xref: /freebsd/sys/net/iflib.c (revision f37852c17391fdf0e8309bcf684384dd0d854e43)
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 #include <sys/bitstring.h>
97 /*
98  * enable accounting of every mbuf as it comes in to and goes out of
99  * iflib's software descriptor references
100  */
101 #define MEMORY_LOGGING 0
102 /*
103  * Enable mbuf vectors for compressing long mbuf chains
104  */
105 
106 /*
107  * NB:
108  * - Prefetching in tx cleaning should perhaps be a tunable. The distance ahead
109  *   we prefetch needs to be determined by the time spent in m_free vis a vis
110  *   the cost of a prefetch. This will of course vary based on the workload:
111  *      - NFLX's m_free path is dominated by vm-based M_EXT manipulation which
112  *        is quite expensive, thus suggesting very little prefetch.
113  *      - small packet forwarding which is just returning a single mbuf to
114  *        UMA will typically be very fast vis a vis the cost of a memory
115  *        access.
116  */
117 
118 
119 /*
120  * File organization:
121  *  - private structures
122  *  - iflib private utility functions
123  *  - ifnet functions
124  *  - vlan registry and other exported functions
125  *  - iflib public core functions
126  *
127  *
128  */
129 static MALLOC_DEFINE(M_IFLIB, "iflib", "ifnet library");
130 
131 struct iflib_txq;
132 typedef struct iflib_txq *iflib_txq_t;
133 struct iflib_rxq;
134 typedef struct iflib_rxq *iflib_rxq_t;
135 struct iflib_fl;
136 typedef struct iflib_fl *iflib_fl_t;
137 
138 struct iflib_ctx;
139 
140 typedef struct iflib_filter_info {
141 	driver_filter_t *ifi_filter;
142 	void *ifi_filter_arg;
143 	struct grouptask *ifi_task;
144 	void *ifi_ctx;
145 } *iflib_filter_info_t;
146 
147 struct iflib_ctx {
148 	KOBJ_FIELDS;
149    /*
150    * Pointer to hardware driver's softc
151    */
152 	void *ifc_softc;
153 	device_t ifc_dev;
154 	if_t ifc_ifp;
155 
156 	cpuset_t ifc_cpus;
157 	if_shared_ctx_t ifc_sctx;
158 	struct if_softc_ctx ifc_softc_ctx;
159 
160 	struct mtx ifc_mtx;
161 
162 	uint16_t ifc_nhwtxqs;
163 	uint16_t ifc_nhwrxqs;
164 
165 	iflib_txq_t ifc_txqs;
166 	iflib_rxq_t ifc_rxqs;
167 	uint32_t ifc_if_flags;
168 	uint32_t ifc_flags;
169 	uint32_t ifc_max_fl_buf_size;
170 	int ifc_in_detach;
171 
172 	int ifc_link_state;
173 	int ifc_link_irq;
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 	bitstr_t 	*ifl_rx_bitmap;
386 	qidx_t		ifl_fragidx;
387 	/* constant */
388 	qidx_t		ifl_size;
389 	uint16_t	ifl_buf_size;
390 	uint16_t	ifl_cltype;
391 	uma_zone_t	ifl_zone;
392 	iflib_rxsd_array_t	ifl_sds;
393 	iflib_rxq_t	ifl_rxq;
394 	uint8_t		ifl_id;
395 	bus_dma_tag_t           ifl_desc_tag;
396 	iflib_dma_info_t	ifl_ifdi;
397 	uint64_t	ifl_bus_addrs[IFLIB_MAX_RX_REFRESH] __aligned(CACHE_LINE_SIZE);
398 	caddr_t		ifl_vm_addrs[IFLIB_MAX_RX_REFRESH];
399 	qidx_t	ifl_rxd_idxs[IFLIB_MAX_RX_REFRESH];
400 }  __aligned(CACHE_LINE_SIZE);
401 
402 static inline qidx_t
403 get_inuse(int size, qidx_t cidx, qidx_t pidx, uint8_t gen)
404 {
405 	qidx_t used;
406 
407 	if (pidx > cidx)
408 		used = pidx - cidx;
409 	else if (pidx < cidx)
410 		used = size - cidx + pidx;
411 	else if (gen == 0 && pidx == cidx)
412 		used = 0;
413 	else if (gen == 1 && pidx == cidx)
414 		used = size;
415 	else
416 		panic("bad state");
417 
418 	return (used);
419 }
420 
421 #define TXQ_AVAIL(txq) (txq->ift_size - get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx, txq->ift_gen))
422 
423 #define IDXDIFF(head, tail, wrap) \
424 	((head) >= (tail) ? (head) - (tail) : (wrap) - (tail) + (head))
425 
426 struct iflib_rxq {
427 	/* If there is a separate completion queue -
428 	 * these are the cq cidx and pidx. Otherwise
429 	 * these are unused.
430 	 */
431 	qidx_t		ifr_size;
432 	qidx_t		ifr_cq_cidx;
433 	qidx_t		ifr_cq_pidx;
434 	uint8_t		ifr_cq_gen;
435 	uint8_t		ifr_fl_offset;
436 
437 	if_ctx_t	ifr_ctx;
438 	iflib_fl_t	ifr_fl;
439 	uint64_t	ifr_rx_irq;
440 	uint16_t	ifr_id;
441 	uint8_t		ifr_lro_enabled;
442 	uint8_t		ifr_nfl;
443 	uint8_t		ifr_ntxqirq;
444 	uint8_t		ifr_txqid[IFLIB_MAX_TX_SHARED_INTR];
445 	struct lro_ctrl			ifr_lc;
446 	struct grouptask        ifr_task;
447 	struct iflib_filter_info ifr_filter_info;
448 	iflib_dma_info_t		ifr_ifdi;
449 
450 	/* dynamically allocate if any drivers need a value substantially larger than this */
451 	struct if_rxd_frag	ifr_frags[IFLIB_MAX_RX_SEGS] __aligned(CACHE_LINE_SIZE);
452 #ifdef IFLIB_DIAGNOSTICS
453 	uint64_t ifr_cpu_exec_count[256];
454 #endif
455 }  __aligned(CACHE_LINE_SIZE);
456 
457 typedef struct if_rxsd {
458 	caddr_t *ifsd_cl;
459 	struct mbuf **ifsd_m;
460 	iflib_fl_t ifsd_fl;
461 	qidx_t ifsd_cidx;
462 } *if_rxsd_t;
463 
464 /* multiple of word size */
465 #ifdef __LP64__
466 #define PKT_INFO_SIZE	6
467 #define RXD_INFO_SIZE	5
468 #define PKT_TYPE uint64_t
469 #else
470 #define PKT_INFO_SIZE	11
471 #define RXD_INFO_SIZE	8
472 #define PKT_TYPE uint32_t
473 #endif
474 #define PKT_LOOP_BOUND  ((PKT_INFO_SIZE/3)*3)
475 #define RXD_LOOP_BOUND  ((RXD_INFO_SIZE/4)*4)
476 
477 typedef struct if_pkt_info_pad {
478 	PKT_TYPE pkt_val[PKT_INFO_SIZE];
479 } *if_pkt_info_pad_t;
480 typedef struct if_rxd_info_pad {
481 	PKT_TYPE rxd_val[RXD_INFO_SIZE];
482 } *if_rxd_info_pad_t;
483 
484 CTASSERT(sizeof(struct if_pkt_info_pad) == sizeof(struct if_pkt_info));
485 CTASSERT(sizeof(struct if_rxd_info_pad) == sizeof(struct if_rxd_info));
486 
487 
488 static inline void
489 pkt_info_zero(if_pkt_info_t pi)
490 {
491 	if_pkt_info_pad_t pi_pad;
492 
493 	pi_pad = (if_pkt_info_pad_t)pi;
494 	pi_pad->pkt_val[0] = 0; pi_pad->pkt_val[1] = 0; pi_pad->pkt_val[2] = 0;
495 	pi_pad->pkt_val[3] = 0; pi_pad->pkt_val[4] = 0; pi_pad->pkt_val[5] = 0;
496 #ifndef __LP64__
497 	pi_pad->pkt_val[6] = 0; pi_pad->pkt_val[7] = 0; pi_pad->pkt_val[8] = 0;
498 	pi_pad->pkt_val[9] = 0; pi_pad->pkt_val[10] = 0;
499 #endif
500 }
501 
502 static inline void
503 rxd_info_zero(if_rxd_info_t ri)
504 {
505 	if_rxd_info_pad_t ri_pad;
506 	int i;
507 
508 	ri_pad = (if_rxd_info_pad_t)ri;
509 	for (i = 0; i < RXD_LOOP_BOUND; i += 4) {
510 		ri_pad->rxd_val[i] = 0;
511 		ri_pad->rxd_val[i+1] = 0;
512 		ri_pad->rxd_val[i+2] = 0;
513 		ri_pad->rxd_val[i+3] = 0;
514 	}
515 #ifdef __LP64__
516 	ri_pad->rxd_val[RXD_INFO_SIZE-1] = 0;
517 #endif
518 }
519 
520 /*
521  * Only allow a single packet to take up most 1/nth of the tx ring
522  */
523 #define MAX_SINGLE_PACKET_FRACTION 12
524 #define IF_BAD_DMA (bus_addr_t)-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, frag_idx = fl->ifl_fragidx;
1804         int pidx = fl->ifl_pidx;
1805 	caddr_t cl, *sd_cl;
1806 	struct mbuf **sd_m;
1807 	uint8_t *sd_flags;
1808 	struct if_rxd_update iru;
1809 	bus_dmamap_t *sd_map;
1810 	int n, i = 0;
1811 	uint64_t bus_addr;
1812 	int err;
1813 
1814 	sd_m = fl->ifl_sds.ifsd_m;
1815 	sd_map = fl->ifl_sds.ifsd_map;
1816 	sd_cl = fl->ifl_sds.ifsd_cl;
1817 	sd_flags = fl->ifl_sds.ifsd_flags;
1818 	idx = pidx;
1819 
1820 	n  = count;
1821 	MPASS(n > 0);
1822 	MPASS(fl->ifl_credits + n <= fl->ifl_size);
1823 
1824 	if (pidx < fl->ifl_cidx)
1825 		MPASS(pidx + n <= fl->ifl_cidx);
1826 	if (pidx == fl->ifl_cidx && (fl->ifl_credits < fl->ifl_size))
1827 		MPASS(fl->ifl_gen == 0);
1828 	if (pidx > fl->ifl_cidx)
1829 		MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx);
1830 
1831 	DBG_COUNTER_INC(fl_refills);
1832 	if (n > 8)
1833 		DBG_COUNTER_INC(fl_refills_large);
1834 	iru.iru_paddrs = fl->ifl_bus_addrs;
1835 	iru.iru_vaddrs = &fl->ifl_vm_addrs[0];
1836 	iru.iru_idxs = fl->ifl_rxd_idxs;
1837 	iru.iru_qsidx = fl->ifl_rxq->ifr_id;
1838 	iru.iru_buf_size = fl->ifl_buf_size;
1839 	iru.iru_flidx = fl->ifl_id;
1840 	while (n--) {
1841 		/*
1842 		 * We allocate an uninitialized mbuf + cluster, mbuf is
1843 		 * initialized after rx.
1844 		 *
1845 		 * If the cluster is still set then we know a minimum sized packet was received
1846 		 */
1847 		bit_ffc_at(fl->ifl_rx_bitmap, frag_idx, fl->ifl_size,  &frag_idx);
1848 		if ((frag_idx < 0) || (frag_idx >= fl->ifl_size))
1849                 	bit_ffc(fl->ifl_rx_bitmap, fl->ifl_size, &frag_idx);
1850 		if ((cl = sd_cl[frag_idx]) == NULL) {
1851                        if ((cl = sd_cl[frag_idx] = m_cljget(NULL, M_NOWAIT, fl->ifl_buf_size)) == NULL)
1852 				break;
1853 #if MEMORY_LOGGING
1854 			fl->ifl_cl_enqueued++;
1855 #endif
1856 		}
1857 		if ((m = m_gethdr(M_NOWAIT, MT_NOINIT)) == NULL) {
1858 			break;
1859 		}
1860 #if MEMORY_LOGGING
1861 		fl->ifl_m_enqueued++;
1862 #endif
1863 
1864 		DBG_COUNTER_INC(rx_allocs);
1865 #if defined(__i386__) || defined(__amd64__)
1866 		if (!IS_DMAR(ctx)) {
1867 			bus_addr = pmap_kextract((vm_offset_t)cl);
1868 		} else
1869 #endif
1870 		{
1871 			struct rxq_refill_cb_arg cb_arg;
1872 			iflib_rxq_t q;
1873 
1874 			cb_arg.error = 0;
1875 			q = fl->ifl_rxq;
1876 			MPASS(sd_map != NULL);
1877 			MPASS(sd_map[frag_idx] != NULL);
1878 			err = bus_dmamap_load(fl->ifl_desc_tag, sd_map[frag_idx],
1879 		         cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg, 0);
1880 			bus_dmamap_sync(fl->ifl_desc_tag, sd_map[frag_idx],
1881 					BUS_DMASYNC_PREREAD);
1882 
1883 			if (err != 0 || cb_arg.error) {
1884 				/*
1885 				 * !zone_pack ?
1886 				 */
1887 				if (fl->ifl_zone == zone_pack)
1888 					uma_zfree(fl->ifl_zone, cl);
1889 				m_free(m);
1890 				n = 0;
1891 				goto done;
1892 			}
1893 			bus_addr = cb_arg.seg.ds_addr;
1894 		}
1895                 bit_set(fl->ifl_rx_bitmap, frag_idx);
1896 		sd_flags[frag_idx] |= RX_SW_DESC_INUSE;
1897 
1898 		MPASS(sd_m[frag_idx] == NULL);
1899 		sd_cl[frag_idx] = cl;
1900 		sd_m[frag_idx] = m;
1901 		fl->ifl_rxd_idxs[i] = frag_idx;
1902 		fl->ifl_bus_addrs[i] = bus_addr;
1903 		fl->ifl_vm_addrs[i] = cl;
1904 		fl->ifl_credits++;
1905 		i++;
1906 		MPASS(fl->ifl_credits <= fl->ifl_size);
1907 		if (++idx == fl->ifl_size) {
1908 			fl->ifl_gen = 1;
1909 			idx = 0;
1910 		}
1911 		if (n == 0 || i == IFLIB_MAX_RX_REFRESH) {
1912 			iru.iru_pidx = pidx;
1913 			iru.iru_count = i;
1914 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
1915 			i = 0;
1916 			pidx = idx;
1917 			fl->ifl_pidx = idx;
1918 		}
1919 
1920 	}
1921 done:
1922 	DBG_COUNTER_INC(rxd_flush);
1923 	if (fl->ifl_pidx == 0)
1924 		pidx = fl->ifl_size - 1;
1925 	else
1926 		pidx = fl->ifl_pidx - 1;
1927 
1928 	if (sd_map)
1929 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
1930 				BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1931 	ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id, fl->ifl_id, pidx);
1932 	fl->ifl_fragidx = frag_idx;
1933 }
1934 
1935 static __inline void
1936 __iflib_fl_refill_lt(if_ctx_t ctx, iflib_fl_t fl, int max)
1937 {
1938 	/* we avoid allowing pidx to catch up with cidx as it confuses ixl */
1939 	int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1;
1940 #ifdef INVARIANTS
1941 	int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1;
1942 #endif
1943 
1944 	MPASS(fl->ifl_credits <= fl->ifl_size);
1945 	MPASS(reclaimable == delta);
1946 
1947 	if (reclaimable > 0)
1948 		_iflib_fl_refill(ctx, fl, min(max, reclaimable));
1949 }
1950 
1951 static void
1952 iflib_fl_bufs_free(iflib_fl_t fl)
1953 {
1954 	iflib_dma_info_t idi = fl->ifl_ifdi;
1955 	uint32_t i;
1956 
1957 	for (i = 0; i < fl->ifl_size; i++) {
1958 		struct mbuf **sd_m = &fl->ifl_sds.ifsd_m[i];
1959 		uint8_t *sd_flags = &fl->ifl_sds.ifsd_flags[i];
1960 		caddr_t *sd_cl = &fl->ifl_sds.ifsd_cl[i];
1961 
1962 		if (*sd_flags & RX_SW_DESC_INUSE) {
1963 			if (fl->ifl_sds.ifsd_map != NULL) {
1964 				bus_dmamap_t sd_map = fl->ifl_sds.ifsd_map[i];
1965 				bus_dmamap_unload(fl->ifl_desc_tag, sd_map);
1966 				bus_dmamap_destroy(fl->ifl_desc_tag, sd_map);
1967 			}
1968 			if (*sd_m != NULL) {
1969 				m_init(*sd_m, M_NOWAIT, MT_DATA, 0);
1970 				uma_zfree(zone_mbuf, *sd_m);
1971 			}
1972 			if (*sd_cl != NULL)
1973 				uma_zfree(fl->ifl_zone, *sd_cl);
1974 			*sd_flags = 0;
1975 		} else {
1976 			MPASS(*sd_cl == NULL);
1977 			MPASS(*sd_m == NULL);
1978 		}
1979 #if MEMORY_LOGGING
1980 		fl->ifl_m_dequeued++;
1981 		fl->ifl_cl_dequeued++;
1982 #endif
1983 		*sd_cl = NULL;
1984 		*sd_m = NULL;
1985 	}
1986 #ifdef INVARIANTS
1987 	for (i = 0; i < fl->ifl_size; i++) {
1988 		MPASS(fl->ifl_sds.ifsd_flags[i] == 0);
1989 		MPASS(fl->ifl_sds.ifsd_cl[i] == NULL);
1990 		MPASS(fl->ifl_sds.ifsd_m[i] == NULL);
1991 	}
1992 #endif
1993 	/*
1994 	 * Reset free list values
1995 	 */
1996 	fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = fl->ifl_fragidx = 0;
1997 	bzero(idi->idi_vaddr, idi->idi_size);
1998 }
1999 
2000 /*********************************************************************
2001  *
2002  *  Initialize a receive ring and its buffers.
