xref: /freebsd/sys/net/iflib.c (revision 85a2ea3f5785a299db029b84eee6d33d174e38da)
1 /*-
2  * Copyright (c) 2014-2018, Matthew Macy <mmacy@mattmacy.io>
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions are met:
7  *
8  *  1. Redistributions of source code must retain the above copyright notice,
9  *     this list of conditions and the following disclaimer.
10  *
11  *  2. Neither the name of Matthew Macy nor the names of its
12  *     contributors may be used to endorse or promote products derived from
13  *     this software without specific prior written permission.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
16  * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18  * ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
19  * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
20  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
21  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
22  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
23  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
24  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
25  * POSSIBILITY OF SUCH DAMAGE.
26  */
27 
28 #include <sys/cdefs.h>
29 __FBSDID("$FreeBSD$");
30 
31 #include "opt_inet.h"
32 #include "opt_inet6.h"
33 #include "opt_acpi.h"
34 #include "opt_sched.h"
35 
36 #include <sys/param.h>
37 #include <sys/types.h>
38 #include <sys/bus.h>
39 #include <sys/eventhandler.h>
40 #include <sys/kernel.h>
41 #include <sys/lock.h>
42 #include <sys/mutex.h>
43 #include <sys/module.h>
44 #include <sys/kobj.h>
45 #include <sys/rman.h>
46 #include <sys/sbuf.h>
47 #include <sys/smp.h>
48 #include <sys/socket.h>
49 #include <sys/sockio.h>
50 #include <sys/sysctl.h>
51 #include <sys/syslog.h>
52 #include <sys/taskqueue.h>
53 #include <sys/limits.h>
54 
55 #include <net/if.h>
56 #include <net/if_var.h>
57 #include <net/if_types.h>
58 #include <net/if_media.h>
59 #include <net/bpf.h>
60 #include <net/ethernet.h>
61 #include <net/mp_ring.h>
62 #include <net/debugnet.h>
63 #include <net/pfil.h>
64 #include <net/vnet.h>
65 
66 #include <netinet/in.h>
67 #include <netinet/in_pcb.h>
68 #include <netinet/tcp_lro.h>
69 #include <netinet/in_systm.h>
70 #include <netinet/if_ether.h>
71 #include <netinet/ip.h>
72 #include <netinet/ip6.h>
73 #include <netinet/tcp.h>
74 #include <netinet/ip_var.h>
75 #include <netinet6/ip6_var.h>
76 
77 #include <machine/bus.h>
78 #include <machine/in_cksum.h>
79 
80 #include <vm/vm.h>
81 #include <vm/pmap.h>
82 
83 #include <dev/led/led.h>
84 #include <dev/pci/pcireg.h>
85 #include <dev/pci/pcivar.h>
86 #include <dev/pci/pci_private.h>
87 
88 #include <net/iflib.h>
89 #include <net/iflib_private.h>
90 
91 #include "ifdi_if.h"
92 
93 #ifdef PCI_IOV
94 #include <dev/pci/pci_iov.h>
95 #endif
96 
97 #include <sys/bitstring.h>
98 /*
99  * enable accounting of every mbuf as it comes in to and goes out of
100  * iflib's software descriptor references
101  */
102 #define MEMORY_LOGGING 0
103 /*
104  * Enable mbuf vectors for compressing long mbuf chains
105  */
106 
107 /*
108  * NB:
109  * - Prefetching in tx cleaning should perhaps be a tunable. The distance ahead
110  *   we prefetch needs to be determined by the time spent in m_free vis a vis
111  *   the cost of a prefetch. This will of course vary based on the workload:
112  *      - NFLX's m_free path is dominated by vm-based M_EXT manipulation which
113  *        is quite expensive, thus suggesting very little prefetch.
114  *      - small packet forwarding which is just returning a single mbuf to
115  *        UMA will typically be very fast vis a vis the cost of a memory
116  *        access.
117  */
118 
119 
120 /*
121  * File organization:
122  *  - private structures
123  *  - iflib private utility functions
124  *  - ifnet functions
125  *  - vlan registry and other exported functions
126  *  - iflib public core functions
127  *
128  *
129  */
130 MALLOC_DEFINE(M_IFLIB, "iflib", "ifnet library");
131 
132 #define	IFLIB_RXEOF_MORE (1U << 0)
133 #define	IFLIB_RXEOF_EMPTY (2U << 0)
134 
135 struct iflib_txq;
136 typedef struct iflib_txq *iflib_txq_t;
137 struct iflib_rxq;
138 typedef struct iflib_rxq *iflib_rxq_t;
139 struct iflib_fl;
140 typedef struct iflib_fl *iflib_fl_t;
141 
142 struct iflib_ctx;
143 
144 static void iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid);
145 static void iflib_timer(void *arg);
146 
147 typedef struct iflib_filter_info {
148 	driver_filter_t *ifi_filter;
149 	void *ifi_filter_arg;
150 	struct grouptask *ifi_task;
151 	void *ifi_ctx;
152 } *iflib_filter_info_t;
153 
154 struct iflib_ctx {
155 	KOBJ_FIELDS;
156 	/*
157 	 * Pointer to hardware driver's softc
158 	 */
159 	void *ifc_softc;
160 	device_t ifc_dev;
161 	if_t ifc_ifp;
162 
163 	cpuset_t ifc_cpus;
164 	if_shared_ctx_t ifc_sctx;
165 	struct if_softc_ctx ifc_softc_ctx;
166 
167 	struct sx ifc_ctx_sx;
168 	struct mtx ifc_state_mtx;
169 
170 	iflib_txq_t ifc_txqs;
171 	iflib_rxq_t ifc_rxqs;
172 	uint32_t ifc_if_flags;
173 	uint32_t ifc_flags;
174 	uint32_t ifc_max_fl_buf_size;
175 	uint32_t ifc_rx_mbuf_sz;
176 
177 	int ifc_link_state;
178 	int ifc_watchdog_events;
179 	struct cdev *ifc_led_dev;
180 	struct resource *ifc_msix_mem;
181 
182 	struct if_irq ifc_legacy_irq;
183 	struct grouptask ifc_admin_task;
184 	struct grouptask ifc_vflr_task;
185 	struct iflib_filter_info ifc_filter_info;
186 	struct ifmedia	ifc_media;
187 	struct ifmedia	*ifc_mediap;
188 
189 	struct sysctl_oid *ifc_sysctl_node;
190 	uint16_t ifc_sysctl_ntxqs;
191 	uint16_t ifc_sysctl_nrxqs;
192 	uint16_t ifc_sysctl_qs_eq_override;
193 	uint16_t ifc_sysctl_rx_budget;
194 	uint16_t ifc_sysctl_tx_abdicate;
195 	uint16_t ifc_sysctl_core_offset;
196 #define	CORE_OFFSET_UNSPECIFIED	0xffff
197 	uint8_t  ifc_sysctl_separate_txrx;
198 
199 	qidx_t ifc_sysctl_ntxds[8];
200 	qidx_t ifc_sysctl_nrxds[8];
201 	struct if_txrx ifc_txrx;
202 #define isc_txd_encap  ifc_txrx.ift_txd_encap
203 #define isc_txd_flush  ifc_txrx.ift_txd_flush
204 #define isc_txd_credits_update  ifc_txrx.ift_txd_credits_update
205 #define isc_rxd_available ifc_txrx.ift_rxd_available
206 #define isc_rxd_pkt_get ifc_txrx.ift_rxd_pkt_get
207 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
208 #define isc_rxd_flush ifc_txrx.ift_rxd_flush
209 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
210 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
211 #define isc_legacy_intr ifc_txrx.ift_legacy_intr
212 	eventhandler_tag ifc_vlan_attach_event;
213 	eventhandler_tag ifc_vlan_detach_event;
214 	struct ether_addr ifc_mac;
215 };
216 
217 void *
218 iflib_get_softc(if_ctx_t ctx)
219 {
220 
221 	return (ctx->ifc_softc);
222 }
223 
224 device_t
225 iflib_get_dev(if_ctx_t ctx)
226 {
227 
228 	return (ctx->ifc_dev);
229 }
230 
231 if_t
232 iflib_get_ifp(if_ctx_t ctx)
233 {
234 
235 	return (ctx->ifc_ifp);
236 }
237 
238 struct ifmedia *
239 iflib_get_media(if_ctx_t ctx)
240 {
241 
242 	return (ctx->ifc_mediap);
243 }
244 
245 uint32_t
246 iflib_get_flags(if_ctx_t ctx)
247 {
248 	return (ctx->ifc_flags);
249 }
250 
251 void
252 iflib_set_mac(if_ctx_t ctx, uint8_t mac[ETHER_ADDR_LEN])
253 {
254 
255 	bcopy(mac, ctx->ifc_mac.octet, ETHER_ADDR_LEN);
256 }
257 
258 if_softc_ctx_t
259 iflib_get_softc_ctx(if_ctx_t ctx)
260 {
261 
262 	return (&ctx->ifc_softc_ctx);
263 }
264 
265 if_shared_ctx_t
266 iflib_get_sctx(if_ctx_t ctx)
267 {
268 
269 	return (ctx->ifc_sctx);
270 }
271 
272 #define IP_ALIGNED(m) ((((uintptr_t)(m)->m_data) & 0x3) == 0x2)
273 #define CACHE_PTR_INCREMENT (CACHE_LINE_SIZE/sizeof(void*))
274 #define CACHE_PTR_NEXT(ptr) ((void *)(((uintptr_t)(ptr)+CACHE_LINE_SIZE-1) & (CACHE_LINE_SIZE-1)))
275 
276 #define LINK_ACTIVE(ctx) ((ctx)->ifc_link_state == LINK_STATE_UP)
277 #define CTX_IS_VF(ctx) ((ctx)->ifc_sctx->isc_flags & IFLIB_IS_VF)
278 
279 typedef struct iflib_sw_rx_desc_array {
280 	bus_dmamap_t	*ifsd_map;         /* bus_dma maps for packet */
281 	struct mbuf	**ifsd_m;           /* pkthdr mbufs */
282 	caddr_t		*ifsd_cl;          /* direct cluster pointer for rx */
283 	bus_addr_t	*ifsd_ba;          /* bus addr of cluster for rx */
284 } iflib_rxsd_array_t;
285 
286 typedef struct iflib_sw_tx_desc_array {
287 	bus_dmamap_t    *ifsd_map;         /* bus_dma maps for packet */
288 	bus_dmamap_t	*ifsd_tso_map;     /* bus_dma maps for TSO packet */
289 	struct mbuf    **ifsd_m;           /* pkthdr mbufs */
290 } if_txsd_vec_t;
291 
292 /* magic number that should be high enough for any hardware */
293 #define IFLIB_MAX_TX_SEGS		128
294 #define IFLIB_RX_COPY_THRESH		128
295 #define IFLIB_MAX_RX_REFRESH		32
296 /* The minimum descriptors per second before we start coalescing */
297 #define IFLIB_MIN_DESC_SEC		16384
298 #define IFLIB_DEFAULT_TX_UPDATE_FREQ	16
299 #define IFLIB_QUEUE_IDLE		0
300 #define IFLIB_QUEUE_HUNG		1
301 #define IFLIB_QUEUE_WORKING		2
302 /* maximum number of txqs that can share an rx interrupt */
303 #define IFLIB_MAX_TX_SHARED_INTR	4
304 
305 /* this should really scale with ring size - this is a fairly arbitrary value */
306 #define TX_BATCH_SIZE			32
307 
308 #define IFLIB_RESTART_BUDGET		8
309 
310 #define CSUM_OFFLOAD		(CSUM_IP_TSO|CSUM_IP6_TSO|CSUM_IP| \
311 				 CSUM_IP_UDP|CSUM_IP_TCP|CSUM_IP_SCTP| \
312 				 CSUM_IP6_UDP|CSUM_IP6_TCP|CSUM_IP6_SCTP)
313 
314 struct iflib_txq {
315 	qidx_t		ift_in_use;
316 	qidx_t		ift_cidx;
317 	qidx_t		ift_cidx_processed;
318 	qidx_t		ift_pidx;
319 	uint8_t		ift_gen;
320 	uint8_t		ift_br_offset;
321 	uint16_t	ift_npending;
322 	uint16_t	ift_db_pending;
323 	uint16_t	ift_rs_pending;
324 	/* implicit pad */
325 	uint8_t		ift_txd_size[8];
326 	uint64_t	ift_processed;
327 	uint64_t	ift_cleaned;
328 	uint64_t	ift_cleaned_prev;
329 #if MEMORY_LOGGING
330 	uint64_t	ift_enqueued;
331 	uint64_t	ift_dequeued;
332 #endif
333 	uint64_t	ift_no_tx_dma_setup;
334 	uint64_t	ift_no_desc_avail;
335 	uint64_t	ift_mbuf_defrag_failed;
336 	uint64_t	ift_mbuf_defrag;
337 	uint64_t	ift_map_failed;
338 	uint64_t	ift_txd_encap_efbig;
339 	uint64_t	ift_pullups;
340 	uint64_t	ift_last_timer_tick;
341 
342 	struct mtx	ift_mtx;
343 	struct mtx	ift_db_mtx;
344 
345 	/* constant values */
346 	if_ctx_t	ift_ctx;
347 	struct ifmp_ring        *ift_br;
348 	struct grouptask	ift_task;
349 	qidx_t		ift_size;
350 	uint16_t	ift_id;
351 	struct callout	ift_timer;
352 
353 	if_txsd_vec_t	ift_sds;
354 	uint8_t		ift_qstatus;
355 	uint8_t		ift_closed;
356 	uint8_t		ift_update_freq;
357 	struct iflib_filter_info ift_filter_info;
358 	bus_dma_tag_t	ift_buf_tag;
359 	bus_dma_tag_t	ift_tso_buf_tag;
360 	iflib_dma_info_t	ift_ifdi;
361 #define	MTX_NAME_LEN	32
362 	char                    ift_mtx_name[MTX_NAME_LEN];
363 	bus_dma_segment_t	ift_segs[IFLIB_MAX_TX_SEGS]  __aligned(CACHE_LINE_SIZE);
364 #ifdef IFLIB_DIAGNOSTICS
365 	uint64_t ift_cpu_exec_count[256];
366 #endif
367 } __aligned(CACHE_LINE_SIZE);
368 
369 struct iflib_fl {
370 	qidx_t		ifl_cidx;
371 	qidx_t		ifl_pidx;
372 	qidx_t		ifl_credits;
373 	uint8_t		ifl_gen;
374 	uint8_t		ifl_rxd_size;
375 #if MEMORY_LOGGING
376 	uint64_t	ifl_m_enqueued;
377 	uint64_t	ifl_m_dequeued;
378 	uint64_t	ifl_cl_enqueued;
379 	uint64_t	ifl_cl_dequeued;
380 #endif
381 	/* implicit pad */
382 	bitstr_t 	*ifl_rx_bitmap;
383 	qidx_t		ifl_fragidx;
384 	/* constant */
385 	qidx_t		ifl_size;
386 	uint16_t	ifl_buf_size;
387 	uint16_t	ifl_cltype;
388 	uma_zone_t	ifl_zone;
389 	iflib_rxsd_array_t	ifl_sds;
390 	iflib_rxq_t	ifl_rxq;
391 	uint8_t		ifl_id;
392 	bus_dma_tag_t	ifl_buf_tag;
393 	iflib_dma_info_t	ifl_ifdi;
394 	uint64_t	ifl_bus_addrs[IFLIB_MAX_RX_REFRESH] __aligned(CACHE_LINE_SIZE);
395 	caddr_t		ifl_vm_addrs[IFLIB_MAX_RX_REFRESH];
396 	qidx_t	ifl_rxd_idxs[IFLIB_MAX_RX_REFRESH];
397 }  __aligned(CACHE_LINE_SIZE);
398 
399 static inline qidx_t
400 get_inuse(int size, qidx_t cidx, qidx_t pidx, uint8_t gen)
401 {
402 	qidx_t used;
403 
404 	if (pidx > cidx)
405 		used = pidx - cidx;
406 	else if (pidx < cidx)
407 		used = size - cidx + pidx;
408 	else if (gen == 0 && pidx == cidx)
409 		used = 0;
410 	else if (gen == 1 && pidx == cidx)
411 		used = size;
412 	else
413 		panic("bad state");
414 
415 	return (used);
416 }
417 
418 #define TXQ_AVAIL(txq) (txq->ift_size - get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx, txq->ift_gen))
419 
420 #define IDXDIFF(head, tail, wrap) \
421 	((head) >= (tail) ? (head) - (tail) : (wrap) - (tail) + (head))
422 
423 struct iflib_rxq {
424 	if_ctx_t	ifr_ctx;
425 	iflib_fl_t	ifr_fl;
426 	uint64_t	ifr_rx_irq;
427 	struct pfil_head	*pfil;
428 	/*
429 	 * If there is a separate completion queue (IFLIB_HAS_RXCQ), this is
430 	 * the command queue consumer index.  Otherwise it's unused.
431 	 */
432 	qidx_t		ifr_cq_cidx;
433 	uint16_t	ifr_id;
434 	uint8_t		ifr_nfl;
435 	uint8_t		ifr_ntxqirq;
436 	uint8_t		ifr_txqid[IFLIB_MAX_TX_SHARED_INTR];
437 	uint8_t		ifr_fl_offset;
438 	struct lro_ctrl			ifr_lc;
439 	struct grouptask        ifr_task;
440 	struct callout		ifr_watchdog;
441 	struct iflib_filter_info ifr_filter_info;
442 	iflib_dma_info_t		ifr_ifdi;
443 
444 	/* dynamically allocate if any drivers need a value substantially larger than this */
445 	struct if_rxd_frag	ifr_frags[IFLIB_MAX_RX_SEGS] __aligned(CACHE_LINE_SIZE);
446 #ifdef IFLIB_DIAGNOSTICS
447 	uint64_t ifr_cpu_exec_count[256];
448 #endif
449 }  __aligned(CACHE_LINE_SIZE);
450 
451 typedef struct if_rxsd {
452 	caddr_t *ifsd_cl;
453 	iflib_fl_t ifsd_fl;
454 } *if_rxsd_t;
455 
456 /* multiple of word size */
457 #ifdef __LP64__
458 #define PKT_INFO_SIZE	6
459 #define RXD_INFO_SIZE	5
460 #define PKT_TYPE uint64_t
461 #else
462 #define PKT_INFO_SIZE	11
463 #define RXD_INFO_SIZE	8
464 #define PKT_TYPE uint32_t
465 #endif
466 #define PKT_LOOP_BOUND  ((PKT_INFO_SIZE/3)*3)
467 #define RXD_LOOP_BOUND  ((RXD_INFO_SIZE/4)*4)
468 
469 typedef struct if_pkt_info_pad {
470 	PKT_TYPE pkt_val[PKT_INFO_SIZE];
471 } *if_pkt_info_pad_t;
472 typedef struct if_rxd_info_pad {
473 	PKT_TYPE rxd_val[RXD_INFO_SIZE];
474 } *if_rxd_info_pad_t;
475 
476 CTASSERT(sizeof(struct if_pkt_info_pad) == sizeof(struct if_pkt_info));
477 CTASSERT(sizeof(struct if_rxd_info_pad) == sizeof(struct if_rxd_info));
478 
479 
480 static inline void
481 pkt_info_zero(if_pkt_info_t pi)
482 {
483 	if_pkt_info_pad_t pi_pad;
484 
485 	pi_pad = (if_pkt_info_pad_t)pi;
486 	pi_pad->pkt_val[0] = 0; pi_pad->pkt_val[1] = 0; pi_pad->pkt_val[2] = 0;
487 	pi_pad->pkt_val[3] = 0; pi_pad->pkt_val[4] = 0; pi_pad->pkt_val[5] = 0;
488 #ifndef __LP64__
489 	pi_pad->pkt_val[6] = 0; pi_pad->pkt_val[7] = 0; pi_pad->pkt_val[8] = 0;
490 	pi_pad->pkt_val[9] = 0; pi_pad->pkt_val[10] = 0;
491 #endif
492 }
493 
494 static device_method_t iflib_pseudo_methods[] = {
495 	DEVMETHOD(device_attach, noop_attach),
496 	DEVMETHOD(device_detach, iflib_pseudo_detach),
497 	DEVMETHOD_END
498 };
499 
500 driver_t iflib_pseudodriver = {
501 	"iflib_pseudo", iflib_pseudo_methods, sizeof(struct iflib_ctx),
502 };
503 
504 static inline void
505 rxd_info_zero(if_rxd_info_t ri)
506 {
507 	if_rxd_info_pad_t ri_pad;
508 	int i;
509 
510 	ri_pad = (if_rxd_info_pad_t)ri;
511 	for (i = 0; i < RXD_LOOP_BOUND; i += 4) {
512 		ri_pad->rxd_val[i] = 0;
513 		ri_pad->rxd_val[i+1] = 0;
514 		ri_pad->rxd_val[i+2] = 0;
515 		ri_pad->rxd_val[i+3] = 0;
516 	}
517 #ifdef __LP64__
518 	ri_pad->rxd_val[RXD_INFO_SIZE-1] = 0;
519 #endif
520 }
521 
522 /*
523  * Only allow a single packet to take up most 1/nth of the tx ring
524  */
525 #define MAX_SINGLE_PACKET_FRACTION 12
526 #define IF_BAD_DMA (bus_addr_t)-1
527 
528 #define CTX_ACTIVE(ctx) ((if_getdrvflags((ctx)->ifc_ifp) & IFF_DRV_RUNNING))
529 
530 #define CTX_LOCK_INIT(_sc)  sx_init(&(_sc)->ifc_ctx_sx, "iflib ctx lock")
531 #define CTX_LOCK(ctx) sx_xlock(&(ctx)->ifc_ctx_sx)
532 #define CTX_UNLOCK(ctx) sx_xunlock(&(ctx)->ifc_ctx_sx)
533 #define CTX_LOCK_DESTROY(ctx) sx_destroy(&(ctx)->ifc_ctx_sx)
534 
535 #define STATE_LOCK_INIT(_sc, _name)  mtx_init(&(_sc)->ifc_state_mtx, _name, "iflib state lock", MTX_DEF)
536 #define STATE_LOCK(ctx) mtx_lock(&(ctx)->ifc_state_mtx)
537 #define STATE_UNLOCK(ctx) mtx_unlock(&(ctx)->ifc_state_mtx)
538 #define STATE_LOCK_DESTROY(ctx) mtx_destroy(&(ctx)->ifc_state_mtx)
539 
540 #define CALLOUT_LOCK(txq)	mtx_lock(&txq->ift_mtx)
541 #define CALLOUT_UNLOCK(txq) 	mtx_unlock(&txq->ift_mtx)
542 
543 void
544 iflib_set_detach(if_ctx_t ctx)
545 {
546 	STATE_LOCK(ctx);
547 	ctx->ifc_flags |= IFC_IN_DETACH;
548 	STATE_UNLOCK(ctx);
549 }
550 
551 /* Our boot-time initialization hook */
552 static int	iflib_module_event_handler(module_t, int, void *);
553 
554 static moduledata_t iflib_moduledata = {
555 	"iflib",
556 	iflib_module_event_handler,
557 	NULL
558 };
559 
560 DECLARE_MODULE(iflib, iflib_moduledata, SI_SUB_INIT_IF, SI_ORDER_ANY);
561 MODULE_VERSION(iflib, 1);
562 
563 MODULE_DEPEND(iflib, pci, 1, 1, 1);
564 MODULE_DEPEND(iflib, ether, 1, 1, 1);
565 
566 TASKQGROUP_DEFINE(if_io_tqg, mp_ncpus, 1);
567 TASKQGROUP_DEFINE(if_config_tqg, 1, 1);
568 
569 #ifndef IFLIB_DEBUG_COUNTERS
570 #ifdef INVARIANTS
571 #define IFLIB_DEBUG_COUNTERS 1
572 #else
573 #define IFLIB_DEBUG_COUNTERS 0
574 #endif /* !INVARIANTS */
575 #endif
576 
577 static SYSCTL_NODE(_net, OID_AUTO, iflib, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
578     "iflib driver parameters");
579 
580 /*
581  * XXX need to ensure that this can't accidentally cause the head to be moved backwards
582  */
583 static int iflib_min_tx_latency = 0;
584 SYSCTL_INT(_net_iflib, OID_AUTO, min_tx_latency, CTLFLAG_RW,
585 		   &iflib_min_tx_latency, 0, "minimize transmit latency at the possible expense of throughput");
586 static int iflib_no_tx_batch = 0;
587 SYSCTL_INT(_net_iflib, OID_AUTO, no_tx_batch, CTLFLAG_RW,
588 		   &iflib_no_tx_batch, 0, "minimize transmit latency at the possible expense of throughput");
589 
590 
591 #if IFLIB_DEBUG_COUNTERS
592 
593 static int iflib_tx_seen;
594 static int iflib_tx_sent;
595 static int iflib_tx_encap;
596 static int iflib_rx_allocs;
597 static int iflib_fl_refills;
598 static int iflib_fl_refills_large;
599 static int iflib_tx_frees;
600 
601 SYSCTL_INT(_net_iflib, OID_AUTO, tx_seen, CTLFLAG_RD,
602 		   &iflib_tx_seen, 0, "# TX mbufs seen");
603 SYSCTL_INT(_net_iflib, OID_AUTO, tx_sent, CTLFLAG_RD,
604 		   &iflib_tx_sent, 0, "# TX mbufs sent");
605 SYSCTL_INT(_net_iflib, OID_AUTO, tx_encap, CTLFLAG_RD,
606 		   &iflib_tx_encap, 0, "# TX mbufs encapped");
607 SYSCTL_INT(_net_iflib, OID_AUTO, tx_frees, CTLFLAG_RD,
608 		   &iflib_tx_frees, 0, "# TX frees");
609 SYSCTL_INT(_net_iflib, OID_AUTO, rx_allocs, CTLFLAG_RD,
610 		   &iflib_rx_allocs, 0, "# RX allocations");
611 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills, CTLFLAG_RD,
612 		   &iflib_fl_refills, 0, "# refills");
613 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills_large, CTLFLAG_RD,
614 		   &iflib_fl_refills_large, 0, "# large refills");
615 
616 
617 static int iflib_txq_drain_flushing;
618 static int iflib_txq_drain_oactive;
619 static int iflib_txq_drain_notready;
620 
621 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_flushing, CTLFLAG_RD,
622 		   &iflib_txq_drain_flushing, 0, "# drain flushes");
623 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_oactive, CTLFLAG_RD,
624 		   &iflib_txq_drain_oactive, 0, "# drain oactives");
625 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_notready, CTLFLAG_RD,
626 		   &iflib_txq_drain_notready, 0, "# drain notready");
627 
628 
629 static int iflib_encap_load_mbuf_fail;
630 static int iflib_encap_pad_mbuf_fail;
631 static int iflib_encap_txq_avail_fail;
632 static int iflib_encap_txd_encap_fail;
633 
634 SYSCTL_INT(_net_iflib, OID_AUTO, encap_load_mbuf_fail, CTLFLAG_RD,
635 		   &iflib_encap_load_mbuf_fail, 0, "# busdma load failures");
636 SYSCTL_INT(_net_iflib, OID_AUTO, encap_pad_mbuf_fail, CTLFLAG_RD,
637 		   &iflib_encap_pad_mbuf_fail, 0, "# runt frame pad failures");
638 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txq_avail_fail, CTLFLAG_RD,
639 		   &iflib_encap_txq_avail_fail, 0, "# txq avail failures");
640 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txd_encap_fail, CTLFLAG_RD,
641 		   &iflib_encap_txd_encap_fail, 0, "# driver encap failures");
642 
643 static int iflib_task_fn_rxs;
644 static int iflib_rx_intr_enables;
645 static int iflib_fast_intrs;
646 static int iflib_rx_unavail;
647 static int iflib_rx_ctx_inactive;
648 static int iflib_rx_if_input;
649 static int iflib_rxd_flush;
650 
651 static int iflib_verbose_debug;
652 
653 SYSCTL_INT(_net_iflib, OID_AUTO, task_fn_rx, CTLFLAG_RD,
654 		   &iflib_task_fn_rxs, 0, "# task_fn_rx calls");
655 SYSCTL_INT(_net_iflib, OID_AUTO, rx_intr_enables, CTLFLAG_RD,
656 		   &iflib_rx_intr_enables, 0, "# RX intr enables");
657 SYSCTL_INT(_net_iflib, OID_AUTO, fast_intrs, CTLFLAG_RD,
658 		   &iflib_fast_intrs, 0, "# fast_intr calls");
659 SYSCTL_INT(_net_iflib, OID_AUTO, rx_unavail, CTLFLAG_RD,
660 		   &iflib_rx_unavail, 0, "# times rxeof called with no available data");
661 SYSCTL_INT(_net_iflib, OID_AUTO, rx_ctx_inactive, CTLFLAG_RD,
662 		   &iflib_rx_ctx_inactive, 0, "# times rxeof called with inactive context");
663 SYSCTL_INT(_net_iflib, OID_AUTO, rx_if_input, CTLFLAG_RD,
664 		   &iflib_rx_if_input, 0, "# times rxeof called if_input");
665 SYSCTL_INT(_net_iflib, OID_AUTO, rxd_flush, CTLFLAG_RD,
666 	         &iflib_rxd_flush, 0, "# times rxd_flush called");
667 SYSCTL_INT(_net_iflib, OID_AUTO, verbose_debug, CTLFLAG_RW,
668 		   &iflib_verbose_debug, 0, "enable verbose debugging");
669 
670 #define DBG_COUNTER_INC(name) atomic_add_int(&(iflib_ ## name), 1)
671 static void
672 iflib_debug_reset(void)
673 {
674 	iflib_tx_seen = iflib_tx_sent = iflib_tx_encap = iflib_rx_allocs =
675 		iflib_fl_refills = iflib_fl_refills_large = iflib_tx_frees =
676 		iflib_txq_drain_flushing = iflib_txq_drain_oactive =
677 		iflib_txq_drain_notready =
678 		iflib_encap_load_mbuf_fail = iflib_encap_pad_mbuf_fail =
679 		iflib_encap_txq_avail_fail = iflib_encap_txd_encap_fail =
680 		iflib_task_fn_rxs = iflib_rx_intr_enables = iflib_fast_intrs =
681 		iflib_rx_unavail =
682 		iflib_rx_ctx_inactive = iflib_rx_if_input =
683 		iflib_rxd_flush = 0;
684 }
685 
686 #else
687 #define DBG_COUNTER_INC(name)
688 static void iflib_debug_reset(void) {}
689 #endif
690 
691 #define IFLIB_DEBUG 0
692 
693 static void iflib_tx_structures_free(if_ctx_t ctx);
694 static void iflib_rx_structures_free(if_ctx_t ctx);
695 static int iflib_queues_alloc(if_ctx_t ctx);
696 static int iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq);
697 static int iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget);
698 static int iflib_qset_structures_setup(if_ctx_t ctx);
699 static int iflib_msix_init(if_ctx_t ctx);
700 static int iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filterarg, int *rid, const char *str);
701 static void iflib_txq_check_drain(iflib_txq_t txq, int budget);
702 static uint32_t iflib_txq_can_drain(struct ifmp_ring *);
703 #ifdef ALTQ
704 static void iflib_altq_if_start(if_t ifp);
705 static int iflib_altq_if_transmit(if_t ifp, struct mbuf *m);
706 #endif
707 static int iflib_register(if_ctx_t);
708 static void iflib_deregister(if_ctx_t);
709 static void iflib_unregister_vlan_handlers(if_ctx_t ctx);
710 static uint16_t iflib_get_mbuf_size_for(unsigned int size);
711 static void iflib_init_locked(if_ctx_t ctx);
712 static void iflib_add_device_sysctl_pre(if_ctx_t ctx);
713 static void iflib_add_device_sysctl_post(if_ctx_t ctx);
714 static void iflib_ifmp_purge(iflib_txq_t txq);
715 static void _iflib_pre_assert(if_softc_ctx_t scctx);
716 static void iflib_if_init_locked(if_ctx_t ctx);
717 static void iflib_free_intr_mem(if_ctx_t ctx);
718 #ifndef __NO_STRICT_ALIGNMENT
719 static struct mbuf * iflib_fixup_rx(struct mbuf *m);
720 #endif
721 
722 static SLIST_HEAD(cpu_offset_list, cpu_offset) cpu_offsets =
723     SLIST_HEAD_INITIALIZER(cpu_offsets);
724 struct cpu_offset {
725 	SLIST_ENTRY(cpu_offset) entries;
726 	cpuset_t	set;
727 	unsigned int	refcount;
728 	uint16_t	offset;
729 };
730 static struct mtx cpu_offset_mtx;
731 MTX_SYSINIT(iflib_cpu_offset, &cpu_offset_mtx, "iflib_cpu_offset lock",
732     MTX_DEF);
733 
734 DEBUGNET_DEFINE(iflib);
735 
736 #ifdef DEV_NETMAP
737 #include <sys/selinfo.h>
738 #include <net/netmap.h>
739 #include <dev/netmap/netmap_kern.h>
740 
741 MODULE_DEPEND(iflib, netmap, 1, 1, 1);
742 
743 static int netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, uint32_t nm_i, bool init);
744 
745 /*
746  * device-specific sysctl variables:
747  *
748  * iflib_crcstrip: 0: keep CRC in rx frames (default), 1: strip it.
749  *	During regular operations the CRC is stripped, but on some
750  *	hardware reception of frames not multiple of 64 is slower,
751  *	so using crcstrip=0 helps in benchmarks.
752  *
753  * iflib_rx_miss, iflib_rx_miss_bufs:
754  *	count packets that might be missed due to lost interrupts.
