xref: /freebsd/sys/net/iflib.c (revision 83d9fd40d5cb67ff1e805523a6fae48c3389f396)
1 /*-
2  * Copyright (c) 2014-2017, Matthew Macy <mmacy@nextbsd.org>
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 
35 #include <sys/param.h>
36 #include <sys/types.h>
37 #include <sys/bus.h>
38 #include <sys/eventhandler.h>
39 #include <sys/sockio.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/sysctl.h>
50 #include <sys/syslog.h>
51 #include <sys/taskqueue.h>
52 #include <sys/limits.h>
53 
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 
63 #include <netinet/in.h>
64 #include <netinet/in_pcb.h>
65 #include <netinet/tcp_lro.h>
66 #include <netinet/in_systm.h>
67 #include <netinet/if_ether.h>
68 #include <netinet/ip.h>
69 #include <netinet/ip6.h>
70 #include <netinet/tcp.h>
71 
72 #include <machine/bus.h>
73 #include <machine/in_cksum.h>
74 
75 #include <vm/vm.h>
76 #include <vm/pmap.h>
77 
78 #include <dev/led/led.h>
79 #include <dev/pci/pcireg.h>
80 #include <dev/pci/pcivar.h>
81 #include <dev/pci/pci_private.h>
82 
83 #include <net/iflib.h>
84 
85 #include "ifdi_if.h"
86 
87 #if defined(__i386__) || defined(__amd64__)
88 #include <sys/memdesc.h>
89 #include <machine/bus.h>
90 #include <machine/md_var.h>
91 #include <machine/specialreg.h>
92 #include <x86/include/busdma_impl.h>
93 #include <x86/iommu/busdma_dmar.h>
94 #endif
95 
96 /*
97  * enable accounting of every mbuf as it comes in to and goes out of
98  * iflib's software descriptor references
99  */
100 #define MEMORY_LOGGING 0
101 /*
102  * Enable mbuf vectors for compressing long mbuf chains
103  */
104 
105 /*
106  * NB:
107  * - Prefetching in tx cleaning should perhaps be a tunable. The distance ahead
108  *   we prefetch needs to be determined by the time spent in m_free vis a vis
109  *   the cost of a prefetch. This will of course vary based on the workload:
110  *      - NFLX's m_free path is dominated by vm-based M_EXT manipulation which
111  *        is quite expensive, thus suggesting very little prefetch.
112  *      - small packet forwarding which is just returning a single mbuf to
113  *        UMA will typically be very fast vis a vis the cost of a memory
114  *        access.
115  */
116 
117 
118 /*
119  * File organization:
120  *  - private structures
121  *  - iflib private utility functions
122  *  - ifnet functions
123  *  - vlan registry and other exported functions
124  *  - iflib public core functions
125  *
126  *
127  */
128 static MALLOC_DEFINE(M_IFLIB, "iflib", "ifnet library");
129 
130 struct iflib_txq;
131 typedef struct iflib_txq *iflib_txq_t;
132 struct iflib_rxq;
133 typedef struct iflib_rxq *iflib_rxq_t;
134 struct iflib_fl;
135 typedef struct iflib_fl *iflib_fl_t;
136 
137 struct iflib_ctx;
138 
139 typedef struct iflib_filter_info {
140 	driver_filter_t *ifi_filter;
141 	void *ifi_filter_arg;
142 	struct grouptask *ifi_task;
143 	void *ifi_ctx;
144 } *iflib_filter_info_t;
145 
146 struct iflib_ctx {
147 	KOBJ_FIELDS;
148    /*
149    * Pointer to hardware driver's softc
150    */
151 	void *ifc_softc;
152 	device_t ifc_dev;
153 	if_t ifc_ifp;
154 
155 	cpuset_t ifc_cpus;
156 	if_shared_ctx_t ifc_sctx;
157 	struct if_softc_ctx ifc_softc_ctx;
158 
159 	struct mtx ifc_mtx;
160 
161 	uint16_t ifc_nhwtxqs;
162 	uint16_t ifc_nhwrxqs;
163 
164 	iflib_txq_t ifc_txqs;
165 	iflib_rxq_t ifc_rxqs;
166 	uint32_t ifc_if_flags;
167 	uint32_t ifc_flags;
168 	uint32_t ifc_max_fl_buf_size;
169 	int ifc_in_detach;
170 
171 	int ifc_link_state;
172 	int ifc_link_irq;
173 	int ifc_pause_frames;
174 	int ifc_watchdog_events;
175 	struct cdev *ifc_led_dev;
176 	struct resource *ifc_msix_mem;
177 
178 	struct if_irq ifc_legacy_irq;
179 	struct grouptask ifc_admin_task;
180 	struct grouptask ifc_vflr_task;
181 	struct iflib_filter_info ifc_filter_info;
182 	struct ifmedia	ifc_media;
183 
184 	struct sysctl_oid *ifc_sysctl_node;
185 	uint16_t ifc_sysctl_ntxqs;
186 	uint16_t ifc_sysctl_nrxqs;
187 	uint16_t ifc_sysctl_qs_eq_override;
188 
189 	qidx_t ifc_sysctl_ntxds[8];
190 	qidx_t ifc_sysctl_nrxds[8];
191 	struct if_txrx ifc_txrx;
192 #define isc_txd_encap  ifc_txrx.ift_txd_encap
193 #define isc_txd_flush  ifc_txrx.ift_txd_flush
194 #define isc_txd_credits_update  ifc_txrx.ift_txd_credits_update
195 #define isc_rxd_available ifc_txrx.ift_rxd_available
196 #define isc_rxd_pkt_get ifc_txrx.ift_rxd_pkt_get
197 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
198 #define isc_rxd_flush ifc_txrx.ift_rxd_flush
199 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
200 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
201 #define isc_legacy_intr ifc_txrx.ift_legacy_intr
202 	eventhandler_tag ifc_vlan_attach_event;
203 	eventhandler_tag ifc_vlan_detach_event;
204 	uint8_t ifc_mac[ETHER_ADDR_LEN];
205 	char ifc_mtx_name[16];
206 };
207 
208 
209 void *
210 iflib_get_softc(if_ctx_t ctx)
211 {
212 
213 	return (ctx->ifc_softc);
214 }
215 
216 device_t
217 iflib_get_dev(if_ctx_t ctx)
218 {
219 
220 	return (ctx->ifc_dev);
221 }
222 
223 if_t
224 iflib_get_ifp(if_ctx_t ctx)
225 {
226 
227 	return (ctx->ifc_ifp);
228 }
229 
230 struct ifmedia *
231 iflib_get_media(if_ctx_t ctx)
232 {
233 
234 	return (&ctx->ifc_media);
235 }
236 
237 void
238 iflib_set_mac(if_ctx_t ctx, uint8_t mac[ETHER_ADDR_LEN])
239 {
240 
241 	bcopy(mac, ctx->ifc_mac, ETHER_ADDR_LEN);
242 }
243 
244 if_softc_ctx_t
245 iflib_get_softc_ctx(if_ctx_t ctx)
246 {
247 
248 	return (&ctx->ifc_softc_ctx);
249 }
250 
251 if_shared_ctx_t
252 iflib_get_sctx(if_ctx_t ctx)
253 {
254 
255 	return (ctx->ifc_sctx);
256 }
257 
258 #define IP_ALIGNED(m) ((((uintptr_t)(m)->m_data) & 0x3) == 0x2)
259 #define CACHE_PTR_INCREMENT (CACHE_LINE_SIZE/sizeof(void*))
260 #define CACHE_PTR_NEXT(ptr) ((void *)(((uintptr_t)(ptr)+CACHE_LINE_SIZE-1) & (CACHE_LINE_SIZE-1)))
261 
262 #define LINK_ACTIVE(ctx) ((ctx)->ifc_link_state == LINK_STATE_UP)
263 #define CTX_IS_VF(ctx) ((ctx)->ifc_sctx->isc_flags & IFLIB_IS_VF)
264 
265 #define RX_SW_DESC_MAP_CREATED	(1 << 0)
266 #define TX_SW_DESC_MAP_CREATED	(1 << 1)
267 #define RX_SW_DESC_INUSE        (1 << 3)
268 #define TX_SW_DESC_MAPPED       (1 << 4)
269 
270 typedef struct iflib_sw_rx_desc_array {
271 	bus_dmamap_t	*ifsd_map;         /* bus_dma maps for packet */
272 	struct mbuf	**ifsd_m;           /* pkthdr mbufs */
273 	caddr_t		*ifsd_cl;          /* direct cluster pointer for rx */
274 	uint8_t		*ifsd_flags;
275 } iflib_rxsd_array_t;
276 
277 typedef struct iflib_sw_tx_desc_array {
278 	bus_dmamap_t    *ifsd_map;         /* bus_dma maps for packet */
279 	struct mbuf    **ifsd_m;           /* pkthdr mbufs */
280 	uint8_t		*ifsd_flags;
281 } if_txsd_vec_t;
282 
283 
284 /* magic number that should be high enough for any hardware */
285 #define IFLIB_MAX_TX_SEGS		128
286 #define IFLIB_MAX_RX_SEGS		32
287 #define IFLIB_RX_COPY_THRESH		128
288 #define IFLIB_MAX_RX_REFRESH		32
289 /* The minimum descriptors per second before we start coalescing */
290 #define IFLIB_MIN_DESC_SEC		16384
291 #define IFLIB_DEFAULT_TX_UPDATE_FREQ	16
292 #define IFLIB_QUEUE_IDLE		0
293 #define IFLIB_QUEUE_HUNG		1
294 #define IFLIB_QUEUE_WORKING		2
295 /* maximum number of txqs that can share an rx interrupt */
296 #define IFLIB_MAX_TX_SHARED_INTR	4
297 
298 /* this should really scale with ring size - this is a fairly arbitrary value */
299 #define TX_BATCH_SIZE			32
300 
301 #define IFLIB_RESTART_BUDGET		8
302 
303 #define	IFC_LEGACY		0x001
304 #define	IFC_QFLUSH		0x002
305 #define	IFC_MULTISEG		0x004
306 #define	IFC_DMAR		0x008
307 #define	IFC_SC_ALLOCATED	0x010
308 #define	IFC_INIT_DONE		0x020
309 #define	IFC_PREFETCH		0x040
310 #define	IFC_DO_RESET		0x080
311 #define	IFC_CHECK_HUNG		0x100
312 
313 #define CSUM_OFFLOAD		(CSUM_IP_TSO|CSUM_IP6_TSO|CSUM_IP| \
314 				 CSUM_IP_UDP|CSUM_IP_TCP|CSUM_IP_SCTP| \
315 				 CSUM_IP6_UDP|CSUM_IP6_TCP|CSUM_IP6_SCTP)
316 struct iflib_txq {
317 	qidx_t		ift_in_use;
318 	qidx_t		ift_cidx;
319 	qidx_t		ift_cidx_processed;
320 	qidx_t		ift_pidx;
321 	uint8_t		ift_gen;
322 	uint8_t		ift_br_offset;
323 	uint16_t	ift_npending;
324 	uint16_t	ift_db_pending;
325 	uint16_t	ift_rs_pending;
326 	/* implicit pad */
327 	uint8_t		ift_txd_size[8];
328 	uint64_t	ift_processed;
329 	uint64_t	ift_cleaned;
330 	uint64_t	ift_cleaned_prev;
331 #if MEMORY_LOGGING
332 	uint64_t	ift_enqueued;
333 	uint64_t	ift_dequeued;
334 #endif
335 	uint64_t	ift_no_tx_dma_setup;
336 	uint64_t	ift_no_desc_avail;
337 	uint64_t	ift_mbuf_defrag_failed;
338 	uint64_t	ift_mbuf_defrag;
339 	uint64_t	ift_map_failed;
340 	uint64_t	ift_txd_encap_efbig;
341 	uint64_t	ift_pullups;
342 
343 	struct mtx	ift_mtx;
344 	struct mtx	ift_db_mtx;
345 
346 	/* constant values */
347 	if_ctx_t	ift_ctx;
348 	struct ifmp_ring        *ift_br;
349 	struct grouptask	ift_task;
350 	qidx_t		ift_size;
351 	uint16_t	ift_id;
352 	struct callout	ift_timer;
353 
354 	if_txsd_vec_t	ift_sds;
355 	uint8_t		ift_qstatus;
356 	uint8_t		ift_closed;
357 	uint8_t		ift_update_freq;
358 	struct iflib_filter_info ift_filter_info;
359 	bus_dma_tag_t		ift_desc_tag;
360 	bus_dma_tag_t		ift_tso_desc_tag;
361 	iflib_dma_info_t	ift_ifdi;
362 #define MTX_NAME_LEN 16
363 	char                    ift_mtx_name[MTX_NAME_LEN];
364 	char                    ift_db_mtx_name[MTX_NAME_LEN];
365 	bus_dma_segment_t	ift_segs[IFLIB_MAX_TX_SEGS]  __aligned(CACHE_LINE_SIZE);
366 #ifdef IFLIB_DIAGNOSTICS
367 	uint64_t ift_cpu_exec_count[256];
368 #endif
369 } __aligned(CACHE_LINE_SIZE);
370 
371 struct iflib_fl {
372 	qidx_t		ifl_cidx;
373 	qidx_t		ifl_pidx;
374 	qidx_t		ifl_credits;
375 	uint8_t		ifl_gen;
376 	uint8_t		ifl_rxd_size;
377 #if MEMORY_LOGGING
378 	uint64_t	ifl_m_enqueued;
379 	uint64_t	ifl_m_dequeued;
380 	uint64_t	ifl_cl_enqueued;
381 	uint64_t	ifl_cl_dequeued;
382 #endif
383 	/* implicit pad */
384 
385 	/* constant */
386 	qidx_t		ifl_size;
387 	uint16_t	ifl_buf_size;
388 	uint16_t	ifl_cltype;
389 	uma_zone_t	ifl_zone;
390 	iflib_rxsd_array_t	ifl_sds;
391 	iflib_rxq_t	ifl_rxq;
392 	uint8_t		ifl_id;
393 	bus_dma_tag_t           ifl_desc_tag;
394 	iflib_dma_info_t	ifl_ifdi;
395 	uint64_t	ifl_bus_addrs[IFLIB_MAX_RX_REFRESH] __aligned(CACHE_LINE_SIZE);
396 	caddr_t		ifl_vm_addrs[IFLIB_MAX_RX_REFRESH];
397 	qidx_t	ifl_rxd_idxs[IFLIB_MAX_RX_REFRESH];
398 }  __aligned(CACHE_LINE_SIZE);
399 
400 static inline qidx_t
401 get_inuse(int size, qidx_t cidx, qidx_t pidx, uint8_t gen)
402 {
403 	qidx_t used;
404 
405 	if (pidx > cidx)
406 		used = pidx - cidx;
407 	else if (pidx < cidx)
408 		used = size - cidx + pidx;
409 	else if (gen == 0 && pidx == cidx)
410 		used = 0;
411 	else if (gen == 1 && pidx == cidx)
412 		used = size;
413 	else
414 		panic("bad state");
415 
416 	return (used);
417 }
418 
419 #define TXQ_AVAIL(txq) (txq->ift_size - get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx, txq->ift_gen))
420 
421 #define IDXDIFF(head, tail, wrap) \
422 	((head) >= (tail) ? (head) - (tail) : (wrap) - (tail) + (head))
423 
424 struct iflib_rxq {
425 	/* If there is a separate completion queue -
426 	 * these are the cq cidx and pidx. Otherwise
427 	 * these are unused.
428 	 */
429 	qidx_t		ifr_size;
430 	qidx_t		ifr_cq_cidx;
431 	qidx_t		ifr_cq_pidx;
432 	uint8_t		ifr_cq_gen;
433 	uint8_t		ifr_fl_offset;
434 
435 	if_ctx_t	ifr_ctx;
436 	iflib_fl_t	ifr_fl;
437 	uint64_t	ifr_rx_irq;
438 	uint16_t	ifr_id;
439 	uint8_t		ifr_lro_enabled;
440 	uint8_t		ifr_nfl;
441 	uint8_t		ifr_ntxqirq;
442 	uint8_t		ifr_txqid[IFLIB_MAX_TX_SHARED_INTR];
443 	struct lro_ctrl			ifr_lc;
444 	struct grouptask        ifr_task;
445 	struct iflib_filter_info ifr_filter_info;
446 	iflib_dma_info_t		ifr_ifdi;
447 
448 	/* dynamically allocate if any drivers need a value substantially larger than this */
449 	struct if_rxd_frag	ifr_frags[IFLIB_MAX_RX_SEGS] __aligned(CACHE_LINE_SIZE);
450 #ifdef IFLIB_DIAGNOSTICS
451 	uint64_t ifr_cpu_exec_count[256];
452 #endif
453 }  __aligned(CACHE_LINE_SIZE);
454 
455 typedef struct if_rxsd {
456 	caddr_t *ifsd_cl;
457 	struct mbuf **ifsd_m;
458 	iflib_fl_t ifsd_fl;
459 	qidx_t ifsd_cidx;
460 } *if_rxsd_t;
461 
462 /* multiple of word size */
463 #ifdef __LP64__
464 #define PKT_INFO_SIZE	6
465 #define RXD_INFO_SIZE	5
466 #define PKT_TYPE uint64_t
467 #else
468 #define PKT_INFO_SIZE	11
469 #define RXD_INFO_SIZE	8
470 #define PKT_TYPE uint32_t
471 #endif
472 #define PKT_LOOP_BOUND  ((PKT_INFO_SIZE/3)*3)
473 #define RXD_LOOP_BOUND  ((RXD_INFO_SIZE/4)*4)
474 
475 typedef struct if_pkt_info_pad {
476 	PKT_TYPE pkt_val[PKT_INFO_SIZE];
477 } *if_pkt_info_pad_t;
478 typedef struct if_rxd_info_pad {
479 	PKT_TYPE rxd_val[RXD_INFO_SIZE];
480 } *if_rxd_info_pad_t;
481 
482 CTASSERT(sizeof(struct if_pkt_info_pad) == sizeof(struct if_pkt_info));
483 CTASSERT(sizeof(struct if_rxd_info_pad) == sizeof(struct if_rxd_info));
484 
485 
486 static inline void
487 pkt_info_zero(if_pkt_info_t pi)
488 {
489 	if_pkt_info_pad_t pi_pad;
490 
491 	pi_pad = (if_pkt_info_pad_t)pi;
492 	pi_pad->pkt_val[0] = 0; pi_pad->pkt_val[1] = 0; pi_pad->pkt_val[2] = 0;
493 	pi_pad->pkt_val[3] = 0; pi_pad->pkt_val[4] = 0; pi_pad->pkt_val[5] = 0;
494 #ifndef __LP64__
495 	pi_pad->pkt_val[6] = 0; pi_pad->pkt_val[7] = 0; pi_pad->pkt_val[8] = 0;
496 	pi_pad->pkt_val[9] = 0; pi_pad->pkt_val[10] = 0;
497 #endif
498 }
499 
500 static inline void
501 rxd_info_zero(if_rxd_info_t ri)
502 {
503 	if_rxd_info_pad_t ri_pad;
504 	int i;
505 
506 	ri_pad = (if_rxd_info_pad_t)ri;
507 	for (i = 0; i < RXD_LOOP_BOUND; i += 4) {
508 		ri_pad->rxd_val[i] = 0;
509 		ri_pad->rxd_val[i+1] = 0;
510 		ri_pad->rxd_val[i+2] = 0;
511 		ri_pad->rxd_val[i+3] = 0;
512 	}
513 #ifdef __LP64__
514 	ri_pad->rxd_val[RXD_INFO_SIZE-1] = 0;
515 #endif
516 }
517 
518 /*
519  * Only allow a single packet to take up most 1/nth of the tx ring
520  */
521 #define MAX_SINGLE_PACKET_FRACTION 12
522 #define IF_BAD_DMA (bus_addr_t)-1
523 
524 static int enable_msix = 1;
525 
526 #define CTX_ACTIVE(ctx) ((if_getdrvflags((ctx)->ifc_ifp) & IFF_DRV_RUNNING))
527 
528 #define CTX_LOCK_INIT(_sc, _name)  mtx_init(&(_sc)->ifc_mtx, _name, "iflib ctx lock", MTX_DEF)
529 
530 #define CTX_LOCK(ctx) mtx_lock(&(ctx)->ifc_mtx)
531 #define CTX_UNLOCK(ctx) mtx_unlock(&(ctx)->ifc_mtx)
532 #define CTX_LOCK_DESTROY(ctx) mtx_destroy(&(ctx)->ifc_mtx)
533 
534 
535 #define CALLOUT_LOCK(txq)	mtx_lock(&txq->ift_mtx)
536 #define CALLOUT_UNLOCK(txq) 	mtx_unlock(&txq->ift_mtx)
537 
538 
539 /* Our boot-time initialization hook */
540 static int	iflib_module_event_handler(module_t, int, void *);
541 
542 static moduledata_t iflib_moduledata = {
543 	"iflib",
544 	iflib_module_event_handler,
545 	NULL
546 };
547 
548 DECLARE_MODULE(iflib, iflib_moduledata, SI_SUB_INIT_IF, SI_ORDER_ANY);
549 MODULE_VERSION(iflib, 1);
550 
551 MODULE_DEPEND(iflib, pci, 1, 1, 1);
552 MODULE_DEPEND(iflib, ether, 1, 1, 1);
553 
554 TASKQGROUP_DEFINE(if_io_tqg, mp_ncpus, 1);
555 TASKQGROUP_DEFINE(if_config_tqg, 1, 1);
556 
557 #ifndef IFLIB_DEBUG_COUNTERS
558 #ifdef INVARIANTS
559 #define IFLIB_DEBUG_COUNTERS 1
560 #else
561 #define IFLIB_DEBUG_COUNTERS 0
562 #endif /* !INVARIANTS */
563 #endif
564 
565 static SYSCTL_NODE(_net, OID_AUTO, iflib, CTLFLAG_RD, 0,
566                    "iflib driver parameters");
567 
568 /*
569  * XXX need to ensure that this can't accidentally cause the head to be moved backwards
570  */
571 static int iflib_min_tx_latency = 0;
572 SYSCTL_INT(_net_iflib, OID_AUTO, min_tx_latency, CTLFLAG_RW,
573 		   &iflib_min_tx_latency, 0, "minimize transmit latency at the possible expense of throughput");
574 static int iflib_no_tx_batch = 0;
575 SYSCTL_INT(_net_iflib, OID_AUTO, no_tx_batch, CTLFLAG_RW,
576 		   &iflib_no_tx_batch, 0, "minimize transmit latency at the possible expense of throughput");
577 
578 
579 #if IFLIB_DEBUG_COUNTERS
580 
581 static int iflib_tx_seen;
582 static int iflib_tx_sent;
583 static int iflib_tx_encap;
584 static int iflib_rx_allocs;
585 static int iflib_fl_refills;
586 static int iflib_fl_refills_large;
587 static int iflib_tx_frees;
588 
589 SYSCTL_INT(_net_iflib, OID_AUTO, tx_seen, CTLFLAG_RD,
590 		   &iflib_tx_seen, 0, "# tx mbufs seen");
591 SYSCTL_INT(_net_iflib, OID_AUTO, tx_sent, CTLFLAG_RD,
592 		   &iflib_tx_sent, 0, "# tx mbufs sent");
593 SYSCTL_INT(_net_iflib, OID_AUTO, tx_encap, CTLFLAG_RD,
594 		   &iflib_tx_encap, 0, "# tx mbufs encapped");
595 SYSCTL_INT(_net_iflib, OID_AUTO, tx_frees, CTLFLAG_RD,
596 		   &iflib_tx_frees, 0, "# tx frees");
597 SYSCTL_INT(_net_iflib, OID_AUTO, rx_allocs, CTLFLAG_RD,
598 		   &iflib_rx_allocs, 0, "# rx allocations");
599 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills, CTLFLAG_RD,
600 		   &iflib_fl_refills, 0, "# refills");
601 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills_large, CTLFLAG_RD,
602 		   &iflib_fl_refills_large, 0, "# large refills");
603 
604 
605 static int iflib_txq_drain_flushing;
606 static int iflib_txq_drain_oactive;
607 static int iflib_txq_drain_notready;
608 static int iflib_txq_drain_encapfail;
609 
610 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_flushing, CTLFLAG_RD,
611 		   &iflib_txq_drain_flushing, 0, "# drain flushes");
612 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_oactive, CTLFLAG_RD,
613 		   &iflib_txq_drain_oactive, 0, "# drain oactives");
614 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_notready, CTLFLAG_RD,
615 		   &iflib_txq_drain_notready, 0, "# drain notready");
616 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_encapfail, CTLFLAG_RD,
617 		   &iflib_txq_drain_encapfail, 0, "# drain encap fails");
618 
619 
620 static int iflib_encap_load_mbuf_fail;
621 static int iflib_encap_txq_avail_fail;
622 static int iflib_encap_txd_encap_fail;
623 
624 SYSCTL_INT(_net_iflib, OID_AUTO, encap_load_mbuf_fail, CTLFLAG_RD,
625 		   &iflib_encap_load_mbuf_fail, 0, "# busdma load failures");
626 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txq_avail_fail, CTLFLAG_RD,
627 		   &iflib_encap_txq_avail_fail, 0, "# txq avail failures");
628 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txd_encap_fail, CTLFLAG_RD,
629 		   &iflib_encap_txd_encap_fail, 0, "# driver encap failures");
630 
631 static int iflib_task_fn_rxs;
632 static int iflib_rx_intr_enables;
633 static int iflib_fast_intrs;
634 static int iflib_intr_link;
635 static int iflib_intr_msix;
636 static int iflib_rx_unavail;
637 static int iflib_rx_ctx_inactive;
638 static int iflib_rx_zero_len;
639 static int iflib_rx_if_input;
640 static int iflib_rx_mbuf_null;
641 static int iflib_rxd_flush;
642 
643 static int iflib_verbose_debug;
644 
645 SYSCTL_INT(_net_iflib, OID_AUTO, intr_link, CTLFLAG_RD,
646 		   &iflib_intr_link, 0, "# intr link calls");
647 SYSCTL_INT(_net_iflib, OID_AUTO, intr_msix, CTLFLAG_RD,
648 		   &iflib_intr_msix, 0, "# intr msix calls");
649 SYSCTL_INT(_net_iflib, OID_AUTO, task_fn_rx, CTLFLAG_RD,
650 		   &iflib_task_fn_rxs, 0, "# task_fn_rx calls");
651 SYSCTL_INT(_net_iflib, OID_AUTO, rx_intr_enables, CTLFLAG_RD,
652 		   &iflib_rx_intr_enables, 0, "# rx intr enables");
653 SYSCTL_INT(_net_iflib, OID_AUTO, fast_intrs, CTLFLAG_RD,
654 		   &iflib_fast_intrs, 0, "# fast_intr calls");
655 SYSCTL_INT(_net_iflib, OID_AUTO, rx_unavail, CTLFLAG_RD,
656 		   &iflib_rx_unavail, 0, "# times rxeof called with no available data");
657 SYSCTL_INT(_net_iflib, OID_AUTO, rx_ctx_inactive, CTLFLAG_RD,
658 		   &iflib_rx_ctx_inactive, 0, "# times rxeof called with inactive context");
659 SYSCTL_INT(_net_iflib, OID_AUTO, rx_zero_len, CTLFLAG_RD,
660 		   &iflib_rx_zero_len, 0, "# times rxeof saw zero len mbuf");
661 SYSCTL_INT(_net_iflib, OID_AUTO, rx_if_input, CTLFLAG_RD,
662 		   &iflib_rx_if_input, 0, "# times rxeof called if_input");
663 SYSCTL_INT(_net_iflib, OID_AUTO, rx_mbuf_null, CTLFLAG_RD,
664 		   &iflib_rx_mbuf_null, 0, "# times rxeof got null mbuf");
665 SYSCTL_INT(_net_iflib, OID_AUTO, rxd_flush, CTLFLAG_RD,
666 	         &iflib_rxd_flush, 0, "# times rxd_flush called");
667 SYSCTL_INT(_net_iflib, OID_AUTO, verbose_debug, CTLFLAG_RW,
668 		   &iflib_verbose_debug, 0, "enable verbose debugging");
669 
670 #define DBG_COUNTER_INC(name) atomic_add_int(&(iflib_ ## name), 1)
671 static void
672 iflib_debug_reset(void)
673 {
674 	iflib_tx_seen = iflib_tx_sent = iflib_tx_encap = iflib_rx_allocs =
675 		iflib_fl_refills = iflib_fl_refills_large = iflib_tx_frees =
676 		iflib_txq_drain_flushing = iflib_txq_drain_oactive =
677 		iflib_txq_drain_notready = iflib_txq_drain_encapfail =
678 		iflib_encap_load_mbuf_fail = iflib_encap_txq_avail_fail =
679 		iflib_encap_txd_encap_fail = iflib_task_fn_rxs = iflib_rx_intr_enables =
680 		iflib_fast_intrs = iflib_intr_link = iflib_intr_msix = iflib_rx_unavail =
681 		iflib_rx_ctx_inactive = iflib_rx_zero_len = iflib_rx_if_input =
682 		iflib_rx_mbuf_null = iflib_rxd_flush = 0;
683 }
684 
685 #else
686 #define DBG_COUNTER_INC(name)
687 static void iflib_debug_reset(void) {}
688 #endif
689 
690 
691 
692 #define IFLIB_DEBUG 0
693 
694 static void iflib_tx_structures_free(if_ctx_t ctx);
695 static void iflib_rx_structures_free(if_ctx_t ctx);
696 static int iflib_queues_alloc(if_ctx_t ctx);
697 static int iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq);
698 static int iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget);
699 static int iflib_qset_structures_setup(if_ctx_t ctx);
700 static int iflib_msix_init(if_ctx_t ctx);
701 static int iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filterarg, int *rid, char *str);
702 static void iflib_txq_check_drain(iflib_txq_t txq, int budget);
703 static uint32_t iflib_txq_can_drain(struct ifmp_ring *);
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_stop(if_ctx_t ctx);
711 static void iflib_if_init_locked(if_ctx_t ctx);
712 #ifndef __NO_STRICT_ALIGNMENT
713 static struct mbuf * iflib_fixup_rx(struct mbuf *m);
714 #endif
715 
716 #ifdef DEV_NETMAP
717 #include <sys/selinfo.h>
718 #include <net/netmap.h>
719 #include <dev/netmap/netmap_kern.h>
720 
721 MODULE_DEPEND(iflib, netmap, 1, 1, 1);
722 
723 /*
724  * device-specific sysctl variables:
725  *
726  * iflib_crcstrip: 0: keep CRC in rx frames (default), 1: strip it.
