err = iflib_tx_structures_setup(ctx)) != 0) {
5959 		device_printf(ctx->ifc_dev, "iflib_tx_structures_setup failed: %d\n", err);
5960 		return (err);
5961 	}
5962 
5963 	if ((err = iflib_rx_structures_setup(ctx)) != 0)
5964 		device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err);
5965 
5966 	return (err);
5967 }
5968 
5969 int
5970 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
5971 		driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, const char *name)
5972 {
5973 
5974 	return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name));
5975 }
5976 
5977 #ifdef SMP
5978 static int
5979 find_nth(if_ctx_t ctx, int qid)
5980 {
5981 	cpuset_t cpus;
5982 	int i, cpuid, eqid, count;
5983 
5984 	CPU_COPY(&ctx->ifc_cpus, &cpus);
5985 	count = CPU_COUNT(&cpus);
5986 	eqid = qid % count;
5987 	/* clear up to the qid'th bit */
5988 	for (i = 0; i < eqid; i++) {
5989 		cpuid = CPU_FFS(&cpus);
5990 		MPASS(cpuid != 0);
5991 		CPU_CLR(cpuid-1, &cpus);
5992 	}
5993 	cpuid = CPU_FFS(&cpus);
5994 	MPASS(cpuid != 0);
5995 	return (cpuid-1);
5996 }
5997 
5998 #ifdef SCHED_ULE
5999 extern struct cpu_group *cpu_top;              /* CPU topology */
6000 
6001 static int
6002 find_child_with_core(int cpu, struct cpu_group *grp)
6003 {
6004 	int i;
6005 
6006 	if (grp->cg_children == 0)
6007 		return -1;
6008 
6009 	MPASS(grp->cg_child);
6010 	for (i = 0; i < grp->cg_children; i++) {
6011 		if (CPU_ISSET(cpu, &grp->cg_child[i].cg_mask))
6012 			return i;
6013 	}
6014 
6015 	return -1;
6016 }
6017 
6018 /*
6019  * Find the nth "close" core to the specified core
6020  * "close" is defined as the deepest level that shares
6021  * at least an L2 cache.  With threads, this will be
6022  * threads on the same core.  If the shared cache is L3
6023  * or higher, simply returns the same core.
6024  */
6025 static int
6026 find_close_core(int cpu, int core_offset)
6027 {
6028 	struct cpu_group *grp;
6029 	int i;
6030 	int fcpu;
6031 	cpuset_t cs;
6032 
6033 	grp = cpu_top;
6034 	if (grp == NULL)
6035 		return cpu;
6036 	i = 0;
6037 	while ((i = find_child_with_core(cpu, grp)) != -1) {
6038 		/* If the child only has one cpu, don't descend */
6039 		if (grp->cg_child[i].cg_count <= 1)
6040 			break;
6041 		grp = &grp->cg_child[i];
6042 	}
6043 
6044 	/* If they don't share at least an L2 cache, use the same CPU */
6045 	if (grp->cg_level > CG_SHARE_L2 || grp->cg_level == CG_SHARE_NONE)
6046 		return cpu;
6047 
6048 	/* Now pick one */
6049 	CPU_COPY(&grp->cg_mask, &cs);
6050 
6051 	/* Add the selected CPU offset to core offset. */
6052 	for (i = 0; (fcpu = CPU_FFS(&cs)) != 0; i++) {
6053 		if (fcpu - 1 == cpu)
6054 			break;
6055 		CPU_CLR(fcpu - 1, &cs);
6056 	}
6057 	MPASS(fcpu);
6058 
6059 	core_offset += i;
6060 
6061 	CPU_COPY(&grp->cg_mask, &cs);
6062 	for (i = core_offset % grp->cg_count; i > 0; i--) {
6063 		MPASS(CPU_FFS(&cs));
6064 		CPU_CLR(CPU_FFS(&cs) - 1, &cs);
6065 	}
6066 	MPASS(CPU_FFS(&cs));
6067 	return CPU_FFS(&cs) - 1;
6068 }
6069 #else
6070 static int
6071 find_close_core(int cpu, int core_offset __unused)
6072 {
6073 	return cpu;
6074 }
6075 #endif
6076 
6077 static int
6078 get_core_offset(if_ctx_t ctx, iflib_intr_type_t type, int qid)
6079 {
6080 	switch (type) {
6081 	case IFLIB_INTR_TX:
6082 		/* TX queues get cores which share at least an L2 cache with the corresponding RX queue */
6083 		/* XXX handle multiple RX threads per core and more than two core per L2 group */
6084 		return qid / CPU_COUNT(&ctx->ifc_cpus) + 1;
6085 	case IFLIB_INTR_RX:
6086 	case IFLIB_INTR_RXTX:
6087 		/* RX queues get the specified core */
6088 		return qid / CPU_COUNT(&ctx->ifc_cpus);
6089 	default:
6090 		return -1;
6091 	}
6092 }
6093 #else
6094 #define get_core_offset(ctx, type, qid)	CPU_FIRST()
6095 #define find_close_core(cpuid, tid)	CPU_FIRST()
6096 #define find_nth(ctx, gid)		CPU_FIRST()
6097 #endif
6098 
6099 /* Just to avoid copy/paste */
6100 static inline int
