xref: /freebsd/sys/net/iflib.c (revision 8554754c7bbe3cda171378ad6305bbe8a1a94cb6)
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, flags;
1445 	struct resource *res;
1446 	void *tag = NULL;
1447 	device_t dev = ctx->ifc_dev;
1448 
1449 	flags = RF_ACTIVE;
1450 	if (ctx->ifc_flags & IFC_LEGACY)
1451 		flags |= RF_SHAREABLE;
1452 	MPASS(rid < 512);
1453 	irq->ii_rid = rid;
1454 	res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &irq->ii_rid, flags);
1455 	if (res == NULL) {
1456 		device_printf(dev,
1457 		    "failed to allocate IRQ for rid %d, name %s.\n", rid, name);
1458 		return (ENOMEM);
1459 	}
1460 	irq->ii_res = res;
1461 	KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL"));
1462 	rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET,
1463 						filter, handler, arg, &tag);
1464 	if (rc != 0) {
1465 		device_printf(dev,
1466 		    "failed to setup interrupt for rid %d, name %s: %d\n",
1467 					  rid, name ? name : "unknown", rc);
1468 		return (rc);
1469 	} else if (name)
1470 		bus_describe_intr(dev, res, tag, "%s", name);
1471 
1472 	irq->ii_tag = tag;
1473 	return (0);
1474 }
1475 
1476 
1477 /*********************************************************************
1478  *
1479  *  Allocate memory for tx_buffer structures. The tx_buffer stores all
1480  *  the information needed to transmit a packet on the wire. This is
1481  *  called only once at attach, setup is done every reset.
1482  *
1483  **********************************************************************/
1484 
1485 static int
1486 iflib_txsd_alloc(iflib_txq_t txq)
1487 {
1488 	if_ctx_t ctx = txq->ift_ctx;
1489 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1490 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1491 	device_t dev = ctx->ifc_dev;
1492 	int err, nsegments, ntsosegments;
1493 
1494 	nsegments = scctx->isc_tx_nsegments;
1495 	ntsosegments = scctx->isc_tx_tso_segments_max;
1496 	MPASS(scctx->isc_ntxd[0] > 0);
1497 	MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0);
1498 	MPASS(nsegments > 0);
1499 	MPASS(ntsosegments > 0);
1500 	/*
1501 	 * Setup DMA descriptor areas.
1502 	 */
1503 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1504 			       1, 0,			/* alignment, bounds */
1505 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1506 			       BUS_SPACE_MAXADDR,	/* highaddr */
1507 			       NULL, NULL,		/* filter, filterarg */
1508 			       sctx->isc_tx_maxsize,		/* maxsize */
1509 			       nsegments,	/* nsegments */
1510 			       sctx->isc_tx_maxsegsize,	/* maxsegsize */
1511 			       0,			/* flags */
1512 			       NULL,			/* lockfunc */
1513 			       NULL,			/* lockfuncarg */
1514 			       &txq->ift_desc_tag))) {
1515 		device_printf(dev,"Unable to allocate TX DMA tag: %d\n", err);
1516 		device_printf(dev,"maxsize: %zd nsegments: %d maxsegsize: %zd\n",
1517 					  sctx->isc_tx_maxsize, nsegments, sctx->isc_tx_maxsegsize);
1518 		goto fail;
1519 	}
1520 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1521 			       1, 0,			/* alignment, bounds */
1522 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1523 			       BUS_SPACE_MAXADDR,	/* highaddr */
1524 			       NULL, NULL,		/* filter, filterarg */
1525 			       scctx->isc_tx_tso_size_max,		/* maxsize */
1526 			       ntsosegments,	/* nsegments */
1527 			       scctx->isc_tx_tso_segsize_max,	/* maxsegsize */
1528 			       0,			/* flags */
1529 			       NULL,			/* lockfunc */
1530 			       NULL,			/* lockfuncarg */
1531 			       &txq->ift_tso_desc_tag))) {
1532 		device_printf(dev,"Unable to allocate TX TSO DMA tag: %d\n", err);
1533 
1534 		goto fail;
1535 	}
1536 	if (!(txq->ift_sds.ifsd_flags =
1537 	    (uint8_t *) malloc(sizeof(uint8_t) *
1538 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1539 		device_printf(dev, "Unable to allocate tx_buffer memory\n");
1540 		err = ENOMEM;
1541 		goto fail;
1542 	}
1543 	if (!(txq->ift_sds.ifsd_m =
1544 	    (struct mbuf **) malloc(sizeof(struct mbuf *) *
1545 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1546 		device_printf(dev, "Unable to allocate tx_buffer memory\n");
1547 		err = ENOMEM;
1548 		goto fail;
1549 	}
1550 
1551         /* Create the descriptor buffer dma maps */
1552 #if defined(ACPI_DMAR) || (! (defined(__i386__) || defined(__amd64__)))
1553 	if ((ctx->ifc_flags & IFC_DMAR) == 0)
1554 		return (0);
1555 
1556 	if (!(txq->ift_sds.ifsd_map =
1557 	    (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1558 		device_printf(dev, "Unable to allocate tx_buffer map memory\n");
1559 		err = ENOMEM;
1560 		goto fail;
1561 	}
1562 
1563 	for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) {
1564 		err = bus_dmamap_create(txq->ift_desc_tag, 0, &txq->ift_sds.ifsd_map[i]);
1565 		if (err != 0) {
1566 			device_printf(dev, "Unable to create TX DMA map\n");
1567 			goto fail;
1568 		}
1569 	}
1570 #endif
1571 	return (0);
1572 fail:
1573 	/* We free all, it handles case where we are in the middle */
1574 	iflib_tx_structures_free(ctx);
1575 	return (err);
1576 }
1577 
1578 static void
1579 iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i)
1580 {
1581 	bus_dmamap_t map;
1582 
1583 	map = NULL;
1584 	if (txq->ift_sds.ifsd_map != NULL)
1585 		map = txq->ift_sds.ifsd_map[i];
1586 	if (map != NULL) {
1587 		bus_dmamap_unload(txq->ift_desc_tag, map);
1588 		bus_dmamap_destroy(txq->ift_desc_tag, map);
1589 		txq->ift_sds.ifsd_map[i] = NULL;
1590 	}
1591 }
1592 
1593 static void
1594 iflib_txq_destroy(iflib_txq_t txq)
1595 {
1596 	if_ctx_t ctx = txq->ift_ctx;
1597 
1598 	for (int i = 0; i < txq->ift_size; i++)
1599 		iflib_txsd_destroy(ctx, txq, i);
1600 	if (txq->ift_sds.ifsd_map != NULL) {
1601 		free(txq->ift_sds.ifsd_map, M_IFLIB);
1602 		txq->ift_sds.ifsd_map = NULL;
1603 	}
1604 	if (txq->ift_sds.ifsd_m != NULL) {
1605 		free(txq->ift_sds.ifsd_m, M_IFLIB);
1606 		txq->ift_sds.ifsd_m = NULL;
1607 	}
1608 	if (txq->ift_sds.ifsd_flags != NULL) {
1609 		free(txq->ift_sds.ifsd_flags, M_IFLIB);
1610 		txq->ift_sds.ifsd_flags = NULL;
1611 	}
1612 	if (txq->ift_desc_tag != NULL) {
1613 		bus_dma_tag_destroy(txq->ift_desc_tag);
1614 		txq->ift_desc_tag = NULL;
1615 	}
1616 	if (txq->ift_tso_desc_tag != NULL) {
1617 		bus_dma_tag_destroy(txq->ift_tso_desc_tag);
1618 		txq->ift_tso_desc_tag = NULL;
1619 	}
1620 }
1621 
1622 static void
1623 iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i)
1624 {
1625 	struct mbuf **mp;
1626 
1627 	mp = &txq->ift_sds.ifsd_m[i];
1628 	if (*mp == NULL)
1629 		return;
1630 
1631 	if (txq->ift_sds.ifsd_map != NULL) {
1632 		bus_dmamap_sync(txq->ift_desc_tag,
1633 				txq->ift_sds.ifsd_map[i],
1634 				BUS_DMASYNC_POSTWRITE);
1635 		bus_dmamap_unload(txq->ift_desc_tag,
1636 				  txq->ift_sds.ifsd_map[i]);
1637 	}
1638 	m_free(*mp);
1639 	DBG_COUNTER_INC(tx_frees);
1640 	*mp = NULL;
1641 }
1642 
1643 static int
1644 iflib_txq_setup(iflib_txq_t txq)
1645 {
1646 	if_ctx_t ctx = txq->ift_ctx;
1647 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1648 	iflib_dma_info_t di;
1649 	int i;
1650 
1651 	/* Set number of descriptors available */
1652 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
1653 	/* XXX make configurable */
1654 	txq->ift_update_freq = IFLIB_DEFAULT_TX_UPDATE_FREQ;
1655 
1656 	/* Reset indices */
1657 	txq->ift_cidx_processed = 0;
1658 	txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0;
1659 	txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset];
1660 
1661 	for (i = 0, di = txq->ift_ifdi; i < ctx->ifc_nhwtxqs; i++, di++)
1662 		bzero((void *)di->idi_vaddr, di->idi_size);
1663 
1664 	IFDI_TXQ_SETUP(ctx, txq->ift_id);
1665 	for (i = 0, di = txq->ift_ifdi; i < ctx->ifc_nhwtxqs; i++, di++)
1666 		bus_dmamap_sync(di->idi_tag, di->idi_map,
1667 						BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1668 	return (0);
1669 }
1670 
1671 /*********************************************************************
1672  *
1673  *  Allocate memory for rx_buffer structures. Since we use one
1674  *  rx_buffer per received packet, the maximum number of rx_buffer's
1675  *  that we'll need is equal to the number of receive descriptors
1676  *  that we've allocated.
1677  *
1678  **********************************************************************/
1679 static int
1680 iflib_rxsd_alloc(iflib_rxq_t rxq)
1681 {
1682 	if_ctx_t ctx = rxq->ifr_ctx;
1683 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1684 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1685 	device_t dev = ctx->ifc_dev;
1686 	iflib_fl_t fl;
1687 	int			err;
1688 
1689 	MPASS(scctx->isc_nrxd[0] > 0);
1690 	MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0);
1691 
1692 	fl = rxq->ifr_fl;
1693 	for (int i = 0; i <  rxq->ifr_nfl; i++, fl++) {
1694 		fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */
1695 		err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1696 					 1, 0,			/* alignment, bounds */
1697 					 BUS_SPACE_MAXADDR,	/* lowaddr */
1698 					 BUS_SPACE_MAXADDR,	/* highaddr */
1699 					 NULL, NULL,		/* filter, filterarg */
1700 					 sctx->isc_rx_maxsize,	/* maxsize */
1701 					 sctx->isc_rx_nsegments,	/* nsegments */
1702 					 sctx->isc_rx_maxsegsize,	/* maxsegsize */
1703 					 0,			/* flags */
1704 					 NULL,			/* lockfunc */
1705 					 NULL,			/* lockarg */
1706 					 &fl->ifl_desc_tag);
1707 		if (err) {
1708 			device_printf(dev, "%s: bus_dma_tag_create failed %d\n",
1709 				__func__, err);
1710 			goto fail;
1711 		}
1712 		if (!(fl->ifl_sds.ifsd_flags =
1713 		      (uint8_t *) malloc(sizeof(uint8_t) *
1714 					 scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1715 			device_printf(dev, "Unable to allocate tx_buffer memory\n");
1716 			err = ENOMEM;
1717 			goto fail;
1718 		}
1719 		if (!(fl->ifl_sds.ifsd_m =
1720 		      (struct mbuf **) malloc(sizeof(struct mbuf *) *
1721 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1722 			device_printf(dev, "Unable to allocate tx_buffer memory\n");
1723 			err = ENOMEM;
1724 			goto fail;
1725 		}
1726 		if (!(fl->ifl_sds.ifsd_cl =
1727 		      (caddr_t *) malloc(sizeof(caddr_t) *
1728 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1729 			device_printf(dev, "Unable to allocate tx_buffer memory\n");
1730 			err = ENOMEM;
1731 			goto fail;
1732 		}
1733 
1734 		/* Create the descriptor buffer dma maps */
1735 #if defined(ACPI_DMAR) || (! (defined(__i386__) || defined(__amd64__)))
1736 		if ((ctx->ifc_flags & IFC_DMAR) == 0)
1737 			continue;
1738 
1739 		if (!(fl->ifl_sds.ifsd_map =
1740 		      (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1741 			device_printf(dev, "Unable to allocate tx_buffer map memory\n");
1742 			err = ENOMEM;
1743 			goto fail;
1744 		}
1745 
1746 		for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++) {
1747 			err = bus_dmamap_create(fl->ifl_desc_tag, 0, &fl->ifl_sds.ifsd_map[i]);
1748 			if (err != 0) {
1749 				device_printf(dev, "Unable to create RX buffer DMA map\n");
1750 				goto fail;
1751 			}
1752 		}
1753 #endif
1754 	}
1755 	return (0);
1756 
1757 fail:
1758 	iflib_rx_structures_free(ctx);
1759 	return (err);
1760 }
1761 
1762 
1763 /*
1764  * Internal service routines
1765  */
1766 
1767 struct rxq_refill_cb_arg {
1768 	int               error;
1769 	bus_dma_segment_t seg;
1770 	int               nseg;
1771 };
1772 
1773 static void
1774 _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
1775 {
1776 	struct rxq_refill_cb_arg *cb_arg = arg;
1777 
1778 	cb_arg->error = error;
1779 	cb_arg->seg = segs[0];
1780 	cb_arg->nseg = nseg;
1781 }
1782 
1783 
1784 #ifdef ACPI_DMAR
1785 #define IS_DMAR(ctx) (ctx->ifc_flags & IFC_DMAR)
1786 #else
1787 #define IS_DMAR(ctx) (0)
1788 #endif
1789 
1790 /**
1791  *	rxq_refill - refill an rxq  free-buffer list
1792  *	@ctx: the iflib context
1793  *	@rxq: the free-list to refill
1794  *	@n: the number of new buffers to allocate
1795  *
1796  *	(Re)populate an rxq free-buffer list with up to @n new packet buffers.
1797  *	The caller must assure that @n does not exceed the queue's capacity.
1798  */
1799 static void
1800 _iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
1801 {
1802 	struct mbuf *m;
1803 	int idx, pidx = fl->ifl_pidx;
1804 	caddr_t cl, *sd_cl;
1805 	struct mbuf **sd_m;
1806 	uint8_t *sd_flags;
1807 	struct if_rxd_update iru;
1808 	bus_dmamap_t *sd_map;
1809 	int n, i = 0;
1810 	uint64_t bus_addr;
1811 	int err;
1812 
1813 	sd_m = fl->ifl_sds.ifsd_m;
1814 	sd_map = fl->ifl_sds.ifsd_map;
1815 	sd_cl = fl->ifl_sds.ifsd_cl;
1816 	sd_flags = fl->ifl_sds.ifsd_flags;
1817 	idx = pidx;
1818 
1819 	n  = count;
1820 	MPASS(n > 0);
1821 	MPASS(fl->ifl_credits + n <= fl->ifl_size);
1822 
1823 	if (pidx < fl->ifl_cidx)
1824 		MPASS(pidx + n <= fl->ifl_cidx);
1825 	if (pidx == fl->ifl_cidx && (fl->ifl_credits < fl->ifl_size))
1826 		MPASS(fl->ifl_gen == 0);
1827 	if (pidx > fl->ifl_cidx)
1828 		MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx);
1829 
1830 	DBG_COUNTER_INC(fl_refills);
1831 	if (n > 8)
1832 		DBG_COUNTER_INC(fl_refills_large);
1833 	iru.iru_paddrs = fl->ifl_bus_addrs;
1834 	iru.iru_vaddrs = &fl->ifl_vm_addrs[0];
1835 	iru.iru_idxs = fl->ifl_rxd_idxs;
1836 	iru.iru_qsidx = fl->ifl_rxq->ifr_id;
1837 	iru.iru_buf_size = fl->ifl_buf_size;
1838 	iru.iru_flidx = fl->ifl_id;
1839 	while (n--) {
1840 		/*
1841 		 * We allocate an uninitialized mbuf + cluster, mbuf is
1842 		 * initialized after rx.
1843 		 *
1844 		 * If the cluster is still set then we know a minimum sized packet was received
1845 		 */
1846 		if ((cl = sd_cl[idx]) == NULL) {
1847 			if ((cl = sd_cl[idx] = m_cljget(NULL, M_NOWAIT, fl->ifl_buf_size)) == NULL)
1848 				break;
1849 #if MEMORY_LOGGING
1850 			fl->ifl_cl_enqueued++;
1851 #endif
1852 		}
1853 		if ((m = m_gethdr(M_NOWAIT, MT_NOINIT)) == NULL) {
1854 			break;
1855 		}
1856 #if MEMORY_LOGGING
1857 		fl->ifl_m_enqueued++;
1858 #endif
1859 
1860 		DBG_COUNTER_INC(rx_allocs);
1861 #if defined(__i386__) || defined(__amd64__)
1862 		if (!IS_DMAR(ctx)) {
1863 			bus_addr = pmap_kextract((vm_offset_t)cl);
1864 		} else
1865 #endif
1866 		{
1867 			struct rxq_refill_cb_arg cb_arg;
1868 			iflib_rxq_t q;
1869 
1870 			cb_arg.error = 0;
1871 			q = fl->ifl_rxq;
1872 			MPASS(sd_map != NULL);
1873 			MPASS(sd_map[idx] != NULL);
1874 			err = bus_dmamap_load(fl->ifl_desc_tag, sd_map[idx],
1875 		         cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg, 0);
1876 			bus_dmamap_sync(fl->ifl_desc_tag, sd_map[idx], BUS_DMASYNC_PREREAD);
1877 
1878 			if (err != 0 || cb_arg.error) {
1879 				/*
1880 				 * !zone_pack ?
