xref: /freebsd/sys/net/iflib.c (revision 54e9e4e72d711fb41f88f793f6c64df1126112f9)
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/pfil.h>
63 #include <net/vnet.h>
64 
65 #include <netinet/in.h>
66 #include <netinet/in_pcb.h>
67 #include <netinet/tcp_lro.h>
68 #include <netinet/in_systm.h>
69 #include <netinet/if_ether.h>
70 #include <netinet/ip.h>
71 #include <netinet/ip6.h>
72 #include <netinet/tcp.h>
73 #include <netinet/ip_var.h>
74 #include <netinet/netdump/netdump.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 struct iflib_txq;
133 typedef struct iflib_txq *iflib_txq_t;
134 struct iflib_rxq;
135 typedef struct iflib_rxq *iflib_rxq_t;
136 struct iflib_fl;
137 typedef struct iflib_fl *iflib_fl_t;
138 
139 struct iflib_ctx;
140 
141 static void iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid);
142 static void iflib_timer(void *arg);
143 
144 typedef struct iflib_filter_info {
145 	driver_filter_t *ifi_filter;
146 	void *ifi_filter_arg;
147 	struct grouptask *ifi_task;
148 	void *ifi_ctx;
149 } *iflib_filter_info_t;
150 
151 struct iflib_ctx {
152 	KOBJ_FIELDS;
153 	/*
154 	 * Pointer to hardware driver's softc
155 	 */
156 	void *ifc_softc;
157 	device_t ifc_dev;
158 	if_t ifc_ifp;
159 
160 	cpuset_t ifc_cpus;
161 	if_shared_ctx_t ifc_sctx;
162 	struct if_softc_ctx ifc_softc_ctx;
163 
164 	struct sx ifc_ctx_sx;
165 	struct mtx ifc_state_mtx;
166 
167 	iflib_txq_t ifc_txqs;
168 	iflib_rxq_t ifc_rxqs;
169 	uint32_t ifc_if_flags;
170 	uint32_t ifc_flags;
171 	uint32_t ifc_max_fl_buf_size;
172 	uint32_t ifc_rx_mbuf_sz;
173 
174 	int ifc_link_state;
175 	int ifc_watchdog_events;
176 	struct cdev *ifc_led_dev;
177 	struct resource *ifc_msix_mem;
178 
179 	struct if_irq ifc_legacy_irq;
180 	struct grouptask ifc_admin_task;
181 	struct grouptask ifc_vflr_task;
182 	struct iflib_filter_info ifc_filter_info;
183 	struct ifmedia	ifc_media;
184 	struct ifmedia	*ifc_mediap;
185 
186 	struct sysctl_oid *ifc_sysctl_node;
187 	uint16_t ifc_sysctl_ntxqs;
188 	uint16_t ifc_sysctl_nrxqs;
189 	uint16_t ifc_sysctl_qs_eq_override;
190 	uint16_t ifc_sysctl_rx_budget;
191 	uint16_t ifc_sysctl_tx_abdicate;
192 	uint16_t ifc_sysctl_core_offset;
193 #define	CORE_OFFSET_UNSPECIFIED	0xffff
194 	uint8_t  ifc_sysctl_separate_txrx;
195 
196 	qidx_t ifc_sysctl_ntxds[8];
197 	qidx_t ifc_sysctl_nrxds[8];
198 	struct if_txrx ifc_txrx;
199 #define isc_txd_encap  ifc_txrx.ift_txd_encap
200 #define isc_txd_flush  ifc_txrx.ift_txd_flush
201 #define isc_txd_credits_update  ifc_txrx.ift_txd_credits_update
202 #define isc_rxd_available ifc_txrx.ift_rxd_available
203 #define isc_rxd_pkt_get ifc_txrx.ift_rxd_pkt_get
204 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
205 #define isc_rxd_flush ifc_txrx.ift_rxd_flush
206 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
207 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
208 #define isc_legacy_intr ifc_txrx.ift_legacy_intr
209 	eventhandler_tag ifc_vlan_attach_event;
210 	eventhandler_tag ifc_vlan_detach_event;
211 	struct ether_addr ifc_mac;
212 };
213 
214 void *
215 iflib_get_softc(if_ctx_t ctx)
216 {
217 
218 	return (ctx->ifc_softc);
219 }
220 
221 device_t
222 iflib_get_dev(if_ctx_t ctx)
223 {
224 
225 	return (ctx->ifc_dev);
226 }
227 
228 if_t
229 iflib_get_ifp(if_ctx_t ctx)
230 {
231 
232 	return (ctx->ifc_ifp);
233 }
234 
235 struct ifmedia *
236 iflib_get_media(if_ctx_t ctx)
237 {
238 
239 	return (ctx->ifc_mediap);
240 }
241 
242 uint32_t
243 iflib_get_flags(if_ctx_t ctx)
244 {
245 	return (ctx->ifc_flags);
246 }
247 
248 void
249 iflib_set_mac(if_ctx_t ctx, uint8_t mac[ETHER_ADDR_LEN])
250 {
251 
252 	bcopy(mac, ctx->ifc_mac.octet, ETHER_ADDR_LEN);
253 }
254 
255 if_softc_ctx_t
256 iflib_get_softc_ctx(if_ctx_t ctx)
257 {
258 
259 	return (&ctx->ifc_softc_ctx);
260 }
261 
262 if_shared_ctx_t
263 iflib_get_sctx(if_ctx_t ctx)
264 {
265 
266 	return (ctx->ifc_sctx);
267 }
268 
269 #define IP_ALIGNED(m) ((((uintptr_t)(m)->m_data) & 0x3) == 0x2)
270 #define CACHE_PTR_INCREMENT (CACHE_LINE_SIZE/sizeof(void*))
271 #define CACHE_PTR_NEXT(ptr) ((void *)(((uintptr_t)(ptr)+CACHE_LINE_SIZE-1) & (CACHE_LINE_SIZE-1)))
272 
273 #define LINK_ACTIVE(ctx) ((ctx)->ifc_link_state == LINK_STATE_UP)
274 #define CTX_IS_VF(ctx) ((ctx)->ifc_sctx->isc_flags & IFLIB_IS_VF)
275 
276 typedef struct iflib_sw_rx_desc_array {
277 	bus_dmamap_t	*ifsd_map;         /* bus_dma maps for packet */
278 	struct mbuf	**ifsd_m;           /* pkthdr mbufs */
279 	caddr_t		*ifsd_cl;          /* direct cluster pointer for rx */
280 	bus_addr_t	*ifsd_ba;          /* bus addr of cluster for rx */
281 } iflib_rxsd_array_t;
282 
283 typedef struct iflib_sw_tx_desc_array {
284 	bus_dmamap_t    *ifsd_map;         /* bus_dma maps for packet */
285 	bus_dmamap_t	*ifsd_tso_map;     /* bus_dma maps for TSO packet */
286 	struct mbuf    **ifsd_m;           /* pkthdr mbufs */
287 } if_txsd_vec_t;
288 
289 /* magic number that should be high enough for any hardware */
290 #define IFLIB_MAX_TX_SEGS		128
291 #define IFLIB_RX_COPY_THRESH		128
292 #define IFLIB_MAX_RX_REFRESH		32
293 /* The minimum descriptors per second before we start coalescing */
294 #define IFLIB_MIN_DESC_SEC		16384
295 #define IFLIB_DEFAULT_TX_UPDATE_FREQ	16
296 #define IFLIB_QUEUE_IDLE		0
297 #define IFLIB_QUEUE_HUNG		1
298 #define IFLIB_QUEUE_WORKING		2
299 /* maximum number of txqs that can share an rx interrupt */
300 #define IFLIB_MAX_TX_SHARED_INTR	4
301 
302 /* this should really scale with ring size - this is a fairly arbitrary value */
303 #define TX_BATCH_SIZE			32
304 
305 #define IFLIB_RESTART_BUDGET		8
306 
307 #define CSUM_OFFLOAD		(CSUM_IP_TSO|CSUM_IP6_TSO|CSUM_IP| \
308 				 CSUM_IP_UDP|CSUM_IP_TCP|CSUM_IP_SCTP| \
309 				 CSUM_IP6_UDP|CSUM_IP6_TCP|CSUM_IP6_SCTP)
310 
311 struct iflib_txq {
312 	qidx_t		ift_in_use;
313 	qidx_t		ift_cidx;
314 	qidx_t		ift_cidx_processed;
315 	qidx_t		ift_pidx;
316 	uint8_t		ift_gen;
317 	uint8_t		ift_br_offset;
318 	uint16_t	ift_npending;
319 	uint16_t	ift_db_pending;
320 	uint16_t	ift_rs_pending;
321 	/* implicit pad */
322 	uint8_t		ift_txd_size[8];
323 	uint64_t	ift_processed;
324 	uint64_t	ift_cleaned;
325 	uint64_t	ift_cleaned_prev;
326 #if MEMORY_LOGGING
327 	uint64_t	ift_enqueued;
328 	uint64_t	ift_dequeued;
329 #endif
330 	uint64_t	ift_no_tx_dma_setup;
331 	uint64_t	ift_no_desc_avail;
332 	uint64_t	ift_mbuf_defrag_failed;
333 	uint64_t	ift_mbuf_defrag;
334 	uint64_t	ift_map_failed;
335 	uint64_t	ift_txd_encap_efbig;
336 	uint64_t	ift_pullups;
337 	uint64_t	ift_last_timer_tick;
338 
339 	struct mtx	ift_mtx;
340 	struct mtx	ift_db_mtx;
341 
342 	/* constant values */
343 	if_ctx_t	ift_ctx;
344 	struct ifmp_ring        *ift_br;
345 	struct grouptask	ift_task;
346 	qidx_t		ift_size;
347 	uint16_t	ift_id;
348 	struct callout	ift_timer;
349 
350 	if_txsd_vec_t	ift_sds;
351 	uint8_t		ift_qstatus;
352 	uint8_t		ift_closed;
353 	uint8_t		ift_update_freq;
354 	struct iflib_filter_info ift_filter_info;
355 	bus_dma_tag_t	ift_buf_tag;
356 	bus_dma_tag_t	ift_tso_buf_tag;
357 	iflib_dma_info_t	ift_ifdi;
358 #define MTX_NAME_LEN 16
359 	char                    ift_mtx_name[MTX_NAME_LEN];
360 	bus_dma_segment_t	ift_segs[IFLIB_MAX_TX_SEGS]  __aligned(CACHE_LINE_SIZE);
361 #ifdef IFLIB_DIAGNOSTICS
362 	uint64_t ift_cpu_exec_count[256];
363 #endif
364 } __aligned(CACHE_LINE_SIZE);
365 
366 struct iflib_fl {
367 	qidx_t		ifl_cidx;
368 	qidx_t		ifl_pidx;
369 	qidx_t		ifl_credits;
370 	uint8_t		ifl_gen;
371 	uint8_t		ifl_rxd_size;
372 #if MEMORY_LOGGING
373 	uint64_t	ifl_m_enqueued;
374 	uint64_t	ifl_m_dequeued;
375 	uint64_t	ifl_cl_enqueued;
376 	uint64_t	ifl_cl_dequeued;
377 #endif
378 	/* implicit pad */
379 	bitstr_t 	*ifl_rx_bitmap;
380 	qidx_t		ifl_fragidx;
381 	/* constant */
382 	qidx_t		ifl_size;
383 	uint16_t	ifl_buf_size;
384 	uint16_t	ifl_cltype;
385 	uma_zone_t	ifl_zone;
386 	iflib_rxsd_array_t	ifl_sds;
387 	iflib_rxq_t	ifl_rxq;
388 	uint8_t		ifl_id;
389 	bus_dma_tag_t	ifl_buf_tag;
390 	iflib_dma_info_t	ifl_ifdi;
391 	uint64_t	ifl_bus_addrs[IFLIB_MAX_RX_REFRESH] __aligned(CACHE_LINE_SIZE);
392 	caddr_t		ifl_vm_addrs[IFLIB_MAX_RX_REFRESH];
393 	qidx_t	ifl_rxd_idxs[IFLIB_MAX_RX_REFRESH];
394 }  __aligned(CACHE_LINE_SIZE);
395 
396 static inline qidx_t
397 get_inuse(int size, qidx_t cidx, qidx_t pidx, uint8_t gen)
398 {
399 	qidx_t used;
400 
401 	if (pidx > cidx)
402 		used = pidx - cidx;
403 	else if (pidx < cidx)
404 		used = size - cidx + pidx;
405 	else if (gen == 0 && pidx == cidx)
406 		used = 0;
407 	else if (gen == 1 && pidx == cidx)
408 		used = size;
409 	else
410 		panic("bad state");
411 
412 	return (used);
413 }
414 
415 #define TXQ_AVAIL(txq) (txq->ift_size - get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx, txq->ift_gen))
416 
417 #define IDXDIFF(head, tail, wrap) \
418 	((head) >= (tail) ? (head) - (tail) : (wrap) - (tail) + (head))
419 
420 struct iflib_rxq {
421 	if_ctx_t	ifr_ctx;
422 	iflib_fl_t	ifr_fl;
423 	uint64_t	ifr_rx_irq;
424 	struct pfil_head	*pfil;
425 	/*
426 	 * If there is a separate completion queue (IFLIB_HAS_RXCQ), this is
427 	 * the command queue consumer index.  Otherwise it's unused.
428 	 */
429 	qidx_t		ifr_cq_cidx;
430 	uint16_t	ifr_id;
431 	uint8_t		ifr_nfl;
432 	uint8_t		ifr_ntxqirq;
433 	uint8_t		ifr_txqid[IFLIB_MAX_TX_SHARED_INTR];
434 	uint8_t		ifr_fl_offset;
435 	struct lro_ctrl			ifr_lc;
436 	struct grouptask        ifr_task;
437 	struct iflib_filter_info ifr_filter_info;
438 	iflib_dma_info_t		ifr_ifdi;
439 
440 	/* dynamically allocate if any drivers need a value substantially larger than this */
441 	struct if_rxd_frag	ifr_frags[IFLIB_MAX_RX_SEGS] __aligned(CACHE_LINE_SIZE);
442 #ifdef IFLIB_DIAGNOSTICS
443 	uint64_t ifr_cpu_exec_count[256];
444 #endif
445 }  __aligned(CACHE_LINE_SIZE);
446 
447 typedef struct if_rxsd {
448 	caddr_t *ifsd_cl;
449 	iflib_fl_t ifsd_fl;
450 	qidx_t ifsd_cidx;
451 } *if_rxsd_t;
452 
453 /* multiple of word size */
454 #ifdef __LP64__
455 #define PKT_INFO_SIZE	6
456 #define RXD_INFO_SIZE	5
457 #define PKT_TYPE uint64_t
458 #else
459 #define PKT_INFO_SIZE	11
460 #define RXD_INFO_SIZE	8
461 #define PKT_TYPE uint32_t
462 #endif
463 #define PKT_LOOP_BOUND  ((PKT_INFO_SIZE/3)*3)
464 #define RXD_LOOP_BOUND  ((RXD_INFO_SIZE/4)*4)
465 
466 typedef struct if_pkt_info_pad {
467 	PKT_TYPE pkt_val[PKT_INFO_SIZE];
468 } *if_pkt_info_pad_t;
469 typedef struct if_rxd_info_pad {
470 	PKT_TYPE rxd_val[RXD_INFO_SIZE];
471 } *if_rxd_info_pad_t;
472 
473 CTASSERT(sizeof(struct if_pkt_info_pad) == sizeof(struct if_pkt_info));
474 CTASSERT(sizeof(struct if_rxd_info_pad) == sizeof(struct if_rxd_info));
475 
476 
477 static inline void
478 pkt_info_zero(if_pkt_info_t pi)
479 {
480 	if_pkt_info_pad_t pi_pad;
481 
482 	pi_pad = (if_pkt_info_pad_t)pi;
483 	pi_pad->pkt_val[0] = 0; pi_pad->pkt_val[1] = 0; pi_pad->pkt_val[2] = 0;
484 	pi_pad->pkt_val[3] = 0; pi_pad->pkt_val[4] = 0; pi_pad->pkt_val[5] = 0;
485 #ifndef __LP64__
486 	pi_pad->pkt_val[6] = 0; pi_pad->pkt_val[7] = 0; pi_pad->pkt_val[8] = 0;
487 	pi_pad->pkt_val[9] = 0; pi_pad->pkt_val[10] = 0;
488 #endif
489 }
490 
491 static device_method_t iflib_pseudo_methods[] = {
492 	DEVMETHOD(device_attach, noop_attach),
493 	DEVMETHOD(device_detach, iflib_pseudo_detach),
494 	DEVMETHOD_END
495 };
496 
497 driver_t iflib_pseudodriver = {
498 	"iflib_pseudo", iflib_pseudo_methods, sizeof(struct iflib_ctx),
499 };
500 
501 static inline void
502 rxd_info_zero(if_rxd_info_t ri)
503 {
504 	if_rxd_info_pad_t ri_pad;
505 	int i;
506 
507 	ri_pad = (if_rxd_info_pad_t)ri;
508 	for (i = 0; i < RXD_LOOP_BOUND; i += 4) {
509 		ri_pad->rxd_val[i] = 0;
510 		ri_pad->rxd_val[i+1] = 0;
511 		ri_pad->rxd_val[i+2] = 0;
512 		ri_pad->rxd_val[i+3] = 0;
513 	}
514 #ifdef __LP64__
515 	ri_pad->rxd_val[RXD_INFO_SIZE-1] = 0;
516 #endif
517 }
518 
519 /*
520  * Only allow a single packet to take up most 1/nth of the tx ring
521  */
522 #define MAX_SINGLE_PACKET_FRACTION 12
523 #define IF_BAD_DMA (bus_addr_t)-1
524 
525 #define CTX_ACTIVE(ctx) ((if_getdrvflags((ctx)->ifc_ifp) & IFF_DRV_RUNNING))
526 
527 #define CTX_LOCK_INIT(_sc)  sx_init(&(_sc)->ifc_ctx_sx, "iflib ctx lock")
528 #define CTX_LOCK(ctx) sx_xlock(&(ctx)->ifc_ctx_sx)
529 #define CTX_UNLOCK(ctx) sx_xunlock(&(ctx)->ifc_ctx_sx)
530 #define CTX_LOCK_DESTROY(ctx) sx_destroy(&(ctx)->ifc_ctx_sx)
531 
532 #define STATE_LOCK_INIT(_sc, _name)  mtx_init(&(_sc)->ifc_state_mtx, _name, "iflib state lock", MTX_DEF)
533 #define STATE_LOCK(ctx) mtx_lock(&(ctx)->ifc_state_mtx)
534 #define STATE_UNLOCK(ctx) mtx_unlock(&(ctx)->ifc_state_mtx)
535 #define STATE_LOCK_DESTROY(ctx) mtx_destroy(&(ctx)->ifc_state_mtx)
536 
537 #define CALLOUT_LOCK(txq)	mtx_lock(&txq->ift_mtx)
538 #define CALLOUT_UNLOCK(txq) 	mtx_unlock(&txq->ift_mtx)
539 
540 void
541 iflib_set_detach(if_ctx_t ctx)
542 {
543 	STATE_LOCK(ctx);
544 	ctx->ifc_flags |= IFC_IN_DETACH;
545 	STATE_UNLOCK(ctx);
546 }
547 
548 /* Our boot-time initialization hook */
549 static int	iflib_module_event_handler(module_t, int, void *);
550 
551 static moduledata_t iflib_moduledata = {
552 	"iflib",
553 	iflib_module_event_handler,
554 	NULL
555 };
556 
557 DECLARE_MODULE(iflib, iflib_moduledata, SI_SUB_INIT_IF, SI_ORDER_ANY);
558 MODULE_VERSION(iflib, 1);
559 
560 MODULE_DEPEND(iflib, pci, 1, 1, 1);
561 MODULE_DEPEND(iflib, ether, 1, 1, 1);
562 
563 TASKQGROUP_DEFINE(if_io_tqg, mp_ncpus, 1);
564 TASKQGROUP_DEFINE(if_config_tqg, 1, 1);
565 
566 #ifndef IFLIB_DEBUG_COUNTERS
567 #ifdef INVARIANTS
568 #define IFLIB_DEBUG_COUNTERS 1
569 #else
570 #define IFLIB_DEBUG_COUNTERS 0
571 #endif /* !INVARIANTS */
572 #endif
573 
574 static SYSCTL_NODE(_net, OID_AUTO, iflib, CTLFLAG_RD, 0,
575                    "iflib driver parameters");
576 
577 /*
578  * XXX need to ensure that this can't accidentally cause the head to be moved backwards
579  */
580 static int iflib_min_tx_latency = 0;
581 SYSCTL_INT(_net_iflib, OID_AUTO, min_tx_latency, CTLFLAG_RW,
582 		   &iflib_min_tx_latency, 0, "minimize transmit latency at the possible expense of throughput");
583 static int iflib_no_tx_batch = 0;
584 SYSCTL_INT(_net_iflib, OID_AUTO, no_tx_batch, CTLFLAG_RW,
585 		   &iflib_no_tx_batch, 0, "minimize transmit latency at the possible expense of throughput");
586 
587 
588 #if IFLIB_DEBUG_COUNTERS
589 
590 static int iflib_tx_seen;
591 static int iflib_tx_sent;
592 static int iflib_tx_encap;
593 static int iflib_rx_allocs;
594 static int iflib_fl_refills;
595 static int iflib_fl_refills_large;
596 static int iflib_tx_frees;
597 
598 SYSCTL_INT(_net_iflib, OID_AUTO, tx_seen, CTLFLAG_RD,
599 		   &iflib_tx_seen, 0, "# TX mbufs seen");
600 SYSCTL_INT(_net_iflib, OID_AUTO, tx_sent, CTLFLAG_RD,
601 		   &iflib_tx_sent, 0, "# TX mbufs sent");
602 SYSCTL_INT(_net_iflib, OID_AUTO, tx_encap, CTLFLAG_RD,
603 		   &iflib_tx_encap, 0, "# TX mbufs encapped");
604 SYSCTL_INT(_net_iflib, OID_AUTO, tx_frees, CTLFLAG_RD,
605 		   &iflib_tx_frees, 0, "# TX frees");
606 SYSCTL_INT(_net_iflib, OID_AUTO, rx_allocs, CTLFLAG_RD,
607 		   &iflib_rx_allocs, 0, "# RX allocations");
608 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills, CTLFLAG_RD,
609 		   &iflib_fl_refills, 0, "# refills");
610 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills_large, CTLFLAG_RD,
611 		   &iflib_fl_refills_large, 0, "# large refills");
612 
613 
614 static int iflib_txq_drain_flushing;
615 static int iflib_txq_drain_oactive;
616 static int iflib_txq_drain_notready;
617 
618 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_flushing, CTLFLAG_RD,
619 		   &iflib_txq_drain_flushing, 0, "# drain flushes");
620 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_oactive, CTLFLAG_RD,
621 		   &iflib_txq_drain_oactive, 0, "# drain oactives");
622 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_notready, CTLFLAG_RD,
623 		   &iflib_txq_drain_notready, 0, "# drain notready");
624 
625 
626 static int iflib_encap_load_mbuf_fail;
627 static int iflib_encap_pad_mbuf_fail;
628 static int iflib_encap_txq_avail_fail;
629 static int iflib_encap_txd_encap_fail;
630 
631 SYSCTL_INT(_net_iflib, OID_AUTO, encap_load_mbuf_fail, CTLFLAG_RD,
632 		   &iflib_encap_load_mbuf_fail, 0, "# busdma load failures");
633 SYSCTL_INT(_net_iflib, OID_AUTO, encap_pad_mbuf_fail, CTLFLAG_RD,
634 		   &iflib_encap_pad_mbuf_fail, 0, "# runt frame pad failures");
635 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txq_avail_fail, CTLFLAG_RD,
636 		   &iflib_encap_txq_avail_fail, 0, "# txq avail failures");
637 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txd_encap_fail, CTLFLAG_RD,
638 		   &iflib_encap_txd_encap_fail, 0, "# driver encap failures");
639 
640 static int iflib_task_fn_rxs;
641 static int iflib_rx_intr_enables;
642 static int iflib_fast_intrs;
643 static int iflib_rx_unavail;
644 static int iflib_rx_ctx_inactive;
645 static int iflib_rx_if_input;
646 static int iflib_rxd_flush;
647 
648 static int iflib_verbose_debug;
649 
650 SYSCTL_INT(_net_iflib, OID_AUTO, task_fn_rx, CTLFLAG_RD,
651 		   &iflib_task_fn_rxs, 0, "# task_fn_rx calls");
652 SYSCTL_INT(_net_iflib, OID_AUTO, rx_intr_enables, CTLFLAG_RD,
653 		   &iflib_rx_intr_enables, 0, "# RX intr enables");
654 SYSCTL_INT(_net_iflib, OID_AUTO, fast_intrs, CTLFLAG_RD,
655 		   &iflib_fast_intrs, 0, "# fast_intr calls");
656 SYSCTL_INT(_net_iflib, OID_AUTO, rx_unavail, CTLFLAG_RD,
657 		   &iflib_rx_unavail, 0, "# times rxeof called with no available data");
658 SYSCTL_INT(_net_iflib, OID_AUTO, rx_ctx_inactive, CTLFLAG_RD,
659 		   &iflib_rx_ctx_inactive, 0, "# times rxeof called with inactive context");
660 SYSCTL_INT(_net_iflib, OID_AUTO, rx_if_input, CTLFLAG_RD,
661 		   &iflib_rx_if_input, 0, "# times rxeof called if_input");
662 SYSCTL_INT(_net_iflib, OID_AUTO, rxd_flush, CTLFLAG_RD,
663 	         &iflib_rxd_flush, 0, "# times rxd_flush called");
664 SYSCTL_INT(_net_iflib, OID_AUTO, verbose_debug, CTLFLAG_RW,
665 		   &iflib_verbose_debug, 0, "enable verbose debugging");
666 
667 #define DBG_COUNTER_INC(name) atomic_add_int(&(iflib_ ## name), 1)
668 static void
669 iflib_debug_reset(void)
670 {
671 	iflib_tx_seen = iflib_tx_sent = iflib_tx_encap = iflib_rx_allocs =
672 		iflib_fl_refills = iflib_fl_refills_large = iflib_tx_frees =
673 		iflib_txq_drain_flushing = iflib_txq_drain_oactive =
674 		iflib_txq_drain_notready =
675 		iflib_encap_load_mbuf_fail = iflib_encap_pad_mbuf_fail =
676 		iflib_encap_txq_avail_fail = iflib_encap_txd_encap_fail =
677 		iflib_task_fn_rxs = iflib_rx_intr_enables = iflib_fast_intrs =
678 		iflib_rx_unavail =
679 		iflib_rx_ctx_inactive = iflib_rx_if_input =
680 		iflib_rxd_flush = 0;
681 }
682 
683 #else
684 #define DBG_COUNTER_INC(name)
685 static void iflib_debug_reset(void) {}
686 #endif
687 
688 #define IFLIB_DEBUG 0
689 
690 static void iflib_tx_structures_free(if_ctx_t ctx);
691 static void iflib_rx_structures_free(if_ctx_t ctx);
692 static int iflib_queues_alloc(if_ctx_t ctx);
693 static int iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq);
694 static int iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget);
695 static int iflib_qset_structures_setup(if_ctx_t ctx);
696 static int iflib_msix_init(if_ctx_t ctx);
697 static int iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filterarg, int *rid, const char *str);
698 static void iflib_txq_check_drain(iflib_txq_t txq, int budget);
699 static uint32_t iflib_txq_can_drain(struct ifmp_ring *);
700 #ifdef ALTQ
701 static void iflib_altq_if_start(if_t ifp);
702 static int iflib_altq_if_transmit(if_t ifp, struct mbuf *m);
703 #endif
704 static int iflib_register(if_ctx_t);
705 static void iflib_init_locked(if_ctx_t ctx);
706 static void iflib_add_device_sysctl_pre(if_ctx_t ctx);
707 static void iflib_add_device_sysctl_post(if_ctx_t ctx);
708 static void iflib_ifmp_purge(iflib_txq_t txq);
709 static void _iflib_pre_assert(if_softc_ctx_t scctx);
710 static void iflib_if_init_locked(if_ctx_t ctx);
711 static void iflib_free_intr_mem(if_ctx_t ctx);
712 #ifndef __NO_STRICT_ALIGNMENT
713 static struct mbuf * iflib_fixup_rx(struct mbuf *m);
714 #endif
715 
716 static SLIST_HEAD(cpu_offset_list, cpu_offset) cpu_offsets =
717     SLIST_HEAD_INITIALIZER(cpu_offsets);
718 struct cpu_offset {
719 	SLIST_ENTRY(cpu_offset) entries;
720 	cpuset_t	set;
721 	unsigned int	refcount;
722 	uint16_t	offset;
723 };
724 static struct mtx cpu_offset_mtx;
725 MTX_SYSINIT(iflib_cpu_offset, &cpu_offset_mtx, "iflib_cpu_offset lock",
726     MTX_DEF);
727 
728 NETDUMP_DEFINE(iflib);
729 
730 #ifdef DEV_NETMAP
731 #include <sys/selinfo.h>
732 #include <net/netmap.h>
733 #include <dev/netmap/netmap_kern.h>
734 
735 MODULE_DEPEND(iflib, netmap, 1, 1, 1);
736 
737 static int netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, uint32_t nm_i, bool init);
738 
739 /*
740  * device-specific sysctl variables:
741  *
742  * iflib_crcstrip: 0: keep CRC in rx frames (default), 1: strip it.
743  *	During regular operations the CRC is stripped, but on some
744  *	hardware reception of frames not multiple of 64 is slower,
745  *	so using crcstrip=0 helps in benchmarks.
746  *
747  * iflib_rx_miss, iflib_rx_miss_bufs:
748  *	count packets that might be missed due to lost interrupts.
749  */
750 SYSCTL_DECL(_dev_netmap);
751 /*
752  * The xl driver by default strips CRCs and we do not override it.
753  */
754 
755 int iflib_crcstrip = 1;
756 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_crcstrip,
757     CTLFLAG_RW, &iflib_crcstrip, 1, "strip CRC on RX frames");
758 
759 int iflib_rx_miss, iflib_rx_miss_bufs;
760 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss,
761     CTLFLAG_RW, &iflib_rx_miss, 0, "potentially missed RX intr");
762 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss_bufs,
763     CTLFLAG_RW, &iflib_rx_miss_bufs, 0, "potentially missed RX intr bufs");
764 
765 /*
766  * Register/unregister. We are already under netmap lock.
767  * Only called on the first register or the last unregister.