727  *	During regular operations the CRC is stripped, but on some
728  *	hardware reception of frames not multiple of 64 is slower,
729  *	so using crcstrip=0 helps in benchmarks.
730  *
731  * iflib_rx_miss, iflib_rx_miss_bufs:
732  *	count packets that might be missed due to lost interrupts.
733  */
734 SYSCTL_DECL(_dev_netmap);
735 /*
736  * The xl driver by default strips CRCs and we do not override it.
737  */
738 
739 int iflib_crcstrip = 1;
740 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_crcstrip,
741     CTLFLAG_RW, &iflib_crcstrip, 1, "strip CRC on rx frames");
742 
743 int iflib_rx_miss, iflib_rx_miss_bufs;
744 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss,
745     CTLFLAG_RW, &iflib_rx_miss, 0, "potentially missed rx intr");
746 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss_bufs,
747     CTLFLAG_RW, &iflib_rx_miss_bufs, 0, "potentially missed rx intr bufs");
748 
749 /*
750  * Register/unregister. We are already under netmap lock.
751  * Only called on the first register or the last unregister.
752  */
753 static int
754 iflib_netmap_register(struct netmap_adapter *na, int onoff)
755 {
756 	struct ifnet *ifp = na->ifp;
757 	if_ctx_t ctx = ifp->if_softc;
758 	int status;
759 
760 	CTX_LOCK(ctx);
761 	IFDI_INTR_DISABLE(ctx);
762 
763 	/* Tell the stack that the interface is no longer active */
764 	ifp->if_drv_flags &= ~(IFF_DRV_RUNNING | IFF_DRV_OACTIVE);
765 
766 	if (!CTX_IS_VF(ctx))
767 		IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip);
768 
769 	/* enable or disable flags and callbacks in na and ifp */
770 	if (onoff) {
771 		nm_set_native_flags(na);
772 	} else {
773 		nm_clear_native_flags(na);
774 	}
775 	iflib_stop(ctx);
776 	iflib_init_locked(ctx);
777 	IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip); // XXX why twice ?
778 	status = ifp->if_drv_flags & IFF_DRV_RUNNING ? 0 : 1;
779 	if (status)
780 		nm_clear_native_flags(na);
781 	CTX_UNLOCK(ctx);
782 	return (status);
783 }
784 
785 /*
786  * Reconcile kernel and user view of the transmit ring.
787  *
788  * All information is in the kring.
789  * Userspace wants to send packets up to the one before kring->rhead,
790  * kernel knows kring->nr_hwcur is the first unsent packet.
791  *
792  * Here we push packets out (as many as possible), and possibly
793  * reclaim buffers from previously completed transmission.
794  *
795  * The caller (netmap) guarantees that there is only one instance
796  * running at any time. Any interference with other driver
797  * methods should be handled by the individual drivers.
798  */
799 static int
800 iflib_netmap_txsync(struct netmap_kring *kring, int flags)
801 {
802 	struct netmap_adapter *na = kring->na;
803 	struct ifnet *ifp = na->ifp;
804 	struct netmap_ring *ring = kring->ring;
805 	u_int nm_i;	/* index into the netmap ring */
806 	u_int nic_i;	/* index into the NIC ring */
807 	u_int n;
808 	u_int const lim = kring->nkr_num_slots - 1;
809 	u_int const head = kring->rhead;
810 	struct if_pkt_info pi;
811 
812 	/*
813 	 * interrupts on every tx packet are expensive so request
814 	 * them every half ring, or where NS_REPORT is set
815 	 */
816 	u_int report_frequency = kring->nkr_num_slots >> 1;
817 	/* device-specific */
818 	if_ctx_t ctx = ifp->if_softc;
819 	iflib_txq_t txq = &ctx->ifc_txqs[kring->ring_id];
820 
821 	if (txq->ift_sds.ifsd_map)
822 		bus_dmamap_sync(txq->ift_desc_tag, txq->ift_ifdi->idi_map,
823 				BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
824 
825 
826 	/*
827 	 * First part: process new packets to send.
828 	 * nm_i is the current index in the netmap ring,
829 	 * nic_i is the corresponding index in the NIC ring.
830 	 *
831 	 * If we have packets to send (nm_i != head)
832 	 * iterate over the netmap ring, fetch length and update
833 	 * the corresponding slot in the NIC ring. Some drivers also
834 	 * need to update the buffer's physical address in the NIC slot
835 	 * even NS_BUF_CHANGED is not set (PNMB computes the addresses).
836 	 *
837 	 * The netmap_reload_map() calls is especially expensive,
838 	 * even when (as in this case) the tag is 0, so do only
839 	 * when the buffer has actually changed.
840 	 *
841 	 * If possible do not set the report/intr bit on all slots,
842 	 * but only a few times per ring or when NS_REPORT is set.
843 	 *
844 	 * Finally, on 10G and faster drivers, it might be useful
845 	 * to prefetch the next slot and txr entry.
846 	 */
847 
848 	nm_i = kring->nr_hwcur;
849 	pkt_info_zero(&pi);
850 	pi.ipi_segs = txq->ift_segs;
851 	pi.ipi_qsidx = kring->ring_id;
852 	if (nm_i != head) {	/* we have new packets to send */
853 		nic_i = netmap_idx_k2n(kring, nm_i);
854 
855 		__builtin_prefetch(&ring->slot[nm_i]);
856 		__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i]);
857 		if (txq->ift_sds.ifsd_map)
858 			__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i]);
859 
860 		for (n = 0; nm_i != head; n++) {
861 			struct netmap_slot *slot = &ring->slot[nm_i];
862 			u_int len = slot->len;
863 			uint64_t paddr;
864 			void *addr = PNMB(na, slot, &paddr);
865 			int flags = (slot->flags & NS_REPORT ||
866 				nic_i == 0 || nic_i == report_frequency) ?
867 				IPI_TX_INTR : 0;
868 
869 			/* device-specific */
870 			pi.ipi_len = len;
871 			pi.ipi_segs[0].ds_addr = paddr;
872 			pi.ipi_segs[0].ds_len = len;
873 			pi.ipi_nsegs = 1;
874 			pi.ipi_ndescs = 0;
875 			pi.ipi_pidx = nic_i;
876 			pi.ipi_flags = flags;
877 
878 			/* Fill the slot in the NIC ring. */
879 			ctx->isc_txd_encap(ctx->ifc_softc, &pi);
880 
881 			/* prefetch for next round */
882 			__builtin_prefetch(&ring->slot[nm_i + 1]);
883 			__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i + 1]);
884 			if (txq->ift_sds.ifsd_map) {
885 				__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i + 1]);
886 
887 				NM_CHECK_ADDR_LEN(na, addr, len);
888 
889 				if (slot->flags & NS_BUF_CHANGED) {
890 					/* buffer has changed, reload map */
891 					netmap_reload_map(na, txq->ift_desc_tag, txq->ift_sds.ifsd_map[nic_i], addr);
892 				}
893 				/* make sure changes to the buffer are synced */
894 				bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_sds.ifsd_map[nic_i],
895 						BUS_DMASYNC_PREWRITE);
896 			}
897 			slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED);
898 			nm_i = nm_next(nm_i, lim);
899 			nic_i = nm_next(nic_i, lim);
900 		}
901 		kring->nr_hwcur = head;
902 
903 		/* synchronize the NIC ring */
904 		if (txq->ift_sds.ifsd_map)
905 			bus_dmamap_sync(txq->ift_desc_tag, txq->ift_ifdi->idi_map,
906 						BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
907 
908 		/* (re)start the tx unit up to slot nic_i (excluded) */
909 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, nic_i);
910 	}
911 
912 	/*
913 	 * Second part: reclaim buffers for completed transmissions.
914 	 */
915 	if (iflib_tx_credits_update(ctx, txq)) {
916 		/* some tx completed, increment avail */
917 		nic_i = txq->ift_cidx_processed;
918 		kring->nr_hwtail = nm_prev(netmap_idx_n2k(kring, nic_i), lim);
919 	}
920 	return (0);
921 }
922 
923 /*
924  * Reconcile kernel and user view of the receive ring.
925  * Same as for the txsync, this routine must be efficient.
926  * The caller guarantees a single invocations, but races against
927  * the rest of the driver should be handled here.
928  *
929  * On call, kring->rhead is the first packet that userspace wants
930  * to keep, and kring->rcur is the wakeup point.
931  * The kernel has previously reported packets up to kring->rtail.
932  *
933  * If (flags & NAF_FORCE_READ) also check for incoming packets irrespective
934  * of whether or not we received an interrupt.
935  */
936 static int
937 iflib_netmap_rxsync(struct netmap_kring *kring, int flags)
938 {
939 	struct netmap_adapter *na = kring->na;
940 	struct netmap_ring *ring = kring->ring;
941 	uint32_t nm_i;	/* index into the netmap ring */
942 	uint32_t nic_i, nic_i_start;	/* index into the NIC ring */
943 	u_int i, n;
944 	u_int const lim = kring->nkr_num_slots - 1;
945 	u_int const head = kring->rhead;
946 	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
947 	struct if_rxd_info ri;
948 	struct if_rxd_update iru;
949 
950 	struct ifnet *ifp = na->ifp;
951 	if_ctx_t ctx = ifp->if_softc;
952 	iflib_rxq_t rxq = &ctx->ifc_rxqs[kring->ring_id];
953 	iflib_fl_t fl = rxq->ifr_fl;
954 	if (head > lim)
955 		return netmap_ring_reinit(kring);
956 
957 	/* XXX check sync modes */
958 	for (i = 0, fl = rxq->ifr_fl; i < rxq->ifr_nfl; i++, fl++) {
959 		if (fl->ifl_sds.ifsd_map == NULL)
960 			continue;
961 		bus_dmamap_sync(rxq->ifr_fl[i].ifl_desc_tag, fl->ifl_ifdi->idi_map,
962 				BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
963 	}
964 	/*
965 	 * First part: import newly received packets.
966 	 *
967 	 * nm_i is the index of the next free slot in the netmap ring,
968 	 * nic_i is the index of the next received packet in the NIC ring,
969 	 * and they may differ in case if_init() has been called while
970 	 * in netmap mode. For the receive ring we have
971 	 *
972 	 *	nic_i = rxr->next_check;
973 	 *	nm_i = kring->nr_hwtail (previous)
974 	 * and
975 	 *	nm_i == (nic_i + kring->nkr_hwofs) % ring_size
976 	 *
977 	 * rxr->next_check is set to 0 on a ring reinit
978 	 */
979 	if (netmap_no_pendintr || force_update) {
980 		int crclen = iflib_crcstrip ? 0 : 4;
981 		int error, avail;
982 		uint16_t slot_flags = kring->nkr_slot_flags;
983 
984 		for (fl = rxq->ifr_fl, i = 0; i < rxq->ifr_nfl; i++, fl++) {
985 			nic_i = fl->ifl_cidx;
986 			nm_i = netmap_idx_n2k(kring, nic_i);
987 			avail = iflib_rxd_avail(ctx, rxq, nic_i, USHRT_MAX);
988 			for (n = 0; avail > 0; n++, avail--) {
989 				rxd_info_zero(&ri);
990 				ri.iri_frags = rxq->ifr_frags;
991 				ri.iri_qsidx = kring->ring_id;
992 				ri.iri_ifp = ctx->ifc_ifp;
993 				ri.iri_cidx = nic_i;
994 
995 				error = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
996 				ring->slot[nm_i].len = error ? 0 : ri.iri_len - crclen;
997 				ring->slot[nm_i].flags = slot_flags;
998 				if (fl->ifl_sds.ifsd_map)
999 					bus_dmamap_sync(fl->ifl_ifdi->idi_tag,
1000 							fl->ifl_sds.ifsd_map[nic_i], BUS_DMASYNC_POSTREAD);
1001 				nm_i = nm_next(nm_i, lim);
1002 				nic_i = nm_next(nic_i, lim);
1003 			}
1004 			if (n) { /* update the state variables */
1005 				if (netmap_no_pendintr && !force_update) {
1006 					/* diagnostics */
1007 					iflib_rx_miss ++;
1008 					iflib_rx_miss_bufs += n;
1009 				}
1010 				fl->ifl_cidx = nic_i;
1011 				kring->nr_hwtail = nm_i;
1012 			}
1013 			kring->nr_kflags &= ~NKR_PENDINTR;
1014 		}
1015 	}
1016 	/*
1017 	 * Second part: skip past packets that userspace has released.
1018 	 * (kring->nr_hwcur to head excluded),
1019 	 * and make the buffers available for reception.
1020 	 * As usual nm_i is the index in the netmap ring,
1021 	 * nic_i is the index in the NIC ring, and
1022 	 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1023 	 */
1024 	/* XXX not sure how this will work with multiple free lists */
1025 	nm_i = kring->nr_hwcur;
1026 	if (nm_i == head)
1027 		return (0);
1028 
1029 	iru.iru_paddrs = fl->ifl_bus_addrs;
1030 	iru.iru_vaddrs = &fl->ifl_vm_addrs[0];
1031 	iru.iru_idxs = fl->ifl_rxd_idxs;
1032 	iru.iru_qsidx = rxq->ifr_id;
1033 	iru.iru_buf_size = fl->ifl_buf_size;
1034 	iru.iru_flidx = fl->ifl_id;
1035 	nic_i_start = nic_i = netmap_idx_k2n(kring, nm_i);
1036 	for (i = 0; nm_i != head; i++) {
1037 		struct netmap_slot *slot = &ring->slot[nm_i];
1038 		void *addr = PNMB(na, slot, &fl->ifl_bus_addrs[i]);
1039 
1040 		if (addr == NETMAP_BUF_BASE(na)) /* bad buf */
1041 			goto ring_reset;
1042 
1043 		fl->ifl_vm_addrs[i] = addr;
1044 		if (fl->ifl_sds.ifsd_map && (slot->flags & NS_BUF_CHANGED)) {
1045 			/* buffer has changed, reload map */
1046 			netmap_reload_map(na, fl->ifl_ifdi->idi_tag, fl->ifl_sds.ifsd_map[nic_i], addr);
1047 		}
1048 		slot->flags &= ~NS_BUF_CHANGED;
1049 
1050 		nm_i = nm_next(nm_i, lim);
1051 		fl->ifl_rxd_idxs[i] = nic_i = nm_next(nic_i, lim);
1052 		if (nm_i != head && i < IFLIB_MAX_RX_REFRESH)
1053 			continue;
1054 
1055 		iru.iru_pidx = nic_i_start;
1056 		iru.iru_count = i;
1057 		i = 0;
1058 		ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
1059 		if (fl->ifl_sds.ifsd_map == NULL) {
1060 			nic_i_start = nic_i;
1061 			continue;
1062 		}
1063 		nic_i = nic_i_start;
1064 		for (n = 0; n < iru.iru_count; n++) {
1065 			bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_sds.ifsd_map[nic_i],
1066 					BUS_DMASYNC_PREREAD);
1067 			nic_i = nm_next(nic_i, lim);
1068 		}
1069 		nic_i_start = nic_i;
1070 	}
1071 	kring->nr_hwcur = head;
1072 
1073 	if (fl->ifl_sds.ifsd_map)
1074 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
1075 				BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1076 	/*
1077 	 * IMPORTANT: we must leave one free slot in the ring,
1078 	 * so move nic_i back by one unit
1079 	 */
1080 	nic_i = nm_prev(nic_i, lim);
1081 	ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id, nic_i);
1082 	return 0;
1083 
1084 ring_reset:
1085 	return netmap_ring_reinit(kring);
1086 }
1087 
1088 static void
1089 iflib_netmap_intr(struct netmap_adapter *na, int onoff)
1090 {
1091 	struct ifnet *ifp = na->ifp;
1092 	if_ctx_t ctx = ifp->if_softc;
1093 
1094 	CTX_LOCK(ctx);
1095 	if (onoff) {
1096 		IFDI_INTR_ENABLE(ctx);
1097 	} else {
1098 		IFDI_INTR_DISABLE(ctx);
1099 	}
1100 	CTX_UNLOCK(ctx);
1101 }
1102 
1103 
1104 static int
1105 iflib_netmap_attach(if_ctx_t ctx)
1106 {
1107 	struct netmap_adapter na;
1108 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1109 
1110 	bzero(&na, sizeof(na));
1111 
1112 	na.ifp = ctx->ifc_ifp;
1113 	na.na_flags = NAF_BDG_MAYSLEEP;
1114 	MPASS(ctx->ifc_softc_ctx.isc_ntxqsets);
1115 	MPASS(ctx->ifc_softc_ctx.isc_nrxqsets);
1116 
1117 	na.num_tx_desc = scctx->isc_ntxd[0];
1118 	na.num_rx_desc = scctx->isc_nrxd[0];
1119 	na.nm_txsync = iflib_netmap_txsync;
1120 	na.nm_rxsync = iflib_netmap_rxsync;
1121 	na.nm_register = iflib_netmap_register;
1122 	na.nm_intr = iflib_netmap_intr;
1123 	na.num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
1124 	na.num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
1125 	return (netmap_attach(&na));
1126 }
1127 
1128 static void
1129 iflib_netmap_txq_init(if_ctx_t ctx, iflib_txq_t txq)
1130 {
1131 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1132 	struct netmap_slot *slot;
1133 
1134 	slot = netmap_reset(na, NR_TX, txq->ift_id, 0);
1135 	if (slot == NULL)
1136 		return;
1137 	if (txq->ift_sds.ifsd_map == NULL)
1138 		return;
1139 
1140 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxd[0]; i++) {
1141 
1142 		/*
1143 		 * In netmap mode, set the map for the packet buffer.
1144 		 * NOTE: Some drivers (not this one) also need to set
1145 		 * the physical buffer address in the NIC ring.
1146 		 * netmap_idx_n2k() maps a nic index, i, into the corresponding
1147 		 * netmap slot index, si
1148 		 */
1149 		int si = netmap_idx_n2k(&na->tx_rings[txq->ift_id], i);
1150 		netmap_load_map(na, txq->ift_desc_tag, txq->ift_sds.ifsd_map[i], NMB(na, slot + si));
1151 	}
1152 }
1153 static void
1154 iflib_netmap_rxq_init(if_ctx_t ctx, iflib_rxq_t rxq)
1155 {
1156 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1157 	struct netmap_slot *slot;
1158 	struct if_rxd_update iru;
1159 	iflib_fl_t fl;
1160 	bus_dmamap_t *map;
1161 	int nrxd;
1162 	uint32_t i, j, pidx_start;
1163 
1164 	slot = netmap_reset(na, NR_RX, rxq->ifr_id, 0);
1165 	if (slot == NULL)
1166 		return;
1167 	fl = &rxq->ifr_fl[0];
1168 	map = fl->ifl_sds.ifsd_map;
1169 	nrxd = ctx->ifc_softc_ctx.isc_nrxd[0];
1170 	iru.iru_paddrs = fl->ifl_bus_addrs;
1171 	iru.iru_vaddrs = &fl->ifl_vm_addrs[0];
1172 	iru.iru_idxs = fl->ifl_rxd_idxs;
1173 	iru.iru_qsidx = rxq->ifr_id;
1174 	iru.iru_buf_size = rxq->ifr_fl[0].ifl_buf_size;
1175 	iru.iru_flidx = 0;
1176 
1177 	for (pidx_start = i = j = 0; i < nrxd; i++, j++) {
1178 		int sj = netmap_idx_n2k(&na->rx_rings[rxq->ifr_id], i);
1179 		void *addr;
1180 
1181 		fl->ifl_rxd_idxs[j] = i;
1182 		addr = fl->ifl_vm_addrs[j] = PNMB(na, slot + sj, &fl->ifl_bus_addrs[j]);
1183 		if (map) {
1184 			netmap_load_map(na, rxq->ifr_fl[0].ifl_ifdi->idi_tag, *map, addr);
1185 			map++;
1186 		}
1187 
1188 		if (j < IFLIB_MAX_RX_REFRESH && i < nrxd - 1)
1189 			continue;
1190 
1191 		iru.iru_pidx = pidx_start;
1192 		pidx_start = i;
1193 		iru.iru_count = j;
1194 		j = 0;
1195 		MPASS(pidx_start + j <= nrxd);
1196 		/* Update descriptors and the cached value */
1197 		ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
1198 	}
1199 	/* preserve queue */
1200 	if (ctx->ifc_ifp->if_capenable & IFCAP_NETMAP) {
1201 		struct netmap_kring *kring = &na->rx_rings[rxq->ifr_id];
1202 		int t = na->num_rx_desc - 1 - nm_kr_rxspace(kring);
1203 		ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, 0 /* fl_id */, t);
1204 	} else
1205 		ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, 0 /* fl_id */, nrxd-1);
1206 }
1207 
1208 #define iflib_netmap_detach(ifp) netmap_detach(ifp)
1209 
1210 #else
1211 #define iflib_netmap_txq_init(ctx, txq)
1212 #define iflib_netmap_rxq_init(ctx, rxq)
1213 #define iflib_netmap_detach(ifp)
1214 
1215 #define iflib_netmap_attach(ctx) (0)
1216 #define netmap_rx_irq(ifp, qid, budget) (0)
1217 #define netmap_tx_irq(ifp, qid) do {} while (0)
1218 
1219 #endif
1220 
1221 #if defined(__i386__) || defined(__amd64__)
1222 static __inline void
1223 prefetch(void *x)
1224 {
1225 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1226 }
1227 #else
1228 #define prefetch(x)
1229 #endif
1230 
1231 static void
1232 _iflib_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int err)
1233 {
1234 	if (err)
1235 		return;
1236 	*(bus_addr_t *) arg = segs[0].ds_addr;
1237 }
1238 
1239 int
1240 iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags)
1241 {
1242 	int err;
1243 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1244 	device_t dev = ctx->ifc_dev;
1245 
1246 	KASSERT(sctx->isc_q_align != 0, ("alignment value not initialized"));
1247 
1248 	err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1249 				sctx->isc_q_align, 0,	/* alignment, bounds */
1250 				BUS_SPACE_MAXADDR,	/* lowaddr */
1251 				BUS_SPACE_MAXADDR,	/* highaddr */
1252 				NULL, NULL,		/* filter, filterarg */
1253 				size,			/* maxsize */
1254 				1,			/* nsegments */
1255 				size,			/* maxsegsize */
1256 				BUS_DMA_ALLOCNOW,	/* flags */
1257 				NULL,			/* lockfunc */
1258 				NULL,			/* lockarg */
1259 				&dma->idi_tag);
1260 	if (err) {
1261 		device_printf(dev,
1262 		    "%s: bus_dma_tag_create failed: %d\n",
1263 		    __func__, err);
1264 		goto fail_0;
1265 	}
1266 
1267 	err = bus_dmamem_alloc(dma->idi_tag, (void**) &dma->idi_vaddr,
1268 	    BUS_DMA_NOWAIT | BUS_DMA_COHERENT | BUS_DMA_ZERO, &dma->idi_map);
1269 	if (err) {
1270 		device_printf(dev,
1271 		    "%s: bus_dmamem_alloc(%ju) failed: %d\n",
1272 		    __func__, (uintmax_t)size, err);
1273 		goto fail_1;
1274 	}
1275 
1276 	dma->idi_paddr = IF_BAD_DMA;
1277 	err = bus_dmamap_load(dma->idi_tag, dma->idi_map, dma->idi_vaddr,
1278 	    size, _iflib_dmamap_cb, &dma->idi_paddr, mapflags | BUS_DMA_NOWAIT);
1279 	if (err || dma->idi_paddr == IF_BAD_DMA) {
1280 		device_printf(dev,
1281 		    "%s: bus_dmamap_load failed: %d\n",
1282 		    __func__, err);
1283 		goto fail_2;
1284 	}
1285 
1286 	dma->idi_size = size;
1287 	return (0);
1288 
1289 fail_2:
1290 	bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1291 fail_1:
1292 	bus_dma_tag_destroy(dma->idi_tag);
1293 fail_0:
1294 	dma->idi_tag = NULL;
1295 
1296 	return (err);
1297 }
1298 
1299 int
1300 iflib_dma_alloc_multi(if_ctx_t ctx, int *sizes, iflib_dma_info_t *dmalist, int mapflags, int count)
1301 {
1302 	int i, err;
1303 	iflib_dma_info_t *dmaiter;
1304 
1305 	dmaiter = dmalist;
1306 	for (i = 0; i < count; i++, dmaiter++) {
1307 		if ((err = iflib_dma_alloc(ctx, sizes[i], *dmaiter, mapflags)) != 0)
1308 			break;
1309 	}
1310 	if (err)
1311 		iflib_dma_free_multi(dmalist, i);
1312 	return (err);
1313 }
1314 
1315 void
1316 iflib_dma_free(iflib_dma_info_t dma)
1317 {
1318 	if (dma->idi_tag == NULL)
1319 		return;
1320 	if (dma->idi_paddr != IF_BAD_DMA) {
1321 		bus_dmamap_sync(dma->idi_tag, dma->idi_map,
1322 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1323 		bus_dmamap_unload(dma->idi_tag, dma->idi_map);
1324 		dma->idi_paddr = IF_BAD_DMA;
1325 	}
1326 	if (dma->idi_vaddr != NULL) {
1327 		bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1328 		dma->idi_vaddr = NULL;
1329 	}
1330 	bus_dma_tag_destroy(dma->idi_tag);
1331 	dma->idi_tag = NULL;
1332 }
1333 
1334 void
1335 iflib_dma_free_multi(iflib_dma_info_t *dmalist, int count)
1336 {
1337 	int i;
1338 	iflib_dma_info_t *dmaiter = dmalist;
1339 
1340 	for (i = 0; i < count; i++, dmaiter++)
1341 		iflib_dma_free(*dmaiter);
1342 }
1343 
1344 #ifdef EARLY_AP_STARTUP
1345 static const int iflib_started = 1;
1346 #else
1347 /*
1348  * We used to abuse the smp_started flag to decide if the queues have been
1349  * fully initialized (by late taskqgroup_adjust() calls in a SYSINIT()).