1881 				 */
1882 				if (fl->ifl_zone == zone_pack)
1883 					uma_zfree(fl->ifl_zone, cl);
1884 				m_free(m);
1885 				n = 0;
1886 				goto done;
1887 			}
1888 			bus_addr = cb_arg.seg.ds_addr;
1889 		}
1890 		sd_flags[idx] |= RX_SW_DESC_INUSE;
1891 
1892 		MPASS(sd_m[idx] == NULL);
1893 		sd_cl[idx] = cl;
1894 		sd_m[idx] = m;
1895 		fl->ifl_rxd_idxs[i] = idx;
1896 		fl->ifl_bus_addrs[i] = bus_addr;
1897 		fl->ifl_vm_addrs[i] = cl;
1898 		fl->ifl_credits++;
1899 		i++;
1900 		MPASS(fl->ifl_credits <= fl->ifl_size);
1901 		if (++idx == fl->ifl_size) {
1902 			fl->ifl_gen = 1;
1903 			idx = 0;
1904 		}
1905 		if (n == 0 || i == IFLIB_MAX_RX_REFRESH) {
1906 			iru.iru_pidx = pidx;
1907 			iru.iru_count = i;
1908 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
1909 			i = 0;
1910 			pidx = idx;
1911 		}
1912 		fl->ifl_pidx = idx;
1913 
1914 	}
1915 done:
1916 	DBG_COUNTER_INC(rxd_flush);
1917 	if (fl->ifl_pidx == 0)
1918 		pidx = fl->ifl_size - 1;
1919 	else
1920 		pidx = fl->ifl_pidx - 1;
1921 
1922 	if (sd_map)
1923 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
1924 				BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1925 	ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id, fl->ifl_id, pidx);
1926 }
1927 
1928 static __inline void
1929 __iflib_fl_refill_lt(if_ctx_t ctx, iflib_fl_t fl, int max)
1930 {
1931 	/* we avoid allowing pidx to catch up with cidx as it confuses ixl */
1932 	int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1;
1933 #ifdef INVARIANTS
1934 	int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1;
1935 #endif
1936 
1937 	MPASS(fl->ifl_credits <= fl->ifl_size);
1938 	MPASS(reclaimable == delta);
1939 
1940 	if (reclaimable > 0)
1941 		_iflib_fl_refill(ctx, fl, min(max, reclaimable));
1942 }
1943 
1944 static void
1945 iflib_fl_bufs_free(iflib_fl_t fl)
1946 {
1947 	iflib_dma_info_t idi = fl->ifl_ifdi;
1948 	uint32_t i;
1949 
1950 	for (i = 0; i < fl->ifl_size; i++) {
1951 		struct mbuf **sd_m = &fl->ifl_sds.ifsd_m[i];
1952 		uint8_t *sd_flags = &fl->ifl_sds.ifsd_flags[i];
1953 		caddr_t *sd_cl = &fl->ifl_sds.ifsd_cl[i];
1954 
1955 		if (*sd_flags & RX_SW_DESC_INUSE) {
1956 			if (fl->ifl_sds.ifsd_map != NULL) {
1957 				bus_dmamap_t sd_map = fl->ifl_sds.ifsd_map[i];
1958 				bus_dmamap_unload(fl->ifl_desc_tag, sd_map);
1959 				bus_dmamap_destroy(fl->ifl_desc_tag, sd_map);
1960 			}
1961 			if (*sd_m != NULL) {
1962 				m_init(*sd_m, M_NOWAIT, MT_DATA, 0);
1963 				uma_zfree(zone_mbuf, *sd_m);
1964 			}
1965 			if (*sd_cl != NULL)
1966 				uma_zfree(fl->ifl_zone, *sd_cl);
1967 			*sd_flags = 0;
1968 		} else {
1969 			MPASS(*sd_cl == NULL);
1970 			MPASS(*sd_m == NULL);
1971 		}
1972 #if MEMORY_LOGGING
1973 		fl->ifl_m_dequeued++;
1974 		fl->ifl_cl_dequeued++;
1975 #endif
1976 		*sd_cl = NULL;
1977 		*sd_m = NULL;
1978 	}
1979 #ifdef INVARIANTS
1980 	for (i = 0; i < fl->ifl_size; i++) {
1981 		MPASS(fl->ifl_sds.ifsd_flags[i] == 0);
1982 		MPASS(fl->ifl_sds.ifsd_cl[i] == NULL);
1983 		MPASS(fl->ifl_sds.ifsd_m[i] == NULL);
1984 	}
1985 #endif
1986 	/*
1987 	 * Reset free list values
1988 	 */
1989 	fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = 0;;
1990 	bzero(idi->idi_vaddr, idi->idi_size);
1991 }
1992 
1993 /*********************************************************************
1994  *
1995  *  Initialize a receive ring and its buffers.
1996  *
1997  **********************************************************************/
1998 static int
1999 iflib_fl_setup(iflib_fl_t fl)
2000 {
2001 	iflib_rxq_t rxq = fl->ifl_rxq;
2002 	if_ctx_t ctx = rxq->ifr_ctx;
2003 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2004 
2005 	/*
2006 	** Free current RX buffer structs and their mbufs
2007 	*/
2008 	iflib_fl_bufs_free(fl);
2009 	/* Now replenish the mbufs */
2010 	MPASS(fl->ifl_credits == 0);
2011 	/*
2012 	 * XXX don't set the max_frame_size to larger
2013 	 * than the hardware can handle
2014 	 */
2015 	if (sctx->isc_max_frame_size <= 2048)
2016 		fl->ifl_buf_size = MCLBYTES;
2017 #ifndef CONTIGMALLOC_WORKS
2018 	else
2019 		fl->ifl_buf_size = MJUMPAGESIZE;
2020 #else
2021 	else if (sctx->isc_max_frame_size <= 4096)
2022 		fl->ifl_buf_size = MJUMPAGESIZE;
2023 	else if (sctx->isc_max_frame_size <= 9216)
2024 		fl->ifl_buf_size = MJUM9BYTES;
2025 	else
2026 		fl->ifl_buf_size = MJUM16BYTES;
2027 #endif
2028 	if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size)
2029 		ctx->ifc_max_fl_buf_size = fl->ifl_buf_size;
2030 	fl->ifl_cltype = m_gettype(fl->ifl_buf_size);
2031 	fl->ifl_zone = m_getzone(fl->ifl_buf_size);
2032 
2033 
2034 	/* avoid pre-allocating zillions of clusters to an idle card
2035 	 * potentially speeding up attach
2036 	 */
2037 	_iflib_fl_refill(ctx, fl, min(128, fl->ifl_size));
2038 	MPASS(min(128, fl->ifl_size) == fl->ifl_credits);
2039 	if (min(128, fl->ifl_size) != fl->ifl_credits)
2040 		return (ENOBUFS);
2041 	/*
2042 	 * handle failure
2043 	 */
2044 	MPASS(rxq != NULL);
2045 	MPASS(fl->ifl_ifdi != NULL);
2046 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2047 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2048 	return (0);
2049 }
2050 
2051 /*********************************************************************
2052  *
2053  *  Free receive ring data structures
2054  *
2055  **********************************************************************/
2056 static void
2057 iflib_rx_sds_free(iflib_rxq_t rxq)
2058 {
2059 	iflib_fl_t fl;
2060 	int i;
2061 
2062 	if (rxq->ifr_fl != NULL) {
2063 		for (i = 0; i < rxq->ifr_nfl; i++) {
2064 			fl = &rxq->ifr_fl[i];
2065 			if (fl->ifl_desc_tag != NULL) {
2066 				bus_dma_tag_destroy(fl->ifl_desc_tag);
2067 				fl->ifl_desc_tag = NULL;
2068 			}
2069 			free(fl->ifl_sds.ifsd_m, M_IFLIB);
2070 			free(fl->ifl_sds.ifsd_cl, M_IFLIB);
2071 			/* XXX destroy maps first */
2072 			free(fl->ifl_sds.ifsd_map, M_IFLIB);
2073 			fl->ifl_sds.ifsd_m = NULL;
2074 			fl->ifl_sds.ifsd_cl = NULL;
2075 			fl->ifl_sds.ifsd_map = NULL;
2076 		}
2077 		free(rxq->ifr_fl, M_IFLIB);
2078 		rxq->ifr_fl = NULL;
2079 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
2080 	}
2081 }
2082 
2083 /*
2084  * MI independent logic
2085  *
2086  */
2087 static void
2088 iflib_timer(void *arg)
2089 {
2090 	iflib_txq_t txq = arg;
2091 	if_ctx_t ctx = txq->ift_ctx;
2092 
2093 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
2094 		return;
2095 	/*
2096 	** Check on the state of the TX queue(s), this
2097 	** can be done without the lock because its RO
2098 	** and the HUNG state will be static if set.
2099 	*/
2100 	IFDI_TIMER(ctx, txq->ift_id);
2101 	if ((txq->ift_qstatus == IFLIB_QUEUE_HUNG) &&
2102 	    ((txq->ift_cleaned_prev == txq->ift_cleaned) ||
2103 	     (ctx->ifc_pause_frames == 0)))
2104 		goto hung;
2105 
2106 	if (ifmp_ring_is_stalled(txq->ift_br))
2107 		txq->ift_qstatus = IFLIB_QUEUE_HUNG;
2108 	txq->ift_cleaned_prev = txq->ift_cleaned;
2109 	/* handle any laggards */
2110 	if (txq->ift_db_pending)
2111 		GROUPTASK_ENQUEUE(&txq->ift_task);
2112 
2113 	ctx->ifc_pause_frames = 0;
2114 	if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)
2115 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq, txq->ift_timer.c_cpu);
2116 	return;
2117 hung:
2118 	CTX_LOCK(ctx);
2119 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2120 	device_printf(ctx->ifc_dev,  "TX(%d) desc avail = %d, pidx = %d\n",
2121 				  txq->ift_id, TXQ_AVAIL(txq), txq->ift_pidx);
2122 
2123 	IFDI_WATCHDOG_RESET(ctx);
2124 	ctx->ifc_watchdog_events++;
2125 	ctx->ifc_pause_frames = 0;
2126 
2127 	ctx->ifc_flags |= IFC_DO_RESET;
2128 	iflib_admin_intr_deferred(ctx);
2129 	CTX_UNLOCK(ctx);
2130 }
2131 
2132 static void
2133 iflib_init_locked(if_ctx_t ctx)
2134 {
2135 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2136 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2137 	if_t ifp = ctx->ifc_ifp;
2138 	iflib_fl_t fl;
2139 	iflib_txq_t txq;
2140 	iflib_rxq_t rxq;
2141 	int i, j, tx_ip_csum_flags, tx_ip6_csum_flags;
2142 
2143 
2144 	if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2145 	IFDI_INTR_DISABLE(ctx);
2146 
2147 	tx_ip_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP);
2148 	tx_ip6_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | CSUM_IP6_SCTP);
2149 	/* Set hardware offload abilities */
2150 	if_clearhwassist(ifp);
2151 	if (if_getcapenable(ifp) & IFCAP_TXCSUM)
2152 		if_sethwassistbits(ifp, tx_ip_csum_flags, 0);
2153 	if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6)
2154 		if_sethwassistbits(ifp,  tx_ip6_csum_flags, 0);
2155 	if (if_getcapenable(ifp) & IFCAP_TSO4)
2156 		if_sethwassistbits(ifp, CSUM_IP_TSO, 0);
2157 	if (if_getcapenable(ifp) & IFCAP_TSO6)
2158 		if_sethwassistbits(ifp, CSUM_IP6_TSO, 0);
2159 
2160 	for (i = 0, txq = ctx->ifc_txqs; i < sctx->isc_ntxqsets; i++, txq++) {
2161 		CALLOUT_LOCK(txq);
2162 		callout_stop(&txq->ift_timer);
2163 		CALLOUT_UNLOCK(txq);
2164 		iflib_netmap_txq_init(ctx, txq);
2165 	}
2166 	for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
2167 		MPASS(rxq->ifr_id == i);
2168 		iflib_netmap_rxq_init(ctx, rxq);
2169 	}
2170 #ifdef INVARIANTS
2171 	i = if_getdrvflags(ifp);
2172 #endif
2173 	IFDI_INIT(ctx);
2174 	MPASS(if_getdrvflags(ifp) == i);
2175 	for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
2176 		/* XXX this should really be done on a per-queue basis */
2177 		if (if_getcapenable(ifp) & IFCAP_NETMAP)
2178 			continue;
2179 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
2180 			if (iflib_fl_setup(fl)) {
2181 				device_printf(ctx->ifc_dev, "freelist setup failed - check cluster settings\n");
2182 				goto done;
2183 			}
2184 		}
2185 	}
2186 	done:
2187 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
2188 	IFDI_INTR_ENABLE(ctx);
2189 	txq = ctx->ifc_txqs;
2190 	for (i = 0; i < sctx->isc_ntxqsets; i++, txq++)
2191 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq,
2192 			txq->ift_timer.c_cpu);
2193 }
2194 
2195 static int
2196 iflib_media_change(if_t ifp)
2197 {
2198 	if_ctx_t ctx = if_getsoftc(ifp);
2199 	int err;
2200 
2201 	CTX_LOCK(ctx);
2202 	if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0)
2203 		iflib_init_locked(ctx);
2204 	CTX_UNLOCK(ctx);
2205 	return (err);
2206 }
2207 
2208 static void
2209 iflib_media_status(if_t ifp, struct ifmediareq *ifmr)
2210 {
2211 	if_ctx_t ctx = if_getsoftc(ifp);
2212 
2213 	CTX_LOCK(ctx);
2214 	IFDI_UPDATE_ADMIN_STATUS(ctx);
2215 	IFDI_MEDIA_STATUS(ctx, ifmr);
2216 	CTX_UNLOCK(ctx);
2217 }
2218 
2219 static void
2220 iflib_stop(if_ctx_t ctx)
2221 {
2222 	iflib_txq_t txq = ctx->ifc_txqs;
2223 	iflib_rxq_t rxq = ctx->ifc_rxqs;
2224 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2225 	iflib_dma_info_t di;
2226 	iflib_fl_t fl;
2227 	int i, j;
2228 
2229 	/* Tell the stack that the interface is no longer active */
2230 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2231 
2232 	IFDI_INTR_DISABLE(ctx);
2233 	DELAY(1000);
2234 	IFDI_STOP(ctx);
2235 	DELAY(1000);
2236 
2237 	iflib_debug_reset();
2238 	/* Wait for current tx queue users to exit to disarm watchdog timer. */
2239 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) {
2240 		/* make sure all transmitters have completed before proceeding XXX */
2241 
2242 		/* clean any enqueued buffers */
2243 		iflib_ifmp_purge(txq);
2244 		/* Free any existing tx buffers. */
2245 		for (j = 0; j < txq->ift_size; j++) {
2246 			iflib_txsd_free(ctx, txq, j);
2247 		}
2248 		txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0;
2249 		txq->ift_in_use = txq->ift_gen = txq->ift_cidx = txq->ift_pidx = txq->ift_no_desc_avail = 0;
2250 		txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0;
2251 		txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0;
2252 		txq->ift_pullups = 0;
2253 		ifmp_ring_reset_stats(txq->ift_br);
2254 		for (j = 0, di = txq->ift_ifdi; j < ctx->ifc_nhwtxqs; j++, di++)
2255 			bzero((void *)di->idi_vaddr, di->idi_size);
2256 	}
2257 	for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) {
2258 		/* make sure all transmitters have completed before proceeding XXX */
2259 
2260 		for (j = 0, di = txq->ift_ifdi; j < ctx->ifc_nhwrxqs; j++, di++)
2261 			bzero((void *)di->idi_vaddr, di->idi_size);
2262 		/* also resets the free lists pidx/cidx */
2263 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
2264 			iflib_fl_bufs_free(fl);
2265 	}
2266 }
2267 
2268 static inline caddr_t
2269 calc_next_rxd(iflib_fl_t fl, int cidx)
2270 {
2271 	qidx_t size;
2272 	int nrxd;
2273 	caddr_t start, end, cur, next;
2274 
2275 	nrxd = fl->ifl_size;
2276 	size = fl->ifl_rxd_size;
2277 	start = fl->ifl_ifdi->idi_vaddr;
2278 
2279 	if (__predict_false(size == 0))
2280 		return (start);
2281 	cur = start + size*cidx;
2282 	end = start + size*nrxd;
2283 	next = CACHE_PTR_NEXT(cur);
2284 	return (next < end ? next : start);
2285 }
2286 
2287 static inline void
2288 prefetch_pkts(iflib_fl_t fl, int cidx)
2289 {
2290 	int nextptr;
2291 	int nrxd = fl->ifl_size;
2292 	caddr_t next_rxd;
2293 
2294 
2295 	nextptr = (cidx + CACHE_PTR_INCREMENT) & (nrxd-1);
2296 	prefetch(&fl->ifl_sds.ifsd_m[nextptr]);
2297 	prefetch(&fl->ifl_sds.ifsd_cl[nextptr]);
2298 	next_rxd = calc_next_rxd(fl, cidx);
2299 	prefetch(next_rxd);
2300 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 1) & (nrxd-1)]);
2301 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 2) & (nrxd-1)]);
2302 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 3) & (nrxd-1)]);
2303 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 4) & (nrxd-1)]);
2304 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 1) & (nrxd-1)]);
2305 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 2) & (nrxd-1)]);
2306 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 3) & (nrxd-1)]);
2307 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 4) & (nrxd-1)]);
2308 }
2309 
2310 static void
2311 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, int unload, if_rxsd_t sd)
2312 {
2313 	int flid, cidx;
2314 	bus_dmamap_t map;
2315 	iflib_fl_t fl;
2316 	iflib_dma_info_t di;
2317 	int next;
2318 
2319 	map = NULL;
2320 	flid = irf->irf_flid;
2321 	cidx = irf->irf_idx;
2322 	fl = &rxq->ifr_fl[flid];
2323 	sd->ifsd_fl = fl;
2324 	sd->ifsd_cidx = cidx;
2325 	sd->ifsd_m = &fl->ifl_sds.ifsd_m[cidx];
2326 	sd->ifsd_cl = &fl->ifl_sds.ifsd_cl[cidx];
2327 	fl->ifl_credits--;
2328 #if MEMORY_LOGGING
2329 	fl->ifl_m_dequeued++;
2330 #endif
2331 	if (rxq->ifr_ctx->ifc_flags & IFC_PREFETCH)
2332 		prefetch_pkts(fl, cidx);
2333 	if (fl->ifl_sds.ifsd_map != NULL) {
2334 		next = (cidx + CACHE_PTR_INCREMENT) & (fl->ifl_size-1);
2335 		prefetch(&fl->ifl_sds.ifsd_map[next]);
2336 		map = fl->ifl_sds.ifsd_map[cidx];
2337 		di = fl->ifl_ifdi;
2338 		next = (cidx + CACHE_LINE_SIZE) & (fl->ifl_size-1);
2339 		prefetch(&fl->ifl_sds.ifsd_flags[next]);
2340 		bus_dmamap_sync(di->idi_tag, di->idi_map,
2341 				BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
2342 
2343 	/* not valid assert if bxe really does SGE from non-contiguous elements */
2344 		MPASS(fl->ifl_cidx == cidx);
2345 		if (unload)
2346 			bus_dmamap_unload(fl->ifl_desc_tag, map);