755  */
756 SYSCTL_DECL(_dev_netmap);
757 /*
758  * The xl driver by default strips CRCs and we do not override it.
759  */
760 
761 int iflib_crcstrip = 1;
762 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_crcstrip,
763     CTLFLAG_RW, &iflib_crcstrip, 1, "strip CRC on RX frames");
764 
765 int iflib_rx_miss, iflib_rx_miss_bufs;
766 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss,
767     CTLFLAG_RW, &iflib_rx_miss, 0, "potentially missed RX intr");
768 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss_bufs,
769     CTLFLAG_RW, &iflib_rx_miss_bufs, 0, "potentially missed RX intr bufs");
770 
771 /*
772  * Register/unregister. We are already under netmap lock.
773  * Only called on the first register or the last unregister.
774  */
775 static int
776 iflib_netmap_register(struct netmap_adapter *na, int onoff)
777 {
778 	if_t ifp = na->ifp;
779 	if_ctx_t ctx = ifp->if_softc;
780 	int status;
781 
782 	CTX_LOCK(ctx);
783 	IFDI_INTR_DISABLE(ctx);
784 
785 	/* Tell the stack that the interface is no longer active */
786 	ifp->if_drv_flags &= ~(IFF_DRV_RUNNING | IFF_DRV_OACTIVE);
787 
788 	if (!CTX_IS_VF(ctx))
789 		IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip);
790 
791 	iflib_stop(ctx);
792 
793 	/*
794 	 * Enable (or disable) netmap flags, and intercept (or restore)
795 	 * ifp->if_transmit. This is done once the device has been stopped
796 	 * to prevent race conditions.
797 	 */
798 	if (onoff) {
799 		nm_set_native_flags(na);
800 	} else {
801 		nm_clear_native_flags(na);
802 	}
803 
804 	iflib_init_locked(ctx);
805 	IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip); // XXX why twice ?
806 	status = ifp->if_drv_flags & IFF_DRV_RUNNING ? 0 : 1;
807 	if (status)
808 		nm_clear_native_flags(na);
809 	CTX_UNLOCK(ctx);
810 	return (status);
811 }
812 
813 static int
814 netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, uint32_t nm_i, bool init)
815 {
816 	struct netmap_adapter *na = kring->na;
817 	u_int const lim = kring->nkr_num_slots - 1;
818 	u_int head = kring->rhead;
819 	struct netmap_ring *ring = kring->ring;
820 	bus_dmamap_t *map;
821 	struct if_rxd_update iru;
822 	if_ctx_t ctx = rxq->ifr_ctx;
823 	iflib_fl_t fl = &rxq->ifr_fl[0];
824 	uint32_t refill_pidx, nic_i;
825 #if IFLIB_DEBUG_COUNTERS
826 	int rf_count = 0;
827 #endif
828 
829 	if (nm_i == head && __predict_true(!init))
830 		return 0;
831 	iru_init(&iru, rxq, 0 /* flid */);
832 	map = fl->ifl_sds.ifsd_map;
833 	refill_pidx = netmap_idx_k2n(kring, nm_i);
834 	/*
835 	 * IMPORTANT: we must leave one free slot in the ring,
836 	 * so move head back by one unit
837 	 */
838 	head = nm_prev(head, lim);
839 	nic_i = UINT_MAX;
840 	DBG_COUNTER_INC(fl_refills);
841 	while (nm_i != head) {
842 #if IFLIB_DEBUG_COUNTERS
843 		if (++rf_count == 9)
844 			DBG_COUNTER_INC(fl_refills_large);
845 #endif
846 		for (int tmp_pidx = 0; tmp_pidx < IFLIB_MAX_RX_REFRESH && nm_i != head; tmp_pidx++) {
847 			struct netmap_slot *slot = &ring->slot[nm_i];
848 			void *addr = PNMB(na, slot, &fl->ifl_bus_addrs[tmp_pidx]);
849 			uint32_t nic_i_dma = refill_pidx;
850 			nic_i = netmap_idx_k2n(kring, nm_i);
851 
852 			MPASS(tmp_pidx < IFLIB_MAX_RX_REFRESH);
853 
854 			if (addr == NETMAP_BUF_BASE(na)) /* bad buf */
855 			        return netmap_ring_reinit(kring);
856 
857 			fl->ifl_vm_addrs[tmp_pidx] = addr;
858 			if (__predict_false(init)) {
859 				netmap_load_map(na, fl->ifl_buf_tag,
860 				    map[nic_i], addr);
861 			} else if (slot->flags & NS_BUF_CHANGED) {
862 				/* buffer has changed, reload map */
863 				netmap_reload_map(na, fl->ifl_buf_tag,
864 				    map[nic_i], addr);
865 			}
866 			slot->flags &= ~NS_BUF_CHANGED;
867 
868 			nm_i = nm_next(nm_i, lim);
869 			fl->ifl_rxd_idxs[tmp_pidx] = nic_i = nm_next(nic_i, lim);
870 			if (nm_i != head && tmp_pidx < IFLIB_MAX_RX_REFRESH-1)
871 				continue;
872 
873 			iru.iru_pidx = refill_pidx;
874 			iru.iru_count = tmp_pidx+1;
875 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
876 			refill_pidx = nic_i;
877 			for (int n = 0; n < iru.iru_count; n++) {
878 				bus_dmamap_sync(fl->ifl_buf_tag, map[nic_i_dma],
879 						BUS_DMASYNC_PREREAD);
880 				/* XXX - change this to not use the netmap func*/
881 				nic_i_dma = nm_next(nic_i_dma, lim);
882 			}
883 		}
884 	}
885 	kring->nr_hwcur = head;
886 
887 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
888 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
889 	if (__predict_true(nic_i != UINT_MAX)) {
890 		ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id, nic_i);
891 		DBG_COUNTER_INC(rxd_flush);
892 	}
893 	return (0);
894 }
895 
896 /*
897  * Reconcile kernel and user view of the transmit ring.
898  *
899  * All information is in the kring.
900  * Userspace wants to send packets up to the one before kring->rhead,
901  * kernel knows kring->nr_hwcur is the first unsent packet.
902  *
903  * Here we push packets out (as many as possible), and possibly
904  * reclaim buffers from previously completed transmission.
905  *
906  * The caller (netmap) guarantees that there is only one instance
907  * running at any time. Any interference with other driver
908  * methods should be handled by the individual drivers.
909  */
910 static int
911 iflib_netmap_txsync(struct netmap_kring *kring, int flags)
912 {
913 	struct netmap_adapter *na = kring->na;
914 	if_t ifp = na->ifp;
915 	struct netmap_ring *ring = kring->ring;
916 	u_int nm_i;	/* index into the netmap kring */
917 	u_int nic_i;	/* index into the NIC ring */
918 	u_int n;
919 	u_int const lim = kring->nkr_num_slots - 1;
920 	u_int const head = kring->rhead;
921 	struct if_pkt_info pi;
922 
923 	/*
924 	 * interrupts on every tx packet are expensive so request
925 	 * them every half ring, or where NS_REPORT is set
926 	 */
927 	u_int report_frequency = kring->nkr_num_slots >> 1;
928 	/* device-specific */
929 	if_ctx_t ctx = ifp->if_softc;
930 	iflib_txq_t txq = &ctx->ifc_txqs[kring->ring_id];
931 
932 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
933 	    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
934 
935 	/*
936 	 * First part: process new packets to send.
937 	 * nm_i is the current index in the netmap kring,
938 	 * nic_i is the corresponding index in the NIC ring.
939 	 *
940 	 * If we have packets to send (nm_i != head)
941 	 * iterate over the netmap ring, fetch length and update
942 	 * the corresponding slot in the NIC ring. Some drivers also
943 	 * need to update the buffer's physical address in the NIC slot
944 	 * even NS_BUF_CHANGED is not set (PNMB computes the addresses).
945 	 *
946 	 * The netmap_reload_map() calls is especially expensive,
947 	 * even when (as in this case) the tag is 0, so do only
948 	 * when the buffer has actually changed.
949 	 *
950 	 * If possible do not set the report/intr bit on all slots,
951 	 * but only a few times per ring or when NS_REPORT is set.
952 	 *
953 	 * Finally, on 10G and faster drivers, it might be useful
954 	 * to prefetch the next slot and txr entry.
955 	 */
956 
957 	nm_i = kring->nr_hwcur;
958 	if (nm_i != head) {	/* we have new packets to send */
959 		pkt_info_zero(&pi);
960 		pi.ipi_segs = txq->ift_segs;
961 		pi.ipi_qsidx = kring->ring_id;
962 		nic_i = netmap_idx_k2n(kring, nm_i);
963 
964 		__builtin_prefetch(&ring->slot[nm_i]);
965 		__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i]);
966 		__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i]);
967 
968 		for (n = 0; nm_i != head; n++) {
969 			struct netmap_slot *slot = &ring->slot[nm_i];
970 			u_int len = slot->len;
971 			uint64_t paddr;
972 			void *addr = PNMB(na, slot, &paddr);
973 			int flags = (slot->flags & NS_REPORT ||
974 				nic_i == 0 || nic_i == report_frequency) ?
975 				IPI_TX_INTR : 0;
976 
977 			/* device-specific */
978 			pi.ipi_len = len;
979 			pi.ipi_segs[0].ds_addr = paddr;
980 			pi.ipi_segs[0].ds_len = len;
981 			pi.ipi_nsegs = 1;
982 			pi.ipi_ndescs = 0;
983 			pi.ipi_pidx = nic_i;
984 			pi.ipi_flags = flags;
985 
986 			/* Fill the slot in the NIC ring. */
987 			ctx->isc_txd_encap(ctx->ifc_softc, &pi);
988 			DBG_COUNTER_INC(tx_encap);
989 
990 			/* prefetch for next round */
991 			__builtin_prefetch(&ring->slot[nm_i + 1]);
992 			__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i + 1]);
993 			__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i + 1]);
994 
995 			NM_CHECK_ADDR_LEN(na, addr, len);
996 
997 			if (slot->flags & NS_BUF_CHANGED) {
998 				/* buffer has changed, reload map */
999 				netmap_reload_map(na, txq->ift_buf_tag,
1000 				    txq->ift_sds.ifsd_map[nic_i], addr);
1001 			}
1002 			/* make sure changes to the buffer are synced */
1003 			bus_dmamap_sync(txq->ift_buf_tag,
1004 			    txq->ift_sds.ifsd_map[nic_i],
1005 			    BUS_DMASYNC_PREWRITE);
1006 
1007 			slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED);
1008 			nm_i = nm_next(nm_i, lim);
1009 			nic_i = nm_next(nic_i, lim);
1010 		}
1011 		kring->nr_hwcur = nm_i;
1012 
1013 		/* synchronize the NIC ring */
1014 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1015 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1016 
1017 		/* (re)start the tx unit up to slot nic_i (excluded) */
1018 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, nic_i);
1019 	}
1020 
1021 	/*
1022 	 * Second part: reclaim buffers for completed transmissions.
1023 	 *
1024 	 * If there are unclaimed buffers, attempt to reclaim them.
1025 	 * If none are reclaimed, and TX IRQs are not in use, do an initial
1026 	 * minimal delay, then trigger the tx handler which will spin in the
1027 	 * group task queue.
1028 	 */
1029 	if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1030 		if (iflib_tx_credits_update(ctx, txq)) {
1031 			/* some tx completed, increment avail */
1032 			nic_i = txq->ift_cidx_processed;
1033 			kring->nr_hwtail = nm_prev(netmap_idx_n2k(kring, nic_i), lim);
1034 		}
1035 	}
1036 	if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ))
1037 		if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1038 			callout_reset_on(&txq->ift_timer, hz < 2000 ? 1 : hz / 1000,
1039 			    iflib_timer, txq, txq->ift_timer.c_cpu);
1040 	}
1041 	return (0);
1042 }
1043 
1044 /*
1045  * Reconcile kernel and user view of the receive ring.
1046  * Same as for the txsync, this routine must be efficient.
1047  * The caller guarantees a single invocations, but races against
1048  * the rest of the driver should be handled here.
1049  *
1050  * On call, kring->rhead is the first packet that userspace wants
1051  * to keep, and kring->rcur is the wakeup point.
1052  * The kernel has previously reported packets up to kring->rtail.
1053  *
1054  * If (flags & NAF_FORCE_READ) also check for incoming packets irrespective
1055  * of whether or not we received an interrupt.
1056  */
1057 static int
1058 iflib_netmap_rxsync(struct netmap_kring *kring, int flags)
1059 {
1060 	struct netmap_adapter *na = kring->na;
1061 	struct netmap_ring *ring = kring->ring;
1062 	if_t ifp = na->ifp;
1063 	iflib_fl_t fl;
1064 	uint32_t nm_i;	/* index into the netmap ring */
1065 	uint32_t nic_i;	/* index into the NIC ring */
1066 	u_int i, n;
1067 	u_int const lim = kring->nkr_num_slots - 1;
1068 	u_int const head = kring->rhead;
1069 	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
1070 	struct if_rxd_info ri;
1071 
1072 	if_ctx_t ctx = ifp->if_softc;
1073 	iflib_rxq_t rxq = &ctx->ifc_rxqs[kring->ring_id];
1074 	if (head > lim)
1075 		return netmap_ring_reinit(kring);
1076 
1077 	/*
1078 	 * XXX netmap_fl_refill() only ever (re)fills free list 0 so far.
1079 	 */
1080 
1081 	for (i = 0, fl = rxq->ifr_fl; i < rxq->ifr_nfl; i++, fl++) {
1082 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
1083 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1084 	}
1085 
1086 	/*
1087 	 * First part: import newly received packets.
1088 	 *
1089 	 * nm_i is the index of the next free slot in the netmap ring,
1090 	 * nic_i is the index of the next received packet in the NIC ring,
1091 	 * and they may differ in case if_init() has been called while
1092 	 * in netmap mode. For the receive ring we have
1093 	 *
1094 	 *	nic_i = rxr->next_check;
1095 	 *	nm_i = kring->nr_hwtail (previous)
1096 	 * and
1097 	 *	nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1098 	 *
1099 	 * rxr->next_check is set to 0 on a ring reinit
1100 	 */
1101 	if (netmap_no_pendintr || force_update) {
1102 		uint32_t hwtail_lim = nm_prev(kring->nr_hwcur, lim);
1103 		int crclen = iflib_crcstrip ? 0 : 4;
1104 		int error, avail;
1105 
1106 		for (i = 0; i < rxq->ifr_nfl; i++) {
1107 			fl = &rxq->ifr_fl[i];
1108 			nic_i = fl->ifl_cidx;
1109 			nm_i = netmap_idx_n2k(kring, nic_i);
1110 			avail = ctx->isc_rxd_available(ctx->ifc_softc,
1111 			    rxq->ifr_id, nic_i, USHRT_MAX);
1112 			for (n = 0; avail > 0 && nm_i != hwtail_lim; n++, avail--) {
1113 				rxd_info_zero(&ri);
1114 				ri.iri_frags = rxq->ifr_frags;
1115 				ri.iri_qsidx = kring->ring_id;
1116 				ri.iri_ifp = ctx->ifc_ifp;
1117 				ri.iri_cidx = nic_i;
1118 
1119 				error = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
1120 				ring->slot[nm_i].len = error ? 0 : ri.iri_len - crclen;
1121 				ring->slot[nm_i].flags = 0;
1122 				bus_dmamap_sync(fl->ifl_buf_tag,
1123 				    fl->ifl_sds.ifsd_map[nic_i], BUS_DMASYNC_POSTREAD);
1124 				nm_i = nm_next(nm_i, lim);
1125 				nic_i = nm_next(nic_i, lim);
1126 			}
1127 			if (n) { /* update the state variables */
1128 				if (netmap_no_pendintr && !force_update) {
1129 					/* diagnostics */
1130 					iflib_rx_miss ++;
1131 					iflib_rx_miss_bufs += n;
1132 				}
1133 				fl->ifl_cidx = nic_i;
1134 				kring->nr_hwtail = nm_i;
1135 			}
1136 			kring->nr_kflags &= ~NKR_PENDINTR;
1137 		}
1138 	}
1139 	/*
1140 	 * Second part: skip past packets that userspace has released.
1141 	 * (kring->nr_hwcur to head excluded),
1142 	 * and make the buffers available for reception.
1143 	 * As usual nm_i is the index in the netmap ring,
1144 	 * nic_i is the index in the NIC ring, and
1145 	 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1146 	 */
1147 	/* XXX not sure how this will work with multiple free lists */
1148 	nm_i = kring->nr_hwcur;
1149 
1150 	return (netmap_fl_refill(rxq, kring, nm_i, false));
1151 }
1152 
1153 static void
1154 iflib_netmap_intr(struct netmap_adapter *na, int onoff)
1155 {
1156 	if_ctx_t ctx = na->ifp->if_softc;
1157 
1158 	CTX_LOCK(ctx);
1159 	if (onoff) {
1160 		IFDI_INTR_ENABLE(ctx);
1161 	} else {
1162 		IFDI_INTR_DISABLE(ctx);
1163 	}
1164 	CTX_UNLOCK(ctx);
1165 }
1166 
1167 
1168 static int
1169 iflib_netmap_attach(if_ctx_t ctx)
1170 {
1171 	struct netmap_adapter na;
1172 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1173 
1174 	bzero(&na, sizeof(na));
1175 
1176 	na.ifp = ctx->ifc_ifp;
1177 	na.na_flags = NAF_BDG_MAYSLEEP;
1178 	MPASS(ctx->ifc_softc_ctx.isc_ntxqsets);
1179 	MPASS(ctx->ifc_softc_ctx.isc_nrxqsets);
1180 
1181 	na.num_tx_desc = scctx->isc_ntxd[0];
1182 	na.num_rx_desc = scctx->isc_nrxd[0];
1183 	na.nm_txsync = iflib_netmap_txsync;
1184 	na.nm_rxsync = iflib_netmap_rxsync;
1185 	na.nm_register = iflib_netmap_register;
1186 	na.nm_intr = iflib_netmap_intr;
1187 	na.num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
1188 	na.num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
1189 	return (netmap_attach(&na));
1190 }
1191 
1192 static int
1193 iflib_netmap_txq_init(if_ctx_t ctx, iflib_txq_t txq)
1194 {
1195 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1196 	struct netmap_slot *slot;
1197 
1198 	slot = netmap_reset(na, NR_TX, txq->ift_id, 0);
1199 	if (slot == NULL)
1200 		return (0);
1201 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxd[0]; i++) {
1202 
1203 		/*
1204 		 * In netmap mode, set the map for the packet buffer.
1205 		 * NOTE: Some drivers (not this one) also need to set
1206 		 * the physical buffer address in the NIC ring.
1207 		 * netmap_idx_n2k() maps a nic index, i, into the corresponding
1208 		 * netmap slot index, si
1209 		 */
1210 		int si = netmap_idx_n2k(na->tx_rings[txq->ift_id], i);
1211 		netmap_load_map(na, txq->ift_buf_tag, txq->ift_sds.ifsd_map[i],
1212 		    NMB(na, slot + si));
1213 	}
1214 	return (1);
1215 }
1216 
1217 static int
1218 iflib_netmap_rxq_init(if_ctx_t ctx, iflib_rxq_t rxq)
1219 {
1220 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1221 	struct netmap_kring *kring;
1222 	struct netmap_slot *slot;
1223 	uint32_t nm_i;
1224 
1225 	slot = netmap_reset(na, NR_RX, rxq->ifr_id, 0);
1226 	if (slot == NULL)
1227 		return (0);
1228 	kring = na->rx_rings[rxq->ifr_id];
1229 	nm_i = netmap_idx_n2k(kring, 0);
1230 	netmap_fl_refill(rxq, kring, nm_i, true);
1231 	return (1);
1232 }
1233 
1234 static void
1235 iflib_netmap_timer_adjust(if_ctx_t ctx, iflib_txq_t txq, uint32_t *reset_on)
1236 {
1237 	struct netmap_kring *kring;
1238 	uint16_t txqid;
1239 
1240 	txqid = txq->ift_id;
1241 	kring = netmap_kring_on(NA(ctx->ifc_ifp), txqid, NR_TX);
1242 	if (kring == NULL)
1243 		return;
1244 
1245 	if (kring->nr_hwcur != nm_next(kring->nr_hwtail, kring->nkr_num_slots - 1)) {
1246 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1247 		    BUS_DMASYNC_POSTREAD);
1248 		if (ctx->isc_txd_credits_update(ctx->ifc_softc, txqid, false))
1249 			netmap_tx_irq(ctx->ifc_ifp, txqid);
1250 		if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ)) {
1251 			if (hz < 2000)
1252 				*reset_on = 1;
1253 			else
1254 				*reset_on = hz / 1000;
1255 		}
1256 	}
1257 }
1258 
1259 #define iflib_netmap_detach(ifp) netmap_detach(ifp)
1260 
1261 #else
1262 #define iflib_netmap_txq_init(ctx, txq) (0)
1263 #define iflib_netmap_rxq_init(ctx, rxq) (0)
1264 #define iflib_netmap_detach(ifp)
1265 
1266 #define iflib_netmap_attach(ctx) (0)
1267 #define netmap_rx_irq(ifp, qid, budget) (0)
1268 #define netmap_tx_irq(ifp, qid) do {} while (0)
1269 #define iflib_netmap_timer_adjust(ctx, txq, reset_on)
1270 #endif
1271 
1272 #if defined(__i386__) || defined(__amd64__)
1273 static __inline void
1274 prefetch(void *x)
1275 {
1276 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1277 }
1278 static __inline void
1279 prefetch2cachelines(void *x)
1280 {
1281 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1282 #if (CACHE_LINE_SIZE < 128)
1283 	__asm volatile("prefetcht0 %0" :: "m" (*(((unsigned long *)x)+CACHE_LINE_SIZE/(sizeof(unsigned long)))));
1284 #endif
1285 }
1286 #else
1287 #define prefetch(x)
1288 #define prefetch2cachelines(x)
1289 #endif
1290 
1291 static void
1292 iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid)
1293 {
1294 	iflib_fl_t fl;
1295 
1296 	fl = &rxq->ifr_fl[flid];
1297 	iru->iru_paddrs = fl->ifl_bus_addrs;
1298 	iru->iru_vaddrs = &fl->ifl_vm_addrs[0];
1299 	iru->iru_idxs = fl->ifl_rxd_idxs;
1300 	iru->iru_qsidx = rxq->ifr_id;
1301 	iru->iru_buf_size = fl->ifl_buf_size;
1302 	iru->iru_flidx = fl->ifl_id;
1303 }
1304 
1305 static void
1306 _iflib_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int err)
1307 {
1308 	if (err)
1309 		return;
1310 	*(bus_addr_t *) arg = segs[0].ds_addr;
1311 }
1312 
1313 int
1314 iflib_dma_alloc_align(if_ctx_t ctx, int size, int align, iflib_dma_info_t dma, int mapflags)
1315 {
1316 	int err;
1317 	device_t dev = ctx->ifc_dev;
1318 
1319 	err = bus_dma_tag_create(bus_get_dma_tag(dev),	/* parent */
1320 				align, 0,		/* alignment, bounds */
1321 				BUS_SPACE_MAXADDR,	/* lowaddr */
1322 				BUS_SPACE_MAXADDR,	/* highaddr */
1323 				NULL, NULL,		/* filter, filterarg */
1324 				size,			/* maxsize */
1325 				1,			/* nsegments */
1326 				size,			/* maxsegsize */
1327 				BUS_DMA_ALLOCNOW,	/* flags */
1328 				NULL,			/* lockfunc */
1329 				NULL,			/* lockarg */
1330 				&dma->idi_tag);
1331 	if (err) {
1332 		device_printf(dev,
1333 		    "%s: bus_dma_tag_create failed: %d\n",
1334 		    __func__, err);
1335 		goto fail_0;
1336 	}
1337 
1338 	err = bus_dmamem_alloc(dma->idi_tag, (void**) &dma->idi_vaddr,
1339 	    BUS_DMA_NOWAIT | BUS_DMA_COHERENT | BUS_DMA_ZERO, &dma->idi_map);
1340 	if (err) {
1341 		device_printf(dev,
1342 		    "%s: bus_dmamem_alloc(%ju) failed: %d\n",
1343 		    __func__, (uintmax_t)size, err);
1344 		goto fail_1;
1345 	}
1346 
1347 	dma->idi_paddr = IF_BAD_DMA;
1348 	err = bus_dmamap_load(dma->idi_tag, dma->idi_map, dma->idi_vaddr,
1349 	    size, _iflib_dmamap_cb, &dma->idi_paddr, mapflags | BUS_DMA_NOWAIT);
1350 	if (err || dma->idi_paddr == IF_BAD_DMA) {
1351 		device_printf(dev,
1352 		    "%s: bus_dmamap_load failed: %d\n",
1353 		    __func__, err);
1354 		goto fail_2;
1355 	}
1356 
1357 	dma->idi_size = size;
1358 	return (0);
1359 
1360 fail_2:
1361 	bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1362 fail_1:
1363 	bus_dma_tag_destroy(dma->idi_tag);
1364 fail_0:
1365 	dma->idi_tag = NULL;
1366 
1367 	return (err);
1368 }
1369 
1370 int
1371 iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags)
1372 {
1373 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1374 
1375 	KASSERT(sctx->isc_q_align != 0, ("alignment value not initialized"));
1376 
1377 	return (iflib_dma_alloc_align(ctx, size, sctx->isc_q_align, dma, mapflags));
1378 }
1379 
1380 int
1381 iflib_dma_alloc_multi(if_ctx_t ctx, int *sizes, iflib_dma_info_t *dmalist, int mapflags, int count)
1382 {
1383 	int i, err;
1384 	iflib_dma_info_t *dmaiter;
1385 
1386 	dmaiter = dmalist;
1387 	for (i = 0; i < count; i++, dmaiter++) {
1388 		if ((err = iflib_dma_alloc(ctx, sizes[i], *dmaiter, mapflags)) != 0)
1389 			break;
1390 	}
1391 	if (err)
1392 		iflib_dma_free_multi(dmalist, i);
1393 	return (err);
1394 }
1395 
1396 void
1397 iflib_dma_free(iflib_dma_info_t dma)
1398 {
1399 	if (dma->idi_tag == NULL)
1400 		return;
1401 	if (dma->idi_paddr != IF_BAD_DMA) {
1402 		bus_dmamap_sync(dma->idi_tag, dma->idi_map,
1403 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1404 		bus_dmamap_unload(dma->idi_tag, dma->idi_map);
1405 		dma->idi_paddr = IF_BAD_DMA;
1406 	}
1407 	if (dma->idi_vaddr != NULL) {
1408 		bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1409 		dma->idi_vaddr = NULL;
1410 	}
1411 	bus_dma_tag_destroy(dma->idi_tag);
1412 	dma->idi_tag = NULL;
1413 }
1414 
1415 void
1416 iflib_dma_free_multi(iflib_dma_info_t *dmalist, int count)
1417 {
1418 	int i;
1419 	iflib_dma_info_t *dmaiter = dmalist;
1420 
1421 	for (i = 0; i < count; i++, dmaiter++)
1422 		iflib_dma_free(*dmaiter);
1423 }
1424 
1425 static int
1426 iflib_fast_intr(void *arg)
1427 {
1428 	iflib_filter_info_t info = arg;
1429 	struct grouptask *gtask = info->ifi_task;
1430 	int result;
1431 
1432 	DBG_COUNTER_INC(fast_intrs);
1433 	if (info->ifi_filter != NULL) {
1434 		result = info->ifi_filter(info->ifi_filter_arg);
1435 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1436 			return (result);
1437 	}
1438 
1439 	GROUPTASK_ENQUEUE(gtask);
1440 	return (FILTER_HANDLED);
1441 }
1442 
1443 static int
1444 iflib_fast_intr_rxtx(void *arg)
1445 {
1446 	iflib_filter_info_t info = arg;
1447 	struct grouptask *gtask = info->ifi_task;
1448 	if_ctx_t ctx;
1449 	iflib_rxq_t rxq = (iflib_rxq_t)info->ifi_ctx;
1450 	iflib_txq_t txq;
1451 	void *sc;
1452 	int i, cidx, result;
1453 	qidx_t txqid;
1454 	bool intr_enable, intr_legacy;
1455 
1456 	DBG_COUNTER_INC(fast_intrs);
1457 	if (info->ifi_filter != NULL) {
1458 		result = info->ifi_filter(info->ifi_filter_arg);
1459 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1460 			return (result);
1461 	}
1462 
1463 	ctx = rxq->ifr_ctx;
1464 	sc = ctx->ifc_softc;
1465 	intr_enable = false;
1466 	intr_legacy = !!(ctx->ifc_flags & IFC_LEGACY);
1467 	MPASS(rxq->ifr_ntxqirq);
1468 	for (i = 0; i < rxq->ifr_ntxqirq; i++) {
1469 		txqid = rxq->ifr_txqid[i];
1470 		txq = &ctx->ifc_txqs[txqid];
1471 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1472 		    BUS_DMASYNC_POSTREAD);
1473 		if (!ctx->isc_txd_credits_update(sc, txqid, false)) {
1474 			if (intr_legacy)
1475 				intr_enable = true;
1476 			else
1477 				IFDI_TX_QUEUE_INTR_ENABLE(ctx, txqid);
1478 			continue;
1479 		}
1480 		GROUPTASK_ENQUEUE(&txq->ift_task);
1481 	}
1482 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_RXCQ)
1483 		cidx = rxq->ifr_cq_cidx;
1484 	else
1485 		cidx = rxq->ifr_fl[0].ifl_cidx;
1486 	if (iflib_rxd_avail(ctx, rxq, cidx, 1))
1487 		GROUPTASK_ENQUEUE(gtask);
1488 	else {
1489 		if (intr_legacy)
1490 			intr_enable = true;
1491 		else
1492 			IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
1493 		DBG_COUNTER_INC(rx_intr_enables);
1494 	}
1495 	if (intr_enable)
1496 		IFDI_INTR_ENABLE(ctx);
1497 	return (FILTER_HANDLED);
1498 }
1499 
1500 
1501 static int
1502 iflib_fast_intr_ctx(void *arg)
1503 {
1504 	iflib_filter_info_t info = arg;
1505 	struct grouptask *gtask = info->ifi_task;
1506 	int result;
1507 
1508 	DBG_COUNTER_INC(fast_intrs);
1509 	if (info->ifi_filter != NULL) {
1510 		result = info->ifi_filter(info->ifi_filter_arg);
1511 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1512 			return (result);
1513 	}
1514 
1515 	GROUPTASK_ENQUEUE(gtask);
1516 	return (FILTER_HANDLED);
1517 }
1518 
1519 static int
1520 _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
1521 		 driver_filter_t filter, driver_intr_t handler, void *arg,
1522 		 const char *name)
1523 {
1524 	struct resource *res;
1525 	void *tag = NULL;
1526 	device_t dev = ctx->ifc_dev;
1527 	int flags, i, rc;
1528 
1529 	flags = RF_ACTIVE;
1530 	if (ctx->ifc_flags & IFC_LEGACY)
1531 		flags |= RF_SHAREABLE;
1532 	MPASS(rid < 512);
1533 	i = rid;
1534 	res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &i, flags);
1535 	if (res == NULL) {
1536 		device_printf(dev,
1537 		    "failed to allocate IRQ for rid %d, name %s.\n", rid, name);
1538 		return (ENOMEM);
1539 	}
1540 	irq->ii_res = res;
1541 	KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL"));
1542 	rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET,
1543 						filter, handler, arg, &tag);
1544 	if (rc != 0) {
1545 		device_printf(dev,
1546 		    "failed to setup interrupt for rid %d, name %s: %d\n",
1547 					  rid, name ? name : "unknown", rc);
1548 		return (rc);
1549 	} else if (name)
1550 		bus_describe_intr(dev, res, tag, "%s", name);
1551 
1552 	irq->ii_tag = tag;
1553 	return (0);
1554 }
1555 
1556 /*********************************************************************
1557  *
1558  *  Allocate DMA resources for TX buffers as well as memory for the TX
1559  *  mbuf map.  TX DMA maps (non-TSO/TSO) and TX mbuf map are kept in a
1560  *  iflib_sw_tx_desc_array structure, storing all the information that
1561  *  is needed to transmit a packet on the wire.  This is called only
1562  *  once at attach, setup is done every reset.
1563  *
1564  **********************************************************************/
1565 static int
1566 iflib_txsd_alloc(iflib_txq_t txq)
1567 {
1568 	if_ctx_t ctx = txq->ift_ctx;
1569 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1570 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1571 	device_t dev = ctx->ifc_dev;
1572 	bus_size_t tsomaxsize;
1573 	int err, nsegments, ntsosegments;
1574 	bool tso;
1575 
1576 	nsegments = scctx->isc_tx_nsegments;
1577 	ntsosegments = scctx->isc_tx_tso_segments_max;
1578 	tsomaxsize = scctx->isc_tx_tso_size_max;
1579 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_VLAN_MTU)
1580 		tsomaxsize += sizeof(struct ether_vlan_header);
1581 	MPASS(scctx->isc_ntxd[0] > 0);
1582 	MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0);
1583 	MPASS(nsegments > 0);
1584 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) {
1585 		MPASS(ntsosegments > 0);
1586 		MPASS(sctx->isc_tso_maxsize >= tsomaxsize);
1587 	}
1588 
1589 	/*
1590 	 * Set up DMA tags for TX buffers.