6101 iflib_irq_set_affinity(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type,
6102     int qid, struct grouptask *gtask, struct taskqgroup *tqg, void *uniq,
6103     const char *name)
6104 {
6105 	device_t dev;
6106 	int co, cpuid, err, tid;
6107 
6108 	dev = ctx->ifc_dev;
6109 	co = ctx->ifc_sysctl_core_offset;
6110 	if (ctx->ifc_sysctl_separate_txrx && type == IFLIB_INTR_TX)
6111 		co += ctx->ifc_softc_ctx.isc_nrxqsets;
6112 	cpuid = find_nth(ctx, qid + co);
6113 	tid = get_core_offset(ctx, type, qid);
6114 	if (tid < 0) {
6115 		device_printf(dev, "get_core_offset failed\n");
6116 		return (EOPNOTSUPP);
6117 	}
6118 	cpuid = find_close_core(cpuid, tid);
6119 	err = taskqgroup_attach_cpu(tqg, gtask, uniq, cpuid, dev, irq->ii_res,
6120 	    name);
6121 	if (err) {
6122 		device_printf(dev, "taskqgroup_attach_cpu failed %d\n", err);
6123 		return (err);
6124 	}
6125 #ifdef notyet
6126 	if (cpuid > ctx->ifc_cpuid_highest)
6127 		ctx->ifc_cpuid_highest = cpuid;
6128 #endif
6129 	return (0);
6130 }
6131 
6132 int
6133 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid,
6134 			iflib_intr_type_t type, driver_filter_t *filter,
6135 			void *filter_arg, int qid, const char *name)
6136 {
6137 	device_t dev;
6138 	struct grouptask *gtask;
6139 	struct taskqgroup *tqg;
6140 	iflib_filter_info_t info;
6141 	gtask_fn_t *fn;
6142 	int tqrid, err;
6143 	driver_filter_t *intr_fast;
6144 	void *q;
6145 
6146 	info = &ctx->ifc_filter_info;
6147 	tqrid = rid;
6148 
6149 	switch (type) {
6150 	/* XXX merge tx/rx for netmap? */
6151 	case IFLIB_INTR_TX:
6152 		q = &ctx->ifc_txqs[qid];
6153 		info = &ctx->ifc_txqs[qid].ift_filter_info;
6154 		gtask = &ctx->ifc_txqs[qid].ift_task;
6155 		tqg = qgroup_if_io_tqg;
6156 		fn = _task_fn_tx;
6157 		intr_fast = iflib_fast_intr;
6158 		GROUPTASK_INIT(gtask, 0, fn, q);
6159 		ctx->ifc_flags |= IFC_NETMAP_TX_IRQ;
6160 		break;
6161 	case IFLIB_INTR_RX:
6162 		q = &ctx->ifc_rxqs[qid];
6163 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
6164 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6165 		tqg = qgroup_if_io_tqg;
6166 		fn = _task_fn_rx;
6167 		intr_fast = iflib_fast_intr;
6168 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6169 		break;
6170 	case IFLIB_INTR_RXTX:
6171 		q = &ctx->ifc_rxqs[qid];
6172 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
6173 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6174 		tqg = qgroup_if_io_tqg;
6175 		fn = _task_fn_rx;
6176 		intr_fast = iflib_fast_intr_rxtx;
6177 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6178 		break;
6179 	case IFLIB_INTR_ADMIN:
6180 		q = ctx;
6181 		tqrid = -1;
6182 		info = &ctx->ifc_filter_info;
6183 		gtask = &ctx->ifc_admin_task;
6184 		tqg = qgroup_if_config_tqg;
6185 		fn = _task_fn_admin;
6186 		intr_fast = iflib_fast_intr_ctx;
6187 		break;
6188 	default:
6189 		device_printf(ctx->ifc_dev, "%s: unknown net intr type\n",
6190 		    __func__);
6191 		return (EINVAL);
6192 	}
6193 
6194 	info->ifi_filter = filter;
6195 	info->ifi_filter_arg = filter_arg;
6196 	info->ifi_task = gtask;
6197 	info->ifi_ctx = q;
6198 
6199 	dev = ctx->ifc_dev;
6200 	err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info,  name);
6201 	if (err != 0) {
6202 		device_printf(dev, "_iflib_irq_alloc failed %d\n", err);
6203 		return (err);
6204 	}
6205 	if (type == IFLIB_INTR_ADMIN)
6206 		return (0);
6207 
6208 	if (tqrid != -1) {
6209 		err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg,
6210 		    q, name);
6211 		if (err)
6212 			return (err);
6213 	} else {
6214 		taskqgroup_attach(tqg, gtask, q, dev, irq->ii_res, name);
6215 	}
6216 
6217 	return (0);
6218 }
6219 
6220 void
6221 iflib_softirq_alloc_generic(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type, void *arg, int qid, const char *name)
6222 {
6223 	struct grouptask *gtask;
6224 	struct taskqgroup *tqg;
6225 	gtask_fn_t *fn;
6226 	void *q;
6227 	int err;