2003  *
2004  **********************************************************************/
2005 static int
2006 iflib_fl_setup(iflib_fl_t fl)
2007 {
2008 	iflib_rxq_t rxq = fl->ifl_rxq;
2009 	if_ctx_t ctx = rxq->ifr_ctx;
2010 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2011 
2012 	bit_nclear(fl->ifl_rx_bitmap, 0, fl->ifl_size);
2013 	/*
2014 	** Free current RX buffer structs and their mbufs
2015 	*/
2016 	iflib_fl_bufs_free(fl);
2017 	/* Now replenish the mbufs */
2018 	MPASS(fl->ifl_credits == 0);
2019 	/*
2020 	 * XXX don't set the max_frame_size to larger
2021 	 * than the hardware can handle
2022 	 */
2023 	if (sctx->isc_max_frame_size <= 2048)
2024 		fl->ifl_buf_size = MCLBYTES;
2025 #ifndef CONTIGMALLOC_WORKS
2026 	else
2027 		fl->ifl_buf_size = MJUMPAGESIZE;
2028 #else
2029 	else if (sctx->isc_max_frame_size <= 4096)
2030 		fl->ifl_buf_size = MJUMPAGESIZE;
2031 	else if (sctx->isc_max_frame_size <= 9216)
2032 		fl->ifl_buf_size = MJUM9BYTES;
2033 	else
2034 		fl->ifl_buf_size = MJUM16BYTES;
2035 #endif
2036 	if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size)
2037 		ctx->ifc_max_fl_buf_size = fl->ifl_buf_size;
2038 	fl->ifl_cltype = m_gettype(fl->ifl_buf_size);
2039 	fl->ifl_zone = m_getzone(fl->ifl_buf_size);
2040 
2041 
2042 	/* avoid pre-allocating zillions of clusters to an idle card
2043 	 * potentially speeding up attach
2044 	 */
2045 	_iflib_fl_refill(ctx, fl, min(128, fl->ifl_size));
2046 	MPASS(min(128, fl->ifl_size) == fl->ifl_credits);
2047 	if (min(128, fl->ifl_size) != fl->ifl_credits)
2048 		return (ENOBUFS);
2049 	/*
2050 	 * handle failure
2051 	 */
2052 	MPASS(rxq != NULL);
2053 	MPASS(fl->ifl_ifdi != NULL);
2054 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2055 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2056 	return (0);
2057 }
2058 
2059 /*********************************************************************
2060  *
2061  *  Free receive ring data structures
2062  *
2063  **********************************************************************/
2064 static void
2065 iflib_rx_sds_free(iflib_rxq_t rxq)
2066 {
2067 	iflib_fl_t fl;
2068 	int i;
2069 
2070 	if (rxq->ifr_fl != NULL) {
2071 		for (i = 0; i < rxq->ifr_nfl; i++) {
2072 			fl = &rxq->ifr_fl[i];
2073 			if (fl->ifl_desc_tag != NULL) {
2074 				bus_dma_tag_destroy(fl->ifl_desc_tag);
2075 				fl->ifl_desc_tag = NULL;
2076 			}
2077 			free(fl->ifl_sds.ifsd_m, M_IFLIB);
2078 			free(fl->ifl_sds.ifsd_cl, M_IFLIB);
2079 			/* XXX destroy maps first */
2080 			free(fl->ifl_sds.ifsd_map, M_IFLIB);
2081 			fl->ifl_sds.ifsd_m = NULL;
2082 			fl->ifl_sds.ifsd_cl = NULL;
2083 			fl->ifl_sds.ifsd_map = NULL;
2084 		}
2085 		free(rxq->ifr_fl, M_IFLIB);
2086 		rxq->ifr_fl = NULL;
2087 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
2088 	}
2089 }
2090 
2091 /*
2092  * MI independent logic
2093  *
2094  */
2095 static void
2096 iflib_timer(void *arg)
2097 {
2098 	iflib_txq_t txq = arg;
2099 	if_ctx_t ctx = txq->ift_ctx;
2100 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2101 
2102 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
2103 		return;
2104 	/*
2105 	** Check on the state of the TX queue(s), this
2106 	** can be done without the lock because its RO
2107 	** and the HUNG state will be static if set.
2108 	*/
2109 	IFDI_TIMER(ctx, txq->ift_id);
2110 	if ((txq->ift_qstatus == IFLIB_QUEUE_HUNG) &&
2111 	    ((txq->ift_cleaned_prev == txq->ift_cleaned) ||
2112 	     (sctx->isc_pause_frames == 0)))
2113 		goto hung;
2114 
2115 	if (ifmp_ring_is_stalled(txq->ift_br))
2116 		txq->ift_qstatus = IFLIB_QUEUE_HUNG;
2117 	txq->ift_cleaned_prev = txq->ift_cleaned;
2118 	/* handle any laggards */
2119 	if (txq->ift_db_pending)
2120 		GROUPTASK_ENQUEUE(&txq->ift_task);
2121 
2122 	sctx->isc_pause_frames = 0;
2123 	if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)
2124 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq, txq->ift_timer.c_cpu);
2125 	return;
2126 hung:
2127 	CTX_LOCK(ctx);
2128 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2129 	device_printf(ctx->ifc_dev,  "TX(%d) desc avail = %d, pidx = %d\n",
2130 				  txq->ift_id, TXQ_AVAIL(txq), txq->ift_pidx);
2131 
2132 	IFDI_WATCHDOG_RESET(ctx);
2133 	ctx->ifc_watchdog_events++;
2134 
2135 	ctx->ifc_flags |= IFC_DO_RESET;
2136 	iflib_admin_intr_deferred(ctx);
2137 	CTX_UNLOCK(ctx);
2138 }
2139 
2140 static void
2141 iflib_init_locked(if_ctx_t ctx)
2142 {
2143 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2144 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2145 	if_t ifp = ctx->ifc_ifp;
2146 	iflib_fl_t fl;
2147 	iflib_txq_t txq;
2148 	iflib_rxq_t rxq;
2149 	int i, j, tx_ip_csum_flags, tx_ip6_csum_flags;
2150 
2151 
2152 	if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2153 	IFDI_INTR_DISABLE(ctx);
2154 
2155 	tx_ip_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP);
2156 	tx_ip6_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | CSUM_IP6_SCTP);
2157 	/* Set hardware offload abilities */
2158 	if_clearhwassist(ifp);
2159 	if (if_getcapenable(ifp) & IFCAP_TXCSUM)
2160 		if_sethwassistbits(ifp, tx_ip_csum_flags, 0);
2161 	if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6)
2162 		if_sethwassistbits(ifp,  tx_ip6_csum_flags, 0);
2163 	if (if_getcapenable(ifp) & IFCAP_TSO4)
2164 		if_sethwassistbits(ifp, CSUM_IP_TSO, 0);
2165 	if (if_getcapenable(ifp) & IFCAP_TSO6)
2166 		if_sethwassistbits(ifp, CSUM_IP6_TSO, 0);
2167 
2168 	for (i = 0, txq = ctx->ifc_txqs; i < sctx->isc_ntxqsets; i++, txq++) {
2169 		CALLOUT_LOCK(txq);
2170 		callout_stop(&txq->ift_timer);
2171 		CALLOUT_UNLOCK(txq);
2172 		iflib_netmap_txq_init(ctx, txq);
2173 	}
2174 	for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
2175 		MPASS(rxq->ifr_id == i);
2176 		iflib_netmap_rxq_init(ctx, rxq);
2177 	}
2178 #ifdef INVARIANTS
2179 	i = if_getdrvflags(ifp);
2180 #endif
2181 	IFDI_INIT(ctx);
2182 	MPASS(if_getdrvflags(ifp) == i);
2183 	for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
2184 		/* XXX this should really be done on a per-queue basis */
2185 		if (if_getcapenable(ifp) & IFCAP_NETMAP)
2186 			continue;
2187 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
2188 			if (iflib_fl_setup(fl)) {
2189 				device_printf(ctx->ifc_dev, "freelist setup failed - check cluster settings\n");
2190 				goto done;
2191 			}
2192 		}
2193 	}
2194 	done:
2195 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
2196 	IFDI_INTR_ENABLE(ctx);
2197 	txq = ctx->ifc_txqs;
2198 	for (i = 0; i < sctx->isc_ntxqsets; i++, txq++)
2199 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq,
2200 			txq->ift_timer.c_cpu);
2201 }
2202 
2203 static int
2204 iflib_media_change(if_t ifp)
2205 {
2206 	if_ctx_t ctx = if_getsoftc(ifp);
2207 	int err;
2208 
2209 	CTX_LOCK(ctx);
2210 	if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0)
2211 		iflib_init_locked(ctx);
2212 	CTX_UNLOCK(ctx);
2213 	return (err);
2214 }
2215 
2216 static void
2217 iflib_media_status(if_t ifp, struct ifmediareq *ifmr)
2218 {
2219 	if_ctx_t ctx = if_getsoftc(ifp);
2220 
2221 	CTX_LOCK(ctx);
2222 	IFDI_UPDATE_ADMIN_STATUS(ctx);
2223 	IFDI_MEDIA_STATUS(ctx, ifmr);
2224 	CTX_UNLOCK(ctx);
2225 }
2226 
2227 static void
2228 iflib_stop(if_ctx_t ctx)
2229 {
2230 	iflib_txq_t txq = ctx->ifc_txqs;
2231 	iflib_rxq_t rxq = ctx->ifc_rxqs;
2232 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2233 	iflib_dma_info_t di;
2234 	iflib_fl_t fl;
2235 	int i, j;
2236 
2237 	/* Tell the stack that the interface is no longer active */
2238 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2239 
2240 	IFDI_INTR_DISABLE(ctx);
2241 	DELAY(1000);
2242 	IFDI_STOP(ctx);
2243 	DELAY(1000);
2244 
2245 	iflib_debug_reset();
2246 	/* Wait for current tx queue users to exit to disarm watchdog timer. */
2247 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) {
2248 		/* make sure all transmitters have completed before proceeding XXX */
2249 
2250 		/* clean any enqueued buffers */
2251 		iflib_ifmp_purge(txq);
2252 		/* Free any existing tx buffers. */
2253 		for (j = 0; j < txq->ift_size; j++) {
2254 			iflib_txsd_free(ctx, txq, j);
2255 		}
2256 		txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0;
2257 		txq->ift_in_use = txq->ift_gen = txq->ift_cidx = txq->ift_pidx = txq->ift_no_desc_avail = 0;
2258 		txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0;
2259 		txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0;
2260 		txq->ift_pullups = 0;
2261 		ifmp_ring_reset_stats(txq->ift_br);
2262 		for (j = 0, di = txq->ift_ifdi; j < ctx->ifc_nhwtxqs; j++, di++)
2263 			bzero((void *)di->idi_vaddr, di->idi_size);
2264 	}
2265 	for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) {
2266 		/* make sure all transmitters have completed before proceeding XXX */
2267 
2268 		for (j = 0, di = txq->ift_ifdi; j < ctx->ifc_nhwrxqs; j++, di++)
2269 			bzero((void *)di->idi_vaddr, di->idi_size);
2270 		/* also resets the free lists pidx/cidx */
2271 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
2272 			iflib_fl_bufs_free(fl);
2273 	}
2274 }
2275 
2276 static inline caddr_t
2277 calc_next_rxd(iflib_fl_t fl, int cidx)
2278 {
2279 	qidx_t size;
2280 	int nrxd;
2281 	caddr_t start, end, cur, next;
2282 
2283 	nrxd = fl->ifl_size;
2284 	size = fl->ifl_rxd_size;
2285 	start = fl->ifl_ifdi->idi_vaddr;
2286 
2287 	if (__predict_false(size == 0))
2288 		return (start);
2289 	cur = start + size*cidx;
2290 	end = start + size*nrxd;
2291 	next = CACHE_PTR_NEXT(cur);
2292 	return (next < end ? next : start);
2293 }
2294 
2295 static inline void
2296 prefetch_pkts(iflib_fl_t fl, int cidx)
2297 {
2298 	int nextptr;
2299 	int nrxd = fl->ifl_size;
2300 	caddr_t next_rxd;
2301 
2302 
2303 	nextptr = (cidx + CACHE_PTR_INCREMENT) & (nrxd-1);
2304 	prefetch(&fl->ifl_sds.ifsd_m[nextptr]);
2305 	prefetch(&fl->ifl_sds.ifsd_cl[nextptr]);
2306 	next_rxd = calc_next_rxd(fl, cidx);
2307 	prefetch(next_rxd);
2308 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 1) & (nrxd-1)]);
2309 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 2) & (nrxd-1)]);
2310 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 3) & (nrxd-1)]);
2311 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 4) & (nrxd-1)]);
2312 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 1) & (nrxd-1)]);
2313 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 2) & (nrxd-1)]);
2314 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 3) & (nrxd-1)]);
2315 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 4) & (nrxd-1)]);
2316 }
2317 
2318 static void
2319 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, int unload, if_rxsd_t sd)
2320 {
2321 	int flid, cidx;
2322 	bus_dmamap_t map;
2323 	iflib_fl_t fl;
2324 	iflib_dma_info_t di;
2325 	int next;
2326 
2327 	map = NULL;
2328 	flid = irf->irf_flid;
2329 	cidx = irf->irf_idx;
2330 	fl = &rxq->ifr_fl[flid];
2331 	sd->ifsd_fl = fl;
2332 	sd->ifsd_cidx = cidx;
2333 	sd->ifsd_m = &fl->ifl_sds.ifsd_m[cidx];
2334 	sd->ifsd_cl = &fl->ifl_sds.ifsd_cl[cidx];
2335 	fl->ifl_credits--;
2336 #if MEMORY_LOGGING
2337 	fl->ifl_m_dequeued++;
2338 #endif
2339 	if (rxq->ifr_ctx->ifc_flags & IFC_PREFETCH)
2340 		prefetch_pkts(fl, cidx);
2341 	if (fl->ifl_sds.ifsd_map != NULL) {
2342 		next = (cidx + CACHE_PTR_INCREMENT) & (fl->ifl_size-1);
2343 		prefetch(&fl->ifl_sds.ifsd_map[next]);
2344 		map = fl->ifl_sds.ifsd_map[cidx];
2345 		di = fl->ifl_ifdi;
2346 		next = (cidx + CACHE_LINE_SIZE) & (fl->ifl_size-1);
2347 		prefetch(&fl->ifl_sds.ifsd_flags[next]);
2348 		bus_dmamap_sync(di->idi_tag, di->idi_map,
2349 				BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
2350 
2351 	/* not valid assert if bxe really does SGE from non-contiguous elements */
2352 		MPASS(fl->ifl_cidx == cidx);
2353 		if (unload)
2354 			bus_dmamap_unload(fl->ifl_desc_tag, map);
2355 	}
2356 	fl->ifl_cidx = (fl->ifl_cidx + 1) & (fl->ifl_size-1);
2357 	if (__predict_false(fl->ifl_cidx == 0))
2358 		fl->ifl_gen = 0;
2359 	if (map != NULL)
2360 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2361 			BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2362         bit_clear(fl->ifl_rx_bitmap, cidx);
2363 }
2364 
2365 static struct mbuf *
2366 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri, if_rxsd_t sd)
2367 {
2368 	int i, padlen , flags;
2369 	struct mbuf *m, *mh, *mt;
2370 	caddr_t cl;
2371 
2372 	i = 0;
2373 	mh = NULL;
2374 	do {
2375 		rxd_frag_to_sd(rxq, &ri->iri_frags[i], TRUE, sd);
2376 
2377 		MPASS(*sd->ifsd_cl != NULL);
2378 		MPASS(*sd->ifsd_m != NULL);
2379 
2380 		/* Don't include zero-length frags */
2381 		if (ri->iri_frags[i].irf_len == 0) {
2382 			/* XXX we can save the cluster here, but not the mbuf */
2383 			m_init(*sd->ifsd_m, M_NOWAIT, MT_DATA, 0);
2384 			m_free(*sd->ifsd_m);
2385 			*sd->ifsd_m = NULL;
2386 			continue;
2387 		}
2388 		m = *sd->ifsd_m;
2389 		*sd->ifsd_m = NULL;
2390 		if (mh == NULL) {
2391 			flags = M_PKTHDR|M_EXT;
2392 			mh = mt = m;
2393 			padlen = ri->iri_pad;
2394 		} else {
2395 			flags = M_EXT;
2396 			mt->m_next = m;
2397 			mt = m;
2398 			/* assuming padding is only on the first fragment */
2399 			padlen = 0;
2400 		}
2401 		cl = *sd->ifsd_cl;
2402 		*sd->ifsd_cl = NULL;
2403 
2404 		/* Can these two be made one ? */
2405 		m_init(m, M_NOWAIT, MT_DATA, flags);
2406 		m_cljset(m, cl, sd->ifsd_fl->ifl_cltype);
2407 		/*
2408 		 * These must follow m_init and m_cljset
2409 		 */
2410 		m->m_data += padlen;
2411 		ri->iri_len -= padlen;
2412 		m->m_len = ri->iri_frags[i].irf_len;
2413 	} while (++i < ri->iri_nfrags);
2414 
2415 	return (mh);
2416 }
2417 
2418 /*
2419  * Process one software descriptor
2420  */
2421 static struct mbuf *
2422 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri)
2423 {
2424 	struct if_rxsd sd;
2425 	struct mbuf *m;
2426 
2427 	/* should I merge this back in now that the two paths are basically duplicated? */
2428 	if (ri->iri_nfrags == 1 &&
2429 	    ri->iri_frags[0].irf_len <= IFLIB_RX_COPY_THRESH) {
2430 		rxd_frag_to_sd(rxq, &ri->iri_frags[0], FALSE, &sd);
2431 		m = *sd.ifsd_m;
2432 		*sd.ifsd_m = NULL;
2433 		m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR);
2434 #ifndef __NO_STRICT_ALIGNMENT
2435 		if (!IP_ALIGNED(m))
2436 			m->m_data += 2;
2437 #endif
2438 		memcpy(m->m_data, *sd.ifsd_cl, ri->iri_len);
2439 		m->m_len = ri->iri_frags[0].irf_len;
2440        } else {
2441 		m = assemble_segments(rxq, ri, &sd);
2442 	}
2443 	m->m_pkthdr.len = ri->iri_len;
2444 	m->m_pkthdr.rcvif = ri->iri_ifp;
2445 	m->m_flags |= ri->iri_flags;
2446 	m->m_pkthdr.ether_vtag = ri->iri_vtag;
2447 	m->m_pkthdr.flowid = ri->iri_flowid;
2448 	M_HASHTYPE_SET(m, ri->iri_rsstype);
2449 	m->m_pkthdr.csum_flags = ri->iri_csum_flags;
2450 	m->m_pkthdr.csum_data = ri->iri_csum_data;
2451 	return (m);
2452 }
2453 
2454 static bool
2455 iflib_rxeof(iflib_rxq_t rxq, qidx_t budget)
2456 {
2457 	if_ctx_t ctx = rxq->ifr_ctx;
2458 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2459 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2460 	int avail, i;
2461 	qidx_t *cidxp;
2462 	struct if_rxd_info ri;
2463 	int err, budget_left, rx_bytes, rx_pkts;
2464 	iflib_fl_t fl;
2465 	struct ifnet *ifp;
2466 	int lro_enabled;
2467 
2468 	/*
2469 	 * XXX early demux data packets so that if_input processing only handles
2470 	 * acks in interrupt context
2471 	 */
2472 	struct mbuf *m, *mh, *mt;
2473 
2474 	ifp = ctx->ifc_ifp;
2475 #ifdef DEV_NETMAP
2476 	if (ifp->if_capenable & IFCAP_NETMAP) {
2477 		u_int work = 0;
2478 		if (netmap_rx_irq(ifp, rxq->ifr_id, &work))
2479 			return (FALSE);
2480 	}
2481 #endif
2482 
2483 	mh = mt = NULL;
2484 	MPASS(budget > 0);
2485 	rx_pkts	= rx_bytes = 0;
2486 	if (sctx->isc_flags & IFLIB_HAS_RXCQ)
2487 		cidxp = &rxq->ifr_cq_cidx;
2488 	else
2489 		cidxp = &rxq->ifr_fl[0].ifl_cidx;
2490 	if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) {
2491 		for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2492 			__iflib_fl_refill_lt(ctx, fl, budget + 8);
2493 		DBG_COUNTER_INC(rx_unavail);
2494 		return (false);
2495 	}
2496 
2497 	for (budget_left = budget; (budget_left > 0) && (avail > 0); budget_left--, avail--) {
2498 		if (__predict_false(!CTX_ACTIVE(ctx))) {
2499 			DBG_COUNTER_INC(rx_ctx_inactive);
2500 			break;
2501 		}
2502 		/*
2503 		 * Reset client set fields to their default values
2504 		 */
2505 		rxd_info_zero(&ri);
2506 		ri.iri_qsidx = rxq->ifr_id;
2507 		ri.iri_cidx = *cidxp;
2508 		ri.iri_ifp = ifp;
2509 		ri.iri_frags = rxq->ifr_frags;
2510 		err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
2511 
2512 		if (err)
2513 			goto err;
2514 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
2515 			*cidxp = ri.iri_cidx;
2516 			/* Update our consumer index */
2517 			/* XXX NB: shurd - check if this is still safe */
2518 			while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0]) {
2519 				rxq->ifr_cq_cidx -= scctx->isc_nrxd[0];
2520 				rxq->ifr_cq_gen = 0;
2521 			}
2522 			/* was this only a completion queue message? */
2523 			if (__predict_false(ri.iri_nfrags == 0))
2524 				continue;
2525 		}
2526 		MPASS(ri.iri_nfrags != 0);
2527 		MPASS(ri.iri_len != 0);
2528 
2529 		/* will advance the cidx on the corresponding free lists */
2530 		m = iflib_rxd_pkt_get(rxq, &ri);
2531 		if (avail == 0 && budget_left)
2532 			avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left);
2533 
2534 		if (__predict_false(m == NULL)) {
2535 			DBG_COUNTER_INC(rx_mbuf_null);
2536 			continue;
2537 		}
2538 		/* imm_pkt: -- cxgb */
2539 		if (mh == NULL)
2540 			mh = mt = m;
2541 		else {
2542 			mt->m_nextpkt = m;
2543 			mt = m;
2544 		}
2545 	}
2546 	/* make sure that we can refill faster than drain */
2547 	for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2548 		__iflib_fl_refill_lt(ctx, fl, budget + 8);