1591 	 */
1592 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1593 			       1, 0,			/* alignment, bounds */
1594 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1595 			       BUS_SPACE_MAXADDR,	/* highaddr */
1596 			       NULL, NULL,		/* filter, filterarg */
1597 			       sctx->isc_tx_maxsize,		/* maxsize */
1598 			       nsegments,	/* nsegments */
1599 			       sctx->isc_tx_maxsegsize,	/* maxsegsize */
1600 			       0,			/* flags */
1601 			       NULL,			/* lockfunc */
1602 			       NULL,			/* lockfuncarg */
1603 			       &txq->ift_buf_tag))) {
1604 		device_printf(dev,"Unable to allocate TX DMA tag: %d\n", err);
1605 		device_printf(dev,"maxsize: %ju nsegments: %d maxsegsize: %ju\n",
1606 		    (uintmax_t)sctx->isc_tx_maxsize, nsegments, (uintmax_t)sctx->isc_tx_maxsegsize);
1607 		goto fail;
1608 	}
1609 	tso = (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) != 0;
1610 	if (tso && (err = bus_dma_tag_create(bus_get_dma_tag(dev),
1611 			       1, 0,			/* alignment, bounds */
1612 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1613 			       BUS_SPACE_MAXADDR,	/* highaddr */
1614 			       NULL, NULL,		/* filter, filterarg */
1615 			       tsomaxsize,		/* maxsize */
1616 			       ntsosegments,	/* nsegments */
1617 			       sctx->isc_tso_maxsegsize,/* maxsegsize */
1618 			       0,			/* flags */
1619 			       NULL,			/* lockfunc */
1620 			       NULL,			/* lockfuncarg */
1621 			       &txq->ift_tso_buf_tag))) {
1622 		device_printf(dev, "Unable to allocate TSO TX DMA tag: %d\n",
1623 		    err);
1624 		goto fail;
1625 	}
1626 
1627 	/* Allocate memory for the TX mbuf map. */
1628 	if (!(txq->ift_sds.ifsd_m =
1629 	    (struct mbuf **) malloc(sizeof(struct mbuf *) *
1630 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1631 		device_printf(dev, "Unable to allocate TX mbuf map memory\n");
1632 		err = ENOMEM;
1633 		goto fail;
1634 	}
1635 
1636 	/*
1637 	 * Create the DMA maps for TX buffers.
1638 	 */
1639 	if ((txq->ift_sds.ifsd_map = (bus_dmamap_t *)malloc(
1640 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1641 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1642 		device_printf(dev,
1643 		    "Unable to allocate TX buffer DMA map memory\n");
1644 		err = ENOMEM;
1645 		goto fail;
1646 	}
1647 	if (tso && (txq->ift_sds.ifsd_tso_map = (bus_dmamap_t *)malloc(
1648 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1649 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1650 		device_printf(dev,
1651 		    "Unable to allocate TSO TX buffer map memory\n");
1652 		err = ENOMEM;
1653 		goto fail;
1654 	}
1655 	for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) {
1656 		err = bus_dmamap_create(txq->ift_buf_tag, 0,
1657 		    &txq->ift_sds.ifsd_map[i]);
1658 		if (err != 0) {
1659 			device_printf(dev, "Unable to create TX DMA map\n");
1660 			goto fail;
1661 		}
1662 		if (!tso)
1663 			continue;
1664 		err = bus_dmamap_create(txq->ift_tso_buf_tag, 0,
1665 		    &txq->ift_sds.ifsd_tso_map[i]);
1666 		if (err != 0) {
1667 			device_printf(dev, "Unable to create TSO TX DMA map\n");
1668 			goto fail;
1669 		}
1670 	}
1671 	return (0);
1672 fail:
1673 	/* We free all, it handles case where we are in the middle */
1674 	iflib_tx_structures_free(ctx);
1675 	return (err);
1676 }
1677 
1678 static void
1679 iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i)
1680 {
1681 	bus_dmamap_t map;
1682 
1683 	if (txq->ift_sds.ifsd_map != NULL) {
1684 		map = txq->ift_sds.ifsd_map[i];
1685 		bus_dmamap_sync(txq->ift_buf_tag, map, BUS_DMASYNC_POSTWRITE);
1686 		bus_dmamap_unload(txq->ift_buf_tag, map);
1687 		bus_dmamap_destroy(txq->ift_buf_tag, map);
1688 		txq->ift_sds.ifsd_map[i] = NULL;
1689 	}
1690 
1691 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1692 		map = txq->ift_sds.ifsd_tso_map[i];
1693 		bus_dmamap_sync(txq->ift_tso_buf_tag, map,
1694 		    BUS_DMASYNC_POSTWRITE);
1695 		bus_dmamap_unload(txq->ift_tso_buf_tag, map);
1696 		bus_dmamap_destroy(txq->ift_tso_buf_tag, map);
1697 		txq->ift_sds.ifsd_tso_map[i] = NULL;
1698 	}
1699 }
1700 
1701 static void
1702 iflib_txq_destroy(iflib_txq_t txq)
1703 {
1704 	if_ctx_t ctx = txq->ift_ctx;
1705 
1706 	for (int i = 0; i < txq->ift_size; i++)
1707 		iflib_txsd_destroy(ctx, txq, i);
1708 
1709 	if (txq->ift_br != NULL) {
1710 		ifmp_ring_free(txq->ift_br);
1711 		txq->ift_br = NULL;
1712 	}
1713 
1714 	mtx_destroy(&txq->ift_mtx);
1715 
1716 	if (txq->ift_sds.ifsd_map != NULL) {
1717 		free(txq->ift_sds.ifsd_map, M_IFLIB);
1718 		txq->ift_sds.ifsd_map = NULL;
1719 	}
1720 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1721 		free(txq->ift_sds.ifsd_tso_map, M_IFLIB);
1722 		txq->ift_sds.ifsd_tso_map = NULL;
1723 	}
1724 	if (txq->ift_sds.ifsd_m != NULL) {
1725 		free(txq->ift_sds.ifsd_m, M_IFLIB);
1726 		txq->ift_sds.ifsd_m = NULL;
1727 	}
1728 	if (txq->ift_buf_tag != NULL) {
1729 		bus_dma_tag_destroy(txq->ift_buf_tag);
1730 		txq->ift_buf_tag = NULL;
1731 	}
1732 	if (txq->ift_tso_buf_tag != NULL) {
1733 		bus_dma_tag_destroy(txq->ift_tso_buf_tag);
1734 		txq->ift_tso_buf_tag = NULL;
1735 	}
1736 	if (txq->ift_ifdi != NULL) {
1737 		free(txq->ift_ifdi, M_IFLIB);
1738 	}
1739 }
1740 
1741 static void
1742 iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i)
1743 {
1744 	struct mbuf **mp;
1745 
1746 	mp = &txq->ift_sds.ifsd_m[i];
1747 	if (*mp == NULL)
1748 		return;
1749 
1750 	if (txq->ift_sds.ifsd_map != NULL) {
1751 		bus_dmamap_sync(txq->ift_buf_tag,
1752 		    txq->ift_sds.ifsd_map[i], BUS_DMASYNC_POSTWRITE);
1753 		bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[i]);
1754 	}
1755 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1756 		bus_dmamap_sync(txq->ift_tso_buf_tag,
1757 		    txq->ift_sds.ifsd_tso_map[i], BUS_DMASYNC_POSTWRITE);
1758 		bus_dmamap_unload(txq->ift_tso_buf_tag,
1759 		    txq->ift_sds.ifsd_tso_map[i]);
1760 	}
1761 	m_free(*mp);
1762 	DBG_COUNTER_INC(tx_frees);
1763 	*mp = NULL;
1764 }
1765 
1766 static int
1767 iflib_txq_setup(iflib_txq_t txq)
1768 {
1769 	if_ctx_t ctx = txq->ift_ctx;
1770 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1771 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1772 	iflib_dma_info_t di;
1773 	int i;
1774 
1775 	/* Set number of descriptors available */
1776 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
1777 	/* XXX make configurable */
1778 	txq->ift_update_freq = IFLIB_DEFAULT_TX_UPDATE_FREQ;
1779 
1780 	/* Reset indices */
1781 	txq->ift_cidx_processed = 0;
1782 	txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0;
1783 	txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset];
1784 
1785 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1786 		bzero((void *)di->idi_vaddr, di->idi_size);
1787 
1788 	IFDI_TXQ_SETUP(ctx, txq->ift_id);
1789 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1790 		bus_dmamap_sync(di->idi_tag, di->idi_map,
1791 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1792 	return (0);
1793 }
1794 
1795 /*********************************************************************
1796  *
1797  *  Allocate DMA resources for RX buffers as well as memory for the RX
1798  *  mbuf map, direct RX cluster pointer map and RX cluster bus address
1799  *  map.  RX DMA map, RX mbuf map, direct RX cluster pointer map and
1800  *  RX cluster map are kept in a iflib_sw_rx_desc_array structure.
1801  *  Since we use use one entry in iflib_sw_rx_desc_array per received
1802  *  packet, the maximum number of entries we'll need is equal to the
1803  *  number of hardware receive descriptors that we've allocated.
1804  *
1805  **********************************************************************/
1806 static int
1807 iflib_rxsd_alloc(iflib_rxq_t rxq)
1808 {
1809 	if_ctx_t ctx = rxq->ifr_ctx;
1810 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1811 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1812 	device_t dev = ctx->ifc_dev;
1813 	iflib_fl_t fl;
1814 	int			err;
1815 
1816 	MPASS(scctx->isc_nrxd[0] > 0);
1817 	MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0);
1818 
1819 	fl = rxq->ifr_fl;
1820 	for (int i = 0; i <  rxq->ifr_nfl; i++, fl++) {
1821 		fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */
1822 		/* Set up DMA tag for RX buffers. */
1823 		err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1824 					 1, 0,			/* alignment, bounds */
1825 					 BUS_SPACE_MAXADDR,	/* lowaddr */
1826 					 BUS_SPACE_MAXADDR,	/* highaddr */
1827 					 NULL, NULL,		/* filter, filterarg */
1828 					 sctx->isc_rx_maxsize,	/* maxsize */
1829 					 sctx->isc_rx_nsegments,	/* nsegments */
1830 					 sctx->isc_rx_maxsegsize,	/* maxsegsize */
1831 					 0,			/* flags */
1832 					 NULL,			/* lockfunc */
1833 					 NULL,			/* lockarg */
1834 					 &fl->ifl_buf_tag);
1835 		if (err) {
1836 			device_printf(dev,
1837 			    "Unable to allocate RX DMA tag: %d\n", err);
1838 			goto fail;
1839 		}
1840 
1841 		/* Allocate memory for the RX mbuf map. */
1842 		if (!(fl->ifl_sds.ifsd_m =
1843 		      (struct mbuf **) malloc(sizeof(struct mbuf *) *
1844 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1845 			device_printf(dev,
1846 			    "Unable to allocate RX mbuf map memory\n");
1847 			err = ENOMEM;
1848 			goto fail;
1849 		}
1850 
1851 		/* Allocate memory for the direct RX cluster pointer map. */
1852 		if (!(fl->ifl_sds.ifsd_cl =
1853 		      (caddr_t *) malloc(sizeof(caddr_t) *
1854 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1855 			device_printf(dev,
1856 			    "Unable to allocate RX cluster map memory\n");
1857 			err = ENOMEM;
1858 			goto fail;
1859 		}
1860 
1861 		/* Allocate memory for the RX cluster bus address map. */
1862 		if (!(fl->ifl_sds.ifsd_ba =
1863 		      (bus_addr_t *) malloc(sizeof(bus_addr_t) *
1864 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1865 			device_printf(dev,
1866 			    "Unable to allocate RX bus address map memory\n");
1867 			err = ENOMEM;
1868 			goto fail;
1869 		}
1870 
1871 		/*
1872 		 * Create the DMA maps for RX buffers.
1873 		 */
1874 		if (!(fl->ifl_sds.ifsd_map =
1875 		      (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1876 			device_printf(dev,
1877 			    "Unable to allocate RX buffer DMA map memory\n");
1878 			err = ENOMEM;
1879 			goto fail;
1880 		}
1881 		for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++) {
1882 			err = bus_dmamap_create(fl->ifl_buf_tag, 0,
1883 			    &fl->ifl_sds.ifsd_map[i]);
1884 			if (err != 0) {
1885 				device_printf(dev, "Unable to create RX buffer DMA map\n");
1886 				goto fail;
1887 			}
1888 		}
1889 	}
1890 	return (0);
1891 
1892 fail:
1893 	iflib_rx_structures_free(ctx);
1894 	return (err);
1895 }
1896 
1897 
1898 /*
1899  * Internal service routines
1900  */
1901 
1902 struct rxq_refill_cb_arg {
1903 	int               error;
1904 	bus_dma_segment_t seg;
1905 	int               nseg;
1906 };
1907 
1908 static void
1909 _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
1910 {
1911 	struct rxq_refill_cb_arg *cb_arg = arg;
1912 
1913 	cb_arg->error = error;
1914 	cb_arg->seg = segs[0];
1915 	cb_arg->nseg = nseg;
1916 }
1917 
1918 /**
1919  * _iflib_fl_refill - refill an rxq free-buffer list
1920  * @ctx: the iflib context
1921  * @fl: the free list to refill
1922  * @count: the number of new buffers to allocate
1923  *
1924  * (Re)populate an rxq free-buffer list with up to @count new packet buffers.
1925  * The caller must assure that @count does not exceed the queue's capacity.
1926  */
1927 static uint8_t
1928 _iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
1929 {
1930 	struct if_rxd_update iru;
1931 	struct rxq_refill_cb_arg cb_arg;
1932 	struct mbuf *m;
1933 	caddr_t cl, *sd_cl;
1934 	struct mbuf **sd_m;
1935 	bus_dmamap_t *sd_map;
1936 	bus_addr_t bus_addr, *sd_ba;
1937 	int err, frag_idx, i, idx, n, pidx;
1938 	qidx_t credits;
1939 
1940 	sd_m = fl->ifl_sds.ifsd_m;
1941 	sd_map = fl->ifl_sds.ifsd_map;
1942 	sd_cl = fl->ifl_sds.ifsd_cl;
1943 	sd_ba = fl->ifl_sds.ifsd_ba;
1944 	pidx = fl->ifl_pidx;
1945 	idx = pidx;
1946 	frag_idx = fl->ifl_fragidx;
1947 	credits = fl->ifl_credits;
1948 
1949 	i = 0;
1950 	n = count;
1951 	MPASS(n > 0);
1952 	MPASS(credits + n <= fl->ifl_size);
1953 
1954 	if (pidx < fl->ifl_cidx)
1955 		MPASS(pidx + n <= fl->ifl_cidx);
1956 	if (pidx == fl->ifl_cidx && (credits < fl->ifl_size))
1957 		MPASS(fl->ifl_gen == 0);
1958 	if (pidx > fl->ifl_cidx)
1959 		MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx);
1960 
1961 	DBG_COUNTER_INC(fl_refills);
1962 	if (n > 8)
1963 		DBG_COUNTER_INC(fl_refills_large);
1964 	iru_init(&iru, fl->ifl_rxq, fl->ifl_id);
1965 	while (n--) {
1966 		/*
1967 		 * We allocate an uninitialized mbuf + cluster, mbuf is
1968 		 * initialized after rx.
1969 		 *
1970 		 * If the cluster is still set then we know a minimum sized packet was received
1971 		 */
1972 		bit_ffc_at(fl->ifl_rx_bitmap, frag_idx, fl->ifl_size,
1973 		    &frag_idx);
1974 		if (frag_idx < 0)
1975 			bit_ffc(fl->ifl_rx_bitmap, fl->ifl_size, &frag_idx);
1976 		MPASS(frag_idx >= 0);
1977 		if ((cl = sd_cl[frag_idx]) == NULL) {
1978 			if ((cl = m_cljget(NULL, M_NOWAIT, fl->ifl_buf_size)) == NULL)
1979 				break;
1980 
1981 			cb_arg.error = 0;
1982 			MPASS(sd_map != NULL);
1983 			err = bus_dmamap_load(fl->ifl_buf_tag, sd_map[frag_idx],
1984 			    cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg,
1985 			    BUS_DMA_NOWAIT);
1986 			if (err != 0 || cb_arg.error) {
1987 				/*
1988 				 * !zone_pack ?
1989 				 */
1990 				if (fl->ifl_zone == zone_pack)
1991 					uma_zfree(fl->ifl_zone, cl);
1992 				break;
1993 			}
1994 
1995 			sd_ba[frag_idx] =  bus_addr = cb_arg.seg.ds_addr;
1996 			sd_cl[frag_idx] = cl;
1997 #if MEMORY_LOGGING
1998 			fl->ifl_cl_enqueued++;
1999 #endif
2000 		} else {
2001 			bus_addr = sd_ba[frag_idx];
2002 		}
2003 		bus_dmamap_sync(fl->ifl_buf_tag, sd_map[frag_idx],
2004 		    BUS_DMASYNC_PREREAD);
2005 
2006 		if (sd_m[frag_idx] == NULL) {
2007 			if ((m = m_gethdr(M_NOWAIT, MT_NOINIT)) == NULL) {
2008 				break;
2009 			}
2010 			sd_m[frag_idx] = m;
2011 		}
2012 		bit_set(fl->ifl_rx_bitmap, frag_idx);
2013 #if MEMORY_LOGGING
2014 		fl->ifl_m_enqueued++;
2015 #endif
2016 
2017 		DBG_COUNTER_INC(rx_allocs);
2018 		fl->ifl_rxd_idxs[i] = frag_idx;
2019 		fl->ifl_bus_addrs[i] = bus_addr;
2020 		fl->ifl_vm_addrs[i] = cl;
2021 		credits++;
2022 		i++;
2023 		MPASS(credits <= fl->ifl_size);
2024 		if (++idx == fl->ifl_size) {
2025 			fl->ifl_gen = 1;
2026 			idx = 0;
2027 		}
2028 		if (n == 0 || i == IFLIB_MAX_RX_REFRESH) {
2029 			iru.iru_pidx = pidx;
2030 			iru.iru_count = i;
2031 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2032 			i = 0;
2033 			pidx = idx;
2034 			fl->ifl_pidx = idx;
2035 			fl->ifl_credits = credits;
2036 		}
2037 	}
2038 
2039 	if (i) {
2040 		iru.iru_pidx = pidx;
2041 		iru.iru_count = i;
2042 		ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2043 		fl->ifl_pidx = idx;
2044 		fl->ifl_credits = credits;
2045 	}
2046 	DBG_COUNTER_INC(rxd_flush);
2047 	if (fl->ifl_pidx == 0)
2048 		pidx = fl->ifl_size - 1;
2049 	else
2050 		pidx = fl->ifl_pidx - 1;
2051 
2052 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2053 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2054 	ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id, fl->ifl_id, pidx);
2055 	fl->ifl_fragidx = frag_idx + 1;
2056 	if (fl->ifl_fragidx == fl->ifl_size)
2057 		fl->ifl_fragidx = 0;
2058 
2059 	return (n == -1 ? 0 : IFLIB_RXEOF_EMPTY);
2060 }
2061 
2062 static __inline uint8_t
2063 __iflib_fl_refill_all(if_ctx_t ctx, iflib_fl_t fl)
2064 {
2065 	/* we avoid allowing pidx to catch up with cidx as it confuses ixl */
2066 	int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1;
2067 #ifdef INVARIANTS
2068 	int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1;
2069 #endif
2070 
2071 	MPASS(fl->ifl_credits <= fl->ifl_size);
2072 	MPASS(reclaimable == delta);
2073 
2074 	if (reclaimable > 0)
2075 		return (_iflib_fl_refill(ctx, fl, reclaimable));
2076 	return (0);
2077 }
2078 
2079 uint8_t
2080 iflib_in_detach(if_ctx_t ctx)
2081 {
2082 	bool in_detach;
2083 
2084 	STATE_LOCK(ctx);
2085 	in_detach = !!(ctx->ifc_flags & IFC_IN_DETACH);
2086 	STATE_UNLOCK(ctx);
2087 	return (in_detach);
2088 }
2089 
2090 static void
2091 iflib_fl_bufs_free(iflib_fl_t fl)
2092 {
2093 	iflib_dma_info_t idi = fl->ifl_ifdi;
2094 	bus_dmamap_t sd_map;
2095 	uint32_t i;
2096 
2097 	for (i = 0; i < fl->ifl_size; i++) {
2098 		struct mbuf **sd_m = &fl->ifl_sds.ifsd_m[i];
2099 		caddr_t *sd_cl = &fl->ifl_sds.ifsd_cl[i];
2100 
2101 		if (*sd_cl != NULL) {
2102 			sd_map = fl->ifl_sds.ifsd_map[i];
2103 			bus_dmamap_sync(fl->ifl_buf_tag, sd_map,
2104 			    BUS_DMASYNC_POSTREAD);
2105 			bus_dmamap_unload(fl->ifl_buf_tag, sd_map);
2106 			if (*sd_cl != NULL)
2107 				uma_zfree(fl->ifl_zone, *sd_cl);
2108 			if (*sd_m != NULL) {
2109 				m_init(*sd_m, M_NOWAIT, MT_DATA, 0);
2110 				uma_zfree(zone_mbuf, *sd_m);
2111 			}
2112 		} else {
2113 			MPASS(*sd_cl == NULL);
2114 			MPASS(*sd_m == NULL);
2115 		}
2116 #if MEMORY_LOGGING
2117 		fl->ifl_m_dequeued++;
2118 		fl->ifl_cl_dequeued++;
2119 #endif
2120 		*sd_cl = NULL;
2121 		*sd_m = NULL;
2122 	}
2123 #ifdef INVARIANTS
2124 	for (i = 0; i < fl->ifl_size; i++) {
2125 		MPASS(fl->ifl_sds.ifsd_cl[i] == NULL);
2126 		MPASS(fl->ifl_sds.ifsd_m[i] == NULL);
2127 	}
2128 #endif
2129 	/*
2130 	 * Reset free list values
2131 	 */
2132 	fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = fl->ifl_fragidx = 0;
2133 	bzero(idi->idi_vaddr, idi->idi_size);
2134 }
2135 
2136 /*********************************************************************
2137  *
2138  *  Initialize a free list and its buffers.
2139  *
2140  **********************************************************************/
2141 static int
2142 iflib_fl_setup(iflib_fl_t fl)
2143 {
2144 	iflib_rxq_t rxq = fl->ifl_rxq;
2145 	if_ctx_t ctx = rxq->ifr_ctx;
2146 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2147 	int qidx;
2148 
2149 	bit_nclear(fl->ifl_rx_bitmap, 0, fl->ifl_size - 1);
2150 	/*
2151 	** Free current RX buffer structs and their mbufs
2152 	*/
2153 	iflib_fl_bufs_free(fl);
2154 	/* Now replenish the mbufs */
2155 	MPASS(fl->ifl_credits == 0);
2156 	qidx = rxq->ifr_fl_offset + fl->ifl_id;
2157 	if (scctx->isc_rxd_buf_size[qidx] != 0)
2158 		fl->ifl_buf_size = scctx->isc_rxd_buf_size[qidx];
2159 	else
2160 		fl->ifl_buf_size = ctx->ifc_rx_mbuf_sz;
2161 	/*
2162 	 * ifl_buf_size may be a driver-supplied value, so pull it up
2163 	 * to the selected mbuf size.
2164 	 */
2165 	fl->ifl_buf_size = iflib_get_mbuf_size_for(fl->ifl_buf_size);
2166 	if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size)
2167 		ctx->ifc_max_fl_buf_size = fl->ifl_buf_size;
2168 	fl->ifl_cltype = m_gettype(fl->ifl_buf_size);
2169 	fl->ifl_zone = m_getzone(fl->ifl_buf_size);
2170 
2171 
2172 	/* avoid pre-allocating zillions of clusters to an idle card
2173 	 * potentially speeding up attach
2174 	 */
2175 	(void) _iflib_fl_refill(ctx, fl, min(128, fl->ifl_size));
2176 	MPASS(min(128, fl->ifl_size) == fl->ifl_credits);
2177 	if (min(128, fl->ifl_size) != fl->ifl_credits)
2178 		return (ENOBUFS);
2179 	/*
2180 	 * handle failure
2181 	 */
2182 	MPASS(rxq != NULL);
2183 	MPASS(fl->ifl_ifdi != NULL);
2184 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2185 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2186 	return (0);
2187 }
2188 
2189 /*********************************************************************
2190  *
2191  *  Free receive ring data structures
2192  *
2193  **********************************************************************/
2194 static void
2195 iflib_rx_sds_free(iflib_rxq_t rxq)
2196 {
2197 	iflib_fl_t fl;
2198 	int i, j;
2199 
2200 	if (rxq->ifr_fl != NULL) {
2201 		for (i = 0; i < rxq->ifr_nfl; i++) {
2202 			fl = &rxq->ifr_fl[i];
2203 			if (fl->ifl_buf_tag != NULL) {
2204 				if (fl->ifl_sds.ifsd_map != NULL) {
2205 					for (j = 0; j < fl->ifl_size; j++) {
2206 						bus_dmamap_sync(
2207 						    fl->ifl_buf_tag,
2208 						    fl->ifl_sds.ifsd_map[j],
2209 						    BUS_DMASYNC_POSTREAD);
2210 						bus_dmamap_unload(
2211 						    fl->ifl_buf_tag,
2212 						    fl->ifl_sds.ifsd_map[j]);
2213 						bus_dmamap_destroy(
2214 						    fl->ifl_buf_tag,
2215 						    fl->ifl_sds.ifsd_map[j]);
2216 					}
2217 				}
2218 				bus_dma_tag_destroy(fl->ifl_buf_tag);
2219 				fl->ifl_buf_tag = NULL;
2220 			}
2221 			free(fl->ifl_sds.ifsd_m, M_IFLIB);
2222 			free(fl->ifl_sds.ifsd_cl, M_IFLIB);
2223 			free(fl->ifl_sds.ifsd_ba, M_IFLIB);
2224 			free(fl->ifl_sds.ifsd_map, M_IFLIB);
2225 			fl->ifl_sds.ifsd_m = NULL;
2226 			fl->ifl_sds.ifsd_cl = NULL;
2227 			fl->ifl_sds.ifsd_ba = NULL;
2228 			fl->ifl_sds.ifsd_map = NULL;
2229 		}
2230 		free(rxq->ifr_fl, M_IFLIB);
2231 		rxq->ifr_fl = NULL;
2232 		free(rxq->ifr_ifdi, M_IFLIB);
2233 		rxq->ifr_ifdi = NULL;
2234 		rxq->ifr_cq_cidx = 0;
2235 	}
2236 }
2237 
2238 /*
2239  * Timer routine
2240  */
2241 static void
2242 iflib_timer(void *arg)
2243 {
2244 	iflib_txq_t txq = arg;
2245 	if_ctx_t ctx = txq->ift_ctx;
2246 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2247 	uint64_t this_tick = ticks;
2248 	uint32_t reset_on = hz / 2;
2249 
2250 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
2251 		return;
2252 
2253 	/*
2254 	** Check on the state of the TX queue(s), this
2255 	** can be done without the lock because its RO
2256 	** and the HUNG state will be static if set.
2257 	*/
2258 	if (this_tick - txq->ift_last_timer_tick >= hz / 2) {
2259 		txq->ift_last_timer_tick = this_tick;
2260 		IFDI_TIMER(ctx, txq->ift_id);
2261 		if ((txq->ift_qstatus == IFLIB_QUEUE_HUNG) &&
2262 		    ((txq->ift_cleaned_prev == txq->ift_cleaned) ||
2263 		     (sctx->isc_pause_frames == 0)))
2264 			goto hung;
2265 
2266 		if (txq->ift_qstatus != IFLIB_QUEUE_IDLE &&
2267 		    ifmp_ring_is_stalled(txq->ift_br)) {
2268 			KASSERT(ctx->ifc_link_state == LINK_STATE_UP, ("queue can't be marked as hung if interface is down"));
2269 			txq->ift_qstatus = IFLIB_QUEUE_HUNG;
2270 		}
2271 		txq->ift_cleaned_prev = txq->ift_cleaned;
2272 	}
2273 #ifdef DEV_NETMAP
2274 	if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP)
2275 		iflib_netmap_timer_adjust(ctx, txq, &reset_on);
2276 #endif
2277 	/* handle any laggards */
2278 	if (txq->ift_db_pending)
2279 		GROUPTASK_ENQUEUE(&txq->ift_task);
2280 
2281 	sctx->isc_pause_frames = 0;
2282 	if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)
2283 		callout_reset_on(&txq->ift_timer, reset_on, iflib_timer, txq, txq->ift_timer.c_cpu);
2284 	return;
2285 
2286  hung:
2287 	device_printf(ctx->ifc_dev,
2288 	    "Watchdog timeout (TX: %d desc avail: %d pidx: %d) -- resetting\n",
2289 	    txq->ift_id, TXQ_AVAIL(txq), txq->ift_pidx);
2290 	STATE_LOCK(ctx);
2291 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2292 	ctx->ifc_flags |= (IFC_DO_WATCHDOG|IFC_DO_RESET);
2293 	iflib_admin_intr_deferred(ctx);
2294 	STATE_UNLOCK(ctx);
2295 }
2296 
2297 static uint16_t
2298 iflib_get_mbuf_size_for(unsigned int size)
2299 {
2300 
2301 	if (size <= MCLBYTES)
2302 		return (MCLBYTES);
2303 	else
2304 		return (MJUMPAGESIZE);
2305 }
2306 
2307 static void
2308 iflib_calc_rx_mbuf_sz(if_ctx_t ctx)
2309 {
2310 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2311 
2312 	/*
2313 	 * XXX don't set the max_frame_size to larger
2314 	 * than the hardware can handle
2315 	 */
2316 	ctx->ifc_rx_mbuf_sz =
2317 	    iflib_get_mbuf_size_for(sctx->isc_max_frame_size);
2318 }
2319 
2320 uint32_t
2321 iflib_get_rx_mbuf_sz(if_ctx_t ctx)
2322 {
2323 
2324 	return (ctx->ifc_rx_mbuf_sz);
2325 }
2326 
2327 static void
2328 iflib_init_locked(if_ctx_t ctx)
2329 {
2330 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2331 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2332 	if_t ifp = ctx->ifc_ifp;
2333 	iflib_fl_t fl;
2334 	iflib_txq_t txq;
2335 	iflib_rxq_t rxq;
2336 	int i, j, tx_ip_csum_flags, tx_ip6_csum_flags;
2337 
2338 	if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2339 	IFDI_INTR_DISABLE(ctx);
2340 
2341 	tx_ip_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP);
2342 	tx_ip6_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | CSUM_IP6_SCTP);
2343 	/* Set hardware offload abilities */
2344 	if_clearhwassist(ifp);
2345 	if (if_getcapenable(ifp) & IFCAP_TXCSUM)
2346 		if_sethwassistbits(ifp, tx_ip_csum_flags, 0);
2347 	if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6)
2348 		if_sethwassistbits(ifp,  tx_ip6_csum_flags, 0);
2349 	if (if_getcapenable(ifp) & IFCAP_TSO4)
2350 		if_sethwassistbits(ifp, CSUM_IP_TSO, 0);
2351 	if (if_getcapenable(ifp) & IFCAP_TSO6)
2352 		if_sethwassistbits(ifp, CSUM_IP6_TSO, 0);
2353 
2354 	for (i = 0, txq = ctx->ifc_txqs; i < sctx->isc_ntxqsets; i++, txq++) {
2355 		CALLOUT_LOCK(txq);
2356 		callout_stop(&txq->ift_timer);
2357 		CALLOUT_UNLOCK(txq);
2358 		iflib_netmap_txq_init(ctx, txq);
2359 	}
2360 
2361 	/*
2362 	 * Calculate a suitable Rx mbuf size prior to calling IFDI_INIT, so
2363 	 * that drivers can use the value when setting up the hardware receive
2364 	 * buffers.