1350  * That gave bad races, since the SYSINIT() runs strictly after smp_started
1351  * is set.  Run a SYSINIT() strictly after that to just set a usable
1352  * completion flag.
1353  */
1354 
1355 static int iflib_started;
1356 
1357 static void
1358 iflib_record_started(void *arg)
1359 {
1360 	iflib_started = 1;
1361 }
1362 
1363 SYSINIT(iflib_record_started, SI_SUB_SMP + 1, SI_ORDER_FIRST,
1364 	iflib_record_started, NULL);
1365 #endif
1366 
1367 static int
1368 iflib_fast_intr(void *arg)
1369 {
1370 	iflib_filter_info_t info = arg;
1371 	struct grouptask *gtask = info->ifi_task;
1372 	if (!iflib_started)
1373 		return (FILTER_HANDLED);
1374 
1375 	DBG_COUNTER_INC(fast_intrs);
1376 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1377 		return (FILTER_HANDLED);
1378 
1379 	GROUPTASK_ENQUEUE(gtask);
1380 	return (FILTER_HANDLED);
1381 }
1382 
1383 static int
1384 iflib_fast_intr_rxtx(void *arg)
1385 {
1386 	iflib_filter_info_t info = arg;
1387 	struct grouptask *gtask = info->ifi_task;
1388 	iflib_rxq_t rxq = (iflib_rxq_t)info->ifi_ctx;
1389 	if_ctx_t ctx;
1390 	int i, cidx;
1391 
1392 	if (!iflib_started)
1393 		return (FILTER_HANDLED);
1394 
1395 	DBG_COUNTER_INC(fast_intrs);
1396 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1397 		return (FILTER_HANDLED);
1398 
1399 	for (i = 0; i < rxq->ifr_ntxqirq; i++) {
1400 		qidx_t txqid = rxq->ifr_txqid[i];
1401 
1402 		ctx = rxq->ifr_ctx;
1403 
1404 		if (!ctx->isc_txd_credits_update(ctx->ifc_softc, txqid, false)) {
1405 			IFDI_TX_QUEUE_INTR_ENABLE(ctx, txqid);
1406 			continue;
1407 		}
1408 		GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
1409 	}
1410 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_RXCQ)
1411 		cidx = rxq->ifr_cq_cidx;
1412 	else
1413 		cidx = rxq->ifr_fl[0].ifl_cidx;
1414 	if (iflib_rxd_avail(ctx, rxq, cidx, 1))
1415 		GROUPTASK_ENQUEUE(gtask);
1416 	else
1417 		IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
1418 	return (FILTER_HANDLED);
1419 }
1420 
1421 
1422 static int
1423 iflib_fast_intr_ctx(void *arg)
1424 {
1425 	iflib_filter_info_t info = arg;
1426 	struct grouptask *gtask = info->ifi_task;
1427 
1428 	if (!iflib_started)
1429 		return (FILTER_HANDLED);
1430 
1431 	DBG_COUNTER_INC(fast_intrs);
1432 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1433 		return (FILTER_HANDLED);
1434 
1435 	GROUPTASK_ENQUEUE(gtask);
1436 	return (FILTER_HANDLED);
1437 }
1438 
1439 static int
1440 _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
1441 	driver_filter_t filter, driver_intr_t handler, void *arg,
1442 				 char *name)
1443 {
1444 	int rc;
1445 	struct resource *res;
1446 	void *tag;
1447 	device_t dev = ctx->ifc_dev;
1448 
1449 	MPASS(rid < 512);
1450 	irq->ii_rid = rid;
1451 	res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &irq->ii_rid,
1452 				     RF_SHAREABLE | RF_ACTIVE);
1453 	if (res == NULL) {
1454 		device_printf(dev,
1455 		    "failed to allocate IRQ for rid %d, name %s.\n", rid, name);
1456 		return (ENOMEM);
1457 	}
1458 	irq->ii_res = res;
1459 	KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL"));
1460 	rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET,
1461 						filter, handler, arg, &tag);
1462 	if (rc != 0) {
1463 		device_printf(dev,
1464 		    "failed to setup interrupt for rid %d, name %s: %d\n",
1465 					  rid, name ? name : "unknown", rc);
1466 		return (rc);
1467 	} else if (name)
1468 		bus_describe_intr(dev, res, tag, "%s", name);
1469 
1470 	irq->ii_tag = tag;
1471 	return (0);
1472 }
1473 
1474 
1475 /*********************************************************************
1476  *
1477  *  Allocate memory for tx_buffer structures. The tx_buffer stores all
1478  *  the information needed to transmit a packet on the wire. This is
1479  *  called only once at attach, setup is done every reset.
1480  *
1481  **********************************************************************/
1482 
1483 static int
1484 iflib_txsd_alloc(iflib_txq_t txq)
1485 {
1486 	if_ctx_t ctx = txq->ift_ctx;
1487 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1488 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1489 	device_t dev = ctx->ifc_dev;
1490 	int err, nsegments, ntsosegments;
1491 
1492 	nsegments = scctx->isc_tx_nsegments;
1493 	ntsosegments = scctx->isc_tx_tso_segments_max;
1494 	MPASS(scctx->isc_ntxd[0] > 0);
1495 	MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0);
1496 	MPASS(nsegments > 0);
1497 	MPASS(ntsosegments > 0);
1498 	/*
1499 	 * Setup DMA descriptor areas.
1500 	 */
1501 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1502 			       1, 0,			/* alignment, bounds */
1503 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1504 			       BUS_SPACE_MAXADDR,	/* highaddr */
1505 			       NULL, NULL,		/* filter, filterarg */
1506 			       sctx->isc_tx_maxsize,		/* maxsize */
1507 			       nsegments,	/* nsegments */
1508 			       sctx->isc_tx_maxsegsize,	/* maxsegsize */
1509 			       0,			/* flags */
1510 			       NULL,			/* lockfunc */
1511 			       NULL,			/* lockfuncarg */
1512 			       &txq->ift_desc_tag))) {
1513 		device_printf(dev,"Unable to allocate TX DMA tag: %d\n", err);
1514 		device_printf(dev,"maxsize: %zd nsegments: %d maxsegsize: %zd\n",
1515 					  sctx->isc_tx_maxsize, nsegments, sctx->isc_tx_maxsegsize);
1516 		goto fail;
1517 	}
1518 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1519 			       1, 0,			/* alignment, bounds */
1520 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1521 			       BUS_SPACE_MAXADDR,	/* highaddr */
1522 			       NULL, NULL,		/* filter, filterarg */
1523 			       scctx->isc_tx_tso_size_max,		/* maxsize */
1524 			       ntsosegments,	/* nsegments */
1525 			       scctx->isc_tx_tso_segsize_max,	/* maxsegsize */
1526 			       0,			/* flags */
1527 			       NULL,			/* lockfunc */
1528 			       NULL,			/* lockfuncarg */
1529 			       &txq->ift_tso_desc_tag))) {
1530 		device_printf(dev,"Unable to allocate TX TSO DMA tag: %d\n", err);
1531 
1532 		goto fail;
1533 	}
1534 	if (!(txq->ift_sds.ifsd_flags =
1535 	    (uint8_t *) malloc(sizeof(uint8_t) *
1536 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1537 		device_printf(dev, "Unable to allocate tx_buffer memory\n");
1538 		err = ENOMEM;
1539 		goto fail;
1540 	}
1541 	if (!(txq->ift_sds.ifsd_m =
1542 	    (struct mbuf **) malloc(sizeof(struct mbuf *) *
1543 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1544 		device_printf(dev, "Unable to allocate tx_buffer memory\n");
1545 		err = ENOMEM;
1546 		goto fail;
1547 	}
1548 
1549         /* Create the descriptor buffer dma maps */
1550 #if defined(ACPI_DMAR) || (! (defined(__i386__) || defined(__amd64__)))
1551 	if ((ctx->ifc_flags & IFC_DMAR) == 0)
1552 		return (0);
1553 
1554 	if (!(txq->ift_sds.ifsd_map =
1555 	    (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1556 		device_printf(dev, "Unable to allocate tx_buffer map memory\n");
1557 		err = ENOMEM;
1558 		goto fail;
1559 	}
1560 
1561 	for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) {
1562 		err = bus_dmamap_create(txq->ift_desc_tag, 0, &txq->ift_sds.ifsd_map[i]);
1563 		if (err != 0) {
1564 			device_printf(dev, "Unable to create TX DMA map\n");
1565 			goto fail;
1566 		}
1567 	}
1568 #endif
1569 	return (0);
1570 fail:
1571 	/* We free all, it handles case where we are in the middle */
1572 	iflib_tx_structures_free(ctx);
1573 	return (err);
1574 }
1575 
1576 static void
1577 iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i)
1578 {
1579 	bus_dmamap_t map;
1580 
1581 	map = NULL;
1582 	if (txq->ift_sds.ifsd_map != NULL)
1583 		map = txq->ift_sds.ifsd_map[i];
1584 	if (map != NULL) {
1585 		bus_dmamap_unload(txq->ift_desc_tag, map);
1586 		bus_dmamap_destroy(txq->ift_desc_tag, map);
1587 		txq->ift_sds.ifsd_map[i] = NULL;
1588 	}
1589 }
1590 
1591 static void
1592 iflib_txq_destroy(iflib_txq_t txq)
1593 {
1594 	if_ctx_t ctx = txq->ift_ctx;
1595 
1596 	for (int i = 0; i < txq->ift_size; i++)
1597 		iflib_txsd_destroy(ctx, txq, i);
1598 	if (txq->ift_sds.ifsd_map != NULL) {
1599 		free(txq->ift_sds.ifsd_map, M_IFLIB);
1600 		txq->ift_sds.ifsd_map = NULL;
1601 	}
1602 	if (txq->ift_sds.ifsd_m != NULL) {
1603 		free(txq->ift_sds.ifsd_m, M_IFLIB);
1604 		txq->ift_sds.ifsd_m = NULL;
1605 	}
1606 	if (txq->ift_sds.ifsd_flags != NULL) {
1607 		free(txq->ift_sds.ifsd_flags, M_IFLIB);
1608 		txq->ift_sds.ifsd_flags = NULL;
1609 	}
1610 	if (txq->ift_desc_tag != NULL) {
1611 		bus_dma_tag_destroy(txq->ift_desc_tag);
1612 		txq->ift_desc_tag = NULL;
1613 	}
1614 	if (txq->ift_tso_desc_tag != NULL) {
1615 		bus_dma_tag_destroy(txq->ift_tso_desc_tag);
1616 		txq->ift_tso_desc_tag = NULL;
1617 	}
1618 }
1619 
1620 static void
1621 iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i)
1622 {
1623 	struct mbuf **mp;
1624 
1625 	mp = &txq->ift_sds.ifsd_m[i];
1626 	if (*mp == NULL)
1627 		return;
1628 
1629 	if (txq->ift_sds.ifsd_map != NULL) {
1630 		bus_dmamap_sync(txq->ift_desc_tag,
1631 				txq->ift_sds.ifsd_map[i],
1632 				BUS_DMASYNC_POSTWRITE);
1633 		bus_dmamap_unload(txq->ift_desc_tag,
1634 				  txq->ift_sds.ifsd_map[i]);
1635 	}
1636 	m_free(*mp);
1637 	DBG_COUNTER_INC(tx_frees);
1638 	*mp = NULL;
1639 }
1640 
1641 static int
1642 iflib_txq_setup(iflib_txq_t txq)
1643 {
1644 	if_ctx_t ctx = txq->ift_ctx;
1645 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1646 	iflib_dma_info_t di;
1647 	int i;
1648 
1649 	/* Set number of descriptors available */
1650 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
1651 	/* XXX make configurable */
1652 	txq->ift_update_freq = IFLIB_DEFAULT_TX_UPDATE_FREQ;
1653 
1654 	/* Reset indices */
1655 	txq->ift_cidx_processed = 0;
1656 	txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0;
1657 	txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset];
1658 
1659 	for (i = 0, di = txq->ift_ifdi; i < ctx->ifc_nhwtxqs; i++, di++)
1660 		bzero((void *)di->idi_vaddr, di->idi_size);
1661 
1662 	IFDI_TXQ_SETUP(ctx, txq->ift_id);
1663 	for (i = 0, di = txq->ift_ifdi; i < ctx->ifc_nhwtxqs; i++, di++)
1664 		bus_dmamap_sync(di->idi_tag, di->idi_map,
1665 						BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1666 	return (0);
1667 }
1668 
1669 /*********************************************************************
1670  *
1671  *  Allocate memory for rx_buffer structures. Since we use one
1672  *  rx_buffer per received packet, the maximum number of rx_buffer's
1673  *  that we'll need is equal to the number of receive descriptors
1674  *  that we've allocated.
1675  *
1676  **********************************************************************/
1677 static int
1678 iflib_rxsd_alloc(iflib_rxq_t rxq)
1679 {
1680 	if_ctx_t ctx = rxq->ifr_ctx;
1681 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1682 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1683 	device_t dev = ctx->ifc_dev;
1684 	iflib_fl_t fl;
1685 	int			err;
1686 
1687 	MPASS(scctx->isc_nrxd[0] > 0);
1688 	MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0);
1689 
1690 	fl = rxq->ifr_fl;
1691 	for (int i = 0; i <  rxq->ifr_nfl; i++, fl++) {
1692 		fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */
1693 		err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1694 					 1, 0,			/* alignment, bounds */
1695 					 BUS_SPACE_MAXADDR,	/* lowaddr */
1696 					 BUS_SPACE_MAXADDR,	/* highaddr */
1697 					 NULL, NULL,		/* filter, filterarg */
1698 					 sctx->isc_rx_maxsize,	/* maxsize */
1699 					 sctx->isc_rx_nsegments,	/* nsegments */
1700 					 sctx->isc_rx_maxsegsize,	/* maxsegsize */
1701 					 0,			/* flags */
1702 					 NULL,			/* lockfunc */
1703 					 NULL,			/* lockarg */
1704 					 &fl->ifl_desc_tag);
1705 		if (err) {
1706 			device_printf(dev, "%s: bus_dma_tag_create failed %d\n",
1707 				__func__, err);
1708 			goto fail;
1709 		}
1710 		if (!(fl->ifl_sds.ifsd_flags =
1711 		      (uint8_t *) malloc(sizeof(uint8_t) *
1712 					 scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1713 			device_printf(dev, "Unable to allocate tx_buffer memory\n");
1714 			err = ENOMEM;
1715 			goto fail;
1716 		}
1717 		if (!(fl->ifl_sds.ifsd_m =
1718 		      (struct mbuf **) malloc(sizeof(struct mbuf *) *
1719 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1720 			device_printf(dev, "Unable to allocate tx_buffer memory\n");
1721 			err = ENOMEM;
1722 			goto fail;
1723 		}
1724 		if (!(fl->ifl_sds.ifsd_cl =
1725 		      (caddr_t *) malloc(sizeof(caddr_t) *
1726 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1727 			device_printf(dev, "Unable to allocate tx_buffer memory\n");
1728 			err = ENOMEM;
1729 			goto fail;
1730 		}
1731 
1732 		/* Create the descriptor buffer dma maps */
1733 #if defined(ACPI_DMAR) || (! (defined(__i386__) || defined(__amd64__)))
1734 		if ((ctx->ifc_flags & IFC_DMAR) == 0)
1735 			continue;
1736 
1737 		if (!(fl->ifl_sds.ifsd_map =
1738 		      (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1739 			device_printf(dev, "Unable to allocate tx_buffer map memory\n");
1740 			err = ENOMEM;
1741 			goto fail;
1742 		}
1743 
1744 		for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++) {
1745 			err = bus_dmamap_create(fl->ifl_desc_tag, 0, &fl->ifl_sds.ifsd_map[i]);
1746 			if (err != 0) {
1747 				device_printf(dev, "Unable to create RX buffer DMA map\n");
1748 				goto fail;
1749 			}
1750 		}
1751 #endif
1752 	}
1753 	return (0);
1754 
1755 fail:
1756 	iflib_rx_structures_free(ctx);
1757 	return (err);
1758 }
1759 
1760 
1761 /*
1762  * Internal service routines
1763  */
1764 
1765 struct rxq_refill_cb_arg {
1766 	int               error;
1767 	bus_dma_segment_t seg;
1768 	int               nseg;
1769 };
1770 
1771 static void
1772 _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
1773 {
1774 	struct rxq_refill_cb_arg *cb_arg = arg;
1775 
1776 	cb_arg->error = error;
1777 	cb_arg->seg = segs[0];
1778 	cb_arg->nseg = nseg;
1779 }
1780 
1781 
1782 #ifdef ACPI_DMAR
1783 #define IS_DMAR(ctx) (ctx->ifc_flags & IFC_DMAR)
1784 #else
1785 #define IS_DMAR(ctx) (0)
1786 #endif
1787 
1788 /**
1789  *	rxq_refill - refill an rxq  free-buffer list
1790  *	@ctx: the iflib context
1791  *	@rxq: the free-list to refill
1792  *	@n: the number of new buffers to allocate
1793  *
1794  *	(Re)populate an rxq free-buffer list with up to @n new packet buffers.
1795  *	The caller must assure that @n does not exceed the queue's capacity.
1796  */
1797 static void
1798 _iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
1799 {
1800 	struct mbuf *m;
1801 	int idx, pidx = fl->ifl_pidx;
1802 	caddr_t cl, *sd_cl;
1803 	struct mbuf **sd_m;
1804 	uint8_t *sd_flags;
1805 	struct if_rxd_update iru;
1806 	bus_dmamap_t *sd_map;
1807 	int n, i = 0;
1808 	uint64_t bus_addr;
1809 	int err;
1810 
1811 	sd_m = fl->ifl_sds.ifsd_m;
1812 	sd_map = fl->ifl_sds.ifsd_map;
1813 	sd_cl = fl->ifl_sds.ifsd_cl;
1814 	sd_flags = fl->ifl_sds.ifsd_flags;
1815 	idx = pidx;
1816 
1817 	n  = count;
1818 	MPASS(n > 0);
1819 	MPASS(fl->ifl_credits + n <= fl->ifl_size);
1820 
1821 	if (pidx < fl->ifl_cidx)
1822 		MPASS(pidx + n <= fl->ifl_cidx);
1823 	if (pidx == fl->ifl_cidx && (fl->ifl_credits < fl->ifl_size))
1824 		MPASS(fl->ifl_gen == 0);
1825 	if (pidx > fl->ifl_cidx)
1826 		MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx);
1827 
1828 	DBG_COUNTER_INC(fl_refills);
1829 	if (n > 8)
1830 		DBG_COUNTER_INC(fl_refills_large);
1831 	iru.iru_paddrs = fl->ifl_bus_addrs;
1832 	iru.iru_vaddrs = &fl->ifl_vm_addrs[0];
1833 	iru.iru_idxs = fl->ifl_rxd_idxs;
1834 	iru.iru_qsidx = fl->ifl_rxq->ifr_id;
1835 	iru.iru_buf_size = fl->ifl_buf_size;
1836 	iru.iru_flidx = fl->ifl_id;
1837 	while (n--) {
1838 		/*
1839 		 * We allocate an uninitialized mbuf + cluster, mbuf is
1840 		 * initialized after rx.
1841 		 *
1842 		 * If the cluster is still set then we know a minimum sized packet was received
1843 		 */
1844 		if ((cl = sd_cl[idx]) == NULL) {
1845 			if ((cl = sd_cl[idx] = m_cljget(NULL, M_NOWAIT, fl->ifl_buf_size)) == NULL)
1846 				break;
1847 #if MEMORY_LOGGING
1848 			fl->ifl_cl_enqueued++;
1849 #endif
1850 		}
1851 		if ((m = m_gethdr(M_NOWAIT, MT_NOINIT)) == NULL) {
1852 			break;
1853 		}
1854 #if MEMORY_LOGGING
1855 		fl->ifl_m_enqueued++;
1856 #endif
1857 
1858 		DBG_COUNTER_INC(rx_allocs);
1859 #if defined(__i386__) || defined(__amd64__)
1860 		if (!IS_DMAR(ctx)) {
1861 			bus_addr = pmap_kextract((vm_offset_t)cl);
1862 		} else
1863 #endif
1864 		{
1865 			struct rxq_refill_cb_arg cb_arg;
1866 			iflib_rxq_t q;
1867 
1868 			cb_arg.error = 0;
1869 			q = fl->ifl_rxq;
1870 			MPASS(sd_map != NULL);
1871 			MPASS(sd_map[idx] != NULL);
1872 			err = bus_dmamap_load(fl->ifl_desc_tag, sd_map[idx],
1873 		         cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg, 0);
1874 			bus_dmamap_sync(fl->ifl_desc_tag, sd_map[idx], BUS_DMASYNC_PREREAD);
1875 
1876 			if (err != 0 || cb_arg.error) {
1877 				/*
1878 				 * !zone_pack ?
1879 				 */
1880 				if (fl->ifl_zone == zone_pack)
1881 					uma_zfree(fl->ifl_zone, cl);
1882 				m_free(m);
1883 				n = 0;
1884 				goto done;
1885 			}
1886 			bus_addr = cb_arg.seg.ds_addr;
1887 		}
1888 		sd_flags[idx] |= RX_SW_DESC_INUSE;
1889 
1890 		MPASS(sd_m[idx] == NULL);
1891 		sd_cl[idx] = cl;
1892 		sd_m[idx] = m;
1893 		fl->ifl_rxd_idxs[i] = idx;
1894 		fl->ifl_bus_addrs[i] = bus_addr;
1895 		fl->ifl_vm_addrs[i] = cl;
1896 		fl->ifl_credits++;
1897 		i++;
1898 		MPASS(fl->ifl_credits <= fl->ifl_size);
1899 		if (++idx == fl->ifl_size) {
1900 			fl->ifl_gen = 1;
1901 			idx = 0;
1902 		}
1903 		if (n == 0 || i == IFLIB_MAX_RX_REFRESH) {
1904 			iru.iru_pidx = pidx;
1905 			iru.iru_count = i;
1906 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
1907 			i = 0;
1908 			pidx = idx;
1909 		}
1910 		fl->ifl_pidx = idx;
1911 
1912 	}
1913 done:
1914 	DBG_COUNTER_INC(rxd_flush);
1915 	if (fl->ifl_pidx == 0)
1916 		pidx = fl->ifl_size - 1;
1917 	else
1918 		pidx = fl->ifl_pidx - 1;
1919 
1920 	if (sd_map)
1921 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
1922 				BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1923 	ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id, fl->ifl_id, pidx);
1924 }
1925 
1926 static __inline void
1927 __iflib_fl_refill_lt(if_ctx_t ctx, iflib_fl_t fl, int max)
1928 {
1929 	/* we avoid allowing pidx to catch up with cidx as it confuses ixl */
1930 	int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1;
1931 #ifdef INVARIANTS
1932 	int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1;
1933 #endif
1934 
1935 	MPASS(fl->ifl_credits <= fl->ifl_size);
1936 	MPASS(reclaimable == delta);
1937 
1938 	if (reclaimable > 0)
1939 		_iflib_fl_refill(ctx, fl, min(max, reclaimable));
1940 }
1941 
1942 static void
1943 iflib_fl_bufs_free(iflib_fl_t fl)
1944 {
1945 	iflib_dma_info_t idi = fl->ifl_ifdi;
1946 	uint32_t i;
1947 
1948 	for (i = 0; i < fl->ifl_size; i++) {
1949 		struct mbuf **sd_m = &fl->ifl_sds.ifsd_m[i];
1950 		uint8_t *sd_flags = &fl->ifl_sds.ifsd_flags[i];
1951 		caddr_t *sd_cl = &fl->ifl_sds.ifsd_cl[i];
1952 
1953 		if (*sd_flags & RX_SW_DESC_INUSE) {
1954 			if (fl->ifl_sds.ifsd_map != NULL) {
1955 				bus_dmamap_t sd_map = fl->ifl_sds.ifsd_map[i];
1956 				bus_dmamap_unload(fl->ifl_desc_tag, sd_map);
1957 				bus_dmamap_destroy(fl->ifl_desc_tag, sd_map);
1958 			}
1959 			if (*sd_m != NULL) {
1960 				m_init(*sd_m, M_NOWAIT, MT_DATA, 0);
1961 				uma_zfree(zone_mbuf, *sd_m);
1962 			}
1963 			if (*sd_cl != NULL)
1964 				uma_zfree(fl->ifl_zone, *sd_cl);
1965 			*sd_flags = 0;
1966 		} else {
1967 			MPASS(*sd_cl == NULL);
1968 			MPASS(*sd_m == NULL);
1969 		}
1970 #if MEMORY_LOGGING
1971 		fl->ifl_m_dequeued++;
1972 		fl->ifl_cl_dequeued++;
1973 #endif
1974 		*sd_cl = NULL;
1975 		*sd_m = NULL;
1976 	}
1977 #ifdef INVARIANTS
1978 	for (i = 0; i < fl->ifl_size; i++) {
1979 		MPASS(fl->ifl_sds.ifsd_flags[i] == 0);
1980 		MPASS(fl->ifl_sds.ifsd_cl[i] == NULL);
1981 		MPASS(fl->ifl_sds.ifsd_m[i] == NULL);
1982 	}
1983 #endif
1984 	/*
1985 	 * Reset free list values
1986 	 */
1987 	fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = 0;;
1988 	bzero(idi->idi_vaddr, idi->idi_size);
1989 }
1990 
1991 /*********************************************************************
1992  *
1993  *  Initialize a receive ring and its buffers.