2549 
2550 	lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO);
2551 	while (mh != NULL) {
2552 		m = mh;
2553 		mh = mh->m_nextpkt;
2554 		m->m_nextpkt = NULL;
2555 #ifndef __NO_STRICT_ALIGNMENT
2556 		if (!IP_ALIGNED(m) && (m = iflib_fixup_rx(m)) == NULL)
2557 			continue;
2558 #endif
2559 		rx_bytes += m->m_pkthdr.len;
2560 		rx_pkts++;
2561 #if defined(INET6) || defined(INET)
2562 		if (lro_enabled && tcp_lro_rx(&rxq->ifr_lc, m, 0) == 0)
2563 			continue;
2564 #endif
2565 		DBG_COUNTER_INC(rx_if_input);
2566 		ifp->if_input(ifp, m);
2567 	}
2568 
2569 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
2570 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
2571 
2572 	/*
2573 	 * Flush any outstanding LRO work
2574 	 */
2575 #if defined(INET6) || defined(INET)
2576 	tcp_lro_flush_all(&rxq->ifr_lc);
2577 #endif
2578 	if (avail)
2579 		return true;
2580 	return (iflib_rxd_avail(ctx, rxq, *cidxp, 1));
2581 err:
2582 	CTX_LOCK(ctx);
2583 	ctx->ifc_flags |= IFC_DO_RESET;
2584 	iflib_admin_intr_deferred(ctx);
2585 	CTX_UNLOCK(ctx);
2586 	return (false);
2587 }
2588 
2589 #define TXD_NOTIFY_COUNT(txq) (((txq)->ift_size / (txq)->ift_update_freq)-1)
2590 static inline qidx_t
2591 txq_max_db_deferred(iflib_txq_t txq, qidx_t in_use)
2592 {
2593 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2594 	qidx_t minthresh = txq->ift_size / 8;
2595 	if (in_use > 4*minthresh)
2596 		return (notify_count);
2597 	if (in_use > 2*minthresh)
2598 		return (notify_count >> 1);
2599 	if (in_use > minthresh)
2600 		return (notify_count >> 3);
2601 	return (0);
2602 }
2603 
2604 static inline qidx_t
2605 txq_max_rs_deferred(iflib_txq_t txq)
2606 {
2607 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2608 	qidx_t minthresh = txq->ift_size / 8;
2609 	if (txq->ift_in_use > 4*minthresh)
2610 		return (notify_count);
2611 	if (txq->ift_in_use > 2*minthresh)
2612 		return (notify_count >> 1);
2613 	if (txq->ift_in_use > minthresh)
2614 		return (notify_count >> 2);
2615 	return (2);
2616 }
2617 
2618 #define M_CSUM_FLAGS(m) ((m)->m_pkthdr.csum_flags)
2619 #define M_HAS_VLANTAG(m) (m->m_flags & M_VLANTAG)
2620 
2621 #define TXQ_MAX_DB_DEFERRED(txq, in_use) txq_max_db_deferred((txq), (in_use))
2622 #define TXQ_MAX_RS_DEFERRED(txq) txq_max_rs_deferred(txq)
2623 #define TXQ_MAX_DB_CONSUMED(size) (size >> 4)
2624 
2625 /* forward compatibility for cxgb */
2626 #define FIRST_QSET(ctx) 0
2627 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets)
2628 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets)
2629 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx))
2630 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments))
2631 
2632 /* XXX we should be setting this to something other than zero */
2633 #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh)
2634 #define MAX_TX_DESC(ctx) ((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max)
2635 
2636 static inline bool
2637 iflib_txd_db_check(if_ctx_t ctx, iflib_txq_t txq, int ring, qidx_t in_use)
2638 {
2639 	qidx_t dbval, max;
2640 	bool rang;
2641 
2642 	rang = false;
2643 	max = TXQ_MAX_DB_DEFERRED(txq, in_use);
2644 	if (ring || txq->ift_db_pending >= max) {
2645 		dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx;
2646 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval);
2647 		txq->ift_db_pending = txq->ift_npending = 0;
2648 		rang = true;
2649 	}
2650 	return (rang);
2651 }
2652 
2653 #ifdef PKT_DEBUG
2654 static void
2655 print_pkt(if_pkt_info_t pi)
2656 {
2657 	printf("pi len:  %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n",
2658 	       pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx);
2659 	printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n",
2660 	       pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag);
2661 	printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n",
2662 	       pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto);
2663 }
2664 #endif
2665 
2666 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO)
2667 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO)
2668 
2669 static int
2670 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp)
2671 {
2672 	if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx;
2673 	struct ether_vlan_header *eh;
2674 	struct mbuf *m, *n;
2675 
2676 	n = m = *mp;
2677 	if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) &&
2678 	    M_WRITABLE(m) == 0) {
2679 		if ((m = m_dup(m, M_NOWAIT)) == NULL) {
2680 			return (ENOMEM);
2681 		} else {
2682 			m_freem(*mp);
2683 			n = *mp = m;
2684 		}
2685 	}
2686 
2687 	/*
2688 	 * Determine where frame payload starts.
2689 	 * Jump over vlan headers if already present,
2690 	 * helpful for QinQ too.
2691 	 */
2692 	if (__predict_false(m->m_len < sizeof(*eh))) {
2693 		txq->ift_pullups++;
2694 		if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL))
2695 			return (ENOMEM);
2696 	}
2697 	eh = mtod(m, struct ether_vlan_header *);
2698 	if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
2699 		pi->ipi_etype = ntohs(eh->evl_proto);
2700 		pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
2701 	} else {
2702 		pi->ipi_etype = ntohs(eh->evl_encap_proto);
2703 		pi->ipi_ehdrlen = ETHER_HDR_LEN;
2704 	}
2705 
2706 	switch (pi->ipi_etype) {
2707 #ifdef INET
2708 	case ETHERTYPE_IP:
2709 	{
2710 		struct ip *ip = NULL;
2711 		struct tcphdr *th = NULL;
2712 		int minthlen;
2713 
2714 		minthlen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip) + sizeof(*th));
2715 		if (__predict_false(m->m_len < minthlen)) {
2716 			/*
2717 			 * if this code bloat is causing too much of a hit
2718 			 * move it to a separate function and mark it noinline
2719 			 */
2720 			if (m->m_len == pi->ipi_ehdrlen) {
2721 				n = m->m_next;
2722 				MPASS(n);
2723 				if (n->m_len >= sizeof(*ip))  {
2724 					ip = (struct ip *)n->m_data;
2725 					if (n->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2726 						th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2727 				} else {
2728 					txq->ift_pullups++;
2729 					if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
2730 						return (ENOMEM);
2731 					ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2732 				}
2733 			} else {
2734 				txq->ift_pullups++;
2735 				if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
2736 					return (ENOMEM);
2737 				ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2738 				if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2739 					th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2740 			}
2741 		} else {
2742 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2743 			if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2744 				th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2745 		}
2746 		pi->ipi_ip_hlen = ip->ip_hl << 2;
2747 		pi->ipi_ipproto = ip->ip_p;
2748 		pi->ipi_flags |= IPI_TX_IPV4;
2749 
2750 		if (pi->ipi_csum_flags & CSUM_IP)
2751                        ip->ip_sum = 0;
2752 
2753 		if (pi->ipi_ipproto == IPPROTO_TCP) {
2754 			if (__predict_false(th == NULL)) {
2755 				txq->ift_pullups++;
2756 				if (__predict_false((m = m_pullup(m, (ip->ip_hl << 2) + sizeof(*th))) == NULL))
2757 					return (ENOMEM);
2758 				th = (struct tcphdr *)((caddr_t)ip + pi->ipi_ip_hlen);
2759 			}
2760 			pi->ipi_tcp_hflags = th->th_flags;
2761 			pi->ipi_tcp_hlen = th->th_off << 2;
2762 			pi->ipi_tcp_seq = th->th_seq;
2763 		}
2764 		if (IS_TSO4(pi)) {
2765 			if (__predict_false(ip->ip_p != IPPROTO_TCP))
2766 				return (ENXIO);
2767 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
2768 					       ip->ip_dst.s_addr, htons(IPPROTO_TCP));
2769 			pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
2770 			if (sctx->isc_flags & IFLIB_TSO_INIT_IP) {
2771 				ip->ip_sum = 0;
2772 				ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz);
2773 			}
2774 		}
2775 		break;
2776 	}
2777 #endif
2778 #ifdef INET6
2779 	case ETHERTYPE_IPV6:
2780 	{
2781 		struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
2782 		struct tcphdr *th;
2783 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
2784 
2785 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
2786 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
2787 				return (ENOMEM);
2788 		}
2789 		th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
2790 
2791 		/* XXX-BZ this will go badly in case of ext hdrs. */
2792 		pi->ipi_ipproto = ip6->ip6_nxt;
2793 		pi->ipi_flags |= IPI_TX_IPV6;
2794 
2795 		if (pi->ipi_ipproto == IPPROTO_TCP) {
2796 			if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) {
2797 				if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL))
2798 					return (ENOMEM);
2799 			}
2800 			pi->ipi_tcp_hflags = th->th_flags;
2801 			pi->ipi_tcp_hlen = th->th_off << 2;
2802 		}
2803 		if (IS_TSO6(pi)) {
2804 
2805 			if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP))
2806 				return (ENXIO);
2807 			/*
2808 			 * The corresponding flag is set by the stack in the IPv4
2809 			 * TSO case, but not in IPv6 (at least in FreeBSD 10.2).
2810 			 * So, set it here because the rest of the flow requires it.
2811 			 */
2812 			pi->ipi_csum_flags |= CSUM_TCP_IPV6;
2813 			th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0);
2814 			pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
2815 		}
2816 		break;
2817 	}
2818 #endif
2819 	default:
2820 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
2821 		pi->ipi_ip_hlen = 0;
2822 		break;
2823 	}
2824 	*mp = m;
2825 
2826 	return (0);
2827 }
2828 
2829 static  __noinline  struct mbuf *
2830 collapse_pkthdr(struct mbuf *m0)
2831 {
2832 	struct mbuf *m, *m_next, *tmp;
2833 
2834 	m = m0;
2835 	m_next = m->m_next;
2836 	while (m_next != NULL && m_next->m_len == 0) {
2837 		m = m_next;
2838 		m->m_next = NULL;
2839 		m_free(m);
2840 		m_next = m_next->m_next;
2841 	}
2842 	m = m0;
2843 	m->m_next = m_next;
2844 	if ((m_next->m_flags & M_EXT) == 0) {
2845 		m = m_defrag(m, M_NOWAIT);
2846 	} else {
2847 		tmp = m_next->m_next;
2848 		memcpy(m_next, m, MPKTHSIZE);
2849 		m = m_next;
2850 		m->m_next = tmp;
2851 	}
2852 	return (m);
2853 }
2854 
2855 /*
2856  * If dodgy hardware rejects the scatter gather chain we've handed it
2857  * we'll need to remove the mbuf chain from ifsg_m[] before we can add the
2858  * m_defrag'd mbufs
2859  */
2860 static __noinline struct mbuf *
2861 iflib_remove_mbuf(iflib_txq_t txq)
2862 {
2863 	int ntxd, i, pidx;
2864 	struct mbuf *m, *mh, **ifsd_m;
2865 
2866 	pidx = txq->ift_pidx;
2867 	ifsd_m = txq->ift_sds.ifsd_m;
2868 	ntxd = txq->ift_size;
2869 	mh = m = ifsd_m[pidx];
2870 	ifsd_m[pidx] = NULL;
2871 #if MEMORY_LOGGING
2872 	txq->ift_dequeued++;
2873 #endif
2874 	i = 1;
2875 
2876 	while (m) {
2877 		ifsd_m[(pidx + i) & (ntxd -1)] = NULL;
2878 #if MEMORY_LOGGING
2879 		txq->ift_dequeued++;
2880 #endif
2881 		m = m->m_next;
2882 		i++;
2883 	}
2884 	return (mh);
2885 }
2886 
2887 static int
2888 iflib_busdma_load_mbuf_sg(iflib_txq_t txq, bus_dma_tag_t tag, bus_dmamap_t map,
2889 			  struct mbuf **m0, bus_dma_segment_t *segs, int *nsegs,
2890 			  int max_segs, int flags)
2891 {
2892 	if_ctx_t ctx;
2893 	if_shared_ctx_t		sctx;
2894 	if_softc_ctx_t		scctx;
2895 	int i, next, pidx, err, ntxd, count;
2896 	struct mbuf *m, *tmp, **ifsd_m;
2897 
2898 	m = *m0;
2899 
2900 	/*
2901 	 * Please don't ever do this
2902 	 */
2903 	if (__predict_false(m->m_len == 0))
2904 		*m0 = m = collapse_pkthdr(m);
2905 
2906 	ctx = txq->ift_ctx;
2907 	sctx = ctx->ifc_sctx;
2908 	scctx = &ctx->ifc_softc_ctx;
2909 	ifsd_m = txq->ift_sds.ifsd_m;
2910 	ntxd = txq->ift_size;
2911 	pidx = txq->ift_pidx;
2912 	if (map != NULL) {
2913 		uint8_t *ifsd_flags = txq->ift_sds.ifsd_flags;
2914 
2915 		err = bus_dmamap_load_mbuf_sg(tag, map,
2916 					      *m0, segs, nsegs, BUS_DMA_NOWAIT);
2917 		if (err)
2918 			return (err);
2919 		ifsd_flags[pidx] |= TX_SW_DESC_MAPPED;
2920 		count = 0;
2921 		m = *m0;
2922 		do {
2923 			if (__predict_false(m->m_len <= 0)) {
2924 				tmp = m;
2925 				m = m->m_next;
2926 				tmp->m_next = NULL;
2927 				m_free(tmp);
2928 				continue;
2929 			}
2930 			next = (pidx + count) & (ntxd-1);
2931 			MPASS(ifsd_m[next] == NULL);
2932 			ifsd_m[next] = m;
2933 			count++;
2934 			tmp = m;
2935 			m = m->m_next;
2936 		} while (m != NULL);
2937 	} else {
2938 		int buflen, sgsize, maxsegsz, max_sgsize;
2939 		vm_offset_t vaddr;
2940 		vm_paddr_t curaddr;
2941 
2942 		count = i = 0;
2943 		m = *m0;
2944 		if (m->m_pkthdr.csum_flags & CSUM_TSO)
2945 			maxsegsz = scctx->isc_tx_tso_segsize_max;
2946 		else
2947 			maxsegsz = sctx->isc_tx_maxsegsize;
2948 
2949 		do {
2950 			if (__predict_false(m->m_len <= 0)) {
2951 				tmp = m;
2952 				m = m->m_next;
2953 				tmp->m_next = NULL;
2954 				m_free(tmp);
2955 				continue;
2956 			}
2957 			buflen = m->m_len;
2958 			vaddr = (vm_offset_t)m->m_data;
2959 			/*
2960 			 * see if we can't be smarter about physically
2961 			 * contiguous mappings
2962 			 */
2963 			next = (pidx + count) & (ntxd-1);
2964 			MPASS(ifsd_m[next] == NULL);
2965 #if MEMORY_LOGGING
2966 			txq->ift_enqueued++;
2967 #endif
2968 			ifsd_m[next] = m;
2969 			while (buflen > 0) {
2970 				max_sgsize = MIN(buflen, maxsegsz);
2971 				curaddr = pmap_kextract(vaddr);
2972 				sgsize = PAGE_SIZE - (curaddr & PAGE_MASK);
2973 				sgsize = MIN(sgsize, max_sgsize);
2974 				segs[i].ds_addr = curaddr;
2975 				segs[i].ds_len = sgsize;
2976 				vaddr += sgsize;
2977 				buflen -= sgsize;
2978 				i++;
2979 				if (i >= max_segs)
2980 					goto err;
2981 			}
2982 			count++;
2983 			tmp = m;
2984 			m = m->m_next;
2985 		} while (m != NULL);
2986 		*nsegs = i;
2987 	}
2988 	return (0);
2989 err:
2990 	*m0 = iflib_remove_mbuf(txq);
2991 	return (EFBIG);
2992 }
2993 
2994 static inline caddr_t
2995 calc_next_txd(iflib_txq_t txq, int cidx, uint8_t qid)
2996 {
2997 	qidx_t size;
2998 	int ntxd;
2999 	caddr_t start, end, cur, next;
3000 
3001 	ntxd = txq->ift_size;
3002 	size = txq->ift_txd_size[qid];
3003 	start = txq->ift_ifdi[qid].idi_vaddr;
3004 
3005 	if (__predict_false(size == 0))
3006 		return (start);
3007 	cur = start + size*cidx;
3008 	end = start + size*ntxd;
3009 	next = CACHE_PTR_NEXT(cur);
3010 	return (next < end ? next : start);
3011 }
3012 
3013 static int
3014 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp)
3015 {
3016 	if_ctx_t		ctx;
3017 	if_shared_ctx_t		sctx;
3018 	if_softc_ctx_t		scctx;
3019 	bus_dma_segment_t	*segs;
3020 	struct mbuf		*m_head;
3021 	void			*next_txd;
3022 	bus_dmamap_t		map;
3023 	struct if_pkt_info	pi;
3024 	int remap = 0;
3025 	int err, nsegs, ndesc, max_segs, pidx, cidx, next, ntxd;
3026 	bus_dma_tag_t desc_tag;
3027 
3028 	segs = txq->ift_segs;
3029 	ctx = txq->ift_ctx;
3030 	sctx = ctx->ifc_sctx;
3031 	scctx = &ctx->ifc_softc_ctx;
3032 	segs = txq->ift_segs;
3033 	ntxd = txq->ift_size;
3034 	m_head = *m_headp;
3035 	map = NULL;
3036 
3037 	/*
3038 	 * If we're doing TSO the next descriptor to clean may be quite far ahead
3039 	 */
3040 	cidx = txq->ift_cidx;
3041 	pidx = txq->ift_pidx;
3042 	if (ctx->ifc_flags & IFC_PREFETCH) {
3043 		next = (cidx + CACHE_PTR_INCREMENT) & (ntxd-1);
3044 		if (!(ctx->ifc_flags & IFLIB_HAS_TXCQ)) {
3045 			next_txd = calc_next_txd(txq, cidx, 0);
3046 			prefetch(next_txd);
3047 		}
3048 
3049 		/* prefetch the next cache line of mbuf pointers and flags */
3050 		prefetch(&txq->ift_sds.ifsd_m[next]);
3051 		if (txq->ift_sds.ifsd_map != NULL) {
3052 			prefetch(&txq->ift_sds.ifsd_map[next]);
3053 			next = (cidx + CACHE_LINE_SIZE) & (ntxd-1);
3054 			prefetch(&txq->ift_sds.ifsd_flags[next]);
3055 		}
3056 	} else if (txq->ift_sds.ifsd_map != NULL)
3057 		map = txq->ift_sds.ifsd_map[pidx];
3058 
3059 	if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
3060 		desc_tag = txq->ift_tso_desc_tag;
3061 		max_segs = scctx->isc_tx_tso_segments_max;
3062 	} else {
3063 		desc_tag = txq->ift_desc_tag;
3064 		max_segs = scctx->isc_tx_nsegments;
3065 	}
3066 	m_head = *m_headp;
3067 
3068 	pkt_info_zero(&pi);
3069 	pi.ipi_len = m_head->m_pkthdr.len;
3070 	pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG|M_BCAST|M_MCAST));
3071 	pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags;
3072 	pi.ipi_vtag = (m_head->m_flags & M_VLANTAG) ? m_head->m_pkthdr.ether_vtag : 0;
3073 	pi.ipi_pidx = pidx;
3074 	pi.ipi_qsidx = txq->ift_id;
3075 
3076 	/* deliberate bitwise OR to make one condition */
3077 	if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) {
3078 		if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0))
3079 			return (err);
3080 		m_head = *m_headp;
3081 	}
3082 
3083 retry:
3084 	err = iflib_busdma_load_mbuf_sg(txq, desc_tag, map, m_headp, segs, &nsegs, max_segs, BUS_DMA_NOWAIT);
3085 defrag:
3086 	if (__predict_false(err)) {
3087 		switch (err) {
3088 		case EFBIG:
3089 			/* try collapse once and defrag once */
3090 			if (remap == 0)
3091 				m_head = m_collapse(*m_headp, M_NOWAIT, max_segs);
3092 			if (remap == 1)
3093 				m_head = m_defrag(*m_headp, M_NOWAIT);
3094 			remap++;
3095 			if (__predict_false(m_head == NULL))
3096 				goto defrag_failed;
3097 			txq->ift_mbuf_defrag++;
3098 			*m_headp = m_head;
3099 			goto retry;
3100 			break;
3101 		case ENOMEM:
3102 			txq->ift_no_tx_dma_setup++;
3103 			break;
3104 		default:
3105 			txq->ift_no_tx_dma_setup++;
3106 			m_freem(*m_headp);
3107 			DBG_COUNTER_INC(tx_frees);
3108 			*m_headp = NULL;
3109 			break;
3110 		}
3111 		txq->ift_map_failed++;
3112 		DBG_COUNTER_INC(encap_load_mbuf_fail);
3113 		return (err);
3114 	}
3115 
3116 	/*
3117 	 * XXX assumes a 1 to 1 relationship between segments and
3118 	 *        descriptors - this does not hold true on all drivers, e.g.