6228 
6229 	switch (type) {
6230 	case IFLIB_INTR_TX:
6231 		q = &ctx->ifc_txqs[qid];
6232 		gtask = &ctx->ifc_txqs[qid].ift_task;
6233 		tqg = qgroup_if_io_tqg;
6234 		fn = _task_fn_tx;
6235 		GROUPTASK_INIT(gtask, 0, fn, q);
6236 		break;
6237 	case IFLIB_INTR_RX:
6238 		q = &ctx->ifc_rxqs[qid];
6239 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6240 		tqg = qgroup_if_io_tqg;
6241 		fn = _task_fn_rx;
6242 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6243 		break;
6244 	case IFLIB_INTR_IOV:
6245 		q = ctx;
6246 		gtask = &ctx->ifc_vflr_task;
6247 		tqg = qgroup_if_config_tqg;
6248 		fn = _task_fn_iov;
6249 		GROUPTASK_INIT(gtask, 0, fn, q);
6250 		break;
6251 	default:
6252 		panic("unknown net intr type");
6253 	}
6254 	if (irq != NULL) {
6255 		err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg,
6256 		    q, name);
6257 		if (err)
6258 			taskqgroup_attach(tqg, gtask, q, ctx->ifc_dev,
6259 			    irq->ii_res, name);
6260 	} else {
6261 		taskqgroup_attach(tqg, gtask, q, NULL, NULL, name);
6262 	}
6263 }
6264 
6265 void
6266 iflib_irq_free(if_ctx_t ctx, if_irq_t irq)
6267 {
6268 
6269 	if (irq->ii_tag)
6270 		bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag);
6271 
6272 	if (irq->ii_res)
6273 		bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ,
6274 		    rman_get_rid(irq->ii_res), irq->ii_res);
6275 }
6276 
6277 static int
6278 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, const char *name)
6279 {
6280 	iflib_txq_t txq = ctx->ifc_txqs;
6281 	iflib_rxq_t rxq = ctx->ifc_rxqs;
6282 	if_irq_t irq = &ctx->ifc_legacy_irq;
6283 	iflib_filter_info_t info;
6284 	device_t dev;
6285 	struct grouptask *gtask;
6286 	struct resource *res;
6287 	struct taskqgroup *tqg;
6288 	void *q;
6289 	int err, tqrid;
6290 	bool rx_only;
6291 
6292 	q = &ctx->ifc_rxqs[0];
6293 	info = &rxq[0].ifr_filter_info;
6294 	gtask = &rxq[0].ifr_task;
6295 	tqg = qgroup_if_io_tqg;
6296 	tqrid = *rid;
6297 	rx_only = (ctx->ifc_sctx->isc_flags & IFLIB_SINGLE_IRQ_RX_ONLY) != 0;
6298 
6299 	ctx->ifc_flags |= IFC_LEGACY;
6300 	info->ifi_filter = filter;
6301 	info->ifi_filter_arg = filter_arg;
6302 	info->ifi_task = gtask;
6303 	info->ifi_ctx = rx_only ? ctx : q;
6304 
6305 	dev = ctx->ifc_dev;
6306 	/* We allocate a single interrupt resource */
6307 	err = _iflib_irq_alloc(ctx, irq, tqrid, rx_only ? iflib_fast_intr_ctx :
6308 	    iflib_fast_intr_rxtx, NULL, info, name);
6309 	if (err != 0)
6310 		return (err);
6311 	NET_GROUPTASK_INIT(gtask, 0, _task_fn_rx, q);
6312 	res = irq->ii_res;
6313 	taskqgroup_attach(tqg, gtask, q, dev, res, name);
6314 
6315 	GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq);
6316 	taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, dev, res,
6317 	    "tx");
6318 	return (0);
6319 }
6320 
6321 void
6322 iflib_led_create(if_ctx_t ctx)
6323 {
6324 
6325 	ctx->ifc_led_dev = led_create(iflib_led_func, ctx,
6326 	    device_get_nameunit(ctx->ifc_dev));
6327 }
6328 
6329 void
6330 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid)
6331 {
6332 
6333 	GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
6334 }
6335 
6336 void
6337 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid)
6338 {
6339 
6340 	GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task);
6341 }
6342 
6343 void
6344 iflib_admin_intr_deferred(if_ctx_t ctx)
6345 {
6346 
6347 	MPASS(ctx->ifc_admin_task.gt_taskqueue != NULL);
6348 	GROUPTASK_ENQUEUE(&ctx->ifc_admin_task);
6349 }
6350 
6351 void
6352 iflib_iov_intr_deferred(if_ctx_t ctx)
6353 {
6354 
6355 	GROUPTASK_ENQUEUE(&ctx->ifc_vflr_task);
6356 }
6357 
6358 void
6359 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, const char *name)
6360 {
6361 
6362 	taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, NULL, NULL,
6363 	    name);
6364 }
6365 
6366 void
6367 iflib_config_gtask_init(void *ctx, struct grouptask *gtask, gtask_fn_t *fn,
6368 	const char *name)
6369 {
6370 