1994  *
1995  **********************************************************************/
1996 static int
1997 iflib_fl_setup(iflib_fl_t fl)
1998 {
1999 	iflib_rxq_t rxq = fl->ifl_rxq;
2000 	if_ctx_t ctx = rxq->ifr_ctx;
2001 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2002 
2003 	/*
2004 	** Free current RX buffer structs and their mbufs
2005 	*/
2006 	iflib_fl_bufs_free(fl);
2007 	/* Now replenish the mbufs */
2008 	MPASS(fl->ifl_credits == 0);
2009 	/*
2010 	 * XXX don't set the max_frame_size to larger
2011 	 * than the hardware can handle
2012 	 */
2013 	if (sctx->isc_max_frame_size <= 2048)
2014 		fl->ifl_buf_size = MCLBYTES;
2015 #ifndef CONTIGMALLOC_WORKS
2016 	else
2017 		fl->ifl_buf_size = MJUMPAGESIZE;
2018 #else
2019 	else if (sctx->isc_max_frame_size <= 4096)
2020 		fl->ifl_buf_size = MJUMPAGESIZE;
2021 	else if (sctx->isc_max_frame_size <= 9216)
2022 		fl->ifl_buf_size = MJUM9BYTES;
2023 	else
2024 		fl->ifl_buf_size = MJUM16BYTES;
2025 #endif
2026 	if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size)
2027 		ctx->ifc_max_fl_buf_size = fl->ifl_buf_size;
2028 	fl->ifl_cltype = m_gettype(fl->ifl_buf_size);
2029 	fl->ifl_zone = m_getzone(fl->ifl_buf_size);
2030 
2031 
2032 	/* avoid pre-allocating zillions of clusters to an idle card
2033 	 * potentially speeding up attach
2034 	 */
2035 	_iflib_fl_refill(ctx, fl, min(128, fl->ifl_size));
2036 	MPASS(min(128, fl->ifl_size) == fl->ifl_credits);
2037 	if (min(128, fl->ifl_size) != fl->ifl_credits)
2038 		return (ENOBUFS);
2039 	/*
2040 	 * handle failure
2041 	 */
2042 	MPASS(rxq != NULL);
2043 	MPASS(fl->ifl_ifdi != NULL);
2044 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2045 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2046 	return (0);
2047 }
2048 
2049 /*********************************************************************
2050  *
2051  *  Free receive ring data structures
2052  *
2053  **********************************************************************/
2054 static void
2055 iflib_rx_sds_free(iflib_rxq_t rxq)
2056 {
2057 	iflib_fl_t fl;
2058 	int i;
2059 
2060 	if (rxq->ifr_fl != NULL) {
2061 		for (i = 0; i < rxq->ifr_nfl; i++) {
2062 			fl = &rxq->ifr_fl[i];
2063 			if (fl->ifl_desc_tag != NULL) {
2064 				bus_dma_tag_destroy(fl->ifl_desc_tag);
2065 				fl->ifl_desc_tag = NULL;
2066 			}
2067 			free(fl->ifl_sds.ifsd_m, M_IFLIB);
2068 			free(fl->ifl_sds.ifsd_cl, M_IFLIB);
2069 			/* XXX destroy maps first */
2070 			free(fl->ifl_sds.ifsd_map, M_IFLIB);
2071 			fl->ifl_sds.ifsd_m = NULL;
2072 			fl->ifl_sds.ifsd_cl = NULL;
2073 			fl->ifl_sds.ifsd_map = NULL;
2074 		}
2075 		free(rxq->ifr_fl, M_IFLIB);
2076 		rxq->ifr_fl = NULL;
2077 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
2078 	}
2079 }
2080 
2081 /*
2082  * MI independent logic
2083  *
2084  */
2085 static void
2086 iflib_timer(void *arg)
2087 {
2088 	iflib_txq_t txq = arg;
2089 	if_ctx_t ctx = txq->ift_ctx;
2090 
2091 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
2092 		return;
2093 	/*
2094 	** Check on the state of the TX queue(s), this
2095 	** can be done without the lock because its RO
2096 	** and the HUNG state will be static if set.
2097 	*/
2098 	IFDI_TIMER(ctx, txq->ift_id);
2099 	if ((txq->ift_qstatus == IFLIB_QUEUE_HUNG) &&
2100 	    ((txq->ift_cleaned_prev == txq->ift_cleaned) ||
2101 	     (ctx->ifc_pause_frames == 0)))
2102 		goto hung;
2103 
2104 	if (ifmp_ring_is_stalled(txq->ift_br))
2105 		txq->ift_qstatus = IFLIB_QUEUE_HUNG;
2106 	txq->ift_cleaned_prev = txq->ift_cleaned;
2107 	/* handle any laggards */
2108 	if (txq->ift_db_pending)
2109 		GROUPTASK_ENQUEUE(&txq->ift_task);
2110 
2111 	ctx->ifc_pause_frames = 0;
2112 	if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)
2113 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq, txq->ift_timer.c_cpu);
2114 	return;
2115 hung:
2116 	CTX_LOCK(ctx);
2117 	if_setdrvflagbits(ctx->ifc_ifp, 0, IFF_DRV_RUNNING);
2118 	device_printf(ctx->ifc_dev,  "TX(%d) desc avail = %d, pidx = %d\n",
2119 				  txq->ift_id, TXQ_AVAIL(txq), txq->ift_pidx);
2120 
2121 	IFDI_WATCHDOG_RESET(ctx);
2122 	ctx->ifc_watchdog_events++;
2123 	ctx->ifc_pause_frames = 0;
2124 
2125 	iflib_init_locked(ctx);
2126 	CTX_UNLOCK(ctx);
2127 }
2128 
2129 static void
2130 iflib_init_locked(if_ctx_t ctx)
2131 {
2132 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2133 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2134 	if_t ifp = ctx->ifc_ifp;
2135 	iflib_fl_t fl;
2136 	iflib_txq_t txq;
2137 	iflib_rxq_t rxq;
2138 	int i, j, tx_ip_csum_flags, tx_ip6_csum_flags;
2139 
2140 
2141 	if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2142 	IFDI_INTR_DISABLE(ctx);
2143 
2144 	tx_ip_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP);
2145 	tx_ip6_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | CSUM_IP6_SCTP);
2146 	/* Set hardware offload abilities */
2147 	if_clearhwassist(ifp);
2148 	if (if_getcapenable(ifp) & IFCAP_TXCSUM)
2149 		if_sethwassistbits(ifp, tx_ip_csum_flags, 0);
2150 	if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6)
2151 		if_sethwassistbits(ifp,  tx_ip6_csum_flags, 0);
2152 	if (if_getcapenable(ifp) & IFCAP_TSO4)
2153 		if_sethwassistbits(ifp, CSUM_IP_TSO, 0);
2154 	if (if_getcapenable(ifp) & IFCAP_TSO6)
2155 		if_sethwassistbits(ifp, CSUM_IP6_TSO, 0);
2156 
2157 	for (i = 0, txq = ctx->ifc_txqs; i < sctx->isc_ntxqsets; i++, txq++) {
2158 		CALLOUT_LOCK(txq);
2159 		callout_stop(&txq->ift_timer);
2160 		CALLOUT_UNLOCK(txq);
2161 		iflib_netmap_txq_init(ctx, txq);
2162 	}
2163 	for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
2164 		MPASS(rxq->ifr_id == i);
2165 		iflib_netmap_rxq_init(ctx, rxq);
2166 	}
2167 #ifdef INVARIANTS
2168 	i = if_getdrvflags(ifp);
2169 #endif
2170 	IFDI_INIT(ctx);
2171 	MPASS(if_getdrvflags(ifp) == i);
2172 	for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
2173 		/* XXX this should really be done on a per-queue basis */
2174 		if (if_getcapenable(ifp) & IFCAP_NETMAP)
2175 			continue;
2176 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
2177 			if (iflib_fl_setup(fl)) {
2178 				device_printf(ctx->ifc_dev, "freelist setup failed - check cluster settings\n");
2179 				goto done;
2180 			}
2181 		}
2182 	}
2183 	done:
2184 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
2185 	IFDI_INTR_ENABLE(ctx);
2186 	txq = ctx->ifc_txqs;
2187 	for (i = 0; i < sctx->isc_ntxqsets; i++, txq++)
2188 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq,
2189 			txq->ift_timer.c_cpu);
2190 }
2191 
2192 static int
2193 iflib_media_change(if_t ifp)
2194 {
2195 	if_ctx_t ctx = if_getsoftc(ifp);
2196 	int err;
2197 
2198 	CTX_LOCK(ctx);
2199 	if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0)
2200 		iflib_init_locked(ctx);
2201 	CTX_UNLOCK(ctx);
2202 	return (err);
2203 }
2204 
2205 static void
2206 iflib_media_status(if_t ifp, struct ifmediareq *ifmr)
2207 {
2208 	if_ctx_t ctx = if_getsoftc(ifp);
2209 
2210 	CTX_LOCK(ctx);
2211 	IFDI_UPDATE_ADMIN_STATUS(ctx);
2212 	IFDI_MEDIA_STATUS(ctx, ifmr);
2213 	CTX_UNLOCK(ctx);
2214 }
2215 
2216 static void
2217 iflib_stop(if_ctx_t ctx)
2218 {
2219 	iflib_txq_t txq = ctx->ifc_txqs;
2220 	iflib_rxq_t rxq = ctx->ifc_rxqs;
2221 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2222 	iflib_dma_info_t di;
2223 	iflib_fl_t fl;
2224 	int i, j;
2225 
2226 	/* Tell the stack that the interface is no longer active */
2227 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2228 
2229 	IFDI_INTR_DISABLE(ctx);
2230 	DELAY(1000);
2231 	IFDI_STOP(ctx);
2232 	DELAY(1000);
2233 
2234 	iflib_debug_reset();
2235 	/* Wait for current tx queue users to exit to disarm watchdog timer. */
2236 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) {
2237 		/* make sure all transmitters have completed before proceeding XXX */
2238 
2239 		/* clean any enqueued buffers */
2240 		iflib_ifmp_purge(txq);
2241 		/* Free any existing tx buffers. */
2242 		for (j = 0; j < txq->ift_size; j++) {
2243 			iflib_txsd_free(ctx, txq, j);
2244 		}
2245 		txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0;
2246 		txq->ift_in_use = txq->ift_gen = txq->ift_cidx = txq->ift_pidx = txq->ift_no_desc_avail = 0;
2247 		txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0;
2248 		txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0;
2249 		txq->ift_pullups = 0;
2250 		ifmp_ring_reset_stats(txq->ift_br);
2251 		for (j = 0, di = txq->ift_ifdi; j < ctx->ifc_nhwtxqs; j++, di++)
2252 			bzero((void *)di->idi_vaddr, di->idi_size);
2253 	}
2254 	for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) {
2255 		/* make sure all transmitters have completed before proceeding XXX */
2256 
2257 		for (j = 0, di = txq->ift_ifdi; j < ctx->ifc_nhwrxqs; j++, di++)
2258 			bzero((void *)di->idi_vaddr, di->idi_size);
2259 		/* also resets the free lists pidx/cidx */
2260 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
2261 			iflib_fl_bufs_free(fl);
2262 	}
2263 }
2264 
2265 static inline caddr_t
2266 calc_next_rxd(iflib_fl_t fl, int cidx)
2267 {
2268 	qidx_t size;
2269 	int nrxd;
2270 	caddr_t start, end, cur, next;
2271 
2272 	nrxd = fl->ifl_size;
2273 	size = fl->ifl_rxd_size;
2274 	start = fl->ifl_ifdi->idi_vaddr;
2275 
2276 	if (__predict_false(size == 0))
2277 		return (start);
2278 	cur = start + size*cidx;
2279 	end = start + size*nrxd;
2280 	next = CACHE_PTR_NEXT(cur);
2281 	return (next < end ? next : start);
2282 }
2283 
2284 static inline void
2285 prefetch_pkts(iflib_fl_t fl, int cidx)
2286 {
2287 	int nextptr;
2288 	int nrxd = fl->ifl_size;
2289 	caddr_t next_rxd;
2290 
2291 
2292 	nextptr = (cidx + CACHE_PTR_INCREMENT) & (nrxd-1);
2293 	prefetch(&fl->ifl_sds.ifsd_m[nextptr]);
2294 	prefetch(&fl->ifl_sds.ifsd_cl[nextptr]);
2295 	next_rxd = calc_next_rxd(fl, cidx);
2296 	prefetch(next_rxd);
2297 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 1) & (nrxd-1)]);
2298 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 2) & (nrxd-1)]);
2299 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 3) & (nrxd-1)]);
2300 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 4) & (nrxd-1)]);
2301 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 1) & (nrxd-1)]);
2302 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 2) & (nrxd-1)]);
2303 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 3) & (nrxd-1)]);
2304 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 4) & (nrxd-1)]);
2305 }
2306 
2307 static void
2308 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, int unload, if_rxsd_t sd)
2309 {
2310 	int flid, cidx;
2311 	bus_dmamap_t map;
2312 	iflib_fl_t fl;
2313 	iflib_dma_info_t di;
2314 	int next;
2315 
2316 	map = NULL;
2317 	flid = irf->irf_flid;
2318 	cidx = irf->irf_idx;
2319 	fl = &rxq->ifr_fl[flid];
2320 	sd->ifsd_fl = fl;
2321 	sd->ifsd_cidx = cidx;
2322 	sd->ifsd_m = &fl->ifl_sds.ifsd_m[cidx];
2323 	sd->ifsd_cl = &fl->ifl_sds.ifsd_cl[cidx];
2324 	fl->ifl_credits--;
2325 #if MEMORY_LOGGING
2326 	fl->ifl_m_dequeued++;
2327 #endif
2328 	if (rxq->ifr_ctx->ifc_flags & IFC_PREFETCH)
2329 		prefetch_pkts(fl, cidx);
2330 	if (fl->ifl_sds.ifsd_map != NULL) {
2331 		next = (cidx + CACHE_PTR_INCREMENT) & (fl->ifl_size-1);
2332 		prefetch(&fl->ifl_sds.ifsd_map[next]);
2333 		map = fl->ifl_sds.ifsd_map[cidx];
2334 		di = fl->ifl_ifdi;
2335 		next = (cidx + CACHE_LINE_SIZE) & (fl->ifl_size-1);
2336 		prefetch(&fl->ifl_sds.ifsd_flags[next]);
2337 		bus_dmamap_sync(di->idi_tag, di->idi_map,
2338 				BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
2339 
2340 	/* not valid assert if bxe really does SGE from non-contiguous elements */
2341 		MPASS(fl->ifl_cidx == cidx);
2342 		if (unload)
2343 			bus_dmamap_unload(fl->ifl_desc_tag, map);
2344 	}
2345 	fl->ifl_cidx = (fl->ifl_cidx + 1) & (fl->ifl_size-1);
2346 	if (__predict_false(fl->ifl_cidx == 0))
2347 		fl->ifl_gen = 0;
2348 	if (map != NULL)
2349 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2350 			BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2351 }
2352 
2353 static struct mbuf *
2354 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri, if_rxsd_t sd)
2355 {
2356 	int i, padlen , flags;
2357 	struct mbuf *m, *mh, *mt;
2358 	caddr_t cl;
2359 
2360 	i = 0;
2361 	mh = NULL;
2362 	do {
2363 		rxd_frag_to_sd(rxq, &ri->iri_frags[i], TRUE, sd);
2364 
2365 		MPASS(*sd->ifsd_cl != NULL);
2366 		MPASS(*sd->ifsd_m != NULL);
2367 
2368 		/* Don't include zero-length frags */
2369 		if (ri->iri_frags[i].irf_len == 0) {
2370 			/* XXX we can save the cluster here, but not the mbuf */
2371 			m_init(*sd->ifsd_m, M_NOWAIT, MT_DATA, 0);
2372 			m_free(*sd->ifsd_m);
2373 			*sd->ifsd_m = NULL;
2374 			continue;
2375 		}
2376 		m = *sd->ifsd_m;
2377 		*sd->ifsd_m = NULL;
2378 		if (mh == NULL) {
2379 			flags = M_PKTHDR|M_EXT;
2380 			mh = mt = m;
2381 			padlen = ri->iri_pad;
2382 		} else {
2383 			flags = M_EXT;
2384 			mt->m_next = m;
2385 			mt = m;
2386 			/* assuming padding is only on the first fragment */
2387 			padlen = 0;
2388 		}
2389 		cl = *sd->ifsd_cl;
2390 		*sd->ifsd_cl = NULL;
2391 
2392 		/* Can these two be made one ? */
2393 		m_init(m, M_NOWAIT, MT_DATA, flags);
2394 		m_cljset(m, cl, sd->ifsd_fl->ifl_cltype);
2395 		/*
2396 		 * These must follow m_init and m_cljset
2397 		 */
2398 		m->m_data += padlen;
2399 		ri->iri_len -= padlen;
2400 		m->m_len = ri->iri_frags[i].irf_len;
2401 	} while (++i < ri->iri_nfrags);
2402 
2403 	return (mh);
2404 }
2405 
2406 /*
2407  * Process one software descriptor
2408  */
2409 static struct mbuf *
2410 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri)
2411 {
2412 	struct if_rxsd sd;
2413 	struct mbuf *m;
2414 
2415 	/* should I merge this back in now that the two paths are basically duplicated? */
2416 	if (ri->iri_nfrags == 1 &&
2417 	    ri->iri_frags[0].irf_len <= IFLIB_RX_COPY_THRESH) {
2418 		rxd_frag_to_sd(rxq, &ri->iri_frags[0], FALSE, &sd);
2419 		m = *sd.ifsd_m;
2420 		*sd.ifsd_m = NULL;
2421 		m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR);
2422 #ifndef __NO_STRICT_ALIGNMENT
2423 		if (!IP_ALIGNED(m))
2424 			m->m_data += 2;
2425 #endif
2426 		memcpy(m->m_data, *sd.ifsd_cl, ri->iri_len);
2427 		m->m_len = ri->iri_frags[0].irf_len;
2428        } else {
2429 		m = assemble_segments(rxq, ri, &sd);
2430 	}
2431 	m->m_pkthdr.len = ri->iri_len;
2432 	m->m_pkthdr.rcvif = ri->iri_ifp;
2433 	m->m_flags |= ri->iri_flags;
2434 	m->m_pkthdr.ether_vtag = ri->iri_vtag;
2435 	m->m_pkthdr.flowid = ri->iri_flowid;
2436 	M_HASHTYPE_SET(m, ri->iri_rsstype);
2437 	m->m_pkthdr.csum_flags = ri->iri_csum_flags;
2438 	m->m_pkthdr.csum_data = ri->iri_csum_data;
2439 	return (m);
2440 }
2441 
2442 static bool
2443 iflib_rxeof(iflib_rxq_t rxq, qidx_t budget)
2444 {
2445 	if_ctx_t ctx = rxq->ifr_ctx;
2446 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2447 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2448 	int avail, i;
2449 	qidx_t *cidxp;
2450 	struct if_rxd_info ri;
2451 	int err, budget_left, rx_bytes, rx_pkts;
2452 	iflib_fl_t fl;
2453 	struct ifnet *ifp;
2454 	int lro_enabled;
2455 
2456 	/*
2457 	 * XXX early demux data packets so that if_input processing only handles
2458 	 * acks in interrupt context
2459 	 */
2460 	struct mbuf *m, *mh, *mt;
2461 
2462 	ifp = ctx->ifc_ifp;
2463 #ifdef DEV_NETMAP
2464 	if (ifp->if_capenable & IFCAP_NETMAP) {
2465 		u_int work = 0;
2466 		if (netmap_rx_irq(ifp, rxq->ifr_id, &work))
2467 			return (FALSE);
2468 	}
2469 #endif
2470 
2471 	mh = mt = NULL;
2472 	MPASS(budget > 0);
2473 	rx_pkts	= rx_bytes = 0;
2474 	if (sctx->isc_flags & IFLIB_HAS_RXCQ)
2475 		cidxp = &rxq->ifr_cq_cidx;
2476 	else
2477 		cidxp = &rxq->ifr_fl[0].ifl_cidx;
2478 	if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) {
2479 		for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2480 			__iflib_fl_refill_lt(ctx, fl, budget + 8);
2481 		DBG_COUNTER_INC(rx_unavail);
2482 		return (false);
2483 	}
2484 
2485 	for (budget_left = budget; (budget_left > 0) && (avail > 0); budget_left--, avail--) {
2486 		if (__predict_false(!CTX_ACTIVE(ctx))) {
2487 			DBG_COUNTER_INC(rx_ctx_inactive);
2488 			break;
2489 		}
2490 		/*
2491 		 * Reset client set fields to their default values
2492 		 */
2493 		rxd_info_zero(&ri);
2494 		ri.iri_qsidx = rxq->ifr_id;
2495 		ri.iri_cidx = *cidxp;
2496 		ri.iri_ifp = ifp;
2497 		ri.iri_frags = rxq->ifr_frags;
2498 		err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
2499 
2500 		if (err)
2501 			goto err;
2502 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
2503 			*cidxp = ri.iri_cidx;
2504 			/* Update our consumer index */
2505 			/* XXX NB: shurd - check if this is still safe */
2506 			while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0]) {
2507 				rxq->ifr_cq_cidx -= scctx->isc_nrxd[0];
2508 				rxq->ifr_cq_gen = 0;
2509 			}
2510 			/* was this only a completion queue message? */
2511 			if (__predict_false(ri.iri_nfrags == 0))
2512 				continue;
2513 		}
2514 		MPASS(ri.iri_nfrags != 0);
2515 		MPASS(ri.iri_len != 0);
2516 
2517 		/* will advance the cidx on the corresponding free lists */
2518 		m = iflib_rxd_pkt_get(rxq, &ri);
2519 		if (avail == 0 && budget_left)
2520 			avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left);
2521 
2522 		if (__predict_false(m == NULL)) {
2523 			DBG_COUNTER_INC(rx_mbuf_null);
2524 			continue;
2525 		}
2526 		/* imm_pkt: -- cxgb */
2527 		if (mh == NULL)
2528 			mh = mt = m;
2529 		else {
2530 			mt->m_nextpkt = m;
2531 			mt = m;
2532 		}
2533 	}
2534 	/* make sure that we can refill faster than drain */
2535 	for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2536 		__iflib_fl_refill_lt(ctx, fl, budget + 8);
2537 
2538 	lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO);
2539 	while (mh != NULL) {
2540 		m = mh;
2541 		mh = mh->m_nextpkt;
2542 		m->m_nextpkt = NULL;
2543 #ifndef __NO_STRICT_ALIGNMENT
2544 		if (!IP_ALIGNED(m) && (m = iflib_fixup_rx(m)) == NULL)
2545 			continue;
2546 #endif
2547 		rx_bytes += m->m_pkthdr.len;
2548 		rx_pkts++;
2549 #if defined(INET6) || defined(INET)
2550 		if (lro_enabled && tcp_lro_rx(&rxq->ifr_lc, m, 0) == 0)
2551 			continue;
2552 #endif
2553 		DBG_COUNTER_INC(rx_if_input);
2554 		ifp->if_input(ifp, m);
2555 	}
2556 
2557 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
2558 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
2559 
2560 	/*
2561 	 * Flush any outstanding LRO work
2562 	 */
2563 #if defined(INET6) || defined(INET)
2564 	tcp_lro_flush_all(&rxq->ifr_lc);
2565 #endif
2566 	if (avail)
2567 		return true;
2568 	return (iflib_rxd_avail(ctx, rxq, *cidxp, 1));
2569 err:
2570 	CTX_LOCK(ctx);
2571 	ctx->ifc_flags |= IFC_DO_RESET;
2572 	iflib_admin_intr_deferred(ctx);
2573 	CTX_UNLOCK(ctx);
2574 	return (false);
2575 }
2576 
2577 #define TXD_NOTIFY_COUNT(txq) (((txq)->ift_size / (txq)->ift_update_freq)-1)
2578 static inline qidx_t
2579 txq_max_db_deferred(iflib_txq_t txq, qidx_t in_use)
2580 {
2581 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2582 	qidx_t minthresh = txq->ift_size / 8;
2583 	if (in_use > 4*minthresh)
2584 		return (notify_count);
2585 	if (in_use > 2*minthresh)
2586 		return (notify_count >> 1);
2587 	if (in_use > minthresh)
2588 		return (notify_count >> 3);
2589 	return (0);
2590 }
2591 
2592 static inline qidx_t
2593 txq_max_rs_deferred(iflib_txq_t txq)
2594 {
2595 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2596 	qidx_t minthresh = txq->ift_size / 8;
2597 	if (txq->ift_in_use > 4*minthresh)
2598 		return (notify_count);
2599 	if (txq->ift_in_use > 2*minthresh)
2600 		return (notify_count >> 1);
2601 	if (txq->ift_in_use > minthresh)
2602 		return (notify_count >> 2);
2603 	return (notify_count >> 4);
2604 }
2605 
2606 #define M_CSUM_FLAGS(m) ((m)->m_pkthdr.csum_flags)
2607 #define M_HAS_VLANTAG(m) (m->m_flags & M_VLANTAG)
2608 
2609 #define TXQ_MAX_DB_DEFERRED(txq, in_use) txq_max_db_deferred((txq), (in_use))
2610 #define TXQ_MAX_RS_DEFERRED(txq) txq_max_rs_deferred(txq)
2611 #define TXQ_MAX_DB_CONSUMED(size) (size >> 4)
2612 
2613 /* forward compatibility for cxgb */
2614 #define FIRST_QSET(ctx) 0
2615 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets)
2616 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets)
2617 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx))
2618 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments))
2619 
2620 /* XXX we should be setting this to something other than zero */
2621 #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh)
2622 #define MAX_TX_DESC(ctx) ((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max)
2623 
2624 static inline bool
2625 iflib_txd_db_check(if_ctx_t ctx, iflib_txq_t txq, int ring, qidx_t in_use)
2626 {
2627 	qidx_t dbval, max;
2628 	bool rang;
2629 
2630 	rang = false;
2631 	max = TXQ_MAX_DB_DEFERRED(txq, in_use);
2632 	if (ring || txq->ift_db_pending >= max) {
2633 		dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx;
2634 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval);
2635 		txq->ift_db_pending = txq->ift_npending = 0;
2636 		rang = true;
2637 	}
2638 	return (rang);
2639 }
2640 
2641 #ifdef PKT_DEBUG
2642 static void
2643 print_pkt(if_pkt_info_t pi)
2644 {
2645 	printf("pi len:  %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n",
2646 	       pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx);
2647 	printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n",
2648 	       pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag);
2649 	printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n",
2650 	       pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto);
2651 }
2652 #endif
2653 
2654 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO)
2655 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO)
2656 
2657 static int
2658 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp)
2659 {
2660 	if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx;
2661 	struct ether_vlan_header *eh;
2662 	struct mbuf *m, *n;
2663 
2664 	n = m = *mp;
2665 	if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) &&
2666 	    M_WRITABLE(m) == 0) {
2667 		if ((m = m_dup(m, M_NOWAIT)) == NULL) {
2668 			return (ENOMEM);
2669 		} else {
2670 			m_freem(*mp);
2671 			n = *mp = m;
2672 		}
2673 	}
2674 
2675 	/*
2676 	 * Determine where frame payload starts.
2677 	 * Jump over vlan headers if already present,
2678 	 * helpful for QinQ too.