768  */
769 static int
770 iflib_netmap_register(struct netmap_adapter *na, int onoff)
771 {
772 	if_t ifp = na->ifp;
773 	if_ctx_t ctx = ifp->if_softc;
774 	int status;
775 
776 	CTX_LOCK(ctx);
777 	IFDI_INTR_DISABLE(ctx);
778 
779 	/* Tell the stack that the interface is no longer active */
780 	ifp->if_drv_flags &= ~(IFF_DRV_RUNNING | IFF_DRV_OACTIVE);
781 
782 	if (!CTX_IS_VF(ctx))
783 		IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip);
784 
785 	/* enable or disable flags and callbacks in na and ifp */
786 	if (onoff) {
787 		nm_set_native_flags(na);
788 	} else {
789 		nm_clear_native_flags(na);
790 	}
791 	iflib_stop(ctx);
792 	iflib_init_locked(ctx);
793 	IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip); // XXX why twice ?
794 	status = ifp->if_drv_flags & IFF_DRV_RUNNING ? 0 : 1;
795 	if (status)
796 		nm_clear_native_flags(na);
797 	CTX_UNLOCK(ctx);
798 	return (status);
799 }
800 
801 static int
802 netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, uint32_t nm_i, bool init)
803 {
804 	struct netmap_adapter *na = kring->na;
805 	u_int const lim = kring->nkr_num_slots - 1;
806 	u_int head = kring->rhead;
807 	struct netmap_ring *ring = kring->ring;
808 	bus_dmamap_t *map;
809 	struct if_rxd_update iru;
810 	if_ctx_t ctx = rxq->ifr_ctx;
811 	iflib_fl_t fl = &rxq->ifr_fl[0];
812 	uint32_t refill_pidx, nic_i;
813 #if IFLIB_DEBUG_COUNTERS
814 	int rf_count = 0;
815 #endif
816 
817 	if (nm_i == head && __predict_true(!init))
818 		return 0;
819 	iru_init(&iru, rxq, 0 /* flid */);
820 	map = fl->ifl_sds.ifsd_map;
821 	refill_pidx = netmap_idx_k2n(kring, nm_i);
822 	/*
823 	 * IMPORTANT: we must leave one free slot in the ring,
824 	 * so move head back by one unit
825 	 */
826 	head = nm_prev(head, lim);
827 	nic_i = UINT_MAX;
828 	DBG_COUNTER_INC(fl_refills);
829 	while (nm_i != head) {
830 #if IFLIB_DEBUG_COUNTERS
831 		if (++rf_count == 9)
832 			DBG_COUNTER_INC(fl_refills_large);
833 #endif
834 		for (int tmp_pidx = 0; tmp_pidx < IFLIB_MAX_RX_REFRESH && nm_i != head; tmp_pidx++) {
835 			struct netmap_slot *slot = &ring->slot[nm_i];
836 			void *addr = PNMB(na, slot, &fl->ifl_bus_addrs[tmp_pidx]);
837 			uint32_t nic_i_dma = refill_pidx;
838 			nic_i = netmap_idx_k2n(kring, nm_i);
839 
840 			MPASS(tmp_pidx < IFLIB_MAX_RX_REFRESH);
841 
842 			if (addr == NETMAP_BUF_BASE(na)) /* bad buf */
843 			        return netmap_ring_reinit(kring);
844 
845 			fl->ifl_vm_addrs[tmp_pidx] = addr;
846 			if (__predict_false(init)) {
847 				netmap_load_map(na, fl->ifl_buf_tag,
848 				    map[nic_i], addr);
849 			} else if (slot->flags & NS_BUF_CHANGED) {
850 				/* buffer has changed, reload map */
851 				netmap_reload_map(na, fl->ifl_buf_tag,
852 				    map[nic_i], addr);
853 			}
854 			slot->flags &= ~NS_BUF_CHANGED;
855 
856 			nm_i = nm_next(nm_i, lim);
857 			fl->ifl_rxd_idxs[tmp_pidx] = nic_i = nm_next(nic_i, lim);
858 			if (nm_i != head && tmp_pidx < IFLIB_MAX_RX_REFRESH-1)
859 				continue;
860 
861 			iru.iru_pidx = refill_pidx;
862 			iru.iru_count = tmp_pidx+1;
863 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
864 			refill_pidx = nic_i;
865 			for (int n = 0; n < iru.iru_count; n++) {
866 				bus_dmamap_sync(fl->ifl_buf_tag, map[nic_i_dma],
867 						BUS_DMASYNC_PREREAD);
868 				/* XXX - change this to not use the netmap func*/
869 				nic_i_dma = nm_next(nic_i_dma, lim);
870 			}
871 		}
872 	}
873 	kring->nr_hwcur = head;
874 
875 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
876 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
877 	if (__predict_true(nic_i != UINT_MAX)) {
878 		ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id, nic_i);
879 		DBG_COUNTER_INC(rxd_flush);
880 	}
881 	return (0);
882 }
883 
884 /*
885  * Reconcile kernel and user view of the transmit ring.
886  *
887  * All information is in the kring.
888  * Userspace wants to send packets up to the one before kring->rhead,
889  * kernel knows kring->nr_hwcur is the first unsent packet.
890  *
891  * Here we push packets out (as many as possible), and possibly
892  * reclaim buffers from previously completed transmission.
893  *
894  * The caller (netmap) guarantees that there is only one instance
895  * running at any time. Any interference with other driver
896  * methods should be handled by the individual drivers.
897  */
898 static int
899 iflib_netmap_txsync(struct netmap_kring *kring, int flags)
900 {
901 	struct netmap_adapter *na = kring->na;
902 	if_t ifp = na->ifp;
903 	struct netmap_ring *ring = kring->ring;
904 	u_int nm_i;	/* index into the netmap kring */
905 	u_int nic_i;	/* index into the NIC ring */
906 	u_int n;
907 	u_int const lim = kring->nkr_num_slots - 1;
908 	u_int const head = kring->rhead;
909 	struct if_pkt_info pi;
910 
911 	/*
912 	 * interrupts on every tx packet are expensive so request
913 	 * them every half ring, or where NS_REPORT is set
914 	 */
915 	u_int report_frequency = kring->nkr_num_slots >> 1;
916 	/* device-specific */
917 	if_ctx_t ctx = ifp->if_softc;
918 	iflib_txq_t txq = &ctx->ifc_txqs[kring->ring_id];
919 
920 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
921 	    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
922 
923 	/*
924 	 * First part: process new packets to send.
925 	 * nm_i is the current index in the netmap kring,
926 	 * nic_i is the corresponding index in the NIC ring.
927 	 *
928 	 * If we have packets to send (nm_i != head)
929 	 * iterate over the netmap ring, fetch length and update
930 	 * the corresponding slot in the NIC ring. Some drivers also
931 	 * need to update the buffer's physical address in the NIC slot
932 	 * even NS_BUF_CHANGED is not set (PNMB computes the addresses).
933 	 *
934 	 * The netmap_reload_map() calls is especially expensive,
935 	 * even when (as in this case) the tag is 0, so do only
936 	 * when the buffer has actually changed.
937 	 *
938 	 * If possible do not set the report/intr bit on all slots,
939 	 * but only a few times per ring or when NS_REPORT is set.
940 	 *
941 	 * Finally, on 10G and faster drivers, it might be useful
942 	 * to prefetch the next slot and txr entry.
943 	 */
944 
945 	nm_i = kring->nr_hwcur;
946 	if (nm_i != head) {	/* we have new packets to send */
947 		pkt_info_zero(&pi);
948 		pi.ipi_segs = txq->ift_segs;
949 		pi.ipi_qsidx = kring->ring_id;
950 		nic_i = netmap_idx_k2n(kring, nm_i);
951 
952 		__builtin_prefetch(&ring->slot[nm_i]);
953 		__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i]);
954 		__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i]);
955 
956 		for (n = 0; nm_i != head; n++) {
957 			struct netmap_slot *slot = &ring->slot[nm_i];
958 			u_int len = slot->len;
959 			uint64_t paddr;
960 			void *addr = PNMB(na, slot, &paddr);
961 			int flags = (slot->flags & NS_REPORT ||
962 				nic_i == 0 || nic_i == report_frequency) ?
963 				IPI_TX_INTR : 0;
964 
965 			/* device-specific */
966 			pi.ipi_len = len;
967 			pi.ipi_segs[0].ds_addr = paddr;
968 			pi.ipi_segs[0].ds_len = len;
969 			pi.ipi_nsegs = 1;
970 			pi.ipi_ndescs = 0;
971 			pi.ipi_pidx = nic_i;
972 			pi.ipi_flags = flags;
973 
974 			/* Fill the slot in the NIC ring. */
975 			ctx->isc_txd_encap(ctx->ifc_softc, &pi);
976 			DBG_COUNTER_INC(tx_encap);
977 
978 			/* prefetch for next round */
979 			__builtin_prefetch(&ring->slot[nm_i + 1]);
980 			__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i + 1]);
981 			__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i + 1]);
982 
983 			NM_CHECK_ADDR_LEN(na, addr, len);
984 
985 			if (slot->flags & NS_BUF_CHANGED) {
986 				/* buffer has changed, reload map */
987 				netmap_reload_map(na, txq->ift_buf_tag,
988 				    txq->ift_sds.ifsd_map[nic_i], addr);
989 			}
990 			/* make sure changes to the buffer are synced */
991 			bus_dmamap_sync(txq->ift_buf_tag,
992 			    txq->ift_sds.ifsd_map[nic_i],
993 			    BUS_DMASYNC_PREWRITE);
994 
995 			slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED);
996 			nm_i = nm_next(nm_i, lim);
997 			nic_i = nm_next(nic_i, lim);
998 		}
999 		kring->nr_hwcur = nm_i;
1000 
1001 		/* synchronize the NIC ring */
1002 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1003 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1004 
1005 		/* (re)start the tx unit up to slot nic_i (excluded) */
1006 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, nic_i);
1007 	}
1008 
1009 	/*
1010 	 * Second part: reclaim buffers for completed transmissions.
1011 	 *
1012 	 * If there are unclaimed buffers, attempt to reclaim them.
1013 	 * If none are reclaimed, and TX IRQs are not in use, do an initial
1014 	 * minimal delay, then trigger the tx handler which will spin in the
1015 	 * group task queue.
1016 	 */
1017 	if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1018 		if (iflib_tx_credits_update(ctx, txq)) {
1019 			/* some tx completed, increment avail */
1020 			nic_i = txq->ift_cidx_processed;
1021 			kring->nr_hwtail = nm_prev(netmap_idx_n2k(kring, nic_i), lim);
1022 		}
1023 	}
1024 	if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ))
1025 		if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1026 			callout_reset_on(&txq->ift_timer, hz < 2000 ? 1 : hz / 1000,
1027 			    iflib_timer, txq, txq->ift_timer.c_cpu);
1028 	}
1029 	return (0);
1030 }
1031 
1032 /*
1033  * Reconcile kernel and user view of the receive ring.
1034  * Same as for the txsync, this routine must be efficient.
1035  * The caller guarantees a single invocations, but races against
1036  * the rest of the driver should be handled here.
1037  *
1038  * On call, kring->rhead is the first packet that userspace wants
1039  * to keep, and kring->rcur is the wakeup point.
1040  * The kernel has previously reported packets up to kring->rtail.
1041  *
1042  * If (flags & NAF_FORCE_READ) also check for incoming packets irrespective
1043  * of whether or not we received an interrupt.
1044  */
1045 static int
1046 iflib_netmap_rxsync(struct netmap_kring *kring, int flags)
1047 {
1048 	struct netmap_adapter *na = kring->na;
1049 	struct netmap_ring *ring = kring->ring;
1050 	if_t ifp = na->ifp;
1051 	iflib_fl_t fl;
1052 	uint32_t nm_i;	/* index into the netmap ring */
1053 	uint32_t nic_i;	/* index into the NIC ring */
1054 	u_int i, n;
1055 	u_int const lim = kring->nkr_num_slots - 1;
1056 	u_int const head = kring->rhead;
1057 	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
1058 	struct if_rxd_info ri;
1059 
1060 	if_ctx_t ctx = ifp->if_softc;
1061 	iflib_rxq_t rxq = &ctx->ifc_rxqs[kring->ring_id];
1062 	if (head > lim)
1063 		return netmap_ring_reinit(kring);
1064 
1065 	/*
1066 	 * XXX netmap_fl_refill() only ever (re)fills free list 0 so far.
1067 	 */
1068 
1069 	for (i = 0, fl = rxq->ifr_fl; i < rxq->ifr_nfl; i++, fl++) {
1070 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
1071 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1072 	}
1073 
1074 	/*
1075 	 * First part: import newly received packets.
1076 	 *
1077 	 * nm_i is the index of the next free slot in the netmap ring,
1078 	 * nic_i is the index of the next received packet in the NIC ring,
1079 	 * and they may differ in case if_init() has been called while
1080 	 * in netmap mode. For the receive ring we have
1081 	 *
1082 	 *	nic_i = rxr->next_check;
1083 	 *	nm_i = kring->nr_hwtail (previous)
1084 	 * and
1085 	 *	nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1086 	 *
1087 	 * rxr->next_check is set to 0 on a ring reinit
1088 	 */
1089 	if (netmap_no_pendintr || force_update) {
1090 		int crclen = iflib_crcstrip ? 0 : 4;
1091 		int error, avail;
1092 
1093 		for (i = 0; i < rxq->ifr_nfl; i++) {
1094 			fl = &rxq->ifr_fl[i];
1095 			nic_i = fl->ifl_cidx;
1096 			nm_i = netmap_idx_n2k(kring, nic_i);
1097 			avail = ctx->isc_rxd_available(ctx->ifc_softc,
1098 			    rxq->ifr_id, nic_i, USHRT_MAX);
1099 			for (n = 0; avail > 0; n++, avail--) {
1100 				rxd_info_zero(&ri);
1101 				ri.iri_frags = rxq->ifr_frags;
1102 				ri.iri_qsidx = kring->ring_id;
1103 				ri.iri_ifp = ctx->ifc_ifp;
1104 				ri.iri_cidx = nic_i;
1105 
1106 				error = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
1107 				ring->slot[nm_i].len = error ? 0 : ri.iri_len - crclen;
1108 				ring->slot[nm_i].flags = 0;
1109 				bus_dmamap_sync(fl->ifl_buf_tag,
1110 				    fl->ifl_sds.ifsd_map[nic_i], BUS_DMASYNC_POSTREAD);
1111 				nm_i = nm_next(nm_i, lim);
1112 				nic_i = nm_next(nic_i, lim);
1113 			}
1114 			if (n) { /* update the state variables */
1115 				if (netmap_no_pendintr && !force_update) {
1116 					/* diagnostics */
1117 					iflib_rx_miss ++;
1118 					iflib_rx_miss_bufs += n;
1119 				}
1120 				fl->ifl_cidx = nic_i;
1121 				kring->nr_hwtail = nm_i;
1122 			}
1123 			kring->nr_kflags &= ~NKR_PENDINTR;
1124 		}
1125 	}
1126 	/*
1127 	 * Second part: skip past packets that userspace has released.
1128 	 * (kring->nr_hwcur to head excluded),
1129 	 * and make the buffers available for reception.
1130 	 * As usual nm_i is the index in the netmap ring,
1131 	 * nic_i is the index in the NIC ring, and
1132 	 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1133 	 */
1134 	/* XXX not sure how this will work with multiple free lists */
1135 	nm_i = kring->nr_hwcur;
1136 
1137 	return (netmap_fl_refill(rxq, kring, nm_i, false));
1138 }
1139 
1140 static void
1141 iflib_netmap_intr(struct netmap_adapter *na, int onoff)
1142 {
1143 	if_ctx_t ctx = na->ifp->if_softc;
1144 
1145 	CTX_LOCK(ctx);
1146 	if (onoff) {
1147 		IFDI_INTR_ENABLE(ctx);
1148 	} else {
1149 		IFDI_INTR_DISABLE(ctx);
1150 	}
1151 	CTX_UNLOCK(ctx);
1152 }
1153 
1154 
1155 static int
1156 iflib_netmap_attach(if_ctx_t ctx)
1157 {
1158 	struct netmap_adapter na;
1159 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1160 
1161 	bzero(&na, sizeof(na));
1162 
1163 	na.ifp = ctx->ifc_ifp;
1164 	na.na_flags = NAF_BDG_MAYSLEEP;
1165 	MPASS(ctx->ifc_softc_ctx.isc_ntxqsets);
1166 	MPASS(ctx->ifc_softc_ctx.isc_nrxqsets);
1167 
1168 	na.num_tx_desc = scctx->isc_ntxd[0];
1169 	na.num_rx_desc = scctx->isc_nrxd[0];
1170 	na.nm_txsync = iflib_netmap_txsync;
1171 	na.nm_rxsync = iflib_netmap_rxsync;
1172 	na.nm_register = iflib_netmap_register;
1173 	na.nm_intr = iflib_netmap_intr;
1174 	na.num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
1175 	na.num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
1176 	return (netmap_attach(&na));
1177 }
1178 
1179 static void
1180 iflib_netmap_txq_init(if_ctx_t ctx, iflib_txq_t txq)
1181 {
1182 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1183 	struct netmap_slot *slot;
1184 
1185 	slot = netmap_reset(na, NR_TX, txq->ift_id, 0);
1186 	if (slot == NULL)
1187 		return;
1188 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxd[0]; i++) {
1189 
1190 		/*
1191 		 * In netmap mode, set the map for the packet buffer.
1192 		 * NOTE: Some drivers (not this one) also need to set
1193 		 * the physical buffer address in the NIC ring.
1194 		 * netmap_idx_n2k() maps a nic index, i, into the corresponding
1195 		 * netmap slot index, si
1196 		 */
1197 		int si = netmap_idx_n2k(na->tx_rings[txq->ift_id], i);
1198 		netmap_load_map(na, txq->ift_buf_tag, txq->ift_sds.ifsd_map[i],
1199 		    NMB(na, slot + si));
1200 	}
1201 }
1202 
1203 static void
1204 iflib_netmap_rxq_init(if_ctx_t ctx, iflib_rxq_t rxq)
1205 {
1206 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1207 	struct netmap_kring *kring = na->rx_rings[rxq->ifr_id];
1208 	struct netmap_slot *slot;
1209 	uint32_t nm_i;
1210 
1211 	slot = netmap_reset(na, NR_RX, rxq->ifr_id, 0);
1212 	if (slot == NULL)
1213 		return;
1214 	nm_i = netmap_idx_n2k(kring, 0);
1215 	netmap_fl_refill(rxq, kring, nm_i, true);
1216 }
1217 
1218 static void
1219 iflib_netmap_timer_adjust(if_ctx_t ctx, iflib_txq_t txq, uint32_t *reset_on)
1220 {
1221 	struct netmap_kring *kring;
1222 	uint16_t txqid;
1223 
1224 	txqid = txq->ift_id;
1225 	kring = NA(ctx->ifc_ifp)->tx_rings[txqid];
1226 
1227 	if (kring->nr_hwcur != nm_next(kring->nr_hwtail, kring->nkr_num_slots - 1)) {
1228 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1229 		    BUS_DMASYNC_POSTREAD);
1230 		if (ctx->isc_txd_credits_update(ctx->ifc_softc, txqid, false))
1231 			netmap_tx_irq(ctx->ifc_ifp, txqid);
1232 		if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ)) {
1233 			if (hz < 2000)
1234 				*reset_on = 1;
1235 			else
1236 				*reset_on = hz / 1000;
1237 		}
1238 	}
1239 }
1240 
1241 #define iflib_netmap_detach(ifp) netmap_detach(ifp)
1242 
1243 #else
1244 #define iflib_netmap_txq_init(ctx, txq)
1245 #define iflib_netmap_rxq_init(ctx, rxq)
1246 #define iflib_netmap_detach(ifp)
1247 
1248 #define iflib_netmap_attach(ctx) (0)
1249 #define netmap_rx_irq(ifp, qid, budget) (0)
1250 #define netmap_tx_irq(ifp, qid) do {} while (0)
1251 #define iflib_netmap_timer_adjust(ctx, txq, reset_on)
1252 #endif
1253 
1254 #if defined(__i386__) || defined(__amd64__)
1255 static __inline void
1256 prefetch(void *x)
1257 {
1258 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1259 }
1260 static __inline void
1261 prefetch2cachelines(void *x)
1262 {
1263 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1264 #if (CACHE_LINE_SIZE < 128)
1265 	__asm volatile("prefetcht0 %0" :: "m" (*(((unsigned long *)x)+CACHE_LINE_SIZE/(sizeof(unsigned long)))));
1266 #endif
1267 }
1268 #else
1269 #define prefetch(x)
1270 #define prefetch2cachelines(x)
1271 #endif
1272 
1273 static void
1274 iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid)
1275 {
1276 	iflib_fl_t fl;
1277 
1278 	fl = &rxq->ifr_fl[flid];
1279 	iru->iru_paddrs = fl->ifl_bus_addrs;
1280 	iru->iru_vaddrs = &fl->ifl_vm_addrs[0];
1281 	iru->iru_idxs = fl->ifl_rxd_idxs;
1282 	iru->iru_qsidx = rxq->ifr_id;
1283 	iru->iru_buf_size = fl->ifl_buf_size;
1284 	iru->iru_flidx = fl->ifl_id;
1285 }
1286 
1287 static void
1288 _iflib_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int err)
1289 {
1290 	if (err)
1291 		return;
1292 	*(bus_addr_t *) arg = segs[0].ds_addr;
1293 }
1294 
1295 int
1296 iflib_dma_alloc_align(if_ctx_t ctx, int size, int align, iflib_dma_info_t dma, int mapflags)
1297 {
1298 	int err;
1299 	device_t dev = ctx->ifc_dev;
1300 
1301 	err = bus_dma_tag_create(bus_get_dma_tag(dev),	/* parent */
1302 				align, 0,		/* alignment, bounds */
1303 				BUS_SPACE_MAXADDR,	/* lowaddr */
1304 				BUS_SPACE_MAXADDR,	/* highaddr */
1305 				NULL, NULL,		/* filter, filterarg */
1306 				size,			/* maxsize */
1307 				1,			/* nsegments */
1308 				size,			/* maxsegsize */
1309 				BUS_DMA_ALLOCNOW,	/* flags */
1310 				NULL,			/* lockfunc */
1311 				NULL,			/* lockarg */
1312 				&dma->idi_tag);
1313 	if (err) {
1314 		device_printf(dev,
1315 		    "%s: bus_dma_tag_create failed: %d\n",
1316 		    __func__, err);
1317 		goto fail_0;
1318 	}
1319 
1320 	err = bus_dmamem_alloc(dma->idi_tag, (void**) &dma->idi_vaddr,
1321 	    BUS_DMA_NOWAIT | BUS_DMA_COHERENT | BUS_DMA_ZERO, &dma->idi_map);
1322 	if (err) {
1323 		device_printf(dev,
1324 		    "%s: bus_dmamem_alloc(%ju) failed: %d\n",
1325 		    __func__, (uintmax_t)size, err);
1326 		goto fail_1;
1327 	}
1328 
1329 	dma->idi_paddr = IF_BAD_DMA;
1330 	err = bus_dmamap_load(dma->idi_tag, dma->idi_map, dma->idi_vaddr,
1331 	    size, _iflib_dmamap_cb, &dma->idi_paddr, mapflags | BUS_DMA_NOWAIT);
1332 	if (err || dma->idi_paddr == IF_BAD_DMA) {
1333 		device_printf(dev,
1334 		    "%s: bus_dmamap_load failed: %d\n",
1335 		    __func__, err);
1336 		goto fail_2;
1337 	}
1338 
1339 	dma->idi_size = size;
1340 	return (0);
1341 
1342 fail_2:
1343 	bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1344 fail_1:
1345 	bus_dma_tag_destroy(dma->idi_tag);
1346 fail_0:
1347 	dma->idi_tag = NULL;
1348 
1349 	return (err);
1350 }
1351 
1352 int
1353 iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags)
1354 {
1355 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1356 
1357 	KASSERT(sctx->isc_q_align != 0, ("alignment value not initialized"));
1358 
1359 	return (iflib_dma_alloc_align(ctx, size, sctx->isc_q_align, dma, mapflags));
1360 }
1361 
1362 int
1363 iflib_dma_alloc_multi(if_ctx_t ctx, int *sizes, iflib_dma_info_t *dmalist, int mapflags, int count)
1364 {
1365 	int i, err;
1366 	iflib_dma_info_t *dmaiter;
1367 
1368 	dmaiter = dmalist;
1369 	for (i = 0; i < count; i++, dmaiter++) {
1370 		if ((err = iflib_dma_alloc(ctx, sizes[i], *dmaiter, mapflags)) != 0)
1371 			break;
1372 	}
1373 	if (err)
1374 		iflib_dma_free_multi(dmalist, i);
1375 	return (err);
1376 }
1377 
1378 void
1379 iflib_dma_free(iflib_dma_info_t dma)
1380 {
1381 	if (dma->idi_tag == NULL)
1382 		return;
1383 	if (dma->idi_paddr != IF_BAD_DMA) {
1384 		bus_dmamap_sync(dma->idi_tag, dma->idi_map,
1385 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1386 		bus_dmamap_unload(dma->idi_tag, dma->idi_map);
1387 		dma->idi_paddr = IF_BAD_DMA;
1388 	}
1389 	if (dma->idi_vaddr != NULL) {
1390 		bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1391 		dma->idi_vaddr = NULL;
1392 	}
1393 	bus_dma_tag_destroy(dma->idi_tag);
1394 	dma->idi_tag = NULL;
1395 }
1396 
1397 void
1398 iflib_dma_free_multi(iflib_dma_info_t *dmalist, int count)
1399 {
1400 	int i;
1401 	iflib_dma_info_t *dmaiter = dmalist;
1402 
1403 	for (i = 0; i < count; i++, dmaiter++)
1404 		iflib_dma_free(*dmaiter);
1405 }
1406 
1407 #ifdef EARLY_AP_STARTUP
1408 static const int iflib_started = 1;
1409 #else
1410 /*
1411  * We used to abuse the smp_started flag to decide if the queues have been
1412  * fully initialized (by late taskqgroup_adjust() calls in a SYSINIT()).
1413  * That gave bad races, since the SYSINIT() runs strictly after smp_started
1414  * is set.  Run a SYSINIT() strictly after that to just set a usable
1415  * completion flag.
1416  */
1417 
1418 static int iflib_started;
1419 
1420 static void
1421 iflib_record_started(void *arg)
1422 {
1423 	iflib_started = 1;
1424 }
1425 
1426 SYSINIT(iflib_record_started, SI_SUB_SMP + 1, SI_ORDER_FIRST,
1427 	iflib_record_started, NULL);
1428 #endif
1429 
1430 static int
1431 iflib_fast_intr(void *arg)
1432 {
1433 	iflib_filter_info_t info = arg;
1434 	struct grouptask *gtask = info->ifi_task;
1435 	int result;
1436 
1437 	if (!iflib_started)
1438 		return (FILTER_STRAY);
1439 
1440 	DBG_COUNTER_INC(fast_intrs);
1441 	if (info->ifi_filter != NULL) {
1442 		result = info->ifi_filter(info->ifi_filter_arg);
1443 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1444 			return (result);
1445 	}
1446 
1447 	GROUPTASK_ENQUEUE(gtask);
1448 	return (FILTER_HANDLED);
1449 }
1450 
1451 static int
1452 iflib_fast_intr_rxtx(void *arg)
1453 {
1454 	iflib_filter_info_t info = arg;
1455 	struct grouptask *gtask = info->ifi_task;
1456 	if_ctx_t ctx;
1457 	iflib_rxq_t rxq = (iflib_rxq_t)info->ifi_ctx;
1458 	iflib_txq_t txq;
1459 	void *sc;
1460 	int i, cidx, result;
1461 	qidx_t txqid;
1462 	bool intr_enable, intr_legacy;
1463 
1464 	if (!iflib_started)
1465 		return (FILTER_STRAY);
1466 
1467 	DBG_COUNTER_INC(fast_intrs);
1468 	if (info->ifi_filter != NULL) {
1469 		result = info->ifi_filter(info->ifi_filter_arg);
1470 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1471 			return (result);
1472 	}
1473 
1474 	ctx = rxq->ifr_ctx;
1475 	sc = ctx->ifc_softc;
1476 	intr_enable = false;
1477 	intr_legacy = !!(ctx->ifc_flags & IFC_LEGACY);
1478 	MPASS(rxq->ifr_ntxqirq);
1479 	for (i = 0; i < rxq->ifr_ntxqirq; i++) {
1480 		txqid = rxq->ifr_txqid[i];
1481 		txq = &ctx->ifc_txqs[txqid];
1482 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1483 		    BUS_DMASYNC_POSTREAD);
1484 		if (!ctx->isc_txd_credits_update(sc, txqid, false)) {
1485 			if (intr_legacy)
1486 				intr_enable = true;
1487 			else
1488 				IFDI_TX_QUEUE_INTR_ENABLE(ctx, txqid);
1489 			continue;
1490 		}
1491 		GROUPTASK_ENQUEUE(&txq->ift_task);
1492 	}
1493 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_RXCQ)
1494 		cidx = rxq->ifr_cq_cidx;
1495 	else
1496 		cidx = rxq->ifr_fl[0].ifl_cidx;
1497 	if (iflib_rxd_avail(ctx, rxq, cidx, 1))
1498 		GROUPTASK_ENQUEUE(gtask);
1499 	else {
1500 		if (intr_legacy)
1501 			intr_enable = true;
1502 		else
1503 			IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
1504 		DBG_COUNTER_INC(rx_intr_enables);
1505 	}
1506 	if (intr_enable)
1507 		IFDI_INTR_ENABLE(ctx);
1508 	return (FILTER_HANDLED);
1509 }
1510 
1511 
1512 static int
1513 iflib_fast_intr_ctx(void *arg)
1514 {
1515 	iflib_filter_info_t info = arg;
1516 	struct grouptask *gtask = info->ifi_task;
1517 	int result;
1518 
1519 	if (!iflib_started)
1520 		return (FILTER_STRAY);
1521 
1522 	DBG_COUNTER_INC(fast_intrs);
1523 	if (info->ifi_filter != NULL) {
1524 		result = info->ifi_filter(info->ifi_filter_arg);
1525 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1526 			return (result);
1527 	}
1528 
1529 	GROUPTASK_ENQUEUE(gtask);
1530 	return (FILTER_HANDLED);
1531 }
1532 
1533 static int
1534 _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
1535 		 driver_filter_t filter, driver_intr_t handler, void *arg,
1536 		 const char *name)
1537 {
1538 	struct resource *res;
1539 	void *tag = NULL;
1540 	device_t dev = ctx->ifc_dev;
1541 	int flags, i, rc;
1542 
1543 	flags = RF_ACTIVE;
1544 	if (ctx->ifc_flags & IFC_LEGACY)
1545 		flags |= RF_SHAREABLE;
1546 	MPASS(rid < 512);
1547 	i = rid;
1548 	res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &i, flags);
1549 	if (res == NULL) {
1550 		device_printf(dev,
1551 		    "failed to allocate IRQ for rid %d, name %s.\n", rid, name);
1552 		return (ENOMEM);
1553 	}
1554 	irq->ii_res = res;
1555 	KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL"));
1556 	rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET,
1557 						filter, handler, arg, &tag);
1558 	if (rc != 0) {
1559 		device_printf(dev,
1560 		    "failed to setup interrupt for rid %d, name %s: %d\n",
1561 					  rid, name ? name : "unknown", rc);
1562 		return (rc);
1563 	} else if (name)
1564 		bus_describe_intr(dev, res, tag, "%s", name);
1565 
1566 	irq->ii_tag = tag;
1567 	return (0);
1568 }
1569 
1570 /*********************************************************************
1571  *
1572  *  Allocate DMA resources for TX buffers as well as memory for the TX
1573  *  mbuf map.  TX DMA maps (non-TSO/TSO) and TX mbuf map are kept in a
1574  *  iflib_sw_tx_desc_array structure, storing all the information that
1575  *  is needed to transmit a packet on the wire.  This is called only
1576  *  once at attach, setup is done every reset.
1577  *
1578  **********************************************************************/
1579 static int
1580 iflib_txsd_alloc(iflib_txq_t txq)
1581 {
1582 	if_ctx_t ctx = txq->ift_ctx;
1583 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1584 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1585 	device_t dev = ctx->ifc_dev;
1586 	bus_size_t tsomaxsize;
1587 	int err, nsegments, ntsosegments;
1588 	bool tso;
1589 
1590 	nsegments = scctx->isc_tx_nsegments;
1591 	ntsosegments = scctx->isc_tx_tso_segments_max;
1592 	tsomaxsize = scctx->isc_tx_tso_size_max;
1593 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_VLAN_MTU)
1594 		tsomaxsize += sizeof(struct ether_vlan_header);
1595 	MPASS(scctx->isc_ntxd[0] > 0);
1596 	MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0);
1597 	MPASS(nsegments > 0);
1598 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) {
1599 		MPASS(ntsosegments > 0);
1600 		MPASS(sctx->isc_tso_maxsize >= tsomaxsize);
1601 	}
1602 
1603 	/*
1604 	 * Set up DMA tags for TX buffers.
1605 	 */
1606 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1607 			       1, 0,			/* alignment, bounds */
1608 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1609 			       BUS_SPACE_MAXADDR,	/* highaddr */
1610 			       NULL, NULL,		/* filter, filterarg */
1611 			       sctx->isc_tx_maxsize,		/* maxsize */
1612 			       nsegments,	/* nsegments */
1613 			       sctx->isc_tx_maxsegsize,	/* maxsegsize */
1614 			       0,			/* flags */
1615 			       NULL,			/* lockfunc */
1616 			       NULL,			/* lockfuncarg */
1617 			       &txq->ift_buf_tag))) {
1618 		device_printf(dev,"Unable to allocate TX DMA tag: %d\n", err);
1619 		device_printf(dev,"maxsize: %ju nsegments: %d maxsegsize: %ju\n",
1620 		    (uintmax_t)sctx->isc_tx_maxsize, nsegments, (uintmax_t)sctx->isc_tx_maxsegsize);
1621 		goto fail;
1622 	}
1623 	tso = (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) != 0;
1624 	if (tso && (err = bus_dma_tag_create(bus_get_dma_tag(dev),
1625 			       1, 0,			/* alignment, bounds */
1626 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1627 			       BUS_SPACE_MAXADDR,	/* highaddr */
1628 			       NULL, NULL,		/* filter, filterarg */
1629 			       tsomaxsize,		/* maxsize */
1630 			       ntsosegments,	/* nsegments */
1631 			       sctx->isc_tso_maxsegsize,/* maxsegsize */
1632 			       0,			/* flags */
1633 			       NULL,			/* lockfunc */
1634 			       NULL,			/* lockfuncarg */
1635 			       &txq->ift_tso_buf_tag))) {
1636 		device_printf(dev, "Unable to allocate TSO TX DMA tag: %d\n",
1637 		    err);
1638 		goto fail;
1639 	}
1640 
1641 	/* Allocate memory for the TX mbuf map. */
1642 	if (!(txq->ift_sds.ifsd_m =
1643 	    (struct mbuf **) malloc(sizeof(struct mbuf *) *
1644 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1645 		device_printf(dev, "Unable to allocate TX mbuf map memory\n");
1646 		err = ENOMEM;
1647 		goto fail;
1648 	}
1649 
1650 	/*
1651 	 * Create the DMA maps for TX buffers.