2347 	}
2348 	fl->ifl_cidx = (fl->ifl_cidx + 1) & (fl->ifl_size-1);
2349 	if (__predict_false(fl->ifl_cidx == 0))
2350 		fl->ifl_gen = 0;
2351 	if (map != NULL)
2352 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2353 			BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2354 }
2355 
2356 static struct mbuf *
2357 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri, if_rxsd_t sd)
2358 {
2359 	int i, padlen , flags;
2360 	struct mbuf *m, *mh, *mt;
2361 	caddr_t cl;
2362 
2363 	i = 0;
2364 	mh = NULL;
2365 	do {
2366 		rxd_frag_to_sd(rxq, &ri->iri_frags[i], TRUE, sd);
2367 
2368 		MPASS(*sd->ifsd_cl != NULL);
2369 		MPASS(*sd->ifsd_m != NULL);
2370 
2371 		/* Don't include zero-length frags */
2372 		if (ri->iri_frags[i].irf_len == 0) {
2373 			/* XXX we can save the cluster here, but not the mbuf */
2374 			m_init(*sd->ifsd_m, M_NOWAIT, MT_DATA, 0);
2375 			m_free(*sd->ifsd_m);
2376 			*sd->ifsd_m = NULL;
2377 			continue;
2378 		}
2379 		m = *sd->ifsd_m;
2380 		*sd->ifsd_m = NULL;
2381 		if (mh == NULL) {
2382 			flags = M_PKTHDR|M_EXT;
2383 			mh = mt = m;
2384 			padlen = ri->iri_pad;
2385 		} else {
2386 			flags = M_EXT;
2387 			mt->m_next = m;
2388 			mt = m;
2389 			/* assuming padding is only on the first fragment */
2390 			padlen = 0;
2391 		}
2392 		cl = *sd->ifsd_cl;
2393 		*sd->ifsd_cl = NULL;
2394 
2395 		/* Can these two be made one ? */
2396 		m_init(m, M_NOWAIT, MT_DATA, flags);
2397 		m_cljset(m, cl, sd->ifsd_fl->ifl_cltype);
2398 		/*
2399 		 * These must follow m_init and m_cljset
2400 		 */
2401 		m->m_data += padlen;
2402 		ri->iri_len -= padlen;
2403 		m->m_len = ri->iri_frags[i].irf_len;
2404 	} while (++i < ri->iri_nfrags);
2405 
2406 	return (mh);
2407 }
2408 
2409 /*
2410  * Process one software descriptor
2411  */
2412 static struct mbuf *
2413 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri)
2414 {
2415 	struct if_rxsd sd;
2416 	struct mbuf *m;
2417 
2418 	/* should I merge this back in now that the two paths are basically duplicated? */
2419 	if (ri->iri_nfrags == 1 &&
2420 	    ri->iri_frags[0].irf_len <= IFLIB_RX_COPY_THRESH) {
2421 		rxd_frag_to_sd(rxq, &ri->iri_frags[0], FALSE, &sd);
2422 		m = *sd.ifsd_m;
2423 		*sd.ifsd_m = NULL;
2424 		m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR);
2425 #ifndef __NO_STRICT_ALIGNMENT
2426 		if (!IP_ALIGNED(m))
2427 			m->m_data += 2;
2428 #endif
2429 		memcpy(m->m_data, *sd.ifsd_cl, ri->iri_len);
2430 		m->m_len = ri->iri_frags[0].irf_len;
2431        } else {
2432 		m = assemble_segments(rxq, ri, &sd);
2433 	}
2434 	m->m_pkthdr.len = ri->iri_len;
2435 	m->m_pkthdr.rcvif = ri->iri_ifp;
2436 	m->m_flags |= ri->iri_flags;
2437 	m->m_pkthdr.ether_vtag = ri->iri_vtag;
2438 	m->m_pkthdr.flowid = ri->iri_flowid;
2439 	M_HASHTYPE_SET(m, ri->iri_rsstype);
2440 	m->m_pkthdr.csum_flags = ri->iri_csum_flags;
2441 	m->m_pkthdr.csum_data = ri->iri_csum_data;
2442 	return (m);
2443 }
2444 
2445 static bool
2446 iflib_rxeof(iflib_rxq_t rxq, qidx_t budget)
2447 {
2448 	if_ctx_t ctx = rxq->ifr_ctx;
2449 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2450 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2451 	int avail, i;
2452 	qidx_t *cidxp;
2453 	struct if_rxd_info ri;
2454 	int err, budget_left, rx_bytes, rx_pkts;
2455 	iflib_fl_t fl;
2456 	struct ifnet *ifp;
2457 	int lro_enabled;
2458 
2459 	/*
2460 	 * XXX early demux data packets so that if_input processing only handles
2461 	 * acks in interrupt context
2462 	 */
2463 	struct mbuf *m, *mh, *mt;
2464 
2465 	ifp = ctx->ifc_ifp;
2466 #ifdef DEV_NETMAP
2467 	if (ifp->if_capenable & IFCAP_NETMAP) {
2468 		u_int work = 0;
2469 		if (netmap_rx_irq(ifp, rxq->ifr_id, &work))
2470 			return (FALSE);
2471 	}
2472 #endif
2473 
2474 	mh = mt = NULL;
2475 	MPASS(budget > 0);
2476 	rx_pkts	= rx_bytes = 0;
2477 	if (sctx->isc_flags & IFLIB_HAS_RXCQ)
2478 		cidxp = &rxq->ifr_cq_cidx;
2479 	else
2480 		cidxp = &rxq->ifr_fl[0].ifl_cidx;
2481 	if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) {
2482 		for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2483 			__iflib_fl_refill_lt(ctx, fl, budget + 8);
2484 		DBG_COUNTER_INC(rx_unavail);
2485 		return (false);
2486 	}
2487 
2488 	for (budget_left = budget; (budget_left > 0) && (avail > 0); budget_left--, avail--) {
2489 		if (__predict_false(!CTX_ACTIVE(ctx))) {
2490 			DBG_COUNTER_INC(rx_ctx_inactive);
2491 			break;
2492 		}
2493 		/*
2494 		 * Reset client set fields to their default values
2495 		 */
2496 		rxd_info_zero(&ri);
2497 		ri.iri_qsidx = rxq->ifr_id;
2498 		ri.iri_cidx = *cidxp;
2499 		ri.iri_ifp = ifp;
2500 		ri.iri_frags = rxq->ifr_frags;
2501 		err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
2502 
2503 		if (err)
2504 			goto err;
2505 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
2506 			*cidxp = ri.iri_cidx;
2507 			/* Update our consumer index */
2508 			/* XXX NB: shurd - check if this is still safe */
2509 			while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0]) {
2510 				rxq->ifr_cq_cidx -= scctx->isc_nrxd[0];
2511 				rxq->ifr_cq_gen = 0;
2512 			}
2513 			/* was this only a completion queue message? */
2514 			if (__predict_false(ri.iri_nfrags == 0))
2515 				continue;
2516 		}
2517 		MPASS(ri.iri_nfrags != 0);
2518 		MPASS(ri.iri_len != 0);
2519 
2520 		/* will advance the cidx on the corresponding free lists */
2521 		m = iflib_rxd_pkt_get(rxq, &ri);
2522 		if (avail == 0 && budget_left)
2523 			avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left);
2524 
2525 		if (__predict_false(m == NULL)) {
2526 			DBG_COUNTER_INC(rx_mbuf_null);
2527 			continue;
2528 		}
2529 		/* imm_pkt: -- cxgb */
2530 		if (mh == NULL)
2531 			mh = mt = m;
2532 		else {
2533 			mt->m_nextpkt = m;
2534 			mt = m;
2535 		}
2536 	}
2537 	/* make sure that we can refill faster than drain */
2538 	for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2539 		__iflib_fl_refill_lt(ctx, fl, budget + 8);
2540 
2541 	lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO);
2542 	while (mh != NULL) {
2543 		m = mh;
2544 		mh = mh->m_nextpkt;
2545 		m->m_nextpkt = NULL;
2546 #ifndef __NO_STRICT_ALIGNMENT
2547 		if (!IP_ALIGNED(m) && (m = iflib_fixup_rx(m)) == NULL)
2548 			continue;
2549 #endif
2550 		rx_bytes += m->m_pkthdr.len;
2551 		rx_pkts++;
2552 #if defined(INET6) || defined(INET)
2553 		if (lro_enabled && tcp_lro_rx(&rxq->ifr_lc, m, 0) == 0)
2554 			continue;
2555 #endif
2556 		DBG_COUNTER_INC(rx_if_input);
2557 		ifp->if_input(ifp, m);
2558 	}
2559 
2560 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
2561 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
2562 
2563 	/*
2564 	 * Flush any outstanding LRO work
2565 	 */
2566 #if defined(INET6) || defined(INET)
2567 	tcp_lro_flush_all(&rxq->ifr_lc);
2568 #endif
2569 	if (avail)
2570 		return true;
2571 	return (iflib_rxd_avail(ctx, rxq, *cidxp, 1));
2572 err:
2573 	CTX_LOCK(ctx);
2574 	ctx->ifc_flags |= IFC_DO_RESET;
2575 	iflib_admin_intr_deferred(ctx);
2576 	CTX_UNLOCK(ctx);
2577 	return (false);
2578 }
2579 
2580 #define TXD_NOTIFY_COUNT(txq) (((txq)->ift_size / (txq)->ift_update_freq)-1)
2581 static inline qidx_t
2582 txq_max_db_deferred(iflib_txq_t txq, qidx_t in_use)
2583 {
2584 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2585 	qidx_t minthresh = txq->ift_size / 8;
2586 	if (in_use > 4*minthresh)
2587 		return (notify_count);
2588 	if (in_use > 2*minthresh)
2589 		return (notify_count >> 1);
2590 	if (in_use > minthresh)
2591 		return (notify_count >> 3);
2592 	return (0);
2593 }
2594 
2595 static inline qidx_t
2596 txq_max_rs_deferred(iflib_txq_t txq)
2597 {
2598 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2599 	qidx_t minthresh = txq->ift_size / 8;
2600 	if (txq->ift_in_use > 4*minthresh)
2601 		return (notify_count);
2602 	if (txq->ift_in_use > 2*minthresh)
2603 		return (notify_count >> 1);
2604 	if (txq->ift_in_use > minthresh)
2605 		return (notify_count >> 2);
2606 	return (2);
2607 }
2608 
2609 #define M_CSUM_FLAGS(m) ((m)->m_pkthdr.csum_flags)
2610 #define M_HAS_VLANTAG(m) (m->m_flags & M_VLANTAG)
2611 
2612 #define TXQ_MAX_DB_DEFERRED(txq, in_use) txq_max_db_deferred((txq), (in_use))
2613 #define TXQ_MAX_RS_DEFERRED(txq) txq_max_rs_deferred(txq)
2614 #define TXQ_MAX_DB_CONSUMED(size) (size >> 4)
2615 
2616 /* forward compatibility for cxgb */
2617 #define FIRST_QSET(ctx) 0
2618 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets)
2619 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets)
2620 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx))
2621 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments))
2622 
2623 /* XXX we should be setting this to something other than zero */
2624 #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh)
2625 #define MAX_TX_DESC(ctx) ((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max)
2626 
2627 static inline bool
2628 iflib_txd_db_check(if_ctx_t ctx, iflib_txq_t txq, int ring, qidx_t in_use)
2629 {
2630 	qidx_t dbval, max;
2631 	bool rang;
2632 
2633 	rang = false;
2634 	max = TXQ_MAX_DB_DEFERRED(txq, in_use);
2635 	if (ring || txq->ift_db_pending >= max) {
2636 		dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx;
2637 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval);
2638 		txq->ift_db_pending = txq->ift_npending = 0;
2639 		rang = true;
2640 	}
2641 	return (rang);
2642 }
2643 
2644 #ifdef PKT_DEBUG
2645 static void
2646 print_pkt(if_pkt_info_t pi)
2647 {
2648 	printf("pi len:  %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n",
2649 	       pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx);
2650 	printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n",
2651 	       pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag);
2652 	printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n",
2653 	       pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto);
2654 }
2655 #endif
2656 
2657 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO)
2658 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO)
2659 
2660 static int
2661 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp)
2662 {
2663 	if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx;
2664 	struct ether_vlan_header *eh;
2665 	struct mbuf *m, *n;
2666 
2667 	n = m = *mp;
2668 	if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) &&
2669 	    M_WRITABLE(m) == 0) {
2670 		if ((m = m_dup(m, M_NOWAIT)) == NULL) {
2671 			return (ENOMEM);
2672 		} else {
2673 			m_freem(*mp);
2674 			n = *mp = m;
2675 		}
2676 	}
2677 
2678 	/*
2679 	 * Determine where frame payload starts.
2680 	 * Jump over vlan headers if already present,
2681 	 * helpful for QinQ too.
2682 	 */
2683 	if (__predict_false(m->m_len < sizeof(*eh))) {
2684 		txq->ift_pullups++;
2685 		if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL))
2686 			return (ENOMEM);
2687 	}
2688 	eh = mtod(m, struct ether_vlan_header *);
2689 	if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
2690 		pi->ipi_etype = ntohs(eh->evl_proto);
2691 		pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
2692 	} else {
2693 		pi->ipi_etype = ntohs(eh->evl_encap_proto);
2694 		pi->ipi_ehdrlen = ETHER_HDR_LEN;
2695 	}
2696 
2697 	switch (pi->ipi_etype) {
2698 #ifdef INET
2699 	case ETHERTYPE_IP:
2700 	{
2701 		struct ip *ip = NULL;
2702 		struct tcphdr *th = NULL;
2703 		int minthlen;
2704 
2705 		minthlen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip) + sizeof(*th));
2706 		if (__predict_false(m->m_len < minthlen)) {
2707 			/*
2708 			 * if this code bloat is causing too much of a hit
2709 			 * move it to a separate function and mark it noinline
2710 			 */
2711 			if (m->m_len == pi->ipi_ehdrlen) {
2712 				n = m->m_next;
2713 				MPASS(n);
2714 				if (n->m_len >= sizeof(*ip))  {
2715 					ip = (struct ip *)n->m_data;
2716 					if (n->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2717 						th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2718 				} else {
2719 					txq->ift_pullups++;
2720 					if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
2721 						return (ENOMEM);
2722 					ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2723 				}
2724 			} else {
2725 				txq->ift_pullups++;
2726 				if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
2727 					return (ENOMEM);
2728 				ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2729 				if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2730 					th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2731 			}
2732 		} else {
2733 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2734 			if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2735 				th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2736 		}
2737 		pi->ipi_ip_hlen = ip->ip_hl << 2;
2738 		pi->ipi_ipproto = ip->ip_p;
2739 		pi->ipi_flags |= IPI_TX_IPV4;
2740 
2741 		if (pi->ipi_csum_flags & CSUM_IP)
2742                        ip->ip_sum = 0;
2743 
2744 		if (pi->ipi_ipproto == IPPROTO_TCP) {
2745 			if (__predict_false(th == NULL)) {
2746 				txq->ift_pullups++;
2747 				if (__predict_false((m = m_pullup(m, (ip->ip_hl << 2) + sizeof(*th))) == NULL))
2748 					return (ENOMEM);
2749 				th = (struct tcphdr *)((caddr_t)ip + pi->ipi_ip_hlen);
2750 			}
2751 			pi->ipi_tcp_hflags = th->th_flags;
2752 			pi->ipi_tcp_hlen = th->th_off << 2;
2753 			pi->ipi_tcp_seq = th->th_seq;
2754 		}
2755 		if (IS_TSO4(pi)) {
2756 			if (__predict_false(ip->ip_p != IPPROTO_TCP))
2757 				return (ENXIO);
2758 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
2759 					       ip->ip_dst.s_addr, htons(IPPROTO_TCP));
2760 			pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
2761 			if (sctx->isc_flags & IFLIB_TSO_INIT_IP) {
2762 				ip->ip_sum = 0;
2763 				ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz);
2764 			}
2765 		}
2766 		break;
2767 	}
2768 #endif
2769 #ifdef INET6
2770 	case ETHERTYPE_IPV6:
2771 	{
2772 		struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
2773 		struct tcphdr *th;
2774 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
2775 
2776 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
2777 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
2778 				return (ENOMEM);
2779 		}
2780 		th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
2781 
2782 		/* XXX-BZ this will go badly in case of ext hdrs. */
2783 		pi->ipi_ipproto = ip6->ip6_nxt;
2784 		pi->ipi_flags |= IPI_TX_IPV6;
2785 
2786 		if (pi->ipi_ipproto == IPPROTO_TCP) {
2787 			if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) {
2788 				if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL))
2789 					return (ENOMEM);
2790 			}
2791 			pi->ipi_tcp_hflags = th->th_flags;
2792 			pi->ipi_tcp_hlen = th->th_off << 2;
2793 		}
2794 		if (IS_TSO6(pi)) {
2795 
2796 			if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP))
2797 				return (ENXIO);
2798 			/*
2799 			 * The corresponding flag is set by the stack in the IPv4
2800 			 * TSO case, but not in IPv6 (at least in FreeBSD 10.2).
2801 			 * So, set it here because the rest of the flow requires it.