2679 	 */
2680 	if (__predict_false(m->m_len < sizeof(*eh))) {
2681 		txq->ift_pullups++;
2682 		if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL))
2683 			return (ENOMEM);
2684 	}
2685 	eh = mtod(m, struct ether_vlan_header *);
2686 	if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
2687 		pi->ipi_etype = ntohs(eh->evl_proto);
2688 		pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
2689 	} else {
2690 		pi->ipi_etype = ntohs(eh->evl_encap_proto);
2691 		pi->ipi_ehdrlen = ETHER_HDR_LEN;
2692 	}
2693 
2694 	switch (pi->ipi_etype) {
2695 #ifdef INET
2696 	case ETHERTYPE_IP:
2697 	{
2698 		struct ip *ip = NULL;
2699 		struct tcphdr *th = NULL;
2700 		int minthlen;
2701 
2702 		minthlen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip) + sizeof(*th));
2703 		if (__predict_false(m->m_len < minthlen)) {
2704 			/*
2705 			 * if this code bloat is causing too much of a hit
2706 			 * move it to a separate function and mark it noinline
2707 			 */
2708 			if (m->m_len == pi->ipi_ehdrlen) {
2709 				n = m->m_next;
2710 				MPASS(n);
2711 				if (n->m_len >= sizeof(*ip))  {
2712 					ip = (struct ip *)n->m_data;
2713 					if (n->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2714 						th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2715 				} else {
2716 					txq->ift_pullups++;
2717 					if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
2718 						return (ENOMEM);
2719 					ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2720 				}
2721 			} else {
2722 				txq->ift_pullups++;
2723 				if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
2724 					return (ENOMEM);
2725 				ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2726 				if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2727 					th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2728 			}
2729 		} else {
2730 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2731 			if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2732 				th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2733 		}
2734 		pi->ipi_ip_hlen = ip->ip_hl << 2;
2735 		pi->ipi_ipproto = ip->ip_p;
2736 		pi->ipi_flags |= IPI_TX_IPV4;
2737 
2738 		if (pi->ipi_csum_flags & CSUM_IP)
2739                        ip->ip_sum = 0;
2740 
2741 		if (pi->ipi_ipproto == IPPROTO_TCP) {
2742 			if (__predict_false(th == NULL)) {
2743 				txq->ift_pullups++;
2744 				if (__predict_false((m = m_pullup(m, (ip->ip_hl << 2) + sizeof(*th))) == NULL))
2745 					return (ENOMEM);
2746 				th = (struct tcphdr *)((caddr_t)ip + pi->ipi_ip_hlen);
2747 			}
2748 			pi->ipi_tcp_hflags = th->th_flags;
2749 			pi->ipi_tcp_hlen = th->th_off << 2;
2750 			pi->ipi_tcp_seq = th->th_seq;
2751 		}
2752 		if (IS_TSO4(pi)) {
2753 			if (__predict_false(ip->ip_p != IPPROTO_TCP))
2754 				return (ENXIO);
2755 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
2756 					       ip->ip_dst.s_addr, htons(IPPROTO_TCP));
2757 			pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
2758 			if (sctx->isc_flags & IFLIB_TSO_INIT_IP) {
2759 				ip->ip_sum = 0;
2760 				ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz);
2761 			}
2762 		}
2763 		break;
2764 	}
2765 #endif
2766 #ifdef INET6
2767 	case ETHERTYPE_IPV6:
2768 	{
2769 		struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
2770 		struct tcphdr *th;
2771 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
2772 
2773 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
2774 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
2775 				return (ENOMEM);
2776 		}
2777 		th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
2778 
2779 		/* XXX-BZ this will go badly in case of ext hdrs. */
2780 		pi->ipi_ipproto = ip6->ip6_nxt;
2781 		pi->ipi_flags |= IPI_TX_IPV6;
2782 
2783 		if (pi->ipi_ipproto == IPPROTO_TCP) {
2784 			if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) {
2785 				if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL))
2786 					return (ENOMEM);
2787 			}
2788 			pi->ipi_tcp_hflags = th->th_flags;
2789 			pi->ipi_tcp_hlen = th->th_off << 2;
2790 		}
2791 		if (IS_TSO6(pi)) {
2792 
2793 			if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP))
2794 				return (ENXIO);
2795 			/*
2796 			 * The corresponding flag is set by the stack in the IPv4
2797 			 * TSO case, but not in IPv6 (at least in FreeBSD 10.2).
2798 			 * So, set it here because the rest of the flow requires it.
2799 			 */
2800 			pi->ipi_csum_flags |= CSUM_TCP_IPV6;
2801 			th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0);
2802 			pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
2803 		}
2804 		break;
2805 	}
2806 #endif
2807 	default:
2808 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
2809 		pi->ipi_ip_hlen = 0;
2810 		break;
2811 	}
2812 	*mp = m;
2813 
2814 	return (0);
2815 }
2816 
2817 static  __noinline  struct mbuf *
2818 collapse_pkthdr(struct mbuf *m0)
2819 {
2820 	struct mbuf *m, *m_next, *tmp;
2821 
2822 	m = m0;
2823 	m_next = m->m_next;
2824 	while (m_next != NULL && m_next->m_len == 0) {
2825 		m = m_next;
2826 		m->m_next = NULL;
2827 		m_free(m);
2828 		m_next = m_next->m_next;
2829 	}
2830 	m = m0;
2831 	m->m_next = m_next;
2832 	if ((m_next->m_flags & M_EXT) == 0) {
2833 		m = m_defrag(m, M_NOWAIT);
2834 	} else {
2835 		tmp = m_next->m_next;
2836 		memcpy(m_next, m, MPKTHSIZE);
2837 		m = m_next;
2838 		m->m_next = tmp;
2839 	}
2840 	return (m);
2841 }
2842 
2843 /*
2844  * If dodgy hardware rejects the scatter gather chain we've handed it
2845  * we'll need to remove the mbuf chain from ifsg_m[] before we can add the
2846  * m_defrag'd mbufs
2847  */
2848 static __noinline struct mbuf *
2849 iflib_remove_mbuf(iflib_txq_t txq)
2850 {
2851 	int ntxd, i, pidx;
2852 	struct mbuf *m, *mh, **ifsd_m;
2853 
2854 	pidx = txq->ift_pidx;
2855 	ifsd_m = txq->ift_sds.ifsd_m;
2856 	ntxd = txq->ift_size;
2857 	mh = m = ifsd_m[pidx];
2858 	ifsd_m[pidx] = NULL;
2859 #if MEMORY_LOGGING
2860 	txq->ift_dequeued++;
2861 #endif
2862 	i = 1;
2863 
2864 	while (m) {
2865 		ifsd_m[(pidx + i) & (ntxd -1)] = NULL;
2866 #if MEMORY_LOGGING
2867 		txq->ift_dequeued++;
2868 #endif
2869 		m = m->m_next;
2870 		i++;
2871 	}
2872 	return (mh);
2873 }
2874 
2875 static int
2876 iflib_busdma_load_mbuf_sg(iflib_txq_t txq, bus_dma_tag_t tag, bus_dmamap_t map,
2877 			  struct mbuf **m0, bus_dma_segment_t *segs, int *nsegs,
2878 			  int max_segs, int flags)
2879 {
2880 	if_ctx_t ctx;
2881 	if_shared_ctx_t		sctx;
2882 	if_softc_ctx_t		scctx;
2883 	int i, next, pidx, mask, err, maxsegsz, ntxd, count;
2884 	struct mbuf *m, *tmp, **ifsd_m, **mp;
2885 
2886 	m = *m0;
2887 
2888 	/*
2889 	 * Please don't ever do this
2890 	 */
2891 	if (__predict_false(m->m_len == 0))
2892 		*m0 = m = collapse_pkthdr(m);
2893 
2894 	ctx = txq->ift_ctx;
2895 	sctx = ctx->ifc_sctx;
2896 	scctx = &ctx->ifc_softc_ctx;
2897 	ifsd_m = txq->ift_sds.ifsd_m;
2898 	ntxd = txq->ift_size;
2899 	pidx = txq->ift_pidx;
2900 	if (map != NULL) {
2901 		uint8_t *ifsd_flags = txq->ift_sds.ifsd_flags;
2902 
2903 		err = bus_dmamap_load_mbuf_sg(tag, map,
2904 					      *m0, segs, nsegs, BUS_DMA_NOWAIT);
2905 		if (err)
2906 			return (err);
2907 		ifsd_flags[pidx] |= TX_SW_DESC_MAPPED;
2908 		i = 0;
2909 		next = pidx;
2910 		mask = (txq->ift_size-1);
2911 		m = *m0;
2912 		do {
2913 			mp = &ifsd_m[next];
2914 			*mp = m;
2915 			m = m->m_next;
2916 			if (__predict_false((*mp)->m_len == 0)) {
2917 				m_free(*mp);
2918 				*mp = NULL;
2919 			} else
2920 				next = (pidx + i) & (ntxd-1);
2921 		} while (m != NULL);
2922 	} else {
2923 		int buflen, sgsize, max_sgsize;
2924 		vm_offset_t vaddr;
2925 		vm_paddr_t curaddr;
2926 
2927 		count = i = 0;
2928 		maxsegsz = sctx->isc_tx_maxsize;
2929 		m = *m0;
2930 		do {
2931 			if (__predict_false(m->m_len <= 0)) {
2932 				tmp = m;
2933 				m = m->m_next;
2934 				tmp->m_next = NULL;
2935 				m_free(tmp);
2936 				continue;
2937 			}
2938 			buflen = m->m_len;
2939 			vaddr = (vm_offset_t)m->m_data;
2940 			/*
2941 			 * see if we can't be smarter about physically
2942 			 * contiguous mappings
2943 			 */
2944 			next = (pidx + count) & (ntxd-1);
2945 			MPASS(ifsd_m[next] == NULL);
2946 #if MEMORY_LOGGING
2947 			txq->ift_enqueued++;
2948 #endif
2949 			ifsd_m[next] = m;
2950 			while (buflen > 0) {
2951 				max_sgsize = MIN(buflen, maxsegsz);
2952 				curaddr = pmap_kextract(vaddr);
2953 				sgsize = PAGE_SIZE - (curaddr & PAGE_MASK);
2954 				sgsize = MIN(sgsize, max_sgsize);
2955 				segs[i].ds_addr = curaddr;
2956 				segs[i].ds_len = sgsize;
2957 				vaddr += sgsize;
2958 				buflen -= sgsize;
2959 				i++;
2960 				if (i >= max_segs)
2961 					goto err;
2962 			}
2963 			count++;
2964 			tmp = m;
2965 			m = m->m_next;
2966 		} while (m != NULL);
2967 		*nsegs = i;
2968 	}
2969 	return (0);
2970 err:
2971 	*m0 = iflib_remove_mbuf(txq);
2972 	return (EFBIG);
2973 }
2974 
2975 static inline caddr_t
2976 calc_next_txd(iflib_txq_t txq, int cidx, uint8_t qid)
2977 {
2978 	qidx_t size;
2979 	int ntxd;
2980 	caddr_t start, end, cur, next;
2981 
2982 	ntxd = txq->ift_size;
2983 	size = txq->ift_txd_size[qid];
2984 	start = txq->ift_ifdi[qid].idi_vaddr;
2985 
2986 	if (__predict_false(size == 0))
2987 		return (start);
2988 	cur = start + size*cidx;
2989 	end = start + size*ntxd;
2990 	next = CACHE_PTR_NEXT(cur);
2991 	return (next < end ? next : start);
2992 }
2993 
2994 static int
2995 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp)
2996 {
2997 	if_ctx_t		ctx;
2998 	if_shared_ctx_t		sctx;
2999 	if_softc_ctx_t		scctx;
3000 	bus_dma_segment_t	*segs;
3001 	struct mbuf		*m_head;
3002 	void			*next_txd;
3003 	bus_dmamap_t		map;
3004 	struct if_pkt_info	pi;
3005 	int remap = 0;
3006 	int err, nsegs, ndesc, max_segs, pidx, cidx, next, ntxd;
3007 	bus_dma_tag_t desc_tag;
3008 
3009 	segs = txq->ift_segs;
3010 	ctx = txq->ift_ctx;
3011 	sctx = ctx->ifc_sctx;
3012 	scctx = &ctx->ifc_softc_ctx;
3013 	segs = txq->ift_segs;
3014 	ntxd = txq->ift_size;
3015 	m_head = *m_headp;
3016 	map = NULL;
3017 
3018 	/*
3019 	 * If we're doing TSO the next descriptor to clean may be quite far ahead
3020 	 */
3021 	cidx = txq->ift_cidx;
3022 	pidx = txq->ift_pidx;
3023 	if (ctx->ifc_flags & IFC_PREFETCH) {
3024 		next = (cidx + CACHE_PTR_INCREMENT) & (ntxd-1);
3025 		if (!(ctx->ifc_flags & IFLIB_HAS_TXCQ)) {
3026 			next_txd = calc_next_txd(txq, cidx, 0);
3027 			prefetch(next_txd);
3028 		}
3029 
3030 		/* prefetch the next cache line of mbuf pointers and flags */
3031 		prefetch(&txq->ift_sds.ifsd_m[next]);
3032 		if (txq->ift_sds.ifsd_map != NULL) {
3033 			prefetch(&txq->ift_sds.ifsd_map[next]);
3034 			next = (cidx + CACHE_LINE_SIZE) & (ntxd-1);
3035 			prefetch(&txq->ift_sds.ifsd_flags[next]);
3036 		}
3037 	} else if (txq->ift_sds.ifsd_map != NULL)
3038 		map = txq->ift_sds.ifsd_map[pidx];
3039 
3040 	if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
3041 		desc_tag = txq->ift_tso_desc_tag;
3042 		max_segs = scctx->isc_tx_tso_segments_max;
3043 	} else {
3044 		desc_tag = txq->ift_desc_tag;
3045 		max_segs = scctx->isc_tx_nsegments;
3046 	}
3047 	m_head = *m_headp;
3048 
3049 	pkt_info_zero(&pi);
3050 	pi.ipi_len = m_head->m_pkthdr.len;
3051 	pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG|M_BCAST|M_MCAST));
3052 	pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags;
3053 	pi.ipi_vtag = (m_head->m_flags & M_VLANTAG) ? m_head->m_pkthdr.ether_vtag : 0;
3054 	pi.ipi_pidx = pidx;
3055 	pi.ipi_qsidx = txq->ift_id;
3056 
3057 	/* deliberate bitwise OR to make one condition */
3058 	if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) {
3059 		if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0))
3060 			return (err);
3061 		m_head = *m_headp;
3062 	}
3063 
3064 retry:
3065 	err = iflib_busdma_load_mbuf_sg(txq, desc_tag, map, m_headp, segs, &nsegs, max_segs, BUS_DMA_NOWAIT);
3066 defrag:
3067 	if (__predict_false(err)) {
3068 		switch (err) {
3069 		case EFBIG:
3070 			/* try collapse once and defrag once */
3071 			if (remap == 0)
3072 				m_head = m_collapse(*m_headp, M_NOWAIT, max_segs);
3073 			if (remap == 1)
3074 				m_head = m_defrag(*m_headp, M_NOWAIT);
3075 			remap++;
3076 			if (__predict_false(m_head == NULL))
3077 				goto defrag_failed;
3078 			txq->ift_mbuf_defrag++;
3079 			*m_headp = m_head;
3080 			goto retry;
3081 			break;
3082 		case ENOMEM:
3083 			txq->ift_no_tx_dma_setup++;
3084 			break;
3085 		default:
3086 			txq->ift_no_tx_dma_setup++;
3087 			m_freem(*m_headp);
3088 			DBG_COUNTER_INC(tx_frees);
3089 			*m_headp = NULL;
3090 			break;
3091 		}
3092 		txq->ift_map_failed++;
3093 		DBG_COUNTER_INC(encap_load_mbuf_fail);
3094 		return (err);
3095 	}
3096 
3097 	/*
3098 	 * XXX assumes a 1 to 1 relationship between segments and
3099 	 *        descriptors - this does not hold true on all drivers, e.g.
3100 	 *        cxgb
3101 	 */
3102 	if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) {
3103 		txq->ift_no_desc_avail++;
3104 		if (map != NULL)
3105 			bus_dmamap_unload(desc_tag, map);
3106 		DBG_COUNTER_INC(encap_txq_avail_fail);
3107 		if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0)
3108 			GROUPTASK_ENQUEUE(&txq->ift_task);
3109 		return (ENOBUFS);
3110 	}
3111 	/*
3112 	 * On Intel cards we can greatly reduce the number of TX interrupts
3113 	 * we see by only setting report status on every Nth descriptor.
3114 	 * However, this also means that the driver will need to keep track
3115 	 * of the descriptors that RS was set on to check them for the DD bit.
3116 	 */
3117 	txq->ift_rs_pending += nsegs + 1;
3118 	if (txq->ift_rs_pending > TXQ_MAX_RS_DEFERRED(txq) ||
3119 	     iflib_no_tx_batch || (TXQ_AVAIL(txq) - nsegs - 1) <= MAX_TX_DESC(ctx)) {
3120 		pi.ipi_flags |= IPI_TX_INTR;
3121 		txq->ift_rs_pending = 0;
3122 	}
3123 
3124 	pi.ipi_segs = segs;
3125 	pi.ipi_nsegs = nsegs;
3126 
3127 	MPASS(pidx >= 0 && pidx < txq->ift_size);
3128 #ifdef PKT_DEBUG
3129 	print_pkt(&pi);
3130 #endif
3131 	if (map != NULL)
3132 		bus_dmamap_sync(desc_tag, map, BUS_DMASYNC_PREWRITE);
3133 	if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) {
3134 		if (map != NULL)
3135 			bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3136 					BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
3137 		DBG_COUNTER_INC(tx_encap);
3138 		MPASS(pi.ipi_new_pidx < txq->ift_size);
3139 
3140 		ndesc = pi.ipi_new_pidx - pi.ipi_pidx;
3141 		if (pi.ipi_new_pidx < pi.ipi_pidx) {
3142 			ndesc += txq->ift_size;
3143 			txq->ift_gen = 1;
3144 		}
3145 		/*
3146 		 * drivers can need as many as
3147 		 * two sentinels
3148 		 */
3149 		MPASS(ndesc <= pi.ipi_nsegs + 2);
3150 		MPASS(pi.ipi_new_pidx != pidx);
3151 		MPASS(ndesc > 0);
3152 		txq->ift_in_use += ndesc;
3153 
3154 		/*
3155 		 * We update the last software descriptor again here because there may
3156 		 * be a sentinel and/or there may be more mbufs than segments
3157 		 */
3158 		txq->ift_pidx = pi.ipi_new_pidx;
3159 		txq->ift_npending += pi.ipi_ndescs;
3160 	} else if (__predict_false(err == EFBIG && remap < 2)) {
3161 		*m_headp = m_head = iflib_remove_mbuf(txq);
3162 		remap = 1;
3163 		txq->ift_txd_encap_efbig++;
3164 		goto defrag;
3165 	} else
3166 		DBG_COUNTER_INC(encap_txd_encap_fail);
3167 	return (err);
3168 
3169 defrag_failed:
3170 	txq->ift_mbuf_defrag_failed++;
3171 	txq->ift_map_failed++;
3172 	m_freem(*m_headp);
3173 	DBG_COUNTER_INC(tx_frees);
3174 	*m_headp = NULL;
3175 	return (ENOMEM);
3176 }
3177 
3178 static void
3179 iflib_tx_desc_free(iflib_txq_t txq, int n)
3180 {
3181 	int hasmap;
3182 	uint32_t qsize, cidx, mask, gen;
3183 	struct mbuf *m, **ifsd_m;
3184 	uint8_t *ifsd_flags;
3185 	bus_dmamap_t *ifsd_map;
3186 	bool do_prefetch;
3187 
3188 	cidx = txq->ift_cidx;
3189 	gen = txq->ift_gen;
3190 	qsize = txq->ift_size;
3191 	mask = qsize-1;
3192 	hasmap = txq->ift_sds.ifsd_map != NULL;
3193 	ifsd_flags = txq->ift_sds.ifsd_flags;
3194 	ifsd_m = txq->ift_sds.ifsd_m;
3195 	ifsd_map = txq->ift_sds.ifsd_map;
3196 	do_prefetch = (txq->ift_ctx->ifc_flags & IFC_PREFETCH);
3197 
3198 	while (n--) {
3199 		if (do_prefetch) {
3200 			prefetch(ifsd_m[(cidx + 3) & mask]);
3201 			prefetch(ifsd_m[(cidx + 4) & mask]);
3202 		}
3203 		if (ifsd_m[cidx] != NULL) {
3204 			prefetch(&ifsd_m[(cidx + CACHE_PTR_INCREMENT) & mask]);
3205 			prefetch(&ifsd_flags[(cidx + CACHE_PTR_INCREMENT) & mask]);
3206 			if (hasmap && (ifsd_flags[cidx] & TX_SW_DESC_MAPPED)) {
3207 				/*
3208 				 * does it matter if it's not the TSO tag? If so we'll
3209 				 * have to add the type to flags
3210 				 */
3211 				bus_dmamap_unload(txq->ift_desc_tag, ifsd_map[cidx]);
3212 				ifsd_flags[cidx] &= ~TX_SW_DESC_MAPPED;
3213 			}
3214 			if ((m = ifsd_m[cidx]) != NULL) {
3215 				/* XXX we don't support any drivers that batch packets yet */
3216 				MPASS(m->m_nextpkt == NULL);
3217 
3218 				m_free(m);
3219 				ifsd_m[cidx] = NULL;
3220 #if MEMORY_LOGGING
3221 				txq->ift_dequeued++;
3222 #endif
3223 				DBG_COUNTER_INC(tx_frees);
3224 			}
3225 		}
3226 		if (__predict_false(++cidx == qsize)) {
3227 			cidx = 0;
3228 			gen = 0;
3229 		}
3230 	}
3231 	txq->ift_cidx = cidx;
3232 	txq->ift_gen = gen;
3233 }
3234 
3235 static __inline int
3236 iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh)
3237 {
3238 	int reclaim;
3239 	if_ctx_t ctx = txq->ift_ctx;
3240 
3241 	KASSERT(thresh >= 0, ("invalid threshold to reclaim"));
3242 	MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size);
3243 
3244 	/*
3245 	 * Need a rate-limiting check so that this isn't called every time
3246 	 */
3247 	iflib_tx_credits_update(ctx, txq);
3248 	reclaim = DESC_RECLAIMABLE(txq);
3249 
3250 	if (reclaim <= thresh /* + MAX_TX_DESC(txq->ift_ctx) */) {
3251 #ifdef INVARIANTS
3252 		if (iflib_verbose_debug) {
3253 			printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __FUNCTION__,
3254 			       txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments,
3255 			       reclaim, thresh);
3256 
3257 		}
3258 #endif
3259 		return (0);
3260 	}
3261 	iflib_tx_desc_free(txq, reclaim);
3262 	txq->ift_cleaned += reclaim;
3263 	txq->ift_in_use -= reclaim;
3264 
3265 	return (reclaim);
3266 }
3267 
3268 static struct mbuf **
3269 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset, int remaining)
3270 {
3271 	int next, size;
3272 	struct mbuf **items;
3273 
3274 	size = r->size;
3275 	next = (cidx + CACHE_PTR_INCREMENT) & (size-1);
3276 	items = __DEVOLATILE(struct mbuf **, &r->items[0]);
3277 
3278 	prefetch(items[(cidx + offset) & (size-1)]);
3279 	if (remaining > 1) {
3280 		prefetch(&items[next]);
3281 		prefetch(items[(cidx + offset + 1) & (size-1)]);
3282 		prefetch(items[(cidx + offset + 2) & (size-1)]);
3283 		prefetch(items[(cidx + offset + 3) & (size-1)]);
3284 	}
3285 	return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (size-1)]));
3286 }
3287 
3288 static void
3289 iflib_txq_check_drain(iflib_txq_t txq, int budget)
3290 {
3291 
3292 	ifmp_ring_check_drainage(txq->ift_br, budget);
3293 }
3294 
3295 static uint32_t
3296 iflib_txq_can_drain(struct ifmp_ring *r)
3297 {
3298 	iflib_txq_t txq = r->cookie;
3299 	if_ctx_t ctx = txq->ift_ctx;
3300 
3301 	return ((TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2) ||
3302 		ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false));
3303 }
3304 
3305 static uint32_t
3306 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3307 {
3308 	iflib_txq_t txq = r->cookie;
3309 	if_ctx_t ctx = txq->ift_ctx;
3310 	struct ifnet *ifp = ctx->ifc_ifp;
3311 	struct mbuf **mp, *m;
3312 	int i, count, consumed, pkt_sent, bytes_sent, mcast_sent, avail;
3313 	int reclaimed, err, in_use_prev, desc_used;
3314 	bool do_prefetch, ring, rang;
3315 
3316 	if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) ||
3317 			    !LINK_ACTIVE(ctx))) {
3318 		DBG_COUNTER_INC(txq_drain_notready);
3319 		return (0);
3320 	}
3321 	reclaimed = iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
3322 	rang = iflib_txd_db_check(ctx, txq, reclaimed, txq->ift_in_use);
3323 	avail = IDXDIFF(pidx, cidx, r->size);
3324 	if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) {
3325 		DBG_COUNTER_INC(txq_drain_flushing);
3326 		for (i = 0; i < avail; i++) {
3327 			m_free(r->items[(cidx + i) & (r->size-1)]);
3328 			r->items[(cidx + i) & (r->size-1)] = NULL;
3329 		}
3330 		return (avail);
3331 	}
3332 
3333 	if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3334 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3335 		CALLOUT_LOCK(txq);
3336 		callout_stop(&txq->ift_timer);
3337 		CALLOUT_UNLOCK(txq);
3338 		DBG_COUNTER_INC(txq_drain_oactive);
3339 		return (0);
3340 	}
3341 	if (reclaimed)
3342 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3343 	consumed = mcast_sent = bytes_sent = pkt_sent = 0;
3344 	count = MIN(avail, TX_BATCH_SIZE);
3345 #ifdef INVARIANTS
3346 	if (iflib_verbose_debug)
3347 		printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __FUNCTION__,
3348 		       avail, ctx->ifc_flags, TXQ_AVAIL(txq));
3349 #endif
3350 	do_prefetch = (ctx->ifc_flags & IFC_PREFETCH);
3351 	avail = TXQ_AVAIL(txq);
3352 	for (desc_used = i = 0; i < count && avail > MAX_TX_DESC(ctx) + 2; i++) {
3353 		int pidx_prev, rem = do_prefetch ? count - i : 0;
3354 
3355 		mp = _ring_peek_one(r, cidx, i, rem);
3356 		MPASS(mp != NULL && *mp != NULL);
3357 		if (__predict_false(*mp == (struct mbuf *)txq)) {
3358 			consumed++;
3359 			reclaimed++;
3360 			continue;
3361 		}
3362 		in_use_prev = txq->ift_in_use;
3363 		pidx_prev = txq->ift_pidx;
3364 		err = iflib_encap(txq, mp);
3365 		if (__predict_false(err)) {
3366 			DBG_COUNTER_INC(txq_drain_encapfail);
3367 			/* no room - bail out */
3368 			if (err == ENOBUFS)
3369 				break;
3370 			consumed++;
3371 			DBG_COUNTER_INC(txq_drain_encapfail);
3372 			/* we can't send this packet - skip it */
3373 			continue;
3374 		}
3375 		consumed++;
3376 		pkt_sent++;
3377 		m = *mp;
3378 		DBG_COUNTER_INC(tx_sent);
3379 		bytes_sent += m->m_pkthdr.len;
3380 		mcast_sent += !!(m->m_flags & M_MCAST);
3381 		avail = TXQ_AVAIL(txq);
3382 
3383 		txq->ift_db_pending += (txq->ift_in_use - in_use_prev);
3384 		desc_used += (txq->ift_in_use - in_use_prev);
3385 		ETHER_BPF_MTAP(ifp, m);
3386 		if (__predict_false(!(ifp->if_drv_flags & IFF_DRV_RUNNING)))
3387 			break;
3388 		rang = iflib_txd_db_check(ctx, txq, false, in_use_prev);
3389 	}
3390 
3391 	/* deliberate use of bitwise or to avoid gratuitous short-circuit */
3392 	ring = rang ? false  : (iflib_min_tx_latency | err) || (TXQ_AVAIL(txq) < MAX_TX_DESC(ctx));
3393 	iflib_txd_db_check(ctx, txq, ring, txq->ift_in_use);
3394 	if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
3395 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
3396 	if (mcast_sent)
3397 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
3398 #ifdef INVARIANTS
3399 	if (iflib_verbose_debug)
3400 		printf("consumed=%d\n", consumed);
3401 #endif
3402 	return (consumed);
3403 }
3404 
3405 static uint32_t
3406 iflib_txq_drain_always(struct ifmp_ring *r)
3407 {
3408 	return (1);
3409 }
3410 
3411 static uint32_t
3412 iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3413 {
3414 	int i, avail;
3415 	struct mbuf **mp;
3416 	iflib_txq_t txq;
3417 
3418 	txq = r->cookie;
3419 
3420 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3421 	CALLOUT_LOCK(txq);
3422 	callout_stop(&txq->ift_timer);
3423 	CALLOUT_UNLOCK(txq);
3424 
3425 	avail = IDXDIFF(pidx, cidx, r->size);
3426 	for (i = 0; i < avail; i++) {
3427 		mp = _ring_peek_one(r, cidx, i, avail - i);
3428 		if (__predict_false(*mp == (struct mbuf *)txq))
3429 			continue;
3430 		m_freem(*mp);
3431 	}
3432 	MPASS(ifmp_ring_is_stalled(r) == 0);
3433 	return (avail);
3434 }
3435 
3436 static void
3437 iflib_ifmp_purge(iflib_txq_t txq)
3438 {
3439 	struct ifmp_ring *r;
3440 
3441 	r = txq->ift_br;
3442 	r->drain = iflib_txq_drain_free;
3443 	r->can_drain = iflib_txq_drain_always;
3444 
3445 	ifmp_ring_check_drainage(r, r->size);
3446 
3447 	r->drain = iflib_txq_drain;
3448 	r->can_drain = iflib_txq_can_drain;
3449 }
3450 
3451 static void
3452 _task_fn_tx(void *context)
3453 {
3454 	iflib_txq_t txq = context;
3455 	if_ctx_t ctx = txq->ift_ctx;
3456 	struct ifnet *ifp = ctx->ifc_ifp;
3457 	int rc;
3458 
3459 #ifdef IFLIB_DIAGNOSTICS
3460 	txq->ift_cpu_exec_count[curcpu]++;
3461 #endif
3462 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3463 		return;
3464 	if ((ifp->if_capenable & IFCAP_NETMAP)) {
3465 		if (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false))
3466 			netmap_tx_irq(ifp, txq->ift_id);
3467 		IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3468 		return;
3469 	}
3470 	if (txq->ift_db_pending)
3471 		ifmp_ring_enqueue(txq->ift_br, (void **)&txq, 1, TX_BATCH_SIZE);
3472 	else
3473 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3474 	if (ctx->ifc_flags & IFC_LEGACY)
3475 		IFDI_INTR_ENABLE(ctx);
3476 	else {
3477 		rc = IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3478 		KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver"));
3479 	}
3480 }
3481 
3482 static void
3483 _task_fn_rx(void *context)
3484 {
3485 	iflib_rxq_t rxq = context;
3486 	if_ctx_t ctx = rxq->ifr_ctx;
3487 	bool more;
3488 	int rc;
3489 
3490 #ifdef IFLIB_DIAGNOSTICS
3491 	rxq->ifr_cpu_exec_count[curcpu]++;
3492 #endif
3493 	DBG_COUNTER_INC(task_fn_rxs);
3494 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3495 		return;
3496 	if ((more = iflib_rxeof(rxq, 16 /* XXX */)) == false) {
3497 		if (ctx->ifc_flags & IFC_LEGACY)
3498 			IFDI_INTR_ENABLE(ctx);
3499 		else {
3500 			DBG_COUNTER_INC(rx_intr_enables);
3501 			rc = IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
3502 			KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver"));
3503 		}
3504 	}
3505 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3506 		return;
3507 	if (more)
3508 		GROUPTASK_ENQUEUE(&rxq->ifr_task);
3509 }
3510 
3511 static void
3512 _task_fn_admin(void *context)
3513 {
3514 	if_ctx_t ctx = context;
3515 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
3516 	iflib_txq_t txq;
3517 	int i;
3518 
3519 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3520 		return;
3521 
3522 	CTX_LOCK(ctx);
3523 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3524 		CALLOUT_LOCK(txq);
3525 		callout_stop(&txq->ift_timer);
3526 		CALLOUT_UNLOCK(txq);
3527 	}
3528 	IFDI_UPDATE_ADMIN_STATUS(ctx);
3529 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3530 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq, txq->ift_timer.c_cpu);
3531 	IFDI_LINK_INTR_ENABLE(ctx);
3532 	if (ctx->ifc_flags & IFC_DO_RESET) {
3533 		ctx->ifc_flags &= ~IFC_DO_RESET;
3534 		iflib_if_init_locked(ctx);
3535 	}
3536 	CTX_UNLOCK(ctx);
3537 
3538 	if (LINK_ACTIVE(ctx) == 0)
3539 		return;
3540 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3541 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
3542 }
3543 
3544 
3545 static void
3546 _task_fn_iov(void *context)
3547 {
3548 	if_ctx_t ctx = context;
3549 
3550 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3551 		return;
3552 
3553 	CTX_LOCK(ctx);
3554 	IFDI_VFLR_HANDLE(ctx);
3555 	CTX_UNLOCK(ctx);
3556 }
3557 
3558 static int
3559 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)
3560 {
3561 	int err;
3562 	if_int_delay_info_t info;
3563 	if_ctx_t ctx;
3564 
3565 	info = (if_int_delay_info_t)arg1;
3566 	ctx = info->iidi_ctx;
3567 	info->iidi_req = req;
3568 	info->iidi_oidp = oidp;
3569 	CTX_LOCK(ctx);
3570 	err = IFDI_SYSCTL_INT_DELAY(ctx, info);
3571 	CTX_UNLOCK(ctx);
3572 	return (err);
3573 }
3574 
3575 /*********************************************************************
3576  *
3577  *  IFNET FUNCTIONS
3578  *
3579  **********************************************************************/
3580 
3581 static void
3582 iflib_if_init_locked(if_ctx_t ctx)
3583 {
3584 	iflib_stop(ctx);
3585 	iflib_init_locked(ctx);
3586 }
3587 
3588 
3589 static void
3590 iflib_if_init(void *arg)
3591 {
3592 	if_ctx_t ctx = arg;
3593 
3594 	CTX_LOCK(ctx);
3595 	iflib_if_init_locked(ctx);
3596 	CTX_UNLOCK(ctx);
3597 }
3598 
3599 static int
3600 iflib_if_transmit(if_t ifp, struct mbuf *m)
3601 {
3602 	if_ctx_t	ctx = if_getsoftc(ifp);
3603 
3604 	iflib_txq_t txq;
3605 	int err, qidx;
3606 
3607 	if (__predict_false((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) {
3608 		DBG_COUNTER_INC(tx_frees);
3609 		m_freem(m);
3610 		return (ENOBUFS);
3611 	}
3612 
3613 	MPASS(m->m_nextpkt == NULL);
3614 	qidx = 0;
3615 	if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m))
3616 		qidx = QIDX(ctx, m);
3617 	/*
3618 	 * XXX calculate buf_ring based on flowid (divvy up bits?)