1652 	 */
1653 	if ((txq->ift_sds.ifsd_map = (bus_dmamap_t *)malloc(
1654 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1655 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1656 		device_printf(dev,
1657 		    "Unable to allocate TX buffer DMA map memory\n");
1658 		err = ENOMEM;
1659 		goto fail;
1660 	}
1661 	if (tso && (txq->ift_sds.ifsd_tso_map = (bus_dmamap_t *)malloc(
1662 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1663 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1664 		device_printf(dev,
1665 		    "Unable to allocate TSO TX buffer map memory\n");
1666 		err = ENOMEM;
1667 		goto fail;
1668 	}
1669 	for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) {
1670 		err = bus_dmamap_create(txq->ift_buf_tag, 0,
1671 		    &txq->ift_sds.ifsd_map[i]);
1672 		if (err != 0) {
1673 			device_printf(dev, "Unable to create TX DMA map\n");
1674 			goto fail;
1675 		}
1676 		if (!tso)
1677 			continue;
1678 		err = bus_dmamap_create(txq->ift_tso_buf_tag, 0,
1679 		    &txq->ift_sds.ifsd_tso_map[i]);
1680 		if (err != 0) {
1681 			device_printf(dev, "Unable to create TSO TX DMA map\n");
1682 			goto fail;
1683 		}
1684 	}
1685 	return (0);
1686 fail:
1687 	/* We free all, it handles case where we are in the middle */
1688 	iflib_tx_structures_free(ctx);
1689 	return (err);
1690 }
1691 
1692 static void
1693 iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i)
1694 {
1695 	bus_dmamap_t map;
1696 
1697 	map = NULL;
1698 	if (txq->ift_sds.ifsd_map != NULL)
1699 		map = txq->ift_sds.ifsd_map[i];
1700 	if (map != NULL) {
1701 		bus_dmamap_sync(txq->ift_buf_tag, map, BUS_DMASYNC_POSTWRITE);
1702 		bus_dmamap_unload(txq->ift_buf_tag, map);
1703 		bus_dmamap_destroy(txq->ift_buf_tag, map);
1704 		txq->ift_sds.ifsd_map[i] = NULL;
1705 	}
1706 
1707 	map = NULL;
1708 	if (txq->ift_sds.ifsd_tso_map != NULL)
1709 		map = txq->ift_sds.ifsd_tso_map[i];
1710 	if (map != NULL) {
1711 		bus_dmamap_sync(txq->ift_tso_buf_tag, map,
1712 		    BUS_DMASYNC_POSTWRITE);
1713 		bus_dmamap_unload(txq->ift_tso_buf_tag, map);
1714 		bus_dmamap_destroy(txq->ift_tso_buf_tag, map);
1715 		txq->ift_sds.ifsd_tso_map[i] = NULL;
1716 	}
1717 }
1718 
1719 static void
1720 iflib_txq_destroy(iflib_txq_t txq)
1721 {
1722 	if_ctx_t ctx = txq->ift_ctx;
1723 
1724 	for (int i = 0; i < txq->ift_size; i++)
1725 		iflib_txsd_destroy(ctx, txq, i);
1726 	if (txq->ift_sds.ifsd_map != NULL) {
1727 		free(txq->ift_sds.ifsd_map, M_IFLIB);
1728 		txq->ift_sds.ifsd_map = NULL;
1729 	}
1730 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1731 		free(txq->ift_sds.ifsd_tso_map, M_IFLIB);
1732 		txq->ift_sds.ifsd_tso_map = NULL;
1733 	}
1734 	if (txq->ift_sds.ifsd_m != NULL) {
1735 		free(txq->ift_sds.ifsd_m, M_IFLIB);
1736 		txq->ift_sds.ifsd_m = NULL;
1737 	}
1738 	if (txq->ift_buf_tag != NULL) {
1739 		bus_dma_tag_destroy(txq->ift_buf_tag);
1740 		txq->ift_buf_tag = NULL;
1741 	}
1742 	if (txq->ift_tso_buf_tag != NULL) {
1743 		bus_dma_tag_destroy(txq->ift_tso_buf_tag);
1744 		txq->ift_tso_buf_tag = NULL;
1745 	}
1746 }
1747 
1748 static void
1749 iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i)
1750 {
1751 	struct mbuf **mp;
1752 
1753 	mp = &txq->ift_sds.ifsd_m[i];
1754 	if (*mp == NULL)
1755 		return;
1756 
1757 	if (txq->ift_sds.ifsd_map != NULL) {
1758 		bus_dmamap_sync(txq->ift_buf_tag,
1759 		    txq->ift_sds.ifsd_map[i], BUS_DMASYNC_POSTWRITE);
1760 		bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[i]);
1761 	}
1762 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1763 		bus_dmamap_sync(txq->ift_tso_buf_tag,
1764 		    txq->ift_sds.ifsd_tso_map[i], BUS_DMASYNC_POSTWRITE);
1765 		bus_dmamap_unload(txq->ift_tso_buf_tag,
1766 		    txq->ift_sds.ifsd_tso_map[i]);
1767 	}
1768 	m_free(*mp);
1769 	DBG_COUNTER_INC(tx_frees);
1770 	*mp = NULL;
1771 }
1772 
1773 static int
1774 iflib_txq_setup(iflib_txq_t txq)
1775 {
1776 	if_ctx_t ctx = txq->ift_ctx;
1777 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1778 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1779 	iflib_dma_info_t di;
1780 	int i;
1781 
1782 	/* Set number of descriptors available */
1783 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
1784 	/* XXX make configurable */
1785 	txq->ift_update_freq = IFLIB_DEFAULT_TX_UPDATE_FREQ;
1786 
1787 	/* Reset indices */
1788 	txq->ift_cidx_processed = 0;
1789 	txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0;
1790 	txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset];
1791 
1792 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1793 		bzero((void *)di->idi_vaddr, di->idi_size);
1794 
1795 	IFDI_TXQ_SETUP(ctx, txq->ift_id);
1796 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1797 		bus_dmamap_sync(di->idi_tag, di->idi_map,
1798 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1799 	return (0);
1800 }
1801 
1802 /*********************************************************************
1803  *
1804  *  Allocate DMA resources for RX buffers as well as memory for the RX
1805  *  mbuf map, direct RX cluster pointer map and RX cluster bus address
1806  *  map.  RX DMA map, RX mbuf map, direct RX cluster pointer map and
1807  *  RX cluster map are kept in a iflib_sw_rx_desc_array structure.
1808  *  Since we use use one entry in iflib_sw_rx_desc_array per received
1809  *  packet, the maximum number of entries we'll need is equal to the
1810  *  number of hardware receive descriptors that we've allocated.
1811  *
1812  **********************************************************************/
1813 static int
1814 iflib_rxsd_alloc(iflib_rxq_t rxq)
1815 {
1816 	if_ctx_t ctx = rxq->ifr_ctx;
1817 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1818 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1819 	device_t dev = ctx->ifc_dev;
1820 	iflib_fl_t fl;
1821 	int			err;
1822 
1823 	MPASS(scctx->isc_nrxd[0] > 0);
1824 	MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0);
1825 
1826 	fl = rxq->ifr_fl;
1827 	for (int i = 0; i <  rxq->ifr_nfl; i++, fl++) {
1828 		fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */
1829 		/* Set up DMA tag for RX buffers. */
1830 		err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1831 					 1, 0,			/* alignment, bounds */
1832 					 BUS_SPACE_MAXADDR,	/* lowaddr */
1833 					 BUS_SPACE_MAXADDR,	/* highaddr */
1834 					 NULL, NULL,		/* filter, filterarg */
1835 					 sctx->isc_rx_maxsize,	/* maxsize */
1836 					 sctx->isc_rx_nsegments,	/* nsegments */
1837 					 sctx->isc_rx_maxsegsize,	/* maxsegsize */
1838 					 0,			/* flags */
1839 					 NULL,			/* lockfunc */
1840 					 NULL,			/* lockarg */
1841 					 &fl->ifl_buf_tag);
1842 		if (err) {
1843 			device_printf(dev,
1844 			    "Unable to allocate RX DMA tag: %d\n", err);
1845 			goto fail;
1846 		}
1847 
1848 		/* Allocate memory for the RX mbuf map. */
1849 		if (!(fl->ifl_sds.ifsd_m =
1850 		      (struct mbuf **) malloc(sizeof(struct mbuf *) *
1851 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1852 			device_printf(dev,
1853 			    "Unable to allocate RX mbuf map memory\n");
1854 			err = ENOMEM;
1855 			goto fail;
1856 		}
1857 
1858 		/* Allocate memory for the direct RX cluster pointer map. */
1859 		if (!(fl->ifl_sds.ifsd_cl =
1860 		      (caddr_t *) malloc(sizeof(caddr_t) *
1861 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1862 			device_printf(dev,
1863 			    "Unable to allocate RX cluster map memory\n");
1864 			err = ENOMEM;
1865 			goto fail;
1866 		}
1867 
1868 		/* Allocate memory for the RX cluster bus address map. */
1869 		if (!(fl->ifl_sds.ifsd_ba =
1870 		      (bus_addr_t *) malloc(sizeof(bus_addr_t) *
1871 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1872 			device_printf(dev,
1873 			    "Unable to allocate RX bus address map memory\n");
1874 			err = ENOMEM;
1875 			goto fail;
1876 		}
1877 
1878 		/*
1879 		 * Create the DMA maps for RX buffers.
1880 		 */
1881 		if (!(fl->ifl_sds.ifsd_map =
1882 		      (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1883 			device_printf(dev,
1884 			    "Unable to allocate RX buffer DMA map memory\n");
1885 			err = ENOMEM;
1886 			goto fail;
1887 		}
1888 		for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++) {
1889 			err = bus_dmamap_create(fl->ifl_buf_tag, 0,
1890 			    &fl->ifl_sds.ifsd_map[i]);
1891 			if (err != 0) {
1892 				device_printf(dev, "Unable to create RX buffer DMA map\n");
1893 				goto fail;
1894 			}
1895 		}
1896 	}
1897 	return (0);
1898 
1899 fail:
1900 	iflib_rx_structures_free(ctx);
1901 	return (err);
1902 }
1903 
1904 
1905 /*
1906  * Internal service routines
1907  */
1908 
1909 struct rxq_refill_cb_arg {
1910 	int               error;
1911 	bus_dma_segment_t seg;
1912 	int               nseg;
1913 };
1914 
1915 static void
1916 _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
1917 {
1918 	struct rxq_refill_cb_arg *cb_arg = arg;
1919 
1920 	cb_arg->error = error;
1921 	cb_arg->seg = segs[0];
1922 	cb_arg->nseg = nseg;
1923 }
1924 
1925 /**
1926  * _iflib_fl_refill - refill an rxq free-buffer list
1927  * @ctx: the iflib context
1928  * @fl: the free list to refill
1929  * @count: the number of new buffers to allocate
1930  *
1931  * (Re)populate an rxq free-buffer list with up to @count new packet buffers.
1932  * The caller must assure that @count does not exceed the queue's capacity.
1933  */
1934 static void
1935 _iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
1936 {
1937 	struct if_rxd_update iru;
1938 	struct rxq_refill_cb_arg cb_arg;
1939 	struct mbuf *m;
1940 	caddr_t cl, *sd_cl;
1941 	struct mbuf **sd_m;
1942 	bus_dmamap_t *sd_map;
1943 	bus_addr_t bus_addr, *sd_ba;
1944 	int err, frag_idx, i, idx, n, pidx;
1945 	qidx_t credits;
1946 
1947 	sd_m = fl->ifl_sds.ifsd_m;
1948 	sd_map = fl->ifl_sds.ifsd_map;
1949 	sd_cl = fl->ifl_sds.ifsd_cl;
1950 	sd_ba = fl->ifl_sds.ifsd_ba;
1951 	pidx = fl->ifl_pidx;
1952 	idx = pidx;
1953 	frag_idx = fl->ifl_fragidx;
1954 	credits = fl->ifl_credits;
1955 
1956 	i = 0;
1957 	n = count;
1958 	MPASS(n > 0);
1959 	MPASS(credits + n <= fl->ifl_size);
1960 
1961 	if (pidx < fl->ifl_cidx)
1962 		MPASS(pidx + n <= fl->ifl_cidx);
1963 	if (pidx == fl->ifl_cidx && (credits < fl->ifl_size))
1964 		MPASS(fl->ifl_gen == 0);
1965 	if (pidx > fl->ifl_cidx)
1966 		MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx);
1967 
1968 	DBG_COUNTER_INC(fl_refills);
1969 	if (n > 8)
1970 		DBG_COUNTER_INC(fl_refills_large);
1971 	iru_init(&iru, fl->ifl_rxq, fl->ifl_id);
1972 	while (n--) {
1973 		/*
1974 		 * We allocate an uninitialized mbuf + cluster, mbuf is
1975 		 * initialized after rx.
1976 		 *
1977 		 * If the cluster is still set then we know a minimum sized packet was received
1978 		 */
1979 		bit_ffc_at(fl->ifl_rx_bitmap, frag_idx, fl->ifl_size,
1980 		    &frag_idx);
1981 		if (frag_idx < 0)
1982 			bit_ffc(fl->ifl_rx_bitmap, fl->ifl_size, &frag_idx);
1983 		MPASS(frag_idx >= 0);
1984 		if ((cl = sd_cl[frag_idx]) == NULL) {
1985 			if ((cl = m_cljget(NULL, M_NOWAIT, fl->ifl_buf_size)) == NULL)
1986 				break;
1987 
1988 			cb_arg.error = 0;
1989 			MPASS(sd_map != NULL);
1990 			err = bus_dmamap_load(fl->ifl_buf_tag, sd_map[frag_idx],
1991 			    cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg,
1992 			    BUS_DMA_NOWAIT);
1993 			if (err != 0 || cb_arg.error) {
1994 				/*
1995 				 * !zone_pack ?
1996 				 */
1997 				if (fl->ifl_zone == zone_pack)
1998 					uma_zfree(fl->ifl_zone, cl);
1999 				break;
2000 			}
2001 
2002 			sd_ba[frag_idx] =  bus_addr = cb_arg.seg.ds_addr;
2003 			sd_cl[frag_idx] = cl;
2004 #if MEMORY_LOGGING
2005 			fl->ifl_cl_enqueued++;
2006 #endif
2007 		} else {
2008 			bus_addr = sd_ba[frag_idx];
2009 		}
2010 		bus_dmamap_sync(fl->ifl_buf_tag, sd_map[frag_idx],
2011 		    BUS_DMASYNC_PREREAD);
2012 
2013 		if (sd_m[frag_idx] == NULL) {
2014 			if ((m = m_gethdr(M_NOWAIT, MT_NOINIT)) == NULL) {
2015 				break;
2016 			}
2017 			sd_m[frag_idx] = m;
2018 		}
2019 		bit_set(fl->ifl_rx_bitmap, frag_idx);
2020 #if MEMORY_LOGGING
2021 		fl->ifl_m_enqueued++;
2022 #endif
2023 
2024 		DBG_COUNTER_INC(rx_allocs);
2025 		fl->ifl_rxd_idxs[i] = frag_idx;
2026 		fl->ifl_bus_addrs[i] = bus_addr;
2027 		fl->ifl_vm_addrs[i] = cl;
2028 		credits++;
2029 		i++;
2030 		MPASS(credits <= fl->ifl_size);
2031 		if (++idx == fl->ifl_size) {
2032 			fl->ifl_gen = 1;
2033 			idx = 0;
2034 		}
2035 		if (n == 0 || i == IFLIB_MAX_RX_REFRESH) {
2036 			iru.iru_pidx = pidx;
2037 			iru.iru_count = i;
2038 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2039 			i = 0;
2040 			pidx = idx;
2041 			fl->ifl_pidx = idx;
2042 			fl->ifl_credits = credits;
2043 		}
2044 	}
2045 
2046 	if (i) {
2047 		iru.iru_pidx = pidx;
2048 		iru.iru_count = i;
2049 		ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2050 		fl->ifl_pidx = idx;
2051 		fl->ifl_credits = credits;
2052 	}
2053 	DBG_COUNTER_INC(rxd_flush);
2054 	if (fl->ifl_pidx == 0)
2055 		pidx = fl->ifl_size - 1;
2056 	else
2057 		pidx = fl->ifl_pidx - 1;
2058 
2059 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2060 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2061 	ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id, fl->ifl_id, pidx);
2062 	fl->ifl_fragidx = frag_idx;
2063 }
2064 
2065 static __inline void
2066 __iflib_fl_refill_lt(if_ctx_t ctx, iflib_fl_t fl, int max)
2067 {
2068 	/* we avoid allowing pidx to catch up with cidx as it confuses ixl */
2069 	int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1;
2070 #ifdef INVARIANTS
2071 	int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1;
2072 #endif
2073 
2074 	MPASS(fl->ifl_credits <= fl->ifl_size);
2075 	MPASS(reclaimable == delta);
2076 
2077 	if (reclaimable > 0)
2078 		_iflib_fl_refill(ctx, fl, min(max, reclaimable));
2079 }
2080 
2081 uint8_t
2082 iflib_in_detach(if_ctx_t ctx)
2083 {
2084 	bool in_detach;
2085 
2086 	STATE_LOCK(ctx);
2087 	in_detach = !!(ctx->ifc_flags & IFC_IN_DETACH);
2088 	STATE_UNLOCK(ctx);
2089 	return (in_detach);
2090 }
2091 
2092 static void
2093 iflib_fl_bufs_free(iflib_fl_t fl)
2094 {
2095 	iflib_dma_info_t idi = fl->ifl_ifdi;
2096 	bus_dmamap_t sd_map;
2097 	uint32_t i;
2098 
2099 	for (i = 0; i < fl->ifl_size; i++) {
2100 		struct mbuf **sd_m = &fl->ifl_sds.ifsd_m[i];
2101 		caddr_t *sd_cl = &fl->ifl_sds.ifsd_cl[i];
2102 
2103 		if (*sd_cl != NULL) {
2104 			sd_map = fl->ifl_sds.ifsd_map[i];
2105 			bus_dmamap_sync(fl->ifl_buf_tag, sd_map,
2106 			    BUS_DMASYNC_POSTREAD);
2107 			bus_dmamap_unload(fl->ifl_buf_tag, sd_map);
2108 			if (*sd_cl != NULL)
2109 				uma_zfree(fl->ifl_zone, *sd_cl);
2110 			// XXX: Should this get moved out?
2111 			if (iflib_in_detach(fl->ifl_rxq->ifr_ctx))
2112 				bus_dmamap_destroy(fl->ifl_buf_tag, sd_map);
2113 			if (*sd_m != NULL) {
2114 				m_init(*sd_m, M_NOWAIT, MT_DATA, 0);
2115 				uma_zfree(zone_mbuf, *sd_m);
2116 			}
2117 		} else {
2118 			MPASS(*sd_cl == NULL);
2119 			MPASS(*sd_m == NULL);
2120 		}
2121 #if MEMORY_LOGGING
2122 		fl->ifl_m_dequeued++;
2123 		fl->ifl_cl_dequeued++;
2124 #endif
2125 		*sd_cl = NULL;
2126 		*sd_m = NULL;
2127 	}
2128 #ifdef INVARIANTS
2129 	for (i = 0; i < fl->ifl_size; i++) {
2130 		MPASS(fl->ifl_sds.ifsd_cl[i] == NULL);
2131 		MPASS(fl->ifl_sds.ifsd_m[i] == NULL);
2132 	}
2133 #endif
2134 	/*
2135 	 * Reset free list values
2136 	 */
2137 	fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = fl->ifl_fragidx = 0;
2138 	bzero(idi->idi_vaddr, idi->idi_size);
2139 }
2140 
2141 /*********************************************************************
2142  *
2143  *  Initialize a free list and its buffers.
2144  *
2145  **********************************************************************/
2146 static int
2147 iflib_fl_setup(iflib_fl_t fl)
2148 {
2149 	iflib_rxq_t rxq = fl->ifl_rxq;
2150 	if_ctx_t ctx = rxq->ifr_ctx;
2151 
2152 	bit_nclear(fl->ifl_rx_bitmap, 0, fl->ifl_size - 1);
2153 	/*
2154 	** Free current RX buffer structs and their mbufs
2155 	*/
2156 	iflib_fl_bufs_free(fl);
2157 	/* Now replenish the mbufs */
2158 	MPASS(fl->ifl_credits == 0);
2159 	fl->ifl_buf_size = ctx->ifc_rx_mbuf_sz;
2160 	if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size)
2161 		ctx->ifc_max_fl_buf_size = fl->ifl_buf_size;
2162 	fl->ifl_cltype = m_gettype(fl->ifl_buf_size);
2163 	fl->ifl_zone = m_getzone(fl->ifl_buf_size);
2164 
2165 
2166 	/* avoid pre-allocating zillions of clusters to an idle card
2167 	 * potentially speeding up attach
2168 	 */
2169 	_iflib_fl_refill(ctx, fl, min(128, fl->ifl_size));
2170 	MPASS(min(128, fl->ifl_size) == fl->ifl_credits);
2171 	if (min(128, fl->ifl_size) != fl->ifl_credits)
2172 		return (ENOBUFS);
2173 	/*
2174 	 * handle failure
2175 	 */
2176 	MPASS(rxq != NULL);
2177 	MPASS(fl->ifl_ifdi != NULL);
2178 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2179 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2180 	return (0);
2181 }
2182 
2183 /*********************************************************************
2184  *
2185  *  Free receive ring data structures
2186  *
2187  **********************************************************************/
2188 static void
2189 iflib_rx_sds_free(iflib_rxq_t rxq)
2190 {
2191 	iflib_fl_t fl;
2192 	int i, j;
2193 
2194 	if (rxq->ifr_fl != NULL) {
2195 		for (i = 0; i < rxq->ifr_nfl; i++) {
2196 			fl = &rxq->ifr_fl[i];
2197 			if (fl->ifl_buf_tag != NULL) {
2198 				if (fl->ifl_sds.ifsd_map != NULL) {
2199 					for (j = 0; j < fl->ifl_size; j++) {
2200 						if (fl->ifl_sds.ifsd_map[j] ==
2201 						    NULL)
2202 							continue;
2203 						bus_dmamap_sync(
2204 						    fl->ifl_buf_tag,
2205 						    fl->ifl_sds.ifsd_map[j],
2206 						    BUS_DMASYNC_POSTREAD);
2207 						bus_dmamap_unload(
2208 						    fl->ifl_buf_tag,
2209 						    fl->ifl_sds.ifsd_map[j]);
2210 					}
2211 				}
2212 				bus_dma_tag_destroy(fl->ifl_buf_tag);
2213 				fl->ifl_buf_tag = NULL;
2214 			}
2215 			free(fl->ifl_sds.ifsd_m, M_IFLIB);
2216 			free(fl->ifl_sds.ifsd_cl, M_IFLIB);
2217 			free(fl->ifl_sds.ifsd_ba, M_IFLIB);
2218 			free(fl->ifl_sds.ifsd_map, M_IFLIB);
2219 			fl->ifl_sds.ifsd_m = NULL;
2220 			fl->ifl_sds.ifsd_cl = NULL;
2221 			fl->ifl_sds.ifsd_ba = NULL;
2222 			fl->ifl_sds.ifsd_map = NULL;
2223 		}
2224 		free(rxq->ifr_fl, M_IFLIB);
2225 		rxq->ifr_fl = NULL;
2226 		rxq->ifr_cq_cidx = 0;
2227 	}
2228 }
2229 
2230 /*
2231  * Timer routine
2232  */
2233 static void
2234 iflib_timer(void *arg)
2235 {
2236 	iflib_txq_t txq = arg;
2237 	if_ctx_t ctx = txq->ift_ctx;
2238 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2239 	uint64_t this_tick = ticks;
2240 	uint32_t reset_on = hz / 2;
2241 
2242 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
2243 		return;
2244 
2245 	/*
2246 	** Check on the state of the TX queue(s), this
2247 	** can be done without the lock because its RO
2248 	** and the HUNG state will be static if set.
2249 	*/
2250 	if (this_tick - txq->ift_last_timer_tick >= hz / 2) {
2251 		txq->ift_last_timer_tick = this_tick;
2252 		IFDI_TIMER(ctx, txq->ift_id);
2253 		if ((txq->ift_qstatus == IFLIB_QUEUE_HUNG) &&
2254 		    ((txq->ift_cleaned_prev == txq->ift_cleaned) ||
2255 		     (sctx->isc_pause_frames == 0)))
2256 			goto hung;
2257 
2258 		if (ifmp_ring_is_stalled(txq->ift_br))
2259 			txq->ift_qstatus = IFLIB_QUEUE_HUNG;
2260 		txq->ift_cleaned_prev = txq->ift_cleaned;
2261 	}
2262 #ifdef DEV_NETMAP
2263 	if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP)
2264 		iflib_netmap_timer_adjust(ctx, txq, &reset_on);
2265 #endif
2266 	/* handle any laggards */
2267 	if (txq->ift_db_pending)
2268 		GROUPTASK_ENQUEUE(&txq->ift_task);
2269 
2270 	sctx->isc_pause_frames = 0;
2271 	if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)
2272 		callout_reset_on(&txq->ift_timer, reset_on, iflib_timer, txq, txq->ift_timer.c_cpu);
2273 	return;
2274 
2275  hung:
2276 	device_printf(ctx->ifc_dev,
2277 	    "Watchdog timeout (TX: %d desc avail: %d pidx: %d) -- resetting\n",
2278 	    txq->ift_id, TXQ_AVAIL(txq), txq->ift_pidx);
2279 	STATE_LOCK(ctx);
2280 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2281 	ctx->ifc_flags |= (IFC_DO_WATCHDOG|IFC_DO_RESET);
2282 	iflib_admin_intr_deferred(ctx);
2283 	STATE_UNLOCK(ctx);
2284 }
2285 
2286 static void
2287 iflib_calc_rx_mbuf_sz(if_ctx_t ctx)
2288 {
2289 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2290 
2291 	/*
2292 	 * XXX don't set the max_frame_size to larger
2293 	 * than the hardware can handle
2294 	 */
2295 	if (sctx->isc_max_frame_size <= MCLBYTES)
2296 		ctx->ifc_rx_mbuf_sz = MCLBYTES;
2297 	else
2298 		ctx->ifc_rx_mbuf_sz = MJUMPAGESIZE;
2299 }
2300 
2301 uint32_t
2302 iflib_get_rx_mbuf_sz(if_ctx_t ctx)
2303 {
2304 
2305 	return (ctx->ifc_rx_mbuf_sz);
2306 }
2307 
2308 static void
2309 iflib_init_locked(if_ctx_t ctx)
2310 {
2311 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2312 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2313 	if_t ifp = ctx->ifc_ifp;
2314 	iflib_fl_t fl;
2315 	iflib_txq_t txq;
2316 	iflib_rxq_t rxq;
2317 	int i, j, tx_ip_csum_flags, tx_ip6_csum_flags;
2318 
2319 	if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2320 	IFDI_INTR_DISABLE(ctx);
2321 
2322 	tx_ip_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP);
2323 	tx_ip6_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | CSUM_IP6_SCTP);
2324 	/* Set hardware offload abilities */
2325 	if_clearhwassist(ifp);
2326 	if (if_getcapenable(ifp) & IFCAP_TXCSUM)
2327 		if_sethwassistbits(ifp, tx_ip_csum_flags, 0);
2328 	if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6)
2329 		if_sethwassistbits(ifp,  tx_ip6_csum_flags, 0);
2330 	if (if_getcapenable(ifp) & IFCAP_TSO4)
2331 		if_sethwassistbits(ifp, CSUM_IP_TSO, 0);
2332 	if (if_getcapenable(ifp) & IFCAP_TSO6)
2333 		if_sethwassistbits(ifp, CSUM_IP6_TSO, 0);
2334 
2335 	for (i = 0, txq = ctx->ifc_txqs; i < sctx->isc_ntxqsets; i++, txq++) {
2336 		CALLOUT_LOCK(txq);
2337 		callout_stop(&txq->ift_timer);
2338 		CALLOUT_UNLOCK(txq);
2339 		iflib_netmap_txq_init(ctx, txq);
2340 	}
2341 
2342 	/*
2343 	 * Calculate a suitable Rx mbuf size prior to calling IFDI_INIT, so
2344 	 * that drivers can use the value when setting up the hardware receive
2345 	 * buffers.