2802 			 */
2803 			pi->ipi_csum_flags |= CSUM_TCP_IPV6;
2804 			th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0);
2805 			pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
2806 		}
2807 		break;
2808 	}
2809 #endif
2810 	default:
2811 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
2812 		pi->ipi_ip_hlen = 0;
2813 		break;
2814 	}
2815 	*mp = m;
2816 
2817 	return (0);
2818 }
2819 
2820 static  __noinline  struct mbuf *
2821 collapse_pkthdr(struct mbuf *m0)
2822 {
2823 	struct mbuf *m, *m_next, *tmp;
2824 
2825 	m = m0;
2826 	m_next = m->m_next;
2827 	while (m_next != NULL && m_next->m_len == 0) {
2828 		m = m_next;
2829 		m->m_next = NULL;
2830 		m_free(m);
2831 		m_next = m_next->m_next;
2832 	}
2833 	m = m0;
2834 	m->m_next = m_next;
2835 	if ((m_next->m_flags & M_EXT) == 0) {
2836 		m = m_defrag(m, M_NOWAIT);
2837 	} else {
2838 		tmp = m_next->m_next;
2839 		memcpy(m_next, m, MPKTHSIZE);
2840 		m = m_next;
2841 		m->m_next = tmp;
2842 	}
2843 	return (m);
2844 }
2845 
2846 /*
2847  * If dodgy hardware rejects the scatter gather chain we've handed it
2848  * we'll need to remove the mbuf chain from ifsg_m[] before we can add the
2849  * m_defrag'd mbufs
2850  */
2851 static __noinline struct mbuf *
2852 iflib_remove_mbuf(iflib_txq_t txq)
2853 {
2854 	int ntxd, i, pidx;
2855 	struct mbuf *m, *mh, **ifsd_m;
2856 
2857 	pidx = txq->ift_pidx;
2858 	ifsd_m = txq->ift_sds.ifsd_m;
2859 	ntxd = txq->ift_size;
2860 	mh = m = ifsd_m[pidx];
2861 	ifsd_m[pidx] = NULL;
2862 #if MEMORY_LOGGING
2863 	txq->ift_dequeued++;
2864 #endif
2865 	i = 1;
2866 
2867 	while (m) {
2868 		ifsd_m[(pidx + i) & (ntxd -1)] = NULL;
2869 #if MEMORY_LOGGING
2870 		txq->ift_dequeued++;
2871 #endif
2872 		m = m->m_next;
2873 		i++;
2874 	}
2875 	return (mh);
2876 }
2877 
2878 static int
2879 iflib_busdma_load_mbuf_sg(iflib_txq_t txq, bus_dma_tag_t tag, bus_dmamap_t map,
2880 			  struct mbuf **m0, bus_dma_segment_t *segs, int *nsegs,
2881 			  int max_segs, int flags)
2882 {
2883 	if_ctx_t ctx;
2884 	if_shared_ctx_t		sctx;
2885 	if_softc_ctx_t		scctx;
2886 	int i, next, pidx, mask, err, maxsegsz, ntxd, count;
2887 	struct mbuf *m, *tmp, **ifsd_m, **mp;
2888 
2889 	m = *m0;
2890 
2891 	/*
2892 	 * Please don't ever do this
2893 	 */
2894 	if (__predict_false(m->m_len == 0))
2895 		*m0 = m = collapse_pkthdr(m);
2896 
2897 	ctx = txq->ift_ctx;
2898 	sctx = ctx->ifc_sctx;
2899 	scctx = &ctx->ifc_softc_ctx;
2900 	ifsd_m = txq->ift_sds.ifsd_m;
2901 	ntxd = txq->ift_size;
2902 	pidx = txq->ift_pidx;
2903 	if (map != NULL) {
2904 		uint8_t *ifsd_flags = txq->ift_sds.ifsd_flags;
2905 
2906 		err = bus_dmamap_load_mbuf_sg(tag, map,
2907 					      *m0, segs, nsegs, BUS_DMA_NOWAIT);
2908 		if (err)
2909 			return (err);
2910 		ifsd_flags[pidx] |= TX_SW_DESC_MAPPED;
2911 		i = 0;
2912 		next = pidx;
2913 		mask = (txq->ift_size-1);
2914 		m = *m0;
2915 		do {
2916 			mp = &ifsd_m[next];
2917 			*mp = m;
2918 			m = m->m_next;
2919 			if (__predict_false((*mp)->m_len == 0)) {
2920 				m_free(*mp);
2921 				*mp = NULL;
2922 			} else
2923 				next = (pidx + i) & (ntxd-1);
2924 		} while (m != NULL);
2925 	} else {
2926 		int buflen, sgsize, max_sgsize;
2927 		vm_offset_t vaddr;
2928 		vm_paddr_t curaddr;
2929 
2930 		count = i = 0;
2931 		maxsegsz = sctx->isc_tx_maxsize;
2932 		m = *m0;
2933 		do {
2934 			if (__predict_false(m->m_len <= 0)) {
2935 				tmp = m;
2936 				m = m->m_next;
2937 				tmp->m_next = NULL;
2938 				m_free(tmp);
2939 				continue;
2940 			}
2941 			buflen = m->m_len;
2942 			vaddr = (vm_offset_t)m->m_data;
2943 			/*
2944 			 * see if we can't be smarter about physically
2945 			 * contiguous mappings
2946 			 */
2947 			next = (pidx + count) & (ntxd-1);
2948 			MPASS(ifsd_m[next] == NULL);
2949 #if MEMORY_LOGGING
2950 			txq->ift_enqueued++;
2951 #endif
2952 			ifsd_m[next] = m;
2953 			while (buflen > 0) {
2954 				max_sgsize = MIN(buflen, maxsegsz);
2955 				curaddr = pmap_kextract(vaddr);
2956 				sgsize = PAGE_SIZE - (curaddr & PAGE_MASK);
2957 				sgsize = MIN(sgsize, max_sgsize);
2958 				segs[i].ds_addr = curaddr;
2959 				segs[i].ds_len = sgsize;
2960 				vaddr += sgsize;
2961 				buflen -= sgsize;
2962 				i++;
2963 				if (i >= max_segs)
2964 					goto err;
2965 			}
2966 			count++;
2967 			tmp = m;
2968 			m = m->m_next;
2969 		} while (m != NULL);
2970 		*nsegs = i;
2971 	}
2972 	return (0);
2973 err:
2974 	*m0 = iflib_remove_mbuf(txq);
2975 	return (EFBIG);
2976 }
2977 
2978 static inline caddr_t
2979 calc_next_txd(iflib_txq_t txq, int cidx, uint8_t qid)
2980 {
2981 	qidx_t size;
2982 	int ntxd;
2983 	caddr_t start, end, cur, next;
2984 
2985 	ntxd = txq->ift_size;
2986 	size = txq->ift_txd_size[qid];
2987 	start = txq->ift_ifdi[qid].idi_vaddr;
2988 
2989 	if (__predict_false(size == 0))
2990 		return (start);
2991 	cur = start + size*cidx;
2992 	end = start + size*ntxd;
2993 	next = CACHE_PTR_NEXT(cur);
2994 	return (next < end ? next : start);
2995 }
2996 
2997 static int
2998 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp)
2999 {
3000 	if_ctx_t		ctx;
3001 	if_shared_ctx_t		sctx;
3002 	if_softc_ctx_t		scctx;
3003 	bus_dma_segment_t	*segs;
3004 	struct mbuf		*m_head;
3005 	void			*next_txd;
3006 	bus_dmamap_t		map;
3007 	struct if_pkt_info	pi;
3008 	int remap = 0;
3009 	int err, nsegs, ndesc, max_segs, pidx, cidx, next, ntxd;
3010 	bus_dma_tag_t desc_tag;
3011 
3012 	segs = txq->ift_segs;
3013 	ctx = txq->ift_ctx;
3014 	sctx = ctx->ifc_sctx;
3015 	scctx = &ctx->ifc_softc_ctx;
3016 	segs = txq->ift_segs;
3017 	ntxd = txq->ift_size;
3018 	m_head = *m_headp;
3019 	map = NULL;
3020 
3021 	/*
3022 	 * If we're doing TSO the next descriptor to clean may be quite far ahead
3023 	 */
3024 	cidx = txq->ift_cidx;
3025 	pidx = txq->ift_pidx;
3026 	if (ctx->ifc_flags & IFC_PREFETCH) {
3027 		next = (cidx + CACHE_PTR_INCREMENT) & (ntxd-1);
3028 		if (!(ctx->ifc_flags & IFLIB_HAS_TXCQ)) {
3029 			next_txd = calc_next_txd(txq, cidx, 0);
3030 			prefetch(next_txd);
3031 		}
3032 
3033 		/* prefetch the next cache line of mbuf pointers and flags */
3034 		prefetch(&txq->ift_sds.ifsd_m[next]);
3035 		if (txq->ift_sds.ifsd_map != NULL) {
3036 			prefetch(&txq->ift_sds.ifsd_map[next]);
3037 			next = (cidx + CACHE_LINE_SIZE) & (ntxd-1);
3038 			prefetch(&txq->ift_sds.ifsd_flags[next]);
3039 		}
3040 	} else if (txq->ift_sds.ifsd_map != NULL)
3041 		map = txq->ift_sds.ifsd_map[pidx];
3042 
3043 	if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
3044 		desc_tag = txq->ift_tso_desc_tag;
3045 		max_segs = scctx->isc_tx_tso_segments_max;
3046 	} else {
3047 		desc_tag = txq->ift_desc_tag;
3048 		max_segs = scctx->isc_tx_nsegments;
3049 	}
3050 	m_head = *m_headp;
3051 
3052 	pkt_info_zero(&pi);
3053 	pi.ipi_len = m_head->m_pkthdr.len;
3054 	pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG|M_BCAST|M_MCAST));
3055 	pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags;
3056 	pi.ipi_vtag = (m_head->m_flags & M_VLANTAG) ? m_head->m_pkthdr.ether_vtag : 0;
3057 	pi.ipi_pidx = pidx;
3058 	pi.ipi_qsidx = txq->ift_id;
3059 
3060 	/* deliberate bitwise OR to make one condition */
3061 	if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) {
3062 		if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0))
3063 			return (err);
3064 		m_head = *m_headp;
3065 	}
3066 
3067 retry:
3068 	err = iflib_busdma_load_mbuf_sg(txq, desc_tag, map, m_headp, segs, &nsegs, max_segs, BUS_DMA_NOWAIT);
3069 defrag:
3070 	if (__predict_false(err)) {
3071 		switch (err) {
3072 		case EFBIG:
3073 			/* try collapse once and defrag once */
3074 			if (remap == 0)
3075 				m_head = m_collapse(*m_headp, M_NOWAIT, max_segs);
3076 			if (remap == 1)
3077 				m_head = m_defrag(*m_headp, M_NOWAIT);
3078 			remap++;
3079 			if (__predict_false(m_head == NULL))
3080 				goto defrag_failed;
3081 			txq->ift_mbuf_defrag++;
3082 			*m_headp = m_head;
3083 			goto retry;
3084 			break;
3085 		case ENOMEM:
3086 			txq->ift_no_tx_dma_setup++;
3087 			break;
3088 		default:
3089 			txq->ift_no_tx_dma_setup++;
3090 			m_freem(*m_headp);
3091 			DBG_COUNTER_INC(tx_frees);
3092 			*m_headp = NULL;
3093 			break;
3094 		}
3095 		txq->ift_map_failed++;
3096 		DBG_COUNTER_INC(encap_load_mbuf_fail);
3097 		return (err);
3098 	}
3099 
3100 	/*
3101 	 * XXX assumes a 1 to 1 relationship between segments and
3102 	 *        descriptors - this does not hold true on all drivers, e.g.
3103 	 *        cxgb
3104 	 */
3105 	if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) {
3106 		txq->ift_no_desc_avail++;
3107 		if (map != NULL)
3108 			bus_dmamap_unload(desc_tag, map);
3109 		DBG_COUNTER_INC(encap_txq_avail_fail);
3110 		if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0)
3111 			GROUPTASK_ENQUEUE(&txq->ift_task);
3112 		return (ENOBUFS);
3113 	}
3114 	/*
3115 	 * On Intel cards we can greatly reduce the number of TX interrupts
3116 	 * we see by only setting report status on every Nth descriptor.
3117 	 * However, this also means that the driver will need to keep track
3118 	 * of the descriptors that RS was set on to check them for the DD bit.
3119 	 */
3120 	txq->ift_rs_pending += nsegs + 1;
3121 	if (txq->ift_rs_pending > TXQ_MAX_RS_DEFERRED(txq) ||
3122 	     iflib_no_tx_batch || (TXQ_AVAIL(txq) - nsegs - 1) <= MAX_TX_DESC(ctx)) {
3123 		pi.ipi_flags |= IPI_TX_INTR;
3124 		txq->ift_rs_pending = 0;
3125 	}
3126 
3127 	pi.ipi_segs = segs;
3128 	pi.ipi_nsegs = nsegs;
3129 
3130 	MPASS(pidx >= 0 && pidx < txq->ift_size);
3131 #ifdef PKT_DEBUG
3132 	print_pkt(&pi);
3133 #endif
3134 	if (map != NULL)
3135 		bus_dmamap_sync(desc_tag, map, BUS_DMASYNC_PREWRITE);
3136 	if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) {
3137 		if (map != NULL)
3138 			bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3139 					BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
3140 		DBG_COUNTER_INC(tx_encap);
3141 		MPASS(pi.ipi_new_pidx < txq->ift_size);
3142 
3143 		ndesc = pi.ipi_new_pidx - pi.ipi_pidx;
3144 		if (pi.ipi_new_pidx < pi.ipi_pidx) {
3145 			ndesc += txq->ift_size;
3146 			txq->ift_gen = 1;
3147 		}
3148 		/*
3149 		 * drivers can need as many as
3150 		 * two sentinels
3151 		 */
3152 		MPASS(ndesc <= pi.ipi_nsegs + 2);
3153 		MPASS(pi.ipi_new_pidx != pidx);
3154 		MPASS(ndesc > 0);
3155 		txq->ift_in_use += ndesc;
3156 
3157 		/*
3158 		 * We update the last software descriptor again here because there may
3159 		 * be a sentinel and/or there may be more mbufs than segments
3160 		 */
3161 		txq->ift_pidx = pi.ipi_new_pidx;
3162 		txq->ift_npending += pi.ipi_ndescs;
3163 	} else if (__predict_false(err == EFBIG && remap < 2)) {
3164 		*m_headp = m_head = iflib_remove_mbuf(txq);
3165 		remap = 1;
3166 		txq->ift_txd_encap_efbig++;
3167 		goto defrag;
3168 	} else
3169 		DBG_COUNTER_INC(encap_txd_encap_fail);
3170 	return (err);
3171 
3172 defrag_failed:
3173 	txq->ift_mbuf_defrag_failed++;
3174 	txq->ift_map_failed++;
3175 	m_freem(*m_headp);
3176 	DBG_COUNTER_INC(tx_frees);
3177 	*m_headp = NULL;
3178 	return (ENOMEM);
3179 }
3180 
3181 static void
3182 iflib_tx_desc_free(iflib_txq_t txq, int n)
3183 {
3184 	int hasmap;
3185 	uint32_t qsize, cidx, mask, gen;
3186 	struct mbuf *m, **ifsd_m;
3187 	uint8_t *ifsd_flags;
3188 	bus_dmamap_t *ifsd_map;
3189 	bool do_prefetch;
3190 
3191 	cidx = txq->ift_cidx;
3192 	gen = txq->ift_gen;
3193 	qsize = txq->ift_size;
3194 	mask = qsize-1;
3195 	hasmap = txq->ift_sds.ifsd_map != NULL;
3196 	ifsd_flags = txq->ift_sds.ifsd_flags;
3197 	ifsd_m = txq->ift_sds.ifsd_m;
3198 	ifsd_map = txq->ift_sds.ifsd_map;
3199 	do_prefetch = (txq->ift_ctx->ifc_flags & IFC_PREFETCH);
3200 
3201 	while (n--) {
3202 		if (do_prefetch) {
3203 			prefetch(ifsd_m[(cidx + 3) & mask]);
3204 			prefetch(ifsd_m[(cidx + 4) & mask]);
3205 		}
3206 		if (ifsd_m[cidx] != NULL) {
3207 			prefetch(&ifsd_m[(cidx + CACHE_PTR_INCREMENT) & mask]);
3208 			prefetch(&ifsd_flags[(cidx + CACHE_PTR_INCREMENT) & mask]);
3209 			if (hasmap && (ifsd_flags[cidx] & TX_SW_DESC_MAPPED)) {
3210 				/*
3211 				 * does it matter if it's not the TSO tag? If so we'll
3212 				 * have to add the type to flags
3213 				 */
3214 				bus_dmamap_unload(txq->ift_desc_tag, ifsd_map[cidx]);
3215 				ifsd_flags[cidx] &= ~TX_SW_DESC_MAPPED;
3216 			}
3217 			if ((m = ifsd_m[cidx]) != NULL) {
3218 				/* XXX we don't support any drivers that batch packets yet */
3219 				MPASS(m->m_nextpkt == NULL);
3220 
3221 				m_free(m);
3222 				ifsd_m[cidx] = NULL;
3223 #if MEMORY_LOGGING
3224 				txq->ift_dequeued++;
3225 #endif
3226 				DBG_COUNTER_INC(tx_frees);
3227 			}
3228 		}
3229 		if (__predict_false(++cidx == qsize)) {
3230 			cidx = 0;
3231 			gen = 0;
3232 		}
3233 	}
3234 	txq->ift_cidx = cidx;
3235 	txq->ift_gen = gen;
3236 }
3237 
3238 static __inline int
3239 iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh)
3240 {
3241 	int reclaim;
3242 	if_ctx_t ctx = txq->ift_ctx;
3243 
3244 	KASSERT(thresh >= 0, ("invalid threshold to reclaim"));
3245 	MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size);
3246 
3247 	/*
3248 	 * Need a rate-limiting check so that this isn't called every time
3249 	 */
3250 	iflib_tx_credits_update(ctx, txq);
3251 	reclaim = DESC_RECLAIMABLE(txq);
3252 
3253 	if (reclaim <= thresh /* + MAX_TX_DESC(txq->ift_ctx) */) {
3254 #ifdef INVARIANTS
3255 		if (iflib_verbose_debug) {
3256 			printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __FUNCTION__,
3257 			       txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments,
3258 			       reclaim, thresh);
3259 
3260 		}
3261 #endif
3262 		return (0);
3263 	}
3264 	iflib_tx_desc_free(txq, reclaim);
3265 	txq->ift_cleaned += reclaim;
3266 	txq->ift_in_use -= reclaim;
3267 
3268 	return (reclaim);
3269 }
3270 
3271 static struct mbuf **
3272 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset, int remaining)
3273 {
3274 	int next, size;
3275 	struct mbuf **items;
3276 
3277 	size = r->size;
3278 	next = (cidx + CACHE_PTR_INCREMENT) & (size-1);
3279 	items = __DEVOLATILE(struct mbuf **, &r->items[0]);
3280 
3281 	prefetch(items[(cidx + offset) & (size-1)]);
3282 	if (remaining > 1) {
3283 		prefetch(&items[next]);
3284 		prefetch(items[(cidx + offset + 1) & (size-1)]);
3285 		prefetch(items[(cidx + offset + 2) & (size-1)]);
3286 		prefetch(items[(cidx + offset + 3) & (size-1)]);
3287 	}
3288 	return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (size-1)]));
3289 }
3290 
3291 static void
3292 iflib_txq_check_drain(iflib_txq_t txq, int budget)
3293 {
3294 
3295 	ifmp_ring_check_drainage(txq->ift_br, budget);
3296 }
3297 
3298 static uint32_t
3299 iflib_txq_can_drain(struct ifmp_ring *r)
3300 {
3301 	iflib_txq_t txq = r->cookie;
3302 	if_ctx_t ctx = txq->ift_ctx;
3303 
3304 	return ((TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2) ||
3305 		ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false));
3306 }
3307 
3308 static uint32_t
3309 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3310 {
3311 	iflib_txq_t txq = r->cookie;
3312 	if_ctx_t ctx = txq->ift_ctx;
3313 	struct ifnet *ifp = ctx->ifc_ifp;
3314 	struct mbuf **mp, *m;
3315 	int i, count, consumed, pkt_sent, bytes_sent, mcast_sent, avail;
3316 	int reclaimed, err, in_use_prev, desc_used;
3317 	bool do_prefetch, ring, rang;
3318 
3319 	if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) ||
3320 			    !LINK_ACTIVE(ctx))) {
3321 		DBG_COUNTER_INC(txq_drain_notready);
3322 		return (0);
3323 	}
3324 	reclaimed = iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
3325 	rang = iflib_txd_db_check(ctx, txq, reclaimed, txq->ift_in_use);
3326 	avail = IDXDIFF(pidx, cidx, r->size);
3327 	if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) {
3328 		DBG_COUNTER_INC(txq_drain_flushing);
3329 		for (i = 0; i < avail; i++) {
3330 			m_free(r->items[(cidx + i) & (r->size-1)]);
3331 			r->items[(cidx + i) & (r->size-1)] = NULL;
3332 		}
3333 		return (avail);
3334 	}
3335 
3336 	if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3337 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3338 		CALLOUT_LOCK(txq);
3339 		callout_stop(&txq->ift_timer);
3340 		CALLOUT_UNLOCK(txq);
3341 		DBG_COUNTER_INC(txq_drain_oactive);
3342 		return (0);
3343 	}
3344 	if (reclaimed)
3345 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3346 	consumed = mcast_sent = bytes_sent = pkt_sent = 0;
3347 	count = MIN(avail, TX_BATCH_SIZE);
3348 #ifdef INVARIANTS
3349 	if (iflib_verbose_debug)
3350 		printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __FUNCTION__,
3351 		       avail, ctx->ifc_flags, TXQ_AVAIL(txq));
3352 #endif
3353 	do_prefetch = (ctx->ifc_flags & IFC_PREFETCH);
3354 	avail = TXQ_AVAIL(txq);
3355 	for (desc_used = i = 0; i < count && avail > MAX_TX_DESC(ctx) + 2; i++) {
3356 		int pidx_prev, rem = do_prefetch ? count - i : 0;
3357 
3358 		mp = _ring_peek_one(r, cidx, i, rem);
3359 		MPASS(mp != NULL && *mp != NULL);
3360 		if (__predict_false(*mp == (struct mbuf *)txq)) {
3361 			consumed++;
3362 			reclaimed++;
3363 			continue;
3364 		}
3365 		in_use_prev = txq->ift_in_use;
3366 		pidx_prev = txq->ift_pidx;
3367 		err = iflib_encap(txq, mp);
3368 		if (__predict_false(err)) {
3369 			DBG_COUNTER_INC(txq_drain_encapfail);
3370 			/* no room - bail out */
3371 			if (err == ENOBUFS)
3372 				break;
3373 			consumed++;
3374 			DBG_COUNTER_INC(txq_drain_encapfail);
3375 			/* we can't send this packet - skip it */
3376 			continue;
3377 		}
3378 		consumed++;
3379 		pkt_sent++;
3380 		m = *mp;
3381 		DBG_COUNTER_INC(tx_sent);
3382 		bytes_sent += m->m_pkthdr.len;
3383 		mcast_sent += !!(m->m_flags & M_MCAST);
3384 		avail = TXQ_AVAIL(txq);
3385 
3386 		txq->ift_db_pending += (txq->ift_in_use - in_use_prev);
3387 		desc_used += (txq->ift_in_use - in_use_prev);
3388 		ETHER_BPF_MTAP(ifp, m);
3389 		if (__predict_false(!(ifp->if_drv_flags & IFF_DRV_RUNNING)))
3390 			break;
3391 		rang = iflib_txd_db_check(ctx, txq, false, in_use_prev);
3392 	}
3393 
3394 	/* deliberate use of bitwise or to avoid gratuitous short-circuit */
3395 	ring = rang ? false  : (iflib_min_tx_latency | err) || (TXQ_AVAIL(txq) < MAX_TX_DESC(ctx));
3396 	iflib_txd_db_check(ctx, txq, ring, txq->ift_in_use);
3397 	if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
3398 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
3399 	if (mcast_sent)
3400 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
3401 #ifdef INVARIANTS
3402 	if (iflib_verbose_debug)
3403 		printf("consumed=%d\n", consumed);
3404 #endif
3405 	return (consumed);
3406 }
3407 
3408 static uint32_t
3409 iflib_txq_drain_always(struct ifmp_ring *r)
3410 {
3411 	return (1);
3412 }
3413 
3414 static uint32_t
3415 iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3416 {
3417 	int i, avail;
3418 	struct mbuf **mp;
3419 	iflib_txq_t txq;
3420 
3421 	txq = r->cookie;
3422 
3423 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3424 	CALLOUT_LOCK(txq);
3425 	callout_stop(&txq->ift_timer);
3426 	CALLOUT_UNLOCK(txq);
3427 
3428 	avail = IDXDIFF(pidx, cidx, r->size);
3429 	for (i = 0; i < avail; i++) {
3430 		mp = _ring_peek_one(r, cidx, i, avail - i);
3431 		if (__predict_false(*mp == (struct mbuf *)txq))
3432 			continue;
3433 		m_freem(*mp);
3434 	}
3435 	MPASS(ifmp_ring_is_stalled(r) == 0);
3436 	return (avail);
3437 }
3438 
3439 static void
3440 iflib_ifmp_purge(iflib_txq_t txq)
3441 {
3442 	struct ifmp_ring *r;
3443 
3444 	r = txq->ift_br;
3445 	r->drain = iflib_txq_drain_free;
3446 	r->can_drain = iflib_txq_drain_always;
3447 
3448 	ifmp_ring_check_drainage(r, r->size);
3449 
3450 	r->drain = iflib_txq_drain;
3451 	r->can_drain = iflib_txq_can_drain;
3452 }
3453 
3454 static void
3455 _task_fn_tx(void *context)
3456 {
3457 	iflib_txq_t txq = context;
3458 	if_ctx_t ctx = txq->ift_ctx;
3459 	struct ifnet *ifp = ctx->ifc_ifp;
3460 	int rc;
3461 
3462 #ifdef IFLIB_DIAGNOSTICS
3463 	txq->ift_cpu_exec_count[curcpu]++;
3464 #endif
3465 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3466 		return;
3467 	if ((ifp->if_capenable & IFCAP_NETMAP)) {
3468 		if (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false))
3469 			netmap_tx_irq(ifp, txq->ift_id);
3470 		IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3471 		return;
3472 	}
3473 	if (txq->ift_db_pending)
3474 		ifmp_ring_enqueue(txq->ift_br, (void **)&txq, 1, TX_BATCH_SIZE);
3475 	else
3476 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3477 	if (ctx->ifc_flags & IFC_LEGACY)
3478 		IFDI_INTR_ENABLE(ctx);
3479 	else {
3480 		rc = IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3481 		KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver"));
3482 	}
3483 }
3484 
3485 static void
3486 _task_fn_rx(void *context)
3487 {
3488 	iflib_rxq_t rxq = context;
3489 	if_ctx_t ctx = rxq->ifr_ctx;
3490 	bool more;
3491 	int rc;
3492 
3493 #ifdef IFLIB_DIAGNOSTICS
3494 	rxq->ifr_cpu_exec_count[curcpu]++;
3495 #endif
3496 	DBG_COUNTER_INC(task_fn_rxs);
3497 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3498 		return;
3499 	if ((more = iflib_rxeof(rxq, 16 /* XXX */)) == false) {
3500 		if (ctx->ifc_flags & IFC_LEGACY)
3501 			IFDI_INTR_ENABLE(ctx);
3502 		else {
3503 			DBG_COUNTER_INC(rx_intr_enables);
3504 			rc = IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
3505 			KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver"));
3506 		}
3507 	}
3508 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3509 		return;
3510 	if (more)
3511 		GROUPTASK_ENQUEUE(&rxq->ifr_task);
3512 }
3513 
3514 static void
3515 _task_fn_admin(void *context)
3516 {
3517 	if_ctx_t ctx = context;
3518 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
3519 	iflib_txq_t txq;
3520 	int i;
3521 
3522 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)) {
3523 		if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3524 			return;
3525 		}
3526 	}
3527 
3528 	CTX_LOCK(ctx);
3529 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3530 		CALLOUT_LOCK(txq);
3531 		callout_stop(&txq->ift_timer);
3532 		CALLOUT_UNLOCK(txq);
3533 	}
3534 	IFDI_UPDATE_ADMIN_STATUS(ctx);
3535 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3536 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq, txq->ift_timer.c_cpu);
3537 	IFDI_LINK_INTR_ENABLE(ctx);
3538 	if (ctx->ifc_flags & IFC_DO_RESET) {
3539 		ctx->ifc_flags &= ~IFC_DO_RESET;
3540 		iflib_if_init_locked(ctx);
3541 	}
3542 	CTX_UNLOCK(ctx);
3543 
3544 	if (LINK_ACTIVE(ctx) == 0)
3545 		return;
3546 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3547 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
3548 }
3549 
3550 
3551 static void
3552 _task_fn_iov(void *context)
3553 {
3554 	if_ctx_t ctx = context;
3555 
3556 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3557 		return;
3558 
3559 	CTX_LOCK(ctx);
3560 	IFDI_VFLR_HANDLE(ctx);
3561 	CTX_UNLOCK(ctx);
3562 }
3563 
3564 static int
3565 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)
3566 {
3567 	int err;
3568 	if_int_delay_info_t info;
3569 	if_ctx_t ctx;
3570 
3571 	info = (if_int_delay_info_t)arg1;
3572 	ctx = info->iidi_ctx;
3573 	info->iidi_req = req;
3574 	info->iidi_oidp = oidp;
3575 	CTX_LOCK(ctx);
3576 	err = IFDI_SYSCTL_INT_DELAY(ctx, info);
3577 	CTX_UNLOCK(ctx);
3578 	return (err);
3579 }
3580 
3581 /*********************************************************************
3582  *
3583  *  IFNET FUNCTIONS
3584  *
3585  **********************************************************************/
3586 
3587 static void
3588 iflib_if_init_locked(if_ctx_t ctx)
3589 {
3590 	iflib_stop(ctx);
3591 	iflib_init_locked(ctx);
3592 }
3593 
3594 
3595 static void
3596 iflib_if_init(void *arg)
3597 {
3598 	if_ctx_t ctx = arg;
3599 
3600 	CTX_LOCK(ctx);
3601 	iflib_if_init_locked(ctx);
3602 	CTX_UNLOCK(ctx);
3603 }
3604 
3605 static int
3606 iflib_if_transmit(if_t ifp, struct mbuf *m)
3607 {
3608 	if_ctx_t	ctx = if_getsoftc(ifp);
3609 
3610 	iflib_txq_t txq;
3611 	int err, qidx;
3612 
3613 	if (__predict_false((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) {
3614 		DBG_COUNTER_INC(tx_frees);
3615 		m_freem(m);
3616 		return (ENOBUFS);
3617 	}
3618 
3619 	MPASS(m->m_nextpkt == NULL);
3620 	qidx = 0;
3621 	if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m))
3622 		qidx = QIDX(ctx, m);
3623 	/*
3624 	 * XXX calculate buf_ring based on flowid (divvy up bits?)