3619 	 */
3620 	txq = &ctx->ifc_txqs[qidx];
3621 
3622 #ifdef DRIVER_BACKPRESSURE
3623 	if (txq->ift_closed) {
3624 		while (m != NULL) {
3625 			next = m->m_nextpkt;
3626 			m->m_nextpkt = NULL;
3627 			m_freem(m);
3628 			m = next;
3629 		}
3630 		return (ENOBUFS);
3631 	}
3632 #endif
3633 #ifdef notyet
3634 	qidx = count = 0;
3635 	mp = marr;
3636 	next = m;
3637 	do {
3638 		count++;
3639 		next = next->m_nextpkt;
3640 	} while (next != NULL);
3641 
3642 	if (count > nitems(marr))
3643 		if ((mp = malloc(count*sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) {
3644 			/* XXX check nextpkt */
3645 			m_freem(m);
3646 			/* XXX simplify for now */
3647 			DBG_COUNTER_INC(tx_frees);
3648 			return (ENOBUFS);
3649 		}
3650 	for (next = m, i = 0; next != NULL; i++) {
3651 		mp[i] = next;
3652 		next = next->m_nextpkt;
3653 		mp[i]->m_nextpkt = NULL;
3654 	}
3655 #endif
3656 	DBG_COUNTER_INC(tx_seen);
3657 	err = ifmp_ring_enqueue(txq->ift_br, (void **)&m, 1, TX_BATCH_SIZE);
3658 
3659 	if (err) {
3660 		GROUPTASK_ENQUEUE(&txq->ift_task);
3661 		/* support forthcoming later */
3662 #ifdef DRIVER_BACKPRESSURE
3663 		txq->ift_closed = TRUE;
3664 #endif
3665 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3666 		m_freem(m);
3667 	} else if (TXQ_AVAIL(txq) < (txq->ift_size >> 1)) {
3668 		GROUPTASK_ENQUEUE(&txq->ift_task);
3669 	}
3670 
3671 	return (err);
3672 }
3673 
3674 static void
3675 iflib_if_qflush(if_t ifp)
3676 {
3677 	if_ctx_t ctx = if_getsoftc(ifp);
3678 	iflib_txq_t txq = ctx->ifc_txqs;
3679 	int i;
3680 
3681 	CTX_LOCK(ctx);
3682 	ctx->ifc_flags |= IFC_QFLUSH;
3683 	CTX_UNLOCK(ctx);
3684 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
3685 		while (!(ifmp_ring_is_idle(txq->ift_br) || ifmp_ring_is_stalled(txq->ift_br)))
3686 			iflib_txq_check_drain(txq, 0);
3687 	CTX_LOCK(ctx);
3688 	ctx->ifc_flags &= ~IFC_QFLUSH;
3689 	CTX_UNLOCK(ctx);
3690 
3691 	if_qflush(ifp);
3692 }
3693 
3694 
3695 #define IFCAP_FLAGS (IFCAP_TXCSUM_IPV6 | IFCAP_RXCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \
3696 		     IFCAP_TSO4 | IFCAP_TSO6 | IFCAP_VLAN_HWTAGGING |	\
3697 		     IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | IFCAP_VLAN_HWTSO)
3698 
3699 static int
3700 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data)
3701 {
3702 	if_ctx_t ctx = if_getsoftc(ifp);
3703 	struct ifreq	*ifr = (struct ifreq *)data;
3704 #if defined(INET) || defined(INET6)
3705 	struct ifaddr	*ifa = (struct ifaddr *)data;
3706 #endif
3707 	bool		avoid_reset = FALSE;
3708 	int		err = 0, reinit = 0, bits;
3709 
3710 	switch (command) {
3711 	case SIOCSIFADDR:
3712 #ifdef INET
3713 		if (ifa->ifa_addr->sa_family == AF_INET)
3714 			avoid_reset = TRUE;
3715 #endif
3716 #ifdef INET6
3717 		if (ifa->ifa_addr->sa_family == AF_INET6)
3718 			avoid_reset = TRUE;
3719 #endif
3720 		/*
3721 		** Calling init results in link renegotiation,
3722 		** so we avoid doing it when possible.
3723 		*/
3724 		if (avoid_reset) {
3725 			if_setflagbits(ifp, IFF_UP,0);
3726 			if (!(if_getdrvflags(ifp)& IFF_DRV_RUNNING))
3727 				reinit = 1;
3728 #ifdef INET
3729 			if (!(if_getflags(ifp) & IFF_NOARP))
3730 				arp_ifinit(ifp, ifa);
3731 #endif
3732 		} else
3733 			err = ether_ioctl(ifp, command, data);
3734 		break;
3735 	case SIOCSIFMTU:
3736 		CTX_LOCK(ctx);
3737 		if (ifr->ifr_mtu == if_getmtu(ifp)) {
3738 			CTX_UNLOCK(ctx);
3739 			break;
3740 		}
3741 		bits = if_getdrvflags(ifp);
3742 		/* stop the driver and free any clusters before proceeding */
3743 		iflib_stop(ctx);
3744 
3745 		if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) {
3746 			if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size)
3747 				ctx->ifc_flags |= IFC_MULTISEG;
3748 			else
3749 				ctx->ifc_flags &= ~IFC_MULTISEG;
3750 			err = if_setmtu(ifp, ifr->ifr_mtu);
3751 		}
3752 		iflib_init_locked(ctx);
3753 		if_setdrvflags(ifp, bits);
3754 		CTX_UNLOCK(ctx);
3755 		break;
3756 	case SIOCSIFFLAGS:
3757 		CTX_LOCK(ctx);
3758 		if (if_getflags(ifp) & IFF_UP) {
3759 			if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
3760 				if ((if_getflags(ifp) ^ ctx->ifc_if_flags) &
3761 				    (IFF_PROMISC | IFF_ALLMULTI)) {
3762 					err = IFDI_PROMISC_SET(ctx, if_getflags(ifp));
3763 				}
3764 			} else
3765 				reinit = 1;
3766 		} else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
3767 			iflib_stop(ctx);
3768 		}
3769 		ctx->ifc_if_flags = if_getflags(ifp);
3770 		CTX_UNLOCK(ctx);
3771 		break;
3772 	case SIOCADDMULTI:
3773 	case SIOCDELMULTI:
3774 		if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
3775 			CTX_LOCK(ctx);
3776 			IFDI_INTR_DISABLE(ctx);
3777 			IFDI_MULTI_SET(ctx);
3778 			IFDI_INTR_ENABLE(ctx);
3779 			CTX_UNLOCK(ctx);
3780 		}
3781 		break;
3782 	case SIOCSIFMEDIA:
3783 		CTX_LOCK(ctx);
3784 		IFDI_MEDIA_SET(ctx);
3785 		CTX_UNLOCK(ctx);
3786 		/* falls thru */
3787 	case SIOCGIFMEDIA:
3788 		err = ifmedia_ioctl(ifp, ifr, &ctx->ifc_media, command);
3789 		break;
3790 	case SIOCGI2C:
3791 	{
3792 		struct ifi2creq i2c;
3793 
3794 		err = copyin(ifr->ifr_data, &i2c, sizeof(i2c));
3795 		if (err != 0)
3796 			break;
3797 		if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) {
3798 			err = EINVAL;
3799 			break;
3800 		}
3801 		if (i2c.len > sizeof(i2c.data)) {
3802 			err = EINVAL;
3803 			break;
3804 		}
3805 
3806 		if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0)
3807 			err = copyout(&i2c, ifr->ifr_data, sizeof(i2c));
3808 		break;
3809 	}
3810 	case SIOCSIFCAP:
3811 	{
3812 		int mask, setmask;
3813 
3814 		mask = ifr->ifr_reqcap ^ if_getcapenable(ifp);
3815 		setmask = 0;
3816 #ifdef TCP_OFFLOAD
3817 		setmask |= mask & (IFCAP_TOE4|IFCAP_TOE6);
3818 #endif
3819 		setmask |= (mask & IFCAP_FLAGS);
3820 
3821 		if (setmask  & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6))
3822 			setmask |= (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6);
3823 		if ((mask & IFCAP_WOL) &&
3824 		    (if_getcapabilities(ifp) & IFCAP_WOL) != 0)
3825 			setmask |= (mask & (IFCAP_WOL_MCAST|IFCAP_WOL_MAGIC));
3826 		if_vlancap(ifp);
3827 		/*
3828 		 * want to ensure that traffic has stopped before we change any of the flags
3829 		 */
3830 		if (setmask) {
3831 			CTX_LOCK(ctx);
3832 			bits = if_getdrvflags(ifp);
3833 			if (bits & IFF_DRV_RUNNING)
3834 				iflib_stop(ctx);
3835 			if_togglecapenable(ifp, setmask);
3836 			if (bits & IFF_DRV_RUNNING)
3837 				iflib_init_locked(ctx);
3838 			if_setdrvflags(ifp, bits);
3839 			CTX_UNLOCK(ctx);
3840 		}
3841 		break;
3842 	    }
3843 	case SIOCGPRIVATE_0:
3844 	case SIOCSDRVSPEC:
3845 	case SIOCGDRVSPEC:
3846 		CTX_LOCK(ctx);
3847 		err = IFDI_PRIV_IOCTL(ctx, command, data);
3848 		CTX_UNLOCK(ctx);
3849 		break;
3850 	default:
3851 		err = ether_ioctl(ifp, command, data);
3852 		break;
3853 	}
3854 	if (reinit)
3855 		iflib_if_init(ctx);
3856 	return (err);
3857 }
3858 
3859 static uint64_t
3860 iflib_if_get_counter(if_t ifp, ift_counter cnt)
3861 {
3862 	if_ctx_t ctx = if_getsoftc(ifp);
3863 
3864 	return (IFDI_GET_COUNTER(ctx, cnt));
3865 }
3866 
3867 /*********************************************************************
3868  *
3869  *  OTHER FUNCTIONS EXPORTED TO THE STACK
3870  *
3871  **********************************************************************/
3872 
3873 static void
3874 iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag)
3875 {
3876 	if_ctx_t ctx = if_getsoftc(ifp);
3877 
3878 	if ((void *)ctx != arg)
3879 		return;
3880 
3881 	if ((vtag == 0) || (vtag > 4095))
3882 		return;
3883 
3884 	CTX_LOCK(ctx);
3885 	IFDI_VLAN_REGISTER(ctx, vtag);
3886 	/* Re-init to load the changes */
3887 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
3888 		iflib_init_locked(ctx);
3889 	CTX_UNLOCK(ctx);
3890 }
3891 
3892 static void
3893 iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag)
3894 {
3895 	if_ctx_t ctx = if_getsoftc(ifp);
3896 
3897 	if ((void *)ctx != arg)
3898 		return;
3899 
3900 	if ((vtag == 0) || (vtag > 4095))
3901 		return;
3902 
3903 	CTX_LOCK(ctx);
3904 	IFDI_VLAN_UNREGISTER(ctx, vtag);
3905 	/* Re-init to load the changes */
3906 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
3907 		iflib_init_locked(ctx);
3908 	CTX_UNLOCK(ctx);
3909 }
3910 
3911 static void
3912 iflib_led_func(void *arg, int onoff)
3913 {
3914 	if_ctx_t ctx = arg;
3915 
3916 	CTX_LOCK(ctx);
3917 	IFDI_LED_FUNC(ctx, onoff);
3918 	CTX_UNLOCK(ctx);
3919 }
3920 
3921 /*********************************************************************
3922  *
3923  *  BUS FUNCTION DEFINITIONS
3924  *
3925  **********************************************************************/
3926 
3927 int
3928 iflib_device_probe(device_t dev)
3929 {
3930 	pci_vendor_info_t *ent;
3931 
3932 	uint16_t	pci_vendor_id, pci_device_id;
3933 	uint16_t	pci_subvendor_id, pci_subdevice_id;
3934 	uint16_t	pci_rev_id;
3935 	if_shared_ctx_t sctx;
3936 
3937 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
3938 		return (ENOTSUP);
3939 
3940 	pci_vendor_id = pci_get_vendor(dev);
3941 	pci_device_id = pci_get_device(dev);
3942 	pci_subvendor_id = pci_get_subvendor(dev);
3943 	pci_subdevice_id = pci_get_subdevice(dev);
3944 	pci_rev_id = pci_get_revid(dev);
3945 	if (sctx->isc_parse_devinfo != NULL)
3946 		sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id);
3947 
3948 	ent = sctx->isc_vendor_info;
3949 	while (ent->pvi_vendor_id != 0) {
3950 		if (pci_vendor_id != ent->pvi_vendor_id) {
3951 			ent++;
3952 			continue;
3953 		}
3954 		if ((pci_device_id == ent->pvi_device_id) &&
3955 		    ((pci_subvendor_id == ent->pvi_subvendor_id) ||
3956 		     (ent->pvi_subvendor_id == 0)) &&
3957 		    ((pci_subdevice_id == ent->pvi_subdevice_id) ||
3958 		     (ent->pvi_subdevice_id == 0)) &&
3959 		    ((pci_rev_id == ent->pvi_rev_id) ||
3960 		     (ent->pvi_rev_id == 0))) {
3961 
3962 			device_set_desc_copy(dev, ent->pvi_name);
3963 			/* this needs to be changed to zero if the bus probing code
3964 			 * ever stops re-probing on best match because the sctx
3965 			 * may have its values over written by register calls
3966 			 * in subsequent probes
3967 			 */
3968 			return (BUS_PROBE_DEFAULT);
3969 		}
3970 		ent++;
3971 	}
3972 	return (ENXIO);
3973 }
3974 
3975 int
3976 iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp)
3977 {
3978 	int err, rid, msix, msix_bar;
3979 	if_ctx_t ctx;
3980 	if_t ifp;
3981 	if_softc_ctx_t scctx;
3982 	int i;
3983 	uint16_t main_txq;
3984 	uint16_t main_rxq;
3985 
3986 
3987 	ctx = malloc(sizeof(* ctx), M_IFLIB, M_WAITOK|M_ZERO);
3988 
3989 	if (sc == NULL) {
3990 		sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
3991 		device_set_softc(dev, ctx);
3992 		ctx->ifc_flags |= IFC_SC_ALLOCATED;
3993 	}
3994 
3995 	ctx->ifc_sctx = sctx;
3996 	ctx->ifc_dev = dev;
3997 	ctx->ifc_softc = sc;
3998 
3999 	if ((err = iflib_register(ctx)) != 0) {
4000 		device_printf(dev, "iflib_register failed %d\n", err);
4001 		return (err);
4002 	}
4003 	iflib_add_device_sysctl_pre(ctx);
4004 
4005 	scctx = &ctx->ifc_softc_ctx;
4006 	ifp = ctx->ifc_ifp;
4007 
4008 	/*
4009 	 * XXX sanity check that ntxd & nrxd are a power of 2
4010 	 */
4011 	if (ctx->ifc_sysctl_ntxqs != 0)
4012 		scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs;
4013 	if (ctx->ifc_sysctl_nrxqs != 0)
4014 		scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs;
4015 
4016 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4017 		if (ctx->ifc_sysctl_ntxds[i] != 0)
4018 			scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i];
4019 		else
4020 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4021 	}
4022 
4023 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4024 		if (ctx->ifc_sysctl_nrxds[i] != 0)
4025 			scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i];
4026 		else
4027 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4028 	}
4029 
4030 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4031 		if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) {
4032 			device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n",
4033 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]);
4034 			scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i];
4035 		}
4036 		if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) {
4037 			device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n",
4038 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]);
4039 			scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i];
4040 		}
4041 	}
4042 
4043 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4044 		if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) {
4045 			device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n",
4046 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]);
4047 			scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i];
4048 		}
4049 		if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) {
4050 			device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n",
4051 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]);
4052 			scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i];
4053 		}
4054 	}
4055 
4056 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4057 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4058 		return (err);
4059 	}
4060 	_iflib_pre_assert(scctx);
4061 	ctx->ifc_txrx = *scctx->isc_txrx;
4062 
4063 #ifdef INVARIANTS
4064 	MPASS(scctx->isc_capenable);
4065 	if (scctx->isc_capenable & IFCAP_TXCSUM)
4066 		MPASS(scctx->isc_tx_csum_flags);
4067 #endif
4068 
4069 	if_setcapabilities(ifp, scctx->isc_capenable);
4070 	if_setcapenable(ifp, scctx->isc_capenable);
4071 
4072 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4073 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4074 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4075 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4076 
4077 #ifdef ACPI_DMAR
4078 	if (dmar_get_dma_tag(device_get_parent(dev), dev) != NULL)
4079 		ctx->ifc_flags |= IFC_DMAR;
4080 #elif !(defined(__i386__) || defined(__amd64__))
4081 	/* set unconditionally for !x86 */
4082 	ctx->ifc_flags |= IFC_DMAR;
4083 #endif
4084 
4085 	msix_bar = scctx->isc_msix_bar;
4086 	main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4087 	main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4088 
4089 	/* XXX change for per-queue sizes */
4090 	device_printf(dev, "using %d tx descriptors and %d rx descriptors\n",
4091 		      scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4092 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4093 		if (!powerof2(scctx->isc_nrxd[i])) {
4094 			/* round down instead? */
4095 			device_printf(dev, "# rx descriptors must be a power of 2\n");
4096 			err = EINVAL;
4097 			goto fail;
4098 		}
4099 	}
4100 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4101 		if (!powerof2(scctx->isc_ntxd[i])) {
4102 			device_printf(dev,
4103 			    "# tx descriptors must be a power of 2");
4104 			err = EINVAL;
4105 			goto fail;
4106 		}
4107 	}
4108 
4109 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4110 	    MAX_SINGLE_PACKET_FRACTION)
4111 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4112 		    MAX_SINGLE_PACKET_FRACTION);
4113 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4114 	    MAX_SINGLE_PACKET_FRACTION)
4115 		scctx->isc_tx_tso_segments_max = max(1,
4116 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4117 
4118 	/*
4119 	 * Protect the stack against modern hardware
4120 	 */
4121 	if (scctx->isc_tx_tso_size_max > FREEBSD_TSO_SIZE_MAX)
4122 		scctx->isc_tx_tso_size_max = FREEBSD_TSO_SIZE_MAX;
4123 
4124 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4125 	ifp->if_hw_tsomaxsegcount = scctx->isc_tx_tso_segments_max;
4126 	ifp->if_hw_tsomax = scctx->isc_tx_tso_size_max;
4127 	ifp->if_hw_tsomaxsegsize = scctx->isc_tx_tso_segsize_max;
4128 	if (scctx->isc_rss_table_size == 0)
4129 		scctx->isc_rss_table_size = 64;
4130 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4131 
4132 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4133 	/* XXX format name */
4134 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx, -1, "admin");
4135 	/*
4136 	** Now setup MSI or MSI/X, should
4137 	** return us the number of supported
4138 	** vectors. (Will be 1 for MSI)
4139 	*/
4140 	if (sctx->isc_flags & IFLIB_SKIP_MSIX) {
4141 		msix = scctx->isc_vectors;
4142 	} else if (scctx->isc_msix_bar != 0)
4143 	       /*
4144 		* The simple fact that isc_msix_bar is not 0 does not mean we
4145 		* we have a good value there that is known to work.