6371 	GROUPTASK_INIT(gtask, 0, fn, ctx);
6372 	taskqgroup_attach(qgroup_if_config_tqg, gtask, gtask, NULL, NULL,
6373 	    name);
6374 }
6375 
6376 void
6377 iflib_config_gtask_deinit(struct grouptask *gtask)
6378 {
6379 
6380 	taskqgroup_detach(qgroup_if_config_tqg, gtask);
6381 }
6382 
6383 void
6384 iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate)
6385 {
6386 	if_t ifp = ctx->ifc_ifp;
6387 	iflib_txq_t txq = ctx->ifc_txqs;
6388 
6389 	if_setbaudrate(ifp, baudrate);
6390 	if (baudrate >= IF_Gbps(10)) {
6391 		STATE_LOCK(ctx);
6392 		ctx->ifc_flags |= IFC_PREFETCH;
6393 		STATE_UNLOCK(ctx);
6394 	}
6395 	/* If link down, disable watchdog */
6396 	if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) {
6397 		for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++)
6398 			txq->ift_qstatus = IFLIB_QUEUE_IDLE;
6399 	}
6400 	ctx->ifc_link_state = link_state;
6401 	if_link_state_change(ifp, link_state);
6402 }
6403 
6404 static int
6405 iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq)
6406 {
6407 	int credits;
6408 #ifdef INVARIANTS
6409 	int credits_pre = txq->ift_cidx_processed;
6410 #endif
6411 
6412 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
6413 	    BUS_DMASYNC_POSTREAD);
6414 	if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, true)) == 0)
6415 		return (0);
6416 
6417 	txq->ift_processed += credits;
6418 	txq->ift_cidx_processed += credits;
6419 
6420 	MPASS(credits_pre + credits == txq->ift_cidx_processed);
6421 	if (txq->ift_cidx_processed >= txq->ift_size)
6422 		txq->ift_cidx_processed -= txq->ift_size;
6423 	return (credits);
6424 }
6425 
6426 static int
6427 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget)
6428 {
6429 	iflib_fl_t fl;
6430 	u_int i;
6431 
6432 	for (i = 0, fl = &rxq->ifr_fl[0]; i < rxq->ifr_nfl; i++, fl++)
6433 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
6434 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
6435 	return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx,
6436 	    budget));
6437 }
6438 
6439 void
6440 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name,
6441 	const char *description, if_int_delay_info_t info,
6442 	int offset, int value)
6443 {
6444 	info->iidi_ctx = ctx;
6445 	info->iidi_offset = offset;
6446 	info->iidi_value = value;
6447 	SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev),
6448 	    SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)),
6449 	    OID_AUTO, name, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE,
6450 	    info, 0, iflib_sysctl_int_delay, "I", description);
6451 }
6452 
6453 struct sx *
6454 iflib_ctx_lock_get(if_ctx_t ctx)
6455 {
6456 
6457 	return (&ctx->ifc_ctx_sx);
6458 }
6459 
6460 static int
6461 iflib_msix_init(if_ctx_t ctx)
6462 {
6463 	device_t dev = ctx->ifc_dev;
6464 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6465 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6466 	int admincnt, bar, err, iflib_num_rx_queues, iflib_num_tx_queues;
6467 	int msgs, queuemsgs, queues, rx_queues, tx_queues, vectors;
6468 
6469 	iflib_num_tx_queues = ctx->ifc_sysctl_ntxqs;
6470 	iflib_num_rx_queues = ctx->ifc_sysctl_nrxqs;
6471 
6472 	if (bootverbose)
6473 		device_printf(dev, "msix_init qsets capped at %d\n",
6474 		    imax(scctx->isc_ntxqsets, scctx->isc_nrxqsets));
6475 
6476 	/* Override by tuneable */
6477 	if (scctx->isc_disable_msix)
6478 		goto msi;
6479 
6480 	/* First try MSI-X */
6481 	if ((msgs = pci_msix_count(dev)) == 0) {
6482 		if (bootverbose)
6483 			device_printf(dev, "MSI-X not supported or disabled\n");
6484 		goto msi;
6485 	}
6486 
6487 	bar = ctx->ifc_softc_ctx.isc_msix_bar;
6488 	/*
6489 	 * bar == -1 => "trust me I know what I'm doing"
6490 	 * Some drivers are for hardware that is so shoddily
6491 	 * documented that no one knows which bars are which
6492 	 * so the developer has to map all bars. This hack
6493 	 * allows shoddy garbage to use MSI-X in this framework.