3625 	 */
3626 	txq = &ctx->ifc_txqs[qidx];
3627 
3628 #ifdef DRIVER_BACKPRESSURE
3629 	if (txq->ift_closed) {
3630 		while (m != NULL) {
3631 			next = m->m_nextpkt;
3632 			m->m_nextpkt = NULL;
3633 			m_freem(m);
3634 			m = next;
3635 		}
3636 		return (ENOBUFS);
3637 	}
3638 #endif
3639 #ifdef notyet
3640 	qidx = count = 0;
3641 	mp = marr;
3642 	next = m;
3643 	do {
3644 		count++;
3645 		next = next->m_nextpkt;
3646 	} while (next != NULL);
3647 
3648 	if (count > nitems(marr))
3649 		if ((mp = malloc(count*sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) {
3650 			/* XXX check nextpkt */
3651 			m_freem(m);
3652 			/* XXX simplify for now */
3653 			DBG_COUNTER_INC(tx_frees);
3654 			return (ENOBUFS);
3655 		}
3656 	for (next = m, i = 0; next != NULL; i++) {
3657 		mp[i] = next;
3658 		next = next->m_nextpkt;
3659 		mp[i]->m_nextpkt = NULL;
3660 	}
3661 #endif
3662 	DBG_COUNTER_INC(tx_seen);
3663 	err = ifmp_ring_enqueue(txq->ift_br, (void **)&m, 1, TX_BATCH_SIZE);
3664 
3665 	if (err) {
3666 		GROUPTASK_ENQUEUE(&txq->ift_task);
3667 		/* support forthcoming later */
3668 #ifdef DRIVER_BACKPRESSURE
3669 		txq->ift_closed = TRUE;
3670 #endif
3671 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3672 		m_freem(m);
3673 	} else if (TXQ_AVAIL(txq) < (txq->ift_size >> 1)) {
3674 		GROUPTASK_ENQUEUE(&txq->ift_task);
3675 	}
3676 
3677 	return (err);
3678 }
3679 
3680 static void
3681 iflib_if_qflush(if_t ifp)
3682 {
3683 	if_ctx_t ctx = if_getsoftc(ifp);
3684 	iflib_txq_t txq = ctx->ifc_txqs;
3685 	int i;
3686 
3687 	CTX_LOCK(ctx);
3688 	ctx->ifc_flags |= IFC_QFLUSH;
3689 	CTX_UNLOCK(ctx);
3690 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
3691 		while (!(ifmp_ring_is_idle(txq->ift_br) || ifmp_ring_is_stalled(txq->ift_br)))
3692 			iflib_txq_check_drain(txq, 0);
3693 	CTX_LOCK(ctx);
3694 	ctx->ifc_flags &= ~IFC_QFLUSH;
3695 	CTX_UNLOCK(ctx);
3696 
3697 	if_qflush(ifp);
3698 }
3699 
3700 
3701 #define IFCAP_FLAGS (IFCAP_TXCSUM_IPV6 | IFCAP_RXCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \
3702 		     IFCAP_TSO4 | IFCAP_TSO6 | IFCAP_VLAN_HWTAGGING |	\
3703 		     IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | IFCAP_VLAN_HWTSO)
3704 
3705 static int
3706 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data)
3707 {
3708 	if_ctx_t ctx = if_getsoftc(ifp);
3709 	struct ifreq	*ifr = (struct ifreq *)data;
3710 #if defined(INET) || defined(INET6)
3711 	struct ifaddr	*ifa = (struct ifaddr *)data;
3712 #endif
3713 	bool		avoid_reset = FALSE;
3714 	int		err = 0, reinit = 0, bits;
3715 
3716 	switch (command) {
3717 	case SIOCSIFADDR:
3718 #ifdef INET
3719 		if (ifa->ifa_addr->sa_family == AF_INET)
3720 			avoid_reset = TRUE;
3721 #endif
3722 #ifdef INET6
3723 		if (ifa->ifa_addr->sa_family == AF_INET6)
3724 			avoid_reset = TRUE;
3725 #endif
3726 		/*
3727 		** Calling init results in link renegotiation,
3728 		** so we avoid doing it when possible.
3729 		*/
3730 		if (avoid_reset) {
3731 			if_setflagbits(ifp, IFF_UP,0);
3732 			if (!(if_getdrvflags(ifp)& IFF_DRV_RUNNING))
3733 				reinit = 1;
3734 #ifdef INET
3735 			if (!(if_getflags(ifp) & IFF_NOARP))
3736 				arp_ifinit(ifp, ifa);
3737 #endif
3738 		} else
3739 			err = ether_ioctl(ifp, command, data);
3740 		break;
3741 	case SIOCSIFMTU:
3742 		CTX_LOCK(ctx);
3743 		if (ifr->ifr_mtu == if_getmtu(ifp)) {
3744 			CTX_UNLOCK(ctx);
3745 			break;
3746 		}
3747 		bits = if_getdrvflags(ifp);
3748 		/* stop the driver and free any clusters before proceeding */
3749 		iflib_stop(ctx);
3750 
3751 		if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) {
3752 			if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size)
3753 				ctx->ifc_flags |= IFC_MULTISEG;
3754 			else
3755 				ctx->ifc_flags &= ~IFC_MULTISEG;
3756 			err = if_setmtu(ifp, ifr->ifr_mtu);
3757 		}
3758 		iflib_init_locked(ctx);
3759 		if_setdrvflags(ifp, bits);
3760 		CTX_UNLOCK(ctx);
3761 		break;
3762 	case SIOCSIFFLAGS:
3763 		CTX_LOCK(ctx);
3764 		if (if_getflags(ifp) & IFF_UP) {
3765 			if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
3766 				if ((if_getflags(ifp) ^ ctx->ifc_if_flags) &
3767 				    (IFF_PROMISC | IFF_ALLMULTI)) {
3768 					err = IFDI_PROMISC_SET(ctx, if_getflags(ifp));
3769 				}
3770 			} else
3771 				reinit = 1;
3772 		} else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
3773 			iflib_stop(ctx);
3774 		}
3775 		ctx->ifc_if_flags = if_getflags(ifp);
3776 		CTX_UNLOCK(ctx);
3777 		break;
3778 	case SIOCADDMULTI:
3779 	case SIOCDELMULTI:
3780 		if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
3781 			CTX_LOCK(ctx);
3782 			IFDI_INTR_DISABLE(ctx);
3783 			IFDI_MULTI_SET(ctx);
3784 			IFDI_INTR_ENABLE(ctx);
3785 			CTX_UNLOCK(ctx);
3786 		}
3787 		break;
3788 	case SIOCSIFMEDIA:
3789 		CTX_LOCK(ctx);
3790 		IFDI_MEDIA_SET(ctx);
3791 		CTX_UNLOCK(ctx);
3792 		/* falls thru */
3793 	case SIOCGIFMEDIA:
3794 		err = ifmedia_ioctl(ifp, ifr, &ctx->ifc_media, command);
3795 		break;
3796 	case SIOCGI2C:
3797 	{
3798 		struct ifi2creq i2c;
3799 
3800 		err = copyin(ifr->ifr_data, &i2c, sizeof(i2c));
3801 		if (err != 0)
3802 			break;
3803 		if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) {
3804 			err = EINVAL;
3805 			break;
3806 		}
3807 		if (i2c.len > sizeof(i2c.data)) {
3808 			err = EINVAL;
3809 			break;
3810 		}
3811 
3812 		if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0)
3813 			err = copyout(&i2c, ifr->ifr_data, sizeof(i2c));
3814 		break;
3815 	}
3816 	case SIOCSIFCAP:
3817 	{
3818 		int mask, setmask;
3819 
3820 		mask = ifr->ifr_reqcap ^ if_getcapenable(ifp);
3821 		setmask = 0;
3822 #ifdef TCP_OFFLOAD
3823 		setmask |= mask & (IFCAP_TOE4|IFCAP_TOE6);
3824 #endif
3825 		setmask |= (mask & IFCAP_FLAGS);
3826 
3827 		if (setmask  & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6))
3828 			setmask |= (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6);
3829 		if ((mask & IFCAP_WOL) &&
3830 		    (if_getcapabilities(ifp) & IFCAP_WOL) != 0)
3831 			setmask |= (mask & (IFCAP_WOL_MCAST|IFCAP_WOL_MAGIC));
3832 		if_vlancap(ifp);
3833 		/*
3834 		 * want to ensure that traffic has stopped before we change any of the flags
3835 		 */
3836 		if (setmask) {
3837 			CTX_LOCK(ctx);
3838 			bits = if_getdrvflags(ifp);
3839 			if (bits & IFF_DRV_RUNNING)
3840 				iflib_stop(ctx);
3841 			if_togglecapenable(ifp, setmask);
3842 			if (bits & IFF_DRV_RUNNING)
3843 				iflib_init_locked(ctx);
3844 			if_setdrvflags(ifp, bits);
3845 			CTX_UNLOCK(ctx);
3846 		}
3847 		break;
3848 	    }
3849 	case SIOCGPRIVATE_0:
3850 	case SIOCSDRVSPEC:
3851 	case SIOCGDRVSPEC:
3852 		CTX_LOCK(ctx);
3853 		err = IFDI_PRIV_IOCTL(ctx, command, data);
3854 		CTX_UNLOCK(ctx);
3855 		break;
3856 	default:
3857 		err = ether_ioctl(ifp, command, data);
3858 		break;
3859 	}
3860 	if (reinit)
3861 		iflib_if_init(ctx);
3862 	return (err);
3863 }
3864 
3865 static uint64_t
3866 iflib_if_get_counter(if_t ifp, ift_counter cnt)
3867 {
3868 	if_ctx_t ctx = if_getsoftc(ifp);
3869 
3870 	return (IFDI_GET_COUNTER(ctx, cnt));
3871 }
3872 
3873 /*********************************************************************
3874  *
3875  *  OTHER FUNCTIONS EXPORTED TO THE STACK
3876  *
3877  **********************************************************************/
3878 
3879 static void
3880 iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag)
3881 {
3882 	if_ctx_t ctx = if_getsoftc(ifp);
3883 
3884 	if ((void *)ctx != arg)
3885 		return;
3886 
3887 	if ((vtag == 0) || (vtag > 4095))
3888 		return;
3889 
3890 	CTX_LOCK(ctx);
3891 	IFDI_VLAN_REGISTER(ctx, vtag);
3892 	/* Re-init to load the changes */
3893 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
3894 		iflib_init_locked(ctx);
3895 	CTX_UNLOCK(ctx);
3896 }
3897 
3898 static void
3899 iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag)
3900 {
3901 	if_ctx_t ctx = if_getsoftc(ifp);
3902 
3903 	if ((void *)ctx != arg)
3904 		return;
3905 
3906 	if ((vtag == 0) || (vtag > 4095))
3907 		return;
3908 
3909 	CTX_LOCK(ctx);
3910 	IFDI_VLAN_UNREGISTER(ctx, vtag);
3911 	/* Re-init to load the changes */
3912 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
3913 		iflib_init_locked(ctx);
3914 	CTX_UNLOCK(ctx);
3915 }
3916 
3917 static void
3918 iflib_led_func(void *arg, int onoff)
3919 {
3920 	if_ctx_t ctx = arg;
3921 
3922 	CTX_LOCK(ctx);
3923 	IFDI_LED_FUNC(ctx, onoff);
3924 	CTX_UNLOCK(ctx);
3925 }
3926 
3927 /*********************************************************************
3928  *
3929  *  BUS FUNCTION DEFINITIONS
3930  *
3931  **********************************************************************/
3932 
3933 int
3934 iflib_device_probe(device_t dev)
3935 {
3936 	pci_vendor_info_t *ent;
3937 
3938 	uint16_t	pci_vendor_id, pci_device_id;
3939 	uint16_t	pci_subvendor_id, pci_subdevice_id;
3940 	uint16_t	pci_rev_id;
3941 	if_shared_ctx_t sctx;
3942 
3943 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
3944 		return (ENOTSUP);
3945 
3946 	pci_vendor_id = pci_get_vendor(dev);
3947 	pci_device_id = pci_get_device(dev);
3948 	pci_subvendor_id = pci_get_subvendor(dev);
3949 	pci_subdevice_id = pci_get_subdevice(dev);
3950 	pci_rev_id = pci_get_revid(dev);
3951 	if (sctx->isc_parse_devinfo != NULL)
3952 		sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id);
3953 
3954 	ent = sctx->isc_vendor_info;
3955 	while (ent->pvi_vendor_id != 0) {
3956 		if (pci_vendor_id != ent->pvi_vendor_id) {
3957 			ent++;
3958 			continue;
3959 		}
3960 		if ((pci_device_id == ent->pvi_device_id) &&
3961 		    ((pci_subvendor_id == ent->pvi_subvendor_id) ||
3962 		     (ent->pvi_subvendor_id == 0)) &&
3963 		    ((pci_subdevice_id == ent->pvi_subdevice_id) ||
3964 		     (ent->pvi_subdevice_id == 0)) &&
3965 		    ((pci_rev_id == ent->pvi_rev_id) ||
3966 		     (ent->pvi_rev_id == 0))) {
3967 
3968 			device_set_desc_copy(dev, ent->pvi_name);
3969 			/* this needs to be changed to zero if the bus probing code
3970 			 * ever stops re-probing on best match because the sctx
3971 			 * may have its values over written by register calls
3972 			 * in subsequent probes
3973 			 */
3974 			return (BUS_PROBE_DEFAULT);
3975 		}
3976 		ent++;
3977 	}
3978 	return (ENXIO);
3979 }
3980 
3981 int
3982 iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp)
3983 {
3984 	int err, rid, msix, msix_bar;
3985 	if_ctx_t ctx;
3986 	if_t ifp;
3987 	if_softc_ctx_t scctx;
3988 	int i;
3989 	uint16_t main_txq;
3990 	uint16_t main_rxq;
3991 
3992 
3993 	ctx = malloc(sizeof(* ctx), M_IFLIB, M_WAITOK|M_ZERO);
3994 
3995 	if (sc == NULL) {
3996 		sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
3997 		device_set_softc(dev, ctx);
3998 		ctx->ifc_flags |= IFC_SC_ALLOCATED;
3999 	}
4000 
4001 	ctx->ifc_sctx = sctx;
4002 	ctx->ifc_dev = dev;
4003 	ctx->ifc_softc = sc;
4004 
4005 	if ((err = iflib_register(ctx)) != 0) {
4006 		device_printf(dev, "iflib_register failed %d\n", err);
4007 		return (err);
4008 	}
4009 	iflib_add_device_sysctl_pre(ctx);
4010 
4011 	scctx = &ctx->ifc_softc_ctx;
4012 	ifp = ctx->ifc_ifp;
4013 
4014 	/*
4015 	 * XXX sanity check that ntxd & nrxd are a power of 2
4016 	 */
4017 	if (ctx->ifc_sysctl_ntxqs != 0)
4018 		scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs;
4019 	if (ctx->ifc_sysctl_nrxqs != 0)
4020 		scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs;
4021 
4022 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4023 		if (ctx->ifc_sysctl_ntxds[i] != 0)
4024 			scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i];
4025 		else
4026 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4027 	}
4028 
4029 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4030 		if (ctx->ifc_sysctl_nrxds[i] != 0)
4031 			scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i];
4032 		else
4033 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4034 	}
4035 
4036 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4037 		if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) {
4038 			device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n",
4039 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]);
4040 			scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i];
4041 		}
4042 		if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) {
4043 			device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n",
4044 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]);
4045 			scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i];
4046 		}
4047 	}
4048 
4049 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4050 		if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) {
4051 			device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n",
4052 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]);
4053 			scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i];
4054 		}
4055 		if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) {
4056 			device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n",
4057 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]);
4058 			scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i];
4059 		}
4060 	}
4061 
4062 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4063 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4064 		return (err);
4065 	}
4066 	_iflib_pre_assert(scctx);
4067 	ctx->ifc_txrx = *scctx->isc_txrx;
4068 
4069 #ifdef INVARIANTS
4070 	MPASS(scctx->isc_capenable);
4071 	if (scctx->isc_capenable & IFCAP_TXCSUM)
4072 		MPASS(scctx->isc_tx_csum_flags);
4073 #endif
4074 
4075 	if_setcapabilities(ifp, scctx->isc_capenable);
4076 	if_setcapenable(ifp, scctx->isc_capenable);
4077 
4078 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4079 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4080 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4081 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4082 
4083 #ifdef ACPI_DMAR
4084 	if (dmar_get_dma_tag(device_get_parent(dev), dev) != NULL)
4085 		ctx->ifc_flags |= IFC_DMAR;
4086 #elif !(defined(__i386__) || defined(__amd64__))
4087 	/* set unconditionally for !x86 */
4088 	ctx->ifc_flags |= IFC_DMAR;
4089 #endif
4090 
4091 	msix_bar = scctx->isc_msix_bar;
4092 	main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4093 	main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4094 
4095 	/* XXX change for per-queue sizes */
4096 	device_printf(dev, "using %d tx descriptors and %d rx descriptors\n",
4097 		      scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4098 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4099 		if (!powerof2(scctx->isc_nrxd[i])) {
4100 			/* round down instead? */
4101 			device_printf(dev, "# rx descriptors must be a power of 2\n");
4102 			err = EINVAL;
4103 			goto fail;
4104 		}
4105 	}
4106 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4107 		if (!powerof2(scctx->isc_ntxd[i])) {
4108 			device_printf(dev,
4109 			    "# tx descriptors must be a power of 2");
4110 			err = EINVAL;
4111 			goto fail;
4112 		}
4113 	}
4114 
4115 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4116 	    MAX_SINGLE_PACKET_FRACTION)
4117 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4118 		    MAX_SINGLE_PACKET_FRACTION);
4119 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4120 	    MAX_SINGLE_PACKET_FRACTION)
4121 		scctx->isc_tx_tso_segments_max = max(1,
4122 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4123 
4124 	/*
4125 	 * Protect the stack against modern hardware
4126 	 */
4127 	if (scctx->isc_tx_tso_size_max > FREEBSD_TSO_SIZE_MAX)
4128 		scctx->isc_tx_tso_size_max = FREEBSD_TSO_SIZE_MAX;
4129 
4130 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4131 	ifp->if_hw_tsomaxsegcount = scctx->isc_tx_tso_segments_max;
4132 	ifp->if_hw_tsomax = scctx->isc_tx_tso_size_max;
4133 	ifp->if_hw_tsomaxsegsize = scctx->isc_tx_tso_segsize_max;
4134 	if (scctx->isc_rss_table_size == 0)
4135 		scctx->isc_rss_table_size = 64;
4136 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4137 
4138 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4139 	/* XXX format name */
4140 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx, -1, "admin");
4141 	/*
4142 	** Now setup MSI or MSI/X, should
4143 	** return us the number of supported
4144 	** vectors. (Will be 1 for MSI)
4145 	*/
4146 	if (sctx->isc_flags & IFLIB_SKIP_MSIX) {
4147 		msix = scctx->isc_vectors;
4148 	} else if (scctx->isc_msix_bar != 0)
4149 	       /*
4150 		* The simple fact that isc_msix_bar is not 0 does not mean we
4151 		* we have a good value there that is known to work.