4146 		*/
4147 		msix = iflib_msix_init(ctx);
4148 	else {
4149 		scctx->isc_vectors = 1;
4150 		scctx->isc_ntxqsets = 1;
4151 		scctx->isc_nrxqsets = 1;
4152 		scctx->isc_intr = IFLIB_INTR_LEGACY;
4153 		msix = 0;
4154 	}
4155 	/* Get memory for the station queues */
4156 	if ((err = iflib_queues_alloc(ctx))) {
4157 		device_printf(dev, "Unable to allocate queue memory\n");
4158 		goto fail;
4159 	}
4160 
4161 	if ((err = iflib_qset_structures_setup(ctx))) {
4162 		device_printf(dev, "qset structure setup failed %d\n", err);
4163 		goto fail_queues;
4164 	}
4165 
4166 	/*
4167 	 * Group taskqueues aren't properly set up until SMP is started,
4168 	 * so we disable interrupts until we can handle them post
4169 	 * SI_SUB_SMP.
4170 	 *
4171 	 * XXX: disabling interrupts doesn't actually work, at least for
4172 	 * the non-MSI case.  When they occur before SI_SUB_SMP completes,
4173 	 * we do null handling and depend on this not causing too large an
4174 	 * interrupt storm.
4175 	 */
4176 	IFDI_INTR_DISABLE(ctx);
4177 	if (msix > 1 && (err = IFDI_MSIX_INTR_ASSIGN(ctx, msix)) != 0) {
4178 		device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n", err);
4179 		goto fail_intr_free;
4180 	}
4181 	if (msix <= 1) {
4182 		rid = 0;
4183 		if (scctx->isc_intr == IFLIB_INTR_MSI) {
4184 			MPASS(msix == 1);
4185 			rid = 1;
4186 		}
4187 		if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) {
4188 			device_printf(dev, "iflib_legacy_setup failed %d\n", err);
4189 			goto fail_intr_free;
4190 		}
4191 	}
4192 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac);
4193 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4194 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4195 		goto fail_detach;
4196 	}
4197 	if ((err = iflib_netmap_attach(ctx))) {
4198 		device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err);
4199 		goto fail_detach;
4200 	}
4201 	*ctxp = ctx;
4202 
4203 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4204 	iflib_add_device_sysctl_post(ctx);
4205 	ctx->ifc_flags |= IFC_INIT_DONE;
4206 	return (0);
4207 fail_detach:
4208 	ether_ifdetach(ctx->ifc_ifp);
4209 fail_intr_free:
4210 	if (scctx->isc_intr == IFLIB_INTR_MSIX || scctx->isc_intr == IFLIB_INTR_MSI)
4211 		pci_release_msi(ctx->ifc_dev);
4212 fail_queues:
4213 	/* XXX free queues */
4214 fail:
4215 	IFDI_DETACH(ctx);
4216 	return (err);
4217 }
4218 
4219 int
4220 iflib_device_attach(device_t dev)
4221 {
4222 	if_ctx_t ctx;
4223 	if_shared_ctx_t sctx;
4224 
4225 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4226 		return (ENOTSUP);
4227 
4228 	pci_enable_busmaster(dev);
4229 
4230 	return (iflib_device_register(dev, NULL, sctx, &ctx));
4231 }
4232 
4233 int
4234 iflib_device_deregister(if_ctx_t ctx)
4235 {
4236 	if_t ifp = ctx->ifc_ifp;
4237 	iflib_txq_t txq;
4238 	iflib_rxq_t rxq;
4239 	device_t dev = ctx->ifc_dev;
4240 	int i;
4241 	struct taskqgroup *tqg;
4242 
4243 	/* Make sure VLANS are not using driver */
4244 	if (if_vlantrunkinuse(ifp)) {
4245 		device_printf(dev,"Vlan in use, detach first\n");
4246 		return (EBUSY);
4247 	}
4248 
4249 	CTX_LOCK(ctx);
4250 	ctx->ifc_in_detach = 1;
4251 	iflib_stop(ctx);
4252 	CTX_UNLOCK(ctx);
4253 
4254 	/* Unregister VLAN events */
4255 	if (ctx->ifc_vlan_attach_event != NULL)
4256 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
4257 	if (ctx->ifc_vlan_detach_event != NULL)
4258 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
4259 
4260 	iflib_netmap_detach(ifp);
4261 	ether_ifdetach(ifp);
4262 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
4263 	CTX_LOCK_DESTROY(ctx);
4264 	if (ctx->ifc_led_dev != NULL)
4265 		led_destroy(ctx->ifc_led_dev);
4266 	/* XXX drain any dependent tasks */
4267 	tqg = qgroup_if_io_tqg;
4268 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
4269 		callout_drain(&txq->ift_timer);
4270 		if (txq->ift_task.gt_uniq != NULL)
4271 			taskqgroup_detach(tqg, &txq->ift_task);
4272 	}
4273 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4274 		if (rxq->ifr_task.gt_uniq != NULL)
4275 			taskqgroup_detach(tqg, &rxq->ifr_task);
4276 	}
4277 	tqg = qgroup_if_config_tqg;
4278 	if (ctx->ifc_admin_task.gt_uniq != NULL)
4279 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
4280 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
4281 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
4282 
4283 	IFDI_DETACH(ctx);
4284 	device_set_softc(ctx->ifc_dev, NULL);
4285 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) {
4286 		pci_release_msi(dev);
4287 	}
4288 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) {
4289 		iflib_irq_free(ctx, &ctx->ifc_legacy_irq);
4290 	}
4291 	if (ctx->ifc_msix_mem != NULL) {
4292 		bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY,
4293 			ctx->ifc_softc_ctx.isc_msix_bar, ctx->ifc_msix_mem);
4294 		ctx->ifc_msix_mem = NULL;
4295 	}
4296 
4297 	bus_generic_detach(dev);
4298 	if_free(ifp);
4299 
4300 	iflib_tx_structures_free(ctx);
4301 	iflib_rx_structures_free(ctx);
4302 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4303 		free(ctx->ifc_softc, M_IFLIB);
4304 	free(ctx, M_IFLIB);
4305 	return (0);
4306 }
4307 
4308 
4309 int
4310 iflib_device_detach(device_t dev)
4311 {
4312 	if_ctx_t ctx = device_get_softc(dev);
4313 
4314 	return (iflib_device_deregister(ctx));
4315 }
4316 
4317 int
4318 iflib_device_suspend(device_t dev)
4319 {
4320 	if_ctx_t ctx = device_get_softc(dev);
4321 
4322 	CTX_LOCK(ctx);
4323 	IFDI_SUSPEND(ctx);
4324 	CTX_UNLOCK(ctx);
4325 
4326 	return bus_generic_suspend(dev);
4327 }
4328 int
4329 iflib_device_shutdown(device_t dev)
4330 {
4331 	if_ctx_t ctx = device_get_softc(dev);
4332 
4333 	CTX_LOCK(ctx);
4334 	IFDI_SHUTDOWN(ctx);
4335 	CTX_UNLOCK(ctx);
4336 
4337 	return bus_generic_suspend(dev);
4338 }
4339 
4340 
4341 int
4342 iflib_device_resume(device_t dev)
4343 {
4344 	if_ctx_t ctx = device_get_softc(dev);
4345 	iflib_txq_t txq = ctx->ifc_txqs;
4346 
4347 	CTX_LOCK(ctx);
4348 	IFDI_RESUME(ctx);
4349 	iflib_init_locked(ctx);
4350 	CTX_UNLOCK(ctx);
4351 	for (int i = 0; i < NTXQSETS(ctx); i++, txq++)
4352 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
4353 
4354 	return (bus_generic_resume(dev));
4355 }
4356 
4357 int
4358 iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params)
4359 {
4360 	int error;
4361 	if_ctx_t ctx = device_get_softc(dev);
4362 
4363 	CTX_LOCK(ctx);
4364 	error = IFDI_IOV_INIT(ctx, num_vfs, params);
4365 	CTX_UNLOCK(ctx);
4366 
4367 	return (error);
4368 }
4369 
4370 void
4371 iflib_device_iov_uninit(device_t dev)
4372 {
4373 	if_ctx_t ctx = device_get_softc(dev);
4374 
4375 	CTX_LOCK(ctx);
4376 	IFDI_IOV_UNINIT(ctx);
4377 	CTX_UNLOCK(ctx);
4378 }
4379 
4380 int
4381 iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params)
4382 {
4383 	int error;
4384 	if_ctx_t ctx = device_get_softc(dev);
4385 
4386 	CTX_LOCK(ctx);
4387 	error = IFDI_IOV_VF_ADD(ctx, vfnum, params);
4388 	CTX_UNLOCK(ctx);
4389 
4390 	return (error);
4391 }
4392 
4393 /*********************************************************************
4394  *
4395  *  MODULE FUNCTION DEFINITIONS
4396  *
4397  **********************************************************************/
4398 
4399 /*
4400  * - Start a fast taskqueue thread for each core
4401  * - Start a taskqueue for control operations
4402  */
4403 static int
4404 iflib_module_init(void)
4405 {
4406 	return (0);
4407 }
4408 
4409 static int
4410 iflib_module_event_handler(module_t mod, int what, void *arg)
4411 {
4412 	int err;
4413 
4414 	switch (what) {
4415 	case MOD_LOAD:
4416 		if ((err = iflib_module_init()) != 0)
4417 			return (err);
4418 		break;
4419 	case MOD_UNLOAD:
4420 		return (EBUSY);
4421 	default:
4422 		return (EOPNOTSUPP);
4423 	}
4424 
4425 	return (0);
4426 }
4427 
4428 /*********************************************************************
4429  *
4430  *  PUBLIC FUNCTION DEFINITIONS
4431  *     ordered as in iflib.h
4432  *
4433  **********************************************************************/
4434 
4435 
4436 static void
4437 _iflib_assert(if_shared_ctx_t sctx)
4438 {
4439 	MPASS(sctx->isc_tx_maxsize);
4440 	MPASS(sctx->isc_tx_maxsegsize);
4441 
4442 	MPASS(sctx->isc_rx_maxsize);
4443 	MPASS(sctx->isc_rx_nsegments);
4444 	MPASS(sctx->isc_rx_maxsegsize);
4445 
4446 	MPASS(sctx->isc_nrxd_min[0]);
4447 	MPASS(sctx->isc_nrxd_max[0]);
4448 	MPASS(sctx->isc_nrxd_default[0]);
4449 	MPASS(sctx->isc_ntxd_min[0]);
4450 	MPASS(sctx->isc_ntxd_max[0]);
4451 	MPASS(sctx->isc_ntxd_default[0]);
4452 }
4453 
4454 static void
4455 _iflib_pre_assert(if_softc_ctx_t scctx)
4456 {
4457 
4458 	MPASS(scctx->isc_txrx->ift_txd_encap);
4459 	MPASS(scctx->isc_txrx->ift_txd_flush);
4460 	MPASS(scctx->isc_txrx->ift_txd_credits_update);
4461 	MPASS(scctx->isc_txrx->ift_rxd_available);
4462 	MPASS(scctx->isc_txrx->ift_rxd_pkt_get);
4463 	MPASS(scctx->isc_txrx->ift_rxd_refill);
4464 	MPASS(scctx->isc_txrx->ift_rxd_flush);
4465 }
4466 
4467 static int
4468 iflib_register(if_ctx_t ctx)
4469 {
4470 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4471 	driver_t *driver = sctx->isc_driver;
4472 	device_t dev = ctx->ifc_dev;
4473 	if_t ifp;
4474 
4475 	_iflib_assert(sctx);
4476 
4477 	CTX_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev));
4478 
4479 	ifp = ctx->ifc_ifp = if_gethandle(IFT_ETHER);
4480 	if (ifp == NULL) {
4481 		device_printf(dev, "can not allocate ifnet structure\n");
4482 		return (ENOMEM);
4483 	}
4484 
4485 	/*
4486 	 * Initialize our context's device specific methods
4487 	 */
4488 	kobj_init((kobj_t) ctx, (kobj_class_t) driver);
4489 	kobj_class_compile((kobj_class_t) driver);
4490 	driver->refs++;
4491 
4492 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
4493 	if_setsoftc(ifp, ctx);
4494 	if_setdev(ifp, dev);
4495 	if_setinitfn(ifp, iflib_if_init);
4496 	if_setioctlfn(ifp, iflib_if_ioctl);
4497 	if_settransmitfn(ifp, iflib_if_transmit);
4498 	if_setqflushfn(ifp, iflib_if_qflush);
4499 	if_setflags(ifp, IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST);
4500 
4501 	ctx->ifc_vlan_attach_event =
4502 		EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx,
4503 							  EVENTHANDLER_PRI_FIRST);
4504 	ctx->ifc_vlan_detach_event =
4505 		EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx,
4506 							  EVENTHANDLER_PRI_FIRST);
4507 
4508 	ifmedia_init(&ctx->ifc_media, IFM_IMASK,
4509 					 iflib_media_change, iflib_media_status);
4510 
4511 	return (0);
4512 }
4513 
4514 
4515 static int
4516 iflib_queues_alloc(if_ctx_t ctx)
4517 {
4518 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4519 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4520 	device_t dev = ctx->ifc_dev;
4521 	int nrxqsets = scctx->isc_nrxqsets;
4522 	int ntxqsets = scctx->isc_ntxqsets;
4523 	iflib_txq_t txq;
4524 	iflib_rxq_t rxq;
4525 	iflib_fl_t fl = NULL;
4526 	int i, j, cpu, err, txconf, rxconf;
4527 	iflib_dma_info_t ifdip;
4528 	uint32_t *rxqsizes = scctx->isc_rxqsizes;
4529 	uint32_t *txqsizes = scctx->isc_txqsizes;
4530 	uint8_t nrxqs = sctx->isc_nrxqs;
4531 	uint8_t ntxqs = sctx->isc_ntxqs;
4532 	int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1;
4533 	caddr_t *vaddrs;
4534 	uint64_t *paddrs;
4535 	struct ifmp_ring **brscp;
4536 
4537 	KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1"));
4538 	KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1"));
4539 
4540 	brscp = NULL;
4541 	txq = NULL;
4542 	rxq = NULL;
4543 
4544 /* Allocate the TX ring struct memory */
4545 	if (!(txq =
4546 	    (iflib_txq_t) malloc(sizeof(struct iflib_txq) *
4547 	    ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
4548 		device_printf(dev, "Unable to allocate TX ring memory\n");
4549 		err = ENOMEM;
4550 		goto fail;
4551 	}
4552 
4553 	/* Now allocate the RX */
4554 	if (!(rxq =
4555 	    (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) *
4556 	    nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
4557 		device_printf(dev, "Unable to allocate RX ring memory\n");
4558 		err = ENOMEM;
4559 		goto rx_fail;
4560 	}
4561 
4562 	ctx->ifc_txqs = txq;
4563 	ctx->ifc_rxqs = rxq;
4564 
4565 	/*
4566 	 * XXX handle allocation failure
4567 	 */
4568 	for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) {
4569 		/* Set up some basics */
4570 
4571 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs, M_IFLIB, M_WAITOK|M_ZERO)) == NULL) {
4572 			device_printf(dev, "failed to allocate iflib_dma_info\n");
4573 			err = ENOMEM;
4574 			goto err_tx_desc;
4575 		}
4576 		txq->ift_ifdi = ifdip;
4577 		for (j = 0; j < ntxqs; j++, ifdip++) {
4578 			if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, BUS_DMA_NOWAIT)) {
4579 				device_printf(dev, "Unable to allocate Descriptor memory\n");
4580 				err = ENOMEM;
4581 				goto err_tx_desc;
4582 			}
4583 			txq->ift_txd_size[j] = scctx->isc_txd_size[j];
4584 			bzero((void *)ifdip->idi_vaddr, txqsizes[j]);
4585 		}
4586 		txq->ift_ctx = ctx;
4587 		txq->ift_id = i;
4588 		if (sctx->isc_flags & IFLIB_HAS_TXCQ) {
4589 			txq->ift_br_offset = 1;
4590 		} else {
4591 			txq->ift_br_offset = 0;
4592 		}
4593 		/* XXX fix this */
4594 		txq->ift_timer.c_cpu = cpu;
4595 
4596 		if (iflib_txsd_alloc(txq)) {
4597 			device_printf(dev, "Critical Failure setting up TX buffers\n");
4598 			err = ENOMEM;
4599 			goto err_tx_desc;
4600 		}
4601 
4602 		/* Initialize the TX lock */
4603 		snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:tx(%d):callout",
4604 		    device_get_nameunit(dev), txq->ift_id);
4605 		mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF);
4606 		callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0);
4607 
4608 		snprintf(txq->ift_db_mtx_name, MTX_NAME_LEN, "%s:tx(%d):db",
4609 			 device_get_nameunit(dev), txq->ift_id);
4610 
4611 		err = ifmp_ring_alloc(&txq->ift_br, 2048, txq, iflib_txq_drain,
4612 				      iflib_txq_can_drain, M_IFLIB, M_WAITOK);
4613 		if (err) {
4614 			/* XXX free any allocated rings */
4615 			device_printf(dev, "Unable to allocate buf_ring\n");
4616 			goto err_tx_desc;
4617 		}
4618 	}
4619 
4620 	for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) {
4621 		/* Set up some basics */
4622 
4623 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs, M_IFLIB, M_WAITOK|M_ZERO)) == NULL) {
4624 			device_printf(dev, "failed to allocate iflib_dma_info\n");
4625 			err = ENOMEM;
4626 			goto err_tx_desc;
4627 		}
4628 
4629 		rxq->ifr_ifdi = ifdip;
4630 		/* XXX this needs to be changed if #rx queues != #tx queues */
4631 		rxq->ifr_ntxqirq = 1;
4632 		rxq->ifr_txqid[0] = i;
4633 		for (j = 0; j < nrxqs; j++, ifdip++) {
4634 			if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, BUS_DMA_NOWAIT)) {
4635 				device_printf(dev, "Unable to allocate Descriptor memory\n");
4636 				err = ENOMEM;
4637 				goto err_tx_desc;
4638 			}
4639 			bzero((void *)ifdip->idi_vaddr, rxqsizes[j]);
4640 		}
4641 		rxq->ifr_ctx = ctx;
4642 		rxq->ifr_id = i;
4643 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
4644 			rxq->ifr_fl_offset = 1;
4645 		} else {
4646 			rxq->ifr_fl_offset = 0;
4647 		}
4648 		rxq->ifr_nfl = nfree_lists;
4649 		if (!(fl =
4650 			  (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) {
4651 			device_printf(dev, "Unable to allocate free list memory\n");
4652 			err = ENOMEM;
4653 			goto err_tx_desc;
4654 		}
4655 		rxq->ifr_fl = fl;
4656 		for (j = 0; j < nfree_lists; j++) {
4657 			fl[j].ifl_rxq = rxq;
4658 			fl[j].ifl_id = j;
4659 			fl[j].ifl_ifdi = &rxq->ifr_ifdi[j + rxq->ifr_fl_offset];
4660 			fl[j].ifl_rxd_size = scctx->isc_rxd_size[j];
4661 		}
4662         /* Allocate receive buffers for the ring*/
4663 		if (iflib_rxsd_alloc(rxq)) {
4664 			device_printf(dev,
4665 			    "Critical Failure setting up receive buffers\n");
4666 			err = ENOMEM;
4667 			goto err_rx_desc;
4668 		}
4669 	}
4670 
4671 	/* TXQs */
4672 	vaddrs = malloc(sizeof(caddr_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
4673 	paddrs = malloc(sizeof(uint64_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
4674 	for (i = 0; i < ntxqsets; i++) {
4675 		iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi;
4676 
4677 		for (j = 0; j < ntxqs; j++, di++) {
4678 			vaddrs[i*ntxqs + j] = di->idi_vaddr;
4679 			paddrs[i*ntxqs + j] = di->idi_paddr;
4680 		}
4681 	}
4682 	if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) {
4683 		device_printf(ctx->ifc_dev, "device queue allocation failed\n");
4684 		iflib_tx_structures_free(ctx);
4685 		free(vaddrs, M_IFLIB);
4686 		free(paddrs, M_IFLIB);
4687 		goto err_rx_desc;
4688 	}
4689 	free(vaddrs, M_IFLIB);
4690 	free(paddrs, M_IFLIB);
4691 
4692 	/* RXQs */
4693 	vaddrs = malloc(sizeof(caddr_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
4694 	paddrs = malloc(sizeof(uint64_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
4695 	for (i = 0; i < nrxqsets; i++) {
4696 		iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi;
4697 
4698 		for (j = 0; j < nrxqs; j++, di++) {
4699 			vaddrs[i*nrxqs + j] = di->idi_vaddr;
4700 			paddrs[i*nrxqs + j] = di->idi_paddr;
4701 		}
4702 	}
4703 	if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) {
4704 		device_printf(ctx->ifc_dev, "device queue allocation failed\n");
4705 		iflib_tx_structures_free(ctx);
4706 		free(vaddrs, M_IFLIB);
4707 		free(paddrs, M_IFLIB);
4708 		goto err_rx_desc;
4709 	}
4710 	free(vaddrs, M_IFLIB);
4711 	free(paddrs, M_IFLIB);
4712 
4713 	return (0);
4714 
4715 /* XXX handle allocation failure changes */
4716 err_rx_desc:
4717 err_tx_desc:
4718 	if (ctx->ifc_rxqs != NULL)
4719 		free(ctx->ifc_rxqs, M_IFLIB);
4720 	ctx->ifc_rxqs = NULL;
4721 	if (ctx->ifc_txqs != NULL)
4722 		free(ctx->ifc_txqs, M_IFLIB);
4723 	ctx->ifc_txqs = NULL;
4724 rx_fail:
4725 	if (brscp != NULL)
4726 		free(brscp, M_IFLIB);
4727 	if (rxq != NULL)
4728 		free(rxq, M_IFLIB);
4729 	if (txq != NULL)
4730 		free(txq, M_IFLIB);
4731 fail:
4732 	return (err);
4733 }
4734 
4735 static int
4736 iflib_tx_structures_setup(if_ctx_t ctx)
4737 {
4738 	iflib_txq_t txq = ctx->ifc_txqs;
4739 	int i;
4740 
4741 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
4742 		iflib_txq_setup(txq);
4743 
4744 	return (0);
4745 }
4746 
4747 static void
4748 iflib_tx_structures_free(if_ctx_t ctx)
4749 {
4750 	iflib_txq_t txq = ctx->ifc_txqs;
4751 	int i, j;
4752 
4753 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
4754 		iflib_txq_destroy(txq);
4755 		for (j = 0; j < ctx->ifc_nhwtxqs; j++)
4756 			iflib_dma_free(&txq->ift_ifdi[j]);
4757 	}
4758 	free(ctx->ifc_txqs, M_IFLIB);
4759 	ctx->ifc_txqs = NULL;
4760 	IFDI_QUEUES_FREE(ctx);
4761 }
4762 
4763 /*********************************************************************
4764  *
4765  *  Initialize all receive rings.
4766  *
4767  **********************************************************************/
4768 static int
4769 iflib_rx_structures_setup(if_ctx_t ctx)
4770 {
4771 	iflib_rxq_t rxq = ctx->ifc_rxqs;
4772 	int q;
4773 #if defined(INET6) || defined(INET)
4774 	int i, err;
4775 #endif
4776 
4777 	for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) {
4778 #if defined(INET6) || defined(INET)
4779 		tcp_lro_free(&rxq->ifr_lc);
4780 		if ((err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp,
4781 		    TCP_LRO_ENTRIES, min(1024,
4782 		    ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset]))) != 0) {
4783 			device_printf(ctx->ifc_dev, "LRO Initialization failed!\n");
4784 			goto fail;
4785 		}
4786 		rxq->ifr_lro_enabled = TRUE;
4787 #endif
4788 		IFDI_RXQ_SETUP(ctx, rxq->ifr_id);
4789 	}
4790 	return (0);
4791 #if defined(INET6) || defined(INET)
4792 fail:
4793 	/*
4794 	 * Free RX software descriptors allocated so far, we will only handle
4795 	 * the rings that completed, the failing case will have
4796 	 * cleaned up for itself. 'q' failed, so its the terminus.
4797 	 */
4798 	rxq = ctx->ifc_rxqs;
4799 	for (i = 0; i < q; ++i, rxq++) {
4800 		iflib_rx_sds_free(rxq);
4801 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
4802 	}
4803 	return (err);
4804 #endif
4805 }
4806 
4807 /*********************************************************************
4808  *
4809  *  Free all receive rings.