6494 	 */
6495 	if (bar != -1) {
6496 		ctx->ifc_msix_mem = bus_alloc_resource_any(dev,
6497 	            SYS_RES_MEMORY, &bar, RF_ACTIVE);
6498 		if (ctx->ifc_msix_mem == NULL) {
6499 			device_printf(dev, "Unable to map MSI-X table\n");
6500 			goto msi;
6501 		}
6502 	}
6503 
6504 	admincnt = sctx->isc_admin_intrcnt;
6505 #if IFLIB_DEBUG
6506 	/* use only 1 qset in debug mode */
6507 	queuemsgs = min(msgs - admincnt, 1);
6508 #else
6509 	queuemsgs = msgs - admincnt;
6510 #endif
6511 #ifdef RSS
6512 	queues = imin(queuemsgs, rss_getnumbuckets());
6513 #else
6514 	queues = queuemsgs;
6515 #endif
6516 	queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues);
6517 	if (bootverbose)
6518 		device_printf(dev,
6519 		    "intr CPUs: %d queue msgs: %d admincnt: %d\n",
6520 		    CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt);
6521 #ifdef  RSS
6522 	/* If we're doing RSS, clamp at the number of RSS buckets */
6523 	if (queues > rss_getnumbuckets())
6524 		queues = rss_getnumbuckets();
6525 #endif
6526 	if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt)
6527 		rx_queues = iflib_num_rx_queues;
6528 	else
6529 		rx_queues = queues;
6530 
6531 	if (rx_queues > scctx->isc_nrxqsets)
6532 		rx_queues = scctx->isc_nrxqsets;
6533 
6534 	/*
6535 	 * We want this to be all logical CPUs by default
6536 	 */
6537 	if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues)
6538 		tx_queues = iflib_num_tx_queues;
6539 	else
6540 		tx_queues = mp_ncpus;
6541 
6542 	if (tx_queues > scctx->isc_ntxqsets)
6543 		tx_queues = scctx->isc_ntxqsets;
6544 
6545 	if (ctx->ifc_sysctl_qs_eq_override == 0) {
6546 #ifdef INVARIANTS
6547 		if (tx_queues != rx_queues)
6548 			device_printf(dev,
6549 			    "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n",
6550 			    min(rx_queues, tx_queues), min(rx_queues, tx_queues));
6551 #endif
6552 		tx_queues = min(rx_queues, tx_queues);
6553 		rx_queues = min(rx_queues, tx_queues);
6554 	}
6555 
6556 	vectors = rx_queues + admincnt;
6557 	if (msgs < vectors) {
6558 		device_printf(dev,
6559 		    "insufficient number of MSI-X vectors "
6560 		    "(supported %d, need %d)\n", msgs, vectors);
6561 		goto msi;
6562 	}
6563 
6564 	device_printf(dev, "Using %d RX queues %d TX queues\n", rx_queues,
6565 	    tx_queues);
6566 	msgs = vectors;
6567 	if ((err = pci_alloc_msix(dev, &vectors)) == 0) {
6568 		if (vectors != msgs) {
6569 			device_printf(dev,
6570 			    "Unable to allocate sufficient MSI-X vectors "
6571 			    "(got %d, need %d)\n", vectors, msgs);
6572 			pci_release_msi(dev);
6573 			if (bar != -1) {
6574 				bus_release_resource(dev, SYS_RES_MEMORY, bar,
6575 				    ctx->ifc_msix_mem);
6576 				ctx->ifc_msix_mem = NULL;
6577 			}
6578 			goto msi;
6579 		}
6580 		device_printf(dev, "Using MSI-X interrupts with %d vectors\n",
6581 		    vectors);
6582 		scctx->isc_vectors = vectors;
6583 		scctx->isc_nrxqsets = rx_queues;
6584 		scctx->isc_ntxqsets = tx_queues;
6585 		scctx->isc_intr = IFLIB_INTR_MSIX;
6586 
6587 		return (vectors);
6588 	} else {
6589 		device_printf(dev,
6590 		    "failed to allocate %d MSI-X vectors, err: %d\n", vectors,
6591 		    err);
6592 		if (bar != -1) {
6593 			bus_release_resource(dev, SYS_RES_MEMORY, bar,
6594 			    ctx->ifc_msix_mem);
6595 			ctx->ifc_msix_mem = NULL;
6596 		}
6597 	}
6598 
6599 msi:
6600 	vectors = pci_msi_count(dev);
6601 	scctx->isc_nrxqsets = 1;
6602 	scctx->isc_ntxqsets = 1;
6603 	scctx->isc_vectors = vectors;
6604 	if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) {
6605 		device_printf(dev,"Using an MSI interrupt\n");
6606 		scctx->isc_intr = IFLIB_INTR_MSI;
6607 	} else {
6608 		scctx->isc_vectors = 1;
6609 		device_printf(dev,"Using a Legacy interrupt\n");
6610 		scctx->isc_intr = IFLIB_INTR_LEGACY;
6611 	}
6612 
6613 	return (vectors);
6614 }
6615 
6616 static const char *ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" };
6617 
6618 static int
6619 mp_ring_state_handler(SYSCTL_HANDLER_ARGS)
6620 {
6621 	int rc;
6622 	uint16_t *state = ((uint16_t *)oidp->oid_arg1);
6623 	struct sbuf *sb;
6624 	const char *ring_state = "UNKNOWN";
6625 
6626 	/* XXX needed ? */
6627 	rc = sysctl_wire_old_buffer(req, 0);
6628 	MPASS(rc == 0);
6629 	if (rc != 0)
6630 		return (rc);
6631 	sb = sbuf_new_for_sysctl(NULL, NULL, 80, req);
6632 	MPASS(sb != NULL);
6633 	if (sb == NULL)
6634 		return (ENOMEM);
6635 	if (state[3] <= 3)
6636 		ring_state = ring_states[state[3]];
6637 
6638 	sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s",
6639 		    state[0], state[1], state[2], ring_state);
6640 	rc = sbuf_finish(sb);
6641 	sbuf_delete(sb);
6642         return(rc);
6643 }
6644 
6645 enum iflib_ndesc_handler {
6646 	IFLIB_NTXD_HANDLER,
6647 	IFLIB_NRXD_HANDLER,
6648 };
6649 
6650 static int