3119 	 *        cxgb
3120 	 */
3121 	if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) {
3122 		txq->ift_no_desc_avail++;
3123 		if (map != NULL)
3124 			bus_dmamap_unload(desc_tag, map);
3125 		DBG_COUNTER_INC(encap_txq_avail_fail);
3126 		if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0)
3127 			GROUPTASK_ENQUEUE(&txq->ift_task);
3128 		return (ENOBUFS);
3129 	}
3130 	/*
3131 	 * On Intel cards we can greatly reduce the number of TX interrupts
3132 	 * we see by only setting report status on every Nth descriptor.
3133 	 * However, this also means that the driver will need to keep track
3134 	 * of the descriptors that RS was set on to check them for the DD bit.
3135 	 */
3136 	txq->ift_rs_pending += nsegs + 1;
3137 	if (txq->ift_rs_pending > TXQ_MAX_RS_DEFERRED(txq) ||
3138 	     iflib_no_tx_batch || (TXQ_AVAIL(txq) - nsegs - 1) <= MAX_TX_DESC(ctx)) {
3139 		pi.ipi_flags |= IPI_TX_INTR;
3140 		txq->ift_rs_pending = 0;
3141 	}
3142 
3143 	pi.ipi_segs = segs;
3144 	pi.ipi_nsegs = nsegs;
3145 
3146 	MPASS(pidx >= 0 && pidx < txq->ift_size);
3147 #ifdef PKT_DEBUG
3148 	print_pkt(&pi);
3149 #endif
3150 	if (map != NULL)
3151 		bus_dmamap_sync(desc_tag, map, BUS_DMASYNC_PREWRITE);
3152 	if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) {
3153 		if (map != NULL)
3154 			bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3155 					BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
3156 		DBG_COUNTER_INC(tx_encap);
3157 		MPASS(pi.ipi_new_pidx < txq->ift_size);
3158 
3159 		ndesc = pi.ipi_new_pidx - pi.ipi_pidx;
3160 		if (pi.ipi_new_pidx < pi.ipi_pidx) {
3161 			ndesc += txq->ift_size;
3162 			txq->ift_gen = 1;
3163 		}
3164 		/*
3165 		 * drivers can need as many as
3166 		 * two sentinels
3167 		 */
3168 		MPASS(ndesc <= pi.ipi_nsegs + 2);
3169 		MPASS(pi.ipi_new_pidx != pidx);
3170 		MPASS(ndesc > 0);
3171 		txq->ift_in_use += ndesc;
3172 
3173 		/*
3174 		 * We update the last software descriptor again here because there may
3175 		 * be a sentinel and/or there may be more mbufs than segments
3176 		 */
3177 		txq->ift_pidx = pi.ipi_new_pidx;
3178 		txq->ift_npending += pi.ipi_ndescs;
3179 	} else if (__predict_false(err == EFBIG && remap < 2)) {
3180 		*m_headp = m_head = iflib_remove_mbuf(txq);
3181 		remap = 1;
3182 		txq->ift_txd_encap_efbig++;
3183 		goto defrag;
3184 	} else
3185 		DBG_COUNTER_INC(encap_txd_encap_fail);
3186 	return (err);
3187 
3188 defrag_failed:
3189 	txq->ift_mbuf_defrag_failed++;
3190 	txq->ift_map_failed++;
3191 	m_freem(*m_headp);
3192 	DBG_COUNTER_INC(tx_frees);
3193 	*m_headp = NULL;
3194 	return (ENOMEM);
3195 }
3196 
3197 static void
3198 iflib_tx_desc_free(iflib_txq_t txq, int n)
3199 {
3200 	int hasmap;
3201 	uint32_t qsize, cidx, mask, gen;
3202 	struct mbuf *m, **ifsd_m;
3203 	uint8_t *ifsd_flags;
3204 	bus_dmamap_t *ifsd_map;
3205 	bool do_prefetch;
3206 
3207 	cidx = txq->ift_cidx;
3208 	gen = txq->ift_gen;
3209 	qsize = txq->ift_size;
3210 	mask = qsize-1;
3211 	hasmap = txq->ift_sds.ifsd_map != NULL;
3212 	ifsd_flags = txq->ift_sds.ifsd_flags;
3213 	ifsd_m = txq->ift_sds.ifsd_m;
3214 	ifsd_map = txq->ift_sds.ifsd_map;
3215 	do_prefetch = (txq->ift_ctx->ifc_flags & IFC_PREFETCH);
3216 
3217 	while (n--) {
3218 		if (do_prefetch) {
3219 			prefetch(ifsd_m[(cidx + 3) & mask]);
3220 			prefetch(ifsd_m[(cidx + 4) & mask]);
3221 		}
3222 		if (ifsd_m[cidx] != NULL) {
3223 			prefetch(&ifsd_m[(cidx + CACHE_PTR_INCREMENT) & mask]);
3224 			prefetch(&ifsd_flags[(cidx + CACHE_PTR_INCREMENT) & mask]);
3225 			if (hasmap && (ifsd_flags[cidx] & TX_SW_DESC_MAPPED)) {
3226 				/*
3227 				 * does it matter if it's not the TSO tag? If so we'll
3228 				 * have to add the type to flags
3229 				 */
3230 				bus_dmamap_unload(txq->ift_desc_tag, ifsd_map[cidx]);
3231 				ifsd_flags[cidx] &= ~TX_SW_DESC_MAPPED;
3232 			}
3233 			if ((m = ifsd_m[cidx]) != NULL) {
3234 				/* XXX we don't support any drivers that batch packets yet */
3235 				MPASS(m->m_nextpkt == NULL);
3236 
3237 				m_free(m);
3238 				ifsd_m[cidx] = NULL;
3239 #if MEMORY_LOGGING
3240 				txq->ift_dequeued++;
3241 #endif
3242 				DBG_COUNTER_INC(tx_frees);
3243 			}
3244 		}
3245 		if (__predict_false(++cidx == qsize)) {
3246 			cidx = 0;
3247 			gen = 0;
3248 		}
3249 	}
3250 	txq->ift_cidx = cidx;
3251 	txq->ift_gen = gen;
3252 }
3253 
3254 static __inline int
3255 iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh)
3256 {
3257 	int reclaim;
3258 	if_ctx_t ctx = txq->ift_ctx;
3259 
3260 	KASSERT(thresh >= 0, ("invalid threshold to reclaim"));
3261 	MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size);
3262 
3263 	/*
3264 	 * Need a rate-limiting check so that this isn't called every time
3265 	 */
3266 	iflib_tx_credits_update(ctx, txq);
3267 	reclaim = DESC_RECLAIMABLE(txq);
3268 
3269 	if (reclaim <= thresh /* + MAX_TX_DESC(txq->ift_ctx) */) {
3270 #ifdef INVARIANTS
3271 		if (iflib_verbose_debug) {
3272 			printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __FUNCTION__,
3273 			       txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments,
3274 			       reclaim, thresh);
3275 
3276 		}
3277 #endif
3278 		return (0);
3279 	}
3280 	iflib_tx_desc_free(txq, reclaim);
3281 	txq->ift_cleaned += reclaim;
3282 	txq->ift_in_use -= reclaim;
3283 
3284 	return (reclaim);
3285 }
3286 
3287 static struct mbuf **
3288 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset, int remaining)
3289 {
3290 	int next, size;
3291 	struct mbuf **items;
3292 
3293 	size = r->size;
3294 	next = (cidx + CACHE_PTR_INCREMENT) & (size-1);
3295 	items = __DEVOLATILE(struct mbuf **, &r->items[0]);
3296 
3297 	prefetch(items[(cidx + offset) & (size-1)]);
3298 	if (remaining > 1) {
3299 		prefetch(&items[next]);
3300 		prefetch(items[(cidx + offset + 1) & (size-1)]);
3301 		prefetch(items[(cidx + offset + 2) & (size-1)]);
3302 		prefetch(items[(cidx + offset + 3) & (size-1)]);
3303 	}
3304 	return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (size-1)]));
3305 }
3306 
3307 static void
3308 iflib_txq_check_drain(iflib_txq_t txq, int budget)
3309 {
3310 
3311 	ifmp_ring_check_drainage(txq->ift_br, budget);
3312 }
3313 
3314 static uint32_t
3315 iflib_txq_can_drain(struct ifmp_ring *r)
3316 {
3317 	iflib_txq_t txq = r->cookie;
3318 	if_ctx_t ctx = txq->ift_ctx;
3319 
3320 	return ((TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2) ||
3321 		ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false));
3322 }
3323 
3324 static uint32_t
3325 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3326 {
3327 	iflib_txq_t txq = r->cookie;
3328 	if_ctx_t ctx = txq->ift_ctx;
3329 	struct ifnet *ifp = ctx->ifc_ifp;
3330 	struct mbuf **mp, *m;
3331 	int i, count, consumed, pkt_sent, bytes_sent, mcast_sent, avail;
3332 	int reclaimed, err, in_use_prev, desc_used;
3333 	bool do_prefetch, ring, rang;
3334 
3335 	if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) ||
3336 			    !LINK_ACTIVE(ctx))) {
3337 		DBG_COUNTER_INC(txq_drain_notready);
3338 		return (0);
3339 	}
3340 	reclaimed = iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
3341 	rang = iflib_txd_db_check(ctx, txq, reclaimed, txq->ift_in_use);
3342 	avail = IDXDIFF(pidx, cidx, r->size);
3343 	if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) {
3344 		DBG_COUNTER_INC(txq_drain_flushing);
3345 		for (i = 0; i < avail; i++) {
3346 			m_free(r->items[(cidx + i) & (r->size-1)]);
3347 			r->items[(cidx + i) & (r->size-1)] = NULL;
3348 		}
3349 		return (avail);
3350 	}
3351 
3352 	if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3353 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3354 		CALLOUT_LOCK(txq);
3355 		callout_stop(&txq->ift_timer);
3356 		CALLOUT_UNLOCK(txq);
3357 		DBG_COUNTER_INC(txq_drain_oactive);
3358 		return (0);
3359 	}
3360 	if (reclaimed)
3361 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3362 	consumed = mcast_sent = bytes_sent = pkt_sent = 0;
3363 	count = MIN(avail, TX_BATCH_SIZE);
3364 #ifdef INVARIANTS
3365 	if (iflib_verbose_debug)
3366 		printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __FUNCTION__,
3367 		       avail, ctx->ifc_flags, TXQ_AVAIL(txq));
3368 #endif
3369 	do_prefetch = (ctx->ifc_flags & IFC_PREFETCH);
3370 	avail = TXQ_AVAIL(txq);
3371 	for (desc_used = i = 0; i < count && avail > MAX_TX_DESC(ctx) + 2; i++) {
3372 		int pidx_prev, rem = do_prefetch ? count - i : 0;
3373 
3374 		mp = _ring_peek_one(r, cidx, i, rem);
3375 		MPASS(mp != NULL && *mp != NULL);
3376 		if (__predict_false(*mp == (struct mbuf *)txq)) {
3377 			consumed++;
3378 			reclaimed++;
3379 			continue;
3380 		}
3381 		in_use_prev = txq->ift_in_use;
3382 		pidx_prev = txq->ift_pidx;
3383 		err = iflib_encap(txq, mp);
3384 		if (__predict_false(err)) {
3385 			DBG_COUNTER_INC(txq_drain_encapfail);
3386 			/* no room - bail out */
3387 			if (err == ENOBUFS)
3388 				break;
3389 			consumed++;
3390 			DBG_COUNTER_INC(txq_drain_encapfail);
3391 			/* we can't send this packet - skip it */
3392 			continue;
3393 		}
3394 		consumed++;
3395 		pkt_sent++;
3396 		m = *mp;
3397 		DBG_COUNTER_INC(tx_sent);
3398 		bytes_sent += m->m_pkthdr.len;
3399 		mcast_sent += !!(m->m_flags & M_MCAST);
3400 		avail = TXQ_AVAIL(txq);
3401 
3402 		txq->ift_db_pending += (txq->ift_in_use - in_use_prev);
3403 		desc_used += (txq->ift_in_use - in_use_prev);
3404 		ETHER_BPF_MTAP(ifp, m);
3405 		if (__predict_false(!(ifp->if_drv_flags & IFF_DRV_RUNNING)))
3406 			break;
3407 		rang = iflib_txd_db_check(ctx, txq, false, in_use_prev);
3408 	}
3409 
3410 	/* deliberate use of bitwise or to avoid gratuitous short-circuit */
3411 	ring = rang ? false  : (iflib_min_tx_latency | err) || (TXQ_AVAIL(txq) < MAX_TX_DESC(ctx));
3412 	iflib_txd_db_check(ctx, txq, ring, txq->ift_in_use);
3413 	if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
3414 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
3415 	if (mcast_sent)
3416 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
3417 #ifdef INVARIANTS
3418 	if (iflib_verbose_debug)
3419 		printf("consumed=%d\n", consumed);
3420 #endif
3421 	return (consumed);
3422 }
3423 
3424 static uint32_t
3425 iflib_txq_drain_always(struct ifmp_ring *r)
3426 {
3427 	return (1);
3428 }
3429 
3430 static uint32_t
3431 iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3432 {
3433 	int i, avail;
3434 	struct mbuf **mp;
3435 	iflib_txq_t txq;
3436 
3437 	txq = r->cookie;
3438 
3439 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3440 	CALLOUT_LOCK(txq);
3441 	callout_stop(&txq->ift_timer);
3442 	CALLOUT_UNLOCK(txq);
3443 
3444 	avail = IDXDIFF(pidx, cidx, r->size);
3445 	for (i = 0; i < avail; i++) {
3446 		mp = _ring_peek_one(r, cidx, i, avail - i);
3447 		if (__predict_false(*mp == (struct mbuf *)txq))
3448 			continue;
3449 		m_freem(*mp);
3450 	}
3451 	MPASS(ifmp_ring_is_stalled(r) == 0);
3452 	return (avail);
3453 }
3454 
3455 static void
3456 iflib_ifmp_purge(iflib_txq_t txq)
3457 {
3458 	struct ifmp_ring *r;
3459 
3460 	r = txq->ift_br;
3461 	r->drain = iflib_txq_drain_free;
3462 	r->can_drain = iflib_txq_drain_always;
3463 
3464 	ifmp_ring_check_drainage(r, r->size);
3465 
3466 	r->drain = iflib_txq_drain;
3467 	r->can_drain = iflib_txq_can_drain;
3468 }
3469 
3470 static void
3471 _task_fn_tx(void *context)
3472 {
3473 	iflib_txq_t txq = context;
3474 	if_ctx_t ctx = txq->ift_ctx;
3475 	struct ifnet *ifp = ctx->ifc_ifp;
3476 	int rc;
3477 
3478 #ifdef IFLIB_DIAGNOSTICS
3479 	txq->ift_cpu_exec_count[curcpu]++;
3480 #endif
3481 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3482 		return;
3483 	if ((ifp->if_capenable & IFCAP_NETMAP)) {
3484 		if (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false))
3485 			netmap_tx_irq(ifp, txq->ift_id);
3486 		IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3487 		return;
3488 	}
3489 	if (txq->ift_db_pending)
3490 		ifmp_ring_enqueue(txq->ift_br, (void **)&txq, 1, TX_BATCH_SIZE);
3491 	else
3492 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3493 	if (ctx->ifc_flags & IFC_LEGACY)
3494 		IFDI_INTR_ENABLE(ctx);
3495 	else {
3496 		rc = IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3497 		KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver"));
3498 	}
3499 }
3500 
3501 static void
3502 _task_fn_rx(void *context)
3503 {
3504 	iflib_rxq_t rxq = context;
3505 	if_ctx_t ctx = rxq->ifr_ctx;
3506 	bool more;
3507 	int rc;
3508 
3509 #ifdef IFLIB_DIAGNOSTICS
3510 	rxq->ifr_cpu_exec_count[curcpu]++;
3511 #endif
3512 	DBG_COUNTER_INC(task_fn_rxs);
3513 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3514 		return;
3515 	if ((more = iflib_rxeof(rxq, 16 /* XXX */)) == false) {
3516 		if (ctx->ifc_flags & IFC_LEGACY)
3517 			IFDI_INTR_ENABLE(ctx);
3518 		else {
3519 			DBG_COUNTER_INC(rx_intr_enables);
3520 			rc = IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
3521 			KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver"));
3522 		}
3523 	}
3524 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3525 		return;
3526 	if (more)
3527 		GROUPTASK_ENQUEUE(&rxq->ifr_task);
3528 }
3529 
3530 static void
3531 _task_fn_admin(void *context)
3532 {
3533 	if_ctx_t ctx = context;
3534 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
3535 	iflib_txq_t txq;
3536 	int i;
3537 
3538 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)) {
3539 		if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3540 			return;
3541 		}
3542 	}
3543 
3544 	CTX_LOCK(ctx);
3545 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3546 		CALLOUT_LOCK(txq);
3547 		callout_stop(&txq->ift_timer);
3548 		CALLOUT_UNLOCK(txq);
3549 	}
3550 	IFDI_UPDATE_ADMIN_STATUS(ctx);
3551 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3552 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq, txq->ift_timer.c_cpu);
3553 	IFDI_LINK_INTR_ENABLE(ctx);
3554 	if (ctx->ifc_flags & IFC_DO_RESET) {
3555 		ctx->ifc_flags &= ~IFC_DO_RESET;
3556 		iflib_if_init_locked(ctx);
3557 	}
3558 	CTX_UNLOCK(ctx);
3559 
3560 	if (LINK_ACTIVE(ctx) == 0)
3561 		return;
3562 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3563 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
3564 }
3565 
3566 
3567 static void
3568 _task_fn_iov(void *context)
3569 {
3570 	if_ctx_t ctx = context;
3571 
3572 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3573 		return;
3574 
3575 	CTX_LOCK(ctx);
3576 	IFDI_VFLR_HANDLE(ctx);
3577 	CTX_UNLOCK(ctx);
3578 }
3579 
3580 static int
3581 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)
3582 {
3583 	int err;
3584 	if_int_delay_info_t info;
3585 	if_ctx_t ctx;
3586 
3587 	info = (if_int_delay_info_t)arg1;
3588 	ctx = info->iidi_ctx;
3589 	info->iidi_req = req;
3590 	info->iidi_oidp = oidp;
3591 	CTX_LOCK(ctx);
3592 	err = IFDI_SYSCTL_INT_DELAY(ctx, info);
3593 	CTX_UNLOCK(ctx);
3594 	return (err);
3595 }
3596 
3597 /*********************************************************************
3598  *
3599  *  IFNET FUNCTIONS
3600  *
3601  **********************************************************************/
3602 
3603 static void
3604 iflib_if_init_locked(if_ctx_t ctx)
3605 {
3606 	iflib_stop(ctx);
3607 	iflib_init_locked(ctx);
3608 }
3609 
3610 
3611 static void
3612 iflib_if_init(void *arg)
3613 {
3614 	if_ctx_t ctx = arg;
3615 
3616 	CTX_LOCK(ctx);
3617 	iflib_if_init_locked(ctx);
3618 	CTX_UNLOCK(ctx);
3619 }
3620 
3621 static int
3622 iflib_if_transmit(if_t ifp, struct mbuf *m)
3623 {
3624 	if_ctx_t	ctx = if_getsoftc(ifp);
3625 
3626 	iflib_txq_t txq;
3627 	int err, qidx;
3628 
3629 	if (__predict_false((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) {
3630 		DBG_COUNTER_INC(tx_frees);
3631 		m_freem(m);
3632 		return (ENOBUFS);
3633 	}
3634 
3635 	MPASS(m->m_nextpkt == NULL);
3636 	qidx = 0;
3637 	if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m))
3638 		qidx = QIDX(ctx, m);
3639 	/*
3640 	 * XXX calculate buf_ring based on flowid (divvy up bits?)