2365 	 */
2366 	iflib_calc_rx_mbuf_sz(ctx);
2367 
2368 #ifdef INVARIANTS
2369 	i = if_getdrvflags(ifp);
2370 #endif
2371 	IFDI_INIT(ctx);
2372 	MPASS(if_getdrvflags(ifp) == i);
2373 	for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
2374 		if (iflib_netmap_rxq_init(ctx, rxq) > 0) {
2375 			/* This rxq is in netmap mode. Skip normal init. */
2376 			continue;
2377 		}
2378 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
2379 			if (iflib_fl_setup(fl)) {
2380 				device_printf(ctx->ifc_dev,
2381 				    "setting up free list %d failed - "
2382 				    "check cluster settings\n", j);
2383 				goto done;
2384 			}
2385 		}
2386 	}
2387 done:
2388 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
2389 	IFDI_INTR_ENABLE(ctx);
2390 	txq = ctx->ifc_txqs;
2391 	for (i = 0; i < sctx->isc_ntxqsets; i++, txq++)
2392 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq,
2393 			txq->ift_timer.c_cpu);
2394 }
2395 
2396 static int
2397 iflib_media_change(if_t ifp)
2398 {
2399 	if_ctx_t ctx = if_getsoftc(ifp);
2400 	int err;
2401 
2402 	CTX_LOCK(ctx);
2403 	if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0)
2404 		iflib_init_locked(ctx);
2405 	CTX_UNLOCK(ctx);
2406 	return (err);
2407 }
2408 
2409 static void
2410 iflib_media_status(if_t ifp, struct ifmediareq *ifmr)
2411 {
2412 	if_ctx_t ctx = if_getsoftc(ifp);
2413 
2414 	CTX_LOCK(ctx);
2415 	IFDI_UPDATE_ADMIN_STATUS(ctx);
2416 	IFDI_MEDIA_STATUS(ctx, ifmr);
2417 	CTX_UNLOCK(ctx);
2418 }
2419 
2420 void
2421 iflib_stop(if_ctx_t ctx)
2422 {
2423 	iflib_txq_t txq = ctx->ifc_txqs;
2424 	iflib_rxq_t rxq = ctx->ifc_rxqs;
2425 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2426 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2427 	iflib_dma_info_t di;
2428 	iflib_fl_t fl;
2429 	int i, j;
2430 
2431 	/* Tell the stack that the interface is no longer active */
2432 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2433 
2434 	IFDI_INTR_DISABLE(ctx);
2435 	DELAY(1000);
2436 	IFDI_STOP(ctx);
2437 	DELAY(1000);
2438 
2439 	iflib_debug_reset();
2440 	/* Wait for current tx queue users to exit to disarm watchdog timer. */
2441 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) {
2442 		/* make sure all transmitters have completed before proceeding XXX */
2443 
2444 		CALLOUT_LOCK(txq);
2445 		callout_stop(&txq->ift_timer);
2446 		CALLOUT_UNLOCK(txq);
2447 
2448 		/* clean any enqueued buffers */
2449 		iflib_ifmp_purge(txq);
2450 		/* Free any existing tx buffers. */
2451 		for (j = 0; j < txq->ift_size; j++) {
2452 			iflib_txsd_free(ctx, txq, j);
2453 		}
2454 		txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0;
2455 		txq->ift_in_use = txq->ift_gen = txq->ift_cidx = txq->ift_pidx = txq->ift_no_desc_avail = 0;
2456 		txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0;
2457 		txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0;
2458 		txq->ift_pullups = 0;
2459 		ifmp_ring_reset_stats(txq->ift_br);
2460 		for (j = 0, di = txq->ift_ifdi; j < sctx->isc_ntxqs; j++, di++)
2461 			bzero((void *)di->idi_vaddr, di->idi_size);
2462 	}
2463 	for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) {
2464 		/* make sure all transmitters have completed before proceeding XXX */
2465 
2466 		rxq->ifr_cq_cidx = 0;
2467 		for (j = 0, di = rxq->ifr_ifdi; j < sctx->isc_nrxqs; j++, di++)
2468 			bzero((void *)di->idi_vaddr, di->idi_size);
2469 		/* also resets the free lists pidx/cidx */
2470 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
2471 			iflib_fl_bufs_free(fl);
2472 	}
2473 }
2474 
2475 static inline caddr_t
2476 calc_next_rxd(iflib_fl_t fl, int cidx)
2477 {
2478 	qidx_t size;
2479 	int nrxd;
2480 	caddr_t start, end, cur, next;
2481 
2482 	nrxd = fl->ifl_size;
2483 	size = fl->ifl_rxd_size;
2484 	start = fl->ifl_ifdi->idi_vaddr;
2485 
2486 	if (__predict_false(size == 0))
2487 		return (start);
2488 	cur = start + size*cidx;
2489 	end = start + size*nrxd;
2490 	next = CACHE_PTR_NEXT(cur);
2491 	return (next < end ? next : start);
2492 }
2493 
2494 static inline void
2495 prefetch_pkts(iflib_fl_t fl, int cidx)
2496 {
2497 	int nextptr;
2498 	int nrxd = fl->ifl_size;
2499 	caddr_t next_rxd;
2500 
2501 
2502 	nextptr = (cidx + CACHE_PTR_INCREMENT) & (nrxd-1);
2503 	prefetch(&fl->ifl_sds.ifsd_m[nextptr]);
2504 	prefetch(&fl->ifl_sds.ifsd_cl[nextptr]);
2505 	next_rxd = calc_next_rxd(fl, cidx);
2506 	prefetch(next_rxd);
2507 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 1) & (nrxd-1)]);
2508 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 2) & (nrxd-1)]);
2509 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 3) & (nrxd-1)]);
2510 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 4) & (nrxd-1)]);
2511 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 1) & (nrxd-1)]);
2512 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 2) & (nrxd-1)]);
2513 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 3) & (nrxd-1)]);
2514 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 4) & (nrxd-1)]);
2515 }
2516 
2517 static struct mbuf *
2518 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, bool unload, if_rxsd_t sd,
2519     int *pf_rv, if_rxd_info_t ri)
2520 {
2521 	bus_dmamap_t map;
2522 	iflib_fl_t fl;
2523 	caddr_t payload;
2524 	struct mbuf *m;
2525 	int flid, cidx, len, next;
2526 
2527 	map = NULL;
2528 	flid = irf->irf_flid;
2529 	cidx = irf->irf_idx;
2530 	fl = &rxq->ifr_fl[flid];
2531 	sd->ifsd_fl = fl;
2532 	m = fl->ifl_sds.ifsd_m[cidx];
2533 	sd->ifsd_cl = &fl->ifl_sds.ifsd_cl[cidx];
2534 	fl->ifl_credits--;
2535 #if MEMORY_LOGGING
2536 	fl->ifl_m_dequeued++;
2537 #endif
2538 	if (rxq->ifr_ctx->ifc_flags & IFC_PREFETCH)
2539 		prefetch_pkts(fl, cidx);
2540 	next = (cidx + CACHE_PTR_INCREMENT) & (fl->ifl_size-1);
2541 	prefetch(&fl->ifl_sds.ifsd_map[next]);
2542 	map = fl->ifl_sds.ifsd_map[cidx];
2543 
2544 	bus_dmamap_sync(fl->ifl_buf_tag, map, BUS_DMASYNC_POSTREAD);
2545 
2546 	if (rxq->pfil != NULL && PFIL_HOOKED_IN(rxq->pfil) && pf_rv != NULL &&
2547 	    irf->irf_len != 0) {
2548 		payload  = *sd->ifsd_cl;
2549 		payload +=  ri->iri_pad;
2550 		len = ri->iri_len - ri->iri_pad;
2551 		*pf_rv = pfil_run_hooks(rxq->pfil, payload, ri->iri_ifp,
2552 		    len | PFIL_MEMPTR | PFIL_IN, NULL);
2553 		switch (*pf_rv) {
2554 		case PFIL_DROPPED:
2555 		case PFIL_CONSUMED:
2556 			/*
2557 			 * The filter ate it.  Everything is recycled.
2558 			 */
2559 			m = NULL;
2560 			unload = 0;
2561 			break;
2562 		case PFIL_REALLOCED:
2563 			/*
2564 			 * The filter copied it.  Everything is recycled.
2565 			 */
2566 			m = pfil_mem2mbuf(payload);
2567 			unload = 0;
2568 			break;
2569 		case PFIL_PASS:
2570 			/*
2571 			 * Filter said it was OK, so receive like
2572 			 * normal
2573 			 */
2574 			fl->ifl_sds.ifsd_m[cidx] = NULL;
2575 			break;
2576 		default:
2577 			MPASS(0);
2578 		}
2579 	} else {
2580 		fl->ifl_sds.ifsd_m[cidx] = NULL;
2581 		*pf_rv = PFIL_PASS;
2582 	}
2583 
2584 	if (unload && irf->irf_len != 0)
2585 		bus_dmamap_unload(fl->ifl_buf_tag, map);
2586 	fl->ifl_cidx = (fl->ifl_cidx + 1) & (fl->ifl_size-1);
2587 	if (__predict_false(fl->ifl_cidx == 0))
2588 		fl->ifl_gen = 0;
2589 	bit_clear(fl->ifl_rx_bitmap, cidx);
2590 	return (m);
2591 }
2592 
2593 static struct mbuf *
2594 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri, if_rxsd_t sd, int *pf_rv)
2595 {
2596 	struct mbuf *m, *mh, *mt;
2597 	caddr_t cl;
2598 	int  *pf_rv_ptr, flags, i, padlen;
2599 	bool consumed;
2600 
2601 	i = 0;
2602 	mh = NULL;
2603 	consumed = false;
2604 	*pf_rv = PFIL_PASS;
2605 	pf_rv_ptr = pf_rv;
2606 	do {
2607 		m = rxd_frag_to_sd(rxq, &ri->iri_frags[i], !consumed, sd,
2608 		    pf_rv_ptr, ri);
2609 
2610 		MPASS(*sd->ifsd_cl != NULL);
2611 
2612 		/*
2613 		 * Exclude zero-length frags & frags from
2614 		 * packets the filter has consumed or dropped
2615 		 */
2616 		if (ri->iri_frags[i].irf_len == 0 || consumed ||
2617 		    *pf_rv == PFIL_CONSUMED || *pf_rv == PFIL_DROPPED) {
2618 			if (mh == NULL) {
2619 				/* everything saved here */
2620 				consumed = true;
2621 				pf_rv_ptr = NULL;
2622 				continue;
2623 			}
2624 			/* XXX we can save the cluster here, but not the mbuf */
2625 			m_init(m, M_NOWAIT, MT_DATA, 0);
2626 			m_free(m);
2627 			continue;
2628 		}
2629 		if (mh == NULL) {
2630 			flags = M_PKTHDR|M_EXT;
2631 			mh = mt = m;
2632 			padlen = ri->iri_pad;
2633 		} else {
2634 			flags = M_EXT;
2635 			mt->m_next = m;
2636 			mt = m;
2637 			/* assuming padding is only on the first fragment */
2638 			padlen = 0;
2639 		}
2640 		cl = *sd->ifsd_cl;
2641 		*sd->ifsd_cl = NULL;
2642 
2643 		/* Can these two be made one ? */
2644 		m_init(m, M_NOWAIT, MT_DATA, flags);
2645 		m_cljset(m, cl, sd->ifsd_fl->ifl_cltype);
2646 		/*
2647 		 * These must follow m_init and m_cljset
2648 		 */
2649 		m->m_data += padlen;
2650 		ri->iri_len -= padlen;
2651 		m->m_len = ri->iri_frags[i].irf_len;
2652 	} while (++i < ri->iri_nfrags);
2653 
2654 	return (mh);
2655 }
2656 
2657 /*
2658  * Process one software descriptor
2659  */
2660 static struct mbuf *
2661 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri)
2662 {
2663 	struct if_rxsd sd;
2664 	struct mbuf *m;
2665 	int pf_rv;
2666 
2667 	/* should I merge this back in now that the two paths are basically duplicated? */
2668 	if (ri->iri_nfrags == 1 &&
2669 	    ri->iri_frags[0].irf_len != 0 &&
2670 	    ri->iri_frags[0].irf_len <= MIN(IFLIB_RX_COPY_THRESH, MHLEN)) {
2671 		m = rxd_frag_to_sd(rxq, &ri->iri_frags[0], false, &sd,
2672 		    &pf_rv, ri);
2673 		if (pf_rv != PFIL_PASS && pf_rv != PFIL_REALLOCED)
2674 			return (m);
2675 		if (pf_rv == PFIL_PASS) {
2676 			m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR);
2677 #ifndef __NO_STRICT_ALIGNMENT
2678 			if (!IP_ALIGNED(m))
2679 				m->m_data += 2;
2680 #endif
2681 			memcpy(m->m_data, *sd.ifsd_cl, ri->iri_len);
2682 			m->m_len = ri->iri_frags[0].irf_len;
2683 		}
2684 	} else {
2685 		m = assemble_segments(rxq, ri, &sd, &pf_rv);
2686 		if (m == NULL)
2687 			return (NULL);
2688 		if (pf_rv != PFIL_PASS && pf_rv != PFIL_REALLOCED)
2689 			return (m);
2690 	}
2691 	m->m_pkthdr.len = ri->iri_len;
2692 	m->m_pkthdr.rcvif = ri->iri_ifp;
2693 	m->m_flags |= ri->iri_flags;
2694 	m->m_pkthdr.ether_vtag = ri->iri_vtag;
2695 	m->m_pkthdr.flowid = ri->iri_flowid;
2696 	M_HASHTYPE_SET(m, ri->iri_rsstype);
2697 	m->m_pkthdr.csum_flags = ri->iri_csum_flags;
2698 	m->m_pkthdr.csum_data = ri->iri_csum_data;
2699 	return (m);
2700 }
2701 
2702 #if defined(INET6) || defined(INET)
2703 static void
2704 iflib_get_ip_forwarding(struct lro_ctrl *lc, bool *v4, bool *v6)
2705 {
2706 	CURVNET_SET(lc->ifp->if_vnet);
2707 #if defined(INET6)
2708 	*v6 = V_ip6_forwarding;
2709 #endif
2710 #if defined(INET)
2711 	*v4 = V_ipforwarding;
2712 #endif
2713 	CURVNET_RESTORE();
2714 }
2715 
2716 /*
2717  * Returns true if it's possible this packet could be LROed.
2718  * if it returns false, it is guaranteed that tcp_lro_rx()
2719  * would not return zero.
2720  */
2721 static bool
2722 iflib_check_lro_possible(struct mbuf *m, bool v4_forwarding, bool v6_forwarding)
2723 {
2724 	struct ether_header *eh;
2725 
2726 	eh = mtod(m, struct ether_header *);
2727 	switch (eh->ether_type) {
2728 #if defined(INET6)
2729 		case htons(ETHERTYPE_IPV6):
2730 			return (!v6_forwarding);
2731 #endif
2732 #if defined (INET)
2733 		case htons(ETHERTYPE_IP):
2734 			return (!v4_forwarding);
2735 #endif
2736 	}
2737 
2738 	return false;
2739 }
2740 #else
2741 static void
2742 iflib_get_ip_forwarding(struct lro_ctrl *lc __unused, bool *v4 __unused, bool *v6 __unused)
2743 {
2744 }
2745 #endif
2746 
2747 static void
2748 _task_fn_rx_watchdog(void *context)
2749 {
2750 	iflib_rxq_t rxq = context;
2751 
2752 	GROUPTASK_ENQUEUE(&rxq->ifr_task);
2753 }
2754 
2755 static uint8_t
2756 iflib_rxeof(iflib_rxq_t rxq, qidx_t budget)
2757 {
2758 	if_t ifp;
2759 	if_ctx_t ctx = rxq->ifr_ctx;
2760 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2761 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2762 	int avail, i;
2763 	qidx_t *cidxp;
2764 	struct if_rxd_info ri;
2765 	int err, budget_left, rx_bytes, rx_pkts;
2766 	iflib_fl_t fl;
2767 	int lro_enabled;
2768 	bool v4_forwarding, v6_forwarding, lro_possible;
2769 	uint8_t retval = 0;
2770 
2771 	/*
2772 	 * XXX early demux data packets so that if_input processing only handles
2773 	 * acks in interrupt context
2774 	 */
2775 	struct mbuf *m, *mh, *mt, *mf;
2776 
2777 	NET_EPOCH_ASSERT();
2778 
2779 	lro_possible = v4_forwarding = v6_forwarding = false;
2780 	ifp = ctx->ifc_ifp;
2781 	mh = mt = NULL;
2782 	MPASS(budget > 0);
2783 	rx_pkts	= rx_bytes = 0;
2784 	if (sctx->isc_flags & IFLIB_HAS_RXCQ)
2785 		cidxp = &rxq->ifr_cq_cidx;
2786 	else
2787 		cidxp = &rxq->ifr_fl[0].ifl_cidx;
2788 	if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) {
2789 		for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2790 			retval |= __iflib_fl_refill_all(ctx, fl);
2791 		DBG_COUNTER_INC(rx_unavail);
2792 		return (retval);
2793 	}
2794 
2795 	/* pfil needs the vnet to be set */
2796 	CURVNET_SET_QUIET(ifp->if_vnet);
2797 	for (budget_left = budget; budget_left > 0 && avail > 0;) {
2798 		if (__predict_false(!CTX_ACTIVE(ctx))) {
2799 			DBG_COUNTER_INC(rx_ctx_inactive);
2800 			break;
2801 		}
2802 		/*
2803 		 * Reset client set fields to their default values
2804 		 */
2805 		rxd_info_zero(&ri);
2806 		ri.iri_qsidx = rxq->ifr_id;
2807 		ri.iri_cidx = *cidxp;
2808 		ri.iri_ifp = ifp;
2809 		ri.iri_frags = rxq->ifr_frags;
2810 		err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
2811 
2812 		if (err)
2813 			goto err;
2814 		rx_pkts += 1;
2815 		rx_bytes += ri.iri_len;
2816 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
2817 			*cidxp = ri.iri_cidx;
2818 			/* Update our consumer index */
2819 			/* XXX NB: shurd - check if this is still safe */
2820 			while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0])
2821 				rxq->ifr_cq_cidx -= scctx->isc_nrxd[0];
2822 			/* was this only a completion queue message? */
2823 			if (__predict_false(ri.iri_nfrags == 0))
2824 				continue;
2825 		}
2826 		MPASS(ri.iri_nfrags != 0);
2827 		MPASS(ri.iri_len != 0);
2828 
2829 		/* will advance the cidx on the corresponding free lists */
2830 		m = iflib_rxd_pkt_get(rxq, &ri);
2831 		avail--;
2832 		budget_left--;
2833 		if (avail == 0 && budget_left)
2834 			avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left);
2835 
2836 		if (__predict_false(m == NULL))
2837 			continue;
2838 
2839 		/* imm_pkt: -- cxgb */
2840 		if (mh == NULL)
2841 			mh = mt = m;
2842 		else {
2843 			mt->m_nextpkt = m;
2844 			mt = m;
2845 		}
2846 	}
2847 	CURVNET_RESTORE();
2848 	/* make sure that we can refill faster than drain */
2849 	for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2850 		retval |= __iflib_fl_refill_all(ctx, fl);
2851 
2852 	lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO);
2853 	if (lro_enabled)
2854 		iflib_get_ip_forwarding(&rxq->ifr_lc, &v4_forwarding, &v6_forwarding);
2855 	mt = mf = NULL;
2856 	while (mh != NULL) {
2857 		m = mh;
2858 		mh = mh->m_nextpkt;
2859 		m->m_nextpkt = NULL;
2860 #ifndef __NO_STRICT_ALIGNMENT
2861 		if (!IP_ALIGNED(m) && (m = iflib_fixup_rx(m)) == NULL)
2862 			continue;
2863 #endif
2864 		rx_bytes += m->m_pkthdr.len;
2865 		rx_pkts++;
2866 #if defined(INET6) || defined(INET)
2867 		if (lro_enabled) {
2868 			if (!lro_possible) {
2869 				lro_possible = iflib_check_lro_possible(m, v4_forwarding, v6_forwarding);
2870 				if (lro_possible && mf != NULL) {
2871 					ifp->if_input(ifp, mf);
2872 					DBG_COUNTER_INC(rx_if_input);
2873 					mt = mf = NULL;
2874 				}
2875 			}
2876 			if ((m->m_pkthdr.csum_flags & (CSUM_L4_CALC|CSUM_L4_VALID)) ==
2877 			    (CSUM_L4_CALC|CSUM_L4_VALID)) {
2878 				if (lro_possible && tcp_lro_rx(&rxq->ifr_lc, m, 0) == 0)
2879 					continue;
2880 			}
2881 		}
2882 #endif
2883 		if (lro_possible) {
2884 			ifp->if_input(ifp, m);
2885 			DBG_COUNTER_INC(rx_if_input);
2886 			continue;
2887 		}
2888 
2889 		if (mf == NULL)
2890 			mf = m;
2891 		if (mt != NULL)
2892 			mt->m_nextpkt = m;
2893 		mt = m;
2894 	}
2895 	if (mf != NULL) {
2896 		ifp->if_input(ifp, mf);
2897 		DBG_COUNTER_INC(rx_if_input);
2898 	}
2899 
2900 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
2901 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
2902 
2903 	/*
2904 	 * Flush any outstanding LRO work
2905 	 */
2906 #if defined(INET6) || defined(INET)
2907 	tcp_lro_flush_all(&rxq->ifr_lc);
2908 #endif
2909 	if (avail != 0 || iflib_rxd_avail(ctx, rxq, *cidxp, 1) != 0)
2910 		retval |= IFLIB_RXEOF_MORE;
2911 	return (retval);
2912 err:
2913 	STATE_LOCK(ctx);
2914 	ctx->ifc_flags |= IFC_DO_RESET;
2915 	iflib_admin_intr_deferred(ctx);
2916 	STATE_UNLOCK(ctx);
2917 	return (0);
2918 }
2919 
2920 #define TXD_NOTIFY_COUNT(txq) (((txq)->ift_size / (txq)->ift_update_freq)-1)
2921 static inline qidx_t
2922 txq_max_db_deferred(iflib_txq_t txq, qidx_t in_use)
2923 {
2924 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2925 	qidx_t minthresh = txq->ift_size / 8;
2926 	if (in_use > 4*minthresh)
2927 		return (notify_count);
2928 	if (in_use > 2*minthresh)
2929 		return (notify_count >> 1);
2930 	if (in_use > minthresh)
2931 		return (notify_count >> 3);
2932 	return (0);
2933 }
2934 
2935 static inline qidx_t
2936 txq_max_rs_deferred(iflib_txq_t txq)
2937 {
2938 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2939 	qidx_t minthresh = txq->ift_size / 8;
2940 	if (txq->ift_in_use > 4*minthresh)
2941 		return (notify_count);
2942 	if (txq->ift_in_use > 2*minthresh)
2943 		return (notify_count >> 1);
2944 	if (txq->ift_in_use > minthresh)
2945 		return (notify_count >> 2);
2946 	return (2);
2947 }
2948 
2949 #define M_CSUM_FLAGS(m) ((m)->m_pkthdr.csum_flags)
2950 #define M_HAS_VLANTAG(m) (m->m_flags & M_VLANTAG)
2951 
2952 #define TXQ_MAX_DB_DEFERRED(txq, in_use) txq_max_db_deferred((txq), (in_use))
2953 #define TXQ_MAX_RS_DEFERRED(txq) txq_max_rs_deferred(txq)
2954 #define TXQ_MAX_DB_CONSUMED(size) (size >> 4)
2955 
2956 /* forward compatibility for cxgb */
2957 #define FIRST_QSET(ctx) 0
2958 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets)
2959 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets)
2960 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx))
2961 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments))
2962 
2963 /* XXX we should be setting this to something other than zero */
2964 #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh)
2965 #define	MAX_TX_DESC(ctx) max((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max, \
2966     (ctx)->ifc_softc_ctx.isc_tx_nsegments)
2967 
2968 static inline bool
2969 iflib_txd_db_check(if_ctx_t ctx, iflib_txq_t txq, int ring, qidx_t in_use)
2970 {
2971 	qidx_t dbval, max;
2972 	bool rang;
2973 
2974 	rang = false;
2975 	max = TXQ_MAX_DB_DEFERRED(txq, in_use);
2976 	if (ring || txq->ift_db_pending >= max) {
2977 		dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx;
2978 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
2979 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2980 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval);
2981 		txq->ift_db_pending = txq->ift_npending = 0;
2982 		rang = true;
2983 	}
2984 	return (rang);
2985 }
2986 
2987 #ifdef PKT_DEBUG
2988 static void
2989 print_pkt(if_pkt_info_t pi)
2990 {
2991 	printf("pi len:  %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n",
2992 	       pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx);
2993 	printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n",
2994 	       pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag);
2995 	printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n",
2996 	       pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto);
2997 }
2998 #endif
2999 
3000 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO)
3001 #define IS_TX_OFFLOAD4(pi) ((pi)->ipi_csum_flags & (CSUM_IP_TCP | CSUM_IP_TSO))
3002 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO)
3003 #define IS_TX_OFFLOAD6(pi) ((pi)->ipi_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_TSO))
3004 
3005 static int
3006 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp)
3007 {
3008 	if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx;
3009 	struct ether_vlan_header *eh;
3010 	struct mbuf *m;
3011 
3012 	m = *mp;
3013 	if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) &&
3014 	    M_WRITABLE(m) == 0) {
3015 		if ((m = m_dup(m, M_NOWAIT)) == NULL) {
3016 			return (ENOMEM);
3017 		} else {
3018 			m_freem(*mp);
3019 			DBG_COUNTER_INC(tx_frees);
3020 			*mp = m;
3021 		}
3022 	}
3023 
3024 	/*
3025 	 * Determine where frame payload starts.
3026 	 * Jump over vlan headers if already present,
3027 	 * helpful for QinQ too.
3028 	 */
3029 	if (__predict_false(m->m_len < sizeof(*eh))) {
3030 		txq->ift_pullups++;
3031 		if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL))
3032 			return (ENOMEM);
3033 	}
3034 	eh = mtod(m, struct ether_vlan_header *);
3035 	if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
3036 		pi->ipi_etype = ntohs(eh->evl_proto);
3037 		pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
3038 	} else {
3039 		pi->ipi_etype = ntohs(eh->evl_encap_proto);
3040 		pi->ipi_ehdrlen = ETHER_HDR_LEN;
3041 	}
3042 
3043 	switch (pi->ipi_etype) {
3044 #ifdef INET
3045 	case ETHERTYPE_IP:
3046 	{
3047 		struct mbuf *n;
3048 		struct ip *ip = NULL;
3049 		struct tcphdr *th = NULL;
3050 		int minthlen;
3051 
3052 		minthlen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip) + sizeof(*th));
3053 		if (__predict_false(m->m_len < minthlen)) {
3054 			/*
3055 			 * if this code bloat is causing too much of a hit
3056 			 * move it to a separate function and mark it noinline
3057 			 */
3058 			if (m->m_len == pi->ipi_ehdrlen) {
3059 				n = m->m_next;
3060 				MPASS(n);
3061 				if (n->m_len >= sizeof(*ip))  {
3062 					ip = (struct ip *)n->m_data;
3063 					if (n->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3064 						th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3065 				} else {
3066 					txq->ift_pullups++;
3067 					if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
3068 						return (ENOMEM);
3069 					ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3070 				}
3071 			} else {
3072 				txq->ift_pullups++;
3073 				if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
3074 					return (ENOMEM);
3075 				ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3076 				if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3077 					th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3078 			}
3079 		} else {
3080 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3081 			if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3082 				th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3083 		}
3084 		pi->ipi_ip_hlen = ip->ip_hl << 2;
3085 		pi->ipi_ipproto = ip->ip_p;
3086 		pi->ipi_flags |= IPI_TX_IPV4;
3087 
3088 		/* TCP checksum offload may require TCP header length */
3089 		if (IS_TX_OFFLOAD4(pi)) {
3090 			if (__predict_true(pi->ipi_ipproto == IPPROTO_TCP)) {
3091 				if (__predict_false(th == NULL)) {
3092 					txq->ift_pullups++;
3093 					if (__predict_false((m = m_pullup(m, (ip->ip_hl << 2) + sizeof(*th))) == NULL))
3094 						return (ENOMEM);
3095 					th = (struct tcphdr *)((caddr_t)ip + pi->ipi_ip_hlen);
3096 				}
3097 				pi->ipi_tcp_hflags = th->th_flags;
3098 				pi->ipi_tcp_hlen = th->th_off << 2;
3099 				pi->ipi_tcp_seq = th->th_seq;
3100 			}
3101 			if (IS_TSO4(pi)) {
3102 				if (__predict_false(ip->ip_p != IPPROTO_TCP))
3103 					return (ENXIO);
3104 				/*
3105 				 * TSO always requires hardware checksum offload.
3106 				 */
3107 				pi->ipi_csum_flags |= (CSUM_IP_TCP | CSUM_IP);
3108 				th->th_sum = in_pseudo(ip->ip_src.s_addr,
3109 						       ip->ip_dst.s_addr, htons(IPPROTO_TCP));
3110 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3111 				if (sctx->isc_flags & IFLIB_TSO_INIT_IP) {
3112 					ip->ip_sum = 0;
3113 					ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz);
3114 				}
3115 			}
3116 		}
3117 		if ((sctx->isc_flags & IFLIB_NEED_ZERO_CSUM) && (pi->ipi_csum_flags & CSUM_IP))
3118                        ip->ip_sum = 0;
3119 
3120 		break;
3121 	}
3122 #endif
3123 #ifdef INET6
3124 	case ETHERTYPE_IPV6:
3125 	{
3126 		struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
3127 		struct tcphdr *th;
3128 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
3129 
3130 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
3131 			txq->ift_pullups++;
3132 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
3133 				return (ENOMEM);
3134 		}
3135 		th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
3136 
3137 		/* XXX-BZ this will go badly in case of ext hdrs. */
3138 		pi->ipi_ipproto = ip6->ip6_nxt;
3139 		pi->ipi_flags |= IPI_TX_IPV6;
3140 
3141 		/* TCP checksum offload may require TCP header length */
3142 		if (IS_TX_OFFLOAD6(pi)) {
3143 			if (pi->ipi_ipproto == IPPROTO_TCP) {
3144 				if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) {
3145 					txq->ift_pullups++;
3146 					if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL))
3147 						return (ENOMEM);
3148 				}
3149 				pi->ipi_tcp_hflags = th->th_flags;
3150 				pi->ipi_tcp_hlen = th->th_off << 2;
3151 				pi->ipi_tcp_seq = th->th_seq;
3152 			}
3153 			if (IS_TSO6(pi)) {
3154 				if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP))
3155 					return (ENXIO);
3156 				/*
3157 				 * TSO always requires hardware checksum offload.
3158 				 */
3159 				pi->ipi_csum_flags |= CSUM_IP6_TCP;
3160 				th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0);
3161 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3162 			}
3163 		}
3164 		break;
3165 	}
3166 #endif
3167 	default:
3168 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
3169 		pi->ipi_ip_hlen = 0;
3170 		break;
3171 	}
3172 	*mp = m;
3173 
3174 	return (0);
3175 }
3176 
3177 /*
3178  * If dodgy hardware rejects the scatter gather chain we've handed it
3179  * we'll need to remove the mbuf chain from ifsg_m[] before we can add the
3180  * m_defrag'd mbufs
3181  */
3182 static __noinline struct mbuf *
3183 iflib_remove_mbuf(iflib_txq_t txq)
3184 {
3185 	int ntxd, pidx;
3186 	struct mbuf *m, **ifsd_m;
3187 
3188 	ifsd_m = txq->ift_sds.ifsd_m;
3189 	ntxd = txq->ift_size;
3190 	pidx = txq->ift_pidx & (ntxd - 1);
3191 	ifsd_m = txq->ift_sds.ifsd_m;
3192 	m = ifsd_m[pidx];
3193 	ifsd_m[pidx] = NULL;
3194 	bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[pidx]);
3195 	if (txq->ift_sds.ifsd_tso_map != NULL)
3196 		bus_dmamap_unload(txq->ift_tso_buf_tag,
3197 		    txq->ift_sds.ifsd_tso_map[pidx]);
3198 #if MEMORY_LOGGING
3199 	txq->ift_dequeued++;
3200 #endif
3201 	return (m);
3202 }
3203 
3204 static inline caddr_t
3205 calc_next_txd(iflib_txq_t txq, int cidx, uint8_t qid)
3206 {
3207 	qidx_t size;
3208 	int ntxd;
3209 	caddr_t start, end, cur, next;
3210 
3211 	ntxd = txq->ift_size;
3212 	size = txq->ift_txd_size[qid];
3213 	start = txq->ift_ifdi[qid].idi_vaddr;
3214 
3215 	if (__predict_false(size == 0))
3216 		return (start);
3217 	cur = start + size*cidx;
3218 	end = start + size*ntxd;
3219 	next = CACHE_PTR_NEXT(cur);
3220 	return (next < end ? next : start);
3221 }
3222 
3223 /*
3224  * Pad an mbuf to ensure a minimum ethernet frame size.