6651 mp_ndesc_handler(SYSCTL_HANDLER_ARGS)
6652 {
6653 	if_ctx_t ctx = (void *)arg1;
6654 	enum iflib_ndesc_handler type = arg2;
6655 	char buf[256] = {0};
6656 	qidx_t *ndesc;
6657 	char *p, *next;
6658 	int nqs, rc, i;
6659 
6660 	nqs = 8;
6661 	switch(type) {
6662 	case IFLIB_NTXD_HANDLER:
6663 		ndesc = ctx->ifc_sysctl_ntxds;
6664 		if (ctx->ifc_sctx)
6665 			nqs = ctx->ifc_sctx->isc_ntxqs;
6666 		break;
6667 	case IFLIB_NRXD_HANDLER:
6668 		ndesc = ctx->ifc_sysctl_nrxds;
6669 		if (ctx->ifc_sctx)
6670 			nqs = ctx->ifc_sctx->isc_nrxqs;
6671 		break;
6672 	default:
6673 		printf("%s: unhandled type\n", __func__);
6674 		return (EINVAL);
6675 	}
6676 	if (nqs == 0)
6677 		nqs = 8;
6678 
6679 	for (i=0; i<8; i++) {
6680 		if (i >= nqs)
6681 			break;
6682 		if (i)
6683 			strcat(buf, ",");
6684 		sprintf(strchr(buf, 0), "%d", ndesc[i]);
6685 	}
6686 
6687 	rc = sysctl_handle_string(oidp, buf, sizeof(buf), req);
6688 	if (rc || req->newptr == NULL)
6689 		return rc;
6690 
6691 	for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p;
6692 	    i++, p = strsep(&next, " ,")) {
6693 		ndesc[i] = strtoul(p, NULL, 10);
6694 	}
6695 
6696 	return(rc);
6697 }
6698 
6699 #define NAME_BUFLEN 32
6700 static void
6701 iflib_add_device_sysctl_pre(if_ctx_t ctx)
6702 {
6703         device_t dev = iflib_get_dev(ctx);
6704 	struct sysctl_oid_list *child, *oid_list;
6705 	struct sysctl_ctx_list *ctx_list;
6706 	struct sysctl_oid *node;
6707 
6708 	ctx_list = device_get_sysctl_ctx(dev);
6709 	child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev));
6710 	ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, "iflib",
6711 	    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "IFLIB fields");
6712 	oid_list = SYSCTL_CHILDREN(node);
6713 
6714 	SYSCTL_ADD_CONST_STRING(ctx_list, oid_list, OID_AUTO, "driver_version",
6715 		       CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version,
6716 		       "driver version");
6717 
6718 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs",
6719 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0,
6720 			"# of txqs to use, 0 => use default #");
6721 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs",
6722 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0,
6723 			"# of rxqs to use, 0 => use default #");
6724 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_qs_enable",
6725 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0,
6726                        "permit #txq != #rxq");
6727 	SYSCTL_ADD_INT(ctx_list, oid_list, OID_AUTO, "disable_msix",
6728                       CTLFLAG_RWTUN, &ctx->ifc_softc_ctx.isc_disable_msix, 0,
6729                       "disable MSI-X (default 0)");
6730 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "rx_budget",
6731 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_rx_budget, 0,
6732 		       "set the RX budget");
6733 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "tx_abdicate",
6734 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_tx_abdicate, 0,
6735 		       "cause TX to abdicate instead of running to completion");
6736 	ctx->ifc_sysctl_core_offset = CORE_OFFSET_UNSPECIFIED;
6737 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "core_offset",
6738 		       CTLFLAG_RDTUN, &ctx->ifc_sysctl_core_offset, 0,
6739 		       "offset to start using cores at");
6740 	SYSCTL_ADD_U8(ctx_list, oid_list, OID_AUTO, "separate_txrx",
6741 		       CTLFLAG_RDTUN, &ctx->ifc_sysctl_separate_txrx, 0,
6742 		       "use separate cores for TX and RX");
6743 
6744 	/* XXX change for per-queue sizes */
6745 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_ntxds",
6746 	    CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NEEDGIANT, ctx,
6747 	    IFLIB_NTXD_HANDLER, mp_ndesc_handler, "A",
6748 	    "list of # of TX descriptors to use, 0 = use default #");
6749 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_nrxds",
6750 	    CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NEEDGIANT, ctx,
6751 	    IFLIB_NRXD_HANDLER, mp_ndesc_handler, "A",
6752 	    "list of # of RX descriptors to use, 0 = use default #");
6753 }
6754 
6755 static void
6756 iflib_add_device_sysctl_post(if_ctx_t ctx)
6757 {
6758 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6759 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6760         device_t dev = iflib_get_dev(ctx);
6761 	struct sysctl_oid_list *child;
6762 	struct sysctl_ctx_list *ctx_list;
6763 	iflib_fl_t fl;
6764 	iflib_txq_t txq;
6765 	iflib_rxq_t rxq;
6766 	int i, j;
6767 	char namebuf[NAME_BUFLEN];
6768 	char *qfmt;
6769 	struct sysctl_oid *queue_node, *fl_node, *node;
6770 	struct sysctl_oid_list *queue_list, *fl_list;
6771 	ctx_list = device_get_sysctl_ctx(dev);
6772 
6773 	node = ctx->ifc_sysctl_node;
6774 	child = SYSCTL_CHILDREN(node);