3641 	 */
3642 	txq = &ctx->ifc_txqs[qidx];
3643 
3644 #ifdef DRIVER_BACKPRESSURE
3645 	if (txq->ift_closed) {
3646 		while (m != NULL) {
3647 			next = m->m_nextpkt;
3648 			m->m_nextpkt = NULL;
3649 			m_freem(m);
3650 			m = next;
3651 		}
3652 		return (ENOBUFS);
3653 	}
3654 #endif
3655 #ifdef notyet
3656 	qidx = count = 0;
3657 	mp = marr;
3658 	next = m;
3659 	do {
3660 		count++;
3661 		next = next->m_nextpkt;
3662 	} while (next != NULL);
3663 
3664 	if (count > nitems(marr))
3665 		if ((mp = malloc(count*sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) {
3666 			/* XXX check nextpkt */
3667 			m_freem(m);
3668 			/* XXX simplify for now */
3669 			DBG_COUNTER_INC(tx_frees);
3670 			return (ENOBUFS);
3671 		}
3672 	for (next = m, i = 0; next != NULL; i++) {
3673 		mp[i] = next;
3674 		next = next->m_nextpkt;
3675 		mp[i]->m_nextpkt = NULL;
3676 	}
3677 #endif
3678 	DBG_COUNTER_INC(tx_seen);
3679 	err = ifmp_ring_enqueue(txq->ift_br, (void **)&m, 1, TX_BATCH_SIZE);
3680 
3681 	if (err) {
3682 		GROUPTASK_ENQUEUE(&txq->ift_task);
3683 		/* support forthcoming later */
3684 #ifdef DRIVER_BACKPRESSURE
3685 		txq->ift_closed = TRUE;
3686 #endif
3687 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3688 		m_freem(m);
3689 	} else if (TXQ_AVAIL(txq) < (txq->ift_size >> 1)) {
3690 		GROUPTASK_ENQUEUE(&txq->ift_task);
3691 	}
3692 
3693 	return (err);
3694 }
3695 
3696 static void
3697 iflib_if_qflush(if_t ifp)
3698 {
3699 	if_ctx_t ctx = if_getsoftc(ifp);
3700 	iflib_txq_t txq = ctx->ifc_txqs;
3701 	int i;
3702 
3703 	CTX_LOCK(ctx);
3704 	ctx->ifc_flags |= IFC_QFLUSH;
3705 	CTX_UNLOCK(ctx);
3706 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
3707 		while (!(ifmp_ring_is_idle(txq->ift_br) || ifmp_ring_is_stalled(txq->ift_br)))
3708 			iflib_txq_check_drain(txq, 0);
3709 	CTX_LOCK(ctx);
3710 	ctx->ifc_flags &= ~IFC_QFLUSH;
3711 	CTX_UNLOCK(ctx);
3712 
3713 	if_qflush(ifp);
3714 }
3715 
3716 
3717 #define IFCAP_FLAGS (IFCAP_TXCSUM_IPV6 | IFCAP_RXCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \
3718 		     IFCAP_TSO4 | IFCAP_TSO6 | IFCAP_VLAN_HWTAGGING |	\
3719 		     IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | IFCAP_VLAN_HWTSO)
3720 
3721 static int
3722 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data)
3723 {
3724 	if_ctx_t ctx = if_getsoftc(ifp);
3725 	struct ifreq	*ifr = (struct ifreq *)data;
3726 #if defined(INET) || defined(INET6)
3727 	struct ifaddr	*ifa = (struct ifaddr *)data;
3728 #endif
3729 	bool		avoid_reset = FALSE;
3730 	int		err = 0, reinit = 0, bits;
3731 
3732 	switch (command) {
3733 	case SIOCSIFADDR:
3734 #ifdef INET
3735 		if (ifa->ifa_addr->sa_family == AF_INET)
3736 			avoid_reset = TRUE;
3737 #endif
3738 #ifdef INET6
3739 		if (ifa->ifa_addr->sa_family == AF_INET6)
3740 			avoid_reset = TRUE;
3741 #endif
3742 		/*
3743 		** Calling init results in link renegotiation,
3744 		** so we avoid doing it when possible.
3745 		*/
3746 		if (avoid_reset) {
3747 			if_setflagbits(ifp, IFF_UP,0);
3748 			if (!(if_getdrvflags(ifp)& IFF_DRV_RUNNING))
3749 				reinit = 1;
3750 #ifdef INET
3751 			if (!(if_getflags(ifp) & IFF_NOARP))
3752 				arp_ifinit(ifp, ifa);
3753 #endif
3754 		} else
3755 			err = ether_ioctl(ifp, command, data);
3756 		break;
3757 	case SIOCSIFMTU:
3758 		CTX_LOCK(ctx);
3759 		if (ifr->ifr_mtu == if_getmtu(ifp)) {
3760 			CTX_UNLOCK(ctx);
3761 			break;
3762 		}
3763 		bits = if_getdrvflags(ifp);
3764 		/* stop the driver and free any clusters before proceeding */
3765 		iflib_stop(ctx);
3766 
3767 		if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) {
3768 			if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size)
3769 				ctx->ifc_flags |= IFC_MULTISEG;
3770 			else
3771 				ctx->ifc_flags &= ~IFC_MULTISEG;
3772 			err = if_setmtu(ifp, ifr->ifr_mtu);
3773 		}
3774 		iflib_init_locked(ctx);
3775 		if_setdrvflags(ifp, bits);
3776 		CTX_UNLOCK(ctx);
3777 		break;
3778 	case SIOCSIFFLAGS:
3779 		CTX_LOCK(ctx);
3780 		if (if_getflags(ifp) & IFF_UP) {
3781 			if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
3782 				if ((if_getflags(ifp) ^ ctx->ifc_if_flags) &
3783 				    (IFF_PROMISC | IFF_ALLMULTI)) {
3784 					err = IFDI_PROMISC_SET(ctx, if_getflags(ifp));
3785 				}
3786 			} else
3787 				reinit = 1;
3788 		} else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
3789 			iflib_stop(ctx);
3790 		}
3791 		ctx->ifc_if_flags = if_getflags(ifp);
3792 		CTX_UNLOCK(ctx);
3793 		break;
3794 	case SIOCADDMULTI:
3795 	case SIOCDELMULTI:
3796 		if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
3797 			CTX_LOCK(ctx);
3798 			IFDI_INTR_DISABLE(ctx);
3799 			IFDI_MULTI_SET(ctx);
3800 			IFDI_INTR_ENABLE(ctx);
3801 			CTX_UNLOCK(ctx);
3802 		}
3803 		break;
3804 	case SIOCSIFMEDIA:
3805 		CTX_LOCK(ctx);
3806 		IFDI_MEDIA_SET(ctx);
3807 		CTX_UNLOCK(ctx);
3808 		/* falls thru */
3809 	case SIOCGIFMEDIA:
3810 		err = ifmedia_ioctl(ifp, ifr, &ctx->ifc_media, command);
3811 		break;
3812 	case SIOCGI2C:
3813 	{
3814 		struct ifi2creq i2c;
3815 
3816 		err = copyin(ifr->ifr_data, &i2c, sizeof(i2c));
3817 		if (err != 0)
3818 			break;
3819 		if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) {
3820 			err = EINVAL;
3821 			break;
3822 		}
3823 		if (i2c.len > sizeof(i2c.data)) {
3824 			err = EINVAL;
3825 			break;
3826 		}
3827 
3828 		if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0)
3829 			err = copyout(&i2c, ifr->ifr_data, sizeof(i2c));
3830 		break;
3831 	}
3832 	case SIOCSIFCAP:
3833 	{
3834 		int mask, setmask;
3835 
3836 		mask = ifr->ifr_reqcap ^ if_getcapenable(ifp);
3837 		setmask = 0;
3838 #ifdef TCP_OFFLOAD
3839 		setmask |= mask & (IFCAP_TOE4|IFCAP_TOE6);
3840 #endif
3841 		setmask |= (mask & IFCAP_FLAGS);
3842 
3843 		if (setmask  & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6))
3844 			setmask |= (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6);
3845 		if ((mask & IFCAP_WOL) &&
3846 		    (if_getcapabilities(ifp) & IFCAP_WOL) != 0)
3847 			setmask |= (mask & (IFCAP_WOL_MCAST|IFCAP_WOL_MAGIC));
3848 		if_vlancap(ifp);
3849 		/*
3850 		 * want to ensure that traffic has stopped before we change any of the flags
3851 		 */
3852 		if (setmask) {
3853 			CTX_LOCK(ctx);
3854 			bits = if_getdrvflags(ifp);
3855 			if (bits & IFF_DRV_RUNNING)
3856 				iflib_stop(ctx);
3857 			if_togglecapenable(ifp, setmask);
3858 			if (bits & IFF_DRV_RUNNING)
3859 				iflib_init_locked(ctx);
3860 			if_setdrvflags(ifp, bits);
3861 			CTX_UNLOCK(ctx);
3862 		}
3863 		break;
3864 	    }
3865 	case SIOCGPRIVATE_0:
3866 	case SIOCSDRVSPEC:
3867 	case SIOCGDRVSPEC:
3868 		CTX_LOCK(ctx);
3869 		err = IFDI_PRIV_IOCTL(ctx, command, data);
3870 		CTX_UNLOCK(ctx);
3871 		break;
3872 	default:
3873 		err = ether_ioctl(ifp, command, data);
3874 		break;
3875 	}
3876 	if (reinit)
3877 		iflib_if_init(ctx);
3878 	return (err);
3879 }
3880 
3881 static uint64_t
3882 iflib_if_get_counter(if_t ifp, ift_counter cnt)
3883 {
3884 	if_ctx_t ctx = if_getsoftc(ifp);
3885 
3886 	return (IFDI_GET_COUNTER(ctx, cnt));
3887 }
3888 
3889 /*********************************************************************
3890  *
3891  *  OTHER FUNCTIONS EXPORTED TO THE STACK
3892  *
3893  **********************************************************************/
3894 
3895 static void
3896 iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag)
3897 {
3898 	if_ctx_t ctx = if_getsoftc(ifp);
3899 
3900 	if ((void *)ctx != arg)
3901 		return;
3902 
3903 	if ((vtag == 0) || (vtag > 4095))
3904 		return;
3905 
3906 	CTX_LOCK(ctx);
3907 	IFDI_VLAN_REGISTER(ctx, vtag);
3908 	/* Re-init to load the changes */
3909 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
3910 		iflib_init_locked(ctx);
3911 	CTX_UNLOCK(ctx);
3912 }
3913 
3914 static void
3915 iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag)
3916 {
3917 	if_ctx_t ctx = if_getsoftc(ifp);
3918 
3919 	if ((void *)ctx != arg)
3920 		return;
3921 
3922 	if ((vtag == 0) || (vtag > 4095))
3923 		return;
3924 
3925 	CTX_LOCK(ctx);
3926 	IFDI_VLAN_UNREGISTER(ctx, vtag);
3927 	/* Re-init to load the changes */
3928 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
3929 		iflib_init_locked(ctx);
3930 	CTX_UNLOCK(ctx);
3931 }
3932 
3933 static void
3934 iflib_led_func(void *arg, int onoff)
3935 {
3936 	if_ctx_t ctx = arg;
3937 
3938 	CTX_LOCK(ctx);
3939 	IFDI_LED_FUNC(ctx, onoff);
3940 	CTX_UNLOCK(ctx);
3941 }
3942 
3943 /*********************************************************************
3944  *
3945  *  BUS FUNCTION DEFINITIONS
3946  *
3947  **********************************************************************/
3948 
3949 int
3950 iflib_device_probe(device_t dev)
3951 {
3952 	pci_vendor_info_t *ent;
3953 
3954 	uint16_t	pci_vendor_id, pci_device_id;
3955 	uint16_t	pci_subvendor_id, pci_subdevice_id;
3956 	uint16_t	pci_rev_id;
3957 	if_shared_ctx_t sctx;
3958 
3959 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
3960 		return (ENOTSUP);
3961 
3962 	pci_vendor_id = pci_get_vendor(dev);
3963 	pci_device_id = pci_get_device(dev);
3964 	pci_subvendor_id = pci_get_subvendor(dev);
3965 	pci_subdevice_id = pci_get_subdevice(dev);
3966 	pci_rev_id = pci_get_revid(dev);
3967 	if (sctx->isc_parse_devinfo != NULL)
3968 		sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id);
3969 
3970 	ent = sctx->isc_vendor_info;
3971 	while (ent->pvi_vendor_id != 0) {
3972 		if (pci_vendor_id != ent->pvi_vendor_id) {
3973 			ent++;
3974 			continue;
3975 		}
3976 		if ((pci_device_id == ent->pvi_device_id) &&
3977 		    ((pci_subvendor_id == ent->pvi_subvendor_id) ||
3978 		     (ent->pvi_subvendor_id == 0)) &&
3979 		    ((pci_subdevice_id == ent->pvi_subdevice_id) ||
3980 		     (ent->pvi_subdevice_id == 0)) &&
3981 		    ((pci_rev_id == ent->pvi_rev_id) ||
3982 		     (ent->pvi_rev_id == 0))) {
3983 
3984 			device_set_desc_copy(dev, ent->pvi_name);
3985 			/* this needs to be changed to zero if the bus probing code
3986 			 * ever stops re-probing on best match because the sctx
3987 			 * may have its values over written by register calls
3988 			 * in subsequent probes
3989 			 */
3990 			return (BUS_PROBE_DEFAULT);
3991 		}
3992 		ent++;
3993 	}
3994 	return (ENXIO);
3995 }
3996 
3997 int
3998 iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp)
3999 {
4000 	int err, rid, msix, msix_bar;
4001 	if_ctx_t ctx;
4002 	if_t ifp;
4003 	if_softc_ctx_t scctx;
4004 	int i;
4005 	uint16_t main_txq;
4006 	uint16_t main_rxq;
4007 
4008 
4009 	ctx = malloc(sizeof(* ctx), M_IFLIB, M_WAITOK|M_ZERO);
4010 
4011 	if (sc == NULL) {
4012 		sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4013 		device_set_softc(dev, ctx);
4014 		ctx->ifc_flags |= IFC_SC_ALLOCATED;
4015 	}
4016 
4017 	ctx->ifc_sctx = sctx;
4018 	ctx->ifc_dev = dev;
4019 	ctx->ifc_softc = sc;
4020 
4021 	if ((err = iflib_register(ctx)) != 0) {
4022 		device_printf(dev, "iflib_register failed %d\n", err);
4023 		return (err);
4024 	}
4025 	iflib_add_device_sysctl_pre(ctx);
4026 
4027 	scctx = &ctx->ifc_softc_ctx;
4028 	ifp = ctx->ifc_ifp;
4029 
4030 	/*
4031 	 * XXX sanity check that ntxd & nrxd are a power of 2
4032 	 */
4033 	if (ctx->ifc_sysctl_ntxqs != 0)
4034 		scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs;
4035 	if (ctx->ifc_sysctl_nrxqs != 0)
4036 		scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs;
4037 
4038 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4039 		if (ctx->ifc_sysctl_ntxds[i] != 0)
4040 			scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i];
4041 		else
4042 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4043 	}
4044 
4045 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4046 		if (ctx->ifc_sysctl_nrxds[i] != 0)
4047 			scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i];
4048 		else
4049 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4050 	}
4051 
4052 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4053 		if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) {
4054 			device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n",
4055 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]);
4056 			scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i];
4057 		}
4058 		if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) {
4059 			device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n",
4060 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]);
4061 			scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i];
4062 		}
4063 	}
4064 
4065 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4066 		if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) {
4067 			device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n",
4068 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]);
4069 			scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i];
4070 		}
4071 		if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) {
4072 			device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n",
4073 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]);
4074 			scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i];
4075 		}
4076 	}
4077 
4078 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4079 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4080 		return (err);
4081 	}
4082 	_iflib_pre_assert(scctx);
4083 	ctx->ifc_txrx = *scctx->isc_txrx;
4084 
4085 #ifdef INVARIANTS
4086 	MPASS(scctx->isc_capenable);
4087 	if (scctx->isc_capenable & IFCAP_TXCSUM)
4088 		MPASS(scctx->isc_tx_csum_flags);
4089 #endif
4090 
4091 	if_setcapabilities(ifp, scctx->isc_capenable);
4092 	if_setcapenable(ifp, scctx->isc_capenable);
4093 
4094 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4095 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4096 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4097 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4098 
4099 #ifdef ACPI_DMAR
4100 	if (dmar_get_dma_tag(device_get_parent(dev), dev) != NULL)
4101 		ctx->ifc_flags |= IFC_DMAR;
4102 #elif !(defined(__i386__) || defined(__amd64__))
4103 	/* set unconditionally for !x86 */
4104 	ctx->ifc_flags |= IFC_DMAR;
4105 #endif
4106 
4107 	msix_bar = scctx->isc_msix_bar;
4108 	main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4109 	main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4110 
4111 	/* XXX change for per-queue sizes */
4112 	device_printf(dev, "using %d tx descriptors and %d rx descriptors\n",
4113 		      scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4114 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4115 		if (!powerof2(scctx->isc_nrxd[i])) {
4116 			/* round down instead? */
4117 			device_printf(dev, "# rx descriptors must be a power of 2\n");
4118 			err = EINVAL;
4119 			goto fail;
4120 		}
4121 	}
4122 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4123 		if (!powerof2(scctx->isc_ntxd[i])) {
4124 			device_printf(dev,
4125 			    "# tx descriptors must be a power of 2");
4126 			err = EINVAL;
4127 			goto fail;
4128 		}
4129 	}
4130 
4131 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4132 	    MAX_SINGLE_PACKET_FRACTION)
4133 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4134 		    MAX_SINGLE_PACKET_FRACTION);
4135 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4136 	    MAX_SINGLE_PACKET_FRACTION)
4137 		scctx->isc_tx_tso_segments_max = max(1,
4138 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4139 
4140 	/*
4141 	 * Protect the stack against modern hardware
4142 	 */
4143 	if (scctx->isc_tx_tso_size_max > FREEBSD_TSO_SIZE_MAX)
4144 		scctx->isc_tx_tso_size_max = FREEBSD_TSO_SIZE_MAX;
4145 
4146 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4147 	ifp->if_hw_tsomaxsegcount = scctx->isc_tx_tso_segments_max;
4148 	ifp->if_hw_tsomax = scctx->isc_tx_tso_size_max;
4149 	ifp->if_hw_tsomaxsegsize = scctx->isc_tx_tso_segsize_max;
4150 	if (scctx->isc_rss_table_size == 0)
4151 		scctx->isc_rss_table_size = 64;
4152 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4153 
4154 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4155 	/* XXX format name */
4156 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx, -1, "admin");
4157 	/*
4158 	** Now setup MSI or MSI/X, should
4159 	** return us the number of supported
4160 	** vectors. (Will be 1 for MSI)
4161 	*/
4162 	if (sctx->isc_flags & IFLIB_SKIP_MSIX) {
4163 		msix = scctx->isc_vectors;
4164 	} else if (scctx->isc_msix_bar != 0)
4165 	       /*
4166 		* The simple fact that isc_msix_bar is not 0 does not mean we
4167 		* we have a good value there that is known to work.