3225  * min_frame_size is the frame size (less CRC) to pad the mbuf to
3226  */
3227 static __noinline int
3228 iflib_ether_pad(device_t dev, struct mbuf **m_head, uint16_t min_frame_size)
3229 {
3230 	/*
3231 	 * 18 is enough bytes to pad an ARP packet to 46 bytes, and
3232 	 * and ARP message is the smallest common payload I can think of
3233 	 */
3234 	static char pad[18];	/* just zeros */
3235 	int n;
3236 	struct mbuf *new_head;
3237 
3238 	if (!M_WRITABLE(*m_head)) {
3239 		new_head = m_dup(*m_head, M_NOWAIT);
3240 		if (new_head == NULL) {
3241 			m_freem(*m_head);
3242 			device_printf(dev, "cannot pad short frame, m_dup() failed");
3243 			DBG_COUNTER_INC(encap_pad_mbuf_fail);
3244 			DBG_COUNTER_INC(tx_frees);
3245 			return ENOMEM;
3246 		}
3247 		m_freem(*m_head);
3248 		*m_head = new_head;
3249 	}
3250 
3251 	for (n = min_frame_size - (*m_head)->m_pkthdr.len;
3252 	     n > 0; n -= sizeof(pad))
3253 		if (!m_append(*m_head, min(n, sizeof(pad)), pad))
3254 			break;
3255 
3256 	if (n > 0) {
3257 		m_freem(*m_head);
3258 		device_printf(dev, "cannot pad short frame\n");
3259 		DBG_COUNTER_INC(encap_pad_mbuf_fail);
3260 		DBG_COUNTER_INC(tx_frees);
3261 		return (ENOBUFS);
3262 	}
3263 
3264 	return 0;
3265 }
3266 
3267 static int
3268 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp)
3269 {
3270 	if_ctx_t		ctx;
3271 	if_shared_ctx_t		sctx;
3272 	if_softc_ctx_t		scctx;
3273 	bus_dma_tag_t		buf_tag;
3274 	bus_dma_segment_t	*segs;
3275 	struct mbuf		*m_head, **ifsd_m;
3276 	void			*next_txd;
3277 	bus_dmamap_t		map;
3278 	struct if_pkt_info	pi;
3279 	int remap = 0;
3280 	int err, nsegs, ndesc, max_segs, pidx, cidx, next, ntxd;
3281 
3282 	ctx = txq->ift_ctx;
3283 	sctx = ctx->ifc_sctx;
3284 	scctx = &ctx->ifc_softc_ctx;
3285 	segs = txq->ift_segs;
3286 	ntxd = txq->ift_size;
3287 	m_head = *m_headp;
3288 	map = NULL;
3289 
3290 	/*
3291 	 * If we're doing TSO the next descriptor to clean may be quite far ahead
3292 	 */
3293 	cidx = txq->ift_cidx;
3294 	pidx = txq->ift_pidx;
3295 	if (ctx->ifc_flags & IFC_PREFETCH) {
3296 		next = (cidx + CACHE_PTR_INCREMENT) & (ntxd-1);
3297 		if (!(ctx->ifc_flags & IFLIB_HAS_TXCQ)) {
3298 			next_txd = calc_next_txd(txq, cidx, 0);
3299 			prefetch(next_txd);
3300 		}
3301 
3302 		/* prefetch the next cache line of mbuf pointers and flags */
3303 		prefetch(&txq->ift_sds.ifsd_m[next]);
3304 		prefetch(&txq->ift_sds.ifsd_map[next]);
3305 		next = (cidx + CACHE_LINE_SIZE) & (ntxd-1);
3306 	}
3307 	map = txq->ift_sds.ifsd_map[pidx];
3308 	ifsd_m = txq->ift_sds.ifsd_m;
3309 
3310 	if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
3311 		buf_tag = txq->ift_tso_buf_tag;
3312 		max_segs = scctx->isc_tx_tso_segments_max;
3313 		map = txq->ift_sds.ifsd_tso_map[pidx];
3314 		MPASS(buf_tag != NULL);
3315 		MPASS(max_segs > 0);
3316 	} else {
3317 		buf_tag = txq->ift_buf_tag;
3318 		max_segs = scctx->isc_tx_nsegments;
3319 		map = txq->ift_sds.ifsd_map[pidx];
3320 	}
3321 	if ((sctx->isc_flags & IFLIB_NEED_ETHER_PAD) &&
3322 	    __predict_false(m_head->m_pkthdr.len < scctx->isc_min_frame_size)) {
3323 		err = iflib_ether_pad(ctx->ifc_dev, m_headp, scctx->isc_min_frame_size);
3324 		if (err) {
3325 			DBG_COUNTER_INC(encap_txd_encap_fail);
3326 			return err;
3327 		}
3328 	}
3329 	m_head = *m_headp;
3330 
3331 	pkt_info_zero(&pi);
3332 	pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG|M_BCAST|M_MCAST));
3333 	pi.ipi_pidx = pidx;
3334 	pi.ipi_qsidx = txq->ift_id;
3335 	pi.ipi_len = m_head->m_pkthdr.len;
3336 	pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags;
3337 	pi.ipi_vtag = M_HAS_VLANTAG(m_head) ? m_head->m_pkthdr.ether_vtag : 0;
3338 
3339 	/* deliberate bitwise OR to make one condition */
3340 	if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) {
3341 		if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0)) {
3342 			DBG_COUNTER_INC(encap_txd_encap_fail);
3343 			return (err);
3344 		}
3345 		m_head = *m_headp;
3346 	}
3347 
3348 retry:
3349 	err = bus_dmamap_load_mbuf_sg(buf_tag, map, m_head, segs, &nsegs,
3350 	    BUS_DMA_NOWAIT);
3351 defrag:
3352 	if (__predict_false(err)) {
3353 		switch (err) {
3354 		case EFBIG:
3355 			/* try collapse once and defrag once */
3356 			if (remap == 0) {
3357 				m_head = m_collapse(*m_headp, M_NOWAIT, max_segs);
3358 				/* try defrag if collapsing fails */
3359 				if (m_head == NULL)
3360 					remap++;
3361 			}
3362 			if (remap == 1) {
3363 				txq->ift_mbuf_defrag++;
3364 				m_head = m_defrag(*m_headp, M_NOWAIT);
3365 			}
3366 			/*
3367 			 * remap should never be >1 unless bus_dmamap_load_mbuf_sg
3368 			 * failed to map an mbuf that was run through m_defrag
3369 			 */
3370 			MPASS(remap <= 1);
3371 			if (__predict_false(m_head == NULL || remap > 1))
3372 				goto defrag_failed;
3373 			remap++;
3374 			*m_headp = m_head;
3375 			goto retry;
3376 			break;
3377 		case ENOMEM:
3378 			txq->ift_no_tx_dma_setup++;
3379 			break;
3380 		default:
3381 			txq->ift_no_tx_dma_setup++;
3382 			m_freem(*m_headp);
3383 			DBG_COUNTER_INC(tx_frees);
3384 			*m_headp = NULL;
3385 			break;
3386 		}
3387 		txq->ift_map_failed++;
3388 		DBG_COUNTER_INC(encap_load_mbuf_fail);
3389 		DBG_COUNTER_INC(encap_txd_encap_fail);
3390 		return (err);
3391 	}
3392 	ifsd_m[pidx] = m_head;
3393 	/*
3394 	 * XXX assumes a 1 to 1 relationship between segments and
3395 	 *        descriptors - this does not hold true on all drivers, e.g.
3396 	 *        cxgb
3397 	 */
3398 	if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) {
3399 		txq->ift_no_desc_avail++;
3400 		bus_dmamap_unload(buf_tag, map);
3401 		DBG_COUNTER_INC(encap_txq_avail_fail);
3402 		DBG_COUNTER_INC(encap_txd_encap_fail);
3403 		if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0)
3404 			GROUPTASK_ENQUEUE(&txq->ift_task);
3405 		return (ENOBUFS);
3406 	}
3407 	/*
3408 	 * On Intel cards we can greatly reduce the number of TX interrupts
3409 	 * we see by only setting report status on every Nth descriptor.
3410 	 * However, this also means that the driver will need to keep track
3411 	 * of the descriptors that RS was set on to check them for the DD bit.
3412 	 */
3413 	txq->ift_rs_pending += nsegs + 1;
3414 	if (txq->ift_rs_pending > TXQ_MAX_RS_DEFERRED(txq) ||
3415 	     iflib_no_tx_batch || (TXQ_AVAIL(txq) - nsegs) <= MAX_TX_DESC(ctx) + 2) {
3416 		pi.ipi_flags |= IPI_TX_INTR;
3417 		txq->ift_rs_pending = 0;
3418 	}
3419 
3420 	pi.ipi_segs = segs;
3421 	pi.ipi_nsegs = nsegs;
3422 
3423 	MPASS(pidx >= 0 && pidx < txq->ift_size);
3424 #ifdef PKT_DEBUG
3425 	print_pkt(&pi);
3426 #endif
3427 	if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) {
3428 		bus_dmamap_sync(buf_tag, map, BUS_DMASYNC_PREWRITE);
3429 		DBG_COUNTER_INC(tx_encap);
3430 		MPASS(pi.ipi_new_pidx < txq->ift_size);
3431 
3432 		ndesc = pi.ipi_new_pidx - pi.ipi_pidx;
3433 		if (pi.ipi_new_pidx < pi.ipi_pidx) {
3434 			ndesc += txq->ift_size;
3435 			txq->ift_gen = 1;
3436 		}
3437 		/*
3438 		 * drivers can need as many as
3439 		 * two sentinels
3440 		 */
3441 		MPASS(ndesc <= pi.ipi_nsegs + 2);
3442 		MPASS(pi.ipi_new_pidx != pidx);
3443 		MPASS(ndesc > 0);
3444 		txq->ift_in_use += ndesc;
3445 
3446 		/*
3447 		 * We update the last software descriptor again here because there may
3448 		 * be a sentinel and/or there may be more mbufs than segments
3449 		 */
3450 		txq->ift_pidx = pi.ipi_new_pidx;
3451 		txq->ift_npending += pi.ipi_ndescs;
3452 	} else {
3453 		*m_headp = m_head = iflib_remove_mbuf(txq);
3454 		if (err == EFBIG) {
3455 			txq->ift_txd_encap_efbig++;
3456 			if (remap < 2) {
3457 				remap = 1;
3458 				goto defrag;
3459 			}
3460 		}
3461 		goto defrag_failed;
3462 	}
3463 	/*
3464 	 * err can't possibly be non-zero here, so we don't neet to test it
3465 	 * to see if we need to DBG_COUNTER_INC(encap_txd_encap_fail).
3466 	 */
3467 	return (err);
3468 
3469 defrag_failed:
3470 	txq->ift_mbuf_defrag_failed++;
3471 	txq->ift_map_failed++;
3472 	m_freem(*m_headp);
3473 	DBG_COUNTER_INC(tx_frees);
3474 	*m_headp = NULL;
3475 	DBG_COUNTER_INC(encap_txd_encap_fail);
3476 	return (ENOMEM);
3477 }
3478 
3479 static void
3480 iflib_tx_desc_free(iflib_txq_t txq, int n)
3481 {
3482 	uint32_t qsize, cidx, mask, gen;
3483 	struct mbuf *m, **ifsd_m;
3484 	bool do_prefetch;
3485 
3486 	cidx = txq->ift_cidx;
3487 	gen = txq->ift_gen;
3488 	qsize = txq->ift_size;
3489 	mask = qsize-1;
3490 	ifsd_m = txq->ift_sds.ifsd_m;
3491 	do_prefetch = (txq->ift_ctx->ifc_flags & IFC_PREFETCH);
3492 
3493 	while (n-- > 0) {
3494 		if (do_prefetch) {
3495 			prefetch(ifsd_m[(cidx + 3) & mask]);
3496 			prefetch(ifsd_m[(cidx + 4) & mask]);
3497 		}
3498 		if ((m = ifsd_m[cidx]) != NULL) {
3499 			prefetch(&ifsd_m[(cidx + CACHE_PTR_INCREMENT) & mask]);
3500 			if (m->m_pkthdr.csum_flags & CSUM_TSO) {
3501 				bus_dmamap_sync(txq->ift_tso_buf_tag,
3502 				    txq->ift_sds.ifsd_tso_map[cidx],
3503 				    BUS_DMASYNC_POSTWRITE);
3504 				bus_dmamap_unload(txq->ift_tso_buf_tag,
3505 				    txq->ift_sds.ifsd_tso_map[cidx]);
3506 			} else {
3507 				bus_dmamap_sync(txq->ift_buf_tag,
3508 				    txq->ift_sds.ifsd_map[cidx],
3509 				    BUS_DMASYNC_POSTWRITE);
3510 				bus_dmamap_unload(txq->ift_buf_tag,
3511 				    txq->ift_sds.ifsd_map[cidx]);
3512 			}
3513 			/* XXX we don't support any drivers that batch packets yet */
3514 			MPASS(m->m_nextpkt == NULL);
3515 			m_freem(m);
3516 			ifsd_m[cidx] = NULL;
3517 #if MEMORY_LOGGING
3518 			txq->ift_dequeued++;
3519 #endif
3520 			DBG_COUNTER_INC(tx_frees);
3521 		}
3522 		if (__predict_false(++cidx == qsize)) {
3523 			cidx = 0;
3524 			gen = 0;
3525 		}
3526 	}
3527 	txq->ift_cidx = cidx;
3528 	txq->ift_gen = gen;
3529 }
3530 
3531 static __inline int
3532 iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh)
3533 {
3534 	int reclaim;
3535 	if_ctx_t ctx = txq->ift_ctx;
3536 
3537 	KASSERT(thresh >= 0, ("invalid threshold to reclaim"));
3538 	MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size);
3539 
3540 	/*
3541 	 * Need a rate-limiting check so that this isn't called every time
3542 	 */
3543 	iflib_tx_credits_update(ctx, txq);
3544 	reclaim = DESC_RECLAIMABLE(txq);
3545 
3546 	if (reclaim <= thresh /* + MAX_TX_DESC(txq->ift_ctx) */) {
3547 #ifdef INVARIANTS
3548 		if (iflib_verbose_debug) {
3549 			printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __FUNCTION__,
3550 			       txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments,
3551 			       reclaim, thresh);
3552 
3553 		}
3554 #endif
3555 		return (0);
3556 	}
3557 	iflib_tx_desc_free(txq, reclaim);
3558 	txq->ift_cleaned += reclaim;
3559 	txq->ift_in_use -= reclaim;
3560 
3561 	return (reclaim);
3562 }
3563 
3564 static struct mbuf **
3565 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset, int remaining)
3566 {
3567 	int next, size;
3568 	struct mbuf **items;
3569 
3570 	size = r->size;
3571 	next = (cidx + CACHE_PTR_INCREMENT) & (size-1);
3572 	items = __DEVOLATILE(struct mbuf **, &r->items[0]);
3573 
3574 	prefetch(items[(cidx + offset) & (size-1)]);
3575 	if (remaining > 1) {
3576 		prefetch2cachelines(&items[next]);
3577 		prefetch2cachelines(items[(cidx + offset + 1) & (size-1)]);
3578 		prefetch2cachelines(items[(cidx + offset + 2) & (size-1)]);
3579 		prefetch2cachelines(items[(cidx + offset + 3) & (size-1)]);
3580 	}
3581 	return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (size-1)]));
3582 }
3583 
3584 static void
3585 iflib_txq_check_drain(iflib_txq_t txq, int budget)
3586 {
3587 
3588 	ifmp_ring_check_drainage(txq->ift_br, budget);
3589 }
3590 
3591 static uint32_t
3592 iflib_txq_can_drain(struct ifmp_ring *r)
3593 {
3594 	iflib_txq_t txq = r->cookie;
3595 	if_ctx_t ctx = txq->ift_ctx;
3596 
3597 	if (TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2)
3598 		return (1);
3599 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3600 	    BUS_DMASYNC_POSTREAD);
3601 	return (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id,
3602 	    false));
3603 }
3604 
3605 static uint32_t
3606 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3607 {
3608 	iflib_txq_t txq = r->cookie;
3609 	if_ctx_t ctx = txq->ift_ctx;
3610 	if_t ifp = ctx->ifc_ifp;
3611 	struct mbuf *m, **mp;
3612 	int avail, bytes_sent, consumed, count, err, i, in_use_prev;
3613 	int mcast_sent, pkt_sent, reclaimed, txq_avail;
3614 	bool do_prefetch, rang, ring;
3615 
3616 	if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) ||
3617 			    !LINK_ACTIVE(ctx))) {
3618 		DBG_COUNTER_INC(txq_drain_notready);
3619 		return (0);
3620 	}
3621 	reclaimed = iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
3622 	rang = iflib_txd_db_check(ctx, txq, reclaimed, txq->ift_in_use);
3623 	avail = IDXDIFF(pidx, cidx, r->size);
3624 	if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) {
3625 		DBG_COUNTER_INC(txq_drain_flushing);
3626 		for (i = 0; i < avail; i++) {
3627 			if (__predict_true(r->items[(cidx + i) & (r->size-1)] != (void *)txq))
3628 				m_free(r->items[(cidx + i) & (r->size-1)]);
3629 			r->items[(cidx + i) & (r->size-1)] = NULL;
3630 		}
3631 		return (avail);
3632 	}
3633 
3634 	if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3635 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3636 		CALLOUT_LOCK(txq);
3637 		callout_stop(&txq->ift_timer);
3638 		CALLOUT_UNLOCK(txq);
3639 		DBG_COUNTER_INC(txq_drain_oactive);
3640 		return (0);
3641 	}
3642 	if (reclaimed)
3643 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3644 	consumed = mcast_sent = bytes_sent = pkt_sent = 0;
3645 	count = MIN(avail, TX_BATCH_SIZE);
3646 #ifdef INVARIANTS
3647 	if (iflib_verbose_debug)
3648 		printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __FUNCTION__,
3649 		       avail, ctx->ifc_flags, TXQ_AVAIL(txq));
3650 #endif
3651 	do_prefetch = (ctx->ifc_flags & IFC_PREFETCH);
3652 	txq_avail = TXQ_AVAIL(txq);
3653 	err = 0;
3654 	for (i = 0; i < count && txq_avail > MAX_TX_DESC(ctx) + 2; i++) {
3655 		int rem = do_prefetch ? count - i : 0;
3656 
3657 		mp = _ring_peek_one(r, cidx, i, rem);
3658 		MPASS(mp != NULL && *mp != NULL);
3659 		if (__predict_false(*mp == (struct mbuf *)txq)) {
3660 			consumed++;
3661 			continue;
3662 		}
3663 		in_use_prev = txq->ift_in_use;
3664 		err = iflib_encap(txq, mp);
3665 		if (__predict_false(err)) {
3666 			/* no room - bail out */
3667 			if (err == ENOBUFS)
3668 				break;
3669 			consumed++;
3670 			/* we can't send this packet - skip it */
3671 			continue;
3672 		}
3673 		consumed++;
3674 		pkt_sent++;
3675 		m = *mp;
3676 		DBG_COUNTER_INC(tx_sent);
3677 		bytes_sent += m->m_pkthdr.len;
3678 		mcast_sent += !!(m->m_flags & M_MCAST);
3679 		txq_avail = TXQ_AVAIL(txq);
3680 
3681 		txq->ift_db_pending += (txq->ift_in_use - in_use_prev);
3682 		ETHER_BPF_MTAP(ifp, m);
3683 		if (__predict_false(!(ifp->if_drv_flags & IFF_DRV_RUNNING)))
3684 			break;
3685 		rang = iflib_txd_db_check(ctx, txq, false, in_use_prev);
3686 	}
3687 
3688 	/* deliberate use of bitwise or to avoid gratuitous short-circuit */
3689 	ring = rang ? false  : (iflib_min_tx_latency | err) || (TXQ_AVAIL(txq) < MAX_TX_DESC(ctx));
3690 	iflib_txd_db_check(ctx, txq, ring, txq->ift_in_use);
3691 	if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
3692 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
3693 	if (mcast_sent)
3694 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
3695 #ifdef INVARIANTS
3696 	if (iflib_verbose_debug)
3697 		printf("consumed=%d\n", consumed);
3698 #endif
3699 	return (consumed);
3700 }
3701 
3702 static uint32_t
3703 iflib_txq_drain_always(struct ifmp_ring *r)
3704 {
3705 	return (1);
3706 }
3707 
3708 static uint32_t
3709 iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3710 {
3711 	int i, avail;
3712 	struct mbuf **mp;
3713 	iflib_txq_t txq;
3714 
3715 	txq = r->cookie;
3716 
3717 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3718 	CALLOUT_LOCK(txq);
3719 	callout_stop(&txq->ift_timer);
3720 	CALLOUT_UNLOCK(txq);
3721 
3722 	avail = IDXDIFF(pidx, cidx, r->size);
3723 	for (i = 0; i < avail; i++) {
3724 		mp = _ring_peek_one(r, cidx, i, avail - i);
3725 		if (__predict_false(*mp == (struct mbuf *)txq))
3726 			continue;
3727 		m_freem(*mp);
3728 		DBG_COUNTER_INC(tx_frees);
3729 	}
3730 	MPASS(ifmp_ring_is_stalled(r) == 0);
3731 	return (avail);
3732 }
3733 
3734 static void
3735 iflib_ifmp_purge(iflib_txq_t txq)
3736 {
3737 	struct ifmp_ring *r;
3738 
3739 	r = txq->ift_br;
3740 	r->drain = iflib_txq_drain_free;
3741 	r->can_drain = iflib_txq_drain_always;
3742 
3743 	ifmp_ring_check_drainage(r, r->size);
3744 
3745 	r->drain = iflib_txq_drain;
3746 	r->can_drain = iflib_txq_can_drain;
3747 }
3748 
3749 static void
3750 _task_fn_tx(void *context)
3751 {
3752 	iflib_txq_t txq = context;
3753 	if_ctx_t ctx = txq->ift_ctx;
3754 	if_t ifp = ctx->ifc_ifp;
3755 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
3756 
3757 #ifdef IFLIB_DIAGNOSTICS
3758 	txq->ift_cpu_exec_count[curcpu]++;
3759 #endif
3760 	if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
3761 		return;
3762 #ifdef DEV_NETMAP
3763 	if ((if_getcapenable(ifp) & IFCAP_NETMAP) &&
3764 	    netmap_tx_irq(ifp, txq->ift_id))
3765 		goto skip_ifmp;
3766 #endif
3767 #ifdef ALTQ
3768 	if (ALTQ_IS_ENABLED(&ifp->if_snd))
3769 		iflib_altq_if_start(ifp);
3770 #endif
3771 	if (txq->ift_db_pending)
3772 		ifmp_ring_enqueue(txq->ift_br, (void **)&txq, 1, TX_BATCH_SIZE, abdicate);
3773 	else if (!abdicate)
3774 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3775 	/*
3776 	 * When abdicating, we always need to check drainage, not just when we don't enqueue
3777 	 */
3778 	if (abdicate)
3779 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3780 #ifdef DEV_NETMAP
3781 skip_ifmp:
3782 #endif
3783 	if (ctx->ifc_flags & IFC_LEGACY)
3784 		IFDI_INTR_ENABLE(ctx);
3785 	else
3786 		IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3787 }
3788 
3789 static void
3790 _task_fn_rx(void *context)
3791 {
3792 	iflib_rxq_t rxq = context;
3793 	if_ctx_t ctx = rxq->ifr_ctx;
3794 	uint8_t more;
3795 	uint16_t budget;
3796 #ifdef DEV_NETMAP
3797 	u_int work = 0;
3798 	int nmirq;
3799 #endif
3800 
3801 #ifdef IFLIB_DIAGNOSTICS
3802 	rxq->ifr_cpu_exec_count[curcpu]++;
3803 #endif
3804 	DBG_COUNTER_INC(task_fn_rxs);
3805 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3806 		return;
3807 #ifdef DEV_NETMAP
3808 	nmirq = netmap_rx_irq(ctx->ifc_ifp, rxq->ifr_id, &work);
3809 	if (nmirq != NM_IRQ_PASS) {
3810 		more = (nmirq == NM_IRQ_RESCHED) ? IFLIB_RXEOF_MORE : 0;
3811 		goto skip_rxeof;
3812 	}
3813 #endif
3814 	budget = ctx->ifc_sysctl_rx_budget;
3815 	if (budget == 0)
3816 		budget = 16;	/* XXX */
3817 	more = iflib_rxeof(rxq, budget);
3818 #ifdef DEV_NETMAP
3819 skip_rxeof:
3820 #endif
3821 	if ((more & IFLIB_RXEOF_MORE) == 0) {
3822 		if (ctx->ifc_flags & IFC_LEGACY)
3823 			IFDI_INTR_ENABLE(ctx);
3824 		else
3825 			IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
3826 		DBG_COUNTER_INC(rx_intr_enables);
3827 	}
3828 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3829 		return;
3830 
3831 	if (more & IFLIB_RXEOF_MORE)
3832 		GROUPTASK_ENQUEUE(&rxq->ifr_task);
3833 	else if (more & IFLIB_RXEOF_EMPTY)
3834 		callout_reset_curcpu(&rxq->ifr_watchdog, 1, &_task_fn_rx_watchdog, rxq);
3835 }
3836 
3837 static void
3838 _task_fn_admin(void *context)
3839 {
3840 	if_ctx_t ctx = context;
3841 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
3842 	iflib_txq_t txq;
3843 	int i;
3844 	bool oactive, running, do_reset, do_watchdog, in_detach;
3845 	uint32_t reset_on = hz / 2;
3846 
3847 	STATE_LOCK(ctx);
3848 	running = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING);
3849 	oactive = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE);
3850 	do_reset = (ctx->ifc_flags & IFC_DO_RESET);
3851 	do_watchdog = (ctx->ifc_flags & IFC_DO_WATCHDOG);
3852 	in_detach = (ctx->ifc_flags & IFC_IN_DETACH);
3853 	ctx->ifc_flags &= ~(IFC_DO_RESET|IFC_DO_WATCHDOG);
3854 	STATE_UNLOCK(ctx);
3855 
3856 	if ((!running && !oactive) && !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
3857 		return;
3858 	if (in_detach)
3859 		return;
3860 
3861 	CTX_LOCK(ctx);
3862 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3863 		CALLOUT_LOCK(txq);
3864 		callout_stop(&txq->ift_timer);
3865 		CALLOUT_UNLOCK(txq);
3866 	}
3867 	if (do_watchdog) {
3868 		ctx->ifc_watchdog_events++;
3869 		IFDI_WATCHDOG_RESET(ctx);
3870 	}
3871 	IFDI_UPDATE_ADMIN_STATUS(ctx);
3872 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3873 #ifdef DEV_NETMAP
3874 		reset_on = hz / 2;
3875 		if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP)
3876 			iflib_netmap_timer_adjust(ctx, txq, &reset_on);
3877 #endif
3878 		callout_reset_on(&txq->ift_timer, reset_on, iflib_timer, txq, txq->ift_timer.c_cpu);
3879 	}
3880 	IFDI_LINK_INTR_ENABLE(ctx);
3881 	if (do_reset)
3882 		iflib_if_init_locked(ctx);
3883 	CTX_UNLOCK(ctx);
3884 
3885 	if (LINK_ACTIVE(ctx) == 0)
3886 		return;
3887 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3888 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
3889 }
3890 
3891 
3892 static void
3893 _task_fn_iov(void *context)
3894 {
3895 	if_ctx_t ctx = context;
3896 
3897 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING) &&
3898 	    !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
3899 		return;
3900 
3901 	CTX_LOCK(ctx);
3902 	IFDI_VFLR_HANDLE(ctx);
3903 	CTX_UNLOCK(ctx);
3904 }
3905 
3906 static int
3907 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)
3908 {
3909 	int err;
3910 	if_int_delay_info_t info;
3911 	if_ctx_t ctx;
3912 
3913 	info = (if_int_delay_info_t)arg1;
3914 	ctx = info->iidi_ctx;
3915 	info->iidi_req = req;
3916 	info->iidi_oidp = oidp;
3917 	CTX_LOCK(ctx);
3918 	err = IFDI_SYSCTL_INT_DELAY(ctx, info);
3919 	CTX_UNLOCK(ctx);
3920 	return (err);
3921 }
3922 
3923 /*********************************************************************
3924  *
3925  *  IFNET FUNCTIONS
3926  *
3927  **********************************************************************/
3928 
3929 static void
3930 iflib_if_init_locked(if_ctx_t ctx)
3931 {
3932 	iflib_stop(ctx);
3933 	iflib_init_locked(ctx);
3934 }
3935 
3936 
3937 static void
3938 iflib_if_init(void *arg)
3939 {
3940 	if_ctx_t ctx = arg;
3941 
3942 	CTX_LOCK(ctx);
3943 	iflib_if_init_locked(ctx);
3944 	CTX_UNLOCK(ctx);
3945 }
3946 
3947 static int
3948 iflib_if_transmit(if_t ifp, struct mbuf *m)
3949 {
3950 	if_ctx_t	ctx = if_getsoftc(ifp);
3951 
3952 	iflib_txq_t txq;
3953 	int err, qidx;
3954 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
3955 
3956 	if (__predict_false((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) {
3957 		DBG_COUNTER_INC(tx_frees);
3958 		m_freem(m);
3959 		return (ENETDOWN);
3960 	}
3961 
3962 	MPASS(m->m_nextpkt == NULL);
3963 	/* ALTQ-enabled interfaces always use queue 0. */
3964 	qidx = 0;
3965 	if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m) && !ALTQ_IS_ENABLED(&ifp->if_snd))
3966 		qidx = QIDX(ctx, m);
3967 	/*
3968 	 * XXX calculate buf_ring based on flowid (divvy up bits?)
3969 	 */
3970 	txq = &ctx->ifc_txqs[qidx];
3971 
3972 #ifdef DRIVER_BACKPRESSURE
3973 	if (txq->ift_closed) {
3974 		while (m != NULL) {
3975 			next = m->m_nextpkt;
3976 			m->m_nextpkt = NULL;
3977 			m_freem(m);
3978 			DBG_COUNTER_INC(tx_frees);
3979 			m = next;
3980 		}
3981 		return (ENOBUFS);
3982 	}
3983 #endif
3984 #ifdef notyet
3985 	qidx = count = 0;
3986 	mp = marr;
3987 	next = m;
3988 	do {
3989 		count++;
3990 		next = next->m_nextpkt;
3991 	} while (next != NULL);
3992 
3993 	if (count > nitems(marr))
3994 		if ((mp = malloc(count*sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) {
3995 			/* XXX check nextpkt */
3996 			m_freem(m);
3997 			/* XXX simplify for now */
3998 			DBG_COUNTER_INC(tx_frees);
3999 			return (ENOBUFS);
4000 		}
4001 	for (next = m, i = 0; next != NULL; i++) {
4002 		mp[i] = next;
4003 		next = next->m_nextpkt;
4004 		mp[i]->m_nextpkt = NULL;
4005 	}
4006 #endif
4007 	DBG_COUNTER_INC(tx_seen);
4008 	err = ifmp_ring_enqueue(txq->ift_br, (void **)&m, 1, TX_BATCH_SIZE, abdicate);
4009 
4010 	if (abdicate)
4011 		GROUPTASK_ENQUEUE(&txq->ift_task);
4012  	if (err) {
4013 		if (!abdicate)
4014 			GROUPTASK_ENQUEUE(&txq->ift_task);
4015 		/* support forthcoming later */
4016 #ifdef DRIVER_BACKPRESSURE
4017 		txq->ift_closed = TRUE;
4018 #endif
4019 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
4020 		m_freem(m);
4021 		DBG_COUNTER_INC(tx_frees);
4022 	}
4023 
4024 	return (err);
4025 }
4026 
4027 #ifdef ALTQ
4028 /*
4029  * The overall approach to integrating iflib with ALTQ is to continue to use
4030  * the iflib mp_ring machinery between the ALTQ queue(s) and the hardware
4031  * ring.  Technically, when using ALTQ, queueing to an intermediate mp_ring
4032  * is redundant/unnecessary, but doing so minimizes the amount of
4033  * ALTQ-specific code required in iflib.  It is assumed that the overhead of
4034  * redundantly queueing to an intermediate mp_ring is swamped by the
4035  * performance limitations inherent in using ALTQ.
4036  *
4037  * When ALTQ support is compiled in, all iflib drivers will use a transmit
4038  * routine, iflib_altq_if_transmit(), that checks if ALTQ is enabled for the
4039  * given interface.  If ALTQ is enabled for an interface, then all
4040  * transmitted packets for that interface will be submitted to the ALTQ
4041  * subsystem via IFQ_ENQUEUE().  We don't use the legacy if_transmit()
4042  * implementation because it uses IFQ_HANDOFF(), which will duplicatively
4043  * update stats that the iflib machinery handles, and which is sensitve to
4044  * the disused IFF_DRV_OACTIVE flag.  Additionally, iflib_altq_if_start()
4045  * will be installed as the start routine for use by ALTQ facilities that
4046  * need to trigger queue drains on a scheduled basis.
4047  *
4048  */
4049 static void
4050 iflib_altq_if_start(if_t ifp)
4051 {
4052 	struct ifaltq *ifq = &ifp->if_snd;
4053 	struct mbuf *m;
4054 
4055 	IFQ_LOCK(ifq);
4056 	IFQ_DEQUEUE_NOLOCK(ifq, m);
4057 	while (m != NULL) {
4058 		iflib_if_transmit(ifp, m);
4059 		IFQ_DEQUEUE_NOLOCK(ifq, m);
4060 	}
4061 	IFQ_UNLOCK(ifq);
4062 }
4063 
4064 static int
4065 iflib_altq_if_transmit(if_t ifp, struct mbuf *m)
4066 {
4067 	int err;
4068 
4069 	if (ALTQ_IS_ENABLED(&ifp->if_snd)) {
4070 		IFQ_ENQUEUE(&ifp->if_snd, m, err);
4071 		if (err == 0)
4072 			iflib_altq_if_start(ifp);
4073 	} else
4074 		err = iflib_if_transmit(ifp, m);
4075 
4076 	return (err);
4077 }
4078 #endif /* ALTQ */
4079 
4080 static void
4081 iflib_if_qflush(if_t ifp)
4082 {
4083 	if_ctx_t ctx = if_getsoftc(ifp);
4084 	iflib_txq_t txq = ctx->ifc_txqs;
4085 	int i;
4086 
4087 	STATE_LOCK(ctx);
4088 	ctx->ifc_flags |= IFC_QFLUSH;
4089 	STATE_UNLOCK(ctx);
4090 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
4091 		while (!(ifmp_ring_is_idle(txq->ift_br) || ifmp_ring_is_stalled(txq->ift_br)))
4092 			iflib_txq_check_drain(txq, 0);
4093 	STATE_LOCK(ctx);
4094 	ctx->ifc_flags &= ~IFC_QFLUSH;
4095 	STATE_UNLOCK(ctx);
4096 
4097 	/*
4098 	 * When ALTQ is enabled, this will also take care of purging the
4099 	 * ALTQ queue(s).
4100 	 */
4101 	if_qflush(ifp);
4102 }
4103 
4104 
4105 #define IFCAP_FLAGS (IFCAP_HWCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \
4106 		     IFCAP_TSO | IFCAP_VLAN_HWTAGGING | IFCAP_HWSTATS | \
4107 		     IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | \
4108 		     IFCAP_VLAN_HWTSO | IFCAP_VLAN_HWCSUM | IFCAP_NOMAP)
4109 
4110 static int
4111 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data)
4112 {
4113 	if_ctx_t ctx = if_getsoftc(ifp);
4114 	struct ifreq	*ifr = (struct ifreq *)data;
4115 #if defined(INET) || defined(INET6)
4116 	struct ifaddr	*ifa = (struct ifaddr *)data;
4117 #endif
4118 	bool		avoid_reset = false;
4119 	int		err = 0, reinit = 0, bits;
4120 
4121 	switch (command) {
4122 	case SIOCSIFADDR:
4123 #ifdef INET
4124 		if (ifa->ifa_addr->sa_family == AF_INET)
4125 			avoid_reset = true;
4126 #endif
4127 #ifdef INET6
4128 		if (ifa->ifa_addr->sa_family == AF_INET6)
4129 			avoid_reset = true;
4130 #endif
4131 		/*
4132 		** Calling init results in link renegotiation,
4133 		** so we avoid doing it when possible.