6775 
6776 	if (scctx->isc_ntxqsets > 100)
6777 		qfmt = "txq%03d";
6778 	else if (scctx->isc_ntxqsets > 10)
6779 		qfmt = "txq%02d";
6780 	else
6781 		qfmt = "txq%d";
6782 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
6783 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6784 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6785 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "Queue Name");
6786 		queue_list = SYSCTL_CHILDREN(queue_node);
6787 #if MEMORY_LOGGING
6788 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued",
6789 				CTLFLAG_RD,
6790 				&txq->ift_dequeued, "total mbufs freed");
6791 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued",
6792 				CTLFLAG_RD,
6793 				&txq->ift_enqueued, "total mbufs enqueued");
6794 #endif
6795 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag",
6796 				   CTLFLAG_RD,
6797 				   &txq->ift_mbuf_defrag, "# of times m_defrag was called");
6798 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "m_pullups",
6799 				   CTLFLAG_RD,
6800 				   &txq->ift_pullups, "# of times m_pullup was called");
6801 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag_failed",
6802 				   CTLFLAG_RD,
6803 				   &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed");
6804 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_desc_avail",
6805 				   CTLFLAG_RD,
6806 				   &txq->ift_no_desc_avail, "# of times no descriptors were available");
6807 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "tx_map_failed",
6808 				   CTLFLAG_RD,
6809 				   &txq->ift_map_failed, "# of times DMA map failed");
6810 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txd_encap_efbig",
6811 				   CTLFLAG_RD,
6812 				   &txq->ift_txd_encap_efbig, "# of times txd_encap returned EFBIG");
6813 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_tx_dma_setup",
6814 				   CTLFLAG_RD,
6815 				   &txq->ift_no_tx_dma_setup, "# of times map failed for other than EFBIG");
6816 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx",
6817 				   CTLFLAG_RD,
6818 				   &txq->ift_pidx, 1, "Producer Index");
6819 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx",
6820 				   CTLFLAG_RD,
6821 				   &txq->ift_cidx, 1, "Consumer Index");
6822 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx_processed",
6823 				   CTLFLAG_RD,
6824 				   &txq->ift_cidx_processed, 1, "Consumer Index seen by credit update");
6825 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use",
6826 				   CTLFLAG_RD,
6827 				   &txq->ift_in_use, 1, "descriptors in use");
6828 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_processed",
6829 				   CTLFLAG_RD,
6830 				   &txq->ift_processed, "descriptors procesed for clean");
6831 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned",
6832 				   CTLFLAG_RD,
6833 				   &txq->ift_cleaned, "total cleaned");
6834 		SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state",
6835 		    CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_NEEDGIANT,
6836 		    __DEVOLATILE(uint64_t *, &txq->ift_br->state), 0,
6837 		    mp_ring_state_handler, "A", "soft ring state");
6838 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_enqueues",
6839 				       CTLFLAG_RD, &txq->ift_br->enqueues,
6840 				       "# of enqueues to the mp_ring for this queue");
6841 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_drops",
6842 				       CTLFLAG_RD, &txq->ift_br->drops,
6843 				       "# of drops in the mp_ring for this queue");
6844 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_starts",
6845 				       CTLFLAG_RD, &txq->ift_br->starts,
6846 				       "# of normal consumer starts in the mp_ring for this queue");
6847 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_stalls",
6848 				       CTLFLAG_RD, &txq->ift_br->stalls,
6849 					       "# of consumer stalls in the mp_ring for this queue");
6850 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_restarts",
6851 			       CTLFLAG_RD, &txq->ift_br->restarts,
6852 				       "# of consumer restarts in the mp_ring for this queue");
6853 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_abdications",
6854 				       CTLFLAG_RD, &txq->ift_br->abdications,
6855 				       "# of consumer abdications in the mp_ring for this queue");
6856 	}
6857 
6858 	if (scctx->isc_nrxqsets > 100)
6859 		qfmt = "rxq%03d";
6860 	else if (scctx->isc_nrxqsets > 10)
6861 		qfmt = "rxq%02d";
6862 	else
6863 		qfmt = "rxq%d";
6864 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
6865 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6866 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6867 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "Queue Name");
6868 		queue_list = SYSCTL_CHILDREN(queue_node);