4168 		*/
4169 		msix = iflib_msix_init(ctx);
4170 	else {
4171 		scctx->isc_vectors = 1;
4172 		scctx->isc_ntxqsets = 1;
4173 		scctx->isc_nrxqsets = 1;
4174 		scctx->isc_intr = IFLIB_INTR_LEGACY;
4175 		msix = 0;
4176 	}
4177 	/* Get memory for the station queues */
4178 	if ((err = iflib_queues_alloc(ctx))) {
4179 		device_printf(dev, "Unable to allocate queue memory\n");
4180 		goto fail;
4181 	}
4182 
4183 	if ((err = iflib_qset_structures_setup(ctx))) {
4184 		device_printf(dev, "qset structure setup failed %d\n", err);
4185 		goto fail_queues;
4186 	}
4187 
4188 	/*
4189 	 * Group taskqueues aren't properly set up until SMP is started,
4190 	 * so we disable interrupts until we can handle them post
4191 	 * SI_SUB_SMP.
4192 	 *
4193 	 * XXX: disabling interrupts doesn't actually work, at least for
4194 	 * the non-MSI case.  When they occur before SI_SUB_SMP completes,
4195 	 * we do null handling and depend on this not causing too large an
4196 	 * interrupt storm.
4197 	 */
4198 	IFDI_INTR_DISABLE(ctx);
4199 	if (msix > 1 && (err = IFDI_MSIX_INTR_ASSIGN(ctx, msix)) != 0) {
4200 		device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n", err);
4201 		goto fail_intr_free;
4202 	}
4203 	if (msix <= 1) {
4204 		rid = 0;
4205 		if (scctx->isc_intr == IFLIB_INTR_MSI) {
4206 			MPASS(msix == 1);
4207 			rid = 1;
4208 		}
4209 		if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) {
4210 			device_printf(dev, "iflib_legacy_setup failed %d\n", err);
4211 			goto fail_intr_free;
4212 		}
4213 	}
4214 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac);
4215 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4216 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4217 		goto fail_detach;
4218 	}
4219 	if ((err = iflib_netmap_attach(ctx))) {
4220 		device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err);
4221 		goto fail_detach;
4222 	}
4223 	*ctxp = ctx;
4224 
4225 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4226 	iflib_add_device_sysctl_post(ctx);
4227 	ctx->ifc_flags |= IFC_INIT_DONE;
4228 	return (0);
4229 fail_detach:
4230 	ether_ifdetach(ctx->ifc_ifp);
4231 fail_intr_free:
4232 	if (scctx->isc_intr == IFLIB_INTR_MSIX || scctx->isc_intr == IFLIB_INTR_MSI)
4233 		pci_release_msi(ctx->ifc_dev);
4234 fail_queues:
4235 	/* XXX free queues */
4236 fail:
4237 	IFDI_DETACH(ctx);
4238 	return (err);
4239 }
4240 
4241 int
4242 iflib_device_attach(device_t dev)
4243 {
4244 	if_ctx_t ctx;
4245 	if_shared_ctx_t sctx;
4246 
4247 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4248 		return (ENOTSUP);
4249 
4250 	pci_enable_busmaster(dev);
4251 
4252 	return (iflib_device_register(dev, NULL, sctx, &ctx));
4253 }
4254 
4255 int
4256 iflib_device_deregister(if_ctx_t ctx)
4257 {
4258 	if_t ifp = ctx->ifc_ifp;
4259 	iflib_txq_t txq;
4260 	iflib_rxq_t rxq;
4261 	device_t dev = ctx->ifc_dev;
4262 	int i, j;
4263 	struct taskqgroup *tqg;
4264 	iflib_fl_t fl;
4265 
4266 	/* Make sure VLANS are not using driver */
4267 	if (if_vlantrunkinuse(ifp)) {
4268 		device_printf(dev,"Vlan in use, detach first\n");
4269 		return (EBUSY);
4270 	}
4271 
4272 	CTX_LOCK(ctx);
4273 	ctx->ifc_in_detach = 1;
4274 	iflib_stop(ctx);
4275 	CTX_UNLOCK(ctx);
4276 
4277 	/* Unregister VLAN events */
4278 	if (ctx->ifc_vlan_attach_event != NULL)
4279 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
4280 	if (ctx->ifc_vlan_detach_event != NULL)
4281 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
4282 
4283 	iflib_netmap_detach(ifp);
4284 	ether_ifdetach(ifp);
4285 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
4286 	CTX_LOCK_DESTROY(ctx);
4287 	if (ctx->ifc_led_dev != NULL)
4288 		led_destroy(ctx->ifc_led_dev);
4289 	/* XXX drain any dependent tasks */
4290 	tqg = qgroup_if_io_tqg;
4291 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
4292 		callout_drain(&txq->ift_timer);
4293 		if (txq->ift_task.gt_uniq != NULL)
4294 			taskqgroup_detach(tqg, &txq->ift_task);
4295 	}
4296 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4297 		if (rxq->ifr_task.gt_uniq != NULL)
4298 			taskqgroup_detach(tqg, &rxq->ifr_task);
4299 
4300 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
4301 			free(fl->ifl_rx_bitmap, M_IFLIB);
4302 
4303 	}
4304 	tqg = qgroup_if_config_tqg;
4305 	if (ctx->ifc_admin_task.gt_uniq != NULL)
4306 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
4307 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
4308 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
4309 
4310 	IFDI_DETACH(ctx);
4311 	device_set_softc(ctx->ifc_dev, NULL);
4312 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) {
4313 		pci_release_msi(dev);
4314 	}
4315 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) {
4316 		iflib_irq_free(ctx, &ctx->ifc_legacy_irq);
4317 	}
4318 	if (ctx->ifc_msix_mem != NULL) {
4319 		bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY,
4320 			ctx->ifc_softc_ctx.isc_msix_bar, ctx->ifc_msix_mem);
4321 		ctx->ifc_msix_mem = NULL;
4322 	}
4323 
4324 	bus_generic_detach(dev);
4325 	if_free(ifp);
4326 
4327 	iflib_tx_structures_free(ctx);
4328 	iflib_rx_structures_free(ctx);
4329 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4330 		free(ctx->ifc_softc, M_IFLIB);
4331 	free(ctx, M_IFLIB);
4332 	return (0);
4333 }
4334 
4335 
4336 int
4337 iflib_device_detach(device_t dev)
4338 {
4339 	if_ctx_t ctx = device_get_softc(dev);
4340 
4341 	return (iflib_device_deregister(ctx));
4342 }
4343 
4344 int
4345 iflib_device_suspend(device_t dev)
4346 {
4347 	if_ctx_t ctx = device_get_softc(dev);
4348 
4349 	CTX_LOCK(ctx);
4350 	IFDI_SUSPEND(ctx);
4351 	CTX_UNLOCK(ctx);
4352 
4353 	return bus_generic_suspend(dev);
4354 }
4355 int
4356 iflib_device_shutdown(device_t dev)
4357 {
4358 	if_ctx_t ctx = device_get_softc(dev);
4359 
4360 	CTX_LOCK(ctx);
4361 	IFDI_SHUTDOWN(ctx);
4362 	CTX_UNLOCK(ctx);
4363 
4364 	return bus_generic_suspend(dev);
4365 }
4366 
4367 
4368 int
4369 iflib_device_resume(device_t dev)
4370 {
4371 	if_ctx_t ctx = device_get_softc(dev);
4372 	iflib_txq_t txq = ctx->ifc_txqs;
4373 
4374 	CTX_LOCK(ctx);
4375 	IFDI_RESUME(ctx);
4376 	iflib_init_locked(ctx);
4377 	CTX_UNLOCK(ctx);
4378 	for (int i = 0; i < NTXQSETS(ctx); i++, txq++)
4379 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
4380 
4381 	return (bus_generic_resume(dev));
4382 }
4383 
4384 int
4385 iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params)
4386 {
4387 	int error;
4388 	if_ctx_t ctx = device_get_softc(dev);
4389 
4390 	CTX_LOCK(ctx);
4391 	error = IFDI_IOV_INIT(ctx, num_vfs, params);
4392 	CTX_UNLOCK(ctx);
4393 
4394 	return (error);
4395 }
4396 
4397 void
4398 iflib_device_iov_uninit(device_t dev)
4399 {
4400 	if_ctx_t ctx = device_get_softc(dev);
4401 
4402 	CTX_LOCK(ctx);
4403 	IFDI_IOV_UNINIT(ctx);
4404 	CTX_UNLOCK(ctx);
4405 }
4406 
4407 int
4408 iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params)
4409 {
4410 	int error;
4411 	if_ctx_t ctx = device_get_softc(dev);
4412 
4413 	CTX_LOCK(ctx);
4414 	error = IFDI_IOV_VF_ADD(ctx, vfnum, params);
4415 	CTX_UNLOCK(ctx);
4416 
4417 	return (error);
4418 }
4419 
4420 /*********************************************************************
4421  *
4422  *  MODULE FUNCTION DEFINITIONS
4423  *
4424  **********************************************************************/
4425 
4426 /*
4427  * - Start a fast taskqueue thread for each core
4428  * - Start a taskqueue for control operations
4429  */
4430 static int
4431 iflib_module_init(void)
4432 {
4433 	return (0);
4434 }
4435 
4436 static int
4437 iflib_module_event_handler(module_t mod, int what, void *arg)
4438 {
4439 	int err;
4440 
4441 	switch (what) {
4442 	case MOD_LOAD:
4443 		if ((err = iflib_module_init()) != 0)
4444 			return (err);
4445 		break;
4446 	case MOD_UNLOAD:
4447 		return (EBUSY);
4448 	default:
4449 		return (EOPNOTSUPP);
4450 	}
4451 
4452 	return (0);
4453 }
4454 
4455 /*********************************************************************
4456  *
4457  *  PUBLIC FUNCTION DEFINITIONS
4458  *     ordered as in iflib.h
4459  *
4460  **********************************************************************/
4461 
4462 
4463 static void
4464 _iflib_assert(if_shared_ctx_t sctx)
4465 {
4466 	MPASS(sctx->isc_tx_maxsize);
4467 	MPASS(sctx->isc_tx_maxsegsize);
4468 
4469 	MPASS(sctx->isc_rx_maxsize);
4470 	MPASS(sctx->isc_rx_nsegments);
4471 	MPASS(sctx->isc_rx_maxsegsize);
4472 
4473 	MPASS(sctx->isc_nrxd_min[0]);
4474 	MPASS(sctx->isc_nrxd_max[0]);
4475 	MPASS(sctx->isc_nrxd_default[0]);
4476 	MPASS(sctx->isc_ntxd_min[0]);
4477 	MPASS(sctx->isc_ntxd_max[0]);
4478 	MPASS(sctx->isc_ntxd_default[0]);
4479 }
4480 
4481 static void
4482 _iflib_pre_assert(if_softc_ctx_t scctx)
4483 {
4484 
4485 	MPASS(scctx->isc_txrx->ift_txd_encap);
4486 	MPASS(scctx->isc_txrx->ift_txd_flush);
4487 	MPASS(scctx->isc_txrx->ift_txd_credits_update);
4488 	MPASS(scctx->isc_txrx->ift_rxd_available);
4489 	MPASS(scctx->isc_txrx->ift_rxd_pkt_get);
4490 	MPASS(scctx->isc_txrx->ift_rxd_refill);
4491 	MPASS(scctx->isc_txrx->ift_rxd_flush);
4492 }
4493 
4494 static int
4495 iflib_register(if_ctx_t ctx)
4496 {
4497 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4498 	driver_t *driver = sctx->isc_driver;
4499 	device_t dev = ctx->ifc_dev;
4500 	if_t ifp;
4501 
4502 	_iflib_assert(sctx);
4503 
4504 	CTX_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev));
4505 
4506 	ifp = ctx->ifc_ifp = if_gethandle(IFT_ETHER);
4507 	if (ifp == NULL) {
4508 		device_printf(dev, "can not allocate ifnet structure\n");
4509 		return (ENOMEM);
4510 	}
4511 
4512 	/*
4513 	 * Initialize our context's device specific methods
4514 	 */
4515 	kobj_init((kobj_t) ctx, (kobj_class_t) driver);
4516 	kobj_class_compile((kobj_class_t) driver);
4517 	driver->refs++;
4518 
4519 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
4520 	if_setsoftc(ifp, ctx);
4521 	if_setdev(ifp, dev);
4522 	if_setinitfn(ifp, iflib_if_init);
4523 	if_setioctlfn(ifp, iflib_if_ioctl);
4524 	if_settransmitfn(ifp, iflib_if_transmit);
4525 	if_setqflushfn(ifp, iflib_if_qflush);
4526 	if_setflags(ifp, IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST);
4527 
4528 	ctx->ifc_vlan_attach_event =
4529 		EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx,
4530 							  EVENTHANDLER_PRI_FIRST);
4531 	ctx->ifc_vlan_detach_event =
4532 		EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx,
4533 							  EVENTHANDLER_PRI_FIRST);
4534 
4535 	ifmedia_init(&ctx->ifc_media, IFM_IMASK,
4536 					 iflib_media_change, iflib_media_status);
4537 
4538 	return (0);
4539 }
4540 
4541 
4542 static int
4543 iflib_queues_alloc(if_ctx_t ctx)
4544 {
4545 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4546 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4547 	device_t dev = ctx->ifc_dev;
4548 	int nrxqsets = scctx->isc_nrxqsets;
4549 	int ntxqsets = scctx->isc_ntxqsets;
4550 	iflib_txq_t txq;
4551 	iflib_rxq_t rxq;
4552 	iflib_fl_t fl = NULL;
4553 	int i, j, cpu, err, txconf, rxconf;
4554 	iflib_dma_info_t ifdip;
4555 	uint32_t *rxqsizes = scctx->isc_rxqsizes;
4556 	uint32_t *txqsizes = scctx->isc_txqsizes;
4557 	uint8_t nrxqs = sctx->isc_nrxqs;
4558 	uint8_t ntxqs = sctx->isc_ntxqs;
4559 	int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1;
4560 	caddr_t *vaddrs;
4561 	uint64_t *paddrs;
4562 	struct ifmp_ring **brscp;
4563 
4564 	KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1"));
4565 	KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1"));
4566 
4567 	brscp = NULL;
4568 	txq = NULL;
4569 	rxq = NULL;
4570 
4571 /* Allocate the TX ring struct memory */
4572 	if (!(txq =
4573 	    (iflib_txq_t) malloc(sizeof(struct iflib_txq) *
4574 	    ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
4575 		device_printf(dev, "Unable to allocate TX ring memory\n");
4576 		err = ENOMEM;
4577 		goto fail;
4578 	}
4579 
4580 	/* Now allocate the RX */
4581 	if (!(rxq =
4582 	    (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) *
4583 	    nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
4584 		device_printf(dev, "Unable to allocate RX ring memory\n");
4585 		err = ENOMEM;
4586 		goto rx_fail;
4587 	}
4588 
4589 	ctx->ifc_txqs = txq;
4590 	ctx->ifc_rxqs = rxq;
4591 
4592 	/*
4593 	 * XXX handle allocation failure
4594 	 */
4595 	for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) {
4596 		/* Set up some basics */
4597 
4598 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs, M_IFLIB, M_WAITOK|M_ZERO)) == NULL) {
4599 			device_printf(dev, "failed to allocate iflib_dma_info\n");
4600 			err = ENOMEM;
4601 			goto err_tx_desc;
4602 		}
4603 		txq->ift_ifdi = ifdip;
4604 		for (j = 0; j < ntxqs; j++, ifdip++) {
4605 			if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, BUS_DMA_NOWAIT)) {
4606 				device_printf(dev, "Unable to allocate Descriptor memory\n");
4607 				err = ENOMEM;
4608 				goto err_tx_desc;
4609 			}
4610 			txq->ift_txd_size[j] = scctx->isc_txd_size[j];
4611 			bzero((void *)ifdip->idi_vaddr, txqsizes[j]);
4612 		}
4613 		txq->ift_ctx = ctx;
4614 		txq->ift_id = i;
4615 		if (sctx->isc_flags & IFLIB_HAS_TXCQ) {
4616 			txq->ift_br_offset = 1;
4617 		} else {
4618 			txq->ift_br_offset = 0;
4619 		}
4620 		/* XXX fix this */
4621 		txq->ift_timer.c_cpu = cpu;
4622 
4623 		if (iflib_txsd_alloc(txq)) {
4624 			device_printf(dev, "Critical Failure setting up TX buffers\n");
4625 			err = ENOMEM;
4626 			goto err_tx_desc;
4627 		}
4628 
4629 		/* Initialize the TX lock */
4630 		snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:tx(%d):callout",
4631 		    device_get_nameunit(dev), txq->ift_id);
4632 		mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF);
4633 		callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0);
4634 
4635 		snprintf(txq->ift_db_mtx_name, MTX_NAME_LEN, "%s:tx(%d):db",
4636 			 device_get_nameunit(dev), txq->ift_id);
4637 
4638 		err = ifmp_ring_alloc(&txq->ift_br, 2048, txq, iflib_txq_drain,
4639 				      iflib_txq_can_drain, M_IFLIB, M_WAITOK);
4640 		if (err) {
4641 			/* XXX free any allocated rings */
4642 			device_printf(dev, "Unable to allocate buf_ring\n");
4643 			goto err_tx_desc;
4644 		}
4645 	}
4646 
4647 	for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) {
4648 		/* Set up some basics */
4649 
4650 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs, M_IFLIB, M_WAITOK|M_ZERO)) == NULL) {
4651 			device_printf(dev, "failed to allocate iflib_dma_info\n");
4652 			err = ENOMEM;
4653 			goto err_tx_desc;
4654 		}
4655 
4656 		rxq->ifr_ifdi = ifdip;
4657 		/* XXX this needs to be changed if #rx queues != #tx queues */
4658 		rxq->ifr_ntxqirq = 1;
4659 		rxq->ifr_txqid[0] = i;
4660 		for (j = 0; j < nrxqs; j++, ifdip++) {
4661 			if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, BUS_DMA_NOWAIT)) {
4662 				device_printf(dev, "Unable to allocate Descriptor memory\n");
4663 				err = ENOMEM;
4664 				goto err_tx_desc;
4665 			}
4666 			bzero((void *)ifdip->idi_vaddr, rxqsizes[j]);
4667 		}
4668 		rxq->ifr_ctx = ctx;
4669 		rxq->ifr_id = i;
4670 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
4671 			rxq->ifr_fl_offset = 1;
4672 		} else {
4673 			rxq->ifr_fl_offset = 0;
4674 		}
4675 		rxq->ifr_nfl = nfree_lists;
4676 		if (!(fl =
4677 			  (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) {
4678 			device_printf(dev, "Unable to allocate free list memory\n");
4679 			err = ENOMEM;
4680 			goto err_tx_desc;
4681 		}
4682 		rxq->ifr_fl = fl;
4683 		for (j = 0; j < nfree_lists; j++) {
4684 			fl[j].ifl_rxq = rxq;
4685 			fl[j].ifl_id = j;
4686 			fl[j].ifl_ifdi = &rxq->ifr_ifdi[j + rxq->ifr_fl_offset];
4687 			fl[j].ifl_rxd_size = scctx->isc_rxd_size[j];