4134 		*/
4135 		if (avoid_reset) {
4136 			if_setflagbits(ifp, IFF_UP,0);
4137 			if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
4138 				reinit = 1;
4139 #ifdef INET
4140 			if (!(if_getflags(ifp) & IFF_NOARP))
4141 				arp_ifinit(ifp, ifa);
4142 #endif
4143 		} else
4144 			err = ether_ioctl(ifp, command, data);
4145 		break;
4146 	case SIOCSIFMTU:
4147 		CTX_LOCK(ctx);
4148 		if (ifr->ifr_mtu == if_getmtu(ifp)) {
4149 			CTX_UNLOCK(ctx);
4150 			break;
4151 		}
4152 		bits = if_getdrvflags(ifp);
4153 		/* stop the driver and free any clusters before proceeding */
4154 		iflib_stop(ctx);
4155 
4156 		if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) {
4157 			STATE_LOCK(ctx);
4158 			if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size)
4159 				ctx->ifc_flags |= IFC_MULTISEG;
4160 			else
4161 				ctx->ifc_flags &= ~IFC_MULTISEG;
4162 			STATE_UNLOCK(ctx);
4163 			err = if_setmtu(ifp, ifr->ifr_mtu);
4164 		}
4165 		iflib_init_locked(ctx);
4166 		STATE_LOCK(ctx);
4167 		if_setdrvflags(ifp, bits);
4168 		STATE_UNLOCK(ctx);
4169 		CTX_UNLOCK(ctx);
4170 		break;
4171 	case SIOCSIFFLAGS:
4172 		CTX_LOCK(ctx);
4173 		if (if_getflags(ifp) & IFF_UP) {
4174 			if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4175 				if ((if_getflags(ifp) ^ ctx->ifc_if_flags) &
4176 				    (IFF_PROMISC | IFF_ALLMULTI)) {
4177 					err = IFDI_PROMISC_SET(ctx, if_getflags(ifp));
4178 				}
4179 			} else
4180 				reinit = 1;
4181 		} else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4182 			iflib_stop(ctx);
4183 		}
4184 		ctx->ifc_if_flags = if_getflags(ifp);
4185 		CTX_UNLOCK(ctx);
4186 		break;
4187 	case SIOCADDMULTI:
4188 	case SIOCDELMULTI:
4189 		if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4190 			CTX_LOCK(ctx);
4191 			IFDI_INTR_DISABLE(ctx);
4192 			IFDI_MULTI_SET(ctx);
4193 			IFDI_INTR_ENABLE(ctx);
4194 			CTX_UNLOCK(ctx);
4195 		}
4196 		break;
4197 	case SIOCSIFMEDIA:
4198 		CTX_LOCK(ctx);
4199 		IFDI_MEDIA_SET(ctx);
4200 		CTX_UNLOCK(ctx);
4201 		/* FALLTHROUGH */
4202 	case SIOCGIFMEDIA:
4203 	case SIOCGIFXMEDIA:
4204 		err = ifmedia_ioctl(ifp, ifr, ctx->ifc_mediap, command);
4205 		break;
4206 	case SIOCGI2C:
4207 	{
4208 		struct ifi2creq i2c;
4209 
4210 		err = copyin(ifr_data_get_ptr(ifr), &i2c, sizeof(i2c));
4211 		if (err != 0)
4212 			break;
4213 		if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) {
4214 			err = EINVAL;
4215 			break;
4216 		}
4217 		if (i2c.len > sizeof(i2c.data)) {
4218 			err = EINVAL;
4219 			break;
4220 		}
4221 
4222 		if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0)
4223 			err = copyout(&i2c, ifr_data_get_ptr(ifr),
4224 			    sizeof(i2c));
4225 		break;
4226 	}
4227 	case SIOCSIFCAP:
4228 	{
4229 		int mask, setmask, oldmask;
4230 
4231 		oldmask = if_getcapenable(ifp);
4232 		mask = ifr->ifr_reqcap ^ oldmask;
4233 		mask &= ctx->ifc_softc_ctx.isc_capabilities | IFCAP_NOMAP;
4234 		setmask = 0;
4235 #ifdef TCP_OFFLOAD
4236 		setmask |= mask & (IFCAP_TOE4|IFCAP_TOE6);
4237 #endif
4238 		setmask |= (mask & IFCAP_FLAGS);
4239 		setmask |= (mask & IFCAP_WOL);
4240 
4241 		/*
4242 		 * If any RX csum has changed, change all the ones that
4243 		 * are supported by the driver.
4244 		 */
4245 		if (setmask & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6)) {
4246 			setmask |= ctx->ifc_softc_ctx.isc_capabilities &
4247 			    (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6);
4248 		}
4249 
4250 		/*
4251 		 * want to ensure that traffic has stopped before we change any of the flags
4252 		 */
4253 		if (setmask) {
4254 			CTX_LOCK(ctx);
4255 			bits = if_getdrvflags(ifp);
4256 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4257 				iflib_stop(ctx);
4258 			STATE_LOCK(ctx);
4259 			if_togglecapenable(ifp, setmask);
4260 			STATE_UNLOCK(ctx);
4261 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4262 				iflib_init_locked(ctx);
4263 			STATE_LOCK(ctx);
4264 			if_setdrvflags(ifp, bits);
4265 			STATE_UNLOCK(ctx);
4266 			CTX_UNLOCK(ctx);
4267 		}
4268 		if_vlancap(ifp);
4269 		break;
4270 	}
4271 	case SIOCGPRIVATE_0:
4272 	case SIOCSDRVSPEC:
4273 	case SIOCGDRVSPEC:
4274 		CTX_LOCK(ctx);
4275 		err = IFDI_PRIV_IOCTL(ctx, command, data);
4276 		CTX_UNLOCK(ctx);
4277 		break;
4278 	default:
4279 		err = ether_ioctl(ifp, command, data);
4280 		break;
4281 	}
4282 	if (reinit)
4283 		iflib_if_init(ctx);
4284 	return (err);
4285 }
4286 
4287 static uint64_t
4288 iflib_if_get_counter(if_t ifp, ift_counter cnt)
4289 {
4290 	if_ctx_t ctx = if_getsoftc(ifp);
4291 
4292 	return (IFDI_GET_COUNTER(ctx, cnt));
4293 }
4294 
4295 /*********************************************************************
4296  *
4297  *  OTHER FUNCTIONS EXPORTED TO THE STACK
4298  *
4299  **********************************************************************/
4300 
4301 static void
4302 iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag)
4303 {
4304 	if_ctx_t ctx = if_getsoftc(ifp);
4305 
4306 	if ((void *)ctx != arg)
4307 		return;
4308 
4309 	if ((vtag == 0) || (vtag > 4095))
4310 		return;
4311 
4312 	if (iflib_in_detach(ctx))
4313 		return;
4314 
4315 	CTX_LOCK(ctx);
4316 	/* Driver may need all untagged packets to be flushed */
4317 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4318 		iflib_stop(ctx);
4319 	IFDI_VLAN_REGISTER(ctx, vtag);
4320 	/* Re-init to load the changes, if required */
4321 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4322 		iflib_init_locked(ctx);
4323 	CTX_UNLOCK(ctx);
4324 }
4325 
4326 static void
4327 iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag)
4328 {
4329 	if_ctx_t ctx = if_getsoftc(ifp);
4330 
4331 	if ((void *)ctx != arg)
4332 		return;
4333 
4334 	if ((vtag == 0) || (vtag > 4095))
4335 		return;
4336 
4337 	CTX_LOCK(ctx);
4338 	/* Driver may need all tagged packets to be flushed */
4339 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4340 		iflib_stop(ctx);
4341 	IFDI_VLAN_UNREGISTER(ctx, vtag);
4342 	/* Re-init to load the changes, if required */
4343 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4344 		iflib_init_locked(ctx);
4345 	CTX_UNLOCK(ctx);
4346 }
4347 
4348 static void
4349 iflib_led_func(void *arg, int onoff)
4350 {
4351 	if_ctx_t ctx = arg;
4352 
4353 	CTX_LOCK(ctx);
4354 	IFDI_LED_FUNC(ctx, onoff);
4355 	CTX_UNLOCK(ctx);
4356 }
4357 
4358 /*********************************************************************
4359  *
4360  *  BUS FUNCTION DEFINITIONS
4361  *
4362  **********************************************************************/
4363 
4364 int
4365 iflib_device_probe(device_t dev)
4366 {
4367 	const pci_vendor_info_t *ent;
4368 	if_shared_ctx_t sctx;
4369 	uint16_t pci_device_id, pci_rev_id, pci_subdevice_id, pci_subvendor_id;
4370 	uint16_t pci_vendor_id;
4371 
4372 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4373 		return (ENOTSUP);
4374 
4375 	pci_vendor_id = pci_get_vendor(dev);
4376 	pci_device_id = pci_get_device(dev);
4377 	pci_subvendor_id = pci_get_subvendor(dev);
4378 	pci_subdevice_id = pci_get_subdevice(dev);
4379 	pci_rev_id = pci_get_revid(dev);
4380 	if (sctx->isc_parse_devinfo != NULL)
4381 		sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id);
4382 
4383 	ent = sctx->isc_vendor_info;
4384 	while (ent->pvi_vendor_id != 0) {
4385 		if (pci_vendor_id != ent->pvi_vendor_id) {
4386 			ent++;
4387 			continue;
4388 		}
4389 		if ((pci_device_id == ent->pvi_device_id) &&
4390 		    ((pci_subvendor_id == ent->pvi_subvendor_id) ||
4391 		     (ent->pvi_subvendor_id == 0)) &&
4392 		    ((pci_subdevice_id == ent->pvi_subdevice_id) ||
4393 		     (ent->pvi_subdevice_id == 0)) &&
4394 		    ((pci_rev_id == ent->pvi_rev_id) ||
4395 		     (ent->pvi_rev_id == 0))) {
4396 
4397 			device_set_desc_copy(dev, ent->pvi_name);
4398 			/* this needs to be changed to zero if the bus probing code
4399 			 * ever stops re-probing on best match because the sctx
4400 			 * may have its values over written by register calls
4401 			 * in subsequent probes
4402 			 */
4403 			return (BUS_PROBE_DEFAULT);
4404 		}
4405 		ent++;
4406 	}
4407 	return (ENXIO);
4408 }
4409 
4410 int
4411 iflib_device_probe_vendor(device_t dev)
4412 {
4413 	int probe;
4414 
4415 	probe = iflib_device_probe(dev);
4416 	if (probe == BUS_PROBE_DEFAULT)
4417 		return (BUS_PROBE_VENDOR);
4418 	else
4419 		return (probe);
4420 }
4421 
4422 static void
4423 iflib_reset_qvalues(if_ctx_t ctx)
4424 {
4425 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4426 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4427 	device_t dev = ctx->ifc_dev;
4428 	int i;
4429 
4430 	if (ctx->ifc_sysctl_ntxqs != 0)
4431 		scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs;
4432 	if (ctx->ifc_sysctl_nrxqs != 0)
4433 		scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs;
4434 
4435 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4436 		if (ctx->ifc_sysctl_ntxds[i] != 0)
4437 			scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i];
4438 		else
4439 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4440 	}
4441 
4442 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4443 		if (ctx->ifc_sysctl_nrxds[i] != 0)
4444 			scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i];
4445 		else
4446 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4447 	}
4448 
4449 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4450 		if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) {
4451 			device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n",
4452 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]);
4453 			scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i];
4454 		}
4455 		if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) {
4456 			device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n",
4457 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]);
4458 			scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i];
4459 		}
4460 		if (!powerof2(scctx->isc_nrxd[i])) {
4461 			device_printf(dev, "nrxd%d: %d is not a power of 2 - using default value of %d\n",
4462 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_default[i]);
4463 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4464 		}
4465 	}
4466 
4467 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4468 		if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) {
4469 			device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n",
4470 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]);
4471 			scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i];
4472 		}
4473 		if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) {
4474 			device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n",
4475 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]);
4476 			scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i];
4477 		}
4478 		if (!powerof2(scctx->isc_ntxd[i])) {
4479 			device_printf(dev, "ntxd%d: %d is not a power of 2 - using default value of %d\n",
4480 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_default[i]);
4481 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4482 		}
4483 	}
4484 }
4485 
4486 static void
4487 iflib_add_pfil(if_ctx_t ctx)
4488 {
4489 	struct pfil_head *pfil;
4490 	struct pfil_head_args pa;
4491 	iflib_rxq_t rxq;
4492 	int i;
4493 
4494 	pa.pa_version = PFIL_VERSION;
4495 	pa.pa_flags = PFIL_IN;
4496 	pa.pa_type = PFIL_TYPE_ETHERNET;
4497 	pa.pa_headname = ctx->ifc_ifp->if_xname;
4498 	pfil = pfil_head_register(&pa);
4499 
4500 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4501 		rxq->pfil = pfil;
4502 	}
4503 }
4504 
4505 static void
4506 iflib_rem_pfil(if_ctx_t ctx)
4507 {
4508 	struct pfil_head *pfil;
4509 	iflib_rxq_t rxq;
4510 	int i;
4511 
4512 	rxq = ctx->ifc_rxqs;
4513 	pfil = rxq->pfil;
4514 	for (i = 0; i < NRXQSETS(ctx); i++, rxq++) {
4515 		rxq->pfil = NULL;
4516 	}
4517 	pfil_head_unregister(pfil);
4518 }
4519 
4520 static uint16_t
4521 get_ctx_core_offset(if_ctx_t ctx)
4522 {
4523 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4524 	struct cpu_offset *op;
4525 	uint16_t qc;
4526 	uint16_t ret = ctx->ifc_sysctl_core_offset;
4527 
4528 	if (ret != CORE_OFFSET_UNSPECIFIED)
4529 		return (ret);
4530 
4531 	if (ctx->ifc_sysctl_separate_txrx)
4532 		qc = scctx->isc_ntxqsets + scctx->isc_nrxqsets;
4533 	else
4534 		qc = max(scctx->isc_ntxqsets, scctx->isc_nrxqsets);
4535 
4536 	mtx_lock(&cpu_offset_mtx);
4537 	SLIST_FOREACH(op, &cpu_offsets, entries) {
4538 		if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) {
4539 			ret = op->offset;
4540 			op->offset += qc;
4541 			MPASS(op->refcount < UINT_MAX);
4542 			op->refcount++;
4543 			break;
4544 		}
4545 	}
4546 	if (ret == CORE_OFFSET_UNSPECIFIED) {
4547 		ret = 0;
4548 		op = malloc(sizeof(struct cpu_offset), M_IFLIB,
4549 		    M_NOWAIT | M_ZERO);
4550 		if (op == NULL) {
4551 			device_printf(ctx->ifc_dev,
4552 			    "allocation for cpu offset failed.\n");
4553 		} else {
4554 			op->offset = qc;
4555 			op->refcount = 1;
4556 			CPU_COPY(&ctx->ifc_cpus, &op->set);
4557 			SLIST_INSERT_HEAD(&cpu_offsets, op, entries);
4558 		}
4559 	}
4560 	mtx_unlock(&cpu_offset_mtx);
4561 
4562 	return (ret);
4563 }
4564 
4565 static void
4566 unref_ctx_core_offset(if_ctx_t ctx)
4567 {
4568 	struct cpu_offset *op, *top;
4569 
4570 	mtx_lock(&cpu_offset_mtx);
4571 	SLIST_FOREACH_SAFE(op, &cpu_offsets, entries, top) {
4572 		if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) {
4573 			MPASS(op->refcount > 0);
4574 			op->refcount--;
4575 			if (op->refcount == 0) {
4576 				SLIST_REMOVE(&cpu_offsets, op, cpu_offset, entries);
4577 				free(op, M_IFLIB);
4578 			}
4579 			break;
4580 		}
4581 	}
4582 	mtx_unlock(&cpu_offset_mtx);
4583 }
4584 
4585 int
4586 iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp)
4587 {
4588 	if_ctx_t ctx;
4589 	if_t ifp;
4590 	if_softc_ctx_t scctx;
4591 	kobjop_desc_t kobj_desc;
4592 	kobj_method_t *kobj_method;
4593 	int err, msix, rid;
4594 	uint16_t main_rxq, main_txq;
4595 
4596 	ctx = malloc(sizeof(* ctx), M_IFLIB, M_WAITOK|M_ZERO);
4597 
4598 	if (sc == NULL) {
4599 		sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4600 		device_set_softc(dev, ctx);
4601 		ctx->ifc_flags |= IFC_SC_ALLOCATED;
4602 	}
4603 
4604 	ctx->ifc_sctx = sctx;
4605 	ctx->ifc_dev = dev;
4606 	ctx->ifc_softc = sc;
4607 
4608 	if ((err = iflib_register(ctx)) != 0) {
4609 		device_printf(dev, "iflib_register failed %d\n", err);
4610 		goto fail_ctx_free;
4611 	}
4612 	iflib_add_device_sysctl_pre(ctx);
4613 
4614 	scctx = &ctx->ifc_softc_ctx;
4615 	ifp = ctx->ifc_ifp;
4616 
4617 	iflib_reset_qvalues(ctx);
4618 	CTX_LOCK(ctx);
4619 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4620 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4621 		goto fail_unlock;
4622 	}
4623 	_iflib_pre_assert(scctx);
4624 	ctx->ifc_txrx = *scctx->isc_txrx;
4625 
4626 	if (sctx->isc_flags & IFLIB_DRIVER_MEDIA)
4627 		ctx->ifc_mediap = scctx->isc_media;
4628 
4629 #ifdef INVARIANTS
4630 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
4631 		MPASS(scctx->isc_tx_csum_flags);
4632 #endif
4633 
4634 	if_setcapabilities(ifp,
4635 	    scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_NOMAP);
4636 	if_setcapenable(ifp,
4637 	    scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_NOMAP);
4638 
4639 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4640 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4641 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4642 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4643 
4644 	main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4645 	main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4646 
4647 	/* XXX change for per-queue sizes */
4648 	device_printf(dev, "Using %d TX descriptors and %d RX descriptors\n",
4649 	    scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4650 
4651 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4652 	    MAX_SINGLE_PACKET_FRACTION)
4653 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4654 		    MAX_SINGLE_PACKET_FRACTION);
4655 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4656 	    MAX_SINGLE_PACKET_FRACTION)
4657 		scctx->isc_tx_tso_segments_max = max(1,
4658 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4659 
4660 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4661 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
4662 		/*
4663 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
4664 		 * but some MACs do.
4665 		 */
4666 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
4667 		    IP_MAXPACKET));
4668 		/*
4669 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
4670 		 * into account.  In the worst case, each of these calls will
4671 		 * add another mbuf and, thus, the requirement for another DMA
4672 		 * segment.  So for best performance, it doesn't make sense to
4673 		 * advertize a maximum of TSO segments that typically will
4674 		 * require defragmentation in iflib_encap().
4675 		 */
4676 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
4677 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
4678 	}
4679 	if (scctx->isc_rss_table_size == 0)
4680 		scctx->isc_rss_table_size = 64;
4681 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4682 
4683 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4684 	/* XXX format name */
4685 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx,
4686 	    NULL, NULL, "admin");
4687 
4688 	/* Set up cpu set.  If it fails, use the set of all CPUs. */
4689 	if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) != 0) {
4690 		device_printf(dev, "Unable to fetch CPU list\n");
4691 		CPU_COPY(&all_cpus, &ctx->ifc_cpus);
4692 	}
4693 	MPASS(CPU_COUNT(&ctx->ifc_cpus) > 0);
4694 
4695 	/*
4696 	** Now set up MSI or MSI-X, should return us the number of supported
4697 	** vectors (will be 1 for a legacy interrupt and MSI).
4698 	*/
4699 	if (sctx->isc_flags & IFLIB_SKIP_MSIX) {
4700 		msix = scctx->isc_vectors;
4701 	} else if (scctx->isc_msix_bar != 0)
4702 	       /*
4703 		* The simple fact that isc_msix_bar is not 0 does not mean we
4704 		* we have a good value there that is known to work.
4705 		*/
4706 		msix = iflib_msix_init(ctx);
4707 	else {
4708 		scctx->isc_vectors = 1;
4709 		scctx->isc_ntxqsets = 1;
4710 		scctx->isc_nrxqsets = 1;
4711 		scctx->isc_intr = IFLIB_INTR_LEGACY;
4712 		msix = 0;
4713 	}
4714 	/* Get memory for the station queues */
4715 	if ((err = iflib_queues_alloc(ctx))) {
4716 		device_printf(dev, "Unable to allocate queue memory\n");
4717 		goto fail_intr_free;
4718 	}
4719 
4720 	if ((err = iflib_qset_structures_setup(ctx)))
4721 		goto fail_queues;
4722 
4723 	/*
4724 	 * Now that we know how many queues there are, get the core offset.
4725 	 */
4726 	ctx->ifc_sysctl_core_offset = get_ctx_core_offset(ctx);
4727 
4728 	if (msix > 1) {
4729 		/*
4730 		 * When using MSI-X, ensure that ifdi_{r,t}x_queue_intr_enable
4731 		 * aren't the default NULL implementation.
4732 		 */
4733 		kobj_desc = &ifdi_rx_queue_intr_enable_desc;
4734 		kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL,
4735 		    kobj_desc);
4736 		if (kobj_method == &kobj_desc->deflt) {
4737 			device_printf(dev,
4738 			    "MSI-X requires ifdi_rx_queue_intr_enable method");
4739 			err = EOPNOTSUPP;
4740 			goto fail_queues;
4741 		}
4742 		kobj_desc = &ifdi_tx_queue_intr_enable_desc;
4743 		kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL,
4744 		    kobj_desc);
4745 		if (kobj_method == &kobj_desc->deflt) {
4746 			device_printf(dev,
4747 			    "MSI-X requires ifdi_tx_queue_intr_enable method");
4748 			err = EOPNOTSUPP;
4749 			goto fail_queues;
4750 		}
4751 
4752 		/*
4753 		 * Assign the MSI-X vectors.
4754 		 * Note that the default NULL ifdi_msix_intr_assign method will
4755 		 * fail here, too.
4756 		 */
4757 		err = IFDI_MSIX_INTR_ASSIGN(ctx, msix);
4758 		if (err != 0) {
4759 			device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n",
4760 			    err);
4761 			goto fail_queues;
4762 		}
4763 	} else if (scctx->isc_intr != IFLIB_INTR_MSIX) {
4764 		rid = 0;
4765 		if (scctx->isc_intr == IFLIB_INTR_MSI) {
4766 			MPASS(msix == 1);
4767 			rid = 1;
4768 		}
4769 		if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) {
4770 			device_printf(dev, "iflib_legacy_setup failed %d\n", err);
4771 			goto fail_queues;
4772 		}
4773 	} else {
4774 		device_printf(dev,
4775 		    "Cannot use iflib with only 1 MSI-X interrupt!\n");
4776 		err = ENODEV;
4777 		goto fail_intr_free;
4778 	}
4779 
4780 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
4781 
4782 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4783 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4784 		goto fail_detach;
4785 	}
4786 
4787 	/*
4788 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4789 	 * This must appear after the call to ether_ifattach() because
4790 	 * ether_ifattach() sets if_hdrlen to the default value.
4791 	 */
4792 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4793 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
4794 
4795 	if ((err = iflib_netmap_attach(ctx))) {
4796 		device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err);
4797 		goto fail_detach;
4798 	}
4799 	*ctxp = ctx;
4800 
4801 	DEBUGNET_SET(ctx->ifc_ifp, iflib);
4802 
4803 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4804 	iflib_add_device_sysctl_post(ctx);
4805 	iflib_add_pfil(ctx);
4806 	ctx->ifc_flags |= IFC_INIT_DONE;
4807 	CTX_UNLOCK(ctx);
4808 
4809 	return (0);
4810 
4811 fail_detach:
4812 	ether_ifdetach(ctx->ifc_ifp);
4813 fail_intr_free:
4814 	iflib_free_intr_mem(ctx);
4815 fail_queues:
4816 	iflib_tx_structures_free(ctx);
4817 	iflib_rx_structures_free(ctx);
4818 	taskqgroup_detach(qgroup_if_config_tqg, &ctx->ifc_admin_task);
4819 	IFDI_DETACH(ctx);
4820 fail_unlock:
4821 	CTX_UNLOCK(ctx);
4822 	iflib_deregister(ctx);
4823 fail_ctx_free:
4824 	device_set_softc(ctx->ifc_dev, NULL);
4825         if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4826                 free(ctx->ifc_softc, M_IFLIB);
4827         free(ctx, M_IFLIB);
4828 	return (err);
4829 }
4830 
4831 int
4832 iflib_pseudo_register(device_t dev, if_shared_ctx_t sctx, if_ctx_t *ctxp,
4833 					  struct iflib_cloneattach_ctx *clctx)
4834 {
4835 	int err;
4836 	if_ctx_t ctx;
4837 	if_t ifp;
4838 	if_softc_ctx_t scctx;
4839 	int i;
4840 	void *sc;
4841 	uint16_t main_txq;
4842 	uint16_t main_rxq;
4843 
4844 	ctx = malloc(sizeof(*ctx), M_IFLIB, M_WAITOK|M_ZERO);
4845 	sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4846 	ctx->ifc_flags |= IFC_SC_ALLOCATED;
4847 	if (sctx->isc_flags & (IFLIB_PSEUDO|IFLIB_VIRTUAL))
4848 		ctx->ifc_flags |= IFC_PSEUDO;
4849 
4850 	ctx->ifc_sctx = sctx;
4851 	ctx->ifc_softc = sc;
4852 	ctx->ifc_dev = dev;
4853 
4854 	if ((err = iflib_register(ctx)) != 0) {
4855 		device_printf(dev, "%s: iflib_register failed %d\n", __func__, err);
4856 		goto fail_ctx_free;
4857 	}
4858 	iflib_add_device_sysctl_pre(ctx);
4859 
4860 	scctx = &ctx->ifc_softc_ctx;
4861 	ifp = ctx->ifc_ifp;
4862 
4863 	iflib_reset_qvalues(ctx);
4864 	CTX_LOCK(ctx);
4865 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4866 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4867 		goto fail_unlock;
4868 	}
4869 	if (sctx->isc_flags & IFLIB_GEN_MAC)
4870 		ether_gen_addr(ifp, &ctx->ifc_mac);
4871 	if ((err = IFDI_CLONEATTACH(ctx, clctx->cc_ifc, clctx->cc_name,
4872 								clctx->cc_params)) != 0) {
4873 		device_printf(dev, "IFDI_CLONEATTACH failed %d\n", err);
4874 		goto fail_unlock;
4875 	}
4876 #ifdef INVARIANTS
4877 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
4878 		MPASS(scctx->isc_tx_csum_flags);
4879 #endif
4880 
4881 	if_setcapabilities(ifp, scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_LINKSTATE);
4882 	if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_LINKSTATE);
4883 
4884 	ifp->if_flags |= IFF_NOGROUP;
4885 	if (sctx->isc_flags & IFLIB_PSEUDO) {
4886 		ifmedia_add(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO, 0, NULL);
4887 		ifmedia_set(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO);
4888 		if (sctx->isc_flags & IFLIB_PSEUDO_ETHER) {
4889 			ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
4890 		} else {
4891 			if_attach(ctx->ifc_ifp);
4892 			bpfattach(ctx->ifc_ifp, DLT_NULL, sizeof(u_int32_t));
4893 		}
4894 
4895 		if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4896 			device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4897 			goto fail_detach;
4898 		}
4899 		*ctxp = ctx;
4900 
4901 		/*
4902 		 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4903 		 * This must appear after the call to ether_ifattach() because
4904 		 * ether_ifattach() sets if_hdrlen to the default value.
4905 		 */
4906 		if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4907 			if_setifheaderlen(ifp,
4908 			    sizeof(struct ether_vlan_header));
4909 
4910 		if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4911 		iflib_add_device_sysctl_post(ctx);
4912 		ctx->ifc_flags |= IFC_INIT_DONE;
4913 		CTX_UNLOCK(ctx);
4914 		return (0);
4915 	}
4916 	ifmedia_add(ctx->ifc_mediap, IFM_ETHER | IFM_1000_T | IFM_FDX, 0, NULL);
4917 	ifmedia_add(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO, 0, NULL);
4918 	ifmedia_set(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO);
4919 
4920 	_iflib_pre_assert(scctx);
4921 	ctx->ifc_txrx = *scctx->isc_txrx;
4922 
4923 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4924 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4925 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4926 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4927 
4928 	main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4929 	main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4930 
4931 	/* XXX change for per-queue sizes */
4932 	device_printf(dev, "Using %d TX descriptors and %d RX descriptors\n",
4933 	    scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4934 
4935 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4936 	    MAX_SINGLE_PACKET_FRACTION)
4937 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4938 		    MAX_SINGLE_PACKET_FRACTION);
4939 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4940 	    MAX_SINGLE_PACKET_FRACTION)
4941 		scctx->isc_tx_tso_segments_max = max(1,
4942 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4943 
4944 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4945 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
4946 		/*
4947 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
4948 		 * but some MACs do.
4949 		 */
4950 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
4951 		    IP_MAXPACKET));
4952 		/*
4953 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
4954 		 * into account.  In the worst case, each of these calls will
4955 		 * add another mbuf and, thus, the requirement for another DMA
4956 		 * segment.  So for best performance, it doesn't make sense to
4957 		 * advertize a maximum of TSO segments that typically will
4958 		 * require defragmentation in iflib_encap().
4959 		 */
4960 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
4961 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
4962 	}
4963 	if (scctx->isc_rss_table_size == 0)
4964 		scctx->isc_rss_table_size = 64;
4965 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4966 
4967 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4968 	/* XXX format name */
4969 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx,
4970 	    NULL, NULL, "admin");
4971 
4972 	/* XXX --- can support > 1 -- but keep it simple for now */
4973 	scctx->isc_intr = IFLIB_INTR_LEGACY;
4974 
4975 	/* Get memory for the station queues */
4976 	if ((err = iflib_queues_alloc(ctx))) {
4977 		device_printf(dev, "Unable to allocate queue memory\n");
4978 		goto fail_iflib_detach;
4979 	}
4980 
4981 	if ((err = iflib_qset_structures_setup(ctx))) {
4982 		device_printf(dev, "qset structure setup failed %d\n", err);
4983 		goto fail_queues;
4984 	}
4985 
4986 	/*
4987 	 * XXX What if anything do we want to do about interrupts?
4988 	 */
4989 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
4990 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4991 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4992 		goto fail_detach;
4993 	}
4994 
4995 	/*
4996 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4997 	 * This must appear after the call to ether_ifattach() because
4998 	 * ether_ifattach() sets if_hdrlen to the default value.