4152 		*/
4153 		msix = iflib_msix_init(ctx);
4154 	else {
4155 		scctx->isc_vectors = 1;
4156 		scctx->isc_ntxqsets = 1;
4157 		scctx->isc_nrxqsets = 1;
4158 		scctx->isc_intr = IFLIB_INTR_LEGACY;
4159 		msix = 0;
4160 	}
4161 	/* Get memory for the station queues */
4162 	if ((err = iflib_queues_alloc(ctx))) {
4163 		device_printf(dev, "Unable to allocate queue memory\n");
4164 		goto fail;
4165 	}
4166 
4167 	if ((err = iflib_qset_structures_setup(ctx))) {
4168 		device_printf(dev, "qset structure setup failed %d\n", err);
4169 		goto fail_queues;
4170 	}
4171 
4172 	/*
4173 	 * Group taskqueues aren't properly set up until SMP is started,
4174 	 * so we disable interrupts until we can handle them post
4175 	 * SI_SUB_SMP.
4176 	 *
4177 	 * XXX: disabling interrupts doesn't actually work, at least for
4178 	 * the non-MSI case.  When they occur before SI_SUB_SMP completes,
4179 	 * we do null handling and depend on this not causing too large an
4180 	 * interrupt storm.
4181 	 */
4182 	IFDI_INTR_DISABLE(ctx);
4183 	if (msix > 1 && (err = IFDI_MSIX_INTR_ASSIGN(ctx, msix)) != 0) {
4184 		device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n", err);
4185 		goto fail_intr_free;
4186 	}
4187 	if (msix <= 1) {
4188 		rid = 0;
4189 		if (scctx->isc_intr == IFLIB_INTR_MSI) {
4190 			MPASS(msix == 1);
4191 			rid = 1;
4192 		}
4193 		if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) {
4194 			device_printf(dev, "iflib_legacy_setup failed %d\n", err);
4195 			goto fail_intr_free;
4196 		}
4197 	}
4198 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac);
4199 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4200 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4201 		goto fail_detach;
4202 	}
4203 	if ((err = iflib_netmap_attach(ctx))) {
4204 		device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err);
4205 		goto fail_detach;
4206 	}
4207 	*ctxp = ctx;
4208 
4209 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4210 	iflib_add_device_sysctl_post(ctx);
4211 	ctx->ifc_flags |= IFC_INIT_DONE;
4212 	return (0);
4213 fail_detach:
4214 	ether_ifdetach(ctx->ifc_ifp);
4215 fail_intr_free:
4216 	if (scctx->isc_intr == IFLIB_INTR_MSIX || scctx->isc_intr == IFLIB_INTR_MSI)
4217 		pci_release_msi(ctx->ifc_dev);
4218 fail_queues:
4219 	/* XXX free queues */
4220 fail:
4221 	IFDI_DETACH(ctx);
4222 	return (err);
4223 }
4224 
4225 int
4226 iflib_device_attach(device_t dev)
4227 {
4228 	if_ctx_t ctx;
4229 	if_shared_ctx_t sctx;
4230 
4231 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4232 		return (ENOTSUP);
4233 
4234 	pci_enable_busmaster(dev);
4235 
4236 	return (iflib_device_register(dev, NULL, sctx, &ctx));
4237 }
4238 
4239 int
4240 iflib_device_deregister(if_ctx_t ctx)
4241 {
4242 	if_t ifp = ctx->ifc_ifp;
4243 	iflib_txq_t txq;
4244 	iflib_rxq_t rxq;
4245 	device_t dev = ctx->ifc_dev;
4246 	int i;
4247 	struct taskqgroup *tqg;
4248 
4249 	/* Make sure VLANS are not using driver */
4250 	if (if_vlantrunkinuse(ifp)) {
4251 		device_printf(dev,"Vlan in use, detach first\n");
4252 		return (EBUSY);
4253 	}
4254 
4255 	CTX_LOCK(ctx);
4256 	ctx->ifc_in_detach = 1;
4257 	iflib_stop(ctx);
4258 	CTX_UNLOCK(ctx);
4259 
4260 	/* Unregister VLAN events */
4261 	if (ctx->ifc_vlan_attach_event != NULL)
4262 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
4263 	if (ctx->ifc_vlan_detach_event != NULL)
4264 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
4265 
4266 	iflib_netmap_detach(ifp);
4267 	ether_ifdetach(ifp);
4268 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
4269 	CTX_LOCK_DESTROY(ctx);
4270 	if (ctx->ifc_led_dev != NULL)
4271 		led_destroy(ctx->ifc_led_dev);
4272 	/* XXX drain any dependent tasks */
4273 	tqg = qgroup_if_io_tqg;
4274 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
4275 		callout_drain(&txq->ift_timer);
4276 		if (txq->ift_task.gt_uniq != NULL)
4277 			taskqgroup_detach(tqg, &txq->ift_task);
4278 	}
4279 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4280 		if (rxq->ifr_task.gt_uniq != NULL)
4281 			taskqgroup_detach(tqg, &rxq->ifr_task);
4282 	}
4283 	tqg = qgroup_if_config_tqg;
4284 	if (ctx->ifc_admin_task.gt_uniq != NULL)
4285 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
4286 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
4287 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
4288 
4289 	IFDI_DETACH(ctx);
4290 	device_set_softc(ctx->ifc_dev, NULL);
4291 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) {
4292 		pci_release_msi(dev);
4293 	}
4294 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) {
4295 		iflib_irq_free(ctx, &ctx->ifc_legacy_irq);
4296 	}
4297 	if (ctx->ifc_msix_mem != NULL) {
4298 		bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY,
4299 			ctx->ifc_softc_ctx.isc_msix_bar, ctx->ifc_msix_mem);
4300 		ctx->ifc_msix_mem = NULL;
4301 	}
4302 
4303 	bus_generic_detach(dev);
4304 	if_free(ifp);
4305 
4306 	iflib_tx_structures_free(ctx);
4307 	iflib_rx_structures_free(ctx);
4308 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4309 		free(ctx->ifc_softc, M_IFLIB);
4310 	free(ctx, M_IFLIB);
4311 	return (0);
4312 }
4313 
4314 
4315 int
4316 iflib_device_detach(device_t dev)
4317 {
4318 	if_ctx_t ctx = device_get_softc(dev);
4319 
4320 	return (iflib_device_deregister(ctx));
4321 }
4322 
4323 int
4324 iflib_device_suspend(device_t dev)
4325 {
4326 	if_ctx_t ctx = device_get_softc(dev);
4327 
4328 	CTX_LOCK(ctx);
4329 	IFDI_SUSPEND(ctx);
4330 	CTX_UNLOCK(ctx);
4331 
4332 	return bus_generic_suspend(dev);
4333 }
4334 int
4335 iflib_device_shutdown(device_t dev)
4336 {
4337 	if_ctx_t ctx = device_get_softc(dev);
4338 
4339 	CTX_LOCK(ctx);
4340 	IFDI_SHUTDOWN(ctx);
4341 	CTX_UNLOCK(ctx);
4342 
4343 	return bus_generic_suspend(dev);
4344 }
4345 
4346 
4347 int
4348 iflib_device_resume(device_t dev)
4349 {
4350 	if_ctx_t ctx = device_get_softc(dev);
4351 	iflib_txq_t txq = ctx->ifc_txqs;
4352 
4353 	CTX_LOCK(ctx);
4354 	IFDI_RESUME(ctx);
4355 	iflib_init_locked(ctx);
4356 	CTX_UNLOCK(ctx);
4357 	for (int i = 0; i < NTXQSETS(ctx); i++, txq++)
4358 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
4359 
4360 	return (bus_generic_resume(dev));
4361 }
4362 
4363 int
4364 iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params)
4365 {
4366 	int error;
4367 	if_ctx_t ctx = device_get_softc(dev);
4368 
4369 	CTX_LOCK(ctx);
4370 	error = IFDI_IOV_INIT(ctx, num_vfs, params);
4371 	CTX_UNLOCK(ctx);
4372 
4373 	return (error);
4374 }
4375 
4376 void
4377 iflib_device_iov_uninit(device_t dev)
4378 {
4379 	if_ctx_t ctx = device_get_softc(dev);
4380 
4381 	CTX_LOCK(ctx);
4382 	IFDI_IOV_UNINIT(ctx);
4383 	CTX_UNLOCK(ctx);
4384 }
4385 
4386 int
4387 iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params)
4388 {
4389 	int error;
4390 	if_ctx_t ctx = device_get_softc(dev);
4391 
4392 	CTX_LOCK(ctx);
4393 	error = IFDI_IOV_VF_ADD(ctx, vfnum, params);
4394 	CTX_UNLOCK(ctx);
4395 
4396 	return (error);
4397 }
4398 
4399 /*********************************************************************
4400  *
4401  *  MODULE FUNCTION DEFINITIONS
4402  *
4403  **********************************************************************/
4404 
4405 /*
4406  * - Start a fast taskqueue thread for each core
4407  * - Start a taskqueue for control operations
4408  */
4409 static int
4410 iflib_module_init(void)
4411 {
4412 	return (0);
4413 }
4414 
4415 static int
4416 iflib_module_event_handler(module_t mod, int what, void *arg)
4417 {
4418 	int err;
4419 
4420 	switch (what) {
4421 	case MOD_LOAD:
4422 		if ((err = iflib_module_init()) != 0)
4423 			return (err);
4424 		break;
4425 	case MOD_UNLOAD:
4426 		return (EBUSY);
4427 	default:
4428 		return (EOPNOTSUPP);
4429 	}
4430 
4431 	return (0);
4432 }
4433 
4434 /*********************************************************************
4435  *
4436  *  PUBLIC FUNCTION DEFINITIONS
4437  *     ordered as in iflib.h
4438  *
4439  **********************************************************************/
4440 
4441 
4442 static void
4443 _iflib_assert(if_shared_ctx_t sctx)
4444 {
4445 	MPASS(sctx->isc_tx_maxsize);
4446 	MPASS(sctx->isc_tx_maxsegsize);
4447 
4448 	MPASS(sctx->isc_rx_maxsize);
4449 	MPASS(sctx->isc_rx_nsegments);
4450 	MPASS(sctx->isc_rx_maxsegsize);
4451 
4452 	MPASS(sctx->isc_nrxd_min[0]);
4453 	MPASS(sctx->isc_nrxd_max[0]);
4454 	MPASS(sctx->isc_nrxd_default[0]);
4455 	MPASS(sctx->isc_ntxd_min[0]);
4456 	MPASS(sctx->isc_ntxd_max[0]);
4457 	MPASS(sctx->isc_ntxd_default[0]);
4458 }
4459 
4460 static void
4461 _iflib_pre_assert(if_softc_ctx_t scctx)
4462 {
4463 
4464 	MPASS(scctx->isc_txrx->ift_txd_encap);
4465 	MPASS(scctx->isc_txrx->ift_txd_flush);
4466 	MPASS(scctx->isc_txrx->ift_txd_credits_update);
4467 	MPASS(scctx->isc_txrx->ift_rxd_available);
4468 	MPASS(scctx->isc_txrx->ift_rxd_pkt_get);
4469 	MPASS(scctx->isc_txrx->ift_rxd_refill);
4470 	MPASS(scctx->isc_txrx->ift_rxd_flush);
4471 }
4472 
4473 static int
4474 iflib_register(if_ctx_t ctx)
4475 {
4476 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4477 	driver_t *driver = sctx->isc_driver;
4478 	device_t dev = ctx->ifc_dev;
4479 	if_t ifp;
4480 
4481 	_iflib_assert(sctx);
4482 
4483 	CTX_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev));
4484 
4485 	ifp = ctx->ifc_ifp = if_gethandle(IFT_ETHER);
4486 	if (ifp == NULL) {
4487 		device_printf(dev, "can not allocate ifnet structure\n");
4488 		return (ENOMEM);
4489 	}
4490 
4491 	/*
4492 	 * Initialize our context's device specific methods
4493 	 */
4494 	kobj_init((kobj_t) ctx, (kobj_class_t) driver);
4495 	kobj_class_compile((kobj_class_t) driver);
4496 	driver->refs++;
4497 
4498 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
4499 	if_setsoftc(ifp, ctx);
4500 	if_setdev(ifp, dev);
4501 	if_setinitfn(ifp, iflib_if_init);
4502 	if_setioctlfn(ifp, iflib_if_ioctl);
4503 	if_settransmitfn(ifp, iflib_if_transmit);
4504 	if_setqflushfn(ifp, iflib_if_qflush);
4505 	if_setflags(ifp, IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST);
4506 
4507 	ctx->ifc_vlan_attach_event =
4508 		EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx,
4509 							  EVENTHANDLER_PRI_FIRST);
4510 	ctx->ifc_vlan_detach_event =
4511 		EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx,
4512 							  EVENTHANDLER_PRI_FIRST);
4513 
4514 	ifmedia_init(&ctx->ifc_media, IFM_IMASK,
4515 					 iflib_media_change, iflib_media_status);
4516 
4517 	return (0);
4518 }
4519 
4520 
4521 static int
4522 iflib_queues_alloc(if_ctx_t ctx)
4523 {
4524 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4525 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4526 	device_t dev = ctx->ifc_dev;
4527 	int nrxqsets = scctx->isc_nrxqsets;
4528 	int ntxqsets = scctx->isc_ntxqsets;
4529 	iflib_txq_t txq;
4530 	iflib_rxq_t rxq;
4531 	iflib_fl_t fl = NULL;
4532 	int i, j, cpu, err, txconf, rxconf;
4533 	iflib_dma_info_t ifdip;
4534 	uint32_t *rxqsizes = scctx->isc_rxqsizes;
4535 	uint32_t *txqsizes = scctx->isc_txqsizes;
4536 	uint8_t nrxqs = sctx->isc_nrxqs;
4537 	uint8_t ntxqs = sctx->isc_ntxqs;
4538 	int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1;
4539 	caddr_t *vaddrs;
4540 	uint64_t *paddrs;
4541 	struct ifmp_ring **brscp;
4542 
4543 	KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1"));
4544 	KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1"));
4545 
4546 	brscp = NULL;
4547 	txq = NULL;
4548 	rxq = NULL;
4549 
4550 /* Allocate the TX ring struct memory */
4551 	if (!(txq =
4552 	    (iflib_txq_t) malloc(sizeof(struct iflib_txq) *
4553 	    ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
4554 		device_printf(dev, "Unable to allocate TX ring memory\n");
4555 		err = ENOMEM;
4556 		goto fail;
4557 	}
4558 
4559 	/* Now allocate the RX */
4560 	if (!(rxq =
4561 	    (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) *
4562 	    nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
4563 		device_printf(dev, "Unable to allocate RX ring memory\n");
4564 		err = ENOMEM;
4565 		goto rx_fail;
4566 	}
4567 
4568 	ctx->ifc_txqs = txq;
4569 	ctx->ifc_rxqs = rxq;
4570 
4571 	/*
4572 	 * XXX handle allocation failure
4573 	 */
4574 	for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) {
4575 		/* Set up some basics */
4576 
4577 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs, M_IFLIB, M_WAITOK|M_ZERO)) == NULL) {
4578 			device_printf(dev, "failed to allocate iflib_dma_info\n");
4579 			err = ENOMEM;
4580 			goto err_tx_desc;
4581 		}
4582 		txq->ift_ifdi = ifdip;
4583 		for (j = 0; j < ntxqs; j++, ifdip++) {
4584 			if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, BUS_DMA_NOWAIT)) {
4585 				device_printf(dev, "Unable to allocate Descriptor memory\n");
4586 				err = ENOMEM;
4587 				goto err_tx_desc;
4588 			}
4589 			txq->ift_txd_size[j] = scctx->isc_txd_size[j];
4590 			bzero((void *)ifdip->idi_vaddr, txqsizes[j]);
4591 		}
4592 		txq->ift_ctx = ctx;
4593 		txq->ift_id = i;
4594 		if (sctx->isc_flags & IFLIB_HAS_TXCQ) {
4595 			txq->ift_br_offset = 1;
4596 		} else {
4597 			txq->ift_br_offset = 0;
4598 		}
4599 		/* XXX fix this */
4600 		txq->ift_timer.c_cpu = cpu;
4601 
4602 		if (iflib_txsd_alloc(txq)) {
4603 			device_printf(dev, "Critical Failure setting up TX buffers\n");
4604 			err = ENOMEM;
4605 			goto err_tx_desc;
4606 		}
4607 
4608 		/* Initialize the TX lock */
4609 		snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:tx(%d):callout",
4610 		    device_get_nameunit(dev), txq->ift_id);
4611 		mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF);
4612 		callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0);
4613 
4614 		snprintf(txq->ift_db_mtx_name, MTX_NAME_LEN, "%s:tx(%d):db",
4615 			 device_get_nameunit(dev), txq->ift_id);
4616 
4617 		err = ifmp_ring_alloc(&txq->ift_br, 2048, txq, iflib_txq_drain,
4618 				      iflib_txq_can_drain, M_IFLIB, M_WAITOK);
4619 		if (err) {
4620 			/* XXX free any allocated rings */
4621 			device_printf(dev, "Unable to allocate buf_ring\n");
4622 			goto err_tx_desc;
4623 		}
4624 	}
4625 
4626 	for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) {
4627 		/* Set up some basics */
4628 
4629 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs, M_IFLIB, M_WAITOK|M_ZERO)) == NULL) {
4630 			device_printf(dev, "failed to allocate iflib_dma_info\n");
4631 			err = ENOMEM;
4632 			goto err_tx_desc;
4633 		}
4634 
4635 		rxq->ifr_ifdi = ifdip;
4636 		/* XXX this needs to be changed if #rx queues != #tx queues */
4637 		rxq->ifr_ntxqirq = 1;
4638 		rxq->ifr_txqid[0] = i;
4639 		for (j = 0; j < nrxqs; j++, ifdip++) {
4640 			if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, BUS_DMA_NOWAIT)) {
4641 				device_printf(dev, "Unable to allocate Descriptor memory\n");
4642 				err = ENOMEM;
4643 				goto err_tx_desc;
4644 			}
4645 			bzero((void *)ifdip->idi_vaddr, rxqsizes[j]);
4646 		}
4647 		rxq->ifr_ctx = ctx;
4648 		rxq->ifr_id = i;
4649 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
4650 			rxq->ifr_fl_offset = 1;
4651 		} else {
4652 			rxq->ifr_fl_offset = 0;
4653 		}
4654 		rxq->ifr_nfl = nfree_lists;
4655 		if (!(fl =
4656 			  (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) {
4657 			device_printf(dev, "Unable to allocate free list memory\n");
4658 			err = ENOMEM;
4659 			goto err_tx_desc;
4660 		}
4661 		rxq->ifr_fl = fl;
4662 		for (j = 0; j < nfree_lists; j++) {
4663 			fl[j].ifl_rxq = rxq;
4664 			fl[j].ifl_id = j;
4665 			fl[j].ifl_ifdi = &rxq->ifr_ifdi[j + rxq->ifr_fl_offset];
4666 			fl[j].ifl_rxd_size = scctx->isc_rxd_size[j];