6869 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
6870 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_cidx",
6871 				       CTLFLAG_RD,
6872 				       &rxq->ifr_cq_cidx, 1, "Consumer Index");
6873 		}
6874 
6875 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
6876 			snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j);
6877 			fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list, OID_AUTO, namebuf,
6878 			    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "freelist Name");
6879 			fl_list = SYSCTL_CHILDREN(fl_node);
6880 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx",
6881 				       CTLFLAG_RD,
6882 				       &fl->ifl_pidx, 1, "Producer Index");
6883 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx",
6884 				       CTLFLAG_RD,
6885 				       &fl->ifl_cidx, 1, "Consumer Index");
6886 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits",
6887 				       CTLFLAG_RD,
6888 				       &fl->ifl_credits, 1, "credits available");
6889 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "buf_size",
6890 				       CTLFLAG_RD,
6891 				       &fl->ifl_buf_size, 1, "buffer size");
6892 #if MEMORY_LOGGING
6893 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_enqueued",
6894 					CTLFLAG_RD,
6895 					&fl->ifl_m_enqueued, "mbufs allocated");
6896 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_dequeued",
6897 					CTLFLAG_RD,
6898 					&fl->ifl_m_dequeued, "mbufs freed");
6899 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_enqueued",
6900 					CTLFLAG_RD,
6901 					&fl->ifl_cl_enqueued, "clusters allocated");
6902 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_dequeued",
6903 					CTLFLAG_RD,
6904 					&fl->ifl_cl_dequeued, "clusters freed");
6905 #endif
6906 		}
6907 	}
6908 
6909 }
6910 
6911 void
6912 iflib_request_reset(if_ctx_t ctx)
6913 {
6914 
6915 	STATE_LOCK(ctx);
6916 	ctx->ifc_flags |= IFC_DO_RESET;
6917 	STATE_UNLOCK(ctx);
6918 }
6919 
6920 #ifndef __NO_STRICT_ALIGNMENT
6921 static struct mbuf *
6922 iflib_fixup_rx(struct mbuf *m)
6923 {
6924 	struct mbuf *n;
6925 
6926 	if (m->m_len <= (MCLBYTES - ETHER_HDR_LEN)) {
6927 		bcopy(m->m_data, m->m_data + ETHER_HDR_LEN, m->m_len);
6928 		m->m_data += ETHER_HDR_LEN;
6929 		n = m;
6930 	} else {
6931 		MGETHDR(n, M_NOWAIT, MT_DATA);
6932 		if (n == NULL) {
6933 			m_freem(m);
6934 			return (NULL);
6935 		}
6936 		bcopy(m->m_data, n->m_data, ETHER_HDR_LEN);
6937 		m->m_data += ETHER_HDR_LEN;
6938 		m->m_len -= ETHER_HDR_LEN;
6939 		n->m_len = ETHER_HDR_LEN;
6940 		M_MOVE_PKTHDR(n, m);
6941 		n->m_next = m;
6942 	}
6943 	return (n);
6944 }
6945 #endif
6946 
6947 #ifdef DEBUGNET
6948 static void
6949 iflib_debugnet_init(if_t ifp, int *nrxr, int *ncl, int *clsize)
6950 {
6951 	if_ctx_t ctx;
6952 
6953 	ctx = if_getsoftc(ifp);
6954 	CTX_LOCK(ctx);
6955 	*nrxr = NRXQSETS(ctx);
6956 	*ncl = ctx->ifc_rxqs[0].ifr_fl->ifl_size;
6957 	*clsize = ctx->ifc_rxqs[0].ifr_fl->ifl_buf_size;
6958 	CTX_UNLOCK(ctx);
6959 }
6960 
6961 static void
6962 iflib_debugnet_event(if_t ifp, enum debugnet_ev event)
6963 {
6964 	if_ctx_t ctx;
6965 	if_softc_ctx_t scctx;
6966 	iflib_fl_t fl;
6967 	iflib_rxq_t rxq;
6968 	int i, j;
6969 
6970 	ctx = if_getsoftc(ifp);
6971 	scctx = &ctx->ifc_softc_ctx;
6972 
6973 	switch (event) {
6974 	case DEBUGNET_START:
6975 		for (i = 0; i < scctx->isc_nrxqsets; i++) {
6976 			rxq = &ctx->ifc_rxqs[i];
6977 			for (j = 0; j < rxq->ifr_nfl; j++) {
6978 				fl = rxq->ifr_fl;
6979 				fl->ifl_zone = m_getzone(fl->ifl_buf_size);
6980 			}
6981 		}
6982 		iflib_no_tx_batch = 1;
6983 		break;
6984 	default:
6985 		break;
6986 	}
6987 }
6988 
6989 static int
6990 iflib_debugnet_transmit(if_t ifp, struct mbuf *m)
6991 {
6992 	if_ctx_t ctx;
6993 	iflib_txq_t txq;
6994 	int error;
6995 
6996 	ctx = if_getsoftc(ifp);
6997 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
6998 	    IFF_DRV_RUNNING)
6999 		return (EBUSY);
7000 
7001 	txq = &ctx->ifc_txqs[0];
7002 	error = iflib_encap(txq, &m);
7003 	if (error == 0)
7004 		(void)iflib_txd_db_check(txq, true);
7005 	return (error);
7006 }
7007 
7008 static int
7009 iflib_debugnet_poll(if_t ifp, int count)
7010 {
7011 	struct epoch_tracker et;
7012 	if_ctx_t ctx;
7013 	if_softc_ctx_t scctx;
7014 	iflib_txq_t txq;
7015 	int i;
7016 
7017 	ctx = if_getsoftc(ifp);
7018 	scctx = &ctx->ifc_softc_ctx;
7019 
7020 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
7021 	    IFF_DRV_RUNNING)
7022 		return (EBUSY);
7023 
7024 	txq = &ctx->ifc_txqs[0];
7025 	(void)iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
7026 
7027 	NET_EPOCH_ENTER(et);
7028 	for (i = 0; i < scctx->isc_nrxqsets; i++)
7029 		(void)iflib_rxeof(&ctx->ifc_rxqs[i], 16 /* XXX */);
7030 	NET_EPOCH_EXIT(et);
7031 	return (0);
7032 }
7033 #endif /* DEBUGNET */
7034