2346 	 */
2347 	iflib_calc_rx_mbuf_sz(ctx);
2348 
2349 #ifdef INVARIANTS
2350 	i = if_getdrvflags(ifp);
2351 #endif
2352 	IFDI_INIT(ctx);
2353 	MPASS(if_getdrvflags(ifp) == i);
2354 	for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
2355 		/* XXX this should really be done on a per-queue basis */
2356 		if (if_getcapenable(ifp) & IFCAP_NETMAP) {
2357 			MPASS(rxq->ifr_id == i);
2358 			iflib_netmap_rxq_init(ctx, rxq);
2359 			continue;
2360 		}
2361 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
2362 			if (iflib_fl_setup(fl)) {
2363 				device_printf(ctx->ifc_dev,
2364 				    "setting up free list %d failed - "
2365 				    "check cluster settings\n", j);
2366 				goto done;
2367 			}
2368 		}
2369 	}
2370 done:
2371 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
2372 	IFDI_INTR_ENABLE(ctx);
2373 	txq = ctx->ifc_txqs;
2374 	for (i = 0; i < sctx->isc_ntxqsets; i++, txq++)
2375 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq,
2376 			txq->ift_timer.c_cpu);
2377 }
2378 
2379 static int
2380 iflib_media_change(if_t ifp)
2381 {
2382 	if_ctx_t ctx = if_getsoftc(ifp);
2383 	int err;
2384 
2385 	CTX_LOCK(ctx);
2386 	if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0)
2387 		iflib_init_locked(ctx);
2388 	CTX_UNLOCK(ctx);
2389 	return (err);
2390 }
2391 
2392 static void
2393 iflib_media_status(if_t ifp, struct ifmediareq *ifmr)
2394 {
2395 	if_ctx_t ctx = if_getsoftc(ifp);
2396 
2397 	CTX_LOCK(ctx);
2398 	IFDI_UPDATE_ADMIN_STATUS(ctx);
2399 	IFDI_MEDIA_STATUS(ctx, ifmr);
2400 	CTX_UNLOCK(ctx);
2401 }
2402 
2403 void
2404 iflib_stop(if_ctx_t ctx)
2405 {
2406 	iflib_txq_t txq = ctx->ifc_txqs;
2407 	iflib_rxq_t rxq = ctx->ifc_rxqs;
2408 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2409 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2410 	iflib_dma_info_t di;
2411 	iflib_fl_t fl;
2412 	int i, j;
2413 
2414 	/* Tell the stack that the interface is no longer active */
2415 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2416 
2417 	IFDI_INTR_DISABLE(ctx);
2418 	DELAY(1000);
2419 	IFDI_STOP(ctx);
2420 	DELAY(1000);
2421 
2422 	iflib_debug_reset();
2423 	/* Wait for current tx queue users to exit to disarm watchdog timer. */
2424 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) {
2425 		/* make sure all transmitters have completed before proceeding XXX */
2426 
2427 		CALLOUT_LOCK(txq);
2428 		callout_stop(&txq->ift_timer);
2429 		CALLOUT_UNLOCK(txq);
2430 
2431 		/* clean any enqueued buffers */
2432 		iflib_ifmp_purge(txq);
2433 		/* Free any existing tx buffers. */
2434 		for (j = 0; j < txq->ift_size; j++) {
2435 			iflib_txsd_free(ctx, txq, j);
2436 		}
2437 		txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0;
2438 		txq->ift_in_use = txq->ift_gen = txq->ift_cidx = txq->ift_pidx = txq->ift_no_desc_avail = 0;
2439 		txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0;
2440 		txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0;
2441 		txq->ift_pullups = 0;
2442 		ifmp_ring_reset_stats(txq->ift_br);
2443 		for (j = 0, di = txq->ift_ifdi; j < sctx->isc_ntxqs; j++, di++)
2444 			bzero((void *)di->idi_vaddr, di->idi_size);
2445 	}
2446 	for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) {
2447 		/* make sure all transmitters have completed before proceeding XXX */
2448 
2449 		rxq->ifr_cq_cidx = 0;
2450 		for (j = 0, di = rxq->ifr_ifdi; j < sctx->isc_nrxqs; j++, di++)
2451 			bzero((void *)di->idi_vaddr, di->idi_size);
2452 		/* also resets the free lists pidx/cidx */
2453 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
2454 			iflib_fl_bufs_free(fl);
2455 	}
2456 }
2457 
2458 static inline caddr_t
2459 calc_next_rxd(iflib_fl_t fl, int cidx)
2460 {
2461 	qidx_t size;
2462 	int nrxd;
2463 	caddr_t start, end, cur, next;
2464 
2465 	nrxd = fl->ifl_size;
2466 	size = fl->ifl_rxd_size;
2467 	start = fl->ifl_ifdi->idi_vaddr;
2468 
2469 	if (__predict_false(size == 0))
2470 		return (start);
2471 	cur = start + size*cidx;
2472 	end = start + size*nrxd;
2473 	next = CACHE_PTR_NEXT(cur);
2474 	return (next < end ? next : start);
2475 }
2476 
2477 static inline void
2478 prefetch_pkts(iflib_fl_t fl, int cidx)
2479 {
2480 	int nextptr;
2481 	int nrxd = fl->ifl_size;
2482 	caddr_t next_rxd;
2483 
2484 
2485 	nextptr = (cidx + CACHE_PTR_INCREMENT) & (nrxd-1);
2486 	prefetch(&fl->ifl_sds.ifsd_m[nextptr]);
2487 	prefetch(&fl->ifl_sds.ifsd_cl[nextptr]);
2488 	next_rxd = calc_next_rxd(fl, cidx);
2489 	prefetch(next_rxd);
2490 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 1) & (nrxd-1)]);
2491 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 2) & (nrxd-1)]);
2492 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 3) & (nrxd-1)]);
2493 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 4) & (nrxd-1)]);
2494 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 1) & (nrxd-1)]);
2495 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 2) & (nrxd-1)]);
2496 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 3) & (nrxd-1)]);
2497 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 4) & (nrxd-1)]);
2498 }
2499 
2500 static struct mbuf *
2501 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, bool unload, if_rxsd_t sd,
2502     int *pf_rv, if_rxd_info_t ri)
2503 {
2504 	bus_dmamap_t map;
2505 	iflib_fl_t fl;
2506 	caddr_t payload;
2507 	struct mbuf *m;
2508 	int flid, cidx, len, next;
2509 
2510 	map = NULL;
2511 	flid = irf->irf_flid;
2512 	cidx = irf->irf_idx;
2513 	fl = &rxq->ifr_fl[flid];
2514 	sd->ifsd_fl = fl;
2515 	sd->ifsd_cidx = cidx;
2516 	m = fl->ifl_sds.ifsd_m[cidx];
2517 	sd->ifsd_cl = &fl->ifl_sds.ifsd_cl[cidx];
2518 	fl->ifl_credits--;
2519 #if MEMORY_LOGGING
2520 	fl->ifl_m_dequeued++;
2521 #endif
2522 	if (rxq->ifr_ctx->ifc_flags & IFC_PREFETCH)
2523 		prefetch_pkts(fl, cidx);
2524 	next = (cidx + CACHE_PTR_INCREMENT) & (fl->ifl_size-1);
2525 	prefetch(&fl->ifl_sds.ifsd_map[next]);
2526 	map = fl->ifl_sds.ifsd_map[cidx];
2527 	next = (cidx + CACHE_LINE_SIZE) & (fl->ifl_size-1);
2528 
2529 	/* not valid assert if bxe really does SGE from non-contiguous elements */
2530 	MPASS(fl->ifl_cidx == cidx);
2531 	bus_dmamap_sync(fl->ifl_buf_tag, map, BUS_DMASYNC_POSTREAD);
2532 
2533 	if (rxq->pfil != NULL && PFIL_HOOKED_IN(rxq->pfil) && pf_rv != NULL) {
2534 		payload  = *sd->ifsd_cl;
2535 		payload +=  ri->iri_pad;
2536 		len = ri->iri_len - ri->iri_pad;
2537 		*pf_rv = pfil_run_hooks(rxq->pfil, payload, ri->iri_ifp,
2538 		    len | PFIL_MEMPTR | PFIL_IN, NULL);
2539 		switch (*pf_rv) {
2540 		case PFIL_DROPPED:
2541 		case PFIL_CONSUMED:
2542 			/*
2543 			 * The filter ate it.  Everything is recycled.
2544 			 */
2545 			m = NULL;
2546 			unload = 0;
2547 			break;
2548 		case PFIL_REALLOCED:
2549 			/*
2550 			 * The filter copied it.  Everything is recycled.
2551 			 */
2552 			m = pfil_mem2mbuf(payload);
2553 			unload = 0;
2554 			break;
2555 		case PFIL_PASS:
2556 			/*
2557 			 * Filter said it was OK, so receive like
2558 			 * normal
2559 			 */
2560 			fl->ifl_sds.ifsd_m[cidx] = NULL;
2561 			break;
2562 		default:
2563 			MPASS(0);
2564 		}
2565 	} else {
2566 		fl->ifl_sds.ifsd_m[cidx] = NULL;
2567 		*pf_rv = PFIL_PASS;
2568 	}
2569 
2570 	if (unload)
2571 		bus_dmamap_unload(fl->ifl_buf_tag, map);
2572 	fl->ifl_cidx = (fl->ifl_cidx + 1) & (fl->ifl_size-1);
2573 	if (__predict_false(fl->ifl_cidx == 0))
2574 		fl->ifl_gen = 0;
2575 	bit_clear(fl->ifl_rx_bitmap, cidx);
2576 	return (m);
2577 }
2578 
2579 static struct mbuf *
2580 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri, if_rxsd_t sd, int *pf_rv)
2581 {
2582 	struct mbuf *m, *mh, *mt;
2583 	caddr_t cl;
2584 	int  *pf_rv_ptr, flags, i, padlen;
2585 	bool consumed;
2586 
2587 	i = 0;
2588 	mh = NULL;
2589 	consumed = false;
2590 	*pf_rv = PFIL_PASS;
2591 	pf_rv_ptr = pf_rv;
2592 	do {
2593 		m = rxd_frag_to_sd(rxq, &ri->iri_frags[i], !consumed, sd,
2594 		    pf_rv_ptr, ri);
2595 
2596 		MPASS(*sd->ifsd_cl != NULL);
2597 
2598 		/*
2599 		 * Exclude zero-length frags & frags from
2600 		 * packets the filter has consumed or dropped
2601 		 */
2602 		if (ri->iri_frags[i].irf_len == 0 || consumed ||
2603 		    *pf_rv == PFIL_CONSUMED || *pf_rv == PFIL_DROPPED) {
2604 			if (mh == NULL) {
2605 				/* everything saved here */
2606 				consumed = true;
2607 				pf_rv_ptr = NULL;
2608 				continue;
2609 			}
2610 			/* XXX we can save the cluster here, but not the mbuf */
2611 			m_init(m, M_NOWAIT, MT_DATA, 0);
2612 			m_free(m);
2613 			continue;
2614 		}
2615 		if (mh == NULL) {
2616 			flags = M_PKTHDR|M_EXT;
2617 			mh = mt = m;
2618 			padlen = ri->iri_pad;
2619 		} else {
2620 			flags = M_EXT;
2621 			mt->m_next = m;
2622 			mt = m;
2623 			/* assuming padding is only on the first fragment */
2624 			padlen = 0;
2625 		}
2626 		cl = *sd->ifsd_cl;
2627 		*sd->ifsd_cl = NULL;
2628 
2629 		/* Can these two be made one ? */
2630 		m_init(m, M_NOWAIT, MT_DATA, flags);
2631 		m_cljset(m, cl, sd->ifsd_fl->ifl_cltype);
2632 		/*
2633 		 * These must follow m_init and m_cljset
2634 		 */
2635 		m->m_data += padlen;
2636 		ri->iri_len -= padlen;
2637 		m->m_len = ri->iri_frags[i].irf_len;
2638 	} while (++i < ri->iri_nfrags);
2639 
2640 	return (mh);
2641 }
2642 
2643 /*
2644  * Process one software descriptor
2645  */
2646 static struct mbuf *
2647 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri)
2648 {
2649 	struct if_rxsd sd;
2650 	struct mbuf *m;
2651 	int pf_rv;
2652 
2653 	/* should I merge this back in now that the two paths are basically duplicated? */
2654 	if (ri->iri_nfrags == 1 &&
2655 	    ri->iri_frags[0].irf_len <= MIN(IFLIB_RX_COPY_THRESH, MHLEN)) {
2656 		m = rxd_frag_to_sd(rxq, &ri->iri_frags[0], false, &sd,
2657 		    &pf_rv, ri);
2658 		if (pf_rv != PFIL_PASS && pf_rv != PFIL_REALLOCED)
2659 			return (m);
2660 		if (pf_rv == PFIL_PASS) {
2661 			m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR);
2662 #ifndef __NO_STRICT_ALIGNMENT
2663 			if (!IP_ALIGNED(m))
2664 				m->m_data += 2;
2665 #endif
2666 			memcpy(m->m_data, *sd.ifsd_cl, ri->iri_len);
2667 			m->m_len = ri->iri_frags[0].irf_len;
2668 		}
2669 	} else {
2670 		m = assemble_segments(rxq, ri, &sd, &pf_rv);
2671 		if (pf_rv != PFIL_PASS && pf_rv != PFIL_REALLOCED)
2672 			return (m);
2673 	}
2674 	m->m_pkthdr.len = ri->iri_len;
2675 	m->m_pkthdr.rcvif = ri->iri_ifp;
2676 	m->m_flags |= ri->iri_flags;
2677 	m->m_pkthdr.ether_vtag = ri->iri_vtag;
2678 	m->m_pkthdr.flowid = ri->iri_flowid;
2679 	M_HASHTYPE_SET(m, ri->iri_rsstype);
2680 	m->m_pkthdr.csum_flags = ri->iri_csum_flags;
2681 	m->m_pkthdr.csum_data = ri->iri_csum_data;
2682 	return (m);
2683 }
2684 
2685 #if defined(INET6) || defined(INET)
2686 static void
2687 iflib_get_ip_forwarding(struct lro_ctrl *lc, bool *v4, bool *v6)
2688 {
2689 	CURVNET_SET(lc->ifp->if_vnet);
2690 #if defined(INET6)
2691 	*v6 = V_ip6_forwarding;
2692 #endif
2693 #if defined(INET)
2694 	*v4 = V_ipforwarding;
2695 #endif
2696 	CURVNET_RESTORE();
2697 }
2698 
2699 /*
2700  * Returns true if it's possible this packet could be LROed.
2701  * if it returns false, it is guaranteed that tcp_lro_rx()
2702  * would not return zero.
2703  */
2704 static bool
2705 iflib_check_lro_possible(struct mbuf *m, bool v4_forwarding, bool v6_forwarding)
2706 {
2707 	struct ether_header *eh;
2708 
2709 	eh = mtod(m, struct ether_header *);
2710 	switch (eh->ether_type) {
2711 #if defined(INET6)
2712 		case htons(ETHERTYPE_IPV6):
2713 			return (!v6_forwarding);
2714 #endif
2715 #if defined (INET)
2716 		case htons(ETHERTYPE_IP):
2717 			return (!v4_forwarding);
2718 #endif
2719 	}
2720 
2721 	return false;
2722 }
2723 #else
2724 static void
2725 iflib_get_ip_forwarding(struct lro_ctrl *lc __unused, bool *v4 __unused, bool *v6 __unused)
2726 {
2727 }
2728 #endif
2729 
2730 static bool
2731 iflib_rxeof(iflib_rxq_t rxq, qidx_t budget)
2732 {
2733 	if_t ifp;
2734 	if_ctx_t ctx = rxq->ifr_ctx;
2735 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2736 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2737 	int avail, i;
2738 	qidx_t *cidxp;
2739 	struct if_rxd_info ri;
2740 	int err, budget_left, rx_bytes, rx_pkts;
2741 	iflib_fl_t fl;
2742 	int lro_enabled;
2743 	bool v4_forwarding, v6_forwarding, lro_possible;
2744 
2745 	/*
2746 	 * XXX early demux data packets so that if_input processing only handles
2747 	 * acks in interrupt context
2748 	 */
2749 	struct mbuf *m, *mh, *mt, *mf;
2750 
2751 	lro_possible = v4_forwarding = v6_forwarding = false;
2752 	ifp = ctx->ifc_ifp;
2753 	mh = mt = NULL;
2754 	MPASS(budget > 0);
2755 	rx_pkts	= rx_bytes = 0;
2756 	if (sctx->isc_flags & IFLIB_HAS_RXCQ)
2757 		cidxp = &rxq->ifr_cq_cidx;
2758 	else
2759 		cidxp = &rxq->ifr_fl[0].ifl_cidx;
2760 	if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) {
2761 		for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2762 			__iflib_fl_refill_lt(ctx, fl, budget + 8);
2763 		DBG_COUNTER_INC(rx_unavail);
2764 		return (false);
2765 	}
2766 
2767 	/* pfil needs the vnet to be set */
2768 	CURVNET_SET_QUIET(ifp->if_vnet);
2769 	for (budget_left = budget; budget_left > 0 && avail > 0;) {
2770 		if (__predict_false(!CTX_ACTIVE(ctx))) {
2771 			DBG_COUNTER_INC(rx_ctx_inactive);
2772 			break;
2773 		}
2774 		/*
2775 		 * Reset client set fields to their default values
2776 		 */
2777 		rxd_info_zero(&ri);
2778 		ri.iri_qsidx = rxq->ifr_id;
2779 		ri.iri_cidx = *cidxp;
2780 		ri.iri_ifp = ifp;
2781 		ri.iri_frags = rxq->ifr_frags;
2782 		err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
2783 
2784 		if (err)
2785 			goto err;
2786 		rx_pkts += 1;
2787 		rx_bytes += ri.iri_len;
2788 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
2789 			*cidxp = ri.iri_cidx;
2790 			/* Update our consumer index */
2791 			/* XXX NB: shurd - check if this is still safe */
2792 			while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0])
2793 				rxq->ifr_cq_cidx -= scctx->isc_nrxd[0];
2794 			/* was this only a completion queue message? */
2795 			if (__predict_false(ri.iri_nfrags == 0))
2796 				continue;
2797 		}
2798 		MPASS(ri.iri_nfrags != 0);
2799 		MPASS(ri.iri_len != 0);
2800 
2801 		/* will advance the cidx on the corresponding free lists */
2802 		m = iflib_rxd_pkt_get(rxq, &ri);
2803 		avail--;
2804 		budget_left--;
2805 		if (avail == 0 && budget_left)
2806 			avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left);
2807 
2808 		if (__predict_false(m == NULL))
2809 			continue;
2810 
2811 		/* imm_pkt: -- cxgb */
2812 		if (mh == NULL)
2813 			mh = mt = m;
2814 		else {
2815 			mt->m_nextpkt = m;
2816 			mt = m;
2817 		}
2818 	}
2819 	CURVNET_RESTORE();
2820 	/* make sure that we can refill faster than drain */
2821 	for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2822 		__iflib_fl_refill_lt(ctx, fl, budget + 8);
2823 
2824 	lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO);
2825 	if (lro_enabled)
2826 		iflib_get_ip_forwarding(&rxq->ifr_lc, &v4_forwarding, &v6_forwarding);
2827 	mt = mf = NULL;
2828 	while (mh != NULL) {
2829 		m = mh;
2830 		mh = mh->m_nextpkt;
2831 		m->m_nextpkt = NULL;
2832 #ifndef __NO_STRICT_ALIGNMENT
2833 		if (!IP_ALIGNED(m) && (m = iflib_fixup_rx(m)) == NULL)
2834 			continue;
2835 #endif
2836 		rx_bytes += m->m_pkthdr.len;
2837 		rx_pkts++;
2838 #if defined(INET6) || defined(INET)
2839 		if (lro_enabled) {
2840 			if (!lro_possible) {
2841 				lro_possible = iflib_check_lro_possible(m, v4_forwarding, v6_forwarding);
2842 				if (lro_possible && mf != NULL) {
2843 					ifp->if_input(ifp, mf);
2844 					DBG_COUNTER_INC(rx_if_input);
2845 					mt = mf = NULL;
2846 				}
2847 			}
2848 			if ((m->m_pkthdr.csum_flags & (CSUM_L4_CALC|CSUM_L4_VALID)) ==
2849 			    (CSUM_L4_CALC|CSUM_L4_VALID)) {
2850 				if (lro_possible && tcp_lro_rx(&rxq->ifr_lc, m, 0) == 0)
2851 					continue;
2852 			}
2853 		}
2854 #endif
2855 		if (lro_possible) {
2856 			ifp->if_input(ifp, m);
2857 			DBG_COUNTER_INC(rx_if_input);
2858 			continue;
2859 		}
2860 
2861 		if (mf == NULL)
2862 			mf = m;
2863 		if (mt != NULL)
2864 			mt->m_nextpkt = m;
2865 		mt = m;
2866 	}
2867 	if (mf != NULL) {
2868 		ifp->if_input(ifp, mf);
2869 		DBG_COUNTER_INC(rx_if_input);
2870 	}
2871 
2872 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
2873 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
2874 
2875 	/*
2876 	 * Flush any outstanding LRO work
2877 	 */
2878 #if defined(INET6) || defined(INET)
2879 	tcp_lro_flush_all(&rxq->ifr_lc);
2880 #endif
2881 	if (avail)
2882 		return true;
2883 	return (iflib_rxd_avail(ctx, rxq, *cidxp, 1));
2884 err:
2885 	STATE_LOCK(ctx);
2886 	ctx->ifc_flags |= IFC_DO_RESET;
2887 	iflib_admin_intr_deferred(ctx);
2888 	STATE_UNLOCK(ctx);
2889 	return (false);
2890 }
2891 
2892 #define TXD_NOTIFY_COUNT(txq) (((txq)->ift_size / (txq)->ift_update_freq)-1)
2893 static inline qidx_t
2894 txq_max_db_deferred(iflib_txq_t txq, qidx_t in_use)
2895 {
2896 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2897 	qidx_t minthresh = txq->ift_size / 8;
2898 	if (in_use > 4*minthresh)
2899 		return (notify_count);
2900 	if (in_use > 2*minthresh)
2901 		return (notify_count >> 1);
2902 	if (in_use > minthresh)
2903 		return (notify_count >> 3);
2904 	return (0);
2905 }
2906 
2907 static inline qidx_t
2908 txq_max_rs_deferred(iflib_txq_t txq)
2909 {
2910 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2911 	qidx_t minthresh = txq->ift_size / 8;
2912 	if (txq->ift_in_use > 4*minthresh)
2913 		return (notify_count);
2914 	if (txq->ift_in_use > 2*minthresh)
2915 		return (notify_count >> 1);
2916 	if (txq->ift_in_use > minthresh)
2917 		return (notify_count >> 2);
2918 	return (2);
2919 }
2920 
2921 #define M_CSUM_FLAGS(m) ((m)->m_pkthdr.csum_flags)
2922 #define M_HAS_VLANTAG(m) (m->m_flags & M_VLANTAG)
2923 
2924 #define TXQ_MAX_DB_DEFERRED(txq, in_use) txq_max_db_deferred((txq), (in_use))
2925 #define TXQ_MAX_RS_DEFERRED(txq) txq_max_rs_deferred(txq)
2926 #define TXQ_MAX_DB_CONSUMED(size) (size >> 4)
2927 
2928 /* forward compatibility for cxgb */
2929 #define FIRST_QSET(ctx) 0
2930 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets)
2931 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets)
2932 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx))
2933 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments))
2934 
2935 /* XXX we should be setting this to something other than zero */
2936 #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh)
2937 #define	MAX_TX_DESC(ctx) max((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max, \
2938     (ctx)->ifc_softc_ctx.isc_tx_nsegments)
2939 
2940 static inline bool
2941 iflib_txd_db_check(if_ctx_t ctx, iflib_txq_t txq, int ring, qidx_t in_use)
2942 {
2943 	qidx_t dbval, max;
2944 	bool rang;
2945 
2946 	rang = false;
2947 	max = TXQ_MAX_DB_DEFERRED(txq, in_use);
2948 	if (ring || txq->ift_db_pending >= max) {
2949 		dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx;
2950 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
2951 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2952 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval);
2953 		txq->ift_db_pending = txq->ift_npending = 0;
2954 		rang = true;
2955 	}
2956 	return (rang);
2957 }
2958 
2959 #ifdef PKT_DEBUG
2960 static void
2961 print_pkt(if_pkt_info_t pi)
2962 {
2963 	printf("pi len:  %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n",
2964 	       pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx);
2965 	printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n",
2966 	       pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag);
2967 	printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n",
2968 	       pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto);
2969 }
2970 #endif
2971 
2972 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO)
2973 #define IS_TX_OFFLOAD4(pi) ((pi)->ipi_csum_flags & (CSUM_IP_TCP | CSUM_IP_TSO))
2974 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO)
2975 #define IS_TX_OFFLOAD6(pi) ((pi)->ipi_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_TSO))
2976 
2977 static int
2978 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp)
2979 {
2980 	if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx;
2981 	struct ether_vlan_header *eh;
2982 	struct mbuf *m;
2983 
2984 	m = *mp;
2985 	if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) &&
2986 	    M_WRITABLE(m) == 0) {
2987 		if ((m = m_dup(m, M_NOWAIT)) == NULL) {
2988 			return (ENOMEM);
2989 		} else {
2990 			m_freem(*mp);
2991 			DBG_COUNTER_INC(tx_frees);
2992 			*mp = m;
2993 		}
2994 	}
2995 
2996 	/*
2997 	 * Determine where frame payload starts.
2998 	 * Jump over vlan headers if already present,
2999 	 * helpful for QinQ too.
3000 	 */
3001 	if (__predict_false(m->m_len < sizeof(*eh))) {
3002 		txq->ift_pullups++;
3003 		if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL))
3004 			return (ENOMEM);
3005 	}
3006 	eh = mtod(m, struct ether_vlan_header *);
3007 	if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
3008 		pi->ipi_etype = ntohs(eh->evl_proto);
3009 		pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
3010 	} else {
3011 		pi->ipi_etype = ntohs(eh->evl_encap_proto);
3012 		pi->ipi_ehdrlen = ETHER_HDR_LEN;
3013 	}
3014 
3015 	switch (pi->ipi_etype) {
3016 #ifdef INET
3017 	case ETHERTYPE_IP:
3018 	{
3019 		struct mbuf *n;
3020 		struct ip *ip = NULL;
3021 		struct tcphdr *th = NULL;
3022 		int minthlen;
3023 
3024 		minthlen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip) + sizeof(*th));
3025 		if (__predict_false(m->m_len < minthlen)) {
3026 			/*
3027 			 * if this code bloat is causing too much of a hit
3028 			 * move it to a separate function and mark it noinline
3029 			 */
3030 			if (m->m_len == pi->ipi_ehdrlen) {
3031 				n = m->m_next;
3032 				MPASS(n);
3033 				if (n->m_len >= sizeof(*ip))  {
3034 					ip = (struct ip *)n->m_data;
3035 					if (n->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3036 						th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3037 				} else {
3038 					txq->ift_pullups++;
3039 					if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
3040 						return (ENOMEM);
3041 					ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3042 				}
3043 			} else {
3044 				txq->ift_pullups++;
3045 				if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
3046 					return (ENOMEM);
3047 				ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3048 				if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3049 					th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3050 			}
3051 		} else {
3052 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3053 			if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3054 				th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3055 		}
3056 		pi->ipi_ip_hlen = ip->ip_hl << 2;
3057 		pi->ipi_ipproto = ip->ip_p;
3058 		pi->ipi_flags |= IPI_TX_IPV4;
3059 
3060 		/* TCP checksum offload may require TCP header length */
3061 		if (IS_TX_OFFLOAD4(pi)) {
3062 			if (__predict_true(pi->ipi_ipproto == IPPROTO_TCP)) {
3063 				if (__predict_false(th == NULL)) {
3064 					txq->ift_pullups++;
3065 					if (__predict_false((m = m_pullup(m, (ip->ip_hl << 2) + sizeof(*th))) == NULL))
3066 						return (ENOMEM);
3067 					th = (struct tcphdr *)((caddr_t)ip + pi->ipi_ip_hlen);
3068 				}
3069 				pi->ipi_tcp_hflags = th->th_flags;
3070 				pi->ipi_tcp_hlen = th->th_off << 2;
3071 				pi->ipi_tcp_seq = th->th_seq;
3072 			}
3073 			if (IS_TSO4(pi)) {
3074 				if (__predict_false(ip->ip_p != IPPROTO_TCP))
3075 					return (ENXIO);
3076 				/*
3077 				 * TSO always requires hardware checksum offload.
3078 				 */
3079 				pi->ipi_csum_flags |= (CSUM_IP_TCP | CSUM_IP);
3080 				th->th_sum = in_pseudo(ip->ip_src.s_addr,
3081 						       ip->ip_dst.s_addr, htons(IPPROTO_TCP));
3082 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3083 				if (sctx->isc_flags & IFLIB_TSO_INIT_IP) {
3084 					ip->ip_sum = 0;
3085 					ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz);
3086 				}
3087 			}
3088 		}
3089 		if ((sctx->isc_flags & IFLIB_NEED_ZERO_CSUM) && (pi->ipi_csum_flags & CSUM_IP))
3090                        ip->ip_sum = 0;
3091 
3092 		break;
3093 	}
3094 #endif
3095 #ifdef INET6
3096 	case ETHERTYPE_IPV6:
3097 	{
3098 		struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
3099 		struct tcphdr *th;
3100 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
3101 
3102 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
3103 			txq->ift_pullups++;
3104 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
3105 				return (ENOMEM);
3106 		}
3107 		th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
3108 
3109 		/* XXX-BZ this will go badly in case of ext hdrs. */
3110 		pi->ipi_ipproto = ip6->ip6_nxt;
3111 		pi->ipi_flags |= IPI_TX_IPV6;
3112 
3113 		/* TCP checksum offload may require TCP header length */
3114 		if (IS_TX_OFFLOAD6(pi)) {
3115 			if (pi->ipi_ipproto == IPPROTO_TCP) {
3116 				if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) {
3117 					txq->ift_pullups++;
3118 					if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL))
3119 						return (ENOMEM);
3120 				}
3121 				pi->ipi_tcp_hflags = th->th_flags;
3122 				pi->ipi_tcp_hlen = th->th_off << 2;
3123 				pi->ipi_tcp_seq = th->th_seq;
3124 			}
3125 			if (IS_TSO6(pi)) {
3126 				if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP))
3127 					return (ENXIO);
3128 				/*
3129 				 * TSO always requires hardware checksum offload.
3130 				 */
3131 				pi->ipi_csum_flags |= CSUM_IP6_TCP;
3132 				th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0);
3133 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3134 			}
3135 		}
3136 		break;
3137 	}
3138 #endif
3139 	default:
3140 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
3141 		pi->ipi_ip_hlen = 0;
3142 		break;
3143 	}
3144 	*mp = m;
3145 
3146 	return (0);
3147 }
3148 
3149 /*
3150  * If dodgy hardware rejects the scatter gather chain we've handed it
3151  * we'll need to remove the mbuf chain from ifsg_m[] before we can add the
3152  * m_defrag'd mbufs
3153  */
3154 static __noinline struct mbuf *
3155 iflib_remove_mbuf(iflib_txq_t txq)
3156 {
3157 	int ntxd, pidx;
3158 	struct mbuf *m, **ifsd_m;
3159 
3160 	ifsd_m = txq->ift_sds.ifsd_m;
3161 	ntxd = txq->ift_size;
3162 	pidx = txq->ift_pidx & (ntxd - 1);
3163 	ifsd_m = txq->ift_sds.ifsd_m;
3164 	m = ifsd_m[pidx];
3165 	ifsd_m[pidx] = NULL;
3166 	bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[pidx]);
3167 	if (txq->ift_sds.ifsd_tso_map != NULL)
3168 		bus_dmamap_unload(txq->ift_tso_buf_tag,
3169 		    txq->ift_sds.ifsd_tso_map[pidx]);
3170 #if MEMORY_LOGGING
3171 	txq->ift_dequeued++;
3172 #endif
3173 	return (m);
3174 }
3175 
3176 static inline caddr_t
3177 calc_next_txd(iflib_txq_t txq, int cidx, uint8_t qid)
3178 {
3179 	qidx_t size;
3180 	int ntxd;
3181 	caddr_t start, end, cur, next;
3182 
3183 	ntxd = txq->ift_size;
3184 	size = txq->ift_txd_size[qid];
3185 	start = txq->ift_ifdi[qid].idi_vaddr;
3186 
3187 	if (__predict_false(size == 0))
3188 		return (start);
3189 	cur = start + size*cidx;
3190 	end = start + size*ntxd;
3191 	next = CACHE_PTR_NEXT(cur);
3192 	return (next < end ? next : start);
3193 }
3194 
3195 /*
3196  * Pad an mbuf to ensure a minimum ethernet frame size.