4667 		}
4668         /* Allocate receive buffers for the ring*/
4669 		if (iflib_rxsd_alloc(rxq)) {
4670 			device_printf(dev,
4671 			    "Critical Failure setting up receive buffers\n");
4672 			err = ENOMEM;
4673 			goto err_rx_desc;
4674 		}
4675 	}
4676 
4677 	/* TXQs */
4678 	vaddrs = malloc(sizeof(caddr_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
4679 	paddrs = malloc(sizeof(uint64_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
4680 	for (i = 0; i < ntxqsets; i++) {
4681 		iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi;
4682 
4683 		for (j = 0; j < ntxqs; j++, di++) {
4684 			vaddrs[i*ntxqs + j] = di->idi_vaddr;
4685 			paddrs[i*ntxqs + j] = di->idi_paddr;
4686 		}
4687 	}
4688 	if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) {
4689 		device_printf(ctx->ifc_dev, "device queue allocation failed\n");
4690 		iflib_tx_structures_free(ctx);
4691 		free(vaddrs, M_IFLIB);
4692 		free(paddrs, M_IFLIB);
4693 		goto err_rx_desc;
4694 	}
4695 	free(vaddrs, M_IFLIB);
4696 	free(paddrs, M_IFLIB);
4697 
4698 	/* RXQs */
4699 	vaddrs = malloc(sizeof(caddr_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
4700 	paddrs = malloc(sizeof(uint64_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
4701 	for (i = 0; i < nrxqsets; i++) {
4702 		iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi;
4703 
4704 		for (j = 0; j < nrxqs; j++, di++) {
4705 			vaddrs[i*nrxqs + j] = di->idi_vaddr;
4706 			paddrs[i*nrxqs + j] = di->idi_paddr;
4707 		}
4708 	}
4709 	if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) {
4710 		device_printf(ctx->ifc_dev, "device queue allocation failed\n");
4711 		iflib_tx_structures_free(ctx);
4712 		free(vaddrs, M_IFLIB);
4713 		free(paddrs, M_IFLIB);
4714 		goto err_rx_desc;
4715 	}
4716 	free(vaddrs, M_IFLIB);
4717 	free(paddrs, M_IFLIB);
4718 
4719 	return (0);
4720 
4721 /* XXX handle allocation failure changes */
4722 err_rx_desc:
4723 err_tx_desc:
4724 	if (ctx->ifc_rxqs != NULL)
4725 		free(ctx->ifc_rxqs, M_IFLIB);
4726 	ctx->ifc_rxqs = NULL;
4727 	if (ctx->ifc_txqs != NULL)
4728 		free(ctx->ifc_txqs, M_IFLIB);
4729 	ctx->ifc_txqs = NULL;
4730 rx_fail:
4731 	if (brscp != NULL)
4732 		free(brscp, M_IFLIB);
4733 	if (rxq != NULL)
4734 		free(rxq, M_IFLIB);
4735 	if (txq != NULL)
4736 		free(txq, M_IFLIB);
4737 fail:
4738 	return (err);
4739 }
4740 
4741 static int
4742 iflib_tx_structures_setup(if_ctx_t ctx)
4743 {
4744 	iflib_txq_t txq = ctx->ifc_txqs;
4745 	int i;
4746 
4747 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
4748 		iflib_txq_setup(txq);
4749 
4750 	return (0);
4751 }
4752 
4753 static void
4754 iflib_tx_structures_free(if_ctx_t ctx)
4755 {
4756 	iflib_txq_t txq = ctx->ifc_txqs;
4757 	int i, j;
4758 
4759 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
4760 		iflib_txq_destroy(txq);
4761 		for (j = 0; j < ctx->ifc_nhwtxqs; j++)
4762 			iflib_dma_free(&txq->ift_ifdi[j]);
4763 	}
4764 	free(ctx->ifc_txqs, M_IFLIB);
4765 	ctx->ifc_txqs = NULL;
4766 	IFDI_QUEUES_FREE(ctx);
4767 }
4768 
4769 /*********************************************************************
4770  *
4771  *  Initialize all receive rings.
4772  *
4773  **********************************************************************/
4774 static int
4775 iflib_rx_structures_setup(if_ctx_t ctx)
4776 {
4777 	iflib_rxq_t rxq = ctx->ifc_rxqs;
4778 	int q;
4779 #if defined(INET6) || defined(INET)
4780 	int i, err;
4781 #endif
4782 
4783 	for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) {
4784 #if defined(INET6) || defined(INET)
4785 		tcp_lro_free(&rxq->ifr_lc);
4786 		if ((err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp,
4787 		    TCP_LRO_ENTRIES, min(1024,
4788 		    ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset]))) != 0) {
4789 			device_printf(ctx->ifc_dev, "LRO Initialization failed!\n");
4790 			goto fail;
4791 		}
4792 		rxq->ifr_lro_enabled = TRUE;
4793 #endif
4794 		IFDI_RXQ_SETUP(ctx, rxq->ifr_id);
4795 	}
4796 	return (0);
4797 #if defined(INET6) || defined(INET)
4798 fail:
4799 	/*
4800 	 * Free RX software descriptors allocated so far, we will only handle
4801 	 * the rings that completed, the failing case will have
4802 	 * cleaned up for itself. 'q' failed, so its the terminus.
4803 	 */
4804 	rxq = ctx->ifc_rxqs;
4805 	for (i = 0; i < q; ++i, rxq++) {
4806 		iflib_rx_sds_free(rxq);
4807 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
4808 	}
4809 	return (err);
4810 #endif
4811 }
4812 
4813 /*********************************************************************
4814  *
4815  *  Free all receive rings.
4816  *
4817  **********************************************************************/
4818 static void
4819 iflib_rx_structures_free(if_ctx_t ctx)
4820 {
4821 	iflib_rxq_t rxq = ctx->ifc_rxqs;
4822 
4823 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) {
4824 		iflib_rx_sds_free(rxq);
4825 	}
4826 }
4827 
4828 static int
4829 iflib_qset_structures_setup(if_ctx_t ctx)
4830 {
4831 	int err;
4832 
4833 	if ((err = iflib_tx_structures_setup(ctx)) != 0)
4834 		return (err);
4835 
4836 	if ((err = iflib_rx_structures_setup(ctx)) != 0) {
4837 		device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err);
4838 		iflib_tx_structures_free(ctx);
4839 		iflib_rx_structures_free(ctx);
4840 	}
4841 	return (err);
4842 }
4843 
4844 int
4845 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
4846 				driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, char *name)
4847 {
4848 
4849 	return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name));
4850 }
4851 
4852 static int
4853 find_nth(if_ctx_t ctx, cpuset_t *cpus, int qid)
4854 {
4855 	int i, cpuid, eqid, count;
4856 
4857 	CPU_COPY(&ctx->ifc_cpus, cpus);
4858 	count = CPU_COUNT(&ctx->ifc_cpus);
4859 	eqid = qid % count;
4860 	/* clear up to the qid'th bit */
4861 	for (i = 0; i < eqid; i++) {
4862 		cpuid = CPU_FFS(cpus);
4863 		MPASS(cpuid != 0);
4864 		CPU_CLR(cpuid-1, cpus);
4865 	}
4866 	cpuid = CPU_FFS(cpus);
4867 	MPASS(cpuid != 0);
4868 	return (cpuid-1);
4869 }
4870 
4871 int
4872 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid,
4873 						iflib_intr_type_t type, driver_filter_t *filter,
4874 						void *filter_arg, int qid, char *name)
4875 {
4876 	struct grouptask *gtask;
4877 	struct taskqgroup *tqg;
4878 	iflib_filter_info_t info;
4879 	cpuset_t cpus;
4880 	gtask_fn_t *fn;
4881 	int tqrid, err, cpuid;
4882 	driver_filter_t *intr_fast;
4883 	void *q;
4884 
4885 	info = &ctx->ifc_filter_info;
4886 	tqrid = rid;
4887 
4888 	switch (type) {
4889 	/* XXX merge tx/rx for netmap? */
4890 	case IFLIB_INTR_TX:
4891 		q = &ctx->ifc_txqs[qid];
4892 		info = &ctx->ifc_txqs[qid].ift_filter_info;
4893 		gtask = &ctx->ifc_txqs[qid].ift_task;
4894 		tqg = qgroup_if_io_tqg;
4895 		fn = _task_fn_tx;
4896 		intr_fast = iflib_fast_intr;
4897 		GROUPTASK_INIT(gtask, 0, fn, q);
4898 		break;
4899 	case IFLIB_INTR_RX:
4900 		q = &ctx->ifc_rxqs[qid];
4901 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
4902 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
4903 		tqg = qgroup_if_io_tqg;
4904 		fn = _task_fn_rx;
4905 		intr_fast = iflib_fast_intr;
4906 		GROUPTASK_INIT(gtask, 0, fn, q);
4907 		break;
4908 	case IFLIB_INTR_RXTX:
4909 		q = &ctx->ifc_rxqs[qid];
4910 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
4911 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
4912 		tqg = qgroup_if_io_tqg;
4913 		fn = _task_fn_rx;
4914 		intr_fast = iflib_fast_intr_rxtx;
4915 		GROUPTASK_INIT(gtask, 0, fn, q);
4916 		break;
4917 	case IFLIB_INTR_ADMIN:
4918 		q = ctx;
4919 		tqrid = -1;
4920 		info = &ctx->ifc_filter_info;
4921 		gtask = &ctx->ifc_admin_task;
4922 		tqg = qgroup_if_config_tqg;
4923 		fn = _task_fn_admin;
4924 		intr_fast = iflib_fast_intr_ctx;
4925 		break;
4926 	default:
4927 		panic("unknown net intr type");
4928 	}
4929 
4930 	info->ifi_filter = filter;
4931 	info->ifi_filter_arg = filter_arg;
4932 	info->ifi_task = gtask;
4933 	info->ifi_ctx = q;
4934 
4935 	err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info,  name);
4936 	if (err != 0) {
4937 		device_printf(ctx->ifc_dev, "_iflib_irq_alloc failed %d\n", err);
4938 		return (err);
4939 	}
4940 	if (type == IFLIB_INTR_ADMIN)
4941 		return (0);
4942 
4943 	if (tqrid != -1) {
4944 		cpuid = find_nth(ctx, &cpus, qid);
4945 		taskqgroup_attach_cpu(tqg, gtask, q, cpuid, irq->ii_rid, name);
4946 	} else {
4947 		taskqgroup_attach(tqg, gtask, q, tqrid, name);
4948 	}
4949 
4950 	return (0);
4951 }
4952 
4953 void
4954 iflib_softirq_alloc_generic(if_ctx_t ctx, int rid, iflib_intr_type_t type,  void *arg, int qid, char *name)
4955 {
4956 	struct grouptask *gtask;
4957 	struct taskqgroup *tqg;
4958 	gtask_fn_t *fn;
4959 	void *q;
4960 
4961 	switch (type) {
4962 	case IFLIB_INTR_TX:
4963 		q = &ctx->ifc_txqs[qid];
4964 		gtask = &ctx->ifc_txqs[qid].ift_task;
4965 		tqg = qgroup_if_io_tqg;
4966 		fn = _task_fn_tx;
4967 		break;
4968 	case IFLIB_INTR_RX:
4969 		q = &ctx->ifc_rxqs[qid];
4970 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
4971 		tqg = qgroup_if_io_tqg;
4972 		fn = _task_fn_rx;
4973 		break;
4974 	case IFLIB_INTR_IOV:
4975 		q = ctx;
4976 		gtask = &ctx->ifc_vflr_task;
4977 		tqg = qgroup_if_config_tqg;
4978 		rid = -1;
4979 		fn = _task_fn_iov;
4980 		break;
4981 	default:
4982 		panic("unknown net intr type");
4983 	}
4984 	GROUPTASK_INIT(gtask, 0, fn, q);
4985 	taskqgroup_attach(tqg, gtask, q, rid, name);
4986 }
4987 
4988 void
4989 iflib_irq_free(if_ctx_t ctx, if_irq_t irq)
4990 {
4991 	if (irq->ii_tag)
4992 		bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag);
4993 
4994 	if (irq->ii_res)
4995 		bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ, irq->ii_rid, irq->ii_res);
4996 }
4997 
4998 static int
4999 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, char *name)
5000 {
5001 	iflib_txq_t txq = ctx->ifc_txqs;
5002 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5003 	if_irq_t irq = &ctx->ifc_legacy_irq;
5004 	iflib_filter_info_t info;
5005 	struct grouptask *gtask;
5006 	struct taskqgroup *tqg;
5007 	gtask_fn_t *fn;
5008 	int tqrid;
5009 	void *q;
5010 	int err;
5011 
5012 	q = &ctx->ifc_rxqs[0];
5013 	info = &rxq[0].ifr_filter_info;
5014 	gtask = &rxq[0].ifr_task;
5015 	tqg = qgroup_if_io_tqg;
5016 	tqrid = irq->ii_rid = *rid;
5017 	fn = _task_fn_rx;
5018 
5019 	ctx->ifc_flags |= IFC_LEGACY;
5020 	info->ifi_filter = filter;
5021 	info->ifi_filter_arg = filter_arg;
5022 	info->ifi_task = gtask;
5023 	info->ifi_ctx = ctx;
5024 
5025 	/* We allocate a single interrupt resource */
5026 	if ((err = _iflib_irq_alloc(ctx, irq, tqrid, iflib_fast_intr_ctx, NULL, info, name)) != 0)
5027 		return (err);
5028 	GROUPTASK_INIT(gtask, 0, fn, q);
5029 	taskqgroup_attach(tqg, gtask, q, tqrid, name);
5030 
5031 	GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq);
5032 	taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, tqrid, "tx");
5033 	return (0);
5034 }
5035 
5036 void
5037 iflib_led_create(if_ctx_t ctx)
5038 {
5039 
5040 	ctx->ifc_led_dev = led_create(iflib_led_func, ctx,
5041 	    device_get_nameunit(ctx->ifc_dev));
5042 }
5043 
5044 void
5045 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid)
5046 {
5047 
5048 	GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
5049 }
5050 
5051 void
5052 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid)
5053 {
5054 
5055 	GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task);
5056 }
5057 
5058 void
5059 iflib_admin_intr_deferred(if_ctx_t ctx)
5060 {
5061 #ifdef INVARIANTS
5062 	struct grouptask *gtask;
5063 
5064 	gtask = &ctx->ifc_admin_task;
5065 	MPASS(gtask->gt_taskqueue != NULL);
5066 #endif
5067 
5068 	GROUPTASK_ENQUEUE(&ctx->ifc_admin_task);
5069 }
5070 
5071 void
5072 iflib_iov_intr_deferred(if_ctx_t ctx)
5073 {
5074 
5075 	GROUPTASK_ENQUEUE(&ctx->ifc_vflr_task);
5076 }
5077 
5078 void
5079 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, char *name)
5080 {
5081 
5082 	taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, -1, name);
5083 }
5084 
5085 void
5086 iflib_config_gtask_init(if_ctx_t ctx, struct grouptask *gtask, gtask_fn_t *fn,
5087 	char *name)
5088 {
5089 
5090 	GROUPTASK_INIT(gtask, 0, fn, ctx);
5091 	taskqgroup_attach(qgroup_if_config_tqg, gtask, gtask, -1, name);
5092 }
5093 
5094 void
5095 iflib_config_gtask_deinit(struct grouptask *gtask)
5096 {
5097 
5098 	taskqgroup_detach(qgroup_if_config_tqg, gtask);
5099 }
5100 
5101 void
5102 iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate)
5103 {
5104 	if_t ifp = ctx->ifc_ifp;
5105 	iflib_txq_t txq = ctx->ifc_txqs;
5106 
5107 	if_setbaudrate(ifp, baudrate);
5108 	if (baudrate >= IF_Gbps(10))
5109 		ctx->ifc_flags |= IFC_PREFETCH;
5110 
5111 	/* If link down, disable watchdog */
5112 	if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) {
5113 		for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++)
5114 			txq->ift_qstatus = IFLIB_QUEUE_IDLE;
5115 	}
5116 	ctx->ifc_link_state = link_state;
5117 	if_link_state_change(ifp, link_state);
5118 }
5119 
5120 static int
5121 iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq)
5122 {
5123 	int credits;
5124 #ifdef INVARIANTS
5125 	int credits_pre = txq->ift_cidx_processed;
5126 #endif
5127 
5128 	if (ctx->isc_txd_credits_update == NULL)
5129 		return (0);
5130 
5131 	if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, true)) == 0)
5132 		return (0);
5133 
5134 	txq->ift_processed += credits;
5135 	txq->ift_cidx_processed += credits;
5136 
5137 	MPASS(credits_pre + credits == txq->ift_cidx_processed);
5138 	if (txq->ift_cidx_processed >= txq->ift_size)
5139 		txq->ift_cidx_processed -= txq->ift_size;
5140 	return (credits);
5141 }
5142 
5143 static int
5144 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget)
5145 {
5146 
5147 	return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx,
5148 	    budget));
5149 }
5150 
5151 void
5152 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name,
5153 	const char *description, if_int_delay_info_t info,
5154 	int offset, int value)
5155 {
5156 	info->iidi_ctx = ctx;
5157 	info->iidi_offset = offset;
5158 	info->iidi_value = value;
5159 	SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev),
5160 	    SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)),
5161 	    OID_AUTO, name, CTLTYPE_INT|CTLFLAG_RW,
5162 	    info, 0, iflib_sysctl_int_delay, "I", description);
5163 }
5164 
5165 struct mtx *
5166 iflib_ctx_lock_get(if_ctx_t ctx)
5167 {
5168 
5169 	return (&ctx->ifc_mtx);
5170 }
5171 
5172 static int
5173 iflib_msix_init(if_ctx_t ctx)
5174 {
5175 	device_t dev = ctx->ifc_dev;
5176 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5177 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5178 	int vectors, queues, rx_queues, tx_queues, queuemsgs, msgs;
5179 	int iflib_num_tx_queues, iflib_num_rx_queues;
5180 	int err, admincnt, bar;
5181 
5182 	iflib_num_tx_queues = scctx->isc_ntxqsets;
5183 	iflib_num_rx_queues = scctx->isc_nrxqsets;
5184 
5185 	device_printf(dev, "msix_init qsets capped at %d\n", iflib_num_tx_queues);
5186 
5187 	bar = ctx->ifc_softc_ctx.isc_msix_bar;
5188 	admincnt = sctx->isc_admin_intrcnt;
5189 	/* Override by tuneable */
5190 	if (enable_msix == 0)
5191 		goto msi;
5192 
5193 	/*
5194 	** When used in a virtualized environment
5195 	** PCI BUSMASTER capability may not be set
5196 	** so explicity set it here and rewrite
5197 	** the ENABLE in the MSIX control register
5198 	** at this point to cause the host to
5199 	** successfully initialize us.