4999 	 */
5000 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
5001 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
5002 
5003 	/* XXX handle more than one queue */
5004 	for (i = 0; i < scctx->isc_nrxqsets; i++)
5005 		IFDI_RX_CLSET(ctx, 0, i, ctx->ifc_rxqs[i].ifr_fl[0].ifl_sds.ifsd_cl);
5006 
5007 	*ctxp = ctx;
5008 
5009 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
5010 	iflib_add_device_sysctl_post(ctx);
5011 	ctx->ifc_flags |= IFC_INIT_DONE;
5012 	CTX_UNLOCK(ctx);
5013 
5014 	return (0);
5015 fail_detach:
5016 	ether_ifdetach(ctx->ifc_ifp);
5017 fail_queues:
5018 	iflib_tx_structures_free(ctx);
5019 	iflib_rx_structures_free(ctx);
5020 fail_iflib_detach:
5021 	IFDI_DETACH(ctx);
5022 fail_unlock:
5023 	CTX_UNLOCK(ctx);
5024 	iflib_deregister(ctx);
5025 fail_ctx_free:
5026 	free(ctx->ifc_softc, M_IFLIB);
5027 	free(ctx, M_IFLIB);
5028 	return (err);
5029 }
5030 
5031 int
5032 iflib_pseudo_deregister(if_ctx_t ctx)
5033 {
5034 	if_t ifp = ctx->ifc_ifp;
5035 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5036 	iflib_txq_t txq;
5037 	iflib_rxq_t rxq;
5038 	int i, j;
5039 	struct taskqgroup *tqg;
5040 	iflib_fl_t fl;
5041 
5042 	/* Unregister VLAN event handlers early */
5043 	iflib_unregister_vlan_handlers(ctx);
5044 
5045 	if ((sctx->isc_flags & IFLIB_PSEUDO)  &&
5046 		(sctx->isc_flags & IFLIB_PSEUDO_ETHER) == 0) {
5047 		bpfdetach(ifp);
5048 		if_detach(ifp);
5049 	} else {
5050 		ether_ifdetach(ifp);
5051 	}
5052 	/* XXX drain any dependent tasks */
5053 	tqg = qgroup_if_io_tqg;
5054 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
5055 		callout_drain(&txq->ift_timer);
5056 		if (txq->ift_task.gt_uniq != NULL)
5057 			taskqgroup_detach(tqg, &txq->ift_task);
5058 	}
5059 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
5060 		callout_drain(&rxq->ifr_watchdog);
5061 		if (rxq->ifr_task.gt_uniq != NULL)
5062 			taskqgroup_detach(tqg, &rxq->ifr_task);
5063 
5064 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5065 			free(fl->ifl_rx_bitmap, M_IFLIB);
5066 	}
5067 	tqg = qgroup_if_config_tqg;
5068 	if (ctx->ifc_admin_task.gt_uniq != NULL)
5069 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
5070 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
5071 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
5072 
5073 	iflib_tx_structures_free(ctx);
5074 	iflib_rx_structures_free(ctx);
5075 
5076 	iflib_deregister(ctx);
5077 
5078 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5079 		free(ctx->ifc_softc, M_IFLIB);
5080 	free(ctx, M_IFLIB);
5081 	return (0);
5082 }
5083 
5084 int
5085 iflib_device_attach(device_t dev)
5086 {
5087 	if_ctx_t ctx;
5088 	if_shared_ctx_t sctx;
5089 
5090 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
5091 		return (ENOTSUP);
5092 
5093 	pci_enable_busmaster(dev);
5094 
5095 	return (iflib_device_register(dev, NULL, sctx, &ctx));
5096 }
5097 
5098 int
5099 iflib_device_deregister(if_ctx_t ctx)
5100 {
5101 	if_t ifp = ctx->ifc_ifp;
5102 	iflib_txq_t txq;
5103 	iflib_rxq_t rxq;
5104 	device_t dev = ctx->ifc_dev;
5105 	int i, j;
5106 	struct taskqgroup *tqg;
5107 	iflib_fl_t fl;
5108 
5109 	/* Make sure VLANS are not using driver */
5110 	if (if_vlantrunkinuse(ifp)) {
5111 		device_printf(dev, "Vlan in use, detach first\n");
5112 		return (EBUSY);
5113 	}
5114 #ifdef PCI_IOV
5115 	if (!CTX_IS_VF(ctx) && pci_iov_detach(dev) != 0) {
5116 		device_printf(dev, "SR-IOV in use; detach first.\n");
5117 		return (EBUSY);
5118 	}
5119 #endif
5120 
5121 	STATE_LOCK(ctx);
5122 	ctx->ifc_flags |= IFC_IN_DETACH;
5123 	STATE_UNLOCK(ctx);
5124 
5125 	/* Unregister VLAN handlers before calling iflib_stop() */
5126 	iflib_unregister_vlan_handlers(ctx);
5127 
5128 	iflib_netmap_detach(ifp);
5129 	ether_ifdetach(ifp);
5130 
5131 	CTX_LOCK(ctx);
5132 	iflib_stop(ctx);
5133 	CTX_UNLOCK(ctx);
5134 
5135 	iflib_rem_pfil(ctx);
5136 	if (ctx->ifc_led_dev != NULL)
5137 		led_destroy(ctx->ifc_led_dev);
5138 	/* XXX drain any dependent tasks */
5139 	tqg = qgroup_if_io_tqg;
5140 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
5141 		callout_drain(&txq->ift_timer);
5142 		if (txq->ift_task.gt_uniq != NULL)
5143 			taskqgroup_detach(tqg, &txq->ift_task);
5144 	}
5145 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
5146 		if (rxq->ifr_task.gt_uniq != NULL)
5147 			taskqgroup_detach(tqg, &rxq->ifr_task);
5148 
5149 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5150 			free(fl->ifl_rx_bitmap, M_IFLIB);
5151 	}
5152 	tqg = qgroup_if_config_tqg;
5153 	if (ctx->ifc_admin_task.gt_uniq != NULL)
5154 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
5155 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
5156 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
5157 	CTX_LOCK(ctx);
5158 	IFDI_DETACH(ctx);
5159 	CTX_UNLOCK(ctx);
5160 
5161 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
5162 	iflib_free_intr_mem(ctx);
5163 
5164 	bus_generic_detach(dev);
5165 
5166 	iflib_tx_structures_free(ctx);
5167 	iflib_rx_structures_free(ctx);
5168 
5169 	iflib_deregister(ctx);
5170 
5171 	device_set_softc(ctx->ifc_dev, NULL);
5172 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5173 		free(ctx->ifc_softc, M_IFLIB);
5174 	unref_ctx_core_offset(ctx);
5175 	free(ctx, M_IFLIB);
5176 	return (0);
5177 }
5178 
5179 static void
5180 iflib_free_intr_mem(if_ctx_t ctx)
5181 {
5182 
5183 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) {
5184 		iflib_irq_free(ctx, &ctx->ifc_legacy_irq);
5185 	}
5186 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) {
5187 		pci_release_msi(ctx->ifc_dev);
5188 	}
5189 	if (ctx->ifc_msix_mem != NULL) {
5190 		bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY,
5191 		    rman_get_rid(ctx->ifc_msix_mem), ctx->ifc_msix_mem);
5192 		ctx->ifc_msix_mem = NULL;
5193 	}
5194 }
5195 
5196 int
5197 iflib_device_detach(device_t dev)
5198 {
5199 	if_ctx_t ctx = device_get_softc(dev);
5200 
5201 	return (iflib_device_deregister(ctx));
5202 }
5203 
5204 int
5205 iflib_device_suspend(device_t dev)
5206 {
5207 	if_ctx_t ctx = device_get_softc(dev);
5208 
5209 	CTX_LOCK(ctx);
5210 	IFDI_SUSPEND(ctx);
5211 	CTX_UNLOCK(ctx);
5212 
5213 	return bus_generic_suspend(dev);
5214 }
5215 int
5216 iflib_device_shutdown(device_t dev)
5217 {
5218 	if_ctx_t ctx = device_get_softc(dev);
5219 
5220 	CTX_LOCK(ctx);
5221 	IFDI_SHUTDOWN(ctx);
5222 	CTX_UNLOCK(ctx);
5223 
5224 	return bus_generic_suspend(dev);
5225 }
5226 
5227 
5228 int
5229 iflib_device_resume(device_t dev)
5230 {
5231 	if_ctx_t ctx = device_get_softc(dev);
5232 	iflib_txq_t txq = ctx->ifc_txqs;
5233 
5234 	CTX_LOCK(ctx);
5235 	IFDI_RESUME(ctx);
5236 	iflib_if_init_locked(ctx);
5237 	CTX_UNLOCK(ctx);
5238 	for (int i = 0; i < NTXQSETS(ctx); i++, txq++)
5239 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
5240 
5241 	return (bus_generic_resume(dev));
5242 }
5243 
5244 int
5245 iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params)
5246 {
5247 	int error;
5248 	if_ctx_t ctx = device_get_softc(dev);
5249 
5250 	CTX_LOCK(ctx);
5251 	error = IFDI_IOV_INIT(ctx, num_vfs, params);
5252 	CTX_UNLOCK(ctx);
5253 
5254 	return (error);
5255 }
5256 
5257 void
5258 iflib_device_iov_uninit(device_t dev)
5259 {
5260 	if_ctx_t ctx = device_get_softc(dev);
5261 
5262 	CTX_LOCK(ctx);
5263 	IFDI_IOV_UNINIT(ctx);
5264 	CTX_UNLOCK(ctx);
5265 }
5266 
5267 int
5268 iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params)
5269 {
5270 	int error;
5271 	if_ctx_t ctx = device_get_softc(dev);
5272 
5273 	CTX_LOCK(ctx);
5274 	error = IFDI_IOV_VF_ADD(ctx, vfnum, params);
5275 	CTX_UNLOCK(ctx);
5276 
5277 	return (error);
5278 }
5279 
5280 /*********************************************************************
5281  *
5282  *  MODULE FUNCTION DEFINITIONS
5283  *
5284  **********************************************************************/
5285 
5286 /*
5287  * - Start a fast taskqueue thread for each core
5288  * - Start a taskqueue for control operations
5289  */
5290 static int
5291 iflib_module_init(void)
5292 {
5293 	return (0);
5294 }
5295 
5296 static int
5297 iflib_module_event_handler(module_t mod, int what, void *arg)
5298 {
5299 	int err;
5300 
5301 	switch (what) {
5302 	case MOD_LOAD:
5303 		if ((err = iflib_module_init()) != 0)
5304 			return (err);
5305 		break;
5306 	case MOD_UNLOAD:
5307 		return (EBUSY);
5308 	default:
5309 		return (EOPNOTSUPP);
5310 	}
5311 
5312 	return (0);
5313 }
5314 
5315 /*********************************************************************
5316  *
5317  *  PUBLIC FUNCTION DEFINITIONS
5318  *     ordered as in iflib.h
5319  *
5320  **********************************************************************/
5321 
5322 
5323 static void
5324 _iflib_assert(if_shared_ctx_t sctx)
5325 {
5326 	int i;
5327 
5328 	MPASS(sctx->isc_tx_maxsize);
5329 	MPASS(sctx->isc_tx_maxsegsize);
5330 
5331 	MPASS(sctx->isc_rx_maxsize);
5332 	MPASS(sctx->isc_rx_nsegments);
5333 	MPASS(sctx->isc_rx_maxsegsize);
5334 
5335 	MPASS(sctx->isc_nrxqs >= 1 && sctx->isc_nrxqs <= 8);
5336 	for (i = 0; i < sctx->isc_nrxqs; i++) {
5337 		MPASS(sctx->isc_nrxd_min[i]);
5338 		MPASS(powerof2(sctx->isc_nrxd_min[i]));
5339 		MPASS(sctx->isc_nrxd_max[i]);
5340 		MPASS(powerof2(sctx->isc_nrxd_max[i]));
5341 		MPASS(sctx->isc_nrxd_default[i]);
5342 		MPASS(powerof2(sctx->isc_nrxd_default[i]));
5343 	}
5344 
5345 	MPASS(sctx->isc_ntxqs >= 1 && sctx->isc_ntxqs <= 8);
5346 	for (i = 0; i < sctx->isc_ntxqs; i++) {
5347 		MPASS(sctx->isc_ntxd_min[i]);
5348 		MPASS(powerof2(sctx->isc_ntxd_min[i]));
5349 		MPASS(sctx->isc_ntxd_max[i]);
5350 		MPASS(powerof2(sctx->isc_ntxd_max[i]));
5351 		MPASS(sctx->isc_ntxd_default[i]);
5352 		MPASS(powerof2(sctx->isc_ntxd_default[i]));
5353 	}
5354 }
5355 
5356 static void
5357 _iflib_pre_assert(if_softc_ctx_t scctx)
5358 {
5359 
5360 	MPASS(scctx->isc_txrx->ift_txd_encap);
5361 	MPASS(scctx->isc_txrx->ift_txd_flush);
5362 	MPASS(scctx->isc_txrx->ift_txd_credits_update);
5363 	MPASS(scctx->isc_txrx->ift_rxd_available);
5364 	MPASS(scctx->isc_txrx->ift_rxd_pkt_get);
5365 	MPASS(scctx->isc_txrx->ift_rxd_refill);
5366 	MPASS(scctx->isc_txrx->ift_rxd_flush);
5367 }
5368 
5369 static int
5370 iflib_register(if_ctx_t ctx)
5371 {
5372 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5373 	driver_t *driver = sctx->isc_driver;
5374 	device_t dev = ctx->ifc_dev;
5375 	if_t ifp;
5376 	u_char type;
5377 	int iflags;
5378 
5379 	if ((sctx->isc_flags & IFLIB_PSEUDO) == 0)
5380 		_iflib_assert(sctx);
5381 
5382 	CTX_LOCK_INIT(ctx);
5383 	STATE_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev));
5384 	if (sctx->isc_flags & IFLIB_PSEUDO) {
5385 		if (sctx->isc_flags & IFLIB_PSEUDO_ETHER)
5386 			type = IFT_ETHER;
5387 		else
5388 			type = IFT_PPP;
5389 	} else
5390 		type = IFT_ETHER;
5391 	ifp = ctx->ifc_ifp = if_alloc(type);
5392 	if (ifp == NULL) {
5393 		device_printf(dev, "can not allocate ifnet structure\n");
5394 		return (ENOMEM);
5395 	}
5396 
5397 	/*
5398 	 * Initialize our context's device specific methods
5399 	 */
5400 	kobj_init((kobj_t) ctx, (kobj_class_t) driver);
5401 	kobj_class_compile((kobj_class_t) driver);
5402 
5403 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
5404 	if_setsoftc(ifp, ctx);
5405 	if_setdev(ifp, dev);
5406 	if_setinitfn(ifp, iflib_if_init);
5407 	if_setioctlfn(ifp, iflib_if_ioctl);
5408 #ifdef ALTQ
5409 	if_setstartfn(ifp, iflib_altq_if_start);
5410 	if_settransmitfn(ifp, iflib_altq_if_transmit);
5411 	if_setsendqready(ifp);
5412 #else
5413 	if_settransmitfn(ifp, iflib_if_transmit);
5414 #endif
5415 	if_setqflushfn(ifp, iflib_if_qflush);
5416 	iflags = IFF_MULTICAST | IFF_KNOWSEPOCH;
5417 
5418 	if ((sctx->isc_flags & IFLIB_PSEUDO) &&
5419 		(sctx->isc_flags & IFLIB_PSEUDO_ETHER) == 0)
5420 		iflags |= IFF_POINTOPOINT;
5421 	else
5422 		iflags |= IFF_BROADCAST | IFF_SIMPLEX;
5423 	if_setflags(ifp, iflags);
5424 	ctx->ifc_vlan_attach_event =
5425 		EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx,
5426 							  EVENTHANDLER_PRI_FIRST);
5427 	ctx->ifc_vlan_detach_event =
5428 		EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx,
5429 							  EVENTHANDLER_PRI_FIRST);
5430 
5431 	if ((sctx->isc_flags & IFLIB_DRIVER_MEDIA) == 0) {
5432 		ctx->ifc_mediap = &ctx->ifc_media;
5433 		ifmedia_init(ctx->ifc_mediap, IFM_IMASK,
5434 		    iflib_media_change, iflib_media_status);
5435 	}
5436 	return (0);
5437 }
5438 
5439 static void
5440 iflib_unregister_vlan_handlers(if_ctx_t ctx)
5441 {
5442 	/* Unregister VLAN events */
5443 	if (ctx->ifc_vlan_attach_event != NULL) {
5444 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
5445 		ctx->ifc_vlan_attach_event = NULL;
5446 	}
5447 	if (ctx->ifc_vlan_detach_event != NULL) {
5448 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
5449 		ctx->ifc_vlan_detach_event = NULL;
5450 	}
5451 
5452 }
5453 
5454 static void
5455 iflib_deregister(if_ctx_t ctx)
5456 {
5457 	if_t ifp = ctx->ifc_ifp;
5458 
5459 	/* Remove all media */
5460 	ifmedia_removeall(&ctx->ifc_media);
5461 
5462 	/* Ensure that VLAN event handlers are unregistered */
5463 	iflib_unregister_vlan_handlers(ctx);
5464 
5465 	/* Release kobject reference */
5466 	kobj_delete((kobj_t) ctx, NULL);
5467 
5468 	/* Free the ifnet structure */
5469 	if_free(ifp);
5470 
5471 	STATE_LOCK_DESTROY(ctx);
5472 
5473 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
5474 	CTX_LOCK_DESTROY(ctx);
5475 }
5476 
5477 static int
5478 iflib_queues_alloc(if_ctx_t ctx)
5479 {
5480 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5481 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5482 	device_t dev = ctx->ifc_dev;
5483 	int nrxqsets = scctx->isc_nrxqsets;
5484 	int ntxqsets = scctx->isc_ntxqsets;
5485 	iflib_txq_t txq;
5486 	iflib_rxq_t rxq;
5487 	iflib_fl_t fl = NULL;
5488 	int i, j, cpu, err, txconf, rxconf;
5489 	iflib_dma_info_t ifdip;
5490 	uint32_t *rxqsizes = scctx->isc_rxqsizes;
5491 	uint32_t *txqsizes = scctx->isc_txqsizes;
5492 	uint8_t nrxqs = sctx->isc_nrxqs;
5493 	uint8_t ntxqs = sctx->isc_ntxqs;
5494 	int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1;
5495 	caddr_t *vaddrs;
5496 	uint64_t *paddrs;
5497 
5498 	KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1"));
5499 	KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1"));
5500 
5501 	/* Allocate the TX ring struct memory */
5502 	if (!(ctx->ifc_txqs =
5503 	    (iflib_txq_t) malloc(sizeof(struct iflib_txq) *
5504 	    ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5505 		device_printf(dev, "Unable to allocate TX ring memory\n");
5506 		err = ENOMEM;
5507 		goto fail;
5508 	}
5509 
5510 	/* Now allocate the RX */
5511 	if (!(ctx->ifc_rxqs =
5512 	    (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) *
5513 	    nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5514 		device_printf(dev, "Unable to allocate RX ring memory\n");
5515 		err = ENOMEM;
5516 		goto rx_fail;
5517 	}
5518 
5519 	txq = ctx->ifc_txqs;
5520 	rxq = ctx->ifc_rxqs;
5521 
5522 	/*
5523 	 * XXX handle allocation failure
5524 	 */
5525 	for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) {
5526 		/* Set up some basics */
5527 
5528 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs,
5529 		    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
5530 			device_printf(dev,
5531 			    "Unable to allocate TX DMA info memory\n");
5532 			err = ENOMEM;
5533 			goto err_tx_desc;
5534 		}
5535 		txq->ift_ifdi = ifdip;
5536 		for (j = 0; j < ntxqs; j++, ifdip++) {
5537 			if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, 0)) {
5538 				device_printf(dev,
5539 				    "Unable to allocate TX descriptors\n");
5540 				err = ENOMEM;
5541 				goto err_tx_desc;
5542 			}
5543 			txq->ift_txd_size[j] = scctx->isc_txd_size[j];
5544 			bzero((void *)ifdip->idi_vaddr, txqsizes[j]);
5545 		}
5546 		txq->ift_ctx = ctx;
5547 		txq->ift_id = i;
5548 		if (sctx->isc_flags & IFLIB_HAS_TXCQ) {
5549 			txq->ift_br_offset = 1;
5550 		} else {
5551 			txq->ift_br_offset = 0;
5552 		}
5553 		/* XXX fix this */
5554 		txq->ift_timer.c_cpu = cpu;
5555 
5556 		if (iflib_txsd_alloc(txq)) {
5557 			device_printf(dev, "Critical Failure setting up TX buffers\n");
5558 			err = ENOMEM;
5559 			goto err_tx_desc;
5560 		}
5561 
5562 		/* Initialize the TX lock */
5563 		snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:TX(%d):callout",
5564 		    device_get_nameunit(dev), txq->ift_id);
5565 		mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF);
5566 		callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0);
5567 
5568 		err = ifmp_ring_alloc(&txq->ift_br, 2048, txq, iflib_txq_drain,
5569 				      iflib_txq_can_drain, M_IFLIB, M_WAITOK);
5570 		if (err) {
5571 			/* XXX free any allocated rings */
5572 			device_printf(dev, "Unable to allocate buf_ring\n");
5573 			goto err_tx_desc;
5574 		}
5575 	}
5576 
5577 	for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) {
5578 		/* Set up some basics */
5579 		callout_init(&rxq->ifr_watchdog, 1);
5580 
5581 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs,
5582 		   M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
5583 			device_printf(dev,
5584 			    "Unable to allocate RX DMA info memory\n");
5585 			err = ENOMEM;
5586 			goto err_tx_desc;
5587 		}
5588 
5589 		rxq->ifr_ifdi = ifdip;
5590 		/* XXX this needs to be changed if #rx queues != #tx queues */
5591 		rxq->ifr_ntxqirq = 1;
5592 		rxq->ifr_txqid[0] = i;
5593 		for (j = 0; j < nrxqs; j++, ifdip++) {
5594 			if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, 0)) {
5595 				device_printf(dev,
5596 				    "Unable to allocate RX descriptors\n");
5597 				err = ENOMEM;
5598 				goto err_tx_desc;
5599 			}
5600 			bzero((void *)ifdip->idi_vaddr, rxqsizes[j]);
5601 		}
5602 		rxq->ifr_ctx = ctx;
5603 		rxq->ifr_id = i;
5604 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
5605 			rxq->ifr_fl_offset = 1;
5606 		} else {
5607 			rxq->ifr_fl_offset = 0;
5608 		}
5609 		rxq->ifr_nfl = nfree_lists;
5610 		if (!(fl =
5611 			  (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) {
5612 			device_printf(dev, "Unable to allocate free list memory\n");
5613 			err = ENOMEM;
5614 			goto err_tx_desc;
5615 		}
5616 		rxq->ifr_fl = fl;
5617 		for (j = 0; j < nfree_lists; j++) {
5618 			fl[j].ifl_rxq = rxq;
5619 			fl[j].ifl_id = j;
5620 			fl[j].ifl_ifdi = &rxq->ifr_ifdi[j + rxq->ifr_fl_offset];
5621 			fl[j].ifl_rxd_size = scctx->isc_rxd_size[j];
5622 		}
5623 		/* Allocate receive buffers for the ring */
5624 		if (iflib_rxsd_alloc(rxq)) {
5625 			device_printf(dev,
5626 			    "Critical Failure setting up receive buffers\n");
5627 			err = ENOMEM;
5628 			goto err_rx_desc;
5629 		}
5630 
5631 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5632 			fl->ifl_rx_bitmap = bit_alloc(fl->ifl_size, M_IFLIB,
5633 			    M_WAITOK);
5634 	}
5635 
5636 	/* TXQs */
5637 	vaddrs = malloc(sizeof(caddr_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5638 	paddrs = malloc(sizeof(uint64_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5639 	for (i = 0; i < ntxqsets; i++) {
5640 		iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi;
5641 
5642 		for (j = 0; j < ntxqs; j++, di++) {
5643 			vaddrs[i*ntxqs + j] = di->idi_vaddr;
5644 			paddrs[i*ntxqs + j] = di->idi_paddr;
5645 		}
5646 	}
5647 	if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) {
5648 		device_printf(ctx->ifc_dev,
5649 		    "Unable to allocate device TX queue\n");
5650 		iflib_tx_structures_free(ctx);
5651 		free(vaddrs, M_IFLIB);
5652 		free(paddrs, M_IFLIB);
5653 		goto err_rx_desc;
5654 	}
5655 	free(vaddrs, M_IFLIB);
5656 	free(paddrs, M_IFLIB);
5657 
5658 	/* RXQs */
5659 	vaddrs = malloc(sizeof(caddr_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5660 	paddrs = malloc(sizeof(uint64_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5661 	for (i = 0; i < nrxqsets; i++) {
5662 		iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi;
5663 
5664 		for (j = 0; j < nrxqs; j++, di++) {
5665 			vaddrs[i*nrxqs + j] = di->idi_vaddr;
5666 			paddrs[i*nrxqs + j] = di->idi_paddr;
5667 		}
5668 	}
5669 	if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) {
5670 		device_printf(ctx->ifc_dev,
5671 		    "Unable to allocate device RX queue\n");
5672 		iflib_tx_structures_free(ctx);
5673 		free(vaddrs, M_IFLIB);
5674 		free(paddrs, M_IFLIB);
5675 		goto err_rx_desc;
5676 	}
5677 	free(vaddrs, M_IFLIB);
5678 	free(paddrs, M_IFLIB);
5679 
5680 	return (0);
5681 
5682 /* XXX handle allocation failure changes */
5683 err_rx_desc:
5684 err_tx_desc:
5685 rx_fail:
5686 	if (ctx->ifc_rxqs != NULL)
5687 		free(ctx->ifc_rxqs, M_IFLIB);
5688 	ctx->ifc_rxqs = NULL;
5689 	if (ctx->ifc_txqs != NULL)
5690 		free(ctx->ifc_txqs, M_IFLIB);
5691 	ctx->ifc_txqs = NULL;
5692 fail:
5693 	return (err);
5694 }
5695 
5696 static int
5697 iflib_tx_structures_setup(if_ctx_t ctx)
5698 {
5699 	iflib_txq_t txq = ctx->ifc_txqs;
5700 	int i;
5701 
5702 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
5703 		iflib_txq_setup(txq);
5704 
5705 	return (0);
5706 }
5707 
5708 static void
5709 iflib_tx_structures_free(if_ctx_t ctx)
5710 {
5711 	iflib_txq_t txq = ctx->ifc_txqs;
5712 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5713 	int i, j;
5714 
5715 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
5716 		for (j = 0; j < sctx->isc_ntxqs; j++)
5717 			iflib_dma_free(&txq->ift_ifdi[j]);
5718 		iflib_txq_destroy(txq);
5719 	}
5720 	free(ctx->ifc_txqs, M_IFLIB);
5721 	ctx->ifc_txqs = NULL;
5722 	IFDI_QUEUES_FREE(ctx);
5723 }
5724 
5725 /*********************************************************************
5726  *
5727  *  Initialize all receive rings.
5728  *
5729  **********************************************************************/
5730 static int
5731 iflib_rx_structures_setup(if_ctx_t ctx)
5732 {
5733 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5734 	int q;
5735 #if defined(INET6) || defined(INET)
5736 	int err, i;
5737 #endif
5738 
5739 	for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) {
5740 #if defined(INET6) || defined(INET)
5741 		if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_LRO) {
5742 			err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp,
5743 			    TCP_LRO_ENTRIES, min(1024,
5744 			    ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset]));
5745 			if (err != 0) {
5746 				device_printf(ctx->ifc_dev,
5747 				    "LRO Initialization failed!\n");
5748 				goto fail;
5749 			}
5750 		}
5751 #endif
5752 		IFDI_RXQ_SETUP(ctx, rxq->ifr_id);
5753 	}
5754 	return (0);
5755 #if defined(INET6) || defined(INET)
5756 fail:
5757 	/*
5758 	 * Free LRO resources allocated so far, we will only handle
5759 	 * the rings that completed, the failing case will have
5760 	 * cleaned up for itself.  'q' failed, so its the terminus.
5761 	 */
5762 	rxq = ctx->ifc_rxqs;
5763 	for (i = 0; i < q; ++i, rxq++) {
5764 		if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_LRO)
5765 			tcp_lro_free(&rxq->ifr_lc);
5766 	}
5767 	return (err);
5768 #endif
5769 }
5770 
5771 /*********************************************************************
5772  *
5773  *  Free all receive rings.
5774  *
5775  **********************************************************************/
5776 static void
5777 iflib_rx_structures_free(if_ctx_t ctx)
5778 {
5779 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5780 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5781 	int i, j;
5782 
5783 	for (i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) {
5784 		for (j = 0; j < sctx->isc_nrxqs; j++)
5785 			iflib_dma_free(&rxq->ifr_ifdi[j]);
5786 		iflib_rx_sds_free(rxq);
5787 #if defined(INET6) || defined(INET)
5788 		if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_LRO)
5789 			tcp_lro_free(&rxq->ifr_lc);
5790 #endif
5791 	}
5792 	free(ctx->ifc_rxqs, M_IFLIB);
5793 	ctx->ifc_rxqs = NULL;
5794 }
5795 
5796 static int
5797 iflib_qset_structures_setup(if_ctx_t ctx)
5798 {
5799 	int err;
5800 
5801 	/*
5802 	 * It is expected that the caller takes care of freeing queues if this
5803 	 * fails.
5804 	 */
5805 	if ((err = iflib_tx_structures_setup(ctx)) != 0) {
5806 		device_printf(ctx->ifc_dev, "iflib_tx_structures_setup failed: %d\n", err);
5807 		return (err);
5808 	}
5809 
5810 	if ((err = iflib_rx_structures_setup(ctx)) != 0)
5811 		device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err);
5812 
5813 	return (err);
5814 }
5815 
5816 int
5817 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
5818 		driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, const char *name)
5819 {
5820 
5821 	return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name));
5822 }
5823 
5824 #ifdef SMP
5825 static int
5826 find_nth(if_ctx_t ctx, int qid)
5827 {
5828 	cpuset_t cpus;
5829 	int i, cpuid, eqid, count;
5830 
5831 	CPU_COPY(&ctx->ifc_cpus, &cpus);
5832 	count = CPU_COUNT(&cpus);
5833 	eqid = qid % count;
5834 	/* clear up to the qid'th bit */
5835 	for (i = 0; i < eqid; i++) {
5836 		cpuid = CPU_FFS(&cpus);
5837 		MPASS(cpuid != 0);
5838 		CPU_CLR(cpuid-1, &cpus);
5839 	}
5840 	cpuid = CPU_FFS(&cpus);
5841 	MPASS(cpuid != 0);
5842 	return (cpuid-1);
5843 }
5844 
5845 #ifdef SCHED_ULE
5846 extern struct cpu_group *cpu_top;              /* CPU topology */
5847 
5848 static int
5849 find_child_with_core(int cpu, struct cpu_group *grp)
5850 {
5851 	int i;
5852 
5853 	if (grp->cg_children == 0)
5854 		return -1;
5855 
5856 	MPASS(grp->cg_child);
5857 	for (i = 0; i < grp->cg_children; i++) {
5858 		if (CPU_ISSET(cpu, &grp->cg_child[i].cg_mask))
5859 			return i;
5860 	}
5861 
5862 	return -1;
5863 }
5864 
5865 /*
5866  * Find the nth "close" core to the specified core
5867  * "close" is defined as the deepest level that shares
5868  * at least an L2 cache.  With threads, this will be
5869  * threads on the same core.  If the shared cache is L3
5870  * or higher, simply returns the same core.