4688 		}
4689         /* Allocate receive buffers for the ring*/
4690 		if (iflib_rxsd_alloc(rxq)) {
4691 			device_printf(dev,
4692 			    "Critical Failure setting up receive buffers\n");
4693 			err = ENOMEM;
4694 			goto err_rx_desc;
4695 		}
4696 
4697 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
4698 			fl->ifl_rx_bitmap = bit_alloc(fl->ifl_size, M_IFLIB, M_WAITOK|M_ZERO);
4699 	}
4700 
4701 	/* TXQs */
4702 	vaddrs = malloc(sizeof(caddr_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
4703 	paddrs = malloc(sizeof(uint64_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
4704 	for (i = 0; i < ntxqsets; i++) {
4705 		iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi;
4706 
4707 		for (j = 0; j < ntxqs; j++, di++) {
4708 			vaddrs[i*ntxqs + j] = di->idi_vaddr;
4709 			paddrs[i*ntxqs + j] = di->idi_paddr;
4710 		}
4711 	}
4712 	if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) {
4713 		device_printf(ctx->ifc_dev, "device queue allocation failed\n");
4714 		iflib_tx_structures_free(ctx);
4715 		free(vaddrs, M_IFLIB);
4716 		free(paddrs, M_IFLIB);
4717 		goto err_rx_desc;
4718 	}
4719 	free(vaddrs, M_IFLIB);
4720 	free(paddrs, M_IFLIB);
4721 
4722 	/* RXQs */
4723 	vaddrs = malloc(sizeof(caddr_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
4724 	paddrs = malloc(sizeof(uint64_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
4725 	for (i = 0; i < nrxqsets; i++) {
4726 		iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi;
4727 
4728 		for (j = 0; j < nrxqs; j++, di++) {
4729 			vaddrs[i*nrxqs + j] = di->idi_vaddr;
4730 			paddrs[i*nrxqs + j] = di->idi_paddr;
4731 		}
4732 	}
4733 	if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) {
4734 		device_printf(ctx->ifc_dev, "device queue allocation failed\n");
4735 		iflib_tx_structures_free(ctx);
4736 		free(vaddrs, M_IFLIB);
4737 		free(paddrs, M_IFLIB);
4738 		goto err_rx_desc;
4739 	}
4740 	free(vaddrs, M_IFLIB);
4741 	free(paddrs, M_IFLIB);
4742 
4743 	return (0);
4744 
4745 /* XXX handle allocation failure changes */
4746 err_rx_desc:
4747 err_tx_desc:
4748 	if (ctx->ifc_rxqs != NULL)
4749 		free(ctx->ifc_rxqs, M_IFLIB);
4750 	ctx->ifc_rxqs = NULL;
4751 	if (ctx->ifc_txqs != NULL)
4752 		free(ctx->ifc_txqs, M_IFLIB);
4753 	ctx->ifc_txqs = NULL;
4754 rx_fail:
4755 	if (brscp != NULL)
4756 		free(brscp, M_IFLIB);
4757 	if (rxq != NULL)
4758 		free(rxq, M_IFLIB);
4759 	if (txq != NULL)
4760 		free(txq, M_IFLIB);
4761 fail:
4762 	return (err);
4763 }
4764 
4765 static int
4766 iflib_tx_structures_setup(if_ctx_t ctx)
4767 {
4768 	iflib_txq_t txq = ctx->ifc_txqs;
4769 	int i;
4770 
4771 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
4772 		iflib_txq_setup(txq);
4773 
4774 	return (0);
4775 }
4776 
4777 static void
4778 iflib_tx_structures_free(if_ctx_t ctx)
4779 {
4780 	iflib_txq_t txq = ctx->ifc_txqs;
4781 	int i, j;
4782 
4783 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
4784 		iflib_txq_destroy(txq);
4785 		for (j = 0; j < ctx->ifc_nhwtxqs; j++)
4786 			iflib_dma_free(&txq->ift_ifdi[j]);
4787 	}
4788 	free(ctx->ifc_txqs, M_IFLIB);
4789 	ctx->ifc_txqs = NULL;
4790 	IFDI_QUEUES_FREE(ctx);
4791 }
4792 
4793 /*********************************************************************
4794  *
4795  *  Initialize all receive rings.
4796  *
4797  **********************************************************************/
4798 static int
4799 iflib_rx_structures_setup(if_ctx_t ctx)
4800 {
4801 	iflib_rxq_t rxq = ctx->ifc_rxqs;
4802 	int q;
4803 #if defined(INET6) || defined(INET)
4804 	int i, err;
4805 #endif
4806 
4807 	for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) {
4808 #if defined(INET6) || defined(INET)
4809 		tcp_lro_free(&rxq->ifr_lc);
4810 		if ((err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp,
4811 		    TCP_LRO_ENTRIES, min(1024,
4812 		    ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset]))) != 0) {
4813 			device_printf(ctx->ifc_dev, "LRO Initialization failed!\n");
4814 			goto fail;
4815 		}
4816 		rxq->ifr_lro_enabled = TRUE;
4817 #endif
4818 		IFDI_RXQ_SETUP(ctx, rxq->ifr_id);
4819 	}
4820 	return (0);
4821 #if defined(INET6) || defined(INET)
4822 fail:
4823 	/*
4824 	 * Free RX software descriptors allocated so far, we will only handle
4825 	 * the rings that completed, the failing case will have
4826 	 * cleaned up for itself. 'q' failed, so its the terminus.
4827 	 */
4828 	rxq = ctx->ifc_rxqs;
4829 	for (i = 0; i < q; ++i, rxq++) {
4830 		iflib_rx_sds_free(rxq);
4831 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
4832 	}
4833 	return (err);
4834 #endif
4835 }
4836 
4837 /*********************************************************************
4838  *
4839  *  Free all receive rings.
4840  *
4841  **********************************************************************/
4842 static void
4843 iflib_rx_structures_free(if_ctx_t ctx)
4844 {
4845 	iflib_rxq_t rxq = ctx->ifc_rxqs;
4846 
4847 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) {
4848 		iflib_rx_sds_free(rxq);
4849 	}
4850 }
4851 
4852 static int
4853 iflib_qset_structures_setup(if_ctx_t ctx)
4854 {
4855 	int err;
4856 
4857 	if ((err = iflib_tx_structures_setup(ctx)) != 0)
4858 		return (err);
4859 
4860 	if ((err = iflib_rx_structures_setup(ctx)) != 0) {
4861 		device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err);
4862 		iflib_tx_structures_free(ctx);
4863 		iflib_rx_structures_free(ctx);
4864 	}
4865 	return (err);
4866 }
4867 
4868 int
4869 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
4870 				driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, char *name)
4871 {
4872 
4873 	return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name));
4874 }
4875 
4876 static int
4877 find_nth(if_ctx_t ctx, cpuset_t *cpus, int qid)
4878 {
4879 	int i, cpuid, eqid, count;
4880 
4881 	CPU_COPY(&ctx->ifc_cpus, cpus);
4882 	count = CPU_COUNT(&ctx->ifc_cpus);
4883 	eqid = qid % count;
4884 	/* clear up to the qid'th bit */
4885 	for (i = 0; i < eqid; i++) {
4886 		cpuid = CPU_FFS(cpus);
4887 		MPASS(cpuid != 0);
4888 		CPU_CLR(cpuid-1, cpus);
4889 	}
4890 	cpuid = CPU_FFS(cpus);
4891 	MPASS(cpuid != 0);
4892 	return (cpuid-1);
4893 }
4894 
4895 int
4896 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid,
4897 						iflib_intr_type_t type, driver_filter_t *filter,
4898 						void *filter_arg, int qid, char *name)
4899 {
4900 	struct grouptask *gtask;
4901 	struct taskqgroup *tqg;
4902 	iflib_filter_info_t info;
4903 	cpuset_t cpus;
4904 	gtask_fn_t *fn;
4905 	int tqrid, err, cpuid;
4906 	driver_filter_t *intr_fast;
4907 	void *q;
4908 
4909 	info = &ctx->ifc_filter_info;
4910 	tqrid = rid;
4911 
4912 	switch (type) {
4913 	/* XXX merge tx/rx for netmap? */
4914 	case IFLIB_INTR_TX:
4915 		q = &ctx->ifc_txqs[qid];
4916 		info = &ctx->ifc_txqs[qid].ift_filter_info;
4917 		gtask = &ctx->ifc_txqs[qid].ift_task;
4918 		tqg = qgroup_if_io_tqg;
4919 		fn = _task_fn_tx;
4920 		intr_fast = iflib_fast_intr;
4921 		GROUPTASK_INIT(gtask, 0, fn, q);
4922 		break;
4923 	case IFLIB_INTR_RX:
4924 		q = &ctx->ifc_rxqs[qid];
4925 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
4926 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
4927 		tqg = qgroup_if_io_tqg;
4928 		fn = _task_fn_rx;
4929 		intr_fast = iflib_fast_intr;
4930 		GROUPTASK_INIT(gtask, 0, fn, q);
4931 		break;
4932 	case IFLIB_INTR_RXTX:
4933 		q = &ctx->ifc_rxqs[qid];
4934 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
4935 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
4936 		tqg = qgroup_if_io_tqg;
4937 		fn = _task_fn_rx;
4938 		intr_fast = iflib_fast_intr_rxtx;
4939 		GROUPTASK_INIT(gtask, 0, fn, q);
4940 		break;
4941 	case IFLIB_INTR_ADMIN:
4942 		q = ctx;
4943 		tqrid = -1;
4944 		info = &ctx->ifc_filter_info;
4945 		gtask = &ctx->ifc_admin_task;
4946 		tqg = qgroup_if_config_tqg;
4947 		fn = _task_fn_admin;
4948 		intr_fast = iflib_fast_intr_ctx;
4949 		break;
4950 	default:
4951 		panic("unknown net intr type");
4952 	}
4953 
4954 	info->ifi_filter = filter;
4955 	info->ifi_filter_arg = filter_arg;
4956 	info->ifi_task = gtask;
4957 	info->ifi_ctx = q;
4958 
4959 	err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info,  name);
4960 	if (err != 0) {
4961 		device_printf(ctx->ifc_dev, "_iflib_irq_alloc failed %d\n", err);
4962 		return (err);
4963 	}
4964 	if (type == IFLIB_INTR_ADMIN)
4965 		return (0);
4966 
4967 	if (tqrid != -1) {
4968 		cpuid = find_nth(ctx, &cpus, qid);
4969 		taskqgroup_attach_cpu(tqg, gtask, q, cpuid, irq->ii_rid, name);
4970 	} else {
4971 		taskqgroup_attach(tqg, gtask, q, tqrid, name);
4972 	}
4973 
4974 	return (0);
4975 }
4976 
4977 void
4978 iflib_softirq_alloc_generic(if_ctx_t ctx, int rid, iflib_intr_type_t type,  void *arg, int qid, char *name)
4979 {
4980 	struct grouptask *gtask;
4981 	struct taskqgroup *tqg;
4982 	gtask_fn_t *fn;
4983 	void *q;
4984 
4985 	switch (type) {
4986 	case IFLIB_INTR_TX:
4987 		q = &ctx->ifc_txqs[qid];
4988 		gtask = &ctx->ifc_txqs[qid].ift_task;
4989 		tqg = qgroup_if_io_tqg;
4990 		fn = _task_fn_tx;
4991 		break;
4992 	case IFLIB_INTR_RX:
4993 		q = &ctx->ifc_rxqs[qid];
4994 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
4995 		tqg = qgroup_if_io_tqg;
4996 		fn = _task_fn_rx;
4997 		break;
4998 	case IFLIB_INTR_IOV:
4999 		q = ctx;
5000 		gtask = &ctx->ifc_vflr_task;
5001 		tqg = qgroup_if_config_tqg;
5002 		rid = -1;
5003 		fn = _task_fn_iov;
5004 		break;
5005 	default:
5006 		panic("unknown net intr type");
5007 	}
5008 	GROUPTASK_INIT(gtask, 0, fn, q);
5009 	taskqgroup_attach(tqg, gtask, q, rid, name);
5010 }
5011 
5012 void
5013 iflib_irq_free(if_ctx_t ctx, if_irq_t irq)
5014 {
5015 	if (irq->ii_tag)
5016 		bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag);
5017 
5018 	if (irq->ii_res)
5019 		bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ, irq->ii_rid, irq->ii_res);
5020 }
5021 
5022 static int
5023 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, char *name)
5024 {
5025 	iflib_txq_t txq = ctx->ifc_txqs;
5026 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5027 	if_irq_t irq = &ctx->ifc_legacy_irq;
5028 	iflib_filter_info_t info;
5029 	struct grouptask *gtask;
5030 	struct taskqgroup *tqg;
5031 	gtask_fn_t *fn;
5032 	int tqrid;
5033 	void *q;
5034 	int err;
5035 
5036 	q = &ctx->ifc_rxqs[0];
5037 	info = &rxq[0].ifr_filter_info;
5038 	gtask = &rxq[0].ifr_task;
5039 	tqg = qgroup_if_io_tqg;
5040 	tqrid = irq->ii_rid = *rid;
5041 	fn = _task_fn_rx;
5042 
5043 	ctx->ifc_flags |= IFC_LEGACY;
5044 	info->ifi_filter = filter;
5045 	info->ifi_filter_arg = filter_arg;
5046 	info->ifi_task = gtask;
5047 	info->ifi_ctx = ctx;
5048 
5049 	/* We allocate a single interrupt resource */
5050 	if ((err = _iflib_irq_alloc(ctx, irq, tqrid, iflib_fast_intr_ctx, NULL, info, name)) != 0)
5051 		return (err);
5052 	GROUPTASK_INIT(gtask, 0, fn, q);
5053 	taskqgroup_attach(tqg, gtask, q, tqrid, name);
5054 
5055 	GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq);
5056 	taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, tqrid, "tx");
5057 	return (0);
5058 }
5059 
5060 void
5061 iflib_led_create(if_ctx_t ctx)
5062 {
5063 
5064 	ctx->ifc_led_dev = led_create(iflib_led_func, ctx,
5065 	    device_get_nameunit(ctx->ifc_dev));
5066 }
5067 
5068 void
5069 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid)
5070 {
5071 
5072 	GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
5073 }
5074 
5075 void
5076 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid)
5077 {
5078 
5079 	GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task);
5080 }
5081 
5082 void
5083 iflib_admin_intr_deferred(if_ctx_t ctx)
5084 {
5085 #ifdef INVARIANTS
5086 	struct grouptask *gtask;
5087 
5088 	gtask = &ctx->ifc_admin_task;
5089 	MPASS(gtask->gt_taskqueue != NULL);
5090 #endif
5091 
5092 	GROUPTASK_ENQUEUE(&ctx->ifc_admin_task);
5093 }
5094 
5095 void
5096 iflib_iov_intr_deferred(if_ctx_t ctx)
5097 {
5098 
5099 	GROUPTASK_ENQUEUE(&ctx->ifc_vflr_task);
5100 }
5101 
5102 void
5103 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, char *name)
5104 {
5105 
5106 	taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, -1, name);
5107 }
5108 
5109 void
5110 iflib_config_gtask_init(if_ctx_t ctx, struct grouptask *gtask, gtask_fn_t *fn,
5111 	char *name)
5112 {
5113 
5114 	GROUPTASK_INIT(gtask, 0, fn, ctx);
5115 	taskqgroup_attach(qgroup_if_config_tqg, gtask, gtask, -1, name);
5116 }
5117 
5118 void
5119 iflib_config_gtask_deinit(struct grouptask *gtask)
5120 {
5121 
5122 	taskqgroup_detach(qgroup_if_config_tqg, gtask);
5123 }
5124 
5125 void
5126 iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate)
5127 {
5128 	if_t ifp = ctx->ifc_ifp;
5129 	iflib_txq_t txq = ctx->ifc_txqs;
5130 
5131 	if_setbaudrate(ifp, baudrate);
5132 	if (baudrate >= IF_Gbps(10))
5133 		ctx->ifc_flags |= IFC_PREFETCH;
5134 
5135 	/* If link down, disable watchdog */
5136 	if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) {
5137 		for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++)
5138 			txq->ift_qstatus = IFLIB_QUEUE_IDLE;
5139 	}
5140 	ctx->ifc_link_state = link_state;
5141 	if_link_state_change(ifp, link_state);
5142 }
5143 
5144 static int
5145 iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq)
5146 {
5147 	int credits;
5148 #ifdef INVARIANTS
5149 	int credits_pre = txq->ift_cidx_processed;
5150 #endif
5151 
5152 	if (ctx->isc_txd_credits_update == NULL)
5153 		return (0);
5154 
5155 	if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, true)) == 0)
5156 		return (0);
5157 
5158 	txq->ift_processed += credits;
5159 	txq->ift_cidx_processed += credits;
5160 
5161 	MPASS(credits_pre + credits == txq->ift_cidx_processed);
5162 	if (txq->ift_cidx_processed >= txq->ift_size)
5163 		txq->ift_cidx_processed -= txq->ift_size;
5164 	return (credits);
5165 }
5166 
5167 static int
5168 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget)
5169 {
5170 
5171 	return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx,
5172 	    budget));
5173 }
5174 
5175 void
5176 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name,
5177 	const char *description, if_int_delay_info_t info,
5178 	int offset, int value)
5179 {
5180 	info->iidi_ctx = ctx;
5181 	info->iidi_offset = offset;
5182 	info->iidi_value = value;
5183 	SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev),
5184 	    SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)),
5185 	    OID_AUTO, name, CTLTYPE_INT|CTLFLAG_RW,
5186 	    info, 0, iflib_sysctl_int_delay, "I", description);
5187 }
5188 
5189 struct mtx *
5190 iflib_ctx_lock_get(if_ctx_t ctx)
5191 {
5192 
5193 	return (&ctx->ifc_mtx);
5194 }
5195 
5196 static int
5197 iflib_msix_init(if_ctx_t ctx)
5198 {
5199 	device_t dev = ctx->ifc_dev;
5200 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5201 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5202 	int vectors, queues, rx_queues, tx_queues, queuemsgs, msgs;
5203 	int iflib_num_tx_queues, iflib_num_rx_queues;
5204 	int err, admincnt, bar;
5205 
5206 	iflib_num_tx_queues = scctx->isc_ntxqsets;
5207 	iflib_num_rx_queues = scctx->isc_nrxqsets;
5208 
5209 	device_printf(dev, "msix_init qsets capped at %d\n", iflib_num_tx_queues);
5210 
5211 	bar = ctx->ifc_softc_ctx.isc_msix_bar;
5212 	admincnt = sctx->isc_admin_intrcnt;
5213 	/* Override by tuneable */
5214 	if (scctx->isc_disable_msix)
5215 		goto msi;
5216 
5217 	/*
5218 	** When used in a virtualized environment
5219 	** PCI BUSMASTER capability may not be set
5220 	** so explicity set it here and rewrite
5221 	** the ENABLE in the MSIX control register
5222 	** at this point to cause the host to
5223 	** successfully initialize us.