3197  * min_frame_size is the frame size (less CRC) to pad the mbuf to
3198  */
3199 static __noinline int
3200 iflib_ether_pad(device_t dev, struct mbuf **m_head, uint16_t min_frame_size)
3201 {
3202 	/*
3203 	 * 18 is enough bytes to pad an ARP packet to 46 bytes, and
3204 	 * and ARP message is the smallest common payload I can think of
3205 	 */
3206 	static char pad[18];	/* just zeros */
3207 	int n;
3208 	struct mbuf *new_head;
3209 
3210 	if (!M_WRITABLE(*m_head)) {
3211 		new_head = m_dup(*m_head, M_NOWAIT);
3212 		if (new_head == NULL) {
3213 			m_freem(*m_head);
3214 			device_printf(dev, "cannot pad short frame, m_dup() failed");
3215 			DBG_COUNTER_INC(encap_pad_mbuf_fail);
3216 			DBG_COUNTER_INC(tx_frees);
3217 			return ENOMEM;
3218 		}
3219 		m_freem(*m_head);
3220 		*m_head = new_head;
3221 	}
3222 
3223 	for (n = min_frame_size - (*m_head)->m_pkthdr.len;
3224 	     n > 0; n -= sizeof(pad))
3225 		if (!m_append(*m_head, min(n, sizeof(pad)), pad))
3226 			break;
3227 
3228 	if (n > 0) {
3229 		m_freem(*m_head);
3230 		device_printf(dev, "cannot pad short frame\n");
3231 		DBG_COUNTER_INC(encap_pad_mbuf_fail);
3232 		DBG_COUNTER_INC(tx_frees);
3233 		return (ENOBUFS);
3234 	}
3235 
3236 	return 0;
3237 }
3238 
3239 static int
3240 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp)
3241 {
3242 	if_ctx_t		ctx;
3243 	if_shared_ctx_t		sctx;
3244 	if_softc_ctx_t		scctx;
3245 	bus_dma_tag_t		buf_tag;
3246 	bus_dma_segment_t	*segs;
3247 	struct mbuf		*m_head, **ifsd_m;
3248 	void			*next_txd;
3249 	bus_dmamap_t		map;
3250 	struct if_pkt_info	pi;
3251 	int remap = 0;
3252 	int err, nsegs, ndesc, max_segs, pidx, cidx, next, ntxd;
3253 
3254 	ctx = txq->ift_ctx;
3255 	sctx = ctx->ifc_sctx;
3256 	scctx = &ctx->ifc_softc_ctx;
3257 	segs = txq->ift_segs;
3258 	ntxd = txq->ift_size;
3259 	m_head = *m_headp;
3260 	map = NULL;
3261 
3262 	/*
3263 	 * If we're doing TSO the next descriptor to clean may be quite far ahead
3264 	 */
3265 	cidx = txq->ift_cidx;
3266 	pidx = txq->ift_pidx;
3267 	if (ctx->ifc_flags & IFC_PREFETCH) {
3268 		next = (cidx + CACHE_PTR_INCREMENT) & (ntxd-1);
3269 		if (!(ctx->ifc_flags & IFLIB_HAS_TXCQ)) {
3270 			next_txd = calc_next_txd(txq, cidx, 0);
3271 			prefetch(next_txd);
3272 		}
3273 
3274 		/* prefetch the next cache line of mbuf pointers and flags */
3275 		prefetch(&txq->ift_sds.ifsd_m[next]);
3276 		prefetch(&txq->ift_sds.ifsd_map[next]);
3277 		next = (cidx + CACHE_LINE_SIZE) & (ntxd-1);
3278 	}
3279 	map = txq->ift_sds.ifsd_map[pidx];
3280 	ifsd_m = txq->ift_sds.ifsd_m;
3281 
3282 	if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
3283 		buf_tag = txq->ift_tso_buf_tag;
3284 		max_segs = scctx->isc_tx_tso_segments_max;
3285 		map = txq->ift_sds.ifsd_tso_map[pidx];
3286 		MPASS(buf_tag != NULL);
3287 		MPASS(max_segs > 0);
3288 	} else {
3289 		buf_tag = txq->ift_buf_tag;
3290 		max_segs = scctx->isc_tx_nsegments;
3291 		map = txq->ift_sds.ifsd_map[pidx];
3292 	}
3293 	if ((sctx->isc_flags & IFLIB_NEED_ETHER_PAD) &&
3294 	    __predict_false(m_head->m_pkthdr.len < scctx->isc_min_frame_size)) {
3295 		err = iflib_ether_pad(ctx->ifc_dev, m_headp, scctx->isc_min_frame_size);
3296 		if (err) {
3297 			DBG_COUNTER_INC(encap_txd_encap_fail);
3298 			return err;
3299 		}
3300 	}
3301 	m_head = *m_headp;
3302 
3303 	pkt_info_zero(&pi);
3304 	pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG|M_BCAST|M_MCAST));
3305 	pi.ipi_pidx = pidx;
3306 	pi.ipi_qsidx = txq->ift_id;
3307 	pi.ipi_len = m_head->m_pkthdr.len;
3308 	pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags;
3309 	pi.ipi_vtag = M_HAS_VLANTAG(m_head) ? m_head->m_pkthdr.ether_vtag : 0;
3310 
3311 	/* deliberate bitwise OR to make one condition */
3312 	if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) {
3313 		if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0)) {
3314 			DBG_COUNTER_INC(encap_txd_encap_fail);
3315 			return (err);
3316 		}
3317 		m_head = *m_headp;
3318 	}
3319 
3320 retry:
3321 	err = bus_dmamap_load_mbuf_sg(buf_tag, map, m_head, segs, &nsegs,
3322 	    BUS_DMA_NOWAIT);
3323 defrag:
3324 	if (__predict_false(err)) {
3325 		switch (err) {
3326 		case EFBIG:
3327 			/* try collapse once and defrag once */
3328 			if (remap == 0) {
3329 				m_head = m_collapse(*m_headp, M_NOWAIT, max_segs);
3330 				/* try defrag if collapsing fails */
3331 				if (m_head == NULL)
3332 					remap++;
3333 			}
3334 			if (remap == 1) {
3335 				txq->ift_mbuf_defrag++;
3336 				m_head = m_defrag(*m_headp, M_NOWAIT);
3337 			}
3338 			/*
3339 			 * remap should never be >1 unless bus_dmamap_load_mbuf_sg
3340 			 * failed to map an mbuf that was run through m_defrag
3341 			 */
3342 			MPASS(remap <= 1);
3343 			if (__predict_false(m_head == NULL || remap > 1))
3344 				goto defrag_failed;
3345 			remap++;
3346 			*m_headp = m_head;
3347 			goto retry;
3348 			break;
3349 		case ENOMEM:
3350 			txq->ift_no_tx_dma_setup++;
3351 			break;
3352 		default:
3353 			txq->ift_no_tx_dma_setup++;
3354 			m_freem(*m_headp);
3355 			DBG_COUNTER_INC(tx_frees);
3356 			*m_headp = NULL;
3357 			break;
3358 		}
3359 		txq->ift_map_failed++;
3360 		DBG_COUNTER_INC(encap_load_mbuf_fail);
3361 		DBG_COUNTER_INC(encap_txd_encap_fail);
3362 		return (err);
3363 	}
3364 	ifsd_m[pidx] = m_head;
3365 	/*
3366 	 * XXX assumes a 1 to 1 relationship between segments and
3367 	 *        descriptors - this does not hold true on all drivers, e.g.
3368 	 *        cxgb
3369 	 */
3370 	if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) {
3371 		txq->ift_no_desc_avail++;
3372 		bus_dmamap_unload(buf_tag, map);
3373 		DBG_COUNTER_INC(encap_txq_avail_fail);
3374 		DBG_COUNTER_INC(encap_txd_encap_fail);
3375 		if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0)
3376 			GROUPTASK_ENQUEUE(&txq->ift_task);
3377 		return (ENOBUFS);
3378 	}
3379 	/*
3380 	 * On Intel cards we can greatly reduce the number of TX interrupts
3381 	 * we see by only setting report status on every Nth descriptor.
3382 	 * However, this also means that the driver will need to keep track
3383 	 * of the descriptors that RS was set on to check them for the DD bit.
3384 	 */
3385 	txq->ift_rs_pending += nsegs + 1;
3386 	if (txq->ift_rs_pending > TXQ_MAX_RS_DEFERRED(txq) ||
3387 	     iflib_no_tx_batch || (TXQ_AVAIL(txq) - nsegs) <= MAX_TX_DESC(ctx) + 2) {
3388 		pi.ipi_flags |= IPI_TX_INTR;
3389 		txq->ift_rs_pending = 0;
3390 	}
3391 
3392 	pi.ipi_segs = segs;
3393 	pi.ipi_nsegs = nsegs;
3394 
3395 	MPASS(pidx >= 0 && pidx < txq->ift_size);
3396 #ifdef PKT_DEBUG
3397 	print_pkt(&pi);
3398 #endif
3399 	if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) {
3400 		bus_dmamap_sync(buf_tag, map, BUS_DMASYNC_PREWRITE);
3401 		DBG_COUNTER_INC(tx_encap);
3402 		MPASS(pi.ipi_new_pidx < txq->ift_size);
3403 
3404 		ndesc = pi.ipi_new_pidx - pi.ipi_pidx;
3405 		if (pi.ipi_new_pidx < pi.ipi_pidx) {
3406 			ndesc += txq->ift_size;
3407 			txq->ift_gen = 1;
3408 		}
3409 		/*
3410 		 * drivers can need as many as
3411 		 * two sentinels
3412 		 */
3413 		MPASS(ndesc <= pi.ipi_nsegs + 2);
3414 		MPASS(pi.ipi_new_pidx != pidx);
3415 		MPASS(ndesc > 0);
3416 		txq->ift_in_use += ndesc;
3417 
3418 		/*
3419 		 * We update the last software descriptor again here because there may
3420 		 * be a sentinel and/or there may be more mbufs than segments
3421 		 */
3422 		txq->ift_pidx = pi.ipi_new_pidx;
3423 		txq->ift_npending += pi.ipi_ndescs;
3424 	} else {
3425 		*m_headp = m_head = iflib_remove_mbuf(txq);
3426 		if (err == EFBIG) {
3427 			txq->ift_txd_encap_efbig++;
3428 			if (remap < 2) {
3429 				remap = 1;
3430 				goto defrag;
3431 			}
3432 		}
3433 		goto defrag_failed;
3434 	}
3435 	/*
3436 	 * err can't possibly be non-zero here, so we don't neet to test it
3437 	 * to see if we need to DBG_COUNTER_INC(encap_txd_encap_fail).
3438 	 */
3439 	return (err);
3440 
3441 defrag_failed:
3442 	txq->ift_mbuf_defrag_failed++;
3443 	txq->ift_map_failed++;
3444 	m_freem(*m_headp);
3445 	DBG_COUNTER_INC(tx_frees);
3446 	*m_headp = NULL;
3447 	DBG_COUNTER_INC(encap_txd_encap_fail);
3448 	return (ENOMEM);
3449 }
3450 
3451 static void
3452 iflib_tx_desc_free(iflib_txq_t txq, int n)
3453 {
3454 	uint32_t qsize, cidx, mask, gen;
3455 	struct mbuf *m, **ifsd_m;
3456 	bool do_prefetch;
3457 
3458 	cidx = txq->ift_cidx;
3459 	gen = txq->ift_gen;
3460 	qsize = txq->ift_size;
3461 	mask = qsize-1;
3462 	ifsd_m = txq->ift_sds.ifsd_m;
3463 	do_prefetch = (txq->ift_ctx->ifc_flags & IFC_PREFETCH);
3464 
3465 	while (n-- > 0) {
3466 		if (do_prefetch) {
3467 			prefetch(ifsd_m[(cidx + 3) & mask]);
3468 			prefetch(ifsd_m[(cidx + 4) & mask]);
3469 		}
3470 		if ((m = ifsd_m[cidx]) != NULL) {
3471 			prefetch(&ifsd_m[(cidx + CACHE_PTR_INCREMENT) & mask]);
3472 			if (m->m_pkthdr.csum_flags & CSUM_TSO) {
3473 				bus_dmamap_sync(txq->ift_tso_buf_tag,
3474 				    txq->ift_sds.ifsd_tso_map[cidx],
3475 				    BUS_DMASYNC_POSTWRITE);
3476 				bus_dmamap_unload(txq->ift_tso_buf_tag,
3477 				    txq->ift_sds.ifsd_tso_map[cidx]);
3478 			} else {
3479 				bus_dmamap_sync(txq->ift_buf_tag,
3480 				    txq->ift_sds.ifsd_map[cidx],
3481 				    BUS_DMASYNC_POSTWRITE);
3482 				bus_dmamap_unload(txq->ift_buf_tag,
3483 				    txq->ift_sds.ifsd_map[cidx]);
3484 			}
3485 			/* XXX we don't support any drivers that batch packets yet */
3486 			MPASS(m->m_nextpkt == NULL);
3487 			m_freem(m);
3488 			ifsd_m[cidx] = NULL;
3489 #if MEMORY_LOGGING
3490 			txq->ift_dequeued++;
3491 #endif
3492 			DBG_COUNTER_INC(tx_frees);
3493 		}
3494 		if (__predict_false(++cidx == qsize)) {
3495 			cidx = 0;
3496 			gen = 0;
3497 		}
3498 	}
3499 	txq->ift_cidx = cidx;
3500 	txq->ift_gen = gen;
3501 }
3502 
3503 static __inline int
3504 iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh)
3505 {
3506 	int reclaim;
3507 	if_ctx_t ctx = txq->ift_ctx;
3508 
3509 	KASSERT(thresh >= 0, ("invalid threshold to reclaim"));
3510 	MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size);
3511 
3512 	/*
3513 	 * Need a rate-limiting check so that this isn't called every time
3514 	 */
3515 	iflib_tx_credits_update(ctx, txq);
3516 	reclaim = DESC_RECLAIMABLE(txq);
3517 
3518 	if (reclaim <= thresh /* + MAX_TX_DESC(txq->ift_ctx) */) {
3519 #ifdef INVARIANTS
3520 		if (iflib_verbose_debug) {
3521 			printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __FUNCTION__,
3522 			       txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments,
3523 			       reclaim, thresh);
3524 
3525 		}
3526 #endif
3527 		return (0);
3528 	}
3529 	iflib_tx_desc_free(txq, reclaim);
3530 	txq->ift_cleaned += reclaim;
3531 	txq->ift_in_use -= reclaim;
3532 
3533 	return (reclaim);
3534 }
3535 
3536 static struct mbuf **
3537 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset, int remaining)
3538 {
3539 	int next, size;
3540 	struct mbuf **items;
3541 
3542 	size = r->size;
3543 	next = (cidx + CACHE_PTR_INCREMENT) & (size-1);
3544 	items = __DEVOLATILE(struct mbuf **, &r->items[0]);
3545 
3546 	prefetch(items[(cidx + offset) & (size-1)]);
3547 	if (remaining > 1) {
3548 		prefetch2cachelines(&items[next]);
3549 		prefetch2cachelines(items[(cidx + offset + 1) & (size-1)]);
3550 		prefetch2cachelines(items[(cidx + offset + 2) & (size-1)]);
3551 		prefetch2cachelines(items[(cidx + offset + 3) & (size-1)]);
3552 	}
3553 	return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (size-1)]));
3554 }
3555 
3556 static void
3557 iflib_txq_check_drain(iflib_txq_t txq, int budget)
3558 {
3559 
3560 	ifmp_ring_check_drainage(txq->ift_br, budget);
3561 }
3562 
3563 static uint32_t
3564 iflib_txq_can_drain(struct ifmp_ring *r)
3565 {
3566 	iflib_txq_t txq = r->cookie;
3567 	if_ctx_t ctx = txq->ift_ctx;
3568 
3569 	if (TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2)
3570 		return (1);
3571 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3572 	    BUS_DMASYNC_POSTREAD);
3573 	return (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id,
3574 	    false));
3575 }
3576 
3577 static uint32_t
3578 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3579 {
3580 	iflib_txq_t txq = r->cookie;
3581 	if_ctx_t ctx = txq->ift_ctx;
3582 	if_t ifp = ctx->ifc_ifp;
3583 	struct mbuf *m, **mp;
3584 	int avail, bytes_sent, consumed, count, err, i, in_use_prev;
3585 	int mcast_sent, pkt_sent, reclaimed, txq_avail;
3586 	bool do_prefetch, rang, ring;
3587 
3588 	if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) ||
3589 			    !LINK_ACTIVE(ctx))) {
3590 		DBG_COUNTER_INC(txq_drain_notready);
3591 		return (0);
3592 	}
3593 	reclaimed = iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
3594 	rang = iflib_txd_db_check(ctx, txq, reclaimed, txq->ift_in_use);
3595 	avail = IDXDIFF(pidx, cidx, r->size);
3596 	if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) {
3597 		DBG_COUNTER_INC(txq_drain_flushing);
3598 		for (i = 0; i < avail; i++) {
3599 			if (__predict_true(r->items[(cidx + i) & (r->size-1)] != (void *)txq))
3600 				m_free(r->items[(cidx + i) & (r->size-1)]);
3601 			r->items[(cidx + i) & (r->size-1)] = NULL;
3602 		}
3603 		return (avail);
3604 	}
3605 
3606 	if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3607 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3608 		CALLOUT_LOCK(txq);
3609 		callout_stop(&txq->ift_timer);
3610 		CALLOUT_UNLOCK(txq);
3611 		DBG_COUNTER_INC(txq_drain_oactive);
3612 		return (0);
3613 	}
3614 	if (reclaimed)
3615 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3616 	consumed = mcast_sent = bytes_sent = pkt_sent = 0;
3617 	count = MIN(avail, TX_BATCH_SIZE);
3618 #ifdef INVARIANTS
3619 	if (iflib_verbose_debug)
3620 		printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __FUNCTION__,
3621 		       avail, ctx->ifc_flags, TXQ_AVAIL(txq));
3622 #endif
3623 	do_prefetch = (ctx->ifc_flags & IFC_PREFETCH);
3624 	txq_avail = TXQ_AVAIL(txq);
3625 	err = 0;
3626 	for (i = 0; i < count && txq_avail > MAX_TX_DESC(ctx) + 2; i++) {
3627 		int rem = do_prefetch ? count - i : 0;
3628 
3629 		mp = _ring_peek_one(r, cidx, i, rem);
3630 		MPASS(mp != NULL && *mp != NULL);
3631 		if (__predict_false(*mp == (struct mbuf *)txq)) {
3632 			consumed++;
3633 			continue;
3634 		}
3635 		in_use_prev = txq->ift_in_use;
3636 		err = iflib_encap(txq, mp);
3637 		if (__predict_false(err)) {
3638 			/* no room - bail out */
3639 			if (err == ENOBUFS)
3640 				break;
3641 			consumed++;
3642 			/* we can't send this packet - skip it */
3643 			continue;
3644 		}
3645 		consumed++;
3646 		pkt_sent++;
3647 		m = *mp;
3648 		DBG_COUNTER_INC(tx_sent);
3649 		bytes_sent += m->m_pkthdr.len;
3650 		mcast_sent += !!(m->m_flags & M_MCAST);
3651 		txq_avail = TXQ_AVAIL(txq);
3652 
3653 		txq->ift_db_pending += (txq->ift_in_use - in_use_prev);
3654 		ETHER_BPF_MTAP(ifp, m);
3655 		if (__predict_false(!(ifp->if_drv_flags & IFF_DRV_RUNNING)))
3656 			break;
3657 		rang = iflib_txd_db_check(ctx, txq, false, in_use_prev);
3658 	}
3659 
3660 	/* deliberate use of bitwise or to avoid gratuitous short-circuit */
3661 	ring = rang ? false  : (iflib_min_tx_latency | err) || (TXQ_AVAIL(txq) < MAX_TX_DESC(ctx));
3662 	iflib_txd_db_check(ctx, txq, ring, txq->ift_in_use);
3663 	if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
3664 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
3665 	if (mcast_sent)
3666 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
3667 #ifdef INVARIANTS
3668 	if (iflib_verbose_debug)
3669 		printf("consumed=%d\n", consumed);
3670 #endif
3671 	return (consumed);
3672 }
3673 
3674 static uint32_t
3675 iflib_txq_drain_always(struct ifmp_ring *r)
3676 {
3677 	return (1);
3678 }
3679 
3680 static uint32_t
3681 iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3682 {
3683 	int i, avail;
3684 	struct mbuf **mp;
3685 	iflib_txq_t txq;
3686 
3687 	txq = r->cookie;
3688 
3689 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3690 	CALLOUT_LOCK(txq);
3691 	callout_stop(&txq->ift_timer);
3692 	CALLOUT_UNLOCK(txq);
3693 
3694 	avail = IDXDIFF(pidx, cidx, r->size);
3695 	for (i = 0; i < avail; i++) {
3696 		mp = _ring_peek_one(r, cidx, i, avail - i);
3697 		if (__predict_false(*mp == (struct mbuf *)txq))
3698 			continue;
3699 		m_freem(*mp);
3700 		DBG_COUNTER_INC(tx_frees);
3701 	}
3702 	MPASS(ifmp_ring_is_stalled(r) == 0);
3703 	return (avail);
3704 }
3705 
3706 static void
3707 iflib_ifmp_purge(iflib_txq_t txq)
3708 {
3709 	struct ifmp_ring *r;
3710 
3711 	r = txq->ift_br;
3712 	r->drain = iflib_txq_drain_free;
3713 	r->can_drain = iflib_txq_drain_always;
3714 
3715 	ifmp_ring_check_drainage(r, r->size);
3716 
3717 	r->drain = iflib_txq_drain;
3718 	r->can_drain = iflib_txq_can_drain;
3719 }
3720 
3721 static void
3722 _task_fn_tx(void *context)
3723 {
3724 	iflib_txq_t txq = context;
3725 	if_ctx_t ctx = txq->ift_ctx;
3726 #if defined(ALTQ) || defined(DEV_NETMAP)
3727 	if_t ifp = ctx->ifc_ifp;
3728 #endif
3729 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
3730 
3731 #ifdef IFLIB_DIAGNOSTICS
3732 	txq->ift_cpu_exec_count[curcpu]++;
3733 #endif
3734 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3735 		return;
3736 #ifdef DEV_NETMAP
3737 	if (if_getcapenable(ifp) & IFCAP_NETMAP) {
3738 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3739 		    BUS_DMASYNC_POSTREAD);
3740 		if (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false))
3741 			netmap_tx_irq(ifp, txq->ift_id);
3742 		if (ctx->ifc_flags & IFC_LEGACY)
3743 			IFDI_INTR_ENABLE(ctx);
3744 		else
3745 			IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3746 		return;
3747 	}
3748 #endif
3749 #ifdef ALTQ
3750 	if (ALTQ_IS_ENABLED(&ifp->if_snd))
3751 		iflib_altq_if_start(ifp);
3752 #endif
3753 	if (txq->ift_db_pending)
3754 		ifmp_ring_enqueue(txq->ift_br, (void **)&txq, 1, TX_BATCH_SIZE, abdicate);
3755 	else if (!abdicate)
3756 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3757 	/*
3758 	 * When abdicating, we always need to check drainage, not just when we don't enqueue
3759 	 */
3760 	if (abdicate)
3761 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3762 	if (ctx->ifc_flags & IFC_LEGACY)
3763 		IFDI_INTR_ENABLE(ctx);
3764 	else
3765 		IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3766 }
3767 
3768 static void
3769 _task_fn_rx(void *context)
3770 {
3771 	iflib_rxq_t rxq = context;
3772 	if_ctx_t ctx = rxq->ifr_ctx;
3773 	bool more;
3774 	uint16_t budget;
3775 
3776 #ifdef IFLIB_DIAGNOSTICS
3777 	rxq->ifr_cpu_exec_count[curcpu]++;
3778 #endif
3779 	DBG_COUNTER_INC(task_fn_rxs);
3780 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3781 		return;
3782 	more = true;
3783 #ifdef DEV_NETMAP
3784 	if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP) {
3785 		u_int work = 0;
3786 		if (netmap_rx_irq(ctx->ifc_ifp, rxq->ifr_id, &work)) {
3787 			more = false;
3788 		}
3789 	}
3790 #endif
3791 	budget = ctx->ifc_sysctl_rx_budget;
3792 	if (budget == 0)
3793 		budget = 16;	/* XXX */
3794 	if (more == false || (more = iflib_rxeof(rxq, budget)) == false) {
3795 		if (ctx->ifc_flags & IFC_LEGACY)
3796 			IFDI_INTR_ENABLE(ctx);
3797 		else
3798 			IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
3799 		DBG_COUNTER_INC(rx_intr_enables);
3800 	}
3801 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3802 		return;
3803 	if (more)
3804 		GROUPTASK_ENQUEUE(&rxq->ifr_task);
3805 }
3806 
3807 static void
3808 _task_fn_admin(void *context)
3809 {
3810 	if_ctx_t ctx = context;
3811 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
3812 	iflib_txq_t txq;
3813 	int i;
3814 	bool oactive, running, do_reset, do_watchdog, in_detach;
3815 	uint32_t reset_on = hz / 2;
3816 
3817 	STATE_LOCK(ctx);
3818 	running = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING);
3819 	oactive = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE);
3820 	do_reset = (ctx->ifc_flags & IFC_DO_RESET);
3821 	do_watchdog = (ctx->ifc_flags & IFC_DO_WATCHDOG);
3822 	in_detach = (ctx->ifc_flags & IFC_IN_DETACH);
3823 	ctx->ifc_flags &= ~(IFC_DO_RESET|IFC_DO_WATCHDOG);
3824 	STATE_UNLOCK(ctx);
3825 
3826 	if ((!running && !oactive) && !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
3827 		return;
3828 	if (in_detach)
3829 		return;
3830 
3831 	CTX_LOCK(ctx);
3832 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3833 		CALLOUT_LOCK(txq);
3834 		callout_stop(&txq->ift_timer);
3835 		CALLOUT_UNLOCK(txq);
3836 	}
3837 	if (do_watchdog) {
3838 		ctx->ifc_watchdog_events++;
3839 		IFDI_WATCHDOG_RESET(ctx);
3840 	}
3841 	IFDI_UPDATE_ADMIN_STATUS(ctx);
3842 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3843 #ifdef DEV_NETMAP
3844 		reset_on = hz / 2;
3845 		if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP)
3846 			iflib_netmap_timer_adjust(ctx, txq, &reset_on);
3847 #endif
3848 		callout_reset_on(&txq->ift_timer, reset_on, iflib_timer, txq, txq->ift_timer.c_cpu);
3849 	}
3850 	IFDI_LINK_INTR_ENABLE(ctx);
3851 	if (do_reset)
3852 		iflib_if_init_locked(ctx);
3853 	CTX_UNLOCK(ctx);
3854 
3855 	if (LINK_ACTIVE(ctx) == 0)
3856 		return;
3857 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3858 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
3859 }
3860 
3861 
3862 static void
3863 _task_fn_iov(void *context)
3864 {
3865 	if_ctx_t ctx = context;
3866 
3867 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING) &&
3868 	    !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
3869 		return;
3870 
3871 	CTX_LOCK(ctx);
3872 	IFDI_VFLR_HANDLE(ctx);
3873 	CTX_UNLOCK(ctx);
3874 }
3875 
3876 static int
3877 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)
3878 {
3879 	int err;
3880 	if_int_delay_info_t info;
3881 	if_ctx_t ctx;
3882 
3883 	info = (if_int_delay_info_t)arg1;
3884 	ctx = info->iidi_ctx;
3885 	info->iidi_req = req;
3886 	info->iidi_oidp = oidp;
3887 	CTX_LOCK(ctx);
3888 	err = IFDI_SYSCTL_INT_DELAY(ctx, info);
3889 	CTX_UNLOCK(ctx);
3890 	return (err);
3891 }
3892 
3893 /*********************************************************************
3894  *
3895  *  IFNET FUNCTIONS
3896  *
3897  **********************************************************************/
3898 
3899 static void
3900 iflib_if_init_locked(if_ctx_t ctx)
3901 {
3902 	iflib_stop(ctx);
3903 	iflib_init_locked(ctx);
3904 }
3905 
3906 
3907 static void
3908 iflib_if_init(void *arg)
3909 {
3910 	if_ctx_t ctx = arg;
3911 
3912 	CTX_LOCK(ctx);
3913 	iflib_if_init_locked(ctx);
3914 	CTX_UNLOCK(ctx);
3915 }
3916 
3917 static int
3918 iflib_if_transmit(if_t ifp, struct mbuf *m)
3919 {
3920 	if_ctx_t	ctx = if_getsoftc(ifp);
3921 
3922 	iflib_txq_t txq;
3923 	int err, qidx;
3924 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
3925 
3926 	if (__predict_false((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) {
3927 		DBG_COUNTER_INC(tx_frees);
3928 		m_freem(m);
3929 		return (ENETDOWN);
3930 	}
3931 
3932 	MPASS(m->m_nextpkt == NULL);
3933 	/* ALTQ-enabled interfaces always use queue 0. */
3934 	qidx = 0;
3935 	if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m) && !ALTQ_IS_ENABLED(&ifp->if_snd))
3936 		qidx = QIDX(ctx, m);
3937 	/*
3938 	 * XXX calculate buf_ring based on flowid (divvy up bits?)
3939 	 */
3940 	txq = &ctx->ifc_txqs[qidx];
3941 
3942 #ifdef DRIVER_BACKPRESSURE
3943 	if (txq->ift_closed) {
3944 		while (m != NULL) {
3945 			next = m->m_nextpkt;
3946 			m->m_nextpkt = NULL;
3947 			m_freem(m);
3948 			DBG_COUNTER_INC(tx_frees);
3949 			m = next;
3950 		}
3951 		return (ENOBUFS);
3952 	}
3953 #endif
3954 #ifdef notyet
3955 	qidx = count = 0;
3956 	mp = marr;
3957 	next = m;
3958 	do {
3959 		count++;
3960 		next = next->m_nextpkt;
3961 	} while (next != NULL);
3962 
3963 	if (count > nitems(marr))
3964 		if ((mp = malloc(count*sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) {
3965 			/* XXX check nextpkt */
3966 			m_freem(m);
3967 			/* XXX simplify for now */
3968 			DBG_COUNTER_INC(tx_frees);
3969 			return (ENOBUFS);
3970 		}
3971 	for (next = m, i = 0; next != NULL; i++) {
3972 		mp[i] = next;
3973 		next = next->m_nextpkt;
3974 		mp[i]->m_nextpkt = NULL;
3975 	}
3976 #endif
3977 	DBG_COUNTER_INC(tx_seen);
3978 	err = ifmp_ring_enqueue(txq->ift_br, (void **)&m, 1, TX_BATCH_SIZE, abdicate);
3979 
3980 	if (abdicate)
3981 		GROUPTASK_ENQUEUE(&txq->ift_task);
3982  	if (err) {
3983 		if (!abdicate)
3984 			GROUPTASK_ENQUEUE(&txq->ift_task);
3985 		/* support forthcoming later */
3986 #ifdef DRIVER_BACKPRESSURE
3987 		txq->ift_closed = TRUE;
3988 #endif
3989 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3990 		m_freem(m);
3991 		DBG_COUNTER_INC(tx_frees);
3992 	}
3993 
3994 	return (err);
3995 }
3996 
3997 #ifdef ALTQ
3998 /*
3999  * The overall approach to integrating iflib with ALTQ is to continue to use
4000  * the iflib mp_ring machinery between the ALTQ queue(s) and the hardware
4001  * ring.  Technically, when using ALTQ, queueing to an intermediate mp_ring
4002  * is redundant/unnecessary, but doing so minimizes the amount of
4003  * ALTQ-specific code required in iflib.  It is assumed that the overhead of
4004  * redundantly queueing to an intermediate mp_ring is swamped by the
4005  * performance limitations inherent in using ALTQ.
4006  *
4007  * When ALTQ support is compiled in, all iflib drivers will use a transmit
4008  * routine, iflib_altq_if_transmit(), that checks if ALTQ is enabled for the
4009  * given interface.  If ALTQ is enabled for an interface, then all
4010  * transmitted packets for that interface will be submitted to the ALTQ
4011  * subsystem via IFQ_ENQUEUE().  We don't use the legacy if_transmit()
4012  * implementation because it uses IFQ_HANDOFF(), which will duplicatively
4013  * update stats that the iflib machinery handles, and which is sensitve to
4014  * the disused IFF_DRV_OACTIVE flag.  Additionally, iflib_altq_if_start()
4015  * will be installed as the start routine for use by ALTQ facilities that
4016  * need to trigger queue drains on a scheduled basis.
4017  *
4018  */
4019 static void
4020 iflib_altq_if_start(if_t ifp)
4021 {
4022 	struct ifaltq *ifq = &ifp->if_snd;
4023 	struct mbuf *m;
4024 
4025 	IFQ_LOCK(ifq);
4026 	IFQ_DEQUEUE_NOLOCK(ifq, m);
4027 	while (m != NULL) {
4028 		iflib_if_transmit(ifp, m);
4029 		IFQ_DEQUEUE_NOLOCK(ifq, m);
4030 	}
4031 	IFQ_UNLOCK(ifq);
4032 }
4033 
4034 static int
4035 iflib_altq_if_transmit(if_t ifp, struct mbuf *m)
4036 {
4037 	int err;
4038 
4039 	if (ALTQ_IS_ENABLED(&ifp->if_snd)) {
4040 		IFQ_ENQUEUE(&ifp->if_snd, m, err);
4041 		if (err == 0)
4042 			iflib_altq_if_start(ifp);
4043 	} else
4044 		err = iflib_if_transmit(ifp, m);
4045 
4046 	return (err);
4047 }
4048 #endif /* ALTQ */
4049 
4050 static void
4051 iflib_if_qflush(if_t ifp)
4052 {
4053 	if_ctx_t ctx = if_getsoftc(ifp);
4054 	iflib_txq_t txq = ctx->ifc_txqs;
4055 	int i;
4056 
4057 	STATE_LOCK(ctx);
4058 	ctx->ifc_flags |= IFC_QFLUSH;
4059 	STATE_UNLOCK(ctx);
4060 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
4061 		while (!(ifmp_ring_is_idle(txq->ift_br) || ifmp_ring_is_stalled(txq->ift_br)))
4062 			iflib_txq_check_drain(txq, 0);
4063 	STATE_LOCK(ctx);
4064 	ctx->ifc_flags &= ~IFC_QFLUSH;
4065 	STATE_UNLOCK(ctx);
4066 
4067 	/*
4068 	 * When ALTQ is enabled, this will also take care of purging the
4069 	 * ALTQ queue(s).
4070 	 */
4071 	if_qflush(ifp);
4072 }
4073 
4074 
4075 #define IFCAP_FLAGS (IFCAP_HWCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \
4076 		     IFCAP_TSO | IFCAP_VLAN_HWTAGGING | IFCAP_HWSTATS | \
4077 		     IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | \
4078 		     IFCAP_VLAN_HWTSO | IFCAP_VLAN_HWCSUM)
4079 
4080 static int
4081 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data)
4082 {
4083 	if_ctx_t ctx = if_getsoftc(ifp);
4084 	struct ifreq	*ifr = (struct ifreq *)data;
4085 #if defined(INET) || defined(INET6)
4086 	struct ifaddr	*ifa = (struct ifaddr *)data;
4087 #endif
4088 	bool		avoid_reset = false;
4089 	int		err = 0, reinit = 0, bits;
4090 
4091 	switch (command) {
4092 	case SIOCSIFADDR:
4093 #ifdef INET
4094 		if (ifa->ifa_addr->sa_family == AF_INET)
4095 			avoid_reset = true;
4096 #endif
4097 #ifdef INET6
4098 		if (ifa->ifa_addr->sa_family == AF_INET6)
4099 			avoid_reset = true;
4100 #endif
4101 		/*
4102 		** Calling init results in link renegotiation,
4103 		** so we avoid doing it when possible.