5871  */
5872 static int
5873 find_close_core(int cpu, int core_offset)
5874 {
5875 	struct cpu_group *grp;
5876 	int i;
5877 	int fcpu;
5878 	cpuset_t cs;
5879 
5880 	grp = cpu_top;
5881 	if (grp == NULL)
5882 		return cpu;
5883 	i = 0;
5884 	while ((i = find_child_with_core(cpu, grp)) != -1) {
5885 		/* If the child only has one cpu, don't descend */
5886 		if (grp->cg_child[i].cg_count <= 1)
5887 			break;
5888 		grp = &grp->cg_child[i];
5889 	}
5890 
5891 	/* If they don't share at least an L2 cache, use the same CPU */
5892 	if (grp->cg_level > CG_SHARE_L2 || grp->cg_level == CG_SHARE_NONE)
5893 		return cpu;
5894 
5895 	/* Now pick one */
5896 	CPU_COPY(&grp->cg_mask, &cs);
5897 
5898 	/* Add the selected CPU offset to core offset. */
5899 	for (i = 0; (fcpu = CPU_FFS(&cs)) != 0; i++) {
5900 		if (fcpu - 1 == cpu)
5901 			break;
5902 		CPU_CLR(fcpu - 1, &cs);
5903 	}
5904 	MPASS(fcpu);
5905 
5906 	core_offset += i;
5907 
5908 	CPU_COPY(&grp->cg_mask, &cs);
5909 	for (i = core_offset % grp->cg_count; i > 0; i--) {
5910 		MPASS(CPU_FFS(&cs));
5911 		CPU_CLR(CPU_FFS(&cs) - 1, &cs);
5912 	}
5913 	MPASS(CPU_FFS(&cs));
5914 	return CPU_FFS(&cs) - 1;
5915 }
5916 #else
5917 static int
5918 find_close_core(int cpu, int core_offset __unused)
5919 {
5920 	return cpu;
5921 }
5922 #endif
5923 
5924 static int
5925 get_core_offset(if_ctx_t ctx, iflib_intr_type_t type, int qid)
5926 {
5927 	switch (type) {
5928 	case IFLIB_INTR_TX:
5929 		/* TX queues get cores which share at least an L2 cache with the corresponding RX queue */
5930 		/* XXX handle multiple RX threads per core and more than two core per L2 group */
5931 		return qid / CPU_COUNT(&ctx->ifc_cpus) + 1;
5932 	case IFLIB_INTR_RX:
5933 	case IFLIB_INTR_RXTX:
5934 		/* RX queues get the specified core */
5935 		return qid / CPU_COUNT(&ctx->ifc_cpus);
5936 	default:
5937 		return -1;
5938 	}
5939 }
5940 #else
5941 #define get_core_offset(ctx, type, qid)	CPU_FIRST()
5942 #define find_close_core(cpuid, tid)	CPU_FIRST()
5943 #define find_nth(ctx, gid)		CPU_FIRST()
5944 #endif
5945 
5946 /* Just to avoid copy/paste */
5947 static inline int
5948 iflib_irq_set_affinity(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type,
5949     int qid, struct grouptask *gtask, struct taskqgroup *tqg, void *uniq,
5950     const char *name)
5951 {
5952 	device_t dev;
5953 	int co, cpuid, err, tid;
5954 
5955 	dev = ctx->ifc_dev;
5956 	co = ctx->ifc_sysctl_core_offset;
5957 	if (ctx->ifc_sysctl_separate_txrx && type == IFLIB_INTR_TX)
5958 		co += ctx->ifc_softc_ctx.isc_nrxqsets;
5959 	cpuid = find_nth(ctx, qid + co);
5960 	tid = get_core_offset(ctx, type, qid);
5961 	if (tid < 0) {
5962 		device_printf(dev, "get_core_offset failed\n");
5963 		return (EOPNOTSUPP);
5964 	}
5965 	cpuid = find_close_core(cpuid, tid);
5966 	err = taskqgroup_attach_cpu(tqg, gtask, uniq, cpuid, dev, irq->ii_res,
5967 	    name);
5968 	if (err) {
5969 		device_printf(dev, "taskqgroup_attach_cpu failed %d\n", err);
5970 		return (err);
5971 	}
5972 #ifdef notyet
5973 	if (cpuid > ctx->ifc_cpuid_highest)
5974 		ctx->ifc_cpuid_highest = cpuid;
5975 #endif
5976 	return (0);
5977 }
5978 
5979 int
5980 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid,
5981 			iflib_intr_type_t type, driver_filter_t *filter,
5982 			void *filter_arg, int qid, const char *name)
5983 {
5984 	device_t dev;
5985 	struct grouptask *gtask;
5986 	struct taskqgroup *tqg;
5987 	iflib_filter_info_t info;
5988 	gtask_fn_t *fn;
5989 	int tqrid, err;
5990 	driver_filter_t *intr_fast;
5991 	void *q;
5992 
5993 	info = &ctx->ifc_filter_info;
5994 	tqrid = rid;
5995 
5996 	switch (type) {
5997 	/* XXX merge tx/rx for netmap? */
5998 	case IFLIB_INTR_TX:
5999 		q = &ctx->ifc_txqs[qid];
6000 		info = &ctx->ifc_txqs[qid].ift_filter_info;
6001 		gtask = &ctx->ifc_txqs[qid].ift_task;
6002 		tqg = qgroup_if_io_tqg;
6003 		fn = _task_fn_tx;
6004 		intr_fast = iflib_fast_intr;
6005 		GROUPTASK_INIT(gtask, 0, fn, q);
6006 		ctx->ifc_flags |= IFC_NETMAP_TX_IRQ;
6007 		break;
6008 	case IFLIB_INTR_RX:
6009 		q = &ctx->ifc_rxqs[qid];
6010 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
6011 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6012 		tqg = qgroup_if_io_tqg;
6013 		fn = _task_fn_rx;
6014 		intr_fast = iflib_fast_intr;
6015 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6016 		break;
6017 	case IFLIB_INTR_RXTX:
6018 		q = &ctx->ifc_rxqs[qid];
6019 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
6020 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6021 		tqg = qgroup_if_io_tqg;
6022 		fn = _task_fn_rx;
6023 		intr_fast = iflib_fast_intr_rxtx;
6024 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6025 		break;
6026 	case IFLIB_INTR_ADMIN:
6027 		q = ctx;
6028 		tqrid = -1;
6029 		info = &ctx->ifc_filter_info;
6030 		gtask = &ctx->ifc_admin_task;
6031 		tqg = qgroup_if_config_tqg;
6032 		fn = _task_fn_admin;
6033 		intr_fast = iflib_fast_intr_ctx;
6034 		break;
6035 	default:
6036 		device_printf(ctx->ifc_dev, "%s: unknown net intr type\n",
6037 		    __func__);
6038 		return (EINVAL);
6039 	}
6040 
6041 	info->ifi_filter = filter;
6042 	info->ifi_filter_arg = filter_arg;
6043 	info->ifi_task = gtask;
6044 	info->ifi_ctx = q;
6045 
6046 	dev = ctx->ifc_dev;
6047 	err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info,  name);
6048 	if (err != 0) {
6049 		device_printf(dev, "_iflib_irq_alloc failed %d\n", err);
6050 		return (err);
6051 	}
6052 	if (type == IFLIB_INTR_ADMIN)
6053 		return (0);
6054 
6055 	if (tqrid != -1) {
6056 		err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg,
6057 		    q, name);
6058 		if (err)
6059 			return (err);
6060 	} else {
6061 		taskqgroup_attach(tqg, gtask, q, dev, irq->ii_res, name);
6062 	}
6063 
6064 	return (0);
6065 }
6066 
6067 void
6068 iflib_softirq_alloc_generic(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type, void *arg, int qid, const char *name)
6069 {
6070 	struct grouptask *gtask;
6071 	struct taskqgroup *tqg;
6072 	gtask_fn_t *fn;
6073 	void *q;
6074 	int err;
6075 
6076 	switch (type) {
6077 	case IFLIB_INTR_TX:
6078 		q = &ctx->ifc_txqs[qid];
6079 		gtask = &ctx->ifc_txqs[qid].ift_task;
6080 		tqg = qgroup_if_io_tqg;
6081 		fn = _task_fn_tx;
6082 		GROUPTASK_INIT(gtask, 0, fn, q);
6083 		break;
6084 	case IFLIB_INTR_RX:
6085 		q = &ctx->ifc_rxqs[qid];
6086 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6087 		tqg = qgroup_if_io_tqg;
6088 		fn = _task_fn_rx;
6089 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6090 		break;
6091 	case IFLIB_INTR_IOV:
6092 		q = ctx;
6093 		gtask = &ctx->ifc_vflr_task;
6094 		tqg = qgroup_if_config_tqg;
6095 		fn = _task_fn_iov;
6096 		GROUPTASK_INIT(gtask, 0, fn, q);
6097 		break;
6098 	default:
6099 		panic("unknown net intr type");
6100 	}
6101 	if (irq != NULL) {
6102 		err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg,
6103 		    q, name);
6104 		if (err)
6105 			taskqgroup_attach(tqg, gtask, q, ctx->ifc_dev,
6106 			    irq->ii_res, name);
6107 	} else {
6108 		taskqgroup_attach(tqg, gtask, q, NULL, NULL, name);
6109 	}
6110 }
6111 
6112 void
6113 iflib_irq_free(if_ctx_t ctx, if_irq_t irq)
6114 {
6115 
6116 	if (irq->ii_tag)
6117 		bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag);
6118 
6119 	if (irq->ii_res)
6120 		bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ,
6121 		    rman_get_rid(irq->ii_res), irq->ii_res);
6122 }
6123 
6124 static int
6125 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, const char *name)
6126 {
6127 	iflib_txq_t txq = ctx->ifc_txqs;
6128 	iflib_rxq_t rxq = ctx->ifc_rxqs;
6129 	if_irq_t irq = &ctx->ifc_legacy_irq;
6130 	iflib_filter_info_t info;
6131 	device_t dev;
6132 	struct grouptask *gtask;
6133 	struct resource *res;
6134 	struct taskqgroup *tqg;
6135 	void *q;
6136 	int err, tqrid;
6137 	bool rx_only;
6138 
6139 	q = &ctx->ifc_rxqs[0];
6140 	info = &rxq[0].ifr_filter_info;
6141 	gtask = &rxq[0].ifr_task;
6142 	tqg = qgroup_if_io_tqg;
6143 	tqrid = *rid;
6144 	rx_only = (ctx->ifc_sctx->isc_flags & IFLIB_SINGLE_IRQ_RX_ONLY) != 0;
6145 
6146 	ctx->ifc_flags |= IFC_LEGACY;
6147 	info->ifi_filter = filter;
6148 	info->ifi_filter_arg = filter_arg;
6149 	info->ifi_task = gtask;
6150 	info->ifi_ctx = rx_only ? ctx : q;
6151 
6152 	dev = ctx->ifc_dev;
6153 	/* We allocate a single interrupt resource */
6154 	err = _iflib_irq_alloc(ctx, irq, tqrid, rx_only ? iflib_fast_intr_ctx :
6155 	    iflib_fast_intr_rxtx, NULL, info, name);
6156 	if (err != 0)
6157 		return (err);
6158 	NET_GROUPTASK_INIT(gtask, 0, _task_fn_rx, q);
6159 	res = irq->ii_res;
6160 	taskqgroup_attach(tqg, gtask, q, dev, res, name);
6161 
6162 	GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq);
6163 	taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, dev, res,
6164 	    "tx");
6165 	return (0);
6166 }
6167 
6168 void
6169 iflib_led_create(if_ctx_t ctx)
6170 {
6171 
6172 	ctx->ifc_led_dev = led_create(iflib_led_func, ctx,
6173 	    device_get_nameunit(ctx->ifc_dev));
6174 }
6175 
6176 void
6177 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid)
6178 {
6179 
6180 	GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
6181 }
6182 
6183 void
6184 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid)
6185 {
6186 
6187 	GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task);
6188 }
6189 
6190 void
6191 iflib_admin_intr_deferred(if_ctx_t ctx)
6192 {
6193 
6194 	MPASS(ctx->ifc_admin_task.gt_taskqueue != NULL);
6195 	GROUPTASK_ENQUEUE(&ctx->ifc_admin_task);
6196 }
6197 
6198 void
6199 iflib_iov_intr_deferred(if_ctx_t ctx)
6200 {
6201 
6202 	GROUPTASK_ENQUEUE(&ctx->ifc_vflr_task);
6203 }
6204 
6205 void
6206 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, const char *name)
6207 {
6208 
6209 	taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, NULL, NULL,
6210 	    name);
6211 }
6212 
6213 void
6214 iflib_config_gtask_init(void *ctx, struct grouptask *gtask, gtask_fn_t *fn,
6215 	const char *name)
6216 {
6217 
6218 	GROUPTASK_INIT(gtask, 0, fn, ctx);
6219 	taskqgroup_attach(qgroup_if_config_tqg, gtask, gtask, NULL, NULL,
6220 	    name);
6221 }
6222 
6223 void
6224 iflib_config_gtask_deinit(struct grouptask *gtask)
6225 {
6226 
6227 	taskqgroup_detach(qgroup_if_config_tqg, gtask);
6228 }
6229 
6230 void
6231 iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate)
6232 {
6233 	if_t ifp = ctx->ifc_ifp;
6234 	iflib_txq_t txq = ctx->ifc_txqs;
6235 
6236 	if_setbaudrate(ifp, baudrate);
6237 	if (baudrate >= IF_Gbps(10)) {
6238 		STATE_LOCK(ctx);
6239 		ctx->ifc_flags |= IFC_PREFETCH;
6240 		STATE_UNLOCK(ctx);
6241 	}
6242 	/* If link down, disable watchdog */
6243 	if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) {
6244 		for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++)
6245 			txq->ift_qstatus = IFLIB_QUEUE_IDLE;
6246 	}
6247 	ctx->ifc_link_state = link_state;
6248 	if_link_state_change(ifp, link_state);
6249 }
6250 
6251 static int
6252 iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq)
6253 {
6254 	int credits;
6255 #ifdef INVARIANTS
6256 	int credits_pre = txq->ift_cidx_processed;
6257 #endif
6258 
6259 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
6260 	    BUS_DMASYNC_POSTREAD);
6261 	if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, true)) == 0)
6262 		return (0);
6263 
6264 	txq->ift_processed += credits;
6265 	txq->ift_cidx_processed += credits;
6266 
6267 	MPASS(credits_pre + credits == txq->ift_cidx_processed);
6268 	if (txq->ift_cidx_processed >= txq->ift_size)
6269 		txq->ift_cidx_processed -= txq->ift_size;
6270 	return (credits);
6271 }
6272 
6273 static int
6274 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget)
6275 {
6276 	iflib_fl_t fl;
6277 	u_int i;
6278 
6279 	for (i = 0, fl = &rxq->ifr_fl[0]; i < rxq->ifr_nfl; i++, fl++)
6280 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
6281 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
6282 	return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx,
6283 	    budget));
6284 }
6285 
6286 void
6287 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name,
6288 	const char *description, if_int_delay_info_t info,
6289 	int offset, int value)
6290 {
6291 	info->iidi_ctx = ctx;
6292 	info->iidi_offset = offset;
6293 	info->iidi_value = value;
6294 	SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev),
6295 	    SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)),
6296 	    OID_AUTO, name, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE,
6297 	    info, 0, iflib_sysctl_int_delay, "I", description);
6298 }
6299 
6300 struct sx *
6301 iflib_ctx_lock_get(if_ctx_t ctx)
6302 {
6303 
6304 	return (&ctx->ifc_ctx_sx);
6305 }
6306 
6307 static int
6308 iflib_msix_init(if_ctx_t ctx)
6309 {
6310 	device_t dev = ctx->ifc_dev;
6311 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6312 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6313 	int admincnt, bar, err, iflib_num_rx_queues, iflib_num_tx_queues;
6314 	int msgs, queuemsgs, queues, rx_queues, tx_queues, vectors;
6315 
6316 	iflib_num_tx_queues = ctx->ifc_sysctl_ntxqs;
6317 	iflib_num_rx_queues = ctx->ifc_sysctl_nrxqs;
6318 
6319 	if (bootverbose)
6320 		device_printf(dev, "msix_init qsets capped at %d\n",
6321 		    imax(scctx->isc_ntxqsets, scctx->isc_nrxqsets));
6322 
6323 	/* Override by tuneable */
6324 	if (scctx->isc_disable_msix)
6325 		goto msi;
6326 
6327 	/* First try MSI-X */
6328 	if ((msgs = pci_msix_count(dev)) == 0) {
6329 		if (bootverbose)
6330 			device_printf(dev, "MSI-X not supported or disabled\n");
6331 		goto msi;
6332 	}
6333 
6334 	bar = ctx->ifc_softc_ctx.isc_msix_bar;
6335 	/*
6336 	 * bar == -1 => "trust me I know what I'm doing"
6337 	 * Some drivers are for hardware that is so shoddily
6338 	 * documented that no one knows which bars are which
6339 	 * so the developer has to map all bars. This hack
6340 	 * allows shoddy garbage to use MSI-X in this framework.
6341 	 */
6342 	if (bar != -1) {
6343 		ctx->ifc_msix_mem = bus_alloc_resource_any(dev,
6344 	            SYS_RES_MEMORY, &bar, RF_ACTIVE);
6345 		if (ctx->ifc_msix_mem == NULL) {
6346 			device_printf(dev, "Unable to map MSI-X table\n");
6347 			goto msi;
6348 		}
6349 	}
6350 
6351 	admincnt = sctx->isc_admin_intrcnt;
6352 #if IFLIB_DEBUG
6353 	/* use only 1 qset in debug mode */
6354 	queuemsgs = min(msgs - admincnt, 1);
6355 #else
6356 	queuemsgs = msgs - admincnt;
6357 #endif
6358 #ifdef RSS
6359 	queues = imin(queuemsgs, rss_getnumbuckets());
6360 #else
6361 	queues = queuemsgs;
6362 #endif
6363 	queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues);
6364 	if (bootverbose)
6365 		device_printf(dev,
6366 		    "intr CPUs: %d queue msgs: %d admincnt: %d\n",
6367 		    CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt);
6368 #ifdef  RSS
6369 	/* If we're doing RSS, clamp at the number of RSS buckets */
6370 	if (queues > rss_getnumbuckets())
6371 		queues = rss_getnumbuckets();
6372 #endif
6373 	if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt)
6374 		rx_queues = iflib_num_rx_queues;
6375 	else
6376 		rx_queues = queues;
6377 
6378 	if (rx_queues > scctx->isc_nrxqsets)
6379 		rx_queues = scctx->isc_nrxqsets;
6380 
6381 	/*
6382 	 * We want this to be all logical CPUs by default
6383 	 */
6384 	if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues)
6385 		tx_queues = iflib_num_tx_queues;
6386 	else
6387 		tx_queues = mp_ncpus;
6388 
6389 	if (tx_queues > scctx->isc_ntxqsets)
6390 		tx_queues = scctx->isc_ntxqsets;
6391 
6392 	if (ctx->ifc_sysctl_qs_eq_override == 0) {
6393 #ifdef INVARIANTS
6394 		if (tx_queues != rx_queues)
6395 			device_printf(dev,
6396 			    "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n",
6397 			    min(rx_queues, tx_queues), min(rx_queues, tx_queues));
6398 #endif
6399 		tx_queues = min(rx_queues, tx_queues);
6400 		rx_queues = min(rx_queues, tx_queues);
6401 	}
6402 
6403 	vectors = rx_queues + admincnt;
6404 	if (msgs < vectors) {
6405 		device_printf(dev,
6406 		    "insufficient number of MSI-X vectors "
6407 		    "(supported %d, need %d)\n", msgs, vectors);
6408 		goto msi;
6409 	}
6410 
6411 	device_printf(dev, "Using %d RX queues %d TX queues\n", rx_queues,
6412 	    tx_queues);
6413 	msgs = vectors;
6414 	if ((err = pci_alloc_msix(dev, &vectors)) == 0) {
6415 		if (vectors != msgs) {
6416 			device_printf(dev,
6417 			    "Unable to allocate sufficient MSI-X vectors "
6418 			    "(got %d, need %d)\n", vectors, msgs);
6419 			pci_release_msi(dev);
6420 			if (bar != -1) {
6421 				bus_release_resource(dev, SYS_RES_MEMORY, bar,
6422 				    ctx->ifc_msix_mem);
6423 				ctx->ifc_msix_mem = NULL;
6424 			}
6425 			goto msi;
6426 		}
6427 		device_printf(dev, "Using MSI-X interrupts with %d vectors\n",
6428 		    vectors);
6429 		scctx->isc_vectors = vectors;
6430 		scctx->isc_nrxqsets = rx_queues;
6431 		scctx->isc_ntxqsets = tx_queues;
6432 		scctx->isc_intr = IFLIB_INTR_MSIX;
6433 
6434 		return (vectors);
6435 	} else {
6436 		device_printf(dev,
6437 		    "failed to allocate %d MSI-X vectors, err: %d\n", vectors,
6438 		    err);
6439 		if (bar != -1) {
6440 			bus_release_resource(dev, SYS_RES_MEMORY, bar,
6441 			    ctx->ifc_msix_mem);
6442 			ctx->ifc_msix_mem = NULL;
6443 		}
6444 	}
6445 
6446 msi:
6447 	vectors = pci_msi_count(dev);
6448 	scctx->isc_nrxqsets = 1;
6449 	scctx->isc_ntxqsets = 1;
6450 	scctx->isc_vectors = vectors;
6451 	if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) {
6452 		device_printf(dev,"Using an MSI interrupt\n");
6453 		scctx->isc_intr = IFLIB_INTR_MSI;
6454 	} else {
6455 		scctx->isc_vectors = 1;
6456 		device_printf(dev,"Using a Legacy interrupt\n");
6457 		scctx->isc_intr = IFLIB_INTR_LEGACY;
6458 	}
6459 
6460 	return (vectors);
6461 }
6462 
6463 static const char *ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" };
6464 
6465 static int
6466 mp_ring_state_handler(SYSCTL_HANDLER_ARGS)
6467 {
6468 	int rc;
6469 	uint16_t *state = ((uint16_t *)oidp->oid_arg1);
6470 	struct sbuf *sb;
6471 	const char *ring_state = "UNKNOWN";
6472 
6473 	/* XXX needed ? */
6474 	rc = sysctl_wire_old_buffer(req, 0);
6475 	MPASS(rc == 0);
6476 	if (rc != 0)
6477 		return (rc);
6478 	sb = sbuf_new_for_sysctl(NULL, NULL, 80, req);
6479 	MPASS(sb != NULL);
6480 	if (sb == NULL)
6481 		return (ENOMEM);
6482 	if (state[3] <= 3)
6483 		ring_state = ring_states[state[3]];
6484 
6485 	sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s",
6486 		    state[0], state[1], state[2], ring_state);
6487 	rc = sbuf_finish(sb);
6488 	sbuf_delete(sb);
6489         return(rc);
6490 }
6491 
6492 enum iflib_ndesc_handler {
6493 	IFLIB_NTXD_HANDLER,
6494 	IFLIB_NRXD_HANDLER,
6495 };
6496 
6497 static int
6498 mp_ndesc_handler(SYSCTL_HANDLER_ARGS)
6499 {
6500 	if_ctx_t ctx = (void *)arg1;
6501 	enum iflib_ndesc_handler type = arg2;
6502 	char buf[256] = {0};
6503 	qidx_t *ndesc;
6504 	char *p, *next;
6505 	int nqs, rc, i;
6506 
6507 	nqs = 8;
6508 	switch(type) {
6509 	case IFLIB_NTXD_HANDLER:
6510 		ndesc = ctx->ifc_sysctl_ntxds;
6511 		if (ctx->ifc_sctx)
6512 			nqs = ctx->ifc_sctx->isc_ntxqs;
6513 		break;
6514 	case IFLIB_NRXD_HANDLER:
6515 		ndesc = ctx->ifc_sysctl_nrxds;
6516 		if (ctx->ifc_sctx)
6517 			nqs = ctx->ifc_sctx->isc_nrxqs;
6518 		break;
6519 	default:
6520 		printf("%s: unhandled type\n", __func__);
6521 		return (EINVAL);
6522 	}
6523 	if (nqs == 0)
6524 		nqs = 8;
6525 
6526 	for (i=0; i<8; i++) {
6527 		if (i >= nqs)
6528 			break;
6529 		if (i)
6530 			strcat(buf, ",");
6531 		sprintf(strchr(buf, 0), "%d", ndesc[i]);
6532 	}
6533 
6534 	rc = sysctl_handle_string(oidp, buf, sizeof(buf), req);
6535 	if (rc || req->newptr == NULL)
6536 		return rc;
6537 
6538 	for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p;
6539 	    i++, p = strsep(&next, " ,")) {
6540 		ndesc[i] = strtoul(p, NULL, 10);
6541 	}
6542 
6543 	return(rc);
6544 }
6545 
6546 #define NAME_BUFLEN 32
6547 static void
6548 iflib_add_device_sysctl_pre(if_ctx_t ctx)
6549 {
6550         device_t dev = iflib_get_dev(ctx);
6551 	struct sysctl_oid_list *child, *oid_list;
6552 	struct sysctl_ctx_list *ctx_list;
6553 	struct sysctl_oid *node;
6554 
6555 	ctx_list = device_get_sysctl_ctx(dev);
6556 	child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev));
6557 	ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, "iflib",
6558 	    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "IFLIB fields");
6559 	oid_list = SYSCTL_CHILDREN(node);
6560 
6561 	SYSCTL_ADD_CONST_STRING(ctx_list, oid_list, OID_AUTO, "driver_version",
6562 		       CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version,
6563 		       "driver version");
6564 
6565 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs",
6566 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0,
6567 			"# of txqs to use, 0 => use default #");
6568 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs",
6569 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0,
6570 			"# of rxqs to use, 0 => use default #");
6571 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_qs_enable",
6572 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0,
6573                        "permit #txq != #rxq");
6574 	SYSCTL_ADD_INT(ctx_list, oid_list, OID_AUTO, "disable_msix",
6575                       CTLFLAG_RWTUN, &ctx->ifc_softc_ctx.isc_disable_msix, 0,
6576                       "disable MSI-X (default 0)");
6577 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "rx_budget",
6578 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_rx_budget, 0,
6579 		       "set the RX budget");
6580 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "tx_abdicate",
6581 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_tx_abdicate, 0,
6582 		       "cause TX to abdicate instead of running to completion");
6583 	ctx->ifc_sysctl_core_offset = CORE_OFFSET_UNSPECIFIED;
6584 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "core_offset",
6585 		       CTLFLAG_RDTUN, &ctx->ifc_sysctl_core_offset, 0,
6586 		       "offset to start using cores at");
6587 	SYSCTL_ADD_U8(ctx_list, oid_list, OID_AUTO, "separate_txrx",
6588 		       CTLFLAG_RDTUN, &ctx->ifc_sysctl_separate_txrx, 0,
6589 		       "use separate cores for TX and RX");
6590 
6591 	/* XXX change for per-queue sizes */
6592 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_ntxds",
6593 	    CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NEEDGIANT, ctx,
6594 	    IFLIB_NTXD_HANDLER, mp_ndesc_handler, "A",
6595 	    "list of # of TX descriptors to use, 0 = use default #");
6596 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_nrxds",
6597 	    CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NEEDGIANT, ctx,
6598 	    IFLIB_NRXD_HANDLER, mp_ndesc_handler, "A",
6599 	    "list of # of RX descriptors to use, 0 = use default #");
6600 }
6601 
6602 static void
6603 iflib_add_device_sysctl_post(if_ctx_t ctx)
6604 {
6605 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6606 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6607         device_t dev = iflib_get_dev(ctx);
6608 	struct sysctl_oid_list *child;
6609 	struct sysctl_ctx_list *ctx_list;
6610 	iflib_fl_t fl;
6611 	iflib_txq_t txq;
6612 	iflib_rxq_t rxq;
6613 	int i, j;
6614 	char namebuf[NAME_BUFLEN];
6615 	char *qfmt;
6616 	struct sysctl_oid *queue_node, *fl_node, *node;
6617 	struct sysctl_oid_list *queue_list, *fl_list;
6618 	ctx_list = device_get_sysctl_ctx(dev);
6619 
6620 	node = ctx->ifc_sysctl_node;
6621 	child = SYSCTL_CHILDREN(node);
6622 
6623 	if (scctx->isc_ntxqsets > 100)
6624 		qfmt = "txq%03d";
6625 	else if (scctx->isc_ntxqsets > 10)
6626 		qfmt = "txq%02d";
6627 	else
6628 		qfmt = "txq%d";
6629 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
6630 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6631 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6632 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "Queue Name");
6633 		queue_list = SYSCTL_CHILDREN(queue_node);
6634 #if MEMORY_LOGGING
6635 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued",
6636 				CTLFLAG_RD,
6637 				&txq->ift_dequeued, "total mbufs freed");
6638 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued",
6639 				CTLFLAG_RD,
6640 				&txq->ift_enqueued, "total mbufs enqueued");
6641 #endif
6642 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag",
6643 				   CTLFLAG_RD,
6644 				   &txq->ift_mbuf_defrag, "# of times m_defrag was called");
6645 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "m_pullups",
6646 				   CTLFLAG_RD,
6647 				   &txq->ift_pullups, "# of times m_pullup was called");
6648 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag_failed",
6649 				   CTLFLAG_RD,
6650 				   &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed");
6651 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_desc_avail",
6652 				   CTLFLAG_RD,
6653 				   &txq->ift_no_desc_avail, "# of times no descriptors were available");
6654 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "tx_map_failed",
6655 				   CTLFLAG_RD,
6656 				   &txq->ift_map_failed, "# of times DMA map failed");
6657 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txd_encap_efbig",
6658 				   CTLFLAG_RD,
6659 				   &txq->ift_txd_encap_efbig, "# of times txd_encap returned EFBIG");
6660 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_tx_dma_setup",
6661 				   CTLFLAG_RD,
6662 				   &txq->ift_no_tx_dma_setup, "# of times map failed for other than EFBIG");
6663 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx",
6664 				   CTLFLAG_RD,
6665 				   &txq->ift_pidx, 1, "Producer Index");
6666 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx",
6667 				   CTLFLAG_RD,
6668 				   &txq->ift_cidx, 1, "Consumer Index");
6669 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx_processed",
6670 				   CTLFLAG_RD,
6671 				   &txq->ift_cidx_processed, 1, "Consumer Index seen by credit update");
6672 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use",
6673 				   CTLFLAG_RD,
6674 				   &txq->ift_in_use, 1, "descriptors in use");
6675 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_processed",
6676 				   CTLFLAG_RD,
6677 				   &txq->ift_processed, "descriptors procesed for clean");
6678 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned",
6679 				   CTLFLAG_RD,
6680 				   &txq->ift_cleaned, "total cleaned");
6681 		SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state",
6682 		    CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_NEEDGIANT,
6683 		    __DEVOLATILE(uint64_t *, &txq->ift_br->state), 0,
6684 		    mp_ring_state_handler, "A", "soft ring state");
6685 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_enqueues",
6686 				       CTLFLAG_RD, &txq->ift_br->enqueues,
6687 				       "# of enqueues to the mp_ring for this queue");
6688 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_drops",
6689 				       CTLFLAG_RD, &txq->ift_br->drops,
6690 				       "# of drops in the mp_ring for this queue");
6691 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_starts",
6692 				       CTLFLAG_RD, &txq->ift_br->starts,
6693 				       "# of normal consumer starts in the mp_ring for this queue");
6694 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_stalls",
6695 				       CTLFLAG_RD, &txq->ift_br->stalls,
6696 					       "# of consumer stalls in the mp_ring for this queue");
6697 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_restarts",
6698 			       CTLFLAG_RD, &txq->ift_br->restarts,
6699 				       "# of consumer restarts in the mp_ring for this queue");
6700 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_abdications",
6701 				       CTLFLAG_RD, &txq->ift_br->abdications,
6702 				       "# of consumer abdications in the mp_ring for this queue");
6703 	}
6704 
6705 	if (scctx->isc_nrxqsets > 100)
6706 		qfmt = "rxq%03d";
6707 	else if (scctx->isc_nrxqsets > 10)
6708 		qfmt = "rxq%02d";
6709 	else
6710 		qfmt = "rxq%d";
6711 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
6712 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6713 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6714 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "Queue Name");
6715 		queue_list = SYSCTL_CHILDREN(queue_node);
6716 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
6717 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_cidx",
6718 				       CTLFLAG_RD,
6719 				       &rxq->ifr_cq_cidx, 1, "Consumer Index");
6720 		}
6721 
6722 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
6723 			snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j);
6724 			fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list, OID_AUTO, namebuf,
6725 			    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "freelist Name");
6726 			fl_list = SYSCTL_CHILDREN(fl_node);
6727 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx",
6728 				       CTLFLAG_RD,
6729 				       &fl->ifl_pidx, 1, "Producer Index");
6730 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx",
6731 				       CTLFLAG_RD,
6732 				       &fl->ifl_cidx, 1, "Consumer Index");
6733 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits",
6734 				       CTLFLAG_RD,
6735 				       &fl->ifl_credits, 1, "credits available");
6736 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "buf_size",
6737 				       CTLFLAG_RD,
6738 				       &fl->ifl_buf_size, 1, "buffer size");
6739 #if MEMORY_LOGGING
6740 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_enqueued",
6741 					CTLFLAG_RD,
6742 					&fl->ifl_m_enqueued, "mbufs allocated");
6743 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_dequeued",
6744 					CTLFLAG_RD,
6745 					&fl->ifl_m_dequeued, "mbufs freed");
6746 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_enqueued",
6747 					CTLFLAG_RD,
6748 					&fl->ifl_cl_enqueued, "clusters allocated");
6749 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_dequeued",
6750 					CTLFLAG_RD,
6751 					&fl->ifl_cl_dequeued, "clusters freed");
6752 #endif
6753 
6754 		}
6755 	}
6756 
6757 }
6758 
6759 void
6760 iflib_request_reset(if_ctx_t ctx)
6761 {
6762 
6763 	STATE_LOCK(ctx);
6764 	ctx->ifc_flags |= IFC_DO_RESET;
6765 	STATE_UNLOCK(ctx);
6766 }
6767 
6768 #ifndef __NO_STRICT_ALIGNMENT
6769 static struct mbuf *
6770 iflib_fixup_rx(struct mbuf *m)
6771 {
6772 	struct mbuf *n;
6773 
6774 	if (m->m_len <= (MCLBYTES - ETHER_HDR_LEN)) {
6775 		bcopy(m->m_data, m->m_data + ETHER_HDR_LEN, m->m_len);
6776 		m->m_data += ETHER_HDR_LEN;
6777 		n = m;
6778 	} else {
6779 		MGETHDR(n, M_NOWAIT, MT_DATA);
6780 		if (n == NULL) {
6781 			m_freem(m);
6782 			return (NULL);
6783 		}
6784 		bcopy(m->m_data, n->m_data, ETHER_HDR_LEN);
6785 		m->m_data += ETHER_HDR_LEN;
6786 		m->m_len -= ETHER_HDR_LEN;
6787 		n->m_len = ETHER_HDR_LEN;
6788 		M_MOVE_PKTHDR(n, m);
6789 		n->m_next = m;
6790 	}
6791 	return (n);
6792 }
6793 #endif
6794 
6795 #ifdef DEBUGNET
6796 static void
6797 iflib_debugnet_init(if_t ifp, int *nrxr, int *ncl, int *clsize)
6798 {
6799 	if_ctx_t ctx;
6800 
6801 	ctx = if_getsoftc(ifp);
6802 	CTX_LOCK(ctx);
6803 	*nrxr = NRXQSETS(ctx);
6804 	*ncl = ctx->ifc_rxqs[0].ifr_fl->ifl_size;
6805 	*clsize = ctx->ifc_rxqs[0].ifr_fl->ifl_buf_size;
6806 	CTX_UNLOCK(ctx);
6807 }
6808 
6809 static void
6810 iflib_debugnet_event(if_t ifp, enum debugnet_ev event)
6811 {
6812 	if_ctx_t ctx;
6813 	if_softc_ctx_t scctx;
6814 	iflib_fl_t fl;
6815 	iflib_rxq_t rxq;
6816 	int i, j;
6817 
6818 	ctx = if_getsoftc(ifp);
6819 	scctx = &ctx->ifc_softc_ctx;
6820 
6821 	switch (event) {
6822 	case DEBUGNET_START:
6823 		for (i = 0; i < scctx->isc_nrxqsets; i++) {
6824 			rxq = &ctx->ifc_rxqs[i];
6825 			for (j = 0; j < rxq->ifr_nfl; j++) {
6826 				fl = rxq->ifr_fl;
6827 				fl->ifl_zone = m_getzone(fl->ifl_buf_size);
6828 			}
6829 		}
6830 		iflib_no_tx_batch = 1;
6831 		break;
6832 	default:
6833 		break;
6834 	}
6835 }
6836 
6837 static int
6838 iflib_debugnet_transmit(if_t ifp, struct mbuf *m)
6839 {
6840 	if_ctx_t ctx;
6841 	iflib_txq_t txq;
6842 	int error;
6843 
6844 	ctx = if_getsoftc(ifp);
6845 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
6846 	    IFF_DRV_RUNNING)
6847 		return (EBUSY);
6848 
6849 	txq = &ctx->ifc_txqs[0];
6850 	error = iflib_encap(txq, &m);
6851 	if (error == 0)
6852 		(void)iflib_txd_db_check(ctx, txq, true, txq->ift_in_use);
6853 	return (error);
6854 }
6855 
6856 static int
6857 iflib_debugnet_poll(if_t ifp, int count)
6858 {
6859 	struct epoch_tracker et;
6860 	if_ctx_t ctx;
6861 	if_softc_ctx_t scctx;
6862 	iflib_txq_t txq;
6863 	int i;
6864 
6865 	ctx = if_getsoftc(ifp);
6866 	scctx = &ctx->ifc_softc_ctx;
6867 
6868 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
6869 	    IFF_DRV_RUNNING)
6870 		return (EBUSY);
6871 
6872 	txq = &ctx->ifc_txqs[0];
6873 	(void)iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
6874 
6875 	NET_EPOCH_ENTER(et);
6876 	for (i = 0; i < scctx->isc_nrxqsets; i++)
6877 		(void)iflib_rxeof(&ctx->ifc_rxqs[i], 16 /* XXX */);
6878 	NET_EPOCH_EXIT(et);
6879 	return (0);
6880 }
6881 #endif /* DEBUGNET */
6882