5200 	*/
5201 	{
5202 		int msix_ctrl, rid;
5203 
5204  		pci_enable_busmaster(dev);
5205 		rid = 0;
5206 		if (pci_find_cap(dev, PCIY_MSIX, &rid) == 0 && rid != 0) {
5207 			rid += PCIR_MSIX_CTRL;
5208 			msix_ctrl = pci_read_config(dev, rid, 2);
5209 			msix_ctrl |= PCIM_MSIXCTRL_MSIX_ENABLE;
5210 			pci_write_config(dev, rid, msix_ctrl, 2);
5211 		} else {
5212 			device_printf(dev, "PCIY_MSIX capability not found; "
5213 			                   "or rid %d == 0.\n", rid);
5214 			goto msi;
5215 		}
5216 	}
5217 
5218 	/*
5219 	 * bar == -1 => "trust me I know what I'm doing"
5220 	 * https://www.youtube.com/watch?v=nnwWKkNau4I
5221 	 * Some drivers are for hardware that is so shoddily
5222 	 * documented that no one knows which bars are which
5223 	 * so the developer has to map all bars. This hack
5224 	 * allows shoddy garbage to use msix in this framework.
5225 	 */
5226 	if (bar != -1) {
5227 		ctx->ifc_msix_mem = bus_alloc_resource_any(dev,
5228 	            SYS_RES_MEMORY, &bar, RF_ACTIVE);
5229 		if (ctx->ifc_msix_mem == NULL) {
5230 			/* May not be enabled */
5231 			device_printf(dev, "Unable to map MSIX table \n");
5232 			goto msi;
5233 		}
5234 	}
5235 	/* First try MSI/X */
5236 	if ((msgs = pci_msix_count(dev)) == 0) { /* system has msix disabled */
5237 		device_printf(dev, "System has MSIX disabled \n");
5238 		bus_release_resource(dev, SYS_RES_MEMORY,
5239 		    bar, ctx->ifc_msix_mem);
5240 		ctx->ifc_msix_mem = NULL;
5241 		goto msi;
5242 	}
5243 #if IFLIB_DEBUG
5244 	/* use only 1 qset in debug mode */
5245 	queuemsgs = min(msgs - admincnt, 1);
5246 #else
5247 	queuemsgs = msgs - admincnt;
5248 #endif
5249 	if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) == 0) {
5250 #ifdef RSS
5251 		queues = imin(queuemsgs, rss_getnumbuckets());
5252 #else
5253 		queues = queuemsgs;
5254 #endif
5255 		queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues);
5256 		device_printf(dev, "pxm cpus: %d queue msgs: %d admincnt: %d\n",
5257 					  CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt);
5258 	} else {
5259 		device_printf(dev, "Unable to fetch CPU list\n");
5260 		/* Figure out a reasonable auto config value */
5261 		queues = min(queuemsgs, mp_ncpus);
5262 	}
5263 #ifdef  RSS
5264 	/* If we're doing RSS, clamp at the number of RSS buckets */
5265 	if (queues > rss_getnumbuckets())
5266 		queues = rss_getnumbuckets();
5267 #endif
5268 	if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt)
5269 		rx_queues = iflib_num_rx_queues;
5270 	else
5271 		rx_queues = queues;
5272 	/*
5273 	 * We want this to be all logical CPUs by default
5274 	 */
5275 	if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues)
5276 		tx_queues = iflib_num_tx_queues;
5277 	else
5278 		tx_queues = mp_ncpus;
5279 
5280 	if (ctx->ifc_sysctl_qs_eq_override == 0) {
5281 #ifdef INVARIANTS
5282 		if (tx_queues != rx_queues)
5283 			device_printf(dev, "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n",
5284 				      min(rx_queues, tx_queues), min(rx_queues, tx_queues));
5285 #endif
5286 		tx_queues = min(rx_queues, tx_queues);
5287 		rx_queues = min(rx_queues, tx_queues);
5288 	}
5289 
5290 	device_printf(dev, "using %d rx queues %d tx queues \n", rx_queues, tx_queues);
5291 
5292 	vectors = rx_queues + admincnt;
5293 	if ((err = pci_alloc_msix(dev, &vectors)) == 0) {
5294 		device_printf(dev,
5295 					  "Using MSIX interrupts with %d vectors\n", vectors);
5296 		scctx->isc_vectors = vectors;
5297 		scctx->isc_nrxqsets = rx_queues;
5298 		scctx->isc_ntxqsets = tx_queues;
5299 		scctx->isc_intr = IFLIB_INTR_MSIX;
5300 
5301 		return (vectors);
5302 	} else {
5303 		device_printf(dev, "failed to allocate %d msix vectors, err: %d - using MSI\n", vectors, err);
5304 	}
5305 msi:
5306 	vectors = pci_msi_count(dev);
5307 	scctx->isc_nrxqsets = 1;
5308 	scctx->isc_ntxqsets = 1;
5309 	scctx->isc_vectors = vectors;
5310 	if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) {
5311 		device_printf(dev,"Using an MSI interrupt\n");
5312 		scctx->isc_intr = IFLIB_INTR_MSI;
5313 	} else {
5314 		device_printf(dev,"Using a Legacy interrupt\n");
5315 		scctx->isc_intr = IFLIB_INTR_LEGACY;
5316 	}
5317 
5318 	return (vectors);
5319 }
5320 
5321 char * ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" };
5322 
5323 static int
5324 mp_ring_state_handler(SYSCTL_HANDLER_ARGS)
5325 {
5326 	int rc;
5327 	uint16_t *state = ((uint16_t *)oidp->oid_arg1);
5328 	struct sbuf *sb;
5329 	char *ring_state = "UNKNOWN";
5330 
5331 	/* XXX needed ? */
5332 	rc = sysctl_wire_old_buffer(req, 0);
5333 	MPASS(rc == 0);
5334 	if (rc != 0)
5335 		return (rc);
5336 	sb = sbuf_new_for_sysctl(NULL, NULL, 80, req);
5337 	MPASS(sb != NULL);
5338 	if (sb == NULL)
5339 		return (ENOMEM);
5340 	if (state[3] <= 3)
5341 		ring_state = ring_states[state[3]];
5342 
5343 	sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s",
5344 		    state[0], state[1], state[2], ring_state);
5345 	rc = sbuf_finish(sb);
5346 	sbuf_delete(sb);
5347         return(rc);
5348 }
5349 
5350 enum iflib_ndesc_handler {
5351 	IFLIB_NTXD_HANDLER,
5352 	IFLIB_NRXD_HANDLER,
5353 };
5354 
5355 static int
5356 mp_ndesc_handler(SYSCTL_HANDLER_ARGS)
5357 {
5358 	if_ctx_t ctx = (void *)arg1;
5359 	enum iflib_ndesc_handler type = arg2;
5360 	char buf[256] = {0};
5361 	qidx_t *ndesc;
5362 	char *p, *next;
5363 	int nqs, rc, i;
5364 
5365 	MPASS(type == IFLIB_NTXD_HANDLER || type == IFLIB_NRXD_HANDLER);
5366 
5367 	nqs = 8;
5368 	switch(type) {
5369 	case IFLIB_NTXD_HANDLER:
5370 		ndesc = ctx->ifc_sysctl_ntxds;
5371 		if (ctx->ifc_sctx)
5372 			nqs = ctx->ifc_sctx->isc_ntxqs;
5373 		break;
5374 	case IFLIB_NRXD_HANDLER:
5375 		ndesc = ctx->ifc_sysctl_nrxds;
5376 		if (ctx->ifc_sctx)
5377 			nqs = ctx->ifc_sctx->isc_nrxqs;
5378 		break;
5379 	}
5380 	if (nqs == 0)
5381 		nqs = 8;
5382 
5383 	for (i=0; i<8; i++) {
5384 		if (i >= nqs)
5385 			break;
5386 		if (i)
5387 			strcat(buf, ",");
5388 		sprintf(strchr(buf, 0), "%d", ndesc[i]);
5389 	}
5390 
5391 	rc = sysctl_handle_string(oidp, buf, sizeof(buf), req);
5392 	if (rc || req->newptr == NULL)
5393 		return rc;
5394 
5395 	for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p;
5396 	    i++, p = strsep(&next, " ,")) {
5397 		ndesc[i] = strtoul(p, NULL, 10);
5398 	}
5399 
5400 	return(rc);
5401 }
5402 
5403 #define NAME_BUFLEN 32
5404 static void
5405 iflib_add_device_sysctl_pre(if_ctx_t ctx)
5406 {
5407         device_t dev = iflib_get_dev(ctx);
5408 	struct sysctl_oid_list *child, *oid_list;
5409 	struct sysctl_ctx_list *ctx_list;
5410 	struct sysctl_oid *node;
5411 
5412 	ctx_list = device_get_sysctl_ctx(dev);
5413 	child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev));
5414 	ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, "iflib",
5415 						      CTLFLAG_RD, NULL, "IFLIB fields");
5416 	oid_list = SYSCTL_CHILDREN(node);
5417 
5418 	SYSCTL_ADD_STRING(ctx_list, oid_list, OID_AUTO, "driver_version",
5419 		       CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version, 0,
5420 		       "driver version");
5421 
5422 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs",
5423 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0,
5424 			"# of txqs to use, 0 => use default #");
5425 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs",
5426 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0,
5427 			"# of rxqs to use, 0 => use default #");
5428 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_qs_enable",
5429 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0,
5430                        "permit #txq != #rxq");
5431 
5432 	/* XXX change for per-queue sizes */
5433 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_ntxds",
5434 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NTXD_HANDLER,
5435                        mp_ndesc_handler, "A",
5436                        "list of # of tx descriptors to use, 0 = use default #");
5437 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_nrxds",
5438 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NRXD_HANDLER,
5439                        mp_ndesc_handler, "A",
5440                        "list of # of rx descriptors to use, 0 = use default #");
5441 }
5442 
5443 static void
5444 iflib_add_device_sysctl_post(if_ctx_t ctx)
5445 {
5446 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5447 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5448         device_t dev = iflib_get_dev(ctx);
5449 	struct sysctl_oid_list *child;
5450 	struct sysctl_ctx_list *ctx_list;
5451 	iflib_fl_t fl;
5452 	iflib_txq_t txq;
5453 	iflib_rxq_t rxq;
5454 	int i, j;
5455 	char namebuf[NAME_BUFLEN];
5456 	char *qfmt;
5457 	struct sysctl_oid *queue_node, *fl_node, *node;
5458 	struct sysctl_oid_list *queue_list, *fl_list;
5459 	ctx_list = device_get_sysctl_ctx(dev);
5460 
5461 	node = ctx->ifc_sysctl_node;
5462 	child = SYSCTL_CHILDREN(node);
5463 
5464 	if (scctx->isc_ntxqsets > 100)
5465 		qfmt = "txq%03d";
5466 	else if (scctx->isc_ntxqsets > 10)
5467 		qfmt = "txq%02d";
5468 	else
5469 		qfmt = "txq%d";
5470 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
5471 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
5472 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
5473 					     CTLFLAG_RD, NULL, "Queue Name");
5474 		queue_list = SYSCTL_CHILDREN(queue_node);
5475 #if MEMORY_LOGGING
5476 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued",
5477 				CTLFLAG_RD,
5478 				&txq->ift_dequeued, "total mbufs freed");
5479 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued",
5480 				CTLFLAG_RD,
5481 				&txq->ift_enqueued, "total mbufs enqueued");
5482 #endif
5483 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag",
5484 				   CTLFLAG_RD,
5485 				   &txq->ift_mbuf_defrag, "# of times m_defrag was called");
5486 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "m_pullups",
5487 				   CTLFLAG_RD,
5488 				   &txq->ift_pullups, "# of times m_pullup was called");
5489 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag_failed",
5490 				   CTLFLAG_RD,
5491 				   &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed");
5492 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_desc_avail",
5493 				   CTLFLAG_RD,
5494 				   &txq->ift_no_desc_avail, "# of times no descriptors were available");
5495 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "tx_map_failed",
5496 				   CTLFLAG_RD,
5497 				   &txq->ift_map_failed, "# of times dma map failed");
5498 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txd_encap_efbig",
5499 				   CTLFLAG_RD,
5500 				   &txq->ift_txd_encap_efbig, "# of times txd_encap returned EFBIG");
5501 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_tx_dma_setup",
5502 				   CTLFLAG_RD,
5503 				   &txq->ift_no_tx_dma_setup, "# of times map failed for other than EFBIG");
5504 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx",
5505 				   CTLFLAG_RD,
5506 				   &txq->ift_pidx, 1, "Producer Index");
5507 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx",
5508 				   CTLFLAG_RD,
5509 				   &txq->ift_cidx, 1, "Consumer Index");
5510 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx_processed",
5511 				   CTLFLAG_RD,
5512 				   &txq->ift_cidx_processed, 1, "Consumer Index seen by credit update");
5513 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use",
5514 				   CTLFLAG_RD,
5515 				   &txq->ift_in_use, 1, "descriptors in use");
5516 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_processed",
5517 				   CTLFLAG_RD,
5518 				   &txq->ift_processed, "descriptors procesed for clean");
5519 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned",
5520 				   CTLFLAG_RD,
5521 				   &txq->ift_cleaned, "total cleaned");
5522 		SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state",
5523 				CTLTYPE_STRING | CTLFLAG_RD, __DEVOLATILE(uint64_t *, &txq->ift_br->state),
5524 				0, mp_ring_state_handler, "A", "soft ring state");
5525 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_enqueues",
5526 				       CTLFLAG_RD, &txq->ift_br->enqueues,
5527 				       "# of enqueues to the mp_ring for this queue");
5528 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_drops",
5529 				       CTLFLAG_RD, &txq->ift_br->drops,
5530 				       "# of drops in the mp_ring for this queue");
5531 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_starts",
5532 				       CTLFLAG_RD, &txq->ift_br->starts,
5533 				       "# of normal consumer starts in the mp_ring for this queue");
5534 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_stalls",
5535 				       CTLFLAG_RD, &txq->ift_br->stalls,
5536 					       "# of consumer stalls in the mp_ring for this queue");
5537 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_restarts",
5538 			       CTLFLAG_RD, &txq->ift_br->restarts,
5539 				       "# of consumer restarts in the mp_ring for this queue");
5540 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_abdications",
5541 				       CTLFLAG_RD, &txq->ift_br->abdications,
5542 				       "# of consumer abdications in the mp_ring for this queue");
5543 	}
5544 
5545 	if (scctx->isc_nrxqsets > 100)
5546 		qfmt = "rxq%03d";
5547 	else if (scctx->isc_nrxqsets > 10)
5548 		qfmt = "rxq%02d";
5549 	else
5550 		qfmt = "rxq%d";
5551 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
5552 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
5553 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
5554 					     CTLFLAG_RD, NULL, "Queue Name");
5555 		queue_list = SYSCTL_CHILDREN(queue_node);
5556 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
5557 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_pidx",
5558 				       CTLFLAG_RD,
5559 				       &rxq->ifr_cq_pidx, 1, "Producer Index");
5560 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_cidx",
5561 				       CTLFLAG_RD,
5562 				       &rxq->ifr_cq_cidx, 1, "Consumer Index");
5563 		}
5564 
5565 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
5566 			snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j);
5567 			fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list, OID_AUTO, namebuf,
5568 						     CTLFLAG_RD, NULL, "freelist Name");
5569 			fl_list = SYSCTL_CHILDREN(fl_node);
5570 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx",
5571 				       CTLFLAG_RD,
5572 				       &fl->ifl_pidx, 1, "Producer Index");
5573 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx",
5574 				       CTLFLAG_RD,
5575 				       &fl->ifl_cidx, 1, "Consumer Index");
5576 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits",
5577 				       CTLFLAG_RD,
5578 				       &fl->ifl_credits, 1, "credits available");
5579 #if MEMORY_LOGGING
5580 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_enqueued",
5581 					CTLFLAG_RD,
5582 					&fl->ifl_m_enqueued, "mbufs allocated");
5583 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_dequeued",
5584 					CTLFLAG_RD,
5585 					&fl->ifl_m_dequeued, "mbufs freed");
5586 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_enqueued",
5587 					CTLFLAG_RD,
5588 					&fl->ifl_cl_enqueued, "clusters allocated");
5589 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_dequeued",
5590 					CTLFLAG_RD,
5591 					&fl->ifl_cl_dequeued, "clusters freed");
5592 #endif
5593 
5594 		}
5595 	}
5596 
5597 }
5598 
5599 #ifndef __NO_STRICT_ALIGNMENT
5600 static struct mbuf *
5601 iflib_fixup_rx(struct mbuf *m)
5602 {
5603 	struct mbuf *n;
5604 
5605 	if (m->m_len <= (MCLBYTES - ETHER_HDR_LEN)) {
5606 		bcopy(m->m_data, m->m_data + ETHER_HDR_LEN, m->m_len);
5607 		m->m_data += ETHER_HDR_LEN;
5608 		n = m;
5609 	} else {
5610 		MGETHDR(n, M_NOWAIT, MT_DATA);
5611 		if (n == NULL) {
5612 			m_freem(m);
5613 			return (NULL);
5614 		}
5615 		bcopy(m->m_data, n->m_data, ETHER_HDR_LEN);
5616 		m->m_data += ETHER_HDR_LEN;
5617 		m->m_len -= ETHER_HDR_LEN;
5618 		n->m_len = ETHER_HDR_LEN;
5619 		M_MOVE_PKTHDR(n, m);
5620 		n->m_next = m;
5621 	}
5622 	return (n);
5623 }
5624 #endif
5625