4104 		*/
4105 		if (avoid_reset) {
4106 			if_setflagbits(ifp, IFF_UP,0);
4107 			if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
4108 				reinit = 1;
4109 #ifdef INET
4110 			if (!(if_getflags(ifp) & IFF_NOARP))
4111 				arp_ifinit(ifp, ifa);
4112 #endif
4113 		} else
4114 			err = ether_ioctl(ifp, command, data);
4115 		break;
4116 	case SIOCSIFMTU:
4117 		CTX_LOCK(ctx);
4118 		if (ifr->ifr_mtu == if_getmtu(ifp)) {
4119 			CTX_UNLOCK(ctx);
4120 			break;
4121 		}
4122 		bits = if_getdrvflags(ifp);
4123 		/* stop the driver and free any clusters before proceeding */
4124 		iflib_stop(ctx);
4125 
4126 		if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) {
4127 			STATE_LOCK(ctx);
4128 			if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size)
4129 				ctx->ifc_flags |= IFC_MULTISEG;
4130 			else
4131 				ctx->ifc_flags &= ~IFC_MULTISEG;
4132 			STATE_UNLOCK(ctx);
4133 			err = if_setmtu(ifp, ifr->ifr_mtu);
4134 		}
4135 		iflib_init_locked(ctx);
4136 		STATE_LOCK(ctx);
4137 		if_setdrvflags(ifp, bits);
4138 		STATE_UNLOCK(ctx);
4139 		CTX_UNLOCK(ctx);
4140 		break;
4141 	case SIOCSIFFLAGS:
4142 		CTX_LOCK(ctx);
4143 		if (if_getflags(ifp) & IFF_UP) {
4144 			if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4145 				if ((if_getflags(ifp) ^ ctx->ifc_if_flags) &
4146 				    (IFF_PROMISC | IFF_ALLMULTI)) {
4147 					err = IFDI_PROMISC_SET(ctx, if_getflags(ifp));
4148 				}
4149 			} else
4150 				reinit = 1;
4151 		} else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4152 			iflib_stop(ctx);
4153 		}
4154 		ctx->ifc_if_flags = if_getflags(ifp);
4155 		CTX_UNLOCK(ctx);
4156 		break;
4157 	case SIOCADDMULTI:
4158 	case SIOCDELMULTI:
4159 		if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4160 			CTX_LOCK(ctx);
4161 			IFDI_INTR_DISABLE(ctx);
4162 			IFDI_MULTI_SET(ctx);
4163 			IFDI_INTR_ENABLE(ctx);
4164 			CTX_UNLOCK(ctx);
4165 		}
4166 		break;
4167 	case SIOCSIFMEDIA:
4168 		CTX_LOCK(ctx);
4169 		IFDI_MEDIA_SET(ctx);
4170 		CTX_UNLOCK(ctx);
4171 		/* FALLTHROUGH */
4172 	case SIOCGIFMEDIA:
4173 	case SIOCGIFXMEDIA:
4174 		err = ifmedia_ioctl(ifp, ifr, ctx->ifc_mediap, command);
4175 		break;
4176 	case SIOCGI2C:
4177 	{
4178 		struct ifi2creq i2c;
4179 
4180 		err = copyin(ifr_data_get_ptr(ifr), &i2c, sizeof(i2c));
4181 		if (err != 0)
4182 			break;
4183 		if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) {
4184 			err = EINVAL;
4185 			break;
4186 		}
4187 		if (i2c.len > sizeof(i2c.data)) {
4188 			err = EINVAL;
4189 			break;
4190 		}
4191 
4192 		if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0)
4193 			err = copyout(&i2c, ifr_data_get_ptr(ifr),
4194 			    sizeof(i2c));
4195 		break;
4196 	}
4197 	case SIOCSIFCAP:
4198 	{
4199 		int mask, setmask, oldmask;
4200 
4201 		oldmask = if_getcapenable(ifp);
4202 		mask = ifr->ifr_reqcap ^ oldmask;
4203 		mask &= ctx->ifc_softc_ctx.isc_capabilities;
4204 		setmask = 0;
4205 #ifdef TCP_OFFLOAD
4206 		setmask |= mask & (IFCAP_TOE4|IFCAP_TOE6);
4207 #endif
4208 		setmask |= (mask & IFCAP_FLAGS);
4209 		setmask |= (mask & IFCAP_WOL);
4210 
4211 		/*
4212 		 * If any RX csum has changed, change all the ones that
4213 		 * are supported by the driver.
4214 		 */
4215 		if (setmask & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6)) {
4216 			setmask |= ctx->ifc_softc_ctx.isc_capabilities &
4217 			    (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6);
4218 		}
4219 
4220 		/*
4221 		 * want to ensure that traffic has stopped before we change any of the flags
4222 		 */
4223 		if (setmask) {
4224 			CTX_LOCK(ctx);
4225 			bits = if_getdrvflags(ifp);
4226 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4227 				iflib_stop(ctx);
4228 			STATE_LOCK(ctx);
4229 			if_togglecapenable(ifp, setmask);
4230 			STATE_UNLOCK(ctx);
4231 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4232 				iflib_init_locked(ctx);
4233 			STATE_LOCK(ctx);
4234 			if_setdrvflags(ifp, bits);
4235 			STATE_UNLOCK(ctx);
4236 			CTX_UNLOCK(ctx);
4237 		}
4238 		if_vlancap(ifp);
4239 		break;
4240 	}
4241 	case SIOCGPRIVATE_0:
4242 	case SIOCSDRVSPEC:
4243 	case SIOCGDRVSPEC:
4244 		CTX_LOCK(ctx);
4245 		err = IFDI_PRIV_IOCTL(ctx, command, data);
4246 		CTX_UNLOCK(ctx);
4247 		break;
4248 	default:
4249 		err = ether_ioctl(ifp, command, data);
4250 		break;
4251 	}
4252 	if (reinit)
4253 		iflib_if_init(ctx);
4254 	return (err);
4255 }
4256 
4257 static uint64_t
4258 iflib_if_get_counter(if_t ifp, ift_counter cnt)
4259 {
4260 	if_ctx_t ctx = if_getsoftc(ifp);
4261 
4262 	return (IFDI_GET_COUNTER(ctx, cnt));
4263 }
4264 
4265 /*********************************************************************
4266  *
4267  *  OTHER FUNCTIONS EXPORTED TO THE STACK
4268  *
4269  **********************************************************************/
4270 
4271 static void
4272 iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag)
4273 {
4274 	if_ctx_t ctx = if_getsoftc(ifp);
4275 
4276 	if ((void *)ctx != arg)
4277 		return;
4278 
4279 	if ((vtag == 0) || (vtag > 4095))
4280 		return;
4281 
4282 	CTX_LOCK(ctx);
4283 	IFDI_VLAN_REGISTER(ctx, vtag);
4284 	/* Re-init to load the changes */
4285 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
4286 		iflib_if_init_locked(ctx);
4287 	CTX_UNLOCK(ctx);
4288 }
4289 
4290 static void
4291 iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag)
4292 {
4293 	if_ctx_t ctx = if_getsoftc(ifp);
4294 
4295 	if ((void *)ctx != arg)
4296 		return;
4297 
4298 	if ((vtag == 0) || (vtag > 4095))
4299 		return;
4300 
4301 	CTX_LOCK(ctx);
4302 	IFDI_VLAN_UNREGISTER(ctx, vtag);
4303 	/* Re-init to load the changes */
4304 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
4305 		iflib_if_init_locked(ctx);
4306 	CTX_UNLOCK(ctx);
4307 }
4308 
4309 static void
4310 iflib_led_func(void *arg, int onoff)
4311 {
4312 	if_ctx_t ctx = arg;
4313 
4314 	CTX_LOCK(ctx);
4315 	IFDI_LED_FUNC(ctx, onoff);
4316 	CTX_UNLOCK(ctx);
4317 }
4318 
4319 /*********************************************************************
4320  *
4321  *  BUS FUNCTION DEFINITIONS
4322  *
4323  **********************************************************************/
4324 
4325 int
4326 iflib_device_probe(device_t dev)
4327 {
4328 	const pci_vendor_info_t *ent;
4329 	if_shared_ctx_t sctx;
4330 	uint16_t pci_device_id, pci_rev_id, pci_subdevice_id, pci_subvendor_id;
4331 	uint16_t pci_vendor_id;
4332 
4333 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4334 		return (ENOTSUP);
4335 
4336 	pci_vendor_id = pci_get_vendor(dev);
4337 	pci_device_id = pci_get_device(dev);
4338 	pci_subvendor_id = pci_get_subvendor(dev);
4339 	pci_subdevice_id = pci_get_subdevice(dev);
4340 	pci_rev_id = pci_get_revid(dev);
4341 	if (sctx->isc_parse_devinfo != NULL)
4342 		sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id);
4343 
4344 	ent = sctx->isc_vendor_info;
4345 	while (ent->pvi_vendor_id != 0) {
4346 		if (pci_vendor_id != ent->pvi_vendor_id) {
4347 			ent++;
4348 			continue;
4349 		}
4350 		if ((pci_device_id == ent->pvi_device_id) &&
4351 		    ((pci_subvendor_id == ent->pvi_subvendor_id) ||
4352 		     (ent->pvi_subvendor_id == 0)) &&
4353 		    ((pci_subdevice_id == ent->pvi_subdevice_id) ||
4354 		     (ent->pvi_subdevice_id == 0)) &&
4355 		    ((pci_rev_id == ent->pvi_rev_id) ||
4356 		     (ent->pvi_rev_id == 0))) {
4357 
4358 			device_set_desc_copy(dev, ent->pvi_name);
4359 			/* this needs to be changed to zero if the bus probing code
4360 			 * ever stops re-probing on best match because the sctx
4361 			 * may have its values over written by register calls
4362 			 * in subsequent probes
4363 			 */
4364 			return (BUS_PROBE_DEFAULT);
4365 		}
4366 		ent++;
4367 	}
4368 	return (ENXIO);
4369 }
4370 
4371 int
4372 iflib_device_probe_vendor(device_t dev)
4373 {
4374 	int probe;
4375 
4376 	probe = iflib_device_probe(dev);
4377 	if (probe == BUS_PROBE_DEFAULT)
4378 		return (BUS_PROBE_VENDOR);
4379 	else
4380 		return (probe);
4381 }
4382 
4383 static void
4384 iflib_reset_qvalues(if_ctx_t ctx)
4385 {
4386 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4387 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4388 	device_t dev = ctx->ifc_dev;
4389 	int i;
4390 
4391 	if (ctx->ifc_sysctl_ntxqs != 0)
4392 		scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs;
4393 	if (ctx->ifc_sysctl_nrxqs != 0)
4394 		scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs;
4395 
4396 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4397 		if (ctx->ifc_sysctl_ntxds[i] != 0)
4398 			scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i];
4399 		else
4400 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4401 	}
4402 
4403 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4404 		if (ctx->ifc_sysctl_nrxds[i] != 0)
4405 			scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i];
4406 		else
4407 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4408 	}
4409 
4410 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4411 		if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) {
4412 			device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n",
4413 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]);
4414 			scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i];
4415 		}
4416 		if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) {
4417 			device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n",
4418 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]);
4419 			scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i];
4420 		}
4421 		if (!powerof2(scctx->isc_nrxd[i])) {
4422 			device_printf(dev, "nrxd%d: %d is not a power of 2 - using default value of %d\n",
4423 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_default[i]);
4424 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4425 		}
4426 	}
4427 
4428 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4429 		if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) {
4430 			device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n",
4431 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]);
4432 			scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i];
4433 		}
4434 		if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) {
4435 			device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n",
4436 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]);
4437 			scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i];
4438 		}
4439 		if (!powerof2(scctx->isc_ntxd[i])) {
4440 			device_printf(dev, "ntxd%d: %d is not a power of 2 - using default value of %d\n",
4441 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_default[i]);
4442 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4443 		}
4444 	}
4445 }
4446 
4447 static void
4448 iflib_add_pfil(if_ctx_t ctx)
4449 {
4450 	struct pfil_head *pfil;
4451 	struct pfil_head_args pa;
4452 	iflib_rxq_t rxq;
4453 	int i;
4454 
4455 	pa.pa_version = PFIL_VERSION;
4456 	pa.pa_flags = PFIL_IN;
4457 	pa.pa_type = PFIL_TYPE_ETHERNET;
4458 	pa.pa_headname = ctx->ifc_ifp->if_xname;
4459 	pfil = pfil_head_register(&pa);
4460 
4461 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4462 		rxq->pfil = pfil;
4463 	}
4464 }
4465 
4466 static void
4467 iflib_rem_pfil(if_ctx_t ctx)
4468 {
4469 	struct pfil_head *pfil;
4470 	iflib_rxq_t rxq;
4471 	int i;
4472 
4473 	rxq = ctx->ifc_rxqs;
4474 	pfil = rxq->pfil;
4475 	for (i = 0; i < NRXQSETS(ctx); i++, rxq++) {
4476 		rxq->pfil = NULL;
4477 	}
4478 	pfil_head_unregister(pfil);
4479 }
4480 
4481 static uint16_t
4482 get_ctx_core_offset(if_ctx_t ctx)
4483 {
4484 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4485 	struct cpu_offset *op;
4486 	uint16_t qc;
4487 	uint16_t ret = ctx->ifc_sysctl_core_offset;
4488 
4489 	if (ret != CORE_OFFSET_UNSPECIFIED)
4490 		return (ret);
4491 
4492 	if (ctx->ifc_sysctl_separate_txrx)
4493 		qc = scctx->isc_ntxqsets + scctx->isc_nrxqsets;
4494 	else
4495 		qc = max(scctx->isc_ntxqsets, scctx->isc_nrxqsets);
4496 
4497 	mtx_lock(&cpu_offset_mtx);
4498 	SLIST_FOREACH(op, &cpu_offsets, entries) {
4499 		if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) {
4500 			ret = op->offset;
4501 			op->offset += qc;
4502 			MPASS(op->refcount < UINT_MAX);
4503 			op->refcount++;
4504 			break;
4505 		}
4506 	}
4507 	if (ret == CORE_OFFSET_UNSPECIFIED) {
4508 		ret = 0;
4509 		op = malloc(sizeof(struct cpu_offset), M_IFLIB,
4510 		    M_NOWAIT | M_ZERO);
4511 		if (op == NULL) {
4512 			device_printf(ctx->ifc_dev,
4513 			    "allocation for cpu offset failed.\n");
4514 		} else {
4515 			op->offset = qc;
4516 			op->refcount = 1;
4517 			CPU_COPY(&ctx->ifc_cpus, &op->set);
4518 			SLIST_INSERT_HEAD(&cpu_offsets, op, entries);
4519 		}
4520 	}
4521 	mtx_unlock(&cpu_offset_mtx);
4522 
4523 	return (ret);
4524 }
4525 
4526 static void
4527 unref_ctx_core_offset(if_ctx_t ctx)
4528 {
4529 	struct cpu_offset *op, *top;
4530 
4531 	mtx_lock(&cpu_offset_mtx);
4532 	SLIST_FOREACH_SAFE(op, &cpu_offsets, entries, top) {
4533 		if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) {
4534 			MPASS(op->refcount > 0);
4535 			op->refcount--;
4536 			if (op->refcount == 0) {
4537 				SLIST_REMOVE(&cpu_offsets, op, cpu_offset, entries);
4538 				free(op, M_IFLIB);
4539 			}
4540 			break;
4541 		}
4542 	}
4543 	mtx_unlock(&cpu_offset_mtx);
4544 }
4545 
4546 int
4547 iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp)
4548 {
4549 	if_ctx_t ctx;
4550 	if_t ifp;
4551 	if_softc_ctx_t scctx;
4552 	kobjop_desc_t kobj_desc;
4553 	kobj_method_t *kobj_method;
4554 	int err, msix, rid;
4555 	uint16_t main_rxq, main_txq;
4556 
4557 	ctx = malloc(sizeof(* ctx), M_IFLIB, M_WAITOK|M_ZERO);
4558 
4559 	if (sc == NULL) {
4560 		sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4561 		device_set_softc(dev, ctx);
4562 		ctx->ifc_flags |= IFC_SC_ALLOCATED;
4563 	}
4564 
4565 	ctx->ifc_sctx = sctx;
4566 	ctx->ifc_dev = dev;
4567 	ctx->ifc_softc = sc;
4568 
4569 	if ((err = iflib_register(ctx)) != 0) {
4570 		device_printf(dev, "iflib_register failed %d\n", err);
4571 		goto fail_ctx_free;
4572 	}
4573 	iflib_add_device_sysctl_pre(ctx);
4574 
4575 	scctx = &ctx->ifc_softc_ctx;
4576 	ifp = ctx->ifc_ifp;
4577 
4578 	iflib_reset_qvalues(ctx);
4579 	CTX_LOCK(ctx);
4580 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4581 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4582 		goto fail_unlock;
4583 	}
4584 	_iflib_pre_assert(scctx);
4585 	ctx->ifc_txrx = *scctx->isc_txrx;
4586 
4587 	if (sctx->isc_flags & IFLIB_DRIVER_MEDIA)
4588 		ctx->ifc_mediap = scctx->isc_media;
4589 
4590 #ifdef INVARIANTS
4591 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
4592 		MPASS(scctx->isc_tx_csum_flags);
4593 #endif
4594 
4595 	if_setcapabilities(ifp, scctx->isc_capabilities | IFCAP_HWSTATS);
4596 	if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS);
4597 
4598 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4599 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4600 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4601 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4602 
4603 	main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4604 	main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4605 
4606 	/* XXX change for per-queue sizes */
4607 	device_printf(dev, "Using %d TX descriptors and %d RX descriptors\n",
4608 	    scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4609 
4610 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4611 	    MAX_SINGLE_PACKET_FRACTION)
4612 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4613 		    MAX_SINGLE_PACKET_FRACTION);
4614 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4615 	    MAX_SINGLE_PACKET_FRACTION)
4616 		scctx->isc_tx_tso_segments_max = max(1,
4617 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4618 
4619 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4620 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
4621 		/*
4622 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
4623 		 * but some MACs do.
4624 		 */
4625 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
4626 		    IP_MAXPACKET));
4627 		/*
4628 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
4629 		 * into account.  In the worst case, each of these calls will
4630 		 * add another mbuf and, thus, the requirement for another DMA
4631 		 * segment.  So for best performance, it doesn't make sense to
4632 		 * advertize a maximum of TSO segments that typically will
4633 		 * require defragmentation in iflib_encap().
4634 		 */
4635 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
4636 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
4637 	}
4638 	if (scctx->isc_rss_table_size == 0)
4639 		scctx->isc_rss_table_size = 64;
4640 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4641 
4642 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4643 	/* XXX format name */
4644 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx,
4645 	    NULL, NULL, "admin");
4646 
4647 	/* Set up cpu set.  If it fails, use the set of all CPUs. */
4648 	if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) != 0) {
4649 		device_printf(dev, "Unable to fetch CPU list\n");
4650 		CPU_COPY(&all_cpus, &ctx->ifc_cpus);
4651 	}
4652 	MPASS(CPU_COUNT(&ctx->ifc_cpus) > 0);
4653 
4654 	/*
4655 	** Now set up MSI or MSI-X, should return us the number of supported
4656 	** vectors (will be 1 for a legacy interrupt and MSI).
4657 	*/
4658 	if (sctx->isc_flags & IFLIB_SKIP_MSIX) {
4659 		msix = scctx->isc_vectors;
4660 	} else if (scctx->isc_msix_bar != 0)
4661 	       /*
4662 		* The simple fact that isc_msix_bar is not 0 does not mean we
4663 		* we have a good value there that is known to work.
4664 		*/
4665 		msix = iflib_msix_init(ctx);
4666 	else {
4667 		scctx->isc_vectors = 1;
4668 		scctx->isc_ntxqsets = 1;
4669 		scctx->isc_nrxqsets = 1;
4670 		scctx->isc_intr = IFLIB_INTR_LEGACY;
4671 		msix = 0;
4672 	}
4673 	/* Get memory for the station queues */
4674 	if ((err = iflib_queues_alloc(ctx))) {
4675 		device_printf(dev, "Unable to allocate queue memory\n");
4676 		goto fail_intr_free;
4677 	}
4678 
4679 	if ((err = iflib_qset_structures_setup(ctx)))
4680 		goto fail_queues;
4681 
4682 	/*
4683 	 * Now that we know how many queues there are, get the core offset.
4684 	 */
4685 	ctx->ifc_sysctl_core_offset = get_ctx_core_offset(ctx);
4686 
4687 	/*
4688 	 * Group taskqueues aren't properly set up until SMP is started,
4689 	 * so we disable interrupts until we can handle them post
4690 	 * SI_SUB_SMP.
4691 	 *
4692 	 * XXX: disabling interrupts doesn't actually work, at least for
4693 	 * the non-MSI case.  When they occur before SI_SUB_SMP completes,
4694 	 * we do null handling and depend on this not causing too large an
4695 	 * interrupt storm.
4696 	 */
4697 	IFDI_INTR_DISABLE(ctx);
4698 
4699 	if (msix > 1) {
4700 		/*
4701 		 * When using MSI-X, ensure that ifdi_{r,t}x_queue_intr_enable
4702 		 * aren't the default NULL implementation.
4703 		 */
4704 		kobj_desc = &ifdi_rx_queue_intr_enable_desc;
4705 		kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL,
4706 		    kobj_desc);
4707 		if (kobj_method == &kobj_desc->deflt) {
4708 			device_printf(dev,
4709 			    "MSI-X requires ifdi_rx_queue_intr_enable method");
4710 			err = EOPNOTSUPP;
4711 			goto fail_queues;
4712 		}
4713 		kobj_desc = &ifdi_tx_queue_intr_enable_desc;
4714 		kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL,
4715 		    kobj_desc);
4716 		if (kobj_method == &kobj_desc->deflt) {
4717 			device_printf(dev,
4718 			    "MSI-X requires ifdi_tx_queue_intr_enable method");
4719 			err = EOPNOTSUPP;
4720 			goto fail_queues;
4721 		}
4722 
4723 		/*
4724 		 * Assign the MSI-X vectors.
4725 		 * Note that the default NULL ifdi_msix_intr_assign method will
4726 		 * fail here, too.
4727 		 */
4728 		err = IFDI_MSIX_INTR_ASSIGN(ctx, msix);
4729 		if (err != 0) {
4730 			device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n",
4731 			    err);
4732 			goto fail_queues;
4733 		}
4734 	} else if (scctx->isc_intr != IFLIB_INTR_MSIX) {
4735 		rid = 0;
4736 		if (scctx->isc_intr == IFLIB_INTR_MSI) {
4737 			MPASS(msix == 1);
4738 			rid = 1;
4739 		}
4740 		if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) {
4741 			device_printf(dev, "iflib_legacy_setup failed %d\n", err);
4742 			goto fail_queues;
4743 		}
4744 	} else {
4745 		device_printf(dev,
4746 		    "Cannot use iflib with only 1 MSI-X interrupt!\n");
4747 		err = ENODEV;
4748 		goto fail_intr_free;
4749 	}
4750 
4751 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
4752 
4753 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4754 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4755 		goto fail_detach;
4756 	}
4757 
4758 	/*
4759 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4760 	 * This must appear after the call to ether_ifattach() because
4761 	 * ether_ifattach() sets if_hdrlen to the default value.
4762 	 */
4763 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4764 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
4765 
4766 	if ((err = iflib_netmap_attach(ctx))) {
4767 		device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err);
4768 		goto fail_detach;
4769 	}
4770 	*ctxp = ctx;
4771 
4772 	NETDUMP_SET(ctx->ifc_ifp, iflib);
4773 
4774 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4775 	iflib_add_device_sysctl_post(ctx);
4776 	iflib_add_pfil(ctx);
4777 	ctx->ifc_flags |= IFC_INIT_DONE;
4778 	CTX_UNLOCK(ctx);
4779 
4780 	return (0);
4781 
4782 fail_detach:
4783 	ether_ifdetach(ctx->ifc_ifp);
4784 fail_intr_free:
4785 	iflib_free_intr_mem(ctx);
4786 fail_queues:
4787 	iflib_tx_structures_free(ctx);
4788 	iflib_rx_structures_free(ctx);
4789 	taskqgroup_detach(qgroup_if_config_tqg, &ctx->ifc_admin_task);
4790 	IFDI_DETACH(ctx);
4791 fail_unlock:
4792 	CTX_UNLOCK(ctx);
4793 fail_ctx_free:
4794 	device_set_softc(ctx->ifc_dev, NULL);
4795         if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4796                 free(ctx->ifc_softc, M_IFLIB);
4797         free(ctx, M_IFLIB);
4798 	return (err);
4799 }
4800 
4801 int
4802 iflib_pseudo_register(device_t dev, if_shared_ctx_t sctx, if_ctx_t *ctxp,
4803 					  struct iflib_cloneattach_ctx *clctx)
4804 {
4805 	int err;
4806 	if_ctx_t ctx;
4807 	if_t ifp;
4808 	if_softc_ctx_t scctx;
4809 	int i;
4810 	void *sc;
4811 	uint16_t main_txq;
4812 	uint16_t main_rxq;
4813 
4814 	ctx = malloc(sizeof(*ctx), M_IFLIB, M_WAITOK|M_ZERO);
4815 	sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4816 	ctx->ifc_flags |= IFC_SC_ALLOCATED;
4817 	if (sctx->isc_flags & (IFLIB_PSEUDO|IFLIB_VIRTUAL))
4818 		ctx->ifc_flags |= IFC_PSEUDO;
4819 
4820 	ctx->ifc_sctx = sctx;
4821 	ctx->ifc_softc = sc;
4822 	ctx->ifc_dev = dev;
4823 
4824 	if ((err = iflib_register(ctx)) != 0) {
4825 		device_printf(dev, "%s: iflib_register failed %d\n", __func__, err);
4826 		goto fail_ctx_free;
4827 	}
4828 	iflib_add_device_sysctl_pre(ctx);
4829 
4830 	scctx = &ctx->ifc_softc_ctx;
4831 	ifp = ctx->ifc_ifp;
4832 
4833 	iflib_reset_qvalues(ctx);
4834 	CTX_LOCK(ctx);
4835 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4836 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4837 		goto fail_unlock;
4838 	}
4839 	if (sctx->isc_flags & IFLIB_GEN_MAC)
4840 		ether_gen_addr(ifp, &ctx->ifc_mac);
4841 	if ((err = IFDI_CLONEATTACH(ctx, clctx->cc_ifc, clctx->cc_name,
4842 								clctx->cc_params)) != 0) {
4843 		device_printf(dev, "IFDI_CLONEATTACH failed %d\n", err);
4844 		goto fail_ctx_free;
4845 	}
4846 	ifmedia_add(ctx->ifc_mediap, IFM_ETHER | IFM_1000_T | IFM_FDX, 0, NULL);
4847 	ifmedia_add(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO, 0, NULL);
4848 	ifmedia_set(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO);
4849 
4850 #ifdef INVARIANTS
4851 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
4852 		MPASS(scctx->isc_tx_csum_flags);
4853 #endif
4854 
4855 	if_setcapabilities(ifp, scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_LINKSTATE);
4856 	if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_LINKSTATE);
4857 
4858 	ifp->if_flags |= IFF_NOGROUP;
4859 	if (sctx->isc_flags & IFLIB_PSEUDO) {
4860 		ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
4861 
4862 		if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4863 			device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4864 			goto fail_detach;
4865 		}
4866 		*ctxp = ctx;
4867 
4868 		/*
4869 		 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4870 		 * This must appear after the call to ether_ifattach() because
4871 		 * ether_ifattach() sets if_hdrlen to the default value.
4872 		 */
4873 		if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4874 			if_setifheaderlen(ifp,
4875 			    sizeof(struct ether_vlan_header));
4876 
4877 		if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4878 		iflib_add_device_sysctl_post(ctx);
4879 		ctx->ifc_flags |= IFC_INIT_DONE;
4880 		return (0);
4881 	}
4882 	_iflib_pre_assert(scctx);
4883 	ctx->ifc_txrx = *scctx->isc_txrx;
4884 
4885 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4886 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4887 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4888 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4889 
4890 	main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4891 	main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4892 
4893 	/* XXX change for per-queue sizes */
4894 	device_printf(dev, "Using %d TX descriptors and %d RX descriptors\n",
4895 	    scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4896 
4897 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4898 	    MAX_SINGLE_PACKET_FRACTION)
4899 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4900 		    MAX_SINGLE_PACKET_FRACTION);
4901 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4902 	    MAX_SINGLE_PACKET_FRACTION)
4903 		scctx->isc_tx_tso_segments_max = max(1,
4904 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4905 
4906 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4907 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
4908 		/*
4909 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
4910 		 * but some MACs do.
4911 		 */
4912 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
4913 		    IP_MAXPACKET));
4914 		/*
4915 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
4916 		 * into account.  In the worst case, each of these calls will
4917 		 * add another mbuf and, thus, the requirement for another DMA
4918 		 * segment.  So for best performance, it doesn't make sense to
4919 		 * advertize a maximum of TSO segments that typically will
4920 		 * require defragmentation in iflib_encap().
4921 		 */
4922 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
4923 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
4924 	}
4925 	if (scctx->isc_rss_table_size == 0)
4926 		scctx->isc_rss_table_size = 64;
4927 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4928 
4929 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4930 	/* XXX format name */
4931 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx,
4932 	    NULL, NULL, "admin");
4933 
4934 	/* XXX --- can support > 1 -- but keep it simple for now */
4935 	scctx->isc_intr = IFLIB_INTR_LEGACY;
4936 
4937 	/* Get memory for the station queues */
4938 	if ((err = iflib_queues_alloc(ctx))) {
4939 		device_printf(dev, "Unable to allocate queue memory\n");
4940 		goto fail_iflib_detach;
4941 	}
4942 
4943 	if ((err = iflib_qset_structures_setup(ctx))) {
4944 		device_printf(dev, "qset structure setup failed %d\n", err);
4945 		goto fail_queues;
4946 	}
4947 
4948 	/*
4949 	 * XXX What if anything do we want to do about interrupts?
4950 	 */
4951 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
4952 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4953 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4954 		goto fail_detach;
4955 	}
4956 
4957 	/*
4958 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4959 	 * This must appear after the call to ether_ifattach() because
4960 	 * ether_ifattach() sets if_hdrlen to the default value.
4961 	 */
4962 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4963 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
4964 
4965 	/* XXX handle more than one queue */
4966 	for (i = 0; i < scctx->isc_nrxqsets; i++)
4967 		IFDI_RX_CLSET(ctx, 0, i, ctx->ifc_rxqs[i].ifr_fl[0].ifl_sds.ifsd_cl);
4968 
4969 	*ctxp = ctx;
4970 
4971 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4972 	iflib_add_device_sysctl_post(ctx);
4973 	ctx->ifc_flags |= IFC_INIT_DONE;
4974 	CTX_UNLOCK(ctx);
4975 
4976 	return (0);
4977 fail_detach:
4978 	ether_ifdetach(ctx->ifc_ifp);
4979 fail_queues:
4980 	iflib_tx_structures_free(ctx);
4981 	iflib_rx_structures_free(ctx);
4982 fail_iflib_detach:
4983 	IFDI_DETACH(ctx);
4984 fail_unlock:
4985 	CTX_UNLOCK(ctx);
4986 fail_ctx_free:
4987 	free(ctx->ifc_softc, M_IFLIB);
4988 	free(ctx, M_IFLIB);
4989 	return (err);
4990 }
4991 
4992 int
4993 iflib_pseudo_deregister(if_ctx_t ctx)
4994 {
4995 	if_t ifp = ctx->ifc_ifp;
4996 	iflib_txq_t txq;
4997 	iflib_rxq_t rxq;
4998 	int i, j;
4999 	struct taskqgroup *tqg;
5000 	iflib_fl_t fl;
5001 
5002 	/* Unregister VLAN events */
5003 	if (ctx->ifc_vlan_attach_event != NULL)
5004 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
5005 	if (ctx->ifc_vlan_detach_event != NULL)
5006 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
5007 
5008 	ether_ifdetach(ifp);
5009 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
5010 	CTX_LOCK_DESTROY(ctx);
5011 	/* XXX drain any dependent tasks */
5012 	tqg = qgroup_if_io_tqg;
5013 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
5014 		callout_drain(&txq->ift_timer);
5015 		if (txq->ift_task.gt_uniq != NULL)
5016 			taskqgroup_detach(tqg, &txq->ift_task);
5017 	}
5018 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
5019 		if (rxq->ifr_task.gt_uniq != NULL)
5020 			taskqgroup_detach(tqg, &rxq->ifr_task);
5021 
5022 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5023 			free(fl->ifl_rx_bitmap, M_IFLIB);
5024 	}
5025 	tqg = qgroup_if_config_tqg;
5026 	if (ctx->ifc_admin_task.gt_uniq != NULL)
5027 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
5028 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
5029 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
5030 
5031 	if_free(ifp);
5032 
5033 	iflib_tx_structures_free(ctx);
5034 	iflib_rx_structures_free(ctx);
5035 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5036 		free(ctx->ifc_softc, M_IFLIB);
5037 	free(ctx, M_IFLIB);
5038 	return (0);
5039 }
5040 
5041 int
5042 iflib_device_attach(device_t dev)
5043 {
5044 	if_ctx_t ctx;
5045 	if_shared_ctx_t sctx;
5046 
5047 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
5048 		return (ENOTSUP);
5049 
5050 	pci_enable_busmaster(dev);
5051 
5052 	return (iflib_device_register(dev, NULL, sctx, &ctx));
5053 }
5054 
5055 int
5056 iflib_device_deregister(if_ctx_t ctx)
5057 {
5058 	if_t ifp = ctx->ifc_ifp;
5059 	iflib_txq_t txq;
5060 	iflib_rxq_t rxq;
5061 	device_t dev = ctx->ifc_dev;
5062 	int i, j;
5063 	struct taskqgroup *tqg;
5064 	iflib_fl_t fl;
5065 
5066 	/* Make sure VLANS are not using driver */
5067 	if (if_vlantrunkinuse(ifp)) {
5068 		device_printf(dev, "Vlan in use, detach first\n");
5069 		return (EBUSY);
5070 	}
5071 #ifdef PCI_IOV
5072 	if (!CTX_IS_VF(ctx) && pci_iov_detach(dev) != 0) {
5073 		device_printf(dev, "SR-IOV in use; detach first.\n");
5074 		return (EBUSY);
5075 	}
5076 #endif
5077 
5078 	STATE_LOCK(ctx);
5079 	ctx->ifc_flags |= IFC_IN_DETACH;
5080 	STATE_UNLOCK(ctx);
5081 
5082 	CTX_LOCK(ctx);
5083 	iflib_stop(ctx);
5084 	CTX_UNLOCK(ctx);
5085 
5086 	/* Unregister VLAN events */
5087 	if (ctx->ifc_vlan_attach_event != NULL)
5088 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
5089 	if (ctx->ifc_vlan_detach_event != NULL)
5090 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
5091 
5092 	iflib_netmap_detach(ifp);
5093 	ether_ifdetach(ifp);
5094 	iflib_rem_pfil(ctx);
5095 	if (ctx->ifc_led_dev != NULL)
5096 		led_destroy(ctx->ifc_led_dev);
5097 	/* XXX drain any dependent tasks */
5098 	tqg = qgroup_if_io_tqg;
5099 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
5100 		callout_drain(&txq->ift_timer);
5101 		if (txq->ift_task.gt_uniq != NULL)
5102 			taskqgroup_detach(tqg, &txq->ift_task);
5103 	}
5104 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
5105 		if (rxq->ifr_task.gt_uniq != NULL)
5106 			taskqgroup_detach(tqg, &rxq->ifr_task);
5107 
5108 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5109 			free(fl->ifl_rx_bitmap, M_IFLIB);
5110 	}
5111 	tqg = qgroup_if_config_tqg;
5112 	if (ctx->ifc_admin_task.gt_uniq != NULL)
5113 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
5114 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
5115 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
5116 	CTX_LOCK(ctx);
5117 	IFDI_DETACH(ctx);
5118 	CTX_UNLOCK(ctx);
5119 
5120 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
5121 	CTX_LOCK_DESTROY(ctx);
5122 	device_set_softc(ctx->ifc_dev, NULL);
5123 	iflib_free_intr_mem(ctx);
5124 
5125 	bus_generic_detach(dev);
5126 	if_free(ifp);
5127 
5128 	iflib_tx_structures_free(ctx);
5129 	iflib_rx_structures_free(ctx);
5130 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5131 		free(ctx->ifc_softc, M_IFLIB);
5132 	unref_ctx_core_offset(ctx);
5133 	STATE_LOCK_DESTROY(ctx);
5134 	free(ctx, M_IFLIB);
5135 	return (0);
5136 }
5137 
5138 static void
5139 iflib_free_intr_mem(if_ctx_t ctx)
5140 {
5141 
5142 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) {
5143 		iflib_irq_free(ctx, &ctx->ifc_legacy_irq);
5144 	}
5145 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) {
5146 		pci_release_msi(ctx->ifc_dev);
5147 	}
5148 	if (ctx->ifc_msix_mem != NULL) {
5149 		bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY,
5150 		    rman_get_rid(ctx->ifc_msix_mem), ctx->ifc_msix_mem);
5151 		ctx->ifc_msix_mem = NULL;
5152 	}
5153 }
5154 
5155 int
5156 iflib_device_detach(device_t dev)
5157 {
5158 	if_ctx_t ctx = device_get_softc(dev);
5159 
5160 	return (iflib_device_deregister(ctx));
5161 }
5162 
5163 int
5164 iflib_device_suspend(device_t dev)
5165 {
5166 	if_ctx_t ctx = device_get_softc(dev);
5167 
5168 	CTX_LOCK(ctx);
5169 	IFDI_SUSPEND(ctx);
5170 	CTX_UNLOCK(ctx);
5171 
5172 	return bus_generic_suspend(dev);
5173 }
5174 int
5175 iflib_device_shutdown(device_t dev)
5176 {
5177 	if_ctx_t ctx = device_get_softc(dev);
5178 
5179 	CTX_LOCK(ctx);
5180 	IFDI_SHUTDOWN(ctx);
5181 	CTX_UNLOCK(ctx);
5182 
5183 	return bus_generic_suspend(dev);
5184 }
5185 
5186 
5187 int
5188 iflib_device_resume(device_t dev)
5189 {
5190 	if_ctx_t ctx = device_get_softc(dev);
5191 	iflib_txq_t txq = ctx->ifc_txqs;
5192 
5193 	CTX_LOCK(ctx);
5194 	IFDI_RESUME(ctx);
5195 	iflib_if_init_locked(ctx);
5196 	CTX_UNLOCK(ctx);
5197 	for (int i = 0; i < NTXQSETS(ctx); i++, txq++)
5198 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
5199 
5200 	return (bus_generic_resume(dev));
5201 }
5202 
5203 int
5204 iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params)
5205 {
5206 	int error;
5207 	if_ctx_t ctx = device_get_softc(dev);
5208 
5209 	CTX_LOCK(ctx);
5210 	error = IFDI_IOV_INIT(ctx, num_vfs, params);
5211 	CTX_UNLOCK(ctx);
5212 
5213 	return (error);
5214 }
5215 
5216 void
5217 iflib_device_iov_uninit(device_t dev)
5218 {
5219 	if_ctx_t ctx = device_get_softc(dev);
5220 
5221 	CTX_LOCK(ctx);
5222 	IFDI_IOV_UNINIT(ctx);
5223 	CTX_UNLOCK(ctx);
5224 }
5225 
5226 int
5227 iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params)
5228 {
5229 	int error;
5230 	if_ctx_t ctx = device_get_softc(dev);
5231 
5232 	CTX_LOCK(ctx);
5233 	error = IFDI_IOV_VF_ADD(ctx, vfnum, params);
5234 	CTX_UNLOCK(ctx);
5235 
5236 	return (error);
5237 }
5238 
5239 /*********************************************************************
5240  *
5241  *  MODULE FUNCTION DEFINITIONS
5242  *
5243  **********************************************************************/
5244 
5245 /*
5246  * - Start a fast taskqueue thread for each core
5247  * - Start a taskqueue for control operations
5248  */
5249 static int
5250 iflib_module_init(void)
5251 {
5252 	return (0);
5253 }
5254 
5255 static int
5256 iflib_module_event_handler(module_t mod, int what, void *arg)
5257 {
5258 	int err;
5259 
5260 	switch (what) {
5261 	case MOD_LOAD:
5262 		if ((err = iflib_module_init()) != 0)
5263 			return (err);
5264 		break;
5265 	case MOD_UNLOAD:
5266 		return (EBUSY);
5267 	default:
5268 		return (EOPNOTSUPP);
5269 	}
5270 
5271 	return (0);
5272 }
5273 
5274 /*********************************************************************
5275  *
5276  *  PUBLIC FUNCTION DEFINITIONS
5277  *     ordered as in iflib.h
5278  *
5279  **********************************************************************/
5280 
5281 
5282 static void
5283 _iflib_assert(if_shared_ctx_t sctx)
5284 {
5285 	int i;
5286 
5287 	MPASS(sctx->isc_tx_maxsize);
5288 	MPASS(sctx->isc_tx_maxsegsize);
5289 
5290 	MPASS(sctx->isc_rx_maxsize);
5291 	MPASS(sctx->isc_rx_nsegments);
5292 	MPASS(sctx->isc_rx_maxsegsize);
5293 
5294 	MPASS(sctx->isc_nrxqs >= 1 && sctx->isc_nrxqs <= 8);
5295 	for (i = 0; i < sctx->isc_nrxqs; i++) {
5296 		MPASS(sctx->isc_nrxd_min[i]);
5297 		MPASS(powerof2(sctx->isc_nrxd_min[i]));
5298 		MPASS(sctx->isc_nrxd_max[i]);
5299 		MPASS(powerof2(sctx->isc_nrxd_max[i]));
5300 		MPASS(sctx->isc_nrxd_default[i]);
5301 		MPASS(powerof2(sctx->isc_nrxd_default[i]));
5302 	}
5303 
5304 	MPASS(sctx->isc_ntxqs >= 1 && sctx->isc_ntxqs <= 8);
5305 	for (i = 0; i < sctx->isc_ntxqs; i++) {
5306 		MPASS(sctx->isc_ntxd_min[i]);
5307 		MPASS(powerof2(sctx->isc_ntxd_min[i]));
5308 		MPASS(sctx->isc_ntxd_max[i]);
5309 		MPASS(powerof2(sctx->isc_ntxd_max[i]));
5310 		MPASS(sctx->isc_ntxd_default[i]);
5311 		MPASS(powerof2(sctx->isc_ntxd_default[i]));
5312 	}
5313 }
5314 
5315 static void
5316 _iflib_pre_assert(if_softc_ctx_t scctx)
5317 {
5318 
5319 	MPASS(scctx->isc_txrx->ift_txd_encap);
5320 	MPASS(scctx->isc_txrx->ift_txd_flush);
5321 	MPASS(scctx->isc_txrx->ift_txd_credits_update);
5322 	MPASS(scctx->isc_txrx->ift_rxd_available);
5323 	MPASS(scctx->isc_txrx->ift_rxd_pkt_get);
5324 	MPASS(scctx->isc_txrx->ift_rxd_refill);
5325 	MPASS(scctx->isc_txrx->ift_rxd_flush);
5326 }
5327 
5328 static int
5329 iflib_register(if_ctx_t ctx)
5330 {
5331 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5332 	driver_t *driver = sctx->isc_driver;
5333 	device_t dev = ctx->ifc_dev;
5334 	if_t ifp;
5335 
5336 	_iflib_assert(sctx);
5337 
5338 	CTX_LOCK_INIT(ctx);
5339 	STATE_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev));
5340 	ifp = ctx->ifc_ifp = if_alloc(IFT_ETHER);
5341 	if (ifp == NULL) {
5342 		device_printf(dev, "can not allocate ifnet structure\n");
5343 		return (ENOMEM);
5344 	}
5345 
5346 	/*
5347 	 * Initialize our context's device specific methods
5348 	 */
5349 	kobj_init((kobj_t) ctx, (kobj_class_t) driver);
5350 	kobj_class_compile((kobj_class_t) driver);
5351 
5352 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
5353 	if_setsoftc(ifp, ctx);
5354 	if_setdev(ifp, dev);
5355 	if_setinitfn(ifp, iflib_if_init);
5356 	if_setioctlfn(ifp, iflib_if_ioctl);
5357 #ifdef ALTQ
5358 	if_setstartfn(ifp, iflib_altq_if_start);
5359 	if_settransmitfn(ifp, iflib_altq_if_transmit);
5360 	if_setsendqready(ifp);
5361 #else
5362 	if_settransmitfn(ifp, iflib_if_transmit);
5363 #endif
5364 	if_setqflushfn(ifp, iflib_if_qflush);
5365 	if_setflags(ifp, IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST);
5366 
5367 	ctx->ifc_vlan_attach_event =
5368 		EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx,
5369 							  EVENTHANDLER_PRI_FIRST);
5370 	ctx->ifc_vlan_detach_event =
5371 		EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx,
5372 							  EVENTHANDLER_PRI_FIRST);
5373 
5374 	if ((sctx->isc_flags & IFLIB_DRIVER_MEDIA) == 0) {
5375 		ctx->ifc_mediap = &ctx->ifc_media;
5376 		ifmedia_init(ctx->ifc_mediap, IFM_IMASK,
5377 		    iflib_media_change, iflib_media_status);
5378 	}
5379 	return (0);
5380 }
5381 
5382 static int
5383 iflib_queues_alloc(if_ctx_t ctx)
5384 {
5385 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5386 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5387 	device_t dev = ctx->ifc_dev;
5388 	int nrxqsets = scctx->isc_nrxqsets;
5389 	int ntxqsets = scctx->isc_ntxqsets;
5390 	iflib_txq_t txq;
5391 	iflib_rxq_t rxq;
5392 	iflib_fl_t fl = NULL;
5393 	int i, j, cpu, err, txconf, rxconf;
5394 	iflib_dma_info_t ifdip;
5395 	uint32_t *rxqsizes = scctx->isc_rxqsizes;
5396 	uint32_t *txqsizes = scctx->isc_txqsizes;
5397 	uint8_t nrxqs = sctx->isc_nrxqs;
5398 	uint8_t ntxqs = sctx->isc_ntxqs;
5399 	int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1;
5400 	caddr_t *vaddrs;
5401 	uint64_t *paddrs;
5402 
5403 	KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1"));
5404 	KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1"));
5405 
5406 	/* Allocate the TX ring struct memory */
5407 	if (!(ctx->ifc_txqs =
5408 	    (iflib_txq_t) malloc(sizeof(struct iflib_txq) *
5409 	    ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5410 		device_printf(dev, "Unable to allocate TX ring memory\n");
5411 		err = ENOMEM;
5412 		goto fail;
5413 	}
5414 
5415 	/* Now allocate the RX */
5416 	if (!(ctx->ifc_rxqs =
5417 	    (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) *
5418 	    nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5419 		device_printf(dev, "Unable to allocate RX ring memory\n");
5420 		err = ENOMEM;
5421 		goto rx_fail;
5422 	}
5423 
5424 	txq = ctx->ifc_txqs;
5425 	rxq = ctx->ifc_rxqs;
5426 
5427 	/*
5428 	 * XXX handle allocation failure
5429 	 */
5430 	for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) {
5431 		/* Set up some basics */
5432 
5433 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs,
5434 		    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
5435 			device_printf(dev,
5436 			    "Unable to allocate TX DMA info memory\n");
5437 			err = ENOMEM;
5438 			goto err_tx_desc;
5439 		}
5440 		txq->ift_ifdi = ifdip;
5441 		for (j = 0; j < ntxqs; j++, ifdip++) {
5442 			if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, 0)) {
5443 				device_printf(dev,
5444 				    "Unable to allocate TX descriptors\n");
5445 				err = ENOMEM;
5446 				goto err_tx_desc;
5447 			}
5448 			txq->ift_txd_size[j] = scctx->isc_txd_size[j];
5449 			bzero((void *)ifdip->idi_vaddr, txqsizes[j]);
5450 		}
5451 		txq->ift_ctx = ctx;
5452 		txq->ift_id = i;
5453 		if (sctx->isc_flags & IFLIB_HAS_TXCQ) {
5454 			txq->ift_br_offset = 1;
5455 		} else {
5456 			txq->ift_br_offset = 0;
5457 		}
5458 		/* XXX fix this */
5459 		txq->ift_timer.c_cpu = cpu;
5460 
5461 		if (iflib_txsd_alloc(txq)) {
5462 			device_printf(dev, "Critical Failure setting up TX buffers\n");
5463 			err = ENOMEM;
5464 			goto err_tx_desc;
5465 		}
5466 
5467 		/* Initialize the TX lock */
5468 		snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:TX(%d):callout",
5469 		    device_get_nameunit(dev), txq->ift_id);
5470 		mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF);
5471 		callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0);
5472 
5473 		err = ifmp_ring_alloc(&txq->ift_br, 2048, txq, iflib_txq_drain,
5474 				      iflib_txq_can_drain, M_IFLIB, M_WAITOK);
5475 		if (err) {
5476 			/* XXX free any allocated rings */
5477 			device_printf(dev, "Unable to allocate buf_ring\n");
5478 			goto err_tx_desc;
5479 		}
5480 	}
5481 
5482 	for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) {
5483 		/* Set up some basics */
5484 
5485 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs,
5486 		   M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
5487 			device_printf(dev,
5488 			    "Unable to allocate RX DMA info memory\n");
5489 			err = ENOMEM;
5490 			goto err_tx_desc;
5491 		}
5492 
5493 		rxq->ifr_ifdi = ifdip;
5494 		/* XXX this needs to be changed if #rx queues != #tx queues */
5495 		rxq->ifr_ntxqirq = 1;
5496 		rxq->ifr_txqid[0] = i;
5497 		for (j = 0; j < nrxqs; j++, ifdip++) {
5498 			if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, 0)) {
5499 				device_printf(dev,
5500 				    "Unable to allocate RX descriptors\n");
5501 				err = ENOMEM;
5502 				goto err_tx_desc;
5503 			}
5504 			bzero((void *)ifdip->idi_vaddr, rxqsizes[j]);
5505 		}
5506 		rxq->ifr_ctx = ctx;
5507 		rxq->ifr_id = i;
5508 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
5509 			rxq->ifr_fl_offset = 1;
5510 		} else {
5511 			rxq->ifr_fl_offset = 0;
5512 		}
5513 		rxq->ifr_nfl = nfree_lists;
5514 		if (!(fl =
5515 			  (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) {
5516 			device_printf(dev, "Unable to allocate free list memory\n");
5517 			err = ENOMEM;
5518 			goto err_tx_desc;
5519 		}
5520 		rxq->ifr_fl = fl;
5521 		for (j = 0; j < nfree_lists; j++) {
5522 			fl[j].ifl_rxq = rxq;
5523 			fl[j].ifl_id = j;
5524 			fl[j].ifl_ifdi = &rxq->ifr_ifdi[j + rxq->ifr_fl_offset];
5525 			fl[j].ifl_rxd_size = scctx->isc_rxd_size[j];
5526 		}
5527 		/* Allocate receive buffers for the ring */
5528 		if (iflib_rxsd_alloc(rxq)) {
5529 			device_printf(dev,
5530 			    "Critical Failure setting up receive buffers\n");
5531 			err = ENOMEM;
5532 			goto err_rx_desc;
5533 		}
5534 
5535 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5536 			fl->ifl_rx_bitmap = bit_alloc(fl->ifl_size, M_IFLIB,
5537 			    M_WAITOK);
5538 	}
5539 
5540 	/* TXQs */
5541 	vaddrs = malloc(sizeof(caddr_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5542 	paddrs = malloc(sizeof(uint64_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5543 	for (i = 0; i < ntxqsets; i++) {
5544 		iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi;
5545 
5546 		for (j = 0; j < ntxqs; j++, di++) {
5547 			vaddrs[i*ntxqs + j] = di->idi_vaddr;
5548 			paddrs[i*ntxqs + j] = di->idi_paddr;
5549 		}
5550 	}
5551 	if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) {
5552 		device_printf(ctx->ifc_dev,
5553 		    "Unable to allocate device TX queue\n");
5554 		iflib_tx_structures_free(ctx);
5555 		free(vaddrs, M_IFLIB);
5556 		free(paddrs, M_IFLIB);
5557 		goto err_rx_desc;
5558 	}
5559 	free(vaddrs, M_IFLIB);
5560 	free(paddrs, M_IFLIB);
5561 
5562 	/* RXQs */
5563 	vaddrs = malloc(sizeof(caddr_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5564 	paddrs = malloc(sizeof(uint64_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5565 	for (i = 0; i < nrxqsets; i++) {
5566 		iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi;
5567 
5568 		for (j = 0; j < nrxqs; j++, di++) {
5569 			vaddrs[i*nrxqs + j] = di->idi_vaddr;
5570 			paddrs[i*nrxqs + j] = di->idi_paddr;
5571 		}
5572 	}
5573 	if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) {
5574 		device_printf(ctx->ifc_dev,
5575 		    "Unable to allocate device RX queue\n");
5576 		iflib_tx_structures_free(ctx);
5577 		free(vaddrs, M_IFLIB);
5578 		free(paddrs, M_IFLIB);
5579 		goto err_rx_desc;
5580 	}
5581 	free(vaddrs, M_IFLIB);
5582 	free(paddrs, M_IFLIB);
5583 
5584 	return (0);
5585 
5586 /* XXX handle allocation failure changes */
5587 err_rx_desc:
5588 err_tx_desc:
5589 rx_fail:
5590 	if (ctx->ifc_rxqs != NULL)
5591 		free(ctx->ifc_rxqs, M_IFLIB);
5592 	ctx->ifc_rxqs = NULL;
5593 	if (ctx->ifc_txqs != NULL)
5594 		free(ctx->ifc_txqs, M_IFLIB);
5595 	ctx->ifc_txqs = NULL;
5596 fail:
5597 	return (err);
5598 }
5599 
5600 static int
5601 iflib_tx_structures_setup(if_ctx_t ctx)
5602 {
5603 	iflib_txq_t txq = ctx->ifc_txqs;
5604 	int i;
5605 
5606 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
5607 		iflib_txq_setup(txq);
5608 
5609 	return (0);
5610 }
5611 
5612 static void
5613 iflib_tx_structures_free(if_ctx_t ctx)
5614 {
5615 	iflib_txq_t txq = ctx->ifc_txqs;
5616 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5617 	int i, j;
5618 
5619 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
5620 		iflib_txq_destroy(txq);
5621 		for (j = 0; j < sctx->isc_ntxqs; j++)
5622 			iflib_dma_free(&txq->ift_ifdi[j]);
5623 	}
5624 	free(ctx->ifc_txqs, M_IFLIB);
5625 	ctx->ifc_txqs = NULL;
5626 	IFDI_QUEUES_FREE(ctx);
5627 }
5628 
5629 /*********************************************************************
5630  *
5631  *  Initialize all receive rings.
5632  *
5633  **********************************************************************/
5634 static int
5635 iflib_rx_structures_setup(if_ctx_t ctx)
5636 {
5637 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5638 	int q;
5639 #if defined(INET6) || defined(INET)
5640 	int err, i;
5641 #endif
5642 
5643 	for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) {
5644 #if defined(INET6) || defined(INET)
5645 		if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_LRO) {
5646 			err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp,
5647 			    TCP_LRO_ENTRIES, min(1024,
5648 			    ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset]));
5649 			if (err != 0) {
5650 				device_printf(ctx->ifc_dev,
5651 				    "LRO Initialization failed!\n");
5652 				goto fail;
5653 			}
5654 		}
5655 #endif
5656 		IFDI_RXQ_SETUP(ctx, rxq->ifr_id);
5657 	}
5658 	return (0);
5659 #if defined(INET6) || defined(INET)
5660 fail:
5661 	/*
5662 	 * Free LRO resources allocated so far, we will only handle
5663 	 * the rings that completed, the failing case will have
5664 	 * cleaned up for itself.  'q' failed, so its the terminus.
5665 	 */
5666 	rxq = ctx->ifc_rxqs;
5667 	for (i = 0; i < q; ++i, rxq++) {
5668 		if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_LRO)
5669 			tcp_lro_free(&rxq->ifr_lc);
5670 	}
5671 	return (err);
5672 #endif
5673 }
5674 
5675 /*********************************************************************
5676  *
5677  *  Free all receive rings.
5678  *
5679  **********************************************************************/
5680 static void
5681 iflib_rx_structures_free(if_ctx_t ctx)
5682 {
5683 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5684 	int i;
5685 
5686 	for (i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) {
5687 		iflib_rx_sds_free(rxq);
5688 #if defined(INET6) || defined(INET)
5689 		if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_LRO)
5690 			tcp_lro_free(&rxq->ifr_lc);
5691 #endif
5692 	}
5693 	free(ctx->ifc_rxqs, M_IFLIB);
5694 	ctx->ifc_rxqs = NULL;
5695 }
5696 
5697 static int
5698 iflib_qset_structures_setup(if_ctx_t ctx)
5699 {
5700 	int err;
5701 
5702 	/*
5703 	 * It is expected that the caller takes care of freeing queues if this
5704 	 * fails.
5705 	 */
5706 	if ((err = iflib_tx_structures_setup(ctx)) != 0) {
5707 		device_printf(ctx->ifc_dev, "iflib_tx_structures_setup failed: %d\n", err);
5708 		return (err);
5709 	}
5710 
5711 	if ((err = iflib_rx_structures_setup(ctx)) != 0)
5712 		device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err);
5713 
5714 	return (err);
5715 }
5716 
5717 int
5718 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
5719 		driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, const char *name)
5720 {
5721 
5722 	return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name));
5723 }
5724 
5725 #ifdef SMP
5726 static int
5727 find_nth(if_ctx_t ctx, int qid)
5728 {
5729 	cpuset_t cpus;
5730 	int i, cpuid, eqid, count;
5731 
5732 	CPU_COPY(&ctx->ifc_cpus, &cpus);
5733 	count = CPU_COUNT(&cpus);
5734 	eqid = qid % count;
5735 	/* clear up to the qid'th bit */
5736 	for (i = 0; i < eqid; i++) {
5737 		cpuid = CPU_FFS(&cpus);
5738 		MPASS(cpuid != 0);
5739 		CPU_CLR(cpuid-1, &cpus);
5740 	}
5741 	cpuid = CPU_FFS(&cpus);
5742 	MPASS(cpuid != 0);
5743 	return (cpuid-1);
5744 }
5745 
5746 #ifdef SCHED_ULE
5747 extern struct cpu_group *cpu_top;              /* CPU topology */
5748 
5749 static int
5750 find_child_with_core(int cpu, struct cpu_group *grp)
5751 {
5752 	int i;
5753 
5754 	if (grp->cg_children == 0)
5755 		return -1;
5756 
5757 	MPASS(grp->cg_child);
5758 	for (i = 0; i < grp->cg_children; i++) {
5759 		if (CPU_ISSET(cpu, &grp->cg_child[i].cg_mask))
5760 			return i;
5761 	}
5762 
5763 	return -1;
5764 }
5765 
5766 /*
5767  * Find the nth "close" core to the specified core
5768  * "close" is defined as the deepest level that shares
5769  * at least an L2 cache.  With threads, this will be
5770  * threads on the same core.  If the shared cache is L3
5771  * or higher, simply returns the same core.
5772  */
5773 static int
5774 find_close_core(int cpu, int core_offset)
5775 {
5776 	struct cpu_group *grp;
5777 	int i;
5778 	int fcpu;
5779 	cpuset_t cs;
5780 
5781 	grp = cpu_top;
5782 	if (grp == NULL)
5783 		return cpu;
5784 	i = 0;
5785 	while ((i = find_child_with_core(cpu, grp)) != -1) {
5786 		/* If the child only has one cpu, don't descend */
5787 		if (grp->cg_child[i].cg_count <= 1)
5788 			break;
5789 		grp = &grp->cg_child[i];
5790 	}
5791 
5792 	/* If they don't share at least an L2 cache, use the same CPU */
5793 	if (grp->cg_level > CG_SHARE_L2 || grp->cg_level == CG_SHARE_NONE)
5794 		return cpu;
5795 
5796 	/* Now pick one */
5797 	CPU_COPY(&grp->cg_mask, &cs);
5798 
5799 	/* Add the selected CPU offset to core offset. */
5800 	for (i = 0; (fcpu = CPU_FFS(&cs)) != 0; i++) {
5801 		if (fcpu - 1 == cpu)
5802 			break;
5803 		CPU_CLR(fcpu - 1, &cs);
5804 	}
5805 	MPASS(fcpu);
5806 
5807 	core_offset += i;
5808 
5809 	CPU_COPY(&grp->cg_mask, &cs);
5810 	for (i = core_offset % grp->cg_count; i > 0; i--) {
5811 		MPASS(CPU_FFS(&cs));
5812 		CPU_CLR(CPU_FFS(&cs) - 1, &cs);
5813 	}
5814 	MPASS(CPU_FFS(&cs));
5815 	return CPU_FFS(&cs) - 1;
5816 }
5817 #else
5818 static int
5819 find_close_core(int cpu, int core_offset __unused)
5820 {
5821 	return cpu;
5822 }
5823 #endif
5824 
5825 static int
5826 get_core_offset(if_ctx_t ctx, iflib_intr_type_t type, int qid)
5827 {
5828 	switch (type) {
5829 	case IFLIB_INTR_TX:
5830 		/* TX queues get cores which share at least an L2 cache with the corresponding RX queue */
5831 		/* XXX handle multiple RX threads per core and more than two core per L2 group */
5832 		return qid / CPU_COUNT(&ctx->ifc_cpus) + 1;
5833 	case IFLIB_INTR_RX:
5834 	case IFLIB_INTR_RXTX:
5835 		/* RX queues get the specified core */
5836 		return qid / CPU_COUNT(&ctx->ifc_cpus);
5837 	default:
5838 		return -1;
5839 	}
5840 }
5841 #else
5842 #define get_core_offset(ctx, type, qid)	CPU_FIRST()
5843 #define find_close_core(cpuid, tid)	CPU_FIRST()
5844 #define find_nth(ctx, gid)		CPU_FIRST()
5845 #endif
5846 
5847 /* Just to avoid copy/paste */
5848 static inline int
5849 iflib_irq_set_affinity(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type,
5850     int qid, struct grouptask *gtask, struct taskqgroup *tqg, void *uniq,
5851     const char *name)
5852 {
5853 	device_t dev;
5854 	int co, cpuid, err, tid;
5855 
5856 	dev = ctx->ifc_dev;
5857 	co = ctx->ifc_sysctl_core_offset;
5858 	if (ctx->ifc_sysctl_separate_txrx && type == IFLIB_INTR_TX)
5859 		co += ctx->ifc_softc_ctx.isc_nrxqsets;
5860 	cpuid = find_nth(ctx, qid + co);
5861 	tid = get_core_offset(ctx, type, qid);
5862 	if (tid < 0) {
5863 		device_printf(dev, "get_core_offset failed\n");
5864 		return (EOPNOTSUPP);
5865 	}
5866 	cpuid = find_close_core(cpuid, tid);
5867 	err = taskqgroup_attach_cpu(tqg, gtask, uniq, cpuid, dev, irq->ii_res,
5868 	    name);
5869 	if (err) {
5870 		device_printf(dev, "taskqgroup_attach_cpu failed %d\n", err);
5871 		return (err);
5872 	}
5873 #ifdef notyet
5874 	if (cpuid > ctx->ifc_cpuid_highest)
5875 		ctx->ifc_cpuid_highest = cpuid;
5876 #endif
5877 	return (0);
5878 }
5879 
5880 int
5881 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid,
5882 			iflib_intr_type_t type, driver_filter_t *filter,
5883 			void *filter_arg, int qid, const char *name)
5884 {
5885 	device_t dev;
5886 	struct grouptask *gtask;
5887 	struct taskqgroup *tqg;
5888 	iflib_filter_info_t info;
5889 	gtask_fn_t *fn;
5890 	int tqrid, err;
5891 	driver_filter_t *intr_fast;
5892 	void *q;
5893 
5894 	info = &ctx->ifc_filter_info;
5895 	tqrid = rid;
5896 
5897 	switch (type) {
5898 	/* XXX merge tx/rx for netmap? */
5899 	case IFLIB_INTR_TX:
5900 		q = &ctx->ifc_txqs[qid];
5901 		info = &ctx->ifc_txqs[qid].ift_filter_info;
5902 		gtask = &ctx->ifc_txqs[qid].ift_task;
5903 		tqg = qgroup_if_io_tqg;
5904 		fn = _task_fn_tx;
5905 		intr_fast = iflib_fast_intr;
5906 		GROUPTASK_INIT(gtask, 0, fn, q);
5907 		ctx->ifc_flags |= IFC_NETMAP_TX_IRQ;
5908 		break;
5909 	case IFLIB_INTR_RX:
5910 		q = &ctx->ifc_rxqs[qid];
5911 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
5912 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5913 		tqg = qgroup_if_io_tqg;
5914 		fn = _task_fn_rx;
5915 		intr_fast = iflib_fast_intr;
5916 		GROUPTASK_INIT(gtask, 0, fn, q);
5917 		break;
5918 	case IFLIB_INTR_RXTX:
5919 		q = &ctx->ifc_rxqs[qid];
5920 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
5921 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5922 		tqg = qgroup_if_io_tqg;
5923 		fn = _task_fn_rx;
5924 		intr_fast = iflib_fast_intr_rxtx;
5925 		GROUPTASK_INIT(gtask, 0, fn, q);
5926 		break;
5927 	case IFLIB_INTR_ADMIN:
5928 		q = ctx;
5929 		tqrid = -1;
5930 		info = &ctx->ifc_filter_info;
5931 		gtask = &ctx->ifc_admin_task;
5932 		tqg = qgroup_if_config_tqg;
5933 		fn = _task_fn_admin;
5934 		intr_fast = iflib_fast_intr_ctx;
5935 		break;
5936 	default:
5937 		device_printf(ctx->ifc_dev, "%s: unknown net intr type\n",
5938 		    __func__);
5939 		return (EINVAL);
5940 	}
5941 
5942 	info->ifi_filter = filter;
5943 	info->ifi_filter_arg = filter_arg;
5944 	info->ifi_task = gtask;
5945 	info->ifi_ctx = q;
5946 
5947 	dev = ctx->ifc_dev;
5948 	err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info,  name);
5949 	if (err != 0) {
5950 		device_printf(dev, "_iflib_irq_alloc failed %d\n", err);
5951 		return (err);
5952 	}
5953 	if (type == IFLIB_INTR_ADMIN)
5954 		return (0);
5955 
5956 	if (tqrid != -1) {
5957 		err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg,
5958 		    q, name);
5959 		if (err)
5960 			return (err);
5961 	} else {
5962 		taskqgroup_attach(tqg, gtask, q, dev, irq->ii_res, name);
5963 	}
5964 
5965 	return (0);
5966 }
5967 
5968 void
5969 iflib_softirq_alloc_generic(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type, void *arg, int qid, const char *name)
5970 {
5971 	struct grouptask *gtask;
5972 	struct taskqgroup *tqg;
5973 	gtask_fn_t *fn;
5974 	void *q;
5975 	int err;
5976 
5977 	switch (type) {
5978 	case IFLIB_INTR_TX:
5979 		q = &ctx->ifc_txqs[qid];
5980 		gtask = &ctx->ifc_txqs[qid].ift_task;
5981 		tqg = qgroup_if_io_tqg;
5982 		fn = _task_fn_tx;
5983 		break;
5984 	case IFLIB_INTR_RX:
5985 		q = &ctx->ifc_rxqs[qid];
5986 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5987 		tqg = qgroup_if_io_tqg;
5988 		fn = _task_fn_rx;
5989 		break;
5990 	case IFLIB_INTR_IOV:
5991 		q = ctx;
5992 		gtask = &ctx->ifc_vflr_task;
5993 		tqg = qgroup_if_config_tqg;
5994 		fn = _task_fn_iov;
5995 		break;
5996 	default:
5997 		panic("unknown net intr type");
5998 	}
5999 	GROUPTASK_INIT(gtask, 0, fn, q);
6000 	if (irq != NULL) {
6001 		err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg,
6002 		    q, name);
6003 		if (err)
6004 			taskqgroup_attach(tqg, gtask, q, ctx->ifc_dev,
6005 			    irq->ii_res, name);
6006 	} else {
6007 		taskqgroup_attach(tqg, gtask, q, NULL, NULL, name);
6008 	}
6009 }
6010 
6011 void
6012 iflib_irq_free(if_ctx_t ctx, if_irq_t irq)
6013 {
6014 
6015 	if (irq->ii_tag)
6016 		bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag);
6017 
6018 	if (irq->ii_res)
6019 		bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ,
6020 		    rman_get_rid(irq->ii_res), irq->ii_res);
6021 }
6022 
6023 static int
6024 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, const char *name)
6025 {
6026 	iflib_txq_t txq = ctx->ifc_txqs;
6027 	iflib_rxq_t rxq = ctx->ifc_rxqs;
6028 	if_irq_t irq = &ctx->ifc_legacy_irq;
6029 	iflib_filter_info_t info;
6030 	device_t dev;
6031 	struct grouptask *gtask;
6032 	struct resource *res;
6033 	struct taskqgroup *tqg;
6034 	gtask_fn_t *fn;
6035 	void *q;
6036 	int err, tqrid;
6037 
6038 	q = &ctx->ifc_rxqs[0];
6039 	info = &rxq[0].ifr_filter_info;
6040 	gtask = &rxq[0].ifr_task;
6041 	tqg = qgroup_if_io_tqg;
6042 	tqrid = *rid;
6043 	fn = _task_fn_rx;
6044 
6045 	ctx->ifc_flags |= IFC_LEGACY;
6046 	info->ifi_filter = filter;
6047 	info->ifi_filter_arg = filter_arg;
6048 	info->ifi_task = gtask;
6049 	info->ifi_ctx = q;
6050 
6051 	dev = ctx->ifc_dev;
6052 	/* We allocate a single interrupt resource */
6053 	if ((err = _iflib_irq_alloc(ctx, irq, tqrid, iflib_fast_intr_rxtx,
6054 	    NULL, info, name)) != 0)
6055 		return (err);
6056 	GROUPTASK_INIT(gtask, 0, fn, q);
6057 	res = irq->ii_res;
6058 	taskqgroup_attach(tqg, gtask, q, dev, res, name);
6059 
6060 	GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq);
6061 	taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, dev, res,
6062 	    "tx");
6063 	return (0);
6064 }
6065 
6066 void
6067 iflib_led_create(if_ctx_t ctx)
6068 {
6069 
6070 	ctx->ifc_led_dev = led_create(iflib_led_func, ctx,
6071 	    device_get_nameunit(ctx->ifc_dev));
6072 }
6073 
6074 void
6075 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid)
6076 {
6077 
6078 	GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
6079 }
6080 
6081 void
6082 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid)
6083 {
6084 
6085 	GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task);
6086 }
6087 
6088 void
6089 iflib_admin_intr_deferred(if_ctx_t ctx)
6090 {
6091 #ifdef INVARIANTS
6092 	struct grouptask *gtask;
6093 
6094 	gtask = &ctx->ifc_admin_task;
6095 	MPASS(gtask != NULL && gtask->gt_taskqueue != NULL);
6096 #endif
6097 
6098 	GROUPTASK_ENQUEUE(&ctx->ifc_admin_task);
6099 }
6100 
6101 void
6102 iflib_iov_intr_deferred(if_ctx_t ctx)
6103 {
6104 
6105 	GROUPTASK_ENQUEUE(&ctx->ifc_vflr_task);
6106 }
6107 
6108 void
6109 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, const char *name)
6110 {
6111 
6112 	taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, NULL, NULL,
6113 	    name);
6114 }
6115 
6116 void
6117 iflib_config_gtask_init(void *ctx, struct grouptask *gtask, gtask_fn_t *fn,
6118 	const char *name)
6119 {
6120 
6121 	GROUPTASK_INIT(gtask, 0, fn, ctx);
6122 	taskqgroup_attach(qgroup_if_config_tqg, gtask, gtask, NULL, NULL,
6123 	    name);
6124 }
6125 
6126 void
6127 iflib_config_gtask_deinit(struct grouptask *gtask)
6128 {
6129 
6130 	taskqgroup_detach(qgroup_if_config_tqg, gtask);
6131 }
6132 
6133 void
6134 iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate)
6135 {
6136 	if_t ifp = ctx->ifc_ifp;
6137 	iflib_txq_t txq = ctx->ifc_txqs;
6138 
6139 	if_setbaudrate(ifp, baudrate);
6140 	if (baudrate >= IF_Gbps(10)) {
6141 		STATE_LOCK(ctx);
6142 		ctx->ifc_flags |= IFC_PREFETCH;
6143 		STATE_UNLOCK(ctx);
6144 	}
6145 	/* If link down, disable watchdog */
6146 	if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) {
6147 		for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++)
6148 			txq->ift_qstatus = IFLIB_QUEUE_IDLE;
6149 	}
6150 	ctx->ifc_link_state = link_state;
6151 	if_link_state_change(ifp, link_state);
6152 }
6153 
6154 static int
6155 iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq)
6156 {
6157 	int credits;
6158 #ifdef INVARIANTS
6159 	int credits_pre = txq->ift_cidx_processed;
6160 #endif
6161 
6162 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
6163 	    BUS_DMASYNC_POSTREAD);
6164 	if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, true)) == 0)
6165 		return (0);
6166 
6167 	txq->ift_processed += credits;
6168 	txq->ift_cidx_processed += credits;
6169 
6170 	MPASS(credits_pre + credits == txq->ift_cidx_processed);
6171 	if (txq->ift_cidx_processed >= txq->ift_size)
6172 		txq->ift_cidx_processed -= txq->ift_size;
6173 	return (credits);
6174 }
6175 
6176 static int
6177 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget)
6178 {
6179 	iflib_fl_t fl;
6180 	u_int i;
6181 
6182 	for (i = 0, fl = &rxq->ifr_fl[0]; i < rxq->ifr_nfl; i++, fl++)
6183 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
6184 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
6185 	return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx,
6186 	    budget));
6187 }
6188 
6189 void
6190 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name,
6191 	const char *description, if_int_delay_info_t info,
6192 	int offset, int value)
6193 {
6194 	info->iidi_ctx = ctx;
6195 	info->iidi_offset = offset;
6196 	info->iidi_value = value;
6197 	SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev),
6198 	    SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)),
6199 	    OID_AUTO, name, CTLTYPE_INT|CTLFLAG_RW,
6200 	    info, 0, iflib_sysctl_int_delay, "I", description);
6201 }
6202 
6203 struct sx *
6204 iflib_ctx_lock_get(if_ctx_t ctx)
6205 {
6206 
6207 	return (&ctx->ifc_ctx_sx);
6208 }
6209 
6210 static int
6211 iflib_msix_init(if_ctx_t ctx)
6212 {
6213 	device_t dev = ctx->ifc_dev;
6214 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6215 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6216 	int admincnt, bar, err, iflib_num_rx_queues, iflib_num_tx_queues;
6217 	int msgs, queuemsgs, queues, rx_queues, tx_queues, vectors;
6218 
6219 	iflib_num_tx_queues = ctx->ifc_sysctl_ntxqs;
6220 	iflib_num_rx_queues = ctx->ifc_sysctl_nrxqs;
6221 
6222 	if (bootverbose)
6223 		device_printf(dev, "msix_init qsets capped at %d\n",
6224 		    imax(scctx->isc_ntxqsets, scctx->isc_nrxqsets));
6225 
6226 	/* Override by tuneable */
6227 	if (scctx->isc_disable_msix)
6228 		goto msi;
6229 
6230 	/* First try MSI-X */
6231 	if ((msgs = pci_msix_count(dev)) == 0) {
6232 		if (bootverbose)
6233 			device_printf(dev, "MSI-X not supported or disabled\n");
6234 		goto msi;
6235 	}
6236 
6237 	bar = ctx->ifc_softc_ctx.isc_msix_bar;
6238 	/*
6239 	 * bar == -1 => "trust me I know what I'm doing"
6240 	 * Some drivers are for hardware that is so shoddily
6241 	 * documented that no one knows which bars are which
6242 	 * so the developer has to map all bars. This hack
6243 	 * allows shoddy garbage to use MSI-X in this framework.
6244 	 */
6245 	if (bar != -1) {
6246 		ctx->ifc_msix_mem = bus_alloc_resource_any(dev,
6247 	            SYS_RES_MEMORY, &bar, RF_ACTIVE);
6248 		if (ctx->ifc_msix_mem == NULL) {
6249 			device_printf(dev, "Unable to map MSI-X table\n");
6250 			goto msi;
6251 		}
6252 	}
6253 
6254 	admincnt = sctx->isc_admin_intrcnt;
6255 #if IFLIB_DEBUG
6256 	/* use only 1 qset in debug mode */
6257 	queuemsgs = min(msgs - admincnt, 1);
6258 #else
6259 	queuemsgs = msgs - admincnt;
6260 #endif
6261 #ifdef RSS
6262 	queues = imin(queuemsgs, rss_getnumbuckets());
6263 #else
6264 	queues = queuemsgs;
6265 #endif
6266 	queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues);
6267 	if (bootverbose)
6268 		device_printf(dev,
6269 		    "intr CPUs: %d queue msgs: %d admincnt: %d\n",
6270 		    CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt);
6271 #ifdef  RSS
6272 	/* If we're doing RSS, clamp at the number of RSS buckets */
6273 	if (queues > rss_getnumbuckets())
6274 		queues = rss_getnumbuckets();
6275 #endif
6276 	if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt)
6277 		rx_queues = iflib_num_rx_queues;
6278 	else
6279 		rx_queues = queues;
6280 
6281 	if (rx_queues > scctx->isc_nrxqsets)
6282 		rx_queues = scctx->isc_nrxqsets;
6283 
6284 	/*
6285 	 * We want this to be all logical CPUs by default
6286 	 */
6287 	if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues)
6288 		tx_queues = iflib_num_tx_queues;
6289 	else
6290 		tx_queues = mp_ncpus;
6291 
6292 	if (tx_queues > scctx->isc_ntxqsets)
6293 		tx_queues = scctx->isc_ntxqsets;
6294 
6295 	if (ctx->ifc_sysctl_qs_eq_override == 0) {
6296 #ifdef INVARIANTS
6297 		if (tx_queues != rx_queues)
6298 			device_printf(dev,
6299 			    "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n",
6300 			    min(rx_queues, tx_queues), min(rx_queues, tx_queues));
6301 #endif
6302 		tx_queues = min(rx_queues, tx_queues);
6303 		rx_queues = min(rx_queues, tx_queues);
6304 	}
6305 
6306 	vectors = rx_queues + admincnt;
6307 	if (msgs < vectors) {
6308 		device_printf(dev,
6309 		    "insufficient number of MSI-X vectors "
6310 		    "(supported %d, need %d)\n", msgs, vectors);
6311 		goto msi;
6312 	}
6313 
6314 	device_printf(dev, "Using %d RX queues %d TX queues\n", rx_queues,
6315 	    tx_queues);
6316 	msgs = vectors;
6317 	if ((err = pci_alloc_msix(dev, &vectors)) == 0) {
6318 		if (vectors != msgs) {
6319 			device_printf(dev,
6320 			    "Unable to allocate sufficient MSI-X vectors "
6321 			    "(got %d, need %d)\n", vectors, msgs);
6322 			pci_release_msi(dev);
6323 			if (bar != -1) {
6324 				bus_release_resource(dev, SYS_RES_MEMORY, bar,
6325 				    ctx->ifc_msix_mem);
6326 				ctx->ifc_msix_mem = NULL;
6327 			}
6328 			goto msi;
6329 		}
6330 		device_printf(dev, "Using MSI-X interrupts with %d vectors\n",
6331 		    vectors);
6332 		scctx->isc_vectors = vectors;
6333 		scctx->isc_nrxqsets = rx_queues;
6334 		scctx->isc_ntxqsets = tx_queues;
6335 		scctx->isc_intr = IFLIB_INTR_MSIX;
6336 
6337 		return (vectors);
6338 	} else {
6339 		device_printf(dev,
6340 		    "failed to allocate %d MSI-X vectors, err: %d\n", vectors,
6341 		    err);
6342 		if (bar != -1) {
6343 			bus_release_resource(dev, SYS_RES_MEMORY, bar,
6344 			    ctx->ifc_msix_mem);
6345 			ctx->ifc_msix_mem = NULL;
6346 		}
6347 	}
6348 
6349 msi:
6350 	vectors = pci_msi_count(dev);
6351 	scctx->isc_nrxqsets = 1;
6352 	scctx->isc_ntxqsets = 1;
6353 	scctx->isc_vectors = vectors;
6354 	if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) {
6355 		device_printf(dev,"Using an MSI interrupt\n");
6356 		scctx->isc_intr = IFLIB_INTR_MSI;
6357 	} else {
6358 		scctx->isc_vectors = 1;
6359 		device_printf(dev,"Using a Legacy interrupt\n");
6360 		scctx->isc_intr = IFLIB_INTR_LEGACY;
6361 	}
6362 
6363 	return (vectors);
6364 }
6365 
6366 static const char *ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" };
6367 
6368 static int
6369 mp_ring_state_handler(SYSCTL_HANDLER_ARGS)
6370 {
6371 	int rc;
6372 	uint16_t *state = ((uint16_t *)oidp->oid_arg1);
6373 	struct sbuf *sb;
6374 	const char *ring_state = "UNKNOWN";
6375 
6376 	/* XXX needed ? */
6377 	rc = sysctl_wire_old_buffer(req, 0);
6378 	MPASS(rc == 0);
6379 	if (rc != 0)
6380 		return (rc);
6381 	sb = sbuf_new_for_sysctl(NULL, NULL, 80, req);
6382 	MPASS(sb != NULL);
6383 	if (sb == NULL)
6384 		return (ENOMEM);
6385 	if (state[3] <= 3)
6386 		ring_state = ring_states[state[3]];
6387 
6388 	sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s",
6389 		    state[0], state[1], state[2], ring_state);
6390 	rc = sbuf_finish(sb);
6391 	sbuf_delete(sb);
6392         return(rc);
6393 }
6394 
6395 enum iflib_ndesc_handler {
6396 	IFLIB_NTXD_HANDLER,
6397 	IFLIB_NRXD_HANDLER,
6398 };
6399 
6400 static int
6401 mp_ndesc_handler(SYSCTL_HANDLER_ARGS)
6402 {
6403 	if_ctx_t ctx = (void *)arg1;
6404 	enum iflib_ndesc_handler type = arg2;
6405 	char buf[256] = {0};
6406 	qidx_t *ndesc;
6407 	char *p, *next;
6408 	int nqs, rc, i;
6409 
6410 	nqs = 8;
6411 	switch(type) {
6412 	case IFLIB_NTXD_HANDLER:
6413 		ndesc = ctx->ifc_sysctl_ntxds;
6414 		if (ctx->ifc_sctx)
6415 			nqs = ctx->ifc_sctx->isc_ntxqs;
6416 		break;
6417 	case IFLIB_NRXD_HANDLER:
6418 		ndesc = ctx->ifc_sysctl_nrxds;
6419 		if (ctx->ifc_sctx)
6420 			nqs = ctx->ifc_sctx->isc_nrxqs;
6421 		break;
6422 	default:
6423 		printf("%s: unhandled type\n", __func__);
6424 		return (EINVAL);
6425 	}
6426 	if (nqs == 0)
6427 		nqs = 8;
6428 
6429 	for (i=0; i<8; i++) {
6430 		if (i >= nqs)
6431 			break;
6432 		if (i)
6433 			strcat(buf, ",");
6434 		sprintf(strchr(buf, 0), "%d", ndesc[i]);
6435 	}
6436 
6437 	rc = sysctl_handle_string(oidp, buf, sizeof(buf), req);
6438 	if (rc || req->newptr == NULL)
6439 		return rc;
6440 
6441 	for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p;
6442 	    i++, p = strsep(&next, " ,")) {
6443 		ndesc[i] = strtoul(p, NULL, 10);
6444 	}
6445 
6446 	return(rc);
6447 }
6448 
6449 #define NAME_BUFLEN 32
6450 static void
6451 iflib_add_device_sysctl_pre(if_ctx_t ctx)
6452 {
6453         device_t dev = iflib_get_dev(ctx);
6454 	struct sysctl_oid_list *child, *oid_list;
6455 	struct sysctl_ctx_list *ctx_list;
6456 	struct sysctl_oid *node;
6457 
6458 	ctx_list = device_get_sysctl_ctx(dev);
6459 	child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev));
6460 	ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, "iflib",
6461 						      CTLFLAG_RD, NULL, "IFLIB fields");
6462 	oid_list = SYSCTL_CHILDREN(node);
6463 
6464 	SYSCTL_ADD_CONST_STRING(ctx_list, oid_list, OID_AUTO, "driver_version",
6465 		       CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version,
6466 		       "driver version");
6467 
6468 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs",
6469 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0,
6470 			"# of txqs to use, 0 => use default #");
6471 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs",
6472 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0,
6473 			"# of rxqs to use, 0 => use default #");
6474 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_qs_enable",
6475 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0,
6476                        "permit #txq != #rxq");
6477 	SYSCTL_ADD_INT(ctx_list, oid_list, OID_AUTO, "disable_msix",
6478                       CTLFLAG_RWTUN, &ctx->ifc_softc_ctx.isc_disable_msix, 0,
6479                       "disable MSI-X (default 0)");
6480 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "rx_budget",
6481 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_rx_budget, 0,
6482 		       "set the RX budget");
6483 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "tx_abdicate",
6484 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_tx_abdicate, 0,
6485 		       "cause TX to abdicate instead of running to completion");
6486 	ctx->ifc_sysctl_core_offset = CORE_OFFSET_UNSPECIFIED;
6487 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "core_offset",
6488 		       CTLFLAG_RDTUN, &ctx->ifc_sysctl_core_offset, 0,
6489 		       "offset to start using cores at");
6490 	SYSCTL_ADD_U8(ctx_list, oid_list, OID_AUTO, "separate_txrx",
6491 		       CTLFLAG_RDTUN, &ctx->ifc_sysctl_separate_txrx, 0,
6492 		       "use separate cores for TX and RX");
6493 
6494 	/* XXX change for per-queue sizes */
6495 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_ntxds",
6496 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NTXD_HANDLER,
6497                        mp_ndesc_handler, "A",
6498 		       "list of # of TX descriptors to use, 0 = use default #");
6499 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_nrxds",
6500 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NRXD_HANDLER,
6501                        mp_ndesc_handler, "A",
6502 		       "list of # of RX descriptors to use, 0 = use default #");
6503 }
6504 
6505 static void
6506 iflib_add_device_sysctl_post(if_ctx_t ctx)
6507 {
6508 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6509 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6510         device_t dev = iflib_get_dev(ctx);
6511 	struct sysctl_oid_list *child;
6512 	struct sysctl_ctx_list *ctx_list;
6513 	iflib_fl_t fl;
6514 	iflib_txq_t txq;
6515 	iflib_rxq_t rxq;
6516 	int i, j;
6517 	char namebuf[NAME_BUFLEN];
6518 	char *qfmt;
6519 	struct sysctl_oid *queue_node, *fl_node, *node;
6520 	struct sysctl_oid_list *queue_list, *fl_list;
6521 	ctx_list = device_get_sysctl_ctx(dev);
6522 
6523 	node = ctx->ifc_sysctl_node;
6524 	child = SYSCTL_CHILDREN(node);
6525 
6526 	if (scctx->isc_ntxqsets > 100)
6527 		qfmt = "txq%03d";
6528 	else if (scctx->isc_ntxqsets > 10)
6529 		qfmt = "txq%02d";
6530 	else
6531 		qfmt = "txq%d";
6532 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
6533 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6534 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6535 					     CTLFLAG_RD, NULL, "Queue Name");
6536 		queue_list = SYSCTL_CHILDREN(queue_node);
6537 #if MEMORY_LOGGING
6538 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued",
6539 				CTLFLAG_RD,
6540 				&txq->ift_dequeued, "total mbufs freed");
6541 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued",
6542 				CTLFLAG_RD,
6543 				&txq->ift_enqueued, "total mbufs enqueued");
6544 #endif
6545 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag",
6546 				   CTLFLAG_RD,
6547 				   &txq->ift_mbuf_defrag, "# of times m_defrag was called");
6548 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "m_pullups",
6549 				   CTLFLAG_RD,
6550 				   &txq->ift_pullups, "# of times m_pullup was called");
6551 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag_failed",
6552 				   CTLFLAG_RD,
6553 				   &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed");
6554 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_desc_avail",
6555 				   CTLFLAG_RD,
6556 				   &txq->ift_no_desc_avail, "# of times no descriptors were available");
6557 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "tx_map_failed",
6558 				   CTLFLAG_RD,
6559 				   &txq->ift_map_failed, "# of times DMA map failed");
6560 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txd_encap_efbig",
6561 				   CTLFLAG_RD,
6562 				   &txq->ift_txd_encap_efbig, "# of times txd_encap returned EFBIG");
6563 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_tx_dma_setup",
6564 				   CTLFLAG_RD,
6565 				   &txq->ift_no_tx_dma_setup, "# of times map failed for other than EFBIG");
6566 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx",
6567 				   CTLFLAG_RD,
6568 				   &txq->ift_pidx, 1, "Producer Index");
6569 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx",
6570 				   CTLFLAG_RD,
6571 				   &txq->ift_cidx, 1, "Consumer Index");
6572 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx_processed",
6573 				   CTLFLAG_RD,
6574 				   &txq->ift_cidx_processed, 1, "Consumer Index seen by credit update");
6575 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use",
6576 				   CTLFLAG_RD,
6577 				   &txq->ift_in_use, 1, "descriptors in use");
6578 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_processed",
6579 				   CTLFLAG_RD,
6580 				   &txq->ift_processed, "descriptors procesed for clean");
6581 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned",
6582 				   CTLFLAG_RD,
6583 				   &txq->ift_cleaned, "total cleaned");
6584 		SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state",
6585 				CTLTYPE_STRING | CTLFLAG_RD, __DEVOLATILE(uint64_t *, &txq->ift_br->state),
6586 				0, mp_ring_state_handler, "A", "soft ring state");
6587 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_enqueues",
6588 				       CTLFLAG_RD, &txq->ift_br->enqueues,
6589 				       "# of enqueues to the mp_ring for this queue");
6590 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_drops",
6591 				       CTLFLAG_RD, &txq->ift_br->drops,
6592 				       "# of drops in the mp_ring for this queue");
6593 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_starts",
6594 				       CTLFLAG_RD, &txq->ift_br->starts,
6595 				       "# of normal consumer starts in the mp_ring for this queue");
6596 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_stalls",
6597 				       CTLFLAG_RD, &txq->ift_br->stalls,
6598 					       "# of consumer stalls in the mp_ring for this queue");
6599 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_restarts",
6600 			       CTLFLAG_RD, &txq->ift_br->restarts,
6601 				       "# of consumer restarts in the mp_ring for this queue");
6602 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_abdications",
6603 				       CTLFLAG_RD, &txq->ift_br->abdications,
6604 				       "# of consumer abdications in the mp_ring for this queue");
6605 	}
6606 
6607 	if (scctx->isc_nrxqsets > 100)
6608 		qfmt = "rxq%03d";
6609 	else if (scctx->isc_nrxqsets > 10)
6610 		qfmt = "rxq%02d";
6611 	else
6612 		qfmt = "rxq%d";
6613 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
6614 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6615 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6616 					     CTLFLAG_RD, NULL, "Queue Name");
6617 		queue_list = SYSCTL_CHILDREN(queue_node);
6618 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
6619 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_cidx",
6620 				       CTLFLAG_RD,
6621 				       &rxq->ifr_cq_cidx, 1, "Consumer Index");
6622 		}
6623 
6624 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
6625 			snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j);
6626 			fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list, OID_AUTO, namebuf,
6627 						     CTLFLAG_RD, NULL, "freelist Name");
6628 			fl_list = SYSCTL_CHILDREN(fl_node);
6629 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx",
6630 				       CTLFLAG_RD,
6631 				       &fl->ifl_pidx, 1, "Producer Index");
6632 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx",
6633 				       CTLFLAG_RD,
6634 				       &fl->ifl_cidx, 1, "Consumer Index");
6635 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits",
6636 				       CTLFLAG_RD,
6637 				       &fl->ifl_credits, 1, "credits available");
6638 #if MEMORY_LOGGING
6639 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_enqueued",
6640 					CTLFLAG_RD,
6641 					&fl->ifl_m_enqueued, "mbufs allocated");
6642 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_dequeued",
6643 					CTLFLAG_RD,
6644 					&fl->ifl_m_dequeued, "mbufs freed");
6645 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_enqueued",
6646 					CTLFLAG_RD,
6647 					&fl->ifl_cl_enqueued, "clusters allocated");
6648 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_dequeued",
6649 					CTLFLAG_RD,
6650 					&fl->ifl_cl_dequeued, "clusters freed");
6651 #endif
6652 
6653 		}
6654 	}
6655 
6656 }
6657 
6658 void
6659 iflib_request_reset(if_ctx_t ctx)
6660 {
6661 
6662 	STATE_LOCK(ctx);
6663 	ctx->ifc_flags |= IFC_DO_RESET;
6664 	STATE_UNLOCK(ctx);
6665 }
6666 
6667 #ifndef __NO_STRICT_ALIGNMENT
6668 static struct mbuf *
6669 iflib_fixup_rx(struct mbuf *m)
6670 {
6671 	struct mbuf *n;
6672 
6673 	if (m->m_len <= (MCLBYTES - ETHER_HDR_LEN)) {
6674 		bcopy(m->m_data, m->m_data + ETHER_HDR_LEN, m->m_len);
6675 		m->m_data += ETHER_HDR_LEN;
6676 		n = m;
6677 	} else {
6678 		MGETHDR(n, M_NOWAIT, MT_DATA);
6679 		if (n == NULL) {
6680 			m_freem(m);
6681 			return (NULL);
6682 		}
6683 		bcopy(m->m_data, n->m_data, ETHER_HDR_LEN);
6684 		m->m_data += ETHER_HDR_LEN;
6685 		m->m_len -= ETHER_HDR_LEN;
6686 		n->m_len = ETHER_HDR_LEN;
6687 		M_MOVE_PKTHDR(n, m);
6688 		n->m_next = m;
6689 	}
6690 	return (n);
6691 }
6692 #endif
6693 
6694 #ifdef NETDUMP
6695 static void
6696 iflib_netdump_init(if_t ifp, int *nrxr, int *ncl, int *clsize)
6697 {
6698 	if_ctx_t ctx;
6699 
6700 	ctx = if_getsoftc(ifp);
6701 	CTX_LOCK(ctx);
6702 	*nrxr = NRXQSETS(ctx);
6703 	*ncl = ctx->ifc_rxqs[0].ifr_fl->ifl_size;
6704 	*clsize = ctx->ifc_rxqs[0].ifr_fl->ifl_buf_size;
6705 	CTX_UNLOCK(ctx);
6706 }
6707 
6708 static void
6709 iflib_netdump_event(if_t ifp, enum netdump_ev event)
6710 {
6711 	if_ctx_t ctx;
6712 	if_softc_ctx_t scctx;
6713 	iflib_fl_t fl;
6714 	iflib_rxq_t rxq;
6715 	int i, j;
6716 
6717 	ctx = if_getsoftc(ifp);
6718 	scctx = &ctx->ifc_softc_ctx;
6719 
6720 	switch (event) {
6721 	case NETDUMP_START:
6722 		for (i = 0; i < scctx->isc_nrxqsets; i++) {
6723 			rxq = &ctx->ifc_rxqs[i];
6724 			for (j = 0; j < rxq->ifr_nfl; j++) {
6725 				fl = rxq->ifr_fl;
6726 				fl->ifl_zone = m_getzone(fl->ifl_buf_size);
6727 			}
6728 		}
6729 		iflib_no_tx_batch = 1;
6730 		break;
6731 	default:
6732 		break;
6733 	}
6734 }
6735 
6736 static int
6737 iflib_netdump_transmit(if_t ifp, struct mbuf *m)
6738 {
6739 	if_ctx_t ctx;
6740 	iflib_txq_t txq;
6741 	int error;
6742 
6743 	ctx = if_getsoftc(ifp);
6744 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
6745 	    IFF_DRV_RUNNING)
6746 		return (EBUSY);
6747 
6748 	txq = &ctx->ifc_txqs[0];
6749 	error = iflib_encap(txq, &m);
6750 	if (error == 0)
6751 		(void)iflib_txd_db_check(ctx, txq, true, txq->ift_in_use);
6752 	return (error);
6753 }
6754 
6755 static int
6756 iflib_netdump_poll(if_t ifp, int count)
6757 {
6758 	if_ctx_t ctx;
6759 	if_softc_ctx_t scctx;
6760 	iflib_txq_t txq;
6761 	int i;
6762 
6763 	ctx = if_getsoftc(ifp);
6764 	scctx = &ctx->ifc_softc_ctx;
6765 
6766 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
6767 	    IFF_DRV_RUNNING)
6768 		return (EBUSY);
6769 
6770 	txq = &ctx->ifc_txqs[0];
6771 	(void)iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
6772 
6773 	for (i = 0; i < scctx->isc_nrxqsets; i++)
6774 		(void)iflib_rxeof(&ctx->ifc_rxqs[i], 16 /* XXX */);
6775 	return (0);
6776 }
6777 #endif /* NETDUMP */
6778