4810  *
4811  **********************************************************************/
4812 static void
4813 iflib_rx_structures_free(if_ctx_t ctx)
4814 {
4815 	iflib_rxq_t rxq = ctx->ifc_rxqs;
4816 
4817 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) {
4818 		iflib_rx_sds_free(rxq);
4819 	}
4820 }
4821 
4822 static int
4823 iflib_qset_structures_setup(if_ctx_t ctx)
4824 {
4825 	int err;
4826 
4827 	if ((err = iflib_tx_structures_setup(ctx)) != 0)
4828 		return (err);
4829 
4830 	if ((err = iflib_rx_structures_setup(ctx)) != 0) {
4831 		device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err);
4832 		iflib_tx_structures_free(ctx);
4833 		iflib_rx_structures_free(ctx);
4834 	}
4835 	return (err);
4836 }
4837 
4838 int
4839 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
4840 				driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, char *name)
4841 {
4842 
4843 	return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name));
4844 }
4845 
4846 static int
4847 find_nth(if_ctx_t ctx, cpuset_t *cpus, int qid)
4848 {
4849 	int i, cpuid, eqid, count;
4850 
4851 	CPU_COPY(&ctx->ifc_cpus, cpus);
4852 	count = CPU_COUNT(&ctx->ifc_cpus);
4853 	eqid = qid % count;
4854 	/* clear up to the qid'th bit */
4855 	for (i = 0; i < eqid; i++) {
4856 		cpuid = CPU_FFS(cpus);
4857 		MPASS(cpuid != 0);
4858 		CPU_CLR(cpuid-1, cpus);
4859 	}
4860 	cpuid = CPU_FFS(cpus);
4861 	MPASS(cpuid != 0);
4862 	return (cpuid-1);
4863 }
4864 
4865 int
4866 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid,
4867 						iflib_intr_type_t type, driver_filter_t *filter,
4868 						void *filter_arg, int qid, char *name)
4869 {
4870 	struct grouptask *gtask;
4871 	struct taskqgroup *tqg;
4872 	iflib_filter_info_t info;
4873 	cpuset_t cpus;
4874 	gtask_fn_t *fn;
4875 	int tqrid, err, cpuid;
4876 	driver_filter_t *intr_fast;
4877 	void *q;
4878 
4879 	info = &ctx->ifc_filter_info;
4880 	tqrid = rid;
4881 
4882 	switch (type) {
4883 	/* XXX merge tx/rx for netmap? */
4884 	case IFLIB_INTR_TX:
4885 		q = &ctx->ifc_txqs[qid];
4886 		info = &ctx->ifc_txqs[qid].ift_filter_info;
4887 		gtask = &ctx->ifc_txqs[qid].ift_task;
4888 		tqg = qgroup_if_io_tqg;
4889 		fn = _task_fn_tx;
4890 		intr_fast = iflib_fast_intr;
4891 		GROUPTASK_INIT(gtask, 0, fn, q);
4892 		break;
4893 	case IFLIB_INTR_RX:
4894 		q = &ctx->ifc_rxqs[qid];
4895 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
4896 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
4897 		tqg = qgroup_if_io_tqg;
4898 		fn = _task_fn_rx;
4899 		intr_fast = iflib_fast_intr;
4900 		GROUPTASK_INIT(gtask, 0, fn, q);
4901 		break;
4902 	case IFLIB_INTR_RXTX:
4903 		q = &ctx->ifc_rxqs[qid];
4904 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
4905 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
4906 		tqg = qgroup_if_io_tqg;
4907 		fn = _task_fn_rx;
4908 		intr_fast = iflib_fast_intr_rxtx;
4909 		GROUPTASK_INIT(gtask, 0, fn, q);
4910 		break;
4911 	case IFLIB_INTR_ADMIN:
4912 		q = ctx;
4913 		tqrid = -1;
4914 		info = &ctx->ifc_filter_info;
4915 		gtask = &ctx->ifc_admin_task;
4916 		tqg = qgroup_if_config_tqg;
4917 		fn = _task_fn_admin;
4918 		intr_fast = iflib_fast_intr_ctx;
4919 		break;
4920 	default:
4921 		panic("unknown net intr type");
4922 	}
4923 
4924 	info->ifi_filter = filter;
4925 	info->ifi_filter_arg = filter_arg;
4926 	info->ifi_task = gtask;
4927 	info->ifi_ctx = q;
4928 
4929 	err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info,  name);
4930 	if (err != 0) {
4931 		device_printf(ctx->ifc_dev, "_iflib_irq_alloc failed %d\n", err);
4932 		return (err);
4933 	}
4934 	if (type == IFLIB_INTR_ADMIN)
4935 		return (0);
4936 
4937 	if (tqrid != -1) {
4938 		cpuid = find_nth(ctx, &cpus, qid);
4939 		taskqgroup_attach_cpu(tqg, gtask, q, cpuid, irq->ii_rid, name);
4940 	} else {
4941 		taskqgroup_attach(tqg, gtask, q, tqrid, name);
4942 	}
4943 
4944 	return (0);
4945 }
4946 
4947 void
4948 iflib_softirq_alloc_generic(if_ctx_t ctx, int rid, iflib_intr_type_t type,  void *arg, int qid, char *name)
4949 {
4950 	struct grouptask *gtask;
4951 	struct taskqgroup *tqg;
4952 	gtask_fn_t *fn;
4953 	void *q;
4954 
4955 	switch (type) {
4956 	case IFLIB_INTR_TX:
4957 		q = &ctx->ifc_txqs[qid];
4958 		gtask = &ctx->ifc_txqs[qid].ift_task;
4959 		tqg = qgroup_if_io_tqg;
4960 		fn = _task_fn_tx;
4961 		break;
4962 	case IFLIB_INTR_RX:
4963 		q = &ctx->ifc_rxqs[qid];
4964 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
4965 		tqg = qgroup_if_io_tqg;
4966 		fn = _task_fn_rx;
4967 		break;
4968 	case IFLIB_INTR_IOV:
4969 		q = ctx;
4970 		gtask = &ctx->ifc_vflr_task;
4971 		tqg = qgroup_if_config_tqg;
4972 		rid = -1;
4973 		fn = _task_fn_iov;
4974 		break;
4975 	default:
4976 		panic("unknown net intr type");
4977 	}
4978 	GROUPTASK_INIT(gtask, 0, fn, q);
4979 	taskqgroup_attach(tqg, gtask, q, rid, name);
4980 }
4981 
4982 void
4983 iflib_irq_free(if_ctx_t ctx, if_irq_t irq)
4984 {
4985 	if (irq->ii_tag)
4986 		bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag);
4987 
4988 	if (irq->ii_res)
4989 		bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ, irq->ii_rid, irq->ii_res);
4990 }
4991 
4992 static int
4993 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, char *name)
4994 {
4995 	iflib_txq_t txq = ctx->ifc_txqs;
4996 	iflib_rxq_t rxq = ctx->ifc_rxqs;
4997 	if_irq_t irq = &ctx->ifc_legacy_irq;
4998 	iflib_filter_info_t info;
4999 	struct grouptask *gtask;
5000 	struct taskqgroup *tqg;
5001 	gtask_fn_t *fn;
5002 	int tqrid;
5003 	void *q;
5004 	int err;
5005 
5006 	q = &ctx->ifc_rxqs[0];
5007 	info = &rxq[0].ifr_filter_info;
5008 	gtask = &rxq[0].ifr_task;
5009 	tqg = qgroup_if_io_tqg;
5010 	tqrid = irq->ii_rid = *rid;
5011 	fn = _task_fn_rx;
5012 
5013 	ctx->ifc_flags |= IFC_LEGACY;
5014 	info->ifi_filter = filter;
5015 	info->ifi_filter_arg = filter_arg;
5016 	info->ifi_task = gtask;
5017 	info->ifi_ctx = ctx;
5018 
5019 	/* We allocate a single interrupt resource */
5020 	if ((err = _iflib_irq_alloc(ctx, irq, tqrid, iflib_fast_intr_ctx, NULL, info, name)) != 0)
5021 		return (err);
5022 	GROUPTASK_INIT(gtask, 0, fn, q);
5023 	taskqgroup_attach(tqg, gtask, q, tqrid, name);
5024 
5025 	GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq);
5026 	taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, tqrid, "tx");
5027 	return (0);
5028 }
5029 
5030 void
5031 iflib_led_create(if_ctx_t ctx)
5032 {
5033 
5034 	ctx->ifc_led_dev = led_create(iflib_led_func, ctx,
5035 	    device_get_nameunit(ctx->ifc_dev));
5036 }
5037 
5038 void
5039 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid)
5040 {
5041 
5042 	GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
5043 }
5044 
5045 void
5046 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid)
5047 {
5048 
5049 	GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task);
5050 }
5051 
5052 void
5053 iflib_admin_intr_deferred(if_ctx_t ctx)
5054 {
5055 #ifdef INVARIANTS
5056 	struct grouptask *gtask;
5057 
5058 	gtask = &ctx->ifc_admin_task;
5059 	MPASS(gtask->gt_taskqueue != NULL);
5060 #endif
5061 
5062 	GROUPTASK_ENQUEUE(&ctx->ifc_admin_task);
5063 }
5064 
5065 void
5066 iflib_iov_intr_deferred(if_ctx_t ctx)
5067 {
5068 
5069 	GROUPTASK_ENQUEUE(&ctx->ifc_vflr_task);
5070 }
5071 
5072 void
5073 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, char *name)
5074 {
5075 
5076 	taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, -1, name);
5077 }
5078 
5079 void
5080 iflib_config_gtask_init(if_ctx_t ctx, struct grouptask *gtask, gtask_fn_t *fn,
5081 	char *name)
5082 {
5083 
5084 	GROUPTASK_INIT(gtask, 0, fn, ctx);
5085 	taskqgroup_attach(qgroup_if_config_tqg, gtask, gtask, -1, name);
5086 }
5087 
5088 void
5089 iflib_config_gtask_deinit(struct grouptask *gtask)
5090 {
5091 
5092 	taskqgroup_detach(qgroup_if_config_tqg, gtask);
5093 }
5094 
5095 void
5096 iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate)
5097 {
5098 	if_t ifp = ctx->ifc_ifp;
5099 	iflib_txq_t txq = ctx->ifc_txqs;
5100 
5101 	if_setbaudrate(ifp, baudrate);
5102 	if (baudrate >= IF_Gbps(10))
5103 		ctx->ifc_flags |= IFC_PREFETCH;
5104 
5105 	/* If link down, disable watchdog */
5106 	if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) {
5107 		for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++)
5108 			txq->ift_qstatus = IFLIB_QUEUE_IDLE;
5109 	}
5110 	ctx->ifc_link_state = link_state;
5111 	if_link_state_change(ifp, link_state);
5112 }
5113 
5114 static int
5115 iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq)
5116 {
5117 	int credits;
5118 #ifdef INVARIANTS
5119 	int credits_pre = txq->ift_cidx_processed;
5120 #endif
5121 
5122 	if (ctx->isc_txd_credits_update == NULL)
5123 		return (0);
5124 
5125 	if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, true)) == 0)
5126 		return (0);
5127 
5128 	txq->ift_processed += credits;
5129 	txq->ift_cidx_processed += credits;
5130 
5131 	MPASS(credits_pre + credits == txq->ift_cidx_processed);
5132 	if (txq->ift_cidx_processed >= txq->ift_size)
5133 		txq->ift_cidx_processed -= txq->ift_size;
5134 	return (credits);
5135 }
5136 
5137 static int
5138 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget)
5139 {
5140 
5141 	return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx,
5142 	    budget));
5143 }
5144 
5145 void
5146 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name,
5147 	const char *description, if_int_delay_info_t info,
5148 	int offset, int value)
5149 {
5150 	info->iidi_ctx = ctx;
5151 	info->iidi_offset = offset;
5152 	info->iidi_value = value;
5153 	SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev),
5154 	    SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)),
5155 	    OID_AUTO, name, CTLTYPE_INT|CTLFLAG_RW,
5156 	    info, 0, iflib_sysctl_int_delay, "I", description);
5157 }
5158 
5159 struct mtx *
5160 iflib_ctx_lock_get(if_ctx_t ctx)
5161 {
5162 
5163 	return (&ctx->ifc_mtx);
5164 }
5165 
5166 static int
5167 iflib_msix_init(if_ctx_t ctx)
5168 {
5169 	device_t dev = ctx->ifc_dev;
5170 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5171 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5172 	int vectors, queues, rx_queues, tx_queues, queuemsgs, msgs;
5173 	int iflib_num_tx_queues, iflib_num_rx_queues;
5174 	int err, admincnt, bar;
5175 
5176 	iflib_num_tx_queues = scctx->isc_ntxqsets;
5177 	iflib_num_rx_queues = scctx->isc_nrxqsets;
5178 
5179 	device_printf(dev, "msix_init qsets capped at %d\n", iflib_num_tx_queues);
5180 
5181 	bar = ctx->ifc_softc_ctx.isc_msix_bar;
5182 	admincnt = sctx->isc_admin_intrcnt;
5183 	/* Override by tuneable */
5184 	if (enable_msix == 0)
5185 		goto msi;
5186 
5187 	/*
5188 	** When used in a virtualized environment
5189 	** PCI BUSMASTER capability may not be set
5190 	** so explicity set it here and rewrite
5191 	** the ENABLE in the MSIX control register
5192 	** at this point to cause the host to
5193 	** successfully initialize us.
5194 	*/
5195 	{
5196 		int msix_ctrl, rid;
5197 
5198  		pci_enable_busmaster(dev);
5199 		rid = 0;
5200 		if (pci_find_cap(dev, PCIY_MSIX, &rid) == 0 && rid != 0) {
5201 			rid += PCIR_MSIX_CTRL;
5202 			msix_ctrl = pci_read_config(dev, rid, 2);
5203 			msix_ctrl |= PCIM_MSIXCTRL_MSIX_ENABLE;
5204 			pci_write_config(dev, rid, msix_ctrl, 2);
5205 		} else {
5206 			device_printf(dev, "PCIY_MSIX capability not found; "
5207 			                   "or rid %d == 0.\n", rid);
5208 			goto msi;
5209 		}
5210 	}
5211 
5212 	/*
5213 	 * bar == -1 => "trust me I know what I'm doing"
5214 	 * https://www.youtube.com/watch?v=nnwWKkNau4I
5215 	 * Some drivers are for hardware that is so shoddily
5216 	 * documented that no one knows which bars are which
5217 	 * so the developer has to map all bars. This hack
5218 	 * allows shoddy garbage to use msix in this framework.
5219 	 */
5220 	if (bar != -1) {
5221 		ctx->ifc_msix_mem = bus_alloc_resource_any(dev,
5222 	            SYS_RES_MEMORY, &bar, RF_ACTIVE);
5223 		if (ctx->ifc_msix_mem == NULL) {
5224 			/* May not be enabled */
5225 			device_printf(dev, "Unable to map MSIX table \n");
5226 			goto msi;
5227 		}
5228 	}
5229 	/* First try MSI/X */
5230 	if ((msgs = pci_msix_count(dev)) == 0) { /* system has msix disabled */
5231 		device_printf(dev, "System has MSIX disabled \n");
5232 		bus_release_resource(dev, SYS_RES_MEMORY,
5233 		    bar, ctx->ifc_msix_mem);
5234 		ctx->ifc_msix_mem = NULL;
5235 		goto msi;
5236 	}
5237 #if IFLIB_DEBUG
5238 	/* use only 1 qset in debug mode */
5239 	queuemsgs = min(msgs - admincnt, 1);
5240 #else
5241 	queuemsgs = msgs - admincnt;
5242 #endif
5243 	if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) == 0) {
5244 #ifdef RSS
5245 		queues = imin(queuemsgs, rss_getnumbuckets());
5246 #else
5247 		queues = queuemsgs;
5248 #endif
5249 		queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues);
5250 		device_printf(dev, "pxm cpus: %d queue msgs: %d admincnt: %d\n",
5251 					  CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt);
5252 	} else {
5253 		device_printf(dev, "Unable to fetch CPU list\n");
5254 		/* Figure out a reasonable auto config value */
5255 		queues = min(queuemsgs, mp_ncpus);
5256 	}
5257 #ifdef  RSS
5258 	/* If we're doing RSS, clamp at the number of RSS buckets */
5259 	if (queues > rss_getnumbuckets())
5260 		queues = rss_getnumbuckets();
5261 #endif
5262 	if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt)
5263 		rx_queues = iflib_num_rx_queues;
5264 	else
5265 		rx_queues = queues;
5266 	/*
5267 	 * We want this to be all logical CPUs by default
5268 	 */
5269 	if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues)
5270 		tx_queues = iflib_num_tx_queues;
5271 	else
5272 		tx_queues = mp_ncpus;
5273 
5274 	if (ctx->ifc_sysctl_qs_eq_override == 0) {
5275 #ifdef INVARIANTS
5276 		if (tx_queues != rx_queues)
5277 			device_printf(dev, "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n",
5278 				      min(rx_queues, tx_queues), min(rx_queues, tx_queues));
5279 #endif
5280 		tx_queues = min(rx_queues, tx_queues);
5281 		rx_queues = min(rx_queues, tx_queues);
5282 	}
5283 
5284 	device_printf(dev, "using %d rx queues %d tx queues \n", rx_queues, tx_queues);
5285 
5286 	vectors = rx_queues + admincnt;
5287 	if ((err = pci_alloc_msix(dev, &vectors)) == 0) {
5288 		device_printf(dev,
5289 					  "Using MSIX interrupts with %d vectors\n", vectors);
5290 		scctx->isc_vectors = vectors;
5291 		scctx->isc_nrxqsets = rx_queues;
5292 		scctx->isc_ntxqsets = tx_queues;
5293 		scctx->isc_intr = IFLIB_INTR_MSIX;
5294 
5295 		return (vectors);
5296 	} else {
5297 		device_printf(dev, "failed to allocate %d msix vectors, err: %d - using MSI\n", vectors, err);
5298 	}
5299 msi:
5300 	vectors = pci_msi_count(dev);
5301 	scctx->isc_nrxqsets = 1;
5302 	scctx->isc_ntxqsets = 1;
5303 	scctx->isc_vectors = vectors;
5304 	if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) {
5305 		device_printf(dev,"Using an MSI interrupt\n");
5306 		scctx->isc_intr = IFLIB_INTR_MSI;
5307 	} else {
5308 		device_printf(dev,"Using a Legacy interrupt\n");
5309 		scctx->isc_intr = IFLIB_INTR_LEGACY;
5310 	}
5311 
5312 	return (vectors);
5313 }
5314 
5315 char * ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" };
5316 
5317 static int
5318 mp_ring_state_handler(SYSCTL_HANDLER_ARGS)
5319 {
5320 	int rc;
5321 	uint16_t *state = ((uint16_t *)oidp->oid_arg1);
5322 	struct sbuf *sb;
5323 	char *ring_state = "UNKNOWN";
5324 
5325 	/* XXX needed ? */
5326 	rc = sysctl_wire_old_buffer(req, 0);
5327 	MPASS(rc == 0);
5328 	if (rc != 0)
5329 		return (rc);
5330 	sb = sbuf_new_for_sysctl(NULL, NULL, 80, req);
5331 	MPASS(sb != NULL);
5332 	if (sb == NULL)
5333 		return (ENOMEM);
5334 	if (state[3] <= 3)
5335 		ring_state = ring_states[state[3]];
5336 
5337 	sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s",
5338 		    state[0], state[1], state[2], ring_state);
5339 	rc = sbuf_finish(sb);
5340 	sbuf_delete(sb);
5341         return(rc);
5342 }
5343 
5344 enum iflib_ndesc_handler {
5345 	IFLIB_NTXD_HANDLER,
5346 	IFLIB_NRXD_HANDLER,
5347 };
5348 
5349 static int
5350 mp_ndesc_handler(SYSCTL_HANDLER_ARGS)
5351 {
5352 	if_ctx_t ctx = (void *)arg1;
5353 	enum iflib_ndesc_handler type = arg2;
5354 	char buf[256] = {0};
5355 	qidx_t *ndesc;
5356 	char *p, *next;
5357 	int nqs, rc, i;
5358 
5359 	MPASS(type == IFLIB_NTXD_HANDLER || type == IFLIB_NRXD_HANDLER);
5360 
5361 	nqs = 8;
5362 	switch(type) {
5363 	case IFLIB_NTXD_HANDLER:
5364 		ndesc = ctx->ifc_sysctl_ntxds;
5365 		if (ctx->ifc_sctx)
5366 			nqs = ctx->ifc_sctx->isc_ntxqs;
5367 		break;
5368 	case IFLIB_NRXD_HANDLER:
5369 		ndesc = ctx->ifc_sysctl_nrxds;
5370 		if (ctx->ifc_sctx)
5371 			nqs = ctx->ifc_sctx->isc_nrxqs;
5372 		break;
5373 	}
5374 	if (nqs == 0)
5375 		nqs = 8;
5376 
5377 	for (i=0; i<8; i++) {
5378 		if (i >= nqs)
5379 			break;
5380 		if (i)
5381 			strcat(buf, ",");
5382 		sprintf(strchr(buf, 0), "%d", ndesc[i]);
5383 	}
5384 
5385 	rc = sysctl_handle_string(oidp, buf, sizeof(buf), req);
5386 	if (rc || req->newptr == NULL)
5387 		return rc;
5388 
5389 	for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p;
5390 	    i++, p = strsep(&next, " ,")) {
5391 		ndesc[i] = strtoul(p, NULL, 10);
5392 	}
5393 
5394 	return(rc);
5395 }
5396 
5397 #define NAME_BUFLEN 32
5398 static void
5399 iflib_add_device_sysctl_pre(if_ctx_t ctx)
5400 {
5401         device_t dev = iflib_get_dev(ctx);
5402 	struct sysctl_oid_list *child, *oid_list;
5403 	struct sysctl_ctx_list *ctx_list;
5404 	struct sysctl_oid *node;
5405 
5406 	ctx_list = device_get_sysctl_ctx(dev);
5407 	child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev));
5408 	ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, "iflib",
5409 						      CTLFLAG_RD, NULL, "IFLIB fields");
5410 	oid_list = SYSCTL_CHILDREN(node);
5411 
5412 	SYSCTL_ADD_STRING(ctx_list, oid_list, OID_AUTO, "driver_version",
5413 		       CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version, 0,
5414 		       "driver version");
5415 
5416 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs",
5417 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0,
5418 			"# of txqs to use, 0 => use default #");
5419 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs",
5420 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0,
5421 			"# of rxqs to use, 0 => use default #");
5422 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_qs_enable",
5423 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0,
5424                        "permit #txq != #rxq");
5425 
5426 	/* XXX change for per-queue sizes */
5427 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_ntxds",
5428 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NTXD_HANDLER,
5429                        mp_ndesc_handler, "A",
5430                        "list of # of tx descriptors to use, 0 = use default #");
5431 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_nrxds",
5432 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NRXD_HANDLER,
5433                        mp_ndesc_handler, "A",
5434                        "list of # of rx descriptors to use, 0 = use default #");
5435 }
5436 
5437 static void
5438 iflib_add_device_sysctl_post(if_ctx_t ctx)
5439 {
5440 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5441 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5442         device_t dev = iflib_get_dev(ctx);
5443 	struct sysctl_oid_list *child;
5444 	struct sysctl_ctx_list *ctx_list;
5445 	iflib_fl_t fl;
5446 	iflib_txq_t txq;
5447 	iflib_rxq_t rxq;
5448 	int i, j;
5449 	char namebuf[NAME_BUFLEN];
5450 	char *qfmt;
5451 	struct sysctl_oid *queue_node, *fl_node, *node;
5452 	struct sysctl_oid_list *queue_list, *fl_list;
5453 	ctx_list = device_get_sysctl_ctx(dev);
5454 
5455 	node = ctx->ifc_sysctl_node;
5456 	child = SYSCTL_CHILDREN(node);
5457 
5458 	if (scctx->isc_ntxqsets > 100)
5459 		qfmt = "txq%03d";
5460 	else if (scctx->isc_ntxqsets > 10)
5461 		qfmt = "txq%02d";
5462 	else
5463 		qfmt = "txq%d";
5464 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
5465 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
5466 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
5467 					     CTLFLAG_RD, NULL, "Queue Name");
5468 		queue_list = SYSCTL_CHILDREN(queue_node);
5469 #if MEMORY_LOGGING
5470 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued",
5471 				CTLFLAG_RD,
5472 				&txq->ift_dequeued, "total mbufs freed");
5473 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued",
5474 				CTLFLAG_RD,
5475 				&txq->ift_enqueued, "total mbufs enqueued");
5476 #endif
5477 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag",
5478 				   CTLFLAG_RD,
5479 				   &txq->ift_mbuf_defrag, "# of times m_defrag was called");
5480 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "m_pullups",
5481 				   CTLFLAG_RD,
5482 				   &txq->ift_pullups, "# of times m_pullup was called");
5483 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag_failed",
5484 				   CTLFLAG_RD,
5485 				   &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed");
5486 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_desc_avail",
5487 				   CTLFLAG_RD,
5488 				   &txq->ift_no_desc_avail, "# of times no descriptors were available");
5489 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "tx_map_failed",
5490 				   CTLFLAG_RD,
5491 				   &txq->ift_map_failed, "# of times dma map failed");
5492 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txd_encap_efbig",
5493 				   CTLFLAG_RD,
5494 				   &txq->ift_txd_encap_efbig, "# of times txd_encap returned EFBIG");
5495 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_tx_dma_setup",
5496 				   CTLFLAG_RD,
5497 				   &txq->ift_no_tx_dma_setup, "# of times map failed for other than EFBIG");
5498 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx",
5499 				   CTLFLAG_RD,
5500 				   &txq->ift_pidx, 1, "Producer Index");
5501 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx",
5502 				   CTLFLAG_RD,
5503 				   &txq->ift_cidx, 1, "Consumer Index");
5504 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx_processed",
5505 				   CTLFLAG_RD,
5506 				   &txq->ift_cidx_processed, 1, "Consumer Index seen by credit update");
5507 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use",
5508 				   CTLFLAG_RD,
5509 				   &txq->ift_in_use, 1, "descriptors in use");
5510 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_processed",
5511 				   CTLFLAG_RD,
5512 				   &txq->ift_processed, "descriptors procesed for clean");
5513 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned",
5514 				   CTLFLAG_RD,
5515 				   &txq->ift_cleaned, "total cleaned");
5516 		SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state",
5517 				CTLTYPE_STRING | CTLFLAG_RD, __DEVOLATILE(uint64_t *, &txq->ift_br->state),
5518 				0, mp_ring_state_handler, "A", "soft ring state");
5519 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_enqueues",
5520 				       CTLFLAG_RD, &txq->ift_br->enqueues,
5521 				       "# of enqueues to the mp_ring for this queue");
5522 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_drops",
5523 				       CTLFLAG_RD, &txq->ift_br->drops,
5524 				       "# of drops in the mp_ring for this queue");
5525 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_starts",
5526 				       CTLFLAG_RD, &txq->ift_br->starts,
5527 				       "# of normal consumer starts in the mp_ring for this queue");
5528 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_stalls",
5529 				       CTLFLAG_RD, &txq->ift_br->stalls,
5530 					       "# of consumer stalls in the mp_ring for this queue");
5531 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_restarts",
5532 			       CTLFLAG_RD, &txq->ift_br->restarts,
5533 				       "# of consumer restarts in the mp_ring for this queue");
5534 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_abdications",
5535 				       CTLFLAG_RD, &txq->ift_br->abdications,
5536 				       "# of consumer abdications in the mp_ring for this queue");
5537 	}
5538 
5539 	if (scctx->isc_nrxqsets > 100)
5540 		qfmt = "rxq%03d";
5541 	else if (scctx->isc_nrxqsets > 10)
5542 		qfmt = "rxq%02d";
5543 	else
5544 		qfmt = "rxq%d";
5545 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
5546 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
5547 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
5548 					     CTLFLAG_RD, NULL, "Queue Name");
5549 		queue_list = SYSCTL_CHILDREN(queue_node);
5550 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
5551 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_pidx",
5552 				       CTLFLAG_RD,
5553 				       &rxq->ifr_cq_pidx, 1, "Producer Index");
5554 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_cidx",
5555 				       CTLFLAG_RD,
5556 				       &rxq->ifr_cq_cidx, 1, "Consumer Index");
5557 		}
5558 
5559 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
5560 			snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j);
5561 			fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list, OID_AUTO, namebuf,
5562 						     CTLFLAG_RD, NULL, "freelist Name");
5563 			fl_list = SYSCTL_CHILDREN(fl_node);
5564 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx",
5565 				       CTLFLAG_RD,
5566 				       &fl->ifl_pidx, 1, "Producer Index");
5567 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx",
5568 				       CTLFLAG_RD,
5569 				       &fl->ifl_cidx, 1, "Consumer Index");
5570 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits",
5571 				       CTLFLAG_RD,
5572 				       &fl->ifl_credits, 1, "credits available");
5573 #if MEMORY_LOGGING
5574 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_enqueued",
5575 					CTLFLAG_RD,
5576 					&fl->ifl_m_enqueued, "mbufs allocated");
5577 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_dequeued",
5578 					CTLFLAG_RD,
5579 					&fl->ifl_m_dequeued, "mbufs freed");
5580 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_enqueued",
5581 					CTLFLAG_RD,
5582 					&fl->ifl_cl_enqueued, "clusters allocated");
5583 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_dequeued",
5584 					CTLFLAG_RD,
5585 					&fl->ifl_cl_dequeued, "clusters freed");
5586 #endif
5587 
5588 		}
5589 	}
5590 
5591 }
5592 
5593 #ifndef __NO_STRICT_ALIGNMENT
5594 static struct mbuf *
5595 iflib_fixup_rx(struct mbuf *m)
5596 {
5597 	struct mbuf *n;
5598 
5599 	if (m->m_len <= (MCLBYTES - ETHER_HDR_LEN)) {
5600 		bcopy(m->m_data, m->m_data + ETHER_HDR_LEN, m->m_len);
5601 		m->m_data += ETHER_HDR_LEN;
5602 		n = m;
5603 	} else {
5604 		MGETHDR(n, M_NOWAIT, MT_DATA);
5605 		if (n == NULL) {
5606 			m_freem(m);
5607 			return (NULL);
5608 		}
5609 		bcopy(m->m_data, n->m_data, ETHER_HDR_LEN);
5610 		m->m_data += ETHER_HDR_LEN;
5611 		m->m_len -= ETHER_HDR_LEN;
5612 		n->m_len = ETHER_HDR_LEN;
5613 		M_MOVE_PKTHDR(n, m);
5614 		n->m_next = m;
5615 	}
5616 	return (n);
5617 }
5618 #endif
5619