5224 	*/
5225 	{
5226 		int msix_ctrl, rid;
5227 
5228  		pci_enable_busmaster(dev);
5229 		rid = 0;
5230 		if (pci_find_cap(dev, PCIY_MSIX, &rid) == 0 && rid != 0) {
5231 			rid += PCIR_MSIX_CTRL;
5232 			msix_ctrl = pci_read_config(dev, rid, 2);
5233 			msix_ctrl |= PCIM_MSIXCTRL_MSIX_ENABLE;
5234 			pci_write_config(dev, rid, msix_ctrl, 2);
5235 		} else {
5236 			device_printf(dev, "PCIY_MSIX capability not found; "
5237 			                   "or rid %d == 0.\n", rid);
5238 			goto msi;
5239 		}
5240 	}
5241 
5242 	/*
5243 	 * bar == -1 => "trust me I know what I'm doing"
5244 	 * Some drivers are for hardware that is so shoddily
5245 	 * documented that no one knows which bars are which
5246 	 * so the developer has to map all bars. This hack
5247 	 * allows shoddy garbage to use msix in this framework.
5248 	 */
5249 	if (bar != -1) {
5250 		ctx->ifc_msix_mem = bus_alloc_resource_any(dev,
5251 	            SYS_RES_MEMORY, &bar, RF_ACTIVE);
5252 		if (ctx->ifc_msix_mem == NULL) {
5253 			/* May not be enabled */
5254 			device_printf(dev, "Unable to map MSIX table \n");
5255 			goto msi;
5256 		}
5257 	}
5258 	/* First try MSI/X */
5259 	if ((msgs = pci_msix_count(dev)) == 0) { /* system has msix disabled */
5260 		device_printf(dev, "System has MSIX disabled \n");
5261 		bus_release_resource(dev, SYS_RES_MEMORY,
5262 		    bar, ctx->ifc_msix_mem);
5263 		ctx->ifc_msix_mem = NULL;
5264 		goto msi;
5265 	}
5266 #if IFLIB_DEBUG
5267 	/* use only 1 qset in debug mode */
5268 	queuemsgs = min(msgs - admincnt, 1);
5269 #else
5270 	queuemsgs = msgs - admincnt;
5271 #endif
5272 	if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) == 0) {
5273 #ifdef RSS
5274 		queues = imin(queuemsgs, rss_getnumbuckets());
5275 #else
5276 		queues = queuemsgs;
5277 #endif
5278 		queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues);
5279 		device_printf(dev, "pxm cpus: %d queue msgs: %d admincnt: %d\n",
5280 					  CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt);
5281 	} else {
5282 		device_printf(dev, "Unable to fetch CPU list\n");
5283 		/* Figure out a reasonable auto config value */
5284 		queues = min(queuemsgs, mp_ncpus);
5285 	}
5286 #ifdef  RSS
5287 	/* If we're doing RSS, clamp at the number of RSS buckets */
5288 	if (queues > rss_getnumbuckets())
5289 		queues = rss_getnumbuckets();
5290 #endif
5291 	if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt)
5292 		rx_queues = iflib_num_rx_queues;
5293 	else
5294 		rx_queues = queues;
5295 	/*
5296 	 * We want this to be all logical CPUs by default
5297 	 */
5298 	if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues)
5299 		tx_queues = iflib_num_tx_queues;
5300 	else
5301 		tx_queues = mp_ncpus;
5302 
5303 	if (ctx->ifc_sysctl_qs_eq_override == 0) {
5304 #ifdef INVARIANTS
5305 		if (tx_queues != rx_queues)
5306 			device_printf(dev, "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n",
5307 				      min(rx_queues, tx_queues), min(rx_queues, tx_queues));
5308 #endif
5309 		tx_queues = min(rx_queues, tx_queues);
5310 		rx_queues = min(rx_queues, tx_queues);
5311 	}
5312 
5313 	device_printf(dev, "using %d rx queues %d tx queues \n", rx_queues, tx_queues);
5314 
5315 	vectors = rx_queues + admincnt;
5316 	if ((err = pci_alloc_msix(dev, &vectors)) == 0) {
5317 		device_printf(dev,
5318 					  "Using MSIX interrupts with %d vectors\n", vectors);
5319 		scctx->isc_vectors = vectors;
5320 		scctx->isc_nrxqsets = rx_queues;
5321 		scctx->isc_ntxqsets = tx_queues;
5322 		scctx->isc_intr = IFLIB_INTR_MSIX;
5323 
5324 		return (vectors);
5325 	} else {
5326 		device_printf(dev, "failed to allocate %d msix vectors, err: %d - using MSI\n", vectors, err);
5327 	}
5328 msi:
5329 	vectors = pci_msi_count(dev);
5330 	scctx->isc_nrxqsets = 1;
5331 	scctx->isc_ntxqsets = 1;
5332 	scctx->isc_vectors = vectors;
5333 	if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) {
5334 		device_printf(dev,"Using an MSI interrupt\n");
5335 		scctx->isc_intr = IFLIB_INTR_MSI;
5336 	} else {
5337 		device_printf(dev,"Using a Legacy interrupt\n");
5338 		scctx->isc_intr = IFLIB_INTR_LEGACY;
5339 	}
5340 
5341 	return (vectors);
5342 }
5343 
5344 char * ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" };
5345 
5346 static int
5347 mp_ring_state_handler(SYSCTL_HANDLER_ARGS)
5348 {
5349 	int rc;
5350 	uint16_t *state = ((uint16_t *)oidp->oid_arg1);
5351 	struct sbuf *sb;
5352 	char *ring_state = "UNKNOWN";
5353 
5354 	/* XXX needed ? */
5355 	rc = sysctl_wire_old_buffer(req, 0);
5356 	MPASS(rc == 0);
5357 	if (rc != 0)
5358 		return (rc);
5359 	sb = sbuf_new_for_sysctl(NULL, NULL, 80, req);
5360 	MPASS(sb != NULL);
5361 	if (sb == NULL)
5362 		return (ENOMEM);
5363 	if (state[3] <= 3)
5364 		ring_state = ring_states[state[3]];
5365 
5366 	sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s",
5367 		    state[0], state[1], state[2], ring_state);
5368 	rc = sbuf_finish(sb);
5369 	sbuf_delete(sb);
5370         return(rc);
5371 }
5372 
5373 enum iflib_ndesc_handler {
5374 	IFLIB_NTXD_HANDLER,
5375 	IFLIB_NRXD_HANDLER,
5376 };
5377 
5378 static int
5379 mp_ndesc_handler(SYSCTL_HANDLER_ARGS)
5380 {
5381 	if_ctx_t ctx = (void *)arg1;
5382 	enum iflib_ndesc_handler type = arg2;
5383 	char buf[256] = {0};
5384 	qidx_t *ndesc;
5385 	char *p, *next;
5386 	int nqs, rc, i;
5387 
5388 	MPASS(type == IFLIB_NTXD_HANDLER || type == IFLIB_NRXD_HANDLER);
5389 
5390 	nqs = 8;
5391 	switch(type) {
5392 	case IFLIB_NTXD_HANDLER:
5393 		ndesc = ctx->ifc_sysctl_ntxds;
5394 		if (ctx->ifc_sctx)
5395 			nqs = ctx->ifc_sctx->isc_ntxqs;
5396 		break;
5397 	case IFLIB_NRXD_HANDLER:
5398 		ndesc = ctx->ifc_sysctl_nrxds;
5399 		if (ctx->ifc_sctx)
5400 			nqs = ctx->ifc_sctx->isc_nrxqs;
5401 		break;
5402 	}
5403 	if (nqs == 0)
5404 		nqs = 8;
5405 
5406 	for (i=0; i<8; i++) {
5407 		if (i >= nqs)
5408 			break;
5409 		if (i)
5410 			strcat(buf, ",");
5411 		sprintf(strchr(buf, 0), "%d", ndesc[i]);
5412 	}
5413 
5414 	rc = sysctl_handle_string(oidp, buf, sizeof(buf), req);
5415 	if (rc || req->newptr == NULL)
5416 		return rc;
5417 
5418 	for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p;
5419 	    i++, p = strsep(&next, " ,")) {
5420 		ndesc[i] = strtoul(p, NULL, 10);
5421 	}
5422 
5423 	return(rc);
5424 }
5425 
5426 #define NAME_BUFLEN 32
5427 static void
5428 iflib_add_device_sysctl_pre(if_ctx_t ctx)
5429 {
5430         device_t dev = iflib_get_dev(ctx);
5431 	struct sysctl_oid_list *child, *oid_list;
5432 	struct sysctl_ctx_list *ctx_list;
5433 	struct sysctl_oid *node;
5434 
5435 	ctx_list = device_get_sysctl_ctx(dev);
5436 	child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev));
5437 	ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, "iflib",
5438 						      CTLFLAG_RD, NULL, "IFLIB fields");
5439 	oid_list = SYSCTL_CHILDREN(node);
5440 
5441 	SYSCTL_ADD_STRING(ctx_list, oid_list, OID_AUTO, "driver_version",
5442 		       CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version, 0,
5443 		       "driver version");
5444 
5445 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs",
5446 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0,
5447 			"# of txqs to use, 0 => use default #");
5448 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs",
5449 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0,
5450 			"# of rxqs to use, 0 => use default #");
5451 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_qs_enable",
5452 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0,
5453                        "permit #txq != #rxq");
5454        SYSCTL_ADD_INT(ctx_list, oid_list, OID_AUTO, "disable_msix",
5455                       CTLFLAG_RWTUN, &ctx->ifc_softc_ctx.isc_disable_msix, 0,
5456                       "disable MSIX (default 0)");
5457 
5458 	/* XXX change for per-queue sizes */
5459 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_ntxds",
5460 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NTXD_HANDLER,
5461                        mp_ndesc_handler, "A",
5462                        "list of # of tx descriptors to use, 0 = use default #");
5463 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_nrxds",
5464 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NRXD_HANDLER,
5465                        mp_ndesc_handler, "A",
5466                        "list of # of rx descriptors to use, 0 = use default #");
5467 }
5468 
5469 static void
5470 iflib_add_device_sysctl_post(if_ctx_t ctx)
5471 {
5472 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5473 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5474         device_t dev = iflib_get_dev(ctx);
5475 	struct sysctl_oid_list *child;
5476 	struct sysctl_ctx_list *ctx_list;
5477 	iflib_fl_t fl;
5478 	iflib_txq_t txq;
5479 	iflib_rxq_t rxq;
5480 	int i, j;
5481 	char namebuf[NAME_BUFLEN];
5482 	char *qfmt;
5483 	struct sysctl_oid *queue_node, *fl_node, *node;
5484 	struct sysctl_oid_list *queue_list, *fl_list;
5485 	ctx_list = device_get_sysctl_ctx(dev);
5486 
5487 	node = ctx->ifc_sysctl_node;
5488 	child = SYSCTL_CHILDREN(node);
5489 
5490 	if (scctx->isc_ntxqsets > 100)
5491 		qfmt = "txq%03d";
5492 	else if (scctx->isc_ntxqsets > 10)
5493 		qfmt = "txq%02d";
5494 	else
5495 		qfmt = "txq%d";
5496 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
5497 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
5498 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
5499 					     CTLFLAG_RD, NULL, "Queue Name");
5500 		queue_list = SYSCTL_CHILDREN(queue_node);
5501 #if MEMORY_LOGGING
5502 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued",
5503 				CTLFLAG_RD,
5504 				&txq->ift_dequeued, "total mbufs freed");
5505 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued",
5506 				CTLFLAG_RD,
5507 				&txq->ift_enqueued, "total mbufs enqueued");
5508 #endif
5509 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag",
5510 				   CTLFLAG_RD,
5511 				   &txq->ift_mbuf_defrag, "# of times m_defrag was called");
5512 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "m_pullups",
5513 				   CTLFLAG_RD,
5514 				   &txq->ift_pullups, "# of times m_pullup was called");
5515 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag_failed",
5516 				   CTLFLAG_RD,
5517 				   &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed");
5518 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_desc_avail",
5519 				   CTLFLAG_RD,
5520 				   &txq->ift_no_desc_avail, "# of times no descriptors were available");
5521 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "tx_map_failed",
5522 				   CTLFLAG_RD,
5523 				   &txq->ift_map_failed, "# of times dma map failed");
5524 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txd_encap_efbig",
5525 				   CTLFLAG_RD,
5526 				   &txq->ift_txd_encap_efbig, "# of times txd_encap returned EFBIG");
5527 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_tx_dma_setup",
5528 				   CTLFLAG_RD,
5529 				   &txq->ift_no_tx_dma_setup, "# of times map failed for other than EFBIG");
5530 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx",
5531 				   CTLFLAG_RD,
5532 				   &txq->ift_pidx, 1, "Producer Index");
5533 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx",
5534 				   CTLFLAG_RD,
5535 				   &txq->ift_cidx, 1, "Consumer Index");
5536 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx_processed",
5537 				   CTLFLAG_RD,
5538 				   &txq->ift_cidx_processed, 1, "Consumer Index seen by credit update");
5539 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use",
5540 				   CTLFLAG_RD,
5541 				   &txq->ift_in_use, 1, "descriptors in use");
5542 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_processed",
5543 				   CTLFLAG_RD,
5544 				   &txq->ift_processed, "descriptors procesed for clean");
5545 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned",
5546 				   CTLFLAG_RD,
5547 				   &txq->ift_cleaned, "total cleaned");
5548 		SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state",
5549 				CTLTYPE_STRING | CTLFLAG_RD, __DEVOLATILE(uint64_t *, &txq->ift_br->state),
5550 				0, mp_ring_state_handler, "A", "soft ring state");
5551 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_enqueues",
5552 				       CTLFLAG_RD, &txq->ift_br->enqueues,
5553 				       "# of enqueues to the mp_ring for this queue");
5554 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_drops",
5555 				       CTLFLAG_RD, &txq->ift_br->drops,
5556 				       "# of drops in the mp_ring for this queue");
5557 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_starts",
5558 				       CTLFLAG_RD, &txq->ift_br->starts,
5559 				       "# of normal consumer starts in the mp_ring for this queue");
5560 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_stalls",
5561 				       CTLFLAG_RD, &txq->ift_br->stalls,
5562 					       "# of consumer stalls in the mp_ring for this queue");
5563 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_restarts",
5564 			       CTLFLAG_RD, &txq->ift_br->restarts,
5565 				       "# of consumer restarts in the mp_ring for this queue");
5566 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_abdications",
5567 				       CTLFLAG_RD, &txq->ift_br->abdications,
5568 				       "# of consumer abdications in the mp_ring for this queue");
5569 	}
5570 
5571 	if (scctx->isc_nrxqsets > 100)
5572 		qfmt = "rxq%03d";
5573 	else if (scctx->isc_nrxqsets > 10)
5574 		qfmt = "rxq%02d";
5575 	else
5576 		qfmt = "rxq%d";
5577 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
5578 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
5579 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
5580 					     CTLFLAG_RD, NULL, "Queue Name");
5581 		queue_list = SYSCTL_CHILDREN(queue_node);
5582 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
5583 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_pidx",
5584 				       CTLFLAG_RD,
5585 				       &rxq->ifr_cq_pidx, 1, "Producer Index");
5586 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_cidx",
5587 				       CTLFLAG_RD,
5588 				       &rxq->ifr_cq_cidx, 1, "Consumer Index");
5589 		}
5590 
5591 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
5592 			snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j);
5593 			fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list, OID_AUTO, namebuf,
5594 						     CTLFLAG_RD, NULL, "freelist Name");
5595 			fl_list = SYSCTL_CHILDREN(fl_node);
5596 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx",
5597 				       CTLFLAG_RD,
5598 				       &fl->ifl_pidx, 1, "Producer Index");
5599 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx",
5600 				       CTLFLAG_RD,
5601 				       &fl->ifl_cidx, 1, "Consumer Index");
5602 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits",
5603 				       CTLFLAG_RD,
5604 				       &fl->ifl_credits, 1, "credits available");
5605 #if MEMORY_LOGGING
5606 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_enqueued",
5607 					CTLFLAG_RD,
5608 					&fl->ifl_m_enqueued, "mbufs allocated");
5609 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_dequeued",
5610 					CTLFLAG_RD,
5611 					&fl->ifl_m_dequeued, "mbufs freed");
5612 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_enqueued",
5613 					CTLFLAG_RD,
5614 					&fl->ifl_cl_enqueued, "clusters allocated");
5615 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_dequeued",
5616 					CTLFLAG_RD,
5617 					&fl->ifl_cl_dequeued, "clusters freed");
5618 #endif
5619 
5620 		}
5621 	}
5622 
5623 }
5624 
5625 #ifndef __NO_STRICT_ALIGNMENT
5626 static struct mbuf *
5627 iflib_fixup_rx(struct mbuf *m)
5628 {
5629 	struct mbuf *n;
5630 
5631 	if (m->m_len <= (MCLBYTES - ETHER_HDR_LEN)) {
5632 		bcopy(m->m_data, m->m_data + ETHER_HDR_LEN, m->m_len);
5633 		m->m_data += ETHER_HDR_LEN;
5634 		n = m;
5635 	} else {
5636 		MGETHDR(n, M_NOWAIT, MT_DATA);
5637 		if (n == NULL) {
5638 			m_freem(m);
5639 			return (NULL);
5640 		}
5641 		bcopy(m->m_data, n->m_data, ETHER_HDR_LEN);
5642 		m->m_data += ETHER_HDR_LEN;
5643 		m->m_len -= ETHER_HDR_LEN;
5644 		n->m_len = ETHER_HDR_LEN;
5645 		M_MOVE_PKTHDR(n, m);
5646 		n->m_next = m;
5647 	}
5648 	return (n);
5649 }
5650 #